• Aucun résultat trouvé

An Efficient Algorithm for the Configuration Problem of Dominance Graphs

N/A
N/A
Protected

Academic year: 2021

Partager "An Efficient Algorithm for the Configuration Problem of Dominance Graphs"

Copied!
11
0
0

Texte intégral

(1)

HAL Id: inria-00536803

https://hal.inria.fr/inria-00536803

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 abroad, or from public or private research centers.

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 publics ou privés.

An Efficient Algorithm for the Configuration Problem of Dominance Graphs

Ernst Althaus, Denys Duchier, Alexander Koller, Kurt Mehlhorn, Joachim Niehren, Sven Thiel

To cite this version:

Ernst Althaus, Denys Duchier, Alexander Koller, Kurt Mehlhorn, Joachim Niehren, et al.. An Efficient

Algorithm for the Configuration Problem of Dominance Graphs. Proceedings of the 12th ACM-

SIAM Symposium on Discrete Algorithms, 2001, Washington, DC, United States. pp.815–824. �inria-

00536803�

(2)

of Dominane Graphs 1

Ernst Althaus 2

Denys Duhier 3

Alexander Koller 4

Kurt Mehlhorn 2

JoahimNiehren 3

Sven Thiel 2

Abstrat

Dominaneonstraintsarelogialtreedesriptionsoriginat-

ingfromautomatatheorythathavemultipleappliationsin

omputationallinguistis. Thesatisabilityproblemofdom-

inane onstraints is NP-omplete. In most appliations,

however, onlynormal dominaneonstraintsareused. The

satisability problem of normaldominane onstraints an

be redued in linear time to the onguration problem of

dominanegraphs,asshownreently.Inthispaper,wegive

apolynomialtimealgorithmtestingongurabilityofdom-

inanegraphs (andthus satisability of normaldominane

onstraints). Previoustoourworknopolynomialtimealgo-

rithmswereknown.

1 Introdution

Thedominanerelationofatreeistheanestorrelation

between its nodes. Dominane onstraints are logial

desriptionsof treestalkingaboutthedominane rela-

tion. Dominanebasedtreedesriptionswererstused

inautomatatheoryinthesixties[TW67℄andredisov-

ered in omputational linguistis in the early eighties

[MHF83℄. Sinethen,theyhavefoundnumerousappli-

ations: they have been used for grammar formalisms

[VS92, RVSW95℄, insemantis[Mus95,ENRX98℄, and

fordisourseanalysis [GW98℄.

Thesatisabilityproblemofdominaneonstraints

isNP-omplete[KNT98℄. Earlierattempts at proess-

ingdominaneonstraints[Cor94,VSWR95,DN00℄all

suerfromthisfat. Butitturnsoutthatnormaldom-

inaneonstraints, a restrited sublanguage, are suÆ-

ient for most appliations. The starting point of the

graphbasedapproahofthispaperisanotherreentre-

sult[KMN00℄showingthatthesatisabilityproblemof

normaldominaneonstraintsanbereduedinlinear

timetotheongurationproblemofdominanegraphs.

Informally,adominane graph isgivenbyaolle-

tionof rootedtreesand aset ofdominanewishes. (A

1

PartiallysupportedbytheISTProgrammeoftheEUunder

ontratnumberIST-1999-14186(ALCOM-FT)

2

Max-Plank-Institute for Computer Siene, Saarbruken,

Germany

3

ProgrammingSystemsLab,FahbereihInformatik,Univer-

sitatdesSaarlandes,Saarbruken,Germany

4

Department of Computational Linguistis, Universitat des

Saarlandes,Saarbruken,Germany

preise denition follows in Setion 2.) A dominane

wish is a direted edge from the leaf of some tree to

therootofsomeothertree. A onguration ofadomi-

nanegraphisobtainedbyassemblingthetreesof the

graphinto a forest, by hooking roots into leaves suh

thatalldominanewishesaretranslatedinto anestor-

desendantrelationships. Theongurationproblem of

dominane graphs is thequestionwhether there exists

aongurationforagivendominanegraph.

Inthispaper,weshowthattheongurationprob-

lem of dominanegraphs is in polynomial time. This

resultimmediatelyleadsthewayforapolynomialtime

andpratiallymoreeÆientproessingofnormaldom-

inaneonstraintsin omputationallinguistis.

Togetanideaofhowongurationsofdominane

graphsariseinlinguistis,onsiderthe[ENRX98℄anal-

ysisofthefollowingEnglishsentene:

(1.1) Everylinguistspeakstwolanguages.

a. ...,namelyEnglishandGerman.

b. ...,notneessarilythesameones.

Depending on theontext (indiatedby the ontinua-

tionsa. and b.), this sentene anberead in twodif-

ferentways{itexhibitsasopeambiguity. Itanmean

eitherthatthereisasetoftwolanguagesspokenbyev-

erylinguist, oritanmeanthat eah linguistanpik

hisownpairoflanguages.

Thisambiguityanberepresentedompatlybythe

graphin Figure 1, whih wean read asa dominane

graph by removing the node labels. Intuitively, the

ontributions of \every linguist" and \two languages"

to the meaning of the sentene are represented as

the two upper trees; the ontribution of \speaks" is

representedas the lower one. The tree ongurations

of this dominane graphare obtainedby plugging the

twouppertreesinsomeorderontoponthelowertree.

