测品娱乐
您的当前位置:首页Lindstr om Quantifiers and Leaf Language

Lindstr om Quantifiers and Leaf Language

来源:测品娱乐
Lindstr¨omQuantifiersandLeafLanguage

Definability

Hans-J¨orgBurtschickFachbereichInformatik

TUBerlinFranklinstr.28/29D-10587Berlin

HeribertVollmerTheoretischeInformatikUniversit¨atW¨urzburgAmExerzierplatz3D-97072W¨urzburg

Abstract

Weintroducesecond-orderLindstr¨omquantifiersandexamineanalogiesto

theconceptofleaflanguagedefinability.Thequantifierstructureinasecond-ordersentencedefiningalanguageandthequantifierstructureinafirst-ordersentencecharacterizingtheappropriateleaflanguagecorrespondtooneanother.Undersomeassumptions,leaflanguagedefinabilityanddefinabilitywithsecond-orderLindstr¨omquantifiersmaybeseenasequivalent.Alongthewaywetightenthebestuptonowknownleaflanguagecharacterizationoftheclassesofthepolynomialtimehierarchyandgiveanewmodel-theoreticcharacterizationofPSPACE.

1Introduction

InitiatedbyFagin’sseminalpaper[Fag74]characterizingNPastheclassofgener-alizedspectra,therehasbeenalonglineofresearchincharacterizingcomplexityclassesusingnotionsfromfinitemodeltheory(seetherecenttextbook[EF95]).Inclassicalcomplexitytheoryproblemsareclassifiedbydefiningrestrictionsondifferentcomputationalresourcesase.g.spaceortime.Usingincontrasttothisthesyntacticcomplexityofadefiningformulaitispossibletomeasurethedescriptivecomplex-ityofproblems,e.g.byconsideringthecomplexityofquantifierprefixesofprenexformulae.Besidesthewell-knownexistentialanduniversalquantifiersrecentlytheexpressivepowerofacertainkindofgeneralizedquantifiers,thesocalledLindstr¨omquantifiers,hasbeenexamined[Lin66];foranoverviewseeChapter10of[EF95].Anextensiveoverviewovertheliteraturecanbefoundin[Got95].

1

Inthefieldofcomputationalcomplexitycharacterizationsofcomplexityclassesbysocalledleaflanguageshavebeenstudiedintensively.Thisapproachwasin-troducedbyBovet,Crescenzi,andSilvestriin1992[BCS92]andindependentlybyVereshchaginin[Ver93]asaunifiedapproachtodefinecomplexityclasses.Considerapolynomial-timenondeterministicTuringmachinethatprintsavalueoneverypathofitscomputationonsomegiveninput.Thesevaluesareattheleavesofthe(notnecessarilyfullybalanced)binarytreedefinedbythenondeterministicchoicesofoninput.Anorderingofthetuplesintheprogramofdetermines,foreachnondeterministicchoicenodeofthetree,whichchildis“left”andwhichis“right”.Thisinturnfixesaleft-to-rightorderingofalltheleaves.Wecallthecorrespondingsequenceofvaluestheleafstringofoninput.(TheparticularorderingoftuplesintheprogramofMdoesnotmatter.)Givennowa(leaf)language,acomplexityclassLeafPisdefinedasfollows:AsetisinLeafPifandonlyifthereisanondeterministicTuringmachinethatoninputproducesaleafstringwhichsuchthatifandonlyif.

StartingwiththesomewhatsurprisingresultthattheclassPSPACEcanalreadybecharacterizedinthiswayusingonlyregularleaflanguages[HLS93],alotofcharac-terizationsweregiven[JMT94,HVW95,HVW96,CMTV96],leadingtoanumberofremarkableandunexpectednormalformsforsuchclassesasPSPACE(viabottleneckmachines[HLS93]),PP(viamachineswhereacceptingandrejectingcomputationpathsformtwoclusters[HVW96]),and#P(whereallrelativizableclosurepropertieswereidentified[HVW95]).Evenseparationsofcircuitclassescouldbeprovedusingleaflanguagesasatechnicalvehicle:Allendershowedin[All96],buildingonpreviousworkby[CMTV96],thatTCPP.

Itturnsout,asweshowhere,thatbothconcepts—Lindstr¨omquantifiersandleaflanguagedefinability—arecloselyrelated:Leaflanguagedefinedcomplexityclassescan(undercertainassumptions)becharacterizedbysecond-orderLindstr¨omquanti-fiersthatarefollowedbyfirst-orderformulas,andviceversaforeveryformulaofthisformthereisaleaflanguageandanondeterministicpolynomiallytime-boundedTur-ingmachinethatdefinethecorrespondingmodelclass.ThusLindstr¨omquantifiersandleaflanguagedefinabilityaretwopointsofviewofessentiallythesameconcept.RelyingonbothviewsonemighthopetogetnewinsightsintotheexpressivepowerofLindstr¨omquantifiersonorderedfinitestructuresandpropertiesofleaflanguagedefinedcomplexityclasses.

