T he Ex A nt e I ncent ive Compat ible Core of
t he A ssignment Game
¤
Françoise For ges
T H EM A , U niversit é de Cer gy-Pont oise,
and I nst it ut U niver sit aire de Fr ance
¤I wish t o t hank B . D ut t a, J.-F . M ert ens, E. M inelli and most part icular ly, R. Vohra,
Résumé
N ous consi dér ons un marché bipar t it e dans lequel les agent s disposent d’infor -mat ions pr ivées sur un ét at de la nat ur e qui dét er mine leur s ut ilit és d’appa-r iement . L es t d’appa-ransfed’appa-r t s monét aid’appa-r es sont ped’appa-r mis et les fonct ions d’ut ilit é sont quasi-linéair es. L e modèle ét end donc les j eux d’allocat ion int r odui t s par Shapley et Shubik. N ous démont r ons que le coeur ex ant e inci t at if du j eu d’appar iement est non-vide. D es exemples simples illust r ent deux di¤ér ences avec l’infor mat ion compl èt e: t out d’abor d, les mécanismes d’appar iement aléat oi res dé…nissent une fonct ion car act ér ist ique ( à ut ilit é t r ansfér able) plus élevée que les mécanismes dét er minist es; de pl us, les solut ions du coeur ex ant e incit at if ne coï ncident pas nécessair ement avec les r ésult at s st ables cor -r espondant s, et ce même si les valeu-r s sont p-r ivées et indépendant es. N ot -r e appr oche s’ét end au coeur br ut ( int er im) incit at if, qui est lui aussi non-vide.
A bst r act
We consider t wosided mat ching mar ket s in which agent s have pr ivat e infor -mat ion on a st at e of nat ur e which det er mines t he agent s’ ut ilit ies of -mat ching. M onet ar y t r ansfer s ar e all owed and ut ilit y funct ions ar e quasi-linear . T he model t hus ext ends t he assignment game int r oduced by Shapley and Shubik. We pr ove t hat t he ex ant e incent ive compat ible cor e of t he mat ching game is nonempt y. Simpl e examples illust r at e t wo di¤er ences wit h complet e infor -mat ion: …rst , r andom -mat ching mechanisms de…ne a higher T U char act er ist ic funct ion t han det er minist ic ones; fur t her mor e, ex ant e incent ive compat ible cor e solut ions need not coincide wit h ex ant e incent ive compat ible st able out comes, even if values ar e independent and pr ivat e. Our appr oach ext ends t o t he ( int er im) i ncent ive compat ible coar se cor e, which is also non-empt y. JEL classi…cat ion number s: C78, C71, D 82
K eywor ds: assignment game, core, incent ive compat ible mechanism, mat ch-ing, pr ivat e infor mat ion.
1
I n t r o d u ct i on
T wo-sided mat chi ng mar ket s have been ext ensively st udied under t he as-sumpt ion of complete i nfor mati on (see, e.g., Rot h and Sot omayor [23]) . Rot h [22] also invest igat ed ex post st abl e mechanisms in mar r i age pr oblems wi t h incomplet e infor mat ion about ot her s’ prefer ences ( i.e., pr ivat e values) .
We consider t wo-sided mat ching mar ket s wit h ar bit r ar y incomplet e in-for mat ion ( i.e., possibl e common values) . T he agent s ar e divided int o t wo disj oint set s ( e.g., pot ent ial buyer s and seller s, …rms and wor ker s, et c.) . Ev-er y agent has pr ivat e infor mat ion on a st at e of nat ur e which ent Ev-er s agent s’ ( ex post ) ut ilit y of being mat ched wit h a par t ner fr om t he ot her side of t he mar ket . T he ( ex post ) wor t h of a coalit ion is det er mined by pair wise com-binat ions of agent s ( fr om di¤er ent sides of t he mar ket ) and ar bit r ar y money t r ansfer s ( e.g., in t he auct ions fr amewor k, we allow for ent r y fees, bidder r ings, et c.) . U t ilit y funct ions ar e assumed t o be linear in money. Our model is t hus an ext ension of t he ( complet e infor mat ion) assi gnment game int r o-duced by Shapley and Shubik [26]. U nder incomplet e infor mat ion, a number of par t icular cases ( bilat er al t r ading, auct ions, et c.) have been analyzed in great det ails ( see, among ot hers, [7], [8], [18], [21],[29]) .
We ext end Shapley and Shubik [26]’s r esult s by pr oving t hat t he ex ant e i ncenti ve compati ble core of t he assignment game is non-empt y. T his solu-t ion concepsolu-t has been mossolu-t ly applied in di ¤er ensolu-t i al infor masolu-t ion exchange economi es ( see [13]) . I t is appr opr iat e if coalit ions can for m befor e agent s know t heir pr ivat e infor mat ion. T he member s of a coalit ion or ganize mat ch-ings and monet ar y t ransfer s by means of ( r andom) B ayesian incent ive com-pat ible mechanisms. T his gener at es a well -de…ned T U char act er ist ic funct ion, namely t he maximal sum of ( ex ant e) expect ed payo¤s t hat every coal it ion can guar ant ee by r elying on an incent ive compat ible mechanism. T he ex ant e incent ive compat ible cor e is de…ned as t he ( st andar d) cor e of t his char act er is-t ic funcis-t ion. T o esis-t ablish iis-t s non-empis-t iness, we apply is-t he B ondar eva-Shapley t heor em.
A s in most paper s on auct ions and bar gaining under incomplet e infor -mat ion, we assume t hat -mat ching may r esult fr om a lotter y. T his pr ocedur e is nat ur al, especially if one r elies on t he r evel at ion pr inciple ( as will be t he case her e) , but is not needed i n Shapley and Shubik [26]’s or iginal model (see also [23]) . Random assignment pr ocedures can never t hel ess be useful under complet e infor mat ion, for inst ance t o guar ant ee fair ness when money is not available ( see, e.g., [1], [5], [6], [16]) . We shall show on an example t hat ,
when incent ive const r aint s mat t er , t he T U char act er ist ic funct ion associat ed wit h r andom mechanisms can t ake higher values t han t he one associat ed wi t h det er minist ic mechanisms.
U nder complet e informat i on, it is also well -known ( see again [26]) t hat t he cor e coincides wit h t he out comes which ar e stable, i.e., cannot be blocked by any single agent nor pai r of agent s ( fr om di¤er ent sides of t he mar ket ) . H owever , as we show on simple examples, no such pr oper t y hol ds under incomplet e infor mat ion, and t his even under pr ivat e, independent val ues. M or e precisely, one can const r uct i ncent ive compat i ble mechanisms which cannot be blocked by any single agent nor any pair of agent s, but ar e blocked by t hr ee agent s coalit ions.
T he non-empt iness of t he ex ant e incent i ve compat ible cor e means t hat t her e exist s an incent ive compat ible mechanism
¹
for t he gr and coal it ion ( which select s mat chings and t r ansfer s) such t hat no coalit ion can pr opose an incent ive compat ible mechanism impr oving t he expect ed payo¤ of all of it s member s. I n par t icular ,¹
is i ncent ive ex ant e e¢ cient (i n t he sense of H olmst r öm and M yer son [15]) and ex ant e individually r at ional.¹
is t hus also incent ive int er im e¢ cient . B ut¹
is not necessari ly int er im individually r at ional.Given t he impor t ance of t he lat t er pr oper t y in t he mechanism desi gn lit er -at ur e, a n-at ur al quest ion concer ns t he exist ence of incent ive comp-at ible cor e solut ions t hat would be int er im individually rat ional, i.e., such t hat all t ypes of agent s would like t o par t icipat e in t he mechanism at t he int er im st age. A sat isfact or y solut ion concept in t his r espect i s t he i ncent i ve compat i ble coar se core, namely Vohr a [27]’s incent ive compat ible ver sion of W ilson [28]’s coar se cor e. I n t his appr oach, agent s for m coalit ions aft er having lear nt t heir t ypes but can only exchange infor mat i on inside coali t i ons, once t hese have for med. T hey t hus use exact ly t he same incent ive compat ible mechanisms as above. H owever , blocking t akes place at t he int er im st age so t hat , basically, ever y possible t ype of an agent should get a bet t er payo¤. I n par t icular, incent i ve compat ible coar se cor e solut ions ar e int er im individually r at ional.
I n a similar way as in W il son [28], we const r uct an auxiliar y ( N T U ) char act er ist ic funct ion such t hat t he incent ive compat ible coar se cor e is t he st andar d cor e of t his charact er ist ic funct ion. Given t he N T U st r uct ur e of t he game, we rely on on Scarf [24]’s t heor em t o pr ove t hat t he i ncenti ve compat -i ble coar se core -is non-empty. T h-i s r esult guar ant ees t hat , -in a lar ge class of t r ading models wit h incomplet e i nfor mat ion, t here exist incent ive compat i-ble mechanisms t hat ar e not only int er im e¢ cient and int er im individually
r at ional ( as est ablished in a number of paper s) , but cannot even be impr oved by any coalit ion. I n t hat sense, t hese mechanisms ar e “ collusion-pr oof ” .
