Semantic Tableaux Methods for Modal Logics That Include the B(rowerische) and G(each) Axioms From: AAAI Technical Report FS-93-01. Compilation copyright © 1993, AAAI (www.aaai.org). All rights reserved. Francis Jeffry Pelletier Depts. of ComputingScience & Philosophy Univ.Alberta Edmonton,Alberta Canada TtG 2El Introduction: Semantic tableaux are a methodfor determiningvalidity of argumentsin a certain class of logics. Themethodis perhapsnot so weB-known to the automatedtheoremproving communityas resolution, 1but it has long beenpresent in the logic community. As its namesuggests, the methoduses semanticfacts about the languagein an attemptto generatea countermodelfor the argument at hand(or to showthat there is no such counter-model and that the argument is therefore valid). In this, the methodbears a strong similarity to Resolution, but it does not require any conversion to a normal form. The method is also different in being an analytic system:eachnon-initial formulain any proof is a subformulaof someearlier formulain the proof. Naturaldeductionis not analytic in this sense, nor is anyResolutionmethodother than UnitResolution. Thereare various different formatsfor stating the methodof semantictableaux. Participants in this AAAI Symposium are doubtless familiar with the form given by Fitting (1988). Themethodhe uses is quite similar to the methodI wishto employ,and there shouldbe no difficulty in transforming the oneinto the other. In this note, I am interested in extending mymethodto a certain class of modallogics. In this extensionwewill only consider propositional logics - a fact that lends someconsiderablesimplification to the description of the tableaux method. (By not needing to describe quantificationrules andtheir difficult integration into modal logics. See Garson1988 for a survey of the difficulties). I will first lay out the tableauxmethod that I prefer, restricting myattention to classical propositionallogic. I then showhowit is extendedto a certain class of modallogics, the normal modallogics. The idea of suchan extensionis not particularly new;it is already mentionedin Fitting (1983, 1988), althoughnaturally believe that the particular modificationsI maketo the basic methodmakesthe extensionconsiderablyeasier to understand.After developinga methodfor the basic one of these logics I turn myattention to logics which contain the B(rowerische) axiom. Of these logics, Fitting (1983:192; 1988:203)says that the style tableauxhe presentsdoesnot lend itself to logics where the accessibility relation involvessymmetry. Andit is implied that it wouldbe very difficult to construct reasonable tableaux methodsfor such logics. I show that the alteration I makeavoids this difficulty and allowsfor a semi-analytictableauxsystem.After this I considermodificationsto the basic systemso as to yield a tableaux methodsuitable for systemscontaining the G(each) axiom. This axiom is unusual in that it involves embedded modal operators in both its antecedentand its consequent,and this in turn raises interesting challengesfor tableauxsystems.Onceagain wesee that a semi-analytictableauxsystemwill work. Classical Propositional Semantic Tableaux: Semantictableaux, whendone in the methodI prefer, are in fact trees; andthe methodof makinga tableauis in fact a methodof tree construction.Like Resolution, it is a refutation method;andthereforeit starts with a "negatedconclusion"formof an argumentto be tested. (UnlikeResolution,there is no further pre-processingof formulas, such as conversion to a normalform.) The root of the tree contains all the premises and the negationof the conclusion.Theconstructionof the rest of the l~ee is governedby "branching rules", which operate (by decomposition)on formulasalready in the tree. For the propositionallogic, formulasare definedin the usual way,usingsentenceletters (A,B,C....) andthe 1 Themethodgoes backexplicitly to Beth(1952,1958), connectives~, &, v, -->, <->. For describingthe method, it is convenientto divideformulasinto four classes: (i) and has been popularized by such elementary logic literals (sentence letters and their negations), (ii) textbooksas Smullyan(1968), Jeffrey (1968), Bonevac formulaswhichare doublenegations (i.e., start with (1987), Bergmann et al (1980). As has been noted two ~-signs), (iii) formulas whichhave somebinary variousresearchers(e.g. Fitting 1983:4-6,Fitting 1988: connective(&,v,--->,<-->)as their mainconnective,and 191), there is a strong analogybetweenthe tableaux (iv) formulaswhichare negationsof type (iii) formulas. methodand Gentzen’s (1934/5) consecution method. Toeach formulawhichis not a literal, exactly one of (Although most people would wish to maintain the followingninebranchingrules is applicable: distinction betweenthe two methods.) 