Notions of Convexity in Carnot Groups

communications in
analysis and geometry

Volume 11, Number 2, 263-341, 2003
Notions of Convexity in Carnot Groups

Donatella Danielli, Nicola Garofalo, and Duy-Minh Nhieu
CONTENTS
1. Introduction
2. Preliminaries
3. Strongly convex functions: A notion to abandon
4. Mildly convex functions: The appearance of horizontal planes
5. Weakly convex functions: The horizontal Hessian
6. Weak convexity of the gauge in groups of Heisenberg type
7. Convexity of sets
8. Further properties of second derivatives of weakly convex functions
9. Basic properties of first derivatives of weakly convex functions: L∞ − L1
estimates
10. The horizontal Monge-Ampère equation
11. Yet another notion of convexity
References
1. Introduction.
In 1993 motivated by some interesting questions in the theory of fully non-

linear equations, we proposed the study of the following problem. Consider a
stratified, nilpotent Lie group G of step r, with Lie algebra g = V1 ⊕ ... ⊕ Vr ,
and let X1 , ..., Xm be a fixed orthonormal basis of the first layer V1 . Continue
to denote with X1 , ..., Xm the corresponding system of left-invariant vector
fields on G. Given a m × m matrix-valued function g → A(g) = (aij (g)) on
G having real measurable entries aij = aji , suppose in addition that there
exist ν > 0 such that

m
ν |ζ|2 ≤ aij (g) ζi ζj ≤ ν −1 |ζ|2 ,
i,j=1
for a. e. g ∈ G, and for every ζ ∈ Rm . The stratification of the Lie algebra

carries a natural family of non-isotropic dilations ∆λ ξ = λξ1 +λ2 ξ2 +...+λr ξr ,
1
Supported in part by NSF Grant No. DMS-0070492
2
Third author was supported in part by NSC Grant No. 89-2115-M-001-018
263
264 D. Danielli, N. Garofalo, and D.-M. Nhieu

r
for any ξ = ξ1 + ξ2 + ... + ξr ∈ g. Let Q = j dim Vj be the homogeneous
j=1
dimension of G associated with such dilations.
Conjecture: Given a connected, bounded open set Ω ⊂ G, let f ∈ LQ (Ω).
Suppose that u ∈ L2,Q
loc (Ω) ∩ C(Ω) satisfy

def 1
m
Lu = aij XiXj u + Xj Xi u ≥ F
2
i,j=1
in Ω. There exists a constant C = C(G, ν, Ω) > 0 such that
(1.1) sup u ≤ sup u+ + C ||F ||LQ (Ω) .

Ω ∂Ω
Here, L2,Q Q
loc (Ω) indicates the Sobolev space of functions u ∈ Lloc (Ω) having
weak derivatives Xi Xj u ∈ LQloc (Ω). We note explicitly that

2
Lu = tr A X u ,

where we have denoted by X 2 u the symmetrized horizontal Hessian of u,
see (1.3), or Section 5. Concerning the optimality of the LQ norm in the
estimate (1.1), we refer the reader to [DGN4]. In the abelian case, when
G = Rn with the standard 2 homogeneous dilations, one has g = V1 = Rn ,
2
so that Q = n, and X u = D u, the standard Hessian matrix of u.
The above conjecture, in this situation, is in fact the celebrated geometric
maximum principle of Alexandrov-Bakelman-Pucci (except that the matrix
(aij ) need not be uniformly elliptic), see [A], [Ba1], [Pu], and also [Ba3].
Given the pervasive role that convexity plays in such principle, in order
to advance our comprehension of the above conjecture it seems natural to
examine appropriate notions of convexity in the non-abelian setting. This is
the goal of the present paper.
We next give a description of our results. The notion of weakly H-convex
function set forth in Definition 5.5 is central to our work. Sections 3 and 4
are devoted to justifying how we arrived to such notion starting from that of
Pansu differential. A function on a Carnot group u : G → R is called weakly
H-convex if for any g ∈ G and every λ ∈ [0, 1] one has

(1.2) u g δλ (g −1 g ) ≤ (1 − λ) u(g) + λ u(g ) for every g ∈ Hg .
In (1.2) we have denoted by {δλ }λ>0 the non-isotropic group dilations

defined by δλ = exp ◦∆λ ◦ exp−1 , see (2.9), whereas the point gδλ (g −1 g )
Notions of convexity in Carnot groups 265
represents the twisted convex combination of g and g , with base point g. The
notation Hg indicates the horizontal plane passing through g, see Definition
4.1. The collection of all such planes determines the horizontal subbundle
generated by a given orthonormal basis of the Lie algebra g. The trajectories
of the dilations gδλ (g −1 g ) are in general quite twisted and they badly fail
to be sub-unitary. They possess however the important redeeming feature
that when the end-point g belongs to the horizontal plane through g, then
such trajectories are just straight line segments, at least at the level of the
Lie algebra, see Proposition 4.4.
This leads us to our first main result, Theorem 5.12, which states that a
function u in the Folland-Stein class Γ2 (G) is weakly H-convex if and only
if its symmetrized horizontal Hessian
1
(1.3) X 2 u(g) = X 2 u(g) + X 2 u(g)T ,
2
is positive semi-definite at every g ∈ G. Here,
X 2 u(g) = (XiXj u(g))i,j=1,...,m .
This result constitutes the sub-Riemannian counterpart to the classical char-

acterization of convexity for C 2 functions. In this perspective it represents
a crucial witness of the appropriateness of the notion of weak convexity, and
it is central to the rest of the paper.
In Section 6 we study an important example of non-smooth weakly H-
convex function. In Theorem 6.8 we prove that the intrinsic gauge in any
group of Heisenberg type is weakly H-convex. This is a deep property of
such groups, intimately connected with their symmetries, see Theorem 6.6.
It would be interesting to know whether such property extends to arbitrary
Carnot groups of step 2, but we presently ignore the answer.
Section 7 is devoted to studying the convexity of sets. Definition 7.1
derives in a natural way from the notion of weakly H-convex function. With
such definition, we prove that a function is weakly H-convex if and only if
such is its epigraph as a subset of the product Carnot group G = G × R.
The main result of Section 8 is Theorem 8.1, which states that if an
1
Lloc (G) function u is weakly H-convex, then the entries of its symmetrized
horizontal Hessian are signed Radon measure. This property extends the
well-known classical analogue for convex functions in Euclidean space. It
plays an important role in the development of a notion of Alexandrov gen-
eralized solution of the Monge-Ampère equation in the sub-Riemannian set-
ting, see [DGN3].
The notion of weakly H-convex function evolved in a natural way, so

to speak, from our desire to identify a class as wide as possible, while still
retaining the basic properties connected with the notion of Pansu differen-
tiability. Its appropriateness is proved by Theorem 5.12. Yet, such notion
is a major source of complications when one tries to extend the basic theo-
rems of convex analysis and new ideas are necessary. To clarify this point,
we emphasize the main obstructions connected with Definition 5.5. (i) It
involves information about the function u only the lower-dimensional man-
ifold Hg . (ii) Furthermore, such information is one-sided since we cannot
exchange the role of the base point g with that of the point g ∈ Hg . (iii)
More importantly, failure of He at being a subgroup of G, results in failure
of the crucial transitivity property : g2 ∈ Hg1 , g1 ∈ Hgo , implies g2 ∈ Hgo .
This phenomenon of inferring conclusions on a set of full measure from in-
formation on lower-dimensional sets is fully present in the above conjecture
and represents a major obstacle to overcome.
In Section 9 we introduce a new idea in this direction which we be-
lieve will prove important also in the study of the geometric Alexandrov-
Bakelman-Pucci maximum principle conjectured above. First, we establish
Theorem 9.1 which shows that a weakly H-convex continuous function is in
fact Lipschitz continuous with respect to the sub-Riemannian metric of G.
Furthermore, there exists C = C(G) > 0 such that the following estimate
holds
C
(1.4) ||Xu||L∞(B(go,R)) ≤ ||u||L∞(B(go,3R)) ,
R
for any go ∈ G and every R > 0. This result follows from a quantitative
characterization of Lipschitz continuity which was obtained in [GN2], and
by suitably slicing the gauge pseudo-ball B(go , R) with horizontal planes.
The main result of the section, which we also consider the central result
of the paper, is Theorem 9.2. The latter states that there exists an absolute
constant C = C(G) > 0 such that for any function u weakly H-convex in G
and continuous, one has
1
(1.5) sup |u| ≤ C |u| dg ,
B(g,r) |B(g, 5r)| B(g,5r)
and
C 1
(1.6) ess sup |Xu| ≤ |u| dg .
B(g,r) r |B(g, 15r)| B(g,15r)
The main trust of the inequalities (1.5), (1.6) is that their right-hand
sides involve a set of full measure. The way we accomplish this result is to
exploit in an ad hoc way the weak convexity of u along the one-parameter

family of horizontal planes {Hγ(s) | −1 ≤ s ≤ 1}, based at points of the
straight-line “segment” γ(s) = exp sX1 , s ∈ [−1, 1]. Here, X1 is the first
element of an orthonormal basis of the first-layer V1 of g (the specific choice
of X1 , rather than any other element of the basis is, of course, absolutely
immaterial). The proof consists in first establishing a suitable control of
φ(s) = u(γ(s)) on the lower dimensional sets B(γ(s), 1) ∩ Hγ(s) , and then
average the resulting inequality in s ∈ [−1, 1]. The crucial geometric fact
which plays in our favor in this last step is that the union of all the regions
B(γ(s), 1) ∩ Hγ(s) covers a butterfly which is of full measure.
One of the fundamental aspects of the theory of convex functions is their
connection with geometry through the Monge-Ampère equation det D2 u =
f. Such fully nonlinear equation becomes elliptic on strictly convex func-
tions. In this perspective, it is natural to ask if there is a corresponding
fully nonlinear operator attached to the notion of weak convexity in the
sub-Riemannian setting. We discuss this question in Section 10, where we
introduce the notion of horizontal Gaussian curvature for both the graph of
a function u : G → R, and for a smooth manifold S = ∂Ω given as the level
set Ω = {g ∈ G | φ(g) < 0} of a defining function φ. In the former case,
we define the horizontal Gaussian curvature of the graph of u at the point
(go , u(go)) as the classical Gaussian curvature of the graph of the restriction
Ugo of u to the horizontal plane Hgo . Such notion also works well for a level
set, provided that the point go ∈ S be non-characteristic. Such new point
of view was suggested by our approach to the minimal surface equation in
[DGN2], and to the horizontal Bernstein problem, formulated in the same
paper. For level sets one encounters difficulties when the point go is charac-
teristic, and this requires a deeper analysis of the asymptotic properties of
the horizontal principal curvatures. For these aspects we refer the reader to
the papers [DGN2], [CGP].
A systematic study of the horizontal Monge-Ampère equation in Carnot
groups
(1.7) det [(X 2u) ] = F ,
will be carried in a forthcoming paper [DGN3]. We also introduce the no-

tion of horizontal normal map and establish some of its properties, and we
propose a notion of viscosity solution for the equation (1.7). The latter is
inspired to the classical one of Crandall-Ishii-Lions [CIL], but, once again,
it strongly reflects those new difficulties which are intrinsic to the notion of
weak convexity. Section 10 ends with Theorem 10.9 which, in essence, states
that non-isotropic gauge in a group of Heisenberg type is a fundamental gen-

eralized solution of the horizontal Monge-Ampère equation. This result is
one further witness of the deep geometric role played by the gauge in these
ambients.
Section 11 closes the paper. The most elementary properties of classical
convex functions are their being locally bounded and continuous on their
effective domain. Corresponding properties for weakly H-convex functions
are far from obvious and are presently being investigated. The purpose of
this section is to prove that by strengthening Definition 5.5, but of course
without falling back into the notion of strong H-convexity, one obtains an
“intermediate” class of convex functions which, remarkably, share with the
classical ones the above mentioned properties. This implies in particular
that the results in Sections 8 and 9 are valid generically for such functions,
without any additional assumption.
Acknowledgments: It is a pleasure to acknowledge several helpful and

stimulating conversations with Luca Capogna on the subject of this paper.
After this paper was submitted we had a conversation with Luis Caffarelli
in which he brought to our attention that for the Heisenberg group H1 a
notion of convexity essentially equivalent to our Definition 5.5 had been
formulated by him in his 1997 NSF Proposal. We thank him for bringing
this fact to our attention, and also for kindly sharing with us the notes from
his proposal.
This paper was submitted in April 2002, and it was simultaneously posted
to the server CVGMT of the Scuola Normale Superiore, Pisa. A related
paper has been posted to the same server by Lu, Manfredi and Stroffolini
in June 2002. The notion of convexity adopted in [LMS] is a reformulation
of Caffarelli’s, and thereby equivalent to our notion of weak H-convexity.
Several results in [LMS] overlap with ours, but the methods of proof are
different.
2. Preliminaries.
We begin by introducing the relevant ambients and their many properties.

A Carnot group G of step r is a simply connected Lie group whose Lie
algebra g admits a nilpotent stratification of step r. This means that g =
V1 ⊕ V2 ⊕ · · · ⊕ Vr , and that moreover [V1 , Vj ] = Vj+1 for j = 1, ..., r − 1,
whereas [V1 , Vr ] = {0}. We assume that a scalar product < ·, · > is given
on g for which the Vj s are mutually orthogonal. We let mj = dim Vj ,
j = 1, ..., r, and denote by
N = m1 + ... + mr
the topological dimension of G. The notation {Xj,1 , ..., Xj,mj }, j = 1, ..., r,

will indicate a fixed orthonormal basis of the j − th layer Vj . Elements of Vj
are assigned the formal degree j. As a rule, we will use letters g, g , go for
points in G, whereas we will reserve the letters Z, Z , Zo, for elements of the
Lie algebra g. We will denote by
(2.1) Lgo (g) = go g , Rgo (g) = g go ,
respectively, the left- and right-translations on G by an element go ∈ G.

Recall that the exponential map exp : g → G is a global analytic diffeomor-
phism [V]. It allows to define analytic maps ξi : G → Vi , i = 1, ..., r, by
letting g = exp(ξ1 (g) + ... + ξr (g)). The mapping ξ : G → g defined by
ξ(g) = ξ1 (g) + ... + ξr (g) ,
is the inverse of the exponential mapping. For g ∈ G, the projection of the

exponential coordinates of g onto the layer Vj , j = 1, ..., r, are defined as
follows
(2.2) xj,s (g) = < ξj (g), Xj,s >, s = 1, ..., mj.
Since Carnot groups of step 2 will play a special role it will be convenient
to have a separate notation for the first two layers V1 and V2 . For simplicity,
we set m = m1 , k = m2 , and indicate
(2.3) X = {X1 , ... , Xm } = {X1,1 , ... , X1,m },

Y = {Y1 , ... , Yk } = {X2,1, ..., X2,k }.
We continue to denote by X and Y the corresponding systems of left-

invariant vector fields on G defined by
Xj (g) = (Lg )∗ (Xj ) , j = 1, ..., m, Yl (g) = (Lg )∗ (Yl ) , l = 1, ..., k ,
where (Lg )∗ denotes the differential of Lg . The system X defines a basis for
the so-called horizontal sub-bundle HG of the tangent bundle T G. For a
given function f : G → R, the action of Xj on f is specified by the equation
f (g exp (tXj )) − f (g) d

(2.4) Xj f (g) = lim = f (g exp (tXj )) t=0
.
t→0 t dt
A similar formula holds for any left-invariant vector field. We indicate

with
(2.5)
xj (g) = < ξ1 (g), Xj >, j = 1, ..., m, ys (g) = < ξ2 (g), Ys >, s = 1, ..., k.
the projections of the exponential coordinates of g onto V1 and V2 . Letting
x(g) = (x1 (g), ..., xm(g)), y(g) = (y1 (g), ..., yk(g)), we will often identify
g ∈ G with its exponential coordinates
(2.6) g = (x(g), y(g), ...),
where the dots indicate the (N − (m + k))-dimensional vector
(x3,1 (g), ..., x3,m3 (g), ..., xr,1(g), ..., xr,mr (g)).
When G is a group of step 2, then (2.6) simply becomes g = (x(g), y(g)).
Such identification of G with its Lie algebra is justified by the Baker-
Campbell-Hausdorff formula, see, e.g., [V]
1
(2.7) exp Z exp Z = exp (Z + Z + [Z, Z ] + ... ) Z, Z ∈ g,
2
where the dots indicate a finite linear combination of terms containing com-
mutators of order two and higher. For Z ∈ g consider the map θZ : g → g
given by
1
θZ (Z ) = Z + Z + [Z, Z ] + ...
2
where the right-hand side is given by the Baker-Campbell-Hausdorff sum in
(2.7). If we endow the Lie algebra g with the polynomial group law
(2.8) Z ◦ Z = θZ (Z ),
then we can identify the group G with g, via the exponential coordinates.
A Carnot group of step r is naturally equipped with a family of non-
isotropic dilations defined by
(2.9) δλ (g) = exp ◦ ∆λ ◦ exp−1 (g), g ∈ G,
where exp : g → G is the exponential map and ∆λ : g → g is defined by
∆λ (X1 + ... + Xr ) = λX1 + ... + λr Xr .
The most basic partial differential operator in a Carnot group is the sub-
Laplacian associated with X is the second-order partial differential operator
on G given by
m
L = Xj2 .
j=1
By the assumption on the Lie algebra one immediately sees that the
system X satisfies the finite rank condition in Hörmander’s theorem [H1],
therefore the operator L is hypoelliptic. However, when r > 1 it fails to be
elliptic, and the loss of regularity is measured by the step r of the stratifica-
tion of g.
For a function u on G we let
Xu = X1 u X1 + ... + Xm u Xm ,
m
and set |Xu| = ( j=1 (Xj u)2 )1/2 . We denote by dg the push-forward of
Lebesgue measure on g via the exponential map. Such dg defines a bi-
invariant Haar measure on G. One has d(g ◦ δλ) = λQdg, so that the number
Q = m1 + 2 m2 + ... + r mr
plays the role of a dimension with respect to the group dilations. For this
reason Q is called the homogeneous dimension of G. Such number is larger
than the topological dimension N of G defined above.
The Euclidean distance to the origin | · | on g induces a homogeneous
pseudo-norm | · |g on g and (via the exponential map) one on the group G in
the following way (see [F2], [FS]). For ξ ∈ g, with ξ = ξ1 + ... + ξr , ξi ∈ Vi ,
we let
2r!
r
2r!/i
(2.10) |ξ|g = |ξi| ,
i=1
and then define a pseudo-norm on G by the equation

(2.11) N (g) = NG(g) = |ξ|g , if g = exp ξ .
The function N is usually referred to as the non-isotropic gauge. It defines
a pseudo-distance on G
(2.12) d(g, g ) = N (g −1g ) .
This is called the gauge pseudo-distance, and it is equivalent to the
Carnot-Carathéodory distance ρ(·, ·) generated by the system X, i.e., there
exists a constant C = C(G) > 0 such that
(2.13) C ρ(g, g ) ≤ d(g, g ) ≤ C −1 ρ(g, g ), g, g ∈ G ,
see [Be]. If B(g, R) = {g ∈ G | d(g, g ) < R}, then by left-translation and
dilation it is easy to see that the Haar measure of B(g, R) is
(2.14) |B(g, R)| = ω RQ ,
where ω = |B(e, 1)|, with e indicating the group identity. A similar transla-
tion invariance holds for the metric balls Bρ (g, R).
The Heisenberg group. A model of Carnot group of paramount impor-
tance is the Heisenberg group Hn , see [S], [G]. The underlying manifold of
this Lie group is simply R2n+1 , with the non-commutative group law
(2.15)
g g = (x, y, t) (x , y , t ) = (x + x , y + y , t + t + 2(< x , y > − < x, y >)) ,
where we have let x, x , y, y ∈ Rn , t, t ∈ R. Let (Lg )∗ be the differential of
the left-translation (2.15). A simple computation shows that

