O
pen
A
rchive
T
OULOUSE
A
rchive
O
uverte (
OATAO
)
OATAO is an open access repository that collects the work of Toulouse researchers
and makes it freely available over the web where possible.
This is an author-deposited version published in :
http://oatao.univ-toulouse.fr/
Eprints ID : 15472
To link to this article :
Official URL: http://dx.doi.org/10.1016/j.ijar.2014.11.005
To cite this version :
Dubois, Didier and Prade, Henri and Rico, Agnés
Representing qualitative capacities as families of possibility measures.
(2014) International Journal of Approximate Reasoning, vol. 58. pp. 3-24.
ISSN 0888-613X
Any correspondence concerning this service should be sent to the repository
administrator: staff-oatao@listes-diff.inp-toulouse.fr
Representing
qualitative
capacities
as
families
of
possibility
measures
Didier Dubois
a,
Henri Prade
a,∗,
Agnès Rico
baIRIT,UniversitéPaulSabatier,118routedeNarbonne,31062Toulousecedex9,France bERIC,UniversitéClaudeBernardLyon1,43blddu11novembre,69100Villeurbanne,France
a b s t r a c t
Keywords:
Fuzzymeasures Possibilitytheory Modallogic
Thispaperstudiesthestructureofqualitativecapacities,thatis,monotonicset-functions, when they range on a finite totally ordered scale equipped with an order-reversing map.Theseset-functions correspondtogeneralrepresentationsofuncertainty,aswell as importancelevelsofgroupsofcriteriainmultiple-criteriadecision-making.Weshowthat any capacityor fuzzy measure rangingon aqualitative scalecan beviewed bothas the lowerbound ofa setofpossibility measuresand theupper boundof aset ofnecessity measures(asituationsomewhat similartotheoneofquantitative capacitieswithrespect to imprecise probability theory). We show that any capacity is characterized by a non-empty class of possibility measures having the structure of an upper semi-lattice. The lower bounds of this class are enough to reconstruct the capacity, and the number of themischaracteristicofitscomplexity.Analgorithmisprovidedtocomputetheminimal set of possibility measures dominating a given capacity. This algorithm relies on the representationofthecapacitybymeansofitsqualitativeMöbiustransform,andtheuseof selectionfunctions ofthe corresponding focal sets. Weprovide theconnection between Sugeno integrals and lower possibility measures. We introduce a sequence of axioms generalizing the maxitivity property of possibility measures, and related to the number of possibility measures needed for this reconstruction. In the Boolean case, capacities arecloselyrelated tonon-regularmodal logicsand theirneighborhood semantics canbe describedintermsofqualitativeMöbiustransforms.
1. Introduction
Afuzzymeasure [38] (oracapacity[10])isaset-functionthatismonotonicunderinclusion.Inthispaper,thecapacity is saidto bequalitative (or q-capacity, forshort) if itsrangeisafinitetotally ordered set.Itmeans wedo notpresuppose addition isavailableinthecapacityrange,only minimumand maximum.Insuchacontexttheconnectionwithprobability measuresislost.Consequentlyanumberofnotions,meaningfulinthenumericalsetting,arelostaswell,suchastheMöbius transform[34],theconjugate,supermodularity[10] andthelike.Likewise,somenumericalcapacities(suchasconvexones, belieffunctions)canbeviewed asencodingaconvexfamilyofprobabilitydistributions[35,40].Thisconnectiondisappears if wegiveupusingadditionintherangeofthecapacity.
*
Correspondingauthor.Yet,itistemptingtocheckwhethercounterpartsofmanysuchquantitativenotionscanbedefinedforqualitative capac-ities,ifwereplaceprobabilitymeasuresbypossibilitymeasures.Conjugatenesscanberecoverediftherangeofthecapacity is equipped withan order-reversingmap. A qualitativecounterpart ofa Möbiustransformhas been introduced byMesiar [31] and Grabisch [25] in 1997 and further studiedby Grabisch[26]. The qualitativeMöbius transformcan be viewed as the possibilisticcounterpart to abasic probability assignment,wherebya capacityis defined withrespectto thelatterby a qualitative counterpart of thebelief functiondefinition, extending aswell thedefinition ofpossibility measures. In fact the process of generation of belieffunctions, introduced byDempster [9], was applied very early to possibility measures by Dubois and Prade [16,17] so as to generate upper and lower possibilities and necessities. It was noticed that upper possibilitiesand lower necessitiesarestill possibilityand necessitymeasures respectively,but uppernecessitiesand lower possibilitiesarenot.However,asweshallsee,theformalanalogybetweenbelieffunctionsandqualitativecapacitiesviathe qualitativeMöbiustransformcanbemisleadingattheinterpretivelevel.
Thissituationleadstonaturalquestions,namely,whetheraqualitativecapacitycanbeexpressedintermsofafamilyof possibility measures,and ifaqualitativeMöbiustransformcanencodesuchafamily.Previous recentworks[11,32] started addressingthisissue,takingupapioneeringworkbyBanon[4].Inthispaperweshowthatinthefinite(qualitative)setting, specialsubsetsofpossibilitymeasuresplayarolesimilartoconvexsetsofprobabilitymeasures.Weprovethatanycapacity can bedefinedinterms ofafiniteset ofpossibility measures,eitherasalower possibilityoranuppernecessityfunction. Thisresultshouldnotcomeasasurprise.Indeed,ithasbeenshownthatpossibilitymeasurescanberefinedbyprobability measures using alexicographic refinement of thebasic axiom of possibility measures, and that capacities on a finiteset canberefinedbybelieffunctions[13,14].Basedonthisfundamentalresultwecangeneralizethemaxitivityandminitivity axiom ofpossibility theoryso astodefine familiesofqualitativecapacities ofincreasing complexity.Finally, thisproperty enablesqualitativecapacities tobeseen asnecessitymodalitiesinanon-regularclassofmodallogics, extending thelinks betweenpossibilitytheoryandmodallogictoapotentiallylargerrangeofuncertaintytheories.
The structure of thepaper is asfollows.1 Section 2 provides basic definitionspertaining to capacities, recalls and dis-cusses thesimilaritybetweenbelieffunctionsand qualitativecapacities,indicating thelimitationofthisanalogy. Section 3 providesthemaincontributionofthispaper,namelyitshowstheformalanalogybetweenqualitativecapacitiesand impre-ciseprobabilities,provinganycapacitycomes downtoany oftwofamiliesofpossibilitydistributions,andcanbedescribed byfinitesetsthereof,eitherasalowerpossibilityoranuppernecessityfunction.Thissectionalsoextendstheseresults to Sugeno integrals. Section 4 providesanalgorithm that computesthe set ofminimalelements among possibility measures that dominate acapacity fromits qualitativeMöbiustransforms. Section 5 axiomaticallydefines subfamilies ofqualitative capacities of increasing complexity generalizing the maxitive and the minitivity axioms of possibility theory. Finally, in Section 6 we lay bare a connection between capacities and neighborhood semantics in non-regular modal logics, which suggestspotentialapplicationstoreasoningfromconflicting informationcomingfromseveralsources.
2. QualitativecapacitiesandMöbiustransforms
Consider a finite set S and a finite totally ordered scale L= {λ0=0< λ1< · · · < λℓ=1} with top 1 and bottom 0.
