• Aucun résultat trouvé

Do Minkowski averages get progressively more convex?

N/A
N/A
Protected

Academic year: 2021

Partager "Do Minkowski averages get progressively more convex?"

Copied!
6
0
0

Texte intégral

(1)

HAL Id: hal-01590072

https://hal-upec-upem.archives-ouvertes.fr/hal-01590072

Submitted on 19 Sep 2017

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.

Do Minkowski averages get progressively more convex?

Matthieu Fradelizi, Mokshay Madiman, Arnaud Marsiglietti, Artem Zvavitch

To cite this version:

Matthieu Fradelizi, Mokshay Madiman, Arnaud Marsiglietti, Artem Zvavitch. Do Minkowski averages

get progressively more convex?. Comptes Rendus. Mathématique, Académie des sciences (Paris),

2016, 354, pp.185 - 189. �10.1016/j.crma.2015.12.005�. �hal-01590072�

(2)

Contents lists available atScienceDirect

C. R. Acad. Sci. Paris, Ser. I

www.sciencedirect.com

Functional analysis/Geometry

Do Minkowski averages get progressively more convex?

Les moyennes de Minkowski deviennent-elles progressivement plus convexes ?

Matthieu Fradelizia,1,Mokshay Madimanb,2,Arnaud Marsigliettic,3, Artem Zvavitchd,4

aLaboratoired’analyseetdemathématiquesappliquées,UMR8050,UniversitéParis-EstMarne-la-Vallée,5,bdDescartes, Champs-sur-Marne,77454Marne-la-Valléecedex2,France

bUniversityofDelaware,DepartmentofMathematicalSciences,501EwingHall,Newark,DE19716,USA

cInstituteforMathematicsanditsApplications,UniversityofMinnesota,207ChurchStreetSE,434LindHall,Minneapolis,MN 55455,USA dDepartmentofMathematicalSciences,KentStateUniversity,Kent,OH44242,USA

a r t i c l e i n f o a b s t r a c t

Articlehistory:

Received6November2015

Acceptedafterrevision11December2015 Availableonline27January2016 PresentedbyGillesPisier

Letusdefine,foracompactset ARn,theMinkowskiaveragesofA:

A(k)=

a1+ · · · +ak

k :a1, . . . ,akA

=1 k

A+ · · · +A

ktimes

.

We study the monotonicity of the convergence of A(k) towards the convex hull of A, when considering the Hausdorff distance, the volume deficit and a non-convexity index of Schneider as measures ofconvergence. For the volume deficit, we show that monotonicityfailsingeneral,thusdisprovingaconjectureofBobkov,MadimanandWang.

ForSchneider’snon-convexityindex,weprovethatastrongformofmonotonicityholds, andfortheHausdorffdistance,weestablishthatthesequenceiseventuallynonincreasing.

©2015Académiedessciences.PublishedbyElsevierMassonSAS.Thisisanopenaccess articleundertheCCBY-NC-NDlicense (http://creativecommons.org/licenses/by-nc-nd/4.0/).

r é s um é

Pourtoutensemblecompact ARn,définissonssesmoyennesdeMinkowskipar A(k)=

a1+ · · · +ak

k :a1, . . . ,akA

=1 k

A+ · · · +A

kfois

.

Nous étudions lamonotonie dela convergencede A(k) versl’enveloppeconvexe de A, mesuréeparladistancedeHausdorff,ledéficitvolumiqueetparl’indicedenon-convexité

E-mailaddresses:matthieu.fradelizi@u-pem.fr(M. Fradelizi),madiman@udel.edu(M. Madiman),arnaud.marsiglietti@ima.umn.edu(A. Marsiglietti), zvavitch@math.kent.edu(A. Zvavitch).

1 SupportedinpartbytheAgenceNationaledelaRecherche,projectGeMeCoD(ANR2011BS0100701).

2 SupportedinpartbytheU.S.NationalScienceFoundationthroughthegrantsDMS-1409504(CAREER)andCCF-1346564.

3 SupportedinpartbytheInstituteforMathematicsanditsApplicationswithfundsprovidedbytheNationalScienceFoundation.

4 SupportedinpartbytheU.S.NationalScienceFoundationGrantDMS-1101636andbytheSimonsFoundation.

http://dx.doi.org/10.1016/j.crma.2015.12.005

1631-073X/©2015Académiedessciences.PublishedbyElsevierMassonSAS.ThisisanopenaccessarticleundertheCCBY-NC-NDlicense (http://creativecommons.org/licenses/by-nc-nd/4.0/).