T he posit ive r esult s obt ained her e cont r ast wit h t he ones t hat pr evail in di¤er ent ial infor mat ion exchange economies. A s shown by For ges, M er t ens and Vohr a [11], in t his model, bot h t he ex ant e i ncent ive compat ible cor e and t he incent ive compat ible coar se cor e can be empt y, even if random mecha-nisms and monet ar y t r ansfer s ar e allowed. T he li near st r uct ur e of t he assign-ment game is cer t ainly helpful but t his explanat ion of t he r esult s should be used wit h car e. T he incent ive compat ible cor es ar e not empt y in exchange economi es wit h linear ut il it y funct ions ( see [27]) but r andom mechanisms can be dispensed wit h i n t his model. On t he cont r ar y, her e, t hey enable t he agent s t o achieve higher expect ed payo¤s. H ence, t he r esult on linear ex-change economies cannot be applied dir ect ly. I t is also t r ue t hat random mechanisms gener at e li near it y in t he sense t hat expect ed payo¤s ar e linear in t he mechanism pr obabilit ies. H owever , t his does not su¢ ce for t he non-empt iness of t he incent ive compat ible cor es, as t he count er -example in [11] shows.
T he quasi-linear ut i lit y funct ions ar e useful in [11], as i n many paper s on mechanism design ( see, e.g. [2], [3], [4], [7], [8] and [17]) t o const r uct incent ive compat ible, ex post e¢ ci ent mechanisms, which t ur n out t o achieve expect ed payo¤s in t he ex ant e incent ive compat i ble core. Such r esult s also apply her e: under appr opr iat e assumpt i ons, t he ex ant e incent ive compat ible cor e cont ains ex post e¢ cient allocat ions. H owever , no speci…c assumpt ions ( e.g., on t he agent s’ bel iefs) are needed her e t o est ablish t he non-empt iness of t he ex ant e incent ive compat ible cor e. Even mor e, t his r esult ( as well as t he non-empt iness of t he incent ive compat ible coar se cor e) st ill holds if t her e ar e no monet ar y t r ansfer s at all. T he model cover s t hen incomplet e infor mat ion ver sions of t he mar r iage pr oblem and t he indivisible good economy of Shapley and Scar f [25]. Obviously, i n t his case, t he under lying char act er ist ic funct ion is always N T U , even in t he ex ant e fr amewor k. T he main r eason t o consider monet ar y t r ansfer s is t o connect our analysis wit h t he lit er at ur e on t r ading wit h incomplet e infor mat ion (e.g., auct ions) , in which B ayesian incent i ve compat ible mechanisms have become a st andar d t ool.
T he next sect ion descr ibes t he model. Sect ion 3 de…nes t he ex ant e incen-t ive compaincen-t ible cor e and esincen-t abli shes iincen-t s non-empincen-t iness. Secincen-t ion 4 i s devoincen-t ed t o t he incent ive compat ible coar se cor e. Examples illust r at e our solut ion con-cept s in sect ion 5. Sect ion 6 concludes wit h fur t her r emar ks on possible ex post proper t ies.
2
I n cen t i v e com p at i b l e m at ch i n g m ech an i sm s
T he economy consist s of t wo, …nit e, disj oint set s of agent s
I
andJ
:I
\J = ;
; letK = I
[ J
. Ever y agentk
2 K
has pri vat e infor mat ion, r epr esent ed by his typet
k2 T
k, wher eT
k is a …nit e set1; letT =
Q
k2KT
k and letq
be apr obabilit y dist r ibut ion over
T
, t he common pr ior of t he agent s, such t hat , wit hout loss of gener alit y,q(t
k) > 0
for ever yt
k. Ever y agent der ives ( vonN eumann - M or genst er n) ut ili t y fr om being mat ched wit h an agent of t he ot her side of t he mar ket and fr om monet ar y t r ansfer s. For ever y
i
2 I
,j
2 J
and
t
2 T
, letu
i(t; j)
( r esp.,u
j(t; i)
) denot e agenti
( r esp.,j
) ’s ut i lit y of beingmat ched wit h agent
j
( r esp.,i
) when t he t ypes ar et
. L et alsou
k(t; 0)
denot eagent
k
’s ut ilit y of being unmat ched (k
2 K
,t
2 T
) . M onet ar y t r ansfer s are j ust added linearl y t o t hese ut il it y number s. A gent s t hus have addit ively separabl e ut i lit y funct ions for mat ching and t r ansfer s.Except per haps for t he l at t er assumpt ions, t he framewor k is quit e gen-er al: any infor mat ional ext gen-er nalit y is allowed ( t he whole vect or of t ypes
t
ent er s all ut ilit y numbers, so t hat common values ar e possibl e) . Fur t her -mor e, any mat ching ext er nalit y is also al lowed, since ever y agent car es about his par t ner .
A s a t ypical example, ext ending Shapley and Shubik [26]’s housi ng mar -ket , t he
I
-agent s ar e seller s, each wi t h a single it em for sale, and t heJ
-agent s are pot ent ial buyer s;u
i(t; j)
r epr esent s selleri
’s ut ilit y of selling his it em t obuyer
j
at t he st at e of infor mat iont
, whileu
j(t; i)
r epr esent s buyerj
’s ut ilit yof buying an it em fr om seller
j
att
. St andar d assumpt ions ar e t hat t he sell-er s ar e indi¤sell-er ent t o sell t o a buysell-er or anot hsell-er (u
j(t; i) = u
j(t; i
0)
for ever yi; i
02 I
) , or t hat t he pr ivat e infor mat ion of t r ader s coi ncides wit h t hei r r eser-vat ion pr ices ( pr i -vat e values:
u
i(t; j) = u
i(t
i; j)
r epr esent s buyeri
’s r eserva-t ion pr ice for seller
j
’s i t em when his t ype ist
i,u
j(t; i) = u
j(t; i
0) = u
j(t
j)
r epr esent s seller
j
’s r eser vat ion pri ce for his it em) . T his example cover s most auct ion models. M onet ar y t r ansfer s ar e nat ur al in t his example; t hey basi-cally cor r espond t o sale pr ices, but may also involve compensat ions for not get t ing an it em, ent r y fees, et c.T he r est ri ct ions imposed on t he mar ket will be r e‡ect ed by t he feasible mat ching mechanisms. T he out comes of such a mechanism consist of disj oint pair s
fi; jg ; i 2 I; j 2 J
of mat ched agent s, a set of unmat ched agent s and1
T his assumpt ion (also made in, e.g., [7], [8], [11], [17], [27], et c.) i s mainly for simplic-it y. M ost of our r esult s go t hr ough if t he set s of t ypes ar e r eal int er vals and t he beliefs have a densit y, as in [18], [19], [21], et c.
monet ar y t r ansfer s
m
k; k
2 K
such t hatP
k2Km
k· 0
. I n par t icular , anagent can be mat ched t o at most one agent of t he ot her side of t he mar ket and ar bit r ar y t r ansfers ar e allowed, incl uding bet ween agent s who ar e not mat ched t oget her .
We assume t hat agent s exchange infor mat ion and mat ch aft er having for med coalit ions2. A coalit ion
S
is j ust a non-empt y subset ofK
. T hemember s of
S
can r ely on any mechanism ( in par t icular , any non-cooper at i ve bar gaini ng game) t o or ganize mat ching and t r ansfer s. B y t he revelat i on pr i n-ci ple ( see, e.g., [18] for an applicat ion t o auct i ons and [22] for an appli cat ion t o mat ching) , all feasible collect ive decisions ofS
can be repr esent ed as t he out comes of a random, di rect, i ncent i ve compat i ble, mat chi ng mechani sm, which we de…ne pr ecisely below.A s in most paper s on col lect ive decisions under incomplet e infor mat ion ( see [7], [8], [11], [12], [15], [18], [22], et c.) , we allow t hat mat ching r esult s fr om lot t er ies, which is usual in many auct ions mechanisms. L ot t er ies also appear as a consequence of t he r evelat ion pri nciple if agent s ar e allowed t o use mixed bar gaining st r at egies. U nder complet e infor mat ion, it is well-known t hat r andom mat ching mechanisms do not achieve mor e expect ed payo¤s t han t he det er minist ic ones ( see [26], [23]) . T his is no longer t he case under asymmet r ic infor mat ion, as we show in sect ion 5.