Moreover we assume that L is equipped with an order-reversing map, i.e., a strictly decreasingmapping
ν
:L→L withν
(1)=0 andν
(0)=1.Notethatν
isunique,and suchthatν
(λi)= λℓ−i.Definition1. A capacity (or fuzzy measure) is a mapping
γ
:2S →L such thatγ
(∅)=0;γ
(S)=1; and if A⊆B thenγ
(A)≤γ
(B).Theconjugateofγ
isthecapacityγ
c defined asγ
c(A)=ν
(γ
(Ac)),∀A⊆S,where Ac isthecomplementof set A.The value
γ
(A) can be interpreted as the degreeof confidence in a propositionrepresented by the set A of possible states ofthe world, or, if S isa set ofcriteria, the degreeof importanceof thegroup of criteria A [27]. In thispaperwe basically use the first interpretation, unless specified otherwise. We here speak of qualitativecapacity (or q-capacity, for short) tomeanthatweonlyrelyonanordinalstructure,notanadditiveunderlying structure.Remark1. In fact, even if thescale isencoded bymeans ofnumbersin [0,1], we do notassume these figures represent orders ofmagnitude,so thattheiraddition orsubtraction makeno sense.Ofcourse,wecouldconstructaq-capacityusing a probabilitymeasure P on S,and considering{P(A):A⊆S} asthetotallyordered set L byrenamingthenumbersusing symbols λi. It isalways possible todo so using any numericalcapacity on S.However, since thesymbols λi only encode
a ranking, we thenare unableto distinguish betweencapacities that yieldthe sameorderingof events (inparticular the probabilitymeasure P cannolongerbedistinguishedfromthemanynon-additivenumericalcapacitiesthatyieldthesame orderingofeventsas P ).Soqualitative herepresupposesthat the“distance” betweentwoconsecutiveλi’scanbearbitrary.
And inourview,anumericalset-functionisaspecialcaseofaqualitativeonewithadditionalstructureinitsrange. Importantspecialcasesofcapacityarepossibilityandnecessitymeasures.
Definition2.Apossibilitymeasure is acapacitythatsatisfiesthecharacteristicaxiom ofmaxitivity:
γ
(
A∪
B) =
max¡
γ
(
A),
γ
(
B)¢.
Apossibilitymeasure isusuallydenoted byΠ.In possibilitytheory[18,20],theavailableinformationisrepresentedby means ofapossibility distribution [42].This isa function,usually denoted by
π
,from theuniverse of discourse S tothe scale L,suchthatπ
(s)=1 forsomeelements∈S.TheassociatedpossibilitymeasureisdefinedbyΠ (
A) =
max s∈Aπ
(
s)
and in the finite case any possibility measure induces a unique possibility distribution
π
(s)= Π ({s}). The functionπ
is supposedtorank-orderpotentialvaluesof(someaspectof)thestateoftheworld–accordingtotheirplausibility.Thevalueπ
(s)isunderstoodasthepossibilitythat s betheactualstateoftheworld.Preciseinformationcorrespondstothesituation where ∃s∗,π
(s∗)=1,and ∀s6=s∗,π
(s)=0,whilecompleteignorance isrepresentedbythevacuouspossibilitydistributiondenoted by
π
?,suchthat ∀s∈S,π
?(s)=1.Apossibilitydistributionπ
issaidto bemorespecific (inthewidesense)than anotherpossibilitydistributionρ
if∀s∈S,π
(s)≤ρ
(s).Itdefinesapartialorderbetweenpossibilitydistributions,reflecting theirrelativeinformativeness.Definition3.Anecessitymeasure isacapacitythatverifiesthecharacteristicaxiomofminitivity:
γ
(
A∩
B) =
min¡
γ
(
A),
γ
(
B)¢.
A necessity measure is usually denoted by N. Function N is characterized by an “impossibility” distribution
ι
:S→Las
ι
(s)=N(S\ {s}),since N(A)=mins/∈Aι
(s). Notethat the conjugate N(A)= Πc(A)=ν
(Π (Ac)) ofa possibilitymeasureis a necessity measure such that
ι
(s)=ν
(π
(s)), which explains the interpretation ofι
(s) as a degree of impossibility. It is also called a degree of potential surprise by Shackle [36], or a grade of disbelief by Spohn [37] (who use numerical representationsthereof).Possibilitydistributionsseemtoplay inthequalitativesettingarolesimilartoprobabilitydistributionsinthenumerical setting, replacing sumby maximum.In thisscope, it iswell-known that, inthefinite case,the probabilitydistribution of a probability measure isgeneralizedbythe Möbiustransformof thecapacity. In thissectionwe recallqualitative Möbius transforms that were firstproposed by Grabisch[25] and Mesiar[31], and discuss the analogyit suggestsbetween belief functions andqualitativecapacities.
2.1. QualitativeMöbiustransforms
Definition4.Theinner(qualitative)Möbiustransformofacapacity
γ
isamappingγ
#:2S→L definedbyγ
#(
E) =
n
γ
(
E)
ifγ
(
E) >
maxB(E
γ
(
B)
0 otherwise.
Aset E suchthat
γ
#(E)>0 is calledafocalset.Intheabove definition,due to themonotonicitypropertyof
γ
, theconditionγ
(E)>maxB(Eγ
(B) canbereplacedbyγ
(E)>maxx∈Eγ
(E\ {x}).Itiseasytocheck that•
γ
#(∅)=0;maxA⊆Sγ
#(A)=1;• If A⊂B,and
γ
#(B)>0,thenγ
#(A)<γ
#(B).Let F (
γ
)= {E,γ
#(E)>0} be the family of focal sets associated toγ
. The last property says that the inner qualitative Möbiustransformofγ
isstrictlymonotonicwithrespecttoinclusionoffocalsets.Example 1. Let S= {s1,s2,s3} and
γ
({s1})=γ
({s1,s3})=0.3,γ
({s1,s2})=0.7,γ
({s2,s3})=γ
({s1,s2,s3})=1 (so L is chosen arbitrarily as {0,0.3,0.7,1}), andγ
(A)=0 otherwise. Then F (γ
)= {{s1},{s1,s2},{s2,s3}}, withγ
#({s1})=0.3,γ
#({s1,s2})=0.7 andγ
#({s2,s3})=1. ✷ItisclearthattheinnerqualitativeMöbiustransformofapossibilitymeasure coincideswithitspossibilitydistribution: Π#(A)=
π
(s) if A= {s}and0otherwise.HencethesetoffocalsetsF (Π) ismade ofsingletonsonly.Thispropertymakes itclearthatγ
#generalizesthenotion ofpossibilitydistributiontothepowersetof S.Just like in the numerical setting, the focal sets associated to a necessitymeasure N form a collection ofnested sets
E1⊂E2⊂ · · · ⊂Ek suchthat N(A)=maxEi⊆AN#(Ei).If thenecessitymeasureisbasedonapossibilitydistribution
π
suchN#(Ei)=
ν
(λi−1). A set of the form {s:π
(s)≥ λi>0} is called a cut ofπ
. So the focal sets of N are the cuts of thepossibility distributionofitsconjugate.
More generally, the inner (qualitative) Möbius transform contains the minimal information needed to reconstruct the capacity
γ
since,byconstruction[25,13]:γ
(
A) =
maxE⊆A
γ
#(
E).
(1)Thereadercancheckthatifoneofthevalues
γ
#(E)ischanged,thecorrespondingcapacitywillbedifferent,namelythe valuesγ
(A) suchthatγ
(A)=γ
#(E).Remark2. Anotherqualitativecounterpartof Möbiustransforms,namedouterqualitativeMöbiustransforms,canbeused to represent acapacity
γ
.Itisamappingγ
#:2S→L defined byγ
#(
A) =
n
γ
(
A)
ifγ
(
A) <
minA⊂Fγ
(
F)
1 otherwise
.
(2)Inparticular,
γ
#(S)=1.Duetothemonotonicityofγ
,theconditionγ
(A)<minA⊂F
γ
(F)canbeequivalentlyreplacedbyγ
(A)<mins∈A/γ
(A∪ {s}).Theoriginalcapacityisthenretrievedas[14]:
γ
(
A) =
minA⊆F
γ
#
(
F).