∂ def ∂ ∂
(2.16) (Lg )∗ = Xj = + 2 yj , j = 1, ..., n ,
∂xj ∂xj ∂t

∂ def ∂ ∂
(Lg )∗ = Xn+j = − 2 xj , j = 1, ..., n ,
∂yj ∂yj ∂t

∂ def ∂
(Lg )∗ = X2n+1 =
∂t ∂t
We note that the only non-trivial commutator is
[Xj , Xn+k ] = − 4 X2n+1 δjk , j, k = 1, ..., n ,
therefore the vector fields {X1 , ..., X2n} constitute a basis of the Lie algebra
hn = R2n+1 = V1 ⊕ V2 , where V1 = R2n × {0}t, V2 = {0}(x,y) × R. The
non-isotropic group dilations are therefore
(2.17) δλ (g) = (λx, λy, λ2t) .
The associated homogeneous dimension is Q = 2n + 2, and the gauge is
given by
1/4
N (g) = (|x|2 + |y|2 )2 + t2 .
We notice that the sub-Laplacian associated with the basis {X1 , ..., X2n}
is
2n
n
∂2 ∂ ∂ ∂
(2.18) L = Xj2 2 2
= ∆x,y +4 (|x| +|y| ) 2 +4 yj −xj ,
∂t ∂t ∂xj ∂yj
j=1 j=1
which coincides with the real part of the complex Kohn Laplacian, see [S].
Returning to general Carnot groups, we recall that for a given open set
Ω ⊂ G, the classes Γ1 (Ω) (respectively, Γ2 (Ω)) introduced by Folland and
Stein [FS], [F2], represent the collection of all functions u : Ω → R such that
the derivatives Xα u (respectively, XαXα u), α, α = 1, ..., m, exist and are
continuous functions in Ω. We also need the definition of Lipschitz function.
Definition 2.1. Let Ω ⊂ G be an open set. A function u : Ω → R is said

to belong to the class Γ0,1 (Ω) if there exists a positive constant L > 0 such
that
(2.19) |u(g) − u(g )| ≤ L d(g, g ), g, g ∈ Ω.
The meaning of the symbol Γ0,1 0,1

loc (Ω) is the obvious one, i.e., u ∈ Γloc (Ω)
if for every ω ⊂⊂ Ω one has u ∈ Γ0,1 (ω). If F ⊂ G denotes a closed set,
by u ∈ Γ0,1 (F ) we mean that u coincides on the set F with a function
g ∈ Γ0,1 (Ω), where Ω is a bounded open set containing F . The condition
(2.19) in the Definition 2.1 measures the smoothness of u in the intrinsic
pseudo-metric d(g, g ) on G associated with the system X = {X1 ,. . . ,Xm }
(of course, using the distance ρ(g, g ), instead of d(g, g ), in view of (2.13)
would produce the same functional classes). The corresponding regularity of
u with respect to the underlying Riemannian metric dR (g, g ) of G is much
lower (depending on the step r of the group), since one has for ω ⊂⊂ G, for
some constant C = C(G, ω) > 0
C dR (g, g ) ≤ d(g, g ) ≤ C −1 dR (g, g )1/r .
In his foundational paper [P] Pansu proved the following remarkable

result of Rademacher type for maps u : G → G between two Carnot groups.
Theorem 2.2. Let G, G be two Carnot groups, and consider open sets
Ω ⊂ G, Ω ⊂ G . If u : Ω → Ω is a Lipschitz map with respect to the
relative intrinsic metrics, i.e., there exists a constant L > 0 such that
d (u(g1), u(g2)) ≤ L d(g1 , g2), g1 , g2 ∈ Ω,
then the differential of u, Du(g), exists for dg-a.e. point g ∈ Ω.
The notion of differential of a map between Carnot groups was introduced

by Pansu himself. One says that u is Pansu differentiable at g ∈ G if the
maps
(2.20) uλ = δλ −1 ◦ Lu(g)−1 ◦ u ◦ Lg ◦ δλ
converge locally uniformly to a homomorphism Du(g) : G → G , as λ → 0.

Here, Lg (h) = gh denotes the operator of left-translation on G, whereas we
have denoted by Lg = g h the analogous operator on G . Similarly, δλ and
δλ respectively denote the dilations on G and G .
We note explicitly that when the target group is G = (R, +), then the
definition (2.20) of Pansu differential of a function u : G → R becomes
u(gδλ (h)) − u(g)

(2.21) Du(g)(h) = lim ,
λ→0+ λ
the limit being assumed locally uniform in h ∈ G. In the next proposition we

give a characterization of formula (2.21) in the case in which u is sufficiently
smooth.
Proposition 2.3. Let G be a Carnot group of step r. For u ∈ Γ1 (G) the

Pansu differential of u is given by the formula

m
Du(g)(h) = xj (h) Xj u(g) = < ξ1 (h), Xu(g) >, g, h ∈ G.
j=1
When u ∈ C 1 (G) the proof of the above result can be found in [DGS],
or also in [MSC]. The passage from C 1 (G) to Γ1 (G) is easily justified by
mollification with group convolution, see equation (3.5) for the relevant def-
inition.
We will need the following quantitative version of Theorem 2.2.
Theorem 2.4. Let Ω ⊂ G be a bounded open set. If for u ∈ Γ0,1 (Ω) and
(2.19) holds, then the distributional derivatives of u along the vector fields
X1 , ..., Xm belong to L∞ (Ω), and moreover there exists C = C(G) > 0 such
that
||Xαu||L∞ (Ω) ≤ C L α = 1, ..., m .
Theorem 2.4 is just a special case of a result which in [GN2] was proved
for a general family of Lipschitz vector fields in Rn , see also [FSS], where
the same result was independently obtained. We will also need the following
converse to Theorem 2.4 which is also a special case of a general result in
[GN2]. Given a measurable set E ⊂ G, we say that a function u ∈ Lp (E)
belongs to L1,p (E), 1 ≤ p ≤ ∞, if the distributional derivatives Xαu ∈
Lp (E), for α = 1, ...m. The space L1,p (E) becomes a Banach space with the
norm
m
||u||L1,p(E) = ||u||Lp(E) + ||Xαu||Lp(E) .
α=1
Theorem 2.5. Given a Carnot group G, let u ∈ L1,∞ (B(go , 3R)), with
go ∈ G and R > 0, then u can be modified on a set of dg-measure zero in
B̄ = B̄(go, R), so as to satisfy for every g, g ∈ B̄(go , R)
|u(g) − u(g )| ≤ C d(g, g ) u L1,∞ (B(go,3R)) ,
for some C = C(G) > 0. If, furthermore, u ∈ C ∞ (B(go , 3R)), then in
the right hand side of the previous inequality one can replace the term
||u||L1,∞ (B(go,3R)) with ||Xu||L∞(B(go,3R)).
Combining Theorems 2.5 and 2.4 we obtain a characterization of the
Lipschitz class Γ0,1 1,∞
loc (G) as the space Lloc (G).
3. Strongly convex functions: A notion to abandon.
The notion of Pansu differential expressed by equation (2.21), and its connec-
tion with the horizontal gradient established by Proposition 2.3, suggests a
first notion of convexity in a Carnot group, which we call strong H-convexity.
At the beginning of Section 4 we will see that such concept is too restrictive
for the development of a satisfactory theory, and has to be modified. An-
other basic reason to renounce the notion of strong H-convexity lies in the
fact that the gauge balls in the Heisenberg group are not strongly H-convex,
see Remark 7.5 in Section 7. But...we are getting ahead of our story. It is
time to introduce a basic definition.
Definition 3.1. Given two points g, g ∈ G, for λ ∈ [0, 1] we will denote

def
(3.1) gλ = gλ (g; g ) = = g δλ g −1 g
the twisted “convex combination” of g and g based at g.
We note that unlike the standard convex combination in a vector space,
in Definition 3.1 the base point g plays a privileged role. This unavoidable
drawback is connected with the non-commutative structure of the ambient
spaces, and it is a source of complications, such as for instance lack of a
well-behaved Jensen inequality. Different, more symmetric, “convex combi-
nations” could be proposed. For instance, one might take gλ = δ1−λ gδλ g ,
but a careful analysis reveals the unsuitability of the latter.
Before proceeding we note explicitly that since we are using the pseudo-
distance generated by the gauge (2.12) we have for every λ > 0
d(gλ, g)
(3.2) λ = .
d(g , g)
This is easily recognized as follows

d(gλ, g) = N (g −1 gλ ) = N (δλ (g −1 g )) = λ N (g −1 g ) = λ d(g , g) .
Definition 3.2. A function u : G → (−∞, ∞] is called strongly H-convex

if it is proper, which means that {g ∈ G | u(g) = ∞} = G, and if for every
g, g ∈ G one has
u(gλ ) ≤ u(g) + λ (u(g ) − u(g)) ,
for every 0 ≤ λ ≤ 1. The effective domain of u is the set
domH u = {g ∈ G | u(g) < ∞} .
We denote by CH
s (G) the class of all strongly H-convex functions on G.
We note right-away that in the abelian case G = Rn , Definition 3.2

coincides with the classical notion of convexity. The next propositions are
direct consequences of Definition 3.2.
Proposition 3.3. For every h ∈ G the left-translated Lh u(g) = u(hg) of a

strongly H-convex function u is also strongly H-convex.
Proof. For every g, g ∈ G, and any 0 ≤ λ ≤ 1 one has

Lh u(g δλ (g −1 g )) = u(h g δλ ((hg)−1 (hg ))) ≤ Lh u(g) + λ (Lh u(g ) − Lh u(g)) .

Proposition 3.4. Let {uα }α∈A be an arbitrary family of strongly H-convex

functions. If supα∈A uα is proper, then it belongs to CH
s
(G). If u, v ∈ CH
s
(G)
, and c ≥ 0, then u + v and cu are in CH (G) if they are proper.
s
Proposition 3.5. Let h : R → R be convex and non-decreasing, then for

every strongly H-convex function u : G → R, the function h ◦ u is strongly
H-convex.
The following definition, and the subsequent propositions, provide us
with a reasonably rich supply of strongly H-convex functions.
Definition
3.6. A function A : G → R is called H-affine if there exists
p= m α=1 α Xα ∈ V1 , and β ∈ R such that
p

m
(3.3) A(g) = pj xj (g) + β = < p, ξ1 (g) > + β .
j=1
Proposition 3.7. Every H-affine function is strongly H-convex.
Proof. Suppose that A is of the form (3.3), then
A(g δλ (g −1 g )) = < p, ξ1 (g δλ (g −1 g )) > + β .
We now observe that the formula of Baker-Campbell-Hausdorff (2.7)

gives
(3.4) ξ1 (g δλ (g −1 g )) = ξ1 (g) + λ (ξ1 (g ) − ξ1 (g)) ,
therefore
A(g δλ (g −1 g )) = < p, ξ1 (g) + λ (ξ1 (g ) − ξ1 (g)) > + β

= A(g) + λ (A(g ) − A(g)) .
Proposition 3.8. Let φ : Rm → (−∞, ∞], then the function u : G →

(−∞, ∞] defined by
def
u(g) = φ(x(g)) ,
is strongly H-convex if and only if φ is a standard convex function.
Proof. From (3.4) we obtain

u(g δλ (g −1 g )) = φ x(g) + λ (x(g ) − x(g)) .
The conclusion easily follows from this formula.
Convexity and convolution. If u and v are two measurable functions on

G, their convolution is defined by
(3.5) u v(g) = u(h) v(h−1 g) dh = u(gh−1 ) v(h) dh .

G G
Since convolution is a smoothing process it is important to know that

a notion of convexity is preserved under such operation. This follows from
Proposition 3.3.
Proposition 3.9. Let u : G → R be strongly H-convex and K ∈ L1 (G),

with K ≥ 0, then K u is also strongly H-convex if it is proper.
Proof. For 0 ≤ λ ≤ 1 one has from (3.5) and from Definition 3.2
K u(g δλ (g −1 g )) = K(h) u(h−1 g δλ (g −1 g ) dh

G
= K(h) u(h−1 g δλ ((h−1 g)−1 (h−1 g )) dh

G
≤ u(h−1 g) + λ (u(h−1 g ) − u(h−1 g))

K(h) dh
G

= K u(g) + λ K u(g ) − K u(g) .
4. Mildly convex functions: The appearance of horizontal

planes.
Although the notion of strongly H-convex function produces some inter-

esting results, a careful analysis of it reveals its restrictive nature. In the
Abelian case of Rn the most important prototype of convex functions are the
quadratic homogeneous polynomials P (x) = ni,j=1 αij xi xj , with αij = αji ,
and (αij )i,j=1,...,n ≥ 0. If we consider for instance the Heisenberg group
H1 , with the parabolic dilations (2.17), then the appropriate homogeneous
polynomials of non-isotropic degree 2 are of the form
P (x, y, t) = α x2 + 2 β xy + γ y 2 + δ t ,

αβ
where the matrix is assumed ≥ 0. Since, thanks to Proposition 3.8 we
βγ
know that Q(x, y, t) = αx2 +2βxy +γy 2 is strongly H-convex, it is natural to
ask whether such is also the function u(x, y, t) = t. This question is further
motivated by the fact that such “quadratic function” is also L-harmonic on
Hn , i.e., it is an entire solution of the Kohn sub-Laplacian (2.18). To verify
the strong H-convexity of u, let g = (x, y, t), g = (x , y , t ) be two points in
H1 , and consider the point defined by (3.1)

(4.1) gλ = x + λ(x − x) , y + λ(y − y) ,

t + 2λ(x y − xy ) + λ2 (t − t + 2xy − 2x y) .
We have from (4.1)
u(gλ ) = t + 2λ(x y − xy ) + λ2 (t − t + 2xy − 2x y) ,

so that Definition 3.2 now reads
t + 2λ(x y − xy ) + λ2 (t − t + 2xy − 2x y) ≤ t + λ(t − t) .
One readily sees that the latter inequality is equivalent to
(1 − λ) (2xy − 2x y) ≤ (1 − λ)(t − t) ,
which in turn holds if and only if
(4.2) t ≥ t + 2xy − 2x y .
We conclude from (4.2) that the function u(x, y, t) = t is not strongly

H-convex in H1 . This important example prompts us to relax Definition 3.2.
First, we introduce a definition which plays a pervasive role in this paper.
Definition 4.1. Let G be a Carnot group with Lie algebra g = V1 ⊕ ... ⊕ Vr ,

and let m = dim V1 . Given a point go ∈ G we define the horizontal plane
through go as the m-dimensional embedded submanifold of G given by

Hgo = Lgo exp (V1 × {0}) ,
where 0 denotes the N − m-dimensional zero vector in g, with N = dim V1 +

... + dim Vr . In particular, we see that
He = exp (V1 × {0}) , and Hgo = Lgo (He ) .
We remark explicitly that He is not a subgroup of G since it is not true,

in general, that g, g ∈ He implies gg ∈ He . For instance, in the Heisenberg
group Hn we have from (2.15)
g g = (x, y, 0) (x , y , 0) = (x + x , y + y , 2(< x , y > − < x, y >)) ,
so that gg ∈ He only when the two vectors (x , y ), (y, −x) ∈ R2n are orthog-
onal. However, it is true that g ∈ He if and only if g −1 ∈ He , and δλ g ∈ He
for λ > 0. From Definition 4.1 we immediately obtain the following elemen-
tary, yet important, result.
Proposition 4.2. A point g ∈ G belongs to Hgo if and only if go−1 g ∈ He .

Moreover, g ∈ Hg if and only if g ∈ Hg . Finally, if g ∈ Hg , then one also
has gλ ∈ Hg for every λ ∈ [0, 1].
Proof. Only the second and third parts require a proof. If g ∈ Hg , then
g −1 g ∈ He . If ξ(g) = ξ1 (g) + ... + ξr (g) denote the exponential coordinates
of g, we see that the condition g −1 g ∈ He translates into
(4.3) ξ2 (g −1 g ) = 0 , ... , ξr (g −1 g ) = 0 .
These are the general equations of the horizontal plane Hg . Since by
(2.7) we have ξ(g −1 ) = −ξ(g), we conclude that g −1 g ∈ He if and only if
(g )−1 g ∈ He , which is equivalent to g ∈ Hg .
For the third part we observe that for every λ ∈ [0, 1] one has gλ ∈ Hg
if and only if g −1 gλ = δλ (g −1 g ) ∈ He . Since for any point h ∈ G we have
δλ h ∈ He if and only if h ∈ He , we conclude that gλ ∈ Hg is equivalent to
g −1 g ∈ He , i.e., g ∈ Hg . This completes the proof.
In the sequel it will be convenient to have for a given g ∈ G an explicit

description of the horizontal plane Hg given by N −m polynomial equations.
Such equations are given by the r − 1 vector equations (4.3). In the next
proposition we calculate such equations in the interesting setting of groups
of step 2. For a group of higher step one has similar, but more complicated
formulas.
Given a Carnot group G of step 2, with Lie algebra g = V1 ⊕ V2 , consider
the mapping J : V2 → End(V1) defined by
(4.4) < J(η)ξ , ξ > = < [ξ , ξ ], η >, η ∈ V2 , ξ , ξ ∈ V1 .
The algebraic properties of the mapping J have important repercussions
on the geometric and analytic properties of Carnot groups of step 2. An
immediate consequence of the definition of J is that
(4.5) < J(η)ξ, ξ > = 0, for every η ∈ V2 , ξ ∈ V1 .
Proposition 4.3. Let G be a Carnot group of step 2, then for any given
go ∈ G the horizontal plane passing through go is the collection of all points
g ∈ G whose exponential coordinates (x(g), y(g)) verify the k linear equa-
tions
1 β
m
yβ (g) = yβ (go ) + bαα xα (go) xα (g) , β = 1, ..., k ,
2 α,α =1
where bβαα represent the group constants

k
[Xα, Xα ] = bβαα Yβ , α, α = 1, ..., m .
β=1
Proof. If we let go = exp(ξ(go)), g = exp(ξ(g)), keeping in mind that

go−1 = exp(−ξ(go)), we have from (2.7)
1
go−1 g = exp (−ξ(go )) exp (ξ(g)) = exp − ξ(go) + ξ(g) − [ξ(go), ξ(g)] .
2
Writing ξ(go) = ξ1 (go ) + ξ2 (go ), ξ(g) = ξ1 (g) + ξ2 (g), we obtain from the
latter equation
go−1 g = exp (−ξ(go )) exp (ξ(g))
1
= exp − ξ1 (go ) + ξ1 (g) − ξ2 (go ) + ξ2 (g) − [ξ1 (go), ξ1 (g)] .
2
This formula shows that
(4.6)
1
ξ1 (go−1 g) = ξ1 (g) − ξ1 (go) , ξ2 (go−1 g) = ξ2 (g) − ξ2 (go ) − [ξ1 (go), ξ1 (g)] .
2
Since g ∈ Hgo is equivalent to go−1 g ∈ He , the second equation in (4.6)
gives
1
ξ2 (go−1 g) = ξ2 (g) − ξ2 (go ) −
[ξ1 (go ), ξ1(g)] = 0 ∈ V2 .
2
k
Keeping in mind that ξ1 (g) = mα=1 xα (g)Xα, ξ2 (g) = β=1 yβ (g)Yβ , we
conclude
1
m
ξ2 (g) = ξ2 (go) + xα (go)xα (g)[Xα, Xα ] ,
2 α,α =1
which we can rewrite as follows

 

k
k
1  β
k m
yβ (g)Yβ = yβ (go)Yβ + bαα xα (go)xα (g) Yβ .
2
β=1 β=1 β=1 α,α =1
This completes the proof.
The next proposition shows that, interestingly, when g ∈ Hg then the

twisted “convex combination” of g and g given by (3.1) coincides, at the
level of the Lie algebra, with the standard Euclidean convex combination.
Proposition 4.4. Let G be a Carnot group. Given g ∈ G, then for any

g ∈ Hg one has
ξ(gλ ) = (1 − λ) ξ(g) + λ ξ(g ) ,
for every λ ∈ [0, 1].
Proof. We only present the proof for groups of step 2, the details are
similar but more (tediously) complicated in the case of higher step. We have
already observed in Proposition 4.2 that g ∈ Hg if and only if gλ ∈ Hg for
every 0 ≤ λ ≤ 1. Proposition 4.3 gives
1
m
yβ (gλ ) = yβ (g) + bβαα xα (g) xα (gλ ) , β = 1, ..., k ,
2
α,α=1
Using (2.7) we easily find
(4.7) ξ1 (gλ ) = (1 − λ) ξ1 (g) + λ ξ1 (g ) .

