• Aucun résultat trouvé

The Logic WT_mu

N/A
N/A
Protected

Academic year: 2021

Partager "The Logic WT_mu"

Copied!
29
0
0

Texte intégral

(1)

HAL Id: hal-00383062

https://hal.archives-ouvertes.fr/hal-00383062

Preprint submitted on 12 May 2009

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.

The Logic WT_mu

Omer Landry Nguena Timo

To cite this version:

Omer Landry Nguena Timo. The Logic WT_mu. 2009. �hal-00383062�

(2)

The Logi WT

µ

Omer-Landry Nguena-Timo

Université Bordeaux 1,LaBRI, CNRS

351 ours de laLibération, 33400Talene-FRANCE

omer-landry.nguena-timolabri.fr

Abstrat

Thepowerof Model-hekingdepends ontheexpressivepowerofmodelsof systems

andmodelsofspeiations.ThepaperintroduesWT

µ

, areal-timelogiwith theleast

and thegreatestxpointoperators.WT

µ

isaweaktimed extensionofthe

µ

-alulus;it

is losed to L

ν

. As Event-reordinglogi, WT

µ

desribes properties onEvent-reording automata.

WeshowthatWT

µ

ismoreexpressivethanEevent-reordinglogi.Inpartiular,with

WT

µ

formulas, one an require ourrenes of an event at all the time instants that

satises atiming onstraint.We providean exponential-time deisionproedure forthe

model-hekingofWT

µ

.

1 Introdution

ThepowerofModel-hekingdepends ontheexpressivepowerofmodelsof systemsand

modelsofspeiations.Ourgoalistopresentanewexpressivexpointlogifordesribing

propertiesonalassofreal-timesystems.Asigniantpropertythatourlogiisableto

desribethe requirement of the ourrenean event in all the time satisfying a timing

onstraint(neessity modaloperator). Wearguethat suh akindofpropertyannotbe

desribedwithEvent-reordinglogi(ERL)[Sor02℄thathasbeenintroduedbySoreafor

desribingpropertythesamelassofreal-time systems.

Real-timesystemsaremodeledwithtimedproesses.Timedproessesarenothingelse

butevent-reordingautomata[AFH99℄withoutanaeptaneondition.Timedproesses

have loalloks eah assoiatedto anevent andsuh alokgathers thetime elapsed

sine the last ourrene of the orresponding event. A timed proess is a nite state

labelled transition systemwhose transitions (

p −→ g,a p

)are labelledwith onstraintson

loks and events. A onstraint onloks isjust aonjuntionof omparisons of values

ofalokwithanintegeronstant.Clokareinterpretedoverrealnumbers.Thevalueof

eahlokgrowsontinuouslyandwiththesamerateasthetimeunlessitisreset.When

theproessisin somestate,thetimeelapses ontinuously(the valuesofthelokstoo)

untilaneventours.Then,theproessinstantaneouslyseletsatransitionlabelledwith

that eventandhekswhethertheonstraint(

g

)onthe hosentransitionis satisedby

the values of loks before it resets the lok assoiated to the eventand movesto the

targetstateofthetransition.Iftheonstraintisnotsatised,theproessdoesnothange

thestate.

The logi that we introdue in this paperis alled WT

µ

. The logi WT

µ

is a weak

timed extensionofthestandard

µ

-alulus.FormulasofWT

µ

areinterpretedovertimed proesses. Timed proesses are nothing else but event-reording automata without an

(3)

events,whilemodalitiesofERLareindexedwithpairsmadeofaonstraintandanevent.

They are of WT

µ

are of theform

hgi

and

[g]

in addition to the lassial modalities of

the

µ

-alulus indexed with event (

hai

and

[a]

). Intuitively, a state of a timed proess

p

satises

hgiϕ

from a given time-ontext desribed by avaluation

v

if by letting time

elapse in it,itis possibleto reah amomentwhen thevaluesofthelokssatisfy

g

and

inthatmoment,theformula

ϕ

issatised.Astate

p

ofatimedproesssatises

[g]ϕ

from

a time-ontext

v

ifwhenever startingfrom

v

welet thetime pass and reah amoment

when

g

is satisedthen

ϕ

is satisedin that moment. We onsider themodel-heking problem forWT

µ

;that is: Doesatimed proess satisfyaWT

µ

formula.We providean

exponential-timedeisionproedureforthat problem.

Weompare WT

µ

with ERL.ERL is also presented[Sor02℄ asatimed extension of

the

µ

-alulus;andmodelsofERL formulasaretimedproesses.InERL,modalitiesare

indexed bothwith aneventand aonstraint(

[g, a]

,

hg, ai

).A stateofatimedproess

p

satises

hg, aiϕ

from a given time-ontext desribed by avaluation

v

if by letting time

elapseinit,itispossiblethattheevent

a

oursinamomentwhenthevaluesoftheloks

satisfy

g

andaftertheourreneof

a

,theproessgoestoastatethatsatises

ϕ

.Astate

ofatimedproess

p

satises

[g, a]ϕ

fromagiventime-ontextdesribedbyavaluation

v

ifaftertheourreneof

a

inamomentwhenthevaluesoftheloks(obtainedbyletting

time elapses in

v

) satisfy

g

the proess always goes to a statethat satises

ϕ

. We will

showthat WT

µ

ismoreexpressivethanERL aseveryformulaofERL anbetranslated

into an equivalent WT

µ

formula;and there are someformulasof WT

µ

that annot be

translated into formulas of ERL. In partiular with WT

µ

, it is possible to requirethe

ourreneaneventinallthetimesatisfyingatimingonstraint;butitisnotwithERL.

Related results:Logis (TML[HLY91℄,

L t µ

[SS95℄ L

ν

[LLW95℄)that enableto de-

sribethethe neessity modal operator hasbeenonsidered fordesribing propertieson

timedautomatabutthedeidabilityofthesatisabilityproblemhasnotbeenestablished.

Laroussinieetal.[LLW95℄haveintroduedthelogi

L ν

asamorepowerfullogithanthe

onein[HLY91,SS95℄butitssatisabilityproblemisstillopenandnodisjuntivenormal

form hasbeenprovided [BCL05℄. Thelogis

L ν

and WT

µ

are inomparableastheyare not interpretedoverthe samemodel and

L ν

doesnotallowthe least xpoint operator.

