Admissible Tools in the Kitchen of Intuitionistic Logic

10/17/2018 ∙ by Andrea Condoluci, et al. ∙ University of Bologna 0

The usual reading of logical implication "A implies B" as "if A then B" fails in intuitionistic logic: there are formulas A and B such that "A implies B" is not provable, even though B is provable whenever A is provable. Intuitionistic rules apparently do not capture interesting meta-properties of the logic and, from a computational perspective, the programs corresponding to intuitionistic proofs are not powerful enough. Such non-provable implications are nevertheless admissible, and we study their behavior by means of a proof term assignment and related rules of reduction. We introduce V, a calculus that is able to represent admissible inferences, while remaining in the intuitionistic world by having normal forms that are just intuitionistic terms. We then extend intuitionistic logic with principles corresponding to admissible rules. As an example, we consider the Kreisel-Putnam logic KP, for which we prove the strong normalization and the disjunction property through our term assignment. This is our first step in understanding the essence of admissible rules for intuitionistic logic.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

page 2

page 3

page 4

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

Proof systems are usually presented inductively by giving axioms and rules of inference, which are respectively the ingredients and the tools for cooking new proofs. For example, when presenting classical propositional logic (CPC) in natural deduction, for each of the usual connectives one gives a set of standard tools to introduce or remove a connective from a formula in order to obtain a proof.

In their most essential form, we can represent rules as an inference (read “from infer ”) where are schemata of logic formulas. A rule is said to be admissible in a proof system if it is in a way redundant, i.e. whenever are provable, then is already provable without using that rule. Adding or dropping rules may increase or decrease the amount of proofs we can cook in a proof system. The effect can be dramatic: for example, classical propositional logic CPC can be obtained by simply adding the rule of double negation elimination () to intuitionistic propositional logic IPC. Admissible rules are all the opposite: if we decide to utilize one in order to cook something, then we could have just used our ingredients in a different way to reach the same result.

One appealing feature of CPC is the fact that it is structurally complete: all its admissible rules are derivable, in the sense that whenever is an admissible rule, then also the corresponding principle is provable [5]i.e. the system acknowledges that there’s no need for that additional tool, so we can internalize it and use the old tools to complete our reasoning. This is not the case in intuitionistic logic: the mere fact that we know that the tool was not needed, doesn’t give us any way to show inside the system why is that. On the other hand, IPC has other wonderful features. Relevant here is the disjunction property, fundamental for a constructive system: when a disjunction is provable, then one of the disjuncts or is provable as well. Our interest is in these intuitionistic admissible rules that are not derivable, in the computational principles they describe, and in the logic systems obtained by explicitly adding such rules to IPC.

Can one effectively identify all intuitionistic admissible rules? The question of whether that set of rules is recursively enumerable was posed by Friedman in 1975, and answered positively by Rybakov in 1984. It was then de Jongh and Visser who exhibited a numerable set of rules (now known as Visser’s rules) and conjectured that it formed a basis for all the admissible rules of IPC. This conjecture was later proved by Iemhoff in the fundamental [7]. Rozière in his Ph.D. thesis [10] reached the same conclusion with a substantially different technique, independently of Visser and Iemhoff. These works elegantly settled the problem of identifying and building admissible rules. However our question is different: why are these rules superfluous, and what reduction steps can eliminate them from proofs?

Rozière first posed the question of finding a computational correspondence for his basis of the admissible rules in the conclusion of his thesis, but no work has been done on this ever since. Natural deduction provides a powerful tool to analyse the computational behaviour of logical axioms, thanks to the fact that it gives a simple way to translate axioms into rules and to develop correspondences with -calculi. Our plan is therefore to understand the phenomenon of admissibility by equipping proofs with -terms and associated reductions in the spirit of the Curry-Howard correspondence. Normalization will show explicitly what role admissible rules play in a proof.

1.1 Visser’s Basis

The central role in the developement of the paper is played by Visser’s basis of rules. The term basis means that any rule that is admissible for IPC is obtainable by combining some of the rules of the family with other intuitionistic reasoning. It consists of the following sequence of rules:

This is read as: for every natural number , whenever the left part of the rule (a -ary implication) is provable, then the right part (an -ary disjunction) is provable. It forms a basis in the sense that all other admissible rules of IPC can be obtained from the combination of rules from this family with the usual rules of intuitionistic logic. It is an infinite family, since Visser cannot be derived from Visser,…,Visser  [10].

