mirror of
https://github.com/ParisNeo/lollms.git
synced 2024-12-20 05:08:00 +00:00
83 lines
191 KiB
Plaintext
83 lines
191 KiB
Plaintext
#0 A Mechanised Proof of G¨odel ’ s Incompleteness Theorems using Nominal Isabelle Lawrence C. Paulson Abstract An Isabelle/HOL formalisation of G¨odel ’ s two incompleteness theorems is presented . The work follows Swierczkowski ’ s detailed proof of the theorems using hered ´ - itarily finite ( HF ) set theory [ 32 ] . Avoiding the usual arithmetical encodings of syntax eliminates the necessity to formalise elementary number theory within an embedded logical calculus . The Isabelle formalisation uses two separate treatments of variable binding : the nominal package [ 34 ] is shown to scale to a development of this complexity , while de Bruijn indices [ 3 ] turn out to be ideal for coding syntax . Critical details of the Isabelle proof are described , in particular gaps and errors found in the literature . 1 Introduction This paper describes mechanised proofs of G¨odel ’ s incompleteness theorems [ 8 ] , including the first mechanised proof of the second incompleteness theorem . Very informally , these results can be stated as follows : Theorem 1 ( First Incompleteness Theorem ) If L is a consistent theory capable of formalising a sufficient amount of elementary mathematics , then there is a sentence δ such that neither δ nor ¬δ is a theorem of L , and moreover , δ is true.1 Theorem 2 ( Second Incompleteness Theorem ) If L is as above and Con ( L ) is a sentence stating that L is consistent , then Con ( L ) is not a theorem of L. Both of these will be presented formally below . Let us start to examine what these theorems actually assert . They concern a consistent formal system , say L , based on firstorder logic with some additional axioms : G¨odel chose Peano arithmetic ( PA ) [ 7 ] , but hereditarily finite ( HF ) set theory is an alternative [ 32 ] , used here . The first theorem states that any such axiomatic system must be incomplete , in the sense that some sentence can neither be proved nor disproved . The expedient of adding that sentence as an axiom merely creates a new axiomatic system , for which there is another undecidable sentence . The theorem can be strengthened to allow infinitely many additional axioms , Computer Laboratory , University of Cambridge , England E-mail : lp15 @ cl.cam.ac.uk 1 Meaning , δ ( which has no free variables ) is true in the standard model for L. 2 provided there is an effective procedure to recognise whether a given formula is an axiom or not . The second incompleteness theorem asserts that the consistency of L can not be proved in L itself . Even to state this theorem rigorously requires first defining the concept of provability in L ; the necessary series of definitions amounts to a computer program that occupies many pages . Although the same definitions are used to prove
|
||
#1 : G¨odel chose Peano arithmetic ( PA ) [ 7 ] , but hereditarily finite ( HF ) set theory is an alternative [ 32 ] , used here . The first theorem states that any such axiomatic system must be incomplete , in the sense that some sentence can neither be proved nor disproved . The expedient of adding that sentence as an axiom merely creates a new axiomatic system , for which there is another undecidable sentence . The theorem can be strengthened to allow infinitely many additional axioms , Computer Laboratory , University of Cambridge , England E-mail : lp15 @ cl.cam.ac.uk 1 Meaning , δ ( which has no free variables ) is true in the standard model for L. 2 provided there is an effective procedure to recognise whether a given formula is an axiom or not . The second incompleteness theorem asserts that the consistency of L can not be proved in L itself . Even to state this theorem rigorously requires first defining the concept of provability in L ; the necessary series of definitions amounts to a computer program that occupies many pages . Although the same definitions are used to prove the first incompleteness theorem , they are at least not needed to state that theorem . The original rationale for this project was a logician ’ s suggestion that the second incompleteness theorem had never been proved rigorously . Having completed this project , I sympathise with his view ; most published proofs contain substantial gaps and use cryptic notation . Both incompleteness theorems are widely misinterpreted , both in popular culture and even by some mathematicians . The first incompleteness theorem is often taken to imply that mathematics can not be formalised , when evidently it has been , this paper being one of numerous instances . It has also been used to assert that human intelligence can perceive truths ( in particular , the truth of δ , the undecidable sentence ) that no computer will ever understand . Franz´en [ 5 ] surveys and demolishes many of these fallacies . The second incompleteness theorem destroyed Hilbert ’ s hope that the consistency of quite strong theories might be proved even in Peano arithmetic . It also tells us , for example , that the axioms of set theory do not imply the existence of an inaccessible cardinal , as that would yield a model for set theory itself . The first incompleteness theorem has been proved with machine assistance at least three times before . The first time ( surprisingly early : 1986 ) was by Shankar [ 29 , 30 ] , using Nqthm . Then in 2004 , O ’ Connor [ 22 ] ( using Coq ) and Harrison ( using HOL Light ) 2 each proved versions of the theorem . The present proof , conducted using Isabelle/HOL , is novel in adopting nominal syntax [ 34 ] for formalising variable binding in
|
||
#2 this paper being one of numerous instances . It has also been used to assert that human intelligence can perceive truths ( in particular , the truth of δ , the undecidable sentence ) that no computer will ever understand . Franz´en [ 5 ] surveys and demolishes many of these fallacies . The second incompleteness theorem destroyed Hilbert ’ s hope that the consistency of quite strong theories might be proved even in Peano arithmetic . It also tells us , for example , that the axioms of set theory do not imply the existence of an inaccessible cardinal , as that would yield a model for set theory itself . The first incompleteness theorem has been proved with machine assistance at least three times before . The first time ( surprisingly early : 1986 ) was by Shankar [ 29 , 30 ] , using Nqthm . Then in 2004 , O ’ Connor [ 22 ] ( using Coq ) and Harrison ( using HOL Light ) 2 each proved versions of the theorem . The present proof , conducted using Isabelle/HOL , is novel in adopting nominal syntax [ 34 ] for formalising variable binding in the syntax of L , while using de Bruijn notation [ 3 ] for coding those formulas . Despite using two different treatments of variable binding , the necessary representation theorem for formulas is not difficult to prove . It is not clear that other treatments of higher-order abstract syntax could make this claim . These proofs can be seen as an extended demonstration of the power of nominal syntax , while at the same time vindicating de Bruijn indexing in some situations . The machine proofs are fairly concise at under 12,400 lines for both theorems.3 The paper presents highlights of the proof , commenting on the advantages and disadvantages of the nominal framework and HF set theory . An overview of the project from a logician ’ s perspective has appeared elsewhere [ 27 ] . The proof reported here closely follows a detailed exposition by Swierczkowski [ 32 ] . His careful and detailed proofs were ´ indispensable , despite some errors and omissions , which are reported below . For the first time , we have complete , formal proofs of both theorems . They take the form of structured Isar proof scripts [ 26 ] that can be examined interactively . The remainder of the paper presents background material ( Sect . 2 ) before outlining the development itself : the proof calculus ( Sect . 3 ) , the coding of the calculus within itself ( Sect . 4 ) and finally the first theorem ( Sect . 5 ) . Technical material relating to the second theorem are developed ( Sect . 6 ) then the theorem is presented and discussed ( Sect . 7 ) . Finally , the paper concludes ( Sect . 8 ) . 2 Proof files at http
|
||
#3 highlights of the proof , commenting on the advantages and disadvantages of the nominal framework and HF set theory . An overview of the project from a logician ’ s perspective has appeared elsewhere [ 27 ] . The proof reported here closely follows a detailed exposition by Swierczkowski [ 32 ] . His careful and detailed proofs were ´ indispensable , despite some errors and omissions , which are reported below . For the first time , we have complete , formal proofs of both theorems . They take the form of structured Isar proof scripts [ 26 ] that can be examined interactively . The remainder of the paper presents background material ( Sect . 2 ) before outlining the development itself : the proof calculus ( Sect . 3 ) , the coding of the calculus within itself ( Sect . 4 ) and finally the first theorem ( Sect . 5 ) . Technical material relating to the second theorem are developed ( Sect . 6 ) then the theorem is presented and discussed ( Sect . 7 ) . Finally , the paper concludes ( Sect . 8 ) . 2 Proof files at http : //code.google.com/p/hol-light/ , directory Arithmetic 3 This is approximately as long as Isabelle ’ s theory of Kurzweil-Henstock gauge integration . 3 2 Background Isabelle/HOL [ 20 ] is an interactive theorem prover for higher-order logic . This formalism can be seen as extending a polymorphic typed first-order logic with a functional language in which recursive datatypes and functions can be defined . Extensive documentation is available.4 For interpreting the theorem statements presented below , it is important to note that a theorem that concludes ψ from the premises φ1 , . . . , φn may be expressed in a variety of equivalent forms . The following denote precisely the same theorem : [ [ φ1 ; ... ; φn ] ] =⇒ ψ φ1 =⇒ · · · =⇒ φn =⇒ ψ assumes φ1 and ... and φn shows ψ If the conclusion of a theorem involves the keyword obtains , then there is an implicit existential quantification . The following two theorems are logically equivalent . φ =⇒ ∃x . ψ1 ∧ . . . ∧ ψn assumes φ obtains x where ψ1 ... ψn Other background material for this paper includes an outline of G¨odel ’ s proof , an introduction to hereditarily finite set theory and some examples of Nominal Isabelle . 2.1 G¨odel ’ s Proof Much of G¨odel ’ s proof may be known to many readers , but it will be useful to list the milestones here , for reference . 1 . A first-order deductive calculus is formalised , including the syntax of terms and formulas , substitution , and semantics ( evaluation ) . The calculus includes axioms to define Peano arithmetic or some alternative , such as the HF set theory used here . There are inference rules for
|
||
#4 a variety of equivalent forms . The following denote precisely the same theorem : [ [ φ1 ; ... ; φn ] ] =⇒ ψ φ1 =⇒ · · · =⇒ φn =⇒ ψ assumes φ1 and ... and φn shows ψ If the conclusion of a theorem involves the keyword obtains , then there is an implicit existential quantification . The following two theorems are logically equivalent . φ =⇒ ∃x . ψ1 ∧ . . . ∧ ψn assumes φ obtains x where ψ1 ... ψn Other background material for this paper includes an outline of G¨odel ’ s proof , an introduction to hereditarily finite set theory and some examples of Nominal Isabelle . 2.1 G¨odel ’ s Proof Much of G¨odel ’ s proof may be known to many readers , but it will be useful to list the milestones here , for reference . 1 . A first-order deductive calculus is formalised , including the syntax of terms and formulas , substitution , and semantics ( evaluation ) . The calculus includes axioms to define Peano arithmetic or some alternative , such as the HF set theory used here . There are inference rules for propositional and quantifier reasoning . We write H ⊢ A to mean that A can be proved from H ( a set of formulas ) in the calculus . 2 . Meta-theory is developed to relate truth and provability . The need for tedious proof constructions in the deductive calculus is minimised through a meta-theorem stating that a class of true formulas are theorems of that calculus . One way to do this is through the notion of Σ formulas , which are built up from atomic formulas using ∨ , ∧ , ∃ and bounded universal quantification . Then the key result is If α is a true Σ sentence , then ⊢ α . ( 1 ) 3 . A system of coding is set up within the formalised first-order theory . The code of a formula α is written pαq and is a term of the calculus itself . 4 . In order to formalise the calculus within itself , predicates to recognise codes are defined , including terms and formulas , and syntactic operations on them . Next , predicates are defined to recognise individual axioms and inference rules , then a sequence of such logical steps . We obtain a predicate Pf , where ⊢ Pf pαq expresses that the formula α has a proof . The key result is ⊢ α ⇐⇒ ⊢ Pf pαq . ( 2 ) 4 http : //isabelle.in.tum.de/documentation.html 4 All of these developments must be completed before the second incompleteness theorem can even be stated . 5 . G¨odel ’ s first incompleteness theorem is obtained by constructing a formula δ that is provably equivalent ( within the calculus ) to the formal statement that δ is not provable . It follows ( provided the calculus is consistent )
|
||
#5 the key result is If α is a true Σ sentence , then ⊢ α . ( 1 ) 3 . A system of coding is set up within the formalised first-order theory . The code of a formula α is written pαq and is a term of the calculus itself . 4 . In order to formalise the calculus within itself , predicates to recognise codes are defined , including terms and formulas , and syntactic operations on them . Next , predicates are defined to recognise individual axioms and inference rules , then a sequence of such logical steps . We obtain a predicate Pf , where ⊢ Pf pαq expresses that the formula α has a proof . The key result is ⊢ α ⇐⇒ ⊢ Pf pαq . ( 2 ) 4 http : //isabelle.in.tum.de/documentation.html 4 All of these developments must be completed before the second incompleteness theorem can even be stated . 5 . G¨odel ’ s first incompleteness theorem is obtained by constructing a formula δ that is provably equivalent ( within the calculus ) to the formal statement that δ is not provable . It follows ( provided the calculus is consistent ) that neither δ nor its negation can be proved , although δ is true in the semantics . 6 . G¨odel ’ s second incompleteness theorem requires the following crucial lemma : If α is a Σ sentence , then ⊢ α → Pf pαq . This is an internalisation of theorem ( 1 ) above . It is proved by induction over the construction of α as a Σ formula . This requires generalising the statement above to allow the formula α to contain free variables . The technical details are complicated , and lengthy deductions in the calculus seem to be essential . The proof sketched above incorporates numerous improvements over G¨odel ’ s original version . G¨odel proved only the left-to-right direction of the equivalence ( 2 ) and required a stronger assumption than consistency , namely ω-consistency . 2.2 Hereditarily Finite Set Theory G¨odel first proved his incompleteness theorems in a first-order theory of Peano arithmetic [ 7 ] . O ’ Connor and Harrison do the same , while Shankar and I have both chosen a formalisation of the hereditarily finite ( HF ) sets . Although each theory can be formally interpreted in the other , meaning that they are of equivalent strength , the HF theory is more convenient , as it can express all set-theoretic constructions that do not require infinite sets . An HF set is a finite set of HF sets , and this recursive definition can be captured by the following three axioms : z = 0 ↔ ∀x [ x 6∈ z ] ( HF1 ) z = x ⊳ y ↔ ∀u [ u ∈ z ↔ u ∈ x ∨ u = y ] ( HF2 ) φ ( 0 ) ∧ ∀xy [ φ (
|
||
#6 seem to be essential . The proof sketched above incorporates numerous improvements over G¨odel ’ s original version . G¨odel proved only the left-to-right direction of the equivalence ( 2 ) and required a stronger assumption than consistency , namely ω-consistency . 2.2 Hereditarily Finite Set Theory G¨odel first proved his incompleteness theorems in a first-order theory of Peano arithmetic [ 7 ] . O ’ Connor and Harrison do the same , while Shankar and I have both chosen a formalisation of the hereditarily finite ( HF ) sets . Although each theory can be formally interpreted in the other , meaning that they are of equivalent strength , the HF theory is more convenient , as it can express all set-theoretic constructions that do not require infinite sets . An HF set is a finite set of HF sets , and this recursive definition can be captured by the following three axioms : z = 0 ↔ ∀x [ x 6∈ z ] ( HF1 ) z = x ⊳ y ↔ ∀u [ u ∈ z ↔ u ∈ x ∨ u = y ] ( HF2 ) φ ( 0 ) ∧ ∀xy [ φ ( x ) ∧ φ ( y ) → φ ( x ⊳ y ) ] → ∀x [ φ ( x ) ] ( HF3 ) The first axiom states that 0 denotes the empty set . The second axiom defines the binary operation symbol ⊳ ( pronounced “ eats ” ) to denote insertion into a set , so that x ⊳ y = x ∪ { y } . The third axiom is an induction scheme , and states that every set is created by a finite number of applications of 0 and ⊳ . The machine proofs of the incompleteness theorems rest on an Isabelle theory of the hereditarily finite sets . To illustrate the syntax , here are the three basic axioms as formalised in Isabelle . The type of such sets is called hf , and is constructed such that the axioms above can be proved . lemma hempty iff : `` z=0 ←→ ( ∀ x . ¬ x ∈ z ) '' lemma hinsert iff : `` z = x ⊳ y ←→ ( ∀ u. u ∈ z ←→ u ∈ x | u = y ) '' lemma hf induct ax : `` [ [ P 0 ; ∀ x. P x −→ ( ∀ y. P y −→ P ( x ⊳ y ) ) ] ] =⇒ P x '' The same three axioms , formalised within Isabelle as a deep embedding , form the basis for the incompleteness proofs . Type hf and its associated operators serve as the standard model for the embedded HF set theory . HF set theory is equivalent to Zermelo-Frankel ( ZF ) set theory with the axiom of infinity negated . Many of the Isabelle definitions and theorems were taken , with
|
||
#7 of the incompleteness theorems rest on an Isabelle theory of the hereditarily finite sets . To illustrate the syntax , here are the three basic axioms as formalised in Isabelle . The type of such sets is called hf , and is constructed such that the axioms above can be proved . lemma hempty iff : `` z=0 ←→ ( ∀ x . ¬ x ∈ z ) '' lemma hinsert iff : `` z = x ⊳ y ←→ ( ∀ u. u ∈ z ←→ u ∈ x | u = y ) '' lemma hf induct ax : `` [ [ P 0 ; ∀ x. P x −→ ( ∀ y. P y −→ P ( x ⊳ y ) ) ] ] =⇒ P x '' The same three axioms , formalised within Isabelle as a deep embedding , form the basis for the incompleteness proofs . Type hf and its associated operators serve as the standard model for the embedded HF set theory . HF set theory is equivalent to Zermelo-Frankel ( ZF ) set theory with the axiom of infinity negated . Many of the Isabelle definitions and theorems were taken , with minor 5 modifications , from Isabelle/ZF [ 24 ] . Familiar concepts such as union , intersection , set difference , power set , replacement , ordered pairing and foundation can be derived in terms of the axioms shown above [ 32 ] . A few of these derivations need to be repeated— with infinitely greater effort—in the internal calculus . Ordinals in HF are simply natural numbers , where n = { 0 , 1 , . . . , n − 1 } . Their typical properties ( for example , that they are linearly ordered ) have the same proofs as in ZF set theory . Swierczkowski ’ s proofs [ 32 ] are sometimes more elegant , an ´ d addition on ordinals is obtained through a general notion of addition of sets [ 15 ] . Finally , there are about 400 lines of material concerned with relations , functions and finite sequences . This is needed to reason about the coding of syntax for the incompleteness theorem . 2.3 Isabelle ’ s Nominal Package For the incompleteness theorems , we are concerned with formalising the syntax of firstorder logic . Variable binding is a particular issue : it is well known that technicalities relating to bound variables and substitution have caused errors in published proofs and complicate formal treatments . O ’ Connor [ 23 ] reports severe difficulties in his proofs . The nominal approach [ 6 , 28 ] to formalising variable binding ( and other sophisticated uses of variable names ) is based on a theory of permutations over names . A primitive concept is support : supp ( x ) has a rather technical definition involving permutations , but is equivalent in typical situations to the set of free names
|
||
#8 the same proofs as in ZF set theory . Swierczkowski ’ s proofs [ 32 ] are sometimes more elegant , an ´ d addition on ordinals is obtained through a general notion of addition of sets [ 15 ] . Finally , there are about 400 lines of material concerned with relations , functions and finite sequences . This is needed to reason about the coding of syntax for the incompleteness theorem . 2.3 Isabelle ’ s Nominal Package For the incompleteness theorems , we are concerned with formalising the syntax of firstorder logic . Variable binding is a particular issue : it is well known that technicalities relating to bound variables and substitution have caused errors in published proofs and complicate formal treatments . O ’ Connor [ 23 ] reports severe difficulties in his proofs . The nominal approach [ 6 , 28 ] to formalising variable binding ( and other sophisticated uses of variable names ) is based on a theory of permutations over names . A primitive concept is support : supp ( x ) has a rather technical definition involving permutations , but is equivalent in typical situations to the set of free names in x . We also write a ♯ x for a 6∈ supp ( x ) , saying “ a is fresh for x ” . Isabelle ’ s nominal package [ 33 , 34 ] supports these concepts through commands such as nominal datatype to introduce types , nominal primrec to declare primitive recursive functions and nominal induct to perform structural induction . Syntactic constructions involving variable binding are identified up to α-conversion , using a quotient construction . These mechanisms generally succeed at emulating informal standard conventions for variable names . In particular , we can usually assume that the bound variables we encounter never clash with other variables . The best way to illustrate these ideas is by examples . The following datatype defines the syntax of terms in the HF theory : nominal datatype tm = Zero | Var name | Eats tm tm The type name ( of variable names ) has been created using the nominal framework . The members of this type constitute a countable set of uninterpreted atoms . The function nat of name is a bijection between this type and the type of natural numbers . Here is the syntax of HF formulas , which are t ∈ u , t = u , φ ∨ ψ , ¬φ or ∃x [ φ ] : nominal datatype fm = Mem tm tm ( infixr `` IN '' 150 ) | Eq tm tm ( infixr `` EQ '' 150 ) | Disj fm fm ( infixr `` OR '' 130 ) | Neg fm | Ex x : :name f : :fm binds x in f In most respects , this nominal datatype behaves exactly like a standard algebraic datatype . However , the bound variable name designated by x above
|
||
#9 that the bound variables we encounter never clash with other variables . The best way to illustrate these ideas is by examples . The following datatype defines the syntax of terms in the HF theory : nominal datatype tm = Zero | Var name | Eats tm tm The type name ( of variable names ) has been created using the nominal framework . The members of this type constitute a countable set of uninterpreted atoms . The function nat of name is a bijection between this type and the type of natural numbers . Here is the syntax of HF formulas , which are t ∈ u , t = u , φ ∨ ψ , ¬φ or ∃x [ φ ] : nominal datatype fm = Mem tm tm ( infixr `` IN '' 150 ) | Eq tm tm ( infixr `` EQ '' 150 ) | Disj fm fm ( infixr `` OR '' 130 ) | Neg fm | Ex x : :name f : :fm binds x in f In most respects , this nominal datatype behaves exactly like a standard algebraic datatype . However , the bound variable name designated by x above is not significant : no function can be defined to return the name of a variable bound using Ex . Substitution of a term x for a variable i is defined as follows : 6 nominal primrec subst : : `` name ⇒ tm ⇒ tm ⇒ tm '' where '' subst i x Zero = Zero '' | `` subst i x ( Var k ) = ( if i=k then x else Var k ) '' | `` subst i x ( Eats t u ) = Eats ( subst i x t ) ( subst i x u ) '' Unfortunately , most recursive definitions involving nominal primrec must be followed by a series of proof steps , verifying that the function uses names legitimately . Occasionally , these proofs ( omitted here ) require subtle reasoning involving nominal primitives . Substituting the term x for the variable i in the formula A is written A ( i : :=x ) . nominal primrec subst fm : : `` fm ⇒ name ⇒ tm ⇒ fm '' where Mem : `` ( Mem t u ) ( i : :=x ) = Mem ( subst i x t ) ( subst i x u ) '' | Eq : `` ( Eq t u ) ( i : :=x ) = Eq ( subst i x t ) ( subst i x u ) '' | Disj : `` ( Disj A B ) ( i : :=x ) = Disj ( A ( i : :=x ) ) ( B ( i : :=x ) ) '' | Neg : `` ( Neg A ) ( i : :=x ) = Neg ( A ( i : :=x ) ) '' | Ex : `` atom
|
||
#10 u ) '' Unfortunately , most recursive definitions involving nominal primrec must be followed by a series of proof steps , verifying that the function uses names legitimately . Occasionally , these proofs ( omitted here ) require subtle reasoning involving nominal primitives . Substituting the term x for the variable i in the formula A is written A ( i : :=x ) . nominal primrec subst fm : : `` fm ⇒ name ⇒ tm ⇒ fm '' where Mem : `` ( Mem t u ) ( i : :=x ) = Mem ( subst i x t ) ( subst i x u ) '' | Eq : `` ( Eq t u ) ( i : :=x ) = Eq ( subst i x t ) ( subst i x u ) '' | Disj : `` ( Disj A B ) ( i : :=x ) = Disj ( A ( i : :=x ) ) ( B ( i : :=x ) ) '' | Neg : `` ( Neg A ) ( i : :=x ) = Neg ( A ( i : :=x ) ) '' | Ex : `` atom j ♯ ( i , x ) =⇒ ( Ex j A ) ( i : :=x ) = Ex j ( A ( i : :=x ) ) '' Note that the first seven cases ( considering the two substitution functions collectively ) are straightforward structural recursion . In the final case , we see a precondition , atom j ♯ ( i , x ) , to ensure that the substitution is legitimate within the formula Ex j A . There is no way to define the contrary case , where the bound variable clashes . We would have to eliminate any such clash , renaming the bound variable by applying an appropriate permutation to the formula . Thanks to the nominal framework , such explicit renaming steps are rare . This style of formalisation is more elegant than traditional textbook definitions that do include the variable-clashing case . It is much more elegant than including renaming of the bound variable as part of the definition itself . Such “ definitions ” are really implementations , and greatly complicate proofs . The commutativity of substitution ( two distinct variables , each fresh for the opposite term ) is easily proved . lemma subst fm commute2 [ simp ] : '' [ [ atom j ♯ t ; atom i ♯ u ; i 6= j ] ] =⇒ ( A ( i : :=t ) ) ( j : :=u ) = ( A ( j : :=u ) ) ( i : :=t ) '' by ( nominal induct A avoiding : i j t u rule : fm.strong induct ) auto The proof is by nominal induction on the formula A : the proof method for structural induction over a nominal datatype . Compared with ordinary
|
||
