Contents lists available atScienceDirect
Advances in Mathematics
www.elsevier.com/locate/aim
Weak differentiability for fractional maximal functions of general L
pfunctions on domains
JoãoP.G. Ramosa, Olli Saaria,∗, Julian Weigtb
aMathematicalInstitute,UniversityofBonn,EndenicherAllee60,53115,Bonn, Germany
bDepartmentofMathematicsandSystemsAnalysis,AaltoUniversity,Finland
a r t i c l e i n f o a b s t r a c t
Articlehistory:
Received24November2019 Receivedinrevisedform26March 2020
Accepted1April2020 Availableonlinexxxx CommunicatedbyLarsOlsen
MSC:
primary42B25,46E35
Keywords:
Maximalfunction Sobolevspace Sphericalmeans Domains
LetΩ⊂Rn beboundedadomain.Weproveundercertain structuralassumptionsthatthefractionalmaximaloperator relativetoΩ mapsLp(Ω)→W1,p(Ω) forallp>1,whenthe smoothnessindexα≥1.Inparticular,theresultsarevalidin therangep∈(1,n/(n−1)] thatwaspreviouslyunknown.As anapplication,weproveanendpointregularityresultinthe domainsetting.
©2020ElsevierInc.Allrightsreserved.
1. Introduction
RegularityoftheHardy–LittlewoodmaximalfunctionofaSobolevfunctionwasfirst studiedin[13].It wasshownthatthemaximaloperatorpreserves W1,p(Rn) regularity for p > 1. This continues to hold true at the derivative level when p = 1 and n = 1
* Correspondingauthor.
E-mailaddresses:jpgramos@math.uni-bonn.de(J.P.G. Ramos),saari@math.uni-bonn.de(O. Saari), julian.weigt@aalto.fi(J. Weigt).
https://doi.org/10.1016/j.aim.2020.107144 0001-8708/©2020ElsevierInc. Allrightsreserved.
[21,16] andforradialfunctions[18].ExtendingsuchastatementtomoregeneralSobolev functionsofseveralvariablesisadifficultopenproblem,whichhasinspiredmanyresults in relatedtopics. Forinstance, slightlystrongerbounds havebeen proved formaximal operators withmorespecialconvolution kernels(see[7],[3],[4] and[20]),thecontinuity of the mappinghas been studiedin[17] and [6], and apartof thetechniques used for continuity,alsorelevantforthecurrentpaper,havebeenextendedto p= 1 in[11].
Another aspect of the problem is the fractional endpoint question proposed by CarneiroandMadrid[5].Thefractionalmaximal functionisgiven by
Mαf(x) = sup
r>0
rα
|B(x, r)| ˆ
B(x,r)
|f(y)|dy,
anditdefinesaboundedoperatorLp(Rn)→Lq(Rn) whenq=np/(n−αp),0< α < n/p and p > 1. This boundedness fails at the endpoint p = 1, but the question about boundednessof∇Mα fromW1,1(Rn) toLn/(n−α)(Rn) hasnotbeenansweredso farfor α < 1 (see [19], [1] and [2] for related research and partial results). The case α ≥ 1 turned outtobeverysimple,andthereasoncanbetracedbacktotheinequality
|∇Mαf(x)| ≤cα,nMα−1f(x) (1.1) of Kinnunenand Saksman[15].CarneiroandMadrid[5] notedthat(1.1) togetherwith theGagliardo–Sobolev–NirenberginequalityandtheLp→Lq boundsforthefractional maximal functionimplytheexpected endpointbound whenα≥1.
In thepresent paper,we studythese problems ingeneralopen subsetsof Rn,which is anaturalcontext for analysis from thepoint of viewof potential theoryand partial differential equations. Regularity ofthe local Hardy–Littlewood maximal functionof a SobolevfunctiononanopenΩ⊂Rn wasfirststudiedbyKinnunenandLindqvist[14], andalocalvariantoftheinequalityforthederivativeofthefractionalmaximalfunction (1.1) wasprovedin[12]. Thisisourstartingpoint,and formorethoroughdiscussionof whatwasprovedandwhatisunknown, weintroducesomemorenotation.
IfΩ⊂Rn isanopenset,thelocal fractionalmaximalfunctionisdefinedas MαΩf(x) = sup
0<r<dist(x,Ωc)
rα
|B(x, r)| ˆ
B(x,r)
f(y)dy.
As theboundaryofΩ restrictsthechoiceofrinthedefinition,onecannotexpect (1.1) to triviallycarryovertothelocalsetting. Indeed,suchapointwiseinequalityisfalsein general(Example4.1in[12]).Ontheotherhand,ifoneaddsacorrectionterminvolving thesurfacemeasureofthespheretotheright handsideof(1.1),oneobtains
|∇MαΩf(x)| ≤cα,n
Mα−1f(x) + sup
r>0|rα−1σr∗f(x)|
, (1.2)
which is valid in all domains. This was used in [12] to prove that Lp functions with p> n/(n−1) large enoughhaveMαΩf inafirstorder Sobolev class.The lower bound onprules outfunctionstoo singularforanapplicationofasphericalmaximalfunction argument.
Our main theorem shows that under suitable assumptions on the domain Ω, the maximalfunctionMαΩmaps Lp(Ω) intoafirstorderSobolevspaceforallp>1.
