• Aucun résultat trouvé

en fr

N/A
N/A
Protected

Academic year: 2021

Partager "en fr "

Copied!
8
0
0

Texte intégral

(1)

HAL Id: hal-01727357

https://hal.sorbonne-universite.fr/hal-01727357

Submitted on 9 Mar 2018

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.

Distributed under a Creative Commons Attribution| 4.0 International License

Remarks on the Monge–Kantorovich problem in the

discrete setting

Haim Brezis

To cite this version:

Haim Brezis. Remarks on the Monge–Kantorovich problem in the discrete setting. Comptes

Ren-dus Mathématique, Elsevier Masson, 2018, 356 (2), pp.207-213. �10.1016/j.crma.2017.12.008�.

�hal-01727357�

(2)

Contents lists available atScienceDirect

C. R.

Acad.

Sci.

Paris,

Ser. I

www.sciencedirect.com

Functional analysis/Mathematical analysis

Remarks

on

the

Monge–Kantorovich

problem

in

the

discrete

setting

Remarques

sur

le

problème

de

Monge–Kantorovich

dans

le

cas

discret

Haïm Brezis

a,b,c

aDepartmentofMathematics,HillCenter,BuschCampus,RutgersUniversity,110FrelinghuysenRoad,Piscataway,NJ08854,USA bDepartmentsofMathematicsandComputerScience,Technion,IsraelInstituteofTechnology,32000Haifa,Israel

cLaboratoireJacques-Louis-Lions,UniversitéPierre-et-Marie-Curie,4,placeJussieu,75252Pariscedex05,France

a

r

t

i

c

l

e

i

n

f

o

a

b

s

t

r

a

c

t

Articlehistory:

Received19December2017 Accepted19December2017 Availableonline5January2018 PresentedbyHaïmBrezis

In Optimal Transport theory, three quantities play a central role: the minimal cost of transport,originallyintroducedbyMonge,itsrelaxedversionintroducedbyKantorovich, and adualformulation alsoduetoKantorovich. The goalofthisNote isto publicizea veryelementary,self-containedargument extractedfrom[9],whichshowsthatallthree quantitiescoincideinthediscretecase.

©2017Académiedessciences.PublishedbyElsevierMassonSAS.Thisisanopenaccess articleundertheCCBY-NC-NDlicense (http://creativecommons.org/licenses/by-nc-nd/4.0/).

r

é

s

u

m

é

Enthéoriedutransportoptimal,troisquantités jouentunrôlecentral :lecoûtminimal de transport, introduit par Monge, sa version relaxée, introduitepar Kantorovich, et la formulationduale,dueaussiàKantorovich. L’objetdecettenoteest demettreen avant unedémonstrationtotalementélémentaire,extraitede[9],dufaitquecestroisquantités coïncidentdanslecasdiscret ;cettepreuvenerequiertaucuneconnaissancepréalable.

©2017Académiedessciences.PublishedbyElsevierMassonSAS.Thisisanopenaccess articleundertheCCBY-NC-NDlicense (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction

Considertwosets X

,

Y consistingofm points

(

Pi

)

and

(

Ni

),

1

i

m,i.e.

X

= {

Pi

,

P2

, . . . ,

Pm

}

and Y

= {

N1

,

N2

, . . . ,

Nm

} .

Letc

:

X

×

Y

→ R

beanyfunction(c standsfor“cost”).Weintroducethreequantities.Thefirstonedenoted M (forMonge) isdefinedby

E-mailaddress:brezis@math.rutgers.edu.

https://doi.org/10.1016/j.crma.2017.12.008

1631-073X/©2017Académiedessciences.PublishedbyElsevierMassonSAS.ThisisanopenaccessarticleundertheCCBY-NC-NDlicense (http://creativecommons.org/licenses/by-nc-nd/4.0/).

(3)

208 H. Brezis / C. R. Acad. Sci. Paris, Ser. I 356 (2018) 207–213 M

:=

min σ∈Sm m



i=1 c

(

Pi

,

Nσ(i)

),

(1)

wheretheminimumistakenovertheset

S

m ofallpermutationsoftheintegers

{

1

,

2

,

. . . ,

m

}

.Thesecondone,denoted K

(forKantorovich),isdefinedby

K

:=

min A

m



i,j=1 ai jc

(

Pi

,

Nj

)

;

A

= (

ai j

)

is doubly stochastic

.

