Marcello DAgostino

A Note on the Complexity of Proofs in the Propositional Fragments of KE and KI

1. 2. 3. 4. 5.

Introduction Preliminary Notions KE and KI KE versus the Tableaux Method KE versus Natural Deduction

Notes Summary (English and Italian) References

Ann. Univ. Ferrara. Sez.III. Filosoa Discussion Paper n.9. 1989

Marcello DAgostino

A Note on the Complexity of Proofs in the Propositional Fragments of KE and KI

Introduction

Proof systems are considered to be equivalent as long as they recognize exactly the same set of valid inferences. However, equivalent proof systems can represent totally dierent ways of formalizing the notion of valid inference, and some ways can be better than others with respect to certain purposes. The purpose I have in mind, as far as this note is concerned, is eciency, and my aim is to compare the propositional fragments of the systems KE and KI [Mond 88a, Mond 89a, Mond 88b, Mond 88c, Mond 89b] with the propositional fragments of other well-known proof systems. In section 2 I give some preliminary notions. In section 3 I show that proofs in KE and KI have basically the same complexity. Then, in section 4, I compare KE to the tableaux method and show that the former is essentially more ecient than the latter. Finally, in section 5, I show that KE can eciently simulate Natural Deduction, and discuss a general requirement on proof systems which is satised by KE, KI and natural deduction methods but not by the tableaux method.

Marcello DAgostino

Preliminary Notions

We assume a zero order language consisting of (i) denumerably many propositional variables P, Q, R, . . .; (ii) the usual logical constants (not), (and), (or), (if . . . , then . . . ); (iii) the two parentheses ( and ). The notion of well-formed formula is dened in the usual way. We shall use the letters A, B, C, . . . as schematic letters denoting arbitrary formulas. We shall also adopt the usual conventions about the binding power of the logical constants, namely: binds more tightly than any other logical constant, binds more tightly than and binds more tightly than . Subformulas are dened in the usual way. We indicate sets of formulas by the capital Greek letters , , , . . . (possibly with subscripts). We shall often write , instead of and , A instead of {A}. All the proof systems I shall examine in this note enjoy the so-called (weak) subformula property, that is: to prove a formula A one only needs to consider its weak subformulas, where a weak subformula of A is either a subformula of A or the negation of a subformula of A. Systems enjoying such a property allow for easy proof-search procedures, truly mechanical ones, by dramatically reducing the search space. Proofs with the subformula property can be said to be analytic in that they employ no elements which are not already contained in the formulation of the problem. Some systems of deduction (like the tableaux method and Gentzens sequent calculus without cut) yield only analytic proofs. Others (like Natural Deduction, Gentzens sequent calculus with cut and Mondadoris recently-proposed systems KE and KI) allow a more general notion of proof which includes non-analytic proofs, although in all these cases there is an eective procedure to transform every non-analytic proof into an analytic one. We take the complexity of a proof in S, denoted by () as the number of occurences of formulas in the proof. Let us consider the relation S A meaning that there is a proof of A from in the system S such that () n. Suppose that, given two systems S and S , there is a function g
n

A Note on the Complexity of Proofs such that

n

g(n)

S A = S A we are interested in the rate of growth of g for particular systems S and S . Positive results about the above relation are usually obtained by means of simulation procedures, namely computable functions mapping proofs of one system to proofs of the other. Negative results consist of lower bounds for the function g. If one is interested in the relative length of proofs, namely in the total number of symbols occurring in them, our -measure is obviously adequate for establishing negative results but not, in general, for positive results. It may, however, be adequate also for positive results whenever one can show that the length of formulas is not signicantly increased by the simulation procedure under consideration. All the procedures described in this note will be of this kind, so that our results admit of a simple, though tedious, translation in terms of length of proofs. For certain purposes we may be interested in the complexity not of proofs in general but of analytic proofs. We shall then appeal to the notion of analytic restriction of a system: let S be a system which enjoys the subformula property. We can take S as dened not by its inference rules, but extensionally, by the set of proofs which it recognizes as sound. Then we can denote by S its analytic restriction, i.e. the subset of S consisting of the proofs which enjoy the subformula property. All the notions dened for proof systems can be extended to analytic restrictions in a natural way.