The two waysto arrangethem orrespond to the two

dierentreadingsof(1.1).

Our paperisorganizedasfollows. InSetion 2we

denethetermsdominanegraph,solvedform,andon-

(3)

every ling

x

two

speak y

x lang

y

Figure1: Dominanegraphforasopeambiguity.

the sueedingsetions. InSetion 3, weshowhowto

enumerate all ongurations of a dominane graph in

exponentialtime;thisprovidesaframeworkfortheap-

pliationofthelaterresults. InSetion4,weharater-

ize ongurabledominanegraphs(adominanegraph

hasaongurationiitontainsnohypernormalyle),

and then weshowin Setion 5howthe existene of a

hypernormalyleanbedeidedbysolvingaweighted

mathing problemin anauxiliarygraph. Thisgivesus

apolynomial-timeongurabilitytest whih weuse in

Setion6tomaketheenumerationalgorithmfromSe-

tion3eÆient. Setion 7showsthat aslightextension

of the ongurationproblem by losed leaves is again

NP-omplete. Finally,weoerashort onlusion.

2 Denitions

Adominanegraph isdenedbyadiretedgraphG=

(V;E _

[D) satisfyingthe followingtwoonditions: (1)

the graph G = (V;E) denes 5

a olletion T of node

disjointtrees of height at least 1and (2) eah edge in

D goesfromaleafofsometreein theolletiontothe

root of some tree in the olletion. In our gures, we

drawthe edgesin E solid and the edgesin D dashed.

WealltheedgesinE solidedges ortreeedges andwe

alltheedgesin Ddashededges ordominaneedges or

dominane wishes. A leaf is a node with no outgoing

tree edge and a root is a node with no inoming tree

edge.

Nowtheideaisthatwewanttoassemblethetrees

in T by plugging roots into leaves. We say that a

dominane graphG is in solved form i it is a forest.

If G = (V;E _

[D) is a dominane graph, we all a

dominane graph G 0

= (V 0

;E 0

_

[D 0

) a solved form of

G i V = V 0

, E = E 0

, G 0

is in solved form, and G 0

realizesall dominane wishesin G {that is, for every

dominanewish(v;w)2D thereisapathfromv tow

in G 0

.

Inpartiular,weall asolvedformofGwherethe

dominaneedgesinD 0

areamathingaongurationof

5

G. Rootshaveatmostoneinomingdominaneedgein

ongurations;theintuitionisthattherootshavebeen

\plugged"intotheleaves,andtheremainingdominane

edgesindiate whihrootispluggedintowhihleaves.

A dominane graphis ongurable ifit has aon-

gurationandsolvable ifithasasolvedform. Figure2

showsaongurablegraphandoneonguration. Fig-

ure3displaysanunongurablegraph,theheavyedges

indiatean\unongurableyle",asweshallseelater.

1

3 4

2

1 2

3 4

Figure 2: A ongurable dominanegraphand aon-

gurationofit.

Figure 3: An unongurable dominane graph; the

(4)

l l

0

r

t t

0

z

1

zk l

0

r

l t t

0

z1

z

k

Figure4: AppliationofRule1: All dominanewishes

ofl 0

exeptfor(l 0

;r)areshifteddowntotheleafl.

Theproblemweinvestigateinthispaperistodeide

whether agiven dominanegraphhas aonguration.

Morepreisely,wearegoingtoonsidertheproblemof

whetherit hasasolvedform; butthe followinglemma

expressesthat thisisthesameproblem.

Lemma2.1. Every dominane graph in solved form is

ongurable.

Conversely, every ongurable graph is trivially solv-

able.

Proof. For the proof, we dene a problem leaf to be

a leaf with more than one outgoing dominane edge;

ouraimwillbetoeliminateproblemleavesfromsolved

forms.

Theproofisbyindutiononweights(d;a)ofgraphs

G,wheredisthenegativeminimumdepthofaproblem

leafofG(or 1iftherearen'tany),andaisthetotal

number of dominane edges emanating from problem

leavesof minimum depth (potentially0). Weonsider

thelexiographiorderonthese weights.

Solved forms without problem leaves (i.e. with

weight ( 1;0)) are ongurations, so the lemma is

trivially true in this ase. So let G be a solved form

thatdoeshaveproblemleaves.LetGhaveweight(d;a),

andassumethatweknowthatallsolvedformsoflower

weightdo haveongurations. Then weanapply the

followingruletoaproblemleafl 0

ofminimumdepth:

Simplifiation Rule1. Let e = (l 0

;r) be a domi-

nane edge from the leaf l 0

of a tree t 0

to the root r

of atreet. Letl be an arbitrary leaf of t. Change any

dominane edge (l 0

;z) with z 6= r into (l;z), see Fig-

ure4.

TheresultG 0

isstill insolvedform,anditsweight

hypothesis, G has a onguration G

. But G

also

realizesall dominanewishesofG. This isobviousfor

(l 0

;r)and forallwisheswhih donotstartin l 0

. Fora

wish(l 0

;z)withz6=rwenotethatthiswishisrealized

beausethereisapathfroml 0

tolinG 0

andG

realizes

thewish(l;z). SoGhasaongurationaswell. ut

Finally, we all a dominane wish d = (v;w)

redundant if there is a path from v to w in Gnd.

