• Non ci sono risultati.

Sobolev inequalities in arbitrary domains

N/A
N/A
Protected

Academic year: 2021

Condividi "Sobolev inequalities in arbitrary domains"

Copied!
53
0
0

Testo completo

(1)

Contents lists available atScienceDirect

Advances

in

Mathematics

www.elsevier.com/locate/aim

Sobolev

inequalities

in

arbitrary

domains

Andrea Cianchia,∗, Vladimir Maz’yab,c

a

DipartimentodiMatematicaeInformatica“U.Dini”,UniversitàdiFirenze, VialeMorgagni67/A,50134Firenze,Italy

bDepartmentofMathematics,LinköpingUniversity,SE-58183Linköping,Sweden cDepartmentofMathematicalSciences,M&OBuilding,UniversityofLiverpool,

LiverpoolL693BX,UK a r t i c l e i n f o a bs t r a c t Article history: Received15September2015 Accepted1February2016 Availableonlinexxxx

CommunicatedbyErwinLutwak MSC: 46E35 46E30 Keywords: Sobolevinequalities Irregulardomains Boundarytraces Optimalnorms Representationformulas

AtheoryofSobolevinequalitiesinarbitraryopensetsinRnis established.Boundaryregularityofdomainsisreplacedwith informationonboundarytracesoftrialfunctionsandoftheir derivatives upto someexplicitminimal order.Therelevant Sobolev inequalities exhibit the same critical exponents as intheclassicalframework.Moreover,they involveconstants independentofthegeometryofthedomain,andhenceyield genuinely new results even in the case when just smooth domains are considered. Our approach relies upon new representationformulasforSobolevfunctions,andonensuing pointwiseestimateswhichholdinanyopenset.

© 2016ElsevierInc.All rights reserved.

This research was partly supported by the Research Project of Italian Ministry of University and

Research (MIUR) Prin 2012 (grant number 2012TC7588) “Elliptic and parabolic partial differential equations:geometricaspects,relatedinequalities,andapplications”,andbyGNAMPAoftheItalianINdAM (NationalInstituteofHighMathematics).

* Correspondingauthor.

E-mailaddress:cianchi@unifi.it(A. Cianchi). http://dx.doi.org/10.1016/j.aim.2016.02.012 0001-8708/© 2016ElsevierInc.All rights reserved.

(2)

1. Introduction

The aim of this paper is to develop a theory of Sobolev embeddings, of any order

m∈ N,inarbitrary opensets Ω inRn,withn≥ 2.Asusual, byanm-th orderSobolev embeddingwemeananestimateforanormoftheh-thorderweakderivatives(0≤ h≤

m− 1) ofanym timesweaklydifferentiablefunctioninΩ in termsofnormsofsomeof itsderivativesuptotheorder m.

The classical theory of Sobolev embeddings involves ground domains Ω satisfying suitableregularity assumptions. For instance,a formulationof theoriginal theorem by Sobolevreadsasfollows. AssumethatΩ isaboundeddomainsatisfyingthecone prop-erty,m∈ N,1< p< mn, andF(·) is any continuousseminorm inWm,p(Ω) whichdoes notvanishonany(non-zero)polynomialofdegreenotexceedingm−1.Thenthereexists aconstantC = C(Ω) suchthat

u L np n−mp(Ω)≤ C  ∇mu Lp(Ω)+F(u)  (1.1) foreveryu∈ Wm,p(Ω).Here,Wm,p(Ω) denotes theusual Sobolevspace ofthose func-tionsinΩ whoseweakderivativesupto theorder m belongtoLp(Ω),andmu stands forthevectorofall(weak)derivativesof u oforder m.

ItiswellknownthatstandardSobolevembeddingsarespoiledinpresenceofdomains with“bad”boundaries.Inparticular,inequalitiesoftheform(1.1)donothold,atleast withthesamecriticalexponent n−mpnp ,inirregulardomains. Theinterplaybetweenthe geometryofthedomainandSobolevinequalities,eveninframeworksmoregeneralthan theEuclideanone,hasovertheyearsbeenthesubjectofextensiveinvestigations,along diversedirections,byanumberofauthors.Theirresultsaretheobjectofarichliterature, which includes the papers [3,5,6,16,7,8,12,13,18,19,25–27,29,28,31,32,37,39,43,49,47,48, 50–52,54–56,60–62,64,65,68,69] and the monographs [21–24,41,62,66]. Various classical problemsinthetheoryof spacesofdifferentiable functions,involvingpossiblyirregular sets, have attracted newinterest inthe last two decades, and have been considered in suchcontributionsas[10,17,20,33,34,45,40,46,63,67,70].

In order to avoid any assumption on Ω, we deal with Sobolev inequalities from an unconventionalperspective.Theunderlingideaofourresultsisthatsuitableinformation onboundarytraces oftrialfunctionscanreplaceboundaryregularityof Ω.

Theinequalitiesthatwillbe establishedhavetheform

∇hu Y (Ω,μ)≤ C  ∇mu X(Ω)+N∂Ω(u)  , (1.2)

wherem∈ N,h∈ N0,·X(Ω)isaBanachfunctionnormonΩ withrespecttoLebesgue measureLn,·

Y (Ω,μ)isaBanachfunctionnormwithrespecttoapossiblymoregeneral measure μ,andN∂Ω(·) isa(non-standard)seminormon∂Ω,dependingonthetraceof u, andofitsderivativesuptotheorderm−12 .Here,N0=N∪ {0},and[·] denotesinteger part.Moreover,0u standsjustforu,andweshalldenote1u alsoby∇u.

(3)

Somedistinctivetraitsoftheinequalitiestobe presentedcanbe itemizedasfollows:

• No regularity on Ω is a priori assumed. In particular, the constants in (1.2) are independentofthegeometryofΩ.

• ThecriticalSobolevexponents,or,moregenerally,theoptimaltargetnorms,arethe sameasinthecaseofregulardomains.

• The order m2−1 of the derivatives, on which the seminorm N∂Ω(·) depends, is minimalforaninequalityoftheform(1.2)toholdwithoutanyadditionalassumption onΩ.

Let us emphasize that, in view of these features, our inequalities provide original conclusionsevenwhenjustappliedtosmoothtrialfunctionsu onregulardomains Ω.

A first-order Sobolev inequality on arbitrary domains Ω ⊂ Rn, of the form (1.2), where X(Ω)= Lp(Ω),Y (Ω,μ)= Lq(Ω),andN∂Ω(·)=· Lr(∂Ω),with1≤ p< n,r≥ 1

and q = min{ rn

n−1,n−pnp } was established in [60] via isoperimetricinequalities. Sobolev inequalities of this kind, but still involving only first-order derivatives and Lebesgue measure, haverecentlyreceived renewedattention. Inparticular, the paper[57] makes useofmasstransportationtechniquestoaddresstheproblemoftheoptimalconstantsfor

p∈ (1,n),theproblem whenp= 1 having alreadybeen solvedin[60].Sharp constants ininequalitiesintheborderlinecasewhenp= n areexhibitedin[58].

In the present paper, we develop a completely different approach, which not only enables us to establish arbitrary-order inequalities, which cannot just be derived via iterationoffirst-orderones,butalsoaugmentsthefirst-ordertheory,inthatmoregeneral measures andnormsareallowed.

Ourpointofdepartureisanewpointwiseestimateforfunctions,andtheirderivatives, onarbitrary –possiblyunboundedandwithinfinitemeasure –domains Ω.Suchestimate involves a novel class of double-integral operators, where integration is extended over Ω× Sn−1. Therelevant operators act on akind of higher-order difference quotientsof thetraces offunctionsandoftheirderivativeson∂Ω.

In view of applications to norm inequalities, the next step calls for an analysis of boundednesspropertiesoftheseoperatorsinfunctionspaces.Tothispurpose,weprove their boundednessbetweenoptimalendpointspaces.Incombination withinterpolation argumentsbasedontheuseofPeetreK-functional,theseendpointresultsleadto point-wise bounds, forSobolevfunctions, inrearrangementform. Asaconsequence, Sobolev inequalities onan arbitraryn-dimensional domain arereduced to considerably simpler one-dimensional inequalitiesforHardytypeoperators.

With this apparatus at disposal, we are able to establish inequalities involving Lebesguenorms,withrespecttoquitegeneralmeasures,aswellasYudovich–Pohozaev– Trudinger type inequalities in exponential Orlicz spaces for limiting situations. The compactness of corresponding Rellich–Kondrashov type embeddings, with subcritical exponents,is alsoshown.Inequalities forotherrearrangement-invariantnorms, suchas

(4)

Lorentz and Orlicz norms, could be derived. However, in order to avoid unnecessary additionaltechnicalcomplications,thisissueisnotaddressedhere.

