Automata-based Axiom Pinpointing
FranzBaader RafaelPe~naloza
thedateofreeiptandaeptaneshouldbeinsertedlater
Abstrat Axiompinpointinghasbeenintrodued indesription logis(DL)tohelp
theuserunderstandthereasonswhyonsequenesholdbyomputingminimalsubsets
of the knowledge base that have the onsequene in question (MinA). Most of the
pinpointing algorithms desribed inthe DL literature are obtained as extensions of
tableau-based reasoning algorithmsfor omputing onsequenes fromDL knowledge
bases. In this paper, we show that automata-based algorithmsfor reasoning inDLs
andotherlogisanalsobeextendedtopinpointingalgorithms.Theideaisthatthe
treeautomatononstrutedbytheautomata-basedapproahanbetransformedinto
a weighted tree automaton whose so-alled behaviour yields a pinpointing formula,
i.e.,amonotoneBooleanformulawhoseminimalvaluationsorrespondtotheMinAs.
We also develop an approah for omputing the behaviour of a givenweighted tree
automaton. Weuse theDL SI as wellas LinearTemporalLogi (LTL)to illustrate
ournewpinpointingapproah.
1Introdution
Desriptionlogis (DLs)[2℄ are afamilyof logi-basedknowledgerepresentationfor-
malisms,whihareemployedinvariousappliationdomains,suhasnaturallanguage
proessing, onguration, databases, and bio-medial ontologies, buttheir most no-
tablesuess so far isthe adoptionof theDL-basedlanguageOWL[21℄ asstandard
ontology language for the semanti web. As the size of DL-based ontologies grows,
toolsthat supportimprovingthe quality ofsuhontologiesbeome moreimportant.
DLreasoners[20,19,38℄anbeusedtodetetinonsisteniesandtoinferotherimpliit
onsequenes,suhassubsumptionrelationshipsbetweenoneptsorinstanerelation-
shipsbetweenindividualsandonepts.However,foradeveloperoruserofaDL-based
ontology,itisoftenquitehardtounderstandwhyaertainonsequeneomputedby
the reasoner atually follows from the knowledgebase. For example, in the urrent
FirstauthorpartiallysupportedbyNICTA,CanberraResearhLab.,andseondauthorfunded
bytheGermanResearhFoundation(DFG)undergrantGRK446.
TheoretialComputerSiene,TUDresden,Germany
DLversionofthemedialontologySNOMEDCT, 1
theoneptAmputation-of-Finger
is lassied as asubonept of Amputation-of-Arm. Finding thesix axiomsthat are
responsible forthis error[10℄ amongthemore than350,000 terminologialaxiomsof
SNOMEDwithoutsupportbyanautomatedreasoningtoolisnoteasy.
Axiom pinpointing [34℄ has been introdued to help developers or users of DL-
basedontologiesunderstandthereasonswhyaertainonsequeneholdsbyomputing
minimalsubsetsoftheknowledgebasethathavetheonsequeneinquestion(MinA).
There aretwogeneralapproahes for omputingMinAs:theblak-box approahand
the glass-box approah. Themost nave variant of the blak-box approah onsiders
all subsets of the ontology, and omputes for eah of them whether it still has the
onsequene ornot.More sophistiated versions[35,22℄ useavariant ofReiter's [32℄
hittingsettreealgorithmtoomputeallMinAs.Insteadofapplyingsuhablak-box
approahtoalargeontology,oneanalsorsttrytondasmallandeasytoompute
subsetoftheontologythatontainsallMinAs,andthenapplytheblak-boxapproah
to thissubset[10℄. Themainadvantageof theblak-boxapproah isthat itanuse
existing highly-optimized DLreasoners unhanged.However, itmay be neessaryto
allthereasoneranexponentialnumberoftimes.Inontrast,theglass-box approah
triestondall MinAsbyasinglerunofamodiedreasoner.
Mostoftheglass-boxpinpointingalgorithmsdesribedintheDLliterature (e.g.,
[4,34,33,27,25℄) areobtained as extensionsoftableau-based reasoningalgorithms[9℄
foromputingonsequenesfromDLknowledgebases.Thepinpointingalgorithmsand
proofs oftheir orretnessin thesepapersare given for a spei DL anda spei
typeofknowledgebaseonly,and itisnotlear towhihoftheknowntableau-based
algorithmsforDLstheapproahesreallygeneralize.Forexample,thepinpointingex-
tension desribed in [25℄, whih an deal with general onept inlusions (GCIs) in
the DL ALC, follows the approah introdued in[4℄, butsine GCIs require thein-
trodutionofso-alledblokingonditionsintothetableau-basedalgorithmtoensure
termination[9℄,therearesomenewnon-trivialproblemstobesolved.
Tooverometheproblemofhavingtodesignanewpinpointingextensionforevery
tableau-basedalgorithm, wehaveintrodued in[5℄ ageneralapproahfor extending
tableau-basedalgorithmstopinpointingalgorithms.Thisapproahhas,however,some
annoyinglimitations.First,itonlyappliestotableau-basedalgorithmsthatterminate
withoutrequiringanyyle-hekingmehanismsuhasbloking.Seond,termination
ofthetableau-basedalgorithmonestartswithdoesnotneessarilytransfertoitspin-
pointingextension. Thoughtheseproblemsan, inpriniple,besolvedbyrestriting
thegeneralframeworktoso-alledforesttableaux[8,7℄,thissolutionmakesthedeni-
tionsandproofsquiteompliatedandlessintuitive.Also,theapproahanstillonly
handlethe mostsimpleversion ofbloking,usuallyalledsubsetblokingintheDL
literature.
Inthepresentpaper,weproposeadierentgeneralapproahforobtainingglass-box
pinpointingalgorithms,whihalsoappliestoDLsforwhihtheterminationoftableau-
basedalgorithmsrequirestheuseofappropriateblokingonditions.Itiswell-known
thatautomataworkingoninnitetreesanoftenbeusedtoonstrutworst-aseopti-
maldeisionproeduresforsuhDLs[13,26,11,14,3℄.Inthisautomata-basedapproah,
theinputinferene problem is translatedintoatreeautomatonA ,whihis then
testedfor emptiness.Basially, ourapproahtransformsthetreeautomatonA into
aweightedtreeautomatonworkingoninnitetrees,whoseso-alledbehaviouryields
1
apinpointingformula, i.e., amonotone Boolean formulathat enodes all the MinAs
of .Toobtain anatualpinpointingalgorithm,we hadtodevelop analgorithm for
omputingthebehaviourofweightedtreeautomataworkingoninnitetrees.Whenwe
startedourwork,weouldnotndsuhanalgorithminthequiteextensiveliterature
on weighted automata. In fat, although weighted automata working onnite trees
[37℄ and weighted automataworking oninnite words [16℄ have beenonsidered for
quiteawhile,theresearhonweightedautomataworkingoninnitetrees hasstarted
onlyreently [23,15℄. During thedevelopment ofour work, analternative algorithm
for omputingthebehaviour ofweighted treeautomataworking oninnite trees has
independentlybeendevelopedin[15℄.Itturnsout,however,thatusingthisalgorithm
inour pinpointingappliation basiallyyields a blak-box approah for pinpointing,
ratherthanaglass-boxapproah,asouralgorithmdoes(seeSetion5.4).
We will use the DL SI, whih extends the basi DL ALC [36℄ with transitive
and inverse roles, as well as Linear Temporal Logi (LTL) [28,17℄ to illustrate our
newpinpointingapproah.Theuseof SI is,ontheonehand,motivatedby thefat
thatthepreseneofinversesinSIrequirestableau-basedalgorithmstouseabloking
onditionthatismoresophistiatedthansubsetbloking[9℄.Consequently,ourgeneral
resultsontableau-basedapproahforpinpointing[8,7℄donotapplytothisDL.Onthe
otherhand,theextensionoftheirapproahtoSIismentionedasanopenproblemin
[25℄.TheautomatausedtodeidesatisabilityinSIareso-alledloopingautomata,
whih do notuse anaeptane ondition. Our hoie of LTL as a seond example
is, onthe onehand,motivatedby the fat that automata-based algorithmsfor LTL
requiretheuseofautomatawithaBuhiaeptaneondition.
2
Onetheotherhand,
we believe that pinpointingan also be a useful inferene servie in appliations of
LTL.InLTLmodelheking[12℄, itdoesnotmakesenseto hek whetherasystem
desriptionsatisesagivenLTLformulaifthisformulaoritsnegationisunsatisable.
Pinpointing ould help the userto nd the reasons for the unsatisability and thus
orrettheformula.InLTLsynthesis[29,24℄onetriestogenerateareativenite-state
systemfrom aformalspeiation, whihisgivenas anLTLformula. Iftheformula
is unsatisable, then the speiation is obviously faulty, and needs to be repaired.
Pinpointingouldbeusedtosupporttherepair proess bylarifying thereasonsfor
unsatisability.
