• Aucun résultat trouvé

Dominance Constraints with Set Operators

N/A
N/A
Protected

Academic year: 2021

Partager "Dominance Constraints with Set Operators"

Copied!
28
0
0

Texte intégral

(1)

HAL Id: inria-00536806

https://hal.inria.fr/inria-00536806

Submitted on 16 Nov 2010

HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or

L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires

Dominance Constraints with Set Operators

Denys Duchier, Joachim Niehren

To cite this version:

Denys Duchier, Joachim Niehren. Dominance Constraints with Set Operators. Proceedings of the First International Conference on Computational Logic, 2000, London, United Kingdom. pp.326-341.

�inria-00536806�

(2)

DenysDuhierandJoahimNiehren

ProgrammingSystemsLab,UniversitatdesSaarlandesSaarbruken

Abstrat. Dominaneonstraintsarewidelyusedinomputationallin-

guistisasalanguagefortalkingandreasoningabouttrees.Inthispaper,

weextenddominaneonstraintsbyadmittingsetoperators.Wepresent

asolverfordominaneonstraintswithsetoperators,whihisbasedon

propagation and distribution rules, and prove its soundness and om-

pleteness.Fromthissolver,wederiveanimplementationinaonstraint

programminglanguagewithnitesetsandproveitsfaithfullness.

1 Introdution

The dominane relation of a tree is the anestor relation between its nodes.

Logialdesriptionsoftreesviadominanewereinvestigatedinomputersiene

sinethe beginningof thesixties,for instanein the logis(W)SkS [15,16℄. In

omputationallinguistis,theimportane ofdominanebasedtreedesriptions

fordeterministiparsingwasdisoveredatthebeginningoftheeighties[9℄.Sine

then,treedesriptionsbasedondominaneonstraintshavebeomeinreasingly

popular[14,1℄.Meanwhile,theyareusedfortree-adjoiningandD-treegrammars

[17,13,3℄,forunderspeiedrepresentationofsopeambiguitiesinsemantis[12,

4℄andforunderspeieddesriptionsofdisoursestruture[5℄.

Adominaneonstraintdesribesanitetreebyonjuntionsofliteralswith

variables for nodes. A dominane literal xC

y requires x to denote oneof the

anestorsofthedenotationofy.Alabelingliteralx: f(x

1

;:::;x

n

)expressesthat

thenodedenotedbyxislabeledwithsymbolf andhasthesequeneofhildren

referredto by x

1

;:::;x

n

. Solvingdominane onstraintsisan essentialservie

required by appliations in e.g.semantis and disourse.Eventhough satisa-

bility of dominane onstraintsis NP-omplete [8℄,it appears that dominane

onstraintsourringinthese appliationsanbesolvedrathereÆiently[2,7℄.

Foratypialappliationof dominaneonstraintsinsemantiunderspei-

ationofsopeweonsiderthesentene:everyyogihasaguru.Thissenteneis

semantially ambiguous, even though itssyntatistruture is uniquely deter-

mined.ThetreesinFigure1speifybothmeanings:eitherthereexistsaommon

guruforeveryyogi,oreveryyogihashisownguru.Bothtrees(andthusmean-

ings) an be represented in an underspeied manner through the dominane

onstraintin Figure2.

In this paper, we propose to extend dominane onstraints by admitting

set operators:union, intersetion,and omplementation anbeapplied to the

relationsofdominane C

and inversedominane B

.Set operatorsontribute

aontrolled form of disjuntionand negation that is eminently well-suited for

(3)

guru forall

yogi has

yogi exists

guru has

Fig.1.Setsoftrees representsetsofmeanings.

forall x0

yogi x1 x2

exists y0

guru y1

has z

y2

x0:forall(x1;x2)^

y0: exists(y1;y2)^

x

1

:yogi^x

2 C

z^

y1: guru^y2C

z^

z:has

Fig.2.Asingletreedesriptionasunderspeiedrepresentationofall meanings.

onstraintpropagation while lessexpressivethan generalBoolean onnetives.

Setoperatorsallowto expressproperdominane,disjointness,nondisjointness,

nondominane,andunions thereof.Suharihset ofrelationsisimportantfor

speifying powerful onstraint propagation rules for dominane onstraints as

wewillargueinthepaper.

Werst present asystem of abstrat saturation rulesfor propagation and

distribution, whih solve dominane onstraints with set operators.We illus-

trate the power of the propagation rules and prove soundness, ompleteness,

and termination in nondeterministi polynomial time. We then derive a on-

reteimplementationinaonstraintprogramminglanguagewithnitesets[11,

6℄andproveitsfaithfulnesstotheabstratsaturationrules.Theresultingsolver

isnotonlywellsuitedforformalreasoningbutalsoimprovesinexpressivenesson

thesaturationbasedsolverfor puredominaneonstraintsof[8℄ andprodues

smallersearhtreesthan theearliersetbased implementation of[2℄beauseit

requires lessexpliit solvedforms. Foromittedproofs,weglobally referto the

extendedversionofthispaperavailablefromhttp://www.ps.uni-sb.de/Paper s/.

2 Dominane Constraints

Werstdenetreestruturesandthendominaneonstraintswithsetoperators

whih areinterpretedin the lassof tree strutures.We assumeasignature

of funtion symbolsranged overby f;g;:::, eah ofwhih isequippedwith an

arity ar(f)0. Constants{ funtion symbolsof arity 0 {are ranged overby

a;b.Weassumethatontainsatleastoneonstantandonesymbolofarityat

least2.Weareinterestedin niteonstrutortreesthatanbeseenasground

termsoversuhasf(g(a;b))in Fig.3.

Wedenean(unlabeled)treetobeanitediretedgraph(V;E).V isanite

setsofnodes rangedoverbyu;v;w,andEV V isanitesetofedges.The

in-degree ofeah node isat most 1;eah tree hasexatlyoneroot, i.e. anode

within-degree0.Weallthenodeswithout-degree0theleaves ofthetree.

(4)

g

a b

Fig.3.f(g(a;b)) A(nite)onstrutortree isatriple(V;E;L)onsist-

ingofatree(V;E),andlabelingsL:V !fornodesand

L:E!Nforedges,suhthatanynodeu2V hasexatly

oneoutgoingedgewithlabelk foreah1kar(()),

and no other outgoing edges. We drawonstrutor trees

as in Fig. 3, by annotatingnodes with their labels and ordering the edges by

inreasinglabelsfromleft toright.If =(V;E;L),wewrite V

=V,E

=E,

L

=L.

Denition1. The treestruture M

of anite onstrutor tree over is

the rst-orderstruturewith domainV

whih provides the dominanerelation

C

anda labelingrelation of arityar(f)+1for eah funtion symbol f 2.

These relations aredenedsuhthat for allu;v;u

1

;:::;u

n 2V

:

uC

v ithereisapathfromutov withegdes inE

;

u:f

(v

1

;:::;v

n )iL

(u)=f;ar(f)=n; andL(u;v

i

)=ifor all1in

Weonsiderthefollowingsetoperators onbinaryrelations:inversion 1

,union

[,intersetion\,andomplementation:.WewriteB

fortheinverseofdom-

inane C

, equality =

for theintersetion C

\B

, inequality 6=

for the

omplementofequality,properdominaneC +

asdominanebut notequality,

B +

fortheinverseofproperdominane,anddisjointness?

for:C

\:B

.

Mostimportantly, thefollowingpartition holdsinalltreestruturesM

.

V

V

= ℄f=

;C +

;B +

;?

g

Thus,allrelationsthatsetoperatorsangeneratefromdominaneC

havethe

form[fr

jr2R gforsomesetofrelationsymbolsRf=;C +

;B +

;?g.

Fordeningtheonstraintlanguage,weletx;y;zrangeoveraninnitesetof

node variables.Adominane onstraints withset operators 'hasthefollowing

abstratsyntax(thatleavesset operatorsimpliit).

'::=xRy j x: f(x

1

; ::: ;x

n

) j '^' 0

j false

whereRf=;C +

;B +

;?gisasetofrelationsymbolsandn=ar(f).Constraints

areinterpretedin thelassoftreestruturesover.Forinstane,aonstraint

xf=;?gy expressesthat the nodesdenoted by x andy areeither equalorlie

indisjointsubtrees.Ingeneral,asetRofrelationsymbolsisinterpretedin M

astheunion[fr

jr2R g.

We write Vars(') for the set of variables ourring in '. A solution of a

onstraint ' onsists of a tree struture M

and a variable assignment :

Vars(') ! V

. We write (M

;) j= ' if all onstraints of ' are satised by

(M

;) intheusualTarskiansense.Foronvenieneweadmitsyntatisugar

andallowtowriteonstraintsof theform xSy whereS isasetexpression:

S ::=Rj=jC

jB

j=j 6=jC +

jB +

j?j:SjS

1 [S

2 jS

1

\S

2 jS

1

Clearly,everysetexpressionS anbetranslatedtoasetR ofrelationsymbols

denotingthesamerelation.Inalltreestrutures,x:Sy isequivalentto:xSy

andxS

1 [S

2

y toxS

1 y_xS

2

y.Thusourformalismallowsaontrolledform

ofnegationanddisjuntionwithoutadmittingfullBooleanonnetives.