126 From: AAAI Technical Report FS-93-01. Compilation copyright © 1993, AAAI (www.aaai.org). All rights reserved. [for type(ii) formulas]: Example1: Given(PvQ)&(P--->~Q),does it follow that (Pv(Q&--~P)) -~A I A ~/ (PvQ)&(Q--+--~P) q ~(Pv(Q&--,P)) I q (PvQ) q (Q--->~P) I ~P ~(Q&--~) [for type(iii) formulas]: (AvB) (A-->B) / \ / \ A B -~A B (A&B) I A B (A<-->B) / \ A --~ B ~B [for type(iv) formnl~]: ~(AvB) ~(~->B) ~A ~B A ~B [ t ~(A&B) / ~A \ ~B \ / P x Q / \ ~Q ~(A<-->B) I A ~B \ ~A B Sincethe constructionof a tableauproofin this system is really the constructionof a tree, to define a proof formallyrequires defining various notions relevant to trees -- such as ’dominates’, ’branch’, ’leaf’, ’immediatelydominates’,and the like. For the present audience it is best just to allow one’s background knowledge of trees comeinto play; andI will therefore just presuppose such notions. One terminological point: If a formulaandits negationare on a branch,we say that the branchis closed;otherwiseit is open. Onebegins the tree construction with the set of premisesand negated conclusionat the root. Onemay close a branchas soonas one notices that it contains complementary formulas(i.e., anyformulaand also its negation on the samebranch). Otherwiseone chooses complex formula(of type (ii), (iii), or (iv) above) somenode on an open branchand applies the relevant branchingrule to it, placingthe result at the bottomof all open branches dominatedby that node. (Done by extendingeachsuch openbranchin the mannerindicated by the rule.) Whenthis is done, the initial formulais "checkedoff" (by placing a ’~/’ marknext to it), indicatingthat it needsto be considered no further. This lends itself to the followingalgorithm: Continuethe followingtwosteps until either all branchesare closed or else there are no un-checkedcomplex formulasin the tree. (i) determinewhetherany just-altered branches close; if so, close the branchby placing an ’x’ at the bottom of the branch. (ii) choose an un-checked complex formula from an open branch, apply the relevantbranchingrule to it, andcheckoff the formula. Thetwo halting conditions in the abovealgorithm correspondto whetherthe initial argument is valid or is invalid, respectively. Aproof of the completenessand soundnessof the algorithmdependson K0nig’sLemma, and can be found in Jeffrey (1968:56). Here are two simpleexamples of a proof/disproofin this system. 127 x x Example2: Given~(A<--->B),does ~(B--->A)follow? q -~(A<-->B) ~/ ~(A<--->B) I q (B-~A) l -~B / \ A -~A ~B B X \ A / \ A -~A --i]3 B x The first exampleis valid, as evidencedby all the branchesclosing, while the secondexampleis invalid, as evidenced by branches remaining open yet all complex formulasbeingcheckedoff. It is pretty clear in these examples which branches were the result of working with which previously-appearing formulas; other examples maynot be so clear, and so some textbooks employa bookkeepingmethodto record this information.It should be noted that a branchingrule operates on a formulaat a node in the tree, and the result of applyingthe rule is to extend every (open) branchin the tree that the node dominatesby adding newnode(s)as called for by the particular rule being used. Theformulato whichthe rule has beenapplied is then checkedoff, to indicate that it never need be consideredagain. (Propositional tableaux systemsare therefore decision procedures: they alwaysterminate whenall complexformulas have been checkedoff or whenthe tree is closed). In the classical propositional calculusit doesnot