Inthenextsetion,werstintroduetheDLSIandthetemporallogiLTL,and
thenrealltherelevantdenitionsregardingpinpointing.Setion3denesgeneralized
Buhitree automata, their restritions to Buhitree automataand looping treeau-
tomata, andtheir generalization to the weighted ase. InSetion 4,we rstpresent
ourgeneralapproahforautomata-basedpinpointing,whihisbasedonthenotionof
anaxiomatiautomaton anditstransformation intoapinpointingautomaton.Then,
weshowthat thisapproahanbeappliedto SIand LTLby introduingaxiomati
automatafortheselogis.Thepinpointingautomatonisaweightedautomatonwhose
behaviour is the pinpointingformula. Thus, to apply our approah inpratie, one
needs to be able to ompute the behaviour of weighted generalized Buhi tree au-
tomata.InSetion5,we rstshowhowto omputethebehaviourofweightedBuhi
treeautomata.Seond,weexplainhowthisomputationanbesimpliedforthease
of weighted looping tree automata. For the DL SI, the pinpointingautomaton on-
2
Weould,ofourse,alsohaveusedaDLwithtransitivelosureofroles[1℄forthispurpose.
However,suhDLsareuntilnownotusedinappliations,andwealsowantedtomakelear
strutedbyourapproahissuhaweightedloopingtreeautomaton.Third,wedene
abehaviour-preservingpolynomial-timeredutionof weighted generalized Buhi tree
automata to weighted Buhi tree automata, whih yields an approah for omput-
ingthebehaviourofweightedgeneralizedBuhitreeautomata.Forthetemporallogi
LTL,thepinpointingautomatononstrutedbyourapproahisaweightedgeneralized
Buhitreeautomaton.Fourth,weompareourapproahforomputingthebehaviour
ofweightedBuhitreeautomatawiththeonedevelopedin[15℄.Setion6summarizes
theresultsofthepaperandgivessomeperspetivesonfurtherresearh.
Thisworkextendsthe resultsin[6℄ (theonfereneversion ofthis paper),whih
apply toloopingautomataonly,tothe aseof automatawithBuhiaeptaneon-
ditions.
2Preliminaries
Inthis setion, werst introduethe DL SI and the temporallogi LTL, and then
realltherelevantdenitionsregardingpinpointingfrom[5℄.
2.1 TheDesriptionLogiSI
Asmentionedabove,SIextendsthebasiDLALC withtransitiveandinverseroles.
Anexampleofarolethat shouldbeinterpretedas transitive ishas-desendant,while
has-anestorshouldbeinterpretedastheinverseofhas-desendant.Insteadofemploying
theusual approahof\hard-oding" inverseand transitiverolesinto thesyntaxand
semantisofoneptdesriptions,weallowtheuseofinverseandtransitivityaxiomsin
theknowledgebase.Thisenablesustopinpoint alsothesekindsofaxiomsasreasons
forertain onsequenes.Thus,theoneptdesriptions thatweonsiderinthisase
aresimplyALC oneptdesriptions.
Denition1 (ALC onept desriptions) LetN
C
beasetofoneptnamesand
N
R
asetofrolenames. Theset of ALC oneptdesriptions isthe smallestsetsuh
that
{ alloneptnamesareALConeptdesriptions;
{ ifC andD areALConeptdesriptions,thensoare:C,CtD,andCuD;
{ if C is anALC onept desription and r 2 N
R
, then 9r:C and 8r:C are ALC
oneptdesriptions.
Aninterpretation isapairI=( I
; I
)wherethedomain I
isanon-emptysetand
I
isafuntionthatassignstoeveryoneptnameAasetA I
I
andtoeveryrole
name r abinaryrelation r I
I
I
. Thisfuntionisextended toALC onept
desriptionsasfollows:
{ (CuD) I
=C I
\D I
; (CtD) I
=C I
[D I
; (:C) I
=
I
nC I
;
{ (9r:C) I
=fx2 I
jthereisay2 I
with(x;y)2r I
andy2C I
g;
{ (8r:C) I
=fx2 I
jforally2 I
,(x;y)2r I
impliesy2C I
g.
Inthispaperwerestritourattentiontoterminologialknowledge,whihisgiven
Denition2 (SI TBoxes) AnSI TBox isa niteset of axiomsof the following
form:
(i) CvDwhereCandD areALConeptdesriptions(GCI);
(ii) trans(r)wherer2N
R
(transitivityaxiom);
(iii) inv(r;s),wherer6=s2N
R
(inverseaxiom),
suhthateveryr2N
R
appearsinatmostoneinverseaxiom.
AninterpretationIisalled amodeloftheSITBoxT ifitsatisesallaxiomsin
T,i.e., if
(i) CvD2T impliesC I
D I
;
(ii) trans(r)2T impliesthatr I
istransitive;
(iii) inv(r;s)2T impliesthat(x;y)2r I
i(y;x)2s I
.
The main inferene problemsfor terminologial knowledge are satisability and
subsumption
Denition3 (satisability,subsumption)LetCandDbeALConeptdesrip-
tions andT anSITBox.We saythat C issatisable w.r.t. T if thereisamodelI
of T suhthat C I
6=;.Inthis ase, I is alsoalled amodel of C w.r.t. T.We all
Cunsatisable w.r.t.T ifitdoesnothaveamodelw.r.t.T.Finally,wesaythatC is
subsumedbyDw.r.t.T ifC I
D I
holdsineverymodelIofT.
Wewanttopinpointreasonsforunsatisability andforsubsumption.SineC issub-
sumedby D w.r.t.T i Cu:Disunsatisablew.r.t. T,itis obviously suÆientto
designapinpointingalgorithmforunsatisability.
Theautomata-basedapproahfordeiding(un)satisabilityusesthefat thatan
ALConeptdesriptionC issatisablew.r.t.anSITBoxT iithasaertaintree-
shapedmodel,alledHintikkatreeforCandT.Itonstrutsaloopingtreeautomaton
workingoninnitetrees whoseruns are exatlytheHintikkatrees for C andT (see
[3℄andSetion4.2),andthenteststhisautomatonforemptiness.
2.2 LinearTemporalLogi
LinearTemporalLogi(LTL)isanextensionofpropositionallogithatallowsreason-
ingabouttemporalproperties,wheretimeisseenasdisreteandlinear.Thesemantis
ofthislogi usethenotionofaomputation,whihintuitivelyorrespondtointerpre-
tationswhosedomainisxedtobethesetofnaturalnumbers.
Denition4 (LTL formulae) LetPbeasetofpropositional variables.Thesetof
LTLformulaeisthesmallestsetsuhthat
{ allpropositionalvariablesareLTLformulae,
{ ifand areLTLformulae,thensoare:;^ ;,andU .
A omputation is a funtion : N !
P
(P), where N represents the set of natural numbers.ThisfuntionisextendedtoLTLformulae asfollows, foreveryi2N :{ :2(i)i2=(i); ^ 2(i)if; g(i);
{ U 2(i)ithereisajisuhthat 2(j)andforallk;ik<j,itholds
that2(k).
TheLTLformulaissatisableifthereisaomputationsuhthat2(0).
OneisusuallyinterestedindeidingwhetheragivenLTLformulaissatisableor
not.Here,wewilllookatthesatisabilityprobleminamorene-grainedmanner.We
areinterestedindetetingwhihpartsoftheformulaatuallyausetheunsatisability.
More preisely, we will assume that our formula is a onjuntion of LTLformulae,
and we want to ndout whih onjuntsare responsible for the unsatisability.We
additionally allow some of these onjunts to be trusted inthe sense that theywill
neverbeonsideredastheausesforunsatisability.Thus,weonsiderLTLformulae
thatare onjuntionsofastati formula, whihmustalwaysbethere,and asetof
refutableformulaeR,whihanberemoved.
Denition5 (axiomatisatisability)LetbeanLTLformulaandRaniteset
ofLTLformulae. We say that is a-satisablew.r.t. Rif ^
V
2R
issatisable,
i.e., there is a omputation suhthat R[fg (0). Inthis ase, is alled a
omputationfor(;R).
Wewill showinSetion4.3 howoneanonstrutaBuhitreeautomaton that has
as its suessful runs all omputations for the input, thus allowing us to redue a-
satisabilitytotheemptinessproblemforBuhitreeautomata.
2.3 BasiDenitionsforPinpointing
Following[5℄,wedenepinpointingnotforaspeilogi andinfereneproblem,but
ratherinamoregeneralsetting. Thetypeofinfereneproblemsthatwewillonsider
isdeidinga so-alled-propertyfor agivensetofaxiomatizedinputs.Toobtain an
intuitive understanding ofthe following denition, justassume that inputsare ALC
oneptdesriptions, admissiblesetsof axiomsareSITBoxes,andthe -propertyis
unsatisablility.
Denition6 (axiomatizedinput, -property)LetIandTbesetsofinputsand
axioms,respetively,andlet
P
admis
(T)
P
n
(T)beasetofnitesubsetsofTsuh
thatT 2
P
admis
(T)impliesT 0
2
P
admis
(T)forallT 0
T.Anaxiomatizedinputfor
Iand
P
admis
(T)isoftheform(I;T)whereI2IandT 2
P
admis (T).
A onsequene property(or-property for short)isasetPI
P
admis
(T)suh
that(I;T)2P implies(I;T 0
)2P foreveryT 0
2
P
admis
(T)withT 0
T.
Thereason why we have introdued the set
P
admis
(T) of admissible subsets of
T (rather than taking all nite subsets of T) is to allow us to impose additional
restritionsonthe sets ofaxiomsthat mustbe onsidered.For instane,SITBoxes
arenotarbitrarynitesetsofaxiomsoftheform(i), (ii),and(iii)(seeDenition2).
Inaddition, we requirethat every role name appearsin at most oneinverse axiom.
Clearly,thisrestritionsatisesourrequirementforadmissiblesetsofaxioms.
