BV functions and sets of finite perimeter in sub-Riemannian manifolds

(1)

ScienceDirect

Ann. I. H. Poincaré – AN 32 (2015) 489–517

www.elsevier.com/locate/anihpc

BV functions and sets of finite perimeter in sub-Riemannian manifolds

^✩

L. Ambrosio

^a

, R. Ghezzi

^a,b,^∗

, V. Magnani

^c

aScuola Normale Superiore, Piazza dei Cavalieri 7, 56126 Pisa, Italy

bInstitut de Mathématiques de Bourgogne, 9 Avenue Alain Savary, BP47870 21078 Dijon Cedex, France cDipartimento Università di Pisa, Largo Bruno Pontecorvo 5, 56127 Pisa, Italy

Received 3 June 2013; received in revised form 21 January 2014; accepted 23 January 2014 Available online 11 February 2014

Abstract

We give a notion of BV function on an oriented manifold where a volume form and a family of lower semicontinuous quadratic formsGp:TpM→ [0,∞]are given. When we consider sub-Riemannian manifolds, our definition coincides with the one given in the more general context of metric measure spaces which are doubling and support a Poincaré inequality. We focus on finite perimeter sets, i.e., sets whose characteristic function is BV, in sub-Riemannian manifolds. Under an assumption on the nilpotent approximation, we prove a blowup theorem, generalizing the one obtained for step-2 Carnot groups in[24].

1. Introduction

Sub-Riemannian manifolds are a class of length spaces of non-Euclidean type having a differentiable structure. Our interest in studying functions of bounded variation in this framework arises from the aim of understanding the structure of finite perimeter sets in the general sub-Riemannian setting. This clearly requires suitable notions of “intrinsically regular” hypersurfaces, rectifiability, reduced boundary and blowups.

Sobolev and BV functions have been investigated inRⁿendowed with the Lebesgue measure and with the Carnot–

Carathéodory distancedccassociated with a family of vector fields. Under the assumption that the family is Lie bracket generating, the Lebesgue measure is doubling with respect to the Carnot–Carathéodory distance[37], a Poincaré inequality holds[30]and(Rⁿ, dcc)is complete[16]. These are the main assumptions which enable the authors in[25]

to establish Sobolev and isoperimetric inequalities as well as an approximation theorem of Meyers–Serrin type (see also[22]for a related result with weaker regularity assumptions on the vector fields).

✩ This work was supported by the European research project AdG ERC “GeMeThNES”, grant agreement number 246923, see alsohttp://

gemethnes.sns.it.

* Corresponding author.

E-mail addresses:l.ambrosio@sns.it(L. Ambrosio),roberta.ghezzi@u-bourgogne.fr,roberta.ghezzi@sns.it(R. Ghezzi), magnani@dm.unipi.it(V. Magnani).

http://dx.doi.org/10.1016/j.anihpc.2014.01.005

(2)

Our main goal is to develop a systematic theory of BV functions and sets of finite perimeter in manifolds with suitable structures. To this aim, one needs two ingredients: first a volume measure (with respect to which an integration by parts formula will hold); second a notion of length of tangent vectors (along which one calculates distributional derivatives). For the volume measure, when a manifoldMis oriented, it suffices to take a non-degeneraten-formω which induces the orientation ofM(wheren=dimM) and consider the measuremdefined on Borel setsB⊂Mby m(B)=

Bω. Given an open setΩ⊂Mand a vector fieldX, a functionu∈L¹(Ω,m)has distributional derivative alongXif there exists a Radon measureDXuonΩsuch that

Ω

ϕ dD_Xu= −

Ω

uϕdivωXω−

Ω

u(Xϕ)ω, ∀ϕ∈Cc^∞(Ω),

where divωX is defined in (5). To define the length of tangent vectors, we use a family of lower semicontinuous quadratic formsG_p:T_pM→ [0,∞]defined on the tangent bundle ofM. Note that the dimension of the vector space D(p)= {v∈T_pM|G_p(v) <∞}may vary with respect to the point. Taking this into account, it is natural to say that a functionu∈L¹(Ω,m)has bounded variation inΩ if, for every smooth vector fieldX such thatG_p(X(p))1, p∈Ω, the distributional derivativeD_Xuis a Radon measure of finite total variation inΩand

D_gu(Ω):=sup|D_Xu|(Ω) <∞, (1)

the supremum being taken among all smooth vector fields such thatG(X)1 onΩ. Thus, we writeu∈BV(Ω, g, ω), wheregis the scalar product associated withG. More precisely, for each p∈Mwe have thatg_p(·,·)is the unique scalar product onD(p)such thatg_p(v, v)=G_p(v)for everyv∈D(p), Section2.3for more information.

In this quite general setting, distributional derivatives can be weakly approximated by derivatives of smooth functions, that is, a Meyers–Serrin theorem holds (seeTheorem 2.4). Moreover, the fact thatDXu is a Radon measure with finite bounded variation is characterized in terms of difference quotients along the flow generated byX (see Theorem 2.5). UsingG, one can define the Carnot–Carathéodory distancedcc between points ofMas the infimum of lengths of absolutely continuous curves connecting the two points, where length of tangent vectors is computed with respect toG. Whendcc is finite,(M, dcc,m)is a metric measure space. It is then natural to compare the space BV(Ω, g, ω)with the notion of BV function in a metric measure space developed in[7,36](see also[6]). Without further assumptions, we can only show that BV(Ω, dcc,m)is embedded in BV(Ω, g, ω)(seeTheorem 2.7).