L et
S
be a coalit ion; let us setT
S=
Q
k2ST
k. For ever yt = (t
k)
k2K2 T
,let
t
S= (t
k)
k2S. A ( r andom) mat ching mechanism3 forS
consist s ofmap-pings
¹
S= (x
S; m
S) : T
! [0; 1]
I£J£ R
K: t
! [(x
ijS(t))
(i;j)2I£J; (m
kS(t))
k2K]
such t hat
² ¹
S is measur abl e w.r .t .T
S, namely¹
S(t) = ¹
S(t
0)
for ever yt; t
02 T :
t
S= t
0S, and for ever yt
2 T; x
ij S(t) = 0
ifi =
2 S
orj =
2 S
andm
kS(t) = 0
ifk =
2 S
²
P
j2Jx
ij S(t) =
P
j2J\Sx
ijS
(t)
· 1
for ever yi
2 I
andP
i2Ix
ij S(t) =
P
i2I\Sx
ij S(t)
· 1
for ever yj
2 J
²
P
k2Km
kS(t) =
P
k2K\Sm
kS(t)
· 0
2T ypically, coalit ions for m at t he ex ant e st age, i.e., before t he agent s know t heir t ypes.
I n sect ion 4, we will assume t hat coali t ions for m at t he int erim st age, but t hat agent s do not exchange infor mat ion unt il t hey ar e in a coali t ion.
3
W e wi ll use t he t er m “ mat ching mechanism” t o r efer t o a “ feasible ( mat ching) mech-anism” sat isfying t he feasibilit y const r aint s.
T he int er pr et at ion of t hese feasibili t y condi t i ons is t hat ever y member of
S
is invit ed t o r epor t his pr ivat e informat i on t o t he mechanism¹
S, which, as a funct ion of t he repor t ed t ypest
, mat chesi
andj
wit h pr obabi lit yx
ijS(t)
, and dist r ibut es t he ( expect ed) t r ansfer sm
kS
(t)
.4 T he feasi bilit y condit ions fur t herasser t t hat a coal it ion can only use t he infor mat ion of it s member s, t hat some agent s may be left unmat ched, and t hat t r ansfer s must balance. Obser ve t hat a mat ching mechanism
¹
S, associat ed wit h an ar bit r ar y coalit ionS
, can be viewed as a mat ching mechanism for t he gr and coalit ionK
, which leaves unmat ched all agent s inK
nS
.I n or der t o de…ne incent ive compat ibilit y of
¹
S, consideri
2 I \ S
; lett
i ( r esp.,t
0i) denot e agenti
’s t r ue ( r esp., repor t ed) t ype; agenti
’s expect edut ilit y is
U
i(¹
Sjt
i; t
0i) =
X
t¡iq(t
¡ijt
i)
2
4
X
j2J[f0gx
ijS(t
0i; t
¡i)u
i((t
i; t
¡i); j) + m
iS(t
0i; t
¡i)
3
5
wher e
t
¡i= t
Knfig and for everyt
2 T
,x
i0S(t) = 1
¡
P
j2J
x
ijS
(t)
is t hepr obabilit y t hat agent
i
is left unmat ched by¹
S att
. L et us denot e asU
i(¹
Sjt
i)
t he ( int er im) expect ed ut ilit y of agenti
when he t rut hfully r epor t shis t ype t o
¹
S, namelyU
i(¹
Sjt
i) = U
i(¹
Sjt
i; t
i)
¹
S is i ncenti ve compati ble ifU
i(¹
Sjt
i)
¸ U
i(¹
Sjt
i; t
0i)
for ever yi
2 I \ S; t
i; t
0i2 T
i ( 1)and similar ly for t he agent s
j
2 J \ S
on t he ot her side of t he mar ket .4Ever y x
S(t)2 [0; 1]I£J sat isfying t he above condit ions is a subst ochast ic mat r ix and
t hus a convex combi nat ion of det er minist ic mat chings inf0; 1gI£J ( t his is a var iant of t he well-known B irkho¤-von N eumann t heor em, see, e.g., [26] or [6] for a r ecent applicat ion) . T he descript i on of r andom mechanisms by mat ching pr obabilit ies (as opposed t o pr oba-bilit y dist r ibut i ons over det ermi nist ic mat chings) su¢ ces here, t hanks t o t he propert ies of t he ut ili t y funct ions (such a simpli …cat ion is not always possible in exchange economies, see [11], [12] and t he discussi on following pr oposit ion 1). I n t he same way, t ransfers ar e det erminist ic wit hout loss of gener alit y.
We denot e as
U
k(¹
S)
t he ( ex ant e) expect ed ut ilit y of agentk
2 S
fr omt he mat ching mechani sm
¹
S, namel yU
k(¹
S) =
X
tk
q(t
k)U
k(¹
Sjt
k)
A ll t he not ions int r oduced in t his sect i on ar e illust r at ed in sect ion 5.
3
T h e ex an t e i n cen t i v e com p at i b l e cor e
We de…ne a cooper at ive game, namel y a charact er ist ic funct ion5
V
¤A, fr om t he
pr evious set up. A coalit ion
S
can achieve a payo¤ vect orv = (v
k)
k2S2 R
S,i.e.,
v
2 V
¤A
(S)
, i¤ t her e exist s an incent ive compat ible mechanism¹
S sucht hat
v
k· U
k(¹
S)
for ever yk
2 S
. T he char act er ist ic funct ionV
A¤ is consi st entwit h t he fol lowing scenar io:
²
Ever y coalit ionS
chooses a mat ching mechanism¹
S, coali t ions possibly for m²
A vect or of t ypest
2 T
is select ed accor ding t oq
; ever y agentk
2 K
is informed of his own t ype
t
k2 T
k²
I fS
has for med,¹
S is implement ed, member s ofS
r epor t t heir t ypes t o¹
S, mat chings and t r anfer s ar e r ealized accor ding t o¹
S.One immediat ely checks t hat if
v = (v
k)
k2S2 V
A¤(S)
andw = (w
k)
k2S2
R
S is such t hatP
k2S
w
k=
P
k2S
v
k, t henw
2 V
A¤(S)
. I ndeed,w
can beachieved by modifying t he t r ansfer s used for
v
independent ly of t ypes, i.e., in an incent ive compat i ble way. T he cooper at i ve game is t hus T U and one can descr ibe t he char act eri st ic funct ion asv
A¤(S) = max
¹S IC
[
X
k2S
U
k(¹
S)]
( 2)wher e t he maximum6 is over all incent ive compat ible ( I C) mat ching mech-anisms
¹
S. T hi s de…nes t he ex ante assi gnment game. Obser ve t hat t heex-pect ed sum of t r ansfer s, namely
P
tq(t)
P
k2Sm
kS
(t)
· 0
, is par t of t he above5
A s in [15], we put a “ * ” for t he incent ive compat ible concept s.
