Contents lists available atScienceDirect
Journal
de
Mathématiques
Pures
et
Appliquées
www.elsevier.com/locate/matpur
Poincaré-type
inequality
for
Lipschitz
continuous
vector
fields
Giovanna Cittia,∗, Maria Manfredinia, Andrea Pinamontib,Francesco Serra Cassanoc
aDipartimento di Matematica, Università di Bologna, Piazza di Porta S. Donato 5, 40126, Bologna, Italy bScuola Normale Superiore di Pisa, Piazza dei Cavalieri, 7, 56126, Pisa, Italy
cDipartimento di Matematica, Università di Trento Via Sommarive 14, 38123, Povo (Trento), Italy
a r t i cl e i n f o a b s t r a c t
Article history:
Received23June2013
Availableonline8September2015
MSC: 43A80 35R45 42B20 Keywords: Poincaréinequalities Nonlinearvectorfields Liegroups
The scope of this paper is to prove a Poincaré type inequality for a family of nonlinearvectorfields,whosecoefficientsareonlyLipschitzcontinuouswithrespect tothedistanceinducedbythevectorfieldsthemselves.
© 2015ElsevierMassonSAS.All rights reserved. r é s u m é
Dans cet article on démontre une inégalité de type Poincaré pour une famille dechamps devecteurs nonlinéaires dontles coefficientssont seulementcontinus lipschitziensparrapportàladistanceinduiteparleschampseux-mêmes.
© 2015ElsevierMassonSAS.All rights reserved.
1. Introduction andstatementoftheresult
The Poincaré inequality is one of the main tools in the proof of regularity of solutions of PDEs in divergenceform.Indeed,asprovedbySaloff-Costein[1]andGrigor’yanin[2](seealso[3]),itisequivalent to the Harnack inequality and to Hölder continuity for solutions. Thus, to prove regularity of solutions, it sufficestoestablishasuitablePoincaréinequality.
ThePoincaréinequalityforsmoothHörmandervectorfieldsiswellknownandwasprovedbyJerison[4]. We recall that a Hörmander family of vector fields in Rn, is defined by m ≤ n smooth vector fields,
say∇= (∇1,. . . ,∇m),suchthatthegeneratedLie algebrahasmaximumrankateverypoint.
DenotebyBr(x)⊂ Rnthemetricballofcenterx andradiusr > 0 associatedtotheCC-distancedefined
intermsofthefamily∇.ThePoincaréinequalityprovedin[4]is:
* Correspondingauthor.
E-mail addresses:giovanna.citti@unibo.it(G. Citti),maria.manfredini@unibo.it(M. Manfredini), andrea.pinamonti@gmail.com(A. Pinamonti),casssano@science.unitn.it(F. Serra Cassano). http://dx.doi.org/10.1016/j.matpur.2015.09.001
Br(x)
|ψ(y) − ψr(y)| dLn(y)≤ C r
Br(x)
|∇ψ(y)| dLn(y) ∀ψ ∈ C∞(B r(x)),
where,asusual,ψr(y):= 1/Ln(Br(y))
Br(y)ψ andC > 0.Thepreviousinequalitycanbealsostatedusing
balls defined with respect to different (but equivalent) distances. We mention here the ball box distance (see [5])andthefrozendistancedefinedbyRothschildandSteinin[6].
ThePoincaréinequalityfornon-smoothvectorfieldswasfirstconsideredin[7].Thelaterworksonrelated questions include the papers by Biroli and Mosco [8], Capogna, Danielli and Garofalo [9,10], Chernikov and Vodop’yanov [11], Danielli, Garofalo,Nhieu[12], Franchi, Gallotand Wheeden [13], Franchi, Luand Wheeden[14]andLu[15,16].Morerecently,in[17],theauthorsprovedageneralPoincaréinequalitywhich wasappliedin[18–20]tofamiliesofLipschitzvectorfieldswithdifferentregularityconditions,anddifferent assumptions ontherankofthegeneratedLie algebra.
In [21], the authors studied the relationship between the validity of the Poincaré inequality and the existence of representation formulas for functions as (fractional) integral transforms of first-order vector fields. Theyshowed thatthePoincaré inequalityleadsto (and infactit is oftenequivalent to) asuitable representation formula.Thisapproachwaslaterdevelopedin[9],inwhichanotherproofhasbeengivenof the representationformula relyingon thePoincaréinequalityprovedbyJerison. Finally,in[22], ageneral representation formulaisprovedintermsofthefundamentalsolutionofaHörmandertypesub-Laplacian. Unfortunately,alltheseresultsareexpressedintermsofvectorfieldswithLipschitzcontinuouscoefficients with respect totheEuclidean distance.Ontheother hand,inorder to studypartial differentialequations withnonlinearvectorfields,thisassumptionisnolongernatural.Atypicaldifferentialequationofthistype canbeof theform
m
i,j=1
∇φ
i(aij(φ)∇φjφ) = f, (1.1)
where (aij) is a smooth, symmetric, uniformly positive definite matrix, f is a fixed function and the
coefficients of the vector fields ∇φ depend on the solution φ. Equations of this type naturally arise while studying curvature equations [23], Monge–Ampére equation [24–26], mathematical finance [27–29] or intrinsic minimalgraphsintheHeisenberggroup(seeforinstance [30–35]).
Aparticular,butveryinterestinginstanceof(1.1),istheso-calledminimalsurfaceequationforintrinsic graphsintheHeisenberggroup(seealso[36,35,37]forthecaseofT-graphs).Inthen-dimensionalHeisenberg groupHn,suchgraphsaredescribed asfollows(see [38,39]):
M ={(φ(x1, . . . , x2n), x2, . . . , x2n, x2n+ 2xnφ(x1, . . . , x2n)), (x1, . . . , x2n)∈ ω},
where ω ⊂ R2n
(x1,...,x2n) isan openset andφ: ω → R isacontinuousfunctionsatisfyingsuitable regularity
properties.Intrinsicgraphshavebeenextensivelystudiedinconnectionwiththenotionofrectifiablesetsin Hn (seeforinstance[40–43]),theregularityproblemforminimalsurfaces(seeforinstance[31,32,44–46,35])
andtheBernsteinprobleminHn[30,33,47–49].Inparticular,in[38]itisprovedthattheso-calledhorizontal
perimeter ofM canbe expressedby
PH(M ) =
ω
1 +|∇φφ|2dL2n,
where ∇φ= (∇φ1,. . . ,∇2nφ −1) isthefamilyofnonlinearvectorfieldsdefinedby:
∇φ i = ∂xi− xi+n∂x2n, ∇ φ n = ∂xn+ 2φ∂x2n,∇ φ i+n= ∂xi+n+ xi∂x2n, (1.2)
wherei= 1,. . . ,n−1.Moreover,aspointedoutin[38,33],onecanshowthattheconditionthattheintrinsic surfaceM beH-minimalisexpressed bythenonlinearequationoftype(1.1):
∇φ· ∇φφ
1 +|∇φφ|2 = 0.
Noticethattheregularityofthesolutionφ of(1.1)canbeobtainedonlyintheHölderspacesdefinedinterms of the distance naturallyassociated to the family∇φ. As aconsequence, the coefficientsof the equation, (whichdependonthesolutionitself),arenotexpectedtobeLipschitzwithrespecttotheEuclideandistance. Moreover,thislackofregularityofthecoefficientsimpliesthattheequivalenceofthedifferentdefinitionsof distancescannotbededucedusing[5].Toovercomethisproblem,adistancemodeledonthefrozendistance ofRothschildandSteinwasdefinedin[23]whilestudyingtheLeviequation.In[38,39],ananalogousfrozen distance dφ hasbeen proposedforthevectorfields in(1.2),as thesymmetrizeddistance associatedto the
frozenvectorfields
∇φ(x0)
i :=∇
φ
i for i = 1, . . . , 2n− 1, i = n, ∇φ(xn 0):= ∂xn+ 2φ(x0)∂x2n, (1.3)
wherex0∈ ω isfixed.Themainadvantageofworkingwiththisfamilyofvectorfieldsreliesonthefactthat
theyhaveC∞ coefficientsandtheycanbe consideredas azeroorderapproximationof thefamily∇φ. We pointoutthattheequivalencebetweenthefrozendistance dφ andtheball boxdistancedefinedin[5]was
provedin[39].Moreover, theequivalencebetween dφ and theCC-distance generated by∇φ canbefound
in[50].
Motivatedbythediscussionabove,inthispaperweproveaPoincarétypeinequalityforthemodelvector fields in(1.2),under theassumption thatthecoefficients areLipschitz continuouswith respect to dφ and
forfunctionswhichbelongtoanintrinsicSobolevspace:
Definition 1.1. Let φ : ω ⊂ R2n −→ R be an intrinsic Lipschitz continuous function, in the sense of Definition 2.1 below.Wesaythatafunctionψ : ω⊂ R2n −→ R belongsto thespaceW
φ(ω) ifthere exist
sequences{ψk}k∈Nand {φk}k∈N inC∞(ω) suchthat
(i) ψk → ψ inL1loc(ω) ask→ +∞;
(ii) φk→ φ uniformlyinω ask→ +∞;
(iii) |∇φkψ
k(x)|≤ M ∀x∈ ω andk andforsomepositiveconstantM ;
(iv) ∇φkψ
k ∗∇φψ ask→ +∞.