#11 have to eliminate any such clash , renaming the bound variable by applying an appropriate permutation to the formula . Thanks to the nominal framework , such explicit renaming steps are rare . This style of formalisation is more elegant than traditional textbook definitions that do include the variable-clashing case . It is much more elegant than including renaming of the bound variable as part of the definition itself . Such “ definitions ” are really implementations , and greatly complicate proofs . The commutativity of substitution ( two distinct variables , each fresh for the opposite term ) is easily proved . lemma subst fm commute2 [ simp ] : '' [ [ atom j ♯ t ; atom i ♯ u ; i 6= j ] ] =⇒ ( A ( i : :=t ) ) ( j : :=u ) = ( A ( j : :=u ) ) ( i : :=t ) '' by ( nominal induct A avoiding : i j t u rule : fm.strong induct ) auto The proof is by nominal induction on the formula A : the proof method for structural induction over a nominal datatype . Compared with ordinary induction , nominal induct takes account of the freshness of bound variable names . The phrase avoiding : i j t u is the formal equivalent of the convention that when we consider the case of the existential formula Ex k A , the bound variable k can be assumed to be fresh with respect to the terms mentioned . This convention is formalised by four additional assumptions atom k ♯ i , atom k ♯ j , etc . ; they ensure that substitution will be well-defined over this existential quantifier , making the proof easy . This and many similar facts have two-step proofs , nominal induct followed by auto . In contrast , O ’ Connor needed to combine three substitution lemmas ( including the one above ) in a giant mutual induction involving 1,900 lines of Coq . He blames the renaming step in substitution and suggests that a form of simultaneous substitution might have avoided these difficulties [ 23§4.3 ] . An alternative , using traditional bound variable names , is to treat substitution not as a function but as a relation that holds only when no renaming is necessary . Bound variable renaming is then an independent operation . I briefly tried this idea , which allowed reasonably straightforward proofs of substitution properties , but ultimately nominal looked like a better option . 7 3 Theorems , Σ Formulas , Provability The first milestone in the proof of the incompleteness theorems is the development of a first-order logical calculus equipped with enough meta-theory to guarantee that some true formulas are theorems . The previous section has already presented the definitions of the terms and formulas of this calculus . The terms are for HF set theory , and the formulas are defined via a
|
||
#12 many similar facts have two-step proofs , nominal induct followed by auto . In contrast , O ’ Connor needed to combine three substitution lemmas ( including the one above ) in a giant mutual induction involving 1,900 lines of Coq . He blames the renaming step in substitution and suggests that a form of simultaneous substitution might have avoided these difficulties [ 23§4.3 ] . An alternative , using traditional bound variable names , is to treat substitution not as a function but as a relation that holds only when no renaming is necessary . Bound variable renaming is then an independent operation . I briefly tried this idea , which allowed reasonably straightforward proofs of substitution properties , but ultimately nominal looked like a better option . 7 3 Theorems , Σ Formulas , Provability The first milestone in the proof of the incompleteness theorems is the development of a first-order logical calculus equipped with enough meta-theory to guarantee that some true formulas are theorems . The previous section has already presented the definitions of the terms and formulas of this calculus . The terms are for HF set theory , and the formulas are defined via a minimal set of connectives from which others can be defined . 3.1 A Sequent Calculus for HF Set Theory Compared with a textbook presentation , a machine development must include an effective proof system , one that can actually be used to prove non-trivial theorems . 3.1.1 Semantics The semantics of terms and formulas are given by the obvious recursive function definitions , which yield sets and Booleans , respectively . These functions accept an environment mapping variables to values . The null environment maps all variables to 0 , and is written e0 . It involves the types finfun ( for finite functions ) [ 17 ] and hf ( for HF sets ) . definition e0 : : `` ( name , hf ) finfun '' — the null environment where `` e0 ≡ finfun const 0 '' nominal primrec eval tm : : `` ( name , hf ) finfun ⇒ tm ⇒ hf '' where '' eval tm e Zero = 0 '' | `` eval tm e ( Var k ) = finfun apply e k '' | `` eval tm e ( Eats t u ) = eval tm e t ⊳ eval tm e u '' There are two things to note in the semantics of formulas . First , the special syntax [ [ t ] ] e abbreviates eval tm e t. Second , in the semantics of the quantifier Ex , note how the formula atom k ♯ e asserts that the bound variable , k , is not currently given a value by the environment , e. nominal primrec eval fm : : `` ( name , hf ) finfun ⇒ fm ⇒ bool '' where '' eval fm e ( t IN u ) ←→ [ [ t
|
||
#13 ( for finite functions ) [ 17 ] and hf ( for HF sets ) . definition e0 : : `` ( name , hf ) finfun '' — the null environment where `` e0 ≡ finfun const 0 '' nominal primrec eval tm : : `` ( name , hf ) finfun ⇒ tm ⇒ hf '' where '' eval tm e Zero = 0 '' | `` eval tm e ( Var k ) = finfun apply e k '' | `` eval tm e ( Eats t u ) = eval tm e t ⊳ eval tm e u '' There are two things to note in the semantics of formulas . First , the special syntax [ [ t ] ] e abbreviates eval tm e t. Second , in the semantics of the quantifier Ex , note how the formula atom k ♯ e asserts that the bound variable , k , is not currently given a value by the environment , e. nominal primrec eval fm : : `` ( name , hf ) finfun ⇒ fm ⇒ bool '' where '' eval fm e ( t IN u ) ←→ [ [ t ] ] e ∈ [ [ u ] ] e '' | `` eval fm e ( t EQ u ) ←→ [ [ t ] ] e = [ [ u ] ] e '' | `` eval fm e ( A OR B ) ←→ eval fm e A ∨ eval fm e B '' | `` eval fm e ( Neg A ) ←→ ( ~ eval fm e A ) '' | `` atom k ♯ e =⇒ eval fm e ( Ex k A ) ←→ ( ∃ x. eval fm ( finfun update e k x ) A ) '' This yields the Tarski truth definition for the standard model of HF set theory . In particular , eval fm e0 A denotes the truth of the sentence A . 3.1.2 Axioms Swierczkowski [ 32 ] specifies a standard Hilbert-style calc ´ ulus , with two rules of inference and several axioms or axiom schemes . The latter include sentential axioms , defining the behaviour of disjunction and negation : 8 inductive set boolean axioms : : `` fm set '' where Ident : `` A IMP A ∈ boolean axioms '' | DisjI1 : `` A IMP ( A OR B ) ∈ boolean axioms '' | DisjCont : `` ( A OR A ) IMP A ∈ boolean axioms '' | DisjAssoc : `` ( A OR ( B OR C ) ) IMP ( ( A OR B ) OR C ) ∈ boolean axioms '' | DisjConj : `` ( C OR A ) IMP ( ( ( Neg C ) OR B ) IMP ( A OR B ) ) ∈ boolean axioms '' Here Swierczkowski makes a tiny error , expressing the last axiom ´ scheme as (
|
||
#14 k x ) A ) '' This yields the Tarski truth definition for the standard model of HF set theory . In particular , eval fm e0 A denotes the truth of the sentence A . 3.1.2 Axioms Swierczkowski [ 32 ] specifies a standard Hilbert-style calc ´ ulus , with two rules of inference and several axioms or axiom schemes . The latter include sentential axioms , defining the behaviour of disjunction and negation : 8 inductive set boolean axioms : : `` fm set '' where Ident : `` A IMP A ∈ boolean axioms '' | DisjI1 : `` A IMP ( A OR B ) ∈ boolean axioms '' | DisjCont : `` ( A OR A ) IMP A ∈ boolean axioms '' | DisjAssoc : `` ( A OR ( B OR C ) ) IMP ( ( A OR B ) OR C ) ∈ boolean axioms '' | DisjConj : `` ( C OR A ) IMP ( ( ( Neg C ) OR B ) IMP ( A OR B ) ) ∈ boolean axioms '' Here Swierczkowski makes a tiny error , expressing the last axiom ´ scheme as ( φ ∨ ψ ) ∧ ( ¬φ ∨ µ ) → ψ ∨ µ . Because ∧ is defined in terms of ∨ , while this axiom helps to define ∨ , this formulation is unlikely to work . The Isabelle version eliminates ∧ in favour of nested implication . There are four primitive equality axioms , shown below in mathematical notation . They express reflexivity as well as substitutivity for equality , membership and the eats operator . They are not schemes but single formulas containing specific free variables . Creating an instance of an axiom for specific terms ( which might involve the same variables ) requires many renaming steps to insert fresh variables , before substituting for them one term at a time . x1 = x1 ( x1 = x2 ) ∧ ( x3 = x4 ) → [ ( x1 = x3 ) → ( x2 = x4 ) ] ( x1 = x2 ) ∧ ( x3 = x4 ) → [ ( x1 ∈ x3 ) → ( x2 ∈ x4 ) ] ( x1 = x2 ) ∧ ( x3 = x4 ) → [ x1 ⊳ x3 = x2 ⊳ x4 ] There is also a specialisation axiom scheme , of the form φ ( t/x ) → ∃x φ : inductive set special axioms : : `` fm set '' where I : `` A ( i : :=x ) IMP ( Ex i A ) ∈ special axioms '' There are the axioms HF1 and HF2 for the set theory , while HF3 ( induction ) is formalised as an axiom scheme : inductive set induction axioms : : `` fm set '' where ind : '' atom ( j : :name ) ♯ ( i , A )
|
||
#15 ( which might involve the same variables ) requires many renaming steps to insert fresh variables , before substituting for them one term at a time . x1 = x1 ( x1 = x2 ) ∧ ( x3 = x4 ) → [ ( x1 = x3 ) → ( x2 = x4 ) ] ( x1 = x2 ) ∧ ( x3 = x4 ) → [ ( x1 ∈ x3 ) → ( x2 ∈ x4 ) ] ( x1 = x2 ) ∧ ( x3 = x4 ) → [ x1 ⊳ x3 = x2 ⊳ x4 ] There is also a specialisation axiom scheme , of the form φ ( t/x ) → ∃x φ : inductive set special axioms : : `` fm set '' where I : `` A ( i : :=x ) IMP ( Ex i A ) ∈ special axioms '' There are the axioms HF1 and HF2 for the set theory , while HF3 ( induction ) is formalised as an axiom scheme : inductive set induction axioms : : `` fm set '' where ind : '' atom ( j : :name ) ♯ ( i , A ) =⇒ A ( i : :=Zero ) IMP ( ( All i ( All j ( A IMP ( A ( i : := Var j ) IMP A ( i : := Eats ( Var i ) ( Var j ) ) ) ) ) ) IMP ( All i A ) ) ∈ induction axioms '' Axiom schemes are conveniently introduced using inductive set , simply to express set comprehensions , even though there is no actual induction . 3.1.3 Inference System The axiom schemes shown above , along with inference rules for modus ponens and existential instantiation,5 are combined to form the following inductive definition of theorems : 5 From A → B infer ∃xA → B , for x not free in B . 9 inductive hfthm : : `` fm set ⇒ fm ⇒ bool '' ( infixl `` ⊢ '' 55 ) where Hyp : `` A ∈ H =⇒ H ⊢ A '' | Extra : `` H ⊢ extra axiom '' | Bool : `` A ∈ boolean axioms =⇒ H ⊢ A '' | Eq : `` A ∈ equality axioms =⇒ H ⊢ A '' | Spec : `` A ∈ special axioms =⇒ H ⊢ A '' | HF : `` A ∈ HF axioms =⇒ H ⊢ A '' | Ind : `` A ∈ induction axioms =⇒ H ⊢ A '' | MP : `` H ⊢ A IMP B =⇒ H ’ ⊢ A =⇒ H ∪ H ’ ⊢ B '' | Exists : `` H ⊢ A IMP B =⇒ atom i♯B =⇒ ∀ C∈H . atom i♯C =⇒ H ⊢ ( Ex i A ) IMP B '' A minor deviation from Swierczkowski is ´ extra axiom , which is abstractly specified to
|
||
#16 are combined to form the following inductive definition of theorems : 5 From A → B infer ∃xA → B , for x not free in B . 9 inductive hfthm : : `` fm set ⇒ fm ⇒ bool '' ( infixl `` ⊢ '' 55 ) where Hyp : `` A ∈ H =⇒ H ⊢ A '' | Extra : `` H ⊢ extra axiom '' | Bool : `` A ∈ boolean axioms =⇒ H ⊢ A '' | Eq : `` A ∈ equality axioms =⇒ H ⊢ A '' | Spec : `` A ∈ special axioms =⇒ H ⊢ A '' | HF : `` A ∈ HF axioms =⇒ H ⊢ A '' | Ind : `` A ∈ induction axioms =⇒ H ⊢ A '' | MP : `` H ⊢ A IMP B =⇒ H ’ ⊢ A =⇒ H ∪ H ’ ⊢ B '' | Exists : `` H ⊢ A IMP B =⇒ atom i♯B =⇒ ∀ C∈H . atom i♯C =⇒ H ⊢ ( Ex i A ) IMP B '' A minor deviation from Swierczkowski is ´ extra axiom , which is abstractly specified to be an arbitrary true formula . This means that the proofs will be conducted with respect to an arbitrary finite extension of the HF theory . The first major deviation from Swierczkowski is the introduction of rule ´ Hyp , with a set of assumptions . It would be virtually impossible to prove anything in his Hilbert-style proof system , and it was clear from the outset that lengthy proofs within the calculus might be necessary . Introducing Hyp generalises the notion of provability , allowing the development of a sort of sequent calculus , in which long but tolerably natural proofs can be constructed . It is worth mentioning that Swierczkowski ’ s definitions and proofs fit together very ´ tightly , deviations often being a cause for later regret . One example , concerning an inference rule for substitution , is mentioned at the end of Sect . 4.4 . Another example is that some tricks that simplify the proof of the first incompleteness theorem turn out to complicate the proof of the second . The soundness of the calculus above is trivial to prove by induction . The deduction theorem is also straightforward , the only non-trivial case being the one for the Exists inference rule . The induction formula is stated as follows : lemma deduction Diff : assumes `` H ⊢ B '' shows `` H - { C } ⊢ C IMP B '' This directly yields the standard formulation of the deduction theorem : theorem deduction : assumes `` insert A H ⊢ B '' shows `` H ⊢ A IMP B '' And this is a sequent rule for implication . Setting up a usable sequent calculus requires much work . The corresponding Isabelle theory file , which starts with the definitions
|
||
#17 natural proofs can be constructed . It is worth mentioning that Swierczkowski ’ s definitions and proofs fit together very ´ tightly , deviations often being a cause for later regret . One example , concerning an inference rule for substitution , is mentioned at the end of Sect . 4.4 . Another example is that some tricks that simplify the proof of the first incompleteness theorem turn out to complicate the proof of the second . The soundness of the calculus above is trivial to prove by induction . The deduction theorem is also straightforward , the only non-trivial case being the one for the Exists inference rule . The induction formula is stated as follows : lemma deduction Diff : assumes `` H ⊢ B '' shows `` H - { C } ⊢ C IMP B '' This directly yields the standard formulation of the deduction theorem : theorem deduction : assumes `` insert A H ⊢ B '' shows `` H ⊢ A IMP B '' And this is a sequent rule for implication . Setting up a usable sequent calculus requires much work . The corresponding Isabelle theory file , which starts with the definitions of terms and formulas and ends with a sequent formulation of the HF induction rule , is nearly 1,600 lines long . Deriving natural sequent calculus rules from the sentential and equality axioms requires lengthy chains of steps . Even in the final derived sequent calculus , equalities can only be applied one step at a time . For another example of difficulty , consider the following definition : definition Fls where `` Fls ≡ Zero IN Zero '' Proving that Fls has the properties of falsehood is surprisingly tricky . The relevant axiom , HF1 , is formulated using universal quantifiers , which are defined as negated existentials ; deriving the expected properties of universal quantification seems to require something like Fls itself . The derived sequent calculus has specialised rules to operate on conjunctions , disjunctions , etc. , in the hypothesis part of a sequent . They are crude , but good enough . Used with Isabelle ’ s automatic tactics , they ease somewhat the task of constructing formal HF proofs . Users can extend Isabelle with proof procedures coded in ML , and better automation for the calculus might thereby be achieved . At the time , such a side-project did not seem to be worth the effort . 10 3.2 A Formal Theory of Functions Recursion is not available in HF set theory , and recursive functions must be constructed explicitly . Each recursive computation is expressed in terms of the existence of a sequence ( si ) i≤k such that si is related to sm and sn for m , n < i . Moreover , a sequence is formally a relation rather than a function . In the metalanguage , we write app s k for sk , governed by the theorem
|
||
#18 universal quantifiers , which are defined as negated existentials ; deriving the expected properties of universal quantification seems to require something like Fls itself . The derived sequent calculus has specialised rules to operate on conjunctions , disjunctions , etc. , in the hypothesis part of a sequent . They are crude , but good enough . Used with Isabelle ’ s automatic tactics , they ease somewhat the task of constructing formal HF proofs . Users can extend Isabelle with proof procedures coded in ML , and better automation for the calculus might thereby be achieved . At the time , such a side-project did not seem to be worth the effort . 10 3.2 A Formal Theory of Functions Recursion is not available in HF set theory , and recursive functions must be constructed explicitly . Each recursive computation is expressed in terms of the existence of a sequence ( si ) i≤k such that si is related to sm and sn for m , n < i . Moreover , a sequence is formally a relation rather than a function . In the metalanguage , we write app s k for sk , governed by the theorem lemma app equality : `` hfunction s =⇒ hx , yi ∈ s =⇒ app s x = y '' The following two functions express the recursive definition of sequences , as needed for the G¨odel development : '' Builds B C s l ≡ B ( app s l ) ∨ ( ∃ m∈l . ∃ n∈l . C ( app s l ) ( app s m ) ( app s n ) ) '' '' BuildSeq B C s k y ≡ LstSeq s k y ∧ ( ∀ l∈succ k. Builds B C s l ) '' The statement Builds B C s l constrains element l of sequence s , namely app s l. We have either B ( app s l ) , or C ( app s l ) ( app s m ) ( app s n ) ) where m∈l and n∈l . For the natural numbers , set membership coincides with the less-than relation . Therefore , we are referring to a sequence s and element sl where either the base case B ( sl ) holds , or else the recursive step C ( sl , sm , sn ) for m , n < l. The statement BuildSeq B C s k y states that the sequence s has been constructed in this way right up to the value app s k , or in other words , sk , where y = sk . To formalise the basis for this approach requires a series of definitions in the HF calculus , introducing the subset relation , ordinals ( which are simply natural numbers ) , ordered pairs , relations with a given domain , etc . Foundation ( the well-foundedness of the membership relation ) must also
|
||
#19 '' The statement Builds B C s l constrains element l of sequence s , namely app s l. We have either B ( app s l ) , or C ( app s l ) ( app s m ) ( app s n ) ) where m∈l and n∈l . For the natural numbers , set membership coincides with the less-than relation . Therefore , we are referring to a sequence s and element sl where either the base case B ( sl ) holds , or else the recursive step C ( sl , sm , sn ) for m , n < l. The statement BuildSeq B C s k y states that the sequence s has been constructed in this way right up to the value app s k , or in other words , sk , where y = sk . To formalise the basis for this approach requires a series of definitions in the HF calculus , introducing the subset relation , ordinals ( which are simply natural numbers ) , ordered pairs , relations with a given domain , etc . Foundation ( the well-foundedness of the membership relation ) must also be proved , which in turn requires additional definitions . A few highlights are shown below . The subset relation is defined , with infix syntax SUBS , with the help of All2 , the bounded universal quantifier . nominal primrec Subset : : `` tm ⇒ tm ⇒ fm '' ( infixr `` SUBS '' 150 ) where `` atom z ♯ ( t , u ) =⇒ t SUBS u = All2 z t ( ( Var z ) IN u ) '' In standard notation , this says t ⊆ u = ( ∀z ∈ t ) [ z ∈ u ] . The definition uses nominal primrec , even though it is not recursive , because it requires z to be fresh with respect to the terms t and u , among other nominal-related technicalities . Extensionality is taken as an axiom in traditional set theories , but in HF it can be proved by induction . However , many straightforward properties of the subset relation must first be derived . lemma Extensionality : `` H ⊢ x EQ y IFF x SUBS y AND y SUBS x '' Ordinals will be familiar to set theorists . The definition is the usual one , and shown below mainly as an example of a slightly more complicated HF formula . Two variables , y and z , must be fresh for each other and x. nominal primrec OrdP : : `` tm ⇒ fm '' where `` [ [ atom y ♯ ( x , z ) ; atom z ♯ x ] ] =⇒ OrdP x = All2 y x ( ( Var y ) SUBS x AND All2 z ( Var y ) ( ( Var z ) SUBS ( Var y ) ) )
|
||
#20 [ z ∈ u ] . The definition uses nominal primrec , even though it is not recursive , because it requires z to be fresh with respect to the terms t and u , among other nominal-related technicalities . Extensionality is taken as an axiom in traditional set theories , but in HF it can be proved by induction . However , many straightforward properties of the subset relation must first be derived . lemma Extensionality : `` H ⊢ x EQ y IFF x SUBS y AND y SUBS x '' Ordinals will be familiar to set theorists . The definition is the usual one , and shown below mainly as an example of a slightly more complicated HF formula . Two variables , y and z , must be fresh for each other and x. nominal primrec OrdP : : `` tm ⇒ fm '' where `` [ [ atom y ♯ ( x , z ) ; atom z ♯ x ] ] =⇒ OrdP x = All2 y x ( ( Var y ) SUBS x AND All2 z ( Var y ) ( ( Var z ) SUBS ( Var y ) ) ) '' The formal definition of a function ( as a single-valued set of pairs ) is subject to several complications . As we shall see in Sect . 3.3 below , all definitions must use Σ formulas , which requires certain non-standard formulations . In particular , x 6= y is not a Σ formula in general , but it can be expressed as x < y∨y < x if x and y are ordinals . The following primitive is used extensively when coding the syntax of HF within itself . 11 nominal primrec LstSeqP : : `` tm ⇒ tm ⇒ tm ⇒ fm '' where '' LstSeqP s k y = OrdP k AND HDomain Incl s ( SUCC k ) AND HFun Sigma s AND HPair k y IN s '' Informally , LstSeqP s k y means that s is a non-empty sequence whose domain includes the set { 0 , . . . , k } ( which is the ordinal k + 1 : the sequence is at least that long ) . Moreover , y = sk ; that would be written hk , yi ∈ s in the metalanguage , but becomes HPair k y IN s in the HF calculus , as seen above . Swierczkowski [ 32 ] prefers slightly different definitions , ´ specifying the domain to be exactly k , where k > 0 and y = sk−1 . The definition shown above simplifies the proof of the first incompleteness theorem , but complicates the proof of the second , in particular because they allow a sequence to be longer than necessary . This part of the development consists mainly of proofs in the HF calculus , and is nearly 1,300 lines long . 3.3 Σ Formulas and
|
||
#21 tm ⇒ tm ⇒ fm '' where '' LstSeqP s k y = OrdP k AND HDomain Incl s ( SUCC k ) AND HFun Sigma s AND HPair k y IN s '' Informally , LstSeqP s k y means that s is a non-empty sequence whose domain includes the set { 0 , . . . , k } ( which is the ordinal k + 1 : the sequence is at least that long ) . Moreover , y = sk ; that would be written hk , yi ∈ s in the metalanguage , but becomes HPair k y IN s in the HF calculus , as seen above . Swierczkowski [ 32 ] prefers slightly different definitions , ´ specifying the domain to be exactly k , where k > 0 and y = sk−1 . The definition shown above simplifies the proof of the first incompleteness theorem , but complicates the proof of the second , in particular because they allow a sequence to be longer than necessary . This part of the development consists mainly of proofs in the HF calculus , and is nearly 1,300 lines long . 3.3 Σ Formulas and Provability G¨odel had the foresight to recognise the value of minimising the need to write explicit formal proofs , without relying on the assumption that certain proofs could “ obviously ” be formalised . Instead , he developed enough meta-theory to prove that these proofs existed . One approach for this [ 2 , 32 ] relies on the concept of Σ formulas . These are inductively defined to include all formulas of the form t ∈ u , t = u , α ∨ β , α ∧ β , ∃x α and ( ∀x ∈ t ) α . ( These are closely related to the Σ1 formulas of the arithmetical hierarchy . ) It follows by induction on this construction that every true Σ sentence has a formal proof . Intuitively , the reasoning is that the atomic cases can be calculated , the Boolean cases can be done recursively , and the bounded universal quantifier can be replaced by a finite conjunction . The existential case holds because the semantics of ∃x α yields a specific witnessing value , again allowing an appeal to the induction hypothesis . The Σ formula approach is a good fit to the sort of formulas used in the coding of syntax . In these formulas , universal quantifiers have simple upper bounds , typically a variable giving the length of a sequence , while existential variables are unbounded . G¨odel ’ s original proofs required all quantifiers to be bounded . Existential quantifiers were bounded by complicated expressions requiring deep and difficult arithmetic justifications . Boolos presents similar material in a more modern form [ 2 , p. 41 ] . Relying exclusively on Σ formulas avoids these complications , but instead some straightforward properties have to be proven formally
|
||
#22 . ( These are closely related to the Σ1 formulas of the arithmetical hierarchy . ) It follows by induction on this construction that every true Σ sentence has a formal proof . Intuitively , the reasoning is that the atomic cases can be calculated , the Boolean cases can be done recursively , and the bounded universal quantifier can be replaced by a finite conjunction . The existential case holds because the semantics of ∃x α yields a specific witnessing value , again allowing an appeal to the induction hypothesis . The Σ formula approach is a good fit to the sort of formulas used in the coding of syntax . In these formulas , universal quantifiers have simple upper bounds , typically a variable giving the length of a sequence , while existential variables are unbounded . G¨odel ’ s original proofs required all quantifiers to be bounded . Existential quantifiers were bounded by complicated expressions requiring deep and difficult arithmetic justifications . Boolos presents similar material in a more modern form [ 2 , p. 41 ] . Relying exclusively on Σ formulas avoids these complications , but instead some straightforward properties have to be proven formally in the HF calculus . A complication is that proving the second incompleteness theorem requires another induction over Σ formulas . To minimise that proof effort , it helps to use as restrictive a definition as possible . The strict Σ formulas consist of all formulas of the form x ∈ y , α ∨ β , α ∧ β , ∃x α and ( ∀x ∈ y ) α . Here , x and y are not general terms , but variables . We further stipulate y not free in α in ( ∀x ∈ y ) α ; then in the main induction leading to the second incompleteness theorem , Case 2 of Lemma 9.7 [ 32 ] can be omitted . inductive ss fm : : `` fm ⇒ bool '' where MemI : `` ss fm ( Var i IN Var j ) '' | DisjI : `` ss fm A =⇒ ss fm B =⇒ ss fm ( A OR B ) '' | ConjI : `` ss fm A =⇒ ss fm B =⇒ ss fm ( A AND B ) '' | ExI : `` ss fm A =⇒ ss fm ( Ex i A ) '' | All2I : `` ss fm A =⇒ atom j ♯ ( i , A ) =⇒ ss fm ( All2 i ( Var j ) A ) '' 12 Now , a Σ formula is by definition one that is provably equivalent ( in HF ) to some strict Σ formula containing no additional free variables . In another minor oversight , Swierczkowski omits the free variable condition , but it is n ´ ecessary . definition Sigma fm : : `` fm ⇒ bool '' where `` Sigma fm A ←→ ( ∃ B.
