Contents lists available atScienceDirect
Games
and
Economic
Behavior
www.elsevier.com/locate/geb
Efficiency
and
stability
of
probabilistic
assignments
in marriage
problems
✩
Battal Do˘gan
a,
Kemal Yıldız
b,∗
aFacultyofBusinessandEconomics,UniversityofLausanne,Switzerland bDepartmentofEconomics,BilkentUniversity,Turkey
a
r
t
i
c
l
e
i
n
f
o
a
b
s
t
r
a
c
t
Articlehistory:
Received20July2014
Availableonline17December2015
JELclassification: C60 C71 C78 D61 Keywords: Marriageproblems Probabilisticassignment Efficiency Stability
We study marriage problems where two groups of agents, men and women, match each other and probabilistic assignments are possible. When only ordinal preferences are observable, stochastic dominance efficiency (sd-efficiency) is commonly used. First, we provide a characterization of sd-efficient allocations in terms of a property of an orderrelationdefined onthesetofman–womanpairs.Then,usingthischaracterization, we constructivelyprovethatforeachprobabilisticassignmentthatissd-efficientforsome ordinal preferences,there isavon Neumann–Morgensternutility profileconsistentwith the ordinal preferences for whichthe assignment is Pareto efficient. Second, we show that when the preferences are strict, for each ordinal preference profile and each ex-poststableprobabilisticassignment,thereisavon Neumann–Morgensternutilityprofile, consistentwiththeordinal preferences,forwhichtheassignmentbelongstothecoreof theassociatedtransferableutilitygame.
©2015ElsevierInc.All rights reserved.
1. Introduction
Thetheoryoftwo-sidedmatchingproblemshasbeenusefulinprovidingsolutionstomanyreal-lifeeconomicproblems (seeRoth andSotomayor,1990).A marriageproblem,whichconstitutesabasisfortwo-sidedmatchingproblems,consists of two equal-sized groups of agents: menand women. Each man has ordinal preferences over women and vice versa. A majorityoftheliteratureisfocused ondeterministicassignments,wheremenandwomenarematchedone-to-one.One canalsothinkofprobabilisticassignments,whicharelotteriesoverdeterministicassignments,andtheseareinterestingfor atleast two reasons: (1) agents maymatchin fractions; forexample, a consultantmay allocatehis time among several firms,1 (2) probabilisticassignments mayhelpusachieve fairnesswhenit isnot possiblewithdeterministic assignments (e.g. Bogomolnaiaand Moulin, 2001). Here, we consider probabilistic assignments, andin particular their “stability” and “efficiency”.
Weconsidera modelwhereonlyordinalpreferenceinformationisavailable.That is,foreach man,all weknowishis preferenceorderingoverwomen,andviceversa,whichisinlinewiththeapplicationsandthetheoreticalliterature.Most allocationrulesthathavebeendiscussedinthisliteratureelicitonlyordinalpreferences,as opposed toutility information
✩ WearegratefultoPauloBarelli,SrihariGovindan,BettinaKlaus,EfeOk,ArielRubinstein,WilliamThomson,RakeshVohra,twoanonymousreferees,the
associateeditor,seminarparticipantsatBilkentUniversityandUniversityofRochester,andparticipantsattheBosphorousWorkshoponEconomicDesign andWorkshoponSocialChoiceandMechanismDesignattheUniversityofManchesterfortheirhelpfulcomments.
*
Correspondingauthor.E-mailaddresses:battaldogan@gmail.com(B. Do˘gan),kemal.yildiz@bilkent.edu.tr(K. Yıldız).
1 SeeManjunath (2014)andBogomolnaiaandMoulin (2004)formoreexamplesoffractionalmatchingmarkets. http://dx.doi.org/10.1016/j.geb.2015.12.001
over possible mates.2 When only ordinal preference information is available, extensively studied efficiency and stability notionsarestochasticdominanceefficiency (sd-efficiency)andex-poststability.We askwhetherprobabilisticassignmentsthat aresd-efficientorex-poststableforordinalpreferencesarepossiblyefficientorstableforcardinalpreferences.
The maincontributionofthispaperpertainstostability,whichisa centralrobustness conditionfortwo-sided match-ings. A deterministicassignment is stableifno pair ofa man and awoman prefers each other to their assignedmates. A probabilisticassignmentisex-poststableifitcanbeexpressedasalotteryoverstabledeterministicassignments.In the deterministiccase,if anassignmentisstable,thenitisalsointhecoreofanassociatednon-transferableutility(NTU)game (RothandSotomayor,1990).It followsthatinanordinalenvironmentwheremonetarytransfersarenotallowed,if an as-signmentisex-poststable,thenthereisnoincentiveforex-post3 groupdeviation.However,ex-ante,4 agroupofmenand
women maybreakawayfromtherestofsocietyandmayformmatchesamongthemselvesthatmakeeachofthembetter off in terms ofexpected utility.This possibility motivatesus to ask whetherex-post stability implieswelfare properties deduciblefromtheordinalpreferencesthatwouldpreventsuchbreakaways.
Weshowthatwhenpreferencesarestrict(noagentisindifferentbetweentwodifferentagents),foreachex-poststable probabilistic assignment, there is a utility profile consistent with the ordinal preferences such that no group of agents consistingofequalnumbersofmenandwomencandeviatetoaprobabilisticassignmentamongthemselvesandmakeeach memberbetteroff.In fact,we proveanevenstrongerresult(Theorem 2):foreachex-poststableprobabilisticassignment, thereisautilityprofileconsistentwiththeordinalpreferencessuchthatnogroupofagentsconsistingofequalnumbersof menandwomencandeviatetoaprobabilistic assignmentamongthemselvesinwhichthesumoftheir expectedutilities is greater.Put differently,we associatewitheachutility profilea transferableutility (TU)game,as inShapleyandShubik (1971),in whichacoalitionisadmissibleifitconsistsofequalnumbersofmenandwomen,andtheworth ofeachsuch coalitionisthemaximaltotalexpectedutilityitcanachievebyformingmatchesamongthemselves.5We showthatforeach
ex-poststableassignment,thereisautilityprofilesuchthattheassignmentbelongstothecoreoftheassociatedTU-game (Theorem 2).Thatis,evenifweweretoallowmonetarytransfers,therewouldbenoprofitableex-antegroupdeviation.