Oriented sub-Riemannian manifolds, whereGis induced locally by bracket generating families of vector fields, cast in the framework above. In this case, on coordinate charts,Gis given by

G_p(v)=inf _m

i=1

c²_i v= m i=1

c_iX_i(p)

(with the convention inf∅ = ∞) whereX₁, . . . , X_mplay the role of orthonormal frame and dim(X1(p), . . . , X_m(p)) may vary with respect top. In particular, the aforementioned notion of BV space encompasses the classical one in oriented Riemannian manifolds, the one associated with a Lie bracket generating family of vector fields inRⁿand it also includes the rank-varying case, e.g. the Grushin case and almost-Riemannian manifolds (see Section3.1).

In this setting, the approximation result (Theorem 2.4) allows to show that the metric and differential notion of bounded variation coincide and the corresponding spaces BV(Ω, g, ω) and BV(Ω, dcc,m) are isometric (Theo- rem 3.1). A first consequence of this fact is that the set functionDgu defined on open sets as in(1) is a Borel measure. Moreover, the Riesz theorem of Euclidean setting (see for instance[20, Theorem 1, p. 167]) can be gener- alized. More precisely, ifu∈BV(Ω, g, ω), then there exists a Borel vector fieldν_usatisfyingG(ν_u)=1,D_gu-a.e.

inΩ. Moreover, for every smooth vector fieldXsuch thatG(X)1 inΩ, the distributional derivative ofualongX can be represented as

DXu=g(X, νu)Dgu. (2)

Without assumptions on the dimension ofD(p), even if a local basisX1, . . . , XminducingGis given, some care is needed, due to the fact that a smooth vector fieldXsatisfyingG(X)1 is in general a linear combination of theX_i with coefficient inL^∞only.

(3)

When we consider sets of finite perimeter, i.e., sets whose characteristic function has bounded variation, this result provides a notion of geometric normal (which corresponds to its Euclidean analogue) and which is a horizontal Borel vector field ofG-length 1.

In Euclidean metric spaces, the structure of finite perimeter sets has been completely understood since De Giorgi’s seminal works[18,19]. In this context, ifEhas finite perimeter, then the perimeter measure ofEis concentrated on a set which is rectifiable and it has codimension one. The main step behind this result is a blowup theorem, showing that whenpbelongs to the reduced boundary ofE, the sequence of blowupsE_r=δ_1/r(E+p)converges to a halfspace inL¹_loc.

In non-Euclidean metric spaces, after[31], rectifiability theory has been developed in Banach spaces[12,13], and in Carnot groups[23,35,38], which are Lie groups whose algebra admits a stratification with respect to a one parameter group of dilationsδ_r. More precisely, in[24]the authors generalize De Giorgi’s theorem in Carnot groups of step 2.

Their proof is inspired by the one in the Euclidean case. Moreover, Carnot groups are homogeneous and this makes the perimeter measure both homogeneous with respect to dilations and invariant by translations. Joining these properties with isoperimetric inequalities along with the compact embedding in BV shows that bounded sequences of rescaled sets have converging subsequences and the blowups are both monotone along a horizontal direction and invariant along all orthogonal directions. The techniques of[24]have been further extended in[17]to a special class of Lie groups that do not possess dilations, seeExample 3.3, where the “linearization” of the left invariant vector fields is obtained by the Rothschild–Stein lifting theorem,[39].

As a first step toward a generalization of De Giorgi’s theorem in sub-Riemannian manifolds, in this paper we show a blowup theorem which generalizes the one[24, Theorem 3.1]in Carnot groups of step 2. Again, the proof ofTheorem 4.2is inspired by the corresponding one in the Euclidean case. However, the rationale behind our proof is somehow different from the one in[24]. Without a Carnot group structure, the main idea is to exploit two well known facts in sub-Riemannian geometry [15]: a metric tangent cone to the manifold at a point p always exists;

the quasi-isometry between dilated balls centered atpand balls in the metric tangent cone can be given explicitly by a system of suitable coordinates (called privileged, seeDefinition 3.3) centered at p and, in particular, it is a diffeomorphismϕ_p. In this coordinate system, there exists a subgroup of dilationsδ_r intrinsically associated with the sub-Riemannian structure (and centered atp). Hence, given a finite perimeter set Eandp in its reduced boundary (seeDefinition 3.2), readingEthroughϕ_p, it makes sense to consider the blowupsE_r=δ_1/rϕ_p(E). Our result states that if the metric tangent cone to(M, d_cc)atpis a Carnot group of step 2 then

L¹_loc- lim

r↓01E_r=1F,

whereF is the vertical halfspace in the Carnot group associated with the geometric normalν_E(p)(for the precise statement, seeTheorem 4.2.) In particular, we are able to show that the sequence of blowups{1Er}r>0 is compact inL¹_loc and that if 1_F_˜ is a L¹_loc-limit thenF˜ is monotone along the geometric normal ν_E(p)and invariant along orthogonal directions toν_E(p), in the distributional sense. To prove compactness, we exploit the fact that the distance in the metric tangent cone is the limit of Carnot–Carathéodory distances associated with a sequence of “perturbed”

vector fields (seeTheorem 3.5). Properties of limits are consequences of the definition of geometric normal and of the asymptotically doubling property of the perimeter measure. Finally, it is only at this stage of the proof that we invoke the fact that the metric tangent cone atpis a Carnot group of step 2, to show that the limit of the rescaled sets actually exists. The latter assumption is fulfilled of course when the manifold itself is a Carnot group of step 2 but also in more general case, e.g. in the step 2 equiregular case (see alsoExample 3.7). This hypothesis is essential as it was shown in[24]that in Carnot groups of step higher than 2 the blowup at a point in the reduced boundary need not be a vertical hyperplane. Without bounds on the step, the only result available so far is[14], where it has been proved that at almost every point (with respect to the perimeter measure) there exists a subsequence of blowups converging to a vertical halfspace.