Wegivesomeresultsinthisdirectioninthispaper:Thepolynomialtimehierar-chyintroducedbyStockmeyerin[Sto73]isoneofthecentralconceptsincomplexitytheory.Thusofcoursetherehavebeeneffortstogiveleaflanguagecharacteriza-tionsofthishierarchy.In[HLS93]itwasshownthatusingACasleaflanguage

2

class,oneobtainstheunionofallclassesofthepolynomialhierarchy,andusingthe-thleveloftheBrzozowski-(ordot-depth-)hierarchy[Str94]asleaflanguageclass,thebooleanclosureoftheclasspisobtained.In[JMT94]itwasprovedthat

p

LeafP-LOGTIME.Usingfirst-orderdefinableleaflanguagesweproveatighterconnection:Weshowthatthe-thlevelfromthepolynomialtimehierarchy

for-canbecharacterizedusingleaflanguagesthatcanbedefinedusingfirst-order

mulae.Asanotherexampleofhowacombinationofmodel-theoreticandcomplexitytheoreticargumentscanleadtonewresults,weshowhowtouseatheoremfrom[HLS93]toderiveanewmodel-theoreticcharacterizationofPSPACE.

TheseresultsarespecialcasesofageneraltheoremwhichstatesthatifaleaflanguageisdefinedbyaformulainprenexnormalformstartingwithoneLindstr¨omquantifierandfollowedbyasequenceofuniversalandexistentialquantifiers,thenthecharacterizedcomplexityclasscanbedefinedinamodel-theoreticwaybyasecond-orderformulawithexactlythesamequantifierstructure.Thusweseethatusingmeth-odsfromfinitemodeltheoryweareabletotightenexistingcomplexitytheoreticre-sults.Wethinkthatgenerallyitisagoodideatospecifyleaflanguagesinadescriptiveway.Onepurposewhenconsideringleaflanguagedefinabilityistoseparatethecom-putationalaspectinthedefinitionofaclassfromthewaythecomputation(informofthecomputationtreeconstructedbyanondeterministicmachine)isevaluated.Thisevaluationschemeshouldthereforebegiveninawayfreeofcomputationalaspectswhatsoever.Thisgoalisachievedusingfinitemodeltheorytospecifyleaflanguages.Otherconnectionsbetweenleaflanguagedefinabilityandfinitemodeltheoryareestablishedin[Vei96].Heconsiderslogicalreducibilities[EF95]ase.g.first-orderreducibilityandprojectionreducibilitytosuccinctrepresentationsoflanguages.Thisconceptofsuccinctrepresentationsmaybeviewedasaspecialcaseofleaflanguagedefinability(seealso[BL96]).However,theresultsachievedinhispaperareincom-parabletoours.

2Preliminaries

Leaflanguageswereintroducedandusedtodefinecomplexityclassesin[BCS92,Ver93,HLS93](seealsotherecenttextbook[Pap94]).Letbeapolynomialtime-boundednondeterministicTuringmachine.Givenaninput,suchamachineproducesoneverypathasymbolfromsomefinitealphabet.Letbethestringofthesoproducedsymbols(basedonthenaturalorderofpathsofthemachine).Thenfor

,wedefineLeaf.TheclassLeafPisthedef

classofalllanguagesLeafforallnondeterministicpolynomial-timemachines.Here,isthesocalledleaflanguagedefiningtheclassLeafP.Ifisaclass

3

oflanguages,thenwedefineLeafP

def

LeafP

.

Asitturnedout[JMT94,HVW96],itplaysacrucialrolewhetherintheabovedefinitionweallowarbitraryTuringmachinesorwerequirethatthecomputationtreesproducedhaveaspecialshape.ForthecaseoffullbinarytreesweusethenotationFBTLeafP,see[Her95].Note,thatwithrespecttotheproducedleafstringonlythenondeterministicbranches,butnotthelengthofthepathisrelevant.Intherest

havingthefollowingofthepaperwewillusenondeterministicTuringmachines

property:Letbethesmallestnaturalnumbersuchthatistime-bounded.Thenallcomputation-pathsareexactlyoflength.Thismaybeachievedinanuniquewayusingleftbranches,ifthelengthofapathissmallerthan.Foreverypolynomialtime-boundednondeterministicTuringmachinethereexistsapolynomialtime-boundednondeterministicTuringmachinefulfillingtheabovepropertysuchthat

forallinputs.Thus,whenwriteLeafPweuseTuringmachines

asdescribedabove.

Itcanbeseenthatinaspecialcasethedistinctionbetweenfullbinarytreesandthegeneralcaseisunnecessary.Forthis,givenalanguageforsomealphabet,andaletter1,definepad1tobetheregularlanguage11111for;andletpad1pad1.Saythatalanguageclassdefisclosedunderpaddingiffforalsopad1.NowitcanbeseenthatLeafPFBTLeafPwheneverisclosedunderpadding[JMT94].Weremarkthatgenerallyalllanguageclassesdefinedbyautomata(finiteautomata,pushdownautomata,Turingmachines)areclosedunderpadding.Thisalsoapplies,aswewillseelater,tomostofthesubclassesoftheregularlanguagesweconsider.

Laterwewillalsoallowdifferentpaddingsymbols.Forthisletbeanalphabetsuchthat.Thenpadisthelanguagepaddef

for.

Inthispaper,weconsiderclassesthatarecharacterizedvialogicallydefinedleaf

languages.Especially,wedealwiththefollowingclassesofformulae:

FOistheclassofallfirst-orderformulae,SOistheclassofallsecond-orderfor-mulae.Weusethenotationsand(and)todenotethesubclassesofFO(SO,resp.)consistingofprenexformulaeinformorform,see[EF95].ForaformulawedenotetheclassoffiniteorderedstructuresthataremodelsofbyMod.Sinceweareinterestedincharacterizationsofcomplexityclasses,wecon-siderinthispaperonlyorderedfinitestructures,thatis,structuresoversignaturesthatdonotcontainconstantorfunctionsymbolsbutdocontainasymbolforalinearorderontheuniverse.Wedonotusethesuccessorrelation.