A dominane graph is alled redued if it ontains no

redundantdominanewish. Asusual, weusenand m,

respetively, to denote thenumber ofnodes and edges

ofG.

Inthefollowingsetionsweshow:

Congurabilityofadominanegraphhasasimple

haraterization.

Congurability of a dominane graph an be de-

idedinpolynomialtime. Morepreisely,itanbe

deided by solving a weighted mathing problem

in an auxiliary graphwith n 0

= O(m) nodes and

m 0

= P

v2V indeg

2

v

edges;hereindeg

v

isthedegree

ofvin G. Themathingproblemanbesolvedin

timeO(n 0

m 0

logn 0

),see[GMG86℄.

Asolvedformofa(ongurable)dominanegraph

an be onstruted in polynomial time. More

preisely,itanbefoundin timeO(n 2

n 0

m 0

logn 0

).

All solved forms of a dominane graph an be

enumeratedin polynomial time per onguration.

Morepreisely,ifN denotes thenumberof solved

formsthenallongurationsanbeenumeratedin

time O((N +1)T), where T is the time to nd a

singleone.

OurtheoretialresultsleadtoapratiallyeÆient

algorithmforhandlingdominanegraphs. Theal-

gorithmhasbeenimplemented. Inourappliation,

wehavem=O(n),and P

v2V indeg

2

v

=O(n). The

existaneofaongurationanthereforebetested

in time T = O(n 2

logn), and a solved form an

be onstruted in time O(n 2

T). The atual run-

ning times are smaller sine the arising weighted

mathingproblemsseemtobefairlysimpleandthe

number of mathing problems to be solved seems

tobemuhlessthann 2

. Ourimplementationuses

LEDA [MNSU ,MN99℄ and themathing odes of

T.ZieglerandG.Shafer[Zie95,Sh00℄.

3 Enumerationof Solved Forms

Inthis setion, we show howto enumerate the solved

forms of a dominane graph G. The algorithm we

present may take exponential time to produe even a

(5)

ases. In Setion 5, we will present a polynomial al-

gorithm for determining ongurability. By plugging

this algorithm intothe enumerationalgorithm,wean

enumerate ongurations in polynomial time per on-

guration(Setion6).

The enumeration algorithm applies the following

simpliationrules:

Simplifiation Rule 2. (Redundany Elim.) All

redundantdominaneedges,i.e. edges that are implied

by transitivity, an be removed. In partiular, parallel

edges an beombinedintoone.

Simplifiation Rule 3. (Choie) Let v be a root

with atleasttwo inomingdominane wishes (l;v) and

(l 0

;v)andletrandr 0

betherootsofthetreesontaining

leaves l and l 0

, respetively. Generate two new graphs

H and H 0

by adding either (l 0

;r) or (l;r 0

) to D, see

Figure5.

Theenumerationofthesolvedformsanbearried

outbyareursivealgorithm:

1. Makethegraphredued,i.e. applyRule2.

2. Ifthegraphontainsa(direted) yle,terminate

thisreursionsinethegraphhasnoonguration.

3. If the graph is in solved form, report it and

terminatethisreursion.

4. Otherwise, apply the hoie rule and apply the

algorithmtothetwonewlygeneratedgraphs.

Every solved form derived by the algorithm is

learly a solved form of the original graph. On the

other hand, thealgorithm enumeratesall of itssolved

forms. This is beause appliation of Rule 2 to a

dominane graph does not hange the set of solved

forms. Appliation of Rule 3 partitions the set: The

twonewgraphshavedisjointsets of solved forms,but

the union of these sets is the same as the old set of

solved forms. This anbe seenasfollows. Inasolved

formG

s

ofGthenodeslandl 0

arebothanestorsofv

andthereforeeitherl 0

isanestoroflandheneofror

vieversa. Thisimpliesthat G

s

iseitherasolvedform

ofH orofH 0

.

To prove termination, we derive an upper bound

for the maximum reursion depth. Consider for any

dominanegraphGitsreahabilityrelationR

G

{theset

ofallpairs(u;v)ofnodessuhthatthereisa(direted)

path from uto v in G. If G isayli, theardinality

of R

G

is at most n

2

n 2

. Thus, whenever the size

ofthe relationbeomesgreaterthan n

2

,thereursion

relationinreasesstritly, i.e. jR

G

j<min(jR

H j;jR

H 0j).

This is beauseR

H

R

G

, and (l 0

;r) 2R

H but (l

0

;r)

annot be in R

G

, otherwise (l 0

;v) would have been

redundant. AsimilarargumentholdsforH 0

.

4 A Graph-Theoreti Charaterization of

Solvability

Wegiveagraphtheoretiharaterizationofsolvability;

asthisisequivalenttoongurabilitybyLemma2.1,the

resultarriesoverto ongurability. Theharateriza-

tionimpliesthatthesolvabilityproblemfordominane

graphsisin NP\o-NP.

Theundireteddominanegraph G

u

orresponding

tothedominanegraphG=(V;E _

[D)istheundireted

graphobtainedbymakingalledgesofGundireted.

