MEMOIRS of the

American Mathematical Society

Number 688

Sobolev Met Poincare

Piotr Hajlasz
Pekka Koskela

May 2000 Volume 145 Number 688 (first of 4 numbers) ISSN 0065-9266

American Mathematical Society

Providence. Rhode Island
Sobolev met Poincaré
Piotr HajÃlasz and Pekka Koskela
Contents

Abstract ix

1 Introduction 1

2 What are Poincaré and Sobolev inequalities? 7

3 Poincaré inequalities, pointwise estimates, and Sobolev classes 11

4 Examples and necessary conditions 16

5 Sobolev type inequalities by means of Riesz potentials 21

6 Trudinger inequality 29

7 A version of the Sobolev embedding theorem on spheres 33

8 Rellich-Kondrachov 36

9 Sobolev classes in John domains 38

9.1 John domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
9.2 Sobolev type inequalities . . . . . . . . . . . . . . . . . . . . . . . . . 43

10 Poincaré inequality: examples 46

10.1 Riemannian manifolds. . . . . . . . . . . . . . . . . . . . . . . . . . . 46
10.2 Upper gradients. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
10.3 Topological manifolds. . . . . . . . . . . . . . . . . . . . . . . . . . . 49
10.4 Gluing and related constructions. . . . . . . . . . . . . . . . . . . . . 50
10.5 Further examples. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

11 Carnot–Carathéodory spaces 51
11.1 Carnot–Carathéodory metric. . . . . . . . . . . . . . . . . . . . . . . 51
11.2 Upper gradients and Sobolev spaces. . . . . . . . . . . . . . . . . . . 55
11.3 Carnot groups. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
11.4 Hörmander condition. . . . . . . . . . . . . . . . . . . . . . . . . . . 69
11.5 Further generalizations . . . . . . . . . . . . . . . . . . . . . . . . . . 71

12 Graphs 71

vii
viii CONTENTS

13 Applications to P.D.E and nonlinear potential theory 77

13.1 Admissible weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
13.2 Sobolev embedding for 0 < p < 1 . . . . . . . . . . . . . . . . . . . . 81

14 Appendix 83
14.1 Measures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
14.2 Uniform integrability. . . . . . . . . . . . . . . . . . . . . . . . . . . 85
p
14.3 L and Lpw spaces. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
14.4 Covering lemma. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
14.5 Maximal function. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
14.6 Lebesgue differentiation theorem. . . . . . . . . . . . . . . . . . . . . 88

References 89
Abstract

There are several generalizations of the classical theory of Sobolev spaces as they
are necessary for the applications to Carnot–Carathéodory spaces, subelliptic equa-
tions, quasiconformal mappings on Carnot groups and more general Loewner spaces,
analysis on topological manifolds, potential theory on infinite graphs, analysis on
fractals and the theory of Dirichlet forms.
The aim of this paper is to present a unified approach to the theory of Sobolev
spaces that covers applications to many of those areas. The variety of different
areas of applications forces a very general setting.
We are given a metric space X equipped with a doubling measure µ. A general-
ization of a Sobolev function and its gradient is a pair u ∈ L1loc (X), 0 ≤ g ∈ Lp (X)
such that for every ball B ⊂ X the Poincaré-type inequality
Z µZ ¶1/p
p
|u − uB | dµ ≤ Cr g dµ
B σB

holds, where r is the radius of B and σ ≥ 1, C > 0 are fixed constants. Working in
the above setting we show that basically all relevant results from the classical theory
have their counterparts in our general setting. These include Sobolev-Poincaré type
embeddings, Rellich-Kondrachov compact embedding theorem, and even a version
of the Sobolev embedding theorem on spheres. The second part of the paper is
devoted to examples and applications in the above mentioned areas.

———————————
This research was begun while P.H. was visiting the Universities of Helsinki and Jyväskylä
in 1994, continued during his stay in the ICTP in Trieste in 1995 and finished in Max-Planck
Institute (MIS) in Leipzig in 1998. He wishes to thank all the institutes for their hospitality. P.H.
was partially supported by KBN grant no. 2–PO3A–055–14, P.K. by the Academy of Finland
grant number 34082 and by the NSF.
2000 Mathematical Subject Classification: Primary 46E35. Secondary 22E30, 26D10, 28Axx,
30C65, 31C20, 31C99, 35D10, 35H20, 35J60, 35J70, 42B25, 43A85, 46E39, 53C17, 53C21.
Key words and phrases: Sobolev spaces, Poincaré inequality, embedding theorems, metric
spaces, doubling measures, spaces of homogeneous type, upper gradients, nonlinear potential the-
ory, nonlinear degenerate elliptic equations, Hörmander’s condition, Heisenberg’s group, Carnot–
Carathéodory metric, nonlinear subelliptic equations, John domains, analysis on graphs.

ix
1 Introduction

The theory of Sobolev spaces is a central analytic tool in the study of various aspects
of partial differential equations and calculus of variations. However, the scope of its
applications is much wider, including questions in differential geometry, algebraic
topology, complex analysis, and in probability theory.
Let us recall a definition of the Sobolev spaces. Let u ∈ Lp (Ω), where Ω is an
open subset of IRn , and 1 ≤ p ≤ ∞. We say that u belongs to the Sobolev space
W 1,p (Ω) if the distributional derivatives of the first order belong to Lp (Ω). This
definition easily extends to the setting of Riemannian manifolds, as the gradient is
well defined there.
The fundamental results in the theory of Sobolev spaces are the so-called Sobolev
embedding theorem and the Rellich–Kondrachov compact embedding theorem. The
∗
first theorem states that, for 1 ≤ p < n, W 1,p (Ω) ⊂ Lp (Ω), where p∗ = np/(n − p),
provided the boundary of Ω is sufficiently nice. The second theorem states that, for
every q < p∗ , the embedding W 1,p (Ω) ⊂ Lq (Ω) is compact for such a domain Ω.
Since its introduction the theory and applications of Sobolev spaces have been
under intensive study. Recently there have been attempts to generalize Sobolev
spaces to the setting of metric spaces equipped with a measure. Let us indicate
some of the problems that suggest such a generalization.
1) Study of the Carnot–Carathéodory metric generated by a family of vector
fields. 2) Theory of quasiconformal mappings on Carnot groups and more general
Loewner spaces. 3) Analysis on topological manifolds. 4) Potential theory on
infinite graphs. 5) Analysis on fractals.
Let us briefly discuss the above examples. The Carnot–Carathéodory metric
appears in the study of hypoelliptic operators, see Hörmander [128], Fefferman and
Phong [69], Jerison [133], Nagel, Stein and Wainger [204], Rotschild and Stein [218],
Sánchez-Calle [224].
The Sobolev inequality on balls in Carnot–Carathéodory metric plays a cru-
cial role in the so-called Moser iteration technique, [202], used to obtain Harnack
inequalities and Hölder continuity for solutions of various quasilinear degenerate
equations. The proof of the Harnack inequality by means of the Moser technique
can be reduced to verifying a suitable Sobolev inequality. Conversely, a parabolic
Harnack inequality implies a version of the Sobolev inequality as shown by Saloff-
Coste, [221]. It seems that the first to use the Moser technique in the setting of the
Carnot–Carathéodory metric were Franchi and Lanconelli, [78]. The later work on
related questions include the papers by Biroli and Mosco, [8], [9], Buckley, Koskela
and Lu, [19], Capogna, Danielli and Garofalo, [27], [28], [29], [30], [31], Chernikov
and Vodop’yanov, [38], Danielli, Garofalo, Nhieu, [61], Franchi, [74], Franchi, Gal-
lot and Wheeden, [75], Franchi, Gutiérrez and Wheeden, [76], Franchi and Lanco-
———————————
Received by the editor October 1, 1998.

1
2 PIOTR HAJLASZ AND PEKKA KOSKELA

nelli, [79], Franchi, Lu and Wheeden, [80], [81], Franchi and Serapioni, [83], Garofalo
and Lanconelli, [90], Lu, [175], [178], [180], [181], Marchi [189].
The theory of Carnot–Carathéodory metrics and related Sobolev inequalities can
be extended to the setting of Dirichlet forms, see Biroli and Mosco [8], Garattini
[89], Sturm [239].
For connections to the theory of harmonic maps see the papers by Jost, [142],
[143], [144], [145], Jost and Xu, [146], HajÃlasz and Strzelecki, [108].
The theory of quasiconformal mappings on Carnot Groups has been studied by
Heinonen and Holopainen, [116], Margulis and Mostow, [190], Pansu, [210], Koranyi
and Reimann, [159], Heinonen and Koskela, [117], Vodop’yanov and Greshnov,
[253]. Results on Sobolev spaces play an important role in this theory. Very recently
Heinonen and Koskela, [118], extended the theory to the setting of metric spaces
that support a type of a Sobolev inequality.
Semmes, [226], has shown that a large class of topological manifolds admit
Sobolev type inequalities, see Section 10. Sobolev type inequalities on a Riemannian
manifold are of fundamental importance for heat kernel estimates, see the survey
article [55] of Coulhon for a nice exposition.
Discretization of manifolds has lead one to define the gradient on an infinite
graph using finite differences and then to investigate the related Sobolev inequalities,
see Kanai [149], Auscher and Coulhon [2], Coulhon [52], Coulhon and Grigor’yan
[56], Coulhon and Saloff-Coste [59], Delmotte [65], Holopainen and Soardi [126],
[127]. These results have applications to the classification of Riemannian mani-
folds. Also the study of the geometry of finitely generated groups leads to Sobolev
inequalities on associated Cayley graphs, see Varopoulos, Saloff-Coste and Coulhon
[251], and Section 12 for references.
At last, but not least, the Brownian motion on fractals leads to an associated
Laplace operator and Sobolev type functions on fractals, see Barlow and Bass [5],
Jonsson [139], Jonsson and Wallin [141], Kozlov [162], Kigami [151], [152], [153],
Kigami and Lapidus [154], Lapidus [167], [168], Metz and Sturm [197], Mosco [201].
How does one then generalize the notion of Sobolev space to the setting of a
metric space? There are several possible approaches that we briefly describe below.
In general, the concept of a partial derivative is meaningless on a metric space.
However, it is natural to call a measurable function g ≥ 0 an upper gradient of a
function u if
Z
|u(x) − u(y)| ≤ g ds
γ

holds for each pair x, y and all rectifiable curves γ joining x, y. Thus, in the Eu-
clidean setting, we consider the length of the gradient of a smooth function instead
of the actual gradient. The above definition is due to Heinonen and Koskela, [118].
Assume that the metric space is equipped with a measure µ. Then we can ask
if for every pair u, g, where u is continuous and g is an upper gradient of u, the
 SOBOLEV MET POINCARÉ 3

weak version
µZ ¶1/q µZ ¶1/p
q p
|u − uB | dµ ≤ Cr g dµ (1)
B σB

of the Sobolev-Poincaré inequality holds with q > p ≥ 1 whenever B is a ball of

radius r. Here C and σ ≥ 1 are fixed constants, barred integrals over a set A mean
integral averages, and uB is the average value of u over B.
If q = 1, then we call (1) a p-Poincaré inequality. It turns out that a p-Poincaré
inequality implies a Sobolev-Poincaré inequality, see Section 5.
This approach is however limited to metric spaces that are sufficiently regu-
lar. There have to be sufficiently many rectifiable curves, which excludes fractals
and graphs. For more information see Section 4, Section 10.2, Section 11.2, Bour-
don and Pajot [15], Cheeger [34], Franchi, HajÃlasz and Koskela [77], Hanson and
Heinonen [111], Heinonen and Koskela [118], [119], Kallunki and Shanmugalingam
[148], Laakso [164], Semmes [226], Shanmugalingam [228], Tyson [245].
Recently HajÃlasz, [102], introduced a notion of a Sobolev space in the setting of
an arbitrary metric space equipped with a Borel measure that we next describe.
One can prove that u ∈ W 1,p (Ω), 1 < p ≤ ∞, where Ω ⊂ IRn is a bounded set
with nice boundary if and only if u ∈ Lp (Ω) and there is a non-negative function
g ∈ Lp (Ω) such that

|u(x) − u(y)| ≤ |x − y|(g(x) + g(y)). (2)

Since this characterization does not involve the notion of a derivative it can be
used to define Sobolev space on an arbitrary metric space, see HajÃlasz [102]. These
spaces have been investigated or employed in Franchi, HajÃlasz and Koskela [77],
Franchi, Lu and Wheeden [81], HajÃlasz [103], HajÃlasz and Kinnunen [104], HajÃlasz
and Martio [107], Heinonen [115], Heinonen and Koskela [118], KaÃlamajska [147],
Kilpeläinen, Kinnunen and Martio [156], Kinnunen and Martio [157], Koskela and
MacManus [161], Shanmugalingam [228].
Another approach is presented in the paper [105] of HajÃlasz and Koskela, in
which also some of the results from our current work were announced. Given a
metric space equipped with a Borel measure we assume that a pair u and g, (g ≥ 0),
of locally integrable functions satisfies the family (1) of Poincaré inequalities with
q = 1 and a fixed p ≥ 1 on every ball, that is
Z µZ ¶1/p
|u − uB | dµ ≤ CP r g p dµ . (3)
B σB

This family of inequalities is the only relationship between u and g. Then we can
ask for the properties of u that follow.
Yet another approach to Sobolev inequalities on metric spaces is presented in
the paper [11] by Bobkov and Houdré. However it is much different from the above
mentioned setting and it will not be discussed here.
4 PIOTR HAJLASZ AND PEKKA KOSKELA

One of the purposes of this paper is to systematically develop the theory of

Sobolev spaces from inequality (3). This includes the study of the relationships
between (1), (2), and (3). We show that basically all relevant results from the clas-
sical theory have their counterparts in our general setting. These include Sobolev-
Poincaré type embeddings, Rellich-Kondrachov compact embedding theorem, and
even a version of the Sobolev embedding theorem on spheres.
We will work with metric spaces equipped with a doubling measure. Such spaces
are often called spaces of homogeneous type, but we will call them doubling spaces.
The reader may find many important examples of spaces of homogeneous type in
Christ [41], and Stein [234]. The class of such spaces is pretty large. For example
Volberg and Konyagin [255], [256], proved that every compact subset of IRn supports
a doubling measure; see also Wu [257], and Luukkainen and Saksman [182].
Starting from the work of Coifman and Weiss [48], [49], spaces of homogeneous
type have become a standard setting for the harmonic analysis related to singu-
lar integrals and Hardy spaces, see, e.g., Gatto and Vagi [93], [94], Genebahsvili,
Gogatishvili, Kokilashvili and Krbec [95], Han [109], Han and Sawyer [110], Macias
and Segovia [183].
However it seems that the development of the theory of Sobolev spaces in such
generality did not begin until very recently.
There are some papers on Sobolev inequalities on spaces of homogeneous type
related to our work; see Franchi, Lu and Wheeden [81], Franchi, Pérez and Wheeden
[82], MacManus and Pérez [184], [185]; the last three papers were motivated by our
approach. Also the paper [92] of Garofalo and Nhieu provides a similar approach
in the special case of Carnot–Carathéodory spaces.
The reader might wonder why we insist on studying the situation with a fixed
exponent p instead of assuming that (3) holds with p = 1. There is a simple reason
for this. Indeed, for each p > 1 one can construct examples of situations where (3)
holds for each smooth function u with g = |∇u| but where one cannot replace p by
any exponent q < p. Let us give an example to illustrate the dependence on p. Take
two three-dimensional planes in IR5 whose intersection is a line L, and let X be the
union of these two planes. The metrics and measures induced from the planes have
natural extensions to a metric and a measure on X. If u is a smooth function on X
then we define g(x) to be |∇u(x)| whenever x does not belong to L, where ∇u is
the usual gradient of u in the appropriate plane, and define g(x) to be the sum of
the lengths of the two gradients corresponding to the different planes when x ∈ L.
One can then check that (3) holds for p > 2 but fails for p ≤ 2.
As we said, we want to develop the theory of Sobolev spaces assuming a fam-
ily of Poincaré inequalities (3) and the doubling property. Such assumptions have
found many applications in the literature in various areas of analysis and geom-
etry. The applications include the above mentioned Carnot–Carathéodory spaces
[61], [81], [91], [92], graphs [64], [65], [127], Dirichlet forms [8], [9], [10], [239], and
quasiconformal mappings [118].
 SOBOLEV MET POINCARÉ 5

These assumptions have also found many important applications in Riemannian

geometry. The class of open Riemannian manifolds that satisfy both the doubling
condition and the p-Poincaré inequality is under intensive investigation, see Colding
and Minicozzi [50], [51], Grigor’yan [99], Holopainen [124], Holopainen and Rickman
[125], Li and Wang [171], Maheux and Saloff-Coste [186], Rigoli, Salvatori and
Vignati [215], Saloff-Coste [221], [222], [223], and Tam [242], where some global
properties of manifolds were obtained under the assumption that the Riemannian
manifold satisfies a p-Poincaré inequality and the doubling property.
In 1975 Yau, [261], proved that on open Riemannian manifolds of nonnegative
Ricci curvature bounded harmonic functions are constant. Some time later he
conjectured that for such manifolds the space of harmonic functions with polynomial
growth of fixed rate is finite dimensional.
Independently Grigor’yan, [99], and Saloff-Coste, [221], generalized Yau’s theo-
rem by proving that bounded harmonic functions are constant provided the manifold
satisfies the doubling property and the Poincaré inequality (3) with g = |∇u| and
p = 2. It is known that manifolds with nonnegative Ricci curvature satisfy these
two conditions, see Section 10.1. Under the same assumptions the result of Yau has
been extended to harmonic mappings, see Li and Wang, [171], and Tam [242].
Very recently, Colding and Minicozzi, [50], [51], answered the conjecture of Yau
in the affirmative. Again the assumptions were that the manifold is doubling and
that the 2-Poincaré inequality holds.
Many of the above Riemannian results have counterparts in the more general
settings of Carnot–Carathéodory spaces, graphs, or Dirichlet forms, and again the
main common assumption is the same: doubling and Poincaré.
This common feature was guiding us in our work. The first part of the paper is
devoted to general theory and the second part to examples and applications in the
areas mentioned above.
The paper is organized as follows. In Section 2 we present the setting in which
we later on develop the theory of Sobolev inequalities. In Section 3 we discuss the
equivalence of various approaches to Sobolev inequalities on metric spaces. Section 4
is devoted to some basic examples and conditions that necessarily hold for spaces
that satisfy all p-Poincaré inequalities (1) for pairs of a continuous function and
upper gradient. In Section 5 we show that if a pair u, g satisfies a p-Poincaré
inequality (3), then |u − uB | can be estimated by a generalized Riesz potential.
This together with a generalization of the Fractional Integration Theorem implies
a variant of the Sobolev-Poincaré embedding theorem. In Section 6 we impose
the additional condition that the space be connected and improve on one of the
inequalities from Section 5, namely we prove a variant of the Trudinger inequality.
In Section 7 we prove an embedding theorem on almost all spheres centered at a
given point. In Section 8 we generalize the classical Rellich-Kondrachov theorem to
the setting of metric spaces. So far all the results are local in nature. In Section 9 we
introduce the class of John domains and generalize previous results as global results
in John domains. In Section 10 we collect important examples of metric spaces
6 PIOTR HAJLASZ AND PEKKA KOSKELA

where the theory developed in the paper is applicable (including open Riemannian
manifolds, topological manifolds and Loewner spaces). In Section 11 we study
the theory of Carnot–Carathéodory spaces that are associated with a family of
vector fields, from the point of view of Sobolev inequalities on metric spaces. In
Section 12 we discuss Sobolev inequalities on infinite graphs. Section 13 is devoted
to applications of the theory to nonlinear potential theory and degenerate elliptic
equations. Section 14 is an appendix, where we collect all the results in measure
theory and on maximal functions that are needed in the paper.
The exposition is self-contained and the background material needed is the ab-
stract measure theory in metric spaces, some real analysis related to maximal func-
tions and the basic theory of classical Sobolev spaces covered by each of the following
references: Evans and Gariepy [66], Gilbarg and Trudinger [96], Malý and Ziemer
[187], Ziemer [263].
Some examples and applications that illustrate the theory require slightly more.
In Section 11.3-4 some familiarity with Lie groups and commutators of vector fields
is needed and in Section 13 we assume basic facts about quasilinear elliptic equations
in divergence form. One can, however, skip reading Sections 11.3-4 and 13 and it
will not affect understanding of the remaining parts of the paper.
We did make some effort to give comprehensive references to subjects related
to our work. We are however sure that many important references are still missing
and we want to apologize to those whose contribution is not mentioned.
Notation. Throughout the paper X will be a metric space with a metric d, and
a Borel measure µ. The precise assumptions on µ are collected in the appendix. If
not otherwise stated, µ will be doubling which means that

µ(2B) ≤ Cd µ(B) (4)

whenever B is a ball and 2B is the ball with the same center as B and with radius
twice that of B (in the same way we define σB for σ > 0). We will call such a metric
measure space X a doubling space and Cd a doubling constant. Ω ⊆ X will always
denote an open subset. Sometimes we will need the doubling property on a subset
of X only; we will say that the measure µ is doubling on Ω if (4) holds whenever
B = B(x, r), x ∈ Ω and r ≤ 5 diam Ω. By writing v ∈ Lqloc (Ω), we designate that
v belongs to the class Lq (B) with respect to µ for each ball B ⊂ Ω. If Ω = X, we
will simply write v ∈ Lqloc . By Lip (X) we denote the class of Lipschitz functions
on the metric space X.
R R
The average value will be denoted by vA = A v dµ = µ(A)−1 A v dµ. If R > 0
and v is a measurable function, then MR v stands for the restricted Hardy-Littlewood
maximal function
Z
MR v(x) = sup |v| dµ.
0<r≤R B(x,r)

If R = ∞, then we will simply write M v. Another version of the maximal function

 SOBOLEV MET POINCARÉ 7

is Z
1
MΩ v(x) = sup |v| dµ,
r>0 µ(B(x, r)) Ω∩B(x,r)

which applies to v ∈ L1loc (Ω, µ). It is also clear how to define the restricted maximal
function MΩ,R v.
By H k we denote the k-dimensional Hausdorff measure. The symbol χE denotes
the characteristic function of a set E. We reserve B to always denote a ball. Observe
that in some metric spaces it may happen that the center and the radius of the ball
are not uniquely defined. In what follows, when we write B we assume that the
center and the radius are fixed. Otherwise σB is not properly defined. By C we
will denote a general constant which can change even in a single string of estimates.
By writing C = C(p, q, λ) we indicate that the constant C depends on p, q and λ
only. We write u ≈ v to state that there exist two positive constants C1 , and C2
such that C1 u ≤ v ≤ C2 u.
Some further notation and commonly used results are collected in the appendix.

2 What are Poincaré and Sobolev inequalities?

In this section we describe the general framework and give samples of problems
which are treated later on. Until the end of the section we assume that µ is a Borel
measure on a metric space X, but we do not assume that µ is doubling. As before
Ω ⊂ X denotes an open set.
Definition. Assume that u ∈ L1loc (Ω) and a measurable function g ≥ 0 satisfy the
inequality
Z µZ ¶1/p
p
|u − uB | dµ ≤ CP r g dµ , (5)
B σB

on each ball B with σB ⊂ Ω, where r is the radius of B and p > 0, σ ≥ 1, CP > 0 are
fixed constants. We then say that the pair u, g satisfies a p-Poincaré inequality in
Ω. When Ω = X, we simply say that the pair u, g satisfies a p-Poincaré inequality.
Note that if u ∈ Lip (IRn ), g = |∇u| and p ≥ 1, then (5) is a corollary of the
classical Poincaré inequality
µZ ¶1/p µZ ¶1/p
|u − uB |p dx ≤ C(n, p)r |∇u|p dx . (6)
B B

Quite often we will call an inequality weak if both sides involve a ball and the radius
of the ball on the right hand side is greater than the radius of the ball on the left
hand side, like in (5).
Unfortunately, as is easy to see, in general, inequality (6) does not hold with
p < 1 (cf. [16, p. 224]). Nevertheless, there are many important situations where
the p-Poincaré inequalities (5) and (6) hold with p < 1. For example, they hold
8 PIOTR HAJLASZ AND PEKKA KOSKELA

when u is a solution to an elliptic equation of a certain type, see Section 13. For
this reason we include the case p < 1.
It is natural to regard a pair u, g that satisfies a p-Poincaré inequality in Ω as
a Sobolev function and its gradient. In this sense we will develop the theory of
Sobolev functions on metric spaces with “gradient” in Lp for all p > 0.
In the classical approach the Sobolev spaces are defined for p ≥ 1 only. Moreover,
it was expected that there would be no reasonable theory of Sobolev spaces for
0 < p < 1, see Peetre, [211]. We obtain a rich theory of Sobolev spaces for all p > 0.
In the Euclidean setting, when p ≥ 1, our approach is equivalent to the classical
one.
In the literature there are a few papers that deal with the Sobolev inequalities
for p < 1, see Bakry, Coulhon, Ledoux and Saloff-Coste [4], Buckley and Koskela
[16], Buckley, Koskela and Lu [19], Calderón and Scott [24], HajÃlasz and Koskela
[105].
Let us assume that a pair u, g satisfies a p-Poincaré inequality for p > 0 in an
open set Ω ⊂ X. We inquire for properties of u that follow from this assumption.
A typical question is whether the Sobolev embedding theorem holds i.e., whether
the p-Poincaré inequality in Ω implies the global Sobolev-type inequality
µZ ¶1/q µZ ¶1/p
q p
inf |u − c| dµ ≤C g dµ , (7)
c∈IR Ω Ω

with an exponent q > p. We suggest the reader to have a look at our earlier paper
[105], where a result of this type was obtained by an elementary method. In the
current paper we obtain stronger results by more complicated methods.
Note that if µ(Ω) < ∞ and q ≥ 1, then the above inequality is equivalent to
µZ ¶1/q µZ ¶1/p
q 0 p
|u − uΩ | dµ ≤C g dµ , (8)
Ω Ω

as for q ≥ 1 and µ(Ω) < ∞ we have

µZ ¶1/q µZ ¶1/q
q q
inf |u − c| dµ ≤ |u − uΩ | dµ
c∈IR Ω Ω
µZ ¶1/q
q
≤ 2 inf |u − c| dµ . (9)
c∈IR Ω

The classical gradient of a Lipschitz function has a very important property: if the
function is constant in a set E, then the gradient equals zero a.e. in E. To have
a counterpart of this property in the metric setting we introduce the truncation
property.
Given a function v and ∞ > t2 > t1 > 0, we set

vtt12 = min{max{0, v − t1 }, t2 − t1 }.
 SOBOLEV MET POINCARÉ 9

Definition. Let the pair u, g satisfy a p-Poincaré inequality in Ω. Assume that for
every b ∈ IR, ∞ > t2 > t1 > 0, and ε ∈ {−1, 1}, the pair vtt12 , gχ{t1 <v≤t2 } , where
v = ε(u − b), satisfies the p-Poincaré inequality in Ω (with fixed constants CP , σ).
Then we say that the pair u, g has the truncation property.
Let p ≥ 1 and u ∈ Lip (IRn ). Since v = ±(u−b) satisfies |∇vtt12 | = |∇u|χ{t1 <v≤t2 }
a.e., the pair u, |∇u| has the truncation property. More sophisticated examples are
given in Section 10.
We close the section with a result which shows that inequality (7) is equivalent
to a weaker inequality provided the pair u, g has the truncation property. The
result will be used in the sequel.

Theorem 2.1 Let Ω ⊂ X be an open set with µ(Ω) < ∞. Fix ∞ > q ≥ p > 0,
CP > 0 and σ ≥ 1. Assume that every pair u, g, that satisfies a p-Poincaré
inequality in Ω (with given CP and σ) satisfies also the global Marcinkiewicz–Sobolev
inequality
µZ ¶q/p
inf sup µ({x ∈ Ω : |u(x) − c| > t})tq ≤ C1 g p dµ . (10)
c∈IR t≥0 Ω

Then every pair that satisfies the p-Poincaré inequality in Ω (with given CP and σ)
and has the truncation property satisfies also the global Sobolev inequality
µZ ¶1/q µZ ¶1/p
q p
inf |u − c| dµ ≤ C2 g dµ (11)
c∈IR Ω Ω

with C2 = 8 · (4C1 )1/q .

Remarks. 1) We call (10) a Marcinkiewicz–Sobolev inequality, because it implies

that u belongs to the Marcinkiewicz space Lqw .
2) The result is surprising even in the Euclidean case: inequality (10) seems much
weaker than (11) as the inclusion Lq ⊂ Lqw is proper. Similar phenomena have been
discovered by V. G. Maz’ya, [194], (cf. [195, Section 2.3.1], [106, Theorem 1]), who
proved that a Sobolev embedding is equivalent to a capacitary estimate which is
a version of inequality (10). The main idea of Maz’ya was a truncation method
which is also the key argument in our proof. This method mimics the proof of the
equivalence of the Sobolev inequality with the isoperimetric inequality. Inequal-
ity (10) plays the role of the relative isoperimetric inequality and the truncation
argument provides a discrete counterpart of the co-area formula. The truncation
method of Maz’ya has become very useful in proving various versions of the Sobolev
embedding theorem with sharp exponents in the borderline case where interpolation
arguments do not work. To see how the argument works in the case of the classical
Sobolev embedding theorem, we refer the reader to the comments after the state-
ment of Theorem 5.1. Recently the truncation method has been employed and even
rediscovered by many authors; see Adams and Hedberg [1, Theorem 7.2.1], Bakry,
Coulhon, Ledoux and Saloff-Coste [4], Biroli and Mosco [8], [9], Capogna, Danielli
10 PIOTR HAJLASZ AND PEKKA KOSKELA

and Garofalo [29], Coulhon [54], Franchi, Gallot and Wheeden [76], Garofalo and
Nhieu [92], Heinonen and Koskela [118], [119], Long and Nie [173], Maheux and
Saloff-Coste [186], Semmes [226], and Tartar [243].
Proof of Theorem 2.1. Let u, g be a pair that satisfies the p-Poincaré inequality
in Ω and that has the truncation property. Choose b ∈ IR such that

µ(Ω) µ(Ω)
µ({u ≥ b}) ≥ and µ({u ≤ b}) ≥ .
2 2
Let v+ = max{u−b, 0}, v− = − min{u−b, 0}. We will estimate kv+ kLq and kv− kLq
separately. In what follows v will denote either v+ or v− .

Lemma 2.2 Let ν be a finite measure on a set Y . If w ≥ 0 is a ν-measurable

function such that ν({w = 0}) ≥ ν(Y )/2, then for every t > 0

t
ν({w > t}) ≤ 2 inf ν({|w − c| > }).
c∈IR 2

The proof of the lemma is easy and left to the reader.

By the truncation property the pair vtt12 , gχ{t1 <v≤t2 } satisfies the p-Poincaré
inequality and hence it satisfies (10). Moreover, the function vtt12 has the property
µ({vtt12 = 0}) ≥ µ(Ω)/2. Hence, applying the lemma, we conclude that
µ½ ¾¶ µ ¶q
t t
sup µ({vtt12 > t})tq ≤ 2q+1 inf sup µ |vtt12 − c| >
t≥0 c∈IR t≥0 2 2
≤ 2 q+1
C1 kgχ{t1 <v≤t2 } kqLp .

This yields
Z ∞
X
v q dµ ≤ 2kq µ({2k−1 < v ≤ 2k })
Ω k=−∞
X∞
≤ 2kq µ({v ≥ 2k−1 })
k=−∞
X∞
k−1
= 2kq µ({v22k−2 ≥ 2k−2 })
k=−∞
∞ µZ
X ¶q/p
≤ 23q+1 C1 g p χ{2k−2 <v≤2k−1 } dµ
k=−∞ Ω
à ∞ Z
!q/p
X
≤ 23q+1 C1 g p χ{2k−2 <v≤2k−1 } dµ
k=−∞ Ω

≤ 23q+1 C1 kgkqLp (Ω) .

 SOBOLEV MET POINCARÉ 11

In the second to the last step we used the inequality q/p ≥ 1. Finally
Z Z Z
q q
|u − b|q = v+ + v− ≤ 23q+2 C1 kgkqLp (Ω) .
Ω Ω Ω

This completes the proof.

The following theorem is a modification of the above result.