Weuseanencoding

ofstringsintoorderedfinitestructuresdefinedintheusual

4

way[Fag74]asfollows:Letbeafinitesetofsymbolsandlog.Theimagesofareorderedfinitestructuresoverthesignaturewithunarypredicates.Forafixedenumerationofthesymbolsin,letbethe-thsymbolin.Ifthe-thelementinthestringis,thenin

,iffthe-thbitinthebinaryencodingofis,foralland.

isapoweroftwo,thenisabijection.IfisaNote,thatisone-oneandif

language,thenwewritefor.Ifisaclassoflanguages,then

.def

Furthermorewesometimeshavetoencodefiniteorderedstructuresintostrings.

Sinceisone-one,isafunction,too.Wedefineanencodingfromfiniteor-deredstructuresintostringsthatincludesbutalsoextendstomaprelationsofaritygreaterthanoneintostrings.AnalogouslytotheabovewewriteModfortheset,andgenerallyforastructureclasswewritefor

.(Following[Fag74]wechoosetousethesymbolsandforour

encodingsthoughtheyaremaybenotveryintuitive.)

Ifisalogic(ase.g.FOorSO)andisacomplexityclass,thenwesaythat

capturesiffeverypropertyover(encodingsof)structuresdecidablewithinisexpressiblebysentences(i.e.,Mod),andontheotherhandforeveryfixedsentence,determiningwhethercanbedonein(i.e.,Mod).Asanabbreviationwewillmostofthetimesimplywrite.

Ifisalogic,thenLeafPisthecomplexityclasscharacterizedbytheleaflanguageconsistingofallwordssuchthatthecorrespondingstructureisamodelofanformula,insymbols:LeafPifthereexistaandapolynomialtime-boundednondeterministicTuringmachinesuchthatforall,

.

InaseminalpaperFagin[Fag74]connectedthecomplexityofcomputationswith

thesyntacticcomplexityoflogicallanguages.Heshowedthefollowing:LetbeanondeterministicpolynomialtimeTuringmachinethatrecognizesalanguageoversomealphabet.Letusassumethatoneveryinput,’scomputationtreeisafullbinarytree.Faginconstructedanexistentialsecond-orderformulathatdescribesthecomputationof.Theformulaisoftheformwhereisafirst-orderformulathathasnofreefirst-ordervariablesandthatadditionallycontainsthefreesecond-ordervariables.Thearitiesandthenumberofthesecond-ordervariablesaredeterminedbytheconstruction,see[Fag74].

Theorem2.1[Fag74].NP

,i.e.

NP

Mod

and

Mod

NP.

time-boundedandLetusconsideranondeterministicTuringmachinethatis

thatproducesafullbinarycomputationtreeforallinputs.Itispossibletorepresent

5

apathinacomputationtreeofdepthusingastring.Weenumeratethepathsusingthelexicographicalorderingon.Analogously,wecanenumeratetheassignmentsofa-arysecond-ordervariableusingthelexicographicalorderingoncharacteristicsequences.Inthispaperwewant

inthecomputationtreeandthetouseadirectcorrespondencebetweenthepaths

of:assignments

Proposition2.2.Letbeatime-boundednondeterministicTuringmachineas

above.Thereexistsaformulaoftheform,suchthatPathacceptforallinputsoflengthandforall,wehavethatacceptsonpathiff.Here,PathexpressesPathaccept

thatencodesapathcorrespondingtonondeterministicguesses,andacceptexpressesthatthepathencodedbyisaccepting.

Proof.(Sketch.)Weslightlymodifytheconstructionin[Fag74]byintroducinga

-arysecond-orderpredicatesuchthatitsassignmentrepresentsthenondeter-ministicchoiceforeachstepofthecomputation.Thatis,foreachassignment

uthereexistsexactlyoneassignmentsuchthatPath.

3Lindstr¨omQuantifiers

Inordertoextendthelimitedcapabilitiesoffirst-orderpredicatelogicithasbeentried

toenhanceitsexpressivepowerbyaddingdifferentsortsofoperatorsorquantifiers.OnesystematicandformallyelegantapproachtothisistoconsidersocalledLindstr¨omquantifiers.

Afirst-orderformularelationdef

with

freevariablesdefinesforeverystructure

,see[EF95].

the

Nowtheideaunderlyingthedefinitionofgeneralizedquantifiersisthefollowing:Considertheclassicalfirst-orderexistentialquantifierappliedtosomequantifier-freeformulawithfreevariable,i.e.,considertheformula.Givenanorderedfinitestructure,iff.Analogouslywegetiff

(istheuniverseof).[Str94]definesmodularquantifiersasfollows:

iffmod.Thusitisverynaturaltodefinegeneralized

quantifiersbyconsideringarbitraryconditionsonrelationsdefinedbyformulae.

StructoverasignaturePEveryclassofstructures

definesthefirst-orderLindstr¨omquantifierasfollows:Letbefirst-orderformulaeoversignaturesuchthatforthenumberoffree

6

variablesinisequaltothearityof.ThenaFOformula.IfStruct,then

is

iff

Wewanttostressthefollowing

andtheelementsofturesboth

isrelatedtostructure

verse.Inthejustgivendefinition,theorderofinbothcasestobe

asfollows:Ifsemanticsof

point:Sinceweonlydealwithorderedstruc-areordered.Inthejustgivenequivalence

,i.e.astructurewiththesameuni-wemadetheconventionthathereweassumeidentical.Tobemoreexplicit,wedefinetheisanorderedstructure,then

iff.

forFollowing[Got95]wealsowrite

denotethesubclasseswherealltheformulae

FO.Moreoverlet

areprenex(

()

)formulae.