(5)

(Clash) x;y ! false

(Dom:Re) ' ! xC

x (xoursin')

(Dom:Trans)xC

y^yC

z ! xC

z

(Eq:Deom) x:f(x

1

;:::;x

n

)^y:f(y

1

;:::;y

n

)^x=y ! V

n

i=1 x

i

=y

i

(Lab:Ineq) x:f(:::)^y:g(::: ) ! x6=y iff6=g

(Lab:Disj) x:f(:::;xi;:::;xj;::: ) ! xi?xj where1i<jn

(Lab:Dom) x:f(:::;y;::: ) ! xC +

y

(Inter) xR1y^xR2y ! xR y ifR1\R2R

(Inv) xR y ! yR 1

x

(Disj) x?y^yC

z ! x?z

(NegDisj) xC

z^yC

z ! x:?y

(Child:up) xC

y^x:f(x1;:::;xn)^ V

n

i=1 xi:C

y ! y=x

DistributionRules:

(Distr:Child) xC

y^x:f(x1;:::;xn) ! xiC

y_xi:C

y (1in)

(Distr:NegDisj)x:?y ! xC

y_x:C

y

Fig.4.SaturationrulesDoftheBaseSolver

3 A Saturation Algorithm

Wenowpresentasolverfordominaneonstraintswithsetoperators.First,we

give abase solverwhih saturates aonstraintwith respet to aset ofpropa-

gationanddistributionrules, andprovesoundness,ompleteness,andtermina-

tionofsaturationinnondeterministipolynomialtime.Seond,weaddoptional

propagationrules,whihenhanethepropagationpowerofthebase solver.

Thebase solveris speiedby the rule shemes in Figure 4.Let D be the

(innite) set of rules instantiating these shemes. Eah rule is an impliation

betweenaonstraintandadisjuntionofonstraints.Wedistinguishpropagation

rules'

1

!'

2

whiharedeterministianddistributionrules'

1

!'

2 _'

3 whih

arenondeterministi.

Proposition1 (Soundness). Therules ofD arevalid in alltreestrutures.

TheinferenesystemDanbeinterpretedasasaturationalgorithmwhih

deides thesatisability ofa onstraint. A propagation rule'

1

!'

2

applies to

a onstraint ' if all atomi onstraints in '

1

belong to ' but at least one of

the atomi onstraints in '

2

does not. In this ase, saturation proeeds with

'^'

2

.A distributionrule'

1

!'

2 _'

3

appliesto aonstraint'ifbothrules

'

1

!'

2 and'

1

!'

3

ouldbeapplied to '.In thisase, one ofthese tworules

is non-deterministiallyhosen andapplied. A onstraintis alled D-saturated

ifnoneof therulesin Danbeappliedtoit.

Proposition2 (Termination).ThemaximalnumberofiteratedD-saturation

stepson aonstraintispolynomially boundedin thenumber ofitsvariables.

(6)

leastonenewliteralto'.OnlyaO(m 2

)literalsanbeaddedsineallofthem

havetheform xR ywhere x;y2Vars(') andRhas16possiblevalues.

z x

1

x

2 x

f

Fig.5.(Neg.Disj) Next,weillustrateprototypialinonsisteniesandhow

D-saturationdetetsthem.Westartwiththeonstraint

x: f(x

1

;x

2 )^x

1 C

z^x

2 C

zinFig.5whihisunsatis-

ablesinesiblingsannothaveaommondesendant.

Indeed, the disjointnessof the siblings x

1

?x

2

an be

derivedfrom(Lab.Disj)whereasx

1 :?x

2

followsfrom(NegDisj)sinex

1 andx

2

havetheommondesendantz.

f x

b y a x

1

Fig.6.(Distr.Child) Toillustratetherstdistribution rule,weonsider

the unsatisable onstraint x: f(x

1 )^ xC

y ^ x

1 :a ^

y:bin Fig.6where a6=b.Wean deidetheposition

of y with respet to x by applying rule (Distr.Child)

whiheither addsx

1 C

y orx

1 :C

y.(1) Ifx

1 :C

y is

added,propagationwith(Child.up)yieldsx=y.Asxandyarrydistintlabels,

rule(Lab.Ineq)addsx6=y.Now,weandeduex;ybyintersetingequalityand

inequality (Inter). Thus, the (Clash) rule applies. (2) If x

1 C

y is added then

(Child.up)yieldsx

1

=y whih againlashesbeauseofdistintlabels.

z

f x g y

Fig.7.(Distr.NegDisj) The seond distribution rules helps detet the in-

onsistenyofx: f(z)^y:g(z)inFig7wheref6=g.Ina

rststeponeaninferfrom(Lab.Dom)thatxC +

zand

yC +

z.Asthe(Inter)ruleallowstoweakenrelations,we

also havexC

z and yC

z,i.e. x:?y by(NegDisj), so

that(Distr.NegDisj)andedueeitherxC

yorx:C

y.ConsidertheasexC

y,

fromyC +

zderivez:C

y by(Inv,Inter),and(Child.up)infersy=xresultingin

alashdueto thedistintlabels.Similarlyfortheotherase.

Denition2. A D-solvedform isaD-saturatedonstraintwithoutfalse.

f x

1

x

4

x

5

x

6 x

2

x

3

Fig.8.D-solvedform Theintuition is that a D-solvedform has abak-

bone whih is a dominane forest, i.e. a forest with

hild and dominaneedges. Forinstane, Fig 8shows

thedominaneforest underlyingx

1 :f(x

4 )^x

4 C

x

5

^

x

4 C

x

6

^ x

2 C

x

3

^ x

5

?x

6

whih beomes D-solved

whenD-propagation.

Wewould like to note that the set based solver for dominane onstraints

of [2℄ insists on more expliit solved forms: for eah two variables, one of the

relations f=;C +

;B +

;?gmust beseleted.Forthedominaneforest in Fig.8,

thisleadsto63expliitsolutionsinsteadofasingleD-solvedform.Thesituation

isevenworsefortheformulax

1 C

x

2

^x

2 C

x

3

^ :::^x

n 1 C

x

n

.Thisonstraint

anbedeterministiallyD-solvedbyD-propagationwhereastheimplementation

of[2℄omputesasearhtreeofsize 2 n

.

Proposition3 (Completeness).Every D-solvedformhas asolution.

(7)

(Child:down) xC +

y^x:f(x1;:::;xn)^ n

i=1;i6=j xi:C

y ! xjC

y

(NegDom) x?y^y:?z ! x:C

z

(Dom:Ineq) xC

y

1

^y

1 C

+

y

2

^y

2 C

z ! x6=z

(Child:Ineq) x6=y^x:f(:::;x 0

;:::)^y:g(:::;y 0

;:::) ! x 0

6=y 0

(Parent:Ineq)xC +

z^y:f(:::;z;:::) ! xC

y

Fig.10.SomeOptionalPropagationRulesO

g x

root

f x

1

g x

4

a x

5

a x

6 f x

2

a x

3

Fig.9.Asolution.

The proof is given in the Setion 4. The idea for

onstrutingasolutionofaD-solvedformistoturnits

underlying dominaneforestinto atree,byadding la-

belssuhthatdominanehildrenareplaedatdisjoint

positionswheneverpossible.Forinstane,asolutionof

thedominaneforestinFig.8isdrawnin Fig.9.Note

that this solution does also satisfy x

5

?x

6

whih be-

longs to the aboveonstraintbut notto its dominane forest.This solutionis

obtainedfrom the dominaneforest in Fig.8by addinga root node and node

labelsbywhih alldominaneedgesareturnedintohildedges.

Theorem1. Saturationby the inferene rulesin Ddeides the satisability of

adominaneonstraintwithsetoperatorsinnon-deterministipolynomialtime.

Proof. Let ' be a dominane onstraint with set operators. Sine all rules in

D are sound (Proposition 1) and terminate (Proposition 2), ' is equivalent to

thedisjuntionof allD-solved formsreahablefrom'bynon-deterministi D-

saturation. Completeness (Proposition 3) yields that ' is satisable i there

existsaD-solvedreahablefrom '.

WeanreduethesearhspaeofD-saturationbyaddingoptionalpropaga-

tionrules O.Taking advantageofset operators,weandene ratherpowerful

propagationrules.TheshemesinFig 10,forinstane,exploit theomplemen-

tationsetoperators,andareindeedsupportedbythesetbasedimplementation

ofSetion6.WeillustrateOin thesituationbelowwhiharisesnaturallywhen

resolvingsopeambiguitiesasinFigure 2.

x: f(x

1

;x

2

)^y:g(y

1

;y

2 )^x

2 C

z^y

1 C

z^xC

y

f x

x

1 x

2 g y

y

1

z y

2

WederivexC +

yby(Lab.Ineq,Inter),x

1

?x

2

by(Lab.Disj),andx

2

:?yby(Dom.