Theorem1.1. Let Ω⊂Rn be open,n≥2, p>1and f ∈Lp(Ω). ThenMαΩf is weakly differentiableand
∇MαΩf Lp(Ω)≤C f Lp(Ω)
ifany one ofthefollowingholds:
(1) α >1andΩisbounded.
(2) α= 1 andΩc is convex.
(3) α = 1 and Ω is bounded and satisfies a uniform curvature bound in the sense of Section 2.2.
(4) α= 1 andp>1+n1.
Theconstant C dependsonthedimension, andin(1)and(3)italso dependsonαand thedomain.
Unlike[12],weare notabletoproveanLp→Lq smoothingeffectontopofwinning onederivative.However,ourmethoddoesapplytosingularfunctionsinLp spaceswith 1≤p≤n/(n−1) wherethe argumentin[12] failstogiveany result.Inparticular, we havethefollowing endpointregularity result,whichwaspreviouslyoutofreach.
Corollary1.2. LetΩ⊂Rn be aLipschitzdomain. Thenforallf ∈W1,1(Ω) ∇M1Ωf Ln/(n−1)(Ω)≤C f W1,1(Ω)
wheretheconstant C only dependsonΩandthedimension.
Webrieflyoutlinetheproofofthemaintheorem.Themaximalfunctiononadomain behavesdifferently depending onwhether theball attaining themaximum touchesthe boundaryornot.Incaseitdoesnot,thelocalmaximalfunctionbehavesliketheglobal one, and the analysis is very similar. Otherwise it coincides with a linear averaging operator(2.3),whichdependsonthedomain.These twoparts areanalyzedseparately, andthemainpartoftheproofistoestablishLpboundsforthederivativeof(2.3).This leadsto studyingadomaindependentweightedsphericalaveragingoperator(5.1).
InsteadofresortingtomaximalaveragesandtheBourgain–Steintheorem,anangular decompositionoftheoperatoriscarriedout.Theadditionalgeometricinformationallows
Fig. 1.A setP(y) and a tangent line.
instead to establishgood L1 bounds that canbe interpolated with trivialL∞ bounds inorder toobtainadomination of(5.1) byaconvergingsumofLpbounded operators.
ImprovingtheL1boundoverwhatfollowsfromthebehavior ofgenericsphericalmeans is crucial when aiming atLp bounds forall p>1.Such aconclusioncannot be drawn from merepolynomialdecayoftheFouriertransformoftheweightedsphericalmeasure inquestion,ifnoadditionalL1informationistakenintoaccount.Turningthefocusfrom the Littlewood–PaleydecompositionandL2methodsto anangulardecompositionand geometric estimatesinL1istheleadinginsightoftheproof.
Thekey ideaintheL1 estimatescanbedescribed asfollows. Each domainΩ comes endowedwith afamilyofsets(Fig. 1)
{P(y) :y∈Ω}, P(y) ={x∈Ω :y∈∂B(x,dist(x,Ωc))},
which can morallybe used to dualize the spherical averaging operators (5.1) through Fubini’s theorem.The L1 boundsfor the constituentsinthe angular decomposition of thesphericalaveragingoperatorcorrespondtoweightedintegralsoverthepiecesofP(y).
IfΩ isaball,thenthesetsP(y) areellipsoidswithfociatthecenteroftheballandaty.
Inthecasesofthecomplementofaballandahalf-space,theP(y) takethesimpleforms ofhyperboloidsandparaboloids.Onecannothopeforas explicitdescriptionsasthatin more general domains, butall P(y) are boundaries of convex sets. This observationis used extensivelyintheproof.
The structure of the paper is as follows. In the first section, we introduce notation and sometools thatwill be helpfulthroughout theproof. The first sections are about differentiating themaximalfunctionon so-calledunconstrainedpoints andproving the
weakdifferentiabilityofthemaximalfunctionconditionallytotheLpboundednessofthe averagingoperator(2.3).TherestofthepaperisdevotedtoprovingthoseLpboundsby firstcomputingaformulaforthederivativeandthencarryingoutthestrategysketched above.Finally,thereisaconcludingsectionwithremarksonopenproblemsandcertain observationsabouttheproofwhichmightbeofindependentinterestforfutureresearch.
2. Preliminaries
2.1. Notation
We letn ≥1 denote the dimension. For ameasurable set E, we let |E| denote the n-dimensionalLebesguemeasure.Thek-dimensionalHausdorffmeasuresaredenotedby Hk. An Euclidean ball with centerx ∈Rn and radius r > 0 is denoted byB(x,r).A finiteconstantonlydependingonquantitiesthatarenotbeing kepttrackofisdenoted by C. If A ≤ CB for such constant, we denote A B or write A is B. We write A∼B ifbothAB andBAhold.
2.2. Domains
WealwaysassumeΩ⊂Rn to beanopen set,whichwe interchangeablycalldomain asthedistinctionobviouslyplaysnoroleinthispaper.Weassumeittohavenon-empty complement. The distance function is denoted by δ(x) = dist(x,Ωc). As Ωc is closed, thereexists atleast onebx∈Ωc so that|x−bx|=δ(x).Wereserve thenotationbx for suchapoint,whichneednotbeuniqueunlessΩcisconvex.Thedistancefunctionδ: Ω→ [0,∞) is always 1-Lipschitz. The gradient exists almost everywhere by Rademacher’s theorem,and itholdsthat
∇δ(x) =x−bx
δ(x) . (2.1)
This is because clearly the one sided directional derivative of δ(x) in the direction of bx−x always exists and is −1. Where the gradientexists, we canuse |∇δ(x)|≤ 1 to concludethat the directional derivativein all directions orthogonal to x−bx must be zero.