Then,ourmain resultisthefollowing:
Theorem 1.2. Let ω be a bounded and open subset of R2n with n ≥ 2, and let p≥ 1. Let φ : ω → R be
an intrinsic Lipschitz function and ψ ∈ Wφ(ω). Then there exist positive constants C1,C2 with C2 > 1
(dependingcontinuously ontheLipschitzconstant Lφ ofφ) suchthat
Uφ(¯x,r) |ψ(y) − ψUφ(¯x,r)| p dL2n(y)≤ C 1rp Uφ(¯x,C2r) |∇φψ(y)|p dL2n(y), (1.4)
forevery Uφ(¯x,C2r)⊂ ω, where
Uφ(x, r) :={y ∈ ω : dφ(x, y) < r} . (1.5)
ψUφ(¯x,r):= 1 L2n(U φ(¯x, r)) Uφ(¯x,r) ψ(y) dL2n(y) . (1.6)
Corollary 1.3. If φ : ω → R is an intrinsic Lipschitz function, then there exist positive constants C1, C2
with C2> 1 (dependingcontinuously ontheLipschitz constant Lφ of φ)suchthat
Uφ(¯x,r) |φ(y) − φUφ(¯x,r)| dL 2n(y)≤ C 1r Uφ(¯x,C2r) |∇φφ(y)| dL2n(y), (1.7) forevery Uφ(¯x,C2r)⊂ ω.
We briefly describe our approach. Since the coefficients of the vector fields ∇φ are only Lipschitz
continuous (with respect to dφ), we cannot consider the Lie algebra generated by the vector fields.
Nevertheless,theexplicitexpressionofthevectorfields ∇φ1,. . . ,∇φ2n−1 ensuresthat [∇φi,∇φn+i] = 2 ∂x2n ∀ i = 1, . . . , n − 1,
sothatthevectorfieldsandtheircommutatorsspanthewholespaceateverypoint,whichcanbeinterpreted as aHörmandercondition for non-regular vectorfields. This approachcanbe consideredaversion ofthe Rothschildand Steinmethodfornon-smoothvectorfields,andhasbeenused in[23]inadifferent setting. In particular,everyψ∈ C∞(ω) can be representedbymeansof asuitablerepresentationformula (proved in[22])intermsof thevectorfields ∇φ(x0)
i ,thefundamentalsolutionΓx0 of theLaplacianoperator
Lφ(x0):= 2n−1 i=1 (∇φ(x0) i ) 2
and thesuperlevelsets Ωφ(x0)(x0,r) ofΓx0,whichareequivalentto theballsUφ(x0,r).
InordertoproveTheorem 1.2,insection3 wewillfirstmodifytheaforementionedrepresentationformula to obtainanotherrepresentation formulaintermsofthefamily∇φ.Subsequently,usinganapproximation result for intrinsicLipschitz functionscontained in[45], (see also[50] for arefinement)we provethatthe representation formulaprovedinSection3 stillholdsforintrinsicLipschitzfunctions.Finally,inSection4 we willprovidetheproofofTheorem 1.2.
2. Preliminaries
2.1. Theintrinsic distance
Fixn≥ 2.Letω⊂ R2n beanopenandbounded setandφ: ω−→ R beacontinuousfunction.TheLie algebrageneratedbythefamily∇φ:= (∇φ1,. . . ,∇φ2n−1) definedin(1.2)hasmaximumrankateverypoint, hence it is possible to define on ω the exponential ball box distance and the CC-distance, see [5]. These distances are not explicitly computable, therefore, it is convenient to introduce an equivalent, explicitly computable, quasi-distanceonω. Todothis,we usethefreezing methoddevelopedin[6]andsuccessively refined in[39] (see also [38]). Precisely, let us fix x0 ∈ ω and consider the familyof smooth vector fields
∇φ(x0)definedin(1.3)andthenewfamilyoffrozenvectorfields
ˆ
∇φ(x0)= (∇φ(x0), ∂
x2n) . (2.8)
LetusnowintroducetheLiealgebraG generatedbythefamilyofvectorfields∇φ(x0).Noticethat,since
theonlynon-vanishingcommutatoris[∇φ(x0)
i ,∇
φ(x0)
n+i ]= 2∇ φ(x0)
toHn−1×R,asCarnotgroups,whereHn−1≡ R2n−1denotesthe(n−1)-dimensionalHeisenberggroup.Any
elementx˜∈ G canbeidentifiedbyitscoordinateswithrespecttothebasis∇ˆφ(x0),thatisx˜≡ (˜x
1,. . . ,x˜2n),
ifx =˜ 2ni=1x˜i∇ˆiφ(x0).Wecanalso induceanormonG bydefining
(˜x1, . . . , ˜x2n) := max{|(˜x1, . . . , ˜x2n−1)|R2n−1,|˜x2n| 1
2} (2.9)
foreach G.Theexponential map associated to thefamilyof vectorfields ∇ˆφ(x0) iswell defined.Precisely,
foreachx∈ R2n:
Expφ(x0),x:G −→ R2n, Expφ(x0),x(˜y) := exp
2n i=1 ˜ yi∇ˆφ(xi 0) (x) . Incoordinatesweget Expφ(x0),x(˜y) =x1+ ˜y1, . . . , x2n−1+ ˜y2n−1, x2n+ ˜y2n+ 2˜ynφ(x0)− σ(˜y, x) , (2.10) where σ(x, y) := n−1 i=1 (yi+nxi− xi+nyi). (2.11)
TheinversemappingofExpφ(x0),xwill bedenotedbyLogφ(x0),x:R2n−→ G, andtheball-boxexponential
distanceassociated tothevectorfields ∇ˆφ(x0)isdefinedforeveryx,y∈ R2n
dφ(x0)(x, y) :=Logφ(x0),x(y).
Inparticular,asimplecomputationgives:
Logφ(x0),x(y) =
y1− x1, . . . , y2n−1− x2n−1, y2n− x2n− 2φ(x0)(yn− xn)− σ(x, y)
(2.12)
andthereforeusing (2.9)weget:
dφ(x0)(x, y) = max
|ˆx − ˆy|R2n−1, σφ,x0(x, y)
, (2.13)
whereforeveryx= (x1,. . . ,x2n)∈ R2n wehavedenotedx := (xˆ 1,. . . ,x2n−1)∈ R2n−1 and
σφ,x0(x, y) := |y2n− x2n− 2φ(x0)(yn− xn)− σ(x, y)| 1/2 x
0∈ ω, x, y ∈ R2n. (2.14)
Moreoverwewillsimplydenote
σφ(x, y) := σφ,x(x, y) x, y∈ ω. (2.15)
Finally,wedefinethefollowing symmetricfunction:
dφ(x, y) = 1 2 dφ(x)(x, y) + dφ(y)(y, x) ∀x, y ∈ ω. (2.16)
Definition 2.1. We say thatφ : ω ⊂ R2n −→ R isan intrinsic Lipschitz continuous functionin ω andwe
writeφ∈ Lip(ω),ifthere isaconstantL> 0 suchthat
The Lipschitz constant of φ in ω is the infimum of the numbers L such that (2.17) holds and we write
Lφ,ω (or simplyLφ)to denoteit. Wealsosaythatφ is alocally intrinsicLipschitz function,andwe write
φ∈ Liploc(ω) ifφ∈ Lip(ω) foreveryω ω.
Remark2.2.Itimmediatelyfollowsfromtheexplicitexpressionofdφ (seealso[23])that,ifφ∈ Lip(ω) then
dφ isaquasi-distance onω. Precisely,
dφ(x, y) = 0⇐⇒ x = y;
dφ(x, y) = dφ(y, x);
and foreachx,y,z∈ ω:
dφ(x, y)≤
≤ dφ(x, z) + dφ(y, z) +|φ(x) − φ(z)|1/2|xn− zn|1/2+|φ(y) − φ(z)|1/2|yn− zn|1/2 (2.18)
so that
dφ(x, y)≤ (1 + Lφ)1/2(dφ(x, z) + dφ(y, z)) .
Remark2.3. Itiseasy toseethat,if φ∈ Lip(ω), then
σφ(y, x)≤ σφ(x, y) +|φ(x) − φ(y)|1/2|xn− yn|1/2 ∀x, y ∈ ω
whence,by (2.16),
dφ(x, y)≤ |ˆx − ˆy|R2n−1+ σφ(x, y) +|φ(x) − φ(y)|1/2|xn− yn|1/2 ∀x, y ∈ ω. (2.19)
Remark 2.4. Moreover, by a simple calculation, we obtain that there existsC1 > 1 depending only on Lφ
such thatforeach x,y∈ ω
1
C1
dφ(y)(y, x)≤ dφ(x, y)≤ C1dφ(y)(y, x), (2.20)
1
C1
dφ(x)(x, y)≤ dφ(x, y)≤ C1dφ(x)(x, y). (2.21)
Besides,there existsapositiveconstant C2= C2(Lφ) suchthat foreachx,y,z∈ ω
dφ(x)(x, y)≤ C2
dφ(x)(x, z) + dφ(z)(z, y)
. (2.22)
2.2. Lipschitzcontinuous functionswithrespecttononlinearvector fields
A detailed analysis of Lip(ω) canbe found in[50,40],here werecall only those propertiesthatwe will useintheproofofTheorem 1.2.