Onemajormotivationforourresearchistoprovideaneffectivefunctionalframework forboundaryvalueproblemsforpartialdifferentialequations,andvariationalproblems, indomainslackinganyregularity,towhichthetheoryofSobolevspacesavailableinthe existingliteraturedoesnotapply.Inthisconnection,letusmentionthefollowingsimple instance,whichyetgivestheflavorofthegeneralityallowedbytheuseofaninequality like(1.2).Considertheproblem

Δ2u = div F in Ω,

u = 0, B u = 0 on ∂Ω, (1.3)

where Ω is any open set in Rn, n ≥ 3, Δ2 is the bi-Laplace operator, F : Ω → Rn is a given function, and B is the (second-order) boundary operator generated by the minimizationofthequadraticfunctional

 Ω



|∇2u|2+ 2F· ∇udx

among functionsu vanishingon ∂Ω.Here,thedot“·” standsfor scalarproduct inRn. Assumethat F ∈ Ln+22n (Ω),where 2n

n+2 is the Hölderconjugate of the critical Sobolev exponent 2n

n−2. A special case of Theorem 6.3, Section 6, tells us that there exists a constantC = C(n) suchthattheSobolevinequality

∇u L 2n n−2(Ω)≤ C∇ 2u L2(Ω)

holdsforeveryfunctionu vanishingon∂Ω.(Incidentally,noticethat,instead,an inequal-ity of this kind fails if 2u is replaced just with Δu, unless Ω is sufficiently regular.) AstandardargumentrelyinguponRiesz’representationtheoreminHilbertspacesthen yieldstheexistenceofauniquesolutionu to(1.3), whateverΩ is.

Ananalysis ofmoregeneralproblems inarbitrarydomainsgoesbeyond thescope of thepresentcontribution,andisthesubjectof aworkinprogress.

The paper is organized as follows. In the next section we offer a brief overview of some Sobolev type inequalities, in basic cases, which follow from our results, and we discusstheirnoveltyandoptimality.Section3containssomepreliminarydefinitionsand results. The statement of our main results starts with Section 4, which is devoted to our key pointwise inequalities for Sobolev functions on arbitrary open sets. Estimates in rearrangement form are derivedin the subsequent Section5. In Section 6, Sobolev typeinequalitiesinarbitraryopensetsareshowntofollowviasuchestimates.Examples whichdemonstratethesharpnessofourresultsareexhibited inSection7.Inparticular,

Example 7.4shows thatinequalitiesof theform (1.2) maypossiblyfail ifN∂Ω(u) only depends on derivatives of u on ∂Ω up to an order smaller than [m−12 ]. Finally, in the

(5)

Appendix, somenewnotions,whichare introducedinthedefinitionsof theseminorms

N∂Ω(·),arelinkedto classicalpropertiesofSobolevfunctions. 2. Atasteofresults

Inorder togiveanoverallideaofthecontent ofthispaper, weenucleatehereaftera few basicinstancesof theinequalitiesthatcanbe derivedviaourapproach.

We begin with two examples which demonstrate that our conclusions lead to new resultsalsointhecaseoffirst-orderinequalities,namelyinthecasewhenm= 1 in(1.2).

LetΩ beanyopensetinRn,andletμ beaBorelmeasureonΩ suchthatμ(B

r∩Ω)≤

Crα forsomeC > 0,andα∈ (n− 1,n],andfor everyball Br withradius r.Clearly, if

μ=Ln,thenthis conditionholds withα = n.

Assumethat1< p< n andr > 1,andlets= min{n−1 ,n−pαp }.Then

uLs(Ω,μ)≤ C



∇uLp(Ω)+uLr(∂Ω)



(2.1) for some constant C and every function u with bounded support, provided that

Ln(Ω)<∞,μ(Ω)<∞ andHn−1(∂Ω)<∞.Here,Hn−1denotesthe(n−1)-dimensional Hausdorffmeasure.Inparticular,ifr = p(n−1)n−p ,andhences=nαp−p,then(2.1)holdseven iftheassumptiononthefinitenessofthesemeasuresisdropped;inthiscase,theconstant

C depends onlyonn, p,α.Inequality(2.1)follows viaageneralprinciplecontainedin

Theorem 6.1, Section6.It extendsaversionof theSobolevinequalityformeasures,on regulardomains[62,Theorem 1.4.5].Italsoaugments,asfarasmeasuresareconcerned, theresultof[57],whoseapproach,thoughyieldingsharpconstants,isconfinedtonorms evaluated with respect to the Lebesgue measure. Let us point out that, by contrast, ourmethod,beingbasedonrepresentationformulas,neednotleadtooptimalnormsin inequalitieswithp= 1.

Considernowtheborderlinecasecorresponding top= n.Asaconsequenceof Theo-rem 6.1 again,onecanshowthat

uexp L n n−1(Ω,μ)≤ C  ∇uLn(Ω)+u exp Ln−1n (∂Ω) , (2.2)

for some constant C and every function u with bounded support, provided that

Ln(Ω)<∞,μ(Ω)<∞ andHn−1(∂Ω)<∞.Here,· exp L

n

n−1(Ω,μ)and·exp Ln−1n (∂Ω)

denote norms in Orlicz spaces of exponential type on Ω and ∂Ω, respectively. In-equality (2.2) on the one hand extends the Yudovich–Pohozaev–Trudinger inequality to possiblyirregulardomains;ontheotherhand,itenhances,undersomerespect,a re-sult of [58], where optimal constants are exhibited in estimates involving the weaker norm inexp L(Ω),and justfortheLebesguemeasure.

Letusnowturntohigher-orderinequalities.Focusing,forthetimebeing,on second-orderinequalitiesmayhelptograspthequality andsharpnessofourconclusionsinthis

(6)

framework.In theremaining partof this section, wethus assumethat m= 2 in (1.2); wealsoassume,forsimplicity,thatμ=Ln.

First, assume thath = 0. Then we canprove (among other possible choices of the exponents)that,if1< p<n2,then

uL pn n−2p(Ω)≤ C  ∇2u Lp(Ω)+u V1,0L p(n−1) n−p (∂Ω)+uL p(n−1) n−2p (∂Ω)  , (2.3) for some constant C = C(p,n), in particular independent of Ω, and every function u

withbounded support.Note that n−2ppn is thesamecritical Sobolevexponentas inthe caseofregulardomains.Here,· V1,0Lr(∂Ω)denotes,forr∈ [1,∞],theseminorm given

by

uV1,0Lr(∂Ω)= inf

g gLr(∂Ω), (2.4)

wheretheinfimumistakenamongallBorelfunctionsg on ∂Ω suchthat

|u(x) − u(y)| ≤ |x − y|(g(x) + g(y)) for Hn−1-a.e. x, y∈ ∂Ω, (2.5) and Lr(∂Ω) denotes aLebesgue space on ∂Ω with respect to the measure Hn−1. The functiong appearingin(2.5)isanuppergradient,inthesenseof[38],fortherestriction of u to ∂Ω, endowed with the metric inherited from the Euclidean metric in Rn, and with themeasure Hn−1. In[38],a definition of this kind, and anassociated seminorm givenas in(2.4),wereintroduced todefine first-orderSobolevtypespacesonarbitrary metricmeasure spaces.In thelast twodecades, various notionsofupper gradientsand of Sobolevspaces of functionsdefined onmetric measure spaces, havebeen the object ofinvestigationsandapplications.Theyconstitutethetopicofanumberofpapersand monographs,including[4,11,35,39,42,53].

Letus stress that,although thenew termu

V1,0Lp(n−1)n−p (∂Ω) on theright-hand side of(2.3) canbe droppedwhen Ω is aregular,say Lipschitz,domain,it isindispensable ifΩ isarbitrary. This canbe shownby taking into account adomain as in Fig. 1 (see

Example 7.1,Section7).

Asinthecaseofregulardomains,ifp> n2,thentheLebesguenormontheleft-hand sideof(2.3)canbe replacedbythenorminL∞.Indeed,ifr > n− 1,then

uL∞(Ω)≤ C 

∇2u

Lp(Ω)+uV1,0Lr(∂Ω)+uL(∂Ω) (2.6)

foranyopensetΩ suchthatLn(Ω)<∞ andHn−1(∂Ω)<∞,forsomeconstantC,and for any functionu with bounded support. Inparticular, the constantC depends on Ω onlythroughLn(Ω) andHn−1(∂Ω).

Inthe limitingsituationwhen n ≥ 3,p= n2 and r > n− 1, a Yudovich–Pohozaev– Trudingertypeinequalityoftheform

uexp L n

n−2(Ω)≤ C 

∇2u

Ln2(Ω)+uV1,0Lr(∂Ω)+uexp Ln−2n (∂Ω) 

(7)

Fig. 1.Example 7.1, Section7.

holds for some constant C independent of the geometry of Ω, and every function u

with bounded support, provided that Ln(Ω) < ∞ and Hn−1(∂Ω) < ∞. The norms

· exp L n

n−2(Ω) and · exp Ln−2n (∂Ω) are thesame exponential norms appearingin the Yudovich–Pohozaev–Trudingerinequalityonregulardomains,andinitsboundarytrace counterpart.