But, if we restritthe interpretation of

L ν

on timed proesses, we laim that

hgiϕ

will

havethesamemeaningasthe

L ν

formula

hδi(g ∧ ϕ)

and

[g]ϕ

willhavethesamemeaning

asthe

L ν

formula

[δ](g → ϕ)

.

This paper is organisedasfollows:We presentresults forthe model-hekingof the

µ

-alulus in the next setion. We present time proesses in Setion 3. In that setion

wealsopresentwellknownoneptsandresultsonerningregion,onstraint,andtimed

abstrat bisimulation.In Setion 4we presentWT

µ

and itssemantis.Weonsider the

model-hekingproblemforWT

µ

inSetion5.InSetion6,wepresentERLandweshow

that WT

µ

is more expressive than ERL. We onlude the paper with future works on

WT

µ

.

2 Two Player Parity Game and

µ

-alulus Results

2.1 Two Player Parity Games and Multi-Parity Games

We present a omplexity result for heking a winning strategy in a twoplayer games

withparityondition.Wealsopresentthenotionoftwomulti-paritygame.

Denition 1 A two player parity game(see [Zie98℄) is a tuple

G = hN E , N A , T ⊆ N 2 , Acc G i

where

hN, T i

is a graph with the nodes (or positions)

N = N A ∪ N E

par-

titionedinto

N E

and

N A

.

N E

denotesthesetofnodesoftheplayer

Eve

and

N A

denotes

(4)

the set of nodes of the player

Adam

. The winning ondition

Acc G ⊆ N ω

, is a parity

onditiononthenodes.Thegameisnite if

N

isnite.

Aplay between

Eve

and

Adam

fromsomenode

n ∈ N

proeedsasfollows:if

n ∈ N E

then

Eve

makesahoie of asuessorotherwise

Adam

hooses asuessor; from this

suessor the samerule applies and the play goes on foreverunless one of the parties

annotmakeamove.Aplayisniteifaplayerannotmakeamoveandthenheloosethe

play.Intheasethattheplayisaninnitepath

π = n 0 n 1 n 2 · · ·

,

Eve

winsif

π ∈ Acc G

.

Otherwise

Adam

isthe winner.Among winning onditionsintrodued in theliterature, we onsider the parity ondition. A strategy

σ

for

Eve

is afuntion assigning to every

sequeneofnodes

~n

endinginanode

n

from

N E

avertex

σ(~n)

whihisasuessorof

n

.

A play from

n

onsistent with

σ

is anite orinnitesequene

n 0 n 1 n 2 · · ·

suh that

n i+1 = σ(n i )

forall

i

with

n i ∈ N E

. Thestrategy

σ

iswinning for

Eve

from thenode

n

ifandonlyifalltheplaysstartingin

n

andonsistentwith

σ

arewinning.Thestrategies

for

Adam

is are dened similarly. A node is winning ifthere exists a strategy winning

from it.A gameis determined ifeverynodeis winning foroneof theplayer.Astrategy

is positional if itdoesnot depend on the sequenesof nodes that were played till now,

butonlyonthepresentnode.Sosuhastrategyfor

Eve

anberepresentedasafuntion

σ : N E → N

andidentiedwithahoieofedgesinthegraphof thegame.

Now we state the following results on two player games (see [GH82, EJ91, Jur00,

VJ00℄).

Theorem2 Every paritygame isdetermined. Ina twoplayer parity game aplayer has

a winningpositional strategyfrom eahof his nodes.Thereis aneetive proedurethat

deides whois awinner fromagiven node ina nitegame, andthat proedure worksin

time

O |T | ×

2 × |N | d

⌈d/2⌉ !

where,

d

isthe maximalparityindex.

2.2 The

µ

-Calulus

The

µ

-alulusintroduedbyKozen[Koz82℄(seealso[AN01℄) isanexpressivetemporal

logithat extendsmodallogiwiththegreatest(

ν

)andleast(

µ

)xpointoperators.We

presentthesyntaxand thesemantisof the

µ

-alulus.Thenwestatesomewell known

resultsthatinludetheomplexityofthemodel-hekingproblem,theomplexityofthe

satisabilityproblem andadisjuntivenormalform theorem. The omplexityresultfor

themodel-hekingisobtainedbyredutiontohekingifthere isawinning strategyin

atwoplayerparitygame.

2.2.1 Denitionsand Semantis

Denition 3 The syntax of the

µ

-alulus is dened over a set

Var = {X, Y, . . .}

of

variables, aset

Σ

ofevents.Itisgivenbythefollowinggrammar:

ϕ ::= tt |

| X | ϕ ∨ ψ | ϕ ∧ ψ | haiϕ | [a]ϕ | µX.ϕ(X) | νX.ϕ(X )

In theabove,

X ∈ Var

,

a ∈ Σ

;and

tt

and denote theformula that are alwaystrue

and false respetively;

hai

and

[a]

denote the existential and the universal modalities indexedwiththeevent

a

;theyrepresentexists

a

-suessorandall

a

-suessormodalities

respetively.Theformulas

µX.ϕ(X )

and

νX.ϕ(X )

representrespetivelytheleastandthe greatestxpointformula.

Foraformula

ϕ

,thelosure[Koz82℄of

ϕ

,

sub(ϕ)

isdened asfollows:

(5)

Denition 4 Thelosure

sub(ϕ)

of

ϕ

isthesmallestset offormulassuhthat:

• ϕ ∈ sub(ϕ)

if

ψ 1 ∨ ψ 2 ∈ sub(ϕ)

theboth

ψ 1 , ψ 2 ∈ sub(ϕ)

if

ψ 1 ∧ ψ 2 ∈ sub(ϕ)

theboth

ψ 1 , ψ 2 ∈ sub(ϕ)

if

haiψ ∈ sub(ϕ)

then

ψ ∈ sub(ϕ)

if

[a]ψ ∈ sub(ϕ)

then

ψ ∈ sub(ϕ)

if

σX.ψ(X ) ∈ sub(ϕ)

then

ψ(X ) ∈ sub(ϕ)

,where

σ ∈ {ν, µ}

The formulasin

sub(ϕ)

are alled the subformulas of

ϕ

. Fora formula

ϕ

,

sub(ϕ)

is

nite and,bydenition,itisnotlargerthatthenumberofsymbolsusedin

ϕ

.