Notice that Lip(ω) is not a vector space (see [35, Remark 4.2]). Nevertheless, the intrinsic Lipschitz functionsamounttoathickclassoffunctions.Indeed,itholdsthat
LipE(ω) Liploc(ω) Cloc1/2(ω) , (2.23) where,LipE(ω) andCloc1/2(ω) denotetheclassesofreal-valuedEuclideanLipschitzandlocally1/2-Euclidean– Höldercontinuousfunctionsonω respectively,see [40,Propositions4.8and4.11].
Theorem2.5. (See[40].)If φ∈ Lip(ω) then φ is∇φ-differentiableforL2n-a.e. x∈ ω, inthesense defined
in [38].Besides,forL2n-a.e.x∈ ω there isaunique vector∇φφ(x)∈ R2n−1 called ∇φ-gradientof φ such that
φ(y) = φ(x) +∇φφ(x), ˜π(y) + o(dφ(x, y)) as y→ x
where ·, · denotes theEuclidean scalar productin R2n−1 and π :˜ R2n → R2n−1, π(x˜
1,. . . ,x2n−1,x2n):=
(x1,. . . ,x2n−1),∀x∈ R2n−1.
In [50] thefollowing estimates for Lφ are proved.Precisely, for each x¯∈ ω and eachr > 0 sufficiently
smallthere isC1> 0 depending onlyon∇φφL∞(ω) suchthat
Lφ,Uφ(¯x,r)≤ C1∇ φφ
L∞(ω),
andthere isC2= C2(n)> 0 such that
∇φφ
L∞(ω)≤ C2Lφ(Lφ+ 1)
whereUφ(x,r) isdefinedin(1.5).
Moreover,thefollowingapproximationresultforintrinsicLipschitzfunctionsithasbeenrecentlyproved in[45]:
Theorem 2.6. Letω ⊂ R2n be a bounded open set and let φ ∈ Lip(ω). Then there exists asequence {φk}
with φk∈ C∞(ω) suchthat
(i) φk→ φ uniformlyinω as k→ ∞,
(ii) |∇φkφ
k(x)|≤ ∇φφL∞(ω) ∀x∈ ω.
Wealsoquotethepaper[50]whereweprovedthateveryφ∈ Lip(ω) canbeapproximatedbyasequence
{φk}k∈Nofsmoothfunctionssatisfying(i),(ii)andalso
∇φkφ
k(x)→ ∇φφ(x) L2n-a.e. in ω.
2.3. Sub-Laplacianand fundamentalsolution
In order to study the dependence of the vector fields ∇ˆφ(x0) (defined in (2.8)) on the variable x 0 we
recognizethatthemap
Logφ(x0),x0 :R2n−→ G
changesthefamilies∇φ(x0) and∇ˆφ(x0)into thefamily∇ and∇ respectively,ˆ where
∇i:=∇φ(xi 0), ∇ˆj:= ˆ∇φ(xj 0) for i∈ {1, . . . , 2n − 1}, j ∈ {1, . . . , 2n}, i, j = n,
∇n:= ∂xn, ∇ˆn= ∂xn. (2.24)
Precisely,foreachψ∈ C∞(R2n),ifwedefine
˜
ψ(˜x) := ψ(Log−1φ(x
then
ˆ
∇φ(x0)
i ψ(x) = ˆ∇iψ(Log˜ φ(x0),x0(x)), ∀i ∈ {1, . . . , 2n}.
Wecandefineametric˜d onG associatedtothevectorfields∇,ˆ independentofx0.Namely,givenx∈ R2nlet
Exp∇,xˆ : G → R2n, Exp∇,xˆ (˜y) := exp
2n i=1 ˜ yi∇ˆi x if ˜y = 2n i=1 ˜ yi∇ˆi.
WecanalsoidentifyG withR2n,byidentifyinganelementofG withitscoordinateswithrespecttobasis ˆ∇.
Insuchaway,wecandefine
˜
d(˜x, ˜y) :=Exp−1ˆ
∇,˜x(˜y), (2.26)
where · denotesthenormin(2.9).Inparticular itholds that ˜
d(0, ˜x) =˜x, ∀˜x ∈ G ≡ R2n,
dφ(x0)(x, y) = ˜d(Logφ(x0),x0(x), Logφ(x0),x0(y)) ∀x, y, x0∈ ω. (2.27)
Moreoveritfollows thatd turns˜ outto beahomogeneousnormonG ≡ Hn−1× R. Letus callsub-Laplacianthesecondorder differentialoperatordefinedas
Lφ(x0):= 2n−1
i=1
(∇φ(x0)
i )2. (2.28)
It iswellknownthatLφ(x0) admitsafundamentalsolutionwhichwewill denotebyΓφ(x0)(see for[51]for
thedetails).Thisoperatorischangedbythemap Logφ(x0),x0 into thesub-Laplacian operator
L :=
2n−1 i=1
(∇i)2.
Thatis,foreachψ∈ C∞(R2n)
(Lφ(x0)ψ)(x) = (L ˜ψ)(Logφ(x0),x0(x)) ∀ x ∈ R 2n,
where ψ is˜ definedin(2.25).
Clearly, the operator L has a fundamental solution Γ of class C∞ far from the pole ˜x = ˜y, which is homogeneousofdegree2− Q withrespect tothedilationfamilynaturallyassociatedtoG,where Q isthe homogeneous dimension of Hn−1× R (see [51, Section 5.3] and the references therein). This means that there existpositiveconstantsC1,C2 suchthatforeveryx and˜ y in˜ R2n,x˜= ˜y
C1 ˜ d(˜x, ˜y)Q−2 ≤ Γ(˜x, ˜y) ≤ C2 ˜ d(˜x, ˜y)Q−2; |∇iΓ(˜x, ˜y)| ≤ C2 ˜ d(˜x, ˜y)Q−1; (2.29) |∇j∇iΓ(˜x, ˜y)| ≤ C2 ˜ d(˜x, ˜y)Q,
for every i,j = 1,. . . ,2n− 1 (see [52] and [51, Section 5.4]). Besides, the fundamental solution Γφ(x0) of Lφ(x0)canbeexplicitlywrittenintermsofΓ as
Γφ(x0)(x, y) = Γ(Logφ(x0),x0(x), Logφ(x0),x0(y)), (2.30)
and
∇φ(x0)
i Γφ(x0)(x, y) =∇iΓ(Logφ(x0),x0(x), Logφ(x0),x0(y)),
fori= 1,. . . ,2n−1.Itfollowsthattheinequalitiesin(2.29)aresatisfiedalsoforΓφ(x0)(x,y) anddφ(x0)(x,y)
with the sameconstants. In particular, it is clear thatthese constants are independent of x0. Using the
estimatesforΓφ(x0)itfollowsthatthespheresinthemetricdφ(x0)areequivalenttothesuperlevelsofthe
fundamentalsolutionΓφ(x0):
Ωφ(x0)(x, r) =
y∈ R2n | Γφ(x0)(x, y) > r
2−Q, r > 0. (2.31)
Moreover,for everyfixedx0∈ ω,theset Ωφ(x0)(x0,r) hasregularboundary(see [22]).Inparticular, from
(2.21), (2.27)and(2.29), thereexistr0,α > 0 withα = α(Lφ) suchthatforany x0∈ ω andr≤ r0
Ωφ(x0)(x0, r/α)⊂ Uφ(x0, r)⊂ Ωφ(x0)(x0, α r), (2.32)
whereUφ(x0,r) istheballdefinedin(1.5).By(2.30) wehavethat
Ωφ(x0)(x, r) =
y∈ R2n | Γ(Logφ(x0),x0(x), Logφ(x0),x0(y)) > r
2−Q, (2.33)
inparticularthesetsΩφ(x0)(x0,r) canbeexpressedintermsofthesuperlevelsofthefundamentalsolution
Γ asfollows: Ωφ(x0)(x0, r) = y∈ R2n | Γ(0, Logφ(x0),x0(y)) > r2−Q = Expφ(x0),x0( ˜Ω(0, r)), (2.34) where ˜ Ω(0, r) :=y˜∈ R2n | Γ(0, ˜y) > r2−Q. (2.35) Wewillalsodenote
K(˜y) := Γ−(Q−2)1 (0, ˜y), ˜y∈ R2n, (2.36)
sothat,wecanrewriteΩ(0,˜ r) as:
˜
Ω(0, r) =y˜∈ R2n | K(˜y) < r. (2.37) 3. A representationformulaintermsoftheintrinsicgradient
Thissectionisorganizedintwosubsections.Inthefirstone,wefixasmoothfunctionφ: ω⊂ R2n→ R
andweintroducethenotionofintegralmeanm of¯ anotherregularfunctionψ : ω⊂ R2n → R onthesuper
levels Ωφ(x0)(x0,r) of the fundamental solution Γφ(x0). Then, we prove a representation formula for any
regularfunctionψ intermsofitsintegralmeananditsintrinsicgradient.