Considernextthecasewhenstill m= 2 in (1.2), buth= 1.From ourestimatesone can infer that, if 1< p < n and r ≥ 1, and Ω is any open set with Ln(Ω) < ∞ and

Hn−1(∂Ω)<∞,then ∇uLq(Ω)≤ C  ∇2u Lp(Ω)+uV1,0Lr(∂Ω)  (2.8) for some constant C independent of the geometry of Ω, and every function u with

bounded support,where

q = min rn n−1,n−pnp

. (2.9)

In particular, ifr = p(n−1)n−p , and henceq = nnp−p, then theconstantC in (2.8)depends onlyonn andp.

Inequality(2.8)isoptimalundervariousrespects.Forinstance,ifΩ isregular,then,as aconsequenceof(1.1), theseminorm uV1,0Lr(∂Ω) canbereplacedjustwithuLr(∂Ω)

on the right-hand side. Bycontrast, this is impossible ina domain Ω as inFig. 2, for any q∈ [1,n−pnp ], whateverr is –seeExample 7.2, Section7.

(8)

Fig. 2.Example 7.2, Section7.

Fig. 3.Example 7.3, Section7.

The questionof the optimalityof the exponent q given by (2.9)can also be raised. Theanswerisaffirmative.Actually,domainslikethatofFig. 3showthatsuchexponent

q isthelargestpossiblein(2.8)ifnoregularityisimposedonΩ (Example 7.3,Section7). Whenp> n,inequality(2.8)canbe replacedwith

∇uL∞(Ω)≤ C 

∇2u

Lp(Ω)+uV1,0L(∂Ω), (2.10)

for some constant C independent of the geometry of Ω, and every function u with

(9)

Finally,intheborderlinecasecorrespondingtop= n,anexponentialnormisinvolved again. UndertheassumptionthatLn(Ω)<∞ andHn−1(∂Ω)<∞,onehasthat

∇uexp L n n−1(Ω)≤ C  ∇2u Ln(Ω)+u V1,0exp Ln−1n (∂Ω)  (2.11) forsomeconstantC,dependingonΩ onlythroughLn(Ω) andHn−1(∂Ω),andforevery functionu withbounded support.Here,theseminorm · 

V1,0exp Ln−1n (∂Ω) isdefinedas

in(2.4), withthenorm · Lr(∂Ω) replacedwith thenorm · 

exp Ln−1n (∂Ω).Again, the exponential normsin(2.11) are thesameoptimal Orlicz target normsfor Sobolevand traceinequalities,respectively,onregulardomains.

3. Preliminaries

LetΩ be anyopenset inRn,n≥ 2.Givenx∈ Ω,define

Ωx={y ∈ Ω : (1 − t)x + ty ⊂ Ω for every t ∈ (0, 1)}, (3.1) and

(∂Ω)x={y ∈ ∂Ω : (1 − t)x + ty ⊂ Ω for every t ∈ (0, 1)}. (3.2) TheyarethelargestsubsetofΩ and∂Ω,respectively,whichcanbe“seen”from x.Itis easily verifiedthatΩx isanopenset.Thefollowingpropositiontellsusthat(∂Ω)x isa Borelset.

Proposition 3.1. Assume that Ω is an open set in Rn, n≥ 2. Let x∈ Ω. Then theset (∂Ω)x,defined by(3.2),isBorelmeasurable.

Proof. Givenany r∈ Q∩ (0,1),define

(∂Ω)x(r) ={y ∈ ∂Ω : (1 − t)x + ty ⊂ Ω for every t ∈ (0, r)}.

If y ∈ (∂Ω)x(r), then there exists δ > 0 such that Bδ(y)∩ ∂Ω ⊂ (∂Ω)x(r). Thus, for eachr∈ Q∩ (0,1),theset(∂Ω)x(r) isopenin∂Ω,inthetopologyinducedbyRn. The conclusionthenfollows fromthefactthat(∂Ω)x=∩r∈Q∩(0,1)(∂Ω)x(r). 2

Next,wedefine thesets

× Sn−1)0={(x, ϑ) ∈ Ω × Sn−1 : x + tϑ∈ ∂Ω for some t > 0}, (3.3) and

(10)

Clearly,

× Sn−1)0= Ω× Sn−1 if Ω is bounded. (3.5) Let

ζ : (Ω× Sn−1)0→ Rn (3.6) bethefunctiondefinedas

ζ(x, ϑ) = x + tϑ, where t is such that x + tϑ∈ (∂Ω)x.

Inother words,ζ(x,ϑ) is thefirstpoint ofintersectionof thehalf-line {x+ tϑ: t> 0}

with∂Ω.

Given afunctiong : ∂Ω → R,with compact support,we adoptthe conventionthat

g(ζ(x,ϑ)) isdefinedforevery(x,ϑ)∈ Ω× Sn−1, onextendingitby0 on(Ω× Sn−1); namely,weset

g(ζ(x, ϑ)) = 0 if (x, ϑ)∈ (Ω × Sn−1). (3.7) Letusnextintroducethefunctions

a : Ω× Sn−1→ [−∞, 0) and b : Ω × Sn−1→ (0, ∞] (3.8) givenby b(x, ϑ) = |ζ(x,ϑ) − x| if (x,ϑ) ∈ (Ω × S n−1) 0, otherwise, (3.9) and a(x, ϑ) =−b(x, −ϑ) if (x, ϑ) ∈ Ω × Sn−1. (3.10) Proposition 3.2. The functionζ is Borelmeasurable. Hence, the functions a and b are

Borelmeasurableaswell.

Proof. Assume first that Ω is bounded, so that (Ω× Sn−1)0 = Ω× Sn−1. Consider a sequenceofnestedpolyhedra{Qk} invadingΩ,andthecorresponding sequenceof func-tions{ζk},definedasζ,withΩ replacedwithQk.SuchfunctionsareBorelmeasurable, byelementaryconsiderations, andhenceζ is also Borelmeasurable,sinceζk converges toζ pointwise.

Next,assumethatΩ isunbounded.Foreachh∈ N,considerthesetΩh= Ω∩ Bh(0), where Bh(0) is the ball, centered at 0, with radius h.Let ζh and bh be the functions, definedasζ andb,withΩ replacedwithΩh.SinceΩhisbounded,thenwealreadyknow

(11)

that bh is Borel measurable. Moreover, bh converges to b pointwise. Hence, b is Borel measurable as well,andinparticular theset(Ω× Sn−1)0, whichagreeswith {b<∞}, is Borel measurable.Finally, the functionζh is Borel measurable,inasmuch as Ωh is a bounded set.Moreover,ζh convergestoζ pointwiseontheBorelset(Ω× Sn−1)0.Thus,

ζ isBorelmeasurable. 2

Given m ∈ N and p∈ [1,∞], we denote byVm,p(Ω) the Sobolevtypespace defined as

Vm,p(Ω) =u : u is m-times weakly differentiable in Ω, and|∇mu| ∈ Lp(Ω) .

(3.11) Let usnotice that,inthedefinitionofVm,p(Ω),it isonlyrequired thatthederivatives of u of the highest order m belong to Lp(Ω). Replacing Lp(Ω) in (3.11) with a more general Banach function space X(Ω) leads to the notion of m-th order Sobolev type space VmX(Ω) builtuponX(Ω).

For k ∈ N0, we denote as usual byCk(Ω) the space of real-valued functionswhose

k-th orderderivativesinΩ arecontinuousuptotheboundary.Wealsoset

Cbk(Ω) ={u ∈ Ck(Ω) : u has bounded support}. (3.12) Clearly,

Cbk(Ω) = Ck(Ω) if Ω is bounded.

Let α = (α1,. . . ,αn) bea multi-indexwith αi ∈ N0 for i = 1,. . . ,n. Weadopt the notations|α|= α1+· · ·+αn,α!= α1!· · · αn!,andϑα= ϑα11· · · ϑαnnforϑ∈ Rn.Moreover, we setu= ∂|α|u

∂xα11 ...∂xαnn foru: Ω→ R.

We need to extend the notion of upper gradient g for the restriction of u to ∂Ω

appearingin(2.5)tothecaseofhigher-orderderivatives.Tothispurpose,letusdenote bygk,j,wherek∈ N0andj = 0,1,(k,j) = (0,0),any Borelfunctionon∂Ω such that:

(i) Ifk∈ N, j = 0,andu∈ Cbk−1(Ω), |α|≤k−1 (2k− 2 − |α|)! (k− 1 − |α|)!α! (y− x)α |y − x|2k−1  (−1)|α|Dαu(y)− Dαu(x) ≤ gk,0(x) + gk,0(y) (3.13) forHn−1-a.e. x,y∈ ∂Ω.