Trans,NegDisj).We ombine thelatter twousing optional rule (NegDom) into

x

1 :C

y.Finally,optionalrule(Child.down)yieldsx

2 C

ywherebythesituation

isresolved.

4 Completeness Proof

WenowproveProposition3whih statesompleteness in thesense thatevery

D-solvedform is satisable. Weproeed in twosteps.First,weidentify simple

(8)

howtoextendeveryD-solvedformintoasimpleD-solvedformbyaddingfurther

onstraints(Proposition5).

Denition3. Avariablexis labeledin'ifx=yin 'andy:f(y

1

;:::;y

n )in '

for somevariable yandtermf(y

1

;:::;y

n

).Avariable yisarootvariablefor'

ifyC

z in 'forallz2Vars(').Weallaonstraint'simpleifallitsvariables

are labeled, andifthereisarootvariable for'.

Proposition4. Asimple D-solvedformissatisable.

Proof. Byindution on thenumber of literals in asimple D-solved form '. '

has aroot variable z. Sineall variablesin 'are labeledthere is avariable z 0

andatermf(z

1

;:::;z

n

)suh thatz=z 0

^z 0

:f(z

1

;:::;z

n

)2'.Wepose:

V =fx2Vars(')jx=z2'gandV

i

=fx2Vars(')jz

i C

x2'g

forall1in.ToseethatVars(')isoveredbyV[V

1 [:::[V

n

,letx2Vars(')

suh that z

i C

x 2= ' for all 1 i n. Saturation with (Distr.Child) derives

either z

i C

x or z

i :C

x;but z

i C

x 62' by assumption, thereforez

i :C

x 2'

forall1in. (Child.up)infers z=x2',i.e.x2V.ForasetW Vars(')

wedene'

jW

to betheonjuntionof allliterals 2'withVars( )W.

'j=j' 0

holdswhere ' 0

=

def '

jV

^z:f(z

1

;:::;z

n )^'

jV

1

^:::^'

jV

n

'j=' 0

followsfrom ' 0

'.Toshow' 0

j='weprovethateah literalin 'is

entailedby' 0

1. Casex: g(x

1

;:::;x

m

)2'forsomevariablexandtermg(x

1

;:::;x

m

):Ifx2

V

i , i.e.z

i C

x2'forsome1inthenx: g(x

1

;:::;x

m )2'

jVi

sine'is

saturatedunder(Lab.Dom,Dom.Trans).Otherwisex2V,i.e.z=x2',and

thusz=x2'

jV

.Sine'islashfreeandsaturatedunder(Lab.Ineq,Clash),

f=g and n=m must hold. Saturation with respet to (Eq.Deom) implies

z

i

=x

i

2 ' for all 1 i n and hene z

i

=x

i 2 '

jVi

. All together, the

righthandside' 0

ontainsz=x^z:f(z

1

;:::;z

n )^^

n

i=1 z

i

=x

i

whihentails

x: g(x

1

;:::;x

m

)asrequired.

2. CasexR y2'forsomevariablesx;y andrelationset Rf=;C +

;B +

;?g.

Sinex;y2V [V

1

[:::[V

n

wedistinguish4possibilities:

(a) x2V

i ,y2V

j

,where1i6=j n.Here,x?y2'bysaturationunder

(Lab.Disj,Inv,Disj).Clash-freenessandsaturationunder(Inter,Clash)

yield?2R .Finally,' 0

entailsz

i

?z

j

andthusx?ywhihinturnentails

xR y.

(b) Whenx;y2V (resp.V

i

),bydenitionxR y2'

jV

(resp.'

jV

i )

() x 2 V and y 2 V

i

. Here, xC +

y 2 ' by saturation under (Lab.Dom,

Dom.Trans).ThusC +

2Rbysaturationunder(Inter,Clash)andlash-

freenessof'.But' 0

entailszC +

z

i

andthusxC +

y whihinturnentails

xR y.

(d) Theasex2V andy2V

i

issymmetritothepreviousone.

(9)

jV

i

exist solutions(M

i

;

i ) j= '

jV

i

for all 1 i n. Thus (M f(

1

;:::;

n )

;) is a

solutionof 'if

jV

i

=

i

and(x)=(z)istheroot nodeof f(

1

;:::;

n )for

allx2V. ut

An extension of aonstraint 'is aonstraintof the form '^' 0

for some

' 0

. Given a onstraint 'wedene a partial ordering

'

on itsvariables suh

thatx

'

y holdsifandonlyifxC

y in'butnotyC

xin'.Ifxisunlabeled

thenwedenetheseton

'

(x)ofvariablesonnetedtoxin 'asfollows:

on

'

(x)=fyjy is

'

minimalwithx

' yg

Intuitively,avariable y is onnetedto x if itis a\diretdominane hild" of

x.Soforexample,on

'

1

(x)=fygandon

'

1

(y)=fzgfor:

'

1 :=xC

x^xC

y^xC

z^yC

z;

Denition4. WeallV Vars(') a '-disjointnesssetif for any twodistint

variables y

1

;y

2 2V,y

1 :?y

2

notin'.

The idea is that all variables in a '-disjointness set an safely be plaed at

disjointpositionsin atleastoneofthetreessolving'.

Lemma1. Let'beD-saturated,x2Vars(').IfV isamaximal'-disjointness

setin on

'

(x) thenfor ally2on

'

(x)there existsz2V suhthaty=z in'.

Proof. If y:?znotin' for all z 2 V then fyg[V is a disjointness set; thus

y2V bymaximalityofV.Otherwise,thereexists z2V suh thaty:?zin'.

Saturationof'withrespetto rules(Distr.NegDisj,Inter)yieldsyC

z in'or

zC

y in'. In both ases, it follows that z=y in' sine z and y are both

'

minimalelementsin theseton

' (x).

Lemma2 (Extension by Labeling). Every D-solved form ' with an unla-

beledvariable xanbeextendedtoaD-solvedformwith stritlyfewerunlabeled

variables, andinwhih x islabeled.

Proof. Letfx

1

;:::;x

n

gbeamaximal'-disjointnesssetinludedinon

'

(x).Let

f be afuntion symbolofarity n in , whih exists w.l.o.g.Otherwise, f an

beenodedfromaonstantandasymbolofarity2whose existenein we

assumed.Wedenethefollowingextensionext(')of':

ext(')=

def

'^x:f(x

1

;:::;x

n )^

^

fxR z^zR 1

xjC +

2R ; x

i C

z in'; 1ing^ (1)

^

fyR zj?2R ; x

i C

y in'; x

j C

zin'; 1i6=jng (2)

Notethatx islabeledinext(') sinex=x2'bysaturationunder(Dom.Re).

We haveto verify that ext (') is D-solved, i.e. that noneof theD-rules an be

applied toext(').Wegivetheproof onlyfortwoofthemoreomplexases.

(10)

1. (Distr.Child) annotbeappliedto x: f(x

1

;:::;x

n

):suppose xC y in'and

onsidertheaseyC

xnot in'.Thus x

'

y andthere existsz2on

' (x)

with zC

yin'. Lemma 1 and the maximality of the '-disjointness set

fx

1

;:::;x

n

g yield x

j

=zin' for some 1 j n. Thus, x

j C

y in' by

(Dom.Trans) and (Distr.Child) annot be applied with x

j

. For all suh

1i6=jnweanderivex

i

?yby(Lab.Dom,Disj,Inv),thusx

i :C

y by

(Inter)and(Distr.Child)annotbeappliedwithx

i either.

2. (Inter)applieswhenR

1

\R

2

R ,yR

1

zinext('), andyR

2

z inext(').We

proveyR zinext(') for theasewhere yR

1

zin'and yR

2

z is ontributed

to ext(') by (2). Thus, ? 2 R

2

and there exists 1 i6=j n suh that

x

i C

yin' and x

j C

z in'. It is suÆient to prove ? 2 R

1

sine then

? 2 R

1

\R

2

R whih implies yR zin'. We assume ? 62 R

1

and de-

rive aontradition. If ? 62 R

1

then R

1

f=;C +

;B +

g. Thus, weakening

yR

1

z in'with(Inter)yieldsy:?z in'.Next,weanapply(Distr.NegDisj)

whihproveseither yC

z in'ory:C

z in'.

(a) IfyC

zin'thenx

i C

zin 'followsfrom(Dom.Trans)andx

i :?x

j in'

from (NegDisj).This ontraditsourassumption that fx

1

;:::;x

n gisa

'-disjointnessset.

(b) If y:C

z in' then we have y:C

z in ' and y:?z in' from whih

on anderiveyB

zin' with (Inter) and zC

y in'with (Inv). From

(Dom.Trans) wederivex

j C

yin'. Sinewealready knowx

j C

y in'

weanapply(NegDisj)whihshowsx

i :?x

j

in'.Butagain,this on-

traditsthatfx

1

;:::;x

n

gisa'-disjointnessset. ut

Proposition5. EveryD-solvedformanbeextendedtoasimpleD-solvedform.