Denition 5 Theset

f ree(ϕ)

offreevariableofa

µ

-alulusformula

ϕ

isdenedindu-

tivelyasfollows:

• f ree(tt ) = f ree(

) = ∅

• f ree(X ) = {X }

• f ree(ϕ ∨ ψ) = f ree(ϕ) ∪ f ree(ψ)

• f ree([a]ϕ) = f ree(haiϕ) = f ree(ϕ)

• f ree(µX.ϕ(X)) = f ree(νX.ϕ(X )) = f ree(ϕ) \ {X}

A variable

X

isfree in aformula

ϕ

if

X ∈ f ree(ϕ)

.

Denition 6 Avariable

X

isbound inaformula

ϕ

ifthereisasubformula

σX.ψ(X)

of

ϕ

with

σ ∈ {µ, ν}

.

Denition 7(Wellnamed) Weallaformulawell named iftheexpression

µX.ϕ(X)

(or

νX.ϕ(X )

)ours atmostoneforeahvariable

X

.

Byrenamingvariablesifneessary,everyformulaanbetranslatedintoanequivalent

wellnamedformula.Inwhatfollows,withoutlossofgenerality,weassumethatformulas

arewellnamed.

Denition 8(Binding) Thebinding denition ofaboundvariable

X

inawellnamed

formula

ϕ

,

D ϕ (X )

is the unique subformula of

ϕ

of the form

σX.ψ(X)

. We will omit

subsript

ϕ

whenitausesnoambiguity. Weall

X

a

µ

-variable when

σ = µ

, otherwise

we all

X

a

ν

-variable. The funtion

D ϕ

assigning to every bound variable its binding

denition in

ϕ

willbealledthebindingfuntion assoiatedwith

ϕ

.

Denition 9 A sentene isawellnamedformulawithoutfreevariables.

Denition 10 Thedependenyorder

≤ ϕ

overtheboundvariablesofaformula

ϕ

,isthe

leastpartialordersuhthatif

X

oursin

D ϕ (Y )

(and

D ϕ (Y )

isasubformulaof

D ϕ (X )

)

then

X ≤ ϕ Y

. When

X ≤ ϕ Y

,itisalsosaidthat

Y

dependson

X

or

X

isolderthan

Y

.

Denition 11 Variable

X

in

µX.ϕ(X)

isguarded ifeveryourreneofXin

ϕ(X)

isin

the sopeof somemodalityoperator

hi

or

[]

. Wesay thata formula isguarded if every

boundvariable intheformulaisguarded.

Alternation depth desribes the number of alternations between least and greatest

xpointoperators.

Denition 12 The alternation depth of a formula denoted by

alt(ϕ)

is the numberof

nestingbetween

µ

and

ν

in

ϕ

;itisreursivelydenedasfollows:

(6)

• alt(tt ) = alt(

) = alt(X ) = 0

• alt(ϕ ∧ ψ) = alt(ϕ ∨ ψ) = max(alt(ϕ), alt(ψ))

• alt(haiϕ) = alt([a]ϕ) = alt(ϕ)

• alt(µX.ϕ(X)) = max({1, alt(ϕ(X )} ∪ {1 + alt(νY.ψ(Y )) | νY.ψ(Y ) ∈ sub(ϕ); X ≤ ϕ

Y })

• alt(νX.ϕ(X )) = max({1, alt(ϕ(X )} ∪ {1 + alt(µY.ψ(Y )) | µY.ψ(Y ) ∈ sub(ϕ); X ≤ ϕ

Y })

Formulas of the

µ

-alulus are interpreted over

Σ

-labelled transition systems. The

semantis of a

µ

-alulusformula

ϕ

is aset of states of a

Σ

-labelled transition system

S = hS, Σ, s 0 , ∆ S i

where the formula holds under a given valuation of variables

Val : Var → 2 S

, and itis denoted by

[[ϕ]] Val S

. Givena valuation of variables

Val

and aset of

states

T ⊆ S

, the valuation

Val [X/T ]

is the valuation

Val

with the substitution that assoiatesthestatesof

T

withthevariable

X

.Formally,for

Y ∈ Var

,

Val [X/T ](Y ) = T

if

Y = X

and

Val(Y )

otherwise.Wedenetherelation

betweenastate

s

ofatransition

system

S

, avaluation

Val

and aformula

ϕ

. Wewrite

S, s, Val ϕ

when theformula

ϕ

holdsin

s

orequivalently

s

satises

ϕ

.Therelation

isdenedasfollows:

• S, s, Val X

if

s ∈ Val (X )

• S, s, Val ϕ 1 ∨ ϕ 2

if

S, s, Val ϕ 1

or

S, s, Val ϕ 2

• S, s, Val ϕ 1 ∧ ϕ 2

if

S, s, Val ϕ 1

and

S, s, Val ϕ 2

• S, s, Val haiϕ

ifthereis

s −→ a s

suhthat

S, s , Val ϕ

• S, s, Val [a]ϕ

ifforall

s −→ a s

wehave

S, s , Val ϕ

• S, s, Val µX.ϕ(X)

if

s ∈ ∩{T ⊆ S | [[ϕ(X)]] S Val [X/T] ⊆ T }

.

• S, s, Val νX.ϕ(X )

if

s, ∈ ∪{T ⊆ S | T ⊆ [[ϕ(X )]] S Val [X/T] }

Thenwedene

[[ϕ]] S Val = {s ∈ S | S, s, Val ϕ}

.It issaidthata

Σ

-labelledtransition

system

S

is amodel ofaformula

ϕ

when

s 0 ∈ [[ϕ]] S Val

;in this ase wewrite

S, Val ϕ

.

Thevaluation

Val

isomittediftheformuladoesnotontainsfreevariables.

It is known (see [Eme90℄ for a survey) that properties expressed in temporal logis

LTL,CTL,andCTL

anbeenodedas

µ

-alulusformulasandthatthereareformulas

ofthe

µ

-alulus(forinstane

νX.haihaiX

)thatannotbewritteninCTL

.

Given twoformulas

ϕ 1

and

ϕ 2

, we oftenusethe notation

ϕ 1 ≡ ϕ 2

to saythat

ϕ 1

is

equivalent to

ϕ 2

,meaningthat foreverylabelledtransition system

S

andvaluation

Val

,

[[ϕ 1 ]] S Val = [[ϕ 2 ]] S Val

.