Thesecondsubsectionisquitetechnical,andisdevotedtoestablishsomepropertiesoftheintegralmean ofψ.
3.1. Representationformula
Inthis section,wefix ω⊂ R2n openand bounded,n≥ 2 andφ,ψ∈ C∞(ω).Weprovearepresentation
formula forψ onthesuperlevelsΩφ(x0)(x0,r) ofΓφ(x0). Asimilarrepresentation formulahasbeenproved
in[51,22]fortheapproximatedvectorfields ∇φ(x0).From thisformula,we willdeduceanewandintrinsic
representationformulaforψ,expressedintermsofitsintrinsicgradient∇φψ.Inthecaseunderconsideration
theresultof[51,22]canbe statedasfollows:
Proposition3.1.Foreveryx0∈ ω andR > 0 suchthatΩφ(x0)(x0,R)⊂ ω andforeveryψ∈ C∞(ω) wehave
ψ(x0) = Q (Q − 2)(1 −21Q)RQ Ωφ(x0)(x0,R)\Ωφ(x0)(x0,R2) |∇φ(x0)Γ φ(x0)(x0, y)| 2 Γ2(φ(xQ−1)/(Q−2) 0) (x0, y) ψ(y) dL2n(y) + Q (1− 1 2Q)RQ R R 2 rQ−1 Ωφ(x0)(x0,r) ∇φ(x0)Γ φ(x0)(x0, y),∇ φ(x0)ψ(y) dL2n(y)dr. (3.38)
Here·, · denotesthestandard Euclideanscalar productinR2n−1.
Remark3.2. Weexplicitlynotethat,if wechoose ψ≡ 1,then from(3.38) weget
1 = C(Q) RQ Ωφ(x0)(x0,R)\Ωφ(x0)(x0,R2) |∇φ(x0)Γ φ(x0)(x0, y)| 2 Γ2(φ(xQ−1)/(Q−2) 0) (x0, y) dL2n(y), (3.39) where C(Q):=(Q−2)(1−Q 1 2Q) .
This remark allows tosay that(3.38) representsafunction ψ as thesum ofits meanon asuitable set and thegradient∇φ(x0)ψ.Hence,itseemsnaturalto givethefollowing definition
Definition 3.3. Let ψ ∈ L1loc(ω).For every x0 ∈ ω and R > 0 such that Ωφ(x0)(x0,R) ⊂ ω we define the
following meanofψ,ontheset Ωφ(x0)(x0,R)\ Ωφ(x0)(x0,
R
2),intermsofthefundamentalsolutionΓφ(x0):
¯ m(ψ, φ, R)(x0) := C(Q) RQ Ωφ(x0)(x0,R)\Ωφ(x0)(x0,R2) |∇φ(x0)Γ φ(x0)(x0, y)| 2 Γ2(φ(xQ−1)/(Q−2) 0) (x0, y) ψ(y) dL2n(y).
In thesequel wewill need anothermean of ψ on thesameset Ωφ(x0)(x0,R)\ Ωφ(x0)(x0,
R 2). Precisely, we denote: m(ψ, φ, R)(x0) := 2 R R R 2 ¯ m(ψ, φ, r)(x0) dr. (3.40)
Thefollowingremark,whichwillbeveryusefullateron,providesanintegrationformulabypartsforthe derivative∂2n.
Remark3.4.Letg∈ C1(R2n),r > 0 andc
1,c2> 0 wedefine:
Ar,c1,c2 :={y ∈ R 2n : c
1r < g(y) < c2r}.
Then, for every f,ψ ∈ C1(R2n) and R
1,R2 ∈ R with R1 < R2, using the fact that
∂2n= 12(∇φ(x1 0)∇ φ n+1− ∇ φ(x0) n+1 ∇ φ
1) andintegrating by partwehave
R2 R1 rQ−1 Ar,c1,c2
f (y)∂2nψ(y) dL2n(y)dr =
=1 2 R2 R1 rQ−1 {y:g(y)/c2=r} f (y)∇φn+1ψ(y)∇ φ(x0) 1 g(y) |∇Eg(y)| dH2n−1(y) dr −1 2 R2 R1 rQ−1 {y:g(y)/c1=r} f (y)∇φn+1ψ(y)∇ φ(x0) 1 g(y) |∇Eg(y)| dH2n−1(y)dr −1 2 R2 R1 rQ−1 {y:g(y)/c2=r} f (y)∇φ1ψ(y)∇ φ(x0) n+1 g(y) |∇Eg(y)| dH2n−1(y)dr +1 2 R2 R1 rQ−1 {y:g(y)/c1=r} f (y)∇φ1ψ(y)∇ φ(x0) n+1 g(y) |∇Eg(y)| dH2n−1(y)dr −1 2 R2 R1 rQ−1 Ar,c1,c2 ∇φ(x0) 1 f (y)∇ φ n+1ψ(y) dL2n(y)dr +1 2 R2 R1 rQ−1 Ar,c1,c2 ∇φ(x0) n+1 f (y)∇ φ 1ψ(y) dL2n(y)dr,
where∇E denotestheEuclidean gradient. Bythecoarea formulawe inferthat R2 R1 rQ−1 Ar,c1,c2
f (y)∂2nψ(y) dL2n(y)dr =
=1 2 Ar,c2R1,c2R2 gQ−1(y) cQ−12 f (y)∇ φ n+1ψ(y)∇ φ(x0) 1 g(y) dL2n(y) −1 2 Ar,c1R1,c1R2 gQ−1(y) cQ−11 f (y)∇ φ n+1ψ(y)∇ φ(x0) 1 g(y) dL2n(y) −1 2 Ar,c2R1,c2R2 gQ−1(y) cQ−12 f (y)∇ φ 1ψ(y)∇ φ(x0) n+1 g(y) dL 2n(y) +1 2 Ar,c1R1,c1R2 gQ−1(y) cQ−11 f (y)∇ φ 1ψ(y)∇ φ(x0) n+1 g(y) dL2n(y)
−1 2 R2 R1 rQ−1 Ar,c1,c2 ∇φ(x0) 1 f (y)∇ φ n+1ψ(y) dL 2n(y)dr +1 2 R2 R1 rQ−1 Ar,c1,c2 ∇φ(x0) n+1 f (y)∇ φ 1ψ(y) dL2n(y)dr.
If in additionc1= 0 thentheintegralson Ar,c1R1,c1R2 arenot present.
Inthefollowing propositionwewillslightlymodifytherepresentationformulaprovedinProposition 3.1 inorder to obtain a mean representation formula containing only derivatives with respect to the vector fields ∇φ.
Proposition 3.5. Letφ,ψ∈ C∞(ω). Foreveryx0∈ ω andR > 0 such that Ωφ(x0)(x0,R)⊂ ω wehave ψ(x0)− m(ψ, φ, R)(x0) = IR(x0), where IR(x0) = 2 R R R 4 f1 r R Ωφ(x0)(x0,r) K1(x0, y),∇φψ(y) dL2n(y)dr + 2 R R R 2 Ωφ(x0)(x0,r)\Ωφ(x0)(x0,r2) K2(x0, y, r),∇φψ(y) dL2n(y)dr.
Here,f1∈ C0([14,1]) andthevectorvaluedfunctionsK1andK2aredefinedin(3.44)and(3.45)respectively.
Moreover, |K1(x0, y)| ≤ ˜C1(Lφ,Ωφ(x0)(x0,R)+ 1) 2d1−Q φ(x0)(x0, y) ∀y ∈ Ωφ(x0)(x0, R) (3.41) and |K2(x0, y, r)| ≤ ˜C2(Lφ+ 1)2d1φ(x−Q0)(x0, y), ∀y ∈ Ωφ(x0)(x0, R)\ Ωφ(x0)(x0, R 2), r∈ R 2, R , (3.42)
where Lφ meansLφ,Ωφ(x0)(x0,R)\Ωφ(x0)(x0,R/2) and C˜1,C˜2 > 0 are suitable constants depending only on the homogeneous dimensionQ andon thestructureconstantsC1 andC2 in(2.29).
Proof. We will always denote byC a positiveconstantdepending only onQ whichcan be different from line toline.ByProposition 3.1forallr∈ (0,R)
ψ(x0)− ¯m(ψ, φ, r)(x0) = = C rQ r r 2 sQ−1 Ωφ(x0)(x0,s) ∇φ(x0)Γ φ(x0)(x0, y),∇ φ(x0)ψ(y) dL2n(y)ds = C rQ r r 2 sQ−1 Ωφ(x0)(x0,s) ∇φ(x0)Γ φ(x0)(x0, y),∇ φψ(y)dL2n(y)ds
+ C rQ r r 2 sQ−1 Ωφ(x0)(x0,s) ∇φ(x0)
n Γφ(x0)(x0, y)(φ(x0)− φ(y))∂2nψ(y) dL
2n(y)ds.