Let us mention an application of our result in the rank-varying case. As we see inExample 3.5, there exist sub- Riemannian manifolds with the property: for every pointpthe metric tangent cone atpis either the Euclidean space or a Carnot group of step 2. For these manifolds, which are also called almost-Riemannian, the horizontal distribution is rank-varying and it has full rank at points where the tangent cone is the Euclidean space. Combining our theorem with the one in the Euclidean case, we obtain that, for sets of finite perimeter in these manifolds, the blowups at points in the reduced boundary converge to a halfspace.

(4)

Another important corollary of our blowup theorem is that, settingh(B_r(p))=^m(B_r^r^(p)) whereB_r(p)is the open ball ford_cc, the density

limr↓0

D_g1E(B_r(p))

h(B_r(p)) (3)

exists and equals to the perimeter ofFin the unit ball in the Carnot group divided by the Lebesgue measure of the unit ball. This improves the weaker estimates on ^D^g_h(B¹^E^(B^r^(p))

r(p)) which have been proved in[7, Theorem 5.4]in the metric setting. Moreover, denoting byS^h the spherical measure build by the Carathéodory’s construction (see[21, 2.10.1]) withhas gauge function, the existence of the limit in(3)implies upper and lower bounds on the Radon–Nikodym ofDg1Ewith respect toS^h (restricted to the reduced boundary ofE). As the Radon–Nikodym derivative of the perimeter measureDg1Ewith respect toS^h has been shown to exist in the metric context (see[7, Theorem 5.3]), an open question is whether this derivative coincides with (3) for D_g1E-almost every p. In the constant rank (equiregular) case, a related result in[2]computes the density of the spherical top-dimensional Hausdorff measure S_d^Q_cc (whereQis the Hausdorff dimension of any ball) with respect tomin terms of the Lebesgue measure of the unit ball in the metric tangent cone.

The paper is organized as follows. We define distributional derivatives along vector fields in manifolds with a volume form in Section2.2. Using a family of metrics in the tangent bundle we then define the space of BV functions in Section2.3and we prove an approximation results for distributional derivatives. In Section2.4we show that BV(Ω, dcc,m)is continuously embedded in BV(Ω, g, ω). Section3is a primer in sub-Riemannian geometry. Even though our main results are local, we find it useful to recall the general definition of sub-Riemannian structure that relies on images of Euclidean vector bundles and includes the rank-varying case. In Section 3.1we list some sig- nificative examples, including Carnot groups. In Section3.2we analyze the notion of BV space in sub-Riemannian manifolds. First, we specify its relation with the BV space defined in a metric measure space, showing that the two BV spaces are actually isometric. Second, we prove a Riesz theorem which generalize the Euclidean analogue. Section3.3 recalls the notion (and basic properties) of privileged coordinates and nilpotent approximation (for more details we refer the reader to[15].) In Section3.4we explain the relation between nilpotent approximations and metric tangent cones to sub-Riemannian manifold at a point and, under an additional assumption at the point, we recall how to show that the nilpotent approximation is isometric to a Carnot group endowed with the control distance induced by a left invariant metric on the horizontal bundle. In Section4we prove the blowup theorem. We split the proof into two main parts. In Sections4.1and4.2we demonstrate compactness of the dilated sets and monotone and invariance properties of limits. Then, in Section 4.3we use the assumption on the nilpotent approximation to provide the existence and characterize the limit of dilated sets as a vertical halfspace.

2. Preliminaries

2.1. Basic notation and notions

Given a setA⊂B, we will use the notation1A:B→ {0,1}for the characteristic function ofA, equal to 1 onA and equal to 0 onB\A. In a metric space(X, d), the notationB_r(x)will be used to denote the open ball with radius rand centerx.

Differential notions. Throughout this paper,Mdenotes a smooth, oriented, connected andn-dimensional manifold, with tangent bundleT M. The fiberT_xM can be read as the space of derivations on germs of C¹ functionsϕ atx, namely[vϕ](x)=dϕ_x(v), forv∈T_xM. In the same spirit, we read the action of the differentialdF_x:T_xM→T_xN of aC¹mapF:M→N as follows:

df_x(v)(ϕ)=v(ϕ◦F )(x), ϕ∈C¹(N ).

AnyC¹diffeomorphismF:M→Nbetween smooth manifolds induces an operatorF_∗:T M→T N, by the formula (F_∗X)(F (x))=dF_x(X(x)); equivalently, in terms of derivations, we write

(F_∗X)ϕ ◦F=

X(ϕ◦F ) , ∀ϕ∈C¹(N ).

Measure-theoretic notions. IfF is aσ-algebra of subsets ofXandμ:F→R^mis aσ-additive measure, we shall denote by |μ|:F→ [0,∞)its total variation, still a σ-additive measure. By the Radon–Nikodym theorem, μ is

(5)

representable in the formg|μ|for someF-measurable functiong:X→R^msatisfying|g(x)| =1 for|μ|-a.e.x∈X.