(3)Notethat
γ
(A)isrecoveredfromγ
# viaweightsassignedtosupersetsofset A,whichremindsofoutermeasures,whileγ
# canbecalledaninnerqualitativemassfunction.It is worth noticing that the inner qualitative Möbius transform
γ
#c ofγ
c is related to the outer qualitative Möbius transformγ
# [13,14]:γ
#(
E) =
ν
¡
γ
#c¡
Ec
¢¢
(4) since
γ
(A)<minA⊂Fγ
(F) alsowritesγ
c(Ac)>maxFc⊂Acγ
c(Fc). So if we usecapacities and theirconjugates, usingtheinnerand theouterqualitativeMöbiustransformsconjointlyisredundant.Inthispaper,weonlyuseinnerones.
Example 1 (continued). In the previous example, it can be checked that
γ
#({s2})=
γ
#({s3})=0,γ
#({s1,s2})=0.7,γ
#({s1,s3})=0.3 (and
γ
#(E)=1 otherwise). ✷2.2. Mayqualitativecapacitiesbeinterpretedasqualitativebelieffunctions?
The similarity between capacities and belief functions is striking on Eq. (1). A belief function [34] is a set function 2S→ [0,1]definedby
Bel
(
A) =
X
E:E⊆Am
(
E),
∀
A⊆
S (5)wherem isabasicprobabilityassignment,i.e.aprobabilitydistributionover2S\ {∅}.ThedegreeBel(A)isadegreeofbelief or certaintyinthesensethat Bel(A)=1 means A iscertain, whileBel(A)=0 means A isatbestunknown. Theconjugate set function
Pl
(
A) =
1−
Bel¡
Ac
¢ =
X
E:E∩A6=∅m
(
E),
∀
A⊆
S (6)expresses theidea ofpossibility or plausibility, as Pl(A)=0 means that A isimpossible. The mass assignment m canbe reconstructedfromBel byinvertingthesystemofequationsdefiningthelatter:thisistheMöbiustransform[34].InEq. (1), max replacesthesumintheexpressionofBel.
Hence function
γ
# stands as the qualitative counterpart to a basic probability assignment in the theory of evidence, obtained via a kind of Möbius transform [26]. By analogy with the numerical case,γ
# can be called a basicpossibilityassignment [32] associatedwith
γ
sinceγ
#isapossibilitydistribution onthepowerset of S.Thesubsets E thatreceivea positive supportplay thesameroleasthefocalsetsinDempster–Shafertheory:theyaretheprimitiveitemsofknowledge. Theparallelbetweenpossibilityandprobabilityisconfirmed: thefocalsetsofaqualitativepossibilitymeasureinthesense of theinner Möbius massfunctionare singletons, just asthe focalsets of aprobability measure for thestandard Möbius transformaresingletons.HowevertherearedifferencesbetweennumericalandqualitativeMöbiustransforms.WhileaninnerqualitativeMöbius transform canbeviewed asapossibility distribution onthepower set of S, theconverseis nottrue. Indeed, apossibility distribution
π
m on a space 2S is not requested to satisfy the monotonicity constraint of a qualitative Möbius transform.It means that several basic possibility assignments may yield thesame capacity
γ
(A)=maxE⊆Aπ
m(E), in contrast withthesetting ofevidence theory, wherethereisaone-to-one correspondencebetweenbelieffunctions and basic probability assignments.
Remark3. If
π
m isapossibilitydistribution over2S, monotonicitycanberestoredbydeletingthetermsπ
m(F) whenever ∃E, E⊂F ,π
m(E)>π
m(F). Clearly,if for some F ⊆A,there is E1⊂F ,π
m(E1)>π
m(F), thenγ
(A)=maxE⊆A,E6=Fπ
m(E).So we define a unique qualitative Möbiusmass from
π
m as follows:γ
#(E)=π
m(E)>0 if and only if ∀F ⊂E,π
m(E)>π
m(F), otherwiseγ
#(E)=0. The inner qualitative Möbius transform is the canonical (monotonic) possibility assignment associated withthecapacityinthesensethatallotherpossibilityassignmentscontainredundantinformationwithrespect toreconstructingthecapacityandonlyγ
#canberecoveredinanon-ambiguouswayfromγ
.It is also possible to introduce the counterpart of the contourfunction [34]2 of a belief function (it is
µ
:S→ [0,1] defined byµ
(s)=PE∋sm(E)).Namely,
Definition5.Thequalitativecontourfunction ofacapacity
γ
isthepossibilitydistributionπ
γ(
s) =
maxE∋s
γ
#(
E).
Notethatitisapossibilitydistributionsince
π
γ(s)=1 forsomes∈S.Finally,thequalitativecounterpartofaplausibilityfunction,inconformitywithEq. (6) is:
Π
γ(
A) =
maxA∩E6=∅
γ
#(
E).
It is easyto seethat Πγ is just apossibility measure such that Πγ(A)≥
γ
(A), ∀A.In contrast, inthe numerical setting,replacingmax byasum,and usingamassfunction,one getsaplausibilitymeasure,i.e.aspecialcaseofupperprobability functionmoregeneralthanapossibilitymeasureandaprobabilitymeasure.Itcanbeprovedthatthepossibilitydistribution associated withΠγ ispreciselythequalitativecontourfunction:
Proposition1.∀A,Πγ(A)=maxs∈A
π
γ(s).Proof. To seeitjustnotethat{E:A∩E6= ∅}= {E,∃s∈A∩E}.SoΠγ(A)=maxs∈AmaxE∋s
γ
#(E). ✷Inthenumericalsetting,wedonothavethatPl(A)=maxs∈A
µ
(s)(norhavewePl(A)=Ps∈Aµ
(s)).Themeaningofthequalitativecontourfunctionasaspecial upperpossibilisticapproximationof
γ
willbelaidbareattheendofSection 4.Example1 (continued). ForthecapacityinExample 1,thereadercancheckthat
π
γ(s1)=max(γ
#({s1}),γ
#({s1,s2}))=0.7,π
γ(s2)=max(γ
#({s2,s3}),γ
#({s1,s2}))=1,andπ
γ(s3)=γ
#({s2,s3})=1. ✷In the construction of Dempster [9], a belief function is actually a special case of a lower probability bound induced by a probability space (Ω,P) and a multi-valued mapping, Γ : Ω →2S asBel(A)=P({
ω
∈ Ω : Γ (ω
)⊆A}). Considering the selection functions f of Γ (i.e.,ω
∈ Ω, f(ω
)∈ Γ (ω
)) and the probability functions Pf induced by (Ω,P) on S asPf(A)=P(f−1(A)),itholds(inthefinitecase)thatBel(A)=minf∈Γ Pf(A)[8].
