HAL Id: hal-00243002
https://hal.archives-ouvertes.fr/hal-00243002
Preprint submitted on 6 Feb 2008
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.
Stable Allocation Mechanism
Mourad Baïou, Michel Balinski
To cite this version:
Mourad Baïou, Michel Balinski. Stable Allocation Mechanism. 2002. �hal-00243002�
Stable Allocation Mechanism
Mourad Baïou Michel Balinski
Avril 2002
Cahier n° 2002-009
ECOLE POLYTECHNIQUE
CENTRE NATIONAL DE LA RECHERCHE SCIENTIFIQUE
LABORATOIRE D'ECONOMETRIE
1rue Descartes F-75005 Paris (33) 1 55558215 http://ceco.polytechnique.fr/
Stable Allocation Mechanism
Mourad Baïou
1Michel Balinski
2Avril 2002 Cahier n° 2002-009
Résumé: Le problème d’allocations stables généralise les problèmes d’affectations stables (" one-to-one ", " one-to-many " ou " many-to-many ") à l’attribution de quantités réelles ou d’heures. Il existe deux ensembles d’agents distincts, un ensemble I " employés " et un ensemble J " employeurs " où chaque agent a un ordre de préférences sur les agents de l’ensemble opposé et chacun a un certain nombre d’heures. Comme dans les cas spécifiques, le problème d’allocations stables peut contenir un nombre exponentiel de stables (quoique dans le cas " générique " il admet exactement une allocation stable).
Un mécanisme est une fonction qui sélectionne exactement une allocation stable pour n’importe quel problème. Le mécanisme " optimal-employés " qui sélectionne toujours l’allocation stable optimale pour les employés est caractérisé comme étant l ‘unique mécanisme " efficace " ou " monotone " ou
" strategy-proof. "
Abstract: The stable allocation problem is the generalization of the well-known and much studied stable (0,1)-matching problems to the allocation of real numbers (hours or quantities). There are two distinct sets of agents, a set I of
"employees" or "buyers" and a set J of "employers" or "sellers", each agent with preferences over the opposite set and each with a given available time or quantity. In common with its specializations, and allocation problem may have exponentially many stable solutions (though in the "generic" case it has exactly one stable allocation).
A mechanism is a function that selects exactly one stable allocation for any problem. The "employee-optimal" mechanism XI that always selects xI, the
"employee-optimal" stable allocation, is characterized as the unique one that is, for employees, either "efficient", or "monotone", or "strategy-proof."
Mots clés : affectation stable, mariage stable, couplage stable, transport ordinal, problème d'admission, many-to-many matching, two sided market
Key Words : stable assignment, stable marriage, stable matching, ordinal trasportation, university admissions, two-sided market, many-to-many matching.
Classification AMS: 91B68, 91B26, 90B06, 91A35
1 Université Blaise Pascal, CUST, B.P. 206 - 63174 Aubière Cedex, and Ecole Polytechnique, Laborattoire d'Econométrie. e-mail: baiou@custsv.univ-bpclermont.fr.
2 CNRS and Ecole Polytechnique, Laboratoire d'Economiétrie, 1 rue Descartes, 75005 Paris. e-mail:
balinski@poly.polytechnique.fr
Thestablemarriage(orstableone-to-one) problemisthesimplestexampleofa
two-sidedmarket. There aretwodistint setsofagents,e.g.,men andwomen,
andeahagentononesideofthemarkethaspreferenesovertheoppositeset.
Mathings between men and women are sought that are stable in the sense
that no man and woman not mathed (to eah other) an both be better-o
bybeingmathed [7℄. Thestable admissions (or stable one-to-many) problem
isamoregeneralexampleofatwo-sidedmarket,againwithtwosetsofagents
eahhavingpreferenesovertheoppositeset. Ononesideofthemarketthere
are individuals, e.g., prospetive students, interns or employees, and on the
other there are institutions, e.g., universities, hospitals or rms, eah seeking
to enroll some given number of individuals [7℄. A still more general ase is
thestablepolygamouspolyandry(orstablemany-to-many) problemwhereevery
agentseekstoenrollgivennumbersofagentsoftheoppositeset[2℄. Allofthese
areproblemsofassignment: agentsaremathedwithagents[7,9,8,10℄.