Proof. Let'beD-solved.W.l.o.g.,'hasarootvariable,elsewehooseafresh

variable x and onsider insteadtheD-solvedextension '^ V

fxR y^ yR 1

x j

C +

2R ; y2Vars(')g.ByLemma2,weansuessivelylabelallitsvariables.

u t

5 Constraint Programming with Finite Sets

Current onstraint programming tehnology provides no support for our D-

saturationalgorithm.Instead,improvingon[2℄,wereformulatethetaskofnd-

ingsolutionsofatreedesriptionasaonstraintsatisfationproblemsolvableby

onstraintprogramming [11,6℄.In this setion, we dene ourtarget language.

Its propagation rules are given in Fig 12 and are used in proving orretness

ofimplementation.Distributionrules,however,aretypiallyproblemdependent

andweassumethattheyanbeprogrammatiallystipulatedbytheappliation.

Thus,theonretesolverofSetion6speiesitsdistributionrulesinFigure13.

Let =f1 ::: gbe anite set of integersforsome largepratiallimit

suhas134217726.Weassumeasetofintegervariables withvaluesin and

rangedoverbyI andasetofsetvariables withvaluesin2

andrangedoverby

S.Integerandniteset variablesarealso bothdenotedbyX.

(11)

C::=B j S

1

\S

2

=; j S

3 S

1 [S

2 j C

1

^C

2 j C

1 orC

2

Fig.11.FiniteDomainandFiniteSetConstraints

Equality:

X

1

=X

2

^B[X

j

p B[X

k

℄ fj;kg=f1;2g (eq.subst)

Finitedomainintegeronstraints:

I2D

1

^I2D

2 p

I2D

1

\D

2

(fd.onj)

I2;

p

false (fd.lash)

Finitesetsonstraints:

i2S^i62S

p

false (fs.lash)

S1\S2=;^i2Sj p i62Sk fj;kg=f1;2g (fs.disjoint)

S

3 S

1 [S

2

^i62S

1

^i62S

2 p

i62S

3

(fs.subset.neg)

S3S1[S2^i2S3^i62Sj p i2Sk fj;kg=f1;2g (fs.subset.pos)

Disjuntivepropagators:

B^C

~

p false

B ^ (CorC 0

)

p C

0

B^C

0 ~

p false

B ^ (CorC 0

)

p C

(ommit)

Fig.12.PropagationRules

Theabstrat syntax ofour languageis givenin Fig 11. Wedistinguish be-

tween basi onstraints B, diretly representable in the onstraint store, and

non-basi onstraints C ating as propagators and amplifying the store. The

delarativesemantisoftheseonstraintsisobvious(giventhatC

1 orC

2

isinter-

pretedasdisjuntion).Wewritej=Cif isanassignmentofintegervariables

to integersand set variables to sets whih rendersC true(where set operators

andBooleanonnetiveshavetheusualmeaning).

Weuse the following abbreviations: we write I6=i for I 2nfig, S

1 kS

2

for S

1

\S

2

=;,S=D for ^fi2S j i2 Dg^fi62S ji 2nDg, S

1 S

2 for

S

1 S

2 [S

3

^S

3

=;,andS=S

1

℄S

2 forS

1 kS

2

^S S

1 [S

2

^S

1

S^S

2 S

Thepropagationrules

p

forinferenein this languageare summarizedin

Fig12.TheexpressionC

1 orC

2

operatesasadisjuntive propagator whihdoes

notinvokeanyasedistintion.Thepropagation rulesfordisjuntivepropaga-

torsusethesaturationrelation

~

p

induedby

p

whihin turnisdened by

reursion through

~

p

. Clearly, all propagation rules are valid formulas when

seenasimpliationsorasimpliationsbetweenimpliationsinaseof(ommit).

6 Redution to Finite Set Constraints

Wenowreduedominaneonstraintswithsetoperatorstonitesetonstraints

ofthelanguageintroduedabove.Thisredutionyieldsaonreteimplementa-

tion of theabstrat dominaneonstraint solverwhen realized in a onstraint

programmingsystemsuhas[11,6℄.

(12)

y2R (x) where R (x) is a set variable denoting a nite set of nodes in a tree.

Thisideaisfairlygeneralin thatitdoesnotdepend onthepartiularrelations

interpretingtherelationsymbols.Ourenodingonsistsof3parts:

[['℄℄= ^

x2Vars(') A

1

(x) ^

x;y2Vars(') A

2

(x;y) ^ B [['℄℄

A

1

()introduesanoderepresentationpervariable,A

2

()axiomatizesthetree-

nessoftherelationsbetweenthesenodes,andB[['℄℄enodesthespeirestri-

tionsimposedby'.

Down

x Eq

x

Side

x Up

x

x Representation.Whenobservedfromaspeinode

x, the nodes of a solution tree (hene the variables

that theyinterpret)are partitioned into 4regions:x

itself,allnodesabove,allnodesbelow,andallnodestotheside.Themainidea

istointrodueorrespondingsetvariables.

Letmaxbethemaximumonstrutorarityusedin'.Foreahformalvari-

able x in ' wehoose a distint integer

x

to represent it, and introdue 7+

maxonstraintset variableswrittenEq

x ,Up

x ,Down

x ,Side

x ,Equp

x

,Eqdown

x ,

Parent

x , Down

i

x

for 1 i max, and one onstraintinteger variable Label

x .

Firstwestatethat x=x:

x 2Eq

x

(3)

Eq

x ,Up

x ,Down

x ,Side

x

enode the set of variables that are respetively equal,

above, below, and to the side (i.e. disjoint) of x. Thus, posing I = f

x j x 2

Vars(')gforthesetofintegersenodingVars('),wehave:

I=Eq

x

℄Down

x

℄Up

x

℄Side

x

WeanimprovepropagationbyintroduingEqdown

x

andEqup

x

asintermediate

results.Thisimprovementisrequiredby(Dom.Trans):

I=Eqdown

x

℄Up

x

℄Side

x (4)

I=Equp

x

℄Down

x

℄Side

x (5)

Eqdown

x

=Eq

x

℄Down

x (6)

Equp

x

=Eq

x

℄Up

x

(7)

Down i

x

enodesthesetofvariablesinthesubtreerootedatx'sithhild(empty

ifthereisnosuh hild):

Down

x

=℄fDown i

x

j1imaxg (8)

WedeneA

1

(x)astheonjuntionoftheonstraintsintrodued above:

A

1

(x)=(3)^(4)^(5)^(6)^(7)

(13)

Wellformedness. Posing Rel = f=;C ;B ;?g. In a tree, the relationship

that obtainsbetween the nodes denoted by x and y must be one in Rel. We

introdueanintegervariableC

xy

,alledahoievariable,toexpliitlyrepresent

it and ontribute a well-formedness lause A

3

[[xry℄℄ for eah r 2 Rel. Freely

indentifying thesymbolsinRelwiththeintegers1,2,3,4,wewrite:

A

2

(x;y) = C

xy

2Rel^

^

fA

3

[[xry℄℄jr2Relg (9)

A

3

[[xry℄℄ D[[xry℄℄^C

xy

=r or C

xy

6=r^D[[x:ry℄℄ (10)

Forallr2Rel,itremainstodeneD[[xry℄℄andD[[x:ry℄℄enodingtherelations

xry andx:ry resp.bysetonstraintsontherepresentationsofxandy.

D[[x=y℄℄ = Eq

x

=Eq

y

^Up

x

=Up

y

^Down

x

=Down

y

^Side

x

=Side

y

^Eqdown

x

=Eqdown

y

^Equp

x

=Equp

y

^Parent

x

=Parent

y

^Label

x

=Label

y

^

i Down

i

x

=Down i

y

D[[x:=y℄℄ = Eq

x kEq

y

D[[xC +

y℄℄ = Eqdown

y

Down

x

^Equp

x Up

y

^Side

x Side

y

D[[x:C +

y℄℄ = Eq

x kUp

y

^Down

x kEq

y

D[[x?y℄℄ = Eqdown

x Side

y

^Eqdown

y Side

x

D[[x:?y℄℄ = Eq

x kSide

y

^Side

x kEq

y

Problem spei onstraints. Thethird partB[['℄℄of thetranslationforms

the additionalproblem-spei onstraintsthat further restritthe admissibil-

ity of wellformed solutions and only aept those that are models of '. The

translationisgivenbylauses(11,12,13).

B[['^' 0

℄℄ = B [['℄℄^B[[' 0

℄℄ (11)

A pleasantonsequeneof theintrodutionof hoie variables C

xy

is that any

dominane onstraint xRy an be translated as a restrition on the possible

valuesofC

xy

.Forexample,xC

yanbeenodedasC

xy

2f1;2g.Moregenerally:

B[[xRy℄℄ = C

xy

2R (12)

Finallythelabellingonstraintx:f(y

1 ::: y

n

)requiresamoreomplextreat-

ment.Foreah onstrutorf wehooseadistintinteger

f

toenodeit.