(2)

Recallthatamatrix A

= (

ai j

)

isdoublystochasticif

ai j

0

i

,

j

,

m



i=1 ai j

=

1

j

,

and m



j=1 ai j

=

1

i

.

(3)

Finallydefine D (forduality)by

D

:=

sup ψ:Y→R ϕ:X→R

m



i=1

(

ϕ

(

Pi

)

− ψ(

Ni

))

;

ϕ

(

x

)

− ψ(

y

)

c

(

x

,

y

),

x

X

,

y

Y

.

(4) Theorem1.1.Wehave M

=

K

=

D

.

(5)

Moreoverthe“sup”in(4)isachieved.

Equality K

=

D inTheorem 1.1isattheheartofKantorovich’spioneeringdiscoveryconcerningtheMongeproblem(see

[19] and[20]).Equality M

=

K makes totallytransparentthe connectionbetweenKantorovich’sformulationandMonge’s original goal (see item (3)in Section 4below). The purposeofthis note isto advertise the MK(

=

Monge–Kantorovich) theory initsmostelementary(butinitselfstrikinganduseful!) setting,asitappears,e.g.,inBrezis–Coron–Lieb [11] (see Section 3anditem (1) inSection 4 below). This“primitive” caseilluminates thefoundations ofthe MK sagawhich has “exploded” inrecent years;see, e.g.,the remarkableworksof [2], [3],[7],[15],[16],[17], [23], [24],[29],[34],[35], etc. I reproduceinSection2anelementaryself-containedproofofTheorem 1.1(accessibletofirst-yearstudents),extractedfrom apresentationof[11]thatIgavein1985(see[9]).

2. ProofofTheorem 1.1

ChoosingforA in(2)apermutationmatrixyields

K

M

.

(6)

Ontheotherhand,assumethat

ϕ

and

ψ

areasin(4).LetA

= (

ai j

)

beadoublystochasticmatrix.Multiplyingthe

inequal-ities

ϕ

(

Pi

)

− ψ(

Nj

)

c

(

Pi

,

Nj

)

byai j andsummingoveri

,

j yields m



i=1

(

ϕ

(

Pi

)

− ψ(

Ni

))

m



i,j=1 ai jc

(

Pi

,

Nj

).

(7)

MinimizingoverA andmaximizingover

ϕ

,

ψ

gives

D

K

.

(8)

Inviewof(6)and(8),itsufficestoestablishthat

M

D

.

(9)

Proofof(9). Withoutlossofgeneralitywemayrelabelthepoints

(

Nj

)

sothat

M

=

m



i=1 c

(

Pi

,

Ni

)

m



j=1 c

(

Pj

,

Nσ(j)

)

σ

S

m

.

(10)

(4)

m



i=1

(

ϕ

(

Pi

)

− ψ(

Ni

))

M (11) and

ϕ

(

Pi

)

− ψ(

Nj

)

c

(

Pi

,

Nj

)

i

,

j

.

(12) Set di

=

c

(

Pi

,

Ni

)

i (13) and bi j

=

c

(

Pi

,

Nj

)

di

=

c

(

Pi

,

Nj

)

c

(

Pi

,

Ni

)

i

,

j

.

(14)

Considerthenumbers

λ

i

= ψ(

Ni

),

1

i

m,asbeingtheunknowns.Oncethe

λ

i’shavebeendeterminedset

ϕ

(

Pi

)

= ψ(

Ni

)

+

di

= λ

i

+

di

i

.

(15)

(Thischoiceisdictatedby(10),(11),and(12)appliedwithi

=

j.)From(15),(13)and(10)weseethat(11)holds.Wenow rewrite(12)as

λ

i

− λ

j

bi j

i

,

j

.

(16)

Notethatby(13)and(14)

bii

=

0 1

i

m

,

(17)

andthatthehypothesis(10)reads

m



j=1

bjσ(j)

0

σ

S

m

.

(18)

We complete the proof of (12) (and thus the existence of functions

ϕ

and

ψ

satisfying (11)–(12)) via the next lemma essentiallyduetoAfriat[1].

Lemma2.1.Assumethat

(

bi j

)

isageneralmatrixsatisfying(17)(18).Thenthesystemofinequalities(16)admitsasolution.