UsingRemark 3.4withg(y):= Γ
1 2−Q φ(x0)(x0,y) weobtain ψ(x0)− ¯m(ψ, φ, r)(x0) = 1 rQ r r 2 sQ−1 Ωφ(x0)(x0,s) K1(x0, y),∇φψ(y) dL2n(y)ds + Ωφ(x0)(x0,r)\Ωφ(x0)(x0,r2) K2(x0, y, r)∇φψ(y) dL2n(y), (3.43) where K1(x0, y) := C∇φ(x0)Γφ(x0)(x0, y) − C∇φ(x0) 1 ∇φ(xn 0)Γφ(x0)(x0, y)(φ(x0)− φ(y))en+1 + C∇φ(x0) n+1 ∇φ(xn 0)Γφ(x0)(x0, y)(φ(x0)− φ(y))e1, (3.44) and K2(x0, y, r) := C rQ ∇φ(x0) n Γφ(x0)(x0, y) Γ2(φ(xQ−1)/(Q−2) 0) (x0, y) (φ(x0)− φ(y))∇φ(xn+10)Γφ(x0)(x0, y)e1 − C rQ ∇φ(x0) n Γφ(x0)(x0, y) Γ2(φ(xQ−1)/(Q−2) 0) (x0, y) (φ(x0)− φ(y))∇φ(x1 0)Γφ(x0)(x0, y)en+1, (3.45)
whereei isthei-th elementofthecanonicalbasisofR2n−1.
Integrating(3.43)from R2 toR weget
ψ(x0)− m(ψ, φ, R)(x0) = = 2 R R R 2 1 ρQ ρ ρ 2 rQ−1 Ωφ(x0)(x0,r) K1(x0, y),∇φψ(y) dL2n(y)drdρ + 2 R R R 2 Ωφ(x0)(x0,ρ)\Ωφ(x0)(x0,ρ2) K2(x0, y, ρ),∇φψ(y) dL2n(y)dρ.
Exchangingtheorderofintegration inthefirstintegralandsetting:
f1(t) :=
21−Q− (2t)Q−1
1− Q if t∈ [1/4, 1/2], f1(t) :=
tQ−1− 1
1− Q if t∈ [1/2, 1], wegetthethesis.Finally,theestimatesonK1 andK2 aredirectconsequencesof(2.29). 2
3.2. Somepropertiesof theintegralmean
Inthissection wecollectsomepropertiesof theintegralmeanm(ψ,¯ φ,r) tobeused inthenextsection. Wewill seethatinorderto concludetheproof ofthePoincaréinequalitywewillneedadetailedestimate of thedifference
¯
m(ψ, φ, r)(x)− ¯m(ψ, φ, r)(y) (3.46) at twodifferentpointsx,y∈ ω withx= y.This differenceisestimated inProposition 3.11,anditisbased onsometechnicallemmas.
Lemma 3.6. Let ψ ∈ C∞(ω). For each x0 ∈ ω and each R > 0 such that Ωφ(x0)(x0,R) ⊂ ω the mean,
¯
m(ψ,φ,R)(x0),of asmoothfunctionφ canbeexpressed asfollows:
¯ m(ψ, φ, R)(x0) = 1 RQ ˜ Ω(0,R)\˜Ω(0,R 2) K3(0, ˜y)ψ(Expφ(x0),x0(˜y)) dL 2n(˜y),
where Ω(0,˜ R) isdefined in(2.35) and
K3(0, ˜y) := Q
(Q − 2)(1 − 1 2Q)
|∇Γ(0, ˜y)|2
Γ(0, ˜y)2(Q−1)/(Q−2).
Moreover, there exist constants C˜3,C˜4 depending only Q and on the structure constants C1 and C2 in
(2.29) suchthat |K3(0, ˜y)| ≤ ˜C3, |∇K3(0, ˜y)| ≤ ˜ C4 ˜y ∀˜y ∈ ˜Ω(0, R) \ ˜Ω 0,R 2 . (3.47)
Proof. By(2.34) wehavethat
Ωφ(x0)(x0, R) = Expφ(x0),x0( ˜Ω(0, R)).
Sothat,byDefinition 3.3and(2.30)wehave:
¯ m(ψ, φ, R)(x0) = = C(Q) (Q − 2)RQ Ωφ(x0)(x0,R)\Ωφ(x0)(x0,R2) |∇φ(x0)Γ φ(x0)(x0, y)| 2 Γ2(φ(xQ−1)/(Q−2) 0) (x0, y) ψ(y) dL2n(y) = C(Q) (Q − 2)RQ ˜ Ω(0,R)\˜Ω(0,R 2) |∇Γ(0, ˜y)|2 Γ2(Q−1)/(Q−2)(0, ˜y)ψ(Expφ(x0),x0(˜y)) dL 2n(˜y),
where in the last equality we have applied a change of variables and the fact that the determinant of the Jacobian matrixof Expφ(x0),x0 is equalto 1.Finally, we observe that(3.47) follows directly from the
estimatesonΓ in(2.29). 2
In the next proposition we will start studying properties of the difference
ψ(Expφ(x),x(˜y))− ψ(Expφ(x0),x0(˜y)), which, thanks to the previous lemma, can be considered the first stepintheproofof(3.46).
Proposition 3.7. For every x¯ ∈ ω there exist a constant C0 > 0, such that for every R > 0 with
Ωφ(¯x)(¯x,C0R) ω, every x,x0 ∈ Ωφ(¯x)(¯x,R) and every y˜ ∈ ˜Ω(0,R) there is an integral curve of the
vector fields ∇ˆφ(x0),γ ˜
y : [0,1]→ ω joiningExpφ(x),x(˜y) and Expφ(x0),x0(˜y). Moreover, γy˜ can be explicitly written as γy˜(t) := exp t˜h ˆ∇φ(x0) exp(˜y ˆ∇φ(x0))(x 0) t∈ [0, 1], (3.48) where ˜ h = Logφ(x0),Expφ(x0),x0(˜y) Expφ(x),x(˜y) . (3.49)
Proof. Letusfix¯x∈ ω,andasphereΩφ(¯x)(¯x,CR) subsetofω.TheconstantC = C(Lφ,Ωφ(¯x)(¯x,R))> 0 will
bechosenat theend.Wefirstnotethat,thereexists
C = C(Lφ,Ωφ(¯x)(¯x, ¯R)) > 0 (3.50)
such thatfor everyx,x0 ∈ Ωφ(¯x)(¯x,R), y˜∈ ˜Ω(0,R),the points Expφ(x),x(˜y) and Expφ(x0),x0(˜y) belong to
Ωφ(¯x)(¯x,CR).By(2.10)weget Expφ(x),x(˜y) =x1+ ˜y1, . . . , x2n−1+ ˜y2n−1, x2n+ ˜y2n+ 2˜ynφ(x)− σ(˜y, x) , Expφ(x0),x0(˜y) =x0,1+ ˜y1, . . . , x0,2n−1+ ˜y2n−1, x0,2n+ ˜y2n+ 2˜ynφ(x0)− σ(˜y, x0) .
Then,using (3.49)and(2.12), weobtain
˜ hi = (x− x0)i i = 1, . . . , 2n− 1, ˜ h2n = (x− x0)2n− 2φ(x0)(x− x0)n+ 2˜yn(φ(x)− φ(x0))− 2σ(˜y, x − x0) + σ(x, x0) (3.51) andcalling ˜ x := Logφ(x0),x0(x), (3.52) werealizethat ˜ h = ˜x +2˜yn(φ(x)− φ(x0))− 2σ(˜y, ˜x) e2n. (3.53)
By(3.48)andtheBaker–Campbell–Hausdorffformulawehave
γy˜(t) = exp t˜h ˆ∇φ(x0) expy ˆ˜∇φ(x0) (x0) = exp2t˜yn(φ(x)− φ(x0))∂2n+ 2tσ(˜y, ˜x)∂2n− tσ(˜y, ˜x)∂2n+ (t˜x + ˜y) ˆ∇φ(x0) (x0).