The stable alloation problem [3℄ is also a two-sided market with distint
setsofagentswhereeahagenthasstritpreferenesovertheoppositeset. But
hereeah agent is endowed with real numbers quantities or hours of work
andinsteadofmathing(oralloating0'sand1's)theproblemistoalloate
realnumbers. For example,oneset of agentsonsists ofworkmeneahwith a
numberofavailablehoursofwork,theotherofemployerseahseekinganumber
ofhours ofwork. Stability asks that nopairof opposite agentsan inrease
theirhourstogether eitherduetounusedapaityorbygivinguphourswith
lesspreferredpartners.
Here,asisoftentrue,thestudyofthemoregeneralproblemlariesandin
someaspetssimpliestheissuesand viewsthatonernthepartiularases.
Setion 1 the problem presents the model and Setion 2 stable
alloations summarizes the salient fats onerning them, their existene,
strutureandproperties(see[3℄).
Analloationmehanism isafuntionthat seletsauniquestable alloa-
tionforanyalloationproblem. Setion3mehanismsuniquelyharater-
izestheemployee-andemployer-optimal(orrow-andolumn-optimal)alloation
mehanisms in termsof three separateproperties: eieny, monotoniity,
andstrategy-proofness. This generalizesto stable alloationssimilarhara-
terizationsrstestablishedforadmissionsorone-to-manymathing[6,1℄,then
formany-to-manymathing[2℄.
1 The problem
Astablealloationproblem( ;s;d;)isspeiedbyadiretedgraph dened
overagrid, andarraysofreal numberss; d>0and 0,as follows. There
are two distint nite sets of agents, the row-agents I (employees) and the
olumn-agents J (employers), and eah agent has a strit preferene order
overtheagentsoftheoppositeset. Eah employeei2I hass(i)units ofwork
themaximumnumberofunits that i2I mayontratwithj 2J. Thisdata
ismodeledasagraph.
The nodesof thepreferenegraph are the pairsof opposite agents(i;j),
i2I andj2J. TheyaretakentobeloatedontheIJ gridwhereeahrow
orrespondsto anemployee orsupplieri 2I and eaholumn toan employer
or aquirer j 2 J. The (direted) ars of , or orderedpairs of nodes, are of
twotypes: ahorizontalar (i;j); (i;j 0
)
expressessupplieri'sprefereneforj 0
overj (sometimeswritten j 0
>
i
j), symmetriallyavertialar (i;j); (i 0
;j)
expresses aquirer j's preferene for i 0
over i (sometimes written i 0
>
j i). If
(i;j) = 0 for some (i;j) then the node may be omitted. Ars implied by
transitivityare omitted. Figure 1givesan examplewhere the values s(i)are
assoiated withrows,the valuesd(j)with olumns, and the values(i;j)are
arbitrarilylarge.
The stable marriage problem is the stable alloation problem with s(i) =
d(j) = 1 and (i;j) = 0 or 1, for all i 2 I; j 2 J; the stable university
admissionsproblem isthestable alloationproblem withs(i)positiveintegers,
d(j)=1and(i;j)=0or 1,foralli2I; j 2J; andthestable many-to-many
problem is the stable alloation problem with s(i) and d(j) positive integers,
and(i;j)=0or 1,foralli2I; j2J (see[4, 5,2℄).
Itisonvenient,andunambiguous,torefertothesuessorsofanodeor
to sayanode follows another in its rowor olumn, meaningthey or itare
preferredorrankedhigher. And,similarly,torefertothepredeessorsofanode
or to saya nodepreedesanother in itsrow or olumn, meaning theyor
itarelesspreferredor rankedlower. Alsoarst,leastpreferred(orlast, most
preferred)node inaroworolumnhasnopredeessors(nosuessors)and
arst(orlast)nodewithertainpropertieshasnopredeessors(nosuessors)
withthoseproperties.
In general, if S is a set and y(s); s 2 S, a real number, then y(S) def
=
P
s2S
y(s); also(r;S) def
= f(r;s):s2Sg. For(i;j)2 , (i;j
) def
= f(i;l):l
i
jgand(i;j
>
) def
= f(i;l):l>
i
jg;thesets(i
;j)and(i
>
;j)aredenedsimilarly.