Theorem 2.3 Let ∞ > q ≥ p > 0, CP > 0 and σ ≥ 1. Assume that every pair
u, g that satisfies the p-Poincaré inequality (with given CP and σ) satisfies also the
weak Marcinkiewicz–Sobolev inequality
µZ ¶q/p
µ({x ∈ B : |u(x) − c| > t})tq
inf sup ≤ C1 r q g p dµ
c∈IR t≥0 µ(B) σB

for every ball B, where r denotes the radius of B. Then every pair u, g that satisfies
the p-Poincaré inequality (with given CP and σ) and has the truncation property
satisfies also the weak Sobolev inequality
µZ ¶1/q µZ ¶1/p
q p
inf |u − c| dµ ≤ C2 r g dµ
c∈IR B σB

for every ball B with C2 = 8 · (4C1 )1/q .

The proof is essentially the same as that for Theorem 2.1 and we leave it to the
reader.

3 Poincaré inequalities, pointwise estimates, and

Sobolev classes
Our starting point to the theory of Sobolev spaces on metric spaces is to assume
that the pair u, g satisfies a p-Poincaré inequality. There are however also other
possible approaches. Recently HajÃlasz, [102], introduced a notion of a Sobolev space
in the setting of metric space equipped with a Borel measure. In this section we will
compare this approach to that based on Poincaré inequalities (see Theorem 3.1).
The proof is based on pointwise inequalities which are of independent interest and
which we state in a more general version than is needed for the sake of the proof
(see Theorem 3.2 and Theorem 3.3). Finally we compare the class of Lp -pairs of
u, g that satisfy a p-Poincaré inequality to the classical Sobolev space.
For a detailed study on the equivalence of various approaches to Sobolev in-
equalities on metric spaces, see Franchi, HajÃlasz and Koskela [77], and Koskela and
MacManus [161]. Results related to those of this section appear also in Franchi,
Lu and Wheeden [81], HajÃlasz and Kinnunen [104], Heinonen and Koskela [118],
Shanmugalingam [228].
12 PIOTR HAJLASZ AND PEKKA KOSKELA

Given p > 0 and a triple (X, d, µ), where (X, d) is a metric space and µ is a
Borel measure (not necessarily doubling), HajÃlasz, [102], defines the Sobolev space
M 1,p (X, d, µ) as the set of all u ∈ Lp (X) for which there exists 0 ≤ g ∈ Lp (X) such
that
|u(x) − u(y)| ≤ d(x, y)(g(x) + g(y)) a.e. (12)
When we say that an inequality like (12) holds a.e. we mean that there exists a set
E ⊂ X with µ(E) = 0 such that inequality (12) holds for all x, y ∈ X \ E.
If p ≥ 1, then the space is equipped with a Banach norm kukM 1,p = kukLp +
inf g kgkLp , where the infimum is taken over the set of all 0 ≤ g ∈ Lp (X) that satisfy
(12).
The motivation for the above definition comes from the following result.
If Ω = IRn or if Ω ⊂ IRn is a bounded domain with sufficiently regular boundary,
| · | is the Euclidean metric, H n the Lebesgue measure, and 1 < p ≤ ∞, then

W 1,p (Ω) = M 1,p (Ω, | · |, H n ) (13)

as sets and the norms are equivalent, see [102] and also [103], [107], [252]. Here
W 1,p (Ω) denotes the classical Sobolev space of Lp -integrable functions with gener-
alized gradient in Lp . If p = 1, then the equivalence (13) fails, see [103]. However,
for any open set Ω ⊂ IRn and 1 ≤ p < ∞, M 1,p (Ω, | · |, H n ) ⊂ W 1,p (Ω), see [103,
Proposition 1], and also [107, Lemma 6].
For the further development and applications of the above approach to Sobolev
spaces on metric space, see Franchi, HajÃlasz and Koskela [77], Franchi, Lu and
Wheeden [81], HajÃlasz and Kinnunen [104], HajÃlasz and Martio, [107], Heinonen
[115], Heinonen and Koskela [118], KaÃlamajska [147], Kilpeläinen, Kinnunen and
Martio [156], Kinnunen and Martio [157], Koskela and MacManus [161], Shanmu-
galingam [228].
Prior to the work of HajÃlasz, Varopoulos, [250], defined a function space on a
smooth compact manifold, based on an inequality similar to (12). Recently and
independently, Vodop’yanov, [252], used inequality (12) to define a Sobolev space
on a Carnot group.
The following result compares the above definition of the Sobolev space with
the approach based on Poincaré inequalities.

Theorem 3.1 Let X be a doubling space. If 1 < p < ∞, then the following condi-
tions are equivalent.

1. u ∈ M 1,p (X, d, µ).

2. u ∈ Lp (X) and there exist C > 0, σ ≥ 1, 0 ≤ g ∈ Lp (X), and 0 < q < p such
that the Poincaré inequality
Z µZ ¶1/q
¯ ¯
¯u − uB ¯ dµ ≤ C r g q dµ (14)
B σB
 SOBOLEV MET POINCARÉ 13

holds on every ball B of radius r.

Remarks. 1) In fact we prove the implication 2. ⇒ 1. for any p > 0. 2) Under

much more restrictive assumptions on the measure Theorem 3.1 has been proved
by Franchi, Lu and Wheeden, [81], see also [77], [86], [104], [118], [161].
Proof of Theorem 3.1. Integrating inequality (12) over a ball with respect to x
and y we obtain Z Z
|u − uB | dµ ≤ Cr g dµ ,
B B

which proves the implication 1. ⇒ 2. The opposite implication follows from Theo-
rem 3.2 and the Maximal Theorem 14.13.

Theorem 3.2 Let X be a doubling space. Assume that the pair u, g satisfies a
p-Poincaré inequality (5), p > 0. Then
³ ´
|u(x) − u(y)| ≤ Cd(x, y) (M2σd(x,y) g p (x))1/p + (M2σd(x,y) g p (y))1/p (15)
R
for almost every x, y ∈ X, where MR v(x) = sup0<r<R B(x,r)
|v| dµ.

Before we prove Theorem 3.2 we show how to use it to complete the proof of the
implication 2. ⇒ 1. Assume that u, g ∈ Lp (X) satisfy (14). Then inequality (15)
holds with p replaced by q. Note that
¡ ¢1/q 1/q
M2σd(x,y) g q (x) ≤ (M g q (x)) .

Now, g q ∈ Lp/q , p/q > 1, and so the Maximal Theorem 14.13 implies (M g q )1/q ∈ Lp
and hence the claim follows.
Proof of Theorem 3.2. Let x, y ∈ X be Lebesgue points of u; by the Lebesgue
differentiation theorem (see Theorem 14.15) this is true for almost all points. Write
Bi (x) = B(x, ri ) = B(x, 2−i d(x, y)) for each nonnegative integer i. Then uBi (x) →
u(x) as i tends to infinity. Using the triangle inequality, the doubling of µ and the
p-Poincaré inequality we conclude that
∞
X
|u(x) − uB0 (x) | ≤ |uBi (x) − uBi+1 (x) |
i=0
X∞ Z
≤ |u − uBi (x) | dµ
i=0 Bi+1 (x)
X∞ Z
≤ C |u − uBi (x) | dµ
i=0 Bi (x)
∞
ÃZ !1/p
X
p
≤ C ri g dµ
i=0 σBi (x)
14 PIOTR HAJLASZ AND PEKKA KOSKELA

∞
X
≤ C ri (Mσd(x,y) g p (x))1/p
i=0

= Cd(x, y)(Mσd(x,y) g p (x))1/p . (16)

Similarly,
|u(y) − uB0 (y) | ≤ Cd(x, y)(Mσd(x,y) g p (y))1/p .
Moreover,

|uB0 (x) − uB0 (y) | ≤ |uB0 (x) − u2B0 (x) | + |uB0 (y) − u2B0 (x) |
Z
≤ C |u − u2B0 (x) | dµ
2B0 (x)
ÃZ !1/p
≤ Cd(x, y) g p dµ
2σB0 (x)

≤ Cd(x, y)(M2σd(x,y) g p (x))1/p .

The claim follows by combining the above three inequalities. This completes the
proof of Theorem 3.2 and hence that of Theorem 3.1.
It is interesting to observe that Theorem 3.2 can be converted, see also Heinonen
and Koskela [118]. This is the content of the following result.

Theorem 3.3 Let X be a doubling space and u ∈ L1loc (X, µ), 0 ≤ g ∈ Lploc (X, µ),
1 < p < ∞. Suppose that the pointwise inequality
³ ´
|u(x) − u(y)| ≤ Cd(x, y) (Mσd(x,y) g p (x))1/p + (Mσd(x,y) g p (y))1/p

holds for almost all x, y ∈ X with some fixed σ ≥ 1. Then the p-Poincaré inequality
Z µZ ¶1/p
|u − uB | dµ ≤ CP r g p dµ
B 3σB

holds for all balls B. Here CP depends only on p, C, Cd .

Proof. Fix a ball B with radius r. Then for almost all x, y ∈ B we have
³ ´
|u(x) − u(y)| ≤ Cd(x, y) (M (g p χ3σB )(x))1/p + (M (g p χ3σB )(y))1/p .

Fix t0 > 0. Taking an average with respect to x and y, applying Cavalieri’s principle
(see Theorem 14.10) and the weak type estimate for the maximal function (see
Theorem 14.13) we obtain
Z Z
1/p
|u − uB | dµ ≤ Cr (M (g p χ3σB )) dµ
B B
Z ∞
= Crµ(B)−1 µ({x ∈ B : M (g p χ3σB ) > tp }) dt
0
 SOBOLEV MET POINCARÉ 15

µZ t0 Z ∞ µ Z ¶ ¶
C
≤ Crµ(B)−1 µ(B) dt + g p dµ dt
0 t tp 3σB
µ Z0 ¶
−1 1−p
= Crµ(B) t0 µ(B) + Ct0 g p dµ .
3σB
R
The claim follows when we choose t0 = (µ(B)−1 3σB g p dµ)1/p . The proof is com-
plete.
Note that the argument used above is similar to that used in the proof of The-
orem 14.11.
Theorem 3.1 suggests the following question: Is it true that if a pair u, g ∈
Lp (X), 1 < p < ∞, satisfies a p-Poincaré inequality in a doubling space X, then
there exists 1 ≤ q < p such that the pair u, g satisfies a q-Poincaré inequality? This
seems to be a very delicate question, see the discussion in the remark in Section 4
below.
If the answer to the above question were affirmative, Theorem 3.1 would imply
a stronger result: u ∈ M 1,p , p > 1, if and only if u ∈ Lp and there is 0 ≤ g ∈ Lp (X)
such that the pair u, g satisfies a p-Poincaré inequality.
In the special case when X = IRn , d is the Euclidean metric and µ is the
Lebesgue measure, the answer to the above question is in the positive due to the
results of Franchi, HajÃlasz and Koskela [77], and Koskela and MacManus [161].
The following theorem was proved in [77]. The result is a generalization of some
results in [155], [161].

Theorem 3.4 Let u, g ∈ Lp (IRn ), g ≥ 0, p ≥ 1. Suppose that there exist λ ≥ 1

and C such that Z µZ ¶1/p
|u − uB | dx ≤ Cr g p dx ,
B λB
n 1,p n
for all balls B ⊂ IR . Then u ∈ W (IR ) and |∇u| ≤ C1 g a.e. In particular,
Z Z
|u − uB | dx ≤ C2 r g dx,
B B
n
for all balls B ⊂ IR .

Note that it follows from the results stated before Theorem 3.4 that if a pair u, g ∈
1,q
Lp (IRn ) satisfies a p-Poincaré inequality, p > 1, then u ∈ Wloc , for any 1 ≤ q < p.
Indeed, Theorem 3.2 together with the weak type estimate for the maximal function
and the embedding Lpw ⊂ Lqloc for all q < p (see Theorem 14.11) imply that for some
h ∈ Lqloc the inequality |u(x)−u(u)| ≤ |x−y|(h(x)+h(y)) holds a.e. Then the claim
follows from (13). This argument, however, does not guarantee that u ∈ W 1,p (IRn ).
Thus the proof of Theorem 3.4 requires arguments of a completely different nature.
For far reaching generalizations of Theorem 3.4, see [77] and [161] and also
Section 13.
16 PIOTR HAJLASZ AND PEKKA KOSKELA

4 Examples and necessary conditions

We first discuss three examples that indicate the dependence of the validity of
p-Poincaré inequalities on the exponent p. Notice that if a pair u, g satisfies a p-
Poincaré inequality, then it satisfies a q-Poincaré inequality for all q > p by Hölder’s
inequality. The following examples show that this is not the case for q < p.
Example 4.1 Let X = R2 be equipped with the Euclidean metric and let µ be
the measure generated by the density dµ(x) = |x2 |t dx, t > 0, where x2 denotes the
second coordinate of x. Then the Poincaré inequality (5) holds for each Lipschitz
function u with g = |∇u| if and only if p > t + 1.
The Poincaré inequality holds as µ is an Ap -weight for the indicated values of
p; see [67], [39], [120, Theorem 15.26], [102]. On the other hand, the p-Poincaré
inequality fails for p = 1 + t and hence for 1 ≤ p ≤ 1 + t. To see this, let B be the
disk of radius 2 and with center (0, 1). Let us consider a sequence ui of Lipschitz
functions that only depend on x2 and such that ui = 1 if x ∈ B and x2 ≤ 2−i ,
ui = 0 if x ∈ B and x2 ≥ 1, ui (x) = −i−1 log2 (x2 ) if 2−i ≤ x2 ≤ 1. Then
Z Z 1
1+t −(1+t) ds
|∇ui | dµ ≤ 2(i log 2)
B 2−i s

which tends to zero as i approaches infinity. On the other hand, |ui (x) − ui B | ≥ 1/2
for all x either in the part of B above the line x2 = 1 or in the part below the
line x2 = 0. Hence the integral of |ui − ui B | over B is bounded away from zero
independently of i, and so the (1+t)-Poincaré inequality cannot hold for all Lipschitz
functions. Using a standard regularization argument we can then assume that
functions ui in the above example are C ∞ smooth, so the (1+t)-Poincaré inequality
cannot hold for all C ∞ smooth functions either.
Example 4.2 Let X = {(x1 , x2 , . . . , xn ) ∈ IRn : x21 + · · · + x2n−1 ≤ x2n } be equipped
with the Euclidean metric of IRn and with the Lebesgue measure. The set X consists
of two infinite closed cones with a common vertex. Denote the upper cone by X+
and the lower one by X− .
We will prove that the p-Poincaré inequality (5) holds in X for every pair u, g
where u is a continuous function and g an upper gradient of u if and only if p > n.
(For more information about upper gradients, see Section 10.2.)
First we prove that the inequality fails for p = n (and hence for p < n). Fix
ε > 0. Since ϕ(x) = log | log |x|| satisfies ϕ ∈ W 1,n (B n (0, 1/2)) and ϕ(x) → ∞ as
x → 0, we can truncate it to obtain a continuous function uε ∈ W 1,n (X+ ) such that
uε (0) = 1, uε (x) = 0 for |x| ≥ ε and k∇uε kLn (X+ ) < ε. We extend this function
to the lower cone as the constant 1. Fix a ball B centered at the origin. Then
k∇uε kLn (B) < ε while kuε − uε B kL1 (B) > C uniformly with respect to ε, and thus
the n-Poincaré inequality cannot hold.
It remains to prove the inequality for p > n. Since |∇u| ≤ g for an upper
gradient of u (Proposition 10.1), it suffices to prove the p-Poincaré inequality for
 SOBOLEV MET POINCARÉ 17

the pair u, |∇u|. By Theorem 3.3 it suffices to verify the pointwise estimate
|u(x) − u(y)| ≤ C|x − y|(M2|x−y| |∇u|p (x) + M2|x−y| |∇u|p (y))1/p .
We can assume that x and y belong to different cones as the p-Poincaré inequality
holds in each of those two cones. Then, by the triangle inequality, either |u(x) −
u(0)| ≥ |u(x) − u(y)|/2 or |u(y) − u(0)| ≥ |u(x) − u(y)|/2. Assume that the first
inequality holds. Let ∆ = X+ ∩ B(x, |x|). Then by the embedding into Hölder
continuous functions
µZ ¶1/p
|u(x) − u(0)| ≤ C|x − 0|1−n/p |∇u|p ≤ C|x − y|(M2|x−y| |∇u|p (x))1/p .
∆

This ends the proof of the claim.

Modifying the argument used above one can construct many other examples. Let
for example X be the union of two 3-dimensional planes in IR5 whose intersection
is a line. Equip X with the 3-dimensional Lebesgue measure and with the metric
induced by the Euclidean metrics of the planes. Then the p-Poincaré inequality
holds in X for all pairs u, g, where u is a continuous function in X and g an upper
gradient of u, if and only if p > 2.
A much more general result that allows one to build similar examples in the
setting of metric spaces was proven by Heinonen and Koskela, [118, Theorem 6.15].
Example 4.3 For each 1 < p ≤ n there is an open set X ⊂ IRn equipped with the
Euclidean metric and the Lebesgue measure, such that the p-Poincaré inequality (5)
holds for each smooth function u with g = |∇u| but no Poincaré inequality holds
for smaller exponents for all smooth functions.
Such an example was constructed by Koskela, [160]. We will recall the idea of the
example following [160]. Let E ⊂ IRn be a compact set such that W 1,p (IRn \ E) =
W 1,p (IRn ), and W 1,q (IRn ) is a proper subset of W 1,q (IRn \ E) for all 1 ≤ q < p.
In other words, the set E is W 1,p -removable but it is not W 1,q -removable for any
1 ≤ q < p. Such sets were explicitly constructed in [160]. In fact there is a smooth
function u in IRn \ E such that |∇u| ∈ Lq (IRn ), for all q < p, but the pair u,
|∇u| does not satisfy a q-Poincaré inequality in IRn \ E for any q < p. The pair
satisfies the p-Poincaré inequality but |∇u| 6∈ Lp (IRn \ E) as otherwise we would
have u ∈ W 1,p (IRn \ E) = W 1,p (IRn ) and hence it would even satisfy the 1-Poincaré
inequality. Thus this example does not solve the question posed after Theorem 3.3.
Remark. The last example shows that it may happen that a p-Poincaré inequality
holds for all smooth pairs u, |∇u| in X and there is a smooth function u in X such
that no q-Poincaré inequality holds for any q < p for the pair u, |∇u|. However, in
this example |∇u| 6∈ Lp (X) and hence we do not know if a p-Poincaré inequality
for the pair u, g with g ∈ Lp (X) implies a q-Poincaré inequality for some q < p.
As Example 4.1 shows, a q-Poincaré inequality cannot hold for all q < p in general,
but we do not know if it can hold for some q < p sufficiently close to p.
Let us next describe some necessary geometric conditions for the validity of
Poincaré inequalities. The first result says that a Poincaré inequality implies that
18 PIOTR HAJLASZ AND PEKKA KOSKELA

there are short rectifiable curves. We would like to thank Anton Petrunin for his
very clever argument with a “broken length” that we use in the proof of Proposi-
tion 4.4. Our original version of the proposition was proven under more restrictive
assumptions on X.
Definition. If each pair x1 , x2 of points in a metric space X can be joined by a
curve whose length is no more than Cd(x1 , x2 ), then we say that X is quasiconvex.
Recall that X is proper if closed balls are compact. Observe that this is a
stronger condition than being locally compact as the example X = IRn \ {0} shows.

Proposition 4.4 Suppose that X is proper, path connected and doubling. Let p ≥
1. If each pair u, g of a continuous function and its upper gradient satisfies a p-
Poincaré inequality (with fixed σ, CP ), then X is quasiconvex.

Proof. Fix k ∈ IN. Let γ : [0, 1] → X be a continuous path. For any partition τ :
0 = t0 < t1 < . . . < tn = 1 consider the sum
n−1
X
sτ = min{l(γ|[ti ,ti+1 ] ), kd(γ(ti ), γ(ti+1 ))} ,
i=0

where l denotes the length of a curve, and then define

lk (γ) = inf sτ ,
τ

where the infimum is taken over all possible partitions of [0, 1]. We do not require
that γ is rectifiable. Indeed, if no part of γ is rectifiable, then lk (γ) = kd(γ(0), γ(1)).
Fix x0 ∈ X and define
uk (x) = inf lk (γ)
γ

for x ∈ X, where the infimum is taken over all curves that join x to x0 .
It is easy to verify that the function uk is k-Lipschitz and that the function g ≡ 1
is an upper gradient of uk . Thus the pair (uk , g) satisfies the p-Poincaré inequality
and hence by Theorem 3.2
³ ´
|uk (x)−uk (y)| ≤ Cd(x, y) (M2σd(x,y) g p (x))1/p + (M2σd(x,y) g p (y))1/p = 2Cd(x, y) .

In particular, for y = x0 , we obtain

|uk (x)| ≤ Cd(x, x0 ) .

Let γk be a curve that joins x to x0 and such that lk (γk ) ≤ 2uk (x). Then let τk be
a partition of [0, 1] such that sτk ≤ 3uk (x).
We do this for all k ≥ 1. We would like to show that a subsequence γki converges,
in some sense, to a rectifiable curve γ that joins x to x0 and such that

l(γ) ≤ lim inf lki (γki ) ≤ Cd(x, x0 ) .

i→∞
 SOBOLEV MET POINCARÉ 19

This would end the proof.

Recall the following classical argument (cf. Lemma 9.4). Let {δk }∞ k=1 be a se-
quence of rectifiable curves in a proper metric space Y . Assume that all the curves
δk join given two points x, y ∈ Y and that supk l(δk ) < ∞. Parametrize each δk by
arclength. Scaling the arclength parametrizations we may assume that all curves are
defined on the interval [0, 1]. Now it easily follows that the family {δk } is equicon-
tinuous (because of the good parametrization and the fact that supk l(δk ) < ∞).
By a standard diagonal method we can find a subsequence δki which converges on
a dense subset of [0, 1]. The equicontinuity implies the uniform convergence on the
whole interval to a curve δ. Then it is easy to show that l(δ) ≤ lim inf i→∞ l(δki ).
Unfortunately the above argument does not apply to our situation as lk (γk ) is
not the length of γk . We do not even know whether the curves γk are rectifiable.
There is however a trick that makes the above argument work: we modify our
metric space by attaching to it infinitely many segments and then we modify the
family of curves. In this new framework the above argument will work perfectly.
We construct a new metric space X b by attaching to the space X infinitely many
Euclidean segments. We do this as follows. Let τk : 0 = t0 < t1 < t2 < . . . < tn = 1
be a partition associated with γk chosen as above. If

l(γk |[ti ,ti+1 ] ) ≥ kd(γk (ti ), γk (ti+1 )) , (17)

then we glue to the space X a straight Euclidean segment Ik,i of the length
d(γk (ti ), γk (ti+1 )) (i.e. Ik,i is isometric with [0, d(γk (ti ), γk (ti+1 )]) in such a way
that ends of the segment are attached to the space at points γk (ti ) and γk (ti+1 ).
We do this for every curve γk and every i such that (17) is true.
The space Xb is equipped with a natural metric which is induced from the Eu-
b
clidean metric in each segment and the metric d in X. We denote the metric in X
b
by d.
Now denote by γ bk the new curve which is obtained for γk in the following way.
Let 0 ≤ i ≤ n − 1. If (17) does not hold, then

γ
bk |[ti ,ti+1 ] = γk |[ti ,ti+1 ] .

If (17) is true, then γ

bk |[ti ,ti+1 ] is the obvious linear function with values in the
segment Ik,i .
This means that if the part γk |[ti ,ti+1 ] of the curve γk is too long in the sense of
(17), then we replace this part by a shortcut going through the attached segment.
Since supk sτk ≤ Cd(x, x0 ) we conclude that l(Ik,i ) < Cd(x, x0 )/k. Thus, in
the worst case, we attach to the space X infinitely many segments Ik,i of length
converging to zero. This easily implies that the space X b is proper as well. Now
b γk ) ≤ sτk ≤ Cd(x, x0 ), where b
l(b b
l denotes the length in the new metric space X.
Now we are in a situation where the above classical argument applies. Choosing
a good parametrization of γbk we may subtract a subsequence γ bki that converges
20 PIOTR HAJLASZ AND PEKKA KOSKELA

uniformly to a curve γ such that

b
l(γ) ≤ lim inf b
l(b
γki ) ≤ Cd(x, x0 ) .
i→∞

Since the lengths of Ik,i converge to zero as k → ∞ we conclude that the values of γ
belong to X and hence one can easily show that b l(γ) = l(γ). The proof is complete.
Thus the validity of a p-Poincaré inequality guarantees the existence of short
curves. If the doubling measure µ behaves as the Euclidean volume and the expo-
nent p is no more than the growth order of the volume, then X cannot have narrow
parts. This conclusion is a consequence of Proposition 4.5 below. Under the ad-
ditional assumption that the space satisfies a weak local version of quasiconvexity,
this result can be deduced from the results in [118].

Proposition 4.5 Suppose that X is proper, path connected and that µ(B(x, r)) ≈
rs with s > 1 for each x and all r. Assume that each pair u, g of a continuous
function and its upper gradient satisfies an s-Poincaré inequality (with fixed σ, CP ).
If x0 ∈ X, r > 0, and x1 , x2 ∈ B(x0 , r) \ B(x0 , r/2), then x1 , x2 can be joined in
B(x0 , Cr) \ B(x0 , r/C) by a curve whose length does not exceed Cd(x1 , x2 ).

Notice that the claim of the proposition would still be true if we replaced the
s-Poincaré inequality by a p-Poincaré inequality, p < s, as a p-Poincaré inequality
implies an s-Poincaré inequality, by means of the Hölder inequality. However we
cannot replace the s-Poincaré inequality by a p-Poincaré for any p > s as follows
from Example 4.2.
Proof. The proof is very similar to the arguments used in the proof of [118,
Corollary 5.8] and in the proof of Proposition 4.4. Throughout C ≥ 1 denotes a
constant whose value can change from line to line but that only depends on the
given data. By Proposition 4.4 we may assume that d(x1 , x2 ) ≥ C −1 r as x1 and x2
can be connected by a curve of length comparable to d(x1 , x2 ). Cutting pieces of the
curve near x1 and x2 we obtain rectifiable curves F1 , F2 ⊂ B(x0 , 2r) \ B(x0 , r/4),
both of length comparable to r and such that dist (F1 , F2 ) is comparable to r as
well. It suffices to show that F1 , F2 can be joined by a curve of length less than Cr
inside B(x0 , Cr) \ B(x0 , r/C) for sufficiently large C.
If follows from the s-Poincaré inequality and from the volume growth condition
that Z
g s dµ ≥ C −1 (18)
B(x0 ,6σr)

for any upper gradient g of any continuous function u that takes on the constant
value 0 in F1 and takes on a value greater than or equal to 1 at each point of F2 .
Indeed, assume first that |uB(x0 ,2r) | ≤ 1/2. Then slightly modifying the proof of
(16) we get for all x ∈ F2
à Z !1/s
1 1/s −1 s
≤ |u(x) − uB(x0 ,2r) | ≤ Cr sup R g dµ .
2 R<4σr B(x,R)
 SOBOLEV MET POINCARÉ 21

Thus for some Rx < 4σr

Z
C −1 Rx /r ≤ g s dµ .
B(x,Rx )

Now inequality
S (18) follows P
from the covering lemma (Theorem 14.12) and the fact
that if F2 ⊂ Bi (ri ), then i ri ≥ C −1 r. If |uB(x0 ,2r) | ≥ 1/2, then inequality (18)
follows by a symmetric argument. The proof of (18) is complete.
Fix τ > 4. Set g1 (x) = (log( τ4 ))−1 d(x, x0 )−1 in B(x0 , τ r) \ B(x0 , τ −1 r) and
extend g1 as zero to the rest of X. Suppose that F1 , F2 cannot
R be joined in B(x0 , τ r)\
B(x0 , τ −1 r) by a rectifiable curve. Define u1 (x) = inf γx γx g1 ds where the infimum
is taken over all rectifiable curves that join x to F1 . Then g1 is an upper gradient
of u1 , the restriction of u to F1 is zero and u(x) ≥ 1 at each point of F2 . By the
preceding paragraph, we see that the integral of g1s over B(x0 , 6σr) is bounded away
from zero. On the other hand a computation using the volume growth condition and
Cavalieri’s principle (Theorem 14.10) shows that the integral of g1s over B(x0 , 6σr)
goes to 0 as τ goes to infinity. Hence τ is bounded from above. Thus we can fix
τ large enough so that F1 , F2 can be joined by a rectifiable curve in B(x0 , τ r) \
B(x0 , τ −1 r). It remains to find such a curve with a length comparable to r.
Set a = inf γ l(γ), where the infimum is taken over all rectifiable curves that join
F1 to F2 in B(x0 , τ r)\B(x0 , τ −1 r). We define a function u2 similarly as we defined u1
above using g1 + g2 , where g2 (x) = a−1 χU (x), and χU is the characteristic function
of B(x0 , τ r) \ B(x0 , τ −1 r). Observe that u2 |F1 ≡ 0 and u2 |F2 ≥ 1 independently of
τ and hence (18) holds with g = g1 + g2 . As we can make the integral of g1s over
the ball B(x0 , 6σr) as small as we wish by choosing the constant τ large enough,
we obtain that the integral of g2s over B(x0 , 6σr) must be bounded away from zero,
and thus the volume growth condition implies that a ≤ Cr, as desired.

5 Sobolev type inequalities by means of Riesz po-

tentials
As it was pointed out in Section 2, one of the aims of this paper is to prove a global
Sobolev inequality
µZ ¶1/q µZ ¶1/p
q p
inf |u − c| dµ ≤C g dµ , (19)
c∈IR Ω Ω

where q > p, or at least a weak local Sobolev inequality

µZ ¶1/q µZ ¶1/p
q p
inf |u − c| dµ ≤ Cr g dµ , (20)
c∈IR B 5σB

where σ ≥ 1, and B is any ball of radius r, assuming only that the pair u, g satisfies
a p-Poincaré inequality.
22 PIOTR HAJLASZ AND PEKKA KOSKELA

Inequality (19) requires some additional information on Ω, while (20) turns out
to be true in a very general setting.
Another question we deal with is how to determine the best possible Sobolev
exponent q in inequalities (19) and (20).
In the remaining part of the section we will be consider inequalities of the type
(20). The case of the global Sobolev inequality (19) will be treated in Section 9.
Let X be a doubling space. Beside the doubling condition we will sometimes
require that µ ¶s
µ(B) r
≥ Cb (21)
µ(B0 ) r0
whenever B0 is an arbitrary ball of radius r0 and B = B(x, r), x ∈ B0 , r ≤ r0 .
Notice that the doubling condition on µ always implies (21) for some exponent s
that only depends on the doubling constant of µ. This follows by a standard iteration
of the doubling condition, see Lemma 14.6 in the appendix. Inequality (21) could
well hold with exponents smaller than the one following from the doubling condition
and in the following results s refers to any exponent for which (21) is valid.

Theorem 5.1 Assume that the pair u, g satisfies a p-Poincaré inequality (5), p >
0, in a doubling space X. Assume that the measure µ satisfies condition (21).

1. If p < s, then
∗ µZ ¶1/p
µ({x ∈ B : |u(x) − uB | > t})tp ∗
≤ Crp g p dµ ,
µ(B) 5σB

where p = sp/(s−p) and B is any ball of radius r. Hence for every 0 < h < p∗
∗

µZ ¶1/h µZ ¶1/p
h p
|u − uB | dµ ≤ Cr g dµ .
B 5σB

Moreover for every q with p < q < s

µZ ¶1/q∗ µZ ¶1/q
q∗ q
|u − uB | dµ ≤ Cr g dµ ,
B 5σB

where q ∗ = sq/(s − q) and B is any ball of radius r. If, in addition, the pair
u, g has the truncation property, then
µZ ¶1/p∗ µZ ¶1/p
p∗ p
|u − uB | dµ ≤ Cr g dµ . (22)
B 5σB

2. If p = s, then Z µ ¶
C1 µ(B)1/s |u − uB |
exp dµ ≤ C2 . (23)
B rkgkLs (5σB)
 SOBOLEV MET POINCARÉ 23

3. If p > s, then u (after redefinition in a set of measure zero) is locally Hölder

continuous and
µZ ¶1/p
sup |u(x) − uB | ≤ Cr g p dµ . (24)
x∈B 5σB

In particular
µZ ¶1/p
s/p
|u(x) − u(y)| ≤ Cr0 d(x, y)1−s/p g p dµ (25)
5σB0

for all x, y ∈ B0 , where B0 is an arbitrary ball of radius r0 .

The constants in the theorem depend on p, q, h, s, Cd , σ, CP , and Cb only.