Our other result pertains to efficiency. A natural efficiency requirementfor probabilistic assignments is sd-efficiency, which is based on the first-order stochastic dominance relation. A probabilistic assignment stochastically dominates (sd-dominates) another if for each agent, the probability distribution assigned to that agent in the former first-order stochastically dominatestheprobability distributionassignedtothatagent inthelatterassignment.Assignments thatare undominated inthissensearecalled“sd-efficient”.6 We askthefollowingquestion:Consider anordinalpreferenceprofile
andan sd-efficient assignment.Does thereexista utility profileconsistent withtheordinal preferencesatwhich the as-signmentisParetoefficient?We showthatforeachordinalpreferenceprofile,andforeachsd-efficientassignment,onecan constructautilityprofileconsistentwiththeordinalpreferencessuchthatthesumoftheexpectedutilitiesoftheagentsis maximizedatthatassignment(Theorem 1).To provetheresult,we characterizethesd-efficiencyofanassignmentinterms ofapropertyofanorderrelationthatwedefineoverthesetofman–womanpairs.7
Our results regarding efficiency are intimately related to some recent results. Carroll (2011) proves a counterpart of
Theorem 1inamoregeneralsocialchoicesetup,fromwhichTheorem 1canbeobtainedasacorollary.Azizetal. (2015)
provide an interesting non-expectedutility generalizationof Carroll (2011).Ourresultsare basedon anorder theoretical analysis; we characterize the sd-efficiency ofan assignment in terms ofthe acyclicity of abinary relationon the set of man–womanpairs,whichparallelstheresultsbyBogomolnaiaandMoulin (2001)andKattaandSethuraman (2006).In this vein,Azizetal. (2015)notethatitdoesnotseempossibletoextendthischaracterizationtoCarroll’s(2011)ortheirgeneral socialchoicesetting.Moreover,theutility profileconstructedinTheorem 1clearly relatestotheutilityprofileconstructed inTheorem 2,whichshedsfurtherlightontherelationbetweensd-efficiencyandex-poststability.8
Ourresultsforstability,unlike efficiency,cannotbe directlyrelatedtotherecentfindings onthegeneralsocialchoice setup.To proveTheorem 2,we observeaninterestingpropertyofex-poststableassignmentsrelatedtothelatticestructure of stabledeterministic assignments (see Knuth,1976,pp. 92–93, who attributesthe discovery ofthislattice structure to J.H. Conway). InProposition 3,we showthat each ex-poststableprobabilistic assignment canbe decomposedintoa col-lectionofdeterministic stableassignments,whichcanbeorderedinsuch awaythateachman’swelfareisnon-increasing andeachwoman’swelfareisnon-decreasingaswefollowtheassignments fromthefirsttothelast.Thisresult,whichwe show viatheroundingapproach duetoTeoandSethuraman (1998),playsakeyroleintheproofofTheorem 2.A corollary ofTheorem 2isthatex-poststabilityimpliessd-efficiencywhenpreferencesarestrict,whichisindependentlyshownalso byManjunath (2011).
2 Onejustificationforwhysuchmechanismsarecommonisthatitisacomplexprocessforanagenttoformulatehisutilityinformation.SeeBogomolnaia andMoulin (2001)foradetaileddiscussion.
3 Aftertherealizationofadeterministicassignment.
4 Thatis,beforetherealizationofaparticulardeterministicassignment. 5 Weassumepreferencestobequasilinearinmoney.
6 Thisnotionisusually referredtoas“ordinalefficiency”,startingfromBogomolnaiaandMoulin (2001).Here,we usetheterminologyofThomson (2010).
7 Fortheproblemofassigningobjects,similarcharacterizationsaregivenbyBogomolnaiaandMoulin (2001)inthecaseofstrictpreferencesandby KattaandSethuraman (2006)inthecaseofweakpreferences.
8 Infact,as notedintheproofofProposition 2,if theex-poststableassignmenthasadecompositionintostableassignmentssuchthatnoagentmatches
1.1. Relatedliterature
Fortheprobabilisticassignmentsetup,a strandoftheliteraturediscussesdifferentstabilitynotionsandtheirrelationto eachother.Attheheartofthisliteratureliesthenotionofex-poststability,whichisalsoatthecenter ofourstudy.Roth etal. (1993),Rothblum (1992),andVande Vate (1989) provideacharacterizationofthesetofex-poststableassignments thatwesummarizeinLemma 2anduseintheproofofProposition 3.Morerecently,Manjunath (2011)investigatesseveral ex-ante stability and corenotions that are based on first-order stochastic dominance. Kesten and Ünver (2015)propose ex-ante stability notions forthe school choice problem, in whichschools havepriorityorderings ratherthen preferences, which are not takeninto consideration inthe welfare analysis. Second, schoolsare assumedto behave non-strategically. In allthesestudies,thefocusisonstability andcorenotionsbasedonordinalpreferenceinformation.In contrast,we are interested intherelationship betweenex-post stability basedonordinal preferencesandacorenotionbased oncardinal preferences.