|
||
#23 then in the main induction leading to the second incompleteness theorem , Case 2 of Lemma 9.7 [ 32 ] can be omitted . inductive ss fm : : `` fm ⇒ bool '' where MemI : `` ss fm ( Var i IN Var j ) '' | DisjI : `` ss fm A =⇒ ss fm B =⇒ ss fm ( A OR B ) '' | ConjI : `` ss fm A =⇒ ss fm B =⇒ ss fm ( A AND B ) '' | ExI : `` ss fm A =⇒ ss fm ( Ex i A ) '' | All2I : `` ss fm A =⇒ atom j ♯ ( i , A ) =⇒ ss fm ( All2 i ( Var j ) A ) '' 12 Now , a Σ formula is by definition one that is provably equivalent ( in HF ) to some strict Σ formula containing no additional free variables . In another minor oversight , Swierczkowski omits the free variable condition , but it is n ´ ecessary . definition Sigma fm : : `` fm ⇒ bool '' where `` Sigma fm A ←→ ( ∃ B. ss fm B ∧ supp B ⊆ supp A ∧ { } ⊢ A IFF B ) '' As always , Swierczkowski ’ s exposition is valuable , but far from compl ´ ete . Showing that Σ formulas include t ∈ u , t = u and ( ∀x ∈ t ) α for all terms t and u ( and not only for variables ) is far from obvious . These necessary facts are not even stated clearly . A crucial insight is to focus on proving that t ∈ u and t ⊆ u are Σ formulas . Consideration of the cases t ∈ 0 , t ∈ u1 ⊳ u2 , 0 ⊆ u , t1 ⊳ t2 ⊆ u shows that each reduces to false , true or a combination of simpler uses of ∈ or ⊆ . This observation suggests proving that t ∈ u and t ⊆ u are Σ formulas by mutual induction on the combined sizes of t and u. lemma Subset Mem sf lemma : '' size t + size u < n =⇒ Sigma fm ( t SUBS u ) ∧ Sigma fm ( t IN u ) '' The identical argument turns out to be needed for the second incompleteness theorem itself , formalised this time within the HF calculus . This coincidence should not be that surprising , as it is known that the second theorem could in principle be shown by formalising the first theorem within its own calculus . Now that we have taken care of t ⊆ u , proving that t = u is a Σ formula is trivial by extensionality , and the one remaining objective is ( ∀x ∈ t ) α . But with equality available , we can reduce
|
||
#24 . Consideration of the cases t ∈ 0 , t ∈ u1 ⊳ u2 , 0 ⊆ u , t1 ⊳ t2 ⊆ u shows that each reduces to false , true or a combination of simpler uses of ∈ or ⊆ . This observation suggests proving that t ∈ u and t ⊆ u are Σ formulas by mutual induction on the combined sizes of t and u. lemma Subset Mem sf lemma : '' size t + size u < n =⇒ Sigma fm ( t SUBS u ) ∧ Sigma fm ( t IN u ) '' The identical argument turns out to be needed for the second incompleteness theorem itself , formalised this time within the HF calculus . This coincidence should not be that surprising , as it is known that the second theorem could in principle be shown by formalising the first theorem within its own calculus . Now that we have taken care of t ⊆ u , proving that t = u is a Σ formula is trivial by extensionality , and the one remaining objective is ( ∀x ∈ t ) α . But with equality available , we can reduce this case to the strict Σ formula ( ∀x ∈ y ) α with the help of a lemma : lemma All2 term Iff : `` atom i ♯ t =⇒ atom j ♯ ( i , t , A ) =⇒ { } ⊢ ( All2 i t A ) IFF Ex j ( Var j EQ t AND All2 i ( Var j ) A ) '' This is simply ( ∀x ∈ t ) A ↔ ∃y [ y = t∧ ( ∀x ∈ y ) A ] expressed in the HF calculus , where it is easily proved . We could prove that ( ∀x ∈ t ) α is a Σ formula by induction on t , but this approach leads to complications . Virtually all predicates defined for the G¨odel development are carefully designed to take the form of Σ formulas . Here are two examples ; most such lemmas hold immediately by the construction of the given formula . lemma Subset sf : `` Sigma fm ( t SUBS u ) '' lemma LstSeqP sf : `` Sigma fm ( LstSeqP t u v ) '' The next milestone asserts that if α is a ground Σ formula ( and therefore a sentence ) and α evaluates to true , then α is a theorem . The proof is by induction on the size of the formula , and then by case analysis on its outer form . The case t ∈ u falls to a mutual induction with t ⊆ u resembling the one shown above . The case ( ∀x ∈ t ) α is effectively expanded to a conjunction . theorem Sigma fm imp thm : `` [ [ Sigma fm α ; ground fm α ; eval fm e0
|
||
#25 easily proved . We could prove that ( ∀x ∈ t ) α is a Σ formula by induction on t , but this approach leads to complications . Virtually all predicates defined for the G¨odel development are carefully designed to take the form of Σ formulas . Here are two examples ; most such lemmas hold immediately by the construction of the given formula . lemma Subset sf : `` Sigma fm ( t SUBS u ) '' lemma LstSeqP sf : `` Sigma fm ( LstSeqP t u v ) '' The next milestone asserts that if α is a ground Σ formula ( and therefore a sentence ) and α evaluates to true , then α is a theorem . The proof is by induction on the size of the formula , and then by case analysis on its outer form . The case t ∈ u falls to a mutual induction with t ⊆ u resembling the one shown above . The case ( ∀x ∈ t ) α is effectively expanded to a conjunction . theorem Sigma fm imp thm : `` [ [ Sigma fm α ; ground fm α ; eval fm e0 α ] ] =⇒ { } ⊢ α '' Every true Σ sentence is a theorem . This crucial meta-theoretic result is used eight times in the development . Without it , gigantic explicit HF proofs would be necessary . 4 Coding Provability in HF Within Itself The key insight leading to the proof of G¨odel ’ s theorem is that a sufficiently strong logical calculus can represent its syntax within itself , and in particular , the property of a given formula being provable . This task divides into three parts : coding the syntax , defining predicates to describe the coding and finally , defining predicates to describe the inference system . 13 4.1 Coding Terms , Formulas , Abstraction and Substitution In advocating the use of HF over PA , Swierczkowski begins by emphasising the ease of ´ coding syntax : It is at hand to code the variables x1 , x2 , . . . simply by the ordinals 1 , 2 , . . .. The constant 0 can be coded as 0 , and the remaining 6 symbols as n-tuples of 0s , say ∈ as h0 , 0i , etc . And here ends the arbitrariness of coding , which is so unpleasant when languages are arithmetized . [ 32 , p. 5 ] The adequacy of these definitions is easy to prove in HF itself . The full list is as follows : p0q = 0 , pxiq = i + 1 , p∈q = h0 , 0i , p⊳q = h0 , 0 , 0i , p=q = h0 , 0 , 0 , 0i , p∨q = h0 , 0 , 0 , 0 , 0i , p¬q = h0 , 0 , 0 , 0 , 0 , 0i
|
||
#26 describe the coding and finally , defining predicates to describe the inference system . 13 4.1 Coding Terms , Formulas , Abstraction and Substitution In advocating the use of HF over PA , Swierczkowski begins by emphasising the ease of ´ coding syntax : It is at hand to code the variables x1 , x2 , . . . simply by the ordinals 1 , 2 , . . .. The constant 0 can be coded as 0 , and the remaining 6 symbols as n-tuples of 0s , say ∈ as h0 , 0i , etc . And here ends the arbitrariness of coding , which is so unpleasant when languages are arithmetized . [ 32 , p. 5 ] The adequacy of these definitions is easy to prove in HF itself . The full list is as follows : p0q = 0 , pxiq = i + 1 , p∈q = h0 , 0i , p⊳q = h0 , 0 , 0i , p=q = h0 , 0 , 0 , 0i , p∨q = h0 , 0 , 0 , 0 , 0i , p¬q = h0 , 0 , 0 , 0 , 0 , 0i , p∃q = h0 , 0 , 0 , 0 , 0 , 0 , 0i . We have a few differences from Swierczkowski : ´ pxiq = i + 1 because our variables start at zero , and for the kth de Bruijn index we use hh0 , 0 , 0 , 0 , 0 , 0 , 0 , 0i , ki . Obviously ∈ means nothing by itself , so p∈q = h0 , 0i really means pt ∈ uq = hh0 , 0i , ptq , puqi , etc . Note that nests of n-tuples terminated by ordinals can be decomposed uniquely . De Bruijn equivalents of terms and formulas are then declared . To repeat : de Bruijn syntax is used for coding , for which it is ideal , allowing the simplest possible definitions of abstraction and substitution . Although it destroys readability , encodings are never readable anyway . Using nominal here is out of the question . The entire theory of nominal Isabelle would need to be formalised within the embedded calculus . Quite apart from the work involved , the necessary equivalence classes would be infinite sets , which are not available in HF . The strongest argument for HF is that the mathematical basis of its coding scheme is simply ordered pairs defined in the standard set-theoretic way . An elementary formal argument justifies this . In contrast , the usual arithmetic encoding relies on either the Chinese remainder theorem or unique prime factorisation . This fragment of number theory would have to be formalised within the embedded calculus in order to reason about encoded formulas , which is necessary to prove the second incompleteness theorem . It must be emphasised that proving anything in the calculus ( where such
|
||
#27 by ordinals can be decomposed uniquely . De Bruijn equivalents of terms and formulas are then declared . To repeat : de Bruijn syntax is used for coding , for which it is ideal , allowing the simplest possible definitions of abstraction and substitution . Although it destroys readability , encodings are never readable anyway . Using nominal here is out of the question . The entire theory of nominal Isabelle would need to be formalised within the embedded calculus . Quite apart from the work involved , the necessary equivalence classes would be infinite sets , which are not available in HF . The strongest argument for HF is that the mathematical basis of its coding scheme is simply ordered pairs defined in the standard set-theoretic way . An elementary formal argument justifies this . In contrast , the usual arithmetic encoding relies on either the Chinese remainder theorem or unique prime factorisation . This fragment of number theory would have to be formalised within the embedded calculus in order to reason about encoded formulas , which is necessary to prove the second incompleteness theorem . It must be emphasised that proving anything in the calculus ( where such luxuries as a simplifier , recursion and even function symbols are not available ) is much more difficult than proving the same result in a proof assistant . 4.1.1 Introducing de Bruijn Terms and Formulas De Bruijn terms resemble the type tm declared in Sect . 2.3 , but include the DBInd constructor for bound variable indices as well as the DBVar constructor for free variables . nominal datatype dbtm = DBZero | DBVar name | DBInd nat | DBEats dbtm dbtm De Bruijn formula contructors involve no explicit variable binding , creating an apparent similarity between DBNeg and DBEx , although the latter creates an implicit variable binding scope . nominal datatype dbfm = DBMem dbtm dbtm | DBEq dbtm dbtm | DBDisj dbfm dbfm | DBNeg dbfm | DBEx dbfm 14 How this works should become clear as we consider how terms and formulas are translated into their de Bruijn equivalents . To begin with , we need a lookup function taking a list of names ( representing variables bound in the current context , innermost first ) and a name to be looked up . The integer n , initially 0 , is the index to substitute if the name is next in the list . fun lookup : : `` name list ⇒ nat ⇒ name ⇒ dbtm '' where '' lookup [ ] n x = DBVar x '' | `` lookup ( y # ys ) n x = ( if x = y then DBInd n else ( lookup ys ( Suc n ) x ) ) '' To translate a term into de Bruijn format , the key step is to resolve name references using lookup . Names bound in the local environment are replaced by the corresponding indices , while other
|
||
#28 , although the latter creates an implicit variable binding scope . nominal datatype dbfm = DBMem dbtm dbtm | DBEq dbtm dbtm | DBDisj dbfm dbfm | DBNeg dbfm | DBEx dbfm 14 How this works should become clear as we consider how terms and formulas are translated into their de Bruijn equivalents . To begin with , we need a lookup function taking a list of names ( representing variables bound in the current context , innermost first ) and a name to be looked up . The integer n , initially 0 , is the index to substitute if the name is next in the list . fun lookup : : `` name list ⇒ nat ⇒ name ⇒ dbtm '' where '' lookup [ ] n x = DBVar x '' | `` lookup ( y # ys ) n x = ( if x = y then DBInd n else ( lookup ys ( Suc n ) x ) ) '' To translate a term into de Bruijn format , the key step is to resolve name references using lookup . Names bound in the local environment are replaced by the corresponding indices , while other names are left as they were . nominal primrec trans tm : : `` name list ⇒ tm ⇒ dbtm '' where '' trans tm e Zero = DBZero '' | `` trans tm e ( Var k ) = lookup e 0 k '' | `` trans tm e ( Eats t u ) = DBEats ( trans tm e t ) ( trans tm e u ) '' Noteworthy is the final case of trans fm , which requires the bound variable k in the quantified formula Ex k A to be fresh with respect to e , our list of previouslyencountered bound variables . In the recursive call , k is added to the list , which therefore consists of distinct names . nominal primrec trans fm : : `` name list ⇒ fm ⇒ dbfm '' where '' trans fm e ( Mem t u ) = DBMem ( trans tm e t ) ( trans tm e u ) '' | `` trans fm e ( Eq t u ) = DBEq ( trans tm e t ) ( trans tm e u ) '' | `` trans fm e ( Disj A B ) = DBDisj ( trans fm e A ) ( trans fm e B ) '' | `` trans fm e ( Neg A ) = DBNeg ( trans fm e A ) '' | `` atom k ♯ e =⇒ trans fm e ( Ex k A ) = DBEx ( trans fm ( k # e ) A ) '' Syntactic operations for de Bruijn notation tend to be straightforward , as there are no bound variable names that might clash . Comparisons with previous formalisations of the λ-calculus may be illuminating , but the usual lifting operation [
|
||
#29 our list of previouslyencountered bound variables . In the recursive call , k is added to the list , which therefore consists of distinct names . nominal primrec trans fm : : `` name list ⇒ fm ⇒ dbfm '' where '' trans fm e ( Mem t u ) = DBMem ( trans tm e t ) ( trans tm e u ) '' | `` trans fm e ( Eq t u ) = DBEq ( trans tm e t ) ( trans tm e u ) '' | `` trans fm e ( Disj A B ) = DBDisj ( trans fm e A ) ( trans fm e B ) '' | `` trans fm e ( Neg A ) = DBNeg ( trans fm e A ) '' | `` atom k ♯ e =⇒ trans fm e ( Ex k A ) = DBEx ( trans fm ( k # e ) A ) '' Syntactic operations for de Bruijn notation tend to be straightforward , as there are no bound variable names that might clash . Comparisons with previous formalisations of the λ-calculus may be illuminating , but the usual lifting operation [ 18 , 21 ] is unnecessary . That is because the HF calculus does not allow reductions anywhere , as in the λ-calculus . Substitutions only happen at the top level and never within deeper bound variable contexts . For us , substitution is the usual operation of replacing a variable by a term , which contains no bound variables . ( Substitution can alternatively be defined to replace a de Bruijn index by a term . ) The special de Bruijn operation is abstraction . This replaces every occurrence of a given free variable in a term or formula by a de Bruijn index , in preparation for binding . For example , abstracting the formula DBMem ( DBVar x ) ( DBVar y ) over the variable y yields DBMem ( DBVar x ) ( DBInd 0 ) . This is actually ill-formed , but attaching a quantifier yields the de Bruijn formula DBEx ( DBMem ( DBVar x ) ( DBInd 0 ) ) , representing ∃y [ x ∈ y ] . Abstracting this over the free variable x and attaching another quantifier yields DBEx ( DBEx ( DBMem ( DBInd 1 ) ( DBInd 0 ) ) ) , which is the formula ∃xy [ x ∈ y ] . An index of 1 has been substituted in order to skip over the inner binder . 15 4.1.2 Well-Formed de Bruijn Terms and Formulas With the de Bruijn approach , an index of 0 designates the innermost enclosing binder , while an index of 1 designates the next-innermost binder , etc . ( Here , the only binder is DBEx . ) If every index has a matching binder ( the index i must be nested within at least i + 1 binders ) ,
|
||
#30 by a de Bruijn index , in preparation for binding . For example , abstracting the formula DBMem ( DBVar x ) ( DBVar y ) over the variable y yields DBMem ( DBVar x ) ( DBInd 0 ) . This is actually ill-formed , but attaching a quantifier yields the de Bruijn formula DBEx ( DBMem ( DBVar x ) ( DBInd 0 ) ) , representing ∃y [ x ∈ y ] . Abstracting this over the free variable x and attaching another quantifier yields DBEx ( DBEx ( DBMem ( DBInd 1 ) ( DBInd 0 ) ) ) , which is the formula ∃xy [ x ∈ y ] . An index of 1 has been substituted in order to skip over the inner binder . 15 4.1.2 Well-Formed de Bruijn Terms and Formulas With the de Bruijn approach , an index of 0 designates the innermost enclosing binder , while an index of 1 designates the next-innermost binder , etc . ( Here , the only binder is DBEx . ) If every index has a matching binder ( the index i must be nested within at least i + 1 binders ) , then the term or formula is well-formed . Recall the examples of abstraction above , where a binder must be attached afterwards . In particular , as our terms do not contain any binding constructs , a well-formed term may contain no de Bruijn indices . In contrast to more traditional notions of logical syntax , if you take a well-formed formula and view one of its subformulas or subterms in isolation , it will not necessarily be well-formed . The situation is analogous to extracting a fragment of a program , removing it from necessary enclosing declarations . The property of being a well-formed de Bruijn term or formula is defined inductively . The syntactic predicates defined below recognise such well-formed formulas . A well-formed de Bruijn term has no indices ( DBInd ) at all : inductive wf dbtm : : `` dbtm ⇒ bool '' where Zero : `` wf dbtm DBZero '' | Var : `` wf dbtm ( DBVar name ) '' | Eats : `` wf dbtm t1 =⇒ wf dbtm t2 =⇒ wf dbtm ( DBEats t1 t2 ) '' A trivial induction shows that for every well-founded de Bruijn term there is an equivalent standard term . The only cases to be considered ( as per the definition above ) are Zero , Var and Eats . lemma wf dbtm imp is tm : assumes `` wf dbtm x '' shows `` ∃ t : :tm . x = trans tm [ ] t '' A well-formed de Bruijn formula is constructed from other well-formed terms and formulas , and indices can only be introduced by applying abstraction ( abst dbfm ) over a given name to another well-formed formula , in the existential case . Specifically , the Ex clause below
|
||
#31 property of being a well-formed de Bruijn term or formula is defined inductively . The syntactic predicates defined below recognise such well-formed formulas . A well-formed de Bruijn term has no indices ( DBInd ) at all : inductive wf dbtm : : `` dbtm ⇒ bool '' where Zero : `` wf dbtm DBZero '' | Var : `` wf dbtm ( DBVar name ) '' | Eats : `` wf dbtm t1 =⇒ wf dbtm t2 =⇒ wf dbtm ( DBEats t1 t2 ) '' A trivial induction shows that for every well-founded de Bruijn term there is an equivalent standard term . The only cases to be considered ( as per the definition above ) are Zero , Var and Eats . lemma wf dbtm imp is tm : assumes `` wf dbtm x '' shows `` ∃ t : :tm . x = trans tm [ ] t '' A well-formed de Bruijn formula is constructed from other well-formed terms and formulas , and indices can only be introduced by applying abstraction ( abst dbfm ) over a given name to another well-formed formula , in the existential case . Specifically , the Ex clause below states that , starting with a well-formed formula A , abstracting over some name and applying DBEx to the result yields another well-formed formula . inductive wf dbfm : : `` dbfm ⇒ bool '' where Mem : `` wf dbtm t1 =⇒ wf dbtm t2 =⇒ wf dbfm ( DBMem t1 t2 ) '' | Eq : `` wf dbtm t1 =⇒ wf dbtm t2 =⇒ wf dbfm ( DBEq t1 t2 ) '' | Disj : `` wf dbfm A1 =⇒ wf dbfm A2 =⇒ wf dbfm ( DBDisj A1 A2 ) '' | Neg : `` wf dbfm A =⇒ wf dbfm ( DBNeg A ) '' | Ex : `` wf dbfm A =⇒ wf dbfm ( DBEx ( abst dbfm name 0 A ) ) '' This definition formalises the allowed forms of construction , rather than stating explicitly that every index must have a matching binder . A refinement must be mentioned . Strong nominal induction ( already seen above , Sect . 2.3 ) formalises the assumption that bound variables revealed by induction can be assumed not to clash with other variables . This is set up automatically for nominal datatypes , but here requires a manual step . The command nominal inductive sets up strong induction for name in the Ex case of the inductive definition above ; we must prove that name is not significant according to the nominal theory , and then get to assume that name will not clash . This step ( details omitted ) seems to be necessary in order to complete some inductive proofs about wf dbfm . 16 4.1.3 Quoting Terms and Formulas It is essential to remember that G¨odel encodings are terms ( having type tm ) , not sets or numbers . Textbook
|
||
#32 dbfm A =⇒ wf dbfm ( DBNeg A ) '' | Ex : `` wf dbfm A =⇒ wf dbfm ( DBEx ( abst dbfm name 0 A ) ) '' This definition formalises the allowed forms of construction , rather than stating explicitly that every index must have a matching binder . A refinement must be mentioned . Strong nominal induction ( already seen above , Sect . 2.3 ) formalises the assumption that bound variables revealed by induction can be assumed not to clash with other variables . This is set up automatically for nominal datatypes , but here requires a manual step . The command nominal inductive sets up strong induction for name in the Ex case of the inductive definition above ; we must prove that name is not significant according to the nominal theory , and then get to assume that name will not clash . This step ( details omitted ) seems to be necessary in order to complete some inductive proofs about wf dbfm . 16 4.1.3 Quoting Terms and Formulas It is essential to remember that G¨odel encodings are terms ( having type tm ) , not sets or numbers . Textbook presentations identify terms with their denotations for the sake of clarity , but this can be confusing . The undecidable formula contains an encoding of itself in the form of a term . First , we must define codes for de Bruijn terms and formulas . function quot dbtm : : `` dbtm ⇒ tm '' where '' quot dbtm DBZero = Zero '' | `` quot dbtm ( DBVar name ) = ORD OF ( Suc ( nat of name name ) ) '' | `` quot dbtm ( DBInd k ) = HPair ( HTuple 6 ) ( ORD OF k ) '' | `` quot dbtm ( DBEats t u ) = HPair ( HTuple 1 ) ( HPair ( quot dbtm t ) ( quot dbtm u ) ) '' The codes of real terms and formulas ( for which we set up the overloaded ⌈· · ·⌉ syntax ) are obtained by first translating them to their de Bruijn equivalents and then encoding . We finally obtain facts such as the following : lemma quot Zero : `` ⌈Zero⌉ = Zero '' lemma quot Var : `` ⌈Var x⌉ = SUCC ( ORD OF ( nat of name x ) ) '' lemma quot Eats : `` ⌈Eats x y⌉ = HPair ( HTuple 1 ) ( HPair ⌈x⌉ ⌈y⌉ ) '' lemma quot Eq : `` ⌈x EQ y⌉ = HPair ( HTuple 2 ) ( HPair ( ⌈x⌉ ) ( ⌈y⌉ ) ) '' lemma quot Disj : `` ⌈A OR B⌉ = HPair ( HTuple 3 ) ( HPair ( ⌈A⌉ ) ( ⌈B⌉ ) ) '' lemma quot Ex : `` ⌈Ex i A⌉ = HPair ( HTuple 5 ) ( quot dbfm ( trans fm [ i ] A )