B[[x:f(y

1 ::: y

n

)℄℄ = Label

x

=

f

^ j=max

j=n+1 Down

j

x

=;

^ j=n

j=1 Parent

yj

=Eq

x

^Down j

x

=Eqdown

y

j

^Up

y

j

=Equp

x (13)

Denitionof The Conrete Solver. Foreahproblem'wedeneasearh

strategyspeiedbythedistributionrulesofFigure13.Theserulesorrespond

preisely to (Distr.Child, Distr.NegDisj) of algorithm-D and are to be applied

in the same non-deterministi fashion. Posing =

p [

d

, we dene our

onretesolverasthenon-deterministisaturation

~

induedby andwrite

'

1

~

'

2

to mean that '

2 is in a

~

saturation of '

1

. Whilethe abstrat

solver left this point open, in order to avoid unneessary hoies, we further

requirethata

d

stepbetakenonlyifno

p

stepispossible.

(14)

C

xy

2f=;C g

d C

x

i y

2f=;C g_C

x

i y

62f=;C g forx:f(x

1

;:::;x

n )in'

Cxy6=? d Cxy2f=;C +

g_ Cxy62f=;C +

g

Fig.13.Problemspeidistributionrules

7 Proving Corretness of Implementation

Wenowprovethat[['℄℄ombinedwiththesearhstrategydenedaboveyields

a sound and omplete solverfor '. Completeness is demonstrated by showing

that theonretesolverobtainedby[['℄℄providesatleastasmuh propagation

asspeiedbytherulesofalgorithmD,i.e.wheneverxRyisina!

~

saturation

of'thenC

xy

2Risin a

~

saturationof[['℄℄.

Theorem 2. [['℄℄ issatisablei'issatisable.

Thisfollowsfrom Propositions6and 7below.

Proposition6. if' issatisablethen[['℄℄ issatisable.

Weshowhowtoonstrutamodelof[['℄℄fromamodel(M

;)of'.Wedene

thevariableassignment asfollows:(Up

x )=f

y

j(y)C +

(x)gandsimilarly

for Eq

x

;Down

x

;Side

x

;Eqdown

x

;Equp

x

, (Parent

x )= f

y

j 9k (y)k =(x)g,

(Down k

x ) = f

y

j (y)B

(x)kg, (Label

x ) =

L((x))

and (C

xy

) = R if

(x)R(y)inM

.Wehavethat if(M

;)j='thenj=[['℄℄.

Proposition7. if[['℄℄issatisable, then'issatisable.

WeprovethisbyreadingaD-solvedformoamodel of[['℄℄.

' 0

'^

^

x;y

^

R 0

R xR

0

y whereR=(C

xy )

' 0

is aD-solvedformontaining':allrelationshipsbetweenvariables arefully

resolved and all theirgeneralizations havebeen added. The only possibility is

thatD-rulesmightderiveaontradition.However,if' 0

!

~

p

falsethen[[' 0

℄℄

~

p

false(Lemma4)whihwouldontradittheexisteneofasolution.Therefore

' 0

isaO-solvedformand'issatisable.

Wedistinguishpropagation anddistribution rules;in algorithm D theyare

written !

p and !

d

,andin ouronretesolver

p and

d

. Wewrite' 00

4' 0

for ' 00

is stronger than ' 0

and dene it as the smallest relation that holds of

atomionstraintsandsuhthatfalse4falseandxRy 4 xR 0

y iRR 0

.

Proposition8 (StrongerPropagation).Foreahrule'!

p '

0

ofalgorithm

D,there exists' 00

4' 0

suhthat [['℄℄

~

p [['

00

℄℄.

The proof tehniquefollowsthis pattern: eah ' 0

is of the form xRy and we

hoose' 00

=xR 0

ywhereR 0

R .Assume[['℄℄asapremise.Showthat[['℄℄^C

~

p

false.NotiethatalauseC or C 0

isintroduedby[['℄℄asrequiredby(10).Thus

C 0

follows by (ommit).Then showthat [['℄℄^C 0 ~

p [['

00

℄℄. Forwantof spae,

weinlude hereonlytheproofforrule(NegDisj).

(15)

Lemma3. [[xC y℄℄

p

y

2Eqdown

x

(proofomitted)

Proposition9. [[xC

z^yC

z℄℄

~

p

[[x:?y℄℄

Proof. From thepremises[[xC

z℄℄and[[yC

z℄℄,i.e. C

xz

2f=;C +

gandC

yz 2

f=;C +

g, we must show [[x:?y℄℄ i.e. C

xy

6=?. By Lemma 3 we obtain

z 2

Eqdown

x and

z

2 Eqdown

y

. Sine I =Eqdown

y

℄Up

y

℄Side

y

, we have

z 62

Side

y

.Nowonsiderthenon-basionstraintEqdown

x Side

y

whihoursin

D[[x?y℄℄:from

z

2Eqdown

x

itinfers

z 2Side

y

whih ontradits

z 62Side

y .

Therefore,thewell-formednesslauseD[[x?y℄℄^C

xy

=?or C

xy

6=?^D[[x:?y℄℄

infersitsrightalternativebyrule(ommit).HeneC

xy

6=? ut

Lemma4. (1) if '!

~

p '

0

, thenthereexists ' 00

4' 0

suh that [['℄℄

~

p [['

00

℄℄.

(2)if'!

~

p '

1 and'

1

!

d '

2

,thenthereexists' 0

1 4'

1

suhthat[['℄℄

~

p [['

0

1

℄℄

and[[' 0

1

℄℄

d [['

2

℄℄.

(1) followsfrom Proposition8,and(2) from (1)andthefat that theonrete

distributionrulespreisely orrespondtothoseofalgorithmD.

Proposition10 (Simulation).Theonretesolversimulatestheabstratsolver:

if '!

~

' 0

thenthere exists' 00

4' 0

suhthat [['℄℄

~

[[' 00

℄℄.

Followsfrom Lemma4.

Theorem3. (1)every

~

saturationof[['℄℄orrespondstoaD-solvedformof

'and(2) for everyD-solvedformof 'thereisa orresponding

~

saturation

of [['℄℄.

(1) from Proposition 10. (2) Consider a

~

saturation of [['℄℄. As in Propo-

sition 7, wean onstrut a D-solved form ' 0

of ' by reading o the urrent

domains of the hoie variables C

xy . If '

0

was not D-solved, then !

~

ould

inferanewfat,butthenbyProposition10soould

~

anditwouldnotbea

saturation.

8 Conlusion

In this paper, weextended dominaneonstraintsby admitting set operators.

Set operators introdue a ontrolled form of disjuntion and negation that is

less expressivethan general Boolean onnetives and remains espeially well-

suited for onstraintpropagation. Onthe basis of this extension we presented

twosolvers:oneabstrat,oneonrete.

Thedesignoftheabstratsolverisarefullyinformedbytheneedsofprati-

alappliations:itstipulatesinferenerulesrequiredforeÆientlysolvingdom-

inaneonstraintsourringintheseappliations.Therulestakefulladvantage

oftheextraexpressivityaordedbysetoperators.Weprovedtheabstratsolver

soundandompleteandthatitsdistributionstrategyimprovesover[2℄andmay

avoidan exponential number ofhoiepoints.This improvementaruesfrom

admittinglessexpliitsolvedformswhilepreservingsoundness.

(16)

izesthedesiredonstraintpropagationbyredutiontoonstraintprogramming

usingsetonstraints.Weprovedthattheonretesolverfaithfullysimulatesthe

abstrat one, and thereby shed newlight onthe soure ofits observed prati-

al eetiveness.The onrete solverhas beenimplemented in theonurrent

onstraintprogramminglanguageOz[10℄,performseÆientlyinpratialappli-

ations to semantiunderspeiation, andproduessmallersearhtrees than

thesolverof[2℄.

Referenes

1. R.Bakofen,J.Rogers,andK.Vijay-Shanker. Arst-orderaxiomatizationofthe

theoryofnitetrees. JournalofLogi,Language,andInformation,4:5{39,1995.

2. D.DuhierandC.Gardent. Aonstraint-basedtreatmentofdesriptions. InInt.

WorkshoponComputationalSemantis,Tilburg,1999.

3. D. DuhierandS.Thater. Parsingwithtreedesriptions: aonstraint-basedap-

proah.InInt.WorkshoponNaturalLanguageUnderstandingandLogiProgram-

ming,LasCrues,NewMexio,1999.

4. M.Egg,J.Niehren,P.Ruhrberg,andF.Xu.Constraintsoverlambda-struturesin

semantiunderspeiation. InJointConf.COLING/ACL,pages353{359,1998.

5. C.GardentandB.Webber. Desribingdisoursesemantis. InProeedingsofthe

4thTAG+Workshop,Philadelphia,1998.

6. C.Gervet. Intervalpropagationtoreasonaboutsets:Denitionandimplementa-

tionofapratiallanguage. Constraints,1(3):191{244,1997.

7. A. Koller,K. Mehlhorn, and J. Niehren. A polynomial-time fragment of domi-

naneonstraints. Tehnial report, Programming SystemsLab, Universitat des

Saarlandes,Apr.2000. Submitted.

8. A. Koller,J. Niehren, and R. Treinen. Dominane onstraints:Algorithms and

omplexity.InLogialAspetsofComp.Linguistis98,2000. ToappearinLNCS.