6(2) de…nes a “ gener al linear problem” (see, e.g., Gale [14]). I t i s easily checked t hat
ob j ect ive funct ion, and t hat , i mplicit ly,
v
¤A
(
fkg) =
P
t
q(t)u
k(t; 0)
for anysingle agent
k
.A s a benchmar k, we consider t he …rst best char act er ist ic funct ion
v
A inwhich agent s ar e not submit t ed t o incent ive compat ibilit y const r aint s. I n t his case, at an opt imum, t he sum of monet ar y t r ansfer s i s exact l y
0
:v
A(S) = max
¹S[
X
k2SU
k(¹
S)]
= max
xSX
tq(t)
2
4
X
i2IX
j2J[f0gx
ijS(t)u
i(t; j) +
X
j2JX
i2I[f0gx
ijS(t)u
j(t; i)
3
5
Obviously,v
A¤(S)
· v
A(S)
Shapl ey and Shubik [26] ( see also [23]) have shown t hat under complet e infor mat ion, one can r est ri ct oneself on det er minist ic mechanisms wit hout loss of gener alit y. H ence, in t he above expr ession of
v
A(S)
, t he maximum isin fact over all det er minist ic mechanisms
x
S, wit h values inf0; 1g
. A s wealr eady ment ioned, t his pr oper t y is no longer t r ue under incomplet e infor -mat ion. I n example 3 ( sect i on 5) , t he best expect ed payo¤ a coal it ion
S
can get by r elying on det er minist ic incent ive compat i ble mat ching mechanisms is st r ict l y less t hanv
A¤(S)
.A not her int er est ing pr oper t y of t he assignment game wit h complet e in-for mat ion is t hat all gai ns ar e generat ed by pair s of agent s fr om t he di¤er ent sides of t he mar ket ( see [23]) . M or e precisely, t he wor t h of any coal it ion
S
can t hen be comput ed dir ect ly from t he wort hs of al l singlet ons and pair s cont ained in
S
and, as a consequence, t ransfer s ar e onl y needed bet ween pair s of mat ched agent s. T his st r uct ur e does not sur vi ve under incomplet e infor mat ion, even in t he simplest sellersbuyer s models wit h i ndependent pr i -vat e values ( see example 1 sect ion 5) . T he int uit ive r eason for t his is t hat in our model, t he player s exchange infor mat ion aft er coalit ions have for med. I f t hey par t it ion int o pair s at t he coal it ion for mat ion st age, t hey may loose t he int er act ion possibilit ies which ar e o¤er ed by lar ger coalit ions’ mechanisms.Recall t hat t he core of a T U game
v
overK
i s t he set of all vect or payo¤sw = (w
k)
k2K which can be achieved by t he gr and coalit ion, i.e.,P
k2Kw
k·
v(K)
, and cannot be i mpr oved upon by any coal it ion, i.e.,P
k2S
w
k¸ v(S)
cor e” of t he assignment game. U nder complet e infor mat ion, Shapley and Shubik [26] have pr oved t hat t he cor e of t he assignment game is not empt y. A st r aighfor war d applicat ion of t he B ondar eva-Shapley t heor em will enable us t o show t hat Shapl ey and Shubik’s r esult ext ends, namely
C(v
¤A
)
6= ;
.H owever , as we point ed out above, many nice pr oper t ies of t he char act er ist ic funct ion disappear . I n par t icular, t he ex ant e incent ive compat ible cor e does not coinci de wit h t he ( payo¤s of ) stable mat chings ( which ar e individually r at ional and cannot be blocked by any pai r
(i; j)
of agent s inI
£ J
; see examples 1 and 2 i n sect ion 5) .P r op osi t i on 1 T he ( T U) assi gnment game de…ned by
v
A¤ i s balanced, so that the ex ante i ncenti ve compati ble coreC(v
¤A
)
i s non-empty.T he r esult is a dir ect consequence of t he fol lowing lemma:
L em m a 1 Let
S
be a balanced fami ly of coali ti ons wi t h associ at ed wei ght s¸
S,S
2 S
, and let¹
S= (x
S; m
S)
be a mechani sm forS
,S
2 S
. T hen¹ =
P
S2S¸
S¹
S i s a matchi ng mechani sm forK
and sat i s…esU
k(¹
jt
k; t
0k) =
X
S2S:S3k
¸
SU
k(¹
Sjt
k; t
0k)
8k; t
k; t
0k ( 3)I n par ti cular , i f ever y
¹
S i s i ncenti ve compati ble, so i s¹
.P r o of of t h e l em m a: Recal l t hat a mechanism
¹
S for an ar bit r ar y coal it ionS
can be viewed as a mechani sm for t he gr and coalit ion. T he mechanism¹
de…ned i n t he st at ement isT
-measur abl e as a linear combinat ion ofT
-measur able mechanisms. B y de…nit i on of balancedness,P
S2S:S3k¸
S= 1
forever y
k
2 K
;¹ = (x; m)
is not , as such, a convex combi nat i on of t he¹
S’sbut sat is…es
x
ij(t) =
X
S2S:S¶fi;jg¸
Sx
ijS(t)
8i 2 I; j 2 J; t 2 T
( 4)m
k(t) =
X
S2S:S3k¸
Sm
kS(t)
8k 2 K; t 2 T
( 5)so t hat
¹
is feasi ble. I ndeed, letj
2 J
andt
2 T
:X
i2Ix
ij(t) =
X
i2IX
S2S:S¶fi;jg¸
Sx
ijS(t)
=
X
S2S:S3j¸
SX
i2I\Sx
ijS(t)
· 1
and similar ly on t he ot her side of t he mar ket . T he t r ansfer s sat isfy
X
k2Km
k(t) =
X
k2KX
S2S:k2S¸
Sm
kS(t) =
X
S2S¸
SX
k2Sm
k S(t)
· 0
T he equalit ies (3) can be checked in a similar way:
X
j2Jx
ij(t
0i; t
¡i)u
i((t
i; t
¡i); j)
=
X
j2J[
X
S2S:S¶fi;jg¸
Sx
ijS(t
0i; t
¡i)]u
i((t
i; t
¡i); j)
=
X
S2S:S3i¸
SX
j2J\Sx
ijS(t
0i; t
¡i)u
i((t
i; t
¡i); j)
Simil ar ly,x
i0(t
0i; t
¡i)u
i((t
i; t
¡i); 0)
= [1
¡
X
j2JX
S2S:S¶fi;jg¸
Sx
ijS(t
0i; t
¡i)]u
i((t
i; t
¡i); 0)
=
X
S2S:S3i
¸
S[1
¡
X
j2J\S
x
ijS(t
0i; t
¡i)]u
i((t
i; t
¡i); 0)
and t he t r ansfer s sat isfy ( 5) .
Q.E.D . P r o of of t h e p r op osi t i on : L et
S
be a bal anced fami ly of coalit ions wi t h associat ed weight s¸
S,S
2 S
. We must show t hatv
¤A(K)
¸
P
S2S
¸
Sv
A¤(S)
.L et , for ever y
S
2 S
,¹
S be a mechanism achieving t he maximum in ( 2) , namelyv
¤A
(S) =
P
k2S
U
k(¹
S)
and de…ne¹
as in t he lemma.¹
is incent i vecompat ible. B y t he linear it y of ut ilit y funct ions,
v
A¤(K)
¸
X
k2KU
k(¹) =
X
k2KX
S2S:S3k¸
SU
k(¹
S)
=
X
S2S¸
SX
k2SU
k(¹
S) =
X
S2S¸
Sv
¤A(S)
Q.E.D . L et us compar e t he pr evious pr oposit ion wit h t he r esult s obt ained for t he ex ant e incent ive compat ible core of an exchange economy in Vohr a [27],For ges and M inelli [12] and Forges, M er t ens and Vohr a [11]. T he seller s-buyer s example ( wit hout mat ching ext er nalit ies for t he sell er s) is basi cally a par t icular case of t he economies consider ed in [11], wher e ar bit r ar y monet ar y t r ansfer s ar e allowed and ut ili t y funct i ons ar e quasi -linear. A s shown in [11], t hese condi t i ons do not guar ant ee t he non-empt iness of t he ex ant e incent i ve compat ible cor e. H er e, t he ut ili t y funct ions ar e in fact linear , and t his even if one r est ri ct s oneself t o det er mini st ic mechanisms ( in which all pr obabilit ies of mat ching ar e
0
or1
) .T he ex ant e incent ive compat ible cor e of an exchange economy wit h linear ut ilit y funct ions is known t o be non-empt y; r andom mechanisms ar e clear ly useless in t hat fr amewor k ( see, e.g., [27], [13]) . L et us show t hat t his r esult cannot be applied dir ect ly here. I t is t empt i ng t o vi ew t he mat ching pr oba-bilit ies of sect ion 2 as quant it ies of i ndivi sible goods, e.g., by appealing t o a t ime-shar ing int er pr et at ion. H owever , t he st andar d r esour ce const r aint s do not account for t he feasibi lit y const r aint s st at ing t hat each buyer is mat ched wit h ( t ot al) pr obabilit y less t han one at ever y st at e of nat ur e. T his should r esult fr om an assumpt i on on t he buyer s’ init ial ut ilit y funct ions, namely t hat , at every st at e of nat ur e, a buyer ’s ut ilit y for sever al houses is j ust his ut ilit y for his favor it e house. Such ut i lit y funct ions ar e obviously not li near . To r est r ict on t ime-shar es ( as opposed t o pr obabilit y dist r ibut ions over de-t er minisde-t ic al locade-t ions) and ude-t il ide-t y funcde-t ions de-t hade-t ar e linear in de-t hese, as in sect ion 2 ( recall foot not e 4) , a const r uct ion much mor e t edious t han above is needed. T his appr oach gener at es cor e mechanisms which ar e feasi ble in expect at ion. T hese ar e always feasi ble her e ( see again foot not e 4) , unlike in gener al exchange economies wi t h asymmet r ic infor mat ion.