Remarks. 1) Inequality (22) holds also for functions on graphs, see Theorem 12.2.
2) Assuming that the space is connected we can improve on inequality (23); see
Section 6.
3) Instead of assuming that X be doubling we could assume, for instance, that the
doubling condition holds on all balls with radii bounded from above by r0 , (such
a situation occurs for example on Riemannian manifolds with a lower bound on
the Ricci curvature, see Section 10) or that it holds on a given open set. Then the
inequalities of the theorem would hold on balls with radii bounded from above or
on small balls centered at the open set. We leave it to the reader to check that the
proof of Theorem 5.1 gives such a statement.
4) A modification of the proof shows that the ball 5σB can be replaced by (1+ε)σB;
the details are left to the reader.
5) We present only one of the possible proofs of the above theorem. The proof
can also be based on the embedding theorem for Sobolev spaces on metric spaces
from HajÃlasz, [102]. This approach uses the observation that a family of Poincaré
inequalities leads to pointwise inequalities (15); we do not provide the details here.
Since the proof of the theorem is rather complicated, we begin with some com-
ments that will explain the idea.
∗
In one of the proofs of the classical Sobolev embedding W 1,p (B) ⊂ Lp (B),
where 1 ≤ p < n, p∗ = np/(n − p), and B is an n-dimensional Euclidean ball, one
first proves the inequality
|u(x) − uB | ≤ CI1B |∇u|(x), (26)
R
where I1B g(x) = B g(z)|x − z|1−n dz and then applies the Fractional Integration
Theorem which states that
∗
I1B : Lp (B) −→ Lp (B) (27)
is a bounded operator for 1 < p < n. If p = 1 one only gets a weak type estimate
µZ ¶ n−1
n
n
B
|{x ∈ B : I1 g(x) > t}|t n−1 ≤ C |g(z)| dz
B
24 PIOTR HAJLASZ AND PEKKA KOSKELA

n/(n−1)
in place of (27), which, in turn, leads to the embedding W 1,1 (B) ⊂ Lw (B).
Then the embedding W 1,1 (B) ⊂ Ln/(n−1) (B) follows from Theorem 2.1.
The main idea of our proof of inequalities like (19) or (20) is to mimic the above
argument. Thus the proof splits into two steps.
Assume that a pair u, g satisfies a p-Poincaré inequality in a given doubling
space. In the first step we prove the inequality
σ,B
|u − uB | ≤ CJ1,p g, (28)
σ,B
where J1,p is a suitable generalization of the Riesz potential I1B and then, in the
second step, we prove a version of the Fractional Integration Theorem for the op-
σ,B
erator J1,p . This will complete the proof of (20). The proof of (19) will require a
more sophisticated version of the inequality (28); the details will be completed in
Section 9 where we introduce an appropriate class of domains Ω for the Sobolev-
Poincaré embedding (19).
Any inequality of the type (28) will be called a representation formula.
σ,B
Before we define Jα,p we continue with a discussion on Riesz potentials to
explain the motivation. The classical Riesz potential is defined as
Z
g(y)
Iα g(x) = γα,n dy, (29)
IRn |x − y|n−α
where 0 < α < n and γα,n is a suitable constant. In this paper the exact value of
the constant γα,n is irrelevant to us. Moreover, for our purposes, any operator J
such that
C1 Iα g ≤ Jg ≤ C2 Iα g for g ≥ 0 (30)
is as good as Iα .
A natural generalization of the Riesz potential to the setting of doubling spaces
is Z
g(y)dα (x, y)
Iα g(x) = dµ(y) ,
X µ(B(x, d(x, y)))
or its local version Z
g(y)dα (x, y)
IαΩ g(x) = dµ(y) . (31)
Ω µ(B(x, d(x, y)))
We would like to estimate |u − uB | by CI1Ω g, but, in general, this is not possible.
Instead of that we have to consider a potential which is strictly larger than I1Ω g.
Observe that the potential defined by
∞
à Z !
X
eΩ
Iα g(x) = 2iα
µ(Bi (x))−1
|g(y)| dµ(y) , (32)
i=−∞ Ai (x)∩Ω

where Ai (x) = Bi (x) \ Bi−1 (x) = B(x, 2i ) \ B(x, 2i−1 ), is equivalent to IαΩ g in the
sense of (30). Note that if 2i−1 > diam Ω, then Ai (x) ∩ Ω = ∅, so all the summands
 SOBOLEV MET POINCARÉ 25

in (32) for 2i > 2diam Ω vanish. Thus, replacing the integral over Ai (x) ∩ Ω by the
integral over Bi (x) and then taking the sum over 2i ≤ 2diam Ω, we obtain the new
potential ÃZ !
X
Ω iα
Jα g(x) = 2 |g| dµ ,
2i ≤2diam Ω Bi (x)

which satisfies IeαΩ g ≤ JαΩ g. Now we define

ÃZ !1/p
X
σ,Ω iα p
Jα,p g(x) = 2 |g| dµ ,
2i ≤2σdiam Ω Bi (x)

where σ ≥ 1, p > 0 and α > 0 are fixed constants.

Another generalization is

∞
ÃZ !1/p
X |g(y)|p dαp (x, y)
Ω
Iα,p g(x) = dµ(y) .
i=−∞ Ai (x)∩Ω µ(B(x, d(x, y)))

Ω
Observe that Iα,1 g = IαΩ |g|, and Iα,p
Ω
g ≤ CJα,p1,Ω
g a.e. Thus once we prove the
σ,Ω Ω
fractional integration theorem for Jα,p g it is true for Iα,p g as well.
In Section 9 we will obtain a version of the representation formula (28) with
B σ,B
I1,p gin place of J1,p ; see Theorem 9.10.

Theorem 5.2 Let the pair u, g satisfy a p-Poincaré inequality in a doubling space
X. Then for every ball B ⊂ X the representation formula
σ,B
|u(x) − uB | ≤ CJ1,p g(x) (33)

holds almost everywhere in B.

This representation formula together with a suitable Fractional Integration Theorem

(see Theorem 5.3) will lead to embedding (20).
Proof. The argument is very similar to that used in the proof of inequality (16).
Let x ∈ B be a Lebesgue point of u. Put Di (x) = B(x, 2i σ −1 ). Let i0 be the
least integer such that 2i0 ≥ σdiam B. Then B ⊂ Di0 (x). Since uDi (x) → u(x) as
i → −∞ we obtain
i0
X
|u(x) − uB | ≤ |uB − uDi0 (x) | + |uDi (x) − uDi−1 (x) |
i=−∞
i0
ÃZ !1/p
X
−1 i p
≤ C σ 2 g dµ
i=−∞ σDi (x)
σ,B
≤ CJ1,p g(x).
26 PIOTR HAJLASZ AND PEKKA KOSKELA

Theorem 5.3 Let Ω ⊂ X be an open and bounded set and let 0 < p < ∞, 1 ≤
σ < ∞. Assume that the measure µ is doubling on V = {x ∈ X : dist (x, Ω) <
2σdiam Ω}. Moreover, assume that for some constants Cb , s > 0
³ r ´s
µ(B(x, r)) ≥ Cb µ(Ω)
diam Ω
whenever x ∈ Ω and r ≤ σdiam Ω and that g ∈ Lp (V, µ).
∗
σ,Ω
1. If αp < s, then Jα,p g ∈ Lpw (Ω) where p∗ = sp/(s − αp). Moreover
∗ ∗ ∗ ∗
σ,Ω
µ({x ∈ Ω : Jα,p g > t}) ≤ C1 t−p (diam Ω)αp µ(Ω)1−p /p
kgkpLp (V,µ) (34)

for t > 0, and hence for every 0 < r < p∗

σ,Ω
kJα,p gkLr (Ω,µ) ≤ C2 (diam Ω)α µ(Ω)1/r−1/p kgkLp (V,µ) . (35)

Here the constants C1 and C2 depend on α, σ, p, Cb , s and Cd only.

2. If p < q and αq < s, then
σ,Ω
kJα,p gkLq∗ (Ω,µ) ≤ C(diam Ω)α µ(Ω)−α/s kgkLq (V,µ) , (36)

where q ∗ = sq/(s − αq) and C = C(α, σ, p, q, b, s, Cd ).

3. If αp = s, then
Z Ã !
C1 µ(B)1/s Jα,p
σ,Ω
g
exp α
dµ ≤ C2 ,
Ω (diam Ω) kgkLs (V )

where Ci = C(α, σ, p, b, s, Cd ), i = 1, 2.
σ,Ω
4. If αp > s, then Jα,p ∈ L∞ (Ω, µ) and
σ,Ω
kJα,p gkL∞ (Ω,µ) ≤ C(diam Ω)α µ(Ω)−1/p kgkLp (V,µ) ,

where C = C(α, σ, p, b, s, Cd ).

Proof of Theorem 5.3. We modify a standard proof for the the case of usual
Riesz potentials. All the constants C appearing in the proof depend on α, σ, p, q,
b, s, and Cd only.
Case αp < s. Take arbitrary q ≥ p such that αq < s. Fix 0 < r ≤P 2σdiam Ω.
σ,Ω r
Decompose
P the sum which defines J α,p g into J r g + J g, where J r g = 2i ≤r and
J r g = r<2i ≤2σdiam Ω . For x ∈ Ω we have
 
X 1/p
Jr g(x) ≤  2iα  (MV |g|p (x))1/p ≈ rα (MV |g|p (x)) .
2i ≤r
 SOBOLEV MET POINCARÉ 27

Here MV h denotes the maximal function relative to the open set V .

To estimate J r g, we apply the lower bound on µ :
ÃZ !1/p
X
J r g(x) = 2iα |g|p dµ
r<2i ≤2σdiam Ω Bi (x)
ÃZ !1/q
X
iα −1/q q
≤ 2 µ(Bi (x)) |g| dµ
r<2i ≤2σdiam Ω Bi (x)

X µZ ¶1/q
≤ C 2i(α−s/q) (diam Ω)s/q µ(Ω)−1/q |g|q dµ
r<2i ≤2σdiam Ω V
µZ ¶1/q
≤ Cr(α−s/q) (diam Ω)s/q µ(Ω)−1/q |g|q dµ .
V

In the last step we used the fact α − s/q < 0 to estimate the sum of the series by
its first summand. Now
à µZ ¶1/q !
σ,Ω α p 1/p (α−s/q) s/q −1/q q
Jα,p g(x) ≤ C r (MV |g| ) + r (diam Ω) µ(Ω) |g| dµ .
V

Note that
µZ ¶1/q
α p 1/p (α−s/q) s/q −1/q q
r (MV |g| ) ≤r (diam Ω) µ(Ω) |g| dµ
V

if and only if
³ ´q/s
r ≤ (diam Ω)µ(Ω)−1/s kgkLq (V ) /(MV |g|p )1/p . (37)

If the RHS in (37) does not exceed σdiam Ω, then we take r equal to the RHS. In
this case we get
σ,Ω αq/s
Jα,p g(x) ≤ C(diam Ω)α µ(Ω)−α/s kgkLq (V ) (MV |g|p (x))(s−αq)/sp , (38)

and hence
σ,Ω
Jα,p g(x)sp/(s−αq) ≤ C(diam Ω)αsp/(s−αq) µ(Ω)−αp/(s−αq)
αqp/(s−αq)
× kgkLq (V ) MV |g|p (x). (39)

If the RHS in (37) is greater than σdiam Ω, then we take r = σdiam Ω. Then
σ,Ω
Jα,p g(x) ≤ C(diam Ω)α µ(Ω)−1/q kgkLq (V ) . (40)

Let A1 denote the set of points in Ω for which (38) holds and let A2 consist of
σ,Ω
those points in Ω that satisfy (40). Write Ωt = {x ∈ Ω : Jα,p g > t}. Then
µ(Ωt ) ≤ µ(A1 ∩ Ωt ) + µ(A2 ∩ Ωt ). (41)
28 PIOTR HAJLASZ AND PEKKA KOSKELA

If we take q = p, then inequality (34) follows from estimates (39), (40), and (41):
the weak type estimate for the maximal function MV |g|p (see Theorem 14.13) gives
³n o´
αp2 /(s−αp)
µ(A1 ∩ Ωt ) ≤ µ CDkgkLp (V ) MV |g|p > tsp/(s−αp)
∗ αp2 /(s−αp)
≤ CDt−p kgkLp (V ) kgkpLp (V )
∗ ∗
= CDt−p kgkpLp (V )

with D = (diam Ω)αsp/(s−αp) µ(Ω)−αp/(s−αp) , and from (40) we obtain A2 ∩ Ωt = ∅

when t ≥ C(diam Ω)α µ(Ω)−1/p kgkLp (V ) and for all smaller t
∗ ∗ ∗
/p −p∗
µ(A2 ∩ Ωt ) ≤ µ(Ω) ≤ C(diam Ω)αp µ(Ω)1−p t kgkpLq (Ω) .

This completes the proof of inequality (34). Inequality (35) follows from Theo-
rem 14.11.
To prove inequality (36) we take Lq/p -norms on both sides of inequalities (39)
and (40) and after that we apply Maximal Theorem 14.13; we use the fact that
q/p > 1.
Case αp = s. Notice first that
X tk
exp(t) = .
k!
k≥0

Secondly, (35) and the Hölder inequality give for each integer k ≥ 1 the estimate
σ,Ω
kJα,p gkLk (Ω,µ) ≤ C(diam Ω)α µ(Ω)1/k−1/s kgkLs (V,µ) .
By keeping good track of the constants appearing in the proof of (35), one can
check that C = C0 (σ, s, b, Cd )k. The desired inequality follows by summing over k.
We leave the details to the reader as we prove a better estimate in the next section
under slightly stronger assumptions.
Case αp > s. The lower bound on µ gives
ÃZ !1/p
X
σ,Ω iα −1/p p
Jα,p g(x) = 2 µ(Bi (x)) |g| dµ
2i ≤2σdiam Ω Bi (x)
X
≤ C 2i(α−s/p) (diam Ω)s/p µ(Ω)−1/p kgkLp (V )
2i ≤2σ diam Ω

≈ C(diam Ω)α µ(Ω)−1/p kgkLp (V ) .

The proof of Theorem 5.3 is complete.
Proof of Theorem 5.1. All the inequalities but (22) and (25) follow directly from
Theorem 5.2 and Theorem 5.3. The Hölder continuity estimate (25) follows using
(24) and the lower bound (21). If p∗ < 1, then inequality (22) is trivial as it is
weaker than the p-Poincaré inequality. If p∗ ≥ 1, (22) follows from Theorem 2.3
and from the first inequality in Theorem 5.1. The proof is complete.
 SOBOLEV MET POINCARÉ 29

6 Trudinger inequality
When X = IRn and u belongs to the Sobolev class W 1,n (Ω) for a ball Ω, one has
the following Trudinger inequality [241]:
Z µ ¶n/(n−1)
C1 |u − uΩ |
exp dx ≤ C2 .
Ω ||∇u||Ln (Ω)

Here C1 and C2 depend only on the dimension n. As in case of the Poincaré in-
equality, the exact value of uΩ is not crucial. In fact, it is easy to see that we may
replace it by the average of u over some fixed ball B ⊂⊂ Ω. In the previous section
we observed that an s-Poincaré inequality with s not exceeding the lower order of
the doubling measure results in exponential integrability. We do not know if one
could get an analog of the Trudinger inequality in such a general setting but we
doubt it.
In this section we verify an analog of the Trudinger inequality for connected
doubling spaces. Thus the only assumption we need to add is that X be connected.
For related results, see Bakry, Coulhon, Ledoux and Saloff-Coste [4], Buckley and
O’Shea [21], Coulhon [54], and MacManus and Pérez [185].

Theorem 6.1 Assume that X is a connected doubling space and that the measure µ
satisfies condition (21) with s > 1. Suppose that the pair u, g satisfies an s-Poincaré
inequality. Then there are constants C1 and C2 such that
Z µ ¶s/(s−1)
C1 µ(B)1/s |u − uB |
exp dµ ≤ C2 (42)
B diam (B)kgkLs (5σB)

for any ball B ⊂ X.

Remarks. 1) It is easy to deduce from the connectivity of the space that condition
(21) cannot hold with s < 1. We leave the details to the reader. 2) The argument
employed in the proof actually shows that the inequality holds with 5σB replaced
by (1 + ε)σB.
For the proof of this theorem we need a chain condition, a version of which will
also be used later on.
We say that X satisfies a chain condition if for every λ ≥ 1 there is a constant
M such that for each x ∈ X and all 0 < r < R < diam (X)/4 there is a sequence of
balls B0 , B1 , B2 , . . . , Bk for some integer k with

1. B0 ⊂ X \ B(x, R) and Bk ⊂ B(x, r),

2. M −1 diam (Bi ) ≤ dist (x, Bi ) ≤ M diam (Bi ) for i = 0, 1, 2, . . . , k,
3. there is a ball Ri ⊂ Bi ∩Bi+1 , such that Bi ∪Bi+1 ⊂ M Ri for i = 0, 1, 2, . . . , k,
4. no point of X belongs to more than M balls λBi .
30 PIOTR HAJLASZ AND PEKKA KOSKELA

The sequence {Bi } will be called a chain associated with x, r, R.

The existence of a doubling measure on X does not guarantee a chain condition.
In fact, such a space can be badly disconnected, whereas a space with a chain
condition cannot have “large gaps”.
Let us show that each connected doubling space satisfies a chain condition. Fix
ε. Write Aj (x) = B(x, 2j ) \ B(x, 2j−1 ) for r/4 ≤ 2j ≤ 2R. As µ is doubling we
can cover each annulus Aj (x) by at most N balls of radii equal to ε2j with N
independent of x, j. Naturally, N depends on ε, and the smaller the ε, the larger
the number N. Consider the collection of all these balls when r/4 ≤ 2j ≤ 2R. When
ε is sufficiently small, depending only on λ, the balls 2λB with B corresponding to
Aj (x) and 2λB 0 with B 0 corresponding to Ai (x) do not intersect provided |i−j| ≥ 2.
The balls B corresponding to the annuli Aj (x) together with B(x, r/2), X \B(x, 2R)
form an open cover of X. As X is connected and contains a point inside B(x, r/2)
and another point outside B(x, 2R), we can pick a chain of these balls B that joins
B(x, r/2) to X \ B(x, 2R). The required chain is then obtained as the collection of
the balls 2B from the balls B different from B(x, r/2) in this chain.

Lemma 6.2 Assume that X satisfies a chain condition and suppose that a pair
u, g satisfies an s-Poincaré inequality for all balls in X. Then the following holds for
almost every x. Let 0 < R < diam (X)/4. There is r and a chain {Bi } corresponding
to x, r, R with λ = σ, such that
k
X µZ ¶1/s
s
|u(x) − uB0 | ≤ C ri g dµ . (43)
i=0 σBi

Proof. As λ is fixed, conditions 1 and 2 of the definition of the chain and the
Lebesgue differentiation theorem (see Theorem 14.15) guarantee that uBk → u(x)
for almost all x when r tends to zero (here k = kr ). For such a point we have for
appropriate r and corresponding k
k
X
|u(x) − uB0 | ≤ 2 |uBi − uBi+1 |
i=0
k
X
≤ 2 (|uBi − uRi | + |uBi+1 − uRi |)
i=0
k µZ
X Z ¶
≤ 2 |u − uBi | dµ + |u − uBi+1 | dµ
i=0 Ri Ri
k Z
X
≤ C |u − uBi | dµ
i=0 Bi
k
X µZ ¶1/s
≤ C ri g s dµ .
i=0 σBi
 SOBOLEV MET POINCARÉ 31

The proof is complete.

Proof of Theorem 6.1. By the discussion preceding the previous lemma we
know that X satisfies a chain condition. Thus we may assume that the pointwise
inequality (43) holds for a given point x. Write r for the radius of the fixed ball B.
We may assume that diam B0 ≥ r/C and that Bi ⊂ 5B, σBi ⊂ 5σB for each i.
Fix q > max{s, s/(s − 1)}. For 0 < ε < q −1 we have that
k
X µZ ¶1/s
s
|u(x) − uB0 | ≤ C ri g dµ
i=0 σBi

X µ Z ¶1/q µZ ¶1/s−1/q
= C ri1−ε µ(σBi )1/q−1/s riqε g s dµ g s dµ .
i σBi σBi

As (s − 1)/s + 1/q + (1/s − 1/q) = 1, we can use Hölder’s inequality to obtain the
estimate
à ! s−1
X³ ´ s s
1−ε 1/q−1/s s−1
|u(x) − uB0 | ≤ C ri µ(σBi )
i
à !1/q
X 1−s/q
× riqε M5σB g s (x) kgkLs (5σB) ;
i
R
here we replaced σBi g s dµ by CM5σB g s (x) and used the bounded overlap of the
balls σBi to replace the sum of the integrals of g s over these balls by the integral
of g s over 5σB.
To estimate the second term in the product, we sum over the balls Bi that
correspond to an annulus Aj ; let us write Ii,j for the set of indices i corresponding
to Aj . By the construction of the chain we know that we have at most N balls
for each fixed j. Moreover the radii of balls corresponding to different Aj form a
geometric sequence and hence
X qε X X qε
ri M5σB g s (x) = M5σB g s (x) ri
i j Ii,j

≤ CM5σB g s (x)rqε (1 − 2−qε )−1

≤ C(qε)−1 rqε M5σB g s (x),

where C is independent of q, ε. In the last inequality we employed the fact that

qε < 1.
For the first term, we use the lower bound on µ(Bi ) and argue as above:
X³ ´s0 ³ ´s0 X
(s/q−ε)s0
ri1−ε µ(σBi )1/q−1/s ≤ C r1−s/q µ(B)1/q−1/s ri
i i
³ ´s0
≤ Cq(s − εq)−1 r1−ε µ(B)1/q−1/s ,
32 PIOTR HAJLASZ AND PEKKA KOSKELA

where s0 = s/(s − 1) and C is an absolute constant.

If we let ε = sq −2 , then q(s − εq)−1 = q(s − s/q)−1 ≤ q, as q ≥ s/(s − 1). Hence

1−s/q 1/q
|u(x) − uB0 | ≤ CkgkLs (5σB) µ(B)1/q−1/s q 1/q+(s−1)/s r (M5σB g s (x)) ;

here C is an absolute constant.

We proceed to estimate the integrals of |u − uB |. By the triangle inequality

|u − uB | ≤ |u − uB0 | + |uB0 − uB | .

By controlling the second term by the Poincaré inequality and using the above
pointwise estimate for the first term we arrive at
Z
q/2−s/2
|u − uB |q/2 dµ ≤ C q q 1/2+(s−1)q/2s µ(B)1/2−q/2s kgkLs (5σB) rq/2
B
Z
1/2 q/2
× (M5σB g s ) dµ + C q rq/2 µ(B)1−q/2s kgkLs (5σB) .
B

By the Maximal Theorem (see Theorem 14.13) and Theorem 14.11 in the appendix
Z µ Z ¶1/2
1/2
(Mσr g s ) dµ ≤ C µ(B) g s dµ
B 5σB

and hence we conclude that

Z µ Z ¶q/2s
|u − uB |q/2 dµ ≤ C q q 1/2+(s−1)q/2s rs g s dµ ,
B 5σB

where C does not depend on q. Notice that this estimate holds as well for q ≤
max{s, s/(s − 1)} by Theorem 5.1.
Now
s/(s−1)
X (t|u(x) − uB |)ks/(s−1)
exp (t|u(x) − uB |) = .
k!
k≥0

Integrating over B and using the above estimate we obtain

Z X³
s/(s−1)
exp (t|u(x) − uB |) dµ ≤ 1+ (Ck)1/2+k (k!)−1
B k>0
¡ ¢ks/(s−1) ´
× trµ(5σB)−1/s kgkLs (5σB) .

This series converges when trµ(5σB)−1/s kgkLs (5σB) ≤ C0 , where C0 depends only
on C, s, and the claim follows.
 SOBOLEV MET POINCARÉ 33

7 A version of the Sobolev embedding theorem on

spheres
In order to state our version of the Sobolev embedding theorem on spheres we first
have to deal with the problem that u is only defined almost everywhere. To take
care of this matter we define u(x) everywhere by the formula
Z
u(x) := lim sup u(z) dµ(z). (44)
r→0 B(x,r)

As almost every point is a Lebesgue point, we have only modified u in a set

of measure zero. This redefinition of u essentially corresponds to picking a repre-
sentative of u with nice continuity properties; for related results see HajÃlasz and
Kinnunen [104], and Kinnunen and Martio [157].
We again assume that X is a doubling space and
µ ¶s
r
µ(B(x, r)) ≥ Cb µ(B0 )
r0
whenever B(x, r) ⊂ B0 = B(x0 , r0 ). Recall that such an estimate follows from the
doubling condition.

Theorem 7.1 Suppose that the pair u, g satisfies a p-Poincaré inequality and that
p > s − 1. Then the restriction of u to the set {x : d(x, x0 ) = r} is uniformly Hölder
continuous with exponent 1 − (s − 1)/p for almost every 0 < r < r0 . In particular,
there is a constant C1 and a radius r0 /2 < r < r0 such that
µZ ¶1/p
(s−1)/p
|u(x) − u(y)| ≤ C1 d(x, y)1−(s−1)/p r0 g p dµ
5σB0

whenever d(x, x0 ) = d(y, x0 ) = r. The constant C1 only depends on p, s, CP , Cb ,

Cd .

In the case of Carnot groups a related result has been independently obtained by
Vodop’yanov, [252].
The usual Sobolev embedding theorem on spheres (cf. [187, Lemma 2.10]) is
based on showing that the trace of a Sobolev function belongs to a Sobolev class
on almost all spheres. One then uses the Sobolev embedding on the sphere that is
lower dimensional than the ball. In our situation a sphere can be very wild and this
approach cannot be used. We prove the above result by using a maximal function
argument.
The reader may wonder why the integration is taken over all of 5σB0 and not
over an annulus. The reason for this is that points on the sphere cannot necessarily
be joined inside an annulus centered at x0 . For a trivial example, let X be the
real axis. If we assume that X has reasonable connectivity properties, we obtain a
stronger conclusion.
34 PIOTR HAJLASZ AND PEKKA KOSKELA

Theorem 7.2 Suppose that the pair u, g satisfies a p-Poincaré inequality and that
p > s − 1. Assume that any pair of points in B0 \ 12 B0 can be joined by a continuum
F in CB0 \ C −1 B0 with diam F ≤ Cd(x, y). Then there is a constant C1 and a
radius r0 /2 < r < r0 such that
ÃZ !1/p
(s−1)/p
|u(x) − u(y)| ≤ C1 d(x, y)1−(s−1)/p r0 p
g dµ
C1 B0 \C1−1 B0

whenever d(x, x0 ) = d(y, x0 ) = r. The constant C1 only depends on p, s, C, CP ,

Cb , Cd .

By combining Proposition 4.5 and Theorem 7.2 we obtain the following corollary
(recall that a p-Poincaré inequality guarantees a q-Poincaré inequality when q > p).

Corollary 7.3 Suppose that C −1 rs ≤ µ(B(x, r)) ≤ Crs with s > 1 for each x and
all r. Let s − 1 < p ≤ s. Assume that for each pair u, g of a continuous function
and its upper gradient we have a p-Poincaré inequality. Then there is a constant
C1 and a radius r0 /2 < r < r0 such that
ÃZ !1/p
(s−1)/p
|u(x) − u(y)| ≤ C1 d(x, y)1−(s−1)/p r0 p
g dµ
C1 B0 \C1−1 B0

whenever the pair u, g satisfies the p-Poincaré inequality and d(x, x0 ) = d(y, x0 ) = r.
The constant C1 only depends on p, s, C, CP .

In the preceding corollary we assumed that s > 1 and that p ≤ s. Both of these
assumptions are necessary. Indeed, the 1-Poincaré inequality holds for the real axis,
but one needs to integrate over balls instead of annuli. The union of the two closed
cones in IRn with a common vertex of Example 4.2 supports a p-Poincaré inequality
for all p > n and µ(B(x, r)) ≈ rn for each x and all r. One again needs to integrate
over balls instead of annuli.
Before proceeding with the proofs of Theorems 7.1 and 7.2 let us discuss one
more application. We say that u is monotone if

sup |u(x) − u(y)| ≤ sup {|u(y) − u(w)| : d(x, y) = d(x, w) = r}

y,w∈B(x,r)

for each ball B(x, r). Suppose that u is monotone, u has an upper gradient in Ls
and the assumptions of the previous corollary hold. Then u is continuous and
µ ¶−1/s
C1 M
|u(x) − u(y)| ≤ C log kgks,B(x,C1 M )
d(x, y)

for all x, y with d(x, y) ≤ M. This estimate is commonly used in the Euclidean
setting. We leave it to the reader to deduce this conclusion from the above corollary.
 SOBOLEV MET POINCARÉ 35

Proof of Theorem 7.1. Fix x, y ∈ X and 0 ≤ α < 1. Set B0 = B(x, d(x, y)), and
define Bi = B(x, 2i d(x, y)) when i ≤ 1 and Bi = B(y, 2−i d(x, y)) when i > 1.
Then, using the Poincaré inequality and the triangle inequality as in the proof
of Theorem 3.2, we have that
∞ Z
X
|u(x) − u(y)| ≤ C |u − uBi | dµ
i=−∞ Bi
∞
X µZ ¶1/p
≤ C ri gp
i=−∞ 2σBi
∞
X µZ ¶1/p
= C ri1−α riα g p

i=−∞ 2σBi
¡ ¢
≤ Cd(x, y)1−α M2σd(x,y),p,α g(x) + M2σd(x,y),p,α g(y) ,

where ÃZ !1/p
α p
MR,p,α g(x) = sup r g dµ .
r<R B(x,r)

Observe that, according to (44), the above inequality holds everywhere (cf. [104]).
Write Gt = {x ∈ B0 : M4σr0 ,p,α g(x) < t}. Then

|u(x) − u(y)| ≤ Ctd(x, y)1−α (45)

for all x, y ∈ Gt .
By the covering Lemma 14.12 and the lower bound on µ(B(x, r)) we have
Z
s−αp
H∞ (B0 \ Gt ) ≤ Ct−p r0s g p dµ .
5σB0
P
γ
Recall that the Hausdorff content H∞ (E), γ ≥ 0, is defined as the infimum of i riγ ,
where the infimum is taken over the set of all countable coverings of the set E by
balls with radii ri .
Define v : B0 → [0, r0 ) by the formula v(x) = d(x, x0 ). Then v is Lipschitz
continuous with constant 1, and hence, for each set E ⊂ B0 ,
s−αp s−αp
H∞ (v(E)) ≤ H∞ (E).

Let α = (s − 1)/p. Then

Z
1
H∞ (v(B0 \ Gt )) ≤ Ct−p r0s g p dµ .
5σB0

This implies that the length of the set v(B0 \ Gt ) ⊂ [0, r0 ) goes to 0 as t goes to ∞.
Now the theorem follows from the observation that for r ∈ [0, r0 ) \ v(B0 \ Gt ) the
36 PIOTR HAJLASZ AND PEKKA KOSKELA

“sphere” {x : d(x, x0 ) = r} is contained in Gt and hence inequality (45) applies.

The proof is complete.
Proof of Theorem 7.2. Join the points x, y by a continuum F in A = CB0 \B0 /C
with diam (F ) ≤ Cd(x, y). Let r = (100Cσ)−1 d(x, y), and consider the collection of
all balls B(w, r) with w ∈ A ∩ B(x, 2Cd(x, y)). As µ is doubling we find a cover of
A ∩ B(x, 2Cd(x, y)) consisting of k of these balls with k depending only on C, Cd , σ.
Pick those balls from this cover that intersect F and order them into a chain. That
is, denoting the balls by Vi , Vi ∩ Vi+1 6= ∅ for i = 1, ..., l − 1, and x ∈ V1 , y ∈ Vl ,
assuming that we have l balls. The claim of Theorem 7.2 follows repeating the proof
of Theorem 7.1 with the following modification: we define Bi = Vi for i = 1, ..., l,
Bi = B(x, 2i (100Cσ)−1 d(x, y)) for i < 0 and Bi = B(y, 2−i+l (100Cσ)−1 d(x, y)) for
i > l. It is helpful to notice here that the balls Bi , for 1 ≤ i ≤ l, have uniformly
bounded overlap as l ≤ k.