Fromthis andusing(2.10)we get
(γy˜(t))i= t(x− x0)i+ (˜y + x0)i i = 1, . . . , 2n− 1
(γy˜(t))2n= t(x− x0)2n+ (˜y + x0)2n+ 2t˜yn(φ(x)− φ(x0))
+ 2φ(x0)˜yn+ σ(t(x− x0) + x0, ˜y). (3.54)
dφ(x0)(x0, γy˜(t))≤ ˜y + ˜x +
˜y˜x +˜y|φ(x) − φ(x0)|, (3.55)
where x is˜ asin(3.52).Indeed,using(2.13)and (3.54)weget
dφ(x0)(x0, γy˜(t))≤ (t˜x1+ ˜y1, . . . , t˜x2n−1+ ˜y2n−1) R2n−1+ +t˜x2n+ ˜y2n+ 2t˜yn(φ(x)− φ(x0)) + tσ(˜x, ˜y) 1 2 ,
and (3.55)follows usingthetriangleinequality.Since, x,x0∈ Ωφ(¯x)(¯x,R) and˜y≤ R thenby(3.52) and
(2.22) weget
˜x ≤ CR, and dφ(x, x0)≤ CR (3.56)
forsomeconstantC = C(Lφ,Ωφ(¯x)(¯x, ¯R))> 0.Finally,by(3.55),(3.56)and(2.29) weconcludethat
γy˜(t)∈ Ωφ(x0)(x0, C1R) ∀t ∈ [0, 1]
forsomeC1= C1(Lφ,Ωφ(¯x)(¯x,R))> 0 andthethesisfollowswith C0= max(C,C1). 2
Proposition3.8.LetC0beasinProposition 3.7.Foreveryx¯∈ ω suchthatΩφ(¯x)(¯x,C0R) ω andforevery
x,x0∈ Ωφ(¯x)(¯x,R) and forevery y˜∈ ˜Ω(0,R) wehave
ψ(Expφ(x),x(˜y))− ψ(Expφ(x0),x0(˜y)) =
= 1 0 2n i=1 (Logφ(x0),x0(x))i∇ˆiφ(x0)ψ(γy˜(t))dt + K4(x, x0, ˜y) 1 0 ∂2nψ(γy˜(t))dt, (3.57) where K4(x, x0, ˜y) := 2(φ(x)− φ(x0))˜yn− 2σ(˜y, x − x0). (3.58)
The kernel K4 isofclassC∞ with respecttoy and˜ thefollowingestimateshold:
|K4(x, x0, ˜y)| ≤ 2(Lφ,Ωφ(¯x)(¯x,R)+ 1)dφ(x, x0)˜y, (3.59) |∇K4(x, x0, ˜y)| ≤ 2(Lφ,Ωφ(¯x)(¯x,R)+ 1)dφ(x, x0). (3.60)
Proof. Sinceψ∈ C∞(ω) andγy˜ishorizontalwithrespecttothefamilyofvectorfields{ ˆ∇φ(x0)},weobtain
ψ(Expφ(x),x(˜y))− ψ(Expφ(x0),x0(˜y)) = 1 0 (ψ◦ γy˜)(t)dt = 2n i=1 1 0 ˜ hi∇ˆφ(xi 0)ψ(γy˜(t))dt,
so that(3.57) immediately follows using (3.51). In order to prove(3.59) it suffices to observe that σ(x− x0,y)˜ ≤ dφ(x,x0)˜y. Moreover, since ∂y˜2nK4(x,x0,y)˜ = 0 itfollows that to prove(3.60) it is enough to
estimate theEuclidean gradientofK4 (withrespectto thevariable y).˜ Byadirectcomputationand using
∂y˜iK4(x, x0, ˜y) =−2(x − x0)n+i if i = 1, . . . , n− 1, ∂y˜nK4(x, x0, ˜y) = 2(φ(x)− φ(x0)),
∂y˜iK4(x, x0, ˜y) = 2(x− x0)i if i = n + 1, . . . , 2n− 1.
Hence|∇K4(x,x0,y)˜|≤ 2(Lφ,Ωφ(¯x)(¯x,R)+ 1)dφ(x,x0),whichisthethesis. 2
There is anaturalchange of variables, naturallyassociated to the curve γy˜ definedin Proposition 3.7.
Indeedthefollowinglemmaholds:
Lemma3.9. Letx¯∈ ω,C0> 0 andγy˜beasProposition 3.7.Foreach R > 0,suchthatΩφ(¯x)(¯x,C0R)⊂ ω,
x,x0∈ Ωφ(¯x)(¯x,R) and y˜∈ ˜Ω(0,R).Thenthefunction
H : [0, 1]× ˜Ω(0, R) −→ [0, 1] × ω
(t, ˜y)→ (t, γy˜(t))
has inverse function(t,F (z,˜ t)), themap z → (t,F (z,˜ t)) is C∞ andits Jacobian matrix has determinant equalto1.
Proof. Using(3.54),(2.10)and setting(t,z):= (t,γy˜(t)),F can˜ beexpressedas
˜
Fi(z, t) = (z− x0)i− t(x − x0)i i = 1, . . . , 2n− 1,
˜
F2n(z, t) = (z− x0)2n− t(x − x0)2n− 2t((z − x0)n− t(x − x0)n)(φ(x)− φ(x0))
− 2φ(x0)((z− x0)n− t(x − x0)n) + σ(z, t(x− x0) + x0). (3.61)
In particular it is clear from (3.61) that F is˜ of class C∞ as a function of the variable z and that the Jacobiandeterminantof z→ ˜F (z,t) is equalto 1 foreacht∈ [0,1]. 2
Lemma3.10. Letg∈ C∞(Rn) andF (z,˜ t) as inLemma 3.9 then
∇φ(x0) zi (g( ˜F (z, t)) = ⎧ ⎪ ⎨ ⎪ ⎩ (∇y˜ig)( ˜F (z, t)) i = 1, . . . , n− 1, (∇y˜ng)( ˜F (z, t))− 2t(φ(x) − φ(x0))(∂y˜2ng)( ˜F (z, t)) i = n, (∇y˜ig)( ˜F (z, t)) i = n + 1, . . . , 2n− 1, (3.62)
where(∇1,. . . ,∇2n−1) isthefamily ofvector fieldsdefined in(2.24).
Proof. Letusstartcomputing ∇φ(x0)
zi (g( ˜F (z,t))) withi= 1,. . . ,n− 1,thatis
∂zi− zi+n∂z2n
(g( ˜F (z, t))). (3.63)
Tothisend,wecalculate
∂zi(g( ˜F (z, t))) and ∂z2n(g( ˜F (z, t))).
Bytheexplicitexpression ofF (z,˜ t) weobtain:
∂zi(g( ˜F (z, t))) = (∂y˜ig)( ˜F (z, t)) + (∂y˜2ng)( ˜F (z, t))∂ziF˜2n(z, t), (3.64)
henceby(3.63), (3.64)and(3.65)weget ∇φ(x0) zi (g( ˜F (z, t))) = ∂y˜ig− ˜Fi+n(z, t)∂˜y2ng ( ˜F (z, t))+ + ˜ Fi+n(z, t)− zi+n+ ∂ziF˜2n(z, t) ∂y˜2ng( ˜F (z, t)). Since ˜ Fi(z, t) = (z− x0)i− t(x − x0)i i = 1, . . . , 2n− 1 ˜ F2n(z, t) = (z− x0)2n− t(x − x0)2n− 2t((z − x0)n− t(x − x0)n)(φ(x)− φ(x0)) + − 2φ(x0)((z− x0)n− t(x − x0)n) + σ(z, t(x− x0) + x0) (3.66) this implies ∇φ(x0) zi (g( ˜F (z, t))) = (∇y˜ig)( ˜F (z, t)).
Thecomputations for∇φ(x0)
zi (g( ˜F (z,t))) wheni= n+ 1,. . . ,2n− 1 aresimilar.
Finally,letuscompute∇φ(x0)
zn (g( ˜F (z,t))).Bydefinition: ∇φ(x0) zn (g( ˜F (z, t))) = ∂zn+ 2φ(x0)∂z2n (g( ˜F (z, t))) (3.67) and since ∂zn(g( ˜F (z, t))) = (∂y˜ng)( ˜F (z, t))− 2[t(φ(x) − φ(x0)) + φ(x0)](∂y˜2ng)( ˜F (z, t)) (3.68) by(3.67),(3.65)and(3.68) weget ∇φ(x0) zn g( ˜F (z, t)) = (∇y˜ng)( ˜F (z, t))− 2t(φ(x) − φ(x0))(∂y˜2ng)( ˜F (z, t)). 2
Proposition 3.11.Forevery t∈ [0,1],c1,c2> 0 and r > 0 letus define
Dt,c1,c2,r :=
z∈ R2n: (z, t)∈ ˜F−1Ω(0, c˜ 2r)− ˜Ω(0, c1r)
. (3.69)
Let x¯ ∈ ω and C0 > 0 be as in Proposition 3.7, then for every 0< R such that Ωφ(¯x)(¯x,C0R) ⊂ ω and
x,x0∈ Ωφ(¯x)(¯x,R) with x= x0 itholds m(ψ, φ, R)(x)− m(ψ, φ, R)(x0) = 2 R 1 0 R R 2 Dt, 1 2,1,r < K5(x, x0, t, z, r),∇φψ(z) > dL2n(z)drdt + 1 0 Dt, 1 2,1,R < K6(x, x0, t, z, R),∇φψ(z) > dL2n(z)dt − 1 0 Dt, 1 4, 12,R < K7(x, x0, t, z, R),∇φψ(z) > dL2n(z)dt
forsuitablekernelsK5,K6,K7definedin(3.74),(3.75)and(3.76)respectively.Moreover,therearepositive
|K5(x, x0, t, z, r)| ≤ ˜C5(Lφ+ 1)2 dφ(x0)(x0, x) rQ on Dt,12,1,r,∀t ∈ [0, 1], (3.70) |K6(x, x0, t, z, R)| ≤ ˜C6(Lφ+ 1)2 dφ(x0)(x0, x) R ˜F (z, t)Q−1 on Dt,12,1,R,∀t ∈ [0, 1], (3.71) |K7(x, x0, t, z, R)| ≤ ˜C6(Lφ+ 1)2 dφ(x0)(x0, x) R ˜F (z, t)Q−1 on Dt,14, 1 2,R,∀t ∈ [0, 1]. (3.72)
Proof. ByLemma 3.6forevery0< r < R0 suchthatΩφ(x)(x,r),Ωφ(x0)(x0,r) ω,wehave
¯ m(ψ, φ, r)(x)− ¯m(ψ, φ, r)(x0) = = 1 rQ ˜ Ω(0,r)\˜Ω(0,r 2) K3(0, ˜y)
ψ(Expφ(x),x(˜y))− ψ(Expφ(x0),x0(˜y))
dL2n(˜y) byProposition 3.7 = 1 rQ ˜ Ω(0,r)\˜Ω(0,r 2) K3(0, ˜y) 1 0 < Logφ(x0),x0(x), ˆ∇φ(x0)ψ(γ ˜ y(t)) > dtdL2n(˜y) + 1 rQ ˜ Ω(0,r)\˜Ω(0,r 2) K3(0, ˜y) 1 0 K4(x, x0, ˜y)∂2nψ(γy˜(t))dtdL2n(˜y).