ThejustgivendefinitionisessentiallytheoriginaldefinitiongivenbyLindstr¨om[Lin66],whichthereaderwillalsofindintextbooks[Ebb85,EF95]andformulatedwithslightvariationsbutequivalentlyin[Got95].Forourexaminations,thefollowingformulationwillbeuseful(observethatthisonlymakessensefororderedstructures):Givenafirst-orderformulawithfreevariablesandacorrespondingfiniteor-deredstructure,thisdefinesthebinarystringoflength(),thesocalledcharacteristicstringof.Nowgivenasequenceofformulaewithfreevariableseachandastructure,wesimilarlygetthetuple,where.Certainly,thereisaone-onecorrespondencebe-tweensuchtuplesandstringsoflengthoveralargeralphabet(inourcasewithelements)asfollows.Letbesuchanalphabet.Fixanarbitraryenumerationof,i.e..Thencorrespondstothestring

,wherefor,,forthatwhoselength

binaryrepresentation(possiblywithleadingzeroes)isgivenby.Insymbols:.

isinfirst-orderThisleadsustothefollowingdefinition:Asequencewordnormalform,iffthehavethesamenumberoffreevariables.NextweshowthatifwerestrictsuchthattheFOpartisrequiredtobeinfirst-orderwordnormalformwegetnolossinexpressivepower.

Lemma3.1.LetbeaFOformulaoversigna-StructforarbitraryP.Thenthereareasigna-ture.Let

ture,aclassofstructuresStruct,andfirst-orderformulaeoverinwordnormalformsuchthatModMod.

7

Proof.Letbethemaximalarityofpredicatesymbols()in.Definetoconsistofforalinearorderpluspredicatesofarityeach.Definetoconsistofexactlythosestructuresforwhichthereisan

suchthatforallwehave:

(whereisthearityof,andisasymbolfor

theminimalelementin).Forthecharacteristicstringsthishastheconsequencethat

.Nowitisnottoohardtobuildtheformulaesuchthat

forall,.u

ThusthefollowingdefinitionofLindstr¨omquantifiersoverlanguagesisequivalent(w.r.t.expressivepower)totheoriginaldefinitionofquantifiersoverarbitrarystructureclasses:

Let

,andlet

beinfirst-orderwordnormalform.Letbeanalphabetsuchthat.Theniff

.

WeremarkthatourdefinitionofLindstr¨omquantifiersoverlanguagesresembles

verymuchthedefinitionofgroupquantifiersfrom[BIS90].Ifisthewordproblemofsomefinitegroup,theninournotationisexactlythegroupquantifiergivenby.

Tocharacterizethepowerofleaflanguagesdefinedwithintroducesecond-orderLindstr¨omquantifiers.

formulae,wenow

Givenaformulawithfreesecond-ordervariablesandastructure,define,andletbethedef

correspondingcharacteristicstring,theorderofvectorsofrelationsbeingthenaturaloneinducedbytheunderlyingorderoftheuniverse.Ifthearitiesofare

,resp.,thenthelengthofis.

Letbeasignature,wherefor

.Thusisasignatureconsistingofalinearorderplusasequenceofsignatureswithonlypredicatesymbolseach.Letbethearityof.Asecond-orderstructureofsignatureisatuple,whereforevery,

.Givennowasignatureandsecond-orderformulaeoverwhereforeverythenumber

andarityoffreepredicatesincorrespondsto.Letbeaclassofsecond-orderstructuresover.ThenisaSOformula.If

Struct,then

iff

8

Againwewouldliketostressthatweassumetheorderonsidesoftheequivalence.

tobethesameonboth

Again,wewanttotalkaboutsecond-orderLindstr¨omquantifiersdefinedbylan-guages.Thuswedefineanalogouslytotheabove:Asequence

ofsecond-orderformulaeisinsecond-orderwordnormalform,ifthehavethesamepredicatesymbols,i.e.intheaboveterminology

.Letforthearityofbe.Observethatinthiscase,

(for),thuscorre-spondstoawordofthesamelengthoveranalphabetofcardinality.Asbefore,we

canassumesecond-orderwordnormalformwithoutlossofgenerality:

Lemma3.2.LetbeaSOformulaover

signature.Letbeaclassofsecond-orderstructuresoversomesigna-turewherefor.Thenthereareasecond-orderstructureoverasuitablesignatureandformulaeoverinsecond-orderwordnormalformsuchthatModMod.

Proof.TheproofissimilartotheproofofLemma3.1.Definesuchthatforeachcontainspredicatesymbols,whereisthemaximalnumberofpredicatesymbolsinany,.Thethpredicatesymbolinanyisofarity,whereisthemaximalarityofallthepredicatesymbolsinany,

.consistsofthosesecond-orderstructures

forwhichthereisaninsuchthatforwehave.Hereisthedef

numberofpredicatesymbolsin,andisobtainedfromasinLemma3.1.Nowitisanottoohardexercisetodefineformulaesuchthatthefollowingholds:whichimpliesthestatementofthelemma.u

Letbeasequenceofsecond-orderformulaeinsecond-orderwordnormalformasabove.Givennowalanguagewith

,thesecond-orderLindstr¨omquantifiergivenbyisdefinedby

iff.Lemma3.2showsthat

second-orderLindstr¨omquantifiersgivenbylanguagesareequivalent(w.r.t.express-ibilitypower)tothosegivenbyarbitrarysecond-orderstructures.

Similarlytotheabove,()denotesthosesubclassesofSO,wherealltheformulaeareprenex()formulae.Foraclassoflanguagesweusethenotationwiththeobviousmeaning,e.g.FOdenotestheunionofallFOoverall.