|
||
#33 ORD OF k ) '' | `` quot dbtm ( DBEats t u ) = HPair ( HTuple 1 ) ( HPair ( quot dbtm t ) ( quot dbtm u ) ) '' The codes of real terms and formulas ( for which we set up the overloaded ⌈· · ·⌉ syntax ) are obtained by first translating them to their de Bruijn equivalents and then encoding . We finally obtain facts such as the following : lemma quot Zero : `` ⌈Zero⌉ = Zero '' lemma quot Var : `` ⌈Var x⌉ = SUCC ( ORD OF ( nat of name x ) ) '' lemma quot Eats : `` ⌈Eats x y⌉ = HPair ( HTuple 1 ) ( HPair ⌈x⌉ ⌈y⌉ ) '' lemma quot Eq : `` ⌈x EQ y⌉ = HPair ( HTuple 2 ) ( HPair ( ⌈x⌉ ) ( ⌈y⌉ ) ) '' lemma quot Disj : `` ⌈A OR B⌉ = HPair ( HTuple 3 ) ( HPair ( ⌈A⌉ ) ( ⌈B⌉ ) ) '' lemma quot Ex : `` ⌈Ex i A⌉ = HPair ( HTuple 5 ) ( quot dbfm ( trans fm [ i ] A ) ) '' Note that HPair constructs an HF term denoting a pair , while HTuple n constructs an ( n+ 2 ) -tuple of zeros . Proofs often refer to the denotations of terms rather than to the terms themselves , so the functions q Eats , q Mem , q Eq , q Neg , q Disj , q Ex are defined to express these codes . Here are some examples : '' q Var i ≡ succ ( ord of ( nat of name i ) ) '' '' q Eats x y ≡ hhtuple 1 , x , yi '' '' q Disj x y ≡ hhtuple 3 , x , yi '' '' q Ex x ≡ hhtuple 5 , xi '' Note that hx , yi denotes the pair of x and y as sets , in other words , of type hf . 4.2 Predicates for the Coding of Syntax The next and most arduous step is to define logical predicates corresponding to each of the syntactic concepts ( terms , formulas , abstraction , substitution ) mentioned above . Textbooks and articles describe each predicate at varying levels of detail . G¨odel [ 8 ] gives full definitions , as does Swierczkowski . Boolos [ 2 ] gives many details of how coding i ´ s set up , and gives the predicates for terms and formulas , but not for any operations upon them . Hodel [ 13 ] , like many textbook authors , relies heavily on “ algorithms ” written in English . The definitions indeed amount to pages of computer code . Authors typically conclude with a “ theorem ” asserting the correctness of this code . For example , Swierczkowski [ 32 , Sect . 3–4 ] presents
|
||
#34 , yi '' '' q Disj x y ≡ hhtuple 3 , x , yi '' '' q Ex x ≡ hhtuple 5 , xi '' Note that hx , yi denotes the pair of x and y as sets , in other words , of type hf . 4.2 Predicates for the Coding of Syntax The next and most arduous step is to define logical predicates corresponding to each of the syntactic concepts ( terms , formulas , abstraction , substitution ) mentioned above . Textbooks and articles describe each predicate at varying levels of detail . G¨odel [ 8 ] gives full definitions , as does Swierczkowski . Boolos [ 2 ] gives many details of how coding i ´ s set up , and gives the predicates for terms and formulas , but not for any operations upon them . Hodel [ 13 ] , like many textbook authors , relies heavily on “ algorithms ” written in English . The definitions indeed amount to pages of computer code . Authors typically conclude with a “ theorem ” asserting the correctness of this code . For example , Swierczkowski [ 32 , Sect . 3–4 ] presents 34 highly technical ´ definitions , justified by seven lines of proof . Proving the correctness of this lengthy series of definitions requires a substantial effort , and the proofs ( being syntactically oriented ) are tiresome . It is helpful to introduce shadow versions of all predicates in Isabelle/HOL ’ s native logic , as well as in HF . Having two versions of each predicate simplifies the task of relating the HF version of 17 the predicate to the syntactic concept that it is intended to represent ; the first step is to prove that the HF formula is equivalent to the syntactically similar definition written in Isabelle ’ s higher-order logic , which then is proved to satisfy deeper properties . The shadow predicates also give an easy way to refer to the truth of the corresponding HF predicate ; because each is defined to be a Σ formula , that gives a quick way ( using theorem Sigma fm imp thm above ) to show that some ground instance of the predicate can be proved formally in HF . Also , one way to arrive at the correct definition of an HF predicate is to define its shadow equivalent first , since proving that it implies the required properties is much easier in Isabelle/HOL ’ s native logic than in HF . Swierczkowski [ 32 ] defines a full set of syntactic predicate ´ s , leaving nothing as an exercise . Unfortunately , the introduction of de Bruijn syntax necessitates rewriting many of these definitions . Some predicates ( such as the variable occurrence test ) are replaced by others ( abstraction over a variable occurrence ) . The final list includes predicates to recognise the following items : – the codes of well-formed terms
|
||
#35 is equivalent to the syntactically similar definition written in Isabelle ’ s higher-order logic , which then is proved to satisfy deeper properties . The shadow predicates also give an easy way to refer to the truth of the corresponding HF predicate ; because each is defined to be a Σ formula , that gives a quick way ( using theorem Sigma fm imp thm above ) to show that some ground instance of the predicate can be proved formally in HF . Also , one way to arrive at the correct definition of an HF predicate is to define its shadow equivalent first , since proving that it implies the required properties is much easier in Isabelle/HOL ’ s native logic than in HF . Swierczkowski [ 32 ] defines a full set of syntactic predicate ´ s , leaving nothing as an exercise . Unfortunately , the introduction of de Bruijn syntax necessitates rewriting many of these definitions . Some predicates ( such as the variable occurrence test ) are replaced by others ( abstraction over a variable occurrence ) . The final list includes predicates to recognise the following items : – the codes of well-formed terms ( and constant terms , without variables ) – correct instances of abstraction ( of a term or formula ) over a variable – correct instances of substitution ( in a term or formula ) for a variable – the codes of well-formed formulas As explained below , abstraction over a formula needs to be defined before the notion of a formula itself . We also need the property of variable non-occurrence , “ x does not occur in A ” , which can be expressed directly as “ substituting 0 for x in A yields A ” . This little trick eliminates the need for a full definition . Each operation is first defined in its sequence form ( expressing that sequence s is built up in an appropriate way and that sk is a specific value ) ; existential quantification over s and k then yields the final predicate . Formalising the sequence of steps is a primitive way to express recursion . Moreover , it tends to yield Σ formulas . The simplest example is the predicate for constants . The shadow predicate can be defined with the help of BuildSeq , mentioned in Sect . 3.2 above . Note that shadow predicates are written in ordinary higher-order logic , and refer to syntactic codes using set values . We see below that in the sequence buildup , each element is either 0 ( which is the code of the constant zero ) or else has the form q Eats v w , which is the code for v ⊳ w. definition SeqConst : : `` hf ⇒ hf ⇒ hf ⇒ bool '' where `` SeqConst s k t ≡ BuildSeq ( λu . u=0 ) ( λu v w. u = q Eats v w
|
||
#36 little trick eliminates the need for a full definition . Each operation is first defined in its sequence form ( expressing that sequence s is built up in an appropriate way and that sk is a specific value ) ; existential quantification over s and k then yields the final predicate . Formalising the sequence of steps is a primitive way to express recursion . Moreover , it tends to yield Σ formulas . The simplest example is the predicate for constants . The shadow predicate can be defined with the help of BuildSeq , mentioned in Sect . 3.2 above . Note that shadow predicates are written in ordinary higher-order logic , and refer to syntactic codes using set values . We see below that in the sequence buildup , each element is either 0 ( which is the code of the constant zero ) or else has the form q Eats v w , which is the code for v ⊳ w. definition SeqConst : : `` hf ⇒ hf ⇒ hf ⇒ bool '' where `` SeqConst s k t ≡ BuildSeq ( λu . u=0 ) ( λu v w. u = q Eats v w ) s k t '' Thus a constant expression is built up from 0 using the ⊳ operator . The idea is that every element of the sequence is either 0 or has the form px ⊳ yq , where x and y occur earlier in the sequence . Most of the other syntactic predicates fit exactly the same pattern , but with different base cases and constructors . A function must be coded as a relation , and a typical base case might be h0 , 0i , other sequence elements having the form hpx ⊳ yq , px ′ ⊳ y ′ qi , where hx , x′ i and hy , y′ i occur earlier in the sequence . Substitution is codified in this manner . A function taking two arguments is coded as a sequence of triples , etc . The discussion above relates to shadow predicates , which define formulas of Isabelle/HOL relating HF sets . The real predicates , which denote formulas of the HF calculus , are based on exactly the same ideas except that the various set constructions must be expressed using the HF term language . Note that the real predicates typically have names ending with P. 18 The following formula again specifies the notion of a constant term . It is simply the result of expressing the definition of SeqConst using HF syntax , expanding the definition of BuildSeq . The syntactic sugar for a reference to a sequence element sm within some formula φ must now be expanded to its true form : φ ( sm ) becomes ∃y [ hm , yi ∈ s∧φ ( y ) ] . nominal primrec SeqConstP : : `` tm ⇒ tm ⇒ tm ⇒ fm '' where `` [ [ atom
|
||
#37 ′ ⊳ y ′ qi , where hx , x′ i and hy , y′ i occur earlier in the sequence . Substitution is codified in this manner . A function taking two arguments is coded as a sequence of triples , etc . The discussion above relates to shadow predicates , which define formulas of Isabelle/HOL relating HF sets . The real predicates , which denote formulas of the HF calculus , are based on exactly the same ideas except that the various set constructions must be expressed using the HF term language . Note that the real predicates typically have names ending with P. 18 The following formula again specifies the notion of a constant term . It is simply the result of expressing the definition of SeqConst using HF syntax , expanding the definition of BuildSeq . The syntactic sugar for a reference to a sequence element sm within some formula φ must now be expanded to its true form : φ ( sm ) becomes ∃y [ hm , yi ∈ s∧φ ( y ) ] . nominal primrec SeqConstP : : `` tm ⇒ tm ⇒ tm ⇒ fm '' where `` [ [ atom l ♯ ( s , k , sl , m , n , sm , sn ) ; atom sl ♯ ( s , m , n , sm , sn ) ; atom m ♯ ( s , n , sm , sn ) ; atom n ♯ ( s , sm , sn ) ; atom sm ♯ ( s , sn ) ; atom sn ♯ ( s ) ] ] =⇒ SeqConstP s k t = LstSeqP s k t AND All2 l ( SUCC k ) ( Ex sl ( HPair ( Var l ) ( Var sl ) IN s AND ( Var sl EQ Zero OR Ex m ( Ex n ( Ex sm ( Ex sn ( Var m IN Var l AND Var n IN Var l AND HPair ( Var m ) ( Var sm ) IN s AND HPair ( Var n ) ( Var sn ) IN s AND Var sl EQ Q Eats ( Var sm ) ( Var sn ) ) ) ) ) ) ) ) '' We have five bound variables , namely l , sl , m , sm , n , sn , which must be constrained to be distinct from one another using the freshness conditions shown . This nominal boilerplate may seem excessive . However , to define this predicate without nominal syntax , bound variable names might have to be calculated , perhaps by taking the maximum of the bound variables in s , k and t and continuing from there . Nominal constrains the variables more abstractly and flexibly . As mentioned above , sometimes we deal with sequences of pairs or triples , with correspondingly more complicated formulas . Where a predicate describes a function such as
|
||
#38 ( Var sl ) IN s AND ( Var sl EQ Zero OR Ex m ( Ex n ( Ex sm ( Ex sn ( Var m IN Var l AND Var n IN Var l AND HPair ( Var m ) ( Var sm ) IN s AND HPair ( Var n ) ( Var sn ) IN s AND Var sl EQ Q Eats ( Var sm ) ( Var sn ) ) ) ) ) ) ) ) '' We have five bound variables , namely l , sl , m , sm , n , sn , which must be constrained to be distinct from one another using the freshness conditions shown . This nominal boilerplate may seem excessive . However , to define this predicate without nominal syntax , bound variable names might have to be calculated , perhaps by taking the maximum of the bound variables in s , k and t and continuing from there . Nominal constrains the variables more abstractly and flexibly . As mentioned above , sometimes we deal with sequences of pairs or triples , with correspondingly more complicated formulas . Where a predicate describes a function such as substitution , the sequence being built up consists of ordered pairs of arguments and results , and there are typically nine bound variables . To perform abstraction over a formula requires keeping track of the depth of quantifier nesting during recursion . This is a two-argument function , so the sequence being built up consists of ordered triples and there are 12 bound variables . Although the nominal system copes with these complicated expressions , the processing time can be measured in tens of seconds . Now that we have defined the buildup of a sequence of constants , a constant itself is trivial . The existence of any properly formed sequence s of length k culminating with some term t guarantees that t is a constant term . Here are both the shadow and HF calculus versions of the predicate . definition Const : : `` hf ⇒ bool '' where `` Const t ≡ ( ∃ s k. SeqConst s k t ) '' nominal primrec ConstP : : `` tm ⇒ fm '' where `` [ [ atom k ♯ ( s , t ) ; atom s ♯ t ] ] =⇒ ConstP t = Ex s ( Ex k ( SeqConstP ( Var s ) ( Var k ) t ) ) '' Why don ’ t we define the HF predicate BuildSeqP analogously to BuildSeq , which expresses the definition of SeqConst so succinctly ? Then we might expect to avoid the mess above , defining a predicate such as SeqConstP in a single line . This was actually attempted , but the nominal system is not really suitable for formalising higher-order definitions . Complicated auxiliary definitions and proofs are required . It is easier simply to write out the definitions , especially as
|
||
#39 constant itself is trivial . The existence of any properly formed sequence s of length k culminating with some term t guarantees that t is a constant term . Here are both the shadow and HF calculus versions of the predicate . definition Const : : `` hf ⇒ bool '' where `` Const t ≡ ( ∃ s k. SeqConst s k t ) '' nominal primrec ConstP : : `` tm ⇒ fm '' where `` [ [ atom k ♯ ( s , t ) ; atom s ♯ t ] ] =⇒ ConstP t = Ex s ( Ex k ( SeqConstP ( Var s ) ( Var k ) t ) ) '' Why don ’ t we define the HF predicate BuildSeqP analogously to BuildSeq , which expresses the definition of SeqConst so succinctly ? Then we might expect to avoid the mess above , defining a predicate such as SeqConstP in a single line . This was actually attempted , but the nominal system is not really suitable for formalising higher-order definitions . Complicated auxiliary definitions and proofs are required . It is easier simply to write out the definitions , especially as their very repetitiveness allows proof development by cut and paste . One tiny consolidation has been done . We need to define the predicates Term and TermP analogously to Const and ConstP above but allowing variables . The question of whether variables are allowed in a term or not can be governed by a Boolean . The proof development therefore introduces the predicate SeqCTermP , taking a Boolean argument , from which SeqTermP and SeqConstP are trivially obtained . 19 abbreviation SeqTermP : : `` tm ⇒ tm ⇒ tm ⇒ fm '' where `` SeqTermP ≡ SeqCTermP True '' abbreviation SeqConstP : : `` tm ⇒ tm ⇒ tm ⇒ fm '' where `` SeqConstP ≡ SeqCTermP False '' In this way , we can define the very similar predicates TermP and ConstP with a minimum of repeated material . Many other predicates must be defined . Abstraction and substitution must be defined separately for terms , atomic formulas and formulas . Here are the shadow definitions of abstraction and substitution for terms . They are similar enough ( both involve replacing a variable ) that a single function , SeqStTerm , can express both . BuildSeq2 is similar to BuildSeq above , but constructs a sequence of pairs . definition SeqStTerm : : `` hf ⇒ hf ⇒ hf ⇒ hf ⇒ hf ⇒ hf ⇒ bool '' where `` SeqStTerm v u x x ’ s k ≡ is Var v ∧ BuildSeq2 ( λy y ’ . ( is Ind y ∨ Ord y ) ∧ y ’ = ( if y=v then u else y ) ) ( λu u ’ v v ’ w w ’ . u = q Eats v w ∧ u ’ = q Eats v ’ w ’ )
|
||
#40 abbreviation SeqConstP : : `` tm ⇒ tm ⇒ tm ⇒ fm '' where `` SeqConstP ≡ SeqCTermP False '' In this way , we can define the very similar predicates TermP and ConstP with a minimum of repeated material . Many other predicates must be defined . Abstraction and substitution must be defined separately for terms , atomic formulas and formulas . Here are the shadow definitions of abstraction and substitution for terms . They are similar enough ( both involve replacing a variable ) that a single function , SeqStTerm , can express both . BuildSeq2 is similar to BuildSeq above , but constructs a sequence of pairs . definition SeqStTerm : : `` hf ⇒ hf ⇒ hf ⇒ hf ⇒ hf ⇒ hf ⇒ bool '' where `` SeqStTerm v u x x ’ s k ≡ is Var v ∧ BuildSeq2 ( λy y ’ . ( is Ind y ∨ Ord y ) ∧ y ’ = ( if y=v then u else y ) ) ( λu u ’ v v ’ w w ’ . u = q Eats v w ∧ u ’ = q Eats v ’ w ’ ) s k x x ’ '' definition AbstTerm : : `` hf ⇒ hf ⇒ hf ⇒ hf ⇒ bool '' where `` AbstTerm v i x x ’ ≡ Ord i ∧ ( ∃ s k. SeqStTerm v ( q Ind i ) x x ’ s k ) '' definition SubstTerm : : `` hf ⇒ hf ⇒ hf ⇒ hf ⇒ bool '' where `` SubstTerm v u x x ’ ≡ Term u ∧ ( ∃ s k. SeqStTerm v u x x ’ s k ) '' Abstraction over formulas ( AbstForm/AbstFormP ) must be defined before formulas themselves , in order to formalise existential quantification . There is no circularity here : the abstraction operation can be defined independently of the notion of a well-formed formula , and is not restricted to them . The definition involves sequences of triples and is too complicated to present here . With abstraction over formulas defined , we can finally define the concept of a formula itself . An Atomic formula involves two terms , combined using the relations EQ or IN . Then MakeForm combines one or two existing formulas to build more complex ones . It constrains y to be the code of a formula constructed from existing formulas u and v by the disjunction u ∨ v , the negation ¬u or the existential formula ∃ ( u ’ ) , where u ’ has been obtained by abstracting u over some variable , v via the predicate AbstForm . definition MakeForm : : `` hf ⇒ hf ⇒ hf ⇒ bool '' where `` MakeForm y u w ≡ y = q Disj u w ∨ y = q Neg u ∨ ( ∃ v u ’ . AbstForm v 0 u u
|
||
#41 defined before formulas themselves , in order to formalise existential quantification . There is no circularity here : the abstraction operation can be defined independently of the notion of a well-formed formula , and is not restricted to them . The definition involves sequences of triples and is too complicated to present here . With abstraction over formulas defined , we can finally define the concept of a formula itself . An Atomic formula involves two terms , combined using the relations EQ or IN . Then MakeForm combines one or two existing formulas to build more complex ones . It constrains y to be the code of a formula constructed from existing formulas u and v by the disjunction u ∨ v , the negation ¬u or the existential formula ∃ ( u ’ ) , where u ’ has been obtained by abstracting u over some variable , v via the predicate AbstForm . definition MakeForm : : `` hf ⇒ hf ⇒ hf ⇒ bool '' where `` MakeForm y u w ≡ y = q Disj u w ∨ y = q Neg u ∨ ( ∃ v u ’ . AbstForm v 0 u u ’ ∧ y = q Ex u ’ ) '' nominal primrec MakeFormP : : `` tm ⇒ tm ⇒ tm ⇒ fm '' where `` [ [ atom v ♯ ( y , u , w , au ) ; atom au ♯ ( y , u , w ) ] ] =⇒ MakeFormP y u w = y EQ Q Disj u w OR y EQ Q Neg u OR Ex v ( Ex au ( AbstFormP ( Var v ) Zero u ( Var au ) AND y EQ Q Ex ( Var au ) ) ) '' Now , the sequence buildup of a formula can be defined with Atomic covering the base case and MakeForm expressing one level of the construction . Using similar methods to those illustrated above for constant terms , we arrive at the shadow predicate Form and the corresponding HF predicate FormP . 4.3 Verifying the Coding Predicates Most textbook presentations take the correctness of such definitions as obvious , and indeed many properties are not difficult to prove . To show that the predicate Term 20 accepts all coded terms , a necessary lemma is to show the analogous property for well-formed de Bruijn terms : lemma wf Term quot dbtm : assumes `` wf dbtm u '' shows `` Term [ [ quot dbtm u ] ] e '' The proof is by induction on the construction of u ( in other words , on the inductive definition of wf dbtm u ) , and is routine by the definitions of the predicates Term and SeqTerm . This implies the desired result for terms , by the ( overloaded ) definition of ⌈t⌉ . corollary Term quot tm : fixes t : :tm shows `` Term [ [ ⌈t⌉
|
||
#42 '' Now , the sequence buildup of a formula can be defined with Atomic covering the base case and MakeForm expressing one level of the construction . Using similar methods to those illustrated above for constant terms , we arrive at the shadow predicate Form and the corresponding HF predicate FormP . 4.3 Verifying the Coding Predicates Most textbook presentations take the correctness of such definitions as obvious , and indeed many properties are not difficult to prove . To show that the predicate Term 20 accepts all coded terms , a necessary lemma is to show the analogous property for well-formed de Bruijn terms : lemma wf Term quot dbtm : assumes `` wf dbtm u '' shows `` Term [ [ quot dbtm u ] ] e '' The proof is by induction on the construction of u ( in other words , on the inductive definition of wf dbtm u ) , and is routine by the definitions of the predicates Term and SeqTerm . This implies the desired result for terms , by the ( overloaded ) definition of ⌈t⌉ . corollary Term quot tm : fixes t : :tm shows `` Term [ [ ⌈t⌉ ] ] e '' Note that both results concern the shadow predicate Term , not the HF predicate TermP . The argument of Term is a set , denoted using the evaluation operator , [ [ ... ] ] e. Direct proofs about HF predicates are long and tiresome . Fortunately , many such questions can be reduced to the corresponding questions involving shadow predicates , because codes are ground terms ; then the theorem Sigma fm imp thm guarantees the existence of a proof , sparing us the need to construct one . The converse correctness property must also be proved , namely that everything accepted by Term actually is the code of some term . The proof requires a lemma about the predicate SeqTerm . The reasoning is simply that constants and variables are wellformed , and that combining two well-formed terms preserves this property . Such proofs are streamlined through the use of BuildSeq induct , a derived rule for reasoning about sequence construction by induction on the length of the sequence . lemma Term imp wf dbtm : assumes `` Term x '' obtains t : :dbtm where `` wf dbtm t '' `` x = [ [ quot dbtm t ] ] e '' By the meaning of obtains , we see that if Term x then there exists some well-formed de Bruijn term t whose code evaluates to x . Since for every well-formed de Bruijn term there exists an equivalent standard term of type tm , we can conclude that Term x implies that x is the code of some term . corollary Term imp is tm : assumes `` Term x '' obtains t : :tm where `` x = [ [ ⌈t⌉ ] ] e '' Similar theorems—with similar proofs—are necessary