Thechangeofvariablesz = γ˜y(t),changesΩ(0,˜ r)\ ˜Ω(0,r2) inthesetDt,1
2,1,r andtheinversemappinghas
Jacobiandeterminantequalto1.Hence
m(ψ, φ, R)(x)− m(ψ, φ, R)(x0) = = 2 R 1 0 R R 2 1 rQ Dt, 1 2,1,r K3(0, ˜F (z, t)) < Logφ(x0),x0(x), ˆ∇ φ(x0)ψ(z) > dL2n(z)drdt + + 2 R 1 0 R R 2 1 rQ Dt, 1 2,1,r K3(0, ˜F (z, t))K4(x, x0, ˜F (z, t))∂2nψ(z)dL2n(z)drdt. (3.73)
NowapplyingRemark 3.4wegetthethesiscalling:
K5(x, x0, t, z, r) := 1 rQK3(0, ˜F (z, t))Logφ(x0),x0(x) + + 1 2rQ∇ φ(x0) 1 K3(0, ˜F (z, t))K4(x, x0, ˜F (z, t)) en+1+ + 1 2rQ∇ φ(x0) n+1 K3(0, ˜F (z, t))K4(x, x0, ˜F (z, t)) e1; (3.74) K6(x, x0, t, z, R) := 1 R ∇φ(x0) 1 K( ˜F (z, t)) KQ( ˜F (z, t)) K3(0, ˜F (z, t))K4(x, x0, ˜F (z, t))en+1+ + 1 R ∇φ(x0) n+1 K( ˜F (z, t)) KQ( ˜F (z, t)) K3(0, ˜F (z, t))K4(x, x0, ˜F (z, t))e1; (3.75)
K7(x, x0, t, z, R) :=− 1 R ∇φ(x0) 1 K( ˜F (z, t)) 2QKQ( ˜F (z, t)) K3(0, ˜F (z, t))K4(x, x0, ˜F (z, t))en+1 + 1 R ∇φ(x0) n+1 K( ˜F (z, t)) 2QKQ( ˜F (z, t)) K3(0, ˜F (z, t))K4(x, x0, ˜F (z, t))e1, (3.76)
where asusualei denotesthei-thelementofthecanonicalbasisofR2n−1.Toprove(3.70)weobservethat
byLemma 3.10 K5(x, x0, t, z, r) = 1 rQK3(0, ˜F (z, t))Logφ(x0),x0(x) + 1 2rQ (∇1K3)(0, ˜F (z, t))K4(x, x0, ˜F (z, t)) + (∇1K4)(x, x0, ˜F (z, t))K3(0, ˜F (z, t)) en+1 + 1 2rQ (∇n+1K3)(0, ˜F (z, t))K4(x, x0, ˜F (z, t)) + (∇n+1K4)(x, x0, ˜F (z, t))K3(0, ˜F (z, t)) e1
henceusing(3.47),(3.59)and(3.60) weget
|K5(x, x0, t, z, r)| ≤ ˜ C3 rQdφ(x0)(x0, x) + 2 ˜ C4(Lφ+ 1)dφ(x, x0) rQ + 2 ˜ C3(Lφ+ 1)dφ(x, x0) rQ
and theconclusionfollows using(2.20).Finally,(3.71) and(3.72)aredirectconsequencesof(2.29),(3.47), (3.59) andLemma 3.10. 2
4. Poincaréinequality
Thescope ofthissectionis toproveTheorem 1.2.ThePoincarétypeinequalityprovedhereispartially inspired to the Sobolev type inequality for vector fields with non-regular coefficients contained in [23] and successively extended to a more general class of vector fields in [53]. The key point in our strategy, is to establish a representation formula for intrinsic Lipschitz continuous functions. To this end we use Theorem 2.6andtherepresentationformulaprovedinProposition 3.11forC∞functions.
Throughout this sectionwe denote by ω an open and bounded subset of R2n with n ≥ 2 andbyφ an
intrinsic Lipschitzfunctiondefinedonω withLipschitzconstantequaltoLφ.
Let ψ ∈ Wφ and let {ψk}k∈N, {φk}k∈N smooth functions on ω which satisfy conditions (i)− (iv) in
Definition 1.1.Wedenotebydφk(x0)thedistanceintroducedin(2.13),byΓφk(x0)thefundamentalsolution
oftheoperatorLφk(x0)definedin(2.30)andbyΩφk(x0)(x0,r) thesuperlevelsetofΓφk(x0)definedin(2.31).
Westartprovingthattheaveragem(ψ,¯ φ,R)(x0) canbeapproximatedbymeansoftheregularsequence
¯
m(ψk,φk,R)(x0).Precisely:
Lemma 4.1.Letx0∈ ω andR > 0 such that Ωφ(x0)(x0,R)⊂ ω.Then (i) χΩφk(x0)(x0,R)→ χΩφ(x0)(x0,R) uniformlyin ω ask→ +∞;
(ii) ¯m(ψk,φk,R)(x0)→ ¯m(ψ,φ,R)(x0) uniformly inR > 0 as k→ +∞.
HereχA denotesthecharacteristicfunctionof A.
Proof. Werecallthat
˜
Ω(0, R) =y˜∈ R2n | Γ(0, ˜y) > R2−Q,
Ωφk(x0)(x0, R) = Expφk(x0),x0( ˜Ω(0, R)).
Using the explicit form of Expφk(x0),x0 and Expφ(x0),x0 stated in (2.10) we easily conclude that
(Expφk(x0),x0)k∈NuniformlyconvergestoExpφ(x0),x0 inω ask→ +∞.Inordertoprove(i) weobservethat
itissufficienttoprovethatforall > 0 thereexists ¯k = ¯k( )> 0 suchthatforallk > ¯k
Ωφ(x0)(x0, R)⊆ (Ωφk(x0)(x0, R)) (4.77)
where
(Ωφk(x0)(x0, R)):={y ∈ ω | dφk(x0)(∂Ωφk(x0)(x0, R), y) < }. (4.78)
Forsimplifyingthenotationwedefine
Ek( ˜Ω(0, R)) := Expφk(x0),x0( ˜Ω(0, R)), E( ˜Ω(0, R)) := Expφ(x0),x0( ˜Ω(0, R)).
Supposebycontradictionthatthereexists > 0 suchthatforeveryk there¯ arek > ¯k andyk∈ E(˜Ω(0,R))
suchthatyk∈ E/ k( ˜Ω(0,R)).Then,thereexist(kj)j,kj→ +∞ asj→ +∞ and(xkj)j inΩ(0,˜ R) suchthat E(xkj) ∈ E/ kj( ˜Ω(0,R)). Sothat, the distance between E(xkj) and Ekj(xkj) isgreater than and this is
absurdbeing Ek uniformly convergenttoE.Then,(4.77) followsandhence(i).
Toprove(ii) weobservethatbyDefinition 3.3:
lim k→+∞m(ψ¯ k, φk, R)(x0) = = lim k→+∞ C(Q) Q − 2 1 RQ ω |∇φk(x0)Γ φk(x0)(x0, y)| 2 Γφk(x0)(x0, y)2(Q−1)/(Q−2) ψk(y) χΩφk(x0)(x0,R)\Ωφk(x0)(x0,R2)(y) dL 2n(y). By(2.29)and(2.30) lim k→+∞ |∇φk(x0)Γ φk(x0)(x0, y)| 2 Γφk(x0)(x0, y) 2(Q−1)/(Q−2) = |∇φ(x0)Γ φ(x0)(x0, y)| 2 Γφ(x0)(x0, y) 2(Q−1)/(Q−2) ≤ C ∀y = x0 (4.79)
therefore,(ii) followsfrom (i) andthefactthatψk→ ψ inL1loc(ω). 2
Inwhat follows weprove thattherepresentation formulas obtainedinProposition 3.5 and in Proposi-tion 3.11forC∞ functionsstillholdifφ isintrinsicLipschitzandψ∈ Wφ(ω).