9

4Padding

Inthesequelwewilluselogicallydefinedleaflanguages.Asalreadypointedoutinthepreliminaries,paddingsometimesplaysacrucialroleintheleafcontext.Thuswewillconsiderbesidespaddingonlanguagesdefinedearlieralsopaddingonmodelclasses.FromthedefinitionoflogicallydefinedleaflanguagesinSect.2itisclearthatwearemainlyinterestedinstructureswhichareimagesoftheencoding.

Lettheencoding

besuchthat

thsymbolof

for

,

,where

.

and

isthe

Let,whereallpredicatesymbols(

Structthesamearity.Forastructure

.

)areof

let

def

Theideabehindpaddingonstructuresisthatwewanttoinsertapaddingsymbolinatupleofcharacteristicstrings,i.e.in.Forthisend,wehavetoextendouralphabetbyapaddingsymbol.Actually,whatwedoiswetakeasnewalphabet

andallsymbolsinmaybedef

usedaspaddingsymbols.Moreformally:

Let,whereisofthesamearityastheotherdef

predicatesymbols.DefinepadStructasfollows:paddef

pad.(Herepadreferstopaddingonstringsasdefinedin

Sect.2.)Ifisamodelclass,thenpadpad.Ifisaclassofdefmodelclasses,thenpad.defpad

Wesayalogicisclosedunderpaddingiffthecorrespondingclassofmodelclassesis,i.e.ifpadModMod.

In[Bur96]somelogicsclosedunderpaddingandsomenotclosedunderpaddingareidentified.Theinourcontextrelevantresultsarethefollowing.Proposition4.1.

1.For

,

and

areclosedunderpadding.

2.Ifisclosedunderpadding,thenforall.

and

areclosedunderpadding

Proof.(Sketch.)Theprincipalideaisthatonehastorelativizetheoccurringquan-tifiers,e.g.insteadof“forallpositions”wenowhave“forallpositionsuchthatthethletterisnotapaddingsymbol.”Thiscertainlydoesnotaffectthelinearor-der.Concerningtheotherrelations,thesuchobtainedformuladoesexactlywhatisrequired,i.e.itmodelsthepaddedversionoftheoriginalmodel.Moredetailscanbefoundin[Bur96].u

10

5MainResults

Beforewecometoourmainresultweexamineleaflanguagecharacterizationsoftheclassesofthepolynomialtimehierarchy.Lemma5.1.LeafP

forall

.

p

-LOGTIMEwasproved.Thus,sinceIn[JMT94],LeafP

p

-LOGTIMEand[Sto73],ourlemmafollows,butwegiveanewproofhere,sincewewanttogeneralizeitlater.

Proof.ForagivenisticTuringmachinethat

-formulaandapolynomiallytime-boundednondetermin-,weconstructaformulaandthenshowforall

iff

LeafP

Letbeapolynomiallytime-boundednondeterministicTuringmachine.Letus

firstconsiderthecase,that,namelythesetofsymbolsthatproducesoneverycomputation-path(seeSect.2),hastwoelements.Wewilldiscusstheothercaseattheendoftheconstruction.

bethe-formulawhichformalizestheLetPathaccept

computationofasdescribedinSect.2,Proposition2.2.Anassignmentofrep-resentsthenondeterministicchoices.Forafixedassignmentofthereexistsexactlyoneassignmentofthatrepresentsacomputationpathof.ForallsuitablefiniteorderedstructurestheformulaPathistrueon,iffrepresentsacorrectcomputationpathw.r.t.thenondeterministicchoicesgivenby.

Let

bethe

-formulaover

oftheform

Weconstructa-formula

inthefollowingway:

byreplacingthevariablesandatomicpredicatesin

[]Wereplaceallfirst-ordervariablesinfordowntobytuplesofsecond-ordervariablesandthustransformtheexistentialfirst-orderquantifierthatbindsintoasequenceofsecond-orderquantifiersthatbindand.Toensurethatanassignmentofrepresentsacorrectcomputationpath,wereplaceby

.Inordertoobtainaformulainprenexnormalform,wePath

havenovariablesincommon.renamethevariablesinPathsuchthatPathand

11

[]WereplacetheuniversalquantifierwerenamethevariablesinPathsuchthat

byPathand

.Again

havenovariablesincommon.

Path

[]Wereplaceeachoccurrenceofbythe-formula

byasuitablerenamingofthevariables.weobtainfromaccept

accept

,which

[]Wehavetoextendtheorderingthatisgivenintheorderedstructureto

assignmentsof.Thiscanbeachievedbywell-knowntechniques.

AsdescribedinSect.2,itispossibletoencodethesymbolsofinbinary.For,wehavetoconsiderfirst-orderformulaeoversignaturesthathavelogunarypredicates.Wereplaceeachbyaformula.Foreach

weobtainfromacceptaformulawhichistrue,iffproducesontherepresentedpath.Theformulaisadisjunctionoftheformulaeforthose,suchthatthe-thbitinthebinaryencodingofis1.

ToseethatLeafP,iffMod,forall,observethatthereisaone-onecorrespondencebetweentheelementsintheuniverseof(thepositionsofthesymbolsintheleafstring)andtheassignmentsof.

u

Lemma5.2.

LeafP

forall

.

FO,thislemma(incontrasttoLemma5.1)doesnotSinceevenDLOGTIME

p

followfromLeafP-LOGTIMEprovedin[JMT94],butinfactstrengthenstheirresult.