Now, we want to dene the notion of a yle in an

undireted graph, whih may dier from the reader's

usual notion. A yle C in an undireted graph is a

sequene of edges e

0 Æe

1

Æ:::Æe

n 1

with n> 1 suh

that fori=0;:::;n 1thefollowingholds:

thereisanodev

i

inidenttobothe

i ande

(i+1)modn

e

i 6=e

(i+1)modn

We allC edge-simple iftheedges in thesequeneare

pairwisedierent. Cissaidtobesimpleifallthevisited

nodesv

0

;:::;v

n 1

arepairwisedierent.

4.1 HypernormalDominane Graphs

Let us rst investigate a simpler subproblem of the

solvabilityproblem. AdominanegraphG=(V;E _

[D)

is hypernormal if for every leaf l in (V;E) there is at

most one dominane wish (l;:) in D. Hypernormal

dominanegraphsareredued 6

.

Proposition4.1. Let G=(V;E _

[D)bea hypernor-

maldominanegraph. IfG

u

ontainsaylethenGis

unsolvable.

Proof. Theproofisbyindutionontheminimalnumber

kofdominaneedgesinasimpleyleCofG

u

. Clearly,

theasek=0annotour,andifk=1thenGisnot

solvable. Ontheother hand,assumethatweknowthe

result to betrue for k 1. C either does not ontain

anynodesatwhihits edgeshangediretions; thenit

is alsoayle inG andhene,G islearlyunsolvable.

OrC doeshangediretions, thenitmustontaintwo

dominane edges (l;r) and (l 0

;r) into the same root.

Bothresultsofapplyingthehoieruleproduegraphs

6

Analternativedenitionofhypernormalgraphsisasfollows:

Out of two dominane wishes (l ;v) and (l ;v 0

) at least one is

redundant.

(6)

l l 0

v

l l

0 r

v

r r r

l l

0

r r

v

Figure5: Twographsare generatedbyapplyingthehoierule tothegraphonthelefthand side.

withasimpleyle ontainingk 1dominaneedges,

sobothareunsolvable. Butthen,Gmustbeunsolvable

aswell. ut

Theonverseof theaboveproposition isalso true.

If G is not solvable, then G

u

ontains a yle. This

statementwillbeaorollaryofTheorem4.1,whih we

willprovebelow.

4.2 DominaneGraphs

TheProposition4.1doesnotarryoverliterally tothe

general ase: Figure 6 is a ounterexample. In order

to state our theorem for the general ase, we all a

subgraph H

u of G

u

hypernormal if the orresponding

diretedsubgraphH ofGishypernormal.

Theorem4.1. Let G = (V;E _

[D) be a dominane

graph.

(a) GissolvableiG

u

doesnotontainahypernormal

yle.

(b) Gis solvable i every hypernormal subgraph of G

is.

NotethatthisimpliesthatagraphGisongurable

iG

u

hasnohypernormalyle,byLemma2.1.

Proof. Part (b) follows immediately from part (a). If

some hypernormal subgraph of G is unsolvable, G is

unsolvable. If every hypernormal subgraph of G is

solvable,G

u

ontainsnohypernormalyle, andhene

Gissolvablebypart(a). Weturnto part(a).

Assume rstthat G

u

ontainsahypernormalyle

C. LetD 0

bethedominaneedgesof Gorresponding

toedgesin C. ThenG 0

=(V;E _

[D 0

)isahypernormal

dominanegraphsuhthat G 0

u

ontainsC. ByPropo-

sition4.1,G 0

isunsolvableandheneGisunsolvable.

Itremainstoprovetheonverse: IfGisunsolvable,

G

u

ontainsasimplehypernormalyle. Assume that

the statement is false. W.l.o.g. we may restrit our

attention to redued dominane graphs. If the hoie

ruleisnotappliableto agraph,thisgraphiseitherin

solved form or hasa direted yle, so thetheorem is

truefortheseases. Hene,theremustbea\minimal"

ounterexample G to the statement, in the following

sense:

Gisreduedandunsolvable

G

u

doesnotontainahypernormalyle

Thehoierule anbeapplied toG. Bothgraphs

HandH 0

whiharegeneratedbyitareunsolvable,

and both H

u

and H 0

u

do ontain hypernormal

yles.

We will derive a ontradition by showing that this

impliesthat G

u

ontainsahypernormalyle.

Supposethatvistherootandthat(l;v)and(l 0

;v)

arethewishesonsideredintheaboveappliationofthe

hoierule. Letrbetherootofthetreewiththeleafl

andr 0

betherootofthetreewiththeleafl 0

.

Wehavea hypernormal yle C

1

= fl;r 0

gÆP

1 in H

u

andahypernormalyleC

2

=fl 0

;rgÆP

2 in H

0

u . Sine

C

1

ishypernormal,P

1

doesnotuseanydominaneedge

inidentto l. IfP

1

doesnotusesomedominane edge

fl 0

;wginidenttol 0

,wearedonesineG

u

ontainsthe

hypernormal yle fl;vgfv;l 0

gÆ(l 0

! r 0

)ÆP

1 , where

l 0

!r 0

isthetree-pathfroml 0

tor 0

. Soassumewean

deompose P

1

= Q

1 Æfl

0

;wgÆR

1

, then R

1

does not

use any dominane edge inident to lor l 0

. A similar

argumentgivesusadeompositionP

2

=Q

2

Æfl;ugÆR

2

suh that R

2

avoids all dominane edges inident to l

andl 0

.