8 Rellich-Kondrachov
The classical Rellich–Kondrachov embedding theorem states that, given a bounded
domain Ω ⊂ IRn with smooth boundary, the Sobolev space W 1,p (Ω), 1 ≤ p < ∞,
is compactly embedded into Lq (Ω), where q ≥ 1 is any finite exponent when p ≥ n
and any exponent strictly less that np/(n − p) when p < n. Of course, here, the
Sobolev space W 1,p (Ω) is defined in the classical way.
In the case of Sobolev spaces associated with vector fields, some compact em-
bedding theorems have been obtained by Danielli, [60], Franchi, Serapioni and Serra
Cassano, [85], Garofalo and Lanconelli, [90], Garofalo and Nhieu, [92], Lu, [176],
Manfredini, [188], Rothschild and Stein, [218].
In this section we extend the Rellich–Kondrachov theorem to the setting of
metric spaces. As we will see in Section 11, Sobolev inequalities for vector fields
are special cases of Sobolev inequalities on metric spaces. Hence our result covers
many of the above results. It extends also an earlier result of HajÃlasz and Koskela,
[106], from the Euclidean setting. In the case of Sobolev spaces on metric spaces
introduced by HajÃlasz, [102], a related compactness theorem has been proved inde-
pendently by KaÃlamajska, [147].
Let µ be a Borel measure on X, doubling on Ω. As usual, Ω ⊂ X denotes
an open subset of a metric space. In order to prove the compactness theorem for
Sobolev functions on Ω, we need to assume that a kind of embedding theorem holds
on Ω. Thus, until the end of the section, we make the following assumption:
The open set Ω ⊂ X satisfies µ(Ω) < ∞ and there exist exponents p > 0 and
q > 1 such that every pair u, g which satisfies a p-Poincaré inequality (5) in Ω (with
given constants CP , σ) satisfies also the global Sobolev inequality
µZ ¶ ÃZ µZ ¶ !
1/q 1/p
|u|q dµ ≤C |u| dµ + g p dµ . (46)
Ω Ω Ω
 SOBOLEV MET POINCARÉ 37

Observe that (46) follows from the Sobolev–Poincaré inequality (7).

Theorem 8.1 Let X, Ω, µ, p > 0 and q > 1 be as above. Let {ui , gi } be a

sequence of pairs, all of which satisfy the p-Poincaré inequality (5) in Ω with given
constants CP , σ. If the sequence kui kL1 (Ω) +kgi kLp (Ω) is bounded, then {ui } contains
a subsequence that converges in Lα (Ω) for any 1 ≤ α < q to some u ∈ Lq (Ω).

Proof. Let {ui , gi } be a sequence satisfying the assumptions of the theorem. Since
the sequence {ui } is bounded in Lq (Ω), we can select a subsequence (still denoted
by {ui }) weakly convergent in Lq (Ω) to some u ∈ Lq (Ω). It remains to prove that
this sequence converges to u in the norm of Lα (Ω) for every 1 ≤ α < q.

Lemma 8.2 Let Y be a set equipped with a finite measure ν. Assume that {vi } ⊂
Lq (Y ), 1 < q < ∞, is a bounded sequence. If vi converges in measure to v ∈ Lq (Y ),
then vi converges to v in the norm of Lα (Y ) for every 1 ≤ α < q.

The lemma is a variant of Proposition 14.9. We postpone the proof of the lemma
for a moment and we show how to use it to complete the proof of the theorem.
According to the lemma it remains to prove that the functions ui converge to u in
measure.
Assume that Ωc 6= ∅; otherwise the proof is even simpler. For t > 0 set Ωt =
{x ∈ Ω : dist (x, Ωc ) > t}. Fix ε > 0 and t > 0. For h < t/σ (recall that σ appears
in (5)) and x ∈ Ωt we set
Z Z
uh (x) = u dµ and ui,h (x) = ui dµ.
B(x,h) B(x,h)

We have

µ({x ∈ Ωt : |ui − u| > ε}) ≤ µ({x ∈ Ωt : |ui − ui,h | > ε/3})

+ µ({x ∈ Ωt : |ui,h − uh | > ε/3})
+ µ({x ∈ Ωt : |u − uh | > ε/3})
= Ai,h + Bi,h + Ch .

Note first that

1/p 1/p
|ui (x) − ui,h (x)| ≤ Ch (Mσh gip (x)) ≤ Ch (MΩ gip (x))

by (16) for almost every x ∈ Ωt . Thus the maximal theorem (see Theorem 14.13)
gives µ ¶p Z
³n ³ ε ´p o´ h
p h→0
Ai,h ≤ µ MΩ gi > C ≤C gip dµ −→ 0.
h ε Ω

This convergence is uniform with respect to i as the sequence kgi kLp (Ω) is bounded.
It follows from the definition of the weak convergence in Lq (Ω) that for every x ∈ Ω,
ui,h (x) → uh (x) as i → ∞, so Bi,h → 0 as i → ∞. Finally Ch → 0 by the Lebesgue
38 PIOTR HAJLASZ AND PEKKA KOSKELA

differentiation theorem (see Theorem 14.15). Now it easily follows that ui → u in

measure. Thus the proof is completed provided we prove the lemma.
Proof of the lemma. Fix 1 ≤ α < q. It suffices to prove that every subsequence
of {vi } contains a subsequence convergent to v in Lα (Y ). In what follows all the
subsequences of {vi } will be simply denoted by {vi }. Take an arbitrary subsequence
of {vi }. The convergence in measure implies that this subsequence contains a sub-
sequence which is convergent almost everywhere. Then, by Egorov’s theorem, for
any ε > 0 there exists a measurable set E ⊂ Y with the property that ν(Y \ E) < ε
and vi converges to v uniformly on E. Hence
µZ ¶ ÃZ !1/q
1/α
|vk − vj |α dν ≤ ν(Y \ E)1/α−1/q |vk − vj |q dν
Y Y \E
µZ ¶1/α
+ |vk − vj |α dν
E
µZ ¶1/α
1/α−1/q α
≤ Cε + |vk − vj | dν ,
E

which gives lim supj,k→∞ kvk − vj kLα (Y ) ≤ Cε1/α−1/q . Since ε > 0 was arbitrary,
the subsequence {vi } is a Cauchy sequence in Lα (Y ) and hence the lemma follows.
This completes also the proof of the theorem.
Below we state another version of the compactness theorem. The proof follows
by some obvious modifications to the above proof.

Theorem 8.3 Let X be a doubling space and let s be the lower decay order of the
measure from (21). Suppose that all the pairs ui , gi satisfy a p-Poincaré inequal-
ity in X (with fixed constants CP , σ). Fix a ball B and assume that the sequence
kui kL1 (B) + kgi kLp (5σB) is bounded. Then there is a subsequence of {ui } that con-
verges in Lq (B) for each 1 ≤ q < ps/(s − p), when p < s and for each q ≥ 1 when
p ≥ s.

Notice that this theorem gives compactness in the entire space provided the space
has finite diameter.
We would like to thank Agnieszka KaÃlamajska for an argument that simplified
our original proof of the compactness theorem.

9 Sobolev classes in John domains

In the p-Poincaré inequality (5) we have allowed g to be integrated over a larger
ball than u is integrated over. One cannot, in general, reduce the radii of the
balls on the right hand side. To see this, consider the following example: let Ω =
(0, 1)∪(2, 3)∪(4, 5) and u ≡ 0 on (0, 1), u ≡ 1 on (2, 3)∪(4, 5), g ≡ 0 on (0, 1)∪(2, 3)
and g ≡ const. is very large on (4, 5). The details are left to the reader.
 SOBOLEV MET POINCARÉ 39

Hence in the Sobolev type inequalities like Theorem 5.1 or Theorem 6.1 we have
to integrate g over a larger ball as well.
We show in this section that one can use balls of the same size provided the
geometry of balls is sufficiently nice. This leads us to define John domains.
The first subsection is devoted to study of the geometry of John domains and
in the second subsection we study Sobolev inequalities in John domains.

9.1 John domains

When dealing with Sobolev type inequalities in domains in IRn one usually assumes
that the domain is “nice” in the sense that its boundary is locally a graph of a
Lipschitz function. This notion of being “nice” is not appropriate for the setting
of metric spaces and so one has to define a “nice” domain using only its interior
properties. This leads to John domains.
Definition. A bounded open subset Ω of a metric space is called a John domain
provided it satisfies the following “twisted cone” condition: There exist a distin-
guished point x0 ∈ Ω and a constant C > 0 such that, for every x ∈ Ω, there is a
curve γ : [0, l] → Ω parametrized by the arclength and such that γ(0) = x, γ(l) = x0
and
dist (γ(t), Ωc ) ≥ Ct. (47)
(The length l depends on x.)
Notice that every rectifiable curve in a metric space can be parametrized by
arclength, see Busemann [22], or Väisälä [246].
John domains in IRn were introduced by Martio and Sarvas, [192]. They are
named after F. John who considered similar domains in [137].
The class of John domains in IRn is much larger than the class of domains with
the interior cone condition. In general, the Hausdorff dimension of the boundary of
a John domain can be strictly larger than n − 1.
The above definition of John domain is still not appropriate for many metric
spaces, as points in an arc-wise connected metric space may not be joinable by
rectifiable curves. For example, if Γ ⊂ IR2 is the von Koch snowflake curve and
Ω ⊂ Γ is a nontrivial subcurve, then Ω is not a John domain. However, we would
like to include at least some of the metric spaces that lack rectifiable curves in the
class of John domains; see Example 9.1. The following definition seems to give a
proper generalization of John domains.
Definition. A bounded open subset Ω of a metric space (X, d) is called a weak John
domain provided there is a distinguished point x0 ∈ Ω and a constant 1 ≥ CJ > 0
such that for every x ∈ Ω there exists a curve γ : [0, 1] → Ω such that γ(0) = x,
γ(1) = x0 and for every t ∈ [0, 1]

dist (γ(t), Ωc ) ≥ CJ d(x, γ(t)). (48)

40 PIOTR HAJLASZ AND PEKKA KOSKELA

We call such a curve a weak John curve. If Ωc = ∅, then we set dist (γ(t), Ωc ) = +∞
and hence (48) is always satisfied.
Notice that this definition can be also used in the setting of quasimetric spaces
(i.e., when the triangle inequality is replaced by ρ(x, y) ≤ K(ρ(x, z) + ρ(z, y)),
K ≥ 1). In general, it can happen that even a very nice metric space does not
contain nontrivial rectifiable curves. With the metric ρ(x, y) = |x − y|1/2 on the
real axis, any interval is of infinite length.
Example 9.1 If f : S 2 → S 2 is a quasiconformal mapping and γ ⊂ S 2 is any
smooth Jordan curve, then any connected part of Γ = f (γ) is a weak John domain.
This includes the class of Jordan curves Γ ⊂ S 2 such that both components of S 2 \Γ
are John domains, see Näkki and Väisälä, [205].
There are also many other fractal sets whose “nice” subsets are weak John
domains, while they cannot be John domains because of the lack of rectifiable
curves.
Example 9.2 Let X be a bounded arc-wise connected metric space. If we take
Ω = X, then Ω is weak John domain, since (48) is satisfied for any curve joining x
and x0 .
The reader may find the preceding example somewhat artificial. Let us briefly
clarify the issue. Our aim is to deduce a p-Poincaré inequality for Ω whenever
such an inequality holds for all balls in Ω, that is, B(x, r) ⊂ Ω. If we are given an
underlying space X, then we consider the balls of the space X that are contained in
Ω. Otherwise, the collection of the balls can be fairly large. For example, let Ω be
a bounded domain in IRn so that Ω equipped with the restrictions of the Euclidean
distance and volume is a doubling space. If we neglect IRn , and consider Ω as
our entire doubling space, we shall obtain a Poincaré inequality for Ω, provided we
assume a Poincaré inequality for the pair u, g for all balls of Ω. These balls consist
of the intersections of all Euclidean balls centered in Ω with Ω, see Corollary 9.9.
It is known that in the Euclidean case X = IRn the class of weak John domains
coincides with the class of John domains, see Väisälä, [247, Theorem 2.18]. In a
metric space this is no longer true. Clearly every John domain is a weak John
domain, but the converse implication may fail. However, we generalize the result
by proving that under a mild additional condition on the space X, the above two
definitions are equivalent, see Proposition 9.6.
The crucial property of John domains for us is that they satisfy a chain condition
that is essential in order to effectively use the Poincaré inequality on the balls
contained in the domain.
Let us slightly modify the chain condition we employed in connection with the
Trudinger inequality.
Definition. We say that Ω satisfies a C(λ, M ) condition, where λ, M ≥ 1, if there
is a “central” ball B0 ⊂ Ω such that to every x ∈ Ω and every r > 0 there is a
sequence of balls B0 , B1 , B2 , . . . , Bk , (B0 is fixed and k may depend on x, r) with
the following properties
 SOBOLEV MET POINCARÉ 41

1. λBi ⊂ Ω for i = 0, 1, 2, . . . , k and Bk ⊂ B(x, r),

2. M −1 ri ≤ dist (x, λBi ) ≤ M ri for i = 1, 2, . . .,
3. there is a ball Ri ⊂ Bi ∩ Bi+1 , such that Bi ∪ Bi+1 ⊂ M Ri for i = 0, 1, 2, . . .,
4. no point of Ω belongs to more than M balls λBi .

A variant of the above chain condition was employed in HajÃlasz and Koskela,
[105], [106].

Theorem 9.3 Assume that X is a metric space which is doubling on Ω ⊂ X. If

Ω ⊂ X is a weak John domain, then Ω satisfies the C(λ, M ) condition for any λ ≥ 1
with some M depending on λ, CJ and doubling constant only.

Proof. Assume first that Ω 6= X. Let B0 = B(x0 , dist (x0 , Ωc )/(4λ)). Assume that
x is far away from B0 . Say x ∈ Ω \ 2B0 . Let γ be a weak John curve. First we
define a sequence of balls Bi0 as follows. The centers xi of all balls Bi0 lie on γ.
Let B00 = 21 B0 . Assume that we have already defined Bi0 . Then we trace along γ,
starting from the center xi towards x, until we leave Bi0 for the last time. We let
0
this point be the center xi+1 of Bi+1 and we define

0 CJ
ri+1 = d(x, xi+1 ) . (49)
4λ
Now we define Bi = 2Bi0 . Properties 1. and 2. are evidently satisfied provided
we choose k large enough. Property 3. follows from the fact that consecutive balls
0
have comparable radii and xi+1 lies on the boundary of Bi+1 (ball Ri is centered
0 0
at xi+1 and of radius equal to minimum of ri , ri+1 ). For property 4. assume that
y ∈ λBi1 ∩ λBi2 ∩ . . . ∩ λBil . It follows from the construction that the radii of
the balls λBij , j = 1, 2, . . . , l and the distances between centers of the balls are all
comparable to d(x, y). Indeed, the radii are comparable and the distance of the
centers times 2λ are no less than the radii. Thus there exists a constant t > 0
such that tBi1 , tBi2 ,. . . , tBil are pairwise disjoint and all these balls are contained
in a ball centered at y and having radius comparable to d(x, y), so the doubling
condition implies the upper bound for l.
The case when x ∈ 2B0 is similar. If x ∈ Ω \ B0 , we define B1 = 14 B0 and
the rest of the argument goes as above. Otherwise, we consider the union of two
curves: the weak John curve for x and the weak John curve for some point y with
d(y, x0 ) = 12 dist (x0 , Ωc ). This curve, traced first from x to x0 and then from x0 to
y, is easily seen to be a John curve with a new distinguished point y and a new
John constant only depending on CJ . One can then define desired chain by first
replacing the role of x0 in the above argument by y and then adding a chain joining
y to x0 . We leave the details to the reader.
Suppose finally that Ω = X. Then X is bounded. We fix x0 ∈ Ω and define
B0 = B(x0 , diam (X)/4). The rest of the argument is the same as in the case
X 6= Ω except that in (49) we replace CJ by 1. The proof is complete.
42 PIOTR HAJLASZ AND PEKKA KOSKELA

Lemma 9.4 Let (X, d) be an arcwise connected metric space such that bounded and
closed sets in X are compact. Assume that the metric d has the property that for
every two points a, b ∈ X the distance d(a, b) is equal to the infimum of the lengths
of curves that join a and b. Then there exists a shortest curve γ from a to b with
d(a, b) = d(a, z) + d(z, b), for every z ∈ γ.

Remark. Observe that the compactness of the sets that are bounded and closed is
a stronger assumption than the space being locally compact. The two conditions
are equivalent if the space is complete.
This lemma is due to Busemann, [22, page 25], (cf. [76, page 592]). The idea of
the proof is the following. Let {γk }∞k=1 be a sequence of rectifiable curves joining
a with b and such that the length of γk converges to d(a, b). Parametrize each
γk by arclength. Scaling the arclength parametrizations we may assume that all
curves are defined on the interval [0, 1]. Now it easily follows that the family {γk }
is equicontinuous (because of the good parametrization). By a standard diagonal
method we can find a subsequence of {γk } which converges on a dense subset of
[0, 1]. The equicontinuity implies the uniform convergence on the whole interval. It
is easy to prove that the length of the limiting curve is d(a, b). For a detailed proof,
see [22, page 25].

Corollary 9.5 Let the metric space (X, d) satisfy the assumptions of the above
lemma. Then each ball B ⊂ X is a John domain with a universal constant C.

This corollary shows that balls in a Carnot–Carathéodory metric (see Section 11)
are John domains.
The chain condition is closely connected to the concept of a John domain as
the following proposition shows. Analogs of this result can be found in Buckley,
Koskela and Lu, [20], and in Garofalo and Nhieu, [92], where the authors employ a
Boman chain condition that is different from ours.

Proposition 9.6 Let X be a metric space which is doubling on Ω ⊂ X. Assume

that Ω has the following local connectivity property: there exists a constant δ ≥ 1
such that for every ball B with δB ⊂ Ω, every two points x, y ∈ B can be connected
by a rectifiable curve contained in δB and of length less than or equal to δd(x, y).
Then the following three conditions are equivalent:

1. Ω is a John domain,

2. Ω is a weak John domain,

3. Ω is a C(λ, M )-domain for each λ ≥ 2δ and for some M.

Proof. The implications 1 ⇒ 2 ⇒ 3 hold without any local connectivity assumptions

on Ω: the first implication is immediate from the definitions and the second one
follows from Theorem 9.3. We prove the implication 3 ⇒ 1. Fix x ∈ Ω and
 SOBOLEV MET POINCARÉ 43

let {Bi }k0 be an associated chain, for λ = 2δ and r = d(x, Ωc )/λ. We define
Bk+1 = B(x, d(x, Ωc )/2). If the radii of the balls Bi were to increase geometrically
when i decreases we would obtain the John curve simply by joining the centers
of the balls in our chain by curves obtained from the local connectivity condition.
However, this need not be the case. This difficulty can be rectified as follows.
If the entire chain is contained in B(x, Cd(x, Ωc )), the bounded overlap condition
for the chain and the doubling property imply that the length of the subchain (i.e.,
the number of balls) joining the boundary of Bk+1 to x0 does not exceed a uniform
constant that depends on the doubling constant, the constants in the chain condition
and on C. The above connect-the-dots argument applies in this case. Otherwise,
we consider the subchain joining Bk+1 to ∂B(x, Cd(x, Ωc )). Again, the length of
this subchain is bounded by a uniform constant and the radii are bounded from
below by a multiple of d(x, Ωc ). Pick a point y1 with d(y1 , x) = Cd(x, Ωc ) that is
contained in one of the balls of the subchain. If the constant C is sufficiently large,
then d(y1 , Ωc ) ≥ 2d(x, Ωc ). Consider a chain joining y1 to x0 . We now repeat the
above argument for the subchain joining y1 to ∂B(x, C 2 d(x, Ωc )). By continuing
inductively we obtain a new chain with the appropriate geometric behavior. We
leave it to the reader to provide the details.

9.2 Sobolev type inequalities

In the main theorem of this section (Theorem 9.7) we show how the claims of
Theorem 5.1 and Theorem 6.1 extend to the setting of John domains.
The study of the Sobolev type inequalities in John domains in IRn originated in
the papers of Boman, [13], Bojarski, [12], Goldshtein and Reshetnyak, [97], Hurri,
[130], Iwaniec and Nolder, [132], Kohn, [158], and Martio, [191]. It was then gen-
eralized to the Carnot–Carathéodory spaces by Jerison, [133], and then to more
general situations by Franchi, Gutiérrez and Wheeden, [76], Garofalo and Nhieu,
[92], HajÃlasz and Koskela, [105], Lu [175], [178]. Other related references include
Buckley and Koskela, [16], [17], [18], Buckley, Koskela and Lu, [19], [20], Chen and
Li, [37], Chua, [42], Hurri-Syrjänen, [131], HajÃlasz and Koskela, [106], Franchi, [74],
Lu, [178], Maheux and Saloff-Coste, [186], Saloff-Coste, [221], Smith and Stegenga,
[231], [232].
Buckley and Koskela, [17], [18], showed that the class of John domains is nearly
the largest one for which one can prove the Sobolev-Poincaré embedding theorem.

Theorem 9.7 Let X be a metric space equipped with a measure which is doubling
on a weak John domain Ω ⊂ X. Assume that the measure µ satisfies the condition
³ r ´s
µ(B(x, r)) ≥ Cb µ(Ω) ,
diam Ω
whenever x ∈ Ω and r ≤ diam Ω. If the pair u, g satisfies a p-Poincaré inequality
(5), p > 0, in Ω, then all the claims of Theorem 5.1 hold with B and 5σB replaced
44 PIOTR HAJLASZ AND PEKKA KOSKELA

by Ω. For example, we get that if 0 < p < s and the pair u, g has the truncation
property, then
µZ ¶1/p∗ µZ ¶1/p
p∗ p
inf |u − c| dµ ≤ C(diam Ω) g dµ , (50)
c∈IR Ω Ω
∗
where p = sp/(s − p).
If in addition the space is connected and p = s > 1, then the Trudinger inequality
holds in Ω, i.e.,
Z µ ¶s/(s−1)
C1 |Ω|1/s |u − uΩ |
exp dµ ≤ C2 .
Ω diam Ω kgkLs (Ω)
The constants C, C1 , C2 depend on p, s, σ, CP , Cd , Cb and CJ only.

Remarks. 1) As follows from the proof, the above theorem actually holds for any
open set that satisfies the C(σ, M ) condition. 2) We have stated explicitly only a
generalization of one part of Theorem 5.1, inequality (50). It is left to the reader
to formulate generalizations of the other cases. 3) Observe that |u − uΩ | is replaced
by |u − c| and infimum over c ∈ IR is taken. This is necessary if p∗ < 1, as then we
cannot apply inequality (9).
Proof. By Theorem 9.3 the domain Ω satisfies the chain condition for any given
λ = σ. Thus we obtain inequality (43) with balls Bi as in the chain condition; in
particular with σBi ⊂ B. The proofs of Theorem 5.3 and Theorem 6.1 give the
claim.

Corollary 9.8 Let X be a doubling space satisfying the assumptions of Lemma 9.4.
Suppose that the measure µ satisfies condition (21). Then all the claims of Theo-
rem 5.1 hold with the integrals of g taken over B instead of 5σB. If, in addition,
the space is connected and s > 1, then the Trudinger inequality (42) holds with the
integral of g taken over B.

Remark. This corollary applies to the Carnot–Carathéodory spaces, see Proposi-

tion 11.5.
Proof. By Corollary 9.5 every ball is a John domain with a universal constant
C and hence we may apply Theorem 9.7. The proof is complete.
We have already mentioned that any bounded arc-wise connected set Ω = X is
a weak John domain. To illustrate this issue we state the following special case of
the above results.

Corollary 9.9 Let Ω ⊂ IRn be an arbitrary bounded domain. Assume that

|B(x, r) ∩ Ω| ≥ Crn , whenever x ∈ Ω and r ≤ diam Ω. Assume that u ∈ W 1,p (Ω),
1 ≤ p < n, satisfies
Z Z
|u − uB | dx ≤ Cr |∇u| dx ,
Ω∩B 2B∩Ω
 SOBOLEV MET POINCARÉ 45

whenever B = B(x, r), x ∈ Ω and r ≤ diam Ω. Then the global Sobolev inequality
µZ ¶1/p∗ µZ ¶1/p
∗
|u − uΩ |p dx ≤ C(diam Ω) |∇u|p dx
Ω Ω

∗
holds, where p = np/(n − p).

Proof. Take X = Ω. The condition |B(x, r) ∩ Ω| ≥ Crn means that the space X
equipped with the Lebesgue measure and the Euclidean metric is doubling. Since
X = Ω, we conclude that Ω is a weak John domain and hence the claim follows
from Theorem 9.7. The proof is complete.
As the last application of the chain method we improve the so-called representa-
tion formula (33). The result below is a generalization and simplification of earlier
results due to Capogna, Danielli and Garofalo, [30], Franchi, Lu and Wheeden, [80],
[81], and Franchi and Wheeden, [86].

Theorem 9.10 Assume that X is a metric space which is doubling on a weak John
domain Ω ⊂ X. If the pair u, g satisfies a p-Poincaré inequality (5), p > 0, in Ω,
then for almost every x ∈ Ω we have the inequality

∞
ÃZ !1/p
X g(y)p d(x, y)p
|u(x) − uB0 | ≤ C dµ(y) ,
j=−∞ Aj (x)∩Ω µ(B(x, d(x, y)))

where Aj (x) = B(x, 2j ) \ B(x, 2j−1 ) and B0 = B(x0 , dist (x0 , Ωc )/(4σ)), x0 ∈ Ω, is
a fixed ball.

In particular when p = 1 we get

Z
g(y)d(x, y)
|u(x) − uB0 | ≤ C dµ(y) .
Ω µ(B(x, d(x, y)))

Proof of Theorem 9.10. By Theorem 9.3 the domain satisfies the C(λ, M ) condition
with λ = σ. Then we have
k
X
|u(x) − uB0 | ≤ 2 |uBi − uBi+1 |
i=0
k
X µZ ¶1/p
≤ C ri g p dµ .
i=0 σBi

Each ball σBi is covered by a finite number, say no more than l, of the annuli Aj (x).
Hence if σBi ∩ Aj (x) 6= ∅ we get
µZ ¶1/p j+l
ÃZ !1/p
X g p (y)d(x, y)p
p
rj g dµ ≤C dµ(y) .
σBi Aν (x)∩Ω µ(B(x, d(x, y)))
ν=j−l
46 PIOTR HAJLASZ AND PEKKA KOSKELA

Now observe that the doubling condition and the bounded overlapping of the balls
σBi implies that the number of balls σBi with σBi ∩ Aj (x) 6= ∅ is bounded by a
constant not depending on j. This easily implies the claim.
There are several related results when p = 1. For the Euclidean case see Gold-
shtein and Reshetnyak [97], Martio [191] and HajÃlasz and Koskela [106]; for the
Carnot–Carathéodory case see Franchi, Lu and Wheeden [80], Capogna, Danielli
and Garofalo [30], and for the case of doubling spaces see Franchi, Lu and Wheeden
[81], Franchi and Wheeden [86].

10 Poincaré inequality: examples

The purpose of this section is to illustrate the results obtained up to now in the
paper: we collect basic examples of pairs that satisfy p-Poincaré inequalities.
We will pay particular attention to the validity of the truncation property. Recall
that this property is used to prove the Sobolev embedding in the borderline case
with the sharp exponent.
Two classes of examples, Carnot–Carathéodory spaces and graphs, require a
longer presentation, and so we discuss them in Sections 11 and 12.

10.1 Riemannian manifolds.

The pair u, |∇u|, where u ∈ Lip (IRn ), satisfies the 1-Poincaré inequality and hence
all the p-Poincaré inequalities for 1 ≤ p < ∞. Obviously the pair u, |∇u| also has
the truncation property.
This result extends to those Riemannian manifolds whose Ricci curvature is
bounded from below. Let M be a complete Riemannian manifold of dimension n,
and let g denote the Riemannian metric tensor. Denote the canonical measure on
M by µ. Assume that the Ricci curvature is bounded from below i.e. Ric ≥ −Kg
for some K ≥ 0. Then the Bishop–Gromov comparison theorem implies that
p
µ(B(x, 2r)) ≤ 2n exp( (n − 1)K2r)µ(B(x, r)),

see Cheeger, Gromov and Taylor [35]. Moreover Buser’s inequality, [23], implies
that Z Z
√
|u − uB | dµ ≤ C(n) exp( Kr)r |∇u| dµ.
B B

This shows that for any R > 0 both the doubling property and the 1-Poincaré
inequality hold on all balls with radii less than R. If we assume that the Ricci
curvature is nonnegative (i.e. K = 0), then we can take R = ∞. Obviously,
the pair of a Lipschitz function and the length of its gradient has the truncation
property in this setting as well.
 SOBOLEV MET POINCARÉ 47

Thus the results of our paper imply that in the above setting, the Sobolev-
Poincaré inequality holds, see Maheux and Saloff-Coste [186] and also HajÃlasz and
Koskela [105].
An excellent introduction to the Buser inequality and the comparison theorems
can be found in Chavel’s book, [33].
For related and earlier works on Poincaré and Sobolev inequalities on manifolds
with a bound on the Ricci curvature see Chen and Li [37], Gallot [88], Kusuoka and
Stroock [163], Li and Schoen [170], Li and Yau [172], Saloff-Coste [220], [221].

10.2 Upper gradients.

Let (X, d, µ) be a metric space with a Borel measure, not necessarily doubling.
Definition. We say that a Borel function g : Ω → [0, ∞] is an upper gradient on
Ω of another Borel function u : Ω → IR, if for every 1-Lipschitz curve γ : [a, b] → Ω
we have Z b
|u(γ(b)) − u(γ(a))| ≤ g(γ(t)) dt. (51)
a
Note that g ≡ ∞ is an upper gradient of any Borel function u.
Definition. We say that the space supports a p-Poincaré inequality inequality on
Ω, 1 ≤ p < ∞, if every pair u, g of a continuous function u and its upper gradient g
on Ω satisfies the p-Poincaré inequality (5) in Ω (with some fixed constants CP > 0,
σ ≥ 1).
If we say that the space supports a p-Poincaré inequality, then we mean that
above Ω = X.
Since every rectifiable curve admits an arc-length parametrization that makes
the curve 1-Lipschitz, the class of 1-Lipschitz curves coincides with the class of
rectifiable curves, modulo a parameter change.
It is necessary to assume that the function g is defined everywhere, as we require
the condition (51) for all rectifiable curves. We refer the reader to Busemann, [22],
or Väisälä, [246, Chapter 1], for more information on integration over rectifiable
curves.
The notions of an upper gradient and a space supporting a p-Poincaré inequal-
ity were introduced by Heinonen and Koskela, [118], and then applied and further
developed by Bourdon and Pajot, [15], Cheeger, [34], Franchi, HajÃlasz and Koskela,
[77], Heinonen and Koskela, [119], Kallunki and Shanmugalingam, [148], Koskela
and MacManus, [161], Laakso, [164], Semmes, [226], Tyson, [245], and Shanmu-
galingam, [228].
Notice that above we required the p-Poincaré inequality for continuous functions
and their upper gradients. If X is sufficiently nice, say quasiconvex and each closed
ball in X is compact, then the p-Poincaré inequality follows for any measurable
function and its upper gradient. In fact, it would in such a setting suffice to assume
48 PIOTR HAJLASZ AND PEKKA KOSKELA

the p-Poincaré inequality for Lipschitz functions and their upper gradients. For this
see [119].

Proposition 10.1 If u is a Lipschitz function on a Riemannian manifold M , then

any measurable function g such that g ≥ |∇u| everywhere is an upper gradient of
u. On the other hand, if g ∈ Lp (M ) is an upper gradient of u ∈ Lp (M ), then
u ∈ W 1,p (M ) and g ≥ |∇u| almost everywhere.

Proof. The first part of the proposition is immediate; the second one follows from
the ACL characterization of the Sobolev space, see for example Ziemer [263, Theo-
rem 2.1.4].
Remark. It is not true, in general, that any upper gradient g of u ∈ C ∞ (M )
satisfies g ≥ |∇u| a.e., unless we assume that g is locally integrable. As an example
take u(x) ≡ x on [0, 1]. Let E ⊂ [0, 1] be a Cantor set with positive length and set
g(x) = 0 if x ∈ E, g(x) = ∞ if x 6∈ E. One can then improve this example and
even obtain g < ∞ everywhere.

Proposition 10.2 If u is a locally Lipschitz function defined on an open subset of

a metric space X, then the function |∇+ u|(x) = lim supy→x |u(x) − u(y)|/d(x, y) is
an upper gradient of u.

Remark. The proposition is no longer true if we only assume that u is continuous.