FollowingBogomolnaiaandMoulin (2001),one strandofthe literaturestudies theprobabilistic assignmentofobjects. In this literature, a result similar to our Theorem 1, which is due to McLennan (2002),9 follows from our Theorem 1
(Corollary 1). Manea (2008) gives a constructive proof of McLennan’s result. Ourproof technique is similar to Manea’s: An orderrelation,theacyclicityofwhichcharacterizesanassignment’ssd-efficiency,is usedtoconstructtheutilityprofile. As showninManea (2008),if only onesideofthemarkethaspreferencesovertheother,thentherelationratherdirectly deliversthedesiredutilityprofile.However,in marriageproblems,thepreferencesofbothmenandwomenmustbetaken into account,both in sd-efficiencyand utilitariansocial welfare considerations.We show that doing so requires a rather novelconstructionoftheutilityprofilecomparedtothecaseofone-sidedmarkets.
Inthedeterministicassignmentliterature,thetheoryofstableassignmentswithouttransfers(NTU-model)isfirst devel-opedbyGaleandShapley (1962).Givenpreferencerankings,thealgorithmtheyproposeselectsastableassignmentthatis Paretodominantfortheproposingsideamongallstableassignments.Ontheotherhand,thetheoryofstableassignments withtransfers(TU-model)isdevelopedbyShapleyandShubik (1971),whereforgivenvaluationsofagents,thecoreofthe associatedTU-game(theassignmentgame)ischaracterized.10
OurTheorem 2offersa connectionbetweenTUandNTU-models.Echenique (2008) andEcheniqueetal. (2013)offera similarconnectionbetweenTUandNTU-modelsbasedontheobservablecontentofstabilityinbothsetups.Theiranalysis isina revealedpreferenceframework,in whichordinal preferencesarenot observableandare recoveredfromanobserved
aggregatematching.11Echeniqueetal. (2013)arguethatinthissetting,matchingtheorywithtransfersisnestedinmatching
theory withouttransfers,that is,foreach aggregatematching,if thereisautility profilesuch thatthematchingisinthe coreof the associated TU-game,then there is an ordinal preference profile such that the matching is theunique stable matching.For ourTheorem 2, we assume that ordinal preferencesand an ex-post stableassignment are observable,but utilityprofilesarenot.We constructautilityprofilethatisconsistentwiththeordinalpreferencessuchthatthematching isinthecoreoftheassociatedTU-game.Sincetheconversestatementdoesnothold,12bycontrastwithEchenique (2008), in oursettingmatchingtheorywithouttransfersisnestedinmatchingtheorywithtransfers.
Anotherrecentstudy,whichprovides severalinteresting insightsontherelationshipbetweenstability inTU and NTU-models,isEcheniqueandGalichon (2014).Oneoftheir goalsistounderstandforwhichstabledeterministicassignments, availabilityofmonetary transferswouldnot affectstability.Theyshowthat, fora particularsubsetofstabledeterministic assignments, whichthey call“isolated” assignments, foranyordinal preferenceprofile,one can constructautility profile suchthateachisolatedstabledeterministicassignmentremainsstablewhenmonetarytransfersareintroduced.Attheend ofSection3.3,we showthattheirresultfollowsfromourTheorem 2.
2. Theframework
LetM beasetofn menandW asetofn women.Eachi
∈
M haspreferencesoverW ,andeach j∈
W haspreferences over M. Let N=
M∪
W . For each i∈
N, the preferences of i, which we denote by Ri, is a weakorder, that is, Ri is transitiveandcomplete.Let Pi denotetheassociatedstrict preferencerelation,and Ii theassociatedindifference relation. LetR
i denotethesetofall possiblepreferencerelationsfori,andR
≡ ×
i∈NR
i denotethesetofall possiblepreference profiles.A deterministicassignment isaone-to-onefunction
μ
:
M∪
W→
M∪
W suchthatforeach(
m,
w)
∈
M×
W ,we haveμ
(
m)
∈
W ,μ
(
w)
∈
M,andμ
(
m)
=
w ifandonlyifμ
(
w)
=
m.A deterministicassignmentcanberepresentedbyann×
nmatrix,withrowsindexedby menandcolumnsindexedbywomen, andhaving entriesin
{
0,
1}
,suchthat eachrowand eachcolumnhasexactlyone 1.Sucha matrixiscalleda permutationmatrix.Foreach(
m,
w)
∈
M×
W ,having1 inthe(
m,
w)
entryindicates thatm isassignedto w.A probabilisticassignment isaprobabilitydistributionoverdeterministic assignments.A probabilisticassignmentcan berepresentedby ann×
n matrixhavingentriesin[
0,
1]
suchthat thesum9 Athanassoglou (2011)approachestheprobleminMcLennan (2002)byusingduality. 10 ThistheoryisappliedtothemarriageproblembyBecker (1973).
11 SeeEcheniqueetal. (2013)forthedefinitionofanaggregatematching. 12 SeeourExample 2.
of the entries in each row and each column is 1. Such a matrix is a doublystochasticmatrix. For each probabilistic assignment
π
,andeachpair(
m,
w)
∈
M×
W ,theentryπ
mw indicatestheprobabilitythatm isassignedtow atπ
.Since eachdoublystochasticmatrixcanberepresentedasaconvexcombinationofpermutationmatrices(Birkhoff,1946andVon Neumann, 1953),the set ofall doubly stochastic matricesprovides another representationforthe set ofall probabilistic assignments.Letbethesetofalldoublystochasticmatrices.
Wedenotethecollectionofalllotteriesover M by
L(
M)
,andthecollectionofall lotteriesoverW byL(
W)
.Foreachi
∈
M withpreferences Ri∈
R
i,a von Neumann–Morgenstern(vNM)utilityfunctionuiisareal-valuedmappingonW ,i.e.ui
:
W→ R
.We obtainthecorrespondingpreferencesofi overL(
W)
bycomparingexpectedutilities.Foreachi∈
M with preferences Ri∈
R
i,a vNMutilityfunction uiisconsistentwith Riifforeachpair(
w,
w)
∈
W wehaveui(
w)
≥
ui(
w)
if andonlyifwRiw.Foreachwoman,a (vNM)utilityfunctionconsistentwithherordinalpreferencesisdefinedsimilarly.Foreachutilityprofileu
= (
ui)
i∈N andprobabilisticassignmentπ
,the utilitariansocialwelfareat(
u,
π
)
isthesumof theexpectedutilitiesoftheagents,thatis:SW
(
u,
π
)
=
(m,w)∈M×W
π
mw(
um(
w)
+
uw(
m)).