Incontrast,inthequalitativecase,any capacitycanbegeneratedbythisprocess,replacingprobabilitybypossibility.The possibilisticcounterpartofDempsterprocesshasbeenstudiedatlengthbyDuboisandPrade[17] andmorerecentlybyDe Baets and Tsiporkova[39].Given apossibility distribution
π
on Ω,we canthus definean upperpossibility functionΠ =maxf∈ΓΠf(A) and a lower possibility function Π =minf∈ΓΠf(A) where Πf(A)= Π (f−1(A)) and Π is the possibility
measure associatedtoapossibilitydistribution
π
: Ω → [0,1].Sinceanycapacitycanbeexpressedbyamonotonicpossibilityassignment,itcanalwaysbeviewedasalowerpossibility, as weshall see inthe nextsection. Indeed if we choose Ω = F,the set of focalsets of
γ
,and letπ
=γ
#, thenγ
(A)=minf∈ΓΠf(A)= Π, while Π = Πγ. As a consequence, Π and Π are not conjugateset-functions, contrary to the caseof
belieffunctions.Especially,wecannotusetheconjugatefunctionΠc,thatis,thenecessitymeasureinducedbythecontour functionof
γ
asalower boundforthelatter.IndeedNγ
(
A) =
ν
¡Π
γ¡
Ac¢¢ =
min s∈/Aν
³
max E∋sγ
#(
E)
´
=
min E∩Ac6=∅ν
¡
γ
#(
E)¢,
whichcannotbecomparedwith
γ
(A).Summarizing, the situation of qualitative capacities with respect to possibility measures is different from the one of numericalcapacitieswithrespecttoprobabilitiesinseveralrespects:
• Anyqualitativecapacityisalowerpossibility,whilenotany numericalcapacityisalower probability.Someareupper probabilitiesonly (forinstance thevacuouspossibilityfunctionΠ?)and someareneitheruppernorlower probability bounds.Thereisnoprobabilityfunctionactingasthecounterpartofthevacuouspossibilityfunction.Theinterpretation of a belief function as measuring belief or certainty,3 as opposed to plausibility, does not carry over to qualitative capacities,even ifthedefinitionofacapacityvia
γ
# bearsaformalsimilaritywithbelieffunctions, andcapacities can begeneratedaslower boundsofpossibilityfunctions.• Inparticular,sincetheset-function
γ
# canalsobeviewedasapossibilitydistributionoverthepowerset 2S,expression (1) isalsoageneralizationofthedefinitionofthedegreeofpossibilityofasetintermsofapossibilitydistributiononS aspointedoutearlier.
• Theone-to-onecorrespondencebetweenrandomsets(usingMöbiusmassesm)andbelieffunctionsdoesnotcarryover toqualitative capacities:theset of possibilitydistributions on 2S canbe partitionedintoequivalence classes, one per capacity
γ
,eachcontainingtheinnerqualitativeMöbiustransformofsomecapacity,andallpossibilitydistributionsin theclassgeneratingthesameγ
. Itisclearthatγ
# istheleastelement (eventwisely)oftheequivalenceclassattached toγ
.• Thequalitativecounterparttoaplausibilityfunctionisapossibilityfunctionthatdominatesthecapacity,andits possi-bilitydistributionisthequalitativecontourfunctionofthecapacity.
Due to this state of facts, there is a big mathematical and semantic difference between belief functions and qualitative capacities.
Remark 4. Wong et al. [41] have shown that, in order to be representable by a numerical belief function, a qualitative capacitymustsatisfytheaxiom
BEL:
∀
A,
B,
C disjoint, ifγ
(
A∪
B) >
γ
(
A)
thenγ
(
A∪
B∪
C) >
γ
(
A∪
C).
Thisisaweakeningofthewell-known DeFinettiaxiom ofcomparativeprobability: If∀
A∩ (
B∪
C) = ∅
then,
γ
(
B) ≥
γ
(
C)
if and only ifγ
(
A∪
B) >
γ
(
A∪
C),
which is known to beinsufficientto characterize numerical probabilityon finite sets [24].Curiously, even if a capacity
γ
failstosatisfyaxiomBEL,italmostsatisfiesitinthesensethatthefollowingpropertyalwaysholds[14]:∀
A,
B,
C disjoint, ifγ
(
A∪
B) >
γ
(
A)
thenγ
(
A∪
B∪
C) ≥
γ
(
A∪
C).
However, usingcontraposition,thefollowingdualaxiom,anotherspecialization ofthelatterproperty(distinctfromBEL):
PL:
∀
A,
B,
C disjoint, ifγ
(
A∪
B∪
C) >
γ
(
A∪
C)
thenγ
(
A∪
B) >
γ
(
A)
makes
γ
representable byaplausibility function. Itmeans that the weakordering ofevents inducedbyany capacity can berefinedbyaweakorderingrepresentablebyabelieffunctionorbyaplausibility function[14],whichconfirmsthatthe formalanalogybetweencapacitiesandbelieffunctionsproperismisleading.3. Capacitiesaslowerpossibilityanduppernecessityfunctions
It is well-known that a belief function can be equivalently represented by a non-empty convex set of probabilities, namelyP(Bel)= {P,P(A)≥Bel(A),∀A⊆S}.Then itholdsthatBel(A)=minP∈P(Bel)P(A). Inotherwords abelieffunction
is one example of coherent lower probability in the sense of Walley [40] (exact capacity after Schmeidler [35]). The set P(Bel)iscalledthecredalsetrepresenting Bel.4 Iftherangeoftheset-functionisnotequippedwithadditionand product, thisconstructionisimpossible.Thenaturalquestionisthenwhetherasimilarconstructionmaymakesensewithqualitative possibility measures in place of probability measures, using min and max instead of product and sum. From previous preliminary results [17,39,4] we knowtheanswer ispositive. In thispaper,we investigate theproblem indepth, without resortingtoDempstersetting.
3 Bel(A)=1 meansthat A istotallycertain.
3.1. Capacitiesaslowerpossibilities
Thereisalwaysatleastonepossibilitymeasure thatdominates anycapacity:thevacuouspossibilitymeasure,basedon thedistribution
π
? expressingignorance,sinceforanycapacityγ
,∀A6= ∅⊂S,Π?(A)=1≥γ
(A),and Π?(∅)=γ
(∅)=0.Let R(
γ
)= {π
: Π (A)≥γ
(A),∀A⊆ S} be the non-empty set of possibility distributions whose corresponding set-functions Π dominateγ
.Inanalogytothecorrespondingnotioningametheory[35],wecallR(γ
) thepossibilisticcore ofthe capacity
γ
. Clearly, R(γ
), equipped withthepartial orderofspecificity (introducedat thebeginning ofSection 2), is an uppersemi-lattice,that isifπ
1,π
2∈ R(γ
), thenmax(π
1,π
2)∈ R(γ
). Ofcourse,themaximal element ofR(γ
) isthe vacuouspossibilitydistributionπ
?,butithasseveral minimalelements.Wedenote byR∗(γ
)thesetofminimalelements inR(γ
).Any capacity
γ
can bereconstructed usingpossibility measures inducedbypermutations.Let uspresent this resultin details. Letσ
be a permutation of the n= |S| elements in S. The ith element of the permutation is denoted by sσ(i).Moreoverlet Si
σ = {sσ(i),. . . ,sσ(n)}.Definethepossibilitydistribution
π
σγ asfollows:∀
i=
1, . . . ,
n,
π
σγ(
sσ(i)) =
γ
¡
Siσ
¢.
(7)Thereareatmostn!(numberofpermutations)suchpossibilitydistributions,that arecalledthemarginals of
γ
.Example 1 (continued). If we take the permutation
σ
= (1,2,3), we getπ
σγ(s3)=γ
({s3})=0,π
σγ(s2)=γ
({s2,s3})=1,π
σγ(s1)=γ
(S)=1.If we take the permutation
τ
= (3,2,1), we getπ
τγ(s1) =γ
({s1})= 0.3,π
τγ(s2) =γ
({s1,s2})=0.7,π
τγ(s3) =γ
(S)=1. ✷This is similar to the case of a numerical capacity g from which a set of probability distributions pσg of the form
pσg(sσ(i))=g(Sσi )−g(Sσi+1) canbeextracted. If g isabelieffunction,5 theconvexhullofthese probabilitiescoincide with
the(non-empty)credalset P(g).