TheproblemsofunsatisabilityofALConeptdesriptionsw.r.t.SITBoxesand
onsistofallALConeptdesriptions,TofallGCIs,transitivityaxioms,andinverse
axioms,and
P
admis
(T)ofallSITBoxes.Thefollowingisa-property:
P=f(C ;T)jCisunsatisablew.r.t.Tg:
Likewise,ifIandTbothonsistofallLTLformulaeand
P
admis
(T)=
P
n
(T),then
P=f(;R)jisa-unsatisablew.r.t.Rg
isa-property.
Denition7 Given anaxiomatized input =(I;T) and a-property P, a setof
axioms S T is alled a minimal axiom set (MinA) for w.r.t. P if (I;S) 2 P
and (I;S 0
)2=P foreveryS 0
S.Thesetof allMinAs for w.r.t. Pis denotedby
MIN
P( ) .
NotethatthenotionofaMinAisonlyinterestingif 2P;otherwise,themonotoniity
requirementforP entailsthatMIN
P( )
=;.Letusinstantiatethisdenitionforthe
two-propertieswehaveintroduedabove.
Inour SI example, onsider the axiomatized input =(Au8r:C ;T) where T
onsistsof
ax
1
: Av9r:B; ax
2
: Bv8s::A; ax
3
: Cv:B; ax
4
: inv (r;s) (1)
It is easy to see that 2 P, and that the set of all MinAs for is MIN
P( )
=
ffax
1
;ax
2
;ax
4 g;fax
1
;ax
3 gg.
ForthelogiLTL,onsidertheaxiomatizedinput =(q;R)whereRisgivenby
ax
1
: pU:q; ax
2
: :p; ax
3
: q; ax
4
: :(q^p): (2)
ThesetofallMinAsfor isthenMIN
P( )
=ffax
1
;ax
2
;ax
3 g;fax
1
;ax
3
;ax
4
gg.Thus,
in the LTL formula q^pU:q ^:p^q^:(q^p), the MinAs tell us whih
minimal ombinations ofthe last four onjuntsare responsible for unsatiability in
thepreseneofq.
Onemight think that pinpointing (i.e., the omputation of MinAs) an onlybe
appliedintheLTLsettingiftheformulaoneisinterestedinisalargeonjuntionof
small formulae. Atrstsight, itisnotlearhowasubformula thatdoesnotour
as a top-level onjunt ould be pinpointed as a ulprit for unsatisability.This is,
however,possiblebyreplaingsuhasubformula byanewpropositionalvariablep
andaddingthe\denition"(p , )asatop-levelonjunttotheformulaobtained
thisway.
3
InsteadofomputingallMinAs,oneanalsoomputeapinpointingformula.To
denethisformula,weassumethateveryaxiomt2Tislabelledwithauniquepropo-
sitionalvariable,lab(t).Letlab(T)bethesetofallpropositionalvariableslabellingan
axiominT.AmonotoneBooleanformulaoverlab(T)isaBooleanformulausingvari-
ablesinlab(T)andonlytheonnetivesonjuntionanddisjuntion.Inaddition,the
onstants>and?,whihalwaysevaluatetotrueandfalse,respetively,aremonotone
Boolean formulae. Weidentifyapropositionalvaluation withthe setofpropositional
variablesthatitmakestrue.ForavaluationVlab(T),letT
V
=ft2T jlab(t)2Vg.
ReallthatifT 2
P
admis
(T)thenforeveryT 0
T itholdsthatT 0
2
P
admis (T).In
partiularthismeansthatT
V 2
P
admis
(T)foreveryvaluationV.
3
Here,isanabbreviationfor:(>U:)and
1 ,
2
isanabbreviationfor:(
1
^:
2 )^
:(: ^ ).
Denition8 (pinpointing formula)Givena-propertyPandanaxiomatizedin-
put =(I;T), themonotone Boolean formula overlab(T) isalled apinpointing
formula for w.r.t.Pifthefollowingholdsfor everyvaluationV lab(T):
(I;T
V
)2P i V satises:
InourSI example,weantakelab(T)=fax
1
;:::;ax
4
gassetofpropositional vari-
ables. It is easy to see that ax
1
^((ax
2
^ax
4 )_ax
3
) is a pinpointing formula. In
the LTLexample, we antake the sameset of propositional variables.In this ase,
ax
1
^ax
3
^(ax
2 _ax
4
)isapinpointingformula.
Valuationsanbe orderedbysetinlusion.Thefollowing is animmediate onse-
queneofthedenitionofapinpointingformula [4℄:ifapinpointingformula for
w.r.t.P,then
MIN
P( )
=fT
V
jV isaminimalvaluationsatisfyingg:
Thisshows thatitis enoughtodesignanalgorithmfor omputingapinpointingfor-
mula toobtain allMinAs. However,the redutionsuggestedby theabove identityis
not polynomial. Onepossible way to obtain MIN
P( )
from is to rst transform
intodisjuntivenormalform,andthenremovesuperuousdisjunts.Itiswell-known
that this anause anexponentialblow-up. Thisshould, however, notbe viewedas
adisadvantageofapproahesomputingthepinpointingformula ratherthandiretly
MIN
P( )
.If suha blow-uphappens, thenthe pinpointing formulaatually yields a
ompatrepresentationofall MinAs.
3BuhiTree Automata
In this setion, we introdue both unweighted and weighted generalized Buhi tree
automata. These automata reeive innite trees of a xed arity k as inputs. For a
positive integerk,wedenotethe setf1;:::;kgbyK.Thenodes ofourtrees anbe
identiedby words inK
inthe usualway: therootnodeis identiedbythe empty
word",and thei-th suessorofthenodeu isidentiedby uifor1ik.Inthe
ase oflabelledtrees, we will refer tothe labelling ofthe nodeu2K
inthetree r
by r(u). Wewillalsouse
!
r(u)todenotethetuple
!
r(u)=(r(u);r(u1);:::;r(uk)).An
innitetreerwithlabelsfromasetQanberepresentedasamappingr:K
!Q.
For our purpose, it is suÆient to use unlabelled innite trees as inputsfor our
treeautomata.Foraxedarityk,thereisexatlyonesuhtree,whihweanidentify
withthesetofitsnodes,i.e.,withK
.Wewill alsousetheoneptof apathinthis
tree.A pathis asubsetpK
suhthat "2pandfor everyu2pthereis exatly
onei;1ikwithui2p.
Denition9 (Buhitreeautomaton)AgeneralizedB uhitreeautomatonforarity
k is a tuple(Q;;I;F
1
;:::;Fn),where Q is anite setof states, Q k +1
is the
transition relation, I Q isthe set ofinitial states, and F
1
;:::;Fn Q are setsof
nal states. A generalized Buhitree automaton is alled B uhi automaton if it has
onlyonesetofnalstates;i.e.,ifn=1.Itisalledloopingtree automatonifn=0.
Arun of a generalizedBuhiautomaton onthe unlabelled treeK
is alabelled
!
everypathpand everyi;1in,thereareinnitely manynodes u2p suhthat
r(u)2F
i .
TheemptinessproblemforgeneralizedBuhitreeautomataforaritykistheprob-
lemofdeidingwhether agivensuhautomaton hasasuessful runr withr(")2I
ornot.
Letus illustrate the notions introdued inthis denition on asimple Buhiau-
tomaton.
Example1 ConsidertheBuhitreeautomatonA ex
=(Q;;I;F)forarity2,where
{ Q=fq
0
;q
1
;q
2
;q
3
g,I=fq
0
g,andF =fq
1
;q
3 g:
{ =f(q
0
;q
1
;q
1 );(q
0
;q
2
;q
2 );(q
1
;q
1
;q
1 );(q
2
;q
2
;q
2 );(q
2
;q
3
;q
3 )g.
Thisautomaton has tworuns that labelthe root withthe initial stateq
0 :r
1 , whih
labelsallthenon-rootnodeswithq
1 ,andr
2
,whihlabelsallthenon-rootnodeswith
q
2
;thelatterisnotsuessful,buttheformeris.Thus,A ex
hasr
1
asasuessfulrun
thatlabelstheroot withaninitialstate.Thebinarytreer
3
thatlabelstheroot with
q
0
andall thenon-rootnodeswithq
3
isnot arunofA ex
.Finally,therunr
4 ,whih
labelsallnodeswithq
1
,isasuessfulrunofA ex
,butitdoesnotlabeltherootwith
aninitialstate.
Althoughadiretalgorithmfordeidingthe emptinessproblemfor ageneralized
Buhiautomatonisskethedin[40℄,inthejournalversionofthatpaper[41℄,theideais
simpliedbypresentingaredutiontotheemptinessproblemforBuhiautomata.Our
treatmentofweightedautomatawillfollowasimilarapproah.First,wewillshowhow
toomputethebehaviourofweightedBuhiautomatabyanapproahthatisinspired
bytheemptinesstest forBuhiautomata.
4
Then,wewillintroduearedutionfrom
weightedgeneralizedBuhiautomatatoweightedBuhiautomatathat preservesthe
behaviour.
Wewilllaterextendautomata-baseddeisionproeduresintoalgorithmsthatom-
putepinpointingformulae by transforming Buhiautomata intoweighted Buhiau-
tomata.Theweightsofsuhautomataomefromadistributivelattie [18℄.
Denition10 (distributive lattie) Adistributivelattieisapartiallyorderedset
(S;
S
)suhthatinmaandsupremaofarbitrarynitesubsetsofS alwaysexistand
distributeovereahother.Thedistributivelattie(S;
S
)isalledniteifitsarrier
setSisnite.