L et us fur t her illust r at e t he di ¤er ences bet ween t he assignment game and an exchange economy by consider ing a simple example. L et
I =
f1; 2g
andJ =
f3; 4g
; consider t he balanced familyS = ff1; 2; 4g; f1; 3g; f2; 3g; f4gg
wit h all weight s equal t o 12 and t he following det er mi nist ic mat chings for some st at e of nat ur et
:2
and4
inf1; 2; 4g
,1
and3
inf1; 3g
,2
and3
inf2; 3g
( and null t ransfer s) . T he mechanism const r uct ed in t he lemma is, att
,x(t) =
µ
1 20
1 2 1 2¶
=
1
2
µ
1 0
0 1
¶
+
1
2
µ
0 0
1 0
¶
which is feasible as expect ed. L et us now view t he example as an exchange economy, in which agent
1
and agent2
, t he seller s, bot h init iall y have1
unit of good. T he fol lowing det er minist i c mechanisms ar e now feasible: agent2
get s
2
unit s inf1; 2; 4g
; agent3
get s1
unit inf1; 3g
and1
unit as wel l inf2; 3g
. Pr oceeding as in t he lemma yields t he mechani sm which al locat es0
or
2
unit s, each wit h pr obabili t y 12, t o agent2
and1
unit wit h pr obabilit y1
t o agent3
. T his mechanism cannot be feasible since it di st r ibut es3
unit s wit h posit ive pr obabilit y. H owever , t he average mechanism, which gives1
t o agent2
and1
t o agent3
i s feasible7. I f t he ut ilit y funct ions ar e l inear , t heaver age mechanism is obviously equivalent t o t he r andom one, but t hen, by cont r ast wit h t he assignment game, r andom mechanisms ar e not useful at all.
4
T h e i n cen t i v e com p at i b l e coar se cor e
U p t o now, we have assumed t hat coal it ions for m at t he ex ant e st age, be-for e agent s know t heir t ypes. Such a scenar io is obviously not always feasible ( see, e.g.,[15]) : if t ypes r epr esent int r insi c char act er ist ics of t he agent s, t he mat ching pr ocedur e cannot st ar t befor e t he i nter i m st age, in which ever y agent only knows his own t ype. I f we mai nt ain t he assumpt ion t hat agent s do not exchange infor mat ion unt il t hey ar e in a coalit ion, t he incent ive com-pat ible ver sion of W ilson [29]’s coar se core pr oposed by Vohr a [27] is t hen an appr opr iat e not ion of t he cor e. I n par t icular , solut ions in t he incent i ve compat ible coar se cor e ar e i nt er i m i ndi vi dual ly rati onal in t he st andar d sense ( see [18], [19], [21] et c.) and i ncenti ve i nter i m e¢ ci ent ( see [15]) . A gent s wil l exchange infor mat ion only aft er having for med coalit ions, by means of t he same mechanisms as in sect ion 2, but will possi bly block pr oposals at t he int er im st age. A s a consequence, t hey will base t heir ob j ect ions on event s t hat ar e common knowledge inside t he coalit ion at t he int er i m st age.
I n or der t o make t he for mal de…nit ion of common knowledge event s ( and t hus of t he incent i ve compat ible coar se core) as simple as possible, we shal l make, t hr oughout t his sect ion, t he assumpt ion t hat all vect or of t ypes occur wit h posit ive pr obabilit y:
q(t) > 0
for ever yt
2 T
, e.g., t hat t ypes ar e independent .8 I n t his case, t he set of all t ypesT
is t he only event t hat can be common knowledge i n a coalit ionS
of t wo agent s or mor e at t he int er im7T he present example is simplist ic since t her e is only one good. [11] fully de…nes a
T U exchange economy in which t he ex ant e incent i ve compat ible core is empt y. H ence, t he corr esponding T U game cannot be balanced (which is obviously much st r onger t han ver ifying t hat t he most naive const ruct ion t o est ablish balancedness does not work, as we did above) .
st age ( since t he infor mat ion part it ion of agent
k
of t ypet
0kis t he set of al l t ype
vect or s
t
such t hatt
k= t
0k) . Obviously, ifS =
fkg
, “ common knowledge” issynonymous wi t h “ knowledge” , and at t he i nt er im st age, agent
k
knows his own t ypet
k.L et
¹
be a mat ching mechanism for t he gr and coalit ion and letº
S be amat ching mechanism for coalit ion
S
,jSj ¸ 2
, as de…ned in sect ion 2.º
S is acoar se obj ecti on t o
¹
i¤U
k(º
Sjt
k) > U
k(¹
jt
k)
8k 2 S; t
k2 T
k ( 6)I n a similar way, agent
k
of t ypet
k has a coar se ob j ect ion t o¹
i¤X
t¡k
q(t
¡kjt
k)u
k(t; 0) > U
k(¹
jt
k)
( 7)so t hat agent
k
blocks¹
at t he int er im st age as soon as t her e exist st
k2 T
ksat isfying ( 7) .
¹
is an incent ive compat ible coar se cor e mechanism i¤¹
is incent i ve compat ible and no coalit ionS
has an incent ive compat ible coar se ob j ect ion t o¹
. I n par t icular,¹
is incent ive int er im e¢ cient , namely t her e is no incent i ve compat ible mechanismº = º
K such t hat ( 6) is sat is…ed forS = K
, and¹
isint er im i ndivi dually r at i onal, namel y
U
k(¹
jt
k)
¸
X
t¡k
q(t
¡kjt
k)u
k(t; 0)
8k 2 K; t
k2 T
kI n a similar way as in sect ion 3, we de…ne t he incent ive compat ible coar se cor e as t he set of (i nt er im) payo¤ s ( of t he for m
[(w
k(t
k))
tk2Tk]
k2K2 R
N,
N =
P
kjT
kj
) t o incent ive compat ible coar se cor e mechanisms. A s point edout in [27] and [13], t her e is no inclusion r elat ionship bet ween t he ex ant e incent ive compat ible cor e and t he i ncent ive compat ible coar se cor e; indeed, int er im individual r at ionalit y obviously implies ex ant e individual r at ionalit y, but t he implicat ion goes t he ot her way r ound for e¢ ciency ( see [15]) .
W ilson [29] showed t hat , in t he absence of incent ive const r aint s, t he coar se cor e of a well-behaved exchange economy is non-empt y, as t he st andar d cor e of an appr opr iat ely de…ned balanced N T U cooper at i ve game wit h
N
player s(k; t
k)
,k
2 K
,t
k2 T
k ( see also [13]) . Vohr a [27] ext ended his ar gument t ocoarse cor e as in [27] and [13]. Fr om a concept ual point of view, t he concept of t he coarse cor e seems mor e at t r act ive when no st at e of nat ur e has zer o probabilit y.
exchange economies wit h linear ut ilit y funct ions. We pr oceed i n t he same way her e and est ablish t hat t he incent ive compat ible coar se cor e is t he cor e
C(V
I¤)
of a char act er ist ic funct ion
V
¤I. T he player s in t his game ar e, as in W ilson
[29], t he t ypes of t he or iginal model; we call t hem “ auxiliar y pl ayer s” ; t he only vi abl e coal it ions ar e of t he for m
f(k; t
k)
g
for somek
2 K
and some t ypet
k2 T
k, or of t he for mf(k; t
k) : k
2 S; t
k2 T
kg
for some or iginal coal it ionS
µ K
,jSj ¸ 2
. T he int er pr et at ion is clear : at t he int er im st age, a coal it ion consist s eit her of a si ngl e agent who knows his t ype or of sever al agent s who have t r ivial common knowledge ( as a consequence of our assumpt ions) and t hus consider all t heir t ypes as possible. T he viable coali t ions of t wo agent s or mor e will t hus be ident i …ed wit h t he or igi nal coalit ionsS
µ K
,jSj ¸ 2
wit h a lit t le abuse of language.T he char act er ist ic funct ion
V
¤I is now easily de…ned on viable coalit ions.
For ever y
(k; t
k)
,V
I¤(
f(k; t
k)
g)
is t he set of vect or payo¤sv
2 R
N in whichagent
k
of t ypet
kget s at most his individually rat ional levelP
t¡kq(t
¡kjt
k)u
k(t; 0)
.For ever y coalit ion
S
µ K
,jSj ¸ 2
,V
¤I
(S)
is t he set of vect or payo¤sw
2 R
Nfor which t her e exist s an incent ive compat ible mechanism
¹
S for coalit ionS
such t hat
w
k(t
k)
· U
k(¹
jt
k)
for ever yk
2 S
andt
k2 T
k. F inal ly, as usual ,let us de…ne
V
¤I on ar bit r ar y coalit ions by t aking t he super addit ive cover .
T his amount s t o allowing t he auxili ar y player s
(k; t
k)
of a coalit ion t o usean ar bit r ar y incent ive compat ible mat ching mechanism if all
(k; t
0k)
,t
0k2 T
k,are also in t he coalit ion and t o leaving unmat ched t he ot her auxiliar y play-er s.
V
I¤ is a well-de…ned N T U game, but clear ly, t he ar gument showing t hatV
A¤ de…nes a T U game does not ext end her e ( see [11] and [13] for r elat ed comment s) .T he core of t he int er im assignment game
V
I¤is t he set of all vect or payo¤sw = [(w
k(t
k))
tk2Tk]
k2K2 V
I¤(K)
which cannot be impr oved upon by any( viable) coalit ion. We deduce t he non-empt iness of
C(V
¤I
)
fr om Scar f ’s [24]t heor em.