|
||
#43 be proved , namely that everything accepted by Term actually is the code of some term . The proof requires a lemma about the predicate SeqTerm . The reasoning is simply that constants and variables are wellformed , and that combining two well-formed terms preserves this property . Such proofs are streamlined through the use of BuildSeq induct , a derived rule for reasoning about sequence construction by induction on the length of the sequence . lemma Term imp wf dbtm : assumes `` Term x '' obtains t : :dbtm where `` wf dbtm t '' `` x = [ [ quot dbtm t ] ] e '' By the meaning of obtains , we see that if Term x then there exists some well-formed de Bruijn term t whose code evaluates to x . Since for every well-formed de Bruijn term there exists an equivalent standard term of type tm , we can conclude that Term x implies that x is the code of some term . corollary Term imp is tm : assumes `` Term x '' obtains t : :tm where `` x = [ [ ⌈t⌉ ] ] e '' Similar theorems—with similar proofs—are necessary for each of the syntactic predicates . For example , the following result expresses that SubstForm correctly models the result A ( i : :=t ) of substituting t for i in the formula A. lemma SubstForm quot unique : '' SubstForm ( q Var i ) [ [ ⌈t⌉ ] ] e [ [ ⌈A⌉ ] ] e x ’ ←→ x ’ = [ [ ⌈A ( i : :=t ) ⌉ ] ] e '' 4.4 Predicates for the Coding of Deduction On the whole , the formalisation of deduction is quite different from the formalisation of syntactic operations , which mostly involve simulated recursion over terms or formulas . A proof step in the HF calculus is an axiom , an axiom scheme or an inference rule . Axioms and propositional inference rules are straightforward to recognise using the existing syntactic primitives . Simply x EQ ⌈refl ax⌉ tests whether x denotes the reflexivity axiom . More complicated are inference rules involving quantification , where several syntactic conditions including abstraction and substitution need to be checked in sequence . For example , consider specialisation axioms of the form φ ( t/x ) → ∃x φ . 21 nominal primrec Special axP : : `` tm ⇒ fm '' where '' [ [ atom v ♯ ( p , sx , y , ax , x ) ; atom x ♯ ( p , sx , y , ax ) ; atom ax ♯ ( p , sx , y ) ; atom y ♯ ( p , sx ) ; atom sx ♯ p ] ] =⇒ Special axP p = Ex v ( Ex x ( Ex ax ( Ex y ( Ex sx ( FormP ( Var x ) AND VarP ( Var v
|
||
#44 syntactic operations , which mostly involve simulated recursion over terms or formulas . A proof step in the HF calculus is an axiom , an axiom scheme or an inference rule . Axioms and propositional inference rules are straightforward to recognise using the existing syntactic primitives . Simply x EQ ⌈refl ax⌉ tests whether x denotes the reflexivity axiom . More complicated are inference rules involving quantification , where several syntactic conditions including abstraction and substitution need to be checked in sequence . For example , consider specialisation axioms of the form φ ( t/x ) → ∃x φ . 21 nominal primrec Special axP : : `` tm ⇒ fm '' where '' [ [ atom v ♯ ( p , sx , y , ax , x ) ; atom x ♯ ( p , sx , y , ax ) ; atom ax ♯ ( p , sx , y ) ; atom y ♯ ( p , sx ) ; atom sx ♯ p ] ] =⇒ Special axP p = Ex v ( Ex x ( Ex ax ( Ex y ( Ex sx ( FormP ( Var x ) AND VarP ( Var v ) AND TermP ( Var y ) AND AbstFormP ( Var v ) Zero ( Var x ) ( Var ax ) AND SubstFormP ( Var v ) ( Var y ) ( Var x ) ( Var sx ) AND p EQ Q Imp ( Var sx ) ( Q Ex ( Var ax ) ) ) ) ) ) ) '' This definition states that a specialisation axiom is created from a formula x , a variable v and a term y , combined by appropriate abstraction and substitution operations . Correctness means proving that this predicate exactly characterises the elements of the set special axioms , which was used to define the HF calculus . The most complicated such scheme is the induction axiom HF3 ( defined in Sect . 2.2 ) , with its three quantifiers . The treatment of the induction axiom requires nearly 180 lines , of which 120 are devoted to proving correctness with respect to the HF calculus . A proof of a theorem y is a sequence s of axioms and inference rules , ending with y : definition Prf : : `` hf ⇒ hf ⇒ hf ⇒ bool '' where `` Prf s k y ≡ BuildSeq ( λx . x ∈ Axiom ) ( λu v w. ModPon v w u ∨ Exists v u ∨ Subst v u ) s k Finally , y codes a theorem provided it has a proof : definition Pf : : `` hf ⇒ bool '' where `` Pf y ≡ ( ∃ s k. Prf s k y ) '' Having proved the correctness of the predicates formalising the axioms and rules , the correctness of Pf follows easily . ( Swierczkowski ’ s seven lines of proof start here
|
||
#45 characterises the elements of the set special axioms , which was used to define the HF calculus . The most complicated such scheme is the induction axiom HF3 ( defined in Sect . 2.2 ) , with its three quantifiers . The treatment of the induction axiom requires nearly 180 lines , of which 120 are devoted to proving correctness with respect to the HF calculus . A proof of a theorem y is a sequence s of axioms and inference rules , ending with y : definition Prf : : `` hf ⇒ hf ⇒ hf ⇒ bool '' where `` Prf s k y ≡ BuildSeq ( λx . x ∈ Axiom ) ( λu v w. ModPon v w u ∨ Exists v u ∨ Subst v u ) s k Finally , y codes a theorem provided it has a proof : definition Pf : : `` hf ⇒ bool '' where `` Pf y ≡ ( ∃ s k. Prf s k y ) '' Having proved the correctness of the predicates formalising the axioms and rules , the correctness of Pf follows easily . ( Swierczkowski ’ s seven lines of proof start here . ) ´ One direction is proved by induction on the proof of { } ⊢ α. lemma proved imp Pf : assumes `` H ⊢ α '' `` H= { } '' shows `` Pf [ [ ⌈α⌉ ] ] e '' Here , we use the shadow predicates and work directly in Isabelle/HOL . The corresponding HF predicate , PfP , is ( crucially ) a Σ formula . Moreover , codes are ground terms . Therefore PfP ⌈α⌉ is a Σ sentence and is formally provable . This is the main use of the theorem Sigma fm imp thm . corollary proved imp proved PfP : `` { } ⊢ α =⇒ { } ⊢ PfP ⌈α⌉ '' The reverse implication , despite its usefulness , is not always proved . Again using the rule BuildSeq induct , it holds by induction on the length of the coded proof of ⌈α⌉ . The groundwork for this result involves proving , for each coded axiom and inference rule , that there exists a corresponding proof step in the HF calculus . We continue to work at the level of shadow predicates . lemma Prf imp proved : assumes `` Prf s k x '' shows `` ∃ A. x = [ [ ⌈A⌉ ] ] e ∧ { } ⊢ A '' The corresponding result for Pf is immediate : corollary Pf quot imp is proved : `` Pf [ [ ⌈α⌉ ] ] e =⇒ { } ⊢ α '' Now { } ⊢ PfP ⌈α⌉ implies Pf [ [ ⌈α⌉ ] ] e simply by the soundness of the calculus . It is now easy to show that the predicate PfP corresponds exactly to deduction in the HF calculus . Swierczkowski calls this result the ´
|
||
#46 imp thm . corollary proved imp proved PfP : `` { } ⊢ α =⇒ { } ⊢ PfP ⌈α⌉ '' The reverse implication , despite its usefulness , is not always proved . Again using the rule BuildSeq induct , it holds by induction on the length of the coded proof of ⌈α⌉ . The groundwork for this result involves proving , for each coded axiom and inference rule , that there exists a corresponding proof step in the HF calculus . We continue to work at the level of shadow predicates . lemma Prf imp proved : assumes `` Prf s k x '' shows `` ∃ A. x = [ [ ⌈A⌉ ] ] e ∧ { } ⊢ A '' The corresponding result for Pf is immediate : corollary Pf quot imp is proved : `` Pf [ [ ⌈α⌉ ] ] e =⇒ { } ⊢ α '' Now { } ⊢ PfP ⌈α⌉ implies Pf [ [ ⌈α⌉ ] ] e simply by the soundness of the calculus . It is now easy to show that the predicate PfP corresponds exactly to deduction in the HF calculus . Swierczkowski calls this result the ´ proof formalisation condition . theorem proved iff proved PfP : `` { } ⊢ α ←→ { } ⊢ PfP ⌈α⌉ '' 22 Remark : PfP includes an additional primitive inference , substitution : H ⊢ α H ⊢ α ( t/x ) This inference is derivable in the HF calculus , but the second incompleteness theorem requires performing substitution inferences , and reconstructing the derivation of substitution within PfP would be infeasible . Including substitution in the definition of PfP makes such steps immediate without complicating other proofs . Swierczkowski avoids ´ this complication : his HF calculus [ 32 ] includes a substitution rule alongside a simpler specialisation axiom . 4.5 Pseudo-Functions The HF calculus contains no function symbols other than ⊳ . All other “ functions ” must be declared as predicates , which are mere abbreviations of formulas . This abuse of notation is understood in a standard way . The formula φ ( f ( x ) ) can be taken as an abbreviation for ∃y [ F ( x , y ) ∧ φ ( y ) ] where F is the relation representing the function f and provided that F can be proved to be single valued : F ( x , y ) , F ( x , y′ ) ⊢ y ′ = y . Then f is called a pseudo-function and something like f ( x ) is called a pseudo-term . Pseudo-terms do not actually exist , which will cause problems later . G¨odel formalised all syntactic operations as primitive recursive functions , while Boolos [ 2 ] used ∆ formulas . With both approaches , much effort is necessary to admit a function definition in the first place . But then , it is known to be a
|
||
#47 32 ] includes a substitution rule alongside a simpler specialisation axiom . 4.5 Pseudo-Functions The HF calculus contains no function symbols other than ⊳ . All other “ functions ” must be declared as predicates , which are mere abbreviations of formulas . This abuse of notation is understood in a standard way . The formula φ ( f ( x ) ) can be taken as an abbreviation for ∃y [ F ( x , y ) ∧ φ ( y ) ] where F is the relation representing the function f and provided that F can be proved to be single valued : F ( x , y ) , F ( x , y′ ) ⊢ y ′ = y . Then f is called a pseudo-function and something like f ( x ) is called a pseudo-term . Pseudo-terms do not actually exist , which will cause problems later . G¨odel formalised all syntactic operations as primitive recursive functions , while Boolos [ 2 ] used ∆ formulas . With both approaches , much effort is necessary to admit a function definition in the first place . But then , it is known to be a function . Here we see a drawback of Swierczkowski ’ s decision to base the formalisation on the n ´ otion of Σ formulas : they do not cover the property of being single valued . A predicate that corresponds to a function must be proved to be single valued within the HF calculus itself . G¨odel ’ s proof uses substitution as a function . The proof that substitution ( on terms and formulas ) is single valued requires nearly 500 lines in Isabelle/HOL , not counting considerable preparatory material ( such as the partial ordering properties of < ) mentioned in Sect . 3.2 above . Fortunately , these proofs are conceptually simple and highly repetitious , and again much of the proof development can be done by cut and paste . The first step is to prove an unfolding lemma about the sequence buildup : if the predicate holds , then either the base case holds , or else there exist values earlier in the sequence for which one of the recursive cases can be applied . The single valued theorem is proved by complete induction on the length of the sequence , with a fully quantified induction formula ( analogous to ∀xyy′ [ F ( x , y ) → F ( x , y′ ) → y ′ = y ] ) so that the induction hypothesis says that all shorter sequences are single valued for all possible arguments . All that is left is to apply the unfolding lemma to both instances of the relation F , and then to consider all combinations of cases . Most will be trivially contradictory , and in those few cases where the result has the same outer form , an appeal to the induction hypothesis for the
|
||
#48 ) mentioned in Sect . 3.2 above . Fortunately , these proofs are conceptually simple and highly repetitious , and again much of the proof development can be done by cut and paste . The first step is to prove an unfolding lemma about the sequence buildup : if the predicate holds , then either the base case holds , or else there exist values earlier in the sequence for which one of the recursive cases can be applied . The single valued theorem is proved by complete induction on the length of the sequence , with a fully quantified induction formula ( analogous to ∀xyy′ [ F ( x , y ) → F ( x , y′ ) → y ′ = y ] ) so that the induction hypothesis says that all shorter sequences are single valued for all possible arguments . All that is left is to apply the unfolding lemma to both instances of the relation F , and then to consider all combinations of cases . Most will be trivially contradictory , and in those few cases where the result has the same outer form , an appeal to the induction hypothesis for the operands will complete the proof . Overall , the G¨odel development proves single valued theorems for 12 predicates . Five of the theorems are proved by induction as sketched above . Here is an example : lemma SeqSubstFormP unique : '' { SeqSubstFormP v a x y s u , SeqSubstFormP v a x y ’ s ’ u ’ } ⊢ y ’ EQ y '' The remaining results are straightforward corollaries of those inductions : 23 theorem SubstFormP unique : '' { SubstFormP v tm x y , SubstFormP v tm x y ’ } ⊢ y ’ EQ y '' It is worth repeating that these proofs are formally conducted within the HF calculus , essentially by single-step inferences . Meta-theory is no help here . 5 G¨odel ’ s First Incompleteness Theorem Discussions involving encodings frequently need a way to refer to syntactic objects . We often see the convention where if x is a natural number , then a boldface x stands for the corresponding numeral . Then in expressions like x = y → Pf px = yq , we see that the boldface convention actually abbreviates the function x 7→ x , which needs to be formalisable in the HF calculus . Thus , we need to define a function Q such that Q ( x ) = pxq , in other words , Q ( x ) yields some term t whose denotation is x . This is trivial if x ranges over the natural numbers , by primitive recursion . It is problematical when x ranges over sets , because it requires a canonical ordering over the universe of sets . We don ’ t need to solve this problem just yet : the first incompleteness theorem needs only a function
|
||
#49 EQ y '' It is worth repeating that these proofs are formally conducted within the HF calculus , essentially by single-step inferences . Meta-theory is no help here . 5 G¨odel ’ s First Incompleteness Theorem Discussions involving encodings frequently need a way to refer to syntactic objects . We often see the convention where if x is a natural number , then a boldface x stands for the corresponding numeral . Then in expressions like x = y → Pf px = yq , we see that the boldface convention actually abbreviates the function x 7→ x , which needs to be formalisable in the HF calculus . Thus , we need to define a function Q such that Q ( x ) = pxq , in other words , Q ( x ) yields some term t whose denotation is x . This is trivial if x ranges over the natural numbers , by primitive recursion . It is problematical when x ranges over sets , because it requires a canonical ordering over the universe of sets . We don ’ t need to solve this problem just yet : the first incompleteness theorem needs only a function H such that H ( pAq ) = ppAqq for all A . Possible arguments of H are not arbitrary sets , but only nested tuples of ordinals ; these have a canonical form , so a functional relationship is easy to define [ 32 ] . A certain amount of effort establishes the required property:6 lemma prove HRP : fixes A : :fm shows `` { } ⊢ HRP ⌈A⌉ ⌈⌈A⌉⌉ '' Note that the function H has been formalised as the relation HRP ; it is defined using sequence operators , LstSeqP , etc. , as we have seen already . In order to prove G¨odel ’ s diagonal lemma , we need a function Ki to substitute the code of a formula into itself , replacing the variable xi . This function , again , is realised as a relation , combining HRP with SubstFormP . nominal primrec KRP : : `` tm ⇒ tm ⇒ tm ⇒ fm '' where `` atom y ♯ ( v , x , x ’ ) =⇒ KRP v x x ’ = Ex y ( HRP x ( Var y ) AND SubstFormP v ( Var y ) x x ’ ) '' We easily obtain a key result : Ki ( pαq ) = pα ( i : = pαq ) q. lemma prove KRP : `` { } ⊢ KRP ⌈Var i⌉ ⌈α⌉ ⌈α ( i : :=⌈α⌉ ) ⌉ '' However , it is essential to prove that KRP behaves like a function . The predicates KRP and HRP can be proved to be single valued using the techniques discussed in the previous section . Then an appeal to prove KRP uniquely characterises Ki as a function : lemma KRP subst fm : `` { KRP
|
||
#50 seen already . In order to prove G¨odel ’ s diagonal lemma , we need a function Ki to substitute the code of a formula into itself , replacing the variable xi . This function , again , is realised as a relation , combining HRP with SubstFormP . nominal primrec KRP : : `` tm ⇒ tm ⇒ tm ⇒ fm '' where `` atom y ♯ ( v , x , x ’ ) =⇒ KRP v x x ’ = Ex y ( HRP x ( Var y ) AND SubstFormP v ( Var y ) x x ’ ) '' We easily obtain a key result : Ki ( pαq ) = pα ( i : = pαq ) q. lemma prove KRP : `` { } ⊢ KRP ⌈Var i⌉ ⌈α⌉ ⌈α ( i : :=⌈α⌉ ) ⌉ '' However , it is essential to prove that KRP behaves like a function . The predicates KRP and HRP can be proved to be single valued using the techniques discussed in the previous section . Then an appeal to prove KRP uniquely characterises Ki as a function : lemma KRP subst fm : `` { KRP ⌈Var i⌉ ⌈α⌉ ( Var j ) } ⊢ Var j EQ ⌈α ( i : :=⌈α⌉ ) ⌉ '' Twenty five lines of tricky reasoning are needed to reach the next milestone : the diagonal lemma . Swierczkowski writes ´ We replace the variable xi in α by the [ pseudo-term Ki ( xi ) ] , and we denote by β the resulting formula . [ 32 , p. 22 ] The elimination of the pseudo-function Ki in favour of an existential quantifier involving KRP yields the following not-entirely-obvious Isabelle definition : 6 Here fixes A : :fm declares A to be a formula in the overloaded notation ⌈A⌉ . Swierczkowski ´ uses α , β , . . . to denote formulas , but I ’ ve frequently used the traditional A , B , . . . . 24 theorem Goedel I : assumes `` ¬ { } ⊢ Fls '' obtains δ where `` { } ⊢ δ IFF Neg ( PfP ⌈δ⌉ ) '' `` ¬ { } ⊢ δ '' `` ¬ { } ⊢ Neg δ '' '' eval fm e δ '' `` ground fm δ '' proof - fix i : :name obtain δ where `` { } ⊢ δ IFF Neg ( ( PfP ( Var i ) ) ( i : :=⌈δ⌉ ) ) '' and suppd : `` supp δ = supp ( Neg ( PfP ( Var i ) ) ) - { atom i } '' by ( metis SyntaxN.Neg diagonal ) then have diag : `` { } ⊢ δ IFF Neg ( PfP ⌈δ⌉ ) '' by simp then have np : `` ¬ { } ⊢ δ ∧ ¬ { } ⊢ Neg δ '' by ( metis Iff MP same NegNeg
|
||
#51 declares A to be a formula in the overloaded notation ⌈A⌉ . Swierczkowski ´ uses α , β , . . . to denote formulas , but I ’ ve frequently used the traditional A , B , . . . . 24 theorem Goedel I : assumes `` ¬ { } ⊢ Fls '' obtains δ where `` { } ⊢ δ IFF Neg ( PfP ⌈δ⌉ ) '' `` ¬ { } ⊢ δ '' `` ¬ { } ⊢ Neg δ '' '' eval fm e δ '' `` ground fm δ '' proof - fix i : :name obtain δ where `` { } ⊢ δ IFF Neg ( ( PfP ( Var i ) ) ( i : :=⌈δ⌉ ) ) '' and suppd : `` supp δ = supp ( Neg ( PfP ( Var i ) ) ) - { atom i } '' by ( metis SyntaxN.Neg diagonal ) then have diag : `` { } ⊢ δ IFF Neg ( PfP ⌈δ⌉ ) '' by simp then have np : `` ¬ { } ⊢ δ ∧ ¬ { } ⊢ Neg δ '' by ( metis Iff MP same NegNeg D Neg D Neg cong assms proved iff proved PfP ) then have `` eval fm e δ '' using hfthm sound [ where e=e , OF diag ] by simp ( metis Pf quot imp is proved ) moreover have `` ground fm δ '' using suppd by ( simp add : supp conv fresh ground fm aux def subset eq ) ( metis fresh ineq at base ) ultimately show ? thesis by ( metis diag np that ) qed Fig . 1 G¨odel ’ s First Incompleteness Theorem def β ≡ `` Ex j ( KRP ⌈Var i⌉ ( Var i ) ( Var j ) AND α ( i : := Var j ) ) '' Note that one occurrence of Var i is quoted and the other is not . The development is full of pitfalls such as these . The statement of the diagonal lemma is as follows . The second assertion states that the free variables of δ , the diagonal formula , are those of α , the original formula , minus i. lemma diagonal : obtains δ where `` { } ⊢ δ IFF α ( i : :=⌈δ⌉ ) '' `` supp δ = supp α - { atom i } '' Figure 1 presents the proof of the first incompleteness theorem itself . The top level argument is quite simple , given the diagonal lemma . The key steps of the proof should be visible even to somebody who is not an Isabelle expert , thanks to the structured Isar language . Note that if { } ⊢ Neg δ , then { } ⊢ PfP ⌈δ⌉ and therefore { } ⊢ δ by the proof formalisation condition , violating the assumption of consistency . 6 Towards the Second
|
||