Proof. (Copiedfrom[9],inspiredby[1]).Wefirstproposeanansatzforthe

λ

i’sandthenprovethatthisansatzhasallthe

requiredproperties.Achain K connectingi to j isafinitesequenceK

= (

i1

,

. . . ,

ik

)

suchthatk

2

,

il

∈ {

1

,

. . . ,

m

} ∀

l,i1

=

i,

andik

=

j.(Wedonotassumethati1

,

. . . ,

ikaredistinct.)

GivenachainK connectingi

=

i1 to j

=

ik,set

SK

:=

bi1i2

+

bi2i3

+ · · · +

bik−1ik

.

(19)

Suppose nowthatasolution

i

)

to(16)existsandconsiderachain K connectingi to j.Wehave

λ

i1

− λ

i2

bi1i2

,

λ

i2

− λ

i3

bi2i3

,

. . .

λ

ik−1

− λ

ik

bik−1ik

.

Addingtheseinequalitiesyields

λ

i

− λ

j

SK

,

(20) andinparticular

λ

i

− λ

1

inf K

SK

;

K is a chain connecting i to 1

.

(21)

Wenowturntotheexistence ofasolution

i

)

to(16).Sincethe

λ

i’saredefinedmoduloanadditiveconstantitistempting,

(5)

210 H. Brezis / C. R. Acad. Sci. Paris, Ser. I 356 (2018) 207–213

λ

i

:=

inf K

SK

;

K is a chain connecting i to 1

.

(22)

(A priori,it mayhappen that

λ

i

= −∞

,butthiswillbe excluded below.) Fix1

i

,

j

m andlet K

= (

i1

,

. . . ,

ik

)

beany

chainconnecting j to1,then K

˜

:= (

i

,

K

)

connectsi to1 andtherefore(by(22))

λ

i

SK˜

=

bi j

+

SK

.

(23)

TakingtheinfoverK in(23)weobtain

λ

i

bi j

+ λ

j

1

i

,

j

m

.

(24)

This correspondsto thedesiredinequality (16)providedwe establishthat

λ

j

= −∞

j; assumptions (17)and(18)enter

here.Wewillprovethat

λ

1

=

0

.

(25)

Then,combining(24)and(25)wededucethat

0

= λ

1

b1 j

+ λ

j

j

andthus

λ

j

= −∞

j.Wenowturntotheproofof(25).First,wechoosethechain K

= (

1

,

1

)

in(22)andobtain

λ

1

b11

=

0

.

(26)

Nextweestablishthat

λ

1

0.Westartwithsometerminology.AchainK connectingi to j

=

i iscalledacycle (oraloop).

Acycleissimple ifi1

, . . . ,

ik−1aredistinct.Weclaimthat,foreverycycleK ,

SK

0

.

(27)

Indeedwhen K isasimplecycle(27)followsfrom(18)(and(17)) appliedtothepermutation i1

i2

· · · →

ik (theother

integers are invariant).By decomposinga generalcycleinto simplecycleswe findthat(27) holdsforallcycles.Applying

(27)toanychainconnecting1 to1,wededucefrom(22)that

λ

1

0.

2

Remark2.1. The aboveproof providesinfact anecessaryandsufficient conditionfortheexistence ofasolutionto (16).It readsasfollows.



jB

bjσ(j)

0

,

σ

S

k

,

(28)

foreveryinteger1

k

m andforeverysubset B of

{

1

,

. . . ,

m

}

containingk distinctelements,wherethepermutations

σ

actonlyon B.Thisresultappearsalreadyin[1]asaconsequenceofTheorems3.1and7.2in[1].Unfortunately,theproofs in[1]areobscuredbyaflurryofdefinitions!

3. Whenthecostc isadistance

WenowpresentasimpleconsequenceofTheorem 1.1whenthecostc isadistance,whichcorrespondstothesettingof

[20].Letd

(

x

,

y

)

beapseudometric(i.e.thedistancebetweentwodistinctpointscanbezero)onasetZ .Let

(

Pi

),

(

Ni

),

1

i

m bepointsinZ suchthat Pi

=

Nj

i

,

j (butitmayhappenthat Pi

=

PjorNi

=

Njforsome i

=

j).

Corollary3.1.Wehave DLip

:=

sup ζ

m



i=1

(ζ (

Pi

)

− ζ(

Ni

))

; ζ :

Z

→ R, |ζ(

x

)

− ζ(

y

)

| ≤

d

(

x

,

y

)

x

,

y

Z

=

min σSm m



i=1 d

(

Pi

,

Nσ(i)

)

=

M

.