An alloation x = x(i;j)
of a problem ( ;s;d;) is a set of real-valued
numberssatisfying
x(i;J)s(i); alli2I;
x(I;j)d(j); allj2J;
0x(i;j)(i;j); all(i;j)2 ;
alled,respetively, the row, the olumn and the entry onstraints. In Figure
1both y and z are alloations of the example. It may be and will be
assumedthat(i;j)min
s(i);d(j) .
Analloationx isstableifforevery(i;j)2 ,
x(i;j)<(i;j) implies x(i;j
)s(i) or x(i
;j)=d(j):
l 2 J may together, ignoring others, improve the alloation for themselves.
Speially,thevalueofx(k;l)maybeinreasedbyÆ>0,withx(k;j)>0for
somej <
k
l dereasedbyÆ (or x(k;J)<s(k))andx(i;l)>0forsomei<
l k
dereasedbyÆ(orx(I;l)<d(l)). Otherwise,(k;l)isstableforx. Inpartiular,
ifeitherx(k;l)=(k;l)orx(k;l
)=s(k)then(k;l)isrow-stable;andifeither
x(k;l)=(k;l)orx(k
;l)=d(l)then(k;l)isolumn-stablesoanodemay
bebothrow-andolumn-stable.
Inthe speial aseof marriage,(k;l) blokswhen man k andwomanl are
notmathed x(k;l)= 0
, k is not mathed oris mathed to aless desirable
woman thanl x(k;l
)=0
, l is not mathed or mathed to aless desirable
man thank x(k
;l)=0
, and (k;l)=1: thus together kand l an realize
abettersolutionfor themselves. In Figure1,y is notstable (4;3) bloksy
(theothernodesarestablefory)whereasz isstable.
(1)=10
d(1)=
10 1 11
14 1
2 17
y=
z= 11
14 6 12 7 3
2 1 s
(2)=12 s s
(3)=14 (4)=20 d(2)= d(3)= d(4)=
11 13 15 17
s s
Figure1: An alloationproblem(noupperbounds).
2 Stable alloations
Thissetionsummarizes thepertinentfats onerningstable alloations. For
proofsandamoreompletedesription,see[3℄.
Theemployee-orrow-greedysolutionofaproblem( ;s;d;)isdenedby
assigningtoeahrow-agenti2I his/her/itspreferredsolutionatingasifthere
werenootherrowagents. Itisdened reursively,beginningwithi'spreferred
hoie(the lastnodeinrowi):
(i;j)=min
s(i) (i;j
>
); d(j); (i;j) :
Ifnoolumnonstraintisviolated,isastablealloation. Intermsofmarriage,
assignstoeahmanhisfavoriteavailablewoman,andifnowomanisassigned
morethan oneman itis astableassignment. Theemployer-orolumn-greedy
solution( ;s;d;)isdened similarly.
whoreeivesaproposalmaydisardeverymanlowerinherprefereneswithout
hangingthe problem. A similar fat holds for alloationproblems. If x is a
stablealloationofaproblem( ;s;d;),then
x(i;j)
(i;j) def
= max n
0; min
(i;j); d(j) (i
>
;j) o
;
and the problems ( ;s;d;) and ( ;s;d;
) are equivalent in the sense that
theyadmitexatlythesamesetofstable alloations.
This suggests the generalization of the Gale-Shapley algorithm, the row-
greedyalgorithm: trytherow-greedysolution;ifitisanalloationthenitmust
beastablealloation;otherwise,newstrongerboundsmaybededuedandthe
proessrepeated. For disrete problems when s;d and are integer-valued
theproedure mustterminatewith aproblemwhose row-greedysolutionis
stable,soproves:
Theorem 1 There exist stable alloations for every stable alloation problem
( ;s;d;).
The theorem is proven for arbitrary real-valued data via an indutive al-
gorithm [3℄thatisstronglypolynomial: itrequiresat most3jIjjJj+jJj steps
tondastable alloation, wherejKjis theardinalityofK. Infat,the row-
greedyalgorithm is arbitrarily badfor disrete problems, and it is notknown
whetheritonvergesatallinthegeneralase.