P r op osi t i on 2 T he ( NT U) assi gnment game de…ned by
V
¤I i s balanced, so
that the i ncenti ve compati ble coar se core
C(V
¤I
)
i s non-empty.P r o of : L et
B
be a balanced family of viabl e coalit ions i n t he auxiliar yN
player game de…ned above, wit h associat ed weight s
¸
B,B
2 B
. We mustshow t hat
\
B2BV
I¤(B)
µ V
I¤(K)
. A ssume t hat , for somek
andt
k, t hesin-glet on
f(k; t
k)
g
belongs t oB
; t hen by balancedness, all singlet onsf(k; t
0k)
g
,t
0k
2 T
k, must also be inB
, all wit h same weight , say¸
k. We can t husby mer ging al l
B
’s element sf(k; t
k)
g
,t
k2 T
k, int o a single coal it ion, wi t hweight
¸
k ( and keepi ng unchanged t he coali t i ons of at least t wo player s andt heir weight s, since t hese al ready corr espond t o or iginal pl ayer s) . Since by de…nit ion of
V
¤I, for ever y
k
2 K
,V
I¤(
f(k; t
k) : t
k2 T
kg) = \
tkV
I¤(
f(k; t
k)
g)
,we also have
\
B2BV
I¤(B) =
\
S2SV
I¤(S)
. We can t hus pur sue t he r easoningin t er ms of t he or iginal player s, but consider ing vect or payo¤s indexed by t he t ypes.
L et
w = [(w
k(t
k))
tk2Tk]
k2K2 \
S2SV
I¤(S)
: for ever yS
2 S
, t here exist san incent ive compat ible mechanism
¹
S achievingw
. L et us de…ne¹
as in lemma 1.¹
is an incent ive compat ible mechanism for t he gr and coali t ion, such t hatU
k(¹
jt
k)
¸ w
k(t
k)
for ever yk
,t
k. H encew
2 V
I¤(K)
.Q.E.D . R em ar k s:
( i) T he pr evious analysis shows t hat under t he assumpt ion t hat all t ype vect or s have posit ive pr obabilit y, t he r elevant char act er ist ic funct ion at t he int er im st age can be de…ned over t he or iginal coali t i ons
S
µ K
, by consid-er ing as achievable all vect or payo¤s[(w
k(t
k))
tk2Tk]
k2S t hat r esult fr om anincent ive compat ible mechanism. Coal it ion
S
bl ocks a mechanism¹
if t her e is an incent ive compat ible mechanismº
S such t hat (6) hol ds. Of cour se, ifS
is a singlet on, t he incent ive compat ibil it y condit ions ar e vacuous so t hat a single agent blocks any mechanism t hat is not int er im individually r at io-nal. T he const r uct ion of t he auxiliar y game is only necessar y t o use Scar f ’s t heor em, which is not for mulat ed for vect or payo¤s.( ii) U nlike t he cor e of a mat ching game wit h complet e infor mat ion, and as t he ex ant e incent ive compat ible cor e, t he incent i ve compat ible coarse cor e is smaller t han t he set of int eri m payo¤s which cannot be blocked ( in t he sense of ( 6) and ( 7) ) by any single agent nor by any pai r
(i; j)
2 I £ J
(see example 1 in sect ion 5) .5
E x am p l es
I n t hi s sect ion, we mot i vat e t he basic model and t he sol ut ion concept of sec-t ions 2 and 3 on simple housing mar kesec-t s and we illussec-t r asec-t e sec-t wo di¤er ences bet ween mat ching games wit h complet e and incomplet e infor mat ion. Ex-ample 1 consist s of a simple auct i on ( one sell er ) wit h independent pri vat e values, in which ob j ect ions by coalit ions of more t han t wo agent s mat t er .
T he lat t er pr oper t y i s also t r ue i n t he second example, but i s less sur pr ising t her e, because values ar e common and cor r elat ed. T he goal of t his example is t o show t hat t he grand coalit ion deals bet t er wit h incent ive compat ibilit y t han small coalit ions. F inally, in example 3, random mechani sms improve t he char act er ist ic funct i on under incomplet e infor mat ion.
I n t he sequel, for
i
2 I
andj
2 J
,¡u
i(t; j) = ¼
i(t; j)
is int er pret ed ast he mini mum pr ice at which seller
i
is willing t o sell his house t o buyerj
when t he st at e of infor mat ion is
t
andu
j(t; i)
is int er pr et ed as t he maximumpr ice t hat buyer
j
i s willing t o pay for sel leri
’s house. T he ut ilit y of being unmat ched is0
for ever y t r aderk
.E x am p l e 1
A gent
1
, t he seller , has no pr ivat e infor mat ion and a null r eser vat ion pr ice; agent s2
and3
, t he pot ent ial buyer s, have independent pr ivat e values unifor mly dist r ibut ed overf0; 1g
.9 T he basic par amet er s of t he assignment game ar e t hus:I =
f1g
,J =
f2; 3g
,T
j=
f0; 1g
,j = 2; 3
,t
2 andt
3 ar ei.i.d. unifor mly over
f0; 1g
,u
1(t; 2) = u
1(t; 3) = 0
,u
j(t; 1) = u
j(t
j; 1) = t
j,j = 2; 3
.We shal l show t hat in t his example,
v
¤A(
f1; 2g) = v
A¤(
f1; 3g) =
1
2
whil e
v
A¤(
f1; 2; 3g) =
3
4
H ence a feasi ble vect or payo¤ like
(
12; 0; 0)
cannot be bl ocked by any t wo agent coal it ion but can be blocked by t he gr and coalit ion.Since t he seller is not submi t t ed t o incent ive const r aint s, one can r est r ict on t r ansfer s summing up t o
0
wit hout loss of gener alit y. I n par t icul ar , in a seller -buyer coalit ionf1; jg
,j = 2
or3
, a mechanism(x; m)
can be de…ned by t he pr obabilit y of t r adex(t
j)
and t he expect ed t r ansferm(t
j)
fr om buyerj
t o t he seller ,t
j= 0; 1
. T he t ot al expect ed gai ns fr om mechanism(x; m)
ar e 12
x(1)
·
12. H ence
v
¤A(
f1; 2g) = v
A¤(
f1; 3g) ·
12. On t he ot her hand, 12 can be
9T he example cor responds t o t he simplest possible discr et e model. Everyt hi ng goes
t hr ough if for inst ance, t he pair f0; 1g is r eplaced by t he int erval [0; 1]. I n t hi s case,
v¤
achieved wit h a ( const ant ) mechanism sel ling always t he ob j ect at t he pr ice
1 2.
L et us t ur n t o t he grand coalit ion and consider t he feasible, incent i ve compat ible mechanism induced by a second pr ice auct ion, namely
0 1
0 no sale sell t o 3 at pr ice 0
1 sell t o 2 at pr ice 0 sell t o 2 or 3 wit h pr obabilit y 12 at pr ice 1
wher e t he r ows ( r esp., columns) cor r espond t o t he r epor t ed t ypes of agent
2
( r esp.,3
) . T he t ot al expect ed gains fr om t r ade ar e 34, which is t he maxi-mum expect ed value t hat can be achieved, even in t he absence of incent i ve const r aint s.One can also check t hat no seller -buyer coalit i on can bl ock t he pr evi-ous mechanism at t he i nt er i m st age. I ndeed, t his mechanism is classically ( i.e., wit hout incent ive const r ai nt s) ex ant e e¢ cient and int er im individually r at ional.10
E x am p l e 2
T he seller s, agent s
1
and2
, bot h know, at t he int eri m st age, whet her t he quali t y of t heir own house i s high or low:T
1=
fh
1; l
1g
,T
2=
fh
2; l
2g
.T he pr obabi lit y dist r ibut ion over t he seller s’ t ypes is
q(h
1; h
2) = q(l
1; l
2) =
38,q(h
1; l
2) = q(l
1; h
2) =
18. T he pot ent ial buyer s, agent s3
and4
, have no pri vat einfor mat ion. T he r eser vat ion pr ice of a seller for a high (r esp., low) qualit y house is
¼
h ( r esp.,¼
l) , while for a buyer , t he r eser vat i on pr ices ar eu
h andu
l. We assume t hatu
l< ¼
l< ¼
h< u
h and t hat 21(u
l+ u
h) <
12(¼
l+ ¼
h)
( e.g.,0, 9, 12, 20) .11
10
I f t he buyers’ t ypes ar e unifor mly dist ribut ed over [0; 1], coalit ion f1; jg, j = 2; 3, can achieve t he vect or payo¤ (1
4; (tj¡12)I(tj¸ 12))wit h an incent ive compat ible, int erim
individually r at ional, mechanism. Fur t her mor e, one can check t hat 14 is t he maxi mum t he sell er can expect from a mechanism wit h t hese pr opert i es in a t wo-agent coal it ion. I n t he gr and coal it ion, t he second pr ice auct ion mechanism yields t he expect ed payo¤s(1
3; t2 2 2; t2 3 2)
and t hus st ricly impr oves all agent s’ payo¤s.