Proof.Letbeaformulawithsecond-ordervariables.Weconstructapolynomiallytime-boundedNTMandaformuladefiningaleaf-languageMod,suchthatforallorderedfinitestructuresoverthecorrespondingsignature:ModiffLeafP

OninputtheNTMdeterminesnondeterministicallytheassignmentsofthesecond-ordervariablesfromto.Thatis,oneachcomputationpath,writesthecharacteristicsequenceofonepossibleassignmentonitsworktapes.Usingthesecharacteristicsequences,evaluatesthefirst-orderpartof.Weusetwoleafsymbolsandtodenotetheresultoftheevaluation.

Forformulas,thecorrespondingleaflanguageisdefinedbyformulasthecorrespondingleaflanguageisdefinedby.

,andfor

)the

For,wehavetomarkforallsecond-ordervariablespathsthatsharethesameassignmentofasshowninFig.1.

(

12

Figure1:Constructionofmachine

Weintroducenewleafsymbolssymbolsofoftheleafstringbetweenandsharethesameassignmentfor.

and.The

correspondtoallthosepathsthat

Asdescribedinthepreliminaries,thesymbolsofcanbeencodedusinglogunarypredicates.Bybooleancombinationsofthese,weconstructformulaethatex-.Weabbreviatethesepressthatapathproduces,,,orfor

formulaeby,,,andresp.Theformulawillbeabbreviatedbymarked.

For

letRegion

beanabbreviationfortheformula

For

left

weconstructinductivelyasequenceofformulae

betheformularight

.Let

left

right

iftherightmostquantifierisexistentialand

left

left

marked

otherwise.Forthe-thquantifierblockletusfirstconsiderthecase,

thatthequantifierblockitselfconsistsofexistentialquantifiersandthatthequantifierblocktotherightconsistsofuniversalquantifiers.Letbetherightmostvariableinthe-thquantifierblock.Wedefinetobetheformulaleftright

left

right

left

left

right

right

Region

left

right

left

right

13

Transformingthisintoprenexnormalformwegetaquantifierprefixstartingwith

.Letusnowconsiderthecase,thatthe-leftrightleftrightthquantifierblockconsistsofuniversalquantifiersandthatthequantifierblockstotheleftandtotherightconsistofexistentialquantifiers.Letbe

totherightmostvariableinthe-thquantifierblock.Wedefineleftright

betheformula

left

right

left

left

right

right

Region

left

right

left

right

Transformingthisintoprenexnormalformwegetaquantifierprefixstartingwith

.leftrightleftright

Letbetherightmostvariableinthefirstquantifierblock.If

betheformulaquantifiedthenlet

isexistentially

left

right

Region

left

right

left

right

otherwiselet

be

left

right

Region

left

right

left

right

u

FromLemma5.1andLemma5.2wegetthefollowingleaflanguagecharacteriza-tionoftheclassesofthepolynomialtimehierarchy:Theorem5.3.LeafP

forall

.

Acloseinspectionoftheproofin[Fag74]showsthatthereisadirectcorre-spondencebetweencomputation-pathsandassignmentsoftheexistentiallyquantified

second-ordervariables.Thus,usingsecond-orderLindstr¨omquantifierswecande-scribetheconceptofleaflanguagedefinabilityintermsofmodel-theoreticnotions.Note,thatwedonotrequirethecomputationtreetobeafullbinarytree.

Proposition5.4.Letbealanguage,letbeanumbersuchthat,andletbeanondeterministicpolynomialtimemachineasdescribedinSect.2(i.e.withallpathsofthesamelength).ThenthereexistsaformulaandPath

suchthatasequenceofformulaeleafleaf

Leaf

Mod

Path

leaf

Path

leaf

14

Proof.Letbetheformula,whichisdescribedinPropo-Pathacceptsition2.2.Foreachcomputationpathofoninputthereexistsexactlyoneassign-mentsuchthat

Path

Forwemodifytheformulaacceptsuchthatitistrue,iffproducestheleafsymbol.Wedenotetheresultingformulabyleaf.Inthecorrespondingencodingoftheunarypredicateistrue,iffisproduced.Thus,wehave

Leaf

iff

Mod

Path

Forweconstructtheformulaeleaffromacceptaccord-leafingtotheencoding(seeSect.3):Forallletbeaformulathatistruefor,iffproducestheleaf-symbolonthepaththatcorrespondsto.Eachformulaleafnowconsistsofadisjunctionofallsuchthatthe-thbitinthebinaryrepresentationofisone.u

Corollary5.5.Let

beaclassoflanguages.

LeafP

Thesecond-orderLindstr¨omquantifierintheabovepropositionisappliedto-formulaeinordertoruleoutthoseassignmentsofthatdonotcorrespondtocompu-tationpathsof.Forlanguageclassesthatareclosedunderpadding(seeSect.4)wecangetridofthesecond-orderexistentialquantifier.Wemapthoseassignmentsofthatdonotcorrectlyrepresentacomputationpathoftothepaddingsymbol.