matter, logically speaking,in which order formulas are chosen in the proof construction algorithm(althoughit certainly mattersfor efficiencyof tree construction).This aspect will changein the modal propositionalcalculus. It shouldalso be notedthat the this methodof proof is analytic: no formulaappearsin the tree unless it is a subformulaof somepreviouslyappearingformula. Thusall formulasin any proof are subformulasof the initial set of negated-conclusion From: AAAI Technical Report FS-93-01. Compilation copyright © 1993, AAAI (www.aaai.org). All rights reserved. formula. In this, classical semantictableaux stand in sharp contrast with Natural Deductionproof methods and with Resolution proof methods, which allow formulasto be generated that are not subformulasof previously-generated formulas.This aspect of classical tableaux systems will also change with the modal logicsmentioned in the title of this article. Intuitively speaking,branchinga formulaindicates that the formulais true if and only if at least oneof these newly-createdsub-branchesis true (i.e., all the formulas on the newnode of that branch are true). Hence,the initial root nodeis true exactlyif at least one branchremainsopen-- andthe initial argumentis valid just in case no branchremainsopen. Onecan evenread off a counterexample to an invalid argument(in terms of assignmentsto atomic sentences) by traversing an open branchand assigning True to (atomic) unchecked sentences whichappear unnegatedand False to those whichappearnegated. Normal Modal Propositional Logics: Modal propositional logics are formed from classical propositional logic by addingtwonew(interdef’mable) sentence operators: L ("necessarily") and M ("possibly"). In this section, 21 modalpropositional systemsare described. Theyform a subset of the socalled normalmodallogics, and include most of the well-knownmodalsystems. The techniqueused here to describe these systems is that of Chellas (1980). Starting with a fundamentalsystem, K, various axioms are introduced(called suchnamesas ’D’, ’T’, "B’, ’G’, ’4’, ’5’) and more complex systems are namedin accordancewith whichaxiomsare added to K. Thus, the modalsystem KD45results from adding axiomsD, 4, and 5 to system K. As it turns out, someaxioms imply others, some combinations of axioms are equivalent to each other, and somecombinationsare knownby other names. Attention will be called to these casesas theyarise. So, syntactically speaking, the various modal systemsof interest are seen as being built uponK. SystemKcan be axiomatizedas follows: 1. ClassicalPropositionalLogic(formulatedwith modusponensas a rule of inference) 2. 3. ~--k L(~-~F)--~ (L~--~LVF) 4. if ~k¯ then ~---kL~ Typical theoremsof systemK are M(p--u:I)~-~Lp--->Mq) tLt~q~--~pSa.q) Butsuch familiar formulasas Lp~p Lp--->LLp Lp--->Mp are not theoremsof K. To build uponK, we consider the followingsix axioms: D. L~--->M~ T. L~--->~ 128 G. ML~--->LM~ B. ~--->LM~ 4. L~---->LL~ 5. M~--->LM~ Starting with K(whichadds 0 of these axioms),there are 64 different combinationsof these six axioms,and so one might expect 64 different modal systems. However,there are certain implications betweenaxioms and equivalences amongst groups of axioms. The relevantimpficationsare: T~--~>D B~--~>G 5 ~> G andthere are the followingequivalences KIM <~> KB5 KDB4<~>KTB4<~>KT45<~-.m->KT5<~>KTB5 (and anyother implicationsandequivalencesentailed by these). This leaves us with the 21 modal systems pictured in Figure 1 (adaptedfromChellas 1980p.132) presentedat the endof this article. Thelines between systemsindicate properinclusion: all theoremsof the weakersystemare also theoremsof the stronger system, but not conversely. Oneinteresting fact about these systemsis that they have a semantics whichcan be described by a binaryrelation (calledthe "accessibilityrelation", R) "normal possible worlds". The intuition is that a sentenceis (not just simplytrue but rather) true at a possible world, and whenthe sentence has one of the modaloperatorsas a mainconnectivethen it is true at a particular worldjust in case the demodalized formulais true at every (or some-- depending on the modal operator