Similarly,itcanbecheckedthatthemarginalsof
γ
lieinR(γ
)and enableγ
tobereconstructed(alreadyin[4]):Lemma1.Πσγ(A)≥
γ
(A),∀A⊆S.Proof. Consideragivenpermutation
σ
andanevent A.LetiA=max{i,A⊆Sσi }.ThenSσiA isthesmallestsetinthesequence {Siσ}i that contains A andΠσγ(A)= Πσγ(SiσA)=π
σγ(sσ(i)) since A contains sσ(iA) and otherelements in A areof theform sσ(j), j≥iA byconstruction.Hence,byconstruction,Πσγ(A)=γ
(SσiA)≥γ
(A). ✷Proposition2.∀A⊆S,
γ
(A)=minσΠσγ(A).Proof. Fromthelemma,
γ
(A)≤minσΠσγ(A). Conversely, ∀A⊆S, thereis apermutationσ
A such that A=SiσA. Bycon-struction
γ
(A)= ΠσγA(A)≥minσΠγ
σ(A). ✷
Asconsequences, wehavethefollowingresults:
Proposition3.ForallA⊆S,
γ
(A)=minπ∈R(γ)Π (A).Proof. For all
π
∈ R(γ
), Π ≥γ
soγ
(A)≤minπ∈R(γ)Π (A). Moreover there existsσ
such thatγ
(A) = Πσγ(A)≥minπ∈R(γ)Π (A),and, byLemma 1,weknowthat
π
σγ ∈ R(γ
). ✷Proposition4.∀
π
∈ R(γ
),π
(s)≥π
σγ(s),∀s∈S forsomepermutationσ
ofS.Proof. Justconsiderapermutation
σ
inducedbyπ
,that isσ
(i)≥σ
(j) ⇐⇒π
(si)≤π
(sj).Forthispermutation,Π (Sσi)=π
(si)≥γ
(Siσ)=π
σγ(si), ∀i=1,. . . ,n. ✷This resultsays that theset of marginals
π
σγ includes theset R∗(γ
) of leastelements of R(γ
), i.e., themost specificpossibility distributions dominating
γ
. In other terms, R(γ
)= {π
: ∃σ
,π
≥π
σγ}. Not all the n! possibility distributionsπ
σγ are leastelements of R(γ
). As atrivialexample, ifγ
= Π, thisleast element isunique and ispreciselyπ
.But otherpermutations yieldotherless specificpossibilitydistributions.AmoreefficientmethodforgeneratingR∗(
γ
) isprovidedin Section 4.Table 1
Leastelementsofthepossibilisticcore(R∗(γ))for Ex-ample 1.
S s1 s2 s3
π1 0.3 1 0
π2 0.7 0 1
π3 0.3 0.7 1
By constructionR∗(
γ
) isa finiteset ofpossibility distributionsnone ofwhich ismore specificthat another, that is, ifπ
,ρ
∈ R∗(γ
),∃s16=s2∈S,π
(s1)>ρ
(s1) andπ
(s2)<ρ
(s2).Wethus canprovethefollowingbasicresult:Proposition5.∀A⊆S,
γ
(A)=minπ∈R∗(γ)Π (A).Proof. As ∀
π
∈ R∗(γ
),Π ≥γ
, it holdsthatγ
(A)≤minπ∈R∗(γ)Π (A). Nowsuppose ∃A,γ
(A)<minπ∈R∗(γ)Π (A). Then∀
π
∈ R∗(γ
), Π (A)>γ
(A). However there isπ
′∈ R(γ
)\ R∗(γ
) such that Π′(A)=γ
(A)< Π (A). Henceπ
′ is eithermore specific than some
π
∈ R∗(γ
) (which is impossible since R∗(γ
) contains themost specific elements in R(γ
)) orincomparable withall
π
∈ R∗(γ
),whichwouldmeaninsideit,whichiscontrarytotheassumption. ✷Example1 (continued).Forinstance,thethreeleastspecificpossibilitydistributionsthatdominate
γ
inExample 1 aregiven inTable 1.Theyarenotcomparableand onecancheckthatγ
(A)=min(Π1(A),Π2(A),Π3(A)). ✷Thesefindingsalsoshowthatany capacitycanberepresentedbyafinitesetofpossibilitymeasures.Conversely,forany set T ofpossibility distributions,theset-function
γ
(A)=minπ∈T Π (A) isacapacity,and itiseasytoseethat T ⊆ R(γ
).If T onlycontainspossibilitydistributionsthatarenotcomparablewithrespecttospecificity,T = R∗(
γ
),themostspecificelements ofR(
γ
).Remark5.Note that theset function maxπ∈T Π (A) isnot onlya capacity, butalso apossibility measure with possibility distribution
π
max(s)=maxπ∈T
π
(s)[19].Itisinterestingtoconsider theclosureofR∗(
γ
)underthequalitativecounterpartofaconvexcombination[15]:Definition6.If
π
1,. . . ,π
k arepossibilitydistributionsand∀α
1,. . . ,α
k∈L,suchthatmaxi=k 1α
i=1,thequalitativemixtureofthe
π
i’sismaxi=k 1min(α
i,π
i).Theq-convexclosureC(T )oftheset T ofpossibilitydistributionsisthesetofqualitativemixturesofelementsofT.
Note that the qualitative mixture of possibility distributions is a possibility distribution (it is normalized) and maxki=1min(
α
i,Πi) is a possibility measure with distribution maxki=1min(α
i,π
i) [19]. We then define the q-convexcoreof the capacity
γ
by C(R∗(γ
)), which possesses a maximal element, i.e., the possibility measure with distributionmaxπ∈R∗(γ)
π
. ClearlyC(R∗(γ
))is in general aproper q-convex subsetof R(γ
), that is analogousto anumerical credalset.6 The minimalset R
∗(
γ
) plays thesamerole asextreme probabilitiesincredal sets(the verticesofa convexpolyhe-dron), fromwhichthelattercanbereconstructedviaconvexclosure.
3.2. Capacitiesasuppernecessities
We can dually describe capacity functions as uppernecessities by means ofa family of necessityfunctions that stem from thelower possibilitydescriptionof theirconjugates.Clearly,possibility measuresthat dominate
γ
c areconjugatesofnecessitymeasuresdominatedby
γ
.Inotherwordsγ
isalsoanuppernecessitymeasureinthesense thatProposition6.
γ
(
A) =
maxπ∈R∗(γc)
N
(
A).
Proof.
γ
c(A) = minπ∈R∗(γc)Π (A). Hence,γ
(A) can also be expressed as:γ
(A) =ν
(minπ∈R∗(γc)Π (Ac)) =
maxπ∈R∗(γc)
ν
(Π (Ac)). ✷However, this result can be directly obtained from the expression (1) of
γ
. Indeed, if E is a focal set ofγ
, define the necessity measure NE by ∀A6=S, NE(A)=γ
#(E) if E⊆ A and0 otherwise (its focal sets are E, S, with qualitative Möbiusweights NE#(A)=γ
#(E) andNE#(S)=1).Itisclearthatγ
(A)=maxE∈F (γ)NE(A).Thisisnottheminimalformofcourse.Togettheminimalformonemayconsiderallmaximalchainsofnestedfocalsubsets Ci= {E1
i ⊂ · · · ⊂E ki
i }⊆ F (
γ
),i=1,. . . ,n:eachsuchchainCi definesanecessitymeasure Ni whosenestedfocalsetsareasfollows: • if
γ
#(Eiki)=1 thenF (Ni)= Ci and Ni#(Eki)=γ
#(Eki)• otherwiseF (Ni)= Ci∪ {S} and Ni#(Eik)=
γ
#(Eki), Ni#(S)=1.Itiseasytocheck thatiftherearen maximalchainsofsubsetsinF (
γ
),γ
(
A) =
maxn i=1 maxE∈Ciγ
#(
E) =
n max i=1 Ni(
A)
(8) and Ni(E)=γ
#(E), ∀E∈ Ci.Example1 (continued). The set of focal sets of
γ
in Example 1 containstwo maximal chains: ({s1},{s1,s2}) and {s2,s3} withrespectiveweights(0.3,0.7),and1.Henceweneedtwonecessitymeasurestorepresentγ
:N1 withfocalsetsdefined by N1#({s1})=0.3, N1#({s1,s2})=0.7, N1#(S)=1; and a Boolean necessity measure N2 with one focal set defined byN2#({s2,s3})=1.Itcanbecheckedthat
γ
(A)=max(N1(A),N2(A)). ✷By construction, the cuts of possibility distributions
π
i underlying Ni and different from S are focal sets ofγ
. Thesepossibilitydistributionsarenotcomparablewithrespecttospecificity:iftheywere,theywouldnotcorrespondtomaximal nestedchains offocalsets.Inotherwords:
Proposition7.ThesetofpossibilitydistributionswhosecutsarebuiltfromthemaximalnestedchainsinF (
γ
)isthesetR∗(γ
c)ofpossibilitydistributionsdominatingtheconjugateof
γ
.Proof. The
π
i’sinducedbymaximalchains Ci dominateγ
c sincethecorresponding Ni isdominatedbyγ
.Thesedistribu-tionsarenotcomparableand theΠi’sgenerate
γ
c duetoEq. (8). ✷The conjugatecapacity isthus such that
γ
c(A)=minni=1Πi(A) whereΠi are theconjugatesof the Ni inducedbythemaximalchainsinF (
γ
).LetC(π
) bethesetofcutsofπ
.Ofcourse,C(π
i)= F (Ni).Wecanshowthatthefocalsetsofγ
arepreciselythecutsofall
π
i’sinR∗(γ
c)(uptoset S).Proposition8.F (
γ
)⊆Sπ∈R∗(γc)C(
π
)⊆ F (γ
)∪ {S}.Proof. If E is focal for
γ
thenit appears in one maximal chainCi henceit is acut ofπ
i∈ R∗(γ
c). Conversely, if E6=Sis acutof
π
∈ R∗(γ
c), itisafocalset of some N induced byamax-chain C offocalsets ofγ
. If E=S isacut ofsomeπ
∈ R∗(γ
c),π
correspondstoamaximalchainwhoselargest setG maydifferfrom S (whenγ
#(G)<1). ✷Soitisinterestingtonoticethatthereisalmostanidentitybetweenthefocalsetsofacapacityandthecutsofthemost specificpossibilitydistributionsthat generateitsconjugate.