Given a Borel mapF, we shall use the notationF_#for the induced push-forward operator between Borel measures, namely

F_#μ(B):=μ

F⁻¹(B)

, for allBBorel.

2.2. Volume form, divergence and distributional derivatives

We assume throughout this paper thatMis endowed with a smoothn-formω. We assume also that

Mf ω >0 wheneverf ∈C_c¹(M)is nonnegative and not identically 0, so that thevolume measure

m(E)=

E

ω, E⊂MBorel (4)

is well defined and, in local coordinates, has a smooth and positive density with respect to Lebesgue measure. Ac- cordingly, we shall also callωvolume form.

The volume form ωallows to define thedivergencedivωX of a smooth vector fieldX as the smooth function satisfying

divωXω=L_Xω, (5)

whereL_Xdenotes the Lie derivative alongX. Using properties of exterior derivative and differential forms, we remark that divωXis characterized by

−

M

ϕdivωXω=

M

(Xϕ)ω, ∀ϕ∈C_c¹(M). (6)

By applying this identity to a productf ϕwithf ∈C¹(M)andϕ∈C_c^∞(M), the Leibnitz rule gives

−

M

f ϕdivωXω−

M

f (Xϕ)ω=

M

ϕ(Xf )ω, ∀ϕ∈C_c^∞(M). (7)

We can now use this identity to defineXf also as a distribution onM, namelyDXf =gin the sense of distributions in an open setΩ⊂Mmeans

−

Ω

f ϕdivωXω−

Ω

f (Xϕ)ω=

Ω

ϕgω, ∀ϕ∈C^∞c (Ω). (8)

Our main interest focuses on the theory of BV functions along vector fields. For this reason, we say that a measure with finite total variation inΩ, that we shall denote byD_Xf, represents inΩthe derivative off alongXin the sense of distributions if

−

Ω

f ϕdivωXω−

Ω

f (Xϕ)ω=

Ω

ϕ dD_Xf, ∀ϕ∈Cc^∞(Ω). (9)

By(4)and(7), whenf isC¹we haveD_Xf =(Xf )m.

We can now state a simple criterion for the existence ofD_Xf, a direct consequence of Riesz representation theorem of the dual ofCc(Ω). In order to state our integrations by parts formulas(8),(9)in a more compact form we also use the identity

divω(ϕX)=ϕdivωX+Xϕ.

Proposition 2.1.LetΩ⊂M be an open set and f ∈L¹_loc(Ω,m). Then D_Xf is a signed measure with finite total variation inΩ if and only if

(6)

sup

Ω

fdivω(ϕX)ωϕ∈Cc^∞(Ω), |ϕ|1

<∞.

If this happens, the supremum above equals|DXf|(Ω).

A direct consequence of this proposition is the lower semicontinuity inL¹_loc(Ω,m)off → |DXf|(Ω). We also emphasize that, thanks to(9), we have the identity

D_ψXf =ψ D_Xf, ∀ψ∈C^∞(Ω), (10)

and the properties(8)and(9)need only to be checked for test functionsϕwhose support is contained in a chart.

2.3. Distributions, metrics andBVfunctions on manifolds

OnMwe shall consider a family of lower semicontinuous quadratic (i.e. 2-homogeneous, null in 0 and satisfying the parallelogram identity) formsG_x:T_xM→ [0,∞]and the induced familyDof subspaces

D(x):=

v∈T_xMG_x(v) <∞ .

We are not making at this stage any assumption on the dimension ofD(x)(which need not be locally constant) or on the regularity ofx→G_x. We shall only assume that the map(x, v)→G_x(v)is Borel. This notion can be easily introduced, for instance using local coordinates. Obviously G_x induces a scalar productg_x on D(x), namely the unique bilinear form onD(x)satisfying

g_x(v, v)=G_x(v), ∀v∈D(x).

ForΩ⊂Mopen, we shall denote byΓ (Ω,D)the smooth sections ofD, namely:

Γ (Ω,D):=

XXis smooth vector field inΩ, X(x)∈D(x)for allx∈Ω . We shall also denote

Γ^g(Ω,D):=

X∈Γ (Ω,D)gx

X(x), X(x)

1,∀x∈Ω .

Note that bothDandgare somehow encoded byG. Thus the following definition of BV space only depends on Gandω.

Definition 2.1(SpaceBV(Ω, g, ω)and sets of finite perimeter).LetΩ⊂Mbe an open set andu∈L¹(Ω,m). We say thatuhasbounded variationinΩ, and writeu∈BV(Ω, g, ω), ifD_Xuexists for allX∈Γ^g(Ω,D)and

sup

|DXu|(Ω)X∈Γ^g(Ω,D)

<∞. (11)

We denote byDguthe associated Radon measure onΩ. IfE⊂Mis a Borel set, we say thatEhasfinite perimeter inΩif1E∈BV(Ω, g, ω).

Remark 1. Let us point out that replacingΩ in(11)with any of its open subsets, we get a set function on open sets. Thus, the so-called De Giorgi–Letta criterion, see for instance[11, Theorem 1.53], allows us to extend this set function to a Radon measure onΩ.

Equivalently, thanks toProposition 2.1, we can write condition(11)as follows:

sup

Ω

udivω(ϕX)ωX∈Γ^g(Ω,D), ϕ∈C_c^∞(M), |ϕ|1

<∞. (12)

These BV classes can be read in local coordinates thanks to the following result (more generally, one could consider smooth maps between manifolds).