Any weighted automaton uses as weights onlynitely many elements of the under-
lyingdistributive lattie. Sine nitely generated distributive latties are nite [18℄,
the losure of this set under the lattie operations inmum and supremumyields a
nitedistributivelattie.Forthisreason,wewillinthefollowing assumewithoutloss
of generality that the weights of our weighted Buhi automaton ome from a nite
distributivelattie(S;
S ).
Inthe following, we will often simply use the arrier set S to denote the nite
distributive lattie (S;
S
). Theinmum(supremum)of asubsetT S will bede-
notedby
N
t2T t(
L
t2T
t).Wewill oftenomputetheinmum(supremum)
N
i2I t
i
(
L
i2I t
i
)overaninnitesetofindiesI.However,thenitenessofthelattieandthe
4
idempotenyoftheoperatorsinmumandsupremumensurethatthesetsoverwhih
the operators are atually appliedare nite,and hene inmumand supremum are
well-denedinthisase.Fortheinmum(supremum)oftwoelements,wewillalsouse
inxnotation,i.e.,writet
1 t
2 (t
1 t
2
)todenotetheinmum(supremum)oftheset
ft
1
;t
2
g.Theleastelement ofS (i.e.,theinmumofthewholesetS) willbedenoted
by0,andthegreatestelement(i.e.,thesupremumofthewholesetS)by1.
Itshouldbenotedthatourassumptionthattheweightsomefromanitedistribu-
tivelattieis strongerthantheoneusuallyenounteredintheliteratureonweighted
automata. In fat, for automata working onnite words or trees, it is suÆient to
assumethatthe weightsomefromaso-alledsemiring[37℄.Inordertohaveawell-
denedbehaviouralsoforweightedautomataworkingoninniteobjets,theexistene
of inniteproduts and sumsis required[16,31℄. Asmentionedabove, ourniteness
assumptionensuresthat suhinnite produts andsums areatually nite.Thead-
ditionalproperties imposedbyourrequirementto haveadistributivelattie (inpar-
tiular,distributivityandtheidempotenyofprodutandsum)areneessaryfor our
approah of omputing the behaviour of weighted Buhi automata (see Setion 5).
Thesestrongerassumptionsarenotproblematiinourpinpointingappliation:aswe
will see later, the weights we will enounter in our omputation of the pinpointing
formulaatuallyomefromanitelygeneratedfreedistributivelattie.
Denition11 (weightedBuhiautomaton)LetS beanitedistributivelattie.
A weighted generalized B uhi automaton (WGBA)overS for arity k is atuple A=
(Q;in;wt;F
1
;:::;F
n
) where Q is a nite set of states, in : Q ! S is the initial
distribution,wt:Q k +1
!Sassignsweightstotransitions,andF
1
;:::;F
n
Qarethe
setsofnalstates. AWGBAisalledweightedB uhiautomaton(WBA)ifn=1and
weightedlooping automaton(WLA)ifn=0.
A run of the WGBA A is a labelled tree r : K
! Q. The weight of this run
is wt(r) =
N
u2K
wt(
!
r(u)). This run is suessful if, for every path p and every
i;1 i n, thereare innitely manynodes u 2 p suhthat r(u)2 F
i
. Letsu
A
denotethesetofallsuessfulrunsofA.ThebehaviouroftheautomatonAis
kAk:=
M
r2su
A
in(r("))wt(r):
LetusillustratethisdenitionontheexampleofaWBAovertheBooleansemiring
thatsimulatesan(unweighted)Buhitreeautomaton.
Example2 TheBooleansemiringB =(f0;1g;^;_;1;0)isanitedistributivelattie,
wherethepartialorderisdenedas1
B
0.Notethatwehavedened1tobesmaller
than 0, and thus onjuntion yields the supremum (i.e., is the \addition" ) and
disjuntionyieldstheinmum(i.e.,isthe\multipliation").Likewise,1istheleast
element0,and0isthe greatestelement 1.Thereason forthisunorthodox denition
isthatthismakesiteasytotransformagivenBuhitreeautomatonA=(Q;;I;F)
into aWBA Aw onB suhthat thebehaviourof Aw is 0iA hasasuessful run
that labelsthe rootwith aninitialstate. InAw,theinitial distributionmapsinitial
statesto0andall otherstatesto1;atupleinQ k +1
getsweight 0ifitbelongsto,
andweight 1otherwise.
ConsidertheWBAA ex
w
thatisobtainedbyapplyingthisonstrutiontotheBuhi
treeautomatonA ex
ofExample1.Therunr
1
hasweight0sineallthetransitionsit
disjuntion.Sinethisrunissuessful,itontributesthesummandin(q
0 )wt(r
1 )=
0_0 = 0 to the behaviour of A ex
w
. Sine addition is onjuntion, this auses the
behaviour of A ex
w
to be 0.Let usnevertheless onsider someother runs.Therun r
2
also hasweight 0 andstarts withthe initial state q
0
.However, sinethis run is not
suessful,in(q
0
)wt(r
2
)is notusedasasummandwhenomputingthebehaviour
ofA ex
w
.Thetreer
3
isasuessfulrunofA ex
w
,butitisnotarunofA ex
.Sineituses
the transition (q
3
;q
3
;q
3
), whose weight is 1,its overall weight is 1 as well. Thus,it
ontributesthesummandin(q
0 )wt(r
3
)=0_1=1tothebehaviourofA ex
w
,butthis
summandis\eatenup"bythesummand0ontributedtothesum(i.e.,onjuntion)
bytherunr
1
.Finally,therunr
4
,isasuessfulrunofA ex
w
,whihhasweight0.Sine
q
1
isnotaninitialstateofA ex
,itontributesthesummandwt(q
1 )wt(r
4
)=1_0=1
tothebehaviourofA ex
w .
Bygeneralzing theobservationswehavemadefor therunsr
1
;r
2
;r
3
;r
4 ofA
ex
w ,it
iseasyseethatthefollowingholdsfor anyBuhitreeautomatonA:thebehaviourof
Awis0iAhasasuessfulrunthatlabelstheroot withaninitialstate.
InSetion5,wewilldevelopanapproahforomputingthebehaviourofweighted
(generalized)Buhitreeautomatathatgeneralizestheemptinesstestfor(generalized)
Buhitreeautomata.Butrst,weshowhowtoreduetheproblemofomputingthe
pinpointingformulatotheproblemofomputingthebehaviourofaWGBA.
4Automata-basedPinpointing
Inthissetion,werstintrodueourgeneralapproahforautomata-basedpinpointing,
andthenshowhowitanbeappliedtondingapinpointingformulaforunsatisability
inSIandLTL.
4.1 TheGeneralApproah
Basially,theautomata-basedapproahfordeidinga-propertyPtakesaxiomatized
inputs =(I;T)andtranslatesthemintoautomataA suhthat 2PiA does
not have a suessful run. For example, the automaton onstruted from a onept
desriptionCandaTBoxT hasasuessfulruniCissatisablew.r.t.T,wherethe
-propertyisunsatisability.Ifthetranslationfrom toA isanarbitraryfuntion,
then we have no way of knowing how the axioms in T inuene the behaviour of
theautomaton,andthusitisnotlear howtoonstrutaorrespondingpinpointing
automaton.Forthisreason,wewillassumethattheautomatonA for =(I;T)ina
ertainsensealsoontainsautomataforall axiomatizedinputs(I;T 0
)withT 0
T, 5
whihanbe obtainedbyappropriatelyrestriting the statesandtransitions ofA .
Tobemore preise, let A =(Q;;I;F
1
;:::;Fn) be a generalizedBuhiautomaton
foraritykand =(I;T)anaxiomatizedinput.Thefuntionsres:T !
P
(Q k +1)
and Ires:T !
P
(Q)are respetively alled atransition restriting funtionand an initialrestritingfuntion.TherestritingfuntionsresandIresanbeextendedtosetsofaxiomsT 0
T asfollows:
res (T 0
):=
\
t2T 0
res (t) and Ires (T 0
):=
\
t2T 0
Ires(t):
5
ForT 0
T,theT 0
-restritedsubautomatonofAw.r.t.resandIresisdenedas
A
jT
0 :=(Q;\res (T 0
);I\Ires(T 0
);F
1
;:::;F
n ):
Denition12 (axiomati automaton) Let A=(Q;;I;F
1
;:::;Fn) bea gener-
alized Buhi automaton for arity k, = (I;T) an axiomatized input, and res :
T !
P
(Q k +1) andIres:T !
P
(Q)atransitionand aninitial restritingfuntion, respetively.Thenweall(A;res ;Ires )anaxiomatiautomatonfor .Given a-property P, we say that (A;res ;Ires) is orret for w.r.t. P ifthe
followingholdsforeveryT 0
T:(I;T 0
)2PiA
jT
0 doesnothaveasuessfulrunr
withr(")2I\Ires(T 0
).
Givena orret axiomatiautomaton for =(I;T), we andeide(I;T 0
)2 P
forT 0
T byapplyingtheemptinesstestforgeneralizedBuhiautomatatoA
jT 0.
Example3 Let =(I;T)be anaxiomatizedinput, whereT =fax
1
;ax
2
;ax
3 g,and
assumethat,forallT 0
T,the-propertyPholdsfor(I;T 0
)ifax
1
;ax
2 g\T
0
6=;.