KE and KI

Both KE and KI have been designed to reect closely the classical interpretation of the logical operators and of the notion of truth. The latter is characterized by the assumption that every proposition is either true or false (Principle of Bivalence). In Mondadoris approach this assumption is directly translated into an inference rule with zero premises and two (alternative) conclusions: PB t(A) f(A)

where the signed formulas t(A) and f(A) have the expected intuitive meaning A is true and A is false respectively.

Marcello DAgostino

In what follows we shall refer to the versions of KE and KI which make use of unsigned formulas. In these versions the rule PB takes the following form:

PB

Both KE and KI contain the rule PB, which is a structural rule reecting the assumption of a classical notion of truth, and dier in the set of connective rules characterizing the behaviour of the logical operators. In KE these rules are elimination rules: CONJUNCTION A (A B) E1 B DISJUNCTION A AB E1 B CONDITIONAL A AB E1 B B AB E2 A B AB E2 A B (A B) E2 A

AB E1 A

AB E2 B

(A B) E1 A

(A B) E2 B

(A B) E 1 A

(A B) E 2 B

A Note on the Complexity of Proofs NEGATION

A E A We shall also call these rules E-rules. The rules of KI are introduction rules: CONJUNCTION A B I1 AB

A 1 (A B) DISJUNCTION A B I (A B)

B I2 (A B)

A I1 AB CONDITIONAL

B I2 AB

A B I (A B)

Marcello DAgostino

A I1 AB NEGATION

B I2 AB

A I A We shall also call these rules I-rules. All the rules are to be read as branch-extension rules in the usual way we are allowed to append to a branch r containing the premise(s) of a rule the conclusion(s) of that rule as immediate successor(s). A KE(KI)-tree for a set of formulas is a tree generated by the rules of KE (KI) starting from the formulas in . A branch is closed when it contains both C and C for some formula C, otherwise it is open. A tree is closed when all its branches are closed. Proofs and refutations in KE and KI are dened as follows: A KE-refutation of is a closed KE-tree for . A KE-proof of A from is a KE-refutation of , A (KE is a refutation system). A KI-proof of A from is a KI-tree for such that A occurs in all open branches. Both KE and KI are complete for classical logic, but PB is not eliminable. Moreover, it can be shown that they enjoy the subformula property.1 We start by showing that proofs in KE and KI have basically the same complexity: Theorem 1 KE and KI can simulate each other with an increase in proof complexity by at most a constant factor. Moreover, the simulation preserves the subformula property. Proof. First, observe that the rules of KI can be simulated in KE as follows (we show the procedure only for the rules I and I, the other cases being similar):

A Note on the Complexity of Proofs A B

d d d d

7 A B
d d d d

AB

(A B) B

AB B

(A B)

where marks a closed branch. Now, if T is a KI-tree with assumptions , then replace each application of a KI-rule with its KE-simulation (the applications of PB can be left unchanged since PB is also a rule of KE). The result is a KE-tree T containing at most (T ) + c (T ) nodes, where c is the maximum number of additional nodes generated by a KE-simulation of a KI-rule (namely 2). Thus, if T is a KI-proof of A from , T is a KE-tree such that all its open branches have A as terminal node. Moreover, (T ) 3(T ) and T does not contain any formulas which do not occur in T . Simply adding A to the assumptions will provide a closed KE-tree for , A. The inference rules of KE can be easily simulated in KI as follows (we show the procedure only for the rules E1 and E1, the other cases being similar): A AB
d d d d

A (A B)
d d d d

E1

E1

B (A B)

B AB

Marcello DAgostino

Now, if T is a KE-tree with assumptions , A, then replace each application of a KE-rule with its KI-simulation (the applications of PB can be left unchanged since PB is also a rule of KI). The result is a KI-tree T containing at most (T ) + c (T ) nodes, where c is the maximum number of additional nodes generated by a KI-simulation of a KE-rule (namely 2). Thus, if T is a closed KE-tree for , A, then T is a closed KI-tree for , A. Moreover, (T ) 3(T ) and T does not contain any formula which does not occur in T . Hence, by applying PB to A, one obtains the required KI-proof of A from .2