This gives for β = 1, ..., k
(4.8)
1−λ β
m m
λ
yβ (gλ ) = yβ (g)+ bαα xα(g) xα (g) + bβαα xα (g) xα (g ) .
2
2
α,α =1 α,α=1
Since
bβαα = < [Xα, Xα ], Yβ > = < J(Yβ )Xα, Xα > ,
we easily conclude that

m
(4.9) bβαα xα(g) xα (g)
α,α=1
m
= < ξ1 (g), Xα >< ξ1 (g), Xα >< J(Yβ )Xα, Xα >
α,α=1
m
= < J(Yβ )Xα, ξ1 (g) >< ξ1 (g), Xα >
α=1
m
= < J(Yβ )T ξ1 (g), Xα >< ξ1 (g), Xα >
α=1

m
= − < J(Yβ )ξ1 (g), Xα >< ξ1 (g), Xα >
α=1
= − < J(Yβ )ξ1 (g), ξ1(g) > = 0 ,
where in the latter equality we have used (4.5). Substitution of (4.9) in (4.8)
gives
λ
m
(4.10) yβ (gλ ) = yβ (g) + bβαα xα (g) xα (g ) .
2
α,α=1
Recalling that from Proposition 4.3 we have
1
m
bβαα xα (g) xα (g ) = yβ (g ) − yβ (g) ,
2
α,α=1
we conclude from (4.10) that
ξ2 (gλ ) = (1 − λ) ξ2 (g) + λ ξ2 (g ) .
Since ξ(g) = ξ1 (g) + ξ2 (g), the latter equation and (4.7) imply the con-
clusion.
It is easy to recognize that, when G is the Heisenberg group Hn , Defini-

tion 4.1 gives the following.
Proposition 4.5. Consider a point go ∈ Hn . The horizontal plane Hgo

passing through go is the hyperplane in Hn spanned by {X1 (go), ..., X2n(go)}.
One easily verifies that
Hgo = {g = (x, y, t) ∈ Hn | t = to + 2 < x, yo > −2 < xo , y >} .
We denote by Hg+o the horizontal upper half-space at go defined by
Hg+o = {g = (x, y, t) ∈ Hn | t ≥ to + 2 < x, yo > −2 < xo , y >} .
Returning to the example of the function u(x, y, t) = t in Hn , we see from

(4.2) that u satisfies the H-convexity inequality in Definition 3.2 provided
that we confine the point g to belong to the horizontal upper half-space at
g. This leads us to the following definition.
Definition 4.6. A function u : Hn → (−∞, ∞] will be called mildly H-

convex if it is proper, and if for every g ∈ Hn one has for every 0 ≤ λ ≤ 1
u(gλ ) ≤ u(g) + λ (u(g ) − u(g)) ,
for every g ∈ Hg+ . The effective domain of u is the set
domH u = {g ∈ G | u(g) < ∞} .
We denote by CH
m
(Hn ) the collection of all mildly H-convex functions on Hn .
It is clear from the above considerations that u(x, y, t) = t is mildly H-

convex, and, thanks to Proposition 3.5, so is the function v(x, y, t) = (t+ )p,
p ≥ 1, where t+ = max (0, t). However, the function u(g) = |t|p , p ≥ 1
is not in CH m
(Hn ). To see this, assume p = 1, and consider in H1 the point
g = (x, y, 0). Taking g = (x , y , 0) such that x y − xy < 0, from Proposition
4.5 we see that g belongs to the interior of Hg+ . Using the formula (4.1), a
simple computation shows that for u(g) = |t| the inequality
u(gλ ) ≤ u(g) + λ (u(g ) − u(g))
now reads
2 λ (1 − λ) |x y − xy | ≤ 0 ,
which is clearly impossible, except in the trivial cases λ = 0, 1. We observe,
however, that if we impose to g = (x , y , 0) to be on the horizontal plane
through g = (x, y, 0), then |x y − xy | = 0, and therefore the inequality is
satisfied for all λ ∈ [0, 1]. This negative example brings us to the subject of
the next section.
5. Weakly convex functions: The horizontal Hessian.
Our starting point presently is the classical characterization of convexity in

terms of the semi-definite positiveness of the Hessian matrix of a function.
Such result suggests that a satisfactory notion of convexity in Carnot groups
should be based on an analogous property of the intrinsic Hessian. The
latter is defined as follows.
Definition 5.1. Let G be a Carnot group. Given u : G → R, with u ∈

Γ2 (G), the horizontal Hessian of u at g ∈ G is the m × m matrix X 2 u(g) =
(Xi Xj u(g))i,j=1,...,m. The symmetrized horizontal Hessian is defined by
1
X 2 u(g) = X 2 u(g) + X 2 u(g)T .
2
As we will shortly see, the analysis of the semi-definiteness of the matrix
2
[X u] will lead us to yet another notion of convexity. To free the pre-
sentation of technical details, we consider at first the model situation of the
Heisenberg group H1 . Given u ∈ C 2 (H1 ), and two arbitrary, but fixed, points
g = (x, y, t), g = (x , y , t ) in H1 , we introduce the function φ : [0, 1] → R
defined by
(5.1) φ(λ) = u(gλ ) ,

where gλ is like in (4.1). Clearly, φ ∈ C 2 (0, 1) and moreover
(5.2) φ(0) = u(g) , φ(1) = u(g ) , φ (0) = < Xu(g), x(g) − x(g) > .
Keeping in mind (5.2), applying the standard Taylor formula to φ(λ)

gives
1
(5.3) u(g ) − u(g) − < Xu(g), x(g) − x(g) > = (1 − λ) φ (λ) dλ .
0
We are interested in the explicit expression of φ (λ). We have

(5.4)
φ (λ) = X1 u(gλ ) (x −x)+X2 u(gλ ) (y −y)+2λ ut (gλ ) t − t + 2xy − 2x y .
A calculation based on (5.4) now gives
(5.5) φ (λ) = X12 u(gλ )(x − x)2 + X1 X2 u(gλ )(x − x)(y − y)
+ X2 X1 u(gλ )(x − x)(y − y) + X22 u(gλ )(y − y)2

+ 2(t − t + 2(xy − x y)) ut (gλ )

+ 2λ X1 (ut)(gλ )(x − x) + X2 (ut)(gλ )(y − y)

2
+ 2λ (t − t + 2(xy − x y)) utt(gλ ) .
This formula reveals a remarkable property. If we confine the point g

to belong to the horizontal plane Hg passing through g, then in view of
Proposition 4.5 we have t − t + 2(xy − x y) = 0, and the equation (5.5)
reduces to
(5.6) φ (λ) = X12 u(gλ )(x − x)2 + X1 X2 u(gλ )(x − x)(y − y)
+ X2 X1 u(gλ )(x − x)(y − y) + X22 u(gλ )(y − y)2
= < (X 2 u(gλ )) ζ , ζ > ,
where we have let ζ T = (x −x, y −y). We have thus established the following
result.
Proposition 5.2. Let u ∈ Γ2 (Hn ). For any fixed g = (x, y, t) ∈ Hn and

every g = (x , y , t ) ∈ Hg one has
φ (λ) = < (X 2u(gλ )) ζ , ζ > ,
where ζ T = (x − x, y − y) ∈ R2n .

Proof. Consider a family of compactly supported, smooth approximations

to the identity {K}>0 . For every > 0 the function u = K u belongs
to C ∞ . Moreover, thanks to the hypothesis u ∈ Γ2 (Hn ) we have u → u,
Xj u → Xj u, Xi Xj u → Xi Xj u, uniformly on compact subsets of Hn .
Setting φ (λ) = u (gλ ), we obtain as in (5.6)
(5.7) φ (λ) = < (X 2 u (gλ )) ζ , ζ > .
Letting → 0 we reach the conclusion.
Proposition 5.2 continues to be valid for any Carnot group. We will need
the following useful lemma from [GV1].
Lemma 5.3. Let G be a Carnot group of step 2, then for every α, α =

1, ..., m, one has
1
k
∂ ∂
(5.8) Xα = + < [ξ1 , Xα], Yβ >
∂xα 2 ∂yβ
β=1
1 β
k m
∂ ∂
= + bαα xα (g) .
∂xα 2
∂yβ
β=1 α =1
1
k
∂2 ∂
(5.9) Xα X α = + < [Xα, Xα ], Yβ >
∂xα∂xα 2 ∂yβ
β=1
1
k
∂2
+ < [ξ1 , Xα ], Yβ >
2 ∂xα∂yβ
β=1
1
k
∂2
+ < [ξ1 , Xα], Yβ >
2 ∂xα ∂yβ
β=1
1
k
∂2
+ < [ξ1 , Xα], Yβ >< [ξ1 , Xα ], Yβ > .
4 ∂yβ ∂yβ
β,β =1
From (5.9) we obtain the commutator formula

k
∂
(5.10) [Xα , Xα ] u = bβα α .
∂yβ
β=1
Proof. Equation (5.8) is formula (3.3) in [GV1]. The formula (5.9) follows
from (5.8), once we observe that
∂
< [ξ1 , Xα], Yβ > = 0 ,
∂yβ
and that
∂ ∂
m
< [ξ1 , Xα ], Yβ >= xγ < [Xγ , Xα ], Yβ >=< [Xα, Xα ], Yβ > .
∂xα γ=1
∂xα
In the case of arbitrary step one has to replace formulas (5.8), (5.9) with
more complicated ones. Let ms = dim Vs , s = 1, ..., r, and indicate with
X1 , ..., Xm1 an orthonormal basis of V1 . We denote by (x1 (g), ..., xr(g)) the
exponential coordinates of g = exp (ξ1 (g) + ... + ξr (g)), where for each s =
1, ..., r we have presently let xs (g) = (xs,1 (g), ..., xs,ms (g)) for the components
of ξs (g) with respect to a fixed orthonormal basis of Vs . The following
formula holds
(5.11)
∂ r ms
∂
Xα = + bβs,α (x1 (g), ..., xs−1(g)) , α = 1, ..., m1 ,
∂x1,α ∂xs,β
s=2 β=1
where each bβs,α is a homogeneous polynomial of weighted degree s − 1.

By weighted degree we mean that, as previously mentioned, the layer Vs ,
s = 1, ..., r, in the stratification of g is assigned the formal degree s. Cor-
respondingly, each homogeneous monomial xα1 1 xα2 2 ...xαr r , with multi-indices
αs = (α1,s , ..., αms,s ), s = 1, ..., r, is said to have weighted degree p if

r
ms
s( αβ,s ) = p .
s=1 β=1
Formula (5.11) is proved using (2.7), analogously to (5.8), and we leave

the details to the interested reader.
Proposition 5.4. Let G be a Carnot group and let u ∈ Γ2 (G). For every
fixed g ∈ G, and for any g ∈ Hg one has

m
(5.12) φ (λ) = Xαu(gλ ) (xα(g ) − xα (g)) ,
α=1
(5.13) φ (λ) = < (X 2 u(gλ )) ζ , ζ > .
where ζ = ξ1 (g ) − ξ1 (g) ∈ V1 .
Proof. Again, it suffices to consider the case u ∈ C 2 (G). We give the

details in the case of a group of step r = 2. If r ≥ 3 the proof proceeds
along the same lines and we omit it. Suppose then that g ∈ Hg . Using the
equations (4.7) and (4.10) we obtain for the function φ(λ) defined in (5.1)
λ
m

φ(λ) = u xα(g) + λ(xα(g ) − xα(g)) , yβ (g) + bβαα xα (g) xα (g ) ,
2
α,α =1
where we have abused the notation by writing xα (g) + λ(xα(g ) − xα (g))

instead of x1 (g) + λ(x1(g ) − x1 (g)), ..., xm(g) + λ(xm(g ) − xm (g)), and simi-
larly for the variables in the second layer. Differentiating the latter equation
with respect to λ we find

m
∂u 1 β
k m
∂u
φ (λ) = (gλ )(xα(g ) − xα (g)) + bαα xα (g)xα (g ) (gλ )
∂xα 2
∂yβ
α=1 β=1 α,α =1
Using (5.8) we can rewrite the latter equation as follows
(5.14)

m

φ (λ) = Xαu(gλ ) (xα (g ) − xα(g))
α=1
1
k m
∂u
+ bβαα xα(g) xα (g ) (gλ )
2 ∂yβ
β=1 α,α=1
1 β
k m
∂u
− bα α xα (g) + λ(xα (g ) − xα (g)) (xα (g ) − xα(g)) (gλ )
2
∂yβ
β=1 α,α =1
Using (4.9) we see that

(5.15)
1 β
k m
∂u
− bα α xα (g) + λ(xα (g ) − xα (g)) (xα(g ) − xα (g)) (gλ )
2
∂yβ
β=1 α,α =1
1 β
k m
∂u
=− bαα xα (g)xα(g ) (gλ )
2
∂yβ
β=1 α,α =1
λ β
k m
∂u
+ bα α xα (g )xα(g) (gλ )
2
∂yβ
β=1 α,α =1
λ β
k m
∂u
+ bαα xα (g)xα(g ) (gλ )
2
∂yβ
β=1 α,α =1
1 β
k m
∂u
= − bαα xα(g)xα (g ) (gλ ) ,
2
∂yβ
β=1 α,α =1
where we have used the equality bβαα = −bβα α . Substituting (5.15) in (5.14)
we conclude that (5.12) holds.
From (5.12) we now easily obtain

m

φ (λ) = Xα Xα u(gλ ) xα(g ) − xα (g) xα (g ) − xα (g) ,
α,α =1
which completes the proof.
Proposition (5.4) paves the way to the following definition, which plays
a central role in this paper.
Definition 5.5. Let G be a Carnot group. A function u : G → (−∞, ∞]

will be called weakly H-convex if it is proper, and if for any g ∈ G one has
for every 0 ≤ λ ≤ 1
u(gλ ) ≤ u(g) + λ (u(g ) − u(g)) ,
for every g ∈ Hg . When the strict inequality holds we say that u is strictly
weakly H-convex. The effective domain of u is the set
domH u = {g ∈ G | u(g) < ∞} .
We denote by CH
w
(G) the class of all weakly H-convex functions on G.
Remark 5.6. It is worth noting at this point that Definition 5.5 involves
a one-sided information on u on lower-dimensional manifolds, namely the
horizontal planes. This new aspect, which is of course not present in the
classical theory of convex functions, is a source of difficulties when one wants
to pass from such information to conclusions on a set of full measure.
When u is sufficiently smooth we obtain the following basic consequence

of Definition 5.5.
Proposition 5.7. Suppose that u : G → R is weakly H-convex. If u ∈

Γ1 (G) one has for any fixed g ∈ G
< Xu(g), ξ1(g ) − ξ1 (g) > + u(g) ≤ u(g ) for every g ∈ Hg .
Proof. From Definition 5.5 we have
u(gλ ) − u(g)
≤ u(g ) − u(g) .
λ
We now pass to the limit for λ → 0 obtaining the conclusion from Propo-
sition 2.3.
Geometric interpretation of weak H-convexity. Proposition 5.7

suggests the following geometric interpretation of the notion of weak H-
convexity which avoids reference to the sub-Riemannian gradient of u, and
only involves the subbundle of the horizontal planes. Let us assume that
u ∈ C 1 (G), and for every g introduce a new function Ug : Hg → R, defined
by restricting u to Hg . Using exponential coordinates we can think of Ug
as defined on the embedded submanifold of the Lie algebra g corresponding
to Hg . With this identification Ug becomes a function of the coordinates
ξ1 (g ), projection of the exponential coordinates ξ(g ) onto the first layer V1 .
Following the arguments above one easily recognizes that
(5.16) Xu(g) = DUg (ξ1 (g)) ,
where we have denoted by DUg the standard gradient of Ug with respect to

the variables ξ1 (we stress that (5.16) only holds at the base point g, it does
not hold at other points g ∈ Hg ). Therefore, the conclusion of Proposition
5.7 can be reformulated in the following way
(5.17) < DUg (ξ1 ), ξ1 − ξ1 > + Ug (ξ1 ) ≤ Ug (ξ1 ) for every ξ1 ∈ V1 .
In other words, a function u ∈ C 1 (G) is weakly H-convex if and only if

for every g ∈ G the graph of the restriction of u to Hg lies above its tangent
plane at ξ1 (g).
The notion of weak H-convexity has been conceived to include the func-
tion u(g) = t in the Heisenberg group Hn . In this perspective, the next
useful proposition should come as no surprise.
Proposition 5.8. Let ψ : R → R be convex, then the function u : Hn → R

defined by u(g) = ψ(t) belongs to CH
w
(G). In particular, u(g) = |t|p , p ≥ 1
is in CH
w (G).
Proof. It is an immediate application of Definition 5.5, once we observe

that, thanks to Proposition 4.4, for any g ∈ Hn one has g δλ (g −1 g ) =
(1 − λ)g + λg if g ∈ Hg . Here, the right-hand side of the latter equation
represents the Euclidean convex combination of g and g , as points of R2n+1 .
Before proceeding, it is worth noting the following immediate conse-

quence of the definitions and of the examples u(g) = t, and u(g) = |t|, see
Section 4 and Proposition 5.8 :
(5.18) CH
s
CH
m
CH
w
.
Remark 5.9. It is important to notice that Propositions 3.4, 3.5 and 3.9
continue to be valid if in their statements we replace “strongly H-convex”
with “weakly H-convex”.
Proposition 5.10. A function u : G → (−∞, ∞] is weakly H-convex if and

only if such is Lh u(g) = u(hg), for every h ∈ G.
Proof. We argue similarly to the proof of Proposition 3.3, noting in addition

that for any g ∈ Hg , one has hg ∈ Hhg .
The following two theorems provide the main motivation for Definition
5.5 : the weak H-convexity of a function in the class Γ2 is characterized by
the positivity of its horizontal Hessian. To facilitate the presentation we will
first analyze with full details the Heisenberg group model in Theorem 5.11.
Theorem 5.11. A function u ∈ Γ2 (Hn ) is weakly H-convex if and only if

the symmetrized horizontal Hessian (X 2u(g)) is semi-definite positive at
every g ∈ Hn .
Proof. We first prove the necessity. Suppose that u ∈ Γ2 (Hn ) be weakly

H-convex. If we fix g ∈ Hn , from Proposition 5.7 we obtain
(5.19)
u(g ) ≥ u(g) + < Xu(g), ξ1(g ) − ξ1 (g) > , for every g ∈ Hg .
From (5.19), (5.3) and Proposition 5.2 we conclude

1
(5.20) 0 ≤ (1 − λ) < (X 2 u(gλ )) ζ , ζ > dλ ,
0
where ζ T = (x − x, y − y) ∈ R2n .

At this point, given any vector ζ = (ξ, η)T ∈ R2n , and any 0 < τ < 1,
def
we choose g = (x , y , t ) ∈ Hg such that g −1 g = δτ (ξ, η, 0) = δτ ζ.
˜ In other
words, we want to take
x −x = τ ξ , y −y = τ η , t −t+2 (< x, y > − < x , y >) = 0 .
This can of course be accomplished by choosing
x = x + τ ξ , y = y + τ η , t = t +2 τ (< ξ, y > − < x, η >) .
Inserting this choice in (5.20), we find

1
(5.21) 0 ≤ (1 − λ) < (X 2 u(gδλ(δτ ζ̃))) ζ , ζ > dλ .
0
Noting that gδλ (δτ ζ̃) → g, as τ → 0, and using the continuity of the
second horizontal derivatives of u, we conclude from (5.21)
(5.22) 0 ≤ < (X 2u(g)) ζ , ζ > ,
which establishes the first part of the theorem.