Proof. Clearly DLip

M.Afterrelabelingthepoints

(

Nj

)

wemayassume,asin(10),that

M

=

m



i=1 d

(

Pi

,

Ni

)

m



j=1 d

(

Pj

,

Nσ(j)

)

σ

S

m

.

(29)

(6)

ApplyingTheorem 1.1with X

= {

P1

,

P2

,

. . . ,

Pm

},

Y

= {

N1

,

N2

,

. . . ,

Nm

}

andc

(

x

,

y

)

=

d

(

x

,

y

)

weknowthat M

=

D;thuswe

obtainfunctions

ϕ

:

X

→ R

and

ψ

:

Y

→ R

suchthat

ϕ

(

Pi

)

− ψ(

Ni

)

=

d

(

Pi

,

Ni

)

i

,

(30)

and

ϕ

(

Pi

)

− ψ(

Nj

)

d

(

Pi

,

Nj

)

i

,

j

.

(31)

Weclaimthat

|ψ(

Ni

)

− ψ(

Nj

)

| ≤

d

(

Ni

,

Nj

)

i

,

j

.

(32)

Indeed,by(30),(31),andthetriangleinequalitywehave

ψ (

Ni

)

=

ϕ

(

Pi

)

d

(

Pi

,

Ni

)

≤ ψ(

Nj

)

+

d

(

Pi

,

Nj

)

d

(

Pi

,

Ni

)

≤ ψ(

Nj

)

+

d

(

Ni

,

Nj

),

whichimplies(32).Set,forz

Z ,

ζ

0

(

z

)

=

inf j

ψ (

Nj

)

+

d

(

z

,

Nj

)



,

(33) sothat

0

(

x

)

− ζ

0

(

y

)

| ≤

d

(

x

,

y

)

x

,

y

Z

.

From(32)weseethat

ζ

0

(

Ni

)

= ψ(

Ni

)

i

.

(34)

Ontheotherhand,wehave,by(33)and(31),

ζ

0

(

Pi

)

ϕ

(

Pi

)

i,whiletaking j

=

i in(33),andusing(30),yields

ζ

0

(

Pi

)

d

(

Pi

,

Ni

)

+ ψ(

Ni

)

=

ϕ

(

Pi

)

i

.

Therefore,

ζ

0

(

Pi

)

=

ϕ

(

Pi

)

i

.

(35)

Choosing

ζ

= ζ

0 inthedefinitionofDLipandapplying(30),(34),and(35)yields DLip

M.

2

4. Finalcomments

(1) Corollary 3.1istakenfrom[11].Equality M

=

DLip plays animportantrole inprovingthat the“leastenergyrequired

toproduceprescribedsingularities”(inliquidcrystals)coincideswiththe“lengthofaminimalconnectionconnecting thesesingularities”(for subsequentdevelopmentssee,e.g., [5],[8], [12],[13] and[28]). Theproof ofCorollary 3.1 in

[11]takesafewlines,butitreliesheavilyonthreenontrivialtools.TheequalityK

=

DLipisderivedfromKantorovich’s

duality(seeitem(2)below).WhiletheequalityM

=

K reliesonBirkhoff’s theorem[6]ondoublystochastic matrices (also calledBirkhoff–von Neumann’s theorembecause von Neumann [36] rediscovered itindependently a few years later).Itassertsthattheextremepointsoftheconvexset A ofdoublystochasticmatricesarepreciselythepermutation matrices.Applying theKrein–Milman theorem,onededuces thatanymatrixin A is aconvexcombinationof permu-tationmatrices, andconsequently K

M. (ForrecentdevelopmentsrelatedtoBirkhoff’stheorem,Ireferthereaderto

[22]and[14].) Bycontrast,theaboveproofofCorollary 3.1iselementaryandself-contained.Noprerequisiteisneeded andmoreoverityieldsthetwoequalitiesM

=

K andK

=

D inasingleshot!