KE versus the Tableaux Method

In spite of their similarity, KE and the tableaux method are quite dierent refutation systems. While all the tableaux rules can be easily simulated by means of KE-rules, KE includes a rule, namely PB, which cannot be easily simulated by means of the tableaux rules. Although it is well known that the addition of this rule to the tableaux rules does not increase the stock of inferences that can be shown valid (since PB is classically valid and the tableaux method is classically complete), its absence, in some cases, is responsible for an explosive growth in the size of tableaux proofs. The reason for this explosion is not dicult to see. Suppose there is a tableaux proof of A from , i.e. a closed tableau T1 for , A (where A is assumed to be atomic for the sake of simplicity) and a tableaux proof of B from , A, i.e. a closed tableau T2 for , A, B; then it follows from the elimination theorem (see [Smul 68]) that there is also a closed tableaux for , , B. This fact can be seen as a typical cut inference: A , A B , B where stands for the tableaux derivability relation. When a rule like PB is available, simulating this kind of cut inference is relatively inexpensive in terms of proof size as is shown by the diagram below:

A Note on the Complexity of Proofs B

d d d d

A T2

A T1

But if PB is not in our stock of rules, reproducing the cut inference may be much more expensive. Let us assume, for instance, that A is used more than once, say n times, in T1 to close a branch, so that T1 contains n occurrences of A in distinct branches which will be left open if A is removed from the assumptions. Similarly, let A be used more than once, say m times, in T2 to close a branch, so that T2 contains m occurrences of A in distinct branches which will be left open if A is removed from the assumptions. Then, in some cases, the shortest tableau refutation of , , B will have one of the following two forms: B T1
d d d d

B T2
d d d d

A T2

A T2

A T1

A T1

where the subrefutation T2 is repeated n times in the lefthand tree and the subrefutation T1 is repeated m times in the righthand tree. Because of this intrinsic ineciency of the tableaux method, examples can be found which require a great deal of duplication in the construction of a

10

Marcello DAgostino

closed tableau. A class of hard examples is described in [CoRe 74]: Let Hm = {A1 A2 A3 . . . Am } ... where +A means A and A means A, and the subscript of Ai is a string of i 1 +s or s corresponding to the sequence of signs of the preceding Aj , j < i. Thus Hm contains 2m disjunctions and 2m 1 distinct atomic letters. For instance H2 = {A1 A2 , A1 A2 , A1 A2 , A1 A2 }. + + cm In [CoRe 74] Cook and Rechow report a lower bound of 22 on the number of nodes of a closed tableau for the conjunction of all disjunctions in Hm . In contrast, one can show that there is an easy analytic KE-refutation of Hm which contains 2m +2m m2 nodes. Such a refutation has the following form: start with Hm . This will be a set containing n(= 2m ) disjunctions of which n/2 start with A1 and the remaining n/2 with its negation. Then apply PB to A1 . This creates a branching with A1 in one branch and A1 in the other. Now, on the rst branch, by means of n/2 applications of the rule E1 we obtain a set of formulas which is of the same form as Hm1 . Similarly on the second branch we obtain another set of the same form as Hm1 . By reiterating the same procedure we eventually produce a closed tree for the original set Hm . It is easy to see that the number of nodes generated by the refutation can be calculated as follows (where n is the number of formulas in Hm , that is 2m ):
log n1

(T ) = n +
i=1

2i + n = n + 2

1 2log n1 + n (log n 1) 12 = n + n log n 2

Therefore the following theorem holds (where KE stands for the analytic restriction of KE and TM for the tableaux method): Theorem 2 (Speed-up of KE over TM) There is no polynomial p such that KE A =
n p(n) TM

A.

In other words, the tableaux method cannot p-simulate KE .2 Our discussion shows that cuts are sometimes essential for providing short refutations with the subformula property.3 On the other hand, given

A Note on the Complexity of Proofs

11

a tableaux refutation T of we can eectively construct an analytic KErefutation T of which is not essentially longer. Theorem 3 TM A = KE A Proof. First, observe that the connective rules of KE, combined with PB, can easily simulate the branching rules of the tableaux method, as is shown below in the case of the branching rule for eliminating disjunctions (all the other cases are similar): AB
d d d d
n 2n

Such a simulation lengthens the original tableau by one node for each application of a branching rule. Since the linear rules of the tableaux method are also rules of KE, it follows that there is a KE-refutation T of such that (T ) (T ) + k, where k is the number of applications of branching rules in T . Since k is obviously (T ), then (T ) 2(T ).2

KE versus Natural Deduction

It can also be shown that KE can simulate natural deduction (in tree form) with an increase in proof complexity by at most a constant factor. Moreover the simulation procedure preserves the subformula property. We shall sketch this procedure for the natural deduction system given in [Praw 65], the procedure being similar for other formulations. We want to give an eective proof of the following theorem (where ND stands for Natural Deduction):