#52 i⌉ ( Var i ) ( Var j ) AND α ( i : := Var j ) ) '' Note that one occurrence of Var i is quoted and the other is not . The development is full of pitfalls such as these . The statement of the diagonal lemma is as follows . The second assertion states that the free variables of δ , the diagonal formula , are those of α , the original formula , minus i. lemma diagonal : obtains δ where `` { } ⊢ δ IFF α ( i : :=⌈δ⌉ ) '' `` supp δ = supp α - { atom i } '' Figure 1 presents the proof of the first incompleteness theorem itself . The top level argument is quite simple , given the diagonal lemma . The key steps of the proof should be visible even to somebody who is not an Isabelle expert , thanks to the structured Isar language . Note that if { } ⊢ Neg δ , then { } ⊢ PfP ⌈δ⌉ and therefore { } ⊢ δ by the proof formalisation condition , violating the assumption of consistency . 6 Towards the Second Theorem : Pseudo-Coding and Quotations The second incompleteness theorem [ 1 ] has always been more mysterious than the first . G¨odel ’ s original paper [ 8 ] was designated “ Part I ” in anticipation of a subsequent “ Part II ” proving the second theorem , but no second paper appeared . Logicians recognised that the second theorem followed from the first , assuming that the first could itself be formalised in the internal calculus . While this assumption seems to be widely accepted , conducting such a formalisation explicitly remains difficult , even given today ’ s theoremproving technology . 25 A simpler route to the theorem involves the Hilbert-Bernays derivability conditions [ 2 , p. 15 ] [ 9 , p. 73 ] . If ⊢ α , then ⊢ Pf ( pαq ) ( D1 ) If ⊢ Pf ( pα → βq ) and ⊢ Pf ( pαq ) , then ⊢ Pf ( pβq ) ( D2 ) If α is a strict Σ sentence , then ⊢ α → Pf ( pαq ) ( D3 ) ( Where there is no ambiguity , we identify Pf with the formalised predicate PfP ; the latter is the actual HF predicate , but the notation Pf is widely used in the literature , along with G¨odel ’ s original Bew . ) Condition ( D1 ) is none other than the theorem proved iff proved PfP mentioned in Sect . 4.4 above . Condition ( D2 ) seems clear by the definition of the predicate Pf , although all details of the workings of this predicate must be proved using low-level inferences in the HF calculus . Condition ( D3 ) can be regarded as a version of the theorem Sigma fm imp
|
||
#53 s theoremproving technology . 25 A simpler route to the theorem involves the Hilbert-Bernays derivability conditions [ 2 , p. 15 ] [ 9 , p. 73 ] . If ⊢ α , then ⊢ Pf ( pαq ) ( D1 ) If ⊢ Pf ( pα → βq ) and ⊢ Pf ( pαq ) , then ⊢ Pf ( pβq ) ( D2 ) If α is a strict Σ sentence , then ⊢ α → Pf ( pαq ) ( D3 ) ( Where there is no ambiguity , we identify Pf with the formalised predicate PfP ; the latter is the actual HF predicate , but the notation Pf is widely used in the literature , along with G¨odel ’ s original Bew . ) Condition ( D1 ) is none other than the theorem proved iff proved PfP mentioned in Sect . 4.4 above . Condition ( D2 ) seems clear by the definition of the predicate Pf , although all details of the workings of this predicate must be proved using low-level inferences in the HF calculus . Condition ( D3 ) can be regarded as a version of the theorem Sigma fm imp thm ( “ true Σ sentences are theorems ” ) internalised as a theorem of the internal calculus . So while we avoid having to formalise the whole of G¨odel ’ s theorem within the calculus , we end up formalising a key part of it . Condition ( D3 ) is stated in a general form , but we only need one specific instance : ⊢ Pf ( pαq ) → Pf ( pPf ( pαq ) q ) . Despite a superficial resemblance , ( D3 ) does not follow from ( D1 ) , which holds by induction on the proof of ⊢ α . As Swierczkowski explains [ 32 , p. 23 ] , condition ( D3 ) is ´ not general enough to prove by induction . In the sequel , we generalise and prove it . 6.1 Pseudo-Coding Condition ( D3 ) can be proved by induction on α if the assertion is generalised so that α can have free variables , say x1 , . . . , xn : ⊢ α ( x1 , . . . , xn ) → Pf ( pα ( x1 , . . . , xn ) q ) The syntactic constructions used in this formula have to be formalised , and the necessary transformations have to be justified within the HF calculus . As mentioned above ( Sect . 5 ) , the boldface convention needs to be made rigorous . In particular , codings are always ground HF terms , and yet pα ( x1 , . . . , xn ) q has a functional dependence ( as an HF term ) on x1 , . . . , xn . The first step in this process is to generalise coding to allow
|
||
#54 induction on the proof of ⊢ α . As Swierczkowski explains [ 32 , p. 23 ] , condition ( D3 ) is ´ not general enough to prove by induction . In the sequel , we generalise and prove it . 6.1 Pseudo-Coding Condition ( D3 ) can be proved by induction on α if the assertion is generalised so that α can have free variables , say x1 , . . . , xn : ⊢ α ( x1 , . . . , xn ) → Pf ( pα ( x1 , . . . , xn ) q ) The syntactic constructions used in this formula have to be formalised , and the necessary transformations have to be justified within the HF calculus . As mentioned above ( Sect . 5 ) , the boldface convention needs to be made rigorous . In particular , codings are always ground HF terms , and yet pα ( x1 , . . . , xn ) q has a functional dependence ( as an HF term ) on x1 , . . . , xn . The first step in this process is to generalise coding to allow certain variables to be preserved as variables in the coded term . Recall that with normal quotations , every occurrence of a variable is replaced by the code of the variable name , ultimately a positive integer:7 function quot dbtm : : `` dbtm ⇒ tm '' where '' quot dbtm DBZero = Zero '' | `` quot dbtm ( DBVar name ) = ORD OF ( Suc ( nat of name name ) ) '' | ... Now let us define the V -code of a term or formula , where V is a set of variables to be preserved in the code : 7 ORD OF ( Suc n ) denotes an HF term that denotes a positive integer , even if n is a variable . 26 function vquot dbtm : : `` name set ⇒ dbtm ⇒ tm '' where '' vquot dbtm V DBZero = Zero '' | `` vquot dbtm V ( DBVar name ) = ( if name ∈ V then Var name else ORD OF ( Suc ( nat of name name ) ) ) '' | ... V -coding is otherwise identical to standard coding , with the overloaded syntax ⌊A⌋V . The parameter V is necessary because not all variables should be preserved ; it will be necessary to consider a series of V -codes for V = ∅ , { x1 } . . . , { x1 , . . . , xn } . 6.2 Simultaneous Substitution In order to formalise the notation pα ( x1 , . . . , xn ) q , it is convenient to introduce a function for simultaneous substitution . Here Swierczkowski ’ s presentation is a little ´ hard to follow : Suppose β is a theorem , i.e. ,
|
||
#55 preserved in the code : 7 ORD OF ( Suc n ) denotes an HF term that denotes a positive integer , even if n is a variable . 26 function vquot dbtm : : `` name set ⇒ dbtm ⇒ tm '' where '' vquot dbtm V DBZero = Zero '' | `` vquot dbtm V ( DBVar name ) = ( if name ∈ V then Var name else ORD OF ( Suc ( nat of name name ) ) ) '' | ... V -coding is otherwise identical to standard coding , with the overloaded syntax ⌊A⌋V . The parameter V is necessary because not all variables should be preserved ; it will be necessary to consider a series of V -codes for V = ∅ , { x1 } . . . , { x1 , . . . , xn } . 6.2 Simultaneous Substitution In order to formalise the notation pα ( x1 , . . . , xn ) q , it is convenient to introduce a function for simultaneous substitution . Here Swierczkowski ’ s presentation is a little ´ hard to follow : Suppose β is a theorem , i.e. , ⊢ β . If we replace each of the variables at each of its free occurrences in β by some constant term then the formula so obtained is also a theorem ( by the Substitution Rule . . . ) . This just described situation in the meta-theory admits description in HF . [ 32 , p. 24 ] It took me weeks of failed attempts to grasp the meaning of the phrase “ constant term ” . It does not mean a term containing no variables , but a term satisfying the predicate ConstP and thus denoting the code of a constant . Formalising this process seems to require replacing each variable xi by a new variable , x ′ i , intended to denote xi . Later , it will be constrained to do so by a suitable HF predicate . And so we need a function to perform simultaneous substitutions in a term for all variables in a set V . Using a “ fold ” operator over finite sets [ 19 ] eliminates the need to consider the variables in any particular order . definition ssubst : : `` tm ⇒ name set ⇒ ( name ⇒ tm ) ⇒ tm '' where `` ssubst t V F = Finite Set.fold ( λi . subst i ( F i ) ) t V '' The renaming of xi to x ′ i could be done using arithmetic on variable subscripts , but the formalisation instead follows an abstract approach , using nominal primitives . An Isabelle locale defines a proof context containing a permutation p ( mapping all variable names to new ones ) , a finite set Vs of variable names and finally the actual renaming function F , which simply applies the permutation to
|
||
#56 of a constant . Formalising this process seems to require replacing each variable xi by a new variable , x ′ i , intended to denote xi . Later , it will be constrained to do so by a suitable HF predicate . And so we need a function to perform simultaneous substitutions in a term for all variables in a set V . Using a “ fold ” operator over finite sets [ 19 ] eliminates the need to consider the variables in any particular order . definition ssubst : : `` tm ⇒ name set ⇒ ( name ⇒ tm ) ⇒ tm '' where `` ssubst t V F = Finite Set.fold ( λi . subst i ( F i ) ) t V '' The renaming of xi to x ′ i could be done using arithmetic on variable subscripts , but the formalisation instead follows an abstract approach , using nominal primitives . An Isabelle locale defines a proof context containing a permutation p ( mapping all variable names to new ones ) , a finite set Vs of variable names and finally the actual renaming function F , which simply applies the permutation to any variable in Vs. 8 locale quote perm = fixes p : : perm and Vs : : `` name set '' and F : : `` name ⇒ tm '' assumes p : `` atom ‘ ( p · Vs ) ♯ * Vs '' and pinv : `` -p = p '' and Vs : `` finite Vs '' defines `` F ≡ make F Vs p '' Most proofs about ssubst are done within the context of this locale , because it provides sufficient conditions for the simultaneous substitution to be meaningful . The first condition states that p maps all the variables in Vs to variables outside of that set , while second condition states that p is its own inverse . This abstract approach is a little unwieldy at times , but its benefits can be seen in the simple fact below , which states the effect of the simultaneous substitution on a single variable . 8 make F Vs p i = Var ( p · i ) provided i ∈ Vs. 27 lemma ssubst Var if : assumes `` finite V '' shows `` ssubst ( Var i ) V F = ( if i ∈ V then F i else Var i ) '' We need to show that the variables in the set Vs can be renamed , one at a time , in a pseudo-coded de Bruijn term . Let V ⊆ Vs and suppose that the variables in V have already been renamed , and choose one of the remaining variables , w. It will be replaced by a new variable , computed by F w. And something very subtle is happening : the variable w is represented in the term by its code , ⌈Var w⌉ . Its
|
||
#57 that p maps all the variables in Vs to variables outside of that set , while second condition states that p is its own inverse . This abstract approach is a little unwieldy at times , but its benefits can be seen in the simple fact below , which states the effect of the simultaneous substitution on a single variable . 8 make F Vs p i = Var ( p · i ) provided i ∈ Vs. 27 lemma ssubst Var if : assumes `` finite V '' shows `` ssubst ( Var i ) V F = ( if i ∈ V then F i else Var i ) '' We need to show that the variables in the set Vs can be renamed , one at a time , in a pseudo-coded de Bruijn term . Let V ⊆ Vs and suppose that the variables in V have already been renamed , and choose one of the remaining variables , w. It will be replaced by a new variable , computed by F w. And something very subtle is happening : the variable w is represented in the term by its code , ⌈Var w⌉ . Its replacement , F w , is Var ( p · w ) and a variable . lemma SubstTermP vquot dbtm : assumes w : `` w ∈ Vs - V '' and V : `` V ⊆ Vs '' `` V ’ = p · V '' and s : `` supp dbtm ⊆ atom ‘ Vs '' shows '' insert ( ConstP ( F w ) ) { ConstP ( F i ) | i. i ∈ V } ⊢ SubstTermP ⌈Var w⌉ ( F w ) ( ssubst ( vquot dbtm V dbtm ) V F ) ( subst w ( F w ) ( ssubst ( vquot dbtm ( insert w V ) dbtm ) V F ) ) '' This theorem is proved by structural induction on dbtm , the de Bruijn term . The condition supp dbtm ⊆ atom ‘ Vs states that the free variables of dbtm all belong to Vs . The variable case of the induction is tricky ( and is the crux of the entire proof ) . We are working with a coded term that contains both coded variables and real ones ( of the form F i ) ; it is necessary to show that the real variables are preserved by the substitution , because they are the xi that we are trying to introduce . The F i are preserved under the assumption that they denote constants , which is the point of including the formulas ConstP ( F i ) for all i ∈ V on the left side of the turnstile . These assumptions will have to be justified later . Under virtually the same assumptions ( omitted ) , the analogous result holds for pseudo-coded de Bruijn formulas . lemma SubstFormP vquot dbfm : ''
|
||
#58 subst w ( F w ) ( ssubst ( vquot dbtm ( insert w V ) dbtm ) V F ) ) '' This theorem is proved by structural induction on dbtm , the de Bruijn term . The condition supp dbtm ⊆ atom ‘ Vs states that the free variables of dbtm all belong to Vs . The variable case of the induction is tricky ( and is the crux of the entire proof ) . We are working with a coded term that contains both coded variables and real ones ( of the form F i ) ; it is necessary to show that the real variables are preserved by the substitution , because they are the xi that we are trying to introduce . The F i are preserved under the assumption that they denote constants , which is the point of including the formulas ConstP ( F i ) for all i ∈ V on the left side of the turnstile . These assumptions will have to be justified later . Under virtually the same assumptions ( omitted ) , the analogous result holds for pseudo-coded de Bruijn formulas . lemma SubstFormP vquot dbfm : '' insert ( ConstP ( F w ) ) { ConstP ( F i ) | i. i ∈ V } ⊢ SubstFormP ⌈Var w⌉ ( F w ) ( ssubst ( vquot dbfm V dbfm ) V F ) ( subst w ( F w ) ( ssubst ( vquot dbfm ( insert w V ) dbfm ) V F ) ) '' The proof is an easy structural induction on dbfm . Every case holds immediately by properties of substitution and the induction hypotheses or by the previous theorem , for terms . The only difficult case in these two proofs is the variable case discussed above . Using the notation for V -coding , we can see that the substitution predicate SubstFormP can transform the term ssubst ⌊A⌋V V F into ssubst ⌊A⌋ ( insert w V ) ( insert w V ) F. Repeating this step , we can replace any finite set of variables in a coded formula by real ones , realising Swierczkowski ’ s remark quoted at the top of this section , an ´ d in particular his last sentence . That is , if β is a theorem in HF ( if ⊢ Pf β holds ) then the result of substituting constants for its variables is also an HF theorem . More precisely still , we are replacing some subset V of the variables by fresh variables ( the F i ) , constrained by the predicate ConstP . theorem PfP implies PfP ssubst : fixes β : :fm assumes β : `` { } ⊢ PfP ⌈β⌉ '' and V : `` V ⊆ Vs '' and s : `` supp β ⊆ atom ‘ Vs '' shows `` { ConstP ( F i ) | i. i ∈ V
|
||
#59 two proofs is the variable case discussed above . Using the notation for V -coding , we can see that the substitution predicate SubstFormP can transform the term ssubst ⌊A⌋V V F into ssubst ⌊A⌋ ( insert w V ) ( insert w V ) F. Repeating this step , we can replace any finite set of variables in a coded formula by real ones , realising Swierczkowski ’ s remark quoted at the top of this section , an ´ d in particular his last sentence . That is , if β is a theorem in HF ( if ⊢ Pf β holds ) then the result of substituting constants for its variables is also an HF theorem . More precisely still , we are replacing some subset V of the variables by fresh variables ( the F i ) , constrained by the predicate ConstP . theorem PfP implies PfP ssubst : fixes β : :fm assumes β : `` { } ⊢ PfP ⌈β⌉ '' and V : `` V ⊆ Vs '' and s : `` supp β ⊆ atom ‘ Vs '' shows `` { ConstP ( F i ) | i. i ∈ V } ⊢ PfP ( ssubst ⌊β⌋V V F ) '' 28 The effort needed to formalise the results outlined above is relatively modest , at 330 lines of Isabelle/HOL , but this excludes the effort needed to prove some essential lemmas , which state that the various syntactic predicates work correctly . Because these proofs concern non-ground HF formulas , theorem Sigma fm imp thm does not help . Required is an HF formalisation of operations on sequences , such as concatenation . That in turn requires formalising further operations such as addition and set union . These proofs ( which are conducted largely in the HF calculus ) total over 2,800 lines . This total includes a library of results for truncating and concatenating sequences . Here is a selection of the results proved . Substitution preserves the value Zero : theorem SubstTermP Zero : `` { TermP t } ⊢ SubstTermP ⌈Var v⌉ t Zero Zero '' On terms constructed using Eats ( recall that Q Eats constructs the code of an Eats term ) , substitution performs the natural recursion . theorem SubstTermP Eats : '' { SubstTermP v i t1 u1 , SubstTermP v i t2 u2 } ⊢ SubstTermP v i ( Q Eats t1 t2 ) ( Q Eats u1 u2 ) '' This seemingly obvious result takes nearly 150 lines to prove . The sequences for both substitution computations are combined to form a new sequence , which must be extended to yield the claimed result and shown to be properly constructed . Substitution preserves constants . This fact is proved by induction on the sequence buildup of the constant c , using the previous two facts about SubstTermP . theorem SubstTermP Const : `` { ConstP c , TermP t } ⊢
|
||
#60 which are conducted largely in the HF calculus ) total over 2,800 lines . This total includes a library of results for truncating and concatenating sequences . Here is a selection of the results proved . Substitution preserves the value Zero : theorem SubstTermP Zero : `` { TermP t } ⊢ SubstTermP ⌈Var v⌉ t Zero Zero '' On terms constructed using Eats ( recall that Q Eats constructs the code of an Eats term ) , substitution performs the natural recursion . theorem SubstTermP Eats : '' { SubstTermP v i t1 u1 , SubstTermP v i t2 u2 } ⊢ SubstTermP v i ( Q Eats t1 t2 ) ( Q Eats u1 u2 ) '' This seemingly obvious result takes nearly 150 lines to prove . The sequences for both substitution computations are combined to form a new sequence , which must be extended to yield the claimed result and shown to be properly constructed . Substitution preserves constants . This fact is proved by induction on the sequence buildup of the constant c , using the previous two facts about SubstTermP . theorem SubstTermP Const : `` { ConstP c , TermP t } ⊢ SubstTermP ⌈Var w⌉ t c c '' Each recursive case of a syntactic predicate must be verified using the techniques outlined above for SubstTermP Eats . Even when there is only a single operand , as in the following case for negation , the proof is around 100 lines . theorem SubstFormP Neg : `` { SubstFormP v i x y } ⊢ SubstFormP v i ( Q Neg x ) ( Q Neg y ) '' A complication is that LstSeqP accepts sequences that are longer than necessary , and these must be truncated to a given length before they can be extended . These lengthy arguments must be repeated for each similar proof . So , for the third time , one of our chief tools is cut and paste . Exactly the same sort of reasoning can be used to show that proofs can be combined as expected in order to apply inference rules . The following theorem expresses the Hilbert-Bernays derivability condition ( D2 ) : theorem PfP implies ModPon PfP : `` [ [ H ⊢ PfP ( Q Imp x y ) ; H ⊢ PfP x ] ] =⇒ H ⊢ PfP y '' Now only one task remains : to show condition ( D3 ) . 6.3 Making Sense of Quoted Values As mentioned in Sect . 5 , making sense of expressions like x = y → Pf px = yq requires a function Q such that Q ( x ) = pxq : Q ( 0 ) = p0q = 0 Q ( x ⊳ y ) = hp⊳q , Q ( x ) , Q ( y ) i 29 Trying to make this definition unambiguous , Swierczkowski [ 32 ] sketches a total order- ´ ing on sets
|
||
#61 they can be extended . These lengthy arguments must be repeated for each similar proof . So , for the third time , one of our chief tools is cut and paste . Exactly the same sort of reasoning can be used to show that proofs can be combined as expected in order to apply inference rules . The following theorem expresses the Hilbert-Bernays derivability condition ( D2 ) : theorem PfP implies ModPon PfP : `` [ [ H ⊢ PfP ( Q Imp x y ) ; H ⊢ PfP x ] ] =⇒ H ⊢ PfP y '' Now only one task remains : to show condition ( D3 ) . 6.3 Making Sense of Quoted Values As mentioned in Sect . 5 , making sense of expressions like x = y → Pf px = yq requires a function Q such that Q ( x ) = pxq : Q ( 0 ) = p0q = 0 Q ( x ⊳ y ) = hp⊳q , Q ( x ) , Q ( y ) i 29 Trying to make this definition unambiguous , Swierczkowski [ 32 ] sketches a total order- ´ ing on sets , but the technical details are complicated and incomplete . The same ordering can be defined via the function f : HF → N such that f ( x ) = P { 2 f ( y ) | y ∈ x } . It is intuitively clear , but formalising the required theory in HF would be laborious . It turns out that we do not need a canonical term x or a function Q . We only need a relation : QuoteP relates a set x to ( the codes of ) the terms that denote x . The relation QuoteP can be defined using precisely the same methods as we have seen above for recursive functions , via a sequence buildup . The following facts can be proved using the methods described in the previous sections . lemma QuoteP Zero : `` { } ⊢ QuoteP Zero Zero '' lemma QuoteP Eats : '' { QuoteP t1 u1 , QuoteP t2 u2 } ⊢ QuoteP ( Eats t1 t2 ) ( Q Eats u1 u2 ) '' It is also necessary to prove ( by induction within the HF calculus ) that for every x there exists some term x. lemma exists QuoteP : assumes j : `` atom j ♯ x '' shows `` { } ⊢ Ex j ( QuoteP x ( Var j ) ) '' We need similar results for all of the predicates involved in concatenating two sequences . They essentially prove that the corresponding pseudo-functions are total . Now we need to start connecting these results with those of the previous section , which ( following Swierczkowski ) are proved for constants in general , althou ´ gh they are needed only for the outputs of QuoteP . An induction in
|
||