involved) accessible possible world. For example,L~is true at possible worldwjust in case is true at every possibleworldthat is accessibleto w. In a normal modallogic, to say that a sentence is semantically valid is to say that it is true at every possible worldin every modelin whichR obeyscertain constraints. Obviouslythese definitions put a lot of weighton just whatthe relation R is. It is well-known whatthe relevant requirementson R are for the modalsystemsunder consideration here. In systemK, R has no special requirements;therefore, a ~) sentence ¯ is valid in K (whichwerepresent as ~k just in case for any relation R, ¯ is true at every possibleworld(underthat accessibility relation, R). (It is becausethere are no special requirementsplacedupon R that systemK is "the minimalnormalmodallogic"). Otherlogics addconditionsuponR, for examplethat it mustbe reflexive or that it mustbe connected.Eachof the axiomsto be considered gives rise to its own particular requirementon R, and thus a logic described as KD45(for example) will impose upon R the requirements relevant to D and to 4 and to 5. Anda formula ¯ is valid in KD45( ~kd45~) iff for any relation R whichobeysthe restrictions relevant to D, 4, and 5, ¯ is true at every possible world (under that accessibilityrelation, R). From: AAAI Technical Report FS-93-01. Compilation copyright © 1993, AAAI (www.aaai.org). All rights reserved. Fromthe semanticpoint of view, a standard model Mis a triple <W,R,P> whereWis a (non-empty)set indices ("possibleworlds"),R is a binaryrelation on and P is a total function on WXLA (whereLAis the set of atomicsentencesof the language)into {trne,false }. (Thatis, P says whichatomicsentencesare true/false at whichpossible worlds). A truth-functionally compound sentence¯ is true in the modelat possible worlda in just the "normalway",viz.: ff ¯ = -,W,then~ is true at a iff ~Fis false at o~ if q~= OF&F), then ¯ is true at a iff either ~Fis trne at ¢¢or F is true at ~ if ¢, =OF->r), then¯ is true at a iff, if ~Fis true at a thenr" is true at ¢ andso on for any truth functional connectivewehappen to have. Modallycompound sentences are true in the modelat a possible worlda in a waythat makesuse of the "accessibilityrelation", R: ifO = L~, then ¯ is true at a iff~ is true atall [3 suchthatRa[3 if ¯ = M~F, then¯ is true at a iff ~Fis true at some~ such that Rctl3 Aformulais true in the modelMiff it is true at every possible worldin M. A formulais logically true in M iff it is true at every possible world in each model <W,R,P’>,whereP" might differ from P as to which atomicsentences are true at whichpossible worlds. A formulais valid iff it is logically true in everymodel <W,R’,P>.That is, a sentence is valid iff for all R" andfor all P" the sentenceis true at everyworldin each <W,R’,P’>. Thejust-dcfinod notion characterizes validity in system K. Each other system defines validity with respect to somerestricted set of R’-models. For example,these R" relations might be required to be reflexive, or they mightbe requiredto be serial, etc. As it turns out, each of the axioms which we have considered adding to system K gives rise to its own requirementon what kind of relation R must be. The roquimmonts amstated: (d) [serialityl (Vx)(xeW --+ (3y)(yeW (d), (b) implies (g), and(5) implies (g). Furthermore, the combination (b) and (4) is equivalent the combination (b) and (5). And the following combinations are all equivalentto one another: of (d), (b), (4); of (d), (b), (5); of (t), (4), (5); of andof (t), (b), Thematerial of the preceding paragraphs can be summarized bythe chart in Fig. 2 (adaptedfromChellas 1980,p.164) given at the end of this article. Let us abbreviatethe notionof ¯ beingvalid, with restrictions X placed upon R, as ~xO. Then Fig. 2 reports completenessand soundnesstheoremsfor all the modal logics wehavediscussed,saying where X is comecollection of ’K’ plus any of the letters ’D’, ’G’, ’B’, ’T’, "4’, ’5’ -- thus formingoneof the syntactic (axiomatic)systemsmentionedearlier and x is the corresponding collection chosen from amongst’d’, ’g’, "b’, ’t’, ’4’, "5" given immediately above. These soundness and completeness theorems