It is standard to onsider the negation operator (

¬

) on

µ

-alulus sentenes. This

operatorisdenedasfollows:

• ¬tt ≡

• ¬

≡ tt

• ¬(ϕ 1 ∧ ϕ 2 ) ≡ ¬ϕ 1 ∨ ¬ϕ 2

• ¬(ϕ 1 ∨ ϕ 2 ) ≡ ¬ϕ 1 ∧ ¬ϕ 2

• ¬haiϕ ≡ [a]¬ϕ

• ¬[a]ϕ ≡ hai¬ϕ

• ¬µX.ϕ(X) ≡ νX.¬ϕ(¬X)

• ¬νX.ϕ(X ) ≡ µX.¬ϕ(¬X)

Thefollowingpropositionisstandard.

(7)

Proposition13 Given asentene

ϕ

, a

Σ

-labelledtransition system

S

and avaluation

Val

,

[[¬ϕ]] S Val = S \ [[ϕ]] S Val

Thanks to the proposition just above, we anuse the negation operator anappear in

µ

-alulussentenes.

Letuspresentsomeresultsonthe

µ

-alulus.

Proposition14 ([Koz82℄) Everyformulaisequivalentto someguardedformula.

2.2.2 Model-Cheking Results

Informally,thetaskofhekingwhetheranitestatetransitionsystem,

S = hS, Σ, s 0 , ∆ S i

is amodelofasentene

ϕ

anbeseenastwoplayerparitygamewhosenodesare setof

tuples oftheform

(s, ψ)

where

s ∈ S

and

ψ

isasubformulaof

ϕ

. Positions oftheplayer

Eve

onstainsubformulasofoneoftheforms

tt , ϕ 1 ∨ϕ 2 , haiψ

.Theotherpositionsbelong

to the player

Adam

.The initialposition of thegame is

(s 0 , ϕ)

. Theset of movesofthe

gamesaresuhthat:

There isnomovefromeither

(s, tt)

or

(s,

)

.

From

(s, ϕ ∧ ψ)

aswellas from

(s, ϕ ∨ ψ)

therearemovesto

(s, ϕ)

andto

(s, ψ)

.

From

(s, [a]ϕ)

and from

(s, haiϕ)

there are moves to

(s , ϕ,

for every

s

suh that

s −→ a s

.

There isamovefrom

(s, σX.ϕ(X ))

to

(s, ϕ(X ))

There isamovefrom

X

to

(s, ϕ(X))

where

D(X) = σX.ϕ(X)

Theaeptaneonditionisgivenbytheparityfuntion

rank : Q → N

dened by:

rank(ψ) =

0

if

ψ

isnotavariable

2 × alt(D(X ))

where

ϕ = X

and

X

isa

ν

-variable

2 × alt(D(X )) + 1

where

ϕ = X

and

X

isa

µ

-variable

Oneanshowthat

S

isamodelofaformulaifplayer

Eve

hasawinning strategyin

thethegame.Thisgivesanintuitiveideabehindthefollowingresults.

Theorem15 ([EJ91, Tho97,Jur00℄) Let

S = hS, Σ, s 0 , ∆ S i

bea

Σ

-labelledtransition

system and let

ϕ

be a

µ

-alulus formula. The model-heking problem for

ϕ

and

S

is

solvable in time

O |∆ S | × |sub(ϕ)| × |S| × |sub(ϕ)|

⌊alt(ϕ)/2⌋

⌈alt(ϕ)/2⌉ !!

3 Timed Proesses

We present timed proesses as event-reording automata without aeptane [AFH99℄.

Werstlypresentthenotionsofregion[ACD

+

92,LY97,AFH99,AD94℄anditsproperty.

All theresultspresentedin thissetionarewell-known.

3.1 Cloks and Valuations

Cloksarevariablesevaluatedoverreal numbers.Therearetwooperationsontime,the

timeelapseoperationthatgivesthevalueofthelokafteradelayandtheresetoperation

that setsthevalueofaloksto

0

.

Let

R +

bethesetof nonnegativereal numbers.Weonsider

H = {h 1 , h 2 , . . . }

aset

ofloksvariables(orloksforsimpliity).

(8)

Denition 16 Avaluation onasetoflok

H

isatotalfuntion

v : H → R +

.

The symbol

V

represents the set of valuations. Given a valuation

v ∈ V

, and a lok

h ∈ H

,thevaluation

v + t

isdened by

[v + t](h) = v(h) + t

and,thevaluation

v[h := 0]

isdenedby

v[h := 0](h ) = 0

if

h = h

else

v[h := 0](h ) = v(h )

.Wesaythatavaluation

v

isasuessor ofavaluation

v

if

v = v + t

forsome

t ∈ R +

.

Example:Let

H = {h 1 , h 2 }

beasetoftwoloks.InTable1,wepresentsomevaluations

on

h

aresomevaluationon

H

.

v 0 (h 1 ) = 0 v 0 (h 2 ) = 0

v 1 (h 1 ) = 0.35 v 1 (h 2 ) = 0.35

v 2 (h 1 ) = 0.35 v 2 (h 2 ) = 0

v 3 (h 1 ) = 0.85 v 3 (h 2 ) = 0.50

v 4 (h 1 ) = 0

v 4 (h 2 ) = 0.50

v 5 (h 1 ) = 0.35 v 5 (h 2 ) = 0.85

Table1:Examplesof valuations.

These valuations are suh that

v 1 = v 0 + 0.35

,

v 2 = v 1 [h 2 := 0]

,

v 3 = v 2 + 0.50

,

v 4 = v 3 [h 1 := 0]

,

v 5 = v 4 + 0.35

and

v 2 = v 5 [h 2 := 0]

. In Figure 1 we give another

representationsofthesevaluationsin Cartesianreferene.

0 1

0 1 h 1

h 2

v 0 v 4

v 3

v 2

v 1 v 5

Figure1:RepresentationofvaluationsinCartesianreferene.

3.2 Constraints

Constraintsareonjuntions ofsimpleonstraints;andasimpleonstraintisaompar-

ison of alok with aninteger (diagonalfree simpleonstraint)ora omparisonof the

dierenebetweentwolokswithandinteger.Diagonalfreeonstraintsuseonlydiagonal

free simple onstraints.Constraintsare interpreted over valuations. The semantis of a

onstraintisthesetofvaluationssatisfyingit.Wewillalsoonsidertwotypesofatomi

onstraints :retangular onstraints andtriangularonstraints.