(3)

186 M. Fradelizi et al. / C. R. Acad. Sci. Paris, Ser. I 354 (2016) 185–189

deSchneider.Pourledéficitvolumique,nousdémontronsquelapropriétédemonotonie n’estpassatisfaiteengénéral,réfutantainsiuneconjecturedeBobkov,MadimanetWang.

Pourl’index de non-convexité de Schneider,nous montrons unepropriété renforcée de monotonie, tandis que,pour ladistance de Hausdorff, nousétablissons quela suiteest décroissanteàpartird’uncertainrang.

©2015Académiedessciences.PublishedbyElsevierMassonSAS.Thisisanopenaccess articleundertheCCBY-NC-NDlicense (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Versionfrançaiseabrégée

L’objectifdecettenoteestd’annonceretdedémontrerunepartiedesrésultatsobtenusdans[3]quiportentsurl’étude delamonotoniedelasuite(A(k))k1 définieen(1),mesuréeàtraversdifférentesmesuresdenon-convexité.Intuitivement, les ensembles A(k)deviennentde plusenplus convexesaufuretà mesurequek croît.Cette intuitionestpréciséedans [7,2]ilestdémontréquelasuite(A(k))convergeverssonenveloppeconvexeendistancedeHausdorffdH.

L’origine de notre étude provient d’une conjecture de Bobkov, Madiman et Wang [1], qui affirme que la suite ((A(k)))k1 estdécroissante,

(A):=Voln(conv(A)\A)=Voln(conv(A))Voln(A)

désigneledéficitvolumiqued’unensemblecompactdeRn.Ici,Voln représentelamesuredeLebesguedansRn etconv(A) désigne l’enveloppe convexede A. Nousréfutons cetteconjecture enexhibant un contre-exempleexplicite endimension supérieureouégaleà12.Lecontre-exempleestlaréuniondedeuxensemblesconvexesinclusdansdessous-espacesdedi- mension(presque)moitiédel’espaceambiant(voirFig. 1).Nousdémontronsaussilavaliditédelaconjectureendimension 1enadaptantunedémonstrationde[4]surlecardinaldesommesd’entiers ;celaaaussiétéobservéindépendammentpar F. Barthe.Laconjectureresteouverteendimensionn,pour1<n<12.

Demanière analogueàlaconjecturede Bobkov–Madiman–Wang,nousétudionslamonotoniede lasuite(c(A(k)))k1, c estl’indexdenon-convexitédeSchneider[6]définipar

c(A):=inf{λ≥0:A+λconv(A)est convexe}.

Contrairementaudéficitvolumique,lasuite(c(A(k)))eststrictementdécroissante,àmoinsque A(k)soitdéjàconvexe.Plus précisémentnousmontronsquepourtoutensemblecompact A deRnettoutk∈N

c(A(k+1))k

k+1c(A(k)) .

En outre, nous étudions dans[3] la monotoniede A(k), mesurée par d’autres mesures de non-convexité. Ainsi, nous montronsquesil’onpose

d(A)=dH(A,conv(A))=inf{r>0:conv(A)A+r Bn2},

Bn2estlabouleeuclidiennecentréeen0 derayon1,alorspourtoutcompact A deRn etpourkc(A),

d(A(k+1))k

k+1d(A(k)) .

1. Introduction

Thisnote announcesandprovessome oftheresultsobtainedin[3].Letusdenoteforacompactset ARn andfora positiveintegerk,

A(k)=

a1+ · · · +ak

k :a1, . . . ,akA

=1 k

A+ · · · +A

ktimes

. (1)

Denotingbyconv(A)theconvexhullofA,andby d(A):=inf{r>0:conv(A)A+r Bn2}

the Hausdorff distancebetweenaset A anditsconvexhull, it isa classicalfact (provedindependentlyby [7,2] in1969, and oftencalledthe Shapley–Folkmann–Starrtheorem) that A(k) convergesin Hausdorffdistance to conv(A) ask→ ∞. Furthermore[7,2] alsodetermined therateofconvergence:itturnsout thatd(A(k))=O(1/k) foranycompactset A.For setsofnonemptyinterior,thisconvergenceofMinkowskiaveragestotheconvexhullcanalsobeexpressedintermsofthe volumedeficit(A)ofacompactset A inRn,whichisdefinedas:

(4)

(A):=Voln(conv(A)\A)=Voln(conv(A))Voln(A),

whereVoln denotesLebesgue measureinRn.It was shownby[2]that if A is compactwithnonemptyinterior,thenthe volumedeficitofA(k)alsoconvergesto0;moreprecisely,(A(k))=O(1/k)foranycompactsetAwithnonemptyinterior.

OuroriginalmotivationcamefromaconjecturemadebyBobkov,MadimanandWang[1]:

Conjecture1.(See[1].) LetA beacompactsetinRnforsomen∈N,andletA(k)bedefinedasin(1).Thenthesequence(A(k))is non-increasingink,orequivalently,{Voln(A(k))}k1isnon-decreasing.

WeshowthatConjecture 1failstoholdingeneral,evenformoderatelyhighdimension.

Theorem2.Conjecture 1isfalseinRnforn12,andtrueforR1.

Noticethat Conjecture 1remains openfor1<n<12.Inparticular,theargumentspresentedinthisnotedonotseem towork.InanalogywithConjecture 1,wealsoconsiderwhetheronecanhavemonotonicityof{c(A(k))}k1,wherec isa non-convexityindexdefinedbySchneider[6]asfollows:

c(A):=inf{λ0:A+λconv(A)is convex}.

AnicepropertyofSchneider’sindexisthatitisaffine-invariant,i.e.,c(T A+x)=c(A)foranynonsingularlinearmap T on RnandanyxRn.

Contrarytothevolumedeficit,weprovethatSchneider’snon-convexityindexc satisfiesa strongkindofmonotonicity inanydimension.

Theorem3.LetA beacompactsetinRnandk∈N.Then c(A(k+1))k

k+1c(A(k)) .

Finally,we alsoprove thateventually,forkc(A), theHausdorff distancebetween A(k) andconv(A) isalsostrongly decreasing.

Theorem4.LetA beacompactsetinRnandkc(A)beaninteger.Then d(A(k+1))k

k+1d(A(k)) .

Moreover,Schneiderproved in[6]that c(A)nforevery compactsubset A ofRn.Itfollowsthat theeventual mono- tonicityofthesequenced(A(k))holdstrueforkn.

Itisnaturaltoaskwhattherelationshipisingeneralbetweenconvergenceofc,anddto0,forarbitrarysequences (Ck) ofcompact sets. In fact, none of thesethree notions of approach to convexity are comparable witheach other in general. Toseewhy,observethat whilec isscaling-invariant, neithernord are;so itiseasy toconstructexamples of sequences(Ck)suchthatc(Ck)0 but (Ck)andd(Ck)remainboundedawayfrom0.Thesameargumentenablesusto constructexamples ofsequences (Ck)such that c(Ck) remain boundedaway from0,whereas (Ck) andd(Ck) converge to0.Furthermore,(Ck)remainsboundedawayfrom0foranysequenceCk offinitesets,whereas c(Ck)andd(Ck)could convergeto0ifthefinitesets formafinerandfinergridfillingouta convexset.Anexamplewhere(Ck)0 butboth c(Ck) andd(Ck)are boundedaway from0isgivenby takinga 3-pointset with2ofthepoints getting arbitrarilycloser butstayingawayfromthethird.Onecanobtainfurtherrelationships betweenthesemeasuresofnon-convexity iffurther conditionsareimposedonthesequenceCk;detailsmaybefoundin[3].

Therestofthisnote isdevotedtotheexaminationofwhether A(k)becomesprogressivelymoreconvexaskincreases, whenmeasuredthroughthefunctionals,dandc.Theconcludingsectioncontainssomeadditionaldiscussion.