(12)

(ii) Ifk∈ N,j = 1,andu∈ Cbk(Ω), n i=1 |α|≤k−1 (2k− 2 − |α|)! (k− 1 − |α|)!α! (y− x)α |y − x|2k−1  (−1)|α|Dα ∂u∂x i(y)− D α ∂u ∂xi(x)  ≤ gk,1(x) + gk,1(y) (3.14) forHn−1-a.e.x,y∈ ∂Ω. (iii) Ifk = 0,j = 1,andu∈ C0 b(Ω), |u(x)| ≤ g0,1(x) (3.15) forHn−1-a.e.x∈ ∂Ω.

Notethatinequality(3.13), withk = 1,agrees with(2.5),and henceg1,0 hasthesame roleasg in(2.5).Letusalsopoint outthat,as(2.5)extendsaclassicalpropertyofthe gradientofweaklydifferentiablefunctionsinRn,likewiseitshigher-orderversions(3.13) and(3.14)extendaparallelpropertyoffunctionsinRnendowedwithhigher-orderweak derivatives.ThisisshowninProposition A.1oftheAppendix.

Inanalogywith(2.4),weintroducetheseminormgiven, forr∈ [1,∞], by

uVk,jLr(∂Ω)= inf

gk,jg

k,j

Lr(∂Ω) (3.16)

where k, j and u are as above, and the infimum is extended over all functions gk,j fulfillingtheappropriatedefinitionamong(3.13),(3.14)and(3.15).Moregenerally,given aBanachfunctionspaceZ(∂Ω) on∂Ω withrespecttotheHausdorffmeasure Hn−1,we define

uVk,jZ(∂Ω)= inf

gk,jg

k,j

Z(∂Ω). (3.17)

Observethat,inparticular,

uV0,1Z(∂Ω)=uZ(∂Ω).

4. Pointwiseestimates

In the present section we establish our first main result: a pointwise estimate for Sobolevfunctions,andtheirderivatives,inarbitraryopensets.Inwhatfollowswedefine, fork∈ N,

(k) =



0 if k is odd,

(13)

Theorem 4.1 (Pointwise estimate). Let Ω be any open set in Rn, n ≥ 2. Assume that m∈ N andh∈ N0aresuchthat0< m−h< n.ThenthereexistsaconstantC = C(n,m)

such that |∇hu(x)| ≤ C   Ω |∇mu(y)| |x − y|n−m+hdy + m−h−1 k=1  Ω  Sn−1 g[k+h+12 ],(k+h)(ζ(y, ϑ)) |x − y|n−k dHn−1(ϑ) dy +  Sn−1 g[h+12 ],(h)(ζ(x, ϑ)) dHn−1(ϑ)  for a.e. x∈ Ω, (4.2) for every u ∈ Vm,1(Ω) ∩ C[m−12 ] b (Ω). Here, g[ k+h+1 2 ],(k+h) is any function as in (3.13)–(3.15),andconvention(3.7)is adopted.

Remark 4.2. In the case when m− h = n, and Ω is bounded, an estimate analogous to(4.2)canbeproved,withthekernel |x−y|n−m+h1 inthefirstintegralontheright-hand

side replaced with log|x−y|C . The constant C depends on n andthe diameter of Ω. If

m− h> n,and Ω isbounded, then thekernel isbounded byaconstantdepending on

n, m andthediameterof Ω.

Remark4.3.Undertheassumptionthat

u =∇u = · · · = ∇[m−12 ]u = 0 on ∂Ω, (4.3)

onecanchooseg[h+1

2 ],(h)= 0 fork = 0,. . . ,m− h− 1 in(4.2).Hence, |∇hu(x)| ≤ C  Ω |∇mu(y)| |x − y|n−m+hdy for a.e. x∈ Ω. (4.4) Aspecialcaseof(4.4),correspondingtoh= m− 1,istheobjectof[62,Theorem 1.6.2]. Remark 4.4. As already mentioned in Section 1, the order m2−1 of the derivatives prescribed on∂Ω,whichappearsontheright-handsideof(4.2),isminimal forSobolev type inequalities to hold in arbitrarydomains. This issue isdiscussed inExample 7.4, Section7below.

A keystep inthe proofof Theorem 4.1iscontained inthe nextlemma, whichdeals with thecasewhenh= m− 1 inTheorem 4.1.

Lemma 4.5.LetΩ be any opensetin Rn,n≥ 2. (i) Ifu∈ V2−1,1(Ω)∩ Cb−1(Ω) for some∈ N,then

(14)

|∇2−2u(x)| ≤ C   Ω |∇2−1u(y)| |x − y|n−1 dy +  Sn−1 g−1,1(ζ(x, ϑ)) dHn−1(ϑ)  for a.e. x∈ Ω, (4.5)

forsomeconstant C = C(n,).

(ii) If u∈ V2,1(Ω)∩ C−1

b (Ω) for some∈ N,then

|∇2−1u(x)| ≤ C  Ω |∇2u(y)| |x − y|n−1 dy +  Sn−1 g,0(ζ(x, ϑ)) dHn−1(ϑ)  (4.6)

for some constant C = C(n,). Here, g−1,1 and g,0 are functions as in

(3.13)–(3.15),andconvention (3.7)isadopted.

Ourproof ofLemma 4.5 inturnrequires thefollowingrepresentationformulaforthe (2− 1)-th order derivative of a one-dimensional function in an interval, in terms of its2-th derivativeintherelevant interval, andof itsderivatives upto theorder − 1

evaluatedattheendpoints.

Lemma4.6. Let−∞< a< b<∞.Assumethat ψ∈ W2,1(a,b) for some∈ N.Then ψ(2−1)(t) = t  a Q2−12τ− a − b b− a ψ(2)(τ ) dτ− b  t Q2−1a + b− 2τ b− a ψ(2)(τ ) dτ + (2− 1)!(−1) −1 k=0 (2− k − 2)! k!(− k − 1)! 1 (b− a)2−k−1  (−1)k+1ψ(k)(b) + ψ(k)(a) (4.7)

fort ∈ (a,b). Here, ψ(k) denotes thek-th order derivative of ψ,and Q

2−1 isthe

poly-nomial ofdegree 2− 1, obeying

Q2−1(t) + Q2−1(−t) = 1 for t ∈ R, (4.8)

and

Q2−1(−1) = Q(1)2−1(−1) = · · · = Q(2−1−1)(−1) = 0. (4.9) Proof. Letusrepresentψ as

(15)

where andς arethesolutionstotheproblems  (2)(t) = ψ(2)(t) in (a, b), (k)(a) = (k)(b) = 0 for k = 0, 1, . . . , − 1, (4.11) and  ς(2)(t) = 0 in (a, b),

ς(k)(a) = ψ(k)(a), ς(k)(b) = ψ(k)(b) for k = 0, 1, . . . , − 1, (4.12) respectively.Letusfirstfocus onproblem(4.11).Weclaimthat

(2−1)(t) = t  a Q2−1 − a − b b− a ψ(2)(τ ) dτ b  t Q2−1a + b− 2τ b− a ψ(2)(τ ) dτ for t∈ (a, b), (4.13)

whereQ2−1isasinthestatement.Inordertoverify(4.13),letusconsidertheauxiliary problem



ω2(s) = φ(s) in (−1, 1),

ω(k)(±1) = 0, k = 0, 1, . . . ,  − 1, (4.14) where φ ∈ L1(a,b) is any given function. Let κ: [−1,1]2 → R be the Green function associated withproblem(4.14),sothat

ω(s) =

1  −1

κ(s, r)φ(r) dr for s∈ [−1, 1]. (4.15)

Thefunctionκ takesanexplicitform([14];seealso[36,Section 2.6]),givenby

κ(s, r) = C|s − r|2−1

1−sr |s−r|

1

(t2− 1)−1dt for s = r, (4.16)

whereC = C() isasuitableconstant.Onecaneasilyseefromformula(4.16)thatκ(s,r)

is apolynomialofdegree2− 1 in s forfixedr, andapolynomialofdegree2− 1 inr

forfixeds,bothin{(s,r)∈ [−1,1]2: s> r},andin{(s,r)∈ [−1,1]2: s< r}.Moreover,

(16)

κ(s, r) = C  −1 j=0  − 1 j  (−1)−1−j 2j + 1 (1− sr) 2j+1(s− r)2−2k−2 − (s − r)2−1−1 j=0  − 1 j  (−1)−1−j 2j + 1  . (4.17) Thus,ifs> r, ∂2−1κ ∂s2−1(s, r) = C(2− 1)!  −1 j=0  − 1 j  (−1)+j 2j + 1 r 2j+1 −1 j=0  − 1 j  (−1)−1−j 2j + 1  , (4.18) apolynomialof degree 2− 1 in r, dependingonly on oddpowers of r. Letus denote thispolynomialbyQ2−1(r).It followsfrom(4.18) thatQ2−1(−1)= 0.Moreover,

∂Q2−1 ∂r (r) = ∂r  ∂2−1κ ∂s2−1(s, r)  = C(2− 1)! −1 j=0  − 1 j  (−1)+jr2j =−C(2 − 1)!(r2− 1)−1. (4.19) Thus, Q2−1 vanishes, together with all its derivatives up to the order − 1, at −1, namelyQ2−1 fulfills(4.9).Equation(4.18)alsotellsus thatQ2−1(s)− Q2−1(0) is an oddfunction,and hence

Q2−1(s) + Q2−1(−s) = 2Q2−1(0) for s∈ R. (4.20) Sinceκ isaneven function,

∂2−1κ

∂s2−1(s, r) = 2−1κ

∂s2−1(−s, −r) = −Q2−1(−r) if −1 ≤ s < r ≤ 1.