Lemma4.2.Letφ beaLipschitzcontinuousfunctionandψ∈ Wφ(ω).Foreachx0∈ ω andeachR > 0 such
that Ωφ(x0)(x0,R)⊂ ω, thefollowingformula holds:
ψ(x0) = m(ψ, φ, R)(x0) + IR(x0), wherem(ψ,φ,R)(x0) is asin(3.40) and IR(x0) := 2 R R R 4 f1 r R Ωφ(x0)(x0,r) K1(x0, y),∇φψ(y) dL2n(y)dr +2 R R R 2 Ωφ(x0)(x0,r)\Ωφ(x0)(x0,r2) K2(x0, y, r),∇φψ(y) dL2n(y)dr, (4.80)
where K1, K2 are as in Proposition 3.5. Letx¯ ∈ ω and C0 > 0 be as in Proposition 3.7, then for every
0< R such that Ωφ(¯x)(¯x,C0R)⊂ ω andx,x0∈ Ωφ(¯x)(¯x,R) withx= x0 itholds:
m(ψ, φ, R)(x)− m(ψ, φ, R)(x0) = = 2 R 1 0 R R 2 Dt, 1 2,1,r < K5(x, x0, t, z, r),∇φψ(z) > dL2n(z)drdt + 1 0 Dt, 1 2,1,R < K6(x, x0, t, z, R),∇φψ(z) > dL2n(z)dt − 1 0 Dt, 1 4, 12,R < K7(x, x0, t, z, R),∇φψ(z) > dL2n(z)dt, (4.81)
whereDt,c1,c2,r isasin(3.69)andK5,K6,K7areasin(3.74),(3.75)and(3.76)respectivelyandtheysatisfy thesame estimatesprovedinProposition 3.11 withpossibly different constants.
Proof. Bydefinitionof1.1thereare{ψk}k∈N,{φk}k∈Nsequencesofsmoothfunctionsdefinedonω satisfying
conditions(i)–(iv) ofDefinition 1.1.ByPropositions 3.5 and3.11,thethesisistrueforeveryφk,ψkasabove.
Passing tothelimitasinthepreviousproposition,itholdstruealsoforthelimitfunctionsφ and ψ. 2
It is well known (see for example [14,21]) that the key step in the proof of the Poincaré inequality is a representation formula as the one proved in Lemma 4.2, which is indeed equivalent to the Poincaré inequality itself. For further applications,we note that we can obtainthe representation formula on any family of balls, equivalent to the superlevels Ωφ(¯x)(¯x,R),which canbe Ωφ(¯x)(¯x,R) or Uφ(¯x,R), defined
respectivelyin(2.31)and(1.5).
Let us denote by Bφ(¯x,R) a family of spheres centered at x and¯ radius R, equivalent to the family
Ωφ(¯x)(¯x,R).Let us denote by ψBφ(¯x,R) the mean of ψ on the set Bφ(¯x,R) with respect to the Lebesgue
measure,i.e. ψBφ(¯x,R):= 1 L2n(B φ(¯x, R)) Bφ(¯x,R) ψ(x) dL2n(x) (4.82)
we willprovethefollowing result:
Weuseourrepresentationformulatoproveanupperboundforthequantity|ψ(x0)−ψBφ(¯x,R)|,precisely:
Proposition 4.3. Let φ ∈ Lip(ω) and ψ ∈ Wφ(ω). Let x¯ ∈ ω and C0 > 0 be as in Proposition 3.7.
There are C˜0 > C0, C˜1,C˜2 > 0, depending only on Lφ,ω,Q and the structure constants in (2.29), such
that if Bφ(¯x,C˜0R) ω andx0∈ Bφ(¯x,R) then
|ψ(x0)− ψBφ(¯x,R)| ≤ ≤ ˜C1 Bφ(¯x, ˜C0R) d1φ−Q(x0, y)|∇φψ(y)|dL2n(y) + + C˜2 L2n(B φ(¯x, R)) Bφ(¯x, ˜C0R) Bφ(¯x, ˜C0R) d1φ−Q(x, y)|∇φψ(y)|dL2n(y)dL2n(x). (4.83)
Proof. Let R > 0 such that Bφ(¯x,R) ⊂ Ωφ(¯x)(¯x,C0R) ⊂ Bφ(¯x,C˜0R) ω. By Lemma 4.2, for each x,x0∈ Bφ(¯x,R) we have: ψ(x0) = m(ψ, φ, R)(x0) + IR(x0), ψ(x) = m(ψ, φ, R)(x) + IR(x), hence ψ(x0)− ψ(x) = m(ψ, φ, R)(x0)− m(ψ, φ, R)(x) + IR(x0)− IR(x). (4.84)
Integratingequation(4.84)withrespecttothevariablex onthesphereBφ(¯x,R) andrecallingthedefinition
ofψBφ(¯x,R) weget ψ(x0)− ψBφ(¯x,R)= 1 L2n(B φ(¯x, R)) Bφ(¯x,R) m(ψ, φ, R)(x0)− m(ψ, φ, R)(x) dL2n(x) + 1 L2n(B φ(¯x, R)) Bφ(¯x,R) IR(x0)− IR(x) dL2n(x). Hence: |ψ(x0)− ψBφ(¯x,R)| ≤ 1 L2n(B φ(¯x, R)) Bφ(¯x,R) m(ψ, φ, R)(x0)− m(ψ, φ, R)(x) dL2n(x) +|IR(x0)| + 1 L2n(B φ(¯x, R)) Bφ(¯x,R) |IR(x)| dL2n(x). (4.85)
Now,byLemma 4.2,wehave:
|m(ψ, φ, R)(x0)−m(ψ, φ, R)(x)| ≤ 2 R 1 0 R R 2 Dt, 1 2,1,r | < K5(x, x0, t, z, r),∇φψ(z) >|dL2n(z)drdt + 1 0 Dt, 1 2,1,R | < K6(x, x0, t, z, R),∇φψ(z) >|dL2n(z)dt + 1 0 Dt, 1 4, 12,R | < K7(x, x0, t, z, R),∇φψ(z) >|dL2n(z)dt.
Weclaimthatthere existsC = C(Lφ)> 0 suchthatforallr∈ (R/2,R),t∈ [0,1] itholds
Dt,1
2,1,r⊆ Ωφ(x0)(x0, CR)⊂ Bφ(¯x, ˜C0R). (4.86)
Tothisend letus fixt∈ [0,1] andr∈ (R/2,R) thenforeachy˜∈ ˜Ω(0,r)\ ˜Ω(0,r/2) wehave
r
and, by(3.55),italsoholds
dφ(x0)(x0, γy˜(t))≤ ˜y + ˜x +
˜y˜x +˜y|φ(x) − φ(x0)|. (4.88)
Since x,x0∈ Bφ(¯x,R) by(2.20)and (2.29)wehave
dφ(x, x0)≤ CR and ˜x ≤ CR (4.89)
forsomeC = C(Lφ)> 0.Using(4.87),(4.88)and(4.89)weimmediatelyget(4.86)withpossiblysmaller R.
By Lemma 4.2 we know that the estimates for K5, K6, K7 proved in Proposition 3.11 also hold for
φ∈ Lip(ω).Hence,by(3.70)and (4.87)foreachz∈ Dt,1
2,1,r andt∈ [0,1] wehave |K5(x, x0, t, z, r)| ≤ ˜C dφ(x0)(x0, x) rQ ≤ ˜C dφ(x0)(x0, x) ˜F (z, t)Q
forsomeC = ˜˜ C(Lφ)> 0.Using(4.86) and(4.89)weget
dφ(x0)(x0, x) =˜x ≤ CR ≤ 2Cr ≤ C ˜F (z, t)
forasuitableconstantC = C(Lφ)> 0.Then
|K5(x, x0, t, z, r)| ≤ C 1 ˜F (z, t)Q−1. Moreover, by(4.86),z∈ Ωφ(x0)(x0,CR) and 0 < dφ(x0)(x0, z)≤ CR ≤ 2Cr ≤ C ˜F (z, t). Sothat |K5(x, x0, t, z, r)| ≤ Cdφ(x0)(x0, z) 1−Q. (4.90)
Analogously, wecanprovethatthere existsC = C(Lφ)> 0 such that
|K6(x, x0, t, z, r)|, |K7(x, x0, t, z, r)| ≤ Cdφ(x0)(x0, z)
1−Q. (4.91)
Inconclusionweprovedthat
|m(ψ, φ, R)(x0)− m(ψ, φ, R)(x)| ≤ ≤ ˜C1 Ωφ(x0)x0,CR d1φ(x−Q 0)(x0, z)|∇ φψ(z)|dL2n(z). (4.92)
Furthermore, byLemma 4.2,(3.41) and(3.42)wehave
|IR(x0)| ≤ ˜C2Lφ Ωφ(x0)(x0,R) d1φ(x−Q 0)(x0, y)|∇ φψ(y)|dL2n(y), (4.93) |IR(x)| ≤ ˜C2Lφ Ωφ(x)(x,R) d1φ(x)−Q(x, y)|∇φψ(y)|dL2n(y). (4.94)