Indeed, if u is the familiar non-decreasing continuous function u : [0, 1] → [0, 1]
such that u(0) = 0, u(1) = 1 and u is constant on connected components of the
complement of a Cantor set, then |∇+ u(x)| = 0 a.e. in [0, 1].
Proof of the proposition. Let γ : [a, b] → Ω be 1-Lipschitz. The function u ◦
γ is Lipschitz and hence differentiable a.e. It easily follows that |(u ◦ γ)0 (t)| ≤
|∇+ u(γ(t))| whenever u ◦ γ is differentiable at t. Hence
Z b Z b
0
|u(γ(b)) − u(γ(a))| ≤ |(u ◦ γ) (t)| dt ≤ |∇+ u(γ(t))| dt.
a a

The proof is complete.

Theorem 10.3 Assume that the space X supports a p-Poincaré inequality on Ω.

Then any pair u, g of a continuous function and its upper gradient on Ω has the
truncation property.

Proof. Let g be an upper gradient of a continuous function u. We have to prove a

family of p-Poincaré inequalities for all the pairs vtt12 , gχ{t1 <v≤t2 } , where v = ε(u−b)
(see the definition of the truncation property). Since g is an upper gradient of each
of the functions v, we can assume that u = v. Thus it remains to prove the inequality
Z µZ ¶1/p
|utt21 − utt21 B | dµ ≤ Cr g p χ{t1 <u≤t2 } dµ . (52)
B σB
 SOBOLEV MET POINCARÉ 49

The following lemma is due to Semmes, [226, Lemma C.19]. For reader’s con-
venience we recall the proof.

Lemma 10.4 Let g be an upper gradient of a continuous function u. Let 0 < t1 <
t2 < ∞, and let V be an arbitrary open set such that {t1 ≤ u ≤ t2 } ⊂ V . Then gχV
is an upper gradient of utt21 .

Proof. Let γ be a curve as in the definition of the upper gradient. We have to prove
the analog of (51) for utt21 and gχV . If either γ is contained in V or γ is contained in
X \ {t1 ≤ u ≤ t2 }, the claim is very easy. In the general case the curve γ splits into
a finite number of parts, each of which is contained in V or in X \ {t1 ≤ u ≤ t2 }
and the lemma follows by applying the preceding special cases to those pieces of γ.
Now we can complete the proof of the theorem. Take t1 < s1 < s2 < t2 . Then
{s1 ≤ u ≤ s2 } ⊂ V , where V = {t1 < u < t2 }. Applying the p-Poincaré inequality
to the pair uss21 , gχV and passing to the limit as s1 & t1 , s2 % t2 we obtain the
desired inequality. This completes the proof.
Theorem 10.3 is interesting provided we can find sufficiently many examples
of non-smooth metric spaces that support a p-Poincaré inequality. The rest of
Section 10 is devoted to the discussion of such examples.

10.3 Topological manifolds.

Definition. A metric space X is called Q-regular, Q > 0, if it is a complete metric

space and there is a measure µ on X so that C1 rQ ≤ µ(B(x, r)) ≤ C2 rQ whenever
x ∈ X and r ≤ diam X.
It is well known that one can replace µ in the above definition by the Q-
dimensional Hausdorff measure, see for example [226, Lemma C.3].
Semmes, [226], proved a p-Poincaré inequality on a large class of Q-regular
metric spaces including some topological manifolds.
The following result is a direct consequence of a more general result of Semmes,
[226].

Theorem 10.5 Let X be a connected Q-regular metric space that is also an ori-
entable topological Q-dimensional manifold, Q ≥ 2, integer. Assume that X satisfies
the local linear contractibility condition: there is C ≥ 1 so that, for each x ∈ X and
R ≤ C −1 diam X, the ball B(x, R) can be contracted to a point inside B(x, CR).
Then the space supports a 1-Poincaré inequality.

For related inequalities also see, David and Semmes [62], and Semmes [227].
50 PIOTR HAJLASZ AND PEKKA KOSKELA

10.4 Gluing and related constructions.

Heinonen and Koskela, [118, Theorem 6.15], proved that gluing two spaces that
support a p-Poincaré inequality along a sufficiently large common part results in a
new space that also supports a p-Poincaré inequality. For example, one can glue two
copies of the unit ball of IR2 along the usual 13 -Cantor set and the resulting doubling
space supports a p-Poincaré inequality for all p > 2 − log 2
log 3 . This procedure allows
one to build plethora of examples. Hanson and Heinonen, [111], used this type of a
construction recently to build a space that supports the 1-Poincaré inequality but
has no manifold points.
Laakso, [164], constructed recently for each Q > 1 a Q-regular metric space
that supports the 1-Poincaré inequality. Notice that here Q need not be an integer.
These spaces do not admit bi-Lipschitz imbeddings into any Euclidean space. They
are obtained as quotients by finite to one maps of products of intervals with Cantor
sets.
The first authors to find non-integer dimensional Q-regular spaces that support
a Poincaré inequality were Bourdon and Pajot, [15]. Their examples are boundaries
of certain hyperbolic buildings.

10.5 Further examples.

A huge class of examples of spaces that support p-Poincaré inequalities is con-

tained in the class of so-called Carnot–Carathéodory spaces that are discussed in
Section 11. This class includes the Carnot groups that have been mentioned above.
One can also investigate p-Poincaré inequalities on graphs, see Section 12. Here
the situation is however different. Since the space is disconnected, the notion of an
upper gradient is absurd. Moreover, the truncation property does not hold and it
has to be modified.
There are also many other examples that will not be discussed in the paper. The
main class of such examples is given by Poincaré inequalities on Dirichlet spaces.
Roughly speaking we are given a pair u, g satisfying a p-Poincaré inequality on a
doubling space and in addition g is related to u in terms of a Dirichlet form, see
Biroli and Mosco [8], [9], [10], Jost [143], [144], [145], Sturm [236], [237], [238], [239].
Thus this example fits precisely into the setting of our paper. However, the presence
of the Dirichlet form gives additional structure that may lead to results not under
the scope of our more general setting.
The analysis of Dirichlet forms is closely related to the analysis on fractals and
especially with the spectral theory of Laplace operators, see, e.g., Barlow and Bass
[5], Jonsson [139], Kozlov [162], Kigami [151], [152], [153], Kigami and Lapidus
[154], Lapidus [167], [168], Metz and Sturm [197], Mosco [201]. As it follows from
recent works of Jonsson, [139], and Jonsson and Wallin, [141], the spectral theory of
the Laplace operators on fractals is also related to the theory of function spaces on
 SOBOLEV MET POINCARÉ 51

fractal subsets of IRn developed by Jonsson and Wallin, [138], [140], see also Triebel
[244]. Some connections with the theory presented in this paper seem evident, but
a better understanding of those connections is still lacking.

11 Carnot–Carathéodory spaces
In this section we give an introduction to the analysis of vector fields — one of the
main areas where the theory of Sobolev spaces on metric spaces is applicable.
In the first subsection we define the so called Carnot–Carathéodory metric as-
sociated with a family of vector fields X = (X1 , . . . , Xk ). Then, in the second
subsection, we prove that with respect to this metric the “gradient” |Xu| associ-
ated with the given family of vector fields becomes the smallest upper gradient of
u. We also deal with Poincaré inequalities and Sobolev spaces associated with the
given system of vector fields.
The main novelty in our approach is that we develop the analysis on Carnot–
Carathéodory spaces from the point of view of upper gradients. The prime results
of the section are Theorem 11.7 and Theorem 11.12.
In the last three subsections we consider Carnot groups and vector fields satis-
fying Hörmander’s condition — both are examples where pairs u, |Xu| satisfy such
a Poincaré inequality. We also discuss some other classes of vector fields that do
not satisfy Hörmander’s condition, but still support Poincaré inequalities.

11.1 Carnot–Carathéodory metric.

Let Ω ⊂ IRn be an open and connected set and let X1 , X2 ,. . . ,Xk be vector fields
defined in Ω, with real locally Lipschitz continuous coefficients. We identify the Xj ’s
with the first order differential operators that act on u ∈ Lip (Ω) by the formula

Xj u(x) = hXj (x), ∇u(x)i, j = 1, 2, . . . , k.

We set Xu = (X1 u, . . . , Xk u), and hence

 1/2
Xk
|Xu(x)| =  |Xj u(x)|2  .
j=1

With such a family of vector fieldsP one can associate a suitable degenerate elliptic
k
operator, like for example L = − j=1 Xj∗ Xj , where Xj∗ is the formal adjoint of
R R
Xj in L2 i.e., Xj u v = uXj∗ v for all u, v ∈ C0∞ . Both the Poincaré and Sobolev
inequalities for the pair u, |Xu| are then crucial for the Harnack inequality for
positive solutions to Lu = 0 via Moser’s iteration. Since the Poincaré inequality
implies the Sobolev inequality, one needs only check the validity of a Poincaré
inequality. Of course this requires strong restrictions on the class of vector fields.
52 PIOTR HAJLASZ AND PEKKA KOSKELA

Even if one considers degenerate elliptic equations of the divergence form

Lu(x) = div (A(x)∇u(x)) (53)

with a symmetric, nonnegative semi-definite matrix A with smooth coefficients, it is,

in general, necessary to deal with vectorP
fields that have only Lipschitz coefficients
as they arise in the factorization L = − j Xj∗ Xj , see Oleinik and Radkevic [208].
For more applications to PDE and references, see Section 13.
How does one prove a Poincaré inequality for the pair u, |Xu|? The natural
approach is to bound u by integrals of |Xu| along curves and then average the
resulting one-dimensional integrals to obtain the desired Poincaré inequality.
In order to have such bounds for u in terms of integrals of |Xu|, one would like
to know that |Xu| is an upper gradient of u. Unfortunately this is rarely the case.
For example, if we have only the single vector field X1 = ∂/∂x1 and γ(t) = (0, t),
u(x1 , x2 ) = x2 in IR2 , then |u(γ(1)) − u(γ(0))| = 1, while |Xu| ≡ 0, and so |Xu|
is not an upper gradient of u. It is not an upper gradient even up to a constant
factor. Roughly speaking, the problem is caused by the fact that γ̇ is not spanned
by the Xj ’s.
There is a brilliant idea that allows one to avoid this problem by introducing a
new metric (that is described below) in Ω that makes |Xu| an upper gradient of u
on a new metric space. The metric is such that it restricts the class of 1-Lipschitz
curves to those for which γ̇ is a linear combination of the Xj ’s. To be more precise,
it is not always a metric as it allows the distance to be infinite.
We say that an absolutely continuous curve γ : [a, b] → Ω is admissible if there
Pk
exist measurable functions cj (t), a ≤ t ≤ b, satisfying j=1 cj (t)2 ≤ 1 and γ̇(t) =
Pk
j=1 cj (t)Xj (γ(t)).

Note that if the vector fields are not linearly independent at a point, then the
coefficients cj are not unique.
Then we define the distance ρ(x, y) between x, y ∈ Ω as the infimum of those
T > 0 such that there exists an admissible curve γ : [0, T ] → Ω with γ(0) = x and
γ(T ) = y.
If there is no admissible curve that joins x and y, then we set ρ(x, y) = ∞.
Note
S that the space (Ω, ρ) splits into a (possibly infinite) family of metric spaces
Ω = i∈I Ai , where x, y ∈ Ai if and only if x and y can be connected by an
admissible curve in Ai . Obviously (Ai , ρ) is a metric space and the distance between
distinct Ai ’s is infinite.
If we only have the single vector field X1 = ∂/∂x1 in IR2 , then ρ(x, y) = |x − y|
if x and y lie on a line parallel to the x1 axis; otherwise ρ(x, y) = ∞. On the other
hand, if Xj = ∂/∂xj for j = 1, 2, . . . , n in IRn , then ρ is the Euclidean metric.
The distance function ρ is given many names in the literature. We will use
the name Carnot–Carathéodory distance. A space equipped with the Carnot–
 SOBOLEV MET POINCARÉ 53

Carathéodory distance is called a Carnot–Carathéodory space.

There are several other equivalent definitions for the Carnot–Carathéodory dis-
tance, see, e.g., Jerison and Sanchez-Calle [136] and Nagel, Stein and Wainger
[204]. The Carnot–Carathéodory distance can also be defined for Dirichlet forms,
see Sturm [238].
It has already been mentioned that Lipschitz vector fields arise in connection
with degenerate elliptic equations of the divergence form (53). It seems that Feffer-
man and Phong, [69], where the first to realize that many important properties of the
operator can be read off from the properties of the associated Carnot–Carathéodory
metric. Roughly speaking, they proved that, locally, subellipticity of (53) is equiv-
alent to the estimate ρ(x, y) ≤ C|x − y|ε for some ε > 0.
Other connections with degenerate elliptic equations will be discussed later on
in Section 13. We want to emphasize that the scope of applications of the Carnot–
Carathéodory geometry goes far beyond degenerate elliptic equations and it includes
control theory, CR geometry, and more recently quasiconformal mappings. We refer
the reader to the collection [240] of papers for a comprehensive introduction to the
Carnot–Carathéodory geometry. Other important references include: Franchi, [74],
Franchi and Lanconelli, [78], Garofalo and Nhieu, [92], [91], Gole and Karidi, [98],
Karidi, [150], Liu and Sussman, [174], Nagel, Stein and Wainger, [204], Pansu, [210],
Saloff-Coste, [221], [223], Strichartz, [235], Varopoulos, Saloff-Coste and Coulhon,
[251], to name a few.

Lemma 11.1 Let B(x, R) ⊂⊂ Ω and let supB(x,R) |X| = M . If γ : [0, T ] → Ω,

T < R/M , is an admissible curve with γ(0) = x, then γ([0, T ]) ⊂ B(x, R).

Proof. Assume by contradiction that the image of η is not contained in the ball
B(x, R). Then there is the smallest t0 ∈ (0, T ] such that |x − γ(t0 )| = R. Note that
by the Schwartz inequality, |γ̇(t)| ≤ |X(γ(t))|. Hence
¯Z t 0 ¯ Z t0
¯ ¯
R = |x − γ(t0 )| = ¯¯ γ̇(t) dt¯¯ ≤ |X(γ(t))| dt ≤ M T,
0 0

which contradicts the assumption T < R/M . The proof is complete.

As as a corollary we obtain the following well known result.

Proposition 11.2 Let G ⊂⊂ Ω. Then there is a constant C > 0 such that

ρ(x, y) ≥ C|x − y|,

for all x, y ∈ G.

Proof. Let x, y ∈ G and let γ : [0, T ] → Ω, γ(0) = x, γ(T ) = y, be any admissible

curve. Fix ε > 0 such that Gε = {x ∈ IRn : dist (x, G) < ε} ⊂⊂ Ω and set
M = supGε |X|. Obviously B(x, R) ⊂ Gε , when R = min{|x − y|, ε}, and hence
54 PIOTR HAJLASZ AND PEKKA KOSKELA

Lemma 11.1 implies that T ≥ R/M ≥ min{M −1 , ε(M diam G)−1 }|x − y|. This
completes the proof.
If ρ(x, y) < ∞ for all x, y ∈ Ω, then ρ is a true metric (called the Carnot–
Carathéodory metric). Proposition 11.2 implies that id : (Ω, ρ) → (Ω, | · |) is contin-
uous. However, it need not be a homeomorphism as the simple example of the two
vector fields ∂x and x+ ∂y in IR2 shows.
In order to avoid such pathological situations, it is often assumed in the literature
that
id : (Ω, ρ) → (Ω, | · |) is a homeomorphism. (54)
Fortunately, (54) is true for a large class of vector fields satisfying the so-called
Hörmander condition which includes Carnot groups (see the following subsections),
and the case of Grushin type vector fields like those in Franchi, [74], Franchi,
Gutiérrez and Wheeden, [76], and Franchi and Lanconelli, [78].
To keep the generality we do not assume (54) unless it is explicitly stated.
By a Lipschitz function on Ω we mean Lipschitz continuity with respect to the
Euclidean metric in Ω, but when we say that a function is Lipschitz on (Ω, ρ) we
mean Lipschitz with respect to the distance ρ. The same convention extends to
functions with values in Ω or in (Ω, ρ). Functions that are Lipschitz with respect to
ρ will be also called metric Lipschitz. Balls with respect to ρ will be called metric
balls and denoted by B.e

Lemma 11.3 Every admissible curve γ : [0, T ] → Ω is Lipschitz.

Proof. Use the Schwartz inequality.

Proposition 11.4 A mapping γ : [0, T ] → (Ω, ρ) is an admissible curve if and only

if it is 1-Lipschitz i.e., ρ(γ(b), γ(a)) ≤ |b − a| for all a, b.

Proof. ⇒. This implication directly follows from the definition of ρ.

⇐. Let γ : [0, T ] → (Ω, ρ) be a 1-Lipschitz curve. By Lemma 11.3 it is Lipschitz
with respect to the Euclidean metric on Ω and hence it is differentiable a.e. We have
to prove that γ is admissible. Let t0 ∈ (0, T ) be any point where γ is differentiable.
Since ρ(γ(t0 +ε), γ(t0 )) ≤ ε for ε > 0, there exists an admissible curve η : [0, ε+δ] →
Ω, η(0) = γ(t0 ), η(ε + δ) = γ(t0 + ε) for any δ > 0. We have
Z ε+δ
η̇(t) dt = γ(t0 + ε) − γ(t0 ) = γ̇(t0 )ε + o(ε).
0

By the
P definition of an admissible curve there are measurable functions cj (t) such
that j cj (t)2 ≤ 1 and
k
X k
X ³ ´
η̇(t) = cj (t)Xj (γ(t0 )) + cj (t) Xj (η(t)) − Xj (η(0)
j=1 j=1
 SOBOLEV MET POINCARÉ 55

k
X
= cj (t)Xj (γ(t0 )) + a(t).
j=1

Note that, by Proposition 11.2, C|η(t) − η(0)| ≤ ρ(η(t), η(0)) ≤ t, provided ε and δ
are sufficiently small. Hence |a(t)| ≤ |X(η(t)) − X(η(0))| ≤ Ct, as the vector fields
have locally Lipschitz coefficients. Thus we conclude that
Z ε+δ
o(ε)
γ̇(t0 ) = ε−1 η̇(t) dt +
0 ε
k
ÃZ ! Z
ε+δ X ε+δ ε+δ
o(ε)
= cj (t) dt Xj (γ(t0 )) + ε−1 a(t) dt + .
ε j=1 0 0 ε

Selecting
P suitable P sequences εl → 0 and δl → 0 we conclude that γ̇(t0 ) =
2
b X
j j j (γ(t 0 )), b
j j ≤ 1. This completes the proof.
It is well known that any rectifiable curve in a metric space admits an arc-
length parametrization, see [22] or [246, Chapter 1]. This also holds for the Carnot–
Carathéodory distance as a Carnot–Carathéodory space splits into a family of metric
spaces such that each rectifiable curve is entirely contained in one of these metric
spaces. Note also that the arc-length parametrization makes the curve 1-Lipschitz
and hence admissible. This observation implies the following result.

Proposition 11.5 The Carnot–Carathéodory distance between any two points

equals the infimum of lengths (with respect to ρ) of curves that join those two points.
If the points cannot be connected by a rectifiable curve, then their distance is infinite.

11.2 Upper gradients and Sobolev spaces.

The following two results generalize Proposition 10.1.

Proposition 11.6 |Xu| is an upper gradient of u ∈ C ∞ (Ω) on the space (Ω, ρ).

Proof. Let γ : [a, b] → (Ω, ρ) be a 1-Lipschitz curve. By Lemma 11.3, u ◦ γ is

Lipschitz and hence
¯Z ¯ Z
¯ b ¯ b
¯ ¯
|u(γ(b)) − u(γ(a))| = ¯ h∇u(γ(t)), γ̇(t)i dt¯ ≤ |Xu(γ(t))| dt;
¯ a ¯ a

the inequality follows from the fact that γ is admissible by Lemma 11.4 and from
the Schwartz inequality. The proof is complete.

Theorem 11.7 Let 0 ≤ g ∈ L1loc (Ω) be an upper gradient on (Ω, ρ) of a function

u which is continuous with respect to the Euclidean metric. Then the distributional
derivatives Xj u, j = 1, 2, . . . , k, are locally integrable and |Xu| ≤ g a.e.
56 PIOTR HAJLASZ AND PEKKA KOSKELA

Notice that when (54) holds, the continuity assumption above can as well be
given with respect to the metric ρ.
The proof of the theorem is rather complicated and thus we first make some
comments and give applications and postpone the proof until the end of the sub-
section.
The proof is particularly easy if u ∈ C ∞ (Ω) and the vector fields have C 1 smooth
coefficients. We present it now as it may help to understand the proof for the general
case.
Since u is smooth, we do not have to worry about distributional derivatives and
we simply prove that g ≥ |Xu| a.e.
The set of the points where |Xu(x)| > 0 is open. Since the desired inequality
holds trivially outside this set, we can assume that |Xu| > 0 everywhere in Ω.
Let aP
j (x) = Xj u(x)/|Xu(x)| and let γ beP any integral curve of the vector field
Y = j aj Xj i.e., γ : (−T, T ) → Ω, γ̇(t) = j aj (γ(t))Xj (γ(t)). Obviously γ is an
admissible curve. Thus γ : (−T, T ) → (Ω, ρ) is 1-Lipschitz and hence
Z t2
|u(γ(t2 )) − u(γ(t1 ))| ≤ g(γ(t)) dt,
t1

for any −T < t1 < t2 < T . On the other hand

¯Z t2 ¯ Z t2
¯ ¯
|u(γ(t2 )) − u(γ(t1 ))| = ¯¯ h∇u(γ(t)), γ̇(t)i dt¯¯ = |Xu(γ(t))| dt.
t1 t1

This yields
Z t2 Z t2
|Xu(γ(t))| dt ≤ g(γ(t)) dt. (55)
t1 t1

If the vector field Y were parallel to one of the coordinate axes, then (55) would
imply that g ≥ |Xu| a.e. on almost every line parallel to that axis and hence
g ≥ |Xu| a.e. in Ω. The general case can be reduced to the case of a vector field
of parallel directions by the rectification theorem. This is obvious if the vector field
is C 1 -smooth as it is the usual requirement in the rectification theorem, see Arnold
[3]. However the same argument can be also used in the general Lipschitz case. A
construction of the Lipschitz rectification is provided in the proof of Theorem 11.7
(look for Φ : G → Ω).
Now we give two applications of the theorem.
Note that if u is metric L-Lipschitz, then the constant function L is an upper
gradient of u. However, the function u need not be continuous with respect to the
Euclidean metric, even if the Carnot–Carathéodory distance is a metric. This is
easily seen in the previously discussed example of ∂x and x+ ∂y in IR2 . Thus, in
order to apply Theorem 11.7 to a metric L-Lipschitz function, we need to assume
either that the function is continuous with respect to the Euclidean metric or simply
that (54) holds.
 SOBOLEV MET POINCARÉ 57

The following special case of Theorem 11.7 was proved by Franchi, HajÃlasz and
Koskela, [77], and in a slightly weaker form earlier by Chernikov and Vodop’yanov,
[38], Franchi, Serapioni and Serra Cassano, [85], and Garofalo and Nhieu, [91].

Corollary 11.8 Assume that (54) holds. If u is metric L-Lipschitz, then the dis-
tributional derivatives Xj u, j = 1, 2, . . . , k, are represented by bounded functions
and |Xu| ≤ L a.e.

The following version of Meyers–Serrin’s theorem was discovered in its local form by
Friedrichs, [87], (cf. [214, Lemma 11.27]), and later by Chernikov and Vodop’yanov,
[38, Lemma 1.2], Franchi, Serapioni and Serra Cassano, [84, Proposition 1.2.2], [85,
p. 90], and Garofalo and Nhieu, [92, Lemma 7.6].
Since later on we will need estimates from the proof, rather than the statement
alone, we recall the proof following Friedrich’s argument.

Theorem 11.9 Let X = (X1 , X2 , . . . , Xk ) be a system of vector fields with locally

Lipschitz coefficients in Ω ⊂ IRn and let 1 ≤ p < ∞. If u ∈ Lp (Ω) and (the
distributional derivative) Xu ∈ Lp (Ω), then there exists a sequence uk ∈ C ∞ (Ω)
such that kuk − ukLp (Ω) + kXuk − XukLp (Ω) → 0 as k → ∞.

Proof. We will prove that if u has compact support in Ω, then the standard mollifier
approximation gives a desired approximating sequence. The general case follows by
a partition of unity argument.
Pn
Let Y (x) = j=1 cj (x)∂/∂xj , where the functions cj are locally Lipschitz, de-
R
note one of Xj ’s. Let ϕε (x) = ε−n ϕ(x/ε), 0 ≤ ϕ ∈ C0∞ (B n (0, 1)), ϕ = 1, be a
standard mollifier kernel. For a locally integrable function u we have
Xn Z
∂u
Y (u ∗ ϕε )(x) = cj (x − y) (x − y)ϕε (y) dy
j=1
∂x j

Xn Z
∂u
+ (cj (x) − cj (x − y)) (x − y)ϕε (y) dy
j=1
∂xj
= (Y u) ∗ ϕε (x)
n Z
X ∂ ³¡ ¢ ´
+ (u(x − y) − u(x)) cj (x) − cj (x − y) ϕε (y) dy
j=1
∂yj
= (Y u) ∗ ϕε (x) + Aε u(x), (56)
where the integrals are understood in the sense of distributions. Note that
¯ ´¯¯
¯ ∂ ³¡ ¢
¯ ¯ −n
¯ ∂yj cj (x) − cj (x − y) ϕε (y) ¯ ≤ CL(x, ε)ε χB n (0,ε) , a.e.,

where L(x, ε) is the Lipschitz constant of all cj ’s on B n (x, ε). Hence

Z
|Aε u(x)| ≤ CL(x, ε) |u(y) − u(x)| dy. (57)
B n (x,ε)
58 PIOTR HAJLASZ AND PEKKA KOSKELA

If u ∈ Lp (Ω) has compact support in Ω, then it easily follows that kAε ukLp (Ω) → 0
as ε → 0. Indeed, it is obvious if u is continuous and in the general case we
can approximate u by compactly supported continuous functions in the Lp norm.
If in addition Y u ∈ Lp (Ω), then (Y u) ∗ ϕε → Y u in Lp (Ω) and hence by (56),
Y (u ∗ ϕε ) → Y u in Lp (Ω). The proof is complete.
The following result is a corollary of the above proof.

Proposition 11.10 Assume that (54) holds. Let u be metric L-Lipschitz in Ω.

Then the standard mollifier approximation converges to u uniformly on compact
subsets of Ω and
|X(u ∗ ϕε )(x)| ≤ L + |Aε u(x)|,

where |Aε u| → 0 as ε → 0 uniformly on compact subsets of Ω. The above inequality

holds for all x ∈ Ω of distance at least ε to the boundary.

Proof. Condition (54) is used to guarantee that u is continuous with respect to the
Euclidean metric, which in turn together with (57) implies that |Aε u(x)| converges
to zero uniformly on compact sets. By Corollary 11.8, |Y u| ∗ ϕε ≤ L, for any
Pk P
Y = j=1 cj Xj with j c2j ≤ 1. The desired inequality then results using (56)
with the following choice of the coefficients. Fix an arbitrary point x0 ∈ Ω. If
|X(u∗ϕε )(x0 )| = 0, then we are done; otherwise we take cj = Xj (u∗ϕε )(x0 )/|X(u∗
ϕε )(x0 )|. The proof is complete.
The theorem below shows that, in a certain sense, the analysis of vector fields
is determined by the associated Carnot–Carathéodory metric. The result is also an
affirmative answer to a question posed by Bruno Franchi.

Theorem 11.11 Let X and Y be two families of vector fields with locally Lipschitz
coefficients in Ω and such that (54) holds for the induced Carnot–Carathéodory
metrics ρX and ρY . Then the following conditions are equivalent.

1. There exists a constant C ≥ 1 such that C −1 ρX ≤ ρY ≤ CρX .

2. There exists a constant C ≥ 1 such that C −1 |Xu| ≤ |Y u| ≤ C|Xu| for all

u ∈ C ∞ (Ω).

Proof. 1. ⇒ 2. Note that if g is an upper gradient of u ∈ C ∞ (Ω) on (Ω, ρX ), then

the equivalence of the metrics implies that Cg is an upper gradient of u on (Ω, ρY ).
This fact, Proposition 11.6, and Theorem 11.7 imply that |Y u| ≤ C|Xu|. The
opposite inequality follows by the same argument.
2. ⇒ 1. Fix x, y ∈ Ω, and let u(z) = ρX (x, z). Let γ : [0, T ] → (Ω, ρY ) be an
arbitrary 1-Lipschitz curve such that γ(0) = x, γ(T ) = y.
Let uε = u ∗ ϕε be the standard mollifier approximation. By Proposition 11.6,
|Y uε | is an upper gradient of uε on (Ω, ρY ), and hence, invoking Proposition 11.10,
 SOBOLEV MET POINCARÉ 59

we get
Z T
ε→0
ρX (x, y) ←− |uε (γ(T )) − uε (γ(0))| ≤ |Y uε (γ(t))| dt
0
Z T Z T
ε→0
≤ C |Xuε (γ(t))| dt ≤ CT + C |Aε u(γ(t))| dt −→ CT.
0 0

Now it follows from the definition of ρY that ρX ≤ CρY . The opposite inequality
follows by the same argument. The proof is complete.
Let us come back to the question posed at the beginning of the section. How does
one prove a Poincaré inequality for the pair u, |Xu|? The natural approach is to
bound the oscillation of u by integrals of |Xu| over admissible curves — this can be
done as |Xu| is an upper gradient of u. Then the Poincaré inequality should follow
by averaging the resulting line integrals. Unfortunately, in general, this program
is very difficult to handle, and it turns out that many additional assumptions on
the vector fields are needed. One such a proof of a Poincaré inequality will be
presented later on (see Theorem 11.17). Anyhow, if one succeeds in proving a
Poincaré inequality using the above idea, the resulting inequality holds on metric
balls.
Thus the Poincaré inequality we should expect is

Z µZ ¶1/p
p
|u − uBe| dx ≤ CP r |Xu| dx , (58)
e
B e
σB

whenever σ Be ⊂ Ω and u ∈ C ∞ (σ B).

e Here CP > 0, σ ≥ 1, 1 ≤ p < ∞, are fixed
e denotes a ball with respect to the Carnot–Carathéodory
constants and, as usual, B
metric ρ.
Even if proving inequalities like (58) requires many assumptions on X, there
are sufficiently many important examples where (58) holds. Some of them will be
discussed in the following subsections.

Theorem 11.12 Assume that a system of locally Lipschitz vector fields is such that
condition (54) is satisfied. Fix σ ≥ 1, CP > 0, and 1 ≤ p < ∞. Then the space
(Ω, ρ, H n ) supports a p-Poincaré inequality (with given σ and CP ) if and only if
inequality (58) holds whenever σ B e ⊂ Ω and u ∈ C ∞ (σ B).
e

Proof. The left-to-right implication is easy to obtain as the function |Xu| is an

e then we can extend u from
upper gradient of u, for u ∈ C ∞ (Ω). If u ∈ C ∞ (σ B),
e
the ball (1 − ε)σ B to a continuous function on Ω; next extending |Xu| from the
same ball to Ω by ∞ gives an upper gradient on Ω of the extension of u. Now
e to the extended pair and passing to
applying the p-Poincaré inequality on (1 − ε)B
the limit as ε → 0 yields (58).
60 PIOTR HAJLASZ AND PEKKA KOSKELA

For the right-to-left implication we have to prove that, whenever g is an upper

gradient of a continuous function u, then
Z µZ ¶1/p
|u − uBe| dx ≤ CP r g p dx , (59)
e
B e
σB

e ⊂ Ω. Fix a ball B.
for all σ B e We may assume that g ∈ Lp (σ B); e otherwise the
inequality is obvious.
e ρ), Theorem 11.7 implies that |Xu| ≤ g
Since g is an upper gradient of u on (σ B,
e
a.e. in σ B. Then, by Theorem 11.9, there is a sequence of functions uk ∈ C ∞ (σ B)e
such that kuk − ukLp (σB)e + kXuk − XukLp (σB) e → 0. Thus, if we pass to the limit
in the inequality (58) applied to uk ’s, we obtain the p-Poincaré inequality for the
pair u, |Xu|. This together with the estimate |Xu| ≤ g yields (59). The proof is
complete.
There is an obvious way to define a Sobolev space associated with a system of
1,p
vector fields. Namely, we define WX (Ω), 1 ≤ p ≤ ∞, as the set of those u ∈ Lp (Ω)
p
such that |Xu| ∈ L (Ω), where Xu is defined in the sense of distributions, and we
1,p
equip the space with the norm kukLp (Ω) +kXukLp (Ω) under which WX (Ω) becomes
a Banach space.
According to Theorem 11.9, when 1 ≤ p < ∞, one can equivalently define the
space as the completion of C ∞ (Ω) in the above norm.
We will return to the construction of a Sobolev space associated with the system
of vector fields in Section 13.
Pk
Proof of Theorem 11.7. Let Y = j=1 cj Xj , where cj ’s are arbitrarily chosen
Pk
constant coefficients with j=1 c2j ≤ 1.
Let Φ(x, t) be the function uniquely defined by the conditions Φ(x, 0) = x and
d
dt Φ(x, t) = Y (Φ(x, t)). The properties of Φ are collected in the following lemma.
For the proof, see Franchi, Serapioni and Serra Cassano, [85, p.101], or Hartman,
[112], Hille, [122].