Thus,(2− 1)-timesdifferentiationofequation(4.15) yields

ω(2−1)(s) = s  −1 Q2−1(r)φ(r) dr− 1  s Q2−1(−r)φ(r) dr for s ∈ [−1, 1]. (4.21)

Since ω(2) = φ, an integration by parts in (4.21), equation (4.20), and the fact that

Q2−1(−1)= 0,tellusthat ω(2−1)(s) = 2Q2−1(0)ω(2−1)(s)− s  −1 Q2−1(r)ω(2−1)(r) dr 1  s Q2−1(−r)ω(2−1)(r) dr for s∈ [−1, 1]. (4.22)

(17)

Owing to the arbitrariness of ω, equation (4.22) ensures that 2Q2−1(0) = 1. Equa-tion(4.8)thusfollowsfrom (4.20).

Thefunction definedas

(t) = ω  2t−a−b b−a for t∈ [a, b]

isthusthesolutiontoproblem(4.11),andtherepresentationformula(4.13) followsvia achangeofvariablesin(4.21).

Consider nextproblem (4.12). Thefunction ς is apolynomialof degree 2− 1, and

ς(2−1) is a constant which, owing to the two-point Taylor interpolation formula (see e.g.[30,Chapter 2,Section 2.5,Ex. 3]),isgivenby

ς(2−1)= (2− 1)!  d−1 dt−1  ς(t) (t− b)  |t=a + d −1 dt−1  ς(t) (t− a)  |t=b  = (2− 1)!  d−1 dt−1  ψ(t) (t− b)  |t=a + d −1 dt−1  ψ(t) (t− a)  |t=b  . (4.23)

Leibnitz’differentiationruleforproductsyields  d−1 dt−1  ψ(t) (t− b)  |t=a + d −1 dt−1  ψ(t) (t− a)  |t=b  = (−1) −1 k=0 (2− k − 2)! k!(− k − 1)! 1 (b− a)2−k−1  (−1)k+1ψ(k)(b) + ψ(k)(a). (4.24)

Equation(4.7)followsfrom (4.13),(4.23) and(4.24). 2

Proof of Lemma 4.5. Given x∈ Ω and ϑ∈ Sn−1, leta(x,ϑ) andb(x,ϑ) be definedas in(3.10)and (3.9),respectively.

Webeginwiththeproofof(4.5)for= 1.Ifu∈ V1,1(Ω)∩C0

b(Ω),then,byastandard property ofSobolevfunctions,fora.e.x∈ Ω thefunction

[0, b(x, ϑ)] t → u(x + tϑ) belongsto V1,1(0,b(x,ϑ)) forHn−1-a.e.ϑ∈ Sn−1,and

d

dtu(x + tϑ) =∇u(x + tϑ) · ϑ for a.e. t ∈ [0, b(x, ϑ)].

Hence,forany suchx andϑ,

u(ζ(x, ϑ))− u(x) =

b(x,ϑ)

0

(18)

where convention (3.7)is adopted.Integrating both sidesof equation (4.25) over Sn−1 yields nωnu(x) =  Sn−1 u(ζ(x, ϑ)) dHn−1(ϑ)−  Sn−1 b(x,ϑ) 0 ∇u(x + tϑ) · ϑ dt dHn−1(ϑ), (4.26) whereωn= π n 2/Γ(1+n

2),theLebesguemeasureof theunitballinR

n.Onehasthat  Sn−1 b(x,ϑ) 0 ∇u(x + tϑ) · ϑ dt dHn−1(ϑ) =  Sn−1 b(x,ϑ) 0 1 tn−1∇u(x + ϑ) · ϑ t n−1d dHn−1(ϑ) =  Ωx ∇u(y) · (y − x) |x − y|n dy. (4.27) Inequality(4.5),with= 1,followsfrom(4.26) and(4.27).

Letusnextprove(4.6).Ifu∈ V2,1(Ω)∩ Cb−1(Ω),thenfora.e.x∈ Ω,thefunction [a(x, ϑ), b(x, ϑ)] t → u(x + tϑ)

belongstoV2,1(a(x,ϑ),b(x,ϑ)) forHn−1-a.e. ϑ∈ Sn−1.Consideranysuchx and ϑ.If (x,ϑ)∈ (Ω× Sn−1) 0,then, byLemma 4.6, d2−1 dt2−1u(x + tϑ) = t  a(x,ϑ) Q2−1 − a(x, ϑ) − b(x, ϑ) b(x, ϑ)− a(x, ϑ) d2 dτ2u(x + τ ϑ) dτ b(x,ϑ) t Q2−1a(x, ϑ) + b(x, ϑ)− 2τ b(x, ϑ)− a(x, ϑ) d2 dτ2u(x + τ ϑ) dτ + (2− 1)!(−1) × −1 k=0 (2− k − 2)! k!(− k − 1)!  (−1)k+1dku(x+tϑ) dtk  |t=b(x,ϑ)+ dku(x+tϑ) dtk  |t=a(x,ϑ)  (b(x, ϑ)− a(x, ϑ))2−k−1 (4.28) fort∈ (a(x,ϑ),b(x,ϑ)).If,instead,(x,ϑ)∈ (Ω× Sn−1), then

d2−1 dt2−1u(x + tϑ) =  t∞ d2 dτ2u(x + τ ϑ) dτ, if b(x, ϑ) =∞, t −∞ d 2 dτ2u(x + τ ϑ) dτ, if a(x, ϑ) =−∞, (4.29)

fort∈ (a(x,ϑ),b(x,ϑ)) (ifboth b(x,ϑ)=∞ anda(x,ϑ)=−∞,then eitherexpression ontheright-handsideof(4.29) canbe exploited).

(19)

Wehavethat dk dtku(x + tϑ) = |α|=k k! α!ϑ

αDαu(x + tϑ) for a.e. t∈ (a(x, ϑ), b(x, ϑ)), (4.30)

fork = 1,. . . ,2.From (4.28)–(4.30)weinferthat |α|=2−1 (2− 1)! α! ϑ αDαu(x) =−χ(Ω×Sn−1)(x, ϑ) |α|=2−1 (2− 1)! α! ϑ α ±∞  0 |γ|=1 ϑγDα+γu(x + τ ϑ) dτ + χ(Ω×Sn−1)0(x, ϑ)  0 a(x,ϑ) Q2−1 − a(x, ϑ) − b(x, ϑ) b(x, ϑ)− a(x, ϑ) d2 dτ2u(x + τ ϑ)dτ b(x,ϑ) 0 Q2−1 a(x, ϑ) + b(x, ϑ)− 2τ b(x, ϑ)− a(x, ϑ) d2 dτ2u(x + τ ϑ) dτ + (2− 1)!(−1) −1 k=0 (2− k − 2)! k!(− k − 1)! × |α|=k k! α!ϑ α 

(−1)k+1Dαu(x + b(x, ϑ)ϑ) + Dαu(x + a(x, ϑ)ϑ)

(b(x, ϑ)− a(x, ϑ))2−k−1



, (4.31)

where theupperlimitofintegrationinthefirstintegralontheright-handsideiseither +∞,or −∞ accordingtowhetherb(x,ϑ)=∞ or a(x,ϑ)=−∞.

Denote by{Pβ} the system of allhomogeneous polynomials of degree2− 1 in the variables ϑ1,. . . ,ϑn suchthat

 Sn−1

Pβ(ϑ)ϑαdHn−1(ϑ) = δαβ,

whereδαβstandsfortheKroneckerdelta.Onmultiplyingequation(4.31)by (2−1)!(ϑ),and integrating overSn−1 oneobtainsthat

u(x) β! =  Sn−1 χ×Sn−1)∞(x, ϑ)Pβ(ϑ) |α|=2−1 |γ|=1 ϑα+γ α! ±∞  0 Dα+γu(x + τ ϑ) dτ dHn−1(ϑ)

(20)

+  Sn−1 χ(Ω×Sn−1)0(x, ϑ) Pβ(ϑ) (2− 1)! ×  0 a(x,ϑ) Q2−1 − a(x, ϑ) − b(x, ϑ) b(x, ϑ)− a(x, ϑ) d2 dτ2u(x + τ ϑ)dτ b(x,ϑ) 0 Q2−1a(x, ϑ) + b(x, ϑ)− 2τ b(x, ϑ)− a(x, ϑ) d2 dτ2u(x + τ ϑ) dτ  dHn−1(ϑ) + (−1)  Sn−1 Pβ(ϑ) × |α|≤−1 (2− |α| − 2)! ϑα (− |α| − 1)!α! 