(9)

Denition 17 A simple onstraint dened on aset of loks

H

is an equation of the

form

h − h ⊲⊳ n

or

h ⊲⊳ n

where

n ∈ N

,

⊲⊳

isoneof

{<, ≤, ≥, >}

and

h, h ∈ H

.

A diagonalfree simpleonstraint isasimpleonstraintoftheform

h ⊲⊳ n

.

Denition 18 A lokonstraint overaset ofloks

H

isaonjuntionof simpleon-

straints.

Φ H

,denotesthesetoflokonstraintsover

H

.Adiagonal-free lokonstraint isalokonstraintthatusesonlydiagonalfreesimpleonstraints.

Gds H

denotestheset

ofdiagonal-freelokonstraintsover

H

.

Wewill oftenwrite

h = n

or

h − h = n

asanabbreviationof

h ≤ n ∧ h ≥ n

.Wealso

write

h − h = n

torepresenttheonstraint

h − h ≤ n ∧ h − h ≥ n

.

Laterweonsidertwospeial lokonstraints

tt

and denedby:

tt = V

h∈H h ≥ 0

and

= V

h∈H h < 0

.

The notion of a onstraint satised in a given valuation denoted

v g

is dened

indutivelyasfollows:

• v h ⊲⊳ n

ifandonlyif

v(h) ⊲⊳ n

• v h − h ⊲⊳ n

ifandonlyif

v(h) − v(h ) ⊲⊳ n

• v g 1 ∧ g 2

ifandonlyif

v g 1

and

v g 2

The meaning of a onstraint

g

, denoted

[[g]]

, is the set of valuations in whih it is

satised.Clearly,

[[g]] = {v : v g}

.Itbeomesobviousthat

[[tt]] = H → R +

and

[[

]] = ∅

.

Denition 19 Aonstraint

g

isinonsistentif

[[g]] = ∅

.

Denition 20 The bound of aonstraint

g

, denoted by

M g

, is the maximal onstant

that appears in it. The bound of a set of onstraintsis the maximal value among the

bounds ofonstraintit ontains.A setof onstraintsis

M

-bounded ifeveryonstantin

itissmallerthan

M

.

Nowweonsideratomionstraintsandweshowhowtodeomposeaonstraintinto

anequivalentset ofatomionstraints.

Denition 21 Forainteger

M ∈ N

,a

M

-retangular onstraint isaonjuntionofthe form

V

h∈H g h

where

g h

isaonstraintoftheform

c < h < c + 1

or

h = c

or

h > M

with

c ∈ N ∩ [0..M [

.

Thesetofall

M

-retangularonstraintsisdenotedby

Agds H (M )

.Thesymbol

Agds H

will denotetheset

S

M ∈ N Agds H (M )

Denition 22 A

M

-triangularonstraintisaonjuntionoftheform

V

h∈H g h ∧ V

(h,h )∈H 2 g h,h

where

g h,h

isaonstraintofthe forms

c < h − h < c + 1

or

h − h = c

or

h − h > M

and

g h

isoftheform

c < h < c + 1

or

h = c

or

h > M

with

c ∈ N ∩ [0..M[

.

Thesymbol

T gds H (M )

denotesthesetofallof

M

-triangularonstraints.Thesymbol

T gds H

denotestheset

S

M ∈ N T gds H (M )

.

Notation:Weoftenusethesymbol

g ˆ

todenoteaonstraintin

Agds H (M )

or

T gds H (M )

forsome

M

.Laterthetermsatomionstraintswilloftenbeusedinplaeofretangular

onstraintsortriangularonstraints.

Letus rstreallthefollowingfat resultingfrom denitionsofatomionstraints.

Fat23 (atomiity) Let

M ∈ N

beaonstant.

• ∀ˆ g, g ˆ ∈ T gds H (M )

,if

[[ˆ g]] 6= [[ˆ g ]]

then

[[ˆ g]] ∩ [[ˆ g ]] = ∅

• ∀ˆ g, g ˆ ∈ Agds H (M )

, if

[[ˆ g]] 6= [[ˆ g ]]

then

[[ˆ g]] ∩ [[ˆ g ]] = ∅

(10)

• ∀(ˆ g, ˆ g ) ∈ Agds H (M ) × T gds H (M )

, either

[[ˆ g ]] ∩ [[ˆ g]] = ∅

or

[[ˆ g ]] ⊆ [[ˆ g]]

The rsttwo items statethat either the semantis of twoatomi onstraintsof the

samenatureareequal,ortheyaredisjoint.Thelastitemoftheabovefatstatesthatthe

semantisof atriangular onstraintis either inluded in thesemantisof aretangular

onstraints,orthetwosemantisaredisjoint.

Example: In Figure 2,we illustrate the onepts of onstraints and diagonal free on-

straints. The onstraints

g 1

and

g 3

are general onstraints while the onstraint

g 2

is

diagonal free.Moreover

[[g 3 ]] = [[g 1 ]] ∧ [[g 2 ]]

.The onstraint

g 2

isaretangular onstraint

0 1 2

0 1 2 h a

h b

g 1 = 0 ≤ h a ≤ 3 ∧ 0 ≤ h b ≤ 2 ∧ −1 ≤ h a − h b ≤ 1

g 2 = 1 < h a < 2 ∧ 0 < h b < 1

g 3 = 1 < h a < 2 ∧ 0 < h b ≤ 1 ∧ −1 ≤ h a − h b ≤ 1

Figure2:Illustrationof onstraintsanddiagonalfreeonstraints.

in

Agds H (2)

andtheonstraint

g 3

is atriangularonstraint.

Normalization andRetangularisation Untiltheendofthissubsetionweon-

siderthedeompositionofdiagonalfreeonstraintintosetofretangularonstraints.We

willneedtoonsideronstraintsthatdonotinvolveonstantsgreaterthanaxedbound.

For that purpose, we present the normalisation operation

norm N

that we use later to

deomposeonstraints.

Denition 24 The

N

-normalizationofasimpleonstraint

C

istheonstraint

norm N (C)

dened by:

• norm N (h ⊲⊳ n) = tt

if

⊲⊳∈ {<, ≤}

and

n > N

.

• norm N (h − h ⊲⊳ n) = tt

if

⊲⊳∈ {<, ≤}

and

n > N

.