12

Marcello DAgostino

Theorem 4 If there is an ND-proof T of A from , then there is a KEproof T of A from such that (T ) 3(T ) and T contains only formulas A such that A occurs in T . Proof. By induction on (T ). If (T ) = 1, then the ND-tree consists of only one node which is an assumption, say C. The corresponding KE-tree is the closed sequence C, C. If (T ) = k, with k > 1, then there are several cases depending on which rule has been applied in the last inference of T . We shall consider only the cases in which the rule is elimination of conjunction (E) and elimination of disjunction (E), and leave the others to the reader. If the last rule applied in T is E, then T has the form: T = T1 AB A

By inductive hypothesis there is a KE-refutation T1 of , (AB) such that (T1 ) 3(T1 ). Then the following KE-tree:

A
d d d d

T =

AB A

(A B)
T1

is the required KE-proof and it is easy to verify that (T ) 3(T ). If the last rule applied in T is the rule of elimination of disjunction, then T has the form: 1 2 , [A] 3 , [B] T1 T2 T3 T = AB C C C

A Note on the Complexity of Proofs

13

By inductive hypothesis there are KE-refutations T1 of 1 , (A B), T2 of 2 , A, C, and T3 of 3 , B, C, such that (Ti ) 2(Ti ), i = 1, 2, 3. Then the following KE-tree:

1 2 3 C
d d d d

T =

AB
d d d d

(A B) T1

A T2

A B T3

is the required proof and it is easy to verify that (T ) 3(T ).2 In [Cell 87] Cellucci has proposed a new form of natural deduction which he has shown in several examples to produce shorter proofs than Prawitzs style natural deduction. In Celluccis system, as in natural deduction, assumptions are introduced which may subsequently be discharged; however the inference rules involve sequents rather than single formulas. According to Cellucci, derivations in his system are signicantly simpler (i.e. contain fewer symbols) than in [Prawitzs style] system.4 It is not dicult to show, by means of an argument analogous to the one given above for Prawitzs style natural deduction, that: Theorem 5 If there is a proof in Celluccis system of the sequent from assumptions , then there is a KE-refutation of , , where is equal to {B | B }, such that ( ) 3(). Moreover, the simulation preserves the subformula property.

14

Marcello DAgostino

As seen in section 4, the speed-up of KE over the tableaux method can be traced to the fact that KE, unlike the tableaux method, can easily simulate inference steps based on cuts, like the example at p.9. In other words, KE allows a uniform method for grafting proofs of subsidiary conclusions, or lemmata, in the proof of a theorem. However, the existence of such a uniform method is by no means sucient. In Natural Deduction, for instance, replacing every occurrence of an assumption A with its proof provides an obvious grafting method which, though very perspicuous, is highly inecient, leading to much unwanted duplication in the resulting proof-tree. We should then require that the method be also ecient. A way of making this requirement precise is to ask that the following condition be satised by a proof system S: (C) Let 1 be a proof of A from and let 2 be a proof of B from , A, then there is a uniform method for constructing from 1 and 2 a proof 3 of B from , such that (3 ) (1 ) + (2 ) + c for some constant c. It can be easily seen that condition (C) is satised by KE and KI, the required method being the one described at p.9. In contrast, the standard way of grafting proofs of subsidiary conclusions in Natural Deduction proofs, though providing a uniform method, does not satisfy the further condition on the complexity of the resulting proof. The rules of Natural Deduction, however, permit us to bypass this diculty and produce a method satisfying the whole of condition (C). Consider the rule of Non-Constructive Dilemma (NCD): , [A] . . . B B This is a derived rule in Prawitzs style Natural Deduction which yields classical logic if added to the intuitionistically valid rules (see [Tenn 78, section 4.5]). We can show Natural Deduction to satisfy condition (C) by means of the following construction: , [A] . . . B

A Note on the Complexity of Proofs . . . , [A] . . . B B A B [A]