9. M.P.Marus,D.Hindle,andM.M.Flek.D-theory:Talkingabouttalkingabout

trees. In21stACL,pages129{136,1983.

10. Mozart. Themozartprogrammingsystem. http://www.mozart-oz.org/.

11. T. Muller and M. Muller. Finite set onstraints in Oz. In F. Bry, B. Freitag,

and D. Seipel, editors, 13. Workshop Logishe Programmierung, pages 104{115,

TehnisheUniversitatMunhen, 1997.

12. R.Muskens. Order-IndependeneandUnderspeiation. InJ.Groenendijk,edi-

tor,Ellipsis,Underspeiation,EventsandMoreinDynamiSemantis.DYANA

DeliverableR.2.2.C,1995.

13. O.Rambow,K.Vijay-Shanker,andD.Weir. D-treegrammars. InProeedings of

ACL'95,pages151{158,MIT,Cambridge,1995.

14. J.RogersandK.Vijay-Shanker. Reasoningwithdesriptionsoftrees. InAnnual

Meeting oftheAssoiationforComp.Linguistis(ACL),1992.

15. J. W. Thather and J. B. Wright. Generalized nite automatatheory with an

appliation to a deision problem of seond-order logi. Mathematial Systems

Theory,2(1):57{81,August1967.

16. W.Thomas. AutomataonInniteObjets. InJ.v.Leeuwen,editor,Handbook of

TheoretialComputerSiene,FormalModelsandSemantis,volumeB,hapter4,

pages133{191.TheMITPress,1990.

17. K.Vijay-Shanker. Usingdesriptions oftrees inatreeadjoining grammar. Com-

putational Linguistis,18:481{518, 1992.

(17)

Theproof oforretnessofimplementationrelies onshowingthat theonrete

solversimulatestheabstratsolver.Abstratrulesareeither oftheform'!

p

false (Clash rule) or ' !

p

xRy (all other rules). In the rst ase we show

that [['℄℄

false and in the other asesthat [['℄℄

[[xR 0

y℄℄ where R 0

R .

Inotherwords,weprovethat theonretesolverprovidesasmuhorstronger

propagationthantheabstratsolver.

For every rule ' !

p

xRy, it is the ase that x;y 2 Vars('). Therefore,

posing:

A(')= ^

x2Vars(') A

1

(x) ^

x;y2Vars(') A

2 (x;y)

wehaveA(xRy)A('). Thus, in orderto showthat [['℄℄

p [[xR

0

y℄℄, where

R 0

R ,weonlyneedprovethat [['℄℄

p B[[xR

0

y℄℄.

Ourproofsarein naturaldedutionstyleandpresentedinatabularformat

in 3 olumns; left: the formula, enter: its justiation, right: a line name. A

nameoftheform(j `o)introduesthename(o)butalsoexpliitlyreordsthe

fat thatitwasderivedfromthenon-dishargedassumption(j).

(Clash) x;y ! false

C

xy

2; Premise[[x;y℄℄ (a)

false (a) (fd.lash)

(Dom.Re) xC

x

x 2Eq

x

(3) (a)

Eq

x kEq

x

Assumption (b)

false (a)(b) (b `)

D[[x:=x℄℄

~

p

false () (b`d)

D[[x=x℄℄^C

xx

== or C

xx

6==^D [[x:=x℄℄ (10) (e)

C

xx

== (d)(e)

(Dom.Trans) xC

y^yC

z ! xC

z

C

xy

2f=;C +

g Premise[[xC

y℄℄ (a)

C

yz

2f=;C +

g Premise[[yC

z℄℄ (b)

D[[yC +

x℄℄^C

xy

=B +

or C

xy 6=B

+

^D[[y:C +

x℄℄ (10) ()

D[[y:C +

x℄℄ i.e. Eq

y kUp

x

^Down

y kEq

x

(a) ()(ommit) (d)

D[[x?y℄℄^C

xy

=? or C

xy

6=?^D[[x:?y℄℄ (10) (e)

D[[x:?y℄℄ i.e. Eq

x kSide

y

^Side

x kEq

y

(a) (e)(ommit) (f)

D[[zC +

y℄℄^C

yz

=B +

or C

yz 6=B

+

^D[[z:C +

y℄℄ (10) (g)

D[[z:C +

y℄℄ i.e. Eq

z kUp

y

^Down

z kEq

y

(b)(g)(ommit) (h)

D[[y?z℄℄^C

yz

=? or C

yz

6=?^D [[y:?z℄℄ (10) (i)

D[[y:?z℄℄ i.e. Eq

y kSide

z

^Side

y kEq

z

(b)(i)(ommit) (j)

ontinuedonnextpage

(18)

y 2Eq

y

(3) (k)

y 62Up

x

y 62Side

x

y

62Down

z

y 62Side

z

(d)(f)(g)(j)(k) (l)

I=Eqdown

x

℄Up

x

℄Side

x

(4) (m)

y

2Eqdown

x

(l)(m) (n)

D[[zC +

x℄℄ i.e. Eqdown

x

Down

z

Assumption (o)

false (o)(n)(l) (o`p)

D[[zC +

x℄℄

~

p

false (p) (o` q)

D[[zC +

x℄℄^C

xz

=B +

or C

xz 6=B

+

^D [[z:C +

x℄℄ (10) (r)

C

xz 6=B

+

(r) (q)(ommit) (s)

D[[x:?z℄℄ i.e. Eqdown

x Side

z

Assumption (t)

false (t) (n)(l) (t`u)

D[[x:?z℄℄

~

p

false (u) (t` v)

D[[x?z℄℄^C

xz

=? or C

xz

6=?^D[[x:?z℄℄ (10) (w)

C

xz

6=? (w) (v)(ommit) (x)

C

xz

2f=;C +

g (s) (x)

(Eq.Deom) x:f(x

1

;:::;x

n

)^y:f(y

1

;:::;y

n

)^x=y ! x

i

=y

i

[[x=y℄℄ i.e. C

xy

== Premise (a)

[[x:f(x

1

;:::;x

n

)℄℄ i.e. Down i

x

=Eqdown

x

i

Premise (b)

[[y:f(y

1

;:::;y

n

)℄℄ i.e. Down i

y

=Eqdown

yi

Premise ()

D[[x=y℄℄^C

xy

== or C

xy

6==^D[[x:=y℄℄ (10) (d)

D[[x=y℄℄ i.e. Down i

x

=Down i

y

(d)(a)(ommit)(e)

Eqdown

x

i

=Eqdown

y

i

(e)(b)() (f)

x

i 2Eq

xi

(3) (g)

x

i

2Eqdown

xi

(g)(6) (h)

xi

2Eqdown

y

i

(h)(f) (i)

x

i

2Equp

xi

(g)(7) (j)

I=Equp

xi

℄Down

x

i

℄Side

x

i

(5) (k)

x

i

62Down

x

i

(j)(k) (l)

Eqdown

yi

Down

xi

Assumption (m)

x

i

2Down

x

i

(m)(i) (m` n)

false (l)(m) (m`o)

D[[x

i C

+

y

i

℄℄

~

p

false (o) (m` p)

D[[x

i C

+

y

i

℄℄^C

xiyi

=C +

or C

xiyi 6=C

+

^D [[x

i :C

+

y

i

℄℄ (10) (q)

D[[x

i :C

+

y

i

℄℄ i.e. Down

xi kEq

y

i

(q) (p)(ommit) (r)

y

i 2Eq

yi

(3) (s)

y

i

2Eqdown

yi

(s)(6) (t)

yi

2Eqdown

x

i

(t)(f) (u)

yi

62Down

xi

(s)(r) (v)

yi 2Eq

x

i

(u)(v) (w)

Eq

x

i kEq

yi

Assumption (x)

ontinuedonnextpage

(19)

false (s)(w) (x) (x` y)

D[[x

i :=y

i

℄℄

~

p

false (y) (x`z1)

D[[x

i

=y

i

℄℄^C

xiyi

== or C

xiyi

6==^D[[x

i :=y

i

℄℄ (10) (z2)

C

xiyi

== (z1)(z2)(ommit)

(Lab.Ineq) x:f(:::)^y:g(:::) ! x:=y iff6=g

[[x:f(:::)℄℄ i.e. Label

x

=

f

Premise (a)

[[y:g(:::)℄℄ i.e. Label

y

=

g

Premise (b)

Label

x

=Label

y

Assumption ()

false ()f6=g ( `d)

D[[x=y℄℄

~

p

false (d) ( `e)

D[[x=y℄℄^C

xy

== or C

xy

6==^D[[x:=y℄℄ (10) (f)

C

xy

6== (e)(f)(ommit)

(Lab.Disj) x:f(::: x

i ::: x

j

:::) ! x

i

?x

j

[[x:f(::: x

i ::: x

j

:::)℄℄ i.e. Down i

x

=Eqdown

xi

Premise (a)

[[x:f(::: x

i ::: x

j

:::)℄℄ i.e. Down j

x

=Eqdown

xj

Premise (b)