(2) Equality K

=

D in Theorem 1.1 is at the heart of Kantorovich’sdiscovery (dating back to the late 1930s – see the referencesin[33])andgoesfarbeyond thediscretesettingconsideredhere.NotethatK andD involvetheminimization (resp.maximization)of linearfunctionals onconvexsets. Themostcommonwayto show that K

=

D isvia duality, either inthe sense of linearprogramming orin the sense ofconjugate convex functions(applying for examplethe theoremofFenchel–Rockafellar;see,e.g.,TheoremI.12in[10]).Ireferthereaderto[2],[3],[15],[17],[23],[24],[29],

[34],[35],etc.

(3) According to A. M. Vershik (personal communication), equality M

=

K was known to L. V. Kantorovich, based on Birkhoff’stheorem(seeitem(1)above),whichwas publishedaroundthesametimeas[19]-[20].ApparentlyBirkhoff’s ideas were intheairsince a precursorofBirkhoff’s theoremappeared alreadyin1931(see the historicalnote onp. 25of[14]).Surprisingly,Birkhoff’stheoremishardlyevermentionedinthevastMKliterature.Thereasonforitbeing thattheMKcommunityhasbeenmostlypreoccupiedwiththeequalityM

=

K inthenon-atomic case;inthissetting, theMongeformulationwasnotevenpreciselystateduntilthe1970swhenitwasposedexplicitlyinmoderntermsby A.M. Vershik[32] (seealso[7]). Intheir rushtothecontinuum case, theMK aficionadospaid littleattentiontothe discretecase—whichisinitselfstrikinganduseful!!

(7)

212 H. Brezis / C. R. Acad. Sci. Paris, Ser. I 356 (2018) 207–213

(4) Asalreadymentioned, ourelementaryproof of Theorem 1.1doesnot require anyofthetools described initems(1) and(2)above.Instead,itreliesontheconstruction(22)(copiedfrom[9])involving“chains”and“cycles”.Thisdeviceis reminiscentofRockafellar’scelebratedtheorem[26] oncyclicallymonotone operators.Thesameconstruction appears subsequently, atthesuggestion ofRockafellar,in[25] inthe contextof MathematicalEconomics,andthen in[31] in the MK context. In [31], Smith and Knott introduced the terminology “c-cyclical monotonicity”, which has become very fashionable inthe MK community,see [2], [3], [4], [15], [17], [21], [23],[24],[27],[30],[29],[34],[35], etc. In the literature, one can find two distinct definitions. The original definition says that if X

,

Y are arbitrary sets and c

:

X

×

Y

→ R

isanyfunction,thenaset



X

×

Y isc-cyclicallymonotoneifforeveryintegern,andforanyfinite sequence

(

xi

,

yi

)

,1

i

n,ofpointsin



(notnecessarilydistinct),onehas

n



i=1

{

c

(

xi

,

yi+1

)

c

(

xi

,

yi

)

} ≥

0

,

(36)

where yn+1

:=

y1.Inanotherdefinition,(36)isreplacedby

n



i=1

c

(

xi

,

yσ(i)

)

c

(

xi

,

yi

)



0

σ

S

n

.

(37)

In fact,thetwodefinitions areequivalent. Clearly,(37)implies(36)(justchoose

σ

(

i

)

=

i

+

1 when1

i

n

1 and

σ

(

n

)

=

1).Forthereverseimplication,wereturntotheproof ofLemma 2.1withbi j

=

c

(

xi

,

yj

)

c

(

xi

,

yi

)

.Weclaimthat

foreverycycleK

= (

i1

,

i2

,

. . . ,

ik−1

,

i1

)

,onehasSK

0 (sothattheconclusionofLemma 2.1holds,andclearlyimplies (37)).Applying (36)to

(

xi1

,

yi1

),

. . . ,

(

xik−1

,

yik−1

)

(instead of

(

xi

,

yi

)

) yields SK

0.Ifwetake



= (

Pi

,

Ni

),

1

i

m,

asinthesettingofTheorem 1.1,assumption(37)seems(atleastformally)strongerthanassumption(18)inLemma 2.1

because (37)is assumedforall finite sequences

(

xi

,

yi

)

in



,andmoreover thesepoints are not necessarilydistinct

—buttheconclusionsarethe sameandthusthetwoassumptions areaposterioriequivalent!Finally,observethat if X

=

Y

=

H isaHilbertspaceandc

(

x

,

y

)

= |

x

y

|

2,then aset



H

×

H isc-cyclically monotoneifandonlyifitis

cyclicallymonotoneintheusualsense(coinedbyRockafellar).