15

Notice that the construction does not depend on the number of occurrences of the assumption A in the subproof of B from , A. In analogy to condition (C) we can formulate a condition requiring a proof system to simulate eciently another form of cut which holds for classical systems and is closely related to the rule PB: (C*) Let 1 be a proof of B from , A and 2 be a proof of B from , A. Then there is a uniform method for constructing from 1 and 2 a proof 3 of B from , such that (3 ) (1 )+(2)+c for some constant c. Similarly, the next condition requires that proof systems can eciently simulate the ex falso inference scheme: (XF) Let 1 be a proof of A from and 2 be a proof of A from . Then there is a uniform method for constructing from 1 and 2 a proof 3 of B from , , for any B, such that (3 ) (1 ) + (2 ) + c for some constant c. So, let us say that a classical proof system is a classical cut system if it satises conditions (C*) and (XF). It is easy to show that every classical cut system satises also condition (C) above and, therefore, allows for an ecient implementation of the transitivity property of natural proofs. The next theorem shows that every classical cut system can simulate KE without a signicant increase in proof complexity: Theorem 6 If S is a classical cut system, then S can simulate KE with an increase in proof complexity by at most a constant factor. To prove the theorem it is convenient to assume that our language includes a 0-ary operator (Falsum) and that the proof systems include suitable

16

Marcello DAgostino

rules to deal with it.5 For KE this involves only adding the obvious rule which allows us to append to any branch containing both A and A for some formula A, so that every closed branch in a KE-tree ends with a node labelled with . The assumption is made only for conveniences sake and can be dropped without consequence. Moreover we shall make the obvious assumption that, for every system S, the complexity of a proof of A from A in S is equal to 1. Let () denote the number of nodes generated by a KE-refutation of (i.e. the assumptions are not counted).6 Let S be a classical cut system. If S is complete, then for every rule r of KE there is an S-proof r of the conclusion of r from its premises. Let b1 = Maxr ((r )) and let b2 and b3 be the constants representing, respectively, the -cost of simulating classical cut in S associated with condition (C*) above and the -cost of simulating the ex falso inference scheme in S associated with condition (XF) above. As mentioned before, every classical cut system satises also condition (C) and it is easy to verify that the constant associated with this condition, representing the -cost of simulating absolute cut in S, is b2 + b3 + 1. We set c = b1 + b2 + b3 + 1. The theorem is an immediate consequence of the following Lemma: Lemma 1 For every classical cut system S, if there is a KE-refutation of , then there is an S-proof of from with ( ) c (). Proof. The proof is by induction on (), where is a KE-refutation of . () = 1. Then is explicitly inconsistent, i.e. contains a pair of complementary formulas, say B and B, and the only node generated by the refutation is , which is obtained by means of an application of the KE-rule for to B and B. Since there is an S-proof of the KE-rule for , we can obtain an S-proof of the particular application contained in simply by performing the suitable substitutions and ( ) b1 < c. () > 1. Case 1. The KE-refutation has the form: C T1

A Note on the Complexity of Proofs

17

where C follows from premises in by means of an E-rule. So there is a KErefutation 1 of , C such that () = (1 ) + 1. By inductive hypothesis, there is an S-proof 1 of from , C such that (1 ) c (1 ). Moreover, there is an S-proof 2 of C from the premises from which it is inferred in such that (2 ) b1 . So, from the hypothesis that S is a classical cut system, it follows that there is an S-proof of from such that ( ) c (1 ) + b1 + b2 + b3 + 1 c (1 ) + c c ( (1 ) + 1) c ()

Case 2. has the following form:

d d d d

C T1

C T2

So there are KE-refutations 1 and 2 of , C and , C respectively such that () = (1 ) + (2 ) + 2. Now, by inductive hypothesis there is an S-proof 1 of from , C and an S-proof 2 of from , C with (i ) c (i ), i = 1, 2. Since S is a classical cut system, it follows that there is an S-proof of from such that ( ) < < < c (1 ) + c (2 ) + b2 c (1 ) + c (2 ) + c c ( (1 ) + (2 ) + 2) c () 2

Since (the classical version of) Prawitzs style natural deduction is a classical cut system, it follows that it can simulate KE with an increase in proof complexity by at most a constant factor. The same holds also for Celluccis

18

Marcello DAgostino

system and for the variant of KE described in [Mond 89a] under the name KE. Moreover it can be shown that for all these systems the simulation preserves the subformula property (i.e. it maps analytic proofs to analytic proofs). Therefore theorem 6, together with theorems 4 and 5, imply that Prawitzs style natural deduction, Celluccis natural deduction, KE, KE and KI can simulate each other with an increase in proof complexity by at most a constant factor and with a procedure which preserves the subformula property. Finally theorems 2 and 3 imply that all these systems are essentially more ecient than the tableaux method, even if we restrict our attention to analytic proofs. Marcello DAgostino Wolfson College Oxford OX2 6UD

Notes
I wish to thank Marco Mondadori for many suggestions and Derek Long for reading and commenting a previous draft. This research has been supported by a CNR-NATO grant for the academic year 1989-90 and by MURST (40% Ferrara).
1 See 2 For

[Mond 88a, Mond 89a, Mond 88b, Mond 88c]. the notion of p-simulation see [CoRe 79].