(7)

Proposition 2.2.LetΩ ⊂M and U⊂Rⁿ be open sets, let u∈L¹(Ω)and let φ∈C^∞(Ω, U )be an orientation- preserving diffeomorphism with inverseψ. For eachy∈U, set

G˜_y(w):=G_ψ(y)

d_yψ (w)

, ∀w∈T_xURⁿ, u(y)˜ :=u ψ (y)

(13) and defineD˜ inUand a metricg˜inD˜ accordingly. Then, settingω˜:=ψ^∗ω, the following holds

φ#(DXu)=Dφ_∗Xu,˜ ∀X∈Γ (Ω,D) (14)

in the sense of distributions. In particularu∈BV(Ω, g, ω)if and only ifu˜∈BV(U,g,˜ ω).˜

Proof. LetXbe a smooth section ofDwith compact support contained inΩand consider the change of variable

Ω

udivωXω=

U

˜

u(divωX)◦ψω.˜ ByLemma 2.3below, it follows that

Ω

udivωXω=

U

˜

udivω_˜(φ_∗X)ω,˜

where we denote by divω_˜Y the divergence ofY with respect toω. If we apply this identity to the vector field˜ ϕXand use the relationφ_∗(ϕX)=(ϕ◦φ⁻¹)φ_∗Xwe obtain(14).

Finally, since for ally∈Uwe haveG˜_y(φ_∗X(y))1 if and only if Gψ(y)

dyψ

φ_∗X(y)

=Gψ(y)

X ψ (y)

1

the claim follows. 2

Lemma 2.3.Under the assumptions ofProposition2.2, the following formula holds

(divωX)◦ψ=divω_˜(φ_∗X). (15)

Proof. Ifϕ∈C_c^∞(Ω), then a change of variable in the oriented integral yields

−

Ω

ϕdivωXω=

Ω

Xϕω=

U

(Xϕ)◦ψω˜ =

U

(ψ_∗X)ϕ˜ ◦ψω,˜

where the last equality is a consequence of the definitionX˜ =φ_∗X. The previous equalities give

−

U

ϕdivωXω=

U

X(ϕ˜ ◦ψ )ω˜= −

U

ϕ◦ψdiv_ω_˜X˜ω˜= −

Ω

ϕ(div_ω_˜X)˜ ◦φω.

The arbitrary choice ofϕproves the validity of(15). 2

Distributional derivatives ofL¹functions as defined in(9)can be weakly approximated by derivatives of smooth functions as we prove in the next theorem.

Theorem 2.4(Meyers–Serrin). LetΩ⊂Mbe open,u∈L¹(Ω,m). Then there existu_n∈C^∞(Ω)convergent touin L¹_loc(Ω,m)and satisfyingXu_nm→D_Xuand |Xu_n|m→ |D_Xu|weakly, in the duality withCc(Ω), for all smooth vector fieldsXinΩ such thatD_Xuexists. The same is true if we consider vector-valued measures built with finitely many vector fields.

Proof. By a partition of unity it is not restrictive to assume that Ω is well contained in a local chart. Then, by Proposition 2.2, possibly replacinggandωby their counterpartsg˜andω, we can assume that˜ Ω⊂Rⁿ. In this case, writingω= ¯ω dx₁∧· · ·∧dx_nwithω¯ ∈C^∞(Ω)¯ strictly positive, we also notice that it is not restrictive to assumeω¯≡1

(8)

inΩ; indeed, comparing(6)with the classical integration by parts formula inΩwith no weight, we immediately see that

divωX=divX+Xlogω,¯

where, in the right hand side, divXis the Euclidean divergence ofX. One can then compare the integration by parts formulas in the weighted and in the classical case to obtain thatD_Xudepends onωthrough the factorω. This is also¯ evident in the smooth case, where the functionXuis clearly independent ofω, butD_Xu=(Xu)m.

After this reductions to the standard Euclidean setting, we fix an even, smooth convolution kernelρ inRⁿ with compact support and denote by ρ_ε(x)=ε⁻ⁿρ(x/ε)the rescaled kernels and byu∗ρ_ε the mollified functions. We shall use the so-called commutator lemma (see for instance[9]) which ensures

(D_Xv)∗ρ_ε−X(v∗ρ_ε)→0, asε↓0, strongly inL¹_loc(Ω) (16)

wheneverv∈L¹_loc(Ω)andXis a smooth vector field inΩ. Since(D_Xu)∗ρ_εmand|(D_Xu)∗ρ_ε|mconverge in the duality withC_c(Ω)toD_Xuand|D_Xu|respectively (see for instance[11, Theorem 2.2]),(16)shows that the same is true for X(u∗ρε)mand|X(u∗ρε)|m. The same proof works for vector-valued measures (convergence of total variations, the only thing that cannot be obtained arguing componentwise, is still covered by[11, Theorem 2.2]). 2 The following result provides a characterization of |D_Xu| in terms of difference quotients involving the flows generated by the vector field X (for similar results in the context of doubling metric spaces supporting a Poincaré inequality, see[36]).

Given a smooth vector fieldXinΩ⊂Mopen, we denote byΦ_t^Xthe flow generated byXonM. By compactness, for any compact setK⊂Ω the flow map starting fromKis smooth, remains in a domainΩΩand is defined for everyt∈ [−T , T], withT =T (K, X) >0. Recall that the JacobianJ Φ_t^Xof the flow mapx→Φ_t^X(x)is the smooth functionJ Φ_t^Xsatisfying(Φ_t^X)^∗ω=J Φ_t^Xω, so that the change of variables formula

φω=

φ◦Φ_t^XJ Φ_t^Xω

holds. By smoothness, there is a further constantC depending only onX(andT) such thatJ Φ_t^Xsatisfies

J Φ_t^X(x)−1C|t|, J Φ_t^X(x)−1−tdivωX(x)Ct², ∀x∈K, t∈ [−T , T]. (17) Estimate(17)is a simple consequence of Liouville’s theorem, showing that the time derivative oft→log(J Φ_t^X(x)) equals(divωX)(Φ_t^X(x)).