AdomainissaidtosatisfyauniformcurvatureboundifthereisanR >0 sothatfor everypoint x∈Ω itholds
B
bx+Rx−bx
δ(x) , R
⊂Ω.
AllboundedC2 domains satisfythis condition,butadomainsatisfying auniform cur- vature bound might be non-smooth and have inwards-pointing cusps. A domain with aninteriorballconditionneednotsatisfytheuniformcurvature bound.Anexampleof suchadomain isB(0,1)\({0}∪ {(0,. . . ,0,2k):k∈Z,k <0}).
2.3. Functionspaceson domains
Functionsf ∈Lp(Ω) areapriorionlydefinedinthedomainΩ,butwealwaysextend them by zeroto Rn withoutadditional comments.TheSobolev classW1,p(Ω) consists of functionsf ∈Lp(Ω) suchthat|∇f|∈Lp(Ω).Theweakderivativesaredefinedusing test functionsinCc∞(Ω).
Fortheapplicationofthemaintheoremtotheendpointregularityproblem,weneed aSobolevembedding theoremfor domains.Oneconcrete casewecandealwith isthat of aLipschitzdomain.
Proposition 2.1 (Section 4.4 in [9]).Let Ω ⊂ Rn be a bounded open set so that ∂Ω is Lipschitz. Thenforevery1≤p<∞ thereexistsaboundedextension operator
E:W1,p(Ω)→W1,p(Rn)
such thatsupp(Ef)⊂B(x0,2diam(Ω))forsomex0∈Ωandallf ∈W1,p(Ω).
BytheboundarybeingLipschitz,wemeanthatitcanbecoveredbyafinitenumber of open balls Bi so that for each i the domain Bi∩Ω is the epigraph of a Lipschitz function.
The propositiontogether with the Gagliardo–Nirenberg–Sobolevinequality(see e.g.
Section4.5.1in[9])impliesarudimentarylocal Sobolevembedding
f Lpn/(n−p)(Ω)≤CΩ,p,n f W1,p(Ω) (2.2)
valid for all f ∈ W1,p(Ω) whenever Ω is abounded open set with Lipschitz boundary.
This issufficientforourpurposes.
2.4. Maximal function
Forα∈[1,n),definethelocalfractionalmaximal functionrelative toΩ as MαΩf(x) = sup
0<r<δ(x)
rα − ˆ
B(x,r)
f(y)dy
wheneverf ∈L1loc(Ω).WeomitthesuperscriptwhenΩ isthewholeRn.Inaddition,we define forα∈Rtheauxiliarylinearoperator
Aαf(x) =δ(x)α − ˆ
B(z,δ(x))
f(y)dy. (2.3)
2.5. Constrainedpoints
Letfbecontinuous.Fixx∈Ω.BecausethecomplementofΩ isnon-empty,δ(x)<∞ andthereexistsaconvergentsequencerj∈(0,δ(x)) withlimitr= limj→∞rj∈[0,δ(x)]
suchthat
MαΩf(x) = lim
j→∞rαj ˆ−
B(x,rj)
f(y)dy=rα ˆ−
B(x,r)
f(y)dy
ifr >0.If
MαΩf(x)> δ(x)α − ˆ
B(x,δ(x))
f(y)dy,
thesequence rj mustbe chosen so thatr < δ(x), and thepoint xissaid to be uncon- strained.Allotherpointsarecalled constrained.
3. Theunconstrainedpart
Thelocalmaximal functionbehavessimilarlyto theglobaloneintheunconstrained set,and we reduce the differentiability questionof the unconstrained partaccordingly tothatoftheglobal maximalfunction.Thisisthecontentofthefollowingproposition.
Proposition 3.1. Let p > 1, α ≥ 1 and f ∈ Lploc(Ω) be continuous. The set U of the unconstrained pointsis open,the maximal function MαΩf is weaklydifferentiable in U, andthepointwisebound
|∇MαΩf(x)| ≤cMα−1f(x)
holdsforaconstant c onlydepending on thedimensionandαwheneverx∈U. Proof. Considerthefractionalaveragefunction
A(z, r) :=rα ˆ−
B(z,r)
f(y)dy.
Itiscontinuousin(z,r)∈Ω×R+.Fixnowanunconstrainedpointx.Bydefinition,there exists ε> 0 sothatMαΩf(x)−A(x,δ(x)) > . Moreover, there exists γ >0 so thatif
|(z,r)−(x,δ(x))|< γ,thenMαΩf(x)−A(z,r)> ε/2.SinceMαΩf islowersemicontinuous, onecanfindforeveryz closeenoughtoxasequencerz,j→rz< δ(x)−γ/2 sothat
MαΩf(z) = lim
j→∞rαz,j ˆ−
B(z,rz,j)
f(y)dy.
Inparticular, thereisanopenneighborhood Uxofxsothatforallz∈Ux
MαΩf(z) =Mα(1B(x,δ(x))f)(z).