Anassignment
π
isex-anteutilitarian-welfaremaximizingatautilityprofile u ifitmaximizesthesocialwelfareat u, i.e.π
∈
argmaxπ∈SW(
u,
π
)
.Next,we defineawell-known notionofefficiencythatisindependentofanyvNM utilityspecificationconsistentwith theordinal preferences.Let
π
,
π
∈
and R∈
R
;we saythatπ
first-orderstochasticallydominatesπ
at R ifforeach pair(
m,
w)
∈
M×
W : w:wRmwπ
mw≥
w:wRmwπ
mw and m:mRwmπ
mw≥
m:mRwmπ
mwsuch that for atleast one pair, at least one of the inequalities is strict. An assignment
π
∈
issd-efficientat R if no probabilisticassignmentsd-dominatesπ
atR.Foreach R∈
R
,let Psd(
R)
denotethesetofsd-efficientassignmentsatR.3. Results
3.1. Acharacterizationofsd-efficiency
Foreachpair
(
π
,
R)
∈
×
R
,we definetworelations∼
(π,R)and(π,R)onM×
W inducedby(
π
,
R)
,andcharacterizethesd-efficiencyofanassignment
π
atR intermsofapropertyof∼
(π,R) and(π,R).Foreachpair
(
m,
w),
(
m,
w)
∈
M×
W ,(
m,
w)
∼
(π,R)(
m,
w)
ifandonlyifπ
mw>
0,π
mw>
0,andwImw,mIwm. For each pair(
m,
w),
(
m,
w)
∈
M×
W ,(
m,
w)
(π,R)(
m,
w)
ifand onlyifπ
m,w>
0,π
mw>
0, and wRmw,mRwm with at least one strict preference. Let the relation
(π,R) be the union of the two relations just defined. That is, (π,R)=∼
(π,R)∪
(π,R).Ifassignment
π
isdeterministicandpreferencesarestrict,then(
m,
w)
(π,R)(
m,
w)
impliesthatm andwprefereachother totheir assignedmates.Furthermore,supposethat
(
m,
w)
(π,R)(
m,
w)
(π,R)(
m,
w)
.Then, m andm (or w andw)arebetteroffbyexchangingmates.In general,if therelation
(π,R)hasacycle,agentsinthecyclecanParetoimprovebyexchangingmatesalongthecycle.Next,we formulateanacyclicityrequirementon
(π,R)withthesameimplicationforsd-efficiency.
Notethat,if
(
m,
w)
(π,R)(
m,
w)
,then(
m,
w)
and(
m,
w)
arenotrelatedaccordingto∼
(π,R),thatis,(
m,
w)
(π,R)(
m,
w)
.A strongcycleof(
π,R)isasequenceofpairs(
m1,
w1),
(
m2,
w2)
,. . . ,
(
mk,
wk)
∈
M×
W suchthat(
m1,
w1)
(π,R)(
m2,
w2)
(π,R). . .
(π,R)(
mk,
wk)
(π,R)(
m1,
w1)
. The relation(
π,R) isweaklyacyclic ifand only ifit has no strong cycle. Next,we characterizesd-efficientassignments. Thisresultgeneralizescharacterizationsofsd-efficiencyofobject as-signmentsonthestrictpreferencedomain(BogomolnaiaandMoulin,2001)andontheweakpreferencedomain(Kattaand Sethuraman,2006).Proposition1.Anassignment
π
issd-efficientatapreferenceprofileR ifandonlyif(π,R)isweaklyacyclic. Proof. Letπ
∈
,R∈
R
.Only-ifpart: Weprovethecontrapositivestatement.Supposethat
(π,R)isnotweaklyacyclic,thatis,thereisasequenceof pairs
(
m1,
w1),
(
m2,
w2)
,. . . ,
(
mk,
wk)
∈
M×
W such that(
m1,
w1)
(π,R)(
m2,
w2)
(π,R). . .
(π,R)(
mk,
wk)
(π,R)(
m1,
w1)
.Let≡
mini∈{1,...,k}π
miwi.Letπ
∈
bedefinedbysettingforeachi∈ {
1,
. . . ,
k}
,π
mi wi=
π
miwi−
,
π
mi wi+1=
π
miwi+1+
(withtheconventionthat wk+1
=
w1),andforeachotherpair(
m,
w)
,π
mw=
π
mw.Notethatπ
sd-dominatesπ
atR.Thus,π
∈
Psd(
R)
.Ifpart: We provethe contrapositive statement.Suppose that
π
∈
/
Psd(
R)
, thatis,there isπ
∈
thatsd-dominates it. Without loss of generality, suppose that there is a man, say m1∈
M, who is better off atπ
in stochastic dominancesuchthat m1Rw2m2 and
π
m2w2<
π
m2w2.Notethat(
m1,
w1)
(π,R)(
m2,
w2)
.Now,therearem3∈
M, w3∈
W such thatw3Rm2 w2,
π
m2w3>
π
m2w3,m2Rw3m3,andπ
m3w3<
π
m3w3. Then,(
m1,
w1)
(π,R)(
m2,
w2)
(π,R)(
m3,
w3)
.Proceeding inductively,we canaddpairstothissequence,andsincetherearefinitelymanyman–womanpairs,thissequenceincludes acycleof(π,R).If the cycle includes
(
m1,
w1)
, then it is strong and we are done. Suppose otherwise. Let the cycle consist of(
x1,
y1),
(
x2,
y2),
. . . ,
(
xk,
yk)
∈
M×
W .Note that(
x1,
y1)
∼
(π,R)(
x2,
y2)
∼
(π,R). . .