Lemma 11.13 If Ω0 ⊂⊂ Ω, then there exists T > 0 such that Φ : Ω0 ×(−2T, 2T ) →

Ω. Moreover for every t ∈ (−2T, 2T ), the mapping Φ(·, t) : Ω0 → Ω is bi-Lipschitz
onto the image with the inverse Φ(·, −t); the mapping Φ(·, t) is differentiable a.e.
and
∂Φi
(x, t) = δij + aij (x, t),
∂xj
where δij is the Kronecker symbol and |aij (x, t)| ≤ C|t|, with a constant C which
does not depend neither on x ∈ Ω0 nor on t ∈ (−T, T ). This implies that the
Jacobian of Φ satisfies

JΦ (x, t) = 1 + JeΦ (x, t), |JeΦ (x, t)| ≤ C|t|, (60)

for the given range of x and t.

 SOBOLEV MET POINCARÉ 61

Let Ω0 ⊂⊂ Ω. It suffices to show that |Xu| ≤ g a.e. in Ω0 ; the theorem will follow
then by an exhaustion of the domain Ω.
Define the directional derivative of u in the direction of Y by the formula
Ye u(x) = dt
d
|t=0 u(Φ(x, t)).
The plan of the proof of the theorem is the following. In the first step we prove
that Ye u exists a.e. and that |Ye u| ≤ g a.e. In the second step we prove that Ye u
is actually the distributional derivative and in the last step we show that, by an
appropriate choice of the cj ’s, we get |Ye u| = |Xu|.
Step 1. We show that Ye u exists a.e. and |Ye u| ≤ g a.e.
If Y (x) = 0, then Ye u(x) = 0, and hence |Ye u(x)| ≤ g(x). Thus it remains to
prove the inequality in the open set where Y 6= 0.
Observe that the curves t 7→ Φ(x, t) are admissible, and hence, for −2T < t1 <
t2 < 2T ,
Z t2
|u(Φ(x, t2 )) − u(Φ(x, t1 ))| ≤ g(Φ(x, t)) dt.
t1

Thus, if for given x, the function t 7→ g(Φ(x, t)) is locally integrable, then the above
inequality implies that the function t 7→ u(Φ(x, t)) is absolutely continuous and
¯ ¯
¯d ¯
|Ye u(Φ(x, t))| = ¯¯ u(Φ(x, t))¯¯ ≤ g(Φ(x, t)), (61)
dt

for almost all t ∈ (−2T, 2T ).

Fix x0 ∈ Ω0 with Y (x0 ) 6= 0, and let B n−1 (x0 , δ) be a sufficiently small ball
contained in the hyperplane perpendicular to Y (x0 ). For a moment restrict the
domain of definition of Φ to G = B n−1 (x0 , δ) × (−T, T ). The uniqueness theorem
for ODE implies that Φ is one-to-one on G. Moreover the properties of Φ collected in
Lemma 11.13 imply that Φ is Lipschitz on G and the Jacobian of Φ : G → Ω satisfies
C1 ≥ |JΦ | ≥ C2 > 0 on G provided δ and T are sufficiently small (note that this
is the Jacobian of a different mapping than that in (60)). Hence |JΦ (Φ−1 (z))|−1
is bounded on Φ(G). Note that |JΦ (Φ−1 (z))|−1 is defined almost everywhere on
Φ(G). This follows from the observation R that if E ⊂ Φ(G), |E| > 0, then by the
change of variables formula 0 < |E| = Φ−1 (E) |JΦ | and hence |Φ−1 (E)| > 0. The
last observation implies also that if we prove that some property holds for almost
all (x, t) ∈ B n−1 (x0 , δ) × (−T, T ), then it is equivalent to say that the property
holds for almost all z ∈ Φ(G).
The set Φ(G) is open and it contains x0 . Since we can cover the set where Y 6= 0
with such Φ(G)’s it remains to prove that |Ye u| ≤ g a.e. in Φ(G).
In order to prove that for almost every z ∈ Φ(G) the directional derivative
Y u(z) exists and satisfies |Ye u(z)| ≤ g(z), it suffices to prove that for almost every
e
x ∈ B n−1 (x0 , δ) the function t 7→ g(Φ(x, t)), t ∈ (−T, T ) is integrable, and then the
62 PIOTR HAJLASZ AND PEKKA KOSKELA

claim follows from (61). The integrability follows immediately from the estimate
Z Z T Z
g(Φ(x, t)) dt dx = g(z)|JΦ (Φ−1 (z))|−1 dz
B n−1 (x 0 ,δ) −T Φ(G)
Z
≤ C g(z) dz < ∞
Φ(G)

and the Fubini theorem. Thus we proved that |Ye u| ≤ g a.e. in Φ(G) and hence a.e.
in Ω0 .
Step 2. Now we prove that Ye u = Y u, where Y u is the distributional derivative
defined by its evaluation on ϕ ∈ C0∞ (Ω0 ) by the formula
Z Z Z
hY u, ϕi = − uY ∗ ϕ = − uY ϕ − uϕ div Y.

In the proof we will need a stronger result than just the inequality |Ye u| ≤ g. Let
uσ (z) = (u(Φ(z, σ)) − u(z))/σ. We claim that for every ϕ ∈ C0∞ (Ω0 )
Z Z
uσ (z)ϕ(z) dz −→ Ye u(z)ϕ(z) dz. (62)

Since uσ → Ye u a.e. it suffices to prove that, locally, the family {uσ }0<σ<T is
uniformly integrable. Then the convergence (62) will follow from Proposition 14.9.
According to the Vallée Poussin theorem (see Theorem 14.8), it suffices to prove that
there exists a convex Rfunction F : [0, ∞) → [0, ∞) such that F (0) = 0, F (x)/x → ∞
as x → ∞, and supσ Φ(G) F (|uσ |) < ∞, where G was defined in the first step.
Since g ∈ L1 (Φ(G)), then again by Vallée Poussin’s theorem
R there is a convex
function F with growth properties as above and such that Φ(G) F (g) < ∞. Now
Z t+σ
−1 −1
|uσ (Φ(x, t))| = σ |u(Φ(x, t + σ)) − u(Φ(x, t))| ≤ σ g(Φ(x, s)) ds.
t

Hence Jensen’s inequality implies

Z t+σ ³ ´
F (|uσ (Φ(x, t))|) ≤ σ −1 F g(Φ(x, s)) ds,
t

and thus denoting Bδn−1 = B n−1 (x0 , δ) we get

Z Z T ³ ´ Z Z T Z t+σ ³ ´
−1
F |uσ (Φ(x, t))| dt dx ≤ σ F g(Φ(x, s)) ds dt dx
Bδn−1 −T Bδn−1 −T t
Z Z T +σ ³ ´
≤ F g(Φ(x, s)) ds dx.
Bδn−1 −T
 SOBOLEV MET POINCARÉ 63

Therefore
Z Z Z T ³ ´
sup F (|uσ (z)|) dz = sup F |uσ (Φ(x, t))| |JΦ (x, t)| dt dx
0<σ<T Φ(G) 0<σ<T Bδn−1 −T
Z Z 2T
≤ C F (g(Φ(x, s)) ds dx < ∞,
Bδn−1 −T

which yields desired uniform integrability. This completes the proof of (62).
Now we proceed to prove that Ye u = Y u. Fix an arbitrary x0 ∈ Ω0 . Note that
Y u = Y (u − u(x0 )), and hence
¯Z ¯ Z
¯ ¯
e e
|hY u − Y u, ϕi| ≤ ¯ Y uϕ + (u − u(x0 ))Y ϕ¯¯ + |u − u(x0 )| |ϕ| |div Y | = I1 + I2 .
¯

First we prove that sup |hYe u − Y u, ϕi| = C(x0 , ε) < ∞, where the supremum is
taken over all ϕ ∈ C0∞ (B(x0 , ε)) with kϕk∞ ≤ 1. This inequality implies that
Ye u − Y u is a signed Radon measure with total variation on B n (x0 , ε) equal to
C(x0 , ε).
In what follows we assume that ϕ is compactly supported in B n (x0 , ε) with the
supremum norm no more than 1. As Y has locally Lipschitz coefficients, |div Y | is
locally bounded and hence
I2 ≤ Cεn sup |u − u(x0 )|.
B n (x0 ,ε)

The estimates for I1 are more difficult to handle. In what follows we write u instead
of u − u(x0 ) and simply assume that u(x0 ) = 0. We have
Z µZ Z ¶
1
u(x)Y ϕ(x) dx = lim u(x)ϕ(x) dx − u(x)ϕ(Φ(x, −t)) dx = A.
t→0 t

The change of variables x

e = Φ(x, −t) together with (60) yields
Z Z
u(x)ϕ(Φ(x, −t)) dx = u(Φ(e x)(1 + JeΦ (e
x, t))ϕ(e x, t)) de
x,

and hence by (62)

Z Z
u(x) − u(Φ(x, t)) 1
A = lim ϕ(x) dx − lim u(Φ(x, t))ϕ(x)JeΦ (x, t) dx
t→0 t t→0 t
Z Z
1
= − Ye u(x)ϕ(x) dx − lim u(Φ(x, t))ϕ(x)JeΦ (x, t) dx.
t→0 t

Hence
¯ Z ¯
¯ ¯
I1 = ¯ lim 1 u(Φ(x, t))ϕ(x)JeΦ (x, t) dx¯
¯t→0 t ¯
Z
≤ C lim |u(Φ(x, t))| dx ≤ Cεn sup |u − u(x0 )|.
t→0 B n (x0 ,ε) B n (x0 ,ε)
64 PIOTR HAJLASZ AND PEKKA KOSKELA

This and the estimate for I2 yields that Ye u − Y u is a signed Radon measure whose
total variation on B n (x0 , ε) is estimated from above by

|Ye u − Y u|(B n (x0 , ε)) ≤ Cεn sup |u − u(x0 )|.

B(x0 ,ε)

This in turn implies that the measure |Ye u − Y u| is absolutely continuous with
respect to the Lebesgue measure, so Ye u − Y u ∈ L1loc , and then by the Lebesgue
differentiation theorem
Z
e
|Y u(x0 ) − Y u(x0 )| = lim |Ye u − Y u| ≤ C lim sup |u − u(x0 )| = 0
ε→0 B n (x ,ε) ε→0 B n (x ,ε)
0 0

for almost all x0 . Thus Ye u = Y u, and hence by Step 1, |Y u| ≤ g a.e.

Step 3. Repeating the above arguments for all the rational coefficients cj , we
conclude
P that there is a subset
P of Ω0 of full measure such that for all rational cj ’s
2
with j cj ≤ 1 there is | j cj Xj u| ≤ g at all points of the set. If |Xu| = 0 at a
given point, then |Xu| ≤ g at the point. If |Xu| 6= 0, then approximating coefficients
e
cj = Xj u/|Xu| by rational coefficients
P and passing to 0 with the accuracy of the
approximation yields |Xu| = | j e cj Xj u| ≤ g. The proof of the theorem is complete.

11.3 Carnot groups.

The aim of this subsection is to give a background on the so-called Carnot groups
which are prime examples of spaces that support the p-Poincaré inequality for any
1 ≤ p < ∞. Carnot groups are special cases of Carnot–Carathéodory spaces as-
sociated with a system of vector fields satisfying Hörmander’s condition that will
be described in the next subsection. For a more complete introduction to Carnot
groups, see Folland and Stein [73, Chapter 1] and also Heinonen [114].
Before we give the definition we need to collect some preliminary notions and
results.
Let g be a finite dimensional real Lie algebra. We say that g is nilpotent of step
m if for some positive integer m, g(m) 6= {0}, g(m+1) = {0}, where g(1) = g and
g(j+1) = [g, g(j) ]. A Lie algebra is called nilpotent if it is nilpotent of some step m.
A Lie group G is called nilpotent (of step m) if its Lie algebra is nilpotent (of step
m).
Let V be the underlying vector space of the nilpotent Lie algebra g. Define the
polynomial mapping ◦ : V × V → V by the Campbell–Hausdorff formula
∞
X X
(−1)p+1 (n1 + m1 + . . . + np + mp )−1
X ◦Y =
p=1
p n1 !m1 ! · · · np !mp !
ni +mi ≥1
i=1,2,...,p

×(ad X)n1 (ad Y )m1 . . . (ad X)np (ad Y )mp −1 Y,

 SOBOLEV MET POINCARÉ 65

where (ad A)B = [A, B]. We adopt here the convention that if mp = 0, then the
term in the sum ends with . . . (adX)np −1 X. Note also that if mp > 1, then the
term in the sum is zero.
The formal series on the right hand side of the Campbell–Hausdorff formula is
in fact a polynomial, because the Lie algebra is nilpotent. One can check that the
mapping defines a group structure on V with the Lie algebra g. Since connected
and simply connected Lie groups with isomorphic Lie algebras are isomorphic, we
obtained a full description of simply connected nilpotent Lie groups.
In what follows the group identity will be denoted by 0; however for the group
law we use multiplicative notation: xy.
One can write the Campbell–Hausdorff formula in the form

1
X ◦ Y = X + Y + [X, Y ] + . . .
2
where the dots indicate terms of order greater than or equal to 3. Note that the map
t 7→ tX is a one parameter subgroup of V . Hence the exponential map exp : g → V
is identity. Then one can find a basis in V so that in the induced coordinate system
the Jacobi matrix of the left multiplication by a ∈ V is a lower triangular matrix
with ones on the diagonal. Thus the Jacobi determinant equals one. Hence the
Lebesgue measure is the left invariant Haar measure. The same argument applies
to the right multiplication, and so the Lebesgue measure is the bi-invariant Haar
measure.
A Carnot group is a connected and simply connected Lie group G whose Lie
algebra g admits a stratification g = V1 ⊕ · · · ⊕ Vm , [V1 , Vi ] = Vi+1 , Vi = {0}
for i > m. Obviously a Carnot group is nilpotent. Moreover a Carnot group is
nilpotent of step m if Vm 6= 0. Note that the basis of V1 generates the whole Lie
algebra g. Carnot groups are also known as stratified groups.
Being nilpotent, Carnot group is diffeomorphic to IRn for some n.
Let X1 , X2 , . . . , Xk form a basis of V1 . We identify X1 , X2 ,. . . ,Xk with the left
invariant vector fields.
The following result is due to Chow, [40], and Rashevsky, [213]. For modern
proofs see Bellaı̈che [7], Gromov [101], Herman [121], Nagel, Stein and Wainger
[204], Strichartz [235], Varopoulos, Saloff-Coste and Coulhon [251].

Proposition 11.14 The Carnot–Carathéodory distance associated with the basis

X1 , X2 ,. . . ,Xk of V1 is a metric i.e., every two points of the Carnot group can be
connected by an admissible curve.

The aim of this subsection is to prove that a Carnot group with the above Carnot–
Carathéodory metric supports the p-Poincaré inequalities for all 1 ≤ p < ∞ (see
Theorem 11.17). This is a special case of Jerison’s result (see Theorem 11.20) that
will be described in the next subsection.
66 PIOTR HAJLASZ AND PEKKA KOSKELA

By G we will denote a Carnot group of step m and ρ will be the Carnot–

Carathéodory metric associated with the basis X1 ,. . . , Xk of V1 .
As the Carnot–Carathéodory metric is not given in an explicit form, it is quite
difficult to handle. Therefore it is convenient to introduce new distances that can
be defined explicitly and that are equivalent to the Carnot–Carathéodory metric.
A Carnot group admits a one-parameter family of dilations that we next de-
scribe.
For X ∈ Vi and r > 0 we set δr X = ri X. This extends to a linear map that is
an automorphism of the Lie algebra g. This in turn induces an automorphism of
the Lie group via the exponential map.
Observe that the metric ρ has the two important properties of being left invariant
and commutative with δr in the sense that ρ(δr x, δr y) = rρ(x, y).
A continuous homogeneous norm on G is a continuous function | · | : G → [0, ∞)
that satisfies 1) |x−1 | = |x|, 2) |δr x| = r|x| for all r > 0 and 3) |x| = 0 if and only
if x = 0.
One such homogeneous norm is given by |x| = ρ(0, x).

Proposition 11.15 Let |·| be a continuous homogeneous norm. Then the following
results hold.

1. There exist constants C1 , C2 > 0 such that C1 kxk ≤ |x| ≤ C2 kxk1/m , for
|x| ≤ 1. Here k · k denotes a fixed Euclidean norm.
2. The distance %(x, y) = |x−1 y| is a quasimetric i.e., it has all the properties of
metric but the triangle inequality that is replaced by a weaker condition: there
is a constant C > 0 such that for all x, y, z ∈ G

%(x, y) ≤ C(%(x, z) + %(z, y)). (63)

3. Balls B(x, r) = {y : %(x, y) < r} are the left translates of B(0, r) by x, and
B(0, r) = δr B(0, 1).
Pm
4. The number Q = j=1 j dim Vj will be called the homogeneous dimension. It
satisfies |δr (E)| = rQ |E| and hence |B(x, r)| = CrQ for all x ∈ G and all
r > 0, where |E| denotes the Lebesgue measure of the set E.
5. Any two continuous homogeneous norms are equivalent in the sense that if |·|0
is another continuous homogeneous norm on G, then there exist C1 , C2 > 0
such that C1 |x|0 ≤ |x| ≤ C2 |x|0 for all x ∈ G.

For a proof, see Folland and Stein [73, Chapter 1]. Anyway the proof is easy and it
could be regarded as a very good exercise.
In the literature the concept of a homogeneous norm is defined as above but with
the additional property of being C ∞ -smooth on G \ {0}. This property is irrelevant
 SOBOLEV MET POINCARÉ 67

to us. Thus we delete it and add the adjective “continuous” to indicate that we do
not assume smoothness.
To give an explicit example of a continuous
Pm homogeneous norm note that any
element x ∈ V can be represented as x = j=1 xj , where xj ∈ Vj . Fix an Euclidean
norm k · k in V . Then
m
X
|x| = kxj k1/j ,
j=1

is a continuous homogeneous norm on G (after identification of G with V ).

The continuous homogeneous norm induced by the Carnot–Carathéodory metric
x 7→ ρ(0, x) satisfies (63) with C = 1. For general continuous norms we only have
C ≥ 1. See also Hebisch and Sikora, [113], for a construction of a homogeneous
norm (i.e. smooth on G \ {0}) with C = 1.
Mitchell, [199], proved that the Hausdorff dimension of a Carnot group is equal
to its homogeneous dimension, see also [101, p. 102]. This dimension, in general,
is larger than the Euclidean dimension of the underlying Euclidean space. This
shows that the Carnot–Carathéodory geometry is pretty wild and the metric is not
equivalent to any Riemannian metric.
It is an exercise to show that inequality 1. of the above proposition implies that
for every bounded domain Ω ⊂ G, there are constants C1 , C2 > 0 such that

C1 kx − yk ≤ ρ(x, y) ≤ C2 kx − yk1/m , (64)

whenever x, y ∈ Ω. Note that inequality (64) along with Lemma 9.4 imply that
every two points can be connected by a geodesic — the shortest admissible curve.
So far we have not given any examples of the Carnot group. Let us fill the gap
right now.
Example 11.16 The most simple nontrivial example of a Carnot group is the
Heisenberg group IH1 = C × IR with the group law

(z, t) ◦ (z 0 , t0 ) = (z + z 0 , t + t0 + 2Im zz 0 ).

The basis consisting of the left invariant vector fields X, Y, Z, such that X(0) =
∂/∂x, Y (0) = ∂/∂y, T (0) = ∂/∂t, is given by

∂ ∂ ∂ ∂ ∂
X= + 2y , Y = − 2x , T = .
∂x ∂t ∂y ∂t ∂t

Note that [X, Y ] = −4T and all the other commutators are trivial. Thus the Lie
algebra is stratified, h = V1 ⊕ V2 with V1 = span{X, Y } and V2 = span{T }. The
Carnot–Carathéodory metric is defined with respect to the vector fields X, Y . The
group IH1 is a nilpotent group of step 2 and its homogeneous dimension is 4. The
family of dilations is given by δr (z, t) = (rz, r2 t) for r > 0. Moreover the function
|(z, t)| = (t2 + |z|4 )1/4 is a homogeneous norm.
68 PIOTR HAJLASZ AND PEKKA KOSKELA

The following theorem states that a Carnot group supports a 1-Poincaré in-
equality. This is a corollary of a much more general theorem of Jerison (see The-
orem 11.20). For completeness we provide a clever proof due to Varopoulos, [249]
(see also [223, page 461]).

Proposition 11.17 Any Carnot group equipped with the Lebesgue measure and the
Carnot–Carathéodory metric supports a 1-Poincaré inequality.

Proof. Let G be a Carnot group with the Carnot–Carathéodory metric that we

denote by ρ. Let u, g be a pair of a continuous function and its upper gradient. It
suffices to prove that
Z Z
|u(x) − uB | dx ≤ Cr g(x) dx (65)
B 2B

on every ball of radius r. Obviously we can assume that the ball B is centered at
0. Set |z| = ρ(0, z) and let γz : [0, |z|] → G be a geodesic path that joins 0 with z.
Observe that s 7→ xγz (s) is the shortest path that joins x with xz. Hence
Z |z|
|u(x) − u(xz)| ≤ g(xγz (s)) ds.
0

This and the left invariance of the Lebesgue measure yields

Z Z Z
1
|u(x) − uB | dx ≤ |u(x) − u(y)| dy dx
B |B| B B
Z Z
1
= χB (x)χB (xz)|u(x) − u(xz)| dz dx
|B| G G
Z Z Z |z|
1
≤ χB (x)χB (xz)g(xγz (s)) ds dx dz.
|B| G G 0
Invoking the right invariance of the Lebesgue measure we obtain
Z Z
χB (x)χB (xz)g(xγz (s)) dx = χBγz (s) (ξ)χBz−1 γz (s) (ξ)g(ξ) dξ
G G
Z
≤ χ2B (z) g(ξ) dξ, (66)
2B

Here we denote by Bh the right translation of B by h. The above inequality

requires some explanations. If the expression under the sign of the middle interval
has a nonzero value, then ξ = xγz (s) = yz −1 γz (s) for some x, y ∈ B. Hence
z = x−1 y ∈ 2B. Thus ξ = xγx−1 y (s) lies on a geodesic that joins x with y and
so ρ(x, ξ) + ρ(y, ξ) = ρ(x, y), which together with the triangle inequality implies
ξ ∈ 2B and hence (66) follows. Now
Z Z Z |z| Z
1
|u(x) − uB | dx ≤ χ2B (z) g(ξ) dξ ds dz
B |B| G 0 2B
 SOBOLEV MET POINCARÉ 69

Z Z
1
= |z|g(ξ) dξ dz
|B| 2B 2B
Z
≤ Cr g(ξ) dξ.
2B

The proof is complete.

Remarks. 1) The above proof easily generalizes to more general unimodular
groups, see [249], [223, page 461].
2) Applying Theorem 9.8 to inequality (65) we conclude that the ball 2B on
the right hand side can be replaced by B, and, moreover, the exponent on the left
hand side can be replaced by Q/(Q − 1), where Q is the homogeneous dimension
of the group. This inequality in turn implies the isoperimetric inequality. Such
an isoperimetric inequality was proved first in the case of the Heisenberg group by
Pansu, [209], and in the general case of the Carnot groups by Varopoulos, [251].
For a more complete treatment of Sobolev inequalities on Lie groups with the
Carnot–Carathéodory metric, see Gromov [101], and Varopoulos, Saloff-Coste and
Coulhon [251] and also Folland [71], [72], Nhieu [206], [207].

11.4 Hörmander condition.

Definition. Let Ω ⊂ IRn be an open, connected set, and let X1 , X2 ,. . . ,Xk be
vector fields defined in a neighborhood of Ω, real valued, and C ∞ smooth. We
say that these vector fields satisfy Hörmander’s condition, provided there is an
integer p such that the family of commutators of X1 , X2 , . . . , Xk up to the length
p i.e., the family of vector fields X1 ,. . . ,Xk , [Xi1 , Xi2 ],. . . , [Xi1 , [Xi2 , [. . . , Xip ]] . . .],
ij = 1, 2, . . . , k, span the tangent space IRn at every point of Ω.
The definition easily extends to smooth manifolds, but for simplicity we will
consider the Euclidean space only.
As an example take the vector fields X1 = ∂/∂x1 , X2 = xk1 ∂/∂x2 , where k is
a positive integer. These two vector fields do not span IR2 along the line x1 = 0.
However X1 , X2 and commutators of the length k + 1 do.
Another example is given by vector fields on a Carnot group. Namely, if G is a
Carnot group (see the previous subsection) with the stratification g = V1 ⊕ . . . ⊕ Vm
of its Lie algebra, then the left invariant vector fields associated with a basis of V1
satisfy Hörmander’s condition.
The above condition was used by Hörmander [128], in his celebrated work on hy-
poelliptic operators, see also Bony [14], Chemin and Xu [36], Fefferman and Sánchez-
Calle [70], Hörmander and Melin [129], Jerison [133], Morbidelli [200], Nagel, Stein,
and Wainger [204], Rothschild and Stein [218], Sánchez-Calle [224], Varopoulos,
Saloff-Coste and Coulhon [251]. Related references will also be given in Section 13.
As usual, the Carnot–Carathéodory distance associated with a family of vector
fields satisfying Hörmander’s condition will be denoted by ρ.
70 PIOTR HAJLASZ AND PEKKA KOSKELA

The following result provides the full version of the theorem of Chow and
Raschevsky, whose special case was discussed earlier (see Proposition 11.14). For
the proof, see the references given there. In some more simple settings the theorem
was proved earlier by Carathéodory, [32].

Theorem 11.18 Let an open and connected set Ω ⊂ IRn and a system of vector
fields satisfying Hörmander’s condition in Ω be given. Then any two points in Ω
can be connected by a piecewise smooth admissible curve, and hence the Carnot–
Carathéodory distance is a metric.

Nagel, Stein and Wainger, [204], studied the geometry of Carnot–Carathéodory

spaces in detail and, in particular, they gave a more quantitative version of Chow–
Raschevsky’s theorem. Let us quote some of their results.
In what follows B̃(x, r) will denote a ball with respect to the metric ρ.

Theorem 11.19 Let X1 , . . . , Xk be a system of vector fields satisfying

Hörmander’s condition as above, and let ρ be the associated Carnot–Carathéodory
metric. Then for every compact set K ⊂ Ω there exist constants C1 and C2 such
that
C1 |x − y| ≤ ρ(x, y) ≤ C2 |x − y|1/p (67)
for every x, y ∈ K. Moreover there are r0 > 0 and C ≥ 1 such that

|B̃(x, 2r)| ≤ C|B̃(x, r)| (68)

whenever x ∈ K and r ≤ r0 .

Here, as usual, |B̃| denotes the Lebesgue measure. In the previous subsection we
proved the theorem in the special case of a Carnot group. The general case is
however much more difficult, see also Gromov [101], and Varopoulos, Saloff-Coste
and Coulhon [251, Section IV.5]. Estimate (67) has been obtained independently
by Lanconelli [165].
Assume for a moment that Ω = IRn . If Ω0 ⊂ IRn is bounded with respect to
the Euclidean metric, then by (67) it is also bounded with respect to ρ. However,
if Ω0 is bounded with respect to ρ, then it need not be bounded with respect
to the Euclidean metric. Indeed, if one of the vector fields is x21 ∂/∂x1 , then the
Carnot–Carathéodory distance to infinity is finite because of the rapid growth of
the coefficient. Hence, in general, (68) holds only for r < r0 for some sufficiently
small r0 and r0 cannot be replaced with 5 diam ρ (Ω0 ), even if diam ρ (Ω0 ) < ∞, as
was required for the measure in the definition of doubling in Ω0 .
Proposition 11.17 is a special case of the following Poincaré inequality of Jerison
[133], see also Jerison and Sanchez–Calle [136], and Lanconelli and Morbidelli [166].

Theorem 11.20 Let X1 , . . . , Xk be a system of vector fields satisfying

Hörmander’s condition in Ω. Then for every compact set K ⊂ Ω there are con-
 SOBOLEV MET POINCARÉ 71

stants C > 0 and r0 > 0 such that for u ∈ Lip (B̃)

Z Z
|u − uB̃ | dx ≤ Cr |Xu| dx, (69)
B̃ 2B̃

whenever B̃ is a ball centered at K with radius r < r0 .

In fact, Jerison proved the inequality with the L2 norms on both sides, but the
same argument works with the L1 norm. Then Jerison proved that one can replace
the ball 2B̃ on the right hand side of (69) with B̃. As we have already seen this can
be done in a much more general setting, see Section 9.

11.5 Further generalizations

The results of the previous two subsections concern Poincaré inequalities for smooth
vector fields satisfying Hörmander’s condition. It is a difficult problem to find a large
class of vector fields with Lipschitz coefficients such that the Poincaré type inequal-
ities hold on the associated Carnot–Carathéodory spaces. The lack of smoothness
does not permit one to use a Hörmander type condition. There are few results of
that type, see Franchi [74], Franchi, Gutiérrez and Wheeden [76], Franchi and Ser-
apioni [83], Franchi and Lanconelli [78], Jerison and Sanchez-Calle [136]. It seems
that Franchi and Lanconelli, [78], were the first to prove a Poincaré type inequal-
ity for a Carnot–Carathéodory space. They probably also were the first to prove
estimates of the type as in Theorem 11.19.

12 Graphs
Let G = (V, E) be a graph, where V is the vertex set and E the set of edges. We
say that x, y ∈ V are neighbors if they are joined by an edge; we denote this by
x ∼ y. Assume that the graph is connected in the sense that any two vertices can be
connected by a sequence of neighbors. We let the distance between two neighbors
to be 1. This induces a geodesic metric on V that we denote by %. The graph is
endowed with the counting measure: the measure of a set E ⊂ V is simply the
number V (E) of elements of E. For a ball B = B(x, r) we use also the notation
V (B) = V (x, r). We say that G is locally uniformly finite if d = supx∈V d(x) < ∞,
where d(x) is the number of neighbors of x. The length of the gradient of a function
u on V at a point x is
X
|∇G u|(x) = |u(y) − u(x)|.
y∼x

Many graphs have the following two properties:

1. The counting measure is doubling i.e.,

V (x, 2r) ≤ Cd V (x, r),
72 PIOTR HAJLASZ AND PEKKA KOSKELA

for every x ∈ V and r > 0.

2. The p-Poincaré inequality holds i.e., there are constants C > 0 and σ ≥ 1
such that
à !1/p
1 X 1 X
p
|u(x) − uB | ≤ CP r |∇G u| (x) , (70)
V (B) V (σB)
x∈B x∈σB

for any ball B and any function u : V → IR.

Observe that the doubling condition implies that the graph is locally uniformly
finite.
The Euclidean, or more generally, the upper gradients have the truncation prop-
erty. Unfortunately the truncation property is no longer valid for the length of the
gradient on a graph. This is because, in general, |∇G utt21 | is not supported on the set
{t1 < u ≤ t2 }. However, intuition suggests that |∇G u| should still have properties
similar to those of a gradient with the truncation property.
If v ∈ Lip (IRn ) and p > 0, then
∞
X ∞
X
k
|∇v22k−1 |p = |∇v|p χ{2k−1 <v≤2k } ≤ |∇v|p (71)
k=−∞ k=−∞

almost everywhere. It turns out that a version of inequality (71) is satisfied also
by the length of the gradient on a graph. More precisely we have the following
estimate.

Lemma 12.1 Let G be locally uniformly finite i.e., d = supx∈V d(x) < ∞. If
v : V → IR and p > 0, then
∞ ¯
X ¯p
¯ k ¯
¯∇G v22k−1 ¯ (x) ≤ C(p, d)|∇G v|p (x)
k=−∞

for each x ∈ V.