ByTheorem3.1 in[15],
|∇MαΩf(x)| ≤CMα−1f(x) follows.
4. Thefull maximalfunction
Next we provethe differentiability of the local maximal function conditional to Lp boundsforthederivativeoftheaveragingoperator(2.3).Thisstepmorallyfollowsfrom thelatticeproperty ofSobolevfunctions,butasweonlyknowtheweakdifferentiability of MαΩf intheunconstrainedset,someextraworkisneeded.
Lemma 4.1.Let p >1, α≥ 1 and Ω⊂ Rn be such that ∇Aα and Mα−1 are bounded Lp(Ω) → Lp(Ω). If f ∈ Lp(Ω), then the local fractional maximal function is weakly differentiable and
∇MαΩf Lp(Ω) f Lp(Ω).
Proof. Assumefirstthatf iscontinuous andcompactlysupported.Following theargu- ments in[15], we inferthat MαΩf canbe seenas supremum over radii betweenafixed upperandlowerbound.ThefractionalaveragesareLipschitzcontinuouswithconstants only dependingontheradii,and hencetheirsupremumisalso Lipschitz.Inparticular, we knowthatMαΩf iscontinuous.
Denote byg+ = max(g,0) thepositivepartofafunctiong andwrite MαΩf = (MαΩf−Aαf)++Aαf.
By assumption,the second termadmits the desired Sobolevbounds. To deal with the other term,let>0 anddefine
F(t) =
((t−)2+2)1/2−, t >
0, t≤.
These functions are of class C1(R) and converge pointwise to t → (t)+ as → 0.
Moreover, as MαΩf and Aαf arecontinuous,E ={x∈Ω:F(MαΩf(x)−Aαf(x))>0} hasitsclosurecontainedintheopensetofunconstrainedpointsU.ByProposition3.1, theassumptiononAαandthechainruleforSobolevderivatives(4.2.2in[9]),weobtain forallpartial derivatives∂i
∂iF(MαΩf−Aαf) = (∂iMαΩf−∂iAα)F(MαΩf −Aαf).
Takingatest functionϕandcomputing ˆ
Ω
F(MαΩf−Aαf)∂iϕ dx= ˆ
Ω
(∂iMαΩf−∂iAαf)F(MαΩf−Aαf)ϕ dx,
we see that taking the limit → 0 proves the claim for continuous and compactly supportedf.
Todealwiththegeneralcase,letf ∈Lp(Ω) andletfj becontinuous andcompactly supported functions converging to f in Lp norm. By Lp continuity of the fractional maximaloperator,MαΩfj→MαΩf inLp.As wehaveprovedthefollowinginequality
∇MαΩfj Lp(Ω) fj Lp(Ω),
forcontinuousfunctionsfj,thesequenceMαΩfj isboundedinW1,p(Ω).Wecanextract a weaklyconvergent subsequence. By taking limits along this sequence and using the uniquenessof distributionallimit,weconcludetheproofforgeneralf ∈Lp(Ω).
As the main theorem is a direct consequence of the previous lemma, it remains to investigatetheboundedness oftheoperator∇Aα onLp(Ω).The followingsections are devotedtoestablishingtherequiredLp boundswhenΩ issufficientlywell-behaved.
5. Constrainedpart
Byachangeofvariables, wecanwrite theaveragingoperator(2.3) as
Aαf(x) =δ(x)α ˆ−
B(0,1)
f(x+yδ(x))dy.
This operatoris linear, and as we areaiming for Lp bounds, there is no loss of gener- ality inrestricting theattention to smooth functions.If xis aconstrained point, then MαΩf(x) =Aαf(x), which justifies our reference to Aα as the constrained part. Also, Lemma4.1showedthatLp boundsforthederivativeof Aαf areenoughtoimplyweak differentiabilityofthefullmaximaloperator,sothemaximalfunctiondoesnotplayany role inwhatfollows. A versionof thefollowing proposition wasalready provedin[12], butasweneedaformulamoreprecisethanwhattheystated,weincludetheshortproof forclarity.
Proposition 5.1. Letf ∈C∞(Ω).Thenforalmosteveryx∈Ω
|∇Aαf(x)| ≤cn,α|Aα−1f(x)|+cnδ(x)α−1 − ˆ
∂B(x,δ(x))
|y−bx|
δ(x) f(y)dHn−1(y) where bx∈∂Ωisapointsuchthat |bx−x|=δ(x).
Proof. Fixapointx.AsAαf(x)=δ(x)αA0f(x),itholdsthat
∇Aαf(x) =αδ(x)α−1A0f(x)∇δ(x) +δ(x)α(∇A0f)(x).
Since|∇δ(x)|≤1 (cf. (2.1)),thefirstsummandaboveisboundedbyAα−1f(x).Thusit sufficesto analyzethegradientofA0f.Take theunitvector
e=∇A0f(x)/|∇A0f(x)|. Then
|∇(A0f)(x)|= (e· ∇)A0f(x)
= −
ˆ
B(0,1)
e+y(e· ∇δ(x))
· ∇f(x+δ(x)y)dy
= 1
δ(x) − ˆ
B(0,1)
divy
(e+y(e· ∇δ(x)))f(x+δ(x)y) dy
−ne· ∇δ(x) δ(x) ˆ−
B(0,1)
f(x+δ(x)y)dy=: I + II.