#62 The relation QuoteP can be defined using precisely the same methods as we have seen above for recursive functions , via a sequence buildup . The following facts can be proved using the methods described in the previous sections . lemma QuoteP Zero : `` { } ⊢ QuoteP Zero Zero '' lemma QuoteP Eats : '' { QuoteP t1 u1 , QuoteP t2 u2 } ⊢ QuoteP ( Eats t1 t2 ) ( Q Eats u1 u2 ) '' It is also necessary to prove ( by induction within the HF calculus ) that for every x there exists some term x. lemma exists QuoteP : assumes j : `` atom j ♯ x '' shows `` { } ⊢ Ex j ( QuoteP x ( Var j ) ) '' We need similar results for all of the predicates involved in concatenating two sequences . They essentially prove that the corresponding pseudo-functions are total . Now we need to start connecting these results with those of the previous section , which ( following Swierczkowski ) are proved for constants in general , althou ´ gh they are needed only for the outputs of QuoteP . An induction in HF on the sequence buildup proves that these outputs satisfy ConstP . lemma QuoteP imp ConstP : `` { QuoteP x y } ⊢ ConstP y '' This is obvious , because QuoteP involves only Zero and Q Eats , which construct quoted sets . Unfortunately , the proof requires the usual reasoning about sequences in order to show basic facts about ConstP , which takes hundreds of lines . The main theorem from the previous section included the set of formulas { ConstP ( F i ) | i. i ∈ V } on the left of the turnstile , representing the assumption that all introduced variables denoted constants . Now we can replace this assumption by one expressing that the relation QuoteP holds between each pair of old and new variables . definition quote all : : `` [ perm , name set ] ⇒ fm set '' where `` quote all p V = { QuoteP ( Var i ) ( Var ( p · i ) ) | i. i ∈ V } The relation QuoteP ( Var i ) ( Var ( p · i ) holds between the variable i and the renamed variable p · i , for all i ∈ V. Recall that p is a permutation on variable names . By virtue of the theorem QuoteP imp ConstP , we obtain a key result , which will be used heavily in subsequent proofs for reasoning about coded formulas and the Pf predicate . theorem quote all PfP ssubst : assumes β : `` { } ⊢ β '' and V : `` V ⊆ Vs '' and s : `` supp β ⊆ atom ‘ Vs '' shows `` quote all p V ⊢ PfP ( ssubst ⌊β⌋V V F )
|
||
#63 , representing the assumption that all introduced variables denoted constants . Now we can replace this assumption by one expressing that the relation QuoteP holds between each pair of old and new variables . definition quote all : : `` [ perm , name set ] ⇒ fm set '' where `` quote all p V = { QuoteP ( Var i ) ( Var ( p · i ) ) | i. i ∈ V } The relation QuoteP ( Var i ) ( Var ( p · i ) holds between the variable i and the renamed variable p · i , for all i ∈ V. Recall that p is a permutation on variable names . By virtue of the theorem QuoteP imp ConstP , we obtain a key result , which will be used heavily in subsequent proofs for reasoning about coded formulas and the Pf predicate . theorem quote all PfP ssubst : assumes β : `` { } ⊢ β '' and V : `` V ⊆ Vs '' and s : `` supp β ⊆ atom ‘ Vs '' shows `` quote all p V ⊢ PfP ( ssubst ⌊β⌋V V F ) '' In English : Let ⊢ β be a theorem of HF whose free variables belong to the set Vs. Take the code of this theorem , ⌊β⌋ , and replace some subset V ⊆ Vs of its free variables by 30 new variables constrained by the QuoteP relation . The result , ssubst ⌊β⌋V V F , satisfies the provability predicate . The reader of even a very careful presentation of G¨odel ’ s second incompleteness theorem , such as Grandy [ 9 ] , will look in vain for a clear and rigorous treatment of the x ( or x ) convention . Boolos [ 2 ] comes very close with his Bew [ F ] notation , but he is quite wrong to state “ notice that Bew [ F ] has the same variables free as [ the formula ] F ” [ 2 , p. 45 ] when in fact they have no variables in common . Even Swierczkowski ’ s highly ´ detailed account is at best ambiguous . He consistently uses function notation , but his carefully-stated guidelines for replacing occurrences of pseudo-functions by quantified formulas [ 32 , Sect . 5 ] are not relevant here . ( This problem only became clear after a time-consuming attempt at a formalisation on that basis . ) My companion paper [ 27 ] , which is aimed at logicians , provides a more detailed discussion of these points . It concludes that these various notations obscure not only the formal details of the proof but also the very intuitions they are intended to highlight . 6.4 Proving ⊢ α → Pf ( pαq ) We now have everything necessary to prove condition ( D3 ) : If α is a strict Σ sentence ,
|
||
#64 or x ) convention . Boolos [ 2 ] comes very close with his Bew [ F ] notation , but he is quite wrong to state “ notice that Bew [ F ] has the same variables free as [ the formula ] F ” [ 2 , p. 45 ] when in fact they have no variables in common . Even Swierczkowski ’ s highly ´ detailed account is at best ambiguous . He consistently uses function notation , but his carefully-stated guidelines for replacing occurrences of pseudo-functions by quantified formulas [ 32 , Sect . 5 ] are not relevant here . ( This problem only became clear after a time-consuming attempt at a formalisation on that basis . ) My companion paper [ 27 ] , which is aimed at logicians , provides a more detailed discussion of these points . It concludes that these various notations obscure not only the formal details of the proof but also the very intuitions they are intended to highlight . 6.4 Proving ⊢ α → Pf ( pαq ) We now have everything necessary to prove condition ( D3 ) : If α is a strict Σ sentence , then ⊢ α → Pf ( pαq ) The proof will be by induction on the structure of α . As stated in Sect . 3.3 above , a strict Σ formula has the form x ∈ y , α ∨ β , α ∧ β , ∃x α or ( ∀x ∈ y ) α . Therefore , the induction will include one single base case , x ∈ y → Pf px ∈ yq , ( 3 ) along with inductive steps for disjunction , conjunction , etc . 6.4.1 The Propositional Cases The propositional cases are not difficult , but are worth examining as a warmup exercise . From the induction hypotheses ⊢ α → Pf ( pαq ) and ⊢ β → Pf ( pβq ) , we must show ⊢ α ∨ β → Pf ( pα ∨ βq ) and ⊢ α ∧ β → Pf ( pα ∧ βq ) . Both of these cases are trivial by propositional logic , given the lemmas ⊢ Pf ( pαq ) → Pf ( pα ∨ βq ) , ⊢ Pf ( pβq ) → Pf ( pα ∨ βq ) and ⊢ Pf ( pαq ) → Pf ( pβq ) → Pf ( pα ∧ βq ) ( 4 ) Proving ( 4 ) directly from the definitions would need colossal efforts , but there is a quick proof . The automation of the HF calculus is good enough to prove tautologies , and from ⊢ α → β → α ∧ β , the proof formalisation condition9 yields ⊢ Pf ( pα → β → α ∧ βq ) Finally , the Hilbert-Bernays derivability condition ( D2 ) yields the desired lemma , ( 4 ) . This trick is needed whenever
|
||
#65 difficult , but are worth examining as a warmup exercise . From the induction hypotheses ⊢ α → Pf ( pαq ) and ⊢ β → Pf ( pβq ) , we must show ⊢ α ∨ β → Pf ( pα ∨ βq ) and ⊢ α ∧ β → Pf ( pα ∧ βq ) . Both of these cases are trivial by propositional logic , given the lemmas ⊢ Pf ( pαq ) → Pf ( pα ∨ βq ) , ⊢ Pf ( pβq ) → Pf ( pα ∨ βq ) and ⊢ Pf ( pαq ) → Pf ( pβq ) → Pf ( pα ∧ βq ) ( 4 ) Proving ( 4 ) directly from the definitions would need colossal efforts , but there is a quick proof . The automation of the HF calculus is good enough to prove tautologies , and from ⊢ α → β → α ∧ β , the proof formalisation condition9 yields ⊢ Pf ( pα → β → α ∧ βq ) Finally , the Hilbert-Bernays derivability condition ( D2 ) yields the desired lemma , ( 4 ) . This trick is needed whenever we need to do propositional reasoning with Pf . 9 Of Sect . 4.4 , but via the substitution theorem quote all PfP ssubst proved above . The induction concerns generalised formulas involving pseudo-coding : PfP ( ssubst ⌊α⌋V V F ) . 31 6.4.2 The Equality and Membership Cases Now comes one of the most critical parts of the formalisation . Many published proofs [ 2 , 32 ] of the second completeness theorem use the following lemma : x = y → Pf px = yq ( 5 ) This in turn is proved using a lemma stating that x = y implies x = y , which is false here : we have defined QuoteP only as a relation , and even { 0 , 1 } can be written in two different ways . Nevertheless , the statement ( 5 ) is clearly true : if x and y are constant terms denoting x and y , respectively , where x = y , then Pf px = yq holds . The equivalence of two different ways of writing a finite set should obviously be provable . This problem recalls the situation described in 3.3 above , and the induction used to prove Subset Mem sf lemma . The solution , once again , is to prove the conjunction ( x ∈ y → Pf px ∈ yq ) ∧ ( x ⊆ y → Pf px ⊆ yq ) by induction ( in HF ) on the sum of the sizes of x and y . The proof is huge ( nearly 340 lines ) , with eight universal quantifiers in the induction formula , each of which must be individually instantiated in order to apply an induction hypothesis . ⊢ All i ( All
|
||
#66 that x = y implies x = y , which is false here : we have defined QuoteP only as a relation , and even { 0 , 1 } can be written in two different ways . Nevertheless , the statement ( 5 ) is clearly true : if x and y are constant terms denoting x and y , respectively , where x = y , then Pf px = yq holds . The equivalence of two different ways of writing a finite set should obviously be provable . This problem recalls the situation described in 3.3 above , and the induction used to prove Subset Mem sf lemma . The solution , once again , is to prove the conjunction ( x ∈ y → Pf px ∈ yq ) ∧ ( x ⊆ y → Pf px ⊆ yq ) by induction ( in HF ) on the sum of the sizes of x and y . The proof is huge ( nearly 340 lines ) , with eight universal quantifiers in the induction formula , each of which must be individually instantiated in order to apply an induction hypothesis . ⊢ All i ( All i ’ ( All si ( All li ( All j ( All j ’ ( All sj ( All lj ( SeqQuoteP ( Var i ) ( Var i ’ ) ( Var si ) ( Var li ) IMP SeqQuoteP ( Var j ) ( Var j ’ ) ( Var sj ) ( Var lj ) IMP HaddP ( Var li ) ( Var lj ) ( Var k ) IMP ( ( Var i IN Var j IMP PfP ( Q Mem ( Var i ’ ) ( Var j ’ ) ) ) AND ( Var i SUBS Var j IMP PfP ( Q Subset ( Var i ’ ) ( Var j ’ ) ) ) ) ) ) ) ) ) ) ) ) '' Using SeqQuoteP ( which describes the sequence computation of QuoteP ) gives access to a size measure for the two terms , which are here designated i and j . Their sizes , li and lj , are added using HaddP , which is simply addition as defined in the HF calculus . ( This formalisation of addition is also needed for reasoning about sequences . ) Their sum , k , is used as the induction variable . Although the second half of the conjunction suffices to prove ( 5 ) , it is never needed outside of the induction , and neither is ( 5 ) itself . All we need is ( 3 ) . And so we reach the next milestone . lemma assumes `` atom i ♯ ( j , j ’ , i ’ ) '' `` atom i ’ ♯ ( j , j ’ ) '' `` atom j ♯ ( j ’ ) '' shows QuoteP Mem imp QMem
|
||
#67 Var i SUBS Var j IMP PfP ( Q Subset ( Var i ’ ) ( Var j ’ ) ) ) ) ) ) ) ) ) ) ) ) '' Using SeqQuoteP ( which describes the sequence computation of QuoteP ) gives access to a size measure for the two terms , which are here designated i and j . Their sizes , li and lj , are added using HaddP , which is simply addition as defined in the HF calculus . ( This formalisation of addition is also needed for reasoning about sequences . ) Their sum , k , is used as the induction variable . Although the second half of the conjunction suffices to prove ( 5 ) , it is never needed outside of the induction , and neither is ( 5 ) itself . All we need is ( 3 ) . And so we reach the next milestone . lemma assumes `` atom i ♯ ( j , j ’ , i ’ ) '' `` atom i ’ ♯ ( j , j ’ ) '' `` atom j ♯ ( j ’ ) '' shows QuoteP Mem imp QMem : '' { QuoteP ( Var i ) ( Var i ’ ) , QuoteP ( Var j ) ( Var j ’ ) , Var i IN Var j } ⊢ PfP ( Q Mem ( Var i ’ ) ( Var j ’ ) ) '' and QuoteP Mem imp QSubset : '' { QuoteP ( Var i ) ( Var i ’ ) , QuoteP ( Var j ) ( Var j ’ ) , Var i SUBS Var j } ⊢ PfP ( Q Subset ( Var i ’ ) ( Var j ’ ) ) '' Turning to the main induction on α , the notoriously “ delicate ” [ 2 , p. 48 ] case is bounded universal quantification . Many of the delicate points here are connected with the way the nominal approach is used . We need to maintain and extend a permutation relating old and new variable names . Such complexities are evident in mathematical texts , in their treatment of variable names [ 32 , Lemma 9.7 ] . lemma ( in quote perm ) quote all Mem imp All2 : assumes IH : `` insert ( QuoteP ( Var i ) ( Var i ’ ) ) ( quote all p Vs ) ⊢ α IMP PfP ( ssubst ⌊α⌋ ( insert i Vs ) ( insert i Vs ) Fi ) '' and `` supp ( All2 i ( Var j ) α ) ⊆ atom ‘ Vs '' 32 and j : `` atom j ♯ ( i , α ) '' and i : `` atom i ♯ p '' and i ’ : `` atom i ’ ♯ ( i , p , α ) '' shows `` insert ( All2 i ( Var j
|
||
#68 '' Turning to the main induction on α , the notoriously “ delicate ” [ 2 , p. 48 ] case is bounded universal quantification . Many of the delicate points here are connected with the way the nominal approach is used . We need to maintain and extend a permutation relating old and new variable names . Such complexities are evident in mathematical texts , in their treatment of variable names [ 32 , Lemma 9.7 ] . lemma ( in quote perm ) quote all Mem imp All2 : assumes IH : `` insert ( QuoteP ( Var i ) ( Var i ’ ) ) ( quote all p Vs ) ⊢ α IMP PfP ( ssubst ⌊α⌋ ( insert i Vs ) ( insert i Vs ) Fi ) '' and `` supp ( All2 i ( Var j ) α ) ⊆ atom ‘ Vs '' 32 and j : `` atom j ♯ ( i , α ) '' and i : `` atom i ♯ p '' and i ’ : `` atom i ’ ♯ ( i , p , α ) '' shows `` insert ( All2 i ( Var j ) α ) ( quote all p Vs ) ⊢ PfP ( ssubst ⌊All2 i ( Var j ) α⌋Vs Vs F ) '' The final case , for the existential quantifier , is also somewhat complicated to formalise . The details are again mostly connected with managing free and bound variable names using nominal methods , and are therefore omitted . We can conclude our discussion of the inductive argument by viewing the final result : lemma star : assumes `` ss fm α '' `` finite V '' `` supp α ⊆ atom ‘ V '' shows `` insert α ( quote all p V ) ⊢ PfP ( ssubst ⌊α⌋V V F ) '' Condition ( D3 ) now follows easily , since the formula α is then a sentence . Although some technical conditions ( involving variable names and permutations ) have been omitted from the previous two theorems , our main result below appears exactly as it was proved . Of course , α ⊢ Pf pαq is equivalent to ⊢ α → Pf pαq . theorem Provability : assumes `` Sigma fm α '' `` ground fm α '' shows `` { α } ⊢ PfP ⌈α⌉ '' 7 G¨odel ’ s Second Incompleteness Theorem The final steps of the second incompleteness theorem can be seen in Fig . 2 . The diagonal formula , δ , is obtained via the first incompleteness theorem . Then we can quickly establish both Pf pδq ⊢ Pf ( pPf pδqq ) and Pf pδq ⊢ Pf ( p¬Pf pδqq ) . These together imply Pf pδq ⊢ Pf ( p⊥q ) using a variant of condition ( D2 ) . It follows that if the system proves its own consistency , then it also proves ⊢
|
||
#69 insert α ( quote all p V ) ⊢ PfP ( ssubst ⌊α⌋V V F ) '' Condition ( D3 ) now follows easily , since the formula α is then a sentence . Although some technical conditions ( involving variable names and permutations ) have been omitted from the previous two theorems , our main result below appears exactly as it was proved . Of course , α ⊢ Pf pαq is equivalent to ⊢ α → Pf pαq . theorem Provability : assumes `` Sigma fm α '' `` ground fm α '' shows `` { α } ⊢ PfP ⌈α⌉ '' 7 G¨odel ’ s Second Incompleteness Theorem The final steps of the second incompleteness theorem can be seen in Fig . 2 . The diagonal formula , δ , is obtained via the first incompleteness theorem . Then we can quickly establish both Pf pδq ⊢ Pf ( pPf pδqq ) and Pf pδq ⊢ Pf ( p¬Pf pδqq ) . These together imply Pf pδq ⊢ Pf ( p⊥q ) using a variant of condition ( D2 ) . It follows that if the system proves its own consistency , then it also proves ⊢ ¬Pf pδq and therefore ⊢ δ , a contradiction . Swierczkowski [ 32 ] presents a few other results which have n ´ ot been formalised here . These include a refinement of the incompleteness theorem ( credited to Reinhardt ) and a theory of a linear order on the HF sets , but recall that claim ( 5 ) can be proved without using any such ordering . The approach adopted here undoubtedly involves less effort than formalising the ordering in the HF calculus . The total proof length of nearly 12,400 lines comprises around 4,600 lines for the second theorem and 7,700 lines for the first.10 ( One could also include 2,700 lines for HF set theory itself , but we would not count the standard libraries of natural numbers if they were used as the basis of the proof . ) O ’ Connor ’ s proof comprises 47,000 lines of Coq , while Shankar ’ s takes 20,000 lines [ 30 , p. 139 ] and Harrison ’ s [ 10 ] is a miniscule 4,400 lines of HOL Light . Recall that none of these other proofs include the second incompleteness theorem . But comparisons are almost meaningless because of the enormous differences among the formalisations . Shankar wrote ( and proved to be representable ) a LISP interpreter for coding up the metatheory [ 30 ] ; this was a huge effort , but then the various coding functions could then be written in LISP without further justification . He also used HF for coding , presumably because of its similarity to LISP S-expressions . O ’ Connor formalised a very general syntax for first-order logic [ 22 ] . He introduced a general inductive definition of the primitive recursive functions , but proving
|
||
#70 second theorem and 7,700 lines for the first.10 ( One could also include 2,700 lines for HF set theory itself , but we would not count the standard libraries of natural numbers if they were used as the basis of the proof . ) O ’ Connor ’ s proof comprises 47,000 lines of Coq , while Shankar ’ s takes 20,000 lines [ 30 , p. 139 ] and Harrison ’ s [ 10 ] is a miniscule 4,400 lines of HOL Light . Recall that none of these other proofs include the second incompleteness theorem . But comparisons are almost meaningless because of the enormous differences among the formalisations . Shankar wrote ( and proved to be representable ) a LISP interpreter for coding up the metatheory [ 30 ] ; this was a huge effort , but then the various coding functions could then be written in LISP without further justification . He also used HF for coding , presumably because of its similarity to LISP S-expressions . O ’ Connor formalised a very general syntax for first-order logic [ 22 ] . He introduced a general inductive definition of the primitive recursive functions , but proving specific functions to be primitive recursive turned out to be extremely difficult [ 23 , Sect . 5.3 ] . Harrison has 10 Prior to polishing and removing unused material , the proof totalled 17,000 lines . 33 theorem Goedel II : assumes `` ¬ { } ⊢ Fls '' shows `` ¬ { } ⊢ Neg ( PfP ⌈Fls⌉ ) '' proof - from assms Goedel I obtain δ where diag : `` { } ⊢ δ IFF Neg ( PfP ⌈δ⌉ ) '' `` ¬ { } ⊢ δ '' and gnd : `` ground fm δ '' by metis have `` { PfP ⌈δ⌉ } ⊢ PfP ⌈PfP ⌈δ⌉⌉ '' by ( auto simp : Provability ground fm aux def supp conv fresh ) moreover have `` { PfP ⌈δ⌉ } ⊢ PfP ⌈Neg ( PfP ⌈δ⌉ ) ⌉ '' apply ( rule MonPon PfP implies PfP [ OF gnd ] ) apply ( metis Conj E2 Iff def Iff sym diag ( 1 ) ) apply ( auto simp : ground fm aux def supp conv fresh ) done moreover have `` ground fm ( PfP ⌈δ⌉ ) '' by ( auto simp : ground fm aux def supp conv fresh ) ultimately have `` { PfP ⌈δ⌉ } ⊢ PfP ⌈Fls⌉ '' using PfP quot contra by ( metis ( no types ) anti deduction cut2 ) thus `` ¬ { } ⊢ Neg ( PfP ⌈Fls⌉ ) '' by ( metis Iff MP2 same Neg mono cut1 diag ) qed Fig . 2 G¨odel ’ s Second Incompleteness Theorem not published a paper describing his formalisation , but devotes a few pages to G¨odel ’ s theorems in his Handbook of Practical Logic [ 12 , p. 546–555 ] , including extracts of HOL
|
||
#71 by metis have `` { PfP ⌈δ⌉ } ⊢ PfP ⌈PfP ⌈δ⌉⌉ '' by ( auto simp : Provability ground fm aux def supp conv fresh ) moreover have `` { PfP ⌈δ⌉ } ⊢ PfP ⌈Neg ( PfP ⌈δ⌉ ) ⌉ '' apply ( rule MonPon PfP implies PfP [ OF gnd ] ) apply ( metis Conj E2 Iff def Iff sym diag ( 1 ) ) apply ( auto simp : ground fm aux def supp conv fresh ) done moreover have `` ground fm ( PfP ⌈δ⌉ ) '' by ( auto simp : ground fm aux def supp conv fresh ) ultimately have `` { PfP ⌈δ⌉ } ⊢ PfP ⌈Fls⌉ '' using PfP quot contra by ( metis ( no types ) anti deduction cut2 ) thus `` ¬ { } ⊢ Neg ( PfP ⌈Fls⌉ ) '' by ( metis Iff MP2 same Neg mono cut1 diag ) qed Fig . 2 G¨odel ’ s Second Incompleteness Theorem not published a paper describing his formalisation , but devotes a few pages to G¨odel ’ s theorems in his Handbook of Practical Logic [ 12 , p. 546–555 ] , including extracts of HOL Light proofs . He defines Σ1 and Π1 formulas and quotes some meta-theoretical results relating truth and provability . This project took approximately one year , in time left available after fulfilling teaching and administrative duties . The underlying set theory took only two weeks to formalise . The G¨odel development up to the proof formalisation condition took another five months . From there to the first incompleteness theorem took a further two months , mostly devoted to proving single valued properties . Then the second incompleteness theorem took a further four months , including much time wasted due to misunderstanding this perplexing material . Some material has since been consolidated or streamlined . The final version is available online [ 26 ] . 7.1 The Lengths of Proofs Figure 3 depicts the sizes of the Isabelle/HOL theories making up various sections of the proof development . The first theorem takes up the bulk of the effort . Apart from the massive HF proofs about predicates , which are mostly of obvious properties , the second theorem appears to be a fairly easy step given the first . Why then has it not been formalised until now ? A reasonable guess is that previous researchers were not aware of Swierczkowski ’ s [ 32 ] elaborate development . The most deta ´ iled of the previous proofs [ 2 , 9 ] left too much to the imagination , and even Swierczkowski makes some errors . He ´ devotes much of his Sect . 7 to proofs concerning substitution for ( non-existent ) pseudoterms analogous to x . Recall that pseudo-terms are merely a notational shorthand to allow function syntax , and are not actual terms . Finally , there was the critical issue of the ordering on the HF
|
||