Theorem 2.5. Let Ω⊂M be an open set and let u∈L¹(Ω,m). ThenDXu is a signed measure with finite total variation inΩif and only if

sup

K⊂Ωcompact

K

|u(Φ_t^X)−u|

|t| ω0<|t|T (K, X)

<∞. (18)

Moreover, ifD_Xuis a signed measure with finite total variation inΩ, it holds

|D_Xu|(Ω)=sup

lim inf

t→0

Ω

|u(Φ_t^X)−u|

|t| ωΩΩ

. (19)

Proof. Let us prove that the existence ofDXuimplies(18). For 0<|t|< T (K, X)we will prove the more precise estimate

K

|u(Φ_t^X)−u|

|t| ω

1+C|t|

|D_Xu|(Ω_t) withΩ_t:=

r∈[0,|t|]

Φ_r^X(K), (20)

which also yields the inequalityin(19). In order to prove(20)we can assume with no loss of generality, thanks to Theorem 2.4, thatu∈C^∞(Ω). Under this assumption, since the derivative oft→u(Φ_t^X(x))equalsXu(Φ_t^X(x))we can use Fubini’s theorem and(17)to get, fort >0 (the caset <0 being similar)

(9)

K

u Φ_t^X

−uω t 0

K

|Xu| Φ_r^X

ω dr t 0

(1+Cr)

Ω_t

|Xu|ω dr.

Estimating(1+Cr)with(1+Ct )we obtain(20).

Let us prove the inequalityin(19). By the lower semicontinuity on open sets of the total variation of measures under weak convergence and the inner regularity of|D_Xu|, it suffices to show that for all ΩΩ the difference quotientst⁻¹(u(Φ_t^X)−u)m weakly converge, in the duality with C_c(Ω), toD_Xu. By the upper bound (18)we need only to check the convergence onC_c^∞(Ω)test functions. This latter convergence is a direct consequence of the identity (which comes from the change of variablesx=Φ₋^X_t(y))

Ω

u(Φ_t^X)−u

t ϕ dω(x)= −

Ω

ϕ(Φ₋^X_t)J Φ₋^X_t−ϕ

−t u dω(y), ϕ∈C_c^∞ Ω

,

of the expansion(17)and of the very definition ofD_Xu. The same argument can be used to show that finiteness of the supremum in(18)implies the existence ofDXu. 2

2.4. BV functions in metric measure spaces(X, d,m)

Let(X, d)be a metric space, and define forf:X→Rthelocal Lipschitz constant(also called slope) by

|∇f|(x):=lim sup

y→x

|f (y)−f (x)|

d(y, x) . (21)

If we have also a reference Borel measuremin(X, d), we can define the space BV(Ω, d,m)as follows.

Definition 2.2.LetΩ⊂Xbe open andu∈L¹(Ω,m). We say thatu∈BV(Ω, d,m)if there exist locally Lipschitz functionsunconvergent touinL¹(Ω,m), such that

lim sup

n

Ω

|∇u_n|dm<+∞.

Then, we define Du(Ω):=inf

lim inf

n→∞

Ω

|∇u_n|dmu_n∈Lip_loc(Ω), lim

n

Ω

|u_n−u|dm=0

.

In locally compact spaces, in[36](see also[10]for more general spaces) it is proved that, foru∈BV(Ω, d,m), the set function A→ Du(A) is the restriction to open sets of a finite Borel measure, still denoted by Du. Furthermore, in[36]the following inner regularity is proved:

u∈L¹(Ω)∩BVloc(Ω, d,m), sup

ΩΩ

Du Ω

<∞ ⇒ u∈BV(Ω, d,m). (22)

In the next sections we shall use a fine property of sets of finite perimeters, proved within the metric theory in[7]

(see also[6]for the Ahlfors regular case) to be sure that the set of “good” blow-up points has full measure with respect toD_g1E. The basic assumptions on the metric measure structure needed for the validity of the result are (in local form):

(i) a local doubling assumption, namely for allK⊂Xcompact there existr >¯ 0 andC0 such thatm(B2r(x)) Cm(Br(x))for allx∈Kandr∈(0,r);¯

(ii) a local Poincaré inequality, namely for allK⊂Xcompact there existr, c, λ >¯ 0 such that

B_r(x)

|u−ux,r|dmcr

B_λr(x)

|∇u|dm (23)

for allulocally Lipschitz,x∈Kandr∈(0,r), with¯ u_x,r equal to the mean value ofuonB_r(x).

(10)

Proposition 2.6.(See[7].) Assume that(i),(ii)above hold and letE⊂Ω be such that1E∈BV(Ω, d,m). Then lim inf

r↓0

min{m(Br(x)∩E),m(Br(x)\E)}

m(Br(x)) >0, lim sup

r↓0

D1E(Br(x))

h(B_r(x)) <∞ (24)

forD1E-a.e.x∈Ω, whereh(Br(x))=m(Br(x))/r.