Since |∇δ(x)| ≤ 1, the contribution δ(x)α·II is pointwise bounded by nAα−1f. To estimate theother term,weapplyGauss’stheoremtoobtain
I = cn
δ(x) ˆ
∂B(0,1)
y·(e+y(e· ∇δ(x)))f(x+δ(x)y)dHn−1(y)
= cn
δ(x) ˆ−
∂B(x,δ(x))
(y−bx)·e δ(x) f(y)dy.
Sowereachtheinequality
|∇Aαf(x)| ≤αn|Aα−1(x)|+cnδ(x)α−1 ˆ−
∂B(x,δ(x))
|y−bx|
δ(x) f(y)dHn−1(y), whichproves theclaim.
Because Aα−1f(x) ≤MαΩ−1f(x),and MαΩ−1 satisfies the right Lp →Lq bounds, we havereducedthematterto understandingtheweightedsphericalaverage
Bαf(x) :=δ(x)α−1 − ˆ
B(x,δ(x))
|y−bx|
δ(x) f(y)dHn−1(y) (5.1) on the right hand side of the conclusion of the previousproposition. The weight|y− bx|/δ(x) measurestheangle betweenbx−xand y−xwhen |y−bx|/δ(x) is small. We decomposetheweightedsphericalaveragingoperatoraccordingtotheangleandlocation inthedomain asfollows.Fork∈Z,let
Ωk={x∈Ω : 2k ≤δ(x)<2k+1} andforeverypointx∈Ω andintegerj≥0
ωj(x) = y ∈∂B(x, δ(x)) : 2−j < |y−bx|
δ(x) ≤2−j+1
.
Define
Sjkf(x) = 1Ωk(x) ˆ
ωj(x)
f(y)dHn−1(y).
Then
Bαf(x)
k∈Z
∞ j=1
2k(α−n)−jSjkf(x) (5.2)
anditremains toproveboundsfor Sjk so thattherighthandsidesumsupinLp.This isdonebyinterpolatingboundsonL∞ andL1.
Proposition 5.2. Let Ω be any domain. It holds that Sjk L∞→L∞ 2(n−1)(k−j), and consequently
k2k(1−n)Sjk L∞→L∞ 2−(n−1)j. Proof. Thisfollows fromHn−1(ωj(x))2(n−1)(k−j). 6. L1 bounds
To proveL1 bounds, we introducesome morenotation. For eachinteger j ≥0 and eachpointy∈Ω,define
Pj(y) ={x∈Ω :y∈ωj(x)}, P(y) = ∞ j=0
Pj(y). (6.1)
Inaddition, let
Akj =
x∈Ωk
ωj(x). (6.2)
Formally, certain weighted integrals over P(y) give the adjoint operator of Bα. A naive changeoforderofintegrationisnotjustifiedinthis case,butusingthedecompo- sitionofBα,wecanmaketheideaprecise.Thefollowingtwo propositionsgiveeffective descriptionofP(y) andprovideasubstituteforFubini’stheorem.
Proposition 6.1. LetΩ bean opensetandlety∈Ω.Then E(y) ={x∈Ω :|x−y| ≤δ(x)} is closedandconvex setsuchthat
P(y) =∂E(y).
For each x∈P(y), thesupporting hyperplaneat xbisects theangle betweeny−x and bx−xandisnormal tobx−y.
Proof. RecallthatP(y) consistsofthepointswith{x∈Ω:|y−x|=δ(x)}.Forx∈P(y), itholds that
x+ bx−x
|bx−x| ∈E(y)c, and itiseasytosee∂E(y)=P(y).Considerthehyperplane
{z∈Rn :|z−bx|=|z−y|}.
It dividesthe space into two half spaces H1 ={z : |z−bx| <|z−y|} and H2 ={z :
|z−bx|≥ |z−y|}.Ifx∈P(y),then E(y)⊂H2and x∈H2.Thus∂H2 isasupporting hyperplaneforE(y) atx.AseveryboundarypointofE(y) hasasupportinghyperplane, E(y) isconvex.Theremainingassertionsreadilyfollowfromthedefinitionof∂H2. Proposition6.2.LetΩbeadomain,j≥0andkintegersandf ≥0aboundedcontinuous function onΩ.Then
ˆ
Ω
Sjkf(x)dx2j
|j−j|≤1,|k−k|≤1
ˆ
Akj
f(y)Hn−1(Pjk(y))dy,
where weletPjk(y)=Pj(y)∩Ωk.
Fig. 2.The construction to findx1.
Notethaty∈Akj ifandonlyifPjk(y)=∅.
Proof. The parameterk plays norole inthe following computation,butis includedin thestatementforfuturereference.Letϕ≥0 beasmoothfunctionofonevariable with compactsupportin(0,1) and ϕ L1(R)= 1.Denotethe-dilationbyϕ(t)=−1ϕ(t−1).
Foranyfixedx,wedefine theset ofrelevantdirections
ωjdir(x) =δ(x)−1(ωj(x)−x)⊂∂B(0,1).