Suppose now that (5.22) holds for any g ∈ Hn and for every ζ = (ξ, η)T ∈
R2n . We will prove that u is weakly H-convex. We fix g ∈ Hn and consider
an arbitrary g ∈ Hg . Proposition 5.2 implies
φ (λ) = < (X 2 u(gλ )) ζ , ζ > ,

where now ζ T = (x − x, y − y). According to the hypothesis we have

φ (λ) ≥ 0. Since φ ∈ C 2 (0, 1), we infer that φ : [0, 1] → R is convex. This
gives in particular
φ(λ) ≤ (1 − λ) φ(0) + λ φ(1) , for every 0≤λ≤1.
In terms of u the latter inequality can be written as
u(gλ ) ≤ (1 − λ) u(g) + λ u(g ) ,
which proves that u is weakly H-convex.
Theorem 5.11 continues to be valid in any Carnot group G. The proof is

analogous to that of Theorem 5.11 and we leave the details to the interested
reader. We only mention that in the proof of the necessity, similarly to the
case of Hn one is confronted with the task of choosing, for any given ζ ∈ V1 ,
and any τ ∈ (0, 1), a point g ∈ G such that g −1 g = δτ exp (ζ), and for
which g ∈ Hg . This can be accomplished
similarly to before. For instance,
if G is of step 2, then if ζ = m ∼
α=1 α α we choose g = (x(g ), y(g )) as
ζ X
follows
xα (g ) = xα (g) + τ ζα , α = 1, ..., m ,
τ
m

yβ (g ) = yβ (g) + bβαα xα (g) ζα β = 1, ..., k .
2
α,α=1
Theorem 5.12. In a Carnot group G a function u ∈ Γ2 (G) is weakly H-

convex if and only if the symmetrized horizontal Hessian (X 2 u(g)) is semi-
definite positive at every g ∈ G.
We have seen in Proposition 5.8 that the function u(g) = t2 in Hn is
weakly H-convex. An interesting alternative proof of this result can be given
applying Theorem 5.11. Confining our attention to H1 , a simple computation
gives
8y 2 −8xy
(X 2u(g)) = .
−8xy 8x2
This implies for every ζ = (ξ, η)T
< (X 2 u(g)) ζ, ζ > = 8 (xη − yξ)2 ≥ 0 ,

w (H1 ). We also notice
which, according to Theorem 5.11, proves that u ∈ CH
that
det (X 2u) ≡ 0 .
The geometric meaning of this fact is understood by looking at the

restriction of the function u(g) = t2 to the horizontal plane Hgo , where
go = (xo, yo , to ) is arbitrarily fixed in H1 . Such restriction is given by
Ugo (x, y) = U (x, y) = (to + 2yo x − 2xo y)2 .
The principal curvatures at (xo , yo ) of the graph of Ugo are given by the
equation
κ = H ± H2 − K ,
where H and K respectively denote the mean curvature and the Gaussian
curvature of the graph at (xo , yo ). Since
2
Uxx Uyy − Uxy DU
K = , 2 H = div ,
(1 + Ux2 + Uy2 )2 1 + |DU |2
we easily recognize that K(xo, yo ) = 0, H(xo, yo ) = 4(x2o + yo2 ), and therefore
κ1 (xo, yo ) = 8 (x2o + yo2 ) , κ2 (xo , yo) ≡ 0 .
We stress that when go = (0, 0, to) then the plane Hgo has a characteristic
point at go (i.e., a point where the vector fields X1 and X2 become tangent
to the plane), and we have κ1 = κ2 = 0. We will come back to the equation
det (X 2 u) = F in Section 10.
6. Weak convexity of the gauge in groups of Heisenberg type.
A remarkable class of Carnot groups is that of Heisenberg type introduced

by Kaplan [K] in connection with questions of hypoellipticity. These in-
clude the Heisenberg group and arise as the nilpotent component N in the
Iwasawa decomposition KAN of simple groups of rank one [CDKR]. Due
to their intrinsic symmetries, the natural non-isotropic gauge (6.3) plays a
predominant role in study of the various linear and nonlinear differential
operators, see (6.4), (6.6), Proposition 6.4 and Theorem 10.9. On the other
hand, such function also serves as a true distance [Cy] (not just a pseudo-
distance), equivalent to the Carnot-Carathéodory metric associated with the
basis {X1 , ..., Xm}. It is thus natural to study its convexity properties in view
to applications to problems in optimal control and fully nonlinear equations.
In this section we establish a basic property of groups of Heisenberg type,
namely that their non-isotropic gauge is a weakly H-convex function. This
is the counterpart of the equally basic, yet trivial fact, that the distance from
the origin in a normed linear space is an ordinary convex function. By con-

trast, the weak H-convexity of the gauge is a highly nontrivial phenomenon
which has both geometric and analytic implications. Its proof is crucially
based on Theorem 5.12, but one also needs to develop additional arguments
to take care of the singularity of the gauge. We begin by introducing the
relevant geometric ambients.
Definition 6.1. A Carnot group G of step 2 is called of Heisenberg type if

for every η ∈ V2 , such that |η| = 1, the map J(η) : V1 → V1 is orthogonal.
In Definition 6.1 for a Carnot group G of step 2, with Lie algebra g =
V1 ⊕ V2 , the mapping J : V2 → End(V1) is that defined in (4.4).
Definition 6.1 implies
(6.1) |J(η)ξ| = |η| |ξ|, η ∈ V2 , ξ ∈ V1 ,
(6.2) < J(η )ξ, J(η )ξ > = < η , η > |ξ|2 , η , η ∈ V2 , ξ ∈ V1 .
A basic property of groups of Heisenberg type is that the fundamental

solution of a fixed sub-Laplacian is expressed by the appropriate power of
the renormalized gauge
1/4
4 2
(6.3) N (g) = |x(g)| + 16 |y(g)| .
2
Given the sub-Laplacian L = m j=1 Xj associated with an orthonormal
basis of V1 , Kaplan [K] proved that there exists a constant C = C(G) > 0
such that the locally integrable function
C
(6.4) Γ(g, g ) = ,
N (g −1 g )Q−2
satisfies the equation LΓ(g, ·) = −δg in D (G). This result generalized Fol-
land’s fundamental solution for the Heisenberg group Hn [F1]. In [CDG] it
was discovered that, remarkably, the gauge (6.3) enters also in the expres-
sion of the fundamental solution of the following one-parameter family of
non-linear operators

m
(6.5) Lp u = Xj (|Xu|p−2 Xj u) , 1<p<∞.
j=1
It was proved there that for every 1 < p < ∞, p = Q, there exists
Cp = C(G, p) > 0 such that the fundamental solution of −Lp is given by
Cp
(6.6) Γp (g, g ) = .
N (g −1 g )(Q−p)/(p−1)
In the conformally invariant case p = Q one has instead ΓQ (g, g ) =

C log N (g, g ). This latter result was also independently found in [HH].
Similarly to its Euclidean counterpart, the one-parameter family of quasi-
linear operators Lp tends in a suitable sense, in the limit as p → ∞, to the
following nonlinear operator
def
m
(6.7) L∞ u = XαXα u Xαu Xα u = < (X 2 u) Xu, Xu > .
α,α=1
For the Heisenberg group Hn a notion of viscosity solution of (6.7) has

been recently introduced by T. Bieske in his interesting paper [Bi]. The same
author has also established the uniqueness of such solutions. It is remarkable
that, once again, the gauge N (g) defined by (6.3) constitutes a singular
solution of (6.7), see Proposition 6.4. This result, which we discovered back
in 1995, provides one further witness of the relevance of such function, and
also has applications in various questions connected with unique continuation
[GV2], and with trace inequalities [DGN1]. We recall the following useful
formulas, see [CDG], [GV1].
def
Lemma 6.2. In a Carnot group G of step 2, consider the function ψ(g) =
|x(g)|2. For any s = 1, ..., k, one has for every g ∈ G
(6.8) < Xψ(g), Xys(g) > ≡ 0 .
As a consequence,
(6.9) < Xψ(g), X(|y|2)(g) > = 0 ,
and
(6.10) < ξ1 (g), X(|y|2)(g) > = 0 .
The next lemma expresses some key symmetry properties of groups of

Heisenberg type.
Lemma 6.3. Let G be a group of Heisenberg type. One has
|x(g)|2
(6.11) |XN (g)|2 = .
N (g)2
1
(6.12) |X(yl )(g)|2 = |x(g)|2 , l = 1, ..., k .
4
(6.13) |X(|y|2)(g)|2 = |x(g)|2 |y(g)|2 .
Proposition 6.4. Let G be a group of Heisenberg type, then one has in

the classical sense
L∞ N = 0 , in G \ {e} .
Proof. We observe the following alternative form of (6.7)

1
(6.14) L∞ u = < X(|Xu|2), Xu > .
2
Using (6.14) we find
1 1
L∞ N = < X(|XN |2), XN > = < X(N −2|x|2 ), XN > ,
2 2
where in the latter equation we have used (6.11) in Lemma 6.3. We now
have, again by (6.11),
(6.15) < X(N −2 |x|2 ), XN > = − 2 N −3 |x|2 < XN, XN >

+ 2 N −2 < ξ1 , XN >
= − 2 N −5 |x|4 + 2 N −2 < ξ1 , XN > .
Finally, we obtain
1 −3
XN = N X(ψ 2 + 16|y|2 ) = N −3 |x|2 ξ1 + 4X|y|2 ,
4
which gives
< ξ1 , XN > = N −3 |x|4 + 4 < ξ1 , X|y|2 > = N −3 |x|4 ,
where in the last equality we have used (6.10) in Lemma 6.2. Replacing the
latter equation in (6.15) we reach the conclusion.
Consider the function u(g) = N (g)4 = |x(g)|4 + 16|y(g)|2. From the re-
sults in Section 3, we easily conclude that g → |x(g)|4 is strongly H−convex,
and therefore it is weakly H-convex, see (5.18). Next, we prove that also the
function g → |y(g)|2 is weakly-convex, and therefore the fourth power of the
gauge is weakly H-convex.
def
Proposition 6.5. In a Carnot group of step 2 the function χ(g) = |y(g)|2
w (G), therefore so does u(g) = N (g)4 = |x(g)|4 + 16|y(g)|2.
belongs to CH
Proof. According to Lemma 5.3 we have

k
Xα Xα χ = < [Xα, Xα ], Yβ > yβ
β=1
1
k
+ < [ξ1 , Xα], Yβ >< [ξ1 , Xα ], Yβ > δββ
2
β,β =1

k
= < [Xα, Xα ], Yβ > yβ
β=1
1
k
+ < [ξ1 , Xα], Yβ >< [ξ1 , Xα ], Yβ > .
2
β=1
Since [Xα, Xα ] = −[Xα , Xα], we obtain
(6.16)
1 1
k
Xα Xα χ + Xα Xα χ = < [ξ1 , Xα], Yβ >< [ξ1 , Xα ], Yβ > .
2 2
β=1
This gives for every ζ ∈ V1
(6.17) < (X 2 χ(g))ζ, ζ >

1
k m
= < J(Yβ )ξ1 (g), Xα > ζα < J(Yβ )ξ1 (g), Xα > ζα
2
β=1 α,α =1
1
k
= < J(Yβ )ξ1 (g), ζ >2 ≥ 0 .
2
β=1
In view of Theorem 5.11 we conclude that χ ∈ CH

w
.
Having established the weak H-convexity of N 4 , the question remains

open of whether the gauge itself is weakly H-convex. This is a delicate
problem. The main step toward it consists in establishing the following
important positivity result.
Theorem 6.6. Let N (g) = (|x(g)|4 + 16|y(g)|2)1/4 be the gauge in a group

of Heisenberg type, then for any given g ∈ G \ {e} one has
(6.18) < (X 2N (g))ζ, ζ > ≥ 0 ,
for every ζ ∈ V1 .
Proof. Let χ(g) = |y(g)|2, and consider the function ψ(g) = |x(g)|2. We
thus have u(g) = N (g)4 = ψ(g)2 + 16χ(g). We first compute the quadratic
form associated with the symmetrized horizontal Hessian of u. Since
(6.19) (X 2 u(g)) = (X 2ψ 2 (g)) + 16 (X 2χ(g)) ,
we need to compute < (X 2 ψ 2 (g))ζ, ζ > and < (X 2χ(g)) ζ, ζ >, where
ζ ∈ V1 . The latter of these quantities has been obtained in (6.17), so we are
left with computing the former. From (5.8), (5.9) in Lemma 5.3 we find
(6.20) Xα ψ 2 = 2 ψ Xα ψ = 4 xα ψ ,
(6.21) XαXα ψ 2 = 4 δαα ψ + 8 xα xα .
From (6.21) and (6.17) we conclude

(6.22)

k
2 2 2 2
< (X u) ζ, ζ > = 4 |ξ1 | |ζ| + 8 < ξ1 , ζ >

+8 < J(Yβ )ξ1 , ζ >2 .
β=1
Since N = u1/4 , an elementary computation gives

1 3 XαuXα u
(6.23) Xα Xα N = Xα Xα u − .
4N 3 4 N4
Keeping in mind (6.20) one easily recognizes that

Xα uXα u = 16 ψxα + 4Xαχ ψxα + 4Xα χ .
On the other hand one obtains from (5.8)
(6.24)

k
Xα χ = yβ < [ξ1 , Xα], Yβ >
β=1
(6.25)

k
= < ξ2 , Yβ >< [ξ1 , Xα], Yβ >=< [ξ1 , Xα], ξ2 >=< J(ξ2 )ξ1 , Xα > .
β=1
Substitution in the previous formula gives

XαuXα u = 16 ψxα + 4 < J(ξ2 )ξ1 , Xα > ψxα + 4 < J(ξ2 )ξ1 , Xα >

= 16 |ξ1 |4 xα xα + 4|ξ1 |2 xα < J(ξ2 )ξ1 , Xα >
+ 4|ξ1 |2 xα < J(ξ2 )ξ1 , Xα >

+ 16 < J(ξ2 )ξ1 , Xα >< J(ξ2 )ξ1 , Xα > .
From this expression, from (6.23) and from (6.22) we find
(6.26)
m
2 1 2 3 α,α =1 Xα uXα uζα ζα
< (X N ) ζ, ζ > =
< (X u) ζ, ζ > − .
4N 3 4 N4
k
1 2 2 2
= |ξ 1 | |ζ| + 2 < ξ 1 , ζ > + 2 < J(Yβ )ξ1 , ζ >2
N3
β=1

3
− |ξ1 |4 < ξ1 , ζ >2 + 8|ξ1 |2 < ξ1 , ζ >< J(ξ2 )ξ1 , ζ >
N4

+ 16 < J(ξ2 )ξ1 , ζ >2

k
1
= |ξ1 |2 |ζ|2 + 2 < ξ1 , ζ >2 + 2 < J(Yβ )ξ1 , ζ >2
N3
β=1
2
3
− |ξ1 |2 < ξ1 , ζ > + 4 < J(ξ2 )ξ1 , ζ > .
N4
From (6.26) we see that proving the theorem is equivalent to showing

that
2
3 2
(6.27) |ξ1 | < ξ1 , ζ > + 4 < J(ξ2 )ξ1 , ζ >
N4

k
≤ |ξ1 |2 |ζ|2 + 2 < ξ1 , ζ >2 + 2 < J(Yβ )ξ1 , ζ >2
β=1
k
Recalling that ξ2 = β=1 yβ Yβ , we now have from Cauchy-Schwarz in-
equality
2
2
(6.28) |ξ1 | < ξ1 , ζ > + 4 < J(ξ2 )ξ1 , ζ >

k 2
2
= |ξ1 | < ξ1 , ζ > + 4 yβ < J(Yβ )ξ1 , ζ >
β=1
 

k
≤ |ξ1 |4 + 16 |ξ2 |2 < ξ1 , ζ >2 + < J(Yβ )ξ1 , ζ >2 
β=1
 

k
= N 4 < ξ1 , ζ >2 + < J(Yβ )ξ1 , ζ >2 
β=1
Using (6.28) we see that in order to establish (6.27) it suffices to prove

 
k
3 < ξ1 , ζ >2 + < J(Yβ )ξ1 , ζ >2 
β=1

k
2 2 2
≤ |ξ1 | |ζ| + 2 < ξ1 , ζ > +2 < J(Yβ )ξ1 , ζ >2 .
β=1
The latter inequality is equivalent to

k
(6.29) < ξ1 , ζ >2 + < J(Yβ )ξ1 , ζ >2 ≤ |ξ1 |2 |ζ|2 .
β=1
Now (6.29) is trivially true when ξ1 = 0. If ξ1 = 0, we observe that the

set of vectors
ξ1 J(Y1 )ξ1 J(Yk )ξ1
, , ... , ,
|ξ1 | |ξ1 | |ξ1 |
constitutes an orthonormal system in V1 . This is the only place where

we use the structure of Heisenberg type. Orthogonality follows from
< J(Yβ )ξ1 , ξ1 >= 0, and from (6.2) which gives < J(Yβ )ξ1 , J(Yβ )ξ1 >=
δββ |ξ1 |2 . From (6.1) we have |J(Yβ )ξ1 | = |ξ1 |, for every β = 1, ..., k. Having
observed this, (6.29) is now an immediate consequence of Bessel’s inequality.
Remark 6.7. For the first Heisenberg group H1 the proof of Theorem 6.6
can be simplified to some extent due to the fact that in R1 one has < x, y >=
xy. This causes some special cancellations in the expression of the quadratic
form < (X 2N ) ζ, ζ > which, however, fail to occur in the higher-dimensional
case.
Theorem 6.8. Let N (g) = (|x(g)|4 + 16|y(g)|2)1/4 be the gauge in a group

of Heisenberg type, then N ∈ CH
w (G).
Proof. We need to prove that given any g ∈ G, for any g ∈ Hg one has
for 0 ≤ λ ≤ 1
(6.30) N (gλ) ≤ (1 − λ) N (g) + λ N (g ) .
where as before gλ = gδλ (g −1 g ). When g = e we have gλ = δλ g , and (6.30)

immediately follows from the homogeneity of N . Suppose then that g = e.
We distinguish two cases: (a) g ∈ He ; (b) g ∈ He . In case (a) we see from
(i) in Proposition 4.2 that e ∈ Hg . Therefore, for any g ∈ Hg we also have
gλ = e for 0 ≤ λ ≤ 1. As a consequence of this observation, if we consider the
function φ : [0, 1] → R defined by φ(λ) = N (gλ), we see that φ ∈ C 2 (0, 1).
We can thus look at φ (λ). From Proposition 5.4, we have
φ (λ) = < (X 2N (gλ )) ζ , ζ > ,
where ζ = ξ1 (g −1 g ) = ξ1 (g ) − ξ1 (g). Thanks to Theorem 6.6, we now know

that φ (λ) ≥ 0, thus concluding that φ is a convex function on the interval
[0, 1]. This gives in particular φ(λ) ≤ (1 − λ)φ(0) + λφ(1), which is (6.30).
We are left with considering case (b). Suppose that g ∈ He , so that
y(g) = 0 ∈ Rk . If g = e, then gλ = δλ g and again we obtain (6.30) from
the homogeneity of N . So we can suppose g = e. Consider g ∈ Hg . We
distinguish two cases: (i) g = e; (ii) g = e. In case (i) we have gλ = gδλ g −1 ,
so that x(gλ ) = (1 − λ)x(g), y(gλ ) = 0, which gives
N (gλ) = (1 − λ) |x(g)| = (1 − λ) N (g) + λ N (e) ,

and (6.30) holds. We are thus left with case (ii). If ξ(g ) is not on the
straight line segment joining ξ(g) to the origin in g, then we argue as in case
(a) using Theorem 6.6. The final case is when ξ(g ) belongs to the straight
line segment joining ξ(g) to the origin in g, i.e.,
ξ(g ) = a ξ(g) = a ξ1 (g) ,
for some a ∈ R. The case a ≥ 0 is easily handled by the previous considera-

tions, so let us assume a < 0. We then have
x(gλ ) = (1 − λ)x(g) + λ (a x(g)) , y(gλ ) = 0 ,
which gives
N (gλ) = |(1 − λ) x(g) + λ (a x(g))|

≤ (1 − λ) |x(g)| + λ |a x(g)| = (1 − λ) N (g) + λ N (g ) .
7. Convexity of sets.
In this section we introduce a notion of convexity for sets which turns out
to be equivalent, at level of the epigraph, to that of weakly H-convex func-
tion. Such characterization is very desirable since, quoting from p.25 of [Ro],
“...it emphasizes the geometry which is fundamental to the theory of convex
functions”.
Definition 7.1. A subset A of a Carnot group G is called weakly H-convex

if for any g ∈ A, and every g ∈ A ∩ Hg one has gλ = gλ (g, g ) ∈ A for
every λ ∈ [0, 1]. Here, gλ is as in (4.1).
Remark 7.2. We remark explicitly that, in view of Proposition 4.4, the ge-
ometric interpretation of Definition 7.1 is that for any g ∈ A the intersection
of A with the horizontal plane Hg is starlike in the Euclidean sense with
respect to g at the level of the Lie algebra.
The next proposition is a simple, yet important, consequence of Defini-

tion 7.1.
Proposition 7.3. A set A ⊂ G is weakly H-convex if and only if such is

go A, for every go ∈ G.
A basic collection of weakly H-convex sets is provided by the next result.
Proposition 7.4. Let u ∈ CH

w
(G), then for a ∈ R the level sets
Ωa = {g ∈ G | u(g) ≤ a} ,
are weakly H-convex. In particular, the gauge balls B(g, R) in a group of