[[x:f(::: x

i ::: x

j

:::)℄℄ i.e. Up

xj

=Equp

x

Premise ()

x

i 2Eq

xi

(3) (d)

x

i

2Eqdown

xi

(d)(6) (e)

x

i

2Down i

x

(e)(a) (f)

Down

x

=℄

k Down

k

x

(8) (g)

x

i

62Down i

x

(g)(f) (h)

x

i

62Eqdown

x

j

(h)(b) (i)

x

i

2Down

x

(g)(f) (j)

I=Equp

x

℄Down

x

℄Side

x

(5) (k)

x

i

62Equp

x

(j)(k) (l)

x

i 62Up

xj

(l)() (m)

I=Eqdown

xj

℄Up

xj

℄Side

x

j

(4) (n)

x

i 2Side

x

j

(n)(i)(m) (o)

Eq

xi kSide

x

j

Assumption (p)

false (p)(d)(o) (p` q)

D[[x

i :?x

j

℄℄

~

p

false (q) (p`r)

D[[x

i

?x

j

℄℄^C

x

i x

j

=? or C

x

i x

j

6=?^D[[x

i :?x

j

℄℄ (10) (s)

C

xixj

=? (s) (r)(ommit)

(Lab.Dom) x:f(::: y

i

:::) ! xC +

y

i

[[x:f(::: y

i

:::)℄℄ i.e. Up

yi

=Equp

x

Premise (a)

x 2Eq

x

(3) (b)

x

2Equp

x

(b)(7) ()

x 2Up

yi

()(a) (d)

Eq

x kUp

y

i

Assumption (e)

ontinuedonnextpage

(20)

false (e) (b)(d) (e` f)

D[[x:C +

y

i

℄℄

~

p

false (f) (e `g)

D[[xC +

y

i

℄℄^D[[xy

i

℄℄=C +

or C

xyi 6=C

+

^D [[x:C +

y

i

℄℄ (10) (h)

C

xyi

=C +

(h)(f)(ommit)

(Inter) xR

1

y^xR

2

y ! xRy RR

1

\R

2

[[xR

1

y℄℄ i.e. C

xy 2R

1

Premise (a)

[[xR

2

y℄℄ i.e. C

xy 2R

2

Premise (b)

C

xy 2R

1

\R

2

(a)(b)(fd.onj) ()

C

xy

2R ()RR

1

\R

2

Forrule(Inv)xRy!yR 1

x,weneedtoshowthat[[xRy℄℄

~

p [[yR

0

x℄℄,

whereR 0

R 1

. Weshowthisbyprovingthatwheneverr62R 1

wealsohave

r62R 0

.Wemustonsiderthefollowing4ases:

(ase =62R 1

) [[x:=y℄℄

~

p

[[y:=x℄℄ (i.e. =62R 0

)

(aseB +

62R 1

) [[x:C +

y℄℄

~

p

[[y:B +

x℄℄ (i.e.B +

62R 0

)

(aseC +

62R 1

) [[x:B +

y℄℄

~

p

[[y:C +

x℄℄ (i.e.C +

62R 0

)

(ase ?62R 1

) [[x:?y℄℄

~

p

[[y:?x℄℄ (i.e. ?62R 0

)

(Inv 1) [[x:=y℄℄

~

p

[[y:=x℄℄

C

xy

6== Premise (a)

D[[x=y℄℄^C

xy

== or C

xy

6==^D[[x:=y℄℄ (10) (b)

D[[x:=y℄℄ i.e. Eq

x kEq

y

(b)(ommit) (a `)

x 2Eq

x

(3) (d)

x 62Eq

y

()(d) (e)

Eq

y

=Eq

y

Assumption (f)

false (d)(e)(f) (f `g)

D[[y=x℄℄^C

yx

== or C

yx

6==^D[[y:=x℄℄ (10) (h)

C

yx

6== (h)(g)(ommit)

(Inv 2) x:C

+

y

~

p

y:B +

x

C

xy 6=C

+

Premise (a)

D[[xC +

y℄℄^C

xy

=C +

or C

xy 6=C

+

^D[[x:C +

y℄℄ (10) (b)

D[[x:C +

y℄℄ i.e. Eq

x kUp

y

(a) (b)(ommit) ()

x 2Eq

x

(3) (d)

x

2Equp

x

(d)(7) (e)

x 62Up

y

(d)() (f)

Equp

x Up

y

Assumption (g)

x 2Up

y

(g) (e) (g`h)

false (h)(f) (g`i)

D[[xC +

y℄℄

~

p

false (i) (g `j)

ontinuedonnextpage

(21)

D[[xC +

y℄℄^C

yx

=B +

or C

yx 6=B

+

^D[[x:C +

y℄℄ (10) (k)

C

yx 6=B

+

(k) (j)(ommit)

(Inv 3) x:B

+

y

~

p

y:C +

x

C

xy 6=B

+

Premise (a)

D[[yC +

x℄℄^C

xy

=B +

or C

xy 6=B

+

^D[[y:C +

x℄℄ (10) (b)

D[[y:C +

x℄℄ i.e. Eq

y kUp

x

(a) (b)(ommit) ()

y 2Eq

y

(3) (d)

y Equp

y

(d)(7) (e)

y 62Up

x

(d)() (f)

Equp

y Up

x

Assumption (g)

y 2Up

x

(g) (e) (g`h)

false (h)(f) (g`i)

D[[yC +

x℄℄

~

p

false (i) (g `j)

D[[yC +

x℄℄^C

yx

=C +

or C

yx 6=C

+

^D[[y:C +

x℄℄ (10) (k)

C

yx 6=C

+

(k) (j)(ommit)

(Inv 4) x:?y

~

p

y:?x

C

xy

6=? Premise (a)

D[[x?y℄℄^C

xy

=? or C

xy

6=?^D[[x:?y℄℄ (10) (b)

D[[x:?y℄℄ i.e. Eq

x kSide

y

(a)(b)(ommit) ()

x 2Eq

x

(3) (d)

x 62Side

y

(d)() (e)

x

2Eqdown

x

(d)(6) (f)

Eqdown

x Side

y

Assumption (g)

x 2Side

y

(g)(f) (f`h)

false (h)(e) (f`i)

D[[y?x℄℄

~

p

false (i) (f `j)

D[[y?x℄℄^C

yx

=? or C

yx

6=?^D[[y:?x℄℄ (10) (k)

C

yx

6=? (k)(j)(ommit)

(Disj) x?y^yC

z ! x?z

[[x?y℄℄ i.e. C

xy

=? Premise (a)

[[yC

z℄℄ i.e. C

yz

2f=;C +

g Premise (b)

D[[zC +

y℄℄^C

yz

=B +

or C

yz 6=B

+

^D[[z:C +

y℄℄ (10) ()

D[[z:C +

y℄℄ i.e. Eq

z kUp

y

(b)()(ommit) (d)

D[[y?z℄℄^C

yz

=? or C

yz

6=?^D [[y:?z℄℄ (10) (e)

D[[y:?z℄℄ i.e. Eq

z kSide

y

(e) (b)(ommit) (f)

z 2Eq

z

(3) (g)

z 62Up

y

(g)(d) (h)

z 62Side

y

(g)(f) (i)

I=Eqdown

y

℄Up

y

℄Side

y

(4) (j)

ontinuedonnextpage

(22)

z

2Eqdown

y

(j)(h)(i) (k)

D[[x?y℄℄^C

xy

=? or C

xy

6=?^D[[x:?y℄℄ (10) (l)

D[[x?y℄℄ i.e. Eqdown

y Side

x

(l)(a)(ommit) (m)

z 2Side

x

(k)(m) (n)

Side

x kEq

z

Assumption (o)

false (o)(n)(g) (o`p)

D[[x:?z℄℄

~

p

false (p) (o` q)

D[[x?z℄℄^C

xz

=? or C

xz

6=?^D[[x:?z℄℄ (10) (r)

C

xz

=? (r)(q) (ommit)

Lemma5. [[xC

y℄℄

~

p

y

2Eqdown

x

(Lemma 5) [[xC

y℄℄

~

p

y2Eqdown

x

[[xC

y℄℄ i.e. C

xy

2f=;C +

g Premise (a)

D[[yC +

x℄℄^C

xy

=B +

or C

xy 6=B

+

^D[[y:C +

x℄℄ (10) (b)

D[[y:C +

x℄℄ i.e. Eq

y kUp

x

(a) (b)(ommit) ()

D[[x?y℄℄^C

xy

=? or C

xy

6=?^D[[x:?y℄℄ (10) (d)

D[[x:?y℄℄ i.e. Eq

y kSide

x

(a) (d)(ommit) (e)

y 2Eq

y

(3) (f)

y 62Up

x

(f)() (g)

y 62Side

x

(f)(e) (h)

I=Eqdown

x

℄Up

x

℄Side

x

(4) (i)

y

2Eqdown

x

(i)(g)(h)

(NegDisj) xC

z^yC

z ! x:?y

C

xz

2f=;C +

g Premise (a)