Asf ispositive,theweakconvergence Sjkf(x) =
ˆ
ωj(x)
f(y)dHn−1(y) = lim
→0
ˆ
x+Rωdirj (x)
f(y)ϕ(δ(x)− |x−y|)dy
holds.Integratingoverx,applyingthedominatedconvergencetheorem(thisisjustified, seetheremarkat theend oftheargument),andusingFubini’stheorem,weobtain
ˆ
Ωk
Sjf(x)dx ˆ
Akj
f(y)
→0lim
|{x∈Ωk :y∈ω−j (x)}|
dy (6.3)
wheretheone-sidedneighborhoodisdefinedas(see Fig.2) ωj−(x) =x+ωjdir(x)(δ(x)−, δ(x)).
Nextweestimatethelimitexpressionin(6.3).Asj andkarefixed,wecanassume tobe verysmallrelativetothem.Letx∈Ωk.Assumethaty∈ωj−(x).Then
− <|y−x| −δ(x)<0 (6.4) and bydefinitionxbelongstotheinteriorofE(y).
Set
e= bx−x
|bx−x|
and letr ∈(0,δ(x)) be such thatx+re ∈P(y).Next we give anupper bound for r.
Because y∈ωj−(x),italso holdsthat y−x
|y−x| ∈ωjdir(x).
Themapping
g(ρ) :=|y−(x+ρe)| −δ(x+ρe) =|y−x−ρe| −δ(x) +ρ is Lipschitzandhenceabsolutely continuous.
Forallρ≥0 wehavethelowerbound
g(ρ) =∂ρ[|y−(x+ρe)| −δ(x+ρe)] =−e· y−x−ρe
|y−x−ρe| + 1
= 1−cos(bx−x, y−x−ρe)
≥1−cos(bx−x, y−x) 2−2j
Thelast inequalityis duetoy∈ωj−(x).Recallthatg(0)≥ −andg(r)= 0.Since gis absolutely continuous,weconclude
2−2jr ˆr
0
g(s)ds=g(r)−g(0)≤,
and
r22j.
Denote x0 = x+re ∈ P(y). Consider the 2-planecontaining x, y, bx (and x0). Its intersection with theconvex body E(y) provided by Proposition6.1 is again aconvex set E in the plane. Let be its supporting line at x0. Then x0 ∈ Pjk(y) for some j∈ {j,j−1}, k∈ {k,k−1}because
(bx−x0, y−x0)≥(bx−x, y−x)≥2−j sin(bx−x0, y−x0)≤ |bx−y|
δ(x)−C22j = |bx−y|δ(x)−1
1−Cδ(x)−122j ≤sin 2−j+2
forsmallenough.ByProposition6.1thisalsomeansthaty−x0makesanangle∼2−j with , and hence so does x−x0. Let e be the unit vector perpendicular to and e·(y−x)<0.Thenthere is
s|x−x0|sin 2−j2j
sothatx+se∈.Sincex∈E(y) andintersectsE(y) onlyin∂E(y),thereiss < s withx1=x+se ∈∂E(y),whichmeans
dist(x, P(y))2j. (6.5)
Recallthat
dist(x, Pjk(y))≤ |x−x0|22j. Let
N() =
j∈{j−1,j},k∈{k−1,k}
{x∈P(y) : dist(x, Pjk(y))≤}.
Then
lim→0
|{x∈Ωk :y∈ωj−(x)≤}|
≤ lim
→0lim
→0
|{x∈Ωk : dist(x, P(y))≤cn2j} ∩N()|
lim
→02jHn−1(P(y)∩N())
≤
|j−j|≤1,|k−k|≤1
2jHn−1(Pjk(y)),
where the second inequality follows, for instance, by Theorem 3.2.39 in [10]. The in- tegrable majorant of the sequence above that was needed for the application of the dominatedconvergencetheorembefore canbeobtainedbyanapplicationofthecoarea formula.Thiscompletes theproof.
These two propositions are enough to conclude a general L1 bound for the pieces Sjk. This bound can be refined further,when additional regularity onthe domain Ω is assumed.
Proposition 6.3. LetΩ bean openset.Then Sjk L1→L12k(n−1)+j.
Proof. Ifx∈P(y)∩Ωk,then|x−y|= dist(x,Ωc)≤2k+1.HenceP(y)∩Ωk ⊂B(y,2k+1).
RecallthatP(y)=∂E(y) andthatE(y) isconvex.ThusP(y)∩Ωk⊂∂(B(y,2k+1)∩E(y)) where B(y,2k+1)∩E(y) is convex. Since the perimeter of B(y,2k+1) dominates the perimetersof allconvex setswithnon-emptyinteriorcontainedinit,wecanconclude Hn−1(Pjk(y))≤ Hn−1(P(y)∩B(y,2k+1))≤ Hn−1(∂(B(y,2k+1)∩E(y)))
≤ Hn−1(∂B(y,2k+1))2k(n−1). Now theclaimfollows fromProposition6.2.
Remark6.4.IncaseΩ isboundedand∂Ω isC2 smooth,theestimateforHn−1(Pjk(y)) can be refined as follows. If x ∈ Pjk(y), then |y −bx| ≤ δ(x)2−j+1. This implies dist(y,∂Ω)≤δ(x)·2−j+1 andfurther
|by−bx| ≤ |by−y|+|y−bx| ≤4δ(x)·2−j.