Example1 (continued).In Example 1,
γ
(A)=max(N1(A),N2(A))where N1(A) is inducedbyπ
1(s)=1 if s=s1,ν
(0.3)= 0.7 ifs∈ {s1,s2}\ {s1},i.e.,s=s2 andν
(0.7) otherwise,thatiss=s3.Wecheck thatF (N1)= {{s1},{s1,s2},S}arethecuts ofπ
1.AndN2 isinducedbyπ
2(s)=1 ifs∈ {s2,s3}and 0 otherwise. ✷Oneoftherepresentationsof
γ
bymeans ofR∗(γ
) orR∗(γ
c) maycontainless elementsthantheother.Forinstance,if
γ
isanecessitymeasurebasedonpossibilitydistributionπ
,thenR∗(γ
c)= {π
}whileR∗(
γ
)containsseveralpossibilitydistributions(whosefocalsetsaresingletonsor S). NotethatΠ (A)≥N(A)= Πc(A),sothatitlooksmorenaturaltoreach
N frombelowandΠ fromabove.ForthecapacityinExample 1,weseethatitneedsthreepossibilitymeasurestorepresent it, but only two necessitymeasures areneeded. The results of thissection generalizeto several possibility measures (the leastonesthatdominate
γ
)theobviousremarkthatthecutsofapossibilitydistributioninducingapossibilitymeasureare thefocalsets ofthecorrespondingnecessitymeasure.3.3. Sugenointegralsasupperorlowerpossibilisticexpectations
Thefact that aconvexcapacity (orabelieffunction)coincides withalower probabilitycan becarriedout tointegrals. Namely, theChoquet integral of areal-valued function with respectto a numericalconvex capacity is equalto the lower
expectationwithrespecttoprobabilitiesinthecredalsetinducedbythisconvexcapacity[10].However,thisisnottruefor any capacityand anyconvexprobability set.Inthissection,weexamine thesituationwithqualitativecapacitiesforwhich acounterpartofChoquetintegralexists,namelySugenointegral (see[29] foranoverviewofthisoperator).
Let f :S→L beafunctionthatmayserveasautilityfunctionif S isaset ofattributes.Sugenointegralisoftendefined asfollows[38]:
Sγ
(
f) =
maxλ∈L min
¡λ,
γ
(
Fλ)
¢
(9) where Fλ= {s: f(s)≥ λ} isthe set of attributeshaving best ratingsfor some object, above threshold λ, and
γ
(A) isthedegreeofimportanceoffeatureset A.
When
γ
isapossibilitymeasure Π,Sugenointegralsimplifies intothepossibilityintegral[23]: SΠ(
f) =
maxs∈S min
¡
π
(
s),
f(
s)
¢
(10)whichistheprioritizedmax operator.
AnequivalentexpressionofSugenointegralis[29]: Sγ
(
f) =
max A⊆Smin³
γ
(
A),
min s∈A f(
s)
´
.
In thisdisjunctive form, the set-function
γ
can be replaced without loss of information by the inner qualitative Möbius transformγ
# definedearlier.Sγ
(
f) =
max A∈F (γ)min
¡
γ
#(
A),
fA¢
(11)where fA=mins∈A f(s). This form makesit clear that itsimplifies into the possibility integral (10) if
γ
= Π. The aboveexpressionofSugenointegralhasthestandardmaxminformofapossibilityintegral,viewing
γ
#asapossibilitydistribution over2S.ThesimilarityofSugenointegralintheform(11) withthediscreteformofChoquetintegralusingMöbiustransform [27] ispatent.Andindeed,ithasbeenprovedelsewhere[13,14] thatitispossibletorefinetheorderinginducedbySugeno integral with respectto a qualitativecapacityon the set offunctions f by means ofa Choquet integral withrespectto a belieffunctionthat refinestheorderinginducedbythecapacity.YetanotherequivalentexpressionofSugenointegralis[29]: Sγ
(
f) =
min A⊆Smax¡
γ
¡
Ac¢,
fA¢
(12) where fA=maxs∈A f(s). Likewise whenγ
is a necessity measure N = Πc, based on possibility distributionπ
, Sugenointegral simplifiesintoanecessityintegral[23],asclearintheform(12): SN
(
f) =
mins∈S max
¡
ν
¡
π
(
s)¢,
f(
s)
¢
(13)whichistheprioritizedmin operator.Besidesitiseasytocheckthat theconjugacypropertyextendstoSugenointegralas Sγ
(
f) =
ν
¡
Sγc
¡
ν(
f)¢¢.
(14)Asaconsequenceofresultsinthissection,itcanbeprovedthatSugenointegralisalowerprioritizedmaximumandan upperprioritizedminimum:
Proposition9.Sγ(f)=infπ∈R∗(γ)SΠ(f)=supπ∈R
∗(γc)SN(f).
Proof. Viewing
γ
as a lower possibility (Proposition 5) and using the fact that maxxminy f(x,y) is never greater thanminxmaxy f(x,y),itcomes: Sγ
(
f) =
max A⊆Smin³
min π∈R∗(γ)Π (
A),
fA´
=
max A⊆Sπ∈Rmin∗(γ) min¡Π(
A),
fA¢
≤
min π∈R∗(γ) max A⊆Smin¡Π(
A),
fA¢.
HencewehaveSγ(f)≤infπ∈R∗(γ)SΠ(f).Conversely, let
π
f be the marginal ofγ
obtained from the nested sequence of sets Fλ induced by function f , thenit is clear that Πf(Fλ)=
γ
(Fλ), and thus Sγ(f)= SΠf(f). As ∃π
∈ R∗(γ
),π
f ≥π
, by definition, SΠf(f)≥ SΠ(f)≥infπ∈R∗(γ)SΠ(f). Finally, using conjugacy property (14), the other property showing that Sugeno integral is an upper
Remark6. The obtained equality is not surprising because Sugeno integral with respect to a functioncoincides with the possibilityintegral withrespecttoaspecificmarginalpossibilitydistributioncomputedfromthepermutation
σ
f suchthatf(sσf(1))≤ f(sσf(2))≤ · · · ≤ f(sσf(|S|)),just asthediscreteChoquet integralisaweighted arithmeticmeanwithrespectto
theprobabilitydistributioncomputedfromthecapacitywiththesamepermutation.