• norm N (h ⊲⊳ n) = h > N

if

⊲⊳∈ {>, ≥}

and

n > N

.

• norm N (h − h ⊲⊳ n) = h − h > N

if

⊲⊳∈ {>, ≥}

and

n > N

.

Intheotherases

norm N

doesnotmodifytheonstraint.

Givenaonstraint

g

andaninteger

N

,the

N

-normalizationof

g

,

norm N (g)

isobtained

bynormalizingeahsimpleonstraintourringin

g

.

Lemma25 Let

C

,adiagonal-freesimpleonstraint,thereisaonstant

M

suhthat:

forevery

N ≥ M

,

[[norm M (C)]] = [[norm N (C)]] = [[C]]

forevery

N < M

,

[[norm M (C)]] ( [[norm N (C)]]

Proof

1. When

C

hastheform

h ⊲⊳ n

with

⊲⊳∈ {<, ≤}

andonsider

M = n

,

(11)

(a) Let

N ≥ M

,

norm N (h ⊲⊳ n)

isequalto

norm M (h ⊲⊳ n)

and theyareequalto

h ⊲⊳ n

andwegettheresultthat

[[norm M (C)]] = [[norm N (C)]] = [[C]]

.

(b) Let

N < M

,

norm N (h ⊲⊳ n) = h ≥ 0

.Clearly

[[norm M (C)]] ( [[norm N (C)]]

.

2. When

C

hastheform

h ⊲⊳ n

with

⊲⊳∈ {>, ≥}

andonsider

M = n

,

(a) Let

N ≥ M

,

norm N (h ⊲⊳ n)

isequalto

norm M (h ⊲⊳ n)

and theyareequalto

h ⊲⊳ n

andwegettheresultthat

[[norm M (C)]] = [[norm N (C)]] = [[C]]

.

(b) Let

N < M

, then

norm N (h ⊲⊳ n) = h ⊲⊳ N

and

[[norm M (C)]] = h ⊲⊳ M

.

Clearly,

[[norm M (C)]] ( [[norm N (C)]]

.

Letusreallthatforaonstraint

g

,

M g

denotesthemaximalonstantourringin

g

.

Weuse thelemmaabovetoshowthat the

M

-normalisationofaonstraintdoesmodify itssemantiswhen

M

isgreaterorequalto

M g

.

Proposition26 Let

g ∈ Gds H

,

forevery

M ≥ M g

,

[[norm M (g)]] = [[norm N (g)]] = [[g]]

forevery

M < M g

,

[[norm M (g)]] ( [[norm N (g)]]

Proof

Bydenitions

g = V

i=1..n C i

and,

[[norm M (g)]] = T

i=1..n [[N orm M (C i )]]

.As

M g

isgreater

that theonstant usedin every

C i

, we get, using 25 that for

M ≥ M g

,

[[norm M (g)]] = [[norm N (g)]] = [[g]]

andfor

M < M g

,

[[norm M (g)]] ( [[norm N (g)]]

Example:Consideringtheonstraint

g = 0 ≤ h a ≤ 3 ∧ 0 ≤ h b ≤ 2

,wepresentinTable2

theresultsof

M

-normalisationoperationsdepending onthevalueof

M

.It iseasytosee

M

norm M (g)

0

tt

1

tt

2

0 ≤ h b ≤ 2

3

0 ≤ h a ≤ 3 ∧ 0 ≤ h b ≤ 2

Table2:Illustrationofthenormalisationoperation.

that forevery

M < 2

,

[[g]] ⊆ [[norm M (g)]]

andforevery

M ≥ 2

,

[[g]] = [[norm M (g)]]

To obtain the deomposition of diagonal onstraints,we rstly deompose diagonal

freeonstraintsintoaset(possiblyinnite)ofunboundedretangularonstraints.Then,

weusethenormalisation proedureaboveoneah atomionstraintinthat set to have

anite set of bounded retangular onstraints.The deomposition ofdiagonalfree on-

straints into a set of unbounded retangular onstraints is performed in two steps: in

Lemma27wedeomposesimplediagonalfreeonstraintsandweusethatdeomposition

in Proposition28todeomposediagonalfreeonstraints.

Lemma27 Foreverydiagonalfreesimpleonstraint

C

,thereisaset

Rect(C)

ofatomi

diagonalfreesimpleonstraintssuhthat

[[C]] = S

C ∈Rect(C) [[C ]]

.

Proof

Let

C

beadiagonalfreeonstraint

C

.Weonstrutaset

Rect(C)

dependingontheform

of

C

; andweshowthatfor every

v ∈ V

,

v C

ifand onlyifthereis

C ∈ Rect(C)

suh

that

v C

.

1. if

C

isoftheform

h < n

then set

Rect(C) = {i < h < i + 1, h = i | i = 0..n − 1}

(12)

2. if

C

isoftheform

h ≤ n

thenset

Rect(C) = {i < h < i+1, h = i | i = 0..n−1}∪{h = n}

3. if

C

isoftheform

h > n

then set

Rect(C) = {i < h < i + 1, h = i + 1 | i = n..∞}

4. if

C

is of the form

h ≥ n

then set

Rect(C) = {i < h < i + 1, h = i + 1 | i = n..∞} ∪ {h = n}

Theproofthat ineahase,

[[C]] = ∪ C ′ ∈Rect(C) [[C ]]

,isobvious.

Weobservethatsimpleonstraintsofthe form

h > n

to

h ≥ n

aredeomposed into

innitesetofonstraints.

Proposition28 Foreverydiagonal-freeonstraint

g

,thereisaset

Rect(g)

ofretangular

onstraintssuhthat

[[g]] = S

ˆ

g∈Rect(g) [[ˆ g]]

.

Proof

The resultis aonsequeneofthe Lemma27 aboveasaonstraintsisa onjuntionof

simpleonstraints.

Wesaythat

Rect(g)

istheunboundedretangulardeomposition of

g

.

Now that wehavedeomposed diagonal freeonstraintsinto sets (possibly innite)

ofunbounded retangularonstraints,wewillapply thenormalisationoperationoneah

retangular onstraint in these sets; the result of the appliation of the normalisation

operationwithrespettoaonstant

M

willbenitesetof

M

-retangularonstraints.But weneedtoshowthatthesemantisoftheonstraintresultingfromtheappliationofthe

M

-normalisationoperationonasimplediagonalfreeonstraintisthesameastheunion ofthesemantisofretangularonstraintsinitsunbounded retangulardeomposition.