C

yz

2f=;C +

g Premise (b)

z

2Eqdown

x

(a)(Lemma5) ()

z

2Eqdown

y

(b)(Lemma5) (d)

I=Eqdown

y

℄Up

y

℄Side

y

(4) (e)

z 62Side

y

(d)(e) (f)

Eqdown

x Side

y

Assumption (g)

z 2Side

y

()(g) (g`i)

false (f)(i) (g `j)

D[[x?y℄℄

~

p

false (j) (g` k)

D[[x?y℄℄^C

xy

=? or C

xy

6=?^D[[x:?y℄℄ (10) (l)

C

xy

6=? (l)(k)(ommit)

Lemma6. [[x:C

y℄℄

~

p

y

62Eqdown

x

(Lemma 6) [[x:C

y℄℄

~

p

y

62Eqdown

x

[[x:C

y℄℄ i.e. C

xy 2fB

+

;?g Premise (a)

D[[x=y℄℄^C

xy

== or C

xy

6==^D[[x:=y℄℄ (10) (b)

ontinuedonnextpage

(23)

D[[x:=y℄℄ i.e. Eq

x kEq

y

(a) (b)(ommit) ()

D[[xC +

y℄℄^C

xy

=C +

or C

xy 6=C

+

^D[[x:C +

y℄℄ (10) (d)

D[[x:C +

y℄℄ i.e. Down

x kEq

y

(a) (d)(ommit) (e)

y 2Eq

y

(3) (f)

y 62Eq

x

(f)() (g)

y

62Down

x

(e) () (h)

y

62Eqdown

x

(g) (h)(6)

(Child.up) xC

y^x:f(x

1

;:::;x

n )^

V

n

i=1 x

i :C

y ! x=y

C

xy

2f=;C +

g Premise (a)

[[x:f(x

1

;:::;x

n )℄℄ i.e.

Down i

x

=Eqdown

x

i

1in

Down i

x

=; i>n

Premise (b)

C

x

i y

2fB +

;?g Premise ()

y

62Eqdown

xi

()(Lemma6) (d)

y

62Down i

x

(d)(b) (e)

Down

x

=℄

i Down

i

x

(8) (f)

y

62Down

x

(e)(f) (g)

y

2Eqdown

x

(a)(Lemma5) (h)

y 2Eq

x

(g)(h)(6) (i)

y 2Eq

y

(3) (j)

Eq

x kEq

y

Assumption (k)

false (i)(j)(k) (k`l)

D[[x:=y℄℄

~

p

false (l) (k`m)

D[[x=y℄℄^C

xy

== or C

xy

6==^D[[x:=y℄℄ (10) (n)

C

xy

== (n)(m)(ommit)

Lemma7.

x

2Eqdown

x

(Lemma 7)

x

2Eqdown

x

x 2Eq

x

(3) (a)

Eqdown

x

=Eq

x

℄Down

x

(6) (b)

x

2Eqdown

x

(a)(b)

Lemma8. [[xC +

y℄℄

~

p

y

2Down

x

(Lemma 8) [[xC

+

y℄℄

~

p

y

2Down

x

C

xy

=C +

Premise (a)

C

xy 6=C

+

Assumption (b)

false (a) (b) (b `)

D[[xC +

y℄℄^C

xy

=C +

or C

xy 6=C

+

^D[[x:C +

y℄℄ (10) (d)

D[[xC +

y℄℄ i.e. Eqdown

y

Down

x

() (d)(ommit) (e)

y

2Eqdown

y

(Lemma 7) (f)

ontinuedonnextpage

(24)

y

2Down

x

(e)(f)

Lemma9.

x 62Side

x

(Lemma 9)

x 62Side

x

x 2Eq

x

(3) (a)

Eqdown

x

=Eq

x

℄Down

x

(6) (b)

x

2Eqdown

x

(a)(b) ()

I=Eqdown

x

℄Up

x

℄Side

x

(4) (d)

x 62Side

x

()(d)

Lemma10.

x

62Down

x

(Lemma 10)

x

62Down

x

x 2Eq

x

(3) (a)

Eqdown

x

=Eq

x

℄Down

x

(6) (b)

x

62Down

x

(a)(b)

Lemma11.

y

2Eqdown

x

~

p C

xy

2f=;C +

g

(Lemma 11)

y

2Eqdown

x

~

p C

xy

2f=;C +

g

y

2Eqdown

x

Premise (a)

y 62Side

y

(Lemma 9) (b)

Eqdown

x Side

y

Assumption ()

y 2Side

y

(a) () ( `d)

false (d) ( `e)

D[[x?y℄℄

~

p

false (e) (` f)

D[[x?y℄℄^C

xy

=? or C

xy

6=?^D[[x:?y℄℄ (10) (g)

C

xy

6=? (f)(g) (h)

y

62Down

y

(Lemma 10) (i)

Eqdown

x

Down

y

Assumption (j)

y

2Down

y

(a) (j) (j` k)

false (i)(k) (j`l)

D[[yC +

x℄℄

~

p

false (l) (j `m)

D[[yC +

x℄℄^C

yx

=C +

or C

yx 6=C

+

^D[[y:C +

x℄℄ (10) (n)

C

yx 6=C

+

(m) (n)(ommit) (o)

C

xy :=B

+

(o) (Inv) (p)

C

xy

2f=;C +

g (h)(p)

(Child.down) xC +

y^x:f(x

1

;:::;x

n )^^

n

i=1;i6= j x

i :C

y ! x

j C

y

[[xC +

y℄℄ Premise (a)

[[x:f(x

1

;:::;x

n )℄℄ i.e.

Down i

x

=Eqdown

xi

1in

Down i

x

=; i>n

Premise(13) (b)

ontinuedonnextpage

(25)

[[x

i :C

y℄℄ 1i6=jn Premise ()

y

2Down

x

(a)(Lemma8) (d)

y

62Down i

x

i>n (b) (e)

y

62Eqdown

x

i

1i6=j n ()(Lemma6) (f)

y

62Down i

x

1i6=j n (f)(b) (g)

Down

x

=℄ fDown i

x

j1imaxg (8) (h)

y

2Down j

x

(h)(d)(e)(g) (i)

y

2Eqdown

x

j

(i)(b) (j)

C

x

j y

2f=;C +

g (j)(Lemma11)

Lemma12. [[x:?y℄℄

~

p

y 62Side

x

(Lemma 12) [[x:?y℄℄

~

p

y 62Side

x

[[x:?y℄℄ i.e. C

xy

6=? Premise (a)

D[[x?y℄℄^C

xy

=? or C

xy

6=?^D[[x:?y℄℄ (10) (b)

C

xy

=? Assumption ()

false (a)() ( `d)

D[[x:?y℄℄ i.e. Eq

y kSide

x

(d)(ommit) (e)

y 2Eq

y

(3) (f)

y 62Side

x

(f)(e)

Areallyimportantlemma thatIshould haveprovenmuh earlier

Lemma13. [[xry℄℄

~

p

D[[xry℄℄

(Lemma 13) [[xry℄℄

~

p

D[[xry℄℄

[[xry℄℄ i.e. C

xy

=r Premise (a)

C

xy

6=r Assumption (b)

false (a)(b) (b `)

D[[xry℄℄^C

xy

=r or C

xy

6=r^D[[x:ry℄℄ (10) (d)

D[[xry℄℄ ()(d)(ommit)

(NegDom) x?y^y:?z ! x:C

z

[[x?y℄℄ Premise (a)

[[y:?z℄℄ Premise (b)

z 62Side

y

(b)(Lemma12) ()

Eq

x

=Eq

z

Assumption (d)

z 2Eq

z

(3) (e)

z 2Eq

x

(d)(e) (d` f)

Eqdown

x

=Eq

x

℄Down

x

(6) (g)

z

2Eqdown

x

(f)(g) (d`h)

D[[x?y℄℄ i.e. Eqdown

x Side

y

(a)(Lemma13) (i)

z 2Side

y

(h)(i) (d`j)

ontinuedonnextpage

Références

Documents relatifs

Starting from the mechanical description of the dynamics of discrete mechanical systems subjected to frictionless time-dependent unilateral constraints, we have derived a

Source: Author’s structure-World Bank Data.. Total global assets growth in 2017 was significantly less at around 4.74 percent. Even though strong growth was witnessed in 2018.

In the classical limit, however, p,q are the canonical momenta and coordinates, £i(p,q) the classical dynamical quantity described by the operator fi, and f(p,q) the classical

In his proof, these elliptic modular forms arose form Jacobi forms, which he did not consider directly.. The key to our result is not to pass to elliptic modular forms, but to

are made based on a waveguide filled with ferrite which under certain conditions which means that when a constant magnetic field applied to a ferrite when a

While the con- tact mechanics analysis is simplified, they are nevertheless commonly accepted and experimentally validated models for calculating contact pressure distributions (P˜

Keywords— Pascal matrices, Binomial Recursion, Circulant Oper- ators, Square Matrix Bipartition, Local Network, Parallel networks of agents..

In the following, based on an analogy with (max,+)-linear systems, we extend transfer function matrix models and model reference control to (max, +)-linear systems with