∼
(π,R)(
xk,
yk)
∼
(π,R)(
x1,
y1)
.Remem-ber that
π
x1y1<
π
x1y1,
π
x2y2<
π
x2y2,
. . . ,
π
xkyk<
π
xkyk, andπ
x1y2>
π
x1y2,
π
x2y3>
π
x2y3,
. . . ,
π
xky1>
π
xky1. Let1
≡
min{
π
x1y1−
π
x1y1,
. . . ,
π
xkyk−
π
xkyk}
,2
≡
min{
π
x1y2
−
π
x1y2,
. . . ,
π
xky1−
π
xky1}
and≡
min{
1
,
2
}
.Letπ
∈
be de-fined by settingfor each i∈ {
1,
. . . ,
k}
,π
xiyi=
π
xiyi+
,
π
xiyi+1=
π
xiyi+1−
(with the convention that yk+1
=
y1), and foreach other pair(
m,
w)
,π
mw=
π
mw . Sinceforeach i∈ {
1,
. . . ,
k}
, yiIxi yi+1 and xi−1Iyi xi,π
alsosd-dominatesπ
. Also,by theconstructionofπ
,therearetwo consecutivepairsinthecycle,say(
xt,
yt)
and(
xt+1,
yt+1)
,suchthat eitherπ
xtyt=
π
xtyt orπ
xtyt +1=
π
xtyt+1.Now,because
π
sd-dominatesπ
,as we didabove,we can findacycleof(π,R),say(
x1,
y1)
,(
x2,
y2)
,. . . ,
(
xt,
yt)
∈
M
×
W ,suchthatforeach i∈ {
1,
. . . ,
t}
,π
x iyi<
π
xiyi andπ
xiyi+1>
π
xiyi+1 (withtheconvention that yt+1=
y1).Notethat(
xt,
yt)
and(
xt+1,
yt+1)
cannotbepartofthiscycleconsecutively,implyingthatthisnewcyclemustbedifferentfromthecyclethatwasidentifiedbefore.Continuingsimilarly,we obtainadditionalassignmentsthatsd-dominate
π
andadditional cycles.Noneofthose cyclescaninclude(
xt,
yt)
and(
xt+1,
yt+1)
consecutively;andforeach additionalcycle, we identifyadditionalconsecutivepairsinthecyclethatcannotbepartofanycycleinthefuture,implyingthateachcycleisdifferent fromanycyclethat wasidentifiedbefore.Sincethenumberofcyclesin
(π,R) isfinite,thisprocesseventually leadstoastrongcycle.Thus,
(π,R)isnotweaklyacyclic.2
Given R
∈
R
, for eachπ
∈
Psd(
R)
, the relation(π,R) can have cycles of the form
(
m1,
w1)
(π,R)(
m2,
w2)
(π,R). . .
(π,R)(
mk,
wk)
(π,R)(
m1,
w1)
.However,(π,R)beingweaklyacyclicimpliesthatsuchacycleshouldbelongto∼
(π,R),thatis,thecycleshouldbeoftheform
(
m1,
w1)
∼
(π,R)(
m2,
w2)
∼
(π,R). . .
∼
(π,R)(
mk,
wk)
∼
(π,R)(
m1,
w1)
.Given
(
π
,
R)
∈
×
R
, letCπ,R be thebinary relationon M×
W ,definedasfollows:Foreach pair(
m,
w),
(
m,
w)
∈
M
×
W ,(
m,
w)
Cπ,R(
m,
w)
ifandonlyifthereisacycleof∼
(π,R)thatcontainsboth,thatis,thereisasequenceofpairs(not necessarily distinct)
(
m1,
w1),
(
m2,
w2)
,. . . ,
(
mk,
wk)
∈
M×
W that includes(
m,
w)
and(
m,
w)
, and is such that(
m1,
w1)
∼
(π,R)(
m2,
w2)
∼
(π,R). . .
∼
(π,R)(
mk,
wk)
∼
(π,R)(
m1,
w1)
.Notethat Cπ,R isan equivalence relationon M×
W , that is, it is reflexive, symmetric, and transitive. For each pair(
m,
w)
∈
M×
W , let[
m,
w]
Cπ,R≡ {(
m,
w)
∈
M×
W:
(
m,
w)
Cπ,R(
m,
w)
}
denote the equivalenceclass of(
m,
w)
relative to Cπ,R.Let (π,R) be the relationdefined onthesetofall equivalenceclassesof C(π,R) asfollows:Foreach pair
(
m1,
w1),
(
m2,
w2)
∈
M×
W ,[
m1,
w1]
(π,R)[
m2,
w2]
ifandonlyif
[
m1,
w1]
= [
m2,
w2]
andthereare(
m1,
w1)
∈ [
m1,
w1]
,(
m2,
w2)
∈ [
m2,
w2]
suchthat(
m1,
w1)
(π,R)(
m2,
w2)
.Notethat,if
π
∈
Psd(
R)
,then(π,R)isacyclic.
3.2. Anefficiencytheorem
In thissection we show that for each probabilistic assignment that is sd-efficient ata givenpreference profile,there isa utilityprofileconsistent withthesepreferences suchthat theprobabilistic assignmentmaximizesthesumofthe ex-pected utilities.First,let usconsider thesimplecasewhere thesd-efficient assignment forwhicha utility functionis to be constructedis anefficient deterministicassignment,
μ
.Let u bea utility profileconsistent with R suchthat foreach(
m,
w)
∈
M×
W ,if m and w arematchedatμ
,thenum(
w)
=
uw(
m)
=
1.Further,foreach w∈
W suchthatwmw,let 0<
um(
w)
< δ
forsomeδ >
0 andforeach w∈
W suchthatwmw,let1<
um(
w)
<
1+
forsome
>
0.Letuw be similarlydefined.Notethatforsomesmallenoughselectionofδ
and,theefficientassignment
μ
isawelfaremaximizing assignmentatutilityprofileu.This construction would fail even for the simplest probabilistic assignment, which is obtained as a mixture of two efficient deterministic assignments. However, we show that the same conclusion holds foreach sd-efficient probabilistic assignment
π
byprovidinganexplicitconstructionforautilityprofileatwhichπ
iswelfaremaximizing.Theconstruction inthenextexampleisinstructivetounderstandthegeneralconstructiontofollowinTheorem 1.Example1.LetM
= {
1,
2,
3}
andW= {
a,
b,
c}
.Letthepreferenceprofile R beasfollows:R1 R2 R3 Ra Rb Rc
a b c 3 1 2
c a b 2 3 1
b c a 1 2 3
Let
μ
betheassignmentwhere1 ismatchedwitha,2 with b,and3 withc.Letμ
betheassignmentwhere1 ismatched withb,2 withc,and3 with a.Letπ
assign0.