Lemma29 Foreverydiagonalfreesimpleonstraint

C

oftheform

h ≤ n

or

h ≥ n

,for

every

M ∈ N

,

[[norm M (C)]] = ∪ C ∈Rect(C) [[norm M (C )]]

.

Proof

If

C

isoftheform:

• h ≤ n

,

If

M ≥ n

then

norm M (C) = C

andforevery

C ∈ Rect(C)

,

norm M (C ) = C

.

Thenwegettheresult.

If

M < n

then

norm M (C) = tt

.Let

C h = n

.FromLemma27

C ∈ Rect(C)

and

norm M (C ) = tt

then

C ′ ∈Rect(C) [[norm M (C )]] = tt

and

[[norm M (C)]] =

C ′ ∈Rect(C) [[norm M (C )]]

.

• h ≥ n

,

Theasewhen

M ≥ n

isobviousbeauseeveryonstraintin

Rect(C) ∪ {C}

is

notmodiedby

norm M

.

The ase when

M < n

is also obvious beause

norm(C) = h > M

and

norm M (C ) = h > M

forevery

C ∈ Rect(C)

Nowweaneasilyextendresultsinthelemmaabovetodiagonalfreeonstraints.

Proposition30 Foreverydiagonal-freeonstraint

g

, forevery

M ∈ N

,

[[norm M (g)]] = S

ˆ

g∈Rect(g) [[norm M (ˆ g)]]

.

Proof

It isaonsequeneofLemma 29aboveandProposition28

(13)

Denition 31 Givena

g ∈ Gds

,andandinteger

M ∈ N

wedenetheset

Rect M (g) = {norm M (ˆ g) | g ˆ ∈ Rect(g)}

.

FromProposition26,wegetthateverydiagonal-freeonstraintusingonstantsmaller

thananinteger

M

anbedeomposed intoanitesetof

M

-retangularonstraints.

Proposition32 Foreveryonstraint

g ∈ Gds

,forevery

M ≥ M g

,

[[g]] = S

ˆ

g∈Rect M (g) [[ˆ g]]

.

Proof

FromProposition30

[[norm M (g)]] = S

ˆ

g∈Rect(g) [[norm M (ˆ g)]]

orequivalently

[[norm M (g)]] = S

ˆ

g∈Rect M (g) [[ˆ g]]

. From Proposition 26 for

M ≥ M g

,

[[g]] = [[norm M (g)]]

and weget the

result.

Remark:Thesamekindof propertyanbeestablishedforgeneralonstraintsandtri-

angular onstraints. As retangular onstraintsontain triangularonstraints every

M

-

bounded diagonal free atomi onstraint an be deomposed into a nite union of

M

-

boundedtriangularonstraints.

Fromtheremarkabovewehavethefollowingorollary.

Corollary 33 Every onstraint or diagonal free onstraint an be deomposed into a

nite equivalentsetoftriangularonstraints.

3.3 Regions

We present a partitioning of the valuations into a nite number of equivalene lasses

alled regions. Valuations in the same region must satisfy the same lok onstraints,

theirtime suessorsmustalso satisfythesamelokonstraints,andtheymustsatisfy

thesamelokonstraintsafter alokisreset.

ThedenitionofaregionwepresentherehasbeenintroduedbyAlurandDill[AD94℄

for analysing timed automata using onlydiagonal -freeonstraints.The equivalene re-

lation between valuations is dened with respet to some integer

M

representing the maximal valueusedin onstraints.Thedenition ofthat relationis somehowrelatedto

the denition of atomi onstraintsas atomi onstraints an not be deomposed into

smaller onstraints. Thus, twoequivalent valuations agree on the integral part of eah

lokwhosevaluesaresmallerthan

M

andtheyalsoagreeontheorderonthefrational

partofthevaluesoftheloks.

Forarealnumber

n

let

⌊n⌋

denotetheintegralpartof

n

and

{n}

denotethefrational

partof

n

.

Let

M

beanaturalnumber.Considertheparametrisedbinaryrelation

M ⊆ V H × V H

overvaluationsdened by,

v ∼ M v

if:

1.

v(h) > M

ifandonlyif

v (h) > M

foreah

h ∈ H

;

2. if

v(h) ≤ M

,then

⌊v(h)⌋ = ⌊v (h)⌋

forevery

h ∈ H

;

3. if

v(h) ≤ M

,then

{v(h)} = 0

ifandonlyif

{v (h)} = 0

forevery

h ∈ H

,and;

4. if

v(h) ≤ M

and

v(h ) ≤ M

,then

{v(h)} ≤ {v(h )}

ifandonlyif

{v (h)} ≤ {v (h )}

forevery

h, h ∈ H

.

Proposition34 ([AD94℄) Therelation

M

is an equivalenerelation over the set of

valuationswithat most

2 3|H|−1 × |H|! × (M + 1) |H|

equivalenelasses.

(14)

Therelation

M

isdened asaonjuntionoffour properties.Eahpropertydenesan equivalene relation;letus denote them by

M 1 , . . . , ∼ M 4

, respetively. Foreahof these four relationswewillgiveanupperbound onthenumberofits equivalenelasses.The

produtoftheseboundswill giveanupperbound on

M

asthelateristheintersetion ofthefourequivalenerelations.

The relation dened by the rst ondition has

2 |H|

equivalene lasses, as the only

thing that ountsis whetherthe valueofalok isbiggerthan

M

ornot.Similarlythe

third relationhas

2 |H|

equivalenelasses. Thenumberof lassesof theseond relation

is

(M + 1) |H|

as thereare

M + 1

possibleintegervaluesofinterest.Finally, thenumber

of lassesof thefourth relationis bounded by thenumberofpermutationsof theset of

loksmultipliedby

2 |H|−1

asforeverytwoloksonseutiveinapermutationweneed to deideiftheyareequaloriftheseondisstritlybiggerthantherst.

Summarizing,weget

2 3|H|−1 |H!|(M + 1) |H|

.

Weuse

Reg(M )

(or

Reg

forshort) to representtheset of equivalene lassesofthe

relation

M

.

Denition 35 A region [AD94℄ is an equivalene lass of the relation

M ⊆ V H × V H

dened above.