ThuswehavetheyleC=fl 0

;wgÆR

1

Æfl;ugÆR

2 in

G

u

. This yleis notneessarilyhypernormalbut the

pathsfl 0

;wgÆR

1

Æfl;ugandfl;ugÆR

2 Æfl

0

;wgare.

The following lemma showsthat this suÆes to prove

thatG

u

ontainsahypernormalyle:

Lemma4.1. IfG

u

ontainsayleC=e

S ÆSÆe

T ÆT

where e ;e are edges and S;T are paths suh that

(7)

4 2

5 6

3

7 1

5

4 7

6 1 3 2

Figure6: A solvabledominanegraphand oneofitssolvedforms. Thegraphontainsanundireted yle,but

nohypernormalyle.

e

S ÆSÆe

T and e

T

ÆTÆe

S

are hypernormal, then G

u

ontainsahypernormal yle.

Beforeweprovethelemma, wewantto givesome

intuition ofitsstatement: Ifwean "glue"twohyper-

normalpaths(e

S

ÆSande

T

ÆT)togehtersuhthatthey

formaylewhihishypernormalatthe"glue"nodes,

thenG

u

ontainsahypernormalyle.

Proof. We may assume that C is the smallest yle

(i.e. withtheleastnumberofedges)inG

u

fulllingthe

onditionofthelemma. Weonsidertwoases:

C issimple:

SineeverynodeonC is aninnernodeof atleast

oneofthetwohypernormalpaths,wehavethatC

does not ontaina onseutivepair of dominane

edges inident to the same leaf. And hene, the

simpliityofC impliesthatC ishypernormal.

C isnotsimple:

By the hoie of C the paths P

1

= e

S

ÆS and

P

2

=e

T

ÆT aresimple. Wemayassumethat none

of the twois ayle, otherwiseweare done. Let

v bea node whih is visited twie byC. Then v

mustbeaninner nodeofbothP

1 andP

2

, andwe

andeomposethepathsatv: P

1

=e

S ÆS

0

ÆfÆS 00

andP

2

=e

T ÆT

0

ÆgÆT 00

. Here f andg areedges

andS 0

;S 00

;T 0

;T 00

are(possibly empty)pathssuh

that e

S ÆS

0

ande

T ÆT

0

end atv. Altogehter, we

havethefollowingdeomposition:

C=e

S ÆS

0

ÆfÆS 00

Æe

T ÆT

0

ÆgÆT 00

Wehavetodistinguishtwoases:

1. f isnotadominaneedge:

Then C 0

= e

T ÆT

0

Æf ÆS 00

is a yle and

e

T ÆT

0

Æf and f ÆS 00

Æe

T

are hypernormal

2. f isadominaneedge:

Thene

S ÆS

0

doesnotend with adominane

edge. HeneC 0

=e

S ÆS

0

ÆgÆT 00

isayleand

bothe

S ÆS

0

ÆgandgÆT 00

Æe

S

arehypernomal.

InbothasesC 0

issmaller thanC, andC 0

fullls

theonditionofthelemma,aontradition.

u t

The haraterization theorem has an interesting

onsequene. The solvability problem for dominane

graphsislearlyin NP. Non-solvabilityis tantamount

to the existane of a simple hypernormal yle, and

the existane of suh a yle is learly in NP. Thus

solvabilityisin NP\o-NP.

5 Testing for HypernormalCyles

Nowweshowhowtotestforthepreseneofhypernor-

mal yles in a dominane graph in polynomial time.

This immediately gives us apolynomial algorithm for

testingsolvability(andhene,ongurability)ofdomi-

nane graphs.

Thetest isbysolvingaweightedperfetmathing

problem on anauxiliarygraphA whih is onstruted

asfollows. Foreveryedge e=(v;w)2G wehavetwo

nodes e

v

= ((v;w);v) and e

w

= ((v;w);w) in A. We

alsohavethefollowingtwokindsofedges.

(Type A) Foreveryedgeewehavetheedgefe

v

;e

w g.

(Type B) Foreverynodevanddistintinidentedges

e=(u;v)andf =(v;w)wehavetheedgefe

v

;f

v g

ifeither visnotaleaforv isaleafandeithereor

f isatreeedge.

The type A edges form a perfet mathing in A. We

givetypeAedgesweightzeroandtypeBedgesweight

(8)

inginAispositiveiG

u

ontainsasimplehypernormal

yle.

Proof. Suppose rst that G

u

ontains a hypernormal

yleC. WemayassumethatCissimple. Weonstrut

amathingMofpositiveweight. Foranypaire=(u;v)

and f = (v;w) of onseutive edges in C, we put

fe

v

;f

v

g into M. Observe that fe

v

;f

v

g 2 A sine C

ishypernormal. Foranyedge e=(v;w)2(E[D)nC

we put the fe

v

;e

w

g into M. M is learly a perfet

mathing. It ontains edges of type B and hene has

positiveweight.

Assume nextthat A hasa perfet mathing M of

positive weight. We onstrut a simple hypernormal

yle in G

u

. For any edge fe

v

;f

v

g 2 M we put the

edges e and f into the set C. Sine for any edge

e=(v;w)2G

u

wehavenodese

v ande

w

inAandsine

bothnodesmustbemathedinaperfetmathing,this

rulewillonstrutaolletionofhypernormalylesin

G

u

. ut

We assume that all non-leaves in the dominane

graph G have outdegree at most two 7

. Observe that

we have one edge of type A for every edge of G, a

omplete graph on 2+indeg

v

nodes for every root

v, and a graph of size 1+outdeg

v

for every leaf v.