Proof. Fix x ∈ V and let

vM (x) = max{v(w) : %(w, x) ≤ 1} ,

vm (x) = min{v(w) : %(w, x) ≤ 1} .

Note that |∇G v|(x) ≥ |vM (x) − vm (x)|. Assume for simplicity that vm (x) > 0 (the
case vm (x) ≤ 0 follows by the same argument). Let j ∈ ZZ be the least integer and
i ∈ ZZ the largest integer such that

2j ≥ vM (x) ≥ vm (x) ≥ 2i .
 SOBOLEV MET POINCARÉ 73

We have
j−2
X
|vM (x) − vm (x)| = vM (x) − 2j−1 + (2k+1 − 2k ) + 2i+1 − vm (x).
k=i+1

Hence
à à j−2
!p !
X
p j−1 p k i+1 p
|∇G v| (x) ≥ C |vM (x) − 2 | + 2 + |2 − vm (x)|
k=i+1
à j−2
!
X
≥ C |vM (x) − 2j−1 |p + (1 − 2−p ) 2kp + |2i+1 − vm (x)|p
k=i+1
à j−2
1 j 1−2 −p X k+1
≥ C |∇G v22j−1 |p (x) + |∇G v22k |p (x)
dp dp
k=i+1
¶
1 2i+1 p
+ p |∇G v2i | (x)
d
j−1
C(p) X k+1
≥ |∇G v22k |p (x)
dp
k=i
∞
C(p) X k+1
= p
|∇G v22k |p (x).
d
k=−∞

The proof is complete.

The inequality of the lemma is a good substitute for the truncation property;
it allows one to mimic the proofs of Theorems 2.1 and 2.3. We will generalize
Corollary 9.8. This result deals with sharp inequalities with integrals on the different
sides of the inequality taken over the same domain. As pointed out in Section 9,
a Poincaré inequality does not, in general, guarantee that one could use balls of
the same size on the different sides of the inequality. We described a sufficient
condition in terms of the geometry of balls that, in particular, holds for the Carnot-
Caratheodory metrics. As ρ is a geodesic metric, it should come as no surprise that
we can reduce the size of σ in (70) down to 1.
The following theorem is related to some results in Bakry, Coulhon, Ledoux and
Saloff-Coste [4], Coulhon [54].

Theorem 12.2 Assume that the counting measure is doubling, and that for some
constants Cb > 0, s ≥ 1
µ ¶s
V (x, r) r
≥ Cb
V (x, r0 ) r0
whenever B(x, r) ⊂ B(x0 , r0 ). Suppose that each function u : V → IR satisfies the
p-Poincaré inequality (70) with a fixed p > 0.
74 PIOTR HAJLASZ AND PEKKA KOSKELA

1. If 0 < p < s, then there is a constant C > 0 such that

à !1/p∗ à !1/p
1 X ∗ 1 X
|u(x) − uB |p ≤ Cr p
|∇G u| (x) (72)
V (B) V (B)
x∈B x∈B

for any ball B of radius r and any function u : V → IR, where p∗ = sp/(s−p).

2. If p = s > 1, then there are constants C1 , C2 > 0 such that

µ ¶s/(s−1)
1 X C1 V (B)1/s |u(x) − uB |
exp ≤ C2
V (B) (diam B)k∇G ukLs (B)
x∈B

for any ball B of radius r and any function u : V → IR.

3. If p > s, then there is C > 0 such that

|u(x) − u(y)| ≤ C%(x, y)1−s/p rs/p V (B)−1/p k∇G ukLp (B) ,

for all x, y ∈ B, where B is an arbitrary ball of radius r, and for any function
u : V → IR.

The constants C, C1 , C2 depend on p, s, σ, Cd , Cb and CP only.

Remark. If sp/(s − p) < 1, then we have to replace uB by uB0 in (72), where

B0 = (2σ)−1 B.
Proof. The proof involves arguments similar to those used in the previous sec-
tions, and so we only sketch the main ideas, leaving the details to the reader.
First of all, under the local finiteness of the graph, one may assume that r > 10σ
(as for r ≤ 10σ we only have a finite collection of non-isometric balls). We follow
the line of ideas from Section 9. Given a ball B(x0 , r) and a point x ∈ B(x0 , r),
we join x to x0 by a chain x = x1 , ..., xm of length less than r of vertices. If
we trace along the chain for l steps with l the least integer larger or equal to 4σ,
then B(xl , 2σ) ⊂ B(x0 , r). Following the chain towards x we may construct a chain
B0 = B(x0 , r/(2σ)), B1 , ..., Bk = B(xl , 2) of balls as in the C(σ, M ) condition of
Section 9. Next,
l−1
X k−1
X
|u(x) − uB0 | ≤ |u(xi+1 ) − u(xi )| + |uB(xl ,2) − u(xl )| + |uBi+1 − uBi |
i=1 i=0
l−1
X X
≤ |∇G u|(xi ) + |∇G u|(y)
i=1 y∈B(xl ,2)
 1/p
k
X X
1
+ C ri  |∇G u|p (y) . (73)
i=0
V (σB i )
y∈σBi
 SOBOLEV MET POINCARÉ 75

We employed here the observation that |u(xi+1 ) − u(xi )| ≤ |∇G u|(xi ). Inequality
(73) is a good substitute for (33), (43) as |∇G u|(y) equals to the product of the
radius and the Lp -average of |∇G u| over the ball σB(y, σ −1 ).
Now assume that p < s. Write sp/(s − p) = p∗ . Using a version of Theorem 5.1
as in Section 9 we conclude the weak type inequality
à !p∗ /p
1 X
∗
V ({x ∈ B : |u(x) − uB0 | > t})tp ∗
sup ≤ Crp p
|∇G u| (x) .
t≥0 V (B) V (B)
x∈B

To obtain the desired strong type inequality one reasons as follows.

Define v± as in the proof of Theorem 2.1 (with Ω replaced by B). It suffices to
∗
prove suitable Lp estimates for v+ and v− . In what follows v denotes either v+ or
v− . We have
∗
à !p∗ /p
V {x ∈ B : vtt12 > t}tp ∗ 1 X
sup ≤ Crp |∇G vtt12 |p ,
t≥0 V (B) V (B)
x∈B

and hence
∞ ∗ k−1
1 X p∗ X 2kp V ({x ∈ B : v22k−2 ≥ 2k−2 })
v (x) ≤
V (B) V (B)
x∈B k=−∞
à ∞
!p∗ /p
∗ 1 X X k−1
≤ Crp |∇G v22k−2 |p (x)
V (B)
x∈B k=−∞
à !p∗ /p
p∗ 1 X p
≤ Cr |∇G v| (x)
V (B)
x∈B
à !p∗ /p
∗ 1 X
≤ Crp |∇G u|p (x) .
V (B)
x∈B

If p = s, then the method described above provides us with a chains that are
sufficiently good to mimic the proof of Theorem 6.1.
Once we have good chains also the Hölder continuity with the same balls on
both sides follows when p > s.
The proof is complete.
Remarks. 1) The doubling property 1. and the Poincaré inequality 2. are very
important in the potential theory on graphs. Indeed, independently Delmotte, [64],
[65], Holopainen and Soardi, [127], and Rigoli, Salvatori and Vignati, [216], proved
that 1. and 2. imply the so-called Harnack inequality for p-harmonic functions. As
a consequence, they concluded a Liouville type theorem stating that every bounded
p-harmonic function on G is constant. Recall that u : V → IR is 2-harmonic if
76 PIOTR HAJLASZ AND PEKKA KOSKELA

P
it satisfies the mean value property, i.e., u(x) = d(x)−1 y∼x u(y) for all x ∈ V .
For a definition of p-harmonic functions with p 6= 2, see, e.g., [127]. The proof of
the Harnack inequality employs the Sobolev–Poincaré inequality as it relies on the
Moser iteration. As we have already seen properties 1. and 2. imply the Sobolev–
Poincaré inequality. Papers related to the Harnack inequalities on graphs include
Auscher and Coulhon [2], Chung [43], Chung and Yau [44], Lawler [169], Merkov
[196], Rigoli, Salvatori and Vignati [217], Schinzel [225], Zhou [262].
The Moser iteration was originally employed in the setting of elliptic and
parabolic equations; see the next section.
2) There are many examples of graphs for which both properties 1. and 2. are
satisfied. A very nice example is given by a Cayley graph associated with a finitely
generated group. We say that the group G is finitely generated if there is a finite set
{γi }ki=1 such that every element g ∈ G can be presented as a product g = γiε11 · · · γiεll ,
εi = ±1. Then the vertex set of the Cayley graph is the set of all elements of G and
two elements g1 , g2 ∈ G are connected by an edge if g1 = g2 γi±1 for some generator
γi . Thus we may visualize finitely generated groups as geometric objects. This
point of view has been intensively used after Milnor’s paper, [198].
We say that the group is of polynomial growth if V (r) ≤ CrC for all r > 0
and some C ≥ 1. One of the most beautiful results in the area is due to Gromov
[100]. He proved that a group is of polynomial growth if and only if it is virtually
nilpotent, and hence by the theorem of Bass, [6], V (r) ≈ rd for some positive integer
d.
Thus if the group is of polynomial growth, then it satisfies the doubling property
1. It is also known that it satisfies the 1-Poincaré inequality.

Proposition 12.3 If G is a finitely generated group of polynomial growth V (r) ≈

rd , d positive integer, then there is a constant C > 0 such that
1 X 1 X
|u(x) − uB | ≤ Cr |∇G u(x)|
V (B) V (2B)
x∈B x∈2B

for every ball B ⊂ X.

The reader may prove the proposition as an easy exercise mimicing the proof of
Theorem 11.17.
Other examples of graphs with properties 1. and 2. can be found in Holopainen
and Soardi [127], Coulhon [54], Coulhon and Saloff-Coste [58], [59], Saloff-Coste
[222].
3) The analysis on graphs is also important in the study of open Riemannian
manifolds because, roughly speaking, one can associate with given manifold a graph
with similar global properties. This method of discretization of manifolds has been
in active use after the papers of Kanai, [149], and Mostow, [203]. Related refer-
ences include Coulhon [52] [53], Coulhon and Ledoux [57], Coulhon and Saloff-Coste
[56], [58], [59], Chavel [33], Delmotte [64], [65], Holopainen [123], Holopainen and
 SOBOLEV MET POINCARÉ 77

Soardi [126], [127], Soardi [230], Varopoulos [248], and Varopoulos, Saloff-Coste and
Coulhon [251].

13 Applications to P.D.E and nonlinear potential

theory
The results presented in the paper directly apply to the regularity theory of degen-
erate elliptic equations associated with vector fields. Below we describe some of the
applications.

13.1 Admissible weights

Let A = (A1 , . . . , Am ) : IRn × IRm → IRm be a Carathéodory function satisfying
the growth conditions

|A(x, ξ)| ≤ C1 ω(x)|ξ|p−1 , A(x, ξ) · ξ ≥ C2 ω(x)|ξ|p ,

where 1 < p < ∞, C1 , C2 > 0 are fixed constants and 0 < ω ∈ L1loc (IRn ). We will
denote by dµ = ω dx the measure with the density ω.
Given A, we consider the equation
m
X
Xj∗ Aj (x, X1 u, . . . , Xm u) = 0, (74)
j=1

where X = (X1 , . . . , Xm ) is a family of vector fields with locally Lipschitz co-

n ∗
efficients
R in
R IR ∗. Recall that X j denotes the formal adjoint of Xj , that is,
∞
Xj uv = uXj v for all u, v ∈ C0 .
The theory of nonlinear equations of the type (74), especially when X is a system
of vector fields satisfying Hörmander’s condition, is an area of intensive research; see,
e.g., Buckley, Koskela and Lu [19], Capogna [25], Capogna, Danielli and Garofalo
[27], [31], Chernikov and Vodop’yanov [38], Citti [45], Citti and Di Fazio [46], Citti,
Garofalo and Lanconelli [47], Danielli, Garofalo and Nhieu [61], Franchi, Gutiérrez
and Wheeden [76], Franchi and Lanconelli [78], Franchi and Serapioni [83], Garofalo
and Lanconelli [90], Garofalo and Nhieu [92], HajÃlasz and Strzelecki [108], Jerison
[133], Jerison and Lee [134], [135], Jost and Xu [146], Lu [177], [179], [180], [181],
Marchi [189], Vodop’yanov and Markina [254], Xu [258], [259], Xu and Zuily [260].
The above papers mostly deal with the nonlinear theory. References to the broad
literature on the linear theory can be found in these papers.
Equation (74) is a generalization of the classical weighted p-harmonic equation.
Indeed, if X = ∇ and A(ξ) = ω(x)|ξ|p−2 ξ we get the equation
¡ ¢
div ω(x)|∇u|p−2 ∇u = 0.
78 PIOTR HAJLASZ AND PEKKA KOSKELA

In what follows we assume that the condition (54) is satisfied i.e. we assume that
the Carnot–Carathéodory distance ρ is a metric such that the identity map is a
homeomorphism between the Euclidean metric and ρ.
We call u a weak solution of (74) if
Z m
X
Aj (x, Xu) Xj ϕ(x) dx = 0,
IRn j=1

for any ϕ ∈ C0∞ (IRn ). We assume that the weak solution belongs to the weighted
1,p
Sobolev space WX (IRn , µ) defined as the closure of C ∞ functions in the norm
µZ ¶1/p µZ ¶1/p
p p
kuk1,p,X,ω = |u(x)| ω(x) dx + |Xu(x)| ω(x) dx .
IRn IRn

Already in the “classical” case i.e. when Xu = ∇u one has to put many additional
conditions on the weight ω in order to have a reasonable theory.
The first condition concerns the definition of the Sobolev space. One needs
the so-called uniqueness condition which guarantees that the “gradient” X is well
defined in the Sobolev space associated to X. Later we will clarify this condition.
The regularity results for solutions to (74), like Harnack inequality and Hölder
continuity, are usually obtained via the Moser iteration technique. For that the
essential assumptions are a doubling condition on µ, (with respect to the Carnot–
Carathéodory metric), the Poincaré inequality
µZ ¶1/p µZ ¶1/p
p p
|u − uBe| dµ ≤ Cr |Xu| dµ , (75)
e
B e
B

e and a Sobolev inequality

for all smooth functions u in a metric ball B,
µZ ¶1/q µZ ¶1/p
|u|q dµ ≤ Cr |Xu|p dµ , (76)
e
B e
B

with some q > p for all smooth functions u with compact support in a metric ball
B̃.
Given the above assumptions one can mimic the standard Moser iteration tech-
nique replacing Euclidean balls by metric balls. This leads to the Harnack inequality
e then
which states that if u is a positive solution to (74) on 2B,

sup u ≤ C inf u ,
e
B e
B

where the constant C does not depend on B. e Then the iteration of the Harnack
inequality implies that each weak solution to (74) is locally Hölder continuous with
respect to ρ and hence — if condition (67) is satisfied — locally Hölder continuous
with respect to the Euclidean metric.
 SOBOLEV MET POINCARÉ 79

The fact that the above conditions are essential for the Moser iteration was
observed first by Fabes, Kenig and Serapioni, [67]. They considered the “classical”
linear setting, X = ∇, p = 2.
It seems that Franchi and Lanconelli, [78], were the first to apply the Moser
technique for the Carnot–Carathéodory metric as above. Then the idea was ex-
tended by many authors to more difficult situations; see, e.g., Capogna, Danielli
and Garofalo [27], Chernikov and Vodop’yanov [38], Franchi [74], Franchi and Lan-
conelli [79], Franchi, Lu and Wheeden [81], Franchi and Serapioni [83], Jerison [133],
Lu [180]. There are moreover many other related papers.
Saloff-Coste, [221], and Grigor’yan, [99], independently realized that in certain
settings, a Poincaré inequality implies a Sobolev inequality and hence one can delete
assumption (76) as it follows from (75). This result was extended then to more
general situations by several authors: Biroli and Mosco, [9], Maheux and Saloff-
Coste, [186], HajÃlasz and Koskela, [105], Sturm, [239], Garofalo and Nhieu, [92].
The result presented below (Theorem 13.1) is in the same spirit. This is a
generalization of a result of HajÃlasz and Koskela, [105].
The following definition is due to Heinonen, Kilpeläinen and Martio, [120], when
X = ∇ and due to Chernikov and Vodop’yanov, [38], in the case of general vector
fields.
We say that ω ∈ L1loc (IRn ), ω > 0 a.e., is p-admissible, 1 < p < ∞, if the measure
defined by dµ = ω(x) dx satisfies the following four conditions:

e ≤ Cd µ(B)
1. (Doubling condition) µ(2B) e for all metric balls B
e ⊂ IRn .

2. (Uniqueness condition)
R If Ω is an openR subset of IRn and ϕi ∈ C ∞ (Ω) is a
sequence such that Ω |ϕi |p dµ → 0 and Ω |Xϕi −v|p dµ → 0, where v ∈ Lp (µ),
then v ≡ 0.
3. (Sobolev inequality) There exists a constant k > 1 such that for all metric
e ⊂ IRn and all ϕ ∈ C ∞ (B)
balls B e
0
µZ ¶1/kp µZ ¶1/p
kp p
|ϕ| dµ ≤ C2 r |Xϕ| dµ .
e
B Be

e ⊂ IRn is a metric ball and ϕ ∈ C ∞ (B),

4. (Poincaré inequality) If B e then
Z Z
|ϕ − ϕBe|p dµ ≤ C3 rp |Xϕ|p dµ.
Be Be

One can easily modify the above definition and consider vector fields defined in
an open subset Ω of IRn with the estimates in the above conditions depending on
compact subsets of Ω. However for clarity we assume the global estimates. We do
not care to present the results in their most general form. We aim to present the
method. Various generalizations are then obvious.
80 PIOTR HAJLASZ AND PEKKA KOSKELA

The uniqueness condition guarantees that any function u ∈ Lp (IRn , µ) that

1,p
belongs to WX (IRn , µ) has a uniquely defined gradient Xu as the limit of gradients
Xuk of smooth functions uk which converge to u in the Sobolev norm. If the
uniqueness condition were not true, then we would find uk ∈ C ∞ such that uk → 0
in Lp (µ) and Xu → v 6≡ 0 in Lp (µ). Then the zero function would have at least two
1,p
gradients 0 and v i.e. (0, 0) and (0, v) would be two distinct elements in WX (IRn , ω).

Theorem 13.1 Let 0 < ω ∈ L1loc (IRn ) and let X be a system of vector fields in IRn
satisfying condition (54). Then the weight ω is p-admissible, 1 < p < ∞, if and
only if the measure µ associated with ω is doubling with respect to the metric ρ (i.e.
e ≤ Cd µ(B)
µ(2B) e for all metric balls Be ⊂ IRn ) and there exists σ ≥ 1 such that
Z µZ ¶1/p
|u − uBe| dµ ≤ Cr |Xu|p dµ ,
Be e
σB

e ⊂ IRn is a metric ball of radius r and u ∈ C ∞ (σ B).

whenever B e

Proof. The necessity is obvious. Now we prove the sufficiency. First note that the
uniqueness of the gradient 2. was recently proved by Franchi, HajÃlasz and Koskela,
[77, Corollary 13].
Next, by Corollary 9.8, we conclude the Sobolev–Poincaré inequality
µZ ¶1/p∗ µZ ¶1/p
∗
|ϕ − ϕBe|p dµ ≤ Cr |Xϕ|p dµ ,
e
B e
B

e with some p∗ > p (remember that the doubling condition implies

for all ϕ ∈ C ∞ (B)
(21) with s = log2 Cd ). For our purpose the exact value of p∗ is irrelevant. It is
only important that p∗ > p. This and the Hölder inequality imply the Poincaré
inequality 4.
Now we are left with the Sobolev inequality 3. Since p∗ > p we have p∗ = kp
e we have
for some k > 1. For ϕ ∈ C0∞ (B)
µZ ¶1/kp µZ ¶1/kp
kp kp
|ϕ| dµ ≤ |ϕ − ϕBe| dµ + |ϕBe|.
e
B e
B

The Sobolev–Poincaré inequality provides us with the desired estimate for the first
summand on the right hand side. Thus it suffices to estimate |ϕBe|. The Poincaré
e gives
inequality applied to the ball B
µZ ¶1/p µZ ¶1/p
|ϕ − ϕBe|p dµ ≤ Cr |Xϕ|p dµ , (77)
e
B e
B

and when applied to the ball 2Be gives

µZ ¶1/p µZ ¶1/p
|ϕ − ϕ2Be|p dµ ≤ |ϕ − ϕ2Be|p dµ
e
B e
2B
 SOBOLEV MET POINCARÉ 81

µZ ¶1/p
p
≤ C2r |Xϕ| dµ
e
2B
µZ ¶1/p
= 2Cr |Xϕ|p dµ . (78)
e
B

Thus
à ! µZ ¶1/p µZ ¶1/p
e
µ(B)
1− |ϕBe|p dµ = |ϕBe − ϕ2Be|p dµ
e
µ(2B) e
B e
B
µZ ¶1/p
p
≤ 3Cr |Xϕ| dµ .
e
B

In the proof of the equality we employ the fact that ϕ is supported in Be and the
inequality follows from the triangle inequality and inequalities (77) and (78). It
follows from the doubling property and the geometry of metric balls in IRn that
e
1 − µ(B)/µ(2 e > C > 0 and hence
B)
µZ ¶1/p
0 p
|ϕBe| ≤ C r |Xϕ| dµ .
e
B

The proof is complete.

13.2 Sobolev embedding for 0 < p < 1

The classical Sobolev–Poincaré inequality
µZ ¶1/p∗ µZ ¶1/p
p∗
|u − uB | dx ≤C |∇u|p ,
B B

holds when 1 ≤ p < n. It is easy to see that it fails when 0 < p < 1, and even
a weaker version of the Poincaré inequality fails for the range 0 < p < 1. For an
explicit example, see Buckley and Koskela [16].
However Buckley and Koskela, [16], and in a more general setting Buckley,
Koskela and Lu, [19], proved that if u is a solution to the equation div A(x, Xu) = 0
in a John domain with respect to the Carnot–Carathéodory metric, then u satisfies
a Sobolev–Poincaré inequality for any 0 < p < s, where s is given by condition (21).
As we will see, one of the results of the paper, Theorem 9.7, which states that
for any 0 < p < s, a p-Poincaré inequality implies a Sobolev–Poincaré inequality,
can be regarded as an abstract version of the above result. In particular this gives
a new proof of the result of Buckley, Koskela and Lu.
More precisely, assume that X = (X1 , . . . , Xm ) are locally Lipschitz vector fields
in IRn . Assume that the associated Carnot–Carathéodory metric satisfies condition
(54), the Lebesgue measure is doubling with respect to the Carnot–Carathéodory
82 PIOTR HAJLASZ AND PEKKA KOSKELA

e ≤ Cd |B|
distance i.e. |2B| e for all metric balls B
e ⊂ IRn , and that condition (21) is
satisfied.
In addition we assume that the 1-Poincaré inequality is satisfied i.e. there is
C > 0 and σ ≥ 1 such that
Z Z
|u − uBe| dx ≤ CP r |Xu| dx ,
Be e
σB

e ⊂ IRn and all u ∈ C ∞ (σ B).

for all metric balls B e
Let A : IRn × IRm → IRm be a Carathéodory function such that

|A(x, ξ)| ≤ C1 |ξ|q−1 , A(x, ξ) · ξ ≥ C2 |ξ|q ,

where 1 < q < ∞ is given. (Observe that in contrast with the previous section we
do not allow a weight ω.)
The following result is a variant of the result of Buckley, Koskela and Lu, [19].

Theorem 13.2 Let Ω ⊂ IRn be a John domain with respect to the Carnot–
Carathéodory metric. Then for any 0 < p < s there is a constant C > 0 such
that if u is a solution to equation div A(x, Xu) = 0, in Ω, then
µZ ¶1/p∗ µZ ¶1/p
p∗ p
inf |u − c| dx ≤ Cdiam Ω |Xu| .
c∈IR Ω Ω

The constant C depends on n, p, s, Cb , Cd , C1 , C2 , CP , and CJ only.

Proof. Let u be a solution to div A(x, Xu) = 0 in Ω. The first fact we need is that
the gradient |Xu| of the solution u satisfies a weak reverse Hölder inequality. This
is well known. However, for the sake of completeness, we provide a proof.
e let ηR be a cut-off function such that 0 ≤ ηR ≤ 1,
Given a metric ball B,
ηR |Be ≡ 1, ηR ≡ 0 outside 2Be and |XηR | ≤ 1/R. Using the distance function with
respect to ρ we easily construct a cut-off function with the metric Lipschitz constant
1/R. Then the estimate |XηR | ≤ 1/R follows from Corollary 11.8.
Now, using the test function (u − u2Be)ηR , where 2B e ⊂ Ω is any metric ball and
ηR is the associated cut-off function, we conclude from a standard computation the
Caccioppoli estimate Z Z
C
|Xu|q ≤ q |u − u2Be|q .
e
B R 2Be
Then we estimate the right hand side by the Sobolev–Poincaré inequality and con-
e with 2B
clude that there is p < q such that for all metric balls B e ⊂⊂ Ω
µZ ¶1/q µZ ¶1/p
|Xu|q dx ≤C |Xu|p dx . (79)
B e
2B

This inequality is known under the name weak reverse Hölder inequality.
 SOBOLEV MET POINCARÉ 83

It is well known that the weak reverse Hölder inequality has the self-improving
property: if inequality (79) holds for some 0 < p < q and all B e with 2B
e ⊂ Ω, then
for any 0 < p < q there is a new constant C such that (79) holds for any B e with
2Be ⊂ Ω, see [19, Lemma 1.4]. This together with the 1-Poincaré inequality shows
that the pair u, g satisfies a p-Poincaré inequality in Ω for any p > 0. Hence the
claim follows from Theorem 9.7. The proof is complete.

14 Appendix
Here we collect the results in the measure theory that are needed in the paper. All
the material is standard. Since we could not find a single reference that would cover
the material we need, we have made all the statements precise and sometimes we
have even given proofs. Good references are Federer [68], Mattila [193], and Simon
[229].
In the appendix we do not assume that the measure µ is doubling.

14.1 Measures.
Throughout the paper by a measure we mean an outer measure, and by a Borel
measure, an outer, Borel-regular measure i.e., such a measure µ on a metric space
(X, d) that all Borel sets are µ-measurable and for every set A there exists a Borel
set B such that A ⊂ B and µ(A) = µ(B). In the case of a Borel measure
S∞ we also
assume that the measure of each ball is strictly positive and X = j=1 Uj , where
Uj are open sets with µ(Uj ) < ∞.
Note that if the space X is locally compact, separable and µ(K) < ∞ for every
compact set K, then X can be written as a union of a countable family of open sets
with finite measure.

Theorem 14.1 Suppose that µ is a Borel measure on (X, d). Then

µ(A) = inf µ(U )
U ⊃A
U −open

for all subsets A ⊂ X, and

µ(A) = sup µ(C)
C⊂A
C−closed

for all measurable sets A ⊂ X.

For the proof, see [68, Theorem 2.2.2. and Section 2.2.3], [193, Theorem 1.10] or
[229, Theorem 1.3].
If the space is locally compact and separable, the supremum over closed sets in
the above theorem equals to the supremum over compact sets.
84 PIOTR HAJLASZ AND PEKKA KOSKELA

As a corollary to the above theorem we obtain the following well known result.

Theorem 14.2 If µ is a Borel measure on a metric space (X, d), then for every
1 ≤ p < ∞, continuous functions are dense in Lp (X, µ).

Proof. Simple functions are dense in Lp (X, µ), see [219, Theorem 3.13], so it suffices
to prove that characteristic functions can be approximated by continuous functions.
Fix ε > 0. If A ⊂ X is measurable, µ(A) < ∞, then there exists a closed set C
and an open set U such that C ⊂ A ⊂ U , µ(U \ C) < ε. Now by Urysohn’s lemma,
there exists a continuous function ϕε on X such that 0 ≤ ϕε ≤ 1, ϕε |C = 1 and
ϕε |X\U = 0. Then obviously kχA − ϕε kp → 0 as ε → 0. This completes the proof.
In order to have a variety of Borel measures one usually assumes that the space
be locally compact. In the definition of the doubling measure one does not assume
anything about the metric space. However, as we will see, the existence of a doubling
measure is such a strong condition that the space is “almost” locally compact.
We say that a subset A of a metric space (X, d) is an ε-net if for every x ∈ X
there is y ∈ A with d(x, y) < ε. A metric space (X, d) is called totally bounded if
for each ε > 0 there exists a finite ε-net.
The following two lemmas are well known.

Lemma 14.3 A metric space (X, d) is compact if and only if it is complete and
totally bounded.

Lemma 14.4 Every metric space is isometric to a dense subset of a complete met-
ric space.

The first lemma follows from a direct generalization of the proof that every
bounded sequence of real numbers contains a convergent subsequence, while the
second lemma follows by adding the “abstract limits” of Cauchy sequences to the
space.

Theorem 14.5 If a metric space (X, d) admits a Borel measure µ which is locally
uniformly positive in the sense that for every bounded set A ⊂ X and every ε > 0
inf µ(B(x, ε)) > 0, (80)
x∈A

then (X, d) is isometric to a dense subset of a locally compact separable metric space.

Proof. The fact that X is a union of countably many

S∞open sets of finite measure
and (80) imply that X can be covered by balls X = j=1 Bj with µ(2Bj ) < ∞.
According to Lemma 14.3 and Lemma 14.4 it suffices to prove that for every
j = 1, 2, . . . and every ε > 0 there is a finite ε-net in Bj . This, however, easily
follows from (80) and the condition µ(2Bj ) < ∞. The proof is complete.
It is of fundamental importance to note that the doubling condition implies local
uniform positivity of the measure, as follows from the following result.
 SOBOLEV MET POINCARÉ 85

Lemma 14.6 Let µ be a Borel measure on a metric space X. Assume that µ is

doubling, in the following sense, on a bounded subset Y ⊂ X: there is a constant
Cd ≥ 1 such that
µ(B(x, 2r)) ≤ Cd µ(B(x, r)),
whenever x ∈ Y , and r ≤ diam Y . Then

µ(B(x, r)) ≥ (2 diam Y )−s µ(Y )rs ,

for s = log2 Cd , x ∈ Y and r ≤ diam Y .

The above lemma together with Theorem 14.5 shows that doubling spaces are iso-
metric to dense subsets of locally compact separable metric spaces. The analogous
result holds also when the measure is doubling on some open set only. Note that a
doubling measure is finite on bounded sets.
The above remark together with the following result shows that a doubling
measure can be extended to a doubling measure on the larger locally compact
space.

Proposition 14.7 Let Y ⊂ X be a dense subset of a metric space (X, d). Let µ be
a Borel measure on (Y, d), finite on bounded sets. Then there exists a unique Borel
measure µ̄ on (X, d) such that

µ̄(U ) = µ(U ∩ Y )

for every open set U ⊂ X. Moreover, if µ is doubling on (Y, d), then µ̄ is doubling
on (X, d) with the same doubling constant.

Proof. Set µ̄(A) = inf B⊃A, B−Borel µ(B ∩ Y ) for an arbitrary set A ⊂ X. One
easily verifies that µ̄ is a Borel measure on (X, d). This proves the existence of the
measure. The uniqueness follows form Theorem 14.1.
Assume now that µ is doubling. Then obviously µ̄ is doubling with the same
doubling constant on all balls centered at Y . Since any ball in X can be “approxi-
mated” by balls centered at Y , the result follows.
Remark. If we removed the
S∞assumption that µ be finite on bounded sets, Y would
still have the property Y = j=1 Uj , where the sets Uj are open with µ(Uj ) < ∞.
However then this property would not necessarily be true for µ̄. For example, let
Y be the complement of a Cantor set in [0, 1], and X = [0, 1]. Then Y consists of
countable many intervals. Equip Y with a measure so that the measure of each of
the intervals is 1. Then X cannot be decomposed into a countable number of open
sets with finite µ̄-measure.

14.2 Uniform integrability.

In this section µ is an arbitrary measure on a set X.
86 PIOTR HAJLASZ AND PEKKA KOSKELA

Assume that µ(X) < ∞. We say that a Rfamily {uα }α∈I of µ-measurable func-
tions on X is uniformly integrable if supα∈I X |uα | dµ < ∞ and
Z
lim sup |uα | dµ = 0.
µ(A)→0 α∈I A

The following theorem is due to Vallée Poussin. For a proof, see Dellacherie and
Meyer [63], or Rao and Ren [212].