In order to applyProposition 2.6for our blow-up analysis, we need to provide a bridge between the metric theory outlined above and the differential theory described in Section2.3. As a matter of fact, we will see that in the setting of Section2.3, under mild assumptions, we have always an inclusion between these spaces, and a corresponding inequality betweenD_guandDu. First, given a functionG:T M→ [0,∞]in a smooth manifoldM as in Section2.3, we define the associatedCarnot–Carathéodory distanceby

d_cc(x, y):=inf T

0

√G_γ_t(γ˙_t)T >0, γ0=x, γ_T =y

, (25)

where the infimum runs among all absolutely continuous curvesγ. Notice that sinceG_xis infinite onT_xM\D(x), automatically the minimization is restricted tohorizontalcurves, i.e. the curvesγ satisfyingγ˙∈D(γ )a.e. in[0, T].

We will often use, when convergence arguments are involved, an equivalent definition ofdcc(x, y)in terms of action minimization:

d_cc²(x, y):=inf 1

0

Gγ_t(γ˙t)γ0=x, γ1=y

. (26)

Notice also that we cannot expectd_ccto be finite in general, hence we adopt the convention inf∅ = +∞. However, anyX∈Γ^g(Ω,D)induces, via the flow mapΦ_t^X, admissible curvesγin(25)with speedG(γ )˙ less than 1, for initial points inΩ and|t|sufficiently small. It follows immediately that forΩΩ and|t|sufficiently small, depending only onΩandX, it holds:

d_cc

Φ_t^X(x), x

|t|, ∀x∈Ω. (27)

Theorem 2.7. Let m be defined as in (4) and assume that the distance d_cc defined in (25)is finite and induces the same topology ofM. Then for any open setΩ⊂Mwe have the inclusion BV(Ω, dcc,m)⊂BV(Ω, g, ω)and Dgu(Ω)Du(Ω).

Proof. Take a locally Lipschitz functionuinΩ (with respect tod_cc) with

Ω|∇u|dmfinite, X∈Γ^g(Ω,D)and apply(27)to obtain that|u(Φ_t^X(x))−u(x)|/|t|is uniformly bounded as|t| ↓0 on compact subsets ofΩand

lim sup

t↓0

|u(Φ_t^X(x))−u(x)|

|t| |∇u|(x), ∀x∈Ω.

By integrating onΩΩwe get fromTheorem 2.5thatD_Xuis a measure with finite total variation inΩ and that

|DXu|(Ω)

Ω

|∇u|dm.

Eventually, we apply the very definition of BV(Ω, dcc,m)to extend this inequality to allu∈BV(Ω, dcc,m), in the form|D_Xu|(Ω)Du(Ω). We can now use the arbitrariness ofXto get the conclusion. 2

Remark 1.The main obstacle to the inclusion BV(Ω, g, ω)⊂BV(Ω, dcc,m)is that, wheneveru∈BV(Ω, g, ω), the smooth approximationu_nofugiven inTheorem 2.4need not satisfy the weak convergenceD_gu_n → D_gu. This is due to the fact that the supremum in(11)is taken over an infinite set of generatorsX. Nevertheless, while the general validity of the inclusion BV(Ω, g, ω)⊂BV(Ω, dcc,m)is still an open problem, this difficulty can be bypassed for a large class of metrics G, namely the one of sub-Riemannian structures on manifolds. Indeed inTheorem 3.1we prove that in the sub-Riemannian context the equality BV(Ω, g, ω)=BV(Ω, dcc,m)holds along withD_gu(Ω)= Du(Ω).

(11)

3. Sub-Riemannian manifolds

A frame free approach to describe sub-Riemannian structures locally generated by families of vector fields[37]

relies on images of Euclidean vector bundles. Recall that a Euclidean vector bundle is a vector bundle whose fiber at a pointxis equipped with a scalar product·,·xwhich depends smoothly onx(in particular smooth orthonormal bases locally exist).

Definition 3.1(Sub-Riemannian structure).A sub-Riemannian structure onMis a pair(U, f )whereUis a Euclidean vector bundle overM andf:U→T M is a morphism of vector bundles (i.e., a smooth map, linear on fibers, such thatf (Ux)⊂T_xM, whereUxdenotes the fiber ofUoverx) such that

Liex(D)=T_xM, ∀x∈M, (28)

where we have set D:=

f◦σσ∈Γ (U)

, (29)

andΓ (U):= {σ∈C^∞(M,U)|σ (x)∈Ux}.

Given a sub-Riemannian structure onM, we denote byD(x)the vector spacef (Ux)⊂T_xM and we define the quadratic formsG_x:T_xM→ [0,∞]by

G_x(v)=

min{|u|²x|u∈Ux, f (u)=v}, v∈D(x),

+∞, v /∈D(x). (30)

Notice that in general the dimension ofD(x)need not be constant and(x, v)→G_x(v)is lower semicontinuous. Let g_x:D(x)×D(x)→Mbe the unique scalar product satisfying

g_x(v, v)=G_x(v), ∀v∈D(x).

We shall denote byPx:D(x)→Ux the linear map which associates withv the unique vector u∈f⁻¹(v)having minimal norm. It can be computed intersectingf⁻¹(v)with the orthogonal to the kernel off|Ux.