2. Thebehaviorofvolumedeficit

WeproveTheorem 2inthissection.WestartbyconstructingacounterexampletotheconjectureinRn,forn12.LetF bea p-dimensionalsubspaceofRn,wherep∈ {1,. . . ,n1}.LetusconsiderA=I1I2,whereI1F andI2F,where FdenotestheorthogonalcomplementofF.Onehas(seeFig. 1):

A+A=2I1(I1×I2)2I2,

A+A+A=3I1(2I1×I2)(I1×2I2)3I2.

Noticethat

(5)

188 M. Fradelizi et al. / C. R. Acad. Sci. Paris, Ser. I 354 (2016) 185–189

Fig. 1.A counterexample inR12.

Voln(A+A)=Volp(I1)Volnp(I2),

Voln(A+A+A)=Volp(I1)Volnp(I2)(2p+2np1).

Thus,Voln(A(3))Voln(A(2))ifandonlyif 2p+2np1

3 2

n

. (2)

Noticethatinequality(2)doesnotholdwhenn12 and p= n2.

ForR1,theconjecture maybe provedbyadapting aproof of[4]oncardinalityofintegersumsets;thiswas alsoinde- pendentlyobservedbyF. Barthe.Letk1.SetS=A1+ · · · +Akandfori∈ [k],letai=minAi,bi=maxAi,

Si=

j∈[k]\{i} Aj,

si=

j<iaj+

j>ibj,Si = {xSi;xsi}andS+i = {xSi;x>si}.Foralli∈ [k1],onehas S(ai+Si )(bi+1+S+i+1).

Sinceai+si=

jiaj+

j>ibj=bi+1+si+1,theaboveunionisadisjointunion.Thusfori∈ [k1] Vol1(S)≥Vol1(ai+Si )+Vol1(bi+1+S+i+1)=Vol1(Si )+Vol1(S+i+1).

Noticethat S1 =S1 andSk+=Sk\ {sk},thusaddingtheabovek1 inequalities,weobtain

(k1)Vol1(S)

k1

i=1

Vol1(Si )+Vol1(S+i+1)

=Vol1(S1)+Vol1(Sk+)+

k1

i=2

Vol1(Si)

= k

i=1

Vol1(Si).

Now takingallthesets Ai=A,anddividingthroughbyk(k1),weseethatwehaveestablishedConjecture 1indimen- sion 1.

3. ThebehaviorofSchneider’snon-convexityindexandtheHausdorffdistance

WeestablishTheorems 3 and4inthissection.Thisreliescruciallyontheelementaryobservationsthatconv(A+B)= conv(A)+conv(B)and(t+s)conv(A)=tconv(A)+sconv(A)foranyt,s>0 andanycompactsets A,B.

ProofofTheorem 3. Denote λ=c(A(k)). Since conv(A(k))=conv(A), from the definition of c, one knows that A(k)+ λconv(A)=conv(A)+λconv(A)=(1+λ)conv(A).Usingthat A(k+1)=k+A1+k+k1A(k),onehas

A(k+1)+ k

k+1λconv(A)= A k+1+ k

k+1A(k)+ k

k+1λconv(A)

= A k+1+ k

k+1conv(A)+ k

k+1λconv(A)

(6)

conv(A) k+1 + k

k+1A(k)+ k

k+1λconv(A)

=conv(A) k+1 + k

k+1(1+λ)conv(A)

=

1+ k k+1λ

conv(A).

Sincetheotherinclusionistrivial,wededucethat A(k+1)+k+k1λconv(A)isconvex,whichprovesthat c(A(k+1))k

k+1λ= k

k+1c(A(k)) . 2

ProofofTheorem 4. Letkc(A),then,fromthedefinitionsofc(A)andd(A(k)),onehas conv(A)= A

k+1+ k

k+1conv(A)A k+1+ k

k+1

A(k)+d(A(k))Bn2

=A(k+1)+ k

k+1d(A(k))Bn2.

Weconcludethat d(A(k+1))k

k+1d(A(k)) . 2

4. Discussion

(i) ByrepeatedapplicationofTheorem 3,itisclearthatthe convergenceofc(A(k))isatarate O(1/k)foranycompact set ARn;thisobservationappearstobenew.In[3],westudythequestionofthemonotonicityof A(k),aswell as convergencerates,whenconsidering severaldifferentwaystomeasure non-convexity,includingsome notmentioned inthisnote.