Lemma5.6.Let

beaclassoflanguagesthatisclosedunderpadding.Then

LeafP

LetbeanondeterministicTuringmachine,,and

beformulaeasdescribedinProposition5.4suchthatLeaf

Mod.WemodifythisformulasuchthattheLind-str¨omquantifierbounds,also.Asmentionedabove,wemapthoseassignmentsofthatdonotcorrectlyrepresentancomputationpathoftothepaddingsym-bol.AccordingtoSect.4,wehavetoextendthesequencebyanadditionalformula.Thenforallassignmentsthecorrespondingsymbol

isapaddingsymbol,iffistruefortheassign-in

Proof.():

ment.Theformulaistrue,iffforgivenassignmentsthePath

doesnotcorrectlyrepresentacomputationpathofwithrespecttoassignment

15

thenondeterministicchoicesrepresentedbyindicatethepaddingsymbol.Thuswehave

Leaf

.Wethereforeuse

Path

to

Mod

pad

Path

():Forprovingthisinclusionwedonotneedthefactthatisclosedunderpadding.

andwecanconstructapolynomial-timenon-Forgiven

deterministicTuringmachinewhichnondeterministicallyguessestheassignments

deterministically.Theofandafterwardsevaluatestheformulae

resultingbitstringisusedtodeterminethecorrectleafsymbolaccordingto.u

Combiningourleaflanguagecharacterizationsoftheclassesofthepolynomialtimehierarchy(Theorem5.3)withthejustgivenlemma,wenextexamineleaflan-guagesdefinedbyfirst-orderformulaestartingwithaLindstr¨omquantifier.Lemma5.7.Let

bealanguage.Thenwehaveforall

LeafP

:

pad

Proof.WecombineLemma5.1andLemma5.6toprovethislemma.

Letandbea-formula.AsintheproofofTheorem5.1wemaythinkofuniverseshavingcomputationpathsofaselements.Wetransformthefirst-orderformulaeintosecond-orderformulaeasintheproofof

intosecond-Theorem5.1.Inadditionwetransformthefreevariablesordervariables.ButasintheproofofLemma5.6,wehavetoconsiderthoseassignmentsforanythatdonotrepresentacomputationpathof.Tomapthoseassignmentstopaddingsymbolsweaddaformula

.Weget:Path

Leaf

Mod

pad

u

Lemma5.8.Let

bealanguage.Thenwehaveforall

:

LeafP

pad

Proof.ToprovethislemmawecombineLemma5.2andLemma5.6.andformulaForgiven

weconstructapolynomial-timeTuringmachine

suchthatpad

anda

-formula

16

Mod

Leaf

pad

.

guessesnondeterminis-SimilartotheproofofLemma5.6theTuringmachine

ticallyallassignmentsof.Thereisasubstringoftheresultingleafstringforeachassignmentof.Thesesubstringsstart(ontheleft)withandend(ontheright)with.Thefirst-orderLindstr¨omquantifierboundstwovariablesleftandright.Theassignmentsshouldpointtotheleftandrightboundaryofsomesubstring.WemapassignmentsthatdonotcorrectlyrepresentsuchboundariestopaddingsymbolsasintheproofofLemma5.6.Letbeaformulathatistrueiffthesymbolatpositionintheleafstringis.Wethenconstructtheformulaleftrightthatindicatesthepaddingsymbolasfollows:

left

right

left

right

left

right

WeconstructforeachformulaLemma5.1.simulatesmachinesinFig.2.

aTuringmachineasintheproofof

foreachassignmentofasshown

Figure2:Constructionofmachine

Wenowconstructforeachformulaafirstorderformula.AsintheproofofLemma5.1weconstructinductivelyasequenceofformulae

.Theconstructionofisidenticaltotheconstruc-tiongivenintheproofof5.1.Letbetherightmostvariableinthefirstquantifier

17

blockofandRegionbetheformulagivenintheproofofTheorem5.1thatistrue,ifftheassignmentofandpointtotheleftandrightboundaryofthesubstringthatcorrespondstosomeassignmentof.Bythedefinitionofwemayassumew.l.o.g.thatisexistentiallyquantified.Letleftrightbetheformula

left

right

left

left

right

right

Region

left

right

left

right

u

ThemaintheoremnowfollowsfromLemma5.7andLemma5.8.Theorem5.9.Lethaveforall:

beaclassoflanguagesthatisclosedunderpadding.Thenwe

LeafP

Wecannotgetasimilarresultfor-formulaesincetheformulaleftrightusedintheproofofLemma5.8isexistentiallyquantified.Howeverthefollowingcertainlyholds:

Corollary5.10.Lethaveforall:

beaclassoflanguagesthatisclosedunderpadding.Thenwe

LeafP

Finally,wewanttoaddressaninterestingspecialcase:theclassPSPACE.Bar-rington,Immerman,andStraubingshowedthatfirst-orderlogicwithgroupquantifiersdefinesexactlytheregularlanguages[BIS90].Hertrampfetal.[HLS93]whocharac-terizedPSPACEbyregularleaflanguagesshowedthatinfactforthischaracterizationalreadyonesingleregularlanguage,thewordproblemforthegroup,issufficient.Thus,thisleaflanguagecharacterizationyieldsthefollowing:

Corollary5.11.PSPACE

6Discussion

Aswehaveseen,Lindstr¨omquantifierswhichareawellstudiedlogicalconcepthaveacomplexitytheoreticcounterpart:thesocalledleaflanguagedefinability,whichhasbeenstudiedintensivelyintherecentpast.

18

Second-orderLindstr¨omquantifiersdefine(inamodeltheoreticsense)exactlythoselanguagescharacterizablebyleaflanguagesforpolynomialtimemachines.IfisaLindstr¨omquantifier,thenthelogicdefinesthecomplexityclassLeafP.Thusitmaybepossiblethatresultsaboutleaflanguagescontributetothestudyoftheexpressivepowerofsecond-orderLindstr¨omquantifiersonorderedfinitestructures,andviceversa.