Confronted with any two stable alloations an agent has no hesitation in
deidingwhihhe,sheoritprefers. Formally,anytwostablealloationsxand
ymaybeomparedwiththedenitionthat follows
x def
i
y;i2I; ifx(i;k)<y(i;k)impliesx(i;j)=0forj<
i
k; (1)
read row-agenti prefersx to y or is indierent betweenthem. x def
=
i
y when
x(i;)=y(i;),meaningiisindierentbetweenx andy (impliitly howothers
fareis ofnoimportaneto i),andx def
i
y when x
i
y andx6=
i
y. Symmetri
denitionsholdforolumn-agentsj 2J.
x
i
y implies x(i;j) < y(i;j) is true for at most one x(i;j) > 0. In
partiular, if x
i
y then x(i;k) < y(i;k) and x(i;j) > y(i;j) imply k <
i j.
Sineeah agentis assigned exatlythe sametotalnumber ofhours by every
stablealloation, row-agenti prefersx toy, or x
i
y, implies thaty maybe
transformedinto xbydereasingsomevaluesthat orrespondtoless-preferred
olumn-agentsandinreasingothersthatorrespondtomore-preferredolumn-
agents.
Ineet,thesimplestompletedesriptionofanagent'spreferenesbetween
stablealloationsisthemin-minriterion: thevalueoftheleast-preferredtype
of hourshould be as small as possible. Letting i(x) =j if x(i;j )> 0and
x(i;j)=0forj<j ,thismeans
x
i y if
either i(x)>
i i(y)
or i(x)=i(y)=j andx(i;j )<y(i;j ):
(2)
Theopposition ofinterestsbetweenrowsandolumnsholds heretoo:
Theorem 2 Ifx;y arestable andx(k;l)6=y(k;l),thenx
k
yfor k2I ifand
onlyif x
l
y for l2J.
Morevoer, itis easy to verifythat the set of allstable alloations is adis-
tributivelattie with respet to thepartial order
I
on thepreferenes of all
row-agentsI dened by:
x def
I
y if x
i
y foralli2I:
Surprisingly, onsiderable more is true. When the data s > 0, d > 0 and
(i;j) 0 are arbitrary real numbers it is to be expeted that no sum of a
subsetofthes(i)equalsthesumofasubsetofthed(j),northatsuhsumsare
equalwhenthes(i)andd(j)areeahreduedbyasumofsomeorresponding
(i;j): this isthe generi, strongly nondegenerate problem. In this ase the
problemhasauniquestable alloation.
Aordingly,itisonlyduetothedegeneraiesofthestableone-to-one,one-
to-manyand many-to-manymathing problems that therih lattiestruture
potentiallyinvolvingexponentiallymany stable alloations ours. But
the data of a stable alloation problem is often integer valued and may well
admit degeneraiesand so multiple stable alloations. Thus it is neessaryto
havearationaleforhoosingone stablealloationin thepreseneofmany.
Theexamplegivenaboveissuhaninstane. Ithasexatly7extremestable
alloationsmeaningstable alloationsthatare notaonvexombinationof
others. Theyare givenin Figure 2. The stablealloation z ofFigure 1is not
extreme: z= 3
10 x
3 +
7
10 x
4
. Whenthedata isinteger-valuedthere alwaysexist
stablealloationsin integers.
3 Mehanisms
Analloationmehanismisafuntionthatseletsexatlyonestablealloation
for any problem ( ;s;d;). Three haraterization are given of eah of two
partiularly onspiuous mehanisms. This generalizes known results for the
one-to-many[6,1℄andmany-to-manymathingproblems[2℄.
Theemployee-orrow-optimalstablealloationx
I
ofaproblem( ;s;d;)is
denedby:
x
I
i
x; alli2I; foreverystable alloationx:
x
I
attributes to every row-agentthe best possible alloationamong all stable
alloations. Therow-optimalalgorithm [3℄establishes
xJ= x = 1
10 5 10
12 4 119 5
x = 3
2 5 10
13 14 1
11 x =
2
5 10
13 2 9
14 3 5 10
12 13 2
9 5
14 11
15 10 1
2 3
2 14 1 11
15 3 10
x4= x5=
xI=
14 15
2 3
11 1
Figure2: All extremestable alloation: arrowsshowolletivepreferenes
I
(inthisaseaompleteorder).
Theorem 3 Every problem ( ;s;d;) has a uniquerow-optimal stable alloa-
tionx
I .