(−1)|α|+1Dαu(x + b(x, ϑ)ϑ) + Dαu(x + a(x, ϑ)ϑ)

(b(x, ϑ)− a(x, ϑ))2−|α|−1 dH

n−1(ϑ).

(4.32) Thereexist constants C = C(n,) andC = C(n,) suchthat

 Sn−1 χ×Sn−1)∞(x, ϑ)Pβ(ϑ) |α|=2−1 |γ|=1 ϑα+γ α! ±∞  0 Dα+γu(x + τ ϑ) dτ dHn−1(ϑ) ≤ C  Sn−1  0 |∇2u(x + τ ϑ)| dτ dHn−1(ϑ) = C  Sn−1  0 |∇2u(x + τ ϑ)| τn−1 τn−1dτ dHn−1(ϑ) ≤ C Ω |∇2u(y)| |x − y|n−1 dy. (4.33)

Next,weclaimthatthere existsaconstantC = C(,n) suchthat  Sn−1 χ(Ω×Sn−1)0(x, ϑ)  Pβ(ϑ) 0  a(x,ϑ) Q2−1 − a(x, ϑ) − b(x, ϑ) b(x, ϑ)− a(x, ϑ) d2 dτ2u(x + τ ϑ)dτ − Pβ(ϑ) b(x,ϑ) 0 Q2−1 a(x, ϑ) + b(x, ϑ)− 2τ b(x, ϑ)− a(x, ϑ) d2 dτ2u(x + τ ϑ)dτ  dHn−1(ϑ) ≤ C  Ω |∇2u(y)| |x − y|n−1dy. (4.34)

(21)

 Sn−1 χ(Ω×Sn−1)0(x, ϑ)Pβ(ϑ) × 0  a(x,ϑ) Q2−12τ− a(x, ϑ) − b(x, ϑ) b(x, ϑ)− a(x, ϑ) d2 dr2u(x + τ ϑ)dτ dH n−1(ϑ) =  Sn−1 χ×Sn−1)0(x, ϑ)Pβ(ϑ) × −a(x,ϑ) 0 Q2−1  −2r + a(x, ϑ) + b(x, ϑ) b(x, ϑ)− a(x, ϑ) d2 dτ2u(x− rϑ)dr dH n−1(ϑ) =  Sn−1 χ(Ω×Sn−1)0(x,−θ)Pβ(−θ) × −a(x,−θ) 0 Q2−1  −2r + a(x,−θ) + b(x, −θ) b(x,−θ) − a(x, −θ) d2 dr2u(x + rθ)dr dH n−1(θ) =  Sn−1 χ(Ω×Sn−1)0(x, θ)Pβ(θ) × b(x,θ) 0 Q2−1a(x, θ) + b(x, θ)− 2r b(x, θ)− a(x, θ) d2 dr2u(x + rθ)dr dH n−1(θ),

wherewehavemadeuseofthefactthat(−θ)=−Pβ(θ) if|β|= 2− 1,andof(3.10). Thus,  Sn−1 χ(Ω×Sn−1)0(x, ϑ)  Pβ(ϑ) 0  a(x,ϑ) Q2−1 − a(x, ϑ) − b(x, ϑ) b(x, ϑ)− a(x, ϑ) d2 dτ2u(x + τ ϑ)dτ − Pβ(ϑ) b(x,ϑ) 0 Q2−1a(x, ϑ) + b(x, ϑ)− 2τ b(x, ϑ)− a(x, ϑ) d2 dτ2u(x + τ ϑ)dτ  dHn−1(ϑ) = 2  Sn−1 Pβ(ϑ) b(x,ϑ) 0 Q2−1 a(x, ϑ) + b(x, ϑ)− 2r b(x, ϑ)− a(x, ϑ) d2 dr2u(x + rϑ) drdH n−1(ϑ) = 2  Sn−1 Pβ(ϑ)

(22)

× b(x,ϑ) 0 Q2−1a(x, ϑ) + b(x, ϑ)− 2r b(x, ϑ)− a(x, ϑ) |α|=2 (2)! α! ϑ αDαu(x + rϑ) drdHn−1(ϑ) ≤ C  Sn−1 b(x,ϑ) 0 |∇2u(x + rϑ)| drdHn−1(ϑ)≤ C Ωx |∇2u(y)| |x − y|n−1 dy

forsomeconstants C = C(n,) andC= C(n,).Hence,inequality(4.34) follows. Finally,bydefinition(3.13),thereexists aconstantC = C(n,) suchthat

 Sn−1 χ(ω×Sn−1)0(x, ϑ)(−1)Pβ(ϑ) × |α|≤−1 (2− |α| − 2)! (− |α| − 1)!α!ϑ α × 

(−1)|α|+1Dαu(x + b(x, ϑ)ϑ) + Dαu(x + a(x, ϑ)ϑ)

(b(x, ϑ)− a(x, ϑ))2−|α|−1 dHn−1(ϑ) ≤ C  Sn−1 χ(ω×Sn−1)0(x, ϑ)g,0(x + ϑa(x, ϑ)) + g,0(x + ϑb(x, ϑ))dHn−1(ϑ) = 2C  Sn−1 g,0(ζ(x, ϑ)) dHn−1(ϑ). (4.35) Combining(4.32)–(4.35)yields (4.6).

Inequality(4.5), with≥ 2, follows onapplying (4.6)with u replaced withits first-orderderivatives. 2

ProofofTheorem 4.1. Forsimplicityofnotation,weconsider thecasewhenh= 0,the proof in the generalcase being analogous. Let u ∈ Vm,1(Ω)∩ C[m−12 ]

b (Ω). By inequal-ity(4.5)with= 1, |u(x)| ≤ C   Ω |∇u(y)| |x − y|n−1dy +  Sn−1 g0,1(ζ(x, ϑ)) dHn−1(ϑ)  for a.e. x∈ Ω. (4.36)

From(4.36) andanapplicationofinequality(4.6)with= 1 oneobtainsthat

|u(x)| ≤ C   Ω  Ω |∇2u(z)| |y − z|n−1 dz dy |x − y|n−1 +  Ω  Sn−1 g1,0(ζ(y, ϑ))dH n−1(ϑ) dy |x − y|n−1 +  Sn−1 g0,1(ζ(x, ϑ)) dHn−1(ϑ) 

(23)

≤ C  Ω |∇2u(z)| |x − z|n−2dz +  Ω  Sn−1 g1,0(ζ(y, ϑ))dH n−1(ϑ) dy |x − y|n−1 +  Sn−1 g0,1(ζ(x, ϑ)) dHn−1(ϑ)  for a.e. x∈ Ω, (4.37)

for someconstantsC = C(n) and C = C(n).Notethatinthelast inequalitywehave made useof aspecialcaseofthewellknownidentity

 Rn 1 |x − y|n−σ  Rn f (z) |y − z|n−γdz dy = C  Rn f (z) |x − z|n−σ−γ dz for a.e. x∈ Rn, (4.38) which holds forsomeconstantC = C(n,σ,γ) and foreverycompactly supported inte-grable functionf ,providedthatσ > 0,γ > 0 and σ + γ < n.

Inequality(4.37) inturnyields, viaanapplicationofinequality(4.5)with= 2,

|u(x)| ≤ C   Ω |∇3u(z)| |x − z|n−3 dz +  Ω  Sn−1 g1,1(ζ(y, ϑ))dH n−1(ϑ) dy |x − y|n−2 +  Ω  Sn−1 g1,0(ζ(y, ϑ))dH n−1(ϑ) dy |x − y|n−1 +  Sn−1 g0,1(ζ(x, ϑ)) dHn−1(ϑ)  for a.e. x∈ Ω, (4.39)

for some constant C = C(n). A finite induction argument, relying upon an alternate iterateduseofinequalities(4.6)and(4.5)asabove,eventuallyleadsto (4.2). 2 5. Estimatesinrearrangementform

Thepointwiseboundsestablishedintheprevioussectionenableustoderive rearrange-mentestimatesforfunctions,andtheirderivatives,withrespecttoanyBorelmeasureμ

onΩ suchthat

μ(Br(x)∩ Ω) ≤ Cμrα for x∈ Ω and r > 0, (5.1) forsomeα∈ (n−1,n] andsomeconstantCμ> 0.Here,Br(x) denotestheball,centered at x, withradius r.

Recall that, given a measure space R, endowed with apositive measure ν, the de-creasing rearrangement φ∗ν : [0,∞)→ [0,∞] of a ν-measurable functionφ : R→ R is definedas

φ∗ν(s) = inf{t ≥ 0 : ν(|φ| > t}) ≤ s} for s ∈ [0, ∞).

(24)

(φ + ψ)∗ν(s)≤ φ∗ν(s/2) + ψν∗(s/2) for s≥ 0, (5.2) foreverymeasurable functionsφ andψ onR.

Anyfunctionφ sharesitsintegrabilitypropertieswithitsdecreasingrearrangement φ∗ν, since

ν({|φ| > t}) = L1({φ∗ν> t}) for every t ≥ 0.