11For a concret e example, t hink of t he t er mit es invading some r egions of Fr ance. T he
qualit ies of houses in t he same neighbor hood are highly corr elat ed. M any owners do not know whet her t hei r house is infect ed or not , but wil l go t hrough a t est in case of pot ent ial sale. H ence, t here is an ex ant e st age, befor e t he seller s know t he qualit y of t heir houses.One might ar gue t hat t he qualit y of a house is ver i…able in t his example but , given t he r isk of false cert i…cat es, an incent ive mechanism looks safer .
Since t he buyer s ar e not submit t ed t o incent ive const r ai nt s, we will, wit h-out loss of gener alit y, focus on t r ansfers summing up t o
0
t hr oughout t he example.L et us …rst consider a seller -buyer coali t i on
fi; jg
,i
2 f1; 2g
,j
2 f3; 4g
. T he seller has t hen t wo equipr obable t ypes, which we denot e ash
andl
. We face a simple, di scr et e ver sion of M yer son [19]’s “ lemon pr oblem” . A mecha-nism for a seller -buyer coalit ion consi st s of t he pr obabilit y of t r adex
h ( resp.,x
l) when t he sel ler repor t s t ypeh
( r esp.,l
) and t he cor r esponding expect edt r ansfer s
m
h,m
l fr om t he buyer t o t he seller . B y eliminat ing t he t r ansfer sfr om t he incent ive const r aint s ( see, e.g., [17] or [19]) , t he opt imizat ion pr ob-lem of a seller -buyer coalit ion is
v
A¤(
fi; jg) = max[
1
2
x
h(u
h¡ ¼
h) +
1
2
x
l(u
l¡ ¼
l)] s:t: 0
· x
h· x
l· 1
so t hat , under our assumpt ions,
v
¤A(
fi; jg) = 0 i 2 f1; 2g; j 2 f3; 4g
L et us set
g
h= u
h¡ ¼
h> 0
. Obser ve t hat , in absence of incent i veconst r aint s,
v
A(
fi; jg) =
12g
h. T r ade is indeed bene…ci al in st at eh
, but t heincent ive compat ibilit y condit ions pr event r evelat ion of infor mat ion fr om t he seller .
L et us t ur n t o t he gr and coalit ion. F ir st best e¢ ciency r equir es t o sel l t he high qualit y houses, and only t hose, at ever y st at e of nat ur e. H ence,
v
A(K) = g
hWe will const ruct an incent ive compat ible mechanism achievi ng
g
h as sum ofexpect ed payo¤s, so t hat
v
¤A(K) = g
hSince
v
A(S)
¸ v
A¤(S)
for ever yS
, t his wil l also show t hatC(v
A)
µ C(v
A¤)
and pr ovide a simple pr ocedur e t o const r uct expect ed payo¤s in t he ex ant e incent ive compat ible cor e.12
12T he procedur e can be appli ed t o a large class of assignment pr oblems (see, e.g., [2],
Consider a mat ching mechani sm in which only high qualit y houses ar e sold, e.g.,
x
2 f0; 1g
I£J descr ibed byx(h
1; h
2) =
µ
1
0
0
1
¶
x(h
1; l
2) =
µ
1 0
0 0
¶
x(l
1; h
2) =
µ
0
0
0
1
¶
x(l
1; l
2) =
µ
0 0
0 0
¶
T he sum of expect ed payo¤s fr om
x
i sg
h. Obviously,x
is not incent i vecompat ible, but one can const r uct t r ansfer s
m
such t hat(x; m)
is incent i ve compat ible. For inst ance, t he t r ansfer sm
1 t o t he …rst seller must sat isfy¡¼
h+
3
4
m
1(h
1; h
2) +
1
4
m
1(h
1; l
2)
¸
3
4
m
1(l
1; h
2) +
1
4
m
1(l
1; l
2)
1
4
m
1(l
1; h
2) +
3
4
m
1(l
1; l
2)
¸ ¡¼
l+
1
4
m
1(h
1; h
2) +
3
4
m
1(h
1; l
2)
A possible sol ut ion is
m
1(h
1; h
2) =
3¼h2¡¼l+
g4hm
1(h
1; l
2) =
3¼l¡¼2 h+
g4hm
1(l
1; h
2) =
g4hm
1(l
1; l
2) =
g4hT he t r ansfer s
m
2 t o t he second seller can be chosen in a si milar way. I norder t o balance t he t r ansfers, one can simpl y set
m
3=
¡m
1,m
4=
¡m
2.T he mechanism
(x; m)
t hus associat es buyer3
(r esp.,4
) wit h seller1
( resp.,2
) but sale only t akes place if t he sell er ’s house is of high qualit y.(x; m)
yields t he expect ed payo¤ gh
4 t o each t r ader. T he mechanism re‡ect s t hat
sale pr ices ar e in‡uenced by t he pr esence of l ow qualit y it ems; t he t r ansfer s in t he low st at e should be int er pr et ed as a fee t hat t he pot ent ial buyer s pay t o get infor mat ion and avoid a bad decision. M any ot her mechanisms achieving ex post e¢ ciency can be const r uct ed. I n par t icular , as in [8], it is possible t o design t he t r ansfer s in such a way t hat t he mechanism i s int er im individually r at ional for t he seller s, who fully ext r act t he sur plus13.
13A dding (resp., subt ract ing) gh
4 t o (resp., fr om) all previous t r ansfer s gives t he sur
-plus t o t he seller s (r esp., buyers). A ll t hese mechanisms, including t he lat t er , are int erim individually r at ional for t he seller s. A not her mechanism wit h t he same proper t ies is
m1(h1; h2) =3uh2¡¼l m1(h1; l2) = 3¼l¡u2 h
L et us end t he analysis of t he example by showing t hat t he expect ed payo¤ fr om
(x; m)
, namely,(
gh 4;
gh 4;
gh 4;
gh4
)
, belongs t oC(v
¤A)
. We have evaluat edv
A(
fi; jg)
,i
2 f1; 2g
,j
2 f3; 4g
andv
A(K)
. T o complet e t he descr ipt ion ofv
A, we comput e t hatv
A(
fi; 3; 4g) =
g
h2
i = 1; 2
v
A(
f1; 2; jg) =
5g
h8
j = 3; 4
H ence,(
gh 4;
gh 4;
gh 4;
gh4
)
2 C(v
A)
. I t foll ows fr om our pr evious r emarks t hat(
gh 4;
gh 4;
gh 4;
gh4
)
2 C(v
A¤)
. T he same r easoni ng applies t o(
gh2
;
gh2
; 0; 0)
.A s announced above, t his example shows t hat it may be bet t er for t he agent s ( i n t he sense of gener at ing a higher sum of expect ed payo¤s) t o st ay t oget her at t he ex ant e ( or int er im) st age in or der t o exchange infor mat ion wi t hi n the grand coali ti on. For inst ance, t his enables t he agent s t o exploit t he possible cor r el at ion bet ween t ypes and t o achieve …rst best e¢ ciency t hr ough full r evelat ion.
E x am p l e 3
T his example will con…rm t he advant age of r andom mechanisms, by illus-t r aillus-t ing illus-t haillus-t
v
A¤(K)
can be st r ict ly lar ger t han t he maxi mum sum of payo¤s t hat can be achieved wit h an incent ive compat ible det er mi ni sti c mat ching mechanism ( t o which we will r efer asv
¤AD
(K)
) . T he fr amewor k will consistof one sell er ( agent
1
) and one buyer ( agent2
) , wit h i ndependent typest
1and
t
2, but common values: at t he st at e of infor mat ion(t
1; t
2)
, t he r eserva-t ion pr ice of va-t rader
1
( r esp.,2
) is¼
1(t
1; t
2)
( r esp.,u
2(t
1; t
2)
) . We will fur t herassume t hat each t r ader only has t wo equipr obabl e t ypes (
T
k=
ft
k1; t
k2g
,q(t
k1) = q(t
k2) =
12,k = 1; 2
) .T his seems one of t he simplest models in whi ch one can expect t o have
v
AD¤(K) < v
¤A(K)
. For inst ance, under independent pr ivat e values, …rst best e¢ ciency can be achieved in an incent ive compat ible way, so t hatv
¤A
(K) =
v
A(K)
( see, e.g., [3] and [4] and t he concluding r emar ks below) and t husv
AD¤(K) = v
A¤(K)
in t his case. I t is a r emar kable fact t hatv
¤AD(K)
andv
¤A
(K)
also coincide in seller -buyer models wit h common values in whichonly one t r ader has pr ivat e infor mat ion, as it can be checked by adapt ing M yer son [19]’s resul t s on t he “ lemon pr oblem” of t o our discr et e fr amewor k. L et us complet e our example by assuming t he following possible r eser va-t ion pr ices
¼
1(t
1; t
2)
,u
2(t
1; t
2)
:t
21t
22t
116; 2 2; 3
t
127; 6 0; 4
Since t he t r ader s’ t ypes ar e independent , one can r est r ict on mechanisms in which t he t r ansfer s sum up t o
0
wit hout loss of gener alit y14. Fur t her mor e, byeliminat ing t he t r ansfer s fr om t he incent ive compat ibilit y condit ions ( see e.g., Johnson, Pr at t and Zeckhauser [17]) , we can wr it e t he opt imizat i on pr oblem of t he seller -buyer coali t i on as
v
A¤(
f1; 2g) = maxf¡4®
11+ ®
12¡ ®
21+ 4®
22g
s.t .