Bysymmetry, everyproblem hasauniqueemployer-or olumn-optimalstable
alloation x
J
. So two obvious examples of mehanisms are the employee- or
row-optimalmehanism
I
thatalwaysseletsx
I
,andtheemployer-orolumn-
optimalmehanism
J
thatalwaysseletsx
J .
Eieny
Itwouldbeagreeableifitouldbeassertedthattheemployee-optimalmeh-
anism
I
is eient in that no alloation, stable or not, is ever better for
theemployees thanthe employee-optimalstable alloation x
I
. This depends,
ofourse, on what is meant bybetter: in an intuitivesense thealloationy
giveninFigure1isolletivelypreferredtox
I
bytheemployeesI (y isbloked
by(4;3)). But bythe min-min riterion (2) for omparing stable alloations,
row-agent4wouldbeindierentbetween x
I
and y. Thedenition ofbetter
willextendthemin-minriteriontoarbitraryalloations.
Consider now aproblem where the s(i);i 2 I are generous in omparison
with the d(j);j 2 J, asin Figure 3. The employee-optimal stable alloation
x
I
is viiouslyemployee-ineient: everyalloation,stableor not,that gives
a total of 7 hours to employee 1 and a total of 11 to employee 2 is better
forthe employees. Thus if x
I
isin somesense eient this possibilitymust
be exluded: aordingly, when x
I
(i;J) < s(i) any other alloation y with
y(i;J)=x(i;J)willbeonsideredequallypreferredbyi.
Guidedbytheseexamples,extendthedenitionofanemployee'spreferenes
betweenstablealloations(2)topreferenesbetweenarbitraryalloationsxand
yas follows:
3 4 0 0 0 0 5 6 xI =
6 11
11
3 4 5
Figure3: x
I
ineient.
x def
i y if
i(x)<
i i(y)or
i(x)=i(y)=l ;x(i;j )<y(i;j )
when x(i;J)=y(i;J)=s(i)
x(i;J)>y(i;J) when y(i;J)<s(i):
Also,rowagentiisindierentbetweentwoalloations,x
i
y,ifi(x)=i(y)=
j andx(i;j )=y(i;j )whenx(i;J)=y(i;J)=s(i),orifx(i;J)=y(i;J)<
s(i). Takex
i
y tomeanx
i
y orx
i
y. Asbefore,
x def
I
yifx
i
yforalli2I;
andx def
I
y ifx
I
y andx
i
y isnottrueforalli2I.
Apreliminarylemma onerningstable alloationsisneeded.
Lemma1 Supposethatxisastablealloation andyisanalloation,stableor
not,forwhihy
i
xforalli2I. Thenthereexistsastablealloationy
I x.
Moreover,
(i)x(i;J)=y(i;J)=y
(i;J)=s(i) for everyi2I,
(ii) x(I;j)=y(I;j)=y
(I;j)for every j2J, and
(iii)y
(i;j)>x(i;j)impliesthereexistsh
j
ifor whih y(h;j)>x(h;j).
Proof. Tobegin,supposethatforsomej2J,x(I;j)<y(I;j). Thenx(i;j)<
y(i;j) for some i 2 I. If x(i;J) < s(i) then (i;j) bloks x, a ontradition;
and if x(i;J) = s(i) then y
i
x implies x(i;j 0
) > y(i;j 0
) for some j 0
<
i j,
so (i;j) bloks x, again a ontradition. Therefore, x(I;j) y(I;j) for all
j 2J, andx(I;J)y(I;J). Buty
i
x means x(i;J)y(i;J)for i2I, so
x(I;J)y(I;J)and the inequalitiesare allequations. Finally, x(i;J) <s(i)
impliesx(i;J)<y(i;J)sox(i;J)=y(i;J)=s(i).
Let =
i;i(x)
:i 2 I . Bydenition, x i;i(x)
>y i;i(x)
for every
i;i(x)
2 . From above it follows that for eah i;i(x)
2 there is at
leastone h2 I with x h;i(x)
<y h;i(x)
h;i(x)
. Buty
h
x implies
h;h(x)
preedes h;i(x)
in rowh. Therefore,itmustbethat h<
i(x) i sine
otherwisex wouldnotbestable.