(5) E.Ghys[18]andA.Vershik[33] tellthefascinatingstoriesoftheMongeandKantorovichdiscoveries.Ihighly recom-mendthesepapers.

Acknowledgements

I thankP. Mironescu for catchinga smalloversight in [9]andL. Ambrosio forproviding usefulreferences.I am very gratefultoA.M.VershikforsharingwithmeinformationsconcerningthehistoryandlaterdevelopmentsoftheMKtheory, andtoL.C.Evansforhiswarmencouragements.

References

[1]S.N.Afriat,ThesystemofinequalitiesarsXrXs,Math.Proc.Camb.Philos.Soc.59(1963)125–133.

[2]L.Ambrosio,Lecturenotesonoptimaltransportproblems,in:P.Colli,J.F.Rodrigues(Eds.),MathematicalAspectsofEvolvingInterfaces,in:Lect.Notes Math.,vol. 1812,Springer,2003,pp. 1–52.

[3]L.Ambrosio,N.Gigli,Auser’sguidetooptimaltransport,in:ModellingandOptimisationofFlowsonNetworks,in:Lect.NotesMath.,vol. 2062, Springer,2013,pp. 1–155.

[4]M.Beiglböck,Cyclicalmonotonicityandtheergodictheorem,Ergod.TheoryDyn.Syst.35(2015)710–713.

[5]F.Bethuel,H.Brezis,J.-M.Coron,Relaxedenergiesforharmonicmaps,in:H.Berestycki,J.-M.Coron,I.Ekeland(Eds.),VariationalProblems,Proceedings ofaConferenceHeldinParisin1988,Birkhäuser,1990,pp. 37–52.

[6]G.Birkhoff,Tresobservacionessobreelalgebralineal,Rev.Univ.Nac.TucumánSer.A,Mat.Fis.Teor.5(1946)147–150.

[7]V.I.Bogachev,A.V.Kolesnikov,TheMonge–Kantorovichproblem:achievements,connectionsandperspectives,Usp.Mat.Nauk67(2012)3–110,English translation:Russ.Math.Surv.67(2012)785–890.

[8]J.Bourgain,H.Brezis,P.Mironescu,H1/2mapswithvaluesintothecircle:minimalconnections,lifting,andtheGinzburg–Landauequation,Publ.Math. Inst.HautesÉtudesSci.99(2004)1–115.

[9] H.Brezis,LiquidcrystalsandenergyestimatesforS2-valuedmaps,in:J.Ericksen,D.Kinderlehrer(Eds.),TheoryandApplicationsofLiquidCrystals,

ProceedingsofaConferenceHeldinMinneapolis,MN,USA,in1985,Springer,1987,pp. 31–52.Seealsoarticle112L1inhttp://sites.math.rutgers.edu/ ~brezis/publications.html.

[10]H.Brezis,FunctionalAnalysis,SobolevSpacesandPDEs,Springer,2011.

[11]H.Brezis,J.-M.Coron,E.Lieb,Harmonicmapswithdefects,Commun.Math.Phys.107(1986)649–705.

[12] H.Brezis,P.Mironescu,SobolevMapswithValuesintotheCircle,Birkhäuser,inpreparation. [13] H. Brezis, P. Mironescu, I. Shafrir, Distances between classes in W1,1

(;S1), Calc. Var. Partial Differ. Equ. (2017), https://doi.org/10.1007/ s00526-017-1280-z,toappear.

[14]R.Burkard,M.Dell Amico,S.Martello,AssignmentProblems,revisededition,SIAM,2012.

[15] L.C. Evans,Partial differentialequationsandMonge–Kantorovichmasstransfer,in:S.T.Yau(Ed.),CurrentDevelopmentsinMathematics, Lectures deliveredatHarvardin1997,InternationalPress,Cambridge,MA,USA,1999,pp. 65–126.Seealso“SurveyofapplicationsofPDEmethodstoMonge– Kantorovichmasstransferproblems”https://math.berkeley.edu/evans/.

[16]L.C.Evans,W.Gangbo,DifferentialequationsmethodsfortheMonge–Kantorovichmasstransferproblem,Mem.Amer.Math.Soc.137(1999).

(8)

[18] É. Ghys, Gaspard Monge. Le mémoire sur les déblais et les remblais, CNRS, Images des mathématiques, 2012. See http://images.math.cnrs.fr/ Gaspard-Monge,1094.html?lang=fr.