(ii) SomeoftheresultsinthisnoteareofinterestwhenoneisconsideringMinkowskisumsofdifferentcompactsets,not justsumsof Awithcopiesofitself.Indeed,theoriginalconjectureof[1]wasofthisform,andwouldhaveprovideda strengtheningoftheclassicalBrunn–Minkowskiinequalityformorethan2sets;ofcourse,thatconjectureisfalsesince theweakerConjecture 1is false.Nonethelesswe dohavesome relatedobservationsin[3];forinstance,itturns out thatingeneraldimension,forcompactsets A1,. . . ,Ak,

Voln k

i=1 Ai

1 k1

k i=1

Voln

j∈[k]\{i} Aj

.

Forconvexsets Bi,anevenstrongerfactistrue(thatthisisstrongermaynotbe immediatelyobvious),butiffollows fromwell-knownresults,see,e.g.,[5]:

Voln(B1+B2+B3)+Voln(B1)≥Voln(B1+B2)+Voln(B1+B3).

(iii) ThereisavariantofthestrongmonotonicityofSchneider’sindexwhendealingwithdifferentsets.IfA,B,C aresubsets ofRn,thenitisshownin[3](byasimilarargumenttothatusedforTheorem 3)thatc(A+B+C)max{c(A+B),c(B+ C)}.

Acknowledgements

Weare indebtedtoFedorNazarovforvaluablediscussions,inparticularforhelpintheconstructionofthecounterex- ample inTheorem 2. Wealso thankFranckBarthe, DarioCordero-Erausquin, UriGrupel, JosephLehec, Bo’azKlartag, and Paul-MarieSamsonforinterestingdiscussions.

References

[1]S.G.Bobkov,M.Madiman,L.Wang,FractionalgeneralizationsofYoungandBrunn–Minkowskiinequalities,in:C.Houdré,M.Ledoux,E.Milman,M.

Milman(Eds.),Concentration,FunctionalInequalitiesandIsoperimetry,in:Contemp.Math.,vol. 545,Amer.Math.Soc.,2011,pp. 35–53.

[2]W.R.Emerson,F.P.Greenleaf,AsymptoticbehaviorofproductsCp=C+ · · · +Cinlocallycompactabeliangroups,Trans.Amer.Math.Soc.145(1969) 171–204.

[3] M.Fradelizi,M.Madiman,A.Marsiglietti,A.Zvavitch,OnthemonotonicityofMinkowskisumstowardsconvexity,Preprint.

[4]K.Gyarmati,M.Matolcsi,I.Z.Ruzsa,Asuperadditivityandsubmultiplicativitypropertyforcardinalitiesofsumsets,Combinatorica30 (2)(2010)163–174.

[5]M.Madiman, P.Tetali,Informationinequalitiesforjointdistributions, withinterpretationsand applications,IEEETrans.Inf.Theory56 (6)(2010) 2699–2713.

[6]R.Schneider,Ameasureofconvexityforcompactsets,Pac.J.Math.58 (2)(1975)617–625.

[7]R.M.Starr,Quasi-equilibriainmarketswithnon-convexpreferences,Econometrica37 (1)(1969)25–38.

Références

Documents relatifs

Also use the formulas for det and Trace from the

We prove that arbitrary actions of compact groups on finite dimensional C ∗ -algebras are equivariantly semiprojec- tive, that quasifree actions of compact groups on the Cuntz

The aim of this note is to investigate the relations between these cohomologies and validity of the generalized ∂ ∂-Lemma. Generalized metric and Hodge theory.. [11]), the

Delano¨e [4] proved that the second boundary value problem for the Monge-Amp`ere equation has a unique smooth solution, provided that both domains are uniformly convex.. This result

First introduced by Faddeev and Kashaev [7, 9], the quantum dilogarithm G b (x) and its variants S b (x) and g b (x) play a crucial role in the study of positive representations

For example, he proved independently that two completely regular spaces X and Y have homeomorphic realcompactifications νX and νY whenever there is a biseparating linear maps from

McCoy [17, Section 6] also showed that with sufficiently strong curvature ratio bounds on the initial hypersur- face, the convergence result can be established for α &gt; 1 for a

For general n ≥ 2, G n was first studied by Chai, Wang and the second author [4] from the viewpoint of algebraic geometry. Among other things, they established the