Thus the auxiliary graph A has n 0

= m nodes and

m 0

=O(m)+ P

v2V

(2+indeg

v )

2

edges. ThegraphG

anbereduedintime O(nm),see [AGU72℄. Then we

havenoparalleledges,andhenearootrwithindegree

greater than n must have two dominane edges from

dierent leavesof the sametree to r, whih is trivial

to reognize in time O(nm). So we an assume that

theindegree ofanyroot isat most n. Letus saythat

wehave r n roots and letd

i

bethe indegree of the

i-th root. We have P

r

i=1 d

i

m and d

i

n. What

is the maximum value of S = P

i (2+d

i )

2

? We have

S =O(n+m)+ P

i d

2

i

. Thesum P

i d

2

i

is maximized

ifwemakethed

i

sasunequalaspossible. Soweattain

themaximum ifwe set m=nof thed

i

s equalto n and

all others equal to zero. Thus P

v2V

(2+indeg

v )

2

=

O(n+m)+O(m=nn 2

)=O(mn). Amaximalweighted

mathinginagraphwithn 0

nodesandm 0

edgesanbe

foundin timeO(n 0

m 0

logn 0

),see [GMG86℄.

Theorem5.1. The existene of a hypernormal yle

an be deided in time O(n 0

m 0

logn 0

), where n 0

= m

and m 0

= O(m)+ P

v2V

(2+indeg

v )

2

. A (simple)

7

Weanreplaeeahnon-leafwithoutdegree morethantwo

anditshildrenbyasmallbinarytree.Thisonstrutioninreases

thenumberofnodesandthenumberofedgesonlybyaonstant

fator.

1 2 k

:::

Figure7: Embeddedhainoflengthk.

hypernormalyle(ifitexists)anbefoundinthesame

timebound.

Inourappliationm=O(n)andtheindegreesare

bounded (theoutdegreesare not)andheneongura-

bilityanbedeidedin timeO(n 2

logn). Intheworst

ase,therunningtimeisO(nm 2

logn).

6 EÆient Enumeration

Arstappliationofthepolynomial-timeongurabil-

itytestfromtheprevioussetionistomaketheenumer-

ationofsolvedformsmoreeÆient. Wemodifytheenu-

merationalgorithmfromSetion3bytestingfor(undi-

reted)simple hypernormal yles in step 2instead of

diretedarbitrary yles. Thereursion will terminate

immediately one the graph beomes unongurable,

and we know that the reursion depth is bounded by

n 2

. Thus:

Corollary6.1. A solved form of a solvable domi-

nane graph an be onstruted intime O(n 3

m 2

logn).

If a dominane graph has N solved forms, they an be

enumeratedin timeO(Nn 3

m 2

logn).

NotethatNanstillbeexponentialinn. Alsonote

that we anget ongurationsinstead of solvedforms

in thesameasymptoti time, by applyingLemma 2.1:

Simpliation Rule 1 an only be applied at most

n 2

times either, by a similar argument about the

reahabilityrelation.

Goingbaktotheappliationinomputationallin-

guistisdesribedintheintrodution,thealgorithmfor

enumeratingongurationsthat wehavejust skethed

gives us a straightforward algorithm for enumerating

modelsofanormaldominaneonstraint. Wehaveim-

plementedthisalgorithm,andthisgivesusasigniant

improvement in runtimes over earlier solversfor dom-

inane onstraints. By way of example, onsider the

dominanegraphin Figure 7. Thisgraphisanembed-

dedhain oflengthk. Suhgraphsappearintheappli-

ation;forinstane,thegraphfor\Johnsaysthatevery

linguistspeakstwolanguages"isanembeddedhainof

(9)

3 5 20 180

4 14 190 670

5 42 1210 5900

6 132 4130 12740

7 429 16630 46340

8 1430 255000 n/a

Figure8: Runtimesonembeddedhainsoflengthk. N

istherespetivenumberofongurations.Timesarein

milliseondsCPUtime.

embeddedhainsofvariouslengths(ona550MHzPen-

tiumIII)aredisplayedinFigure8. Inthetable,\new"

refersto the algorithm skethed above; \old" refersto

thedominaneonstraintsolverdesribedin [DG99℄.

7 DominaneGraphs with Closed Leaves

Aslightextensionoftheongurationproblembylosed

leaves beomesNP-ompleteagain. Adominanegraph

with losed leavesis givenby adominanegraphG =

(V;E _

[D)andasetLofleaves. ThemembersofLare

alled losed, all other leaves are alled open. Closed

leaves annot be the soure of dominane wishes. A

solved form of (G;L) with losed leavesL is a solved

form G 0

= (V;E[D 0

) of G whih has the additional

propertythat there is no edge (l;v)2 E 0

with l 2 L,

butthereisanedge(l;v)2D 0

foreveryl2=L. Inother

words, it is not allowed to attah a tree to a losed

leaf,and everyopenleafmustbe\plugged"withsome

other tree. We showthat theongurationproblemof

dominanegraphswithlosedleavesisNP-ompleteby

reduingthe3-partitionproblemto it.