As a consequence, any norm inequality, involving rearrangement-invariant norms, be-tweentherearrangementsofthederivativesofSobolevfunctionsandtherearrangements ofitslower-orderderivatives,immediatelyyieldsacorrespondinginequalityforthe orig-inalSobolevfunctions.Thus,therearrangementinequalitiesto beestablishedhereafter reducetheproblem ofn-dimensionalSobolevtypeinequalitiesinarbitraryopensetsto considerably simpler one-dimensional Hardy type inequalities – see Theorem 6.1, Sec-tion6below.

Theorem 5.1 (Rearrangement estimates). Let Ω be any open bounded open set in Rn,

n≥ 2. Let m∈ N and h∈ N0 be such that 0< m− h< n. Assumethat μ isa Borel

measurein Ω fulfilling (5.1)for some α∈ (n− 1,n] andfor someCμ > 0. Then there

exist constantsc= c(n,m) and C = C(n,m,α,Cμ) suchthat

|∇hu| μ(cs)≤ C  s−n−m+hα snα  0 |∇mu| Ln(r)dr +  snα r−n−m+hn |∇mu|∗ Ln(r)dr + m−h−1 k=1  s−n−1−kα sn−1α 0  g[k+h+12 ],(k+h) Hn−1(r)dr +  sn−1α r−n−1−kn−1 g[k+h+12 ],(k+h) Hn−1(r)dr  + s−n−1α sn−1α 0  g[h+12 ],(h) Hn−1(r)dr  for s > 0, (5.3) foreveryu∈ Vm,1(Ω)∩ C[m2−1]

b (Ω).Here, (·) is definedasin (4.1),andg[

k+h+1 2 ],(k+h)

denotesanyBorelfunctionon∂Ω fulfillingtheappropriateconditionfrom (3.13)–(3.15).

Remark5.2.Ininequality(5.3), andinwhatfollows, whenconsideringrearrangements and norms with respect to ameasure μ, Sobolev functions and their derivatives have tobe interpretedas theirtraces withrespect toμ. Suchtraces arewell defined,thanks to standard (local) Sobolev inequalities with measures, owing to the assumption that

(25)

α∈ (n− 1,n] in(5.1). Ananalogousconvention appliesto theintegraloperatorsto be consideredbelow.

In preparation for theproof of Theorem 5.1, weintroduce afew integral operators, and provepointwise estimatesfortheirrearrangements.

LetΩ be anyopenset inRn.WedefinetheoperatorT as Tg(x) =

 Sn−1

|g(ζ(x, ϑ))| dHn−1(ϑ) for x∈ Ω, (5.4) atanyBorelfunction g : ∂Ω→ R.Here,andinwhatfollows,weadoptconvention(3.7). Note that,owing toFubini’s theorem,Tg isameasurable functionwith respectto any Borelmeasure in Ω.

Forγ∈ (0,n),wedenote by theclassicalRiesz potentialoperatorgivenby

Iγf (x) =  Ω

f (y)

|y − x|n−γ dy for x∈ Ω, (5.5) at anyf ∈ L1(Ω),andwecall N

γ theoperatordefinedas

Nγg(x) =  ∂Ω g(y) |x − y|n−γdHn−1(y) for x∈ Ω, (5.6) at anyfunctiong∈ L1(∂Ω).

Finally,wedefine theoperator asthecomposition

= Iγ◦ T. (5.7) Namely, Qγg(x) =  Ω  Sn−1 |g(ζ(y, ϑ))|dHn−1(ϑ) dy |x − y|n−γ for x∈ Ω, (5.8) forany Borelfunctiong : ∂Ω→ R.

Ouranalysisofthese operatorsrequiresafewnotationsandpropertiesfrom interpo-lationtheory.

AssumethatR isameasurespace,endowedwithapositivemeasureν.Givenapair

X1(R) andX2(R) ofnormedfunctionspaces,a functionφ∈ X1(R)+X2(R),ands∈ R, we denotebyK(φ,s;X1(R),X2(R)) theassociated Peetre’sK-functional, definedas

K(s, φ; X1(R), X2(R)) = inf φ=φ12  1X1(R)+ sφ2X2(R)  for s > 0.

(26)

WeneedanexpressionfortheK-functional(uptoequivalence)inthecasewhenX1(R) andX2(R) arecertainLebesgueorLorentzspaces, andR isoneofthemeasure spaces mentioned above. Recall that, given σ > 1, the Lorentz space Lσ,1(R) is the Banach functionspaceofthosemeasurable functionsφ onR forwhichthenorm

φLσ,1(R)=

 0

φ∗ν(s)s−1+1σds

isfinite.TheLorentzspaceLσ,∞(R),alsocalledMarcinkiewiczspaceorweak-Lσspace, istheBanachfunctionspaceofthosemeasurablefunctionsφ onR forwhichthequantity

φLσ,∞(R)= sup s>0

1φ∗

ν(s)

isfinite.Notethat,inspiteofthenotation,thisisnotanorm. However,itisequivalent to anorm,upto multiplicativeconstants depending onσ, obtainedonreplacing φ∗ν(s) with 1s0sφ∗ν(r)dr.

Itiswellknownthat

K(φ, s; L1(R), L∞(R)) = s  0

φ∗ν(r) dr for s > 0, (5.9)

foreveryφ∈ L1(R)+ L(R)[9,Chapter 5, Theorem 1.6].Ifσ > 1,then

K(φ, s; Lσ,∞(R), L∞(R)) ≈ rσ1φ∗

ν(r)L∞(0,sσ) for s > 0, (5.10)

foreveryφ∈ Lσ,∞(R)+ L∞(R)[44,Equation (4.8)].Moreover,

K(φ, s; L1(R), Lσ,1(R)) ≈ sσ  0 φ∗ν(r) dr + s  sσ r−σ1φ∗ ν(r) dr for s > 0, (5.11)

for every φ ∈ L1(R)+ Lσ,1(R) [44, Theorem 4.2]. In (5.10) and (5.11), the notation≈”meansthatthetwo sidesareboundedbyeachotheruptomultiplicativeconstants dependingonσ.

Let R and S be positive measure spaces. An operator L defined on a linear space ofmeasurablefunctionsonR,andtakingvaluesinto thespaceofmeasurablefunctions onS,iscalledsub-linearif,foreveryφ1andφ2inthedomainofL andeveryλ12∈ R,

|L(λ1φ1+ λ2φ2)| ≤ |λ1||Lφ1| + |λ2||Lφ2|.

A basicresult inthe theory of real interpolation tellsus what follows. Assumethat L

is asub-linear operatoras above,and Xi(R) andYi(S), i= 1,2,are normed function spacesonR suchthat

(27)

L : Xi(R) → Yi(S) (5.12) withnormsnotexceedingNi,i= 1,2.Here,thearrow“→”denotesaboundedoperator. Then,

K(Lφ, s; Y1(S), Y2(S)) ≤ max{N1, N2}K(φ, s; X1(R), X2(R)) for s > 0, (5.13) foreveryφ∈ X1(R)+ X2(R).

Lemma 5.3.LetΩ beanopensetinRn,n≥ 2,andletγ∈ (0,n). Assumethatμ isany

Borel measurein Ω fulfilling (5.1)for someα∈ (n− γ,n] and forsome Cμ> 0. Then

there existsaconstant C = C(n,α,γ,Cμ) such that

Nγg L

α

n−γ,∞(Ω,μ)≤ CgL1(∂Ω) (5.14)

forevery g∈ L1(∂Ω).

Proof. Wemakeuseofanargumentrelatedto[1,2].Giveng∈ L1(∂Ω) andt> 0,define

Et={x ∈ Ω:|Nγg(x)|> t}, anddenote byμt therestriction ofthe measure μ toEt. ByFubini’stheorem, tμ(Lt) = t  Ω dμt(x)≤  Ω |Nγg(x)|dμt(x)≤  Ω  ∂Ω |g(y)| |x − y|n−γdHn−1(y) dμt(x)  ∂Ω |g(y)| Ω dμt(x)dHn−1(y) |x − y|n−γ . (5.15) Next,  Ω dμt(x) |x − y|n−γ = (n− γ)  0 −n+γ−1  {x∈Ω:|x−y|γ−n>γ−n} dμt(x) d ≤ (n − γ)  0 −n+γ−1μt(B(y)) d. (5.16)

From (5.15)and(5.16) wededucethat,foreachfixed r > 0,

tμ(Lt)≤ (n − γ)  ∂Ω |g(y)|  0 −n+γ−1μt(B(y)) d dHn−1(y) = (n− γ) r  0 −n+γ−1  ∂Ω

(28)

+ (n− γ)  r −n+γ−1  ∂Ω

|g(y)|μt(B(y)) dHn−1(y) d. (5.17) Wehavethat r  0 −n+γ−1  ∂Ω

|g(y)|μt(B(y)) dHn−1(y) d

≤ CμgL1(∂Ω) r  0 −n+γ−1+αd = Cμr α−n+γ α−n+γgL1(∂Ω). (5.18)

Ontheotherhand,  r −n+γ−1  ∂Ω

|g(y)|μt(B(y)) dHn−1(y) d≤ μ(Lt)  r −n+γ−1  ∂Ω |g(y)| dHn−1(y) d = μ(Lt)r −n+γ n−γ gL1(∂Ω). (5.19)

Combining(5.17)–(5.19), andchoosingr =μ(Lt)

1 α ,yield tμ(Lt)≤ (n − γ)gL1(∂Ω)  rα−n+γ α− n + γ + μ(Lt) r−n+γ n− γ = α−n+γα gL1(∂Ω)C n−γ α μ μ(Lt)1− n−γ α . (5.20) Thus, tμ(Lt) n−γ α α α−n+γC n−γ α μ gL1(∂Ω) for t > 0. (5.21)

Hence,inequality(5.14) follows. 2

Proposition5.4. LetΩ bean openset inRn.Then  Sn−1 |g(ζ(x, ϑ))| dHn−1(ϑ)≤ 2n  ∂Ω |g(y)| |x − y|n−1dHn−1(y) for x∈ Ω, (5.22)

foreveryBorelfunctiong : ∂Ω→ R.Here, convention(3.7) isadopted.