assureus that the results of anycorrect proofprocedure for a normal modalsystem X can be mirrored by the results of any correct semanticevaluation procedure. Andsince each of these systemshas the finite model property,it followsthat there is a decisionprocedure for eachof the logics. Tableaux Methods for Propositional Modal Logics: In this section wegive tableaux methodsfor any normal modal system containing B and G. Our strategy will be to give a methodsuitable for systemK, and then to explain whatnewrules or modificationsare called for by each of these axioms.Henceto producea tableaux argumentfor KBG,for example,one proceeds as with systemK but adds on the modificationsrelevant to (b) and (g). (The samemethodcan be used to tableauxmethodsfor all the 21 systems, but that more generalmethod is savedfor a different occasion). Therules consideredabovefor classical logic are just that: classical. This meansthat theyare relevantto any one world at a time, but are inapplicable when (0 [reflexivity]CCx)(xe w-->Rxx) considering more than one world. For example,were (b) [symmetry] (Vx)fVy)(x,yeW--+ (Rxy--+Ryx)) formulaAto be Irue in one world and formula~Ato be (g) [incestuality] (Yx)fYy)(Vz)(x,y,ze true in another(evenif it be an accessibleworld),that (Rxy&Rxz--+(3w)(weW&Ryw&Rzw))) doesnot allow us to assert that any branchis or isn’t (4) [transitivity] (Vx)(Yy)(Vz)(x,y,zeW--+ closed. This suggests that wedivide our rules into (Rxy&Ryz ----> Rxz)) differentsorts: (5) [euclidean] (Vx)(Vy)(Vz)(x,y,ze Type 1 rules ("World-Bound Rules’9: This sort of rule (Rxy&Rxz --,Ryz)) containsthe classical rules wehavealreadyintroducedInthemodal system KD,forexample, a formula ¯ is rules that operate only within a specified possible valid just incase itislogically true inevery model in world. In addition to the (classical) worldboundrules which R obeys (d), that is,inwhich R isserial (ffi con- we have already comeacross, wehave two more new nected). O isvalid inKIM5 just incase itislogicallyones that involve the modaloperators. ThesenewType trueinevery model inwhich R obeys (d),(4), and 1 rules are common to all normalmodalsystems. After that is,inwhich R isserial, transitive, andeuclidean. applyingthe rule, wecheckoff the initial formula. Itcanbenoted that, justasinthecaseofthe --LA ~MA axioms, certain conditions imply others andvarious I I combinations areequivalent. Inparticular, (t)implies M~A L~A 129 From: AAAI Technical Report FS-93-01. Compilation copyright © 1993, AAAI (www.aaai.org). All rights reserved. Type 2 Rules ("World-CreatingRules"): This type of rule is distinctive and newto modal logics. When evaluating a modalizedformula, the semantic rules given abovedirect us to determinewhethera certain subformulaof the modalizedformula is true in some other possible world. Type2 rules detect whenthis situation occurs and directs us howto extend the semantictree so as to encompassnewworlds, and how these newworldsare to be initialized. This conception is so that wewill use a doubleline (rather than a single line) to indicate howthe proof tree is extendedwhen these rules are applied.Therule canbe put like this: with anybelowthe double-linebranch,exceptinsofar as permittedto do so by Type3 InterworldCopyRules. Aside effect of the our decree that no formulain one worldcan interact with a formulaof another world (except via Type3 Rules) is that all Type1 Rules must be applied within a given world before any Type 2 Rule can be applied in that world. If we do not have this requirement then it wouldbe possible for a contradictionto be presentin a world but to go unnoticedbecausethe relevant Type1 rule was not applied to the complexformula which wouldmakethe contradiction manifest. Our algorithm for modallogic therefore says that weapply only Type MA 1 rules in whateverworldwehappento be investigating 1 I until either (a) all branchesare closedin that world(and thus that worldhas beendeclaredto be impossible),or /via 2 Co) there is an openbranchon whichall formulasare I either literals, or are checkedoff, or else havea modal operator as their mainconnective.In case (b), if there are anyof the last-mentionedsort