Thesignificanceoftheabove resultstems fromthefactthat iftheset ofminimalpossibility (resp.necessity)measures dominating (resp.dominated by)a qualitativecapacityis very small,thecomputationalcomplexity ofthecapacityis low. Namely, representing acapacityistheoreticallyexponential inthenumberofelements |S| of S,whilerepresentinga pos-sibility oranecessitymeasure islinear(wejust needtoknowthe|S|possibilityvaluesofsingletons).If R∗(
γ
) containskdistributions, onlyk|S| valuesare neededto represent
γ
.Adirect consequenceis tocut downthecomplexity theSugeno integral computation with respectto suchcapacities accordingly (fromexponential to linear),which may beinstrumental inoptimization problemsusingSugenointegralasacriterion.4. Computingtheminimaldominatingpossibilitymeasuresinthequalitativeconvexcore
Inthissectionweconsidertheproblemofdeterminingtheminimalsetofn possibilitydistributions
π
ithataresufficienttogenerateagivenqualitativecapacity
γ
intheformγ
=minni=1Πi.Theaimofthissectionistoshowthat thequalitativeMöbiustransformisinstrumentalinfindingtheseleastelements.
4.1. Selectionfunctions
We needthe notion ofa selection functionfrom F (
γ
), theset of focalsets ofγ
. Aselection function sel: F (γ
)→Sassigns to eachfocalset A anelement s=sel(A)∈A.Wedenote by Σ (F (
γ
)) theset ofselection functionswith domain F (γ
). Given a capacityγ
for any selection function on F (γ
), one can define a possibility distributionπ
selγ by letting max∅=0 andπ
selγ(
s) =
maxE:sel(E)=s
γ
#(
E),
∀
s∈
S.
(15)Note that itparallels a similarconstruction inthe theoryof belieffunctions, wherethe probability assignments pBelsel(s)= P
A:sel(A)=sm(A), ∀s∈S turn out to bethevertices ofthe credalset inducedby Bel.In fact, similarresults for
π
γ sel asforthen! possibilitydistributions
π
σγ canbeobtained. Note that ifγ
= Π,thenthere isonly onepossible selectionfunction(sincefocalsets aresingletons)and
π
Πsel=
π
(asrecalledabove inSection 2.1).Proposition10.Foranyselectionfunctionsel inΣ (F (
γ
)),Πselγ(A)≥γ
(A),∀A⊆S. Proof.Π
selγ(
A) =
max s∈Aπ
γ
sel
(
s) =
maxs∈A E:selmax(E)=sγ
#(
E)
≥
maxs∈A E⊆Amax:sel(E)=s
γ
#(
E) =
γ
(
A)
sinceforafocalset E⊆A,sel(E)∈A too,then{E∈ F (
γ
)⊆A: ∃s∈A:sel(E)=s}= {E∈ F (γ
):E⊆A}. ✷ Infact,notallpossibilitydistributionsπ
selγ areoftheformπ
σγ forapermutationσ
.Example2 (Counterexample). Suppose S= {s1,s2,s3},
γ
#({s1,s3})=1,γ
#({s2,s3})= λ<1. Then consider sel({s1,s3})=s3 and sel({s2,s3})=s2, so thatπ
selγ(s3)=1>π
selγ(s2)= λ>π
selγ(s1)=0. Using the corresponding permutation, note thatπ
σγ(s3)=1 butπ
σγ(s2)=γ
({s1,s2})=0. However if we choose another selection function sel′ such that sel′({s1,s3})=sel′({s2,s3})=s3,
π
selγ′ correspondstotheaboveπ
γ σ. ✷
Neverthelesswedohavethat theset {
π
selγ :sel∈ Σ(F (γ
))}alsoreconstructsγ
.Proposition11.∀A⊆S,
γ
(A)=minsel∈Σ (F (γ))Πselγ(A).Proof. Πselγ(A)=maxs∈AmaxE:sel(E)=s
γ
#(E).Togettheequality,let E⊆A besuchthatγ
(A)=γ
#(E).Choosetheselection function as follows: assignγ
#(E) to an element sE ∈E; then ifγ
#(F)>γ
#(E), assignγ
#(F) to some element not in A. Thisispossible,becauseifγ
#(F)>γ
#(E),thenF *E,and sinceγ
#(F)>γ
(A), F*A either.Forsuchaselectionfunction,Πselγ(A)=
π
selγ(sE) holds since the only elements in A to which a possibility weight is assigned are sE ∈E, and possiblyother sC∈C∩A6= ∅,suchthat
γ
#(C)<γ
#(E). Soγ
(A)≥minsel∈Σ (F (γ))Πselγ(A) anddue toProposition10,theconverseis4.2. Analgorithmbasedonusefulselectionfunctions
DuetoProposition 11,thesetofminimalelements(maximallyspecific)ofR(
γ
)isalsoincludedin{π
selγ :sel∈ Σ(F (γ
))}. Butnotallselectionfunctionsareofinterestsincewecanhaveredundantpossibilitydistributions.Example3.We consider acapacity
γ
suchthat the set offocalsets has acommon intersection I. Wedenote E thefocal set suchthatγ
#(E)=1 and F afocalset suchthatγ
#(F)<1.Wecandefineapossibilitysuchthat
π
selγ(s∗)=1 forsome s∗∈I⊆E andπ
γsel(s)=
γ
#(F) forsome s∈F \I. Butwecandefineamorespecificpossibilitydistributiondominating
γ
:π
selγ(s∗)=1 forsomes∗∈I⊆E and0 otherwise. ✷Definition7.Aselectionfunctionsel issaidtobeuseful foracapacity
γ
ifitisaminimalelementin{π
selγ :sel∈ Σ(F (γ
))}. Letusconsiderthefollowingalgorithmtocalculateselectionfunctionssel andtheassociatedpossibilitydistributionπ
selγ. Algorithm MSUP:MaximalspecificupperpossibilitygenerationInput: n focalsets Ek k=1,. . . ,n rankedindecreasingvaluesof
γ
#(Ek)i.e.,
γ
#(E1)≥ · · · ≥γ
#(En) F ← {E1,· · · ,En}foralls∈S let
π
(s)=0repeat
E←thefirstelementofF
Define sel(E)=s forsomes∈E
Let
π
(s)=γ
#(E)Delete E fromF
T ← {G∈ F such that s∈G}
repeat
G←thefirstelement ofT Definesel(G)=s
DeleteG fromT DeleteG fromF.
untilT = ∅;
untilF = ∅;
Result: Apossibilitydistribution
π
selγ.In fact theabove algorithm appliesthe trick inthe proof ofProposition 11:the value
γ
(A)=γ
#(Ej) is retrievedbyapossibility distribution wheretheweights
γ
#(Ek),k< j areassignedoutside A (which isoneoption offeredbythe aboveprocedure). A selection function sel thus generated satisfies the following property: If sel(E)=s for some E∈ F (
γ
) andπ
sel(s)=γ
#(E),then∀F∈ F (γ
),suchthats∈F :• if
γ
#(E)≥γ
#(F)thensel(F)=s• if
γ
#(E)<γ
#(F) thensel(F)6=s.Conversely, apossibilitydistribution
π
selsatisfyingtheabove propertiesisapossiblesolutionforthealgorithmMSUP.Let Σ∗(F (
γ
))betheset ofselectionfunctionsgeneratedbyrepeatedapplicationsofalgorithm MSUP,and MSUP(γ
)bethecorrespondingsetofpossibilitydistributions.