Proof. Wesplittheproofinsteps.Fixx∈ Ω.

Step1.DenotebyΠ:Rn\ {x}→ Sn−1 theprojectionfunctioninto Sn−1 givenby Π(y) = y− x

(29)

Then

Hn−1(Π(E)) 1

dist(x, E)n−1H

n−1(E) (5.23)

foreveryset E⊂ ∂Ω.

ThefunctionΠ isdifferentiable,and|∇Π(y)|≤ |y − x|−1 fory∈ Rn\ {x}. Thus,the restrictionofΠ toE isLipschitzcontinuous,and

|∇Π(y)| ≤ 1

dist(x, E) for y∈ E. (5.24) Inequality (5.24) implies (5.23), by a standard property of Hausdorff measure – see e.g.[59,Theorem 7.5].

Step 2.Wehavethat

Hn−1(Π(E))≤ 2n  E

dHn−1(y)

|x − y|n−1 for every Borel set E⊂ ∂Ω. (5.25) Thefollowing chainholds:

 E dHn−1(y) |x − y|n−1 =  0 Hn−1({y ∈ E : |x − y|−n+1> t}) dt =  0 Hn−1({y ∈ E : |x − y| < τ}) d(−τ1−n) = k∈Z 2k+1  2k Hn−1({y ∈ E : |x − y| < τ}) d(−τ1−n) k∈Z Hn−1({y ∈ E : |x − y| < 2k}) 2k+1  2k d(−τ1−n) = (1− 2−n+1) k∈Z 2−k(n−1)Hn−1({y ∈ E : |x − y| < 2k}) ≥ (1 − 2−n+1) k∈Z 2−k(n−1)Hn−1({y ∈ E : 2k−1≤ |x − y| < 2k}). (5.26) Since dist(x,{y ∈ E : 2k−1 ≤ |x− y|< 2k})≥ 2k−1,by(5.23) Hn−1({y ∈ E : 2k−1≤ |x − y| < 2k}) ≥ C2(k−1)(n−1)Hn−1(Π({y ∈ E : 2k−1≤ |x − y| < 2k})) (5.27)

(30)

fork∈ Z. From(5.26)and(5.27) wededucethat  E dHn−1(y) |x − y|n−1 12 k∈Z 2−k(n−1)2(k−1)(n−1)Hn−1(Π({y ∈ E : 2k−1≤ |x − y| < 2k})) = 2−n k∈Z Hn−1(Π({y ∈ E : 2k−1 ≤ |x − y| < 2k})) ≥ 2−nHn−1 k∈ZΠ({y ∈ E : 2k−1≤ |x − y| < 2k}) = 2−nHn−1Π(∪k∈Z{y ∈ E : 2k−1≤ |x − y| < 2k}) = 2−nHn−1(Π(E)), (5.28)

whenceinequality(5.25)follows.

Step3.Conclusion. Wehavethat Π({y ∈ (∂Ω)x:|g(y)| > t}) = {ϑ ∈ Sn−1:|g(ζ(x, ϑ))| > t} for t > 0. Thus,by(5.25),  {y∈(∂Ω)x:|g(y)|>t} dHn−1(y) |x − y|n−1 ≥ 2−nHn−1({ϑ ∈ Sn−1:|g(ζ(x, ϑ))| > t}) for t > 0. Hence,  (∂Ω)x |g(y)| |x − y|n−1dHn−1(y) =  0  {y∈(∂Ω)x:|g(y)|>t} dHn−1(y) |x − y|n−1 dt ≥ 2−n  0 Hn−1({ϑ ∈ Sn−1:|g(ζ(x, ϑ))| > t}) dt = 2−n  Sn−1 |g(ζ(x, ϑ))| dHn−1(ϑ). (5.29) Inequality(5.22)isthusestablished. 2

Lemma 5.5. Let Ω be an open set in Rn, n ≥ 2. Assume that μ is any Borel measure

in Ω fulfilling(5.1) for some α∈ (n− 1,n] and forsome > 0. Then there exists a

constant C = C(n,α,Cμ) suchthat

(Tg)∗μ(s)≤ Cs−n−1α

sn−1α

0

g∗Hn−1(r) dr for s > 0, (5.30)

(31)

Proof. ByProposition 5.4,there existsaconstantC = C(n) suchthat

Tg(x)≤ CN1|g|(x) for x ∈ Ω, (5.31) for everyBorelfunction g : ∂Ω→ R. Hence,byLemma 5.3, with γ = 1,there exists a constantC = C(n,α,Cμ) suchthat

T gL α

n−1,∞(Ω,μ)≤ CgL1(∂Ω) (5.32) forevery Borel functiong : ∂Ω→ R.

Ontheother hand,

0≤ T g(x) ≤ gL∞(∂Ω)  Sn−1 dHn−1(ϑ) = nωngL∞(∂Ω) for x∈ Ω, (5.33) and hence T gL∞(Ω,μ)≤ nωngL∞(∂Ω) (5.34) for everyBorel functiong : ∂Ω→ R. We thusdeduce from (5.9), (5.10), (5.13), (5.32)

and (5.34)that s(Tg)∗μ(snα−1)≤ rn−1α (Tg)∗ μ(r)L(0,sn−1α )≈ K(Tg, s; L α n−1,∞(Ω, μ), L(Ω, μ)) ≤ CK(g, s; L1(∂Ω), L(∂Ω)) = C s  0 gH∗n−1(r) dr for s > 0, (5.35)

for some constant C = C(n,α,Cμ), and for every Borel function g : ∂Ω → R. Hence, inequality(5.30)follows. 2

Lemma 5.6.LetΩ beanopensetinRn,n≥ 2,andletγ∈ (0,n). Assumethatμ isany

Borel measureinΩ fulfilling (5.1)forsome α∈ (n− γ,n] and forsomeCμ> 0. Then,

there existsaconstant C = C(n,γ,α,Cμ) such that

(Iγf )∗μ(s)≤ C  s−n−γα snα  0 fL∗n(r) dr +  sαn r−n−γn f∗ Ln(r) dr  for s > 0, (5.36) forevery f ∈ L1(Ω).

Proof. Astandardweak-type inequalityforRiesz potentialstellsus thatthereexists a constantC1= C1(n,γ,α,Cμ) suchthat

Riferimenti

Documenti correlati

such systems are attracting large attention from both Academia and Industry. As a result, DBRs made of both polymers and inorganic mesoporous materials are the first photonic

27-28: «Es sprechen sich in dieser einfachen Schilderung der Landschaft und des Waldlebens Gefühle aus, welche sich mit denen der modernen Zeit inni- ger verschmelzen als alles, was

E’ il tema della TRASVERSALITA’ VERTICALE, che intenderemmo esplorare interrogandoci su cosa significa promuovere competenze nella scuola di base;  Come sviluppare un percorso

La virtù non è una emozione o una capacità, ma una abituale disposizione ad agire bene; essa prende forma sotto la guida continua e critica della ragione (Mortari, 2014, pp.

In RUNT KO melanoma cells, reduced expression of genes involved in metastatic processes (e.g., CD31, MMP9 and VEGFA) was observed, and genes involved in promoting bone metastasis,

Nowadays, estrogen defi ciency (severe or mild) should be considered one of the key steps in the promotion of bone loss in men (Riggs et al 1998) as well as a major determinant in

Copyright Pacini Editore Srl 2021 - Tutti i diritti riservati Estratto per gli Autori a uso esclusivo di valutazione scientifica.... 3 Copyright Pacini Editore Srl 2021 - Tutti

fatti i tuberi di Rossa di Cetica rappresentano un prodotto molto apprezzato sul mercato locale che agisce anche come motore propulsivo per la commercializzazione di altri