InFigure3weillustrateregionfordiagonalfreeonstraintsforthemaximalonstant

M = 2

.InFigure3valuationsearlierpresentedinTable1arenotequivalent.Aregionin thegureiseitheraornerpoint(forexample

(0, 2)

),anopenlinesegment(forexample

0 < h 1 = h 2 < 1

)or anopenbox(forexample

0 < h 1 < h 2 < 1

).

0 1 2

0 1 2 h 1

h 2

v 0

v 4

v 3

v 2

v 1

v 5

Figure3:Regionillustration.

Fromthedenitionof

M

,itomesthatanequivalenelassanberepresentedusing atriangularonstraintin

g

.Aordingtothedenitionof

M

,twovaluationsthatbelong

to thesameequivalenelasssatisfyonstraintoftheform:

(15)

• h = i h

or

i h < h < i h + 1

foreah

h ∈ H

where

i h ∈ {0, 1, . . . , M }

andweassume

M + 1 = ∞

.Thisisaonsequeneof

M 1

,

M 2

,

M 3

.

• h − h = i hh

or

i hh < h − h < i hh + 1

for eah ouple

(h, h ) ∈ H 2

suh that

h ⊲⊳ M

and

h ⊲⊳ M

with

⊲⊳∈ {=, <}

.Thisisaonsequeneof

M 4

.

Given avaluation

v

,

[v]

denotesthe equivalene lass (region)of

v

. We also usethe

letter

r

to represent a region. Given a region

r

, we dene

r + t = {[v + t] | v ∈ r}

,

r↑ = {r + t | t ∈ R ≥0 }

,and

r[h := 0] = {v[h := 0] : v ∈ r}

. Wewrite

r ⊆ g

for

r ⊆ [[g]]

.

Proposition36 Let

G

beasetof

M

-boundedonstraintsthen

Reg(M )

satises:

P1

∀g ∈ G, r ∈ Reg

,either

r ⊆ [[g]]

or

[[g]] ∩ r = ∅

.

P2

∀r, r ∈ Reg

, ifthere exists some

v ∈ r

and

t ∈ R ≥0

suhthat

v + t ∈ r

, thenfor

every

v ∈ r

thereis some

t ∈ R ≥0

suhthat

v + t ∈ r

.

P3

∀r, r ∈ Reg, ∀h ∈ H

,if

r[h := 0] ∩ r 6= ∅

,then

r[h := 0] ⊆ r

.

Proof

WeshowP1intherstitem,P2in theseonditemandP3inthelastitem.

1. Let

g ∈ G

,fromProposition 32let

[[g]] = S

g i ∈Rect M (g) [[ˆ g i ]]

. Eah

ˆ g i

isaretangular

onstraint.

[[g]] ∩ r = S

g i ∈Rect M (g) [[ˆ g i ]] ∩ r)

.FromFat23thereisatmostone

i

suh

that

r

intersets

ˆ g i

. It followsthat

r

intersets a onstraint

ˆ g i

of

Rect M (g)

if and

onlyif

ˆ g i

ontains

r

.Wehavethatif

v r

then

v g

.

2. Let

v, v ∈ r

,adding

t

to

v

maymodifytheintegerpartofthevalue(withrespetto

v

)ofsomeloksormaymodifytheorderonthefrationalpartofthevalue(with

respetto

v

)ofloks.Weaimatndatime

t

suhthat:

- Theintegerpartof thevalueof eah lokwithrespetto

v + t

is equaltothe

integerpartofthevalueofeahlokwithrespetto

v + t

-Theorderofthefrationalpartsofloksin

v + t

isthesamein

v + t

.

-Thesetoflokswithzerofrationalpartin

v + t

isthesamein

v + t

.

Let

|H| = n

andassumeapermutation

π

of

{1, . . . , n}

suhthat

{v(h π 1 )} ⊲⊳ 1 {v(h π 2 )} ⊲⊳ 2 , . . . , ⊲⊳ n−1 {v(h π n )}(∗)

with

⊲⊳ i ∈ {<, =}

.

Let

t ∈ R ≥0

. Itis learthat

{v(h) + t} = {v(h) + {t}}

. Onlythefrationalpartof

t

mayaettheorderin

(∗)

.

There maybealargestindex

j

suhthat

{v(h π j ) + {t}} = {v(h π j )} + {t}

.Inase,nosuh

j

exists, take

j = n

.

Clearly,

{v(h π j ) + {t}} ≥ {v(h π j )}

and;

∀k > j

wehave:

{v(h π k ) + {t}} < {v(h π k )}

and

{v(h π k ) + {t}} < {v(h π j ) + {t}}

.

Wegetthat:

{v(h π j+1 ) + {t}} ⊲⊳ j . . . ⊲⊳ n−1 {v(h π n ) + {t}} < {v(h π j ) + {t}}

Similarly,weestablishthat

{v(h π j ) + {t}} < {v(h π j− 1 ) + {t}}⊲⊳ j−2 . . . ⊲⊳ 1 {v(h π 1 ) + {t}}

where

⊲⊳ k =>

if

⊲⊳ j ∈ {<}

otherwise

⊲⊳ j ∈ {=}

,

∀k ≤ j

Références

Documents relatifs

Keywords : ARCH models, bilinear models, conditional least squares, method of moments, non-linear moving average models.. Classification AMS : 62 M 10, 93

shall be composed of the following members: the Chairman and Vice-Chairmen of the Regional Committee for the Eastern Mediterranean, the Chairman of the Regional

A cet effet, ARITMA propose le Code de Conduite pour la Société, le Jeune et les Ecoles pour faire renaitre cette société de Paix, de Travail et travailleurs qui honorent la Patrie

The resulting matrix is of the same order as A (respectively B), i.e., 2×2; and, by definition, it is obtained via component-wise addition.. The resulting matrix is obtained

We show that any orientable Seifert 3-manifold is diffeomorphic to a connected component of the set of real points of a uniruled real algebraic variety, and prove a conjecture of

While combining time constraints and data structures, we cannot directly rely on the formula for realizability from Section III in the approach outlined above. The crucial

We use three of these no- tions, unitary, dense and planar subcategories in Section 2 to describe the kernels of certain morphisms of categories (functors), called quotient maps, from

Heatmaps of (guess vs. slip) LL of 4 sample real GLOP datasets and the corresponding two simulated datasets that were generated with the best fitting