Ofcourse,itwouldbenicetohavealeaflanguageanalogueforfirst-orderLind-str¨omquantifiers.Tobeabletodothisonewillhavetoconsider“leaflanguagesforFO”insteadofleaflanguagesforpolynomialtime.Tobemoreprecise,whatisanappropriaterestrictionofthecomputationmodelproducingleafwords?

Gottlobin[Got95]showedthatundersomeparticularassumptionsto,first-orderformulaewitharbitrarilynestedandexistentialanduniversalquantifiersyieldsuper-classesofL(thelogspacedecidablesets).Certainly,formulaewithnestedquan-tifiersshouldbeexaminedalongthelinesofthispaper,i.e.whenusedtodefineleaflanguages.

Furthermore,oneshouldconsiderleaflanguagesdefinedbysecond-orderformu-laetoinvestigatethestructureofcomplexityclassesaboveexponentialtime,e.g.theexponentialtimehierarchy.

Acknowledgment.WethankDavidBarringtonforsuggestingtheexaminationoflogicallydefinedleaflanguagesasaninterestingtopic.Thanksarealsoduetotheanonymousrefereesforhelpfulcomments.

References

[All96]

E.ALLENDER.Anoteonuniformcircuitlowerboundsforthecountinghierarchy.InProceedings2ndInternationalComputingandCombina-toricsConference(COCOON),volume1090ofSpringerLectureNotesinComputerScience,pages127–135,1996.

D.P.BOVET,P.CRESCENZI,ANDR.SILVESTRI.Auniformapproachtodefinecomplexityclasses.TheoreticalComputerScience,104:263–283,1992.

D.A.M.BARRINGTON,N.IMMERMAN,ANDH.STRAUBING.Onuni-formitywithinNC.JournalofComputerandSystemSciences,41:274–306,1990.

[BCS92]

[BIS90]

19

[BL96]

B.BORCHERTANDA.LOZANO.Succinctcircuitrepresentationsandleaflanguageclassesarebasicallythesameconcept.InformationPro-cessingLetters,58:211–215,2996.

H.J.BURTSCHICK.Berechnungs-undBeschreibungskomplexit¨atvonZ¨ahlfunktionenundLindstr¨omquantoren.PhDthesis,FachbereichInfor-matik,TU-Berlin,1996.

[Bur96]

´RIEN,ANDH.VOLLMER.Non-[CMTV96]H.CAUSSINUS,P.MCKENZIE,D.THE

[Ebb85]

[EF95][Fag74]

[Got95]

[Her95]

[HLS93][HVW95]

[HVW96]

deterministicNCcomputation.InProceedings11thComputationalComplexity,pages12–21,1996.ToappearinJournalofComputerandSystemSciences.

H.-D.EBBINGHAUS.Extendedlogics:Thegeneralframework.InJ.Bar-wiseandS.Feferman,editors,Model-TheoreticLogics,PerspectivesinMathematicalLogic,chapterII,pages25–76.SpringerVerlag,1985.H.D.EBBINGHAUSANDJ.FLUM.FiniteModelTheory.Springer,1995.R.FAGIN.Generalizedfirst-orderspectraandpolynomialtimerecogniz-ablesets.InR.Karp,editor,ComplexityofComputations,pages43–73,1974.

G.GOTTLOB.Relativizedlogspaceandgeneralizedquantifiersoverfi-nitestructures.TechnicalReportCD-TR-95/76,InstitutforInformationSystems,ViennaUniversityofTechnology,1995.Anextendedabstractappearedintheproceedingsofthe10thSymposiumonLogicinCom-puterScience,1995.

U.HERTRAMPF.Regularleaf-languagesand(non-)regulartreeshapes.TechnicalReportA-95-21,Institutf¨urMathematikundInformatik,Medi-zinischeUniversit¨atzuL¨ubeck,1995.

U.HERTRAMPF,C.LAUTEMANN,T.SCHWENTICK,H.VOLLMER,ANDK.W.WAGNER.Onthepowerofpolynomialtimebit-reductions.InProceedings8thStructureinComplexityTheory,pages200–207,1993.

U.HERTRAMPF,H.VOLLMER,ANDK.W.WAGNER.Onthepowerofnumber-theoreticoperationswithrespecttocounting.InProceedings10thStructureinComplexityTheory,pages299–314,1995.

U.HERTRAMPF,H.VOLLMER,ANDK.W.WAGNER.Onbal-ancedvs.unbalancedcomputationtrees.MathematicalSystemsTheory,29:411–421,1996.

20

[JMT94]

´RIEN.LogspaceandlogtimeB.JENNER,P.MCKENZIE,ANDD.THE

leaflanguages.InProceedings9thStructureinComplexityTheory,pages242–254,1994.

¨.Firstorderpredicatelogicwithgeneralizedquantifiers.P.LINDSTROM

Theoria,32:186–195,1966.

C.H.PAPADIMITRIOU.ComputationalComplexity.Addison-Wesley,1994.

L.STOCKMEYER.Thepolynomial-timehierarchy.TheoreticalComputerScience,3:1–22,1973.

H.STRAUBING.FiniteAutomata,FormalLogic,andCircuitComplexity.Birkh¨auser,1994.H.VEITH.Succinctrepresentation,leaflanguages,andprojectionreduc-tions.InProceedings10thStructureinComplexityTheory,pages118–126,1996.

N.K.VERESHCHAGIN.Relativizableandnon-relativizabletheoremsinthepolynomialtheoryofalgorithms.IzvestijaRossijskojAkademiiNauk,57:51–90,1993.InRussian.

[Lin66]

[Pap94]

[Sto73]

[Str94]

[Vei96]

[Ver93]

21

因篇幅问题不能全部显示,请点此查看更多更全内容