Heisenberg type are weakly H-convex.
Proof. By the assumption u ∈ CHw

(G) we know for any g ∈ G, and every

g ∈ Hg
u(gλ ) ≤ (1 − λ) u(g) + λ u(g ) .
If therefore g ∈ Ωa , and g ∈ Ωa ∩ Hg , we conclude
u(gλ ) ≤ (1 − λ) a + λ a = a .
The second part now follows from the first and from Theorem 6.8.
We remark explicitly that Proposition 7.4 appears so deceitfully easy

because the hard work has already been done in establishing Theorem 6.8.
Remark 7.5. In connection with Remark 7.2 we emphasize that if in Defi-

nition 7.1 we substitute the condition “g ∈ A ∩ Hg ” with the stronger one
“g ∈ A”, then the ensuing notion of convexity means unrestricted starlike-
ness at every point of the set A with respect to the non-isotropic group
dilations. Such notion, which is equivalent to that of strongly H-convex
function in Definition 3.2, is too restrictive, and in fact it fails even for the
gauge ball B(e, 1) in the Heisenberg group Hn . In other words, the ball
B(e, 1) is not strongly H-convex.
We next recall the definition of epigraph of a function u : A → R, where
A ⊂ G. This is the set defined by
(7.1) epi u = {(g, s) ∈ G × R | u(g) ≤ s} .
In order to characterize the weak H-convexity of u in terms of a similar

property of its epigraph, we need to endow G = G × R with a corresponding
Carnot group structure. This is easily accomplished if we define for g̃ =
(g, s), g̃ = (g , s ) ∈ G
def
g̃ g̃ = (g g , s + s ) .
The identity in G is given by ẽ = (e, 0), and for every g̃ = (g, s) we

have g̃ −1 = (g −1 , −s). Denoting by Lg̃ g̃ = g̃g̃ the left-traslation on G, one
recognizes that in exponential coordinates its differential is represented by
the (N + 1) × (N + 1) matrix formed by a N × N block which is just the
differential of Lg on G, and by a remaining row and column in which all
the entries are 0, but the last one which is 1. The Lie algebra of G admits
a nilpotent stratification of step r, g̃ = Ṽ1 ⊕ ... ⊕ Ṽr , where Ṽ1 = V1 × Rs ,
whereas Ṽj = Vj × {0}s , j = 2, ..., r. If X1 , ..., Xm denotes an orthonormal
basis of V1 , then by slightly abusing the notation (in the sense that, although
we continue to use the notation Xα, now this vector field acts on the variable
g̃ = (g, s), not just g), we obtain on the group G the corresponding system
of vector fields X̃α = Xα , α = 1, ..., m, X̃m+1 = ∂/∂s. The natural non-
isotropic dilations on G are defined by
def
δ̃λ g̃ = δ̃λ (g, s) = (δλ g, λs) .
If we denote by ξ˜ : G → g̃ = ξ˜1 (g̃)+...+ξ˜r (g̃) the inverse of the exponential
mapping defined by exp ξ̃(g̃) = g̃, one easily checks that for the projection
onto the first layer Ṽ1 one now has x̃α (g̃) = xα (g), α = 1, ..., m, x̃m+1 (g̃) =
s. From these formulas, it is easy to recognize that the equations of the
horizontal plane Hg̃o in G are unchanged, in the sense that they are the same
equations obtained for Hgo in G, except that now these equations represent
an immersed (m + 1)-dimensional manifold in G. Thus, for instance, if
G = H1 , the first Heisenberg group, and we have go = (xo , yo , to) ∈ H1 ,
so ∈ R, g̃o = (go , so) ∈ G, then the horizontal plane in the product group G
passing through g̃o has equation
t = to + 2 (yo x − xo y) .
After these preliminaries we can state the following characterization of
weak H-convexity.
Proposition 7.6. Given a weakly H-convex subset A ⊂ G of a Carnot

group, a function u : A → R is weakly H-convex if and only if epi u is a
weakly H-convex subset of G.
Proof. For every λ ∈ [0, 1] consider the point

g̃λ = g̃ δ̃λ (g̃ −1 g̃ ) ,
where g̃ = (g, s), g̃ = (g , s ). Suppose that u is weakly H-convex, and fix
g̃ ∈ epi u, so that u(g) ≤ s. Consider now g̃ ∈ epi u ∩ Hg̃ , thus in particular
u(g ) ≤ s . We want to prove that for every λ ∈ [0, 1] one has g̃λ ∈ epi u.
From Proposition 4.4 we obtain

˜ λ ) = (1 − λ)ξ(g) + λξ(g ), (1 − λ)s + λs .
ξ(g̃
This shows that the condition g̃λ ∈ epi u is equivalent to
u(gλ ) ≤ (1 − λ)s + λs ,
with gλ given by (4.1). On the other hand, the weak H-convexity of u implies
u(gλ ) ≤ (1 − λ) u(g) + λ u(g ) ≤ (1 − λ)s + λs .
This proves the necessity. The sufficiency is established by simply re-
versing the above implications, and choosing s = u(g), s = u(g ) in the end.
8. Further properties of second derivatives of weakly convex

functions.
In this section we generalize several basic properties of ordinary convex func-

tions to weakly H-convex ones on a Carnot group. A classical theorem states
that a distribution T ∈ D (RN ) can be identified with a convex function f if
and only if its Hessian matrix {Dij T } is a nonnegative N × N matrix-valued
Radon measure µ, see for instance [R], [Du], or [EG]. Our first result estab-
lishes a similar property for weakly H-convex functions in a Carnot group
G.
Theorem 8.1. Let G be a Carnot group G and consider u ∈ CH w (G) ∩

αα = ν α α
L1loc (G). For α, α = 1, ..., m, there exist signed Radon measures νH

H
such that for every φ ∈ Co∞ (G) one has
1
u(g) {XαXα φ(g) + Xα Xαφ(g)} dg = αα
φ(g) d νH (g) .
2 G G
αα
In addition, the measures νH are nonnegative.
Proof. Let K ∈ Co∞ (G), K ≥ 0, and consider the corresponding approxi-

mation to the identity {K }>0 . By Remark 5.9 for every > 0 the function
u = K u ∈ C ∞ (G) is weakly H-convex. Thanks to Theorem 5.12 we
obtain for every ζ ∈ V1 , with |ζ| = 1
u (g) < (X 2 φ(g))ζ, ζ > dg = φ(g) < (X 2u (g))ζ, ζ > dg ≥ 0 .

G G
Passing to the limit as → 0 we find

def
Tζ (φ) = u(g) < (X 2φ(g))ζ, ζ > dg ≥ 0 ,
G
which shows that Tζ : Co∞ (G) → R defines a nonnegative linear functional.

By the Riesz representation theorem there exists a nonnegative Radon mea-
ζ
sure νH such that
φ ∈ Co∞ (G) .
ζ
Tζ (φ) = φ(g) d νH (g) ,
G
Recalling that {X1 , ..., Xm} is an orthonormal basis of V1 , we now define

αα Xα
νH = νH ,
and observe that if α = α the vector

def Xα + Xα
ζαα = √ ∈ V1
2
is such that |ζαα | = 1. With this choice we have
< (X 2φ(g))ζαα , ζαα >

1
= XαXαφ(g) + XαXα φ(g) + Xα Xα φ(g) + Xα Xα φ(g) ,
2
which gives
(8.1)
ζ 1 αα
φ(g) d νHαα (g) = φ(g) d νH (g)
G 2 G
1 1
+ u(g) {XαXα φ(g) + Xα Xαφ(g)} dg + αα
φ(g) d νH (g) .
2 G 2 G
If we define the signed Radon measure

Xα +X
def √ α 1 αα 1 α α
αα
νH = νH 2
− ν − ν ,
2 H 2 H
then from (8.1) we reach the conclusion
1
u(g) {Xα Xα φ(g) + Xα Xα φ(g)} dg = αα
φ(g) d νH (g) .
2 G G

Remark 8.2. A classical consequence of the Euclidean counterpart of The-

orem 8.1 is that the first derivatives of a convex function f : RN → R are
locally in BV , the space of functions with bounded variation. This result
was first proved by Bakelman [Ba2]. It would be desirable to have a similar
result for a weakly H-convex function u on a Carnot group, but this does
not seem to follow directly from Theorem 8.1. The main obstruction being
that the representability of XαXα u + Xα Xα u by a Radon measure does not
directly guarantee a similar property for XαXα u and Xα Xαu. This is of
course a reflection of the lack of commutation XαXα u = Xα Xα u.
In some cases it is possible to prove that the second horizontal derivatives
XαXα u are Radon measures. This follows from the following elementary
proposition.
Proposition 8.3. Let u ∈ L1loc (G) be a weakly H-convex function in a

Carnot group G. If for every α, α = 1, ..., m, the commutators [Xα, Xα ]u
are Radon measures, then such are also Xα Xα u.
Proof. One has

1 1
Xα Xα u = {XαXα u + Xα Xα u} + [Xα, Xα ] u .
2 2
Since from Theorem 8.1 we know that 12 {XαXα u + Xα Xαu} is a Radon
measure, the conclusion follows.
Corollary 8.1. Under the assumptions of Proposition 8.3, suppose in ad-

dition that Xα u ∈ L1loc (G), for α = 1, ..., m. One then has u ∈ BVH,loc
2 (G).
2 1,1
Here, BVH,loc (G) denotes the space of functions u ∈ Lloc (Ω) whose deriva-
tives Xα u, α = 1, ..., m, are locally of horizontal bounded variation.
Proof. We recall the notion of horizontal bounded variation introduced in

min [GN1]. Let Ω ⊂ G be1 an open set,
[CDG] and subsequently developed
1
and u ∈ Lloc (Ω). Denote by ζ = α=1 ζαXα an element of Co (Ω; HΩ). Let

1
FH (Ω) = ζ ∈ Co (Ω; HΩ) | ||ζ||∞ ≤ 1 .
The H−variation of u in Ω is defined as follows

m
V arH (u; Ω) = sup u Xαζα dg .
ζ∈FH (Ω) Ω α=1
A function u ∈ L1 (Ω) is called of bounded H−variation if V arH (u; Ω) <

∞. In such case, we write u ∈ BVH (Ω), and the collection of all such
functions becomes a Banach space when endowed with the norm
||u||BVH(Ω) = ||u||L1(Ω) + V arH (u; Ω) .
The notation BVH,loc (Ω) indicates the collection of functions u ∈ L1loc (Ω),
such that u ∈ BVH (ω), for every ω ⊂⊂ Ω.
Returning to the proof of the corollary, let then ω ⊂⊂ G, and consider
ζ ∈ FH (ω). For any α = 1, ..., m we have

m
m
Xα u Xα ζα dg = − u Xα Xα ζα dg .
ω α =1 α =1 ω
Using Theorem 8.1 we obtain

m
m

m
Xα u Xα ζα dg = − 2 αα
ζα (g) dνH (g) + (XαXα u, ζα ) ,
ω α=1 α=1 G α =1
where we have denoted by (·, ·) the duality between D (G) and D(G). By
Proposition 8.3 we know that also XαXα u are Radon measures, therefore
we conclude

m
m
m
αα
Xα u Xα ζα dg ≤ 2 νH (ω) + XαXα u(ω) < ∞ .
ω α =1 α =1 α =1
Taking the supremum on all ζ ∈ FH (ω) we reach the conclusion.
An interesting consequence of the above results and of Theorem 6.8 is

provided by the following proposition which finds applications in optimal
control.
Proposition 8.4. Let G be a Carnot group of step 2 with non-isotropic

gauge N , then the commutators [Xα, Xα ]N belong to L1loc (G). When G is
of Heisenberg type G, we obtain
2
N p ∈ BVH,loc (G) , for every p ≥ 1 .
Proof. For convenience we consider the gauge renormalized by the (imma-

terial) factor 16, i.e., N (g) = (|x(g)|4 + 16|y(g)|2)1/4 . Formula (5.10) gives
for g ∈ G \ {e}

k
∂N
(8.2) [Xα, X ] N (g) =
α bβαα
∂yβ
β=1
8
k
= bβαα yβ (g) .
N (g)3
β=1
The trivial observation that |4yβ (g)| ≤ N (g)2 allows to conclude from
(8.2)
Cαα
|[Xα, Xα ] N (g)| ≤ ,
N (g)

where we have let Cαα = 2 kβ=1 |bβαα |. Since N −1 ∈ Lqloc (G) for every
1 ≤ q < Q, where Q represents the homogeneous dimension of G, the first
part of the proposition follows.
Suppose now that G be of Heisenberg type, then by Theorem 6.8 we
know that N is weakly H-convex. Furthermore, by (6.11) in Lemma 6.3 we
have
|x(g)|2
|XN (g)|2 = ≤ 1,
N (g)2
and therefore N ∈ L1,∞ loc (G). All the hypothesis of Proposition 8.3, and
therefore of Corollary 8.1, are thus fulfilled. The conclusion follows.
Remark 8.5. In connection with Remark 8.2, we mention that we have

recently received an interesting preprint from L. Ambrosio and V. Magnani
[AM] in which the authors prove, among other things, that if u ∈ BVH2 (G),
then for dg-a.e. go ∈ G there exists a polynomial P (go ; ·) of weighted degree
≤ 2, such that
1 1
(8.3) lim 2
|u(g) − P (go , g)| dg = 0 .
r→0 r |B(go , r)| B(go,r)
In the classical setting, as a consequence of Bakelman’s result cited in Re-

mark 8.2, the class BVloc2 contains that of convex functions, therefore (8.3)
can be viewed as an integral version for Carnot groups of the Theorem of

Busemann-Feller (n = 2) and Alexandrov (n ≥ 3) on the existence a.e.
of second derivatives of a convex function. However, (8.3) leaves open the
question of the actual sub-Riemannian pointwise version of the Busemann-
Feller-Alexandrov theorem for convex functions. For functions which are in
BVH2 (G) ∩ CH w
(G) ∩ C(G) such result can be obtained as a consequence of
our Theorem 9.2 in the next section, and of (8.3). We plan to come back to
these questions in a forthcoming article.
9. Basic properties of first derivatives of weakly convex

functions: L∞ − L1 estimates.
An essential property of convex functions is that they are locally Lipschitz

continuous. This section is devoted to proving some basic quantitative ver-
sions for continuous weakly H-convex functions in a Carnot group G. The
main result is Theorem 9.2. A remarkable aspect of the latter is that from
an information on lower dimensional manifolds (weak H-convexity), it draws
a conclusion on a set of full measure, see (9.7), (9.8). We mention that if,
instead of assuming continuity, we only suppose that the function u belong
0,1
to L∞loc (G), then we can still conclude membership in Γloc (G), but of course
after redefinition on a set of measure zero.
Theorem 9.1. Let u ∈ CH w

(G) ∩ C(G) in a Carnot group G, then u ∈
0,1
Γloc (G). Furthermore, there exists C = C(G) > 0 such that the following
estimate holds
C
(9.1) ||Xu||L∞(B(go,R)) ≤ ||u||L∞(B(go,3R)) ,
R
for any go ∈ G and every R > 0. In particular, (9.1) shows that there exist
no bounded functions in CH
w
(G) ∩ C(G), other than the constant functions.
Proof.
Given a kernel K ∈ Co∞ (G), K ≥ 0, supp K ⊂ B(e, 1),
G K(g)dg = 1, consider the corresponding approximation to the identity
{K }>0 associated with it. Let u = K u, then u ∈ C ∞ (G), and by the
hypothesis u ∈ C(G) we have u → u uniformly on compact subsets of G.
We now fix a gauge ball B(go , R) ⊂ B(go , 3R) ⊂ G. By Remark 5.9
the functions u are weakly H-convex. Proposition 5.7 implies for any fixed
g ∈ B(go , R) and every g ∈ Hg
< Xu (g), x(g ) − x(g) > ≤ u (g ) − u (g) .
For g ∈ Hg \ {g} we rewrite the latter inequality as follows
x(g ) − x(g) u (g ) − u (g) u (g ) − u (g)

(9.2) < Xu (g), > ≤ = ,
|x(g ) − x(g)| |x(g ) − x(g)| d(g , g)
where we have used the fact that when g ∈ Hg , then
(9.3) d(g , g) = |x(g ) − x(g)| .
This can be easily seen as follows. The condition g ∈ Hg is equivalent to

saying that g = gh, with h ∈ He . Therefore, we obtain from (2.12), (2.11)
and (2.10)
(9.4) d(g , g) = N (g −1 g ) = N (h) = |x(h)| ,
the latter equality being justified by the fact ξ2 (h) = 0, ... , ξr (h) = 0. On
the other hand, (2.7) gives x(h) = x(g −1 g ) = x(g ) − x(g), and (9.3) follows.
If we now take g ∈ ∂B(g, R) ∩ Hg in (9.2), for 0 < < R/2 we obtain
x(g ) − x(g) u (g ) − u (g)

< Xu (g), > ≤
|x(g ) − x(g)| R
2 2
≤ ||u ||L∞ (B(go,2R)) ≤ ||u||L∞(B(go,3R)) .
R R
Passing to the supremum on all g ∈ ∂B(g, R) ∩ Hg we find
2
|Xu(g)| ≤ ||u||L∞(B(go,3R)) .
R
By the arbitrariness of g ∈ B(go , R) we conclude
2
(9.5) ||Xu||L∞ (B(go,R)) ≤ ||u||L∞(B(go,3R)) .
R
From Theorem 2.5, and from (9.5) we obtain
C
|u (g) − u (g )| ≤ ||u||L∞(B(go,3R)) d(g, g ) , g, g ∈ B(go, R) ,
R
where C > 0 is an absolute constant. Passing to the limit as → 0 we find
C
(9.6) |u(g) − u(g )| ≤ ||u||L∞(B(go,3R)) d(g, g ) , g, g ∈ B(go , R) .
R
This shows that the function u belongs to Γ0,1

loc (G). At this point using
Theorem 2.4 we conclude from (9.6) that the weak derivatives X1 u, ..., Xmu
exist dg-a.e. in G, and they belong to L∞ n
loc (H ). Moreover, we obtain the
bound
C
||Xu||L∞(B(go,R)) ≤ ||u||L∞(B(go,3R)) ,
R
which proves (9.1), thus completing the proof.
We next establish an important improvement of Theorem 9.1 in which

we replace the L∞ norm in the right hand side of (9.1) with the L1 average
of the relevant function.
Theorem 9.2. Let u ∈ CH w (G) ∩ C(G), then there exists C = C(G) > 0
such that for every ball B(g, r) one has
1
(9.7) sup |u| ≤ C |u| dg ,
B(g,r) |B(g, 5r)| B(g,5r)
and
C 1
(9.8) ess sup |Xu| ≤ |u| dg .
B(g,r) r |B(g, 15r)| B(g,15r)
In particular, (9.7) implies that there exist no L1 functions in CH

w
(G) ∩ C(G)
other than the trivial one.
Proof. We begin by remarking that, thanks to the estimate (9.1), in order

to prove the theorem we only need to establish (9.7). Similarly to the proof
of Theorem 9.1 it is not restrictive to assume that u ∈ C ∞ (G), since we can
always work with the smooth weakly H-convex functions u = K u, and
then pass to the limit. We observe that u is in CHw
(G) if and only if for any
r > 0 the rescaled function ur (g) = u(δr g) belongs to CH w (G). This follows
from the following equation

δr g δλ (g −1 g ) = δr g δr δλ (g −1 g ) = (δr g) δλ (δr g)−1 δr g ,
and from the observation that g ∈ Hg if and only if δr g ∈ Hδr g . It therefore

suffices to prove (9.7) when r = 1. At this point to simplify the technical
details we present a detailed account of the proof in the case of the first
Heisenberg group H1 . At the end, we briefly indicate the minor adjustements
necessary to treat the general case. Let then go ∈ B(g, 1) be such that
u(go ) = sup u. By left-translation we can suppose that go = e. Consider
B(g,1)
the straight line segment γ(s) = (s, 0, 0), −1 ≤ s ≤ 1, and observe that
such segment lies on He . We define the function φ : [−1, 1] → R by the
equation φ(s) = u(γ(s)). The smoothness of u implies that φ ∈ C ∞ . A
simple computation gives φ (s) = X1 X1 u(s). Applying Theorem 5.11 we
now conclude that φ (s) ≥ 0 for every s ∈ [−1, 1], therefore φ is a convex
function on the interval [−1, 1]. Jensen’s inequality then implies
1
1
φ(0) ≤ φ(s) ds ,
2 −1
which we rewrite
1
1
(9.9) u(e) ≤ u(γ(s)) ds .
2 −1
For any s ∈ [−1, 1], we apply Proposition 5.7 obtaining for every g ∈
Hγ(s)
u(γ(s)) ≤ u(g ) − < Xu(γ(s)), x(g) − x(γ(s)) > .
We integrate the latter inequality with respect to g and to 2-dimensional
Lebesgue measure dσ(g ) on the set B(γ(s), 1) ∩ Hγ(s) , obtaining
(9.10)
u(γ(s)) dσ(g ) ≤ u(g ) dσ(g )
B(γ(s),1)∩Hγ(s) B(γ(s),1)∩Hγ(s)
− < Xu(γ(s)), x(g) − x(γ(s)) > dσ(g ) .