of formulas,then we I can applyType2 rules. MAn To apply the Type2 rule: each M-formulawithin a givenworldcreates a newworldunderat the endof each openbranchthat it dominates,andinitializes that world A1 A2 A3 ... An with the de-mod~liTed formula(i.e., the formulawithout Thewayto understandthis rule is: givena groupof M- the mainM-connective).Havingdone this with a given M-formula,it can be checkedoff (althoughfor ease of formulas(formulaswith ’M’as their mainoperator) all use of someof the Type 3 rules there should be a on the same(open)branch,then simultaneouslystart collection of newbranchesin whicheach newdouble- different style of check-mark,since the Type3 rules sometimes needto makereferenceto these formulasthat line branch is initiated by a different one of these have been decomposed by Type2 rules). Onecontinues formulas(withouttheir ’M’operator). to apply the Type 2 rules to every M-formulaon an Type 3 Rules ("Interworld CopyRules"): Given an open branch in the tree at the given world. Onlyafter openbranchin a world whichalready has beenextended there are no more applications of Type 2 bya doubleline (so that there are newworldscreatedoff that branch),then if this branchcontainsthe L-formulas rules in a given world can one proceed to apply the Type3 rules. (Each of the new worlds LB copyB1, B2.... Brn into each of the 1, LB2,...LBm, introduced by an application of the WorldCreating newworldsthat the L-formuladominates. Rulesis independentof the other, and each of themis (as it were)self-contained, oncethe Type3 Interworld Applicationof the Type1, 2, and3 rules: Theorderof applicationof the rules in the classical CopyRules are applied. This meansthat all of the logic madenological difference--anyorderof application branchesbelowthe double-finescan be investigated in wouldyield the samefinal answer(valid/’mvalid)as any paralleL) After the Type 2 Rules have been applied in a other. This is no longer the case, because of the given possible world, one can apply the Type3 rules, introductionof differing possibleworldswithinthe tree copying the de-modalizedformula into each of the structure. Intuitively, the addition of a double-line already-existing possible worlds dominatedby the Lbranch to an already-existing branch amounts to formula. Whena L-formulahas had each of the relevant constructing a newpossible world accessible from the Type3 rules applied to it [in someof the morecomplex world in whichwe are currently working. Giventhis logics a formulacan satisfy the preconditionsof more understanding of the double-linebranch,it is clear that than one Type3 rule, and each of them needs to be we need somerestriction on howformulas above the applied] then it can be checkedoff. (Althoughagain we doubleline can be interact with those belowthe double line. For instance, werethe formula--,A to be found employa different type of checkmarkthan with the Type1 rules). After each of the L-formulashas been abovethe double-line branch(that is, in the earlier possible world), the appearanceof A belowthe double- dealt withby all the Type3 rules that are relevantto it, one is in a position to moveto a different possible line branch (in the newpossible world) should not worldanddeal with the formnl~sthat are in that world allowus to cancelthat branch.Ourruling will be that by constructing another tree--this time within the new no formulaabovethe double-line branchcan interact possibleworld. In this newworld,oneagain starts with 130 From: AAAI Technical Report FS-93-01. Compilation copyright © 1993, AAAI (www.aaai.org). All rights reserved. the Type1 rules, then moveson to the Type2 rules, This requirementcan be satisfied in the following manner3:After applyingall the rules as indicated for system K (thereby constructing newworlds by Type Theclosingof trees, andof worlds: rules and copying formulas into them by the Type3 Within each world, one uses the World-Bound rules) also copyover the followingformulainto these Rules. If all branchesof the tree that represents a newworldsw: possible world close, then the double-line into that (B-formula): M(A1& A2 & .... & world (and hence out of the previous