The importance of Visser’s basis is not limited to intuitionistic logic but also applies more generally to intermediate logics, as witnessed by the following:

Theorem 1.1 (Iemhoff [8]).

If the rules of Visser’s basis are admissible in a logic, then they form a basis for the admissible rules of that logic

This theorem also gives us a simple argument to prove the structural completeness of CPC: since all the Visser rules are provable in CPC, they are admissible and therefore they constitute a basis for all the admissible rules of CPC; but since the Visser are derivable, all admissible rules are derivable.

1.2 Contributions and Structure of the Paper

In Section 2 we introduce the natural deduction rules corresponding to Visser’s rules, and present the associated -calculus V: we show that proofs in the new calculus normalize to ordinary intuitionistic proofs. In the remaining part of the paper, we push further our idea and start adapting our calculus to intermediate logics characterized by axioms derived from admissible rules. In Section 3 we study the well-known Harrop’s rule, and more precisely the logic KP obtained by adding Harrop’s principle to IPC: we prove good properties like subject reduction, the disjunction property, and strong normalization. In Section 4 we quickly introduce the logic AD (obtained by adding the axiom to IPC) as a candidate for future study, and possible extensions to arithmetic. Proofs can be found in the appendices at the end of the paper.

2 Proof Terms for the Admissible Rules: V

In this section, we are going to assign proof terms to all the inferences of Visser’s basis in a uniform way. First, we give a natural deduction flavor to the Visser rules. Since the conclusion of the left-hand side of the rules is a disjunction, we model the rules as generalized disjunction eliminations ; “generalized” because the main premise will be the disjunction in the antecedent of the Visser, but under implicative assumptions. Therefore the rules of inference Visser have the form:

In order to keep the rules admissible, we need to restrict the usage of the inference: the additional requirement is that the proofs of the main premise (the one on the left with end-formula ) must be closed proofs, i.e. cannot have open assumptions others than the ones discharged by that Visser inference. Otherwise we would be able to go beyond IPC, since for example we would prove all the principles corresponding to the admissible rules (as in system AD, see Section 4). On the other side, it is straightforward to see that our rules directly correspond to rules of Visser’s basis, and that they adequately represent admissibility. We now turn to proof terms:

Figure 1: Proof terms

Since the shape of the rules is the elimination of a disjunction, the proof term associated with this inference will be modeled on the case analysis

. The difference will be in the number of assumptions that are bound, and in the number of possible cases. We use the vector notation

on variables to indicate that a sequence of (indexed) variables is bound, and on terms like to indicate a sequence of (indexed) terms on each of which we are binding the variable . The resulting annotation for a Visser inference is then:

Visser

We call V the calculus obtained by adding this family of rules of inference to IPC. The syntax of V can be found in Figure 1, and it includes the usual proof terms for intuitionistic logic [11], plus the proof terms for the Visser family.

We now turn to the reduction rules. First of all, we need to define contexts: intuitively, contexts are proof terms with a hole, where the hole is denoted by , and means replacing the unique hole in the context with the term .

Definition 2.1 (Weak head Ipc contexts).

contexts are defined by the following grammar:

Reduction rules for IPC Beta Projection Case Additional rules for V Visser-inj Visser-efq Visser-app

Figure 2: Reduction rules (V)

The reduction rules for the proof terms are given in Figure 2: the first block defines by means of the usual rules for IPC, and the second block defines as plus additional reduction rules for the new construct , depending on different shapes that might have. Let us explain the intuition. In the first case (Visser-inj), the term is the injection with possibly free variables of type for ; in that branch one has chosen to prove one of the two disjuncts or , and we may just reduce to the corresponding proof , in which we plug the proof but after binding the free variables . In the second case (Visser-efq), the disjunction is proved by means of a contradiction, and that contradiction may be used to prove any of the cases . In the third case (Visser-app), the term contains an application with one of the variables bound by the Visser rule on the left hand side, i.e. the proof uses one of the Visser assumptions to prove the disjunction. We reduce to the corresponding case , where is substituted for the assumption of type . The reduction relation is obtained as usual as the structural closure of the reduction (and similarly for ).