We will often computeGin local coordinates as follows: letσ₁, . . . , σ_mbe an orthonormal frame forU|Ω (where m=rankU) in an open setΩ⊂M. Then, definingX_j=f ◦σ_j, for everyx∈Ω andv∈D(x)we have

G_x(v)=min _m

i=1

c_i²v= m i=1

c_iX_i(x)

. (31)

In this case we shall also viewPxas anR^m-valued map. Notice that, by polarization, it holds gx(v, w)=

Px(v),Px(w)

, v, w∈D(x). (32)

Remark 2.The above definition includes the cases (see also Section3.1for more examples):

• Uis a subbundle ofT Mandf is the inclusion. In this case the distributionDhas constant rank, i.e., dimD(x)= dimUx=rankU. WhenU=T M andf is the identity, we recover the definition of Riemannian manifold.

• Uis the trivial bundle of rankmonM, i.e.,Uis isomorphic toM×R^mandDis globally generated bymvector fields f ◦e₁, . . . , f ◦e_k, where e_j(x)=(x,e¯_j)ande¯₁, . . . ,e¯_m is the canonical basis ofR^m; in particular, we recover the case whenM=Ω⊂Rⁿand we takemvector fields satisfying the Hörmander condition.

The finiteness of the Carnot–Carathéodory distanced(·,·)induced byGas in(25)(note that we will drop from now on theccin(25)and(26)) is guaranteed by the Lie bracket generating assumption onD(see[5]), as well as the fact thatdinduces the topology ofMas differentiable manifold. The metric space(M, d)is called aCarnot–Carathéodory space.

SinceDis Lie bracket generating, at every pointx∈Mthere existsk_x∈Nsuch that theflag atx associated with Dstabilizes (with stepk_x), that is,

(12)

{0}D¹(x)⊂D²(x)⊂ · · · ⊂D^k^x(x)=T_xM, (33) whereD¹(x)=D(x)andDⁱ⁺¹(x)=(Dⁱ+[D,Dⁱ])(x). The minimum integerk_xsuch that(33)holds is calleddegree of non-holonomyatx. With the flag(33)we associate two nondecreasing sequences of integers defined as follows. The growth vectoris the sequence(n₁(x), . . . , n_k_x(x)), wheren_i(x)=dimDⁱ(x). Notice that for everyx∈M,n_k_x(x)=n.

To define the second sequence, letv1, . . . , vn∈TxMbe a basis ofTxMlinearly adapted to the flag(33). The vector of weightsis the sequence(w1(x), . . . , wn(x))defined bywj(x)=sifvj∈D^s(x)\D^s⁻¹(x). Notice that this definition does not depend on the choice of the adapted basis, and that 1=w₁(x)· · ·w_n(x)=k_x.

We say that a pointx∈Misregularif the growth vector is constant in a neighborhood ofx, otherwise we say that x issingular. If every point is regular, we say that the sub-Riemannian manifold isequiregular.

3.1. Examples

In this section we mention some examples of sub-Riemannian manifolds. A first fundamental class of examples is provided by Carnot groups.

Example 3.1(Carnot groups).Let us consider a connected, simply connected and nilpotent Lie groupG, whose Lie algebragadmits a stepsstratification

g=V₁⊕V₂⊕ · · · ⊕V_s,

namely[V₁, V_j] =V_j₊₁for everyj=1, . . . , s−1 and[V₁, V_s] = {0}. Every layerV_j ofgdefines at each pointx∈G the following fiber of degreej atx

H_x^j=

Y (x)∈TxGY ∈Vj

.

All fibers of degreej are collected into a subbundleH^j ofTGfor everyj =1, . . . , s. Fix a scalar product·,·eon H_e¹. Then this canonically defines a scalar product·,·x onH_x¹by left invariance. In this way we endow the vector bundle H¹with a Euclidean structure. In the language ofDefinition 3.1, the inclusioni:H¹→TGdefines a left invariant sub-Riemannian structure onG. According to(29), the corresponding moduleDis precisely made by all smooth sections ofH¹, the so-called horizontal vector fields andD(x)=H_x¹for everyx∈G. The validity of(28) is ensured by the assumption thatgis stratified. Note that different choices of scalar product onH_e¹define Lipschitz equivalent sub-Riemannian structures onG.

The group Gequipped with a left invariant sub-Riemannian structure is called Carnot group. Let mbe the dimension ofV₁,nthe dimension ofG, and let(y₁, . . . , y_n)be a system of graded coordinates onG. An equivalent way to define a left invariant sub-Riemannian structure onGis the following. Fix a basisX₁, . . . , X_mofV₁with the following form

X_j(y)=∂_j+ n i=m+1

a_{j i}(y)∂_i for everyj=1, . . . , m,

wherea_{j i} is a homogeneous polynomial such thata_{j i}(δ_ry)=r^ωⁱ⁻¹a_{j i}(y)for everyy∈Gandr >0, whereδ_ry= n

j=1r^ω^jej and the degreeωj is defined by the conditionej∈Vω_j for everyj=1, . . . , n. Under these coordinates, we take the Euclidean vector bundleU=G×R^mand

f:G×R^m→TG, f (y, ξ )=

y, m j=1

ξ_jX_j(y)

.

Example 3.2 (Heisenberg group). The Heisenberg group is a special instance of a step 2 Carnot group. It can be represented in the language ofDefinition 3.1asR³equipped with the left invariant vector fields

X1=∂1−x₂

2∂3, X2=∂2+x₁ 2∂3,

with respect to some polynomial group operation. We haveV₁=span{X₁, X₂} ⊂handV₂=span{X₃} ⊂h, where his the 3-dimensional Lie algebra of left invariant vector fields andX₃=∂₃. Following the previous general case of Carnot groups, we setU=R³×R², equipped by the morphismf defined by