As theinward-pointing unitnormalNΩattheboundaryiswell-definedandLipschitz,
|NΩ(by)−NΩ(bx)|diam(Ω)2−j. Because N(bz)= (z−bz)/|z−bz|,thisimplies
NΩ(by)−(x−y) δ(x)
≤ |NΩ(by)−NΩ(bx)|+|y−bx|
δ(x) diam(Ω)·2−j.
Therefore, allvectorsx−y withy ∈ωj(x) are withinanangle∼˜c(Ω)·2−j of NΩ(by).
Hence theset Pjk(y) iscontained inacylinder of height ∼2k and basis ∼˜c(Ω)·2k−j. Bytheinequalityforperimetersofconvex setsasintheproofofProposition6.3
Hn−1(Pjk(y))c(Ω)2k·2(k−j)(n−2).
Thisdependencyonj issharpevenforveryflatdomainsascanbe seenbylettingΩ be asmoothedoutB(0,10)∩ {x1≥0}andy= 2−je1 andk≤0.
However, asthe estimateon Hn−1(Pjk(y)) isnotthe narrowgap ofthe proofof our main theorem,wedonotpursuethisaspectfurther.
TheestimateHn−1(Pjk(y))2k(n−1)cannotbeimprovedingeneral.Iftheboundary of the domain is a single point, the equality is achieved up to a constant. However, focusing on the whole Pj(y) instead of single pieces Pjk(y), onecan obtain adifferent estimate at costof worseningthe dependencyonj. Thefollowing proposition isuseful forsmall valuesofj, anditholdsinverygeneraldomains.
Proposition6.5. LetΩbean openset andy∈Ω.Then ˆ
Pj(y)
1
dist(x, y)n−1dHn−1(x)2j withtheconstant independentof y.Inparticular,
k
2k(1−n)Sjk L1→L1 22j Proof. Wehave
ˆ
Pj(y)
1
dist(x, y)n−1dHn−1(x)lim inf
→0
1
ˆ
{x∈E(y)c:dist(x,Pj(y))≤}
1
dist(x, y)n−1dx.
Given any point x ∈ Pj(y) and a line lx = {y +t(x−y) : t ∈ R}, we see that by Proposition6.1thelinemakesanangle∼2−j withPj(y),andhence
H1(lx∩ {z∈E(y)c : dist(z, Pj(y))≤})2j.
Thefirstclaimedboundfortheintegralfollowsimmediatelyfrompassingtopolarcoor- dinateswithoriginat y.
Toprovethesecond claim,notethatbyProposition6.2 ˆ
Ω
k
2k(1−n)Skjf(x)dx2j
|j−j|≤1
ˆ
Ω
f(y)
k
2k(1−n)Hn−1(Pjk(y))
dy
2j
|j−j|≤1
ˆ
Ω
f(y)
⎛
⎜⎝ ˆ
Pj(y)
1
dist(x, y)n−1dHn−1(x)
⎞
⎟⎠dy
22j f L1,
wherethelaststepwasanapplicationof thefirstclaim.
7. Lp boundsandgeometry
To conclude bounds for the operator Bα, we have to sum up all the pieces in the decomposition.Inordertomakethiswork,onehastoensurethatthereisenoughdecay inj andk.Although the L1 bounds do notsumup, interpolationwith the betterL∞ bounds provides us with enough decay in the angle parameter j. If Ω is bounded, we can take advantage of the Lp(Ω) spaces being nested and use the decay in the scale parameterkneartheboundarytocomplete theproof withnosmoothnessassumptions
ontheboundaryofthedomain.Thisispossibleonlywhen wedonotattempttoprove scalableestimatesthatwouldcaptureLp→Lqsmoothingbeyondonederivativegain.
Theorem 7.1.LetΩbe aboundedopenset,p,α >1.Then Bα Lp(Ω)→Lp(Ω)diam(Ω)α−1
where theimplicit constantonly dependson p,αandthedimension.
Proof. LetSj=
k2k(α−n)Sjk so thatBα=
j2−jSj. ThenbyProposition6.3
Sj L1(Ω)→L1(Ω)≤
log diam(Ω)+1
k=−∞
2k(α−n) Sjk L1(Ω)→L1(Ω)
log diam(Ω)+1
k=−∞
2k(α−1)2j2jdiam(Ω)α−1.
ByProposition5.2
2−jSj L∞(Ω)→L∞(Ω)2−njdiam(Ω)α−1 and byinterpolationweobtain
2−jSj Lp(Ω)→Lp(Ω)2−(p−1)np jdiam(Ω)α−1.
As theexponentisnegative, wecansumupinj toconcludetheproof.
To dealwith thecriticalcaseα= 1 whereourestimateshavethecorrectscaling,we havetotakeintoaccountfinerpropertiesoftheboundary,astheestimationasroughas aboveleadstoalogarithmic blow-upofthek-sumattheboundary.
Proposition 7.2. LetΩ bean openset.
• If Ωsatisfies theuniform curvature boundwith R,thenforall y∈Ωandx∈P(y) with δ(x)≤R,itholdsthat
δ(x)(1−δ(x)
R )(1−cosβ)≤dist(y, ∂Ω) (7.1) where β=(bx−x,y−x).
• If Ωc isconvex,then
δ(x)(1−cosβ)≤dist(y, ∂Ω). (7.2)
Fig. 3.The balls and points appearing in the proof of Proposition7.2.