3 This

can be taken as further evidence in support of Boolos plea for not eliminating cut [Bool 84]. In that paper he gives a natural example of a class of rst order inference schemata which are hard for the tableaux method while admitting of easy (non-analytic) natural deduction proofs. Boolos example is a particularly clear illustration of the well-known fact that the elimination of cuts from proofs in a system in which cuts are eliminable can greatly increase the complexity of proofs. (For a related technical result see [Stat 78].) KE and KI provide an elegant solution to Boolos problem by making cut non-eliminable while preserving the subformula property of proofs.
4 [Cell 5

87, p.29]

Systems which are not already dened over a language containing can usually be redened over such an extended language without diculty.

A Note on the Complexity of Proofs

6

19

The reader should be aware that our -measure applies to KE-refutations and not to trees: the same tree can represent dierent refutations yielding dierent values of the -measure.

Summary In this note I show the propositional fragments of the systems KE and KI [Mond 88a, Mond 89a, Mond 88b, Mond 88c, Mond 89b] to be essentially more ecient than the propositional fragment of the tableaux method. It is also shown that the complexity of both systems is basically the same as that of natural deduction methods. Riassunto In questo lavoro mostro che i frammenti proposizionali dei sistemi KE e KI [Mond 88a, Mond 89a, Mond 88b, Mond 88c, Mond 89b], sono essenzialmente pi` ecienti del frammento proposizionale del metodo dei tableaux. u Mostro anche che la complessit` di entrambi i sistemi ` sostanzialmente la a e stessa di quella dei metodi di deduzione naturale.

References
[Bool 84] G. Boolos, Dont Eliminate Cut, Journal of Philosophical Logic, 7, 1984, pp.373378. C. Cellucci, Ecient Natural Deduction, in Temi e prospettive della logica e della losoa della scienza contemporanee, Vol.I, CLUEB, Bologna 1988, pp.2957. S. A. Cook and R. Rechow, On the Length of Proofs in the Propositional Calculus, Proceedings of the 6th Annual Symposium on the Theory of Computing, 1974, pp. 13548. S. A. Cook and R. Rechow, The Relative Eciency of Propositional Proof Systems, The Journal of Symbolic Logic, 44, 1979, pp. 3650.

[Cell 87]

[CoRe 74]

[CoRe 79]

20

Marcello DAgostino

[Mond 88a] M. Mondadori, Classical Analytical Deduction, Annali dellUniversit` di Ferrara, sez. III, Discussion Paper n. 1, 1988. a [Mond 88b] M. Mondadori, Sulla nozione di dimostrazione classica, Annali dellUniversit` di Ferrara, sez. III, Fil., Discussion Paper n. 3, a 1988. [Mond 88c] M. Mondadori, On the Notion of a Classical Proof, in Temi e prospettive della logica e della losoa della scienza contemporanee, vol. I, CLUEB, Bologna 1988, pp. 211214 (It is a summary in English of [Mond 88b]). [Mond 89a] M. Mondadori, Classical Analytical Deduction, Part II,Annali dellUniversit` di Ferrara, sez. II, Discussion Paper n. 5, 1988. a [Mond 89b] M. Mondadori, An improvement of Jereys Deductive Trees, Annali dellUniversit` di Ferrara, sez. III, Fil., Discussion Paper a n. 7, 1989. [Praw 65] [Smul 68] [Stat 78] D. Prawitz, Natural Deduction. A Proof-Theoretical Study, Almqvist & Wilksell, Uppsala 1965. R. Smullyan, First Order Logic, Springer, Berlin 1968. R. Statman, Bounds for proof-search and speed-up in the predicate calculus, Annals of Mathematical Logic, 15, 1978, p.225 287. N. Tennant, Natural Logic, Edimburgh University Press, Edimburgh 1978.

[Tenn 78]