world) closes wherethe Ai are all the formulason the branchleading indicatingthat it wasnot a possibleworldafter all. If to worldwthat do not havethe check-offmarksrelevant any one of the double-line branches leading outward to the Type1 rules. (So they mayhavebeencheckedoff from a branchwithin somepossible world closes, then by Type2 or 3 rules, but not by Type1 rules). This that leaf is markedas closed. (Notethat this requires new rule will be applied to each of the worlds only that one of the many double-line branches constructedin the earlier world.Theeffect of this will emanatingfroma leaf needs to close in order to close be to place the B-formulainto each world j that is the leaf. Intuitively, from the point of viewof the constructedfromworldi. But then it followsthat when initial possibleworld,the branchasserts the existence branchesin worldj are expanded,it is guaranteedthat of such-and-sopossibleworlds.If it turns out that one one of the worldsthat can be reachedwill be the one of themdoes not exist, then that initial branch is initialiT~d by the de-modalized B-formula.However, that impossible- regardless of what occurs in the other is just worldi - and thus wehavesimulatedsymmetry. possible worlds.) As an example we will show that the B axiom is The goal remains as in the classical logic: validatedby this rule: determinewhetherthe tree that represents the initial world contains an open branch. If it does, then the ~(p---->LMp) argument is invalid; otherwiseit is valid. I P TableauxRules for B and G: The preceding set of --,LMp rules describes system K. All and only the valid I arguments of K will be declared so by the above M~Mp tableaux method.Andsince the methodoperates only II by decompositionof formulasinto subformulas(it is II analytic), the algorithm will detect this in a finite time. from Type2 rule -~Mp So it is a decision procedure. As mentionedat the I outset, the point of this paperis to describehowto add M(p&M--dVlp)from B rule tableaux rules for complexrules such as B and G.2 We I will see that the resulting methodis not analytic. The L~p tableauxrules for both B andGare newType3 rules. II II TableauxRule for B: p&M~Mp fromType2 rule Thesyntactic axiomfor B says that if a formulais ~p from Type3 rule true in a world,thenit is necessarythat it be possible. P Thatis, if it is actually true thenit is possiblytrue in M~Mp every accessible world. The semantic condition on x accessibility for B is symmetry.But in a tableaux setting this raises seriousdifficulties since it apparently Thecontradiction of p and ~p in the third world makes implies that whenwecreate a newworld froma Type2 that world impossible, whichgets "pushedback" into rule, and then in that newworld wish to create yet the secondworld, into the only branchof that world. another newworld, that the newlycreated worldhas to This showsthat the second world is impossible, and either be or somehowinclude the initial starting this gets "pushedback"into the onlybranchof the first world...otherwise it would not be a symmetric world. Thus the fh-st world is impossible and the accessibilityrelation. etc. 2 Afull versionof this papercontainstableauxrules for eachof the six axioms,and hencefor the 21 different systemsindicatedin Figs. 1 and2. 131 3 Thegeneral theorembeing appealedto here is known as the Bulldozer Theorem: any formula that is unsatisf’mble in a (possible worlds) modelwherethe accessibility relation is reflexive or symmetricor transitive is also unsatisfiable in an h-reflexive, asymmetric,and non-transitive model. See Segerberg (1971:80ff). From: AAAI Technical Report FS-93-01. Compilation copyright © 1993, AAAI (www.aaai.org). All rights reserved. argumentis valid: (p-~LMp)is validated by the tableaux method. Note that the B-formula is not necessarily a subformulaof any formulain the tree up to the point of its introduction.So this tableauxmethod for systemscontainingthe B axiomis not analytic. But since the B-formulacan be algorithmeticallyconstructed fromthe formulasin the tree, wemightcall the system semi-analytic. TableauxRule for G: Thesyntactic axiomfor Gsays that if a formulais possiblynecessary,then