5 probabilitytoeachofμ
andμ
.Now,letu besuchthateach agentgetsutility 1 fromhis/hertop-rankedagentandutility0 fromhis/herthird-ranked agent. Further,each agentgetsutilityintheopen interval
(
0,
1/
2)
fromhis/hersecond-rankedagent.Clearly,foreachpair(
m,
w)
,ifπ
mw>
0,thenum(
w)
+
uw(
m)
=
1.Moreover,ifπ
mw=
0,thenum(
w)
+
uw(
m)
<
1.Hence,thesumofexpected utilities,whichis3 atπ
,cannotexceed3 atanyprobabilisticassignment.Intheaboveconstruction,foreachman–womanpair,thesumoftheutilitiestheygetfromeachotheristhesame,and thataspectplaysthecriticalrole.Forourgeneralresult,a similarconstructionworks.
Theorem1.Foreachpreferenceprofileandeachassignmentthatissd-efficientatit,thereisautilityprofileconsistentwiththe preferenceprofilesuchthattheassignmentisex-anteutilitarianwelfaremaximizingatthatutilityprofile.
Proof. Let
(
π
,
R)
∈
×
R
besuchthatπ
∈
Psd(
R)
.Sincewefix(
π
,
R)
throughouttheproof,we remove thereferenceto(
π
,
R)
indenotingthebinaryrelationswehavedefined,andsimplywrite∼
,,,and.Foreachpairm
∈
M, w∈
W ,letsmw denotethelengthofthelongestpathofstartingat[
m,
w]
,andletemw denote thelengthofthelongestpathofendingat[
m,
w]
.13Step 1: Constructinganauxiliaryutilityprofile. Foreachm
∈
M,let vm:
W→ R
bedefinedbysetting,foreach w∈
W , vm(
w)
=
emw emw
+
smw.
Foreachw
∈
W ,letvw:
M→ R
bedefinedbysetting,foreachm∈
M, vw(
m)
=
smw emw
+
smw.
Notethatforeachpair
(
m,
w)
∈
M×
W ,vm(
w)
+
vw(
m)
=
1.Lemma1.Let
(
m,
w),
(
m,
w)
∈
M×
W .If[
m,
w]
[
m,
w]
,thenvw(
m)
>
vw(
m)
. Proof. Sinceemw<
emw andsmw>
smw,thensmw smw
+
emw>
smw smw+
emw,
i.e. 1 1+
emws mw>
1 1+
emw smw.
2
One consequenceof the Lemma is that, foreach pair
(
m,
w),
(
m,
w)
∈
M×
W , if[
m,
w]
[
m,
w]
, then vm(
w)
+
vw
(
m)
<
1.Letz((
m,
w),
(
m,
w))
=
1−
vm(
w)
−
vw(
m)
if[
m,
w]
[
m,
w]
,and1 otherwise.Let min(m,w),(m,w)∈M×Wz
((
m,
w), (
m,
w))
≡
2.
Step 2: Definingtheutilityprofile u. Letm
∈
M.Letumbedefinedasfollows(foreachw∈
W ,we defineuwinasymmetric way):i. For each w
∈
W suchthatπ
mw>
0,set um(
w)
≡
vm(
w)
.We show that um is consistentwith Rm onthe subset of womenforwhomπ
mw>
0.Letm∈
M andw,
w∈
W besuchthatπ
mw>
0 andπ
mw>
0.Withoutlossofgenerality, suppose that wRmw. If w Imw, then note that(
m,
w)
∼ (
m,
w)
, and[
m,
w]
∼ [
m,
w]
. Thus, um(
w)
=
um(
w)
, as desired.If wPmw,then(
m,
w)
(
m,
w)
,and[
m,
w]
[
m,
w]
.Thus,um(
w)
>
um(
w)
,as desired.ii. For each w
∈
W such thatπ
mw=
0 and there is no w∈
W withπ
mw>
0, wRmw, set um(
w)
≤ −
1. Obviously, at this step the utilities can be chosen such that um is consistent with Rm on the subset of women for whom the utilitiesaredefinedsofar.iii. Foreach w
∈
W suchthatπ
mw=
0 andthere is w∈
W withπ
mw>
0, wRmw, considera bestsuch w,that is,π
mw>
0,wRmw,andthereisnosuch w∈
W with wPmw.Setum(
w)
∈ [
vm(
w),
vm(
w)
+
]
.Obviously,at this steptheutilitiescanbechosensuchthatum isconsistentwithRm ontheentiresetofwomen.LetthefunctionSW
(
u,
.)
:
→ R
bedefinedbysetting,foreachπ
∈
, SW(
u,
π
)
=
(m,w)∈M×W
[
π
mw(
um(
w)
+
uw(
m))
].