0
· ®
rs· 1 r; s = 1; 2
and®
11¡ 2®
12¡ ®
21+ 2®
22¸ 0
( 8)¡®
11+ ®
12+ 2®
21¡ 2®
22¸ 0
( 9)wher e
®
rs is t he pr obabilit y of t rade when t he seller ( r esp., buyer ) r epor t shis
r
th ( r esp.,s
th t ype)r; s = 1; 2
. ( 8) is t he seller ’ s incent ive compat ibilit yconst r aint ( aft er eliminat ion of t he t r ansfer s) which expr esses t hat t he sum of his expect ed payo¤s when he always t ells t he t r ut h is larger t han t he sum of his expect ed payo¤s when he always li es. ( 9) has a similar int er pr et at ion for t he buyer .
T he only ex post e¢ ci ent mechani sm in t his example,
® =
0 1
0 1
, i s notincent ive compat ible for t he buyer . T he mechanism
® =
0
2 3 2 3
1
sat is…es ( 8) and ( 9) - wit h an equalit y - so t hat
v
¤A(
f1; 2g) ¸ 4
.1514A s shown in, e.g., [17], if t ypes ar e independent and ¹ = (x; m) is incent ive
com-pat ible, so ar e (x; m), mk(tk) = Pt¡kq(t¡k)mk(t), and (x;m)e , mek(t) = mk(tk)¡
1 jKj¡1
P
l6=kml(tl).m(t)e is exact ly balanced for ever y t. 15I n fact , one can check t hat v¤
A(f1; 2g) = 4. U nder t he feasibilit y const r ai nt s, t he
ob j ect ive is· ®12¡ ®21+ 4while t he sum of t he incent ive compat ibilit y const raint s yields
We now check t hat
v
¤AD
(
f1; 2g) < 4
. L et®
be a det er minist ic mechanism;if
®
11= 1
, t he ob j ect ive¡4®
11+ ®
12¡ ®
21+ 4®
22 is· 1
; and si milar ly if®
22= 0
.v
AD¤(
f1; 2g)
is t he maximum of®
12¡ ®
21+ 4
s.t .®
12,®
212 f0; 1g
,( 8) and (9) ( wit h
®
11= 1
and®
22= 0
) , which is3
.6
C on cl u d i n g r em ar k s
I n t his paper , we have focused on t he ex ante assignment game ( which is de…ned wit hout any ambiguit y) and a possible version of t he int er im assign-ment game, in which agent s do not communicat e unt il t hey ar e in a coali-t ion. W e have escoali-t ablished coali-t hacoali-t coali-t he associacoali-t ed cor es ar e non-empcoali-t y. Ocoali-t her cor e concept s have been pr oposed at t he int er im st age, st ar t ing wit h W ilson [28]’ s …ne core ( de…ned in absence of incent ive const r aint s). T he de…nit ion of an int er im cor e concept t ur ns out t o be quit e delicat e, especially if some communicat ion is allowed at t he coalit ion for mat ion st age ( see, e.g., [9] [10], [15], [13], [20]) and we will not invest i gat e t his pr oblem fur t her her e. B ut t he pr evi ous analysis suggest s t hat t wo-sided mat ching games wit h i ncom-plet e infor mat ion could be an appr opr iat e fr amewor k in which t o appl y or devel op int er im cor e concept s. I ndeed, t he model has a simple, well-behaved st r uct ur e, a number of par t icular cases ar e well-under st ood ( e.g., seller -buyer bar gaini ng, auct ions,...) and t he benchmar k incent ive compat ible coarse cor e is not empt y.
L et us t ur n t o possible ex post pr oper t i es of ex ant e incent ive compat ible cor e solut ions. A number of paper s ( see, e.g., [2], [3], [4], [7], [8], [17]) ident ify assumpt ions ( on beliefs and ut il it y funct ions) which ensur e t he exist ence of mechanisms ( involving exact ly bal anced t r ansfer s) which ar e bot h incent i ve compat ible and ex post e¢ cient . Given t he T U aspect , it is not di¢ cul t t o see t hat t he expect ed payo¤s achieved by such mechanisms ar e in t he ex ant e incent ive compat ible cor e ( t his is illust r at ed on example 2 in t he pr evious sect ion) . I n ot her wor ds, one can easily pr oduce assumpt ions under which t he ex ant e incent ive compat ible core cont ains ex post e¢ cient mechanisms. Such r esult s ar e used in [11] ( wher e an analog of pr oposit ion 1 does not hol d) t o est ablish t he non-empt iness of t he ex ant e incent ive compat ible cor e.
e¢ ciency but not ex post individual r at ionalit y.16 Rot h [22] consider s a much st r onger r equir ement : ex post st abilit y in t he absence of t r ansfer s. H e anal yzes “ mar r iage pr obl ems” ( t ypically wit hout t r ansfer s) in which agent s pr ivat ely know t heir own ut ilit ies ( in our t er minology, “ pr ivat e val ues” ) . A s in t he pr esent paper , he consider s gener al mechanisms ( allowing in par t icular for lot ter i es over mat chings) and applies t he r evelat ion pr inciple in or der t o focus on dir ect r evelat ion mechanisms. H e calls such a mechanism “ st abl e” if it select s a st able mat ching for any st at ed ut ilit ies. To expr ess a similar st abilit y pr oper t y in our fr amewor k, let us de…ne, for ever y
t
2 T
,v
t as t he( T U ) mat ching game ( wit h complet e i nfor mat ion) when t he t ypes ar e
t
. A s shown in [26],v
t is t he super addit ive cover ofv
t(
fkg) = u
k(t; 0)
,k
2 K
,v
t(
fi; jg) = u
i(t; j) + u
j(t; i)
,i
2 I
,j
2 J
. A mat ching mechanism¹
is“ st able” in t he sense of Rot h [22] if for ever y
t
2 T
( int er pr et ed as a vect or of r epor t ed t ypes) ,¹(t)
select s a solut ion inC(v
t)
, i.e., if¹
is ex post st able.Rot h ar gues t hat t his r equir ement is not “ excessively st r ong” i n a model wi t h pr ivat e values. H owever , he shows t hat incent ive compat ible, ex post st able, mat ching mechanisms do not exist in gener al. A s suggest ed above, in our fr amewor k ( wit h t r ansfer s) , it is not di ¢ cult t o const r uct examples in which incent ive compat ible ex post e¢ cient mechanisms do exist , but none of t hem is ex post individually r at ional.
Rot h [22] also obser ves t hat sever al proper t ies involving st rategy-proofness in mar r iage pr oblems wit h complet e i nfor mat ion have an immediat e count er -par t in mar r iage pr oblems wit h incomplet e infor mat ion but pr ivat e values.17 I n par t icular , t here exist mat ching mechanisms such t hat r evealing one’s t r ue pr efer ences is a dominant st r at egy for one side of t he mar ket . Such r esult s are also valid her e, but do not seem t o be hel pful in analyzing t he incent i ve compat ible cor e unless one side of t he mar ket cont ains only one agent ( as in st andar d auct ions) .
T he pr evious comment s r aise t he quest ion of t he validit y of our r esult s in mat ching mar ket s wher e monet ar y t r ansfer s ar e not possible ( like in Shapley and Scar f [25]) . Our solut ion concept s can be used in t his fr amewor k. Except for t he T U aspect in sect ion 3, all our r esult s hold when all monet ary t r ansfer s are imposed t o be null. T he ( N T U ) ex ant e incent i ve compat ible cor e is t hen non-empt y as a consequence of Scar f [24]’s t heorem.
16
Ex ant e and, in some cases like [7] and [8], int er im individual rat ionalit y can be guarant eed.
17A st r at egy-proof mechanism induces a domi nant-strategy incent ive compat ibl e