As expected, V–terms normalize: we prove normalization by providing an evaluation function that reduces V–terms to intuitionistic terms. The idea is to define the evaluator by structural recursion on typed terms, and using normalization for IPC after each recursive call.

Theorem a.1“??” .

“??”

The following is a consequence of Lemma A.2:

Theorem 2.1.

V–terms normalize to IPC–terms.

3 Beyond Ipc: Harrop’s Rule and Kp

In the previous section we have been expecially careful in imposing the restriction on the open assumptions for the application of our new rules, in order to keep our calculus inside the intuitionistic world and to obtain precisely a characterization of admissibility. At this point, however, one can legitimately ask: what happens if we lift such restriction, and allow one or more admissible principles inside an extended logic? The system of rules we introduced assumes then a different role, that is the role of providing a simple and modular way to obtain Curry-Howard systems for semi-classical logics arising from the addition to IPC of axioms corresponding to admissible principles.

The simplest and oldest studied admissible rule of IPC is the rule of independence of premise, also known as Harrop’s rule in its propositional variant [5]:

The logic that arises by adding it to IPC has also been studied, and is known as Kreisel-Putnam logic (KP). It was introduced by G. Kreisel and H. Putnam [9] to show a logic stronger than IPC that still could satisfy the disjunction property, thus providing a counterexample to the conjecture of Łukasiewicz that IPC was the only such logic.

We now proceed to define a Curry-Howard calculus for KP as an instance of the system we presented in the previous section. It suffices to realize that Harrop’s rule is a particular case of Visser where the formula is taken to be (note that the third disjunct in this instance of Visser becomes , that implies both the other hypotheses and ; for this reason we can ignore it). Then we get the following simplified rule in natural deduction:

Harrop

The restriction on the assumptions of the main premise is now gone, and open proofs are allowed. In fact Harrop’s principle is provable in our system:

Harrop

The proof term is a simplified version of the proof term for , where we remove the term corresponding to the trivialized third disjunct:

By inspecting the reduction rules for V, we realize that the rule Visser-app has no counterpart in KP: since the Harrop assumptions have negated type, their use in proof terms is completely encapsulated in exfalso terms (see Classification, Lemma B.1 below). Therefore the reduction rules for KP are the ones for IPC (Figure 2) plus the additional rules Harrop-inj and Harrop-efq in Figure 3. We denote with the toplevel reduction for KP, and with its structural closure.

Harrop-inj Harrop-efq

Figure 3: Reduction rules (KP)

We prove for KP the usual properties of subject reduction, classification, and strong normalization. As expected we denote with the provability in KP, but we use simply when not ambiguous.

Theorem b.1“??” .

“??”

In order to classify normal forms of

KP, we need to consider proof terms with possibly open Harrop assumptions: we denote with a negated typing context, i.e. of the form . We obtain the following classification of normal forms:

Lemma b.1“??” .

“??”

We prove that KP enjoys the strong normalization property, i.e. all typable terms are strongly normalizing. We use a modified version of the method of reducibility candidates by Girard-Tait [3]. The differences with respect to the usual proof are that Harrop and exfalso terms are added to neutral terms, and that the reductions for (which involve terms under binders) require special treatment.

Theorem b.2“??” .

“??”

The complete proof is on the appendix. We can now prove the disjunction property:

Lemma 3.1 (Consistency).

for no .

Proof.

Let us assume that there exists (which we assume in normal form by Theorem B.2) such that , and derive a contradiction. We proceed by induction on the size of . The base case is impossible because by Lemma B.1 cannot be a variable. As for the inductive case, by Lemma B.1, is either an exfalso, or for some . In the former case for some such that , and we use the i.h.; the latter case is not possible, since . ∎

Theorem 3.1 (Disjunction property).

If , then or .

Proof.