13 Apathoflengthk ofconsistsofk pairs(m
Step 3: SW attainsitsmaximumat
π
. For each pair(
m,
w)
∈
M×
W , ifπ
mw>
0, then um(
w)
+
uw(
m)
=
1. Thus,SW
(
u,
π
)
=
n. We showthat foreach pair(
m,
w)
∈
M×
W , ifπ
mw=
0, then um(
w)
+
uw(
m)
≤
1, which implies that themaximalpossibleex-anteutilitariansocialwelfareisn,anditisreachedatπ
.Let
(
m,
w)
∈
M×
W be such thatπ
mw=
0. Suppose that there is no w∈
W such thatπ
mw>
0, wRmw. Then,um
(
w)
≤ −
1.If thereisnom∈
M suchthatπ
mw>
0 andmRwm,thenuw(
m)
≤ −
1 andum(
w)
+
uw(
m)
<
1.If thereism
∈
M suchthatπ
mw>
0 andmRwm,thenuw(
m)
≤
uw(
m)
+
<
2.Thus,um(
w)
+
uw(
m)
<
1.Thecasewhenthereis nom∈
M suchthatπ
mw>
0,mRwm,is symmetric.So,thereisonlyonecaselefttoconsider.Supposethatthereisw
∈
W suchthatπ
mw>
0 and wRmw,andthereism
∈
M suchthatπ
mw>
0 andmRwm.Letw andm bethebestsuch agents.Notethat(
m,
w)
(
m,
w)
.If w Imw andmIwm,thenbyLemma 1,um(
w)
+
uw(
m)
<
1.So,supposethatforatleastoneagent,thepreferenceisstrict.Then,(
m,
w)
(
m,
w)
and[
m,
w]
[
m,
w]
.Recallthatz((
m,
w),
(
m,
w))
=
1−
vm(
w)
−
vw(
m)
≥
2.Now, um
(
w)
+
uw(
m)
≤ [
um(
w)
+
] + [
uw(
m)
+
] =
vm(
w)
+
vw(
m)
+
2whichimpliesum
(
w)
+
uw(
m)
≤
1,as desired.2
The welfare theorem by McLennan (2002) for the problem of allocating objects is a corollary of Theorem 1. To see this,firstconsiderthefollowingcounterpartsofsd-efficiencyandutilitariansocialwelfareforthatmodel.Letuskeepthe men-womennotation.Anassignment
π
∈
men-sidesd-dominatesπ
∈
at R∈ R
ifforeachagent i∈
M,thelottery assignedtoi atπ
sd-dominatestheoneassignedatπ
.Thatis,foreachpair(
m,
w)
∈
M×
W , w:wRmwπ
mw≥
w:wRmwπ
mwsuchthat foratleastonepairtheinequalityisstrict.Anassignment
π
∈
ismen-sidesd-efficientat R∈ R
ifno prob-abilisticassignmentmen-sidesd-dominatesitat R.Foreachutilityprofileu= (
ui)
i∈N andprobabilisticassignmentπ
,the men-sideutilitariansocialwelfareat(
u,
π
)
isthesumoftheutilitiesofthemen,thatis:MSW
(
u,
π
)
=
m∈M
w∈W
π
mwum(
w).
Aprobabilistic assignment
π
isex-antemen-sideutilitarianwelfaremaximizingatautilityprofile u ifit maximizes themen-sidesocialwelfareatu,thatis,π
∈
argmaxπ∈MSW(
u,
π
)
.Corollary1.(SeeMcLennan,2002.)Foreachpreferenceprofileandeachassignmentthatismen-sidesd-efficientatit,thereisautility profileconsistentwiththepreferenceprofilesuchthattheassignmentisex-antemen-sideutilitarianwelfaremaximizingatthatutility profile.
Proof. Let
π
∈
bemen-sidesd-efficientatR∈
R
.LetR∈
R
besuchthatforeachm∈
M, Rm=
Rm,andeachwomanis indifferentbetweenanytwomenat R.Notethatπ
ismen-sidesd-efficientalsoat R.Moreover,π
∈
Psd(
R)
.ByTheorem 1, thereisautilityprofileu consistentwithRsuchthatπ
isex-anteutilitarianwelfaremaximizingatu.Sinceeachwoman getsthesameutility fromanytwomenatu,π
isex-antemen-sideutilitarianwelfaremaximizingatu. Now,let u bea utility profilesuchthatforeachm∈
M, um=
um,andforeach w∈
W ,uw isconsistentwithRw.Notethatπ
isex-ante men-sideutilitarianwelfaremaximizingatu.2
3.3. Astabilitytheorem
Acentral robustness criterion for deterministic assignments is“stability”,which requires thatthere be no unmatched man–womanpair who prefer each other to their assignedmates. A counterpartofstability forprobabilistic assignments is “ex-post” stability,whichrequires that there be atleastone decompositionof theprobabilistic assignment into stable deterministicassignments.
Let
D
denote theset ofdeterministic assignments. An assignmentμ
∈ D
is stable at R∈ R
ifthere isno(
m,
w)
∈
M
×
W such that mPwμ
(
w)
, wPmμ
(
m)
. An assignmentπ
∈
is ex-poststable if it can be expressed as a convex combinationofstabledeterministicassignments.From thispoint on,we restrict ourselvesto strict preferences.Foreach i
∈
N, letP
i⊂
R
i bethe setof all transitive, anti-symmetric,andcompletepreferencerelationsfori.LetP = ×
i∈NP
i bethesetofallstrictpreferenceprofiles.Weshowthatex-poststabilityimplieswelfareproperties,beyondsd-efficiency,whicharealsodeduciblefromtheordinal preferencesandcanavoidex-antebreakawaysofmen-womencoalitionsfromthesociety.
Acoalition S
=
M∪
W⊆
M∪
W is admissible if|
M|
= |
W|
.LetA
bethe setofalladmissiblecoalitions.ForeachS
∈
A
,letS denotethesetofprobabilisticassignmentsdefinedover S.Foreachutilityprofileu
= (
utransferableutilitygame defined by settingforeach S
=
M∪
W∈
A
, Vu(
S)
tobe themaximum totalexpectedutility coalition S canachieveamongitsmembers.Thatis,foreachS=
M∪
W∈
A
,Vu
(
S)
=
max πS∈S (m,w)∈M×Wπ
mwS(
um(
w)
+
uw(
m)).