Thus,MIN
P( )
=ffax
1 g;fax
2
gg,andax
1 _ax
2
isapinpointingformula.
Considertheaxiomatiautomaton(A ex
;res ;Ires),where
{ A ex
istheBuhitreeautomatonintroduedinExample1;
{ thetransitionrestritingfuntionisdenedasres (ax
1
)=nf(q
1
;q
1
;q
1 )g,
res(ax
2
)=,andres (ax
3
)=nf(q
2
;q
2
;q
2 )g;
{ theinitialrestritingfuntionisdenedasIres(ax
1
)=I,Ires(ax
2 )=;,
andIres(ax
3 )=I.
It is easy to see that (A ex
;res ;Ires) is orret for w.r.t. P. In fat, reall that
the onlysuessful run of A ex
isr
1
, whih labels the root with q
0
andall non-root
nodeswithq
1
.Now,assumethatT 0
T.Ifax
1 2T
0
,thenthetransition(q
1
;q
1
;q
1 ),
whihisusedintherunr
1
,isnolongeravailable,andthusr
1
isnotarunofA
jT 0.If
ax
2 2T
0
,thenA
jT
0 doesnothaveaninitialstate,and thusr
1
nolongerstartswith
aninitialstate.Finally,havingax
3 inT
0
doesnotremovetherunr
1
sinethisaxiom
onlyremovesthetransition (q
2
;q
2
;q
2
), whihis notused inr
1
,and italso does not
hangethe set ofinitial states. Consequently,wehaveseen that A
jT 0
doesnothave
arunthatlabelstherootwithaninitialstateifax
1
;ax
2 g\T
0
6=;,andthusiP
holdsfor(I;T 0
).
Now, we show how to transforma orret axiomati automaton into aweighted
generalizedBuhiautomatonwhosebehaviourisapinpointingformulafortheinput.
This weighted automaton uses the T-Boolean semiring, whih is dened as B T
:=
(
^
B(T);^;_;>;?),where
^
B(T) is thequotient set ofall monotone Boolean formulae
overlab(T) by thepropositional equivalenerelation,i.e., twopropositionally equiv-
alent formulae orrespond to the same element of
^
B(T). It is easy to see that this
semiringis indeed adistributive lattie, where the partialorderis dened as
i ! isvalid. Furthermore,as T is nite,this lattie is also nite.
6
Notethat,
similartotheaseoftheBooleansemiringB,onjuntionisthesemiringaddition(i.e.,
yieldsthesupremum)anddisjuntionisthesemiringmultipliation(i.e.,yieldsthe
inmum).Likewise,>istheleastelement0and?isthegreatestelement1.
6 T
Denition13 (pinpointing automaton) Let (A;res ;Ires) be an axiomatiau-
tomaton for = (I;T), with A = (Q;;I;F
1
;:::;F
n
). The violating funtions
vio:Q k +1
!B T
andIvio:Q!B T
aregivenby
vio(q
0
;q
1
;:::;q
k ) :=
_
ft2Tj(q0;q1;:::;qk)=2res (t)g lab(t);
Ivio(q) :=
_
ft2Tjq2Ires(t)g= lab(t);
wheretheemptydisjuntionyields?.
Thepinpointingautomatoninduedby(A;res ;Ires )w.r.t.T istheWGBAover
B T
(A;res ;Ires) pin
=(Q;in;wt;F
1
;:::;Fn),where
in(q):=
(
Ivio(q) ifq2I,
> otherwise;
wt(q
0
;q
1
;:::;q
k ):=
(
vio(q
0
;q
1
;:::;q
k
) if(q
0
;q
1
;:::;q
k )2,
> otherwise.
It is easyto seethat, if r : K
! Q is a runof A, then itsweight is given by
wt(r) =
W
u2K
vio(
!
r(u)); otherwise, wt(r) =>.Intuitively,the violating funtion
vio expresses whih axiomsare not\satised" by agiven transition, and thusthe
weightofarunaumulatesalltheaxiomsviolatedbyanyofthetransitionsappearing
aslabelsinit.Additionally,thefuntionIviorepresentstheaxiomsthatareviolatedby
theinitialstateofthisrun.Removingalltheaxiomsappearinginthesetwoformulae
wouldyield a subsetof axioms whihatually allows for this run;and hene,if the
runissuessfuland theroot islabelledwithaninitialstate, duetoorretness,the
propertydoesnotholdanymore.Conjoiningthisinformationforallpossiblesuessful
runsleadsustoapinpointingformula.
Beforeformulating and provingthis fatmore formally,let usillustrate theon-
strution of the pinpointing automaton on the axiomati automaton introdued in
Example3.
Example4 Let (A ex
;res ;Ires) be the axiomati automaton from Example 3. The
orrespondingpinpointingautomatonhastheinitialdistributionin,where
in(q
0 )=ax
2
and in(q
1 )=in(q
2 )=in(q
3 )=>;
andtheweight funtionwt,where
wt(q
1
;q
1
;q
1 )=ax
1
and wt(q
2
;q
2
;q
2 )=ax
3
;
wt(q;q 0
;q 00
)=? if (q;q 0
;q 00
)2nf(q
1
;q
1
;q
1 );(q
2
;q
2
;q
2 )g;
wt(q;q 0
;q 00
)=> if (q;q 0
;q 00
)62:
The behaviour of this WBA is k(A ex
;res ;Ires) pin
k =
V
r2su
A ex
in(r("))_wt(r).
Obviously,onlysuessfulrunsthat labeltherootwithq
0
anontributeaonjunt
dierentfrom>tothisonjuntion.ThereisasinglesuessfulrunofA ex
thatsatises
thisrestrition:therunr
1
,whihlabelstherootwithq
0
andallothernodeswithq
1 .
Theweightofthisruniswt(r
1
)=wt(q
0
;q
1
;q
1 )_wt(q
1
;q
1
;q
1
)=?_ax
1
=ax
1 .Sine
in(q
0 )=ax
2
,thisshowsthatk(A ex
;res ;Ires) pin
k=ax
2 _ax
1
,whihisapinpointing
Theorem 1 LetP be a -property, and =(I;T) anaxiomatized input. Iftheax-
iomatiautomaton(A;res ;Ires) isorretfor w.r.t.P,thenk(A;res ;Ires) pin
kis
apinpointingformula for w.r.t.P.
Proof Weneedtoshowthat,for everyvaluationV lab(T),itholdsthat V satises
k(A;res ;Ires) pin
ki (I;T
V
) 2P. Let V lab(T). Suppose rstthat (I;T
V ) 2= P.
Sine(A;res ;Ires)isorretfor w.r.t.P,theremustbeasuessfulrunrofA
jTV
with r(") 2 I\Ires(T
V
). Consequently,
!
r(u) 2 res (T
V
) holds for every u 2 K
,
and thus V annot satisfy vio(
!
r(u)), for any u 2 K
. Siner is a suessful run
of A
jTV
, it is also a suessful run of A, whih implies wt(r) =
W
u2K
vio(
!
r(u)).
Thus, V doesnot satisfy wt(r).Sine r(") 2 I,we know that in(r("))=Ivio(r("));
additionally,r(")2Ires (T
V
) impliesthat V doesnotsatisfy Ivio(r(")). Thus,V does
notsatisfyin(r("))_wt(r).ButthenV alsoannotsatisfy
V
r2su
A
in(r("))_wt(r)=
k(A;res ;Ires) pin
k.
Conversely,ifV doesnot satisfy k(A;res ;Ires) pin
k=
V
r2su
A
in(r("))_wt(r),
thentheremustexistasuessfulrunrsuhthatV doesnotsatisfyin(r("))_wt(r).
This implies that r(") 2 I \Ires(T
V
) and that
!
r(u) 2 res (T
V
) for all u 2 K
.
Consequently, r is a suessful run of A
jTV
with r(") 2 I\Ires(T
V
), whih shows
(I;T
V
)2=P,bytheorretnessoftheaxiomatiautomaton. ut
4.2 ConstrutingAxiomatiAutomataforSI
IfwewanttoapplyTheorem1toobtainanautomata-basedapproahforpinpointing
unsatisabilityinSI,wemustshowhow,givenanALConeptdesriptionC andan
SITBoxT,weanonstrutanaxiomatiautomaton(A
C;T
;res
C;T
;Ires
C;T )that
isorretfor(C ;T)w.r.t.unsatisability.Forthispurpose,wemustadapttheknown
onstrutionofaloopingautomatonfor SI from[3℄ suhthat ityieldsanaxiomati
automaton.
7
Asmentionedbefore, theautomata-based approahfordeiding(un)satisability
usesthefatthat aoneptis satisable iithasa so-alledHintikkatree.Theau-
tomatontobeonstrutedwillhaveexatlytheseHintikkatreesasitsruns.Intuitively,
Hintikkatreesareobtainedfromtree-shapedmodelsbylabellingeverynodewiththe
\relevant"oneptdesriptionstowhihitbelongs.
Following[3℄,weassumethatalloneptdesriptionsareinnegationnormalform
(NNF),i.e.,negationappearsonlydiretlyinfrontofoneptnames.AnyALConept
desription anbetransformedintoNNFinlinear timeusingdeMorgan, dualityof
quantiers,andeliminationofdoublenegations.WedenotetheNNFofC by nnf(C)
andnnf(:C)byvC.GivenanALConeptdesriptionCandanSITBoxT,theset
ofrelevantoneptdesriptionsisthesetofallsubdesriptionsofCandoftheonept
desriptionsvDtE forDvE2T.Wedenotethissetbysub(C ;T).Thesetofrole
namesourringinC or T isdenotedby rol(C ;T).Thestatesofour automatonare
so-alled Hintikkasets,whihinadditiontosubdesriptions alsoontaininformation
aboutwhihrolesaresupposedtobetransitive.