Assume for in normal form by Theorem B.2. First note that , because otherwise by inversion , contradicting consistency. By Lemma B.1 (with ) is an injection. Conclude by inversion. ∎

4 Conclusions and Future Work

Our system provides a meaningful explanation of the admissible rules in terms of normalization of natural deduction proofs. In addition, by simply lifting the condition of having closed proofs on the main premise, we can study intermediate logics characterized by the axioms corresponding to some admissible rules; the study of the Kreisel-Putnam logic exemplifies this approach.

We believe that our presentation is well-suited to continue the study of admissibility in intuitionistic systems, a subject that is currently mostly explored with semantic tools. We devised powerful proofs of normalization for our systems KP and V, and we will try to extend these results to other similarly obtained systems. We conclude with some remarks on future generalizations.

4.1 The Logic Ad

Now that we have shown the potential of our system in analysing the extension of IPC with axioms corresponding to admissible rules, we might wonder what could happen when we try to add several of them. We can be even more ambitious: what if we want to add all the Visser rules to IPC? A theorem by Rozière greatly simplifies our task:

Theorem 4.1 (Rozière [10]).

All Visser rules are derivable in the logic AD, obtained by adding the axiom schema to IPC.

Clearly, since the Visser rules are derivable in AD they are also admissible; as we know from Theorem 1.1 this means that they form a basis for all the admissible rules of AD, and since they are derivable we obtain:

Corollary 4.1.

The logic AD is structurally complete.

However, we also know from Iemhoff [6] that IPC is the only logic that has the Visser rules as admissible rules and satisfies the disjunction property. This means that AD cannot satisfy the disjunction property. This was also proved with different techniques by Rozière, who also showed that AD is still weaker than CPC. Given these properties, AD seems the best candidate to be studied with our technique.

4.2 Arithmetic

Since its inception with Harrop [5], the motivation for studying admissible rules of IPC was to understand arithmetical systems. A famous theorem of de Jongh states that the propositional formulas whose arithmetical instances are provable in intuitionistic arithmetic (HA) are exactly the theorems of IPC, and many studies of the admissible rules of HA (like Visser [12], Iemhoff and Artemov [2]) originated from it. In particular Visser shows that the propositional admissible rules of HA coincide with those of IPC, and that rules are also related.

Harrop’s principle, that we have investigated in this paper, is also known as the propositional Independence of Premise principle. Its first order version:

corresponds to an admissible rule of HA that has an important status in the theory of arithmetic, and was given a constructive interpretation for example by Gödel [4] with his well known Dialectica interpretation.

We can assign to IP a proof term and two reduction rules that act in the same way as the ones introduced for Harrop’s rule: that is, we will distinguish the two cases where there is an explicit proof of the existential in the antecedent, and where an exfalso reasoning has been carried on. We believe that a more advanced study of other admissible rules of HA can be carried on similar grounds.

References

  • [1]
  • [2] Sergei N. Artemov & Rosalie Iemhoff (2004): From de Jongh’s theorem to intuitionistic logic of proofs. In: Dick de Jongh’s Festschrift, pp. 1–10. Available at https://istina.msu.ru/publications/article/19375470/.
  • [3] Jean-Yves Girard, Paul Taylor & Yves Lafont (1989): Proofs and types. Cambridge tracts in theoretical computer science 7, Cambridge University Press, Cambridge. Available at http://www.paultaylor.eu/stable/prot.pdf.
  • [4] Kurt Gödel (1958): Über eine bisher noch nicht benützte Erweiterung des finiten Standpunktes. Dialectica 12(3-4), pp. 280–287, doi:10.1111/j.1746-8361.1958.tb01464.x.
  • [5] Ronald Harrop (1956): On disjunctions and existential statements in intuitionistic systems of logic. Mathematische Annalen 132(4), pp. 347–361, doi:10.1007/BF01360048.
  • [6] Rosalie Iemhoff (2001): A(nother) characterization of intuitionistic propositional logic. Annals of Pure and Applied Logic 113(1), pp. 161–173, doi:10.1016/S0168-0072(01)00056-2.
  • [7] Rosalie Iemhoff (2001): On the admissible rules of intuitionistic propositional logic. The Journal of Symbolic Logic 66(1), pp. 281–294, doi:10.2307/2694922.
  • [8] Rosalie Iemhoff (2005): Intermediate logics and Visser’s rules. Notre Dame Journal of Formal Logic 46(1), pp. 65–81, doi:10.1305/ndjfl/1107220674.
  • [9] Georg Kreisel & Hilary Putnam (1957): Eine Unableitbarkeitsbeweismethode für den Intuitionistischen Aussagenkalkül. Archiv für mathematische Logik und Grundlagenforschung 3(3-4), pp. 74–78, doi:10.1007/BF01988049.
  • [10] Paul Rozière (1993): Admissible and Derivable Rules in Intuitionistic Logic. Mathematical Structures in Computer Science 3(2), pp. 129–136, doi:10.1017/S0960129500000165.
  • [11] Morten Heine Sørensen & Pawel Urzyczyn (2006): Lectures on the Curry-Howard isomorphism. Studies in Logic and the Foundations of Mathematics 149, Elsevier, doi:10.1016/s0049-237x(06)x8001-1.
  • [12] Albert Visser (2002): Substitutions of -sentences: explorations between intuitionistic propositional logic and intuitionistic arithmetic. Annals of Pure and Applied Logic 114(1), pp. 227–271, doi:10.1016/S0168-0072(01)00081-1.