it is necessarilyposs~le.That is, if it is necessaryin any accessible world,then it must be possible in every accessible world. The semanticrule for Ois incestuality: if world1 leads to worlds2 and3, then there is a world4 that both 2 and3 lead into. Theeffect of this relation canbe gottenby the followingmethod.Weconsider any tree whichhas been constructed so as to have newworlds Wl, w2.... Wn alreadyin existenceat the end of somebranch.In these newworldsweapply all the relevant Type1 rules. But nowbefore applyingwhateverType2 rules mightapply in these worlds,weapplythis newsort of Type3 rule: If LAi, LBi... are formulasin an openbranch of wi, and LAj, LBj... are formul~in an openbranch of wj, then add M(Ai&Bi&...&Aj&Bj&...) to the end of those open branchesin each of wi and wj. Dothis for all openbranchesin eachof the pairs of worldsof Wl,w2.... Wn. Notethat each pair of worldsin Wl, w2.... Wnare now guaranteedto havea common descendant,and that this descendantwill havetrue all the formulasthat were necessarily true in its parents. So wehavesimulated incest,~lity. A simple example is to prove the G axiom: ~(MLp~LMp) I MLp --~LMp M~Mp // type2 rule: the Grule: type2 rule: \\ Lp --drip :type 2 rule I L~p I I M(p&~p)M(p&~p):G II II p&~p:type 2 p&~p I I P P --,p ~p X tiffs is the only branchin eachof these worlds,they too are each impossible, and this "pushes back" the impossibility into the initial world. (Asmentioned,it is only necessary that one of the two double-line branchesexiting fromthe initial worldbe impossible.) This showsthat the only branchin the initial world closes, and thus that the argument is valid...the unnegatedformulais a theoremin G. Onceagain wenote that this methodis not analytic becausethe formulasconstructedby the G rule is not necessarily subformnh.~of anything in the tree up to that point. Yet since the relevant formulas can be algorithmicallyconstructedfromthe tree, wecan call it a semi-analyticsystemof proof. Acknowledgements: The research reported here was supported in part by the Canadian NSERC grant OPG 5525. Theauthor is also a member of the Institute for Robotics and Intelligent Systems and wishes to acknowledge the support of the Networksof Centres of Excellence Program by NSERC,the Governmentof Cawd~and the participation of PRECARN Associates. References: Bergmann,M., J. Moor,J. Nelson (1980) The Logic Book (McGraw-Hill). Beth, F. (1952)Symbolic Logic. (North-Holland). Beth, F. (1958) "OnMachineswhichProve Theorems", in Siekmann& Wrightsonpp. 79-90. Bonevae,D. (1987) Deduction(MayfieldPub. Co.). Chellas, B. (1980) ModalLogic (CambridgeUniv. Press). Fitting, M. (1983) Proof Methods for Modal and lntuitionistic Logics(Reidel). Fitting, M. (1988) "First Order ModalTableaux" Journal of AutomatedReasoningpp. 191-213. Garson, M. (1988) "Quantified ModalLogic" in Gabbay& F. GuentlmerHandbookof Philosophical Logic Vol 2. Gentzen, G. (1934/5) "Investigations into Logical Deduction". Translation printed in M. SzaboThe Collected Papers of Gerhard Gentzen (NorthHolland)1969,pp. 68-131. Jeffrey, R. (1968) FormalLogic: Its Scopeand Limits (McGraw-HtU). Segerberg, K. (1971) An Essay on Classical Modal Logic.(Uppsala:FilosofiskaStudier). Siekmann,J. & Wrightson,G. (1983) The Automation of Reasoning,Vol. 1. Springer-Verlag. Smullyan, R. (1968) First Order Logic. (Springer- vcr~g). X The closure of at least one of the two worldsat the bottomof the tree will "pushback"the impossibilityof that worldinto the precedingworld. Thuseach of the previousworlds(the oneswhichare initialized with Lp andwith -drip) will get an x at their bottom.Andsince 132 From: AAAI Technical Report FS-93-01. Compilation copyright © 1993, AAAI (www.aaai.org). All rights reserved. Figure I: 21 normalmodalsystems. Arrowsindicate propexinclusion. conditions on R (.~ ¯ (.) ¯ t’~ (-) ¯ (:,~ (./ ¢.~ ¯ t.~ ¯ ¯ ¯ (.) (q C-) (.) (-) ¯ (’) (-) .. (q ¢.~ fq ¯ . . ¯ ¯ (-) (-) ¯ TABLE 1: The semantic requirements on R for 21 normal modal systems.¯ = a required property, (-) = property entailed required properties. KB4and KT5are shownwith different (equivalent) combinations of required vs. entailed properties. 133