7
Onthe onehand,theonstrution in[3℄ismoreomplexthanthe onegivenhere sine
the statesofthe automatain[3℄ontainadditionalinformationneededfordeteting yles
inarunasearlyaspossible,whihisnotrelevantforthepresentpaper.Ontheotherhand,
thestatesoftheautomataonstrutedhereontainadditionalinformationabouttransitivity
Denition14 (Hintikkaset) AsetHsub(C ;T)[rol(C ;T) isalled aHintikka
set for(C ;T)ifthefollowingthreeonditions aresatised:
(i) ifDuE2H,thenfD;EgH;
(ii) ifDtE2H,thenfD;Eg\H6=;;and
(iii) thereisnooneptnameAsuhthatfA;:AgH.
TheHintikkasetH isompatiblewith theGCI DvE 2T if itis theemptysetor
ontains vDtE.It isompatible withthetransitivity axiom trans(r)2T ifitisthe
emptysetorontainsr.Finally,itisompatiblewiththeinverseaxiominv(r;s)2T if
r2Himpliess2Handvieversa.
Thearityk ofour automatonis determinedby thenumberofexistentialrestri-
tions,i.e.,oneptdesriptionsoftheform9r:D,ontainedinsub(C ;T).Sineweneed
toknowwhihsuessorinthetreeorrespondstowhihexistentialrestrition,wex
anarbitrary bijetion':f9r:Dj9r:D2sub(C ;T)g!K.Toobtainfullk-arytrees,
we will use nodes labelled with the empty set (whih is a Hintikka set) as dummy
nodes.Thefollowing Hintikkaonditions willbe usedtodenethetransitionsof our
automaton.
Denition15 (Hintikka ondition) The tupleofHintikkasets (H
0
;H
1
;:::;H
k )
for (C ;T) satises the Hintikka ondition if thefollowing holdsfor everyexistential
restrition9r:D2sub(C ;T):
{ If9r:D 2H
0
, thenH
'(9r:D)
ontains D as wellas everyE for whihthere is a
valuerestrition8r:E2H
0
;if,inaddition,r2H
0
,then8r:EbelongstoH
'(9r:D)
foreveryvaluerestrition8r:E2H
0 .
{ If9r:D2=H
0
,thenH
'(9r:D)
=;.
This tupleis ompatible with the GCI D v E 2 T (ompatible with the transitivity
axiom trans(r)2T)ifallitsomponentsare ompatiblewithD vE(trans(r)).Itis
ompatible with the inverse axiom inv(r;s)2 T if all itsomponentsare ompatible
withinv(r;s),and thefollowingholdsfor allt2fr;sg andt 2fr;sgnftg:for every
8t:F 2H
'(9t :D)
,thesetH
0
ontainsF,andadditionally8t :F ift2H
0 .
WearenowreadytodenetheaxiomatiautomatonforunsatisabilityinSI.
Denition16 (axiomatiautomatonfor SI)LetCbeanALConeptdesrip-
tion,T anSI TBox, and k the numberof existential restritions insub(C ;T). The
axiomatiautomaton(A
C;T
;res
C;T
;Ires
C;T
)hasasitsrstomponentthelooping
automatonA
C;T
:=(Q;;I),where
{ QonsistsofallHintikkasetsfor(C ;T);
{ onsistsofall(H
0
;H
1
;:::;H
k )2Q
k +1
thatsatisfytheHintikkaondition;
{ I:=fH2QjC2Hg.
Thetransition restriting funtionres
C;T
mapseahaxiomt2T tothe setof all
tuplesinthatare ompatiblewitht.Theinitialrestriting funtionIres
C;T maps
eahaxiom t 2 T to the set Q, i.e., there is eetively norestrition onthe initial
statesimposedbytheaxioms.
Corretnessofthisautomatononstrutionanbeshownbyaneasyadaptationof
Theorem 2 LetCbeanALConeptdesriptionandT anSITBox.Theaxiomati
automaton(A
C;T
;res
C;T
;Ires
C;T
)isorretfor(C ;T)w.r.t.unsatisability.
Theorem1shows that itis enough toomputethe behaviourof thepinpointing
automaton(A
C;T
;res
C;T
;Ires
C;T )
pin
induedby(A
C;T
;res
C;T
;Ires
C;T
) inorder
toobtainapinpointingformulafor (C ;T)w.r.t. unsatisability.InSetion5,wewill
show how this behaviour an be omputed, but rst we present an example of an
axiomatiautomatonwheretheuseofaBuhiaeptaneonditionisneessary.
4.3 ConstrutingAxiomatiAutomataforLTL
TheaxiomatiautomatonforLTLa-unsatisabilitywillhaveasstatessetsofformulae
similarto the Hintikkasets introdued for SI, buttheywill need to satisfy slightly
dierentonditions,duetothefatthatwewillnotassumethattheformulaeusedare
innegationnormalform.
8
GivenanLTLformulaandasetofLTLformulaeR,the
losure of(;R)is thesetofall subformulae ofandR, andtheirnegations, where
doublenegationsareanelled.Wedenotethissetasl(;R).
Following[42℄,thestatesofourautomatonareelementarysetsofformulae,whih
playtheroleoftheHintikkasetsoftheprevioussubsetion.Elementarysetsaremax-
imalandonsistentsetsofsubformulaeinl(;R).
Denition17 (elementary set) Theset Hl(;R) is alled anelementaryset
for(;R)ifitsatisesthefollowingonditions:
{ :2Hi2=H, forall:2l(;R);
{ ^ 2Hif; gH, forall^ 2l(;R);
{ 2HimpliesU 2H, forallU 2l(;R);
{ ifU 2Hand 2=H,then2H
Theautomatononstrutedfromagiveninput(;R)takesunarytreesasinput,i.e.,
itsrunsareinnitewordsoverthesetofstates.Thetransitionrelationisthusbinary.
It plays the role of the Hintikka ondition, ensuring that temporal restritions are
transferedtosuessornodeswhenneessary.
Denition18 (ompatible)Atuple(H;H 0
)ofelementarysetsisalledompatible
ifitsatisesthefollowing onditions:
{ forall 2l(;R), 2Hi 2H 0
;and
{ forallU 2l(;R),U 2Hieither(i) 2Hor(ii)2HandU 2H 0
.
Therunsofourautomatonwill besequenesofelementarysets whereeahtwoon-
seutiveonesformaompatibletuple.Inontrastto thease forSI,thepreseneof
arun ofthis automaton doesnot imply theexistene of aomputation. Thereason
is that one an delay the satisfation of an untilformula indenitely; that is, every
nodein the runmay ontain the formula U while none ontains , violating this
waythelastonditioninthedenitionofaomputationfortheinput.Inordertorule
outthesekindsofrunsand makesurethateahuntilformulais eventuallysatised,
wewillimposeageneralizedBuhiondition,whihintroduesasetofnalstatesfor
eahuntilformulainl(;R).
8
AlthoughitispossibletotransformLTLformulaeintonegationnormalform,wedeided
nottodothisinordertostayasloseaspossibletothe knownautomatononstrutionfor
Denition19 (axiomati automatonfor LTL)LetandRbeanLTLformula
and a set of LTLformulae, respetively, and let
1 U
1
;:::;
n U
n
be all the until
formulaeinl(;R).Theaxiomatiautomaton(A
;R
;res
;R
;Ires
;R
)hasasitsrst
omponentthegeneralizedBuhiautomatonA
;R
:=(Q;;I;F
1
;:::;Fn), 9
where
{ Qisthesetofallelementarysetsfor(;R);
{ onsistsofallompatiblepairs(H;H 0
)2QQ;
{ I:=fH2Qj2Hg;
{ F
i
:=fH 2Qj
i
2Hor
i U
i
= 2Hg.
For every 2 R,the transition restriting andinitial restriting funtionsare given
byres
;R
( ):=andIres
;R
( ):=fH2Qj 2Hg,respetively.
Corretnessofthisautomatonanbeshownbyasimpleadaptationoftheproofin[42℄.
Theorem 3 Let bean LTLformula andR aset of LTLformulae. Theaxiomati
automaton(A
;R
;res
;R
;Ires
;R
)isorretfor(;R)w.r.t.a-unsatisability.
From Theorem 1 we know that it suÆes to omputethe behaviour of the pin-
pointingautomaton (A
;R
;res
;R
;Ires
;R )
pin
induedby(A
;R
;res
;R
;Ires
;R )
in orderto obtain a pinpointing formula for (;R) w.r.t. a-unsatisability. We will
shownowhowthisbehaviouranbeomputed.
5Computingthe BehaviourofWeighted TreeAutomata
Inthissetion,werstshowhowthebehaviourofaweightedBuhiautomaton(WBA)
on a nitedistributive lattie anbe omputed by two nested iterations. Then, we
desribehowthisapproahanbesimpliedtoasingle\bottom-up"iterationforthe
speialaseofaweightedloopingautomaton(WLA).Next,weshowthatanyweighted
generalizedBuhiautomaton(WGBA)anberedued,inpolynomialtime,toaWBA
thathasthesamebehaviour.Thisredutionfollowstheideasthathavepreviouslybeen
usedfor thease ofunweightedautomata[41℄.Finally,weompareourapproahfor
omputingthebehaviourofaweighted Buhiautomatonwiththeoneindependently
developedin[15℄.