Appendix A Theorems on V

First some definitions. We denote with the provability in V (but we use when not ambiguous). We denote with an implicative typing context, i.e. of the form . We say that a term is neutral if it has the form or .

Lemma A.1 (Classification for V).

Let for in normal form, and not neutral:

  • Implication: if , then is either an abstraction or a variable in ;

  • Disjunction: if , then is an injection;

  • Conjunction: if , then is a pair;

Proof.

By induction on the type derivation of :

  • () is a variable in . By definition of , the type of is an implication, and we conclude.

  • () is an abstraction, and we conclude.

  • () and with . Because is in normal form, cannot be an abstraction. By i.h., is either a variable in or is neutral; in both cases is neutral.

  • () is an injection, and we conclude.

  • () and with . Because is in normal form, cannot be an injection. By i.h. is neutral, and therefore is neutral.

  • () is a pair, and we conclude.

  • () and with . Because is in normal form, cannot be a pair. By i.h. is neutral, and therefore is neutral.

  • (Visser) not possible. Assume with by inversion, and derive a contradiction. By i.h. is neutral or an injection, but both cases contradict the hypothesis that is a normal form.

In order to prove normalization, we define an evaluation function , mapping each typable term in V to its normal form. We first assume a corresponding function for IPC:

Definition A.1 ().

We call the function mapping each term typable in IPC to its normal form.

Definition A.2 ().

Let a term typable in V. We define its evaluation by structural induction:

Note: the three cases in the definition of on Visser terms are exhaustive by inspection of the normal forms of type disjunction (Lemma A.1) since it holds by inversion that with .

Lemma A.2 ( well-defined).

For every V-term s.t. :

  1. ,

  2. is normal,

  3. .

Proof.

The three points can be proved mutually, by induction on the type derivation :

  1. follows by i.h. and by subject reduction for IPC;

  2. follows by i.h. and from the fact that the output of are only normal forms;

  3. follows by i.h. and from the fact that IPC is a subcalculus of V.

It easily follows:

Theorem A.1 (Normalization for V).

V enjoys the normalization property.

Appendix B Theorems on Kp

Theorem B.1 (Subject reduction for Kp).

If and , then .

Proof.

By the definition of reduction as the closure of under evaluation contexts, we just prove the statement when ; the general case follows because substitution preserves types.

The cases of the usual intuitionistic reductions are standard (see for example [11]); we just prove the cases of the reduction rules associated with .

For the case of the left injection , by inversion we have and for some . Again by inversion
, and by we obtain . By substitutivity we get the desired result . The case of the right injection is analogous.

Finally, if , by inversion we have and for some . It is easy to see, by induction on the definition of weak head contexts and by inversion, that ; by we obtain . By we obtain , and by substitutivity we get the desired result . ∎

We say that a term is neutral if it has the form .

Lemma B.1 (Classification for Kp).