B(γ(s),1)∩Hγ(s)
A simple computation gives

dσ(g ) = π 1 + 4s2 ,
B(γ(s),1)∩Hγ(s)
whereas
< Xu(γ(s)), x(g ) − x(γ(s)) > dσ(g )

B(γ(s),1)∩Hγ(s)

= 1 + 4s2 X1 u(γ(s))(x − s) + X2 u(γ(s))y dx dy
(x −s)2 +(y )2 <1

= 1 + 4s2 {X1 u(γ(s)) x + X2 u(γ(s)) y} dxdy = 0 ,
x2 +y2 <1
the vanishing of the latter integral being justified by the fact that the inte-
grand is an odd function of the variable of integration (x, y). Inserting these
two equations in (9.10) we conclude
1
(9.11) u(γ(s)) ≤ |u(g )| dσ(g ) .
π B(γ(s),1)∩Hγ(s)
At this point we integrate in s ∈ (−1, 1) the inequality (9.11) and obtain

1 1
1
(9.12) u(γ(s)) ds ≤ |u(g )| dσ(g ) .
−1 π −1 B(γ(s),1)∩Hγ(s)
From (9.12), and from (9.9), we conclude

1
1
(9.13) u(e) ≤ |u(g )| dσ(g ) ds .
2π −1 B(γ(s),1)∩Hγ(s)
We now observe that the integral in the right-hand side of the latter in-
equality is finite. To see this we notice that as s varies in [−1, 1] the regions
B(γ(s), 1) ∩ Hγ(s) overlap only along the segment of the x-axis (which has
2-dimensional measure zero) joining the points (−1, 0, 0), (1, 0, 0). Further-
more, one easily sees that
(9.14) B(γ(s), 1) ∩ Hγ(s) ⊂ Ws = {(x, y, t) | |x| < 2, |y| < 1, t = −2sy} ,
and

(9.15) Ws ⊂ B(e, 4) .
s∈[−1,1]
From (9.14), (9.15), and from (9.13), we conclude
(9.16) sup u = u(e) ≤ C1 |u(g )| dg ≤ C1 |u(g )| dg .

B(g,1) B(e,4) B(g,5)
This proves that u is bounded from above in L1 norm.

To establish an analogous bound from below, we argue as follows. Let
g1 ∈ B(g, 1) be such that u(g1 ) = inf u. As before, by left-translation we
B(g,1)
can assume that g1 = e. Let g ∈ B(e, 1) ∩ He be fixed. By Proposition 4.2
we have e ∈ Hg , and therefore Proposition 5.7 allows to conclude
(9.17) u(e) ≥ u(g ) − < Xu(g ), x(g ) > .
We choose ζ ∈ Co∞ (B(e, 1)), 0 ≤ ζ ≤ 1, ζ ≡ 1 on B(e, 1/2), multiply
(9.17) by ζ(g ), and integrate with respect to dσ(g ) over B(e, 1) ∩ He to
obtain
(9.18)
u(e) ζ(g ) dσ(g )
B(e,1)∩He
≥ ζ(g )u(g ) dσ(g ) − < Xu(g ), x(g ) > ζ(g ) dσ(g ) .
B(e,1)∩He B(e,1)∩He
At this point, an important cancellation occurs in the integral involving

Xu
< Xu(g ), x(g ) > ζ(g ) dσ(g )

B(e,1)∩He
= X1 u(g )x + X2 u(g )y } ζ(g ) dσ(g )

B(e,1)∩He

= ux(g ) + 2y ut(g ) x + uy (g ) − 2x ut (g ) y ζ(g ) dσ(g )
B(e,1)∩He
= x ux (g ) + y uy (g ) ζ(g ) dσ(g )
B(e,1)∩He
= − (x ζ)x + (y ζ)y u(g ) dσ(g ) ,

B(e,1)∩He
where we notice that when integrating by parts the boundary integral on

∂B(e, 1) ∩ He vanishes due to our choice of ζ. From the latter equation, the
inequality
σ(B(e, 1/2) ∩ He ) ≤ ζ(g ) dσ(g ) ≤ σ(B(e, 1) ∩ He ) ,

B(e,1)∩He
and from (9.18), we conclude the existence of a constant C2 > 0 such that
(9.19) u(e) ≥ − C2 |u(g )| dσ(g ) .

B(e,1)∩He
We now assert that an estimate similar to (9.19) holds at any point γ(s),
s ∈ [−1, 1], with an absolute constant C2 > 0 independent of s, that is
(9.20) u(γ(s)) ≥ − C2 |u(g )| dσ(g ) .

B(γ(s),1)∩Hγ(s)
The details of the verification of (9.20) are similar to those in the proof
of (9.19) and are left to the reader. If we now let, as in the proof of (9.16),
φ(s) = u(γ(s)), then we can rewrite (9.20) as follows
(9.21) φ(s) ≥ − C2 |u(g )| dσ(g ) , −1 ≤ s ≤ 1 .

B(γ(s),1)∩Hγ(s)
Integrating (9.21) on the interval [0, 1], and using (9.14), (9.15), we obtain
1
(9.22) φ(s) ds ≥ − C2 |u(g )| dg ≥ − C2 |u(g )| dg .
0 B(e,4) B(g,5)
On the other hand, the convexity of φ on [0, 1] gives
φ(s) ≤ (1 − s) φ(0) + s φ(1) , 0≤s≤1.
Integrating the latter inequality on [0, 1] one finds

1
1 1 1 1
φ(s) ds ≤ φ(0) + φ(1) = u(e) + u(go) ,
0 2 2 2 2
where we have let go = (1, 0, 0) ∈ B(e, 1). Invoking (9.16), in which we take
g = e, we obtain
u(go ) ≤ sup u ≤ C1 |u(g )| dg ≤ C1 |u(g )| dg .

B(e,1) B(e,4) B(g,5)
Inserting this estimate in the previous inequality we conclude

1
1 C1
(9.23) φ(s) ds ≤ u(e) + |u(g )| dg .
0 2 2 B(g,5)
Finally, the inequalities (9.22) , (9.23) imply, with C = max {C1 , C2 },
inf u = u(e) ≥ − 3 C |u(g )| dg .

B(g,1) B(g,5)
This establishes the sought for estimate from below, which, together with
(9.16), proves (9.7) when G = H1 . In the case of a general Carnot group G
we define γ : [−1, 1] → G as follows
γ(s) = exp (sX1 ) ,
where as always X1 denotes the first element of a fixed orthonormal basis of

the first layer V1 of g. We note explicitly that the trace of the curve γ lies in
the horizontal plane He . With this choice of γ(s), the function φ(s) is defined
as before, i.e., φ(s) = u(γ(s)). As before we have φ (s) = X1 X1 u(γ(s)) ≥ 0,
thus concluding that φ is convex on [−1, 1]. From this point on the proof
proceeds almost verbatim as for the case of H1 , and we thus leave the details
to the reader.
10. The horizontal Monge-Ampère equation.
One of the most important aspects of the theory of convex functions is its
connection with geometry through the Monge-Ampère equation. Beautiful
accounts of this very rich and well-developed subject, including an exaustive

list of references, are given in the books [Po], [Ba3], [GT], [CC], [Gu]. The
Monge-Ampére equation
det D 2 u = f ,
is fully nonlinear, and becomes elliptic on C 2 strictly convex functions u. It
is deeply connected to the geometry of the graph of u in Rn+1 . For instance,
the Gauss curvature of the graph is given by the equation

n
det D 2 u
(10.1) K = κj = ,
j=1
(1 + |Du|2 )(n+2)/2
where κ1 , ..., κn denote the principal curvatures.

In this perspective, and in consideration of Theorem 5.12, it seems ap-
propriate to introduce a fully nonlinear operator which becomes “elliptic” on
functions which are strictly weakly H-convex. We expect that, similarly to
its classical ancestor, such operator will play an important role in a number
of problems ranging from geometry to optimal control and calculus of vari-
ations. Our intention here is simply to point at some of its basic features,
deferring a systematic analysis to the forthcoming article [DGN3].
Definition 10.1. Let G be a Carnot group. The horizontal Monge-Ampère

operator is defined on functions u ∈ Γ2 (G) by the equation
F [u] = det (X 2 u) .
More in general we can consider nonlinear equations given by
(10.2) F [g, u, Xu, (X 2u) ] = F ,

2
where the variable (g, u, ζ, M ) ranges in the set G × R × Rm × Rm , with
M = (Mij ) belonging to the space of m × m symmetric matrices. The
equation (10.2) is called H-fully nonlinear if it is not linear in the variable
M.
Definition
10.2. The equation (10.2) is called H-elliptic if the matrix
∂F
∂Mij is positive definite, i.e., if for every ζ ∈ Rm \ {0} one has
i,j=1,...,m

m
∂F
ζi ζj > 0 .
∂Mij
i,j=1
According to Definition 10.2, in view of Theorem 5.12 we see that the

horizontal Monge-Ampère equation
(10.3) det (X 2 u) = F
is H-elliptic only for functions u ∈ Γ2 (G) which are strictly weakly H-
convex. In such case, the function F in the right-hand side must be strictly
positive.
We now return to the discussion in the closing of Section 5, with the
purpose of introducing a basic definition.
Definition 10.3. Given a Carnot group G consider a function u ∈ Γ2 (G).

We define the horizontal Gaussian curvature of the graph of u at the point
(go , u(go)) ∈ G × R by the formula

det X 2 u(go)
KH (go ) = (m+2)/2 .
1 + |Xu(go)|2
Definition 10.3 has the following important geometric interpretation.
Consider the horizontal plane Hgo passing through the point go , and as
in (5.16) let Ugo denote the restriction of u to Hgo . Using the inverse of the
exponential mapping we can think of Ugo as a function of the coordinates

x = x(g) = (x1 (g), ..., xm(g)), where ξ1 (g) = m α=1 xα (g)Xα ∈ V1 . Using
considerations similar to those that lead to (5.16), we recognize that

det D2 Ugo (x(go ))
(10.4) KH (go ) = (m+2)/2 .
1 + |DUgo (x(go))|2
In view of (10.1) we conclude from (10.4) that the horizontal
Gaussian curvature KH (go ) is just the classical Gaussian curvature at
(x(go ), Ugo (x(go)) of the graph of the restriction of u to the horizontal plane
Hgo . This point of view was introduced in the paper [DGN2] in connection
with the theory of minimal surfaces and the Bernstein problem in Carnot
groups, see also [GP], and in relation to the equation (10.3) it will be further
developed in [DGN3].
We can also introduce the horizontal Gaussian curvature of an hypersur-
face in G defined as a level set. Consider a function φ ∈ C 2 (G) and suppose
that the level
Ω = {g ∈ G | φ(g) < 0}
is non-critical. By this we mean that there exist a neighborhood U of S =
∂Ω, and δ > 0, such that the Riemaniann gradient of φ satisfy
|Dφ(g)| ≥ δ , for every g ∈ U .
In this case S constitutes a C 2 hypersurface in G. We assume that the

point go ∈ S be non-characteristic, i.e., that we have |Xφ(go)| = 0. By
continuity, this inequality continues to be valid in a full neighborhood of go.
We consider the restriction Φgo of the defining function φ to the horizontal
plane Hgo , so that as above we can think of Φgo as a function of the variable
x = x(g) ∈ Rm . Let
Sgo = {x ∈ Rm | Φgo (x) < 0} ,
and notice that, thanks to the non-characteristic assumption, and to

(5.16), the set Sgo is a C 2 hypersurface in Rm nearby x(go ) ∼ = ξ1 (go) =
m
α=1 xα (go )Xα . Set xo = x(go ) ∈ R , and denote by κi (xo ), i = 1, ..., m − 1
m
the principal curvatures of Sgo at xo . We define the horizontal Gaussian

curvature of S at go as the standard Gaussian curvature of Sgo at xo , i.e.,
def
m−1
KH (S; go) = κi (xo) .
i=1
When the point go is a characteristic point of the hypersurface S, then

equation (5.16) gives
DΦgo (xo ) = Xφ(go) = 0
and we cannot apply the above considerations. In this case the situation
is more complex, as one needs to identify geometric conditions on S near
go which allow to define the horizontal principal curvatures by appropriate
limiting relations. For these aspects we refer the reader to the forthcoming
article [DGN3].
We close this section by introducing some geometric notions of convexity
which will prove important in the study of the equation (10.3), and of its
connections to the problem stated in the introduction. In the process, we
will also propose an appropriate notion of viscosity solution. We begin by
developing a few preliminary properties. We recall Definition 3.6 of an H-
affine function A : G → R
A(g ) = < p, ξ1(g ) > + β g ∈ G .
Given an open set Ω ⊂ G, and a function u : Ω → R, we say that an

H-affine function A supports u at go ∈ Ω if
(10.5)
A(go) = u(go ) , and A(g ) ≤ u(g ) for every g ∈ Ω ∩ Hgo .
Clearly, the first condition in (10.5) implies the existence of p(go) ∈ V1

such that
(10.6) A(g ) = u(go) + < p(go ), ξ1(g ) − ξ1 (go ) > ,
for all g ∈ G. The vector p(go) is called a horizontal subgradient of u at go .
Definition 10.4. We define the horizontal normal mapping of u as the set-

valued function ∂H u : Ω → P(V1 ) given by
∂H u(go ) = {p(go) | p(go) is a horizontal subgradient for u at go } .
For any E ⊂ Ω we set

∂H u(E) = ∂H u(go) .
go ∈E
The horizontal lower contact set of u in Ω is defined as follows
Γ−
H (u; Ω) = {go ∈ Ω | ∂H u(go) = ∅} .
It is clear from (10.5) and (10.6) that ∂H u(go) = ∅ if and only if u admits
a support at go .
Proposition 10.5. Let Ω ⊂ G be an open weakly H-convex set, if
Γ−
H (u; Ω) = Ω ,
then u ∈ CH
w
(Ω).
Proof. Fix arbitrarily go ∈ Ω, we need to show that with gλ = go δλ (go−1 g)
u(gλ ) ≤ (1 − λ) u(go ) + λ u(g) ,
for any g ∈ Ω ∩ Hgo , and every λ ∈ [0, 1]. From Proposition 4.2 we know
gλ ∈ Ω ∩ Hgo . Let A be a support for u at gλ . From (10.6) we have
(10.7) A(g ) = u(gλ ) + < p(gλ ), ξ1 (g ) − ξ1 (gλ ) > g ∈ Ω ,
for some p(gλ ) ∈ V1 , whereas (10.5) gives

(10.8)
u(gλ ) + < p(gλ), ξ1 (g ) − ξ1 (gλ ) > ≤ u(g ) for every g ∈ Ω ∩ Hgλ .
It is now easy to verify that

(10.9) u(gλ ) = (1 − λ) A(go) + λ A(g) .
We have in fact from (2.7)
ξ1 (gλ ) = ξ1 (go) + λ [ξ1 (g) − ξ1 (go)] ,
and therefore (10.7) gives
(1 − λ) A(go ) + λ A(g) = (1 − λ) {u(gλ ) + < p(gλ ), ξ1 (go) − ξ1 (gλ ) >}
+ λ {u(gλ ) + < p(gλ), ξ1 (g) − ξ1 (gλ ) >}
= u(gλ ) ,
which proves (10.9). Using (10.9), we would be finished if we could say that
A(go ) ≤ u(go) , A(g) ≤ u(g) .
These inequalities would follow from (10.8) if we knew that
g o ∈ Hgλ , g ∈ Hgλ .
The former relation is obvious from Proposition 4.2 since g ∈ Hgo implies
gλ ∈ Hgo , and this is equivalent to go ∈ Hgλ . The second relation is less
obvious, but it is equally true. It suffices to observe that g ∈ Hgλ is equivalent
to gλ−1 g ∈ He , which is the same as g −1 gλ ∈ He . Now we can write this as
g −1 gλ = h−1 δλ h ∈ He , where h = go−1 g. To see that this latter relation
is true it is enough to observe that g ∈ Hgo is equivalent to h ∈ He , and
then use the Baker-Campbell-Hausdorff formula (2.7) to verify that for any
h ∈ He one has h−1 δλ h ∈ He . This completes the proof of the proposition.
Proposition 10.6. Let Ω ⊂ G be an open set, and let u ∈ Γ1 (Ω). If

go ∈ Γ−
H (u; Ω), then
∂H u(go ) = {Xu(go)} .
If furthermore u ∈ Γ2 (Ω), then
2
X u(go) ≥ 0.
Proof. By the assumption go ∈ Γ− H (u; Ω) we know that ∂H u(go ) = ∅. To

establish the first part we show that if p(go ) ∈ ∂H u(go), then p = Xu(go).
Now if p(go ) ∈ ∂H u(go) we have
u(go ) + < p(go), ξ1 (g ) − ξ1 (go ) > ≤ u(g ) , g ∈ Ω ∩ Hgo .
Fix g ∈ Ω ∩ Hgo , and consider gλ = go δλ (go−1 g). By Proposition 4.2 we

know gλ ∈ Ω ∩ Hgo , therefore the previous inequality applied to g = gλ gives
< p(go), ξ1(gλ ) − ξ1 (go ) > ≤ u(gλ ) − u(go ) ,
which, using (2.7), we can rewrite
u(gλ ) − u(go)
< p(go), ξ1 (g) − ξ1 (go) > ≤ .
λ
Passing to the limit as λ → 0 in the latter inequality, invoking Proposition
2.3 we conclude
< p(go ), ξ1(g) − ξ1 (go ) > ≤ < Xu(go), ξ1(g) − ξ1 (go ) > g ∈ Ω ∩ Hgo .
By left-translation this inequality can be equivalently rewritten as follows
< p(go ), ξ1(h) > ≤ < Xu(go), ξ1 (h) > h ∈ Ω̃ ∩ He ,
where we have let Ω̃ = go−1 Ω. Since the open set Ω̃ contains the identity,
there exists R > 0 sufficiently small such that B(e, R) ⊂ Ω̃, therefore the
latter inequality holds in particular for every h ∈ B(e, R) ∩ He . However,
B(e, R) ∩ He is closed under the inversion h → h−1 , and on the other hand
ξ1 (h−1 ) = −ξ1 (h). Since when h ∈ B(e, R)∩He the vector ξ1 (h) describes the
set Be (0, R), the m-dimensional Euclidean ball in V1 centered at the origin
and having radius R, we thus conclude that it must be p(go) = Xu(go). This
completes the proof of the first part. The second part now follows from the
first, and from the proof of the necessity in Theorem 5.12, see also Theorem
5.11.
We next compute an important example of normal mapping.
Proposition 10.7. Let G be a Carnot group and consider the gauge

pseudo-ball B(go , R). For any a > 0 let
def a
u(g) = d(g, go) .
R
The graph of u is the cone in G×R with vertex at (go , 0) and basis B(go , R)×
{a}. The horizontal normal mapping of u is given by



a
 R Xd(g, go)