8

Fat7.1. (3-partition) Let A denote a multiset

fa

1

;:::;a

3m

gof integers and B 2 N suhthat B=4 <

a

i

<B=2for alli;and P

3m

i=1 a

i

=mB. Thequestion is

whetherthereisapartitionA

1

℄:::℄A

m

ofAsuhthat

for alli, P

a2Ai

a=B. Theproblem isNP-ompletein

the strongsense [GJ79 ,problem SP15,page 224℄.

We desribethe redutionnow,whih is shown in

Figure 9. The tree T has m leaves. Eah leaf wants

todominateB+1losedsubtrees (i.e.,subtrees whih

haveonly losed leaves). T is requiredto be thehild

ofsomenodel. Thisnodelalsowantstodominatethe

treest

1 , ..., t

3m

. Foralli,thetree t

i hasa

i

+1open

leaves.

8

Exatlythesameredutionworksifwedo notrequireopen

leavestohaveoutgoingdominaneedgesinsolvedforms;sothis

modiedproblemisNP-ompleteaswell.

m 1

B+1 B+1 l

t

1

t

3m t

2 T

Figure 9: The dominane graph onstruted in the

redutionof3-partition.

Theorem 7.1. The ongurability problem for domi-

nanegraphs withlosedleavesis NP-omplete.

Proof. Consider an instane (A;B) of the 3-partition

problem and the dominane graph G onstruted in

the redution. We show that the instane (A;B) has

asolutioniGisongurable.

Assume rst that the 3-partition problem has a

solution. Observethat eah of the sets A

i

must have

ardinalitythree. LetA

i

=fa

x

i

;a

y

i

;a

z

i

gbeoneofthe

sets in the partition. Then a

x

i +a

y

i +a

z

i

= B. We

plug t

xi

ashild into the i-th leaf of T, t

yi

into some

leaf of t

xi and t

zi

into someleaf of t

yi

. Then thetree

T has a

xi

+1+a

yi

+1+a

zi

+1 2 = B +1 open

leavesbelowitsi-thleaf. Theseleavesarepluggedwith

theB+1losedsubtreeswhihthei-thleafofT wants

to dominate. Finally, we plug l with T and obtain a

ongurationofG.

Assume next that the dominane graph G has a

onguration. Consider the subtree plugged to the i-

th leaf of T. It ontains a subset A

i

of the trees

ft

1

;:::;t

3m

g. We must have P

tj2Ai (a

j

+1) B +

1+jA

i

j 1,sineB+1losedsubtreesmustbeplugged

into someopenleaf and sineeverysubtreein A

i also

requiresanopenleaf.

We next show that jA

i

j 3for all i. It is lear that

A

i

annotbeempty(sineB >0). IfA

i

isasingleton,

i.e. A

i

= ft

x

g, we have a ontradition sine t

x has

a

x

+1< B=2+1 B+1 leaves. Now onsider the

ase, where A

i

onsists of two elements t

x and t

y . By

attahingt

x andt

y

belowthei-thleaf ofT, weobtain

a

x +1+a

y

+1 1<B=2+1+B=2=B+1openleaves,

whihisalsoaontradition.

Sine eah set A

i

has ardinality at least three, sine

we have m sets, and sine there are 3m elements to

distribute, we onlude that jA j = 3 for all i. Thus

(10)

tj2Ai a

j

B foreveryi. Finally, weobservethat we

haveequalitysine P

a2A

a=mB. Thus wealsohave

asolutionforthe3-partitionproblem. ut

Note thatfor solvability ofdominanegraphswith

losed leaves, Theorem 4.1 still holds. That is, solv-

ability is still a polynomial problem. The dierene

to theunrestrited problem is that Lemma 2.1 breaks

down: All the graphs we onstrut in the enoding of

3-partition are in solved form, but they may well be

unongurable.

The relevane of this resultis again in its relation

to omputational linguistis. There are alternative

approahestosope[Bos96℄whihrequirethattheholes

and roots of the trees must be paired uniquely: The

roots mustbe\plugged"intotheroots,andeveryhole

mustbeplugged. Thisorrespondstomakingtheholes

openleaves,andallotherslosedleaves. Hene,wean

showthatthesatisabilityproblemsofthesealternative

approahesmustbeNP-ompleteaswell.

8 Conlusion

We have presented a polynomial time algorithm that

solvestheongurationproblem of dominanegraphs.

This problem is of interest to appliations: It an be

usedto enodesatisability of normaldominaneon-

straints,aformalismusedin omputationallinguistis.

Thus, our result establishes a dierene in omplex-

ity between normal dominane onstraints and unre-

strited dominane onstraints, whose satisability is

NP-omplete. Previously, no polynomial time algo-

rithms for any interestingfragment of dominaneon-

straintswereknown. Testswitharstimplementation

showthat thepresentedgraphalgorithmalsoimproves

inruntimeonaprevioussolverfor(unrestrited)dom-

inaneonstraints.

Referenes

[AGU72℄ A.V.Aho,M. R.Garey, andJ.D.Ullman. The

transitiveredutionofadiretedgraph. SIAMJournal

onComputing,1:131{137,1972.

[Bos96℄ JohanBos. Prediatelogiunplugged. InProeed-

ingsofthe10thAmsterdamColloquium,pages133{143,

1996.