Let for in (weak head) normal form and not neutral:

  • Implication: if , then is an abstraction or a variable in ;

  • Disjunction: if , then is an injection;

  • Conjunction: if , then is a pair;

  • Falsity: if , then for some and some .

Proof.

By induction on the type derivation of :

  • () and is a variable in : by definition of , the type of is an implication, and we conclude.

  • () and is an abstraction: trivial.

  • () and with . Because is in normal form, cannot be an abstraction. By i.h., is either a variable in or a neutral term. In the first case, note that we have that and , and the thesis holds; in the second case, is neutral and the thesis holds.

  • () and is an injection: trivial.

  • () and with . By i.h. is either an injection or neutral. The first case is not possible because is in normal form; in the second case, is neutral as required.

  • () and is a pair: trivial.

  • () and with . By i.h. is either a pair or neutral, but the first case contradicts the hypothesis that is in normal form. Therefore is neutral, and also is neutral.

  • () then is immediately neutral.

  • (Harrop) not possible. Assume with , and derive a contradiction. By i.h. is an injection or a neutral term, but both cases contradict the hypothesis that is in normal form.

b.1 Strong Normalization

In this section, we prove the strong normalization property for KP by means of an adapted version of Girard’s method of candidates [3].

Definition B.1 (Weak head Kp contexts).

Let SN be the set of strongly normalizing terms of KP. By abuse of notation, we say that a context – be it an IPC context or a KP context – is strongly normalizing if all its “internal” -terms are strongly normalizing.

Definition B.2 (Weak head reduction , ).

We define as the “strongly normalizing” closure of (Figure 2) under weak head contexts:

for every SN contexts and . As usual, we denote by the reflexive and transitive closure of . A term is a -normal form (in short, nf) if . We say that if and is a nf.

By inspection of the reduction rules, one may prove:

Lemma B.2.

is deterministic.

One of the main properties of reducibility candidates is that they are backward closed under reduction:

Definition B.3 (Backward closure ).

Let be a set of nfs. We define its closure under backward weak head reduction as the set .

Lemma B.3 (Backward closure of Sn).

SN is backward closed under .

Proof.

Let and ; we need show that . By cases on the reduction rules of Definition B.2; we only consider the case of Harrop-inj, as one can proceed in a similar way for the other reduction rules. Let , and let us consider a reduction sequence beginning with . Either the sequence terminates after some internal redutions

which must terminate because all internal terms are SN by definition of , or eventually we have

This term is strongly normalizing because it is a reduct of , and by hypothesis . Therefore the reduction sequence must terminate. ∎

Another key notion are neutral terms, that are intuitively nfs that do not begin with constructors:

Definition B.4 (Neutral terms).

.

Fact B.1.

Neutral terms are strongly normalizing nfs.

We are now ready to define the semantics of formulas:

Definition B.5 (Denotation ).
  1. for every atomic (also ),

  2. ,

  3. ,

  4. .

In fact, we note that our definition produces candidates of reducibility:

Lemma B.4 (Denotations are candidates).

For every , its denotation:

  1. contains only strongly normalizing terms:

  2. contains all neutral terms:

  3. is backward closed: if and , then .

Proof.

Points 2 and 3 are trivial. Before proving Point 1 we note that as shown in the proof of Lemma B.3, if contains only strongly normalizing terms, then does too. We can then prove Point 1 by induction on the structure of types: the case of propositional atoms follows from Definition B.5(1) and Lemma B.3; for the inductive cases, use Fact B.1, the i.h. and Lemma B.3. ∎

We extend the definition of valuation to typing contexts:

Definition B.6.

Let be a typing context; we define as the set of substitutions mapping variables in to terms in the denotation of the corresponding type, i.e.

where when .

A lemma useful in the proof of Lemma B.6:

Lemma B.5.

If and , then .

Proof.

First note that if and , then for some SN context . Therefore, we assume that with , and we prove that by cases on the reduction rules:

  • . By renaming, . Then , with . We conclude because and by the hypothesis that .

  • . Then , and by hypothesis.

  • . By renaming, . Similar to the case below.

  • . By renaming, .
    Then . We conclude because .

  • . By renaming, .
    Then