g = go ,

∂H u(g) =



B 0, a

g = go ,
e R

where we have denoted by Be 0, Ra the m-dimensional Euclidean ball in V1
centered at the origin and having radius a/R.
Proof. From the definition (2.12) of the pseudo-distance we see that g →

d(g, go) belongs to C ∞ (G \ {go}), therefore ∂H u(g) = Ra Xd(g, go) for g = go
immediately follows from Proposition 10.6. Next, we observe that p ∈ V1
belongs to ∂H u(go) if and only if
a
< p, ξ1 (g) − ξ1 (go ) > ≤ d(g, go)
R
for every g ∈ B(go , R) ∩ Hgo . By invariance with respect to left-translation
the latter inequality is equivalent to
a
< p, ξ1 (h) > ≤ N (h)
R
for every h ∈ B(e, R) ∩ He . In view of (9.4) we conclude that p ∈ ∂H u(go)
if and only if
ξ1 (h) a
< p, > ≤ ,
|ξ1 (h)| R
for every h ∈ B(e, R) ∩ He . This is like saying that p ∈ Be (0, a/R).
The converse to Proposition 10.5 is a rather delicate point which we plan

to discuss in [DGN3]. Here, to continue our discussion, we a priori assume
such property. That is we consider, for a fixed weakly H-convex open set
Ω ⊂ G, a function u ∈ CHw
(Ω) ∩ C(Ω), such that
(10.10) ∂H u(go ) = ∅ for every go ∈ Ω .
We stress explicitly that when u ∈ CH

w
(Ω)∩Γ1 (Ω), then thanks to Propo-
sition 10.6 we have ∂H u(go) = {Xu(go)}, and therefore (10.10) is guaranteed.
Suppose now that φ ∈ Γ2 (Ω) and that u − φ restricted to the horizontal
plane Hgo has a local maximum in go ∈ Ω. We claim that it must be
(10.11) det (X 2 φ(go)) ≥ 0 .
To prove (10.11) we argue as follows. We fix arbitrarily g ∈ Ω ∩ Hgo

and consider the function ψ(λ) = φ(gλ ), where gλ = goδλ (go−1 g). ¿From the
assumption on φ we have ψ ∈ C 2 (0, 1), therefore Taylor’s formula gives
ψ (0) 2
ψ(λ) = ψ(0) + ψ (0) λ + λ + o(λ2 ) .
2
From the latter equation, and from (5.12) and (5.13) in Proposition 5.4,
we conclude
(10.12)
φ(gλ ) = φ(go) + < Xφ(go), ξ1 (g) − ξ1 (go ) > λ
1
+ < (X 2 φ(go)) (ξ1 (g) − ξ1 (go)), (ξ1(g) − ξ1 (go)) > λ2 + o(λ2 ) .
2
From Proposition 4.2 and from the weak H-convexity of Ω we know that
gλ ∈ Ω ∩ Hgo for every λ ∈ [0, 1]. Since the restriction of u − φ to Hgo has a
local maximum in go , we conclude that there exists λo > 0 sufficiently small
such that
u(gλ ) ≤ φ(gλ ) + u(go ) − φ(go) for 0 < λ < λo .
Inserting (10.12) in the latter inequality we find
(10.13)
u(gλ ) ≤ u(go ) + < Xφ(go), ξ1(g) − ξ1 (go) > λ
1
+ < (X 2φ(go )) (ξ1 (g) − ξ1 (go )), (ξ1(g) − ξ1 (go )) > λ2 + o(λ2 ) .
2
From the assumption (10.10) we infer the existence of p(go ) ∈ V1 such
that
< p(go), ξ1 (g ) − ξ1 (go ) > ≤ u(g ) − u(go) for every g ∈ Ω ∩ Hgo .
Taking g = gλ in the latter inequality, and substituting in (10.13) we

find
(10.14)
< p(go), ξ1 (gλ ) − ξ1 (go ) > ≤ < Xφ(go), ξ1(g) − ξ1 (go ) > λ
1
+ < (X 2 φ(go)) (ξ1 (g) − ξ1 (go)), (ξ1(g) − ξ1 (go)) > λ2 + o(λ2 ) .
2
By (2.7) one has ξ1 (gλ ) − ξ1 (go) = λ(ξ1(g) − ξ1 (go)), therefore we obtain
from (10.14)
(10.15) < p(go), ξ1 (g1 ) − ξ1 (go) > ≤ < Xφ(go), ξ1 (g) − ξ1 (go ) >
1
+ < (X 2 φ(go )) (ξ1 (g) − ξ1 (go)), (ξ1(g) − ξ1 (go)) > λ + o(λ) .
2
Letting λ → 0 in (10.15) gives

(10.16) < p(go), ξ1 (g) − ξ1 (go) > ≤ < Xφ(go), ξ1 (g) − ξ1 (go) > .
By the arbitrariness of g ∈ Ω ∩ Hgo we conclude from (10.16) that must
be p(go) = Xφ(go). But then (10.15) implies
< (X 2φ(go )) (ξ1 (g) − ξ1 (go )), (ξ1(g) − ξ1 (go )) > ≥ 0 .
Again the arbitrariness of g shows that the matrix (X 2 φ(go )) is semi-
definite positive, and therefore (10.11) holds. This proves the claim.
These considerations, and the results of Sections 5 and 7, suggest an
appropriate notion of viscosity solution for the equation 10.3 which will prove
central in its study. Although such notion is inspired to the by now classical
one due to Crandall-Ishii-Lions [CIL], it is essentially modelled on that of
weakly H-convex function, and therefore one has to necessarily confront
the delicate problems deriving from an information on lower dimensional
manifolds.
Definition 10.8. Given a weakly H-convex open set Ω in a Carnot group

G, and a function F ∈ C(Ω), F ≥ 0, we say that a function u ∈ C(Ω)∩CH w
(Ω)
satisfying (10.10) is a viscosity subsolution (supersolution) to (10.3) if for
any φ ∈ Γ2 (Ω) ∩ CHw
(Ω) and go ∈ Ω such that the restriction of u − φ to the
horizontal plane Hgo has a local maximum (minimum) at go , then it must
be
det [(X 2φ(go )) ] ≥ (≤) F (go ) .
If u is at the same time a viscosity sub/supersolution, then we say that it is
a viscosity solution to (10.3).
The notion of viscosity solution is intimately connected to that of solution
in the sense of the horizontal Monge-Ampère measure associated with the
horizontal normal mapping ∂H u. The latter is defined by the equation
def
MH u(E) = Hm (∂H u(E)) ,
where Hm denotes the standard m-dimensional Hausdorff measure on the
Lie algebra g (we recall that m = dim V1 ), and E ⊂ Ω is such that ∂H u(E) is
Hm -measurable. We will come back in [DGN3] to the fundamental question
of the equivalence of these two notions of solutions.
We close this section with one more witness of the remarkable role played
by the gauge in groups of Heisenberg type. Combined with Theorem 8.1, the
next result states that the gauge is a fundamental solution of the horizontal
Monge-Ampère equation.
Theorem 10.9. Let G be a group of Heisenberg type and N (g) be the

gauge defined in (6.3), then
det (X 2 N (g)) = 0 if g = e .
Proof. To prove the theorem we show that the horizontal Hessian matrix
(X 2 N (g)) is singular for g = e by exhibiting a non-zero vector p = p(g) ∈
Rm V1 in the kernel of the corresponding linear map associated to this
matrix. In fact, we take p = Xu = (X1 u, ..., Xmu) where u(g) = N (g)4.
Since by (6.23),

2 1 2 3
(X N ) =
(X u) − (Xu ⊗ Xu) ,
4 N3 4 N4
to prove the theorem it will suffice to show

3
(10.17) (X 2u) − (Xu ⊗ Xu) (Xu) = 0 in G \ {e} .
4 N4
To this end, we first observe that Xu = 4N 3 XN , hence using (6.11) in

Lemma 6.3 we find
|Xu|2 = 16 N 6 |XN |2 = 16 N 4 ψ .
On the other hand, we have Xu = Xψ 2 + 16Xχ, and so we find from

(6.20), (6.24)
(10.18) Xu = 4 ψ ξ1 + 16 J(ξ2 )ξ1 .
Since (Xu ⊗ Xu)Xu = |Xu|2Xu, we conclude

(10.19)
3
[Xu ⊗ Xu] (Xu) = 12 ψ Xu = 12 ψ 4 ψ ξ1 + 16 J(ξ2 )ξ1 .
4 N4
We next turn to the computation of (X 2 u) (Xu). By (6.19) we have
(10.20) (X 2 u) (Xu) = (X 2 ψ 2 ) (Xu) + 16 (X 2 χ) (Xu)

= (X 2 ψ 2 ) (Xψ 2) + 16 (X 2 ψ 2 ) (Xχ)
+ 16 (X 2χ) (Xψ 2) + 162 (X 2χ) (Xχ) .
Using (6.21) one easily recognizes that
(10.21) (X 2ψ 2 ) (Xψ 2) = 48 ψ 2 ξ1 .
¿From (6.21) and from (6.24) we obtain for the α−th component of
16(X 2ψ 2 ) (Xχ)
m
16 4δαα ψ + 8xαxα < J(ξ2 )ξ1 , Xα >
α =1

= 16 4ψ < J(ξ2 )ξ1 , Xα > + 8xα < J(ξ2 )ξ1 , ξ1 > .
From this formula we conclude
(10.22) 16 (X 2 ψ 2) (Xχ) = 64 ψ J(ξ2 )ξ1 .
Next, from (6.16) and (6.20) we obtain for the α−th component of
16(X 2χ) (Xψ 2)

k
m
32 ψ < [ξ1 , Xα], Yβ >< [ξ1 , Xα ], Yβ > xα
β=1 α =1

k
= 32 ψ < [ξ1 , Xα], Yβ >< J(Yβ )ξ1 , ξ1 > = 0 .
β=1
This gives
(10.23) 16(X 2χ) (Xψ 2) = 0 .
Finally, from (6.16) and (6.24) we find for the α−th component of
16 (X 2 χ) (Xχ)
2
162
k m
< [ξ1 , Xα], Yβ > < [ξ1 , Xα ], Yβ >< J(ξ2 )ξ1 , Xα >
2
β=1 α =1
162
k
= < [ξ1 , Xα], Yβ >< J(Yβ )ξ1 , J(ξ2 )ξ1 >
2
β=1
162
k
= < [ξ1 , Xα], Yβ >< Yβ , ξ2 > |ξ1 |2
2
β=1
2
16 162
= ψ < [ξ1 , Xα], ξ2 > = ψ < J(ξ2 )ξ1 , Xα > .
2 2
This gives
162
(10.24) 162 (X 2 χ) (Xχ) = ψ J(ξ2 )ξ1 .
2
Substituting (10.21) - (10.24) in (10.20), we obtain

2
(10.25) (X u) (Xu) = 12 ψ 4 ψ ξ1 + 16 J(ξ2 )ξ1 .
Comparing (10.25) with (10.19) we conclude that (10.17) holds. This

completes the proof.
11. Yet another notion of convexity.
The most elementary properties of classical convex functions are their be-
ing locally bounded and continuous on their effective domain. Correspond-
ing properties for weakly H-convex functions are far from obvious and are
presently being investigated. The purpose of this section is to prove that
by strengthening Definition 5.5, but of course without falling back into the
notion of strong H-convexity, one obtains an “intermediate” class of convex
functions which, remarkably, share with the classical ones the above men-
tioned properties. This implies in particular that the results in Sections 8
and 9 are valid generically for such functions, without any additional as-
sumption.
To set the stage, consider in the Heisenberg group a function u : Hn → R.
For every go = (xo, yo , to ) ∈ Hn denote by Ugo the restriction of u to the
horizontal plane Hgo , i.e.,
def
Ugo (x, y) = u(x, y, to + 2 (< yo , x > − < xo , y >)) , (x, y) ∈ R2n .
Definition 11.1. A function u : Hn → R is called H-convex if for any fixed

go ∈ Hn the function Ugo is an ordinary convex function of the variables
(x, y) ∈ R2n . The class of all H-convex functions will be denoted by CH (Hn ).
To understand better Definition 11.1, consider u ∈ CH (Hn ). For any
go ∈ Hn , (x, y), (x, y ) ∈ R2n , and 0 ≤ λ ≤ 1 one has
Ugo (x + λ(x − x), y + λ(y − y)) ≤ (1 − λ) Ugo (x, y) + λ Ugo (x , y ) .
Equivalently, we can write

u x + λ(x − x), y + λ(y − y),

to + 2[< yo , x + λ(x − x) > − < xo , y + λ(y − y) >]

≤ (1 − λ) u x, y, to + 2 (< yo , x > − < xo , y >)

+ λ u x , y , to + 2 (< yo , x > − < xo , y >) .
If we now define t = to + 2(< yo , x > − < xo , y >), t = to + 2(< yo , x >

− < xo , y >), then it is clear that the points g = (x, y, t), g = (x , y , t )
belong to Hgo , and since t − t = 2(< yo , x − x > − < xo , y − y >), the latter
inequality is equivalent to writing
(11.1)
u((1 − λ)g + λg ) ≤ (1 − λ) u(g) + λ u(g ) , g, g ∈ Hgo , 0 ≤ λ ≤ 1 ,
where we have denoted with (1 − λ)g + λg the ordinary Euclidean convex

combination of g, g , as points of R2n+1 . We conclude that u is H-convex if
and only if given any horizontal plane Hgo ⊂ Hn , the inequality (11.1) holds.
We remark explicitly that Definition 11.1 continues to make sense in
any Carnot group. It is also immediate to recognize that if a function u
is (Euclidean) convex in the ordinary sense at the level of the Lie algebra,
then it is also H-convex. In fact, a function f : RN → R is convex if and
only if its restriction to any hyperplane in RN is convex. Since the notion of
H-convexity involves restricting only to horizontal planes, it is clear that the
class CH (Hn ) includes Euclidean convex functions. A comparison with our
Definition 5.5 reveals that any function which is H-convex is in particular
weakly H-convex since we only request convexity with respect to the base
point go . The next result shows that the inclusion between the two classes
is strict.
Proposition 11.2. If u : Hn → R is H-convex, then it is also weakly H-

convex. Furthermore, there exists weakly H-convex functions which are not
H-convex. The gauge N (g) is an example of such a function.
Proof. The inclusion CH (Hn ) ⊂ CHw (Hn ) is clear since for any given g ∈ Hn ,

Definition 11.1 gives for any g ∈ Hg and 0 ≤ λ ≤ 1
u((1 − λ)g + λg ) ≤ (1 − λ) u(g) + λ u(g ) .
On the other hand, Proposition 4.4 implies
(1 − λ) g + λ g = gλ ,
which shows that if u ∈ CH , then also u ∈ CH w.
To prove that the inclusion is strict, consider in the Heisenberg group

H1 the gauge N (g) = ((x2 + y 2 )2 + t2 )1/4. According to Theorem 6.8, N (g)
is weakly H-convex. However, it is not H-convex. To see this consider the
point go = (−1/2, 0, 0). The horizontal plane Hgo has equation t = y, and
therefore the restriction of N (g) to Hgo is given by
1/4
Ugo (x, y) = (x2 + y 2 )2 + y 2 , (x, y) ∈ R2 .
If Ugo (x, y) were an ordinary convex function, then its restriction to the
y-axis
1/4
Ugo (0, y) = y 4 + y 2
would be a convex function of y, and this is clearly false since Ugo (0, y) ≈
|y|1/2 for |y| ≈ 0.
A basic property of classical convex functions is that they are locally

bounded in the interior of their effective domain, and furthermore continu-
ous. Such properties are far from obvious in the non abelian setting of this
paper. The purpose of what follows is to prove that H-convex functions
enjoy them generically. Our first theorem in this direction proves the local
boundedness of u. In preparation for it, we recall the following classical
result, see Theorem 1.5.1 in [Sch].
Theorem 11.3. Let f : RN → R be a convex function. For any ball

Be (xo , R), consider a (closed) simplex S formed by the points {x1 , ..., xN +1},
◦
and such that B e (xo , R) ⊂ S. Let M = max {f (xi ) | i = 1, ..., N + 1}, then
one has
|y − xo |
|f (y) − f (xo )| ≤ [M − f (xo )] , for every y ∈ B e (xo , R) .
R
Furthermore,
|y − x|
|f (y)−f (x)| ≤ 4 {[M − f (xo )] + |f (xo )|} x, y ∈ B e (xo , R/2) .
R
In particular, f is Lipschitz continuous on B e (xo, R/2).
We note explicitly that the convexity of f guarantees that M ≥ f (xo ).
Moreover, from the first inequality in Theorem 11.3 one obtains the estimate
(11.2)
|f (y)| ≤ [M−f (xo)]+|f (xo )| ≤ M+|f (xo )|, for every y ∈ B e (xo , R) .
We now return to the Heisenberg group Hn .
Theorem 11.4. Let u : Hn → R be a H-convex function, then u ∈

L∞ n
loc (H ).
Proof. We present the proof in the case of the first Heisenberg group H1 .
The case n ≥ 2 only requires minor modifications, which are left to the
reader. Consider the four points
e+
1 = (1, 0, 0) , e−
1 = (−1, 0, 0) , e+
2 = (0, 1, 0) , e+
2 = (0, 1, 0) .
The four straight line segments
L+
1 = (1, s, −2s) , L−
1 = (−1, s, 2s) ,
L+
2 = (s, 1, 2s) , L−
2 = (s, −1, −2s) ,
with −1 ≤ s ≤ 1, are respectively compact subsets of the horizontal planes

He+ , He− , He+ , He− . By the hypothesis, u restricted to each of these four
1 1 2 2
planes is an ordinary convex function. Appealing to the inequality (11.2) we
infer the existence of four constants Mi± > 0, i = 1, 2, such that
|u(g)| ≤ Mi± , for every g ∈ L±

i , i = 1, 2 .
Setting M = max {M1+ , M1− , M2+, M2− }, we conclude
(11.3) |u(g)| ≤ M , for every g ∈ L+ − + −

1 ∪ L1 ∪ L2 ∪ L2 .
Having established (11.3) we now consider for every −1 ≤ t ≤ 1 the point

gt = (0, 0, t) and the horizontal plane Hgt = {g = (x, y, t) | (x, y) ∈ R2 }.
Such plane intersects the segments L± ±
1 , L2 in the four points
t t
P1+ (t) = 1, − , t , P1− (t) = − 1, , t ,
2 2
+
t −
t
P2 (t) = , 1, t , P2 (t) = − , −1, t .
2 2
We now consider the four additional points
1 1
ẽ±
1 = (± , 0, 0) ẽ±
2 = (0, ± , 0) ,
2 2
and construct corresponding straight lines segments
1 1
L̃±
1 = (± , s, ∓s) , L̃±
2 = (s, ± , ±s) , −1 ≤ s ≤ 1 ,
2 2
as above. Such segments intersect the plane Hgt in the four points
1 1
(11.4) P̃1± (t) = ± , ∓t, t , P̃2± (t) = ± t, ± , t .
2 2
Denote by S̃(t) the simplex in the plane Hgt generated by these four
points. A moment’s thought and elementary geometric considerations will
convince the reader that

U = S̃(t)
−1≤t≤1
constitutes a full neighborhood of the group identity e, hence there exists

δ > 0 such that
B(e, δ) ⊂ U .
Similarly to (11.3), there exists a constant M̃ > 0 such that
(11.5) |u(g)| ≤ M̃ , for every g ∈ L̃+ − + −

1 ∪ L̃1 ∪ L̃2 ∪ L̃2 .
Fix t ∈ [−1, 1]. Applying the estimate (11.2) we infer that
(11.6) sup |u(g)| ≤ C(t) ,

g∈S̃(t)
where C(t) > 0 is a constant which depends only on
max {u(P1+ (t)), u(P1−(t)), u(P2+ (t)), u(P2−(t)), |u(P̃o(t)|} ,
where P̃o (t) is an arbitrarily fixed point in S̃(t), for instance P̃1+ (t). However,
the two estimates (11.3) and (11.5) guarantee that
sup C(t) ≤ C = C(M, M̃ ) < ∞ .

−1≤t≤1
From these considerations, we conclude that
(11.7) sup |u| ≤ C(M, M̃ ) .