Let E(
um|
π
)
=
w∈W
π
mwum(
w)
be the expected value of um atπ
and E(
uw|
π
)
=
m∈M
π
mwuw(
m)
the expected value ofuw atπ
.Givenautility profileu,an assignmentπ
∈
isinthe coreof Vu ifnocoalitioncanincreaseits total expected utilityby deviating toanother probabilistic assignmentwherethey are matchedamong themselves.Thatis,for each S∈
A
, Vu(
S)
≤
m∈M E(
um|
π
)
+
w∈W E(
uw|
π
).
LetC
(
Vu)
bethe setofallassignmentsthatareinthecoreof Vu.Let P
∈
P
.LetPM andPW denotethecommonpreferencesofmenandwomenoverdeterministicassignmentsinduced by P ,definedasfollows:Foreachpairμ
,
μ
∈
D
,μ
PMμ
ifandonlyifforeachm∈
M,μ
Rmμ
withstrictpreferencefor somem.TherelationPW isdefinedsimilarly.An assignment
π
∈
is well-ordered ex-poststable at P∈ P
if it has a decomposition into stable assignmentsμ1
,
. . . ,
μ
T suchthatforeacht,
t∈ {
1,
. . . ,
T}
witht<
t,we haveμ
t PMμ
t andμ
tPWμ
t.Proposition2.Ifanassignment
π
iswell-orderedex-poststableatastrictpreferenceprofile,thenthereisautilityprofileu consistent withthepreferenceprofilesuchthatπ
∈
C(
Vu)
.Proof. Let
π
∈
be well-orderedex-post stableat P∈
P
. Suppose thatπ
hasthe following decompositioninto stable assignments:π
= λ
1μ1+ λ
2μ2+ · · · + λ
Tμ
T.Supposethatforeachpairt,
t∈ {
1,
2,
. . . ,
T}
suchthatt<
t,μ
t PMμ
t andμ
t PWμ
t.Wefirstdefinetheutilitieseachagentgetsfromtheagentsthathe/sheismatchedwithpositiveprobability.Let
(
m,
w)
∈
M
×
W suchthatπ
mw>
0.First note thatifm is matchedto w in twodifferentassignments inthe decomposition,sayμ
t,μ
t,t<
t,thenthey shouldbematchedinall assignmentsbetweenμ
t andμ
t,thatis,foreach t∈ {
t,
t+
1,
. . . ,
t}
,μ
t(
m)
=
w.So,letμ
p,
μ
p+1,
. . . ,
μ
qbethelistofassignmentsatwhichm ismatchedtow.Letλ
mw≡ λ
p+λ
p+1+· · ·+λ
q.Let uw
(
m)
=
λ
pλ
mw·
p T+
1+
λ
p+1λ
mw·
p+
1 T+
1+ · · · +
λ
qλ
mw·
q T+
1,
and um(
w)
=
λ
pλ
mw·
T−
p+
1 T+
1+
λ
p+1λ
mw·
T−
p T+
1+ · · · +
λ
qλ
mw·
T−
q+
1 T+
1.
Ifthereisauniqueassignment
μ
t inthedecompositionsuchthatμ
t(
m)
=
w,we simplyhaveuw(
m)
=
T+t1 andum(
w)
=
T−t+1T+1 .
14
Next, we argue that the utilities each agent gets fromthe agents that he/she is matched with zero probability can be defined in such a way that for each such pair
(
m,
w)
∈
M×
W , um(
w)
+
uw(
m)
<
1. Let(
m,
w)
∈
M×
W be such thatπ
mw=
0.If there isno w∈
W suchthatπ
mw>
0 and wPmw,thenlet um(
w)
≤ −
1.The casewhenthere isnom
∈
W suchthatπ
mw>
0,mPwmisthesame.So,supposethattherearem∈
M,
w∈
W suchthatπ
mw>
0,wPmw, andπ
mw>
0,mPwm.Supposew.l.o.g.that w andmarebestsuchagentsatPw and Pm.We willshowthatum(
w)
+
uw
(
m)
<
1.First,notethatthepairs(
m,
w)
and(
m,
w)
cannotappearinthesameassignmentofthedecomposition,since otherwisethatassignmentwouldnotbestable.Supposethatm and warematchedinassignmentsμ
p,
μ
p+1,
. . . ,
μ
q,andmandw arematchedinassignments
μ
p,
μ
p+1,
. . . ,
μ
q.Eitherq<
porq<
p.In factwecannothaveq<
p;otherwise,m wouldpreferhismatein
μ
q,namelyw,tohismateinμ
p.Butthen,m wouldpreferw tohismateinμ
p,contradicting the assumption thatμ
p is stable. Thus q<
p. Since um(
w)
≤
T−T+p+11 anduw(
m)
≤
q
T+1, um
(
w)
+
uw(
m)
<
1. Then,by argumentssimilartotheproofofTheorem 1,foreachsuchm, w,m,w,letz
((
m,
w),
(
m,
w))
=
1−
um(
w)
−
uw(
m)
and let 2be the minimum of z
((
m,
w),
(
m,
w))
. Now, for the pair(
m,
w)
, let um(
w)
∈ [
um(
w),
um(
w)
+
]
and letuw
(
m)
∈ [
uw(
m),
uw(
m)
+
]
.Thus, utility profileu is consistentwith P andforeach pair(
m,
w)
thatis matchedwith zeroprobabilityum(
w)
+
uw(
m)
<
1.Now,we showthat
π
∈
C(
Vu)
.First,foreachadmissiblecoalition S=
M∪
W,Vu(
S)
≤ |
M|
.Now,let(
m,
w)
∈
M×
W . We showthat E(
um|
π
)
+
E(
uw|
π
)
=
1.ObservethatE