[19]L.V.Kantorovitch,Onthetranslocationofmasses,Dokl.Akad.NaukSSSR37(1942)227–229,Englishtranslation:J.Math.Sci.133(2006)1381–1382.

[20]L.V.Kantorovitch,OnaproblemofMonge,Usp.Mat.Nauk3(1948)225–226,Englishtranslation:J.Math.Sci.133(2006)1383.

[21]M.Knott,C.Smith,Onageneralizationofcyclicmonotonicityanddistancesamongrandomvectors,LinearAlgebraAppl.199(1994)363–371.

[22]N.Linial,Z.Luria,Ontheverticesofthed-dimensionalBirkhoffpolytope,DiscreteComput.Geom.51(2014)161–170.

[23] R.J.McCann,N.Guillen,Fivelecturesonoptimaltransportation:geometry,regularityandapplications,in:G.Dafni,etal.(Eds.),Analysisand Geom-etryofMetricMeasureSpaces,LectureNotesofthe“Séminairedemathématiquessupérieures”(SMS),2011,AmericanMathematicalSociety,2013, pp. 145–180.Seealso[46]inhttp://www.math.toronto.edu/mccann/publications.

[24]S.Rachev,L.Rüschendorf,MassTransportationProblems,Springer,1998.

[25]J.-C.Rochet,Anecessaryandsufficientconditionforrationalizabilityinaquasi-linearcontext,J.Math.Econ.16(1987)191–200.

[26]R.T.Rockafellar,Characterizationofthesubdifferentialsofconvexfunctions,Pac.J.Math.17(1966)497–510.

[27]L.Rüschendorf,Onc-optimalrandomvariables,Stat.Probab.Lett.27(1996)267–270.

[28]E.Sandier,Ginzburg–LandauminimizersfromRn+1toRnandminimalconnections,IndianaUniv.Math.J.50(2001)1807–1844.

[29]F.Santambrogio,OptimalTransportforAppliedMathematicians,Birkhäuser,2015.

[30]C.S.Smith,M.Knott,Noteontheoptimaltransportationofdistributions,J.Optim.TheoryAppl.52(1987)323–329.

[31]C.S.Smith,M.Knott,OnHoeffding–Fréchetboundsandcyclicmonotonerelations,J.Multivar.Anal.40(1992)328–334.

[32]A.M.Vershik,Someremarksontheinfinite-dimensionalproblemsoflinearprogramming,Usp.Mat.Nauk25(1970)117–124,Englishtranslation: Russ.Math.Surv.25(1970)117–124.

[33]A.M.Vershik,LonghistoryoftheMonge–Kantorovichtransportationproblem,Math.Intell.35(2013)1–9.

[34]C.Villani,TopicsinOptimalTransportation,AmericanMathematicalSociety,2003.

[35]C.Villani,OptimalTransport.OldandNew,Springer,2009.

[36]J.vonNeumann,Acertainzero-sumtwo-persongameequivalenttotheoptimalassignmentproblem,in:Contrib.TheoryGames,vol. 2,Princeton UniversityPress,Princeton,NJ,USA,1953,pp. 5–12.

Références

Documents relatifs

The Monge-Kantorovich problem is revisited by means of a variant of the saddle-point method without appealing to cyclical c -monotonicity and c -conjugates.. One gives new proofs of

Dans ce chapitre, nous avons simulé les performances de censure et de détection des détecteurs F/B-ACOSD et F/B-ACCAD, dans des milieux lognormal et Weibull homogènes et en présence

We present a general method, based on conjugate duality, for solving a con- vex minimization problem without assuming unnecessary topological restrictions on the constraint set..

As regards the most basic issue, namely (DG) pertaining to the question whether duality makes sense at all, this is analyzed in detail — building on a lot of previous literature —

The strength of the distortions induced on the signal is characterized using the so-called two-tone test in which the carrier is modulated by two nearby frequencies from which

The Kolmogorov Markov measures, as refer- ence measures in the Schrödinger problem, admit as a limit case the Monge Kantorovich problem with quadratic cost function, namely

Bien que le risque de contamination par la listériose ne semble pas être un sujet d'inquiétude pour les femmes enceintes, 70% d'entre elles ont mis en place la prévention contre

A new abstract characterization of the optimal plans is obtained in the case where the cost function takes infinite values.. It leads us to new explicit sufficient and