[Cor94℄ Thomas Cornell. Ondetermining the onsisteny

of partial desriptions of trees. In Pro. of the 32nd

Annual Meeting of the Assoiation for Computational

Linguistis,pages163{170,1994.

[DG99℄ DenysDuhierand Claire Gardent. A onstraint-

basedtreatment ofdesriptions. InProeedings of the

3 rd

Intern. Workshop on Comp. Semantis, Tilburg,

1999.

[DN00℄ Denys Duhier and Joahim Niehren. Dominane

onstraints with set operators. In Proeedings of the

FirstInternationalConfereneonComputationalLogi

(CL2000),LNCS.Springer,July2000.

[ENRX98℄ M. Egg, J. Niehren, P. Ruhrberg, and F. Xu.

Constraints over Lambda-StruturesinSemanti Un-

derspeiation. In Proeedings COLING/ACL'98,

Montreal,1998.

[GJ79℄ M.R. Garey and D.S Johnson. Computers and

Intratability: A Guide to the Theory of NP-

ompleteness. W.H.FreemanandCompany,1979.

[GMG86℄ Z. Galil, S. Miali, and H.N. Gabow. An

O(EV logV)algorithmforndingamaximalweighted

mathing ingeneral graphs. SIAM Journal on Com-

puting,15:120{130,1986.

[GW98℄ Claire Gardent and Bonnie Webber. Desribing

disoursesemantis. In Proeedings of the 4th TAG+

Workshop, Philadelphia, 1998. University of Pennsyl-

vania.

[KMN00℄ Alexander Koller, Kurt Mehlhorn, and Joahim

Niehren. A polynomial-time fragment of dominane

onstraints. InPro. of the 38thAnn. Meeting of the

Ass.forComputationalLinguistis,2000.

[KNT98℄ Alexander Koller, Joahim Niehren, and Ralf

Treinen. Dominaneonstraints: Algorithmsandom-

plexity. InPro.of the3 th

Conf.onLogialAspetsof

Comp.Linguistis,1998. ToappearasLNCS.

[MHF83℄ Mithell P. Marus, Donald Hindle, and Mar-

garetM.Flek. D-theory: Talkingabouttalkingabout

trees. InPro.of the 21st Ann. Meet. of theAss. for

Comp.Linguistis,pages129{136,1983.

[MN99℄ K.MehlhornandS.Naher. TheLEDAPlatformfor

Combinatorial andGeometri Computing. Cambridge

UniversityPress,1999. 1018pages.

[MNSU℄ K. Mehlhorn, S. Naher, M. Seel, and C. Uhrig.

The LEDA User Manual. Tehnial report, Max-

Plank-Institut fur Informatik. http://www.mpi-

sb.mpg.de/LEDA/leda.html.

[Mus95℄ R.A. Muskens. Order-Independene and Under-

speiation. InJ. Groenendijk, editor, Ellipsis, Un-

derspeiation,EventsandMore inDynamiSeman-

tis.DYANADeliverableR.2.2.C,1995.

[RVSW95℄ Owen Rambow, K. Vijay-Shanker, and David

Weir. D-TreeGrammars. InPro.of the33rdAnnual

MeetingoftheAssoiationforComputational Linguis-

tis,1995.

[Sh00℄ G. Shafer. Max-Weighted-Mathing in general

graphs. Master's thesis,Fahbereih Informatik,Uni-

versitatdesSaarlandes,Saarbruken,2000.

[TW67℄ J. W. Thather and J. B. Wright. Generalized

niteautomatatheorywithanappliationtoadeision

problem of seond-order logi. Mathematial Systems

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

[VS92℄ K.Vijay-Shanker. Using desriptions of trees in a

tree adjoining grammar. Computational Linguistis,

18:481{518, 1992.

[VSWR95℄ K.Vijay-Shanker,DavidWeir,andOwenRam-

(11)

[Zie95℄ T. Ziegler. Max-Weighted-Mathing auf allge-

meinen Graphen. Master's thesis, Fahbereih Infor-

matik,UniversitatdesSaarlandes,Saarbruken,1995.

Références

Documents relatifs

Russell used acquaintance to that effect: ÒI say that I am acquainted with an object when I have a direct cognitive relation to that object, that is when I am directly aware of

The accuracy of the geodesic convexity approach is validated by comparing the prediction provided by the algorithm proposed with the one obtained for similar instances with SVM and

Cheung and Mosca [7] have shown how to compute the decomposition of an abelian group into a direct product of cyclic subgroups in time polynomial in the logarithm of its order on

Binding of human anti-Gal to a-gal epitopes de novo synthesized on lymphoma cells was determined as preliminary evaluation of the tumor cell opsoniza- tion that precedes

In order to allow constraint solvers to handle GIPs in a global way so that they can solve them eciently without loosing CP's exibility, we have in- troduced in [1] a

To solve the problem approximately, we built a constructive heuristic algorithm (CH) to obtain an initial feasible solution for the obtained CVRP with sparse feasibility graph..

Unité de recherche INRIA Futurs : Domaine de Voluceau - Rocquencourt - BP 105 - 78153 Le Chesnay Cedex (France) Unité de recherche INRIA Lorraine : LORIA, Technopôle de Nancy-Brabois

Despite these solutions are effective in producing a proper coloring, generally minimizing the number of colors, they produce highly skewed color classes, undesirable for