#72 misunderstanding this perplexing material . Some material has since been consolidated or streamlined . The final version is available online [ 26 ] . 7.1 The Lengths of Proofs Figure 3 depicts the sizes of the Isabelle/HOL theories making up various sections of the proof development . The first theorem takes up the bulk of the effort . Apart from the massive HF proofs about predicates , which are mostly of obvious properties , the second theorem appears to be a fairly easy step given the first . Why then has it not been formalised until now ? A reasonable guess is that previous researchers were not aware of Swierczkowski ’ s [ 32 ] elaborate development . The most deta ´ iled of the previous proofs [ 2 , 9 ] left too much to the imagination , and even Swierczkowski makes some errors . He ´ devotes much of his Sect . 7 to proofs concerning substitution for ( non-existent ) pseudoterms analogous to x . Recall that pseudo-terms are merely a notational shorthand to allow function syntax , and are not actual terms . Finally , there was the critical issue of the ordering on the HF sets . Solving these mysteries required much thought , and the 34 first completed formalisation included thousands of lines of proofs belonging to aborted attempts . A discussion of the de Bruijn coefficient ( the expansion in size entailed by the process of formalisation ) is always interesting , but difficult to do rigorously . In this case , the formalisation required proving a great many theorems that were not even hinted at in the source text , for example , setting up a usable sequent calculus ( Swierczkowski ´ merely gives the rudiments of a Hilbert system ) , or proving that the various codings of syntax actually work ( virtually all authors take this for granted ) , or proving that “ functions ” are functions . The hundreds of lines of single-step HF calculus proofs are the single largest contributor to the size of this development , and such things never appear in standard presentations of the incompleteness theorems . A crude calculation yields 30 pages at 35 lines per page or 1050 lines for Swier- ´ czkowski ’ s proof , compared with 12,400 lines of Isabelle , for a de Bruijn factor of 12 . Focusing on just the proof of the first incompleteness theorem ( after the preliminary developments ) , we have about 80 lines of informal text and 671 lines of Isabelle , giving a factor of 8.4 ; that includes 150 lines in the Isabelle script to prove that functions are single valued . A further issue is the heavy use of cut and paste in the HF calculus proofs . Better automation for HF could help , but spending time to develop a tactic that will only be called a few times is hard to justify . An alternative idea
|
||
#73 or proving that the various codings of syntax actually work ( virtually all authors take this for granted ) , or proving that “ functions ” are functions . The hundreds of lines of single-step HF calculus proofs are the single largest contributor to the size of this development , and such things never appear in standard presentations of the incompleteness theorems . A crude calculation yields 30 pages at 35 lines per page or 1050 lines for Swier- ´ czkowski ’ s proof , compared with 12,400 lines of Isabelle , for a de Bruijn factor of 12 . Focusing on just the proof of the first incompleteness theorem ( after the preliminary developments ) , we have about 80 lines of informal text and 671 lines of Isabelle , giving a factor of 8.4 ; that includes 150 lines in the Isabelle script to prove that functions are single valued . A further issue is the heavy use of cut and paste in the HF calculus proofs . Better automation for HF could help , but spending time to develop a tactic that will only be called a few times is hard to justify . An alternative idea is to define higher-order operators for the sequence constructions , which could be proved to be functional once and for all . However , higher-order operators are difficult to define using nominal syntax . Perhaps it could be attempted using naive variables . Fig . 3 Sizes of Constituent Theories 7.2 The Formalisation of Variable Binding The role of bound variable syntax remains unclear . Shankar [ 30 , 31 ] used de Bruijn variables to formalise the Church-Rosser theorem but not the incompleteness theorem . Harrison did not use them either . O ’ Connor also used traditional syntactic bound variables , but complained about huge complications concerning substitution ( recall Sect . 2.3 ) . The present development uses the nominal package , easing many proofs , but at a price : over 900 lines involve freshness specifications , and around 70 lemmas involving freshness are proved . Moreover , many proof steps are slow . While the project was 35 underway , proof times taking minutes were not unusual . Even after recent improvements to the nominal package , they can take tens of seconds . Additional performance improvements would be welcome , as well as a more concise notation for freshness conditions when many new names are needed . In fairness to the nominal approach , explicit variable names would also have to be fresh and analogous assertions would be necessary , along with some procedure for calculating new names numerically and proving them to be fresh . The effort may be similar either way , but the nominal approach is more abstract and natural : who after all refers to specific , calculated variable names in textbook proofs ? My first attempt to formalise the incompleteness theorems used explicit names . Then , substitution on
|
||
#74 bound variables , but complained about huge complications concerning substitution ( recall Sect . 2.3 ) . The present development uses the nominal package , easing many proofs , but at a price : over 900 lines involve freshness specifications , and around 70 lemmas involving freshness are proved . Moreover , many proof steps are slow . While the project was 35 underway , proof times taking minutes were not unusual . Even after recent improvements to the nominal package , they can take tens of seconds . Additional performance improvements would be welcome , as well as a more concise notation for freshness conditions when many new names are needed . In fairness to the nominal approach , explicit variable names would also have to be fresh and analogous assertions would be necessary , along with some procedure for calculating new names numerically and proving them to be fresh . The effort may be similar either way , but the nominal approach is more abstract and natural : who after all refers to specific , calculated variable names in textbook proofs ? My first attempt to formalise the incompleteness theorems used explicit names . Then , substitution on formulas was only available as a relation , and many proofs required numerical operations on variable names . This effort would have multiplied considerably for the second incompleteness theorem . Using de Bruijn indices for HF syntax was not attractive : I had previously formalised G¨odel ’ s definition of the constructible sets and his proof of the relative consistency of the axiom of choice [ 25 ] . Here it was also necessary to define a great many predicates within an encoding of first-order logic . This work was done in Isabelle/ZF , a version of Isabelle for axiomatic set theory . I used de Bruijn indices in these definitions , but the loss of readability was a severe impediment to progress . It is worth investigating how this formalisation would be affected by the change to another treatment of variable binding . As regards the G¨odel numbering of formulas , the use of de Bruijn variables can be called an unqualified success . It was easy to set up and all necessary properties were proved without great difficulties . 7.3 On Verifying Proof Assistants In a paper entitled “ Towards Self-verification of HOL Light ” , Harrison says , G¨odel ’ s second incompleteness theorem tells us that [ a logical system ] can not prove its own consistency in any way at all . . . . So , regardless of implementation details , if we want to prove the consistency of a proof checker , we need to use a logic that in at least some respects goes beyond the logic the checker itself supports . [ 11 , p. 179 ] This statement is potentially misleading , and has given rise to the mistaken view that it is impossible to verify a proof checker in
|
||
#75 set theory . I used de Bruijn indices in these definitions , but the loss of readability was a severe impediment to progress . It is worth investigating how this formalisation would be affected by the change to another treatment of variable binding . As regards the G¨odel numbering of formulas , the use of de Bruijn variables can be called an unqualified success . It was easy to set up and all necessary properties were proved without great difficulties . 7.3 On Verifying Proof Assistants In a paper entitled “ Towards Self-verification of HOL Light ” , Harrison says , G¨odel ’ s second incompleteness theorem tells us that [ a logical system ] can not prove its own consistency in any way at all . . . . So , regardless of implementation details , if we want to prove the consistency of a proof checker , we need to use a logic that in at least some respects goes beyond the logic the checker itself supports . [ 11 , p. 179 ] This statement is potentially misleading , and has given rise to the mistaken view that it is impossible to verify a proof checker in its own logic . Harrison ’ s aim is to prove that HOL Light can not prove the theorem FALSE , and this indeed requires proving the consistency of higher-order logic itself . Unfortunately , most consistency proofs are unsatisfactory because they more or less assume the desired conclusion : they are thinly disguised versions of the tautology Con ( L ) ∧ L → Con ( L ) . This is a consequence of the second incompleteness theorem , since the consistency of L can only be proved in a strictly stronger formal system . Mathematicians accept strong formal systems , such as ZF set theory , with little justification other than intuition and experience . Moreover , they examine very strong further axioms . The axiom of constructibility is an instructive case : it is known to be relatively consistent with respect to the axioms of set theory , but it is not generally accepted as true [ 16 , p. 170 ] . The standard ZF axioms are generally regarded as true , although they can not even be proved to be consistent . Thus we have no good way of proving consistency , and yet consistency does not guarantee truth . 36 This situation calls for a separation of concerns . The builders of verification tools should be concerned with the correctness of their code , but the correctness of the underlying formal calculus is the concern of logicians . Harrison notes that “ almost all implementation bugs in HOL Light and other versions of HOL have involved variable renaming ” [ 11 , p. 179 ] , and this type of issue should be our focus . Verifying a proof assistant involves verifying that it implements a data structure for the assertions of the formal
|
||
#76 systems , such as ZF set theory , with little justification other than intuition and experience . Moreover , they examine very strong further axioms . The axiom of constructibility is an instructive case : it is known to be relatively consistent with respect to the axioms of set theory , but it is not generally accepted as true [ 16 , p. 170 ] . The standard ZF axioms are generally regarded as true , although they can not even be proved to be consistent . Thus we have no good way of proving consistency , and yet consistency does not guarantee truth . 36 This situation calls for a separation of concerns . The builders of verification tools should be concerned with the correctness of their code , but the correctness of the underlying formal calculus is the concern of logicians . Harrison notes that “ almost all implementation bugs in HOL Light and other versions of HOL have involved variable renaming ” [ 11 , p. 179 ] , and this type of issue should be our focus . Verifying a proof assistant involves verifying that it implements a data structure for the assertions of the formal calculus and that it satisfies a commuting diagram relating deductions on the implemented assertions with the corresponding deductions in the calculus . G¨odel ’ s theorems have no relevance here . 8 Conclusions The main finding is simply that G¨odel ’ s second incompleteness theorem can be proved with a relatively modest effort , in only a few months starting with a proof of the first incompleteness theorem . While the nominal approach to syntax is clearly not indispensable , it copes convincingly with a development of this size and complexity . The use of HF set theory as an alternative to Peano arithmetic is clearly justified , eliminating the need to formalise basic number theory within the embedded calculus ; the necessary effort to do that would greatly exceed the difficulties ( mentioned in Sect . 6.4 above ) caused by the lack of a simple canonical ordering on HF sets . Many published proofs of the incompleteness theorems replace technical proofs by vague appeals to Church ’ s thesis . Boolos [ 2 ] presents a more detailed and careful exposition , but still leaves substantial gaps . Even the source text [ 32 ] for this project , although written with great care , had problems : a significant gap ( concerning the canonical ordering of HF sets ) , a few minor ones ( concerning Σ formulas , for example ) , and pages of material that are , at the very least , misleading . These remarks are not intended as criticism but as objective observations of the complexity of this material , with its codings of codings . A complete formal proof , written in a fairly readable notation , should greatly clarify the issues involved in these crucially important theorems . Acknowledgment Jesse
|
||
#77 alternative to Peano arithmetic is clearly justified , eliminating the need to formalise basic number theory within the embedded calculus ; the necessary effort to do that would greatly exceed the difficulties ( mentioned in Sect . 6.4 above ) caused by the lack of a simple canonical ordering on HF sets . Many published proofs of the incompleteness theorems replace technical proofs by vague appeals to Church ’ s thesis . Boolos [ 2 ] presents a more detailed and careful exposition , but still leaves substantial gaps . Even the source text [ 32 ] for this project , although written with great care , had problems : a significant gap ( concerning the canonical ordering of HF sets ) , a few minor ones ( concerning Σ formulas , for example ) , and pages of material that are , at the very least , misleading . These remarks are not intended as criticism but as objective observations of the complexity of this material , with its codings of codings . A complete formal proof , written in a fairly readable notation , should greatly clarify the issues involved in these crucially important theorems . Acknowledgment Jesse Alama drew my attention to Swierczkowski [ 32 ] , the source material for this ´ project . Christian Urban assisted with nominal aspects of some of the proofs , even writing code . Brian Huffman provided the core formalisation of type hf . Dana Scott offered advice and drew my attention to Kirby [ 15 ] . Matt Kaufmann and the referees made many insightful comments . References [ 1 ] Joan Bagaria . A short guide to G¨odel ’ s second incompleteness theorem . Teorema , 22 ( 3 ) :5–15 , 2003 . [ 2 ] George Stephen Boolos . The Logic of Provability . Cambridge University Press , 1993 . [ 3 ] N. G. de Bruijn . Lambda calculus notation with nameless dummies , a tool for automatic formula manipulation , with application to the Church-Rosser Theorem . Indagationes Mathematicae , 34:381–392 , 1972 . [ 4 ] S. Feferman , editor . Kurt G¨odel : Collected Works , volume I. Oxford University Press , 1986 . 37 [ 5 ] Torkel Franz´en . G¨odel ’ s Theorem : An Incomplete Guide to Its Use and Abuse . A K Peters , 2005 . [ 6 ] M. J. Gabbay and A. M. Pitts . A new approach to abstract syntax with variable binding . Formal Aspects of Computing , 13:341–363 , 2001 . [ 7 ] Kurt G¨odel . On completeness and consistency . In Feferman [ 4 ] , pages 234–236 . [ 8 ] Kurt G¨odel . On formally undecidable propositions of Principia Mathematica and related systems . In Feferman [ 4 ] , pages 144–195 . First published in 1931 in the Monatshefte f¨ur Mathematik und Physik . [ 9 ] Richard E Grandy . Advanced Logic for Applications . Reidel
|
||
#78 Stephen Boolos . The Logic of Provability . Cambridge University Press , 1993 . [ 3 ] N. G. de Bruijn . Lambda calculus notation with nameless dummies , a tool for automatic formula manipulation , with application to the Church-Rosser Theorem . Indagationes Mathematicae , 34:381–392 , 1972 . [ 4 ] S. Feferman , editor . Kurt G¨odel : Collected Works , volume I. Oxford University Press , 1986 . 37 [ 5 ] Torkel Franz´en . G¨odel ’ s Theorem : An Incomplete Guide to Its Use and Abuse . A K Peters , 2005 . [ 6 ] M. J. Gabbay and A. M. Pitts . A new approach to abstract syntax with variable binding . Formal Aspects of Computing , 13:341–363 , 2001 . [ 7 ] Kurt G¨odel . On completeness and consistency . In Feferman [ 4 ] , pages 234–236 . [ 8 ] Kurt G¨odel . On formally undecidable propositions of Principia Mathematica and related systems . In Feferman [ 4 ] , pages 144–195 . First published in 1931 in the Monatshefte f¨ur Mathematik und Physik . [ 9 ] Richard E Grandy . Advanced Logic for Applications . Reidel , 1977 . [ 10 ] John Harrison . Re : Re : G¨odel ’ s incompleteness theorem . Email dated 15 January 2014 . [ 11 ] John Harrison . Towards self-verification of HOL Light . In Ulrich Furbach and Natarajan Shankar , editors , Automated Reasoning — Third International Joint Conference , IJCAR 2006 , LNAI 4130 , pages 177–191 . Springer , 2006 . [ 12 ] John Harrison . Handbook of Practical Logic and Automated Reasoning . Cambridge University Press , 2009 . [ 13 ] Richard E Hodel . An Introduction to Mathematical Logic . PWS Publishing Company , 1995 . [ 14 ] Joe Hurd and Tom Melham , editors . Theorem Proving in Higher Order Logics : TPHOLs 2005 , LNCS 3603 . Springer , 2005 . [ 15 ] Laurence Kirby . Addition and multiplication of sets . Mathematical Logic Quarterly , 53 ( 1 ) :52–65 , 2007 . [ 16 ] Kenneth Kunen . Set Theory : An Introduction to Independence Proofs . North-Holland , 1980 . [ 17 ] Andreas Lochbihler . Formalising finfuns — generating code for functions as data from Isabelle/HOL . In Stefan Berghofer , Tobias Nipkow , Christian Urban , and Makarius Wenzel , editors , TPHOLs , volume 5674 of Lecture Notes in Computer Science , pages 310–326 . Springer , 2009 . [ 18 ] Tobias Nipkow . More Church-Rosser proofs ( in Isabelle/HOL ) . Journal of Automated Reasoning , 26:51–66 , 2001 . [ 19 ] Tobias Nipkow and Lawrence C. Paulson . Proof pearl : Defining functions over finite sets . In Hurd and Melham [ 14 ] , pages 385–396 . [ 20 ] Tobias Nipkow , Lawrence C. Paulson , and Markus Wenzel . Isabelle/HOL : A Proof
|
||
#79 PWS Publishing Company , 1995 . [ 14 ] Joe Hurd and Tom Melham , editors . Theorem Proving in Higher Order Logics : TPHOLs 2005 , LNCS 3603 . Springer , 2005 . [ 15 ] Laurence Kirby . Addition and multiplication of sets . Mathematical Logic Quarterly , 53 ( 1 ) :52–65 , 2007 . [ 16 ] Kenneth Kunen . Set Theory : An Introduction to Independence Proofs . North-Holland , 1980 . [ 17 ] Andreas Lochbihler . Formalising finfuns — generating code for functions as data from Isabelle/HOL . In Stefan Berghofer , Tobias Nipkow , Christian Urban , and Makarius Wenzel , editors , TPHOLs , volume 5674 of Lecture Notes in Computer Science , pages 310–326 . Springer , 2009 . [ 18 ] Tobias Nipkow . More Church-Rosser proofs ( in Isabelle/HOL ) . Journal of Automated Reasoning , 26:51–66 , 2001 . [ 19 ] Tobias Nipkow and Lawrence C. Paulson . Proof pearl : Defining functions over finite sets . In Hurd and Melham [ 14 ] , pages 385–396 . [ 20 ] Tobias Nipkow , Lawrence C. Paulson , and Markus Wenzel . Isabelle/HOL : A Proof Assistant for Higher-Order Logic . Springer , 2002 . An up-to-date version is distributed with Isabelle . [ 21 ] Michael Norrish and Ren´e Vestergaard . Proof pearl : de Bruijn terms really do work . In Klaus Schneider and Jens Brandt , editors , Theorem Proving in Higher Order Logics : TPHOLs 2007 , LNCS 4732 , pages 207–222 . Springer , 2007 . [ 22 ] Russell O ’ Connor . Essential incompleteness of arithmetic verified by Coq . In Hurd and Melham [ 14 ] , pages 245–260 . [ 23 ] Russell S. S. O ’ Connor . Incompleteness & Completeness : Formalizing Logic and Analysis in Type Theory . PhD thesis , Radboud University Nijmegen , 2009 . [ 24 ] Lawrence C. Paulson . Set theory for verification : I . From foundations to functions . Journal of Automated Reasoning , 11 ( 3 ) :353–389 , 1993 . [ 25 ] Lawrence C. Paulson . The relative consistency of the axiom of choice — mechanized using Isabelle/ZF . LMS Journal of Computation and Mathematics , 6:198–248 , 2003. http : //www.lms.ac.uk/jcm/6/lms2003-001/ . [ 26 ] Lawrence C. Paulson . G¨odel ’ s incompleteness theorems . Archive of Formal Proofs , November 2013. http : //afp.sf.net/entries/Incompleteness.shtml , Formal proof development . [ 27 ] Lawrence C. Paulson . A machine-assisted proof of G¨odel ’ s incompleteness theorems for the theory of hereditarily finite sets . Review of Symbolic Logic , 7 ( 3 ) :484–498 , September 2014 . [ 28 ] Andrew M. Pitts . Nominal Sets : Names and Symmetry in Computer Science . Cambridge University Press , 2013 . [ 29 ] Natarajan Shankar . Proof-checking Metamathematics . PhD thesis , University of Texas at Austin , 1986 . [ 30
|
||
#80 Connor . Incompleteness & Completeness : Formalizing Logic and Analysis in Type Theory . PhD thesis , Radboud University Nijmegen , 2009 . [ 24 ] Lawrence C. Paulson . Set theory for verification : I . From foundations to functions . Journal of Automated Reasoning , 11 ( 3 ) :353–389 , 1993 . [ 25 ] Lawrence C. Paulson . The relative consistency of the axiom of choice — mechanized using Isabelle/ZF . LMS Journal of Computation and Mathematics , 6:198–248 , 2003. http : //www.lms.ac.uk/jcm/6/lms2003-001/ . [ 26 ] Lawrence C. Paulson . G¨odel ’ s incompleteness theorems . Archive of Formal Proofs , November 2013. http : //afp.sf.net/entries/Incompleteness.shtml , Formal proof development . [ 27 ] Lawrence C. Paulson . A machine-assisted proof of G¨odel ’ s incompleteness theorems for the theory of hereditarily finite sets . Review of Symbolic Logic , 7 ( 3 ) :484–498 , September 2014 . [ 28 ] Andrew M. Pitts . Nominal Sets : Names and Symmetry in Computer Science . Cambridge University Press , 2013 . [ 29 ] Natarajan Shankar . Proof-checking Metamathematics . PhD thesis , University of Texas at Austin , 1986 . [ 30 ] Natarajan Shankar . Metamathematics , Machines , and G¨odel ’ s Proof . Cambridge University Press , 1994 . [ 31 ] Natarajan Shankar . Shankar , Boyer , Church-Rosser and de Bruijn indices . E-mail , 2013 . 38 [ 32 ] S. Swierczkowski . Finite sets and G¨odel ’ s incompleteness th ´ eorems . Dissertationes Mathematicae , 422:1–58 , 2003. http : //journals.impan.gov.pl/dm/Inf/422-0-1.html . [ 33 ] Christian Urban . Nominal techniques in Isabelle/HOL . Journal of Automated Reasoning , 40 ( 4 ) :327–356 , 2008 . [ 34 ] Christian Urban and Cezary Kaliszyk . General bindings and alpha-equivalence in Nominal Isabelle . Logical Methods in Computer Science , 8 ( 2:14 ) :1–35 , 2012 .
|
||
#81 . General bindings and alpha-equivalence in Nominal Isabelle . Logical Methods in Computer Science , 8 ( 2:14 ) :1–35 , 2012 .
|