5.1 ComputingtheBehaviourofaWBA
Clearly,the naveapproahthat diretlyusesthe denitionofthebehaviourby rst
omputingandthenaddinguptheweightsofallsuessfulrunswouldnotproduea
resultinnitetimesinethereare potentiallyinnitely manysuessfulruns, whih
are themselves innite. Instead, we will use aniterative method for omputing the
behaviour,whihgeneralizes theemptinesstestfor Buhiautomata
9
If n=0,i.e.,and Rdonotontainuntilformulae,thenthisautomatonisatuallya
TheEmptinessTestforB uhiAutomata
Theemptiness problemfor Buhiautomataanbedeidedintimepolynomialinthe
size oftheautomaton [30,41℄. Thedeisionproedureonstrutsthe setofall states
thatannotour aslabels inanysuessfulrun;we willallthesestatesbad states.
Weantryto disprove thatastate isbad bytryingtoonstrut anitepartialrun
whereeverypathendsinanalstate.
10
Everystatefor whihthisonstrutionfails
islearly bad,buttheremaybebad statesfor whihthis onstrutionsueeds.The
reason is that some of the nal states reahedby the nite runmay themselvesbe
bad.Thus,inorderto omputeallbad stateswe mustiteratethis proess,where in
thenextiterationthepartialrunisrequiredtoreahnalstatesthatarenotalready
knowntobebad.Notie,however,thattheonstrutionofanitepartialrunendingin
non-badnalstatesanitselfberealizedbyaniterativeproedure.Hene,thedeision
proedurefor theemptiness problemusestwonestediterations. Intheinnerloop,we
trytoonstrutanitepartialrunnishingin(non-bad)nalstatesforeverystate.In
theouterloop,weusetheresultoftheinneriterationtoupdatethesetof(known)bad
states,andthenre-starttheinneriterationwiththisnewinformation.Letusallthe
statesfor whihthereisanitepartialrunnishinginnon-badnalstatesadequate.
First, any state q 2 Q for whih thereis a transition leading to onlynon-bad nal
statesislearlyadequate.Then,everystateforwhihthereisatransitionleadingonly
to states that are either(i) nal and not bad or (ii) already knownto beadequate
isalso adequate.Obviously,duringthis iteration,the setofadequatestatesbeomes
stable afteratmost jQjiterations. Theouterloopthenaddsall thestatesthatwere
foundnottobeadequatetothesetofbadstates.Thesetofbadstatesmaintainedin
thisouteriterationbeomesstableafteratmostjQjsteps.Itanbeshownthatthere
isasuessfulrunthatstartswithaninitialstateinotallinitialstatesareontained
inthesetofbadstatesomputedthisway.Thisyieldsanemptinesstestthatrunsin
timepolynomialinthenumberofstates(see[41℄fordetails).
Example5 LetusillustratethisapproahontheBuhiautomatonA ex
ofExample1.
First,wetrytoonstrut,foreverystate,anitepartialrunwhereeverypathendsin
analstate.Thisispossibleforq
0 ,q
1 ,andq
2
,butnotforq
3
.Thus,inthisiteration,
q
0
;q
1
;q
2
aretheadequatestates,andq
3
isnotadequate,whihmeansthatq
3
isadded
tothe setofbad states.Inthe nextiteration, q
2
turnsouttobe nolongeradequate
sineit anonlyreahthebad nalstateq
3
.Thus,itis alsoputintothe setofbad
states. Afterthat, the proess beomes stable, i.e., the set fq
2
;q
3
gis the set ofbad
statesomputedbythealgorithm.Sinetheinitialstateq
0
doesnotbelongtothisset,
weknowthatthereisasuessfulrunthatstartswiththisinitialstate.
EmptinessTestbyBehaviourComputation
BeforetreatingthegeneralaseofaWBA,weonsiderthespeialaseofaweighted
automatonovertheBooleansemiringthatsimulatesanunweightedone.InExample2,
wehavedened,foreveryBuhitreeautomatonAaWBAAwsuhthatthebehaviour
ofAw is0iAhasasuessfulrunthatlabelstherootwithaninitialstate.Inthis
ase, theomputationof thebehaviour ofAw basiallyoinideswiththeemptiness
testappliedtoA.
10
Infat,theemptinesstestforBuhiautomataskethedaboveanbeadaptedsuh
thatitomputesthe behaviourofA
w
asfollows. Weonstrutafuntionbad:Q!
f0;1gsuhthat bad(q)=1 iq is abad state.Theouteriteration of thealgorithm
willupdatethisfuntionateverystep.Inthebeginning,nostateisknowntobebad,
andthuswestart theiterationwith bad
0
(q)=0forall q2Q.Nowassumethatthe
funtionbad
i
:Q!f0;1gfori0hasalreadybeenomputed.Forthenextstepof
theiteration,wealltheinnerlooptoupdatethesetofadequatestates.Inthisloop,
wearegoingtoomputethefuntionadq i
:Q!f0;1g.Here,adq i
(q)=1meansthat
qisnot anadequatestate,i.e.,thatitisnotpossibletoonstrutarunstartingwith
this statewhereeahpathreahesat leastonenon-bad nalstate.Atthe beginning
we know nothing about the adequate states, so we set adq i
0
(q) = 1 for all q 2 Q.
Assumethatwe havealreadyomputedadq i
n
:Q!f0;1g. Toknowwhetherastate
shouldbeomeadequateinthenextstep,weneedtohekforeahtransitionstarting
fromthisstatewhetherthenalstatesreahedbythetransitionarenon-badandthe
non-nalstatesarealreadyknowntobeadequate.Thus,wehave
adq i
n+1 (q)=
^
(q;q
1
;:::;q
k )2Q
k +1 wt(q;q
1
;:::;q
k )_
_
qj2F= adq
i
j (q
j )_
_
q
j 2F
bad
i (q
j ): (3)
Thefuntionadq i
isthelimitofthisinneriteration,whihisreahedafteratmostjQj
steps.Withthisfuntion,wedene
bad
i+1
(q)=bad
i
(q)_adq i
(q):
Thefuntionbadisthelimitofthisouteriteration,whihisalsoreahedafteratmost
jQjsteps.Thisomputationofthefuntionbadbytwonestediterationsbasiallysim-
ulatestheomputationofallbadstatesintheemptinesstestforBuhitreeautomata
skethedabove.Itisthuseasytoshowthatbad(q)=1iqisabadstate,i.e.,annot
ourasalabelinasuessfulrunofA.
Given the denitionof A
w
,it is easy to seethat a run r : K
!Q of A
w has
weight0iitisarunofA(seeExample2).Consequently,Ahasasuessfulrunthat
startswithaninitialstate ikAwk=
V
r2suA
w
in(r("))_wt(r)=0.Putting these
observations together,wethus have:the behaviourof Aw is 0i Ahas asuessful
run that starts with an initial state i there is an initial state q (i.e., in(q) = 0)
that is not bad (i.e.,bad(q) =0). This shows that thebehaviour of A
w
is given by
V
q2Q
in(q)_bad(q).
Next, we show thatthe behaviour ofaWBA analwaysbeomputed by suha
proedurewithtwonestediterations.
ComputingtheBehaviourintheGeneralCaseofanArbitrary WBA
Inthe following, we assumethat A=(Q;in;wt;F) is anarbitrary,butxed,WBA
overthe nitedistributive lattie (S;
S
). We will show that theWBA A induesa
monotoneoperatorQ:S Q
!S Q
,where S Q
isthe setofallmappingsfromQtoS,
andthat thebehaviourofAaneasilybeobtainedfromthegreatestxpointofthis
operator.Thepartialorder
S
anbetransferredtoS Q
intheusualway,byapplying
itomponent-wise:for; 0
2S Q
,wedene
S Q
0
i(q)
S
0
(q)forall q2Q.
Itiseasytoseethat(S Q
;
S
Q)isagainanitedistributivelattie.Wewilluseand
also todenote the inmum andsupremuminS Q
.Theleast (greatest)elementof
S Q
isthefuntion
e
0(e
1)thatmapseveryq2Qto0(1).Thedenitionof theoperatorQwill followtheideaoftheiterativeproedurewe
skethedbeforefor solving the emptinessproblem. We fous rst onthe innerloop,
whihisrealizedbyanothermonotoneoperatorO.Notiethattheinternaliterationof
thealgorithmdependsonthesetofbadstatesomputedsofar.Wewillassumethat
thisinformation isgivenbyafuntionf2S Q
.Thus,weatually deneanoperator
O
f
for eahsuhf.FollowingtheideaofEquation(3),theoperatorO
f
isdenedas
followsfor every2S Q
:
O
f
()(q)=
M
(q;q
1
;:::;q
k )2Q
k +1 wt(q;q
1
;:::;q
k )
O
kj=1 step
f ()(q
j
); (4)
where
step
f
()(q)=
(
f(q) ifq2F
(q) otherwise
Lemma1 For every f 2 S Q
the operator O
f
is monotone, i.e.,
S Q
0
implies
O
f
()
S Q
O
f (
0
).
Proof Let; 0
2S Q
besuhthat
S Q
0
.Thisimpliesalsostep
f
()
S Q
step
f (
0
).