B(e,δ)
This proves that u ∈ L∞ (B(e, δ)). Since by left-translation we can move

any point to the group identity, the conclusion of the theorem follows.
Remark 11.5. It should be clear to the reader that by suitably varying the
choice of the points e±
i in the proof of Theorem 11.4 we can make the radius
δ in the estimate (11.7) arbitrarily large.
Theorem 11.6. Let u : Hn → R be a H-convex function, then u ∈ C(Hn ).

Proof. Again, we present the proof only for H1 , and leave the minor
adjustements for n ≥ 2 to the reader. By left-translation it will suffice
to prove that u is continuous at the group identity e = (0, 0, 0). From
Theorem 11.4, and Remark 11.5 after it, we know that u ∈ L∞ (B(e, 1)).
Set M = ||u||L∞ ((B(e,1)). We will prove that there exist ηo = ηo > 0, and
C = C(M ) > 0 such that for every go = (xo , yo , to) ∈ B(e, ηo) we have
|u(go) − u(e)| ≤ C d(go , e) .
We first consider the case when go belongs to the center of the group, i.e.,
go = (0, 0, to), and without loss of generality we assume to > 0. Furthermore,
we impose that to ≤ ηo2√, where ηo > 0 will be suitably chosen in the sequel.
Consider the point (− 2to , 0, 0) ∈ He , and the straight-line segment L1 (to )
lying on He which joins e to such point. The horizontal plane H(− √to ,0,0)
√ 2
has equation t = to y. On such√plane we consider √
the straight-line segment
√
L2 (to ) which joins the points (− 2to , 0, 0) and (− 2to , to , to ). Finally, denote
by L3 (to ) the straight-line segment lying on the horizontal plane H(0,0,to)
√ √
joining (− 2to , to , to ) to (0, 0, to). We have
√
to √
(11.8) |u(go) − u(e)| ≤ |u(0, 0, to) − u(− , to , to )|
√ 2 √
to √ to
+ |u(− , to , to ) − u(− , 0, 0)|
√2 2
to
+ |u(− , 0, 0) − u(0, 0, 0)|
2 √
to √
= |U(0,0,to)(0, 0) − U(0,0,to) (− , to )|
√ 2 √
to √ to
+ |U(− to ,0,0)(−
√
, to ) − U(− to ,0,0)(−
√
, 0)|
2 2 2 2
√
to
+ |U(0,0,0)(− , 0) − U(0,0,0)(0, 0)| ,
2
where we have denoted by Ugo (x, y) the restriction of the function u(x, y, t)
to the horizontal plane Hgo . By the assumption, each of the functions
U(0,0,to) (x, y), U(− √to ,0,0)(x, y), U(0,0,0)(x, y) is convex, and therefore locally
2
lipschitz. As to varies in the interval [0, ηo2], the segments L1 (to ), L2 (to ) and
L3 (to ) respectively vary in compact subsets K1 (ηo), K2 (ηo ), K3 (ηo ) of the
horizontal planes H(0,0,0), H(− √to ,√t ,t ) , H(0,0,to) . We now choose η in such
2 o o
a way that Ki (ηo) ⊂ B(e, 12 ), i = 1, 2, 3. In view of Theorem 11.3 the local

Lipschitz constants of the functions U ’s on such compact sets are respec-

tively bounded by the maximum of their absolute values on finite simplices
which enclose them. But by our choice of ηo we can choose these simplices
so that they are all contained in B(e, 1), and therefore the local Lipschitz
constants are uniformly bounded by M . We conclude from (11.8)
√
5√ √ 1√
|u(go) − u(e)| ≤ M to + to + to = C(M ) d(go, e) .
2 2
We next consider the case in which go = (xo , yo, to ) does not belong to
the center, i.e., x2o + yo2 = 0. We now assume that x2o + yo2 ≤ ηo2 , 0 ≤ to ≤ ηo2 ,
where as before we are going to choose ηo sufficiently small. In the present
case we first move along the same polygonal path L1 (to ) ∪ L2 (to ) ∪ L3 (to )
as before, joining the identity to the point (0, 0, to). Once we reach (0, 0, to),
we move from there to the point (xo , yo , to ) along a straight-line segment
L4 (xo , yo, to ) lying on the horizontal plane H(0,0,to) . Similarly to before we
obtain
(11.9) |u(go) − u(e)| ≤ |u(xo, yo , to) − u(0, 0, to)|
√
to √
+ |u(0, 0, to) − u(− , to , to)|
√ 2 √
to √ to
+ |u(− , to , to ) − u(− , 0, 0)|
√2 2
to
+ |u(− , 0, 0) − u(0, 0, 0)|
2
= |U(0,0,to) (xo , yo) − |U(0,0,to) (0, 0)|
√
to √
+ |U(0,0,to) (0, 0) − U(0,0,to) (− , to )|
√ 2 √
to √ to
+ |U(− √to ,0,0)(− , to ) − U(− √to ,0,0)(− , 0)|
2 2 2 2
√
to
+ |U(0,0,0)(− , 0) − U(0,0,0)(0, 0)| .
2
At this point, the same argument given in the case in which go is on the
center allows to infer

|U(0,0,to) (xo , yo) − |U(0,0,to) (0, 0)| ≤ M x2o + yo2 ≤ M d(go, e) .
The remaining three terms in the right-hand side of (11.9) are estimated
as before. We conclude
|u(go) − u(e)| ≤ C(M ) d(go, e) .
This proves the continuity of u at e.
The following lemma provides a useful stability property of H-convex

functions with respect to the group convolution.
Lemma 11.7. Let K ∈ L1 (G) with compact support, and assume K ≥ 0.

If u : Hn → R is H-convex, then also K u is H-convex.
Proof. Recall that
K u(g) = K(h) u(h−1 g) dh .

Hn
Let v = K u, and fix go = (xo , yo , to) ∈ Hn . Denoting by Vgo the

restriction of v to Hgo , we want to show that
(11.10) Vgo (x + λ(x − x), y + λ(y − y)) ≤ (1 − λ) Vgo (x, y) + λ Vgo (x , y ) .
If h = (ξ, η, τ ), and if Ug indicates the restriction of u to the horizontal
plane Hg , then a computation shows that
(11.11) Vgo (x, y) = K(h) Uh−1 go (−ξ + x, −η + y) dh

Hn
= K(h) τ(−ξ,−η)Uh−1 go (x, y) dh ,

Hn
where for a function f (x, y) we have let τ(−ξ,−η) f (x, y) = f (−ξ + x, −η + y).
By the hypothesis on u, for any fixed h ∈ Hn the function Uh−1 go is a convex
function of the variables (x, y) ∈ R2n , therefore such is also the function
τ(−ξ,−η) Uh−1 go . This implies
τ(−ξ,−η) Uh−1 go (x + λ(x − x), y + λ(y − y)) ≤ (1 − λ) τ(−ξ,−η)Uh−1 go (x, y)

+ λ τ(−ξ,−η) Uh−1 go (x , y ) ,
for any 0 ≤ λ ≤ 1. Multiplying this inequality by K(h), integrating over Hn ,
and keeping (11.11) in mind, we obtain (11.10). This completes the proof.
It is now clear that in view of Proposition 11.2 and of Theorem 11.6,

all the results in this paper continue to be valid, without any additional
assumption, for functions which are H-convex. Without repeating all the
relevant statements we confine ourselves to state the fundamental L∞ − L1
estimate for H-convex functions.
Theorem 11.8. Let u : Hn → R be H-convex, then there exists C =

C(n) > 0 such that for every ball B(g, r) one has
1
(11.12) sup |u| ≤ C |u| dg ,
B(g,r) |B(g, 5r)| B(g,5r)
and
C 1
(11.13) ess sup |Xu| ≤ |u| dg .
B(g,r) r |B(g, 15r)| B(g,15r)
References.
[A] A. D. Alexandrov, Certain estimates for the Dirichlet problem, So-

viet Math. Dokl., 1 (1960), 1151-1154.
[AM] L. Ambrosio & V. Magnani, Some fine properties of BV functions

on sub-Riemannian groups, preprint, 2002.
[Ba1] I. Ja. Bakelman, Theory of quasilinear elliptic equations, Sibirsk.

Mat. Ž. (in Russian), 2 (1961), 179-186.
[Ba2] , Geometric methods for solving elliptic equations,

“Nauka”, Moscow, 1965.
[Ba3] , Convex Analysis and Nonlinear Geometric Elliptic

Equations, Springer-Verlag, 1994.
[Be] A. Bellaïche & J.-J. Risler, ed., Sub-Riemannian Geometry,

Birkhäuser, 1996.
[Bi] T. Bieske, On ∞-harmonic functions on the Heisenberg Group,

preprint, 2000.
[CC] L. A. Caffarelli, X. Cabré, Fully Nonlinear Elliptic Equations,

Amer. Math. Soc., Colloquium Publ., vol.43, 1991.
[CDG] L. Capogna, D. Danielli & N. Garofalo, Capacitary estimates and

the local behavior of solutions of nonlinear subelliptic equations,
Amer. J. Math., 118 (1996), no.6, 1153-1196.
[CGP] L. Capogna, N. Garofalo & S. Pauls, Asymptotic behavior of the

Gauss map near weakly singular characteristic points, preprint,
2002.
[Ca] C. Carathéodory, Untersuchungen über die Grundlangen der Ther-

modynamik, Math. Ann., 67 (1909), 355-386.
[Ch] W. L. Chow, Über Systeme von linearen partiellen Differentialgle-

ichungen erster Ordnung, Math. Annalen, 117 (1939), 98-105.
[CDKR] M. Cowling, A. H. Dooley, A. Korányi & F. Ricci, H-type groups

and Iwasawa decompositions, Adv. Math., 87 (1991), 1-41.
[CIL] M. Crandall, H. Ishii & P. L. Lions, User’s guide to viscosity so-

lutions of second order partial differential equations, Bull. Amer.
Math. Soc., 27 (1992), no.1, 1-67.
[Cy] J. Cygan, Subadditivity of homogeneous norms on certain nilpotent

Lie groups, Proc. Amer. Math. Soc., 83 (1981), 69-70.
[DGN1] D. Danielli, N. Garofalo & D. M. Nhieu, Non-doubling Ahlfors mea-

sures, perimeter measures, and the characterization of the trace
spaces of Sobolev functions in Carnot-Carathéodory spaces, sub-
mitted paper, 2002.
[DGN2] , Minimal surfaces, area integral, sets of constant

mean curvature and isoperimetry in Carnot groups, preprint, 2000.
[DGN3] , The Monge-Ampère equation in Carnot groups,

work in preparation.
[DGN4] , On the best possible character of the LQ norm

in some a priori estimates for non-divergence form equations in
Carnot groups, submitted paper, 2002.
[DGS] D. Danielli, N. Garofalo & S. Salsa, Variational inequalities with

lack of ellipticity. Part I: optimal interior regularity and non-
degeneracy of the free boundary, Indiana Univ. Math. J., to appear.
[Du] R. M. Dudley, On second derivatives of convex functions, Math.

Scand., 41 (1977), 159-174.
[EG] L. C. Evans & R. F. Gariepy, Measure Theory and Fine Properties

of Functions, CRC press, 1992.
[F1] G. Folland, A fundamental solution for a subelliptic operator, Bull.

Amer. Math. Soc., 79 (1973), 373-376.
[F2] G. B. Folland, Subelliptic estimates and function spaces on nilpo-

tent Lie groups, Ark. Math., 13 (1975), 161-207.
[FS] G. B. Folland & E. M. Stein, Estimates for the ∂¯b Complex and
Analysis on the Heisenberg Group, Comm. Pure Appl. Math.,
27 (1974), 429-522.
[FSS] B. Franchi, R. Serapioni & F. Serra Cassano, Approximation and

imbedding theorems for weighted Sobolev spaces associated with Lip-
schitz continuous vector fields, Boll. Un. Mat. Ital. B (7) 11 (1997),
no. 1, 83-117.
[G] N. Garofalo, Analysis and Geometry of Carnot-Carathéodory

Spaces, With Applications to Pde’s, Birkhäuser, book in prepa-
ration.
[GN1] N. Garofalo & D. M. Nhieu, Isoperimetric and Sobolev inequali-

ties for Carnot-Carathéodory spaces and the existence of minimal
surfaces, Comm. Pure Appl. Math., 49 (1996), 1081-1144.
[GN2] , Lipschitz continuity, global smooth approxima-

tions and extension theorems for Sobolev functions in Carnot-
Carathéodory spaces, J. Anal. Math., 74 (1998), 67-97.
[GP] N. Garofalo & S. D. Pauls, The Bernstein problem in the Heisenberg

group, preprint, 2002.
[GV1] N. Garofalo & D. Vassilev, Symmetry properties of positive entire

solutions of Yamabe-type equations on groups of Heisenberg type,
Duke Math. J., 106 (2001), 411-448.
[GV2] , Strong unique continuation for some subelliptic op-

erators, preprint, 1999.
[GT] D. Gilbarg & N. S. Trudinger, Elliptic Partial Differential Equa-

tions of Second Order, Second Ed., Springer, 1998.
[Gu] C. E. Gutierrez, The Monge-Ampère Equation, Birkhäuser, 2001.
[HH] J. Heinonen & I. Holopainen, Quasiregular maps on Carnot groups,

J. Geom. Anal., 7 (1997), no.1, 109-148.
[H1] H. Hörmander, Hypoelliptic second-order differential equations,

Acta Math., 119 (1967), 147-171.
[H2] , Notions of convexity, Birkhäuser, 1994.
[K] A. Kaplan, Fundamental solutions for a class of hypoelliptic PDE

generated by composition of quadratic forms, Trans. Amer. Math.
Soc., 258 (1980), 147-153.
[LMS] G. Lu, J. Manfredi & B. Stroffolini, Convex functions on the

Heisenberg group, preprint, June 2002.
[MSC] R. Monti & F. Serra Cassano, Surface measures in Carnot-

Carathéodory spaces, Calc. Var. Partial Differential Equations,
13 (2001), no. 3, 339-376.
[P] P. Pansu, Métriques de Carnot-Carathéodory et quasi-isométries

des espaces symétriques de rang un, Ann. of Math. (2)129 (1989),1,
1-60.
[Po] A. V. Pogorelov, Extrinsic Geometry of Convex Surfaces, Amer.

Math. Soc., 1973.
[Pu] C. Pucci, Limitazioni per soluzioni di equazioni ellittiche, Ann.

Mat. Pura Appl., (4) 74 (1966), 15-30.
[R] J. G. Rešetnjak, Generalized derivatives and differentiability almost

everywhere, Math. USSR - Sbornik, 4 (1968), no.3, 293-302.
[Ro] R. T. Rockafellar, Convex Analysis, Princeton Univ. Press, 1970.
[Sch] R. Schneider, Convex Bodies: The Brunn-Minkowski Theory, Cam-

bridge Univ. Press, 1993.
[S] E. M. Stein, Harmonic Analysis: Real Variable Methods, Orthogo-

nality and Oscillatory Integrals, Princeton Univ. Press, 1993.
[V] V. S. Varadarajan, Lie Groups, Lie Algebras, and Their Repre-

sentations, Springer-Verlag, New York, Berlin, Heidelberg, Tokyo,
1974.
Department of Mathematics
Purdue University
West Lafayette, IN 47907
E-mail : [email protected]
Purdue University
West Lafayette, IN 47907

Dipartimento di Metodi e Modelli Matematici per le Scienze Applicate

Università di Padova
35131 Padova, Italy
Georgetown University
Washington DC 20057-1233

Notions of Convexity in Carnot Groups

Uploaded by

Copyright:

Available Formats

Notions of Convexity in Carnot Groups

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Notions of Convexity in Carnot Groups

Uploaded by

Copyright:

Available Formats

communications in

analysis and geometry

Notions of Convexity in Carnot Groups

In 1993 motivated by some interesting questions in the theory of fully non-

for a. e. g ∈ G, and for every ζ ∈ Rm . The stratiﬁcation of the Lie algebra

in Ω. There exists a constant C = C(G, ν, Ω) > 0 such that

(1.1) sup u ≤ sup u+ + C ||F ||LQ (Ω) .

In (1.2) we have denoted by {δλ }λ>0 the non-isotropic group dilations

X 2 u(g) = (XiXj u(g))i,j=1,...,m .

This result constitutes the sub-Riemannian counterpart to the classical char-

The notion of weakly H-convex function evolved in a natural way, so

exploit in an ad hoc way the weak convexity of u along the one-parameter

(1.7) det [(X 2u) ] = F ,

will be carried in a forthcoming paper [DGN3]. We also introduce the no-

that non-isotropic gauge in a group of Heisenberg type is a fundamental gen-

Acknowledgments: It is a pleasure to acknowledge several helpful and

We begin by introducing the relevant ambients and their many properties.

j = 1, ..., r, and denote by

the topological dimension of G. The notation {Xj,1 , ..., Xj,mj }, j = 1, ..., r,

(2.1) Lgo (g) = go g , Rgo (g) = g go ,

respectively, the left- and right-translations on G by an element go ∈ G.

ξ(g) = ξ1 (g) + ... + ξr (g) ,

is the inverse of the exponential mapping. For g ∈ G, the projection of the

(2.2) xj,s (g) = < ξj (g), Xj,s >, s = 1, ..., mj.

(2.3) X = {X1 , ... , Xm } = {X1,1 , ... , X1,m },

We continue to denote by X and Y the corresponding systems of left-

Xj (g) = (Lg )∗ (Xj ) , j = 1, ..., m, Yl (g) = (Lg )∗ (Yl ) , l = 1, ..., k ,

f (g exp (tXj )) − f (g) d

A similar formula holds for any left-invariant vector ﬁeld. We indicate

and then deﬁne a pseudo-norm on G by the equation

Definition 2.1. Let Ω ⊂ G be an open set. A function u : Ω → R is said

(2.19) |u(g) − u(g )| ≤ L d(g, g ), g, g ∈ Ω.

The meaning of the symbol Γ0,1 0,1

C dR (g, g ) ≤ d(g, g ) ≤ C −1 dR (g, g )1/r .

In his foundational paper [P] Pansu proved the following remarkable

d (u(g1), u(g2)) ≤ L d(g1 , g2), g1 , g2 ∈ Ω,

then the diﬀerential of u, Du(g), exists for dg-a.e. point g ∈ Ω.

The notion of diﬀerential of a map between Carnot groups was introduced

(2.20) uλ = δλ −1 ◦ Lu(g)−1 ◦ u ◦ Lg ◦ δλ

converge locally uniformly to a homomorphism Du(g) : G → G , as λ → 0.

u(gδλ (h)) − u(g)

the limit being assumed locally uniform in h ∈ G. In the next proposition we

Proposition 2.3. Let G be a Carnot group of step r. For u ∈ Γ1 (G) the

3. Strongly convex functions: A notion to abandon.

Definition 3.1. Given two points g, g ∈ G, for λ ∈ [0, 1] we will denote

This is easily recognized as follows

Definition 3.2. A function u : G → (−∞, ∞] is called strongly H-convex

We note right-away that in the abelian case G = Rn , Deﬁnition 3.2

Proposition 3.3. For every h ∈ G the left-translated Lh u(g) = u(hg) of a

Proof. For every g, g ∈ G, and any 0 ≤ λ ≤ 1 one has

Proposition 3.4. Let {uα }α∈A be an arbitrary family of strongly H-convex

Proposition 3.5. Let h : R → R be convex and non-decreasing, then for

Proposition 3.7. Every H-aﬃne function is strongly H-convex.

Proof. Suppose that A is of the form (3.3), then

A(g δλ (g −1 g )) = < p, ξ1 (g δλ (g −1 g )) > + β .

We now observe that the formula of Baker-Campbell-Hausdorﬀ (2.7)

(3.4) ξ1 (g δλ (g −1 g )) = ξ1 (g) + λ (ξ1 (g ) − ξ1 (g)) ,

A(g δλ (g −1 g )) = < p, ξ1 (g) + λ (ξ1 (g ) − ξ1 (g)) > + β

Proposition 3.8. Let φ : Rm → (−∞, ∞], then the function u : G →

Proof. From (3.4) we obtain

(3.5) u v(g) = u(h) v(h−1 g) dh = u(gh−1 ) v(h) dh .

K u(g δλ (g −1 g )) = K(h) u(h−1 g δλ (g −1 g ) dh

(5.7) φ (λ) = < (X 2 u (gλ )) ζ , ζ > .

Letting → 0 we reach the conclusion.