Notethatwhilenotallselectionfunctions
π
selyieldapermutationσ
suchthatπ
sel=π
σ,thisistrueforusefulselectionfunctions. Thisis due to the orderingof elements generatedby theprocedure from great to small masses
γ
#(Ej):defineσ
from the sequence of elements sj obtained by one application of the algorithm constructing sel. Forπ
sel, it is clearthat
π
sel(sσ(i))≤π
sel(sσ(j)),∀i> j.Henceπ
σ(i)=γ
(Sσi )= Πsel(Sσi )=π
sel(i). Moreover,if two possibilitydistributionsaregenerated bythealgorithm,theycorrespondtodifferentpermutations (sinceeachtime,differentelements receivepositive masses).So weconclude thatMSUP(
γ
)isasubsetof{π
σ :σ
permutation of S}.Proposition12.If
π
6=ρ
∈MSUP(γ
),thenneitherπ
>ρ
norπ
>ρ
hold.Proof. Suppose
π
>ρ
generated byalgorithm MSUP, i.e.,π
≥ρ
and lets be such thatπ
(s)>ρ
(s). Let selπ and selρ be the selection functions associated withπ
andρ
, respectively. Let E be such that selπ(E)=s andπ
(s)=γ
#(E)>ρ
(s). By construction, selρ(E)=s′6=s. We haveρ
(s′)≥γ
#(E)>
ρ
(s), soπ
(s′)≥ρ
(s′)>0 and ∃G, such that selπ(G)=s′ andπ
(s′)=γ
#(G). But since by construction, s′∈E∩G, it follows that
γ
#(G)≥γ
#(E) and the algorithm MSUP would en-force selπ(E) =s′, which contradicts the assumption that selπ(E)=s. So
π
≥ρ
withπ
(s)>ρ
(s) is in conflict with theSothepossibilitydistributionsobtainedwiththealgorithmMSUParenotpairwisecomparable(theyarebasedonuseful selections).
Example 4. Let S= {s1,s2,s3,s4} and the focal sets F (
γ
)= {{s1,s2},{s1,s3},{s2,s3},{s2,s4}} whereγ
#({s1,s2})=1>γ
#({s1,s3})= .8>γ
#({s2,s3})= .4>γ
#({s2,s4})= .2. We show allpossible applications of the MSUP algorithm, startingwithF = F (
γ
).• sel({s1,s2})=s1,
π
(s1)=γ
#({s1,s2})=1,T = {{s1,s3}},and F = {{s2,s3},{s2,s4}}.– sel({s2,s3})=s2,
π
(s2)=γ
#({s2,s3})= .4,T = {{s2,s4}},and F = ∅.So,
π
(s1)=1,π
(s2)= .4,π
(s3)=π
(s4)=0.Thepermutationσ
canbe(1,2,3,4).– sel({s2,s3})=s3,
π
(s3)=γ
#({s2,s3})= .4,T = ∅and F = {{s2,s4}}.∗ sel({s2,s4})=s2,
π
(s2)=γ
#({s2,s4})= .2,F = ∅.So,
π
(s1)=1,π
(s2)= .2,π
(s3)= .4,π
(s4)=0.Thepermutationσ
is(1,3,2,4).∗ sel({s2,s4})=s4,
π
(s4)=γ
#({s2,s4})= .2,F = ∅.So
π
(s1)=1,π
(s2)=0,π
(s3)= .4,π
(s4)= .2 Thepermutationσ
is(1,3,4,2).• sel({s1,s2})=s2,
π
(s2)=γ
#({s1,s2})=1,T = {{s2,s3},{s2,s4}}and F = {{s1,s3}}.– sel({s1,s3})=s1,
π
(s1)=γ
#({s1,s3})= .8,F = ∅.Wehave
π
(s1)= .8,π
(s2)=1,π
(s3)=π
(s4)=0.Thepermutationσ
canbe(2,1,3,4).– sel({s1,s3})=s3,
π
(s3)=γ
#({s1,s3})= .8,F = ∅.Wehave
π
(s1)=0,π
(s2)=1,π
(s3)= .8,π
(s4)=0.Thepermutationσ
canbe(2,3,4,1). ✷Moreoverthefamilyofpossibilitydistributionswhich canbeobtained withthealgorithmMSUPisnecessaryand suffi-cienttoreconstructthecapacity.Inordertoprovethismainresultthefollowingoneisneeded.
Proposition13.Foranypermutation
σ
,thereexistsaselectionfunctionsel correspondingtoanotherpermutationτ
suchthatπ
σ ≥π
τ =π
sel.Proof. Let
σ
beapermutationof S,andwedenotethecorrespondingelementsassi forsimplicity.Letπ
(si)=γ
(Si)whereSi= {si,. . . ,sn}.Letusshowthat wecanderive ausefulselectionfunctionfrom
σ
.For eachsi weshalldecidewhetherornottomakeitaselectionfrom E∈ F (
γ
).Let U bethesetofelementsnotyetdecided,andF thecurrentset offocalsets.• Let U =S and F = F (
γ
). Ifπ
σ(sj)=0 for j=i,. . . ,n, theelements sj must be discarded from S and will not beselectedfromfocalsets.SetU=: {s1, . . . ,si−1}.
• For j=i−1 downto1
– Select E∈ F such that sj∈E, E⊆Sj,
γ
#(E) ismaximal,and letsel(E)=sj.If thereisnone, letπ
sel(sj)=0.Deletesj from U .
– Letsel(F)=sj forall F suchthat sj∈F and F⊆Sj and
γ
#(F)≤γ
#(E),anddeletesuch F fromF.– Set j=: j−1.
Itisclearthat theobtainedselectionfunctionsel is useful.Moreover, itisalsoobviousthat
π
sel(si)≤π
(si) sinceπ
sel(si)=γ
#(Ei)forsome Ei⊆Si.Nowifweusethepermutationτ
suchthatπ
sel(sτ(i))≤π
sel(sτ(j))wheneveri> j,thenwehavethat
π
sel=π
τ (this isbecauseπ
sel(sτ(i))<π
sel(sτ(j))ifandonlyifγ
#(Ei)<γ
#(Ej)). ✷Example5(Generationofausefulselectionfromapermutation).Consider S= {s1,· · · ,s6}.Let
γ
besuchthatγ
#({s1,s2})=1>γ
#({s2,s3})= .8>γ
#({s5,s6})= .7>γ
#({s3,s4,s5})= .6>γ
#({s1,s6})= .5.Consider the permutation
σ
= (s6,s5,s4,s3,s2,s1). It is clear thatπ
σ(si)=1, ∀i<6 andπ
σ(s6)=0. This possibilitydistributionisaverylooseupperboundof
γ
.Letusfindamoreinformativepermutationviaaselection. WehaveU=S andF = {{s1,s2},{s2,s3},{s5,s6},{s3,s4,s5},{s1,s6}}.1. Firstwecandelete s1 fromU and
π
sel(s1)=0; U= {s2,· · · ,s6}.2. Thenweselect E= {s1,s2} andsel({s1,s2})=s2.Wealsoletsel({s2,s3})=s2.
NowU= {s3,s4,s5,s6} andF = {{s5,s6},{s3,s4,s5},{s1,s6}}.
3. Wecanselectnofocalset in S3= {s1,s2,s3},norin S4= {s4,s3,s2,s1}.So
π
sel(s3)=π
sel(s4)=0.NowU= {s5,s6}and F = {{s5,s6},{s3,s4,s5},{s1,s6}}.
4. Thenweselect E= {s3,s4,s5}⊂S5= {s1,· · · ,s5} andsel({s3,s4,s5})=s5.
NowU= {s6}andF = {{s5,s6},{s1,s6}}.
5. Finallyselect E= {s5,s6} andletsel({s5,s6})=sel({s1,s6})=s6.