Thus,wehaveforeveryq2Q:
O
f
()(q) =
M
(q;q
1
;:::;q
k )2Q
k +1 wt(q;q
1
;:::;q
k )
O
kj=1 step
f ()(q
j )
S
M
(q;q1;:::;q
k )2Q
k +1 wt(q;q
1
;:::;q
k )
O
kj=1 step
f (
0
)(q
j )=O
f (
0
):
u t
Sine we know that S Q
is nite, this in partiular means that the operator O
f is
ontinuous.ByTarski'sxpointtheorem[39℄,thisimpliesthatO
f has
L
n0 O
n
f (
e
0)asitsleastxpoint(lfp).FinitenessofS Q
yieldsthatthislfpisreahedafternitelymany
iterations: thereexists a smallest m;0 m jSj jQj
suhthat O m
f
(
e
0) =Om+1f (0),
e
and for thism wehave
L
n0 O
n
f
(
e
0 )=Omf
(
e
0 ).This yieldsa boundonthe numberof iterations that is exponential in the size of the automaton. We will later show
(seeTheorem6)thatit ispossibleto improvethis boundtoapolynomialnumberof
iterations,measuredinthenumberofstates.
Reallthattheintuitionoftheinternaliterationwastondoutfromwhihstates
it is possible to builda nitepartial runthat nishes innal states. In the general
ase,theoperatorsOwillhelpinomputingtheweightsofallsuhpartialruns.Next,
wegiveaformaldenitionofthenotionofanitepartialrun.
Denition20 (nite run) A nitetree is anitesett K
that islosed under
prexes and suh that, if ui 2 t for some u 2 K
and i 2 K, then uj 2 t for all
j;1jk.Anodeu2tisalledaleaf ifthereisnoj;1jk,suhthatuj2t.
Thesetofallleafnodesofanitetreetis denotedbylnode(t).Thedepthofanite
treetisthelengthofthelargestwordint.
A niterun is a mapping r :t !Q, where tis a nitetree. Givensuha run,
Wedenotebyruns
1
thesetofallniterunsrofdepthatleast1suhthat,forevery
nodeu6=",r(u)2F ifandonlyifuisaleaf.Additionally,foreveryn1,letruns n
1
denotethe setof all niteruns inruns
1
havingdepthat most n.For astate q2 Q,
runs
1
(q)=fr2runs
1
jr(")=qg;analogously,runs n
1
(q)=fr2runs n
1
jr(")=qg.
Theweightofaniterunr:t!Qiswt(r)=
N
u2tnlnode (t)
wt(r(u);r(u1);:::;r(uk)).
Lookingagainatthespeialaseofaweightedautomatonsimulatinganunweighted
one,weseethatduring theinneriterationwedonotwanttoomputetheweightsof
all niterunsinruns
1
butonlythosethatnish instatesthat are notbad.Inother
words,wemultiplytheweightoftherun,bythefuntionbadomputedsofarapplied
toeahofitsleafs.Givenafuntionf:Q!S,wedenethef-weightofaniterun
r aswt
f
(r)=wt(r)
N
q2leaf(r)
f(q).ThelfpoftheoperatorO
f
omputesthesum
ofthef-weightsofallrunsinruns
1 .
Lemma2 Foralln0andallq2Q,O n
f
(
e
0)(q)=L
r2runs n
1 (q)
wt
f (r).
Proof Theproofisbyindutiononn.Forn=0,theresultfollowsfromthefatthat
runs 0
1
=;,andhene
L
r2runs 0
1 (q)
wt
f
(r)=0=
e
0(q)=O0f
(
e
0 )(q).Assumenowthattheidentityholdsforn.Givenatuple(q
1
;:::;q
k )2Q
k
,leti
1
;:::;i
l
bealltheindies
suhthat q
i
j
=
2F forall j;1jl ,andi
l+1
;:::;i
k
thoseindiessuhthatq
i
j 2F
forallj;l+1jk.For1jl ,wewillabbreviateruns n
1 (q
ij )asrn
n
j
andleaf(r
j )
aslf
j
.Inaddition,Fisanabbreviationfortheprodut
N
kj=l+1 f(q
ij
).Then,
O n+1
f
(
e
0 )(q)=M
(q
1
;:::;q
k )2Q
k wt(q;q
1
;:::;q
k )
O
kj=1 step
f (O
n
f (
e
0))(qj) (5)
=
M
(q1;:::;q
k )2Q
k wt(q;q
1
;:::;q
k )
O
lj=1 O
n
f (
e
0)(qij )
O
kj=l+1 f(q
i
j
) (6)
=
M
(q
1
;:::;q
k )2Q
k wt(q;q
1
;:::;q
k )(
O
lj=1
M
r
j 2rn
n
j wt
f (r
j
))F (7)
=
M
(q1;:::;qk)2Q k
wt(q;q
1
;:::;q
k )(
M
r
1 2rn
n
1
;:::;r
l 2rn
n
l
O
lj=1 wt
f (r
j
))F (8)
=
M
(q
1
;:::;q
k )2Q
k wt(q;q
1
;:::;q
k )(
M
r12rn n
1
;:::;rl2rn n
l
O
lj=1 wt(r
j )
O
p2lf
j
f(p))F (9)
=
M
(q
1
;:::;q
k )2Q
k
M
r12rn n
1
;:::;rl2rn n
l wt(q;q
1
;:::;q
k )
O
qj2F= wt(r
j )
O
p2lf
j
f(p)F(10)
=
M
r2runs n+1
1 (q)
wt(r)
O
p2leaf(r)
f(p) (11)
=
M
r2runs n+1
1 (q)
wt
f (r):
Identities(5)and(6)employthedenitionoftheoperatorO
f
andstep
f
,respetively,
and (7)appliesthe indutionhypothesis. Identity(8)uses thefat thatS Q
is adis-
usesthedenitionofthef-weight.Identity(10)usesagainthedistributivitytomulti-
plywt(q;q
1
;:::;q
k
)in.Finally,Identity(11)simpliesthetwosumsbyonstrutinga
runoflargerdepth.Insteadofonsideringrstthe transition(q;q
1
;:::;q
k
)and then
runsofdepthuptonstartingwitheahq
ij
,wesimplytaketheorrespondingrunof
depthn+1startingatq.Thisrunlabelstherootwithqandthesuessornodeiwith
q
i .Ifq
i
isanalstate,thenitremainsasaleaf,otherwise, belowthenodeiwehave
theformerrunstartingwithq
i
.Thus,thesetofleafsofthislargerrunistheunionof
thesets ofleafs ofthe runsr
j
andthe setofthoseq
i
s thatare nalstates. Thelast
identitymerelyappliesthedenitionoff-weightagain. ut
Theorem 4 Letf2S Q
andassume that
0
isthelfpoftheoperator O
f
.Then,for
every q2Q,
0 (q)=
L
r2runs
1 (q)
wt
f (r).
Proof ByLemma2,wehave
M
n0 O
n
f
(0)(q)
e
=M
n0
M
r2runs n
1 (q)
wt
f (r)=
M
r2runs1(q) wt
f (r):
Tarski's xpoint theorem says that the least xpoint of O
f is
L
n0 O
n
f
(
e
0), whihompletestheproofofthetheorem. ut
Beforeturningourattentiontotheouteriterationofthemethodforomputingthe
behaviour,wewill present aboundonthe numberof stepsthatare neessarybefore
reahingthexpointoftheinneriteration.
Denition21 AWBAism-nalisingif,foreveryf2S Q
andeverypartialrunrin
runs
1
(q),thereisapartialrunsrinruns m
1
(q)suhthat wt
f (r)
S wt
f (sr).
WewillrstshowthateveryWBAism-nalisingforanymgreatertothenumber
ofstatesjQj.Afterwardswewillshowhowthispropertyyieldsaboundonthenumber
ofiterationsneededtoreahtheleastxpointofO
f .
Theorem 5 LetAbeaWBAwithless thanmstates.ThenAism-nalising.
Proof Let f 2 S Q
and onsider arun r 2 runs
1
(q). If r 2 runs m
1
(q), then thereis
nothingtoprove.Otherwise,ifr2=runs m
1
(q),thentheremustbeapathinroflength
greaterthanm.Sinetherearelessthanmdierentstates,theremustbetwonon-root
nodesu;vinthispathsuhthatr(u)=r(v).Sinethesenodesareonthesamepath,
weanassumew.l.o.g. thatv=uv 0
forsomev 0
2K
nf"g.Wedeneanewrunsas
follows: foreverynodew,ifthereisnow 0
for whihw=uw 0
,thensets(w):=r(w);
otherwise (thatis, ifw = uw 0
for somew 0
) set s(uw 0
) :=r(vw 0
).This onstrution
denesaninjetivefuntiongfromthenodesofstothenodesofrsuhthat,forevery
nodew of s, we have s(w) =r(g(w)).Notiethat this funtionis notsurjetive, as
thereisnowsuhthatg(w)=u.Thus,shas lessnodes thanr.Furthermore,every
transitioninsisalsoatransition inr,andforeveryw2leaf(s),g(w)2leaf(r).This
impliesthatwt
f (r)
S wt
f
(s).Ifsisstillnotinruns m
1
,thenweanrepeatthesame
proessto produeasmallerruns 0
withasmallerf-weight,untilwe ndonethatis
inruns m
1
. ut
Theorem 6 IfAism-nalising,thenO m
(