Proof. Take x ∈ Ω and y ∈ ∂B(x,δ(x)) and let β be the angle between bx−x and y−x.BecauseΩ satisfiesauniformcurvaturebound, thereisanR >0 independentof xand y so thatwe canfind aball B(z,R)⊂Ω withz =bx+ (x−bx)R/δ(x) sothat B(z, R)∩∂Ω={bx}(seeFig.3). ThePythagoreanidentityreads
|z−y|2= (δ(x) sinβ)2+ (R−δ(x)(1−cosβ))2
=R2(1−2δ(x)
R (1−δ(x)
R )(1−cosβ))
≤R2(1−δ(x)
R (1−δ(x)
R )(1−cosβ))2.
Letw be theclosestpoint toy in∂B(z,R).Since z,y andw areonthe sameline,we get
dist(y, ∂Ω)≥ |y−w|=|z−w| − |z−y|
≥R−R(1−δ(x)
R (1−δ(x)
R )(1−cosβ))
=δ(x)(1−δ(x)
R )(1−cosβ)
asclaimed. IfΩc isconvex, thenthe uniformcurvaturebound issatisfiedwith R=∞, whencethesecond claimfollows.
Theorem7.3. LetΩbean openset.Letα= 1andp>1.Then
• If Ωisbounded andsatisfiestheuniformcurvature bound,then B1 Lp(Ω)→Lp(Ω)log
diam(Ω) R + 1
where Ristheradiusfrom theuniformcurvature bound.
• If Ωc isconvex,then
B1 Lp(Ω)→Lp(Ω)1
andtheoperatornormonly dependsonthedimensionandp.
• If Ωismerelyopen,then
B1 Lp(Ω)→Lp(Ω)1 under therestrictionp>1+n1.
Proof. Proposition6.3implies ˆ
Ω
2k(1−n)−jSjkf(x)dxˆ
f(y)1Ak
j(y)dy. (7.3)
Recallthedefinition(6.2).Thereareonly∼log(diam(Ω)/R+1) valuesofksothatR/8≤ 2k ≤2diam(Ω).Forksuchthat2k+3≤R,wecanusethefirstiteminProposition7.2to seethatforfixedy,thesetPjk(y) isnon-emptyonlyforksuchthat2−2j+kdist(y,∂Ω).
Ontheother hand,theupperbound
dist(y, ∂Ω)≤ |y−bx|2k−j
is always valid, so Pjk(y) isnon-empty only for 2−2j+k dist(y,∂Ω) 2−j+k. Conse- quently,
Akj ⊂ {y∈Ω : 2−2j+k dist(y, ∂Ω)2−j+k}.
For anyy,there areonlyj valuesksuchthattheset aboveisnon-empty,and hence by(7.3)
k
2k(1−n)−jSjk L1→L1 log
diam(Ω) R + 1
+j.
InterpolationasintheproofofTheorem7.1 impliestheclaim.
Toprovetheseconditem,justnotethattheconvexityassumptiononthecomplement meanssendingR→ ∞sothat2k+3≤Ralwaysholds.Toprovethethirditem,westudy Sj as inthe proof of Theorem 7.1 and replace the L1 bound from Proposition 6.3 by thatfromProposition6.5.
Corollary 7.4.LetΩbe adomain,p>1andf ∈Lp.ThenAαf(x)from(2.3) isweakly differentiable and
∇Aαf Lp(Ω) f Lp(Ω)
ifany one ofthefollowingholds:
• α >1andΩisbounded.
• α= 1 andΩisbounded andsatisfiesauniformcurvature bound.
• α= 1 andΩc isconvex.
• α= 1 andp>1+n1
The constantdependson thedomain,αandthedimension.
Proof. By linearity, it suffices to prove the norm inequality for smooth functions. By Proposition5.1,it sufficestobound Bα from (5.1).Thisfollows from Theorem7.1 and Theorem7.3
Theorem1.1follows fromCorollary7.4 andLemma4.1.
8. Remarks
8.1. Roleof thedomain
It isnotclear ifthe conditionsonthe domain inthehypothesis ofTheorem 1.1 are necessary.Onemayask if
∇M1Ω Lp(Ω)→Lp(Ω)1
holdsforalldomains Ω andallp>1.Wearenotawareofany counterexamplesso far.
SinceM0ΩdoessatisfyanLp(Ω) boundindependentofthedomain,thequestionisabout thebehavior ofB1 (seeTheorem7.3)ingeneraldomains.Wepointoutthatoneavenue forimprovingtheLp boundsforB1 couldbe toreplacethestrongL1 boundsforSjk by weaktypeboundsinordertoimprove theoperatornormboundwith respecttoj.
8.2. Endpointregularity indomains
Corollary1.2follows fromTheorem1.1,since
∇M1Ωf Ln/(n−1)(Ω) f Ln/(n−1)(Ω) f W1,1(Ω).
Hereweusedthemaintheoremand(2.2).Thesameobservationwasdoneby[5] tonotice thatthe fractional endpoint regularity problem follows from inequality (1.1) asα ≥1 in the full space Rn. The domain case was not known before as the inequality (1.1) should have been replaced by (1.2). This amounts to changing the Hardy–Littlewood maximal function to thespherical maximal function inthe display above.Thatone is