Theorem 14.8 Let µ be a measure on a set X with µ(X) < ∞ and let {uα }α∈I be a
family of µ-measurable functions. Then the following two conditions are equivalent.

1. The family {uα }α∈I is uniformly integrable.

2. There exists a convex smooth function F : [0, ∞) → [0, ∞) such that F (0) = 0,
F (x)/x → ∞ as x → ∞ and
Z
sup F (|uα |) dµ < ∞.
α∈I X

The following well known result is a very useful criteria for convergence in L1 .

Proposition
R 14.9 Let µ(X) < ∞. If un are uniformly integrable on X and un → u
a.e., then X |un − u| dµ → 0.

Proof. It follows directly from Egorov’s theorem and the definition of uniform
integrability that the sequence un is a Cauchy sequence in the L1 norm, and hence
un converges to u in L1 . The proof is complete.

14.3 Lp and Lpw spaces.

In the following two theorems µ is an arbitrary σ-finite measure on X. The first
result is known as Cavalieri’s principle.

Theorem 14.10 If p > 0 and u is measurable, then

Z Z ∞
|u|p dµ = p tp−1 µ(|u| > t) dt.
X 0

The claim follows from Fubini’s theorem applied to X × [0, ∞).

We say that a measurable function u belongs to the Marcinkiewicz space Lpw (X)
if there is m > 0 such that
µ(|u| > t) ≤ mt−p for all t > 0. (81)
R
If u ∈ Lp (X), then (81) with m = X |u|p dµ is known as Chebyschev’s inequality,
so Lp (X) ⊂ Lpw (X). The converse inclusion does not hold. However, the following,
well known result holds.
 SOBOLEV MET POINCARÉ 87

Theorem 14.11 If µ(X) < ∞ then Lpw (X) ⊂ Lq (X) for all 0 < q < p. Moreover,
if u satisfies (81), then
µ ¶1/p
qm
kukLq (X) ≤ 21/q µ(X)1/q−1/p . (82)
p−q

Proof. Fix t0 > 0. Using Theorem 14.10 and the estimates µ(|u| > t) ≤ µ(X)
for t ≤ t0 and µ(|u| > t) ≤ mt−p for t > t0 we get
Z µZ t0 Z ∞ ¶
qm q−p
q
|u| dµ ≤ q q−1
t µ(X) dt + m tq−p−1
dt = tq0 µ(X) + t .
X 0 t0 p−q 0

Then inequality (82) follows by choosing t0 = (qm/(p − q))1/p µ(X)−1/p .

14.4 Covering lemma.

Theorem 14.12 (5r-covering lemma.) Let B be a family of balls in a metric space
(X, d) with sup{diam B : B ∈ B} < ∞. Then there is a pairwise disjoint subcollec-
tion B 0 ⊂ B such that [ [
B⊂ 5B
B∈B B∈B0

If (X, d) is separable, then B0 is countable and we can represent B0 as a sequence

B0 = {Bi }∞i=1 , and so
[ ∞
[
B⊂ 5Bi .
B∈B i=1

See Federer [68, 2.8.4-6], Simon [229, Theorem 3.3], or Ziemer [263, Theorem 1.3.1]
for a clever proof.

14.5 Maximal function.

Assume that the measure µ is doubling on an open set Ω ⊂ X. The following
theorem is a version of the well known maximal theorem of Hardy, Littlewood and
Wiener.

Theorem 14.13 (Maximal theorem.) If X, Ω and µ are as above, and the maximal
function MΩ u is defined as in the introduction, then
R
1. µ({x ∈ Ω : MΩ u(x) > t}) ≤ Ct−1 Ω
|u| dµ for t > 0 and
2. kMΩ ukLp (Ω) ≤ CkukLp (Ω) for 1 < p ≤ ∞.

In the first inequality the constant C depends on Cd only, while in the second one
it depends on Cd and p.
88 PIOTR HAJLASZ AND PEKKA KOSKELA

For a proof in the case of Lebesgue measure, see Stein, [233, Chapter 1]. We assume
that the reader is familiar with that proof and we show how to modify the argument
in order to cover our setting. It sufices to prove 1.; one then proceeds as in [233].
Inequality 1. would follow from this inequality for the restricted maximal function
MΩ,R u provided we prove it with a constant C that does not depend on R. To this
end, note first that the doubling condition implies that Ω is separable and hence
the second part of Theorem 14.12 applies. Then the argument from the case of the
Lebesgue measure works without any changes. We had to work with the restricted
maximal function in order to obtain a suitable covering consisting of balls with radii
less than R (if we did not have the upper bound for the radii, we could not apply
Theorem 14.12).
We will also need a more general result. For c ≥ 1 and x ∈ Ω define Fc (x) as the
family of all measurable sets E ⊂ Ω such that E ⊂ B(x, r) and µ(B(x, r)) ≤ cµ(E)
for some r > 0. Then we define a new maximal function by
Z
c
MΩ u(x) = sup |u| dµ.
E∈Fc (x) E

Obviously McΩ u ≤ cMΩ u, and thus we obtain as a corollary to Theorem 14.13 the
following result.

Corollary 14.14 Theorem 14.13 holds with MΩ replaced by McΩ . The only differ-
ence is that now the constants C in Theorem 14.13 depend also on c.

14.6 Lebesgue differentiation theorem.

We say that a sequence of nonempty sets {Ei }∞ i=1 converges to x if there exists a
sequence of radii ri > 0 such that Ei ⊂ B(x, ri ) and ri → 0 as i → ∞.

Theorem 14.15 Let µ be doubling on Ω ⊂ X and u ∈ L1loc (Ω, µ). Then for µ-a.e.
x ∈ Ω we have Z
lim u(y) dµ(y) = u(x). (83)
r→0 B(x,r)

Moreover, if we fix c ≥ 1, then for µ-a.e. x ∈ Ω and every sequence of sets Ei ∈

Fc (x), i = 1, 2, . . . that converges to x we have
Z
lim u(y) dµ(y) = u(x). (84)
i→∞ E
i

See [233, Chapter 1] for a proof in the case of the Lebesgue measure in IRn . The
same argument works also in our setting as it only relies on two facts: the weak
type inequality for the maximal function (see Theorem 14.13 and Corollary 14.14)
and the density of continuous functions in L1 (see Theorem 14.2).
 SOBOLEV MET POINCARÉ 89

Let µ be doubling on Ω ⊂ X. Given u ∈ L1loc (Ω, µ) it is often convenient to

identify u with the representative given everywhere by the formula
Z
u(x) := lim sup u(y) dµ(y) . (85)
r→0 B(x,r)

Theorem 14.15 shows that taking the limit above only modifies u on a set of
measure zero. We say that x ∈ Ω is a Lebesgue point of u if
Z
lim |u(y) − u(x)| dµ(y) = 0,
r→0 B(x,r)

where u(x) is given by (85). It follows from Theorem 14.15 that almost all points
of Ω are Lebesgue points of u. Observe that if x ∈ Ω is a Lebesgue point of u, then
both (83) and (84) are true.

References
[1] Adams, D. A., Hedberg, L. I.: Function spaces and potential theory. Grundlehren
der Mathematischen Wissenschaften, 314, Springer-Verlag, Berlin, 1996.
[2] Auscher P., Coulhon T.: Gaussian lower bounds for random walks from elliptic
regularity. Ann. Inst. H. Poincaré Probab. Stat., 35 (1999), 605–630.
[3] Arnol’d, V. I.: Ordinary Differential Equations. Springer, 1992.
[4] Bakry, D., Coulhon, T., Ledoux, M., Saloff-Coste, L.: Sobolev inequalities
in disguise. Indiana Univ. Math. J. 44 (1995), 1033–1074.
[5] Barlow, M. T., Bass, R. F.: Coupling and Harnack inequalities for Sierpinski
carpets. Bull. Amer. Math. Soc. (N.S.) 29 (1993), 208–212.
[6] Bass, H.: The degree of polynomial growth of finitely generated groups. Proc.
London Math. Soc. 25 (1972), 603–614.
[7] Bellaı̈che, A.: The tangent space in sub-Riemannian geometry. In: [240], pp. 1–78.
[8] Biroli, M., Mosco, U.: Formed de Dirichlet et estimationes structurelles dans les
mileux discontinues, C. R. Acad. Sci. Paris, 313 (1991), 593–598.
[9] Biroli, M. Mosco, U.: Sobolev inequalities on homogeneous spaces. Pot. Anal. 4
(1995), 311–324.
[10] Biroli, M., Mosco, U.: A Saint-Venant type principle for Dirichlet forms on
discontinuous media. Ann. Mat. Pura Appl. 169 (1995), 125–181.
[11] Bobkov, S. G., Houdré, C.: Some connections between Sobolev–type inequalities
and isoperimetry. Mem. Amer. Math. Soc. 129 (1997), no. 616, 1–111.
[12] Bojarski, B.: Remarks on Sobolev imbedding inequalities, in: Proc. of the Confer-
ence on Complex Analysis (Joensu 1987), pp. 52–68, Lecture Notes in Math. 1351
(Springer 1988).
[13] Boman, J.: Lp –estimates for very strongly elliptic systems. Department of Mathe-
matics, University of Stockholm, Sweden (1982), Reports no 29.
[14] Bony, J. M.: Principe du maximum, inégalité de Harnack et du probleme de Cauchy
pour les opérateurs elliptiques dégénérés, Ann. Inst. Fourier (Grenoble) 19 (1969),
277–304.
90 PIOTR HAJLASZ AND PEKKA KOSKELA

[15] Bourdon, M., Pajot, H.: Poincaré inequalities and quasiconformal structure on
the boundary of some hyperbolic buildings, Proc. Amer. Math. Soc., 127 (1999),
2315–2324.
[16] Buckley, S. M., Koskela, P.: Sobolev–Poincaré inequalities for p < 1. Indiana
Univ. Math. J. 43 (1994), 221–240.
[17] Buckley, S. M., Koskela, P.: Sobolev–Poincaré implies John. Math. Res. Lett.
2 (1995), 577–593.
[18] Buckley, S. M., Koskela, P.: Geometric criteria for inequalities of Sobolev-
Poincaré type. Internat. Math. Res. Notices 2 (1995), 577–593.
[19] Buckley, S. M., Koskela, P., Lu, G.: Subelliptic Poincaré inequalities: the case
p < 1. Publ. Mat. 39 (1995), 313–334.
[20] Buckley, S. M., Koskela, P., Lu, G.: Boman equals John. in: Proceedings of
the XVI Rolf Nevanlinna Colloquium (Joensuu 1995), pp. 91–99, de Gruyter, Berlin,
1996.
[21] Buckley, S., O’Shea, J.: Weighted Trudinger-type inequalities, Indiana Univ.
Math. J., to appear.
[22] Busemann, H.: The geometry of geodesics, Academic Press, New York 1955.
[23] Buser, P.: A note on the isoperimetric constant. Ann. Sc. Ecole Norm. Sup. 15
(1982), 213–230.
[24] Calderón, A. P., Scott, R.: Sobolev type inequalities for p > 0. Studia Math.
62 (1978), 75–92.
[25] Capogna, L.: Regularity of quasilinear equations in the Heisenberg group. Comm.
Pure Appl. Math. 50 (1997), 867–889.
[26] Capogna, L.: Regularity for quasilinear equations and 1-quasiconformal maps in
Carnot groups. Math. Ann. 313 (1999), 263–295.
[27] Capogna, L., Danielli, D., Garofalo, N.: An imbedding theorem and the Har-
nack inequality for nonlinear subelliptic equations. Comm. Partial Differential Equa-
tions 18 (1993), 1765–1794.
[28] Capogna, L., Danielli, D., Garofalo, N.: Subelliptic mollifiers and character-
ization of Rellich and Poincaré domains. Rend. Sem. Mat. Univ. Polit. Torino 54
(1993), 361–386.
[29] Capogna, L., Danielli, D., Garofalo, N.: The geometric Sobolev embedding for
vector fields and the isoperimetric inequality. Comm. Anal. Geom. 2 (1994), 203–215.
[30] Capogna, L., Danielli, D., Garofalo, N.: Subelliptic mollifiers and a basic
pointwise estimate of Poincaré type. Math. Z. 226 (1997), 147–154.
[31] Capogna, L., Danielli, D., Garofalo, N.: Capacitary estimates and the local
behavior of solutions to nonlinear subelliptic equations. Amer. J. Math. 118 (1996),
1153–1196.
[32] Carathéodory, C.: Untersuchungen über die Grundlagen der Thermodynamik.
Math. Ann. 67 (1909), 355–386.
[33] Chavel, I.: Riemannian Geometry — A modern introduction. Cambridge Tracts in
Mathematics, vol. 108, Cambridge Univ. Press, 1993.
[34] Cheeger, J.: Differentiability of Lipschitz functions on metric measure spaces.
Geom. Funct. Anal. 9 (1999) 428–517.
[35] Cheeger, J., Gromov, M., Taylor, M.: Finite propagation speed, kernel esti-
mates for functions of the Laplace operator, and the geometry of complete Rieman-
nian manifolds. J. Differential Geom. 17 (1982), 15–53.
 SOBOLEV MET POINCARÉ 91

[36] Chemin, J. Y., Xu, C. J.: Inclusions de Sobolev en calcul de Weyl–Hörmander et

champs de vecteurs sous-elliptiques. Ann. Scient. Éc. Norm. Sup. 30 (1997), 719–
751.
[37] Chen, R. Li, P.: On Poincaré type inequalities. Trans. Amer. Math. Soc. 349
(1997), 1561–1585.
[38] Chernikov, V. M., Vodop’yanov, S. K.: Sobolev Spaces and hypoelliptic equa-
tions I,II. Siberian Advances in Mathematics. 6 (1996) no. 3, 27–67, no. 4, 64–96.
Translation from: Trudy In-ta matematiki RAN. Sib. otd-nie. 29 (1995), 7–62.
[39] Chiarenza, F., Frasca, M.: A note on weighted Sobolev inequality. Proc. Amer.
Math. Soc. 93 (1985), 703–704.
[40] Chow, W. L.: Über Systeme non linearen partiellen Differentialgleichungen erster
Ordung. Math. Ann. 117 (1939), 98–105.
[41] Christ, M.: Lectures on Singular Integral Operators, Regional Conference Series in
Math. 77, 1989.
[42] Chua, S. K.: Weighted Sobolev inequalities on domains satisfying the chain condi-
tion. Proc. Amer. Math. Soc. 117 (1993), 449–457.
[43] Chung, F. R. K.: Spectral graph theory. CBMS Regional Conference Series in
Mathematics, 92; American Mathematical Society, 1997
[44] Chung, F. R. K., Yau, S. T.: A Harnack inequality for homogneous graphs and
subgraphs. Comm. Anal. Geom. 2 (1994), 627–640.
[45] Citti, G.: Semilinear Dirichlet problem involving critical exponents for the Kohn
Laplacian. Ann. Mat. Pura Appl. 169 (1995), 375–392.
[46] Citti, G., Di Fazio, G.: Hölder continuity of the solutions for operators which are
sum of squares of vector fields plus a potential. Proc. Amer. Math. Soc. 122 (1994),
741–750.
[47] Citti, G., Garofalo, N., Lanconelli, E.: Harnack’s inequality for sum of squares
plus potential. Amer. J. Math. 115 (1993), 699–734.
[48] Coifman, R., Weiss, G.: Analyse harmonique sur certains espaces homogenes,
Lecture Notes in Math. 242, Springer 1971.
[49] Coifman, R., Weiss, G.: Extensions of Hardy spaces and their use in analysis.
Bull. Amer. Math. Soc. 83 (1977), 569–645.
[50] Colding, T. H., Minicozzi, W. P. II: Harmonic functions on manifolds. Ann. of
Math. 146 (1997), 725–747.
[51] Colding, T. H., Minicozzi, W. P. II: Weyl type bounds for harmonic functions.
Invent. Math. 131 (1988), 257–298.
[52] Coulhon T.: Analysis on graphs with regular volume growth, to appear in Pro-
ceedings of the Cortona conference on Random walks and discrete potential theory,
Cambridge U.P.
[53] Coulhon, T.: Dimension at infinity for Riemannian manifolds. Potential Anal. 4
(1995), 335–344.
[54] Coulhon, T.: Espaces de Lipschitz et inégalités de Poincaré. J. Funct. Anal. 136
(1996), 81–113.
[55] Coulhon T.: Heat kernels on non-compact Riemannian manifolds: a partial survey.
Séminarie de Théorie Spectrale et Géométrie, 15 (1996-1997), Institut Fourier, 1998,
167–187.
[56] Coulhon T., Grigor’yan A.: Random walks on graphs with regular volume.
Geom. Funct. Anal. 8 (1998), 656–701.
92 PIOTR HAJLASZ AND PEKKA KOSKELA

[57] Coulhon, T., Ledoux, M.: Isopérimétrie, décroissance du noyau de la chaleur et

transformations de Riesz: un contre-exemple. Arkiv Mat. 32 (1994), 63–77.
[58] Coulhon, T., Saloff-Coste, L.: Isopérimétrie pour les groupes et les variétés.
Rev. Mat. Iberoamericana 9 (1993), 293–314.
[59] Coulhon, T., Saloff-Coste, L.: Variétés riemanniennes isométriques à l’infini.
Rev. Mat. Iberoamericana. 11 (1995), 687–726.
[60] Danielli, D.: A compact embedding theorem for a class of degenerate Sobolev
spaces. Rend. Sem. Mat. Univ. Pol. Torino 49 (1991), 399–420.
[61] Danielli, D., Garofalo, N., Nhieu, D.-M.: Trace inequalities for Carnot–
Carathéodory spaces and applications. Ann. Scuola Norm. Sup. Pisa 27 (1998),
195–252.
[62] David, G., Semmes, S.: Strong A∞ weights, Sobolev inequalities, and quasiconfor-
mal mappings, in: Analysis and Partial Differential Equations, pp. 101–111, Lecture
Notes on Pure and Applied Math. 122, Marcel Dekker, 1990.
[63] Dellacherie, C., Meyer, P. A.: Probabilities and potential. North-Holland Math-
ematics Studies, 29. North-Holland Publishing Co., 1978.
[64] Delmotte, T.: Inégalité de Harnack elliptique sur les graphes, Colloq. Math. 72
(1997), 19–37.
[65] Delmotte T.: Parabolic Harnack inequality and estimates of Markov chains on
graphs. Rev. Mat. Iberoamericana 15 (1999), 181–232.
[66] Evans, L. C., Gariepy, R. F.: Measure Theory and Fine Properties of Functions.
Studies in Advanced Mathematics, 1992.
[67] Fabes, E. B., Kenig, C. E., Serapioni, R. P.: The local regularity of solutions
of degenerate elliptic equations. Comm. Partial Differential Equations 7 (1982), 77–
116.
[68] Federer, H.: Geometric measure theory (Springer 1969).
[69] Fefferman, C., Phong, D. H.: Subelliptic eigenvalue problems. In Proceedings of
the conference in harmonic analysis in honor of Antoni Zygmund, Wadsworth Math.
Ser., Wadsworth, Belmont, California, 1981, 590–606.
[70] Fefferman, C., Sanchez-Calle, A.: Fundamental solutions for second order
subelliptic operators. Ann. Math. 124 (1986), 247–272.
[71] Folland, G. B.: A fundamental solution for a subelliptic operator. Bull. Amer.
Math. Soc. 79 (1973), 373–376.
[72] Folland, G. B.: Subelliptic estimates and function spaces on nilpotent Lie groups.
Ark. Mat. 13 (1975), 161–207.
[73] Folland, G. B., Stein, E. M.: Hardy spaces on homogeneous groups. Mathemat-
ical Notes vol. 28, Princeton Univ. Press, Princeton, New Jersey, 1982.
[74] Franchi, B.: Weighted Sobolev–Poincaré inequalities and pointwise inequalities
for a class of degenerate elliptic equations. Trans. Amer. Math. Soc. 327 (1991),
125–158.
[75] Franchi, B., Gallot, S., Wheeden, R. L.: Sobolev and isoperimetric inequalities
for degenerate metrics. Math. Ann. 300 (1994), 557–571.
[76] Franchi, B., Gutiérrez , C. E., Wheeden, R. L.: Weighted Sobolev–Poincaré
inequalities for Grushin type operators. Comm. Partial Differential Equations 19
(1994), 523–604.
[77] Franchi, B., Hajlasz, P., Koskela, P.: Definitions of Sobolev classes on metric
spaces. Ann. Inst. Fourier (Grenoble), to appear.
 SOBOLEV MET POINCARÉ 93

[78] Franchi, B., Lanconelli, E.: Hölder regularity theorem for a class of non uni-
formly elliptic operators with measurable coefficients. Ann. Scuola Norm. Sup. Pisa
10 (1983), 523–541.
[79] Franchi, B., Lanconelli, E.: An imbedding theorem for Sobolev spaces related to
non smooth vector fields and Harnack inequality. Comm. Partial Differental Equa-
tions 9 (1984), 1237–1264.
[80] Franchi, B., Lu, G., Wheeden, R.: Representation formulas and weighted
Poincaré inequalities for Hörmander vector fields. Ann. Inst. Fourier (Grenoble)
45 (1995), 577-604.
[81] Franchi, B., Lu, G., Wheeden, R. L.: A relationship between Poincaré type
inequalities and representation formulas in spaces of homogeneous type. Internat.
Math. Res. Notices (1996), No. 1, 1–14.
[82] Franchi, B., Pérez, C., Wheeden R.L.: Self–improving properties of John–
Nirenberg and Poincaré inequalities on spaces of homogeneous type, J. Funct. Anal.
153 (1998), 108–146.
[83] Franchi, B., Serapioni, R.: Pointwise estimates for a class of strongly degenerate
elliptic operators: a geometric approach. Ann. Scuola Norm. Sup. Pisa 14 (1987),
527–568.
[84] Franchi, B., Serapioni R., Serra Cassano, F: Meyers–Serrin type theorems
and relaxation of variational integrals depending on vector fields Houston Math. J.
22 (1996), 859–889.
[85] Franchi, B., Serapioni R., Serra Cassano, F: Approximation and imbedding
theorems for weighted Sobolev spaces associated with Lipschitz continuous vector
fields Boll. Un. Mat. Ital. (7) 11–B (1997), 83–117.
[86] Franchi, B., Wheeden, R.: Some remarks about Poincaré type inequalities and
representation formulas in metric spaces of homogeneous type. J. Inequalities and
Appl., to appear.
[87] Friedrichs, K. O.: The identity of weak and strong extensions of differential op-
erators. Trans. Amer. Math. Soc. 55 (1944), 132–151.
[88] Gallot, S.: Inégalités isopérimetriques, courbure de Ricci et invariants
géométriques. I, II. C. R. Acad. Sci. Paris Ser. I Math. 296 (1983), 333–336, 365–368.
[89] Garattini, R.: Harnack inequality on homogeneous spaces, preprint.
[90] Garofalo, N., Lanconelli, E.: Existence and nonexistence results for semilinear
equations on the Heisenberg group. Indiana Univ. Math. J. 41 (1992), 71–98.
[91] Garofalo, N., Nhieu, D. M.: Lipschitz continuity, global smooth approximation
and extension theorems for Sobolev functions in Carnot–Carathéodory spaces. J.
Analyse Math. 74 (1998), 67–97.
[92] Garofalo, N., Nhieu, D. M.: Isoperimetric and Sobolev inequalities for Carnot-
Carathéodory spaces and the existence of minimal surfaces, Comm. Pure Appl. Math.
49 (1996), 1081–1144.
[93] Gatto, A. E., Vági, S.: On the exponential integrability of fractional integrals on
spaces of homogeneous type. Colloq. Math. 64 (1993), 121–127.
[94] Gatto, A. E., Vági, S.: On functions arising as potentials on spaces of homoge-
neous type. Proc. Amer. Math. Soc. 125 (1997), 1149–1152.
[95] Genebahsvili, I., Gogatishvili, A., Kokilashvili, V., Krbec, M.: Weighted
theory for integral transforms on spaces of homogeneous type. Pitman Monographs
and Surveys in Pure and Applied Math. 92. Addison Wesley Longman Limited 1988.
94 PIOTR HAJLASZ AND PEKKA KOSKELA

[96] Gilbarg, D., Trudinger, N.: Elliptic Partial Differential Equations of Second
Order, Springer–Verlag 1983.
[97] Goldshtein, V. M., Reshetnyak, Yu. G.: Quasiconformal mappings and Sobolev
spaces. Mathematics and its Applications (Soviet Series), 54. Kluwer Academic Pub-
lishers Group, Dordrecht, 1990. (Translated and revised from the 1983 Russian orig-
inal.)
[98] Gole, C., Karidi, R.: A note on Carnot geodesics in nilpotent Lie groups. J.
Dynam. Control Systems 1 (1995), 535–549.
[99] Grigor’yan, A.: The heat equation on noncompact Riemannian manifolds. Mat.
Sb. 182 (1991), 55–87 (in Russian).
[100] Gromov, M.: Groups of polynomial growth and expanding maps. Publ. Math.
I.H.E.S. 53 (1981), 53–78.
[101] Gromov, M.: Carnot–Carathéodory spaces seen from within. In: [240], pp. 79–318.
[102] Hajlasz, P.: Sobolev spaces on an arbitrary metric space. Potential Anal. 5 (1996),
403–415.
[103] Hajlasz, P.: Geometric approach to Sobolev spaces and badly degenerated ellip-
tic equations. The Proceedings of Banach Center Minisemester: Nonlinear Analysis
and Applications. (N.Kenmochi, M.Niezgódka, P.Strzelecki, eds.) GAKUTO Inter-
national Series; Mathematical Sciences and Applications, vol. 7 (1995), 141–168.
[104] Hajlasz, P., Kinnunen, J.: Hölder quasicontinuity of Sobolev functions on metric
spaces. Rev. Mat. Iberoamericana. 14 (1998), 601–622.
[105] Hajlasz, P., Koskela, P.: Sobolev meets Poincaré. C. R. Acad. Sci. Paris Ser. I
Math. 320 (1995), 1211–1215.
[106] Hajlasz, P., Koskela, P.: Isoperimetric inequalities and imbedding theorems in
irregular domains. J. London Math. Soc. 58 (1998), 425–450.
[107] Hajlasz, P., Martio, O.: Traces of Sobolev functions on fractal type sets and
characterization of extension domains. J. Funct. Anal. 143 (1997), 221–246.
[108] Hajlasz, P., Strzelecki, P.: Subelliptic p-harmonic maps into spheres and the
ghost of Hardy spaces, Math. Ann. 312 (1998), 341–362.
[109] Han Y. S.: Triebel–Lizorkin spaces on spaces of homogeneous type, Studia Math.
108 (1994), 247–273.
[110] Han, Y. S. and Sawyer, E. T.: Littlewood–Paley Theory on Spaces of Homoge-
neous Type and Classical Function Spaces. Mem. Amer. Math. Soc. 110 (1994), no.
530, 1–126.
[111] Hanson, B. and Heinonen, J.: An n-dimensional space that admits a Poincaré
inequality but has no manifold points. Proc. Amer. Math. Soc., to appear.
[112] Hartman, P.: Ordinary Differential Equations. Birkhauser, 1982.
[113] Hebisch, W., Sikora, A.: A smooth subadditive homogeneous norm on a homo-
geneous group. Studia Math. 96 (1990), 321–236.
[114] Heinonen, J.: Calculus on Carnot groups. Ber. Univ. Jyväskylä Math. Inst. 68
(1995), 1–31.
[115] Heinonen, J.: Lectures on analysis on metric spaces, preprint.
[116] Heinonen, J., Holopainen, I.: Quasiregular maps on Carnot groups. J. Geom.
Anal. 7 (1997), 109–148.
[117] Heinonen, J., Koskela, P.: Definitions of quasiconformality. Inventiones Math.
120 (1995), 61–79.
 SOBOLEV MET POINCARÉ 95

[118] Heinonen, J., Koskela, P.: Quasiconformal maps on metric spaces with controlled
geometry. Acta Math. 181 (1998), 1–61.
[119] Heinonen, J., Koskela, P.: A note on Lipschitz functions, upper gradients and
the Poincaré inequality. New Zealand J. Math. 28 (1999), 37–42.
[120] Heinonen, J., Kilpeläinen, T., Martio, O.: Nonlinear Potential Theory of De-
generate Elliptic Equations, Oxford Univ. Press, 1993.
[121] Herman, R.: Differential Geometry and the Calculus of Variations, Academic Press,
1968.
[122] Hille, E.: Lectures on Ordinary Differential Equations. Addison-Wesley Pub. Co.
1968.
[123] Holopainen, I.: Rough isometries and p-harmonic functions with finite Dirichlet
integral. Rev. Mat. Iberoamericana 10 (1994), 143–176.
[124] Holopainen, I.: Volume growth, Green’s functions, and parabolicity of ends. Duke
Math. J. 97 (1999), 319–346.
[125] Holopainen, I., Rickman, S.: Ricci curvature, Harnack functions, and Picard type
theorems for quasiregular mappings. Analysis and Topology, pp. 315–326. World Sci.
Publishing, River Edge, NJ, 1998.
[126] Holopainen, I., Soardi, P.: p-harmonic functions on graphs and manifolds.
Manuscripta Math. 94 (1997), 95–110.
[127] Holopainen, I., Soardi, P. M.: A strong Liouville theorem for p-harmonic func-
tions on graphs. Ann. Acad. Sci. Fenn. Math. 22 (1997), 205–226.
[128] Hörmander, L.: Hypoelliptic second order differential equations, Acta Math. 119
(1967), 147–171.
[129] Hörmander, L., Melin, A.: Free systems of vector fields. Ark. Mat. 16 (1978),
83–88.
[130] Hurri, R.: Poincare domains in IRn . Ann. Acad. Sci. Fenn. Ser. A I Math. Disser-
tationes (1988), No. 71, 42 pp.
[131] Hurri-Syrjänen, R.: A weighted Poincaré inequality with a doubling weight. Proc.
Amer. Math. Soc. 126 (1998), 545–552.
[132] Iwaniec, T., Nolder, C.: Hardy–Littlewood inequality for quasiregular mappings
in certain domains in IRn . Ann. Acad. Sci. Fenn. Ser. A I Math. 10 (1985), 267–282.
[133] Jerison, D.: The Poincaré inequality for vector fields satisfying Hörmander’s con-
dition. Duke Math. J. 53 (1986), 503–523.
[134] Jerison, D., Lee, J. M.: Intrinsic CR coordinates and the CR Yamabe problem.
J. Diff. Geom. 29 (1989), 303–343.
[135] Jerison, D., Lee, J. M.: Extremals for the Sobolev inequality on the Heisenberg
group and the CR Yambe problem. J. Amer. Math. Soc. 1 (1988), 1–13.
[136] Jerison, D., Sanchez-Calle, A.: Subelliptic, second order differential operators.
In:Complex analysis, III (College Park, Md., 1985–86). pp. 46–77, Lecture Notes in
Math., 1277, Springer, 1987.
[137] John, F.: Rotation and strain. Comm. Pure Appl. Math. 4 (1961), 391–414.
[138] Jonsson, A.: Besov spaces on closed subsets of IRn . Trans. Amer. Math. Soc. 341
(1994), 355–370.
[139] Jonsson, A.: Brownian motion on fractals and function spaces. Math. Z. 222 (1996),
495–504.
[140] Jonsson, A., Wallin, H.: Function Spaces on Subsets of IRn , Harwood Acad.
Publ., 1984.
96 PIOTR HAJLASZ AND PEKKA KOSKELA

[141] Jonsson, A., Wallin, H.: Boundary value problems and Brownian motion on
fractals. Chaos Solitons Fractals 8 (1997), 191–205.
[142] Jost, J.: Equilibrium maps between metric spaces. Calc. Var. 2 (1994), 173–205.
[143] Jost, J.: Generalized harmonic maps between metric spaces, in: Geometric analysis
and the calculus of variations (J. Jost, ed.), International Press, 1966, pp. 143–174.
[144] Jost, J.: Generalized Dirichlet forms and harmonic maps. Calc. Var. 5 (1997), 1–19.
[145] Jost, J.: Nonlinear Dirichlet forms. In: New directions in Dirichlet forms. pp. 1–47.
AMS/PI Stud. Adv. Math. 8, Amer. Math. Soc., Providence, RI, 1998.
[146] Jost, Xu, C. J.: Subelliptic harmonic maps, Trans. Amer. Math. Soc. 350 (1998),
4633–4649.
[147] Kalamajska, A.: On compactness of embedding for Sobolev spaces defined on
metric spaces. Ann. Acad. Sci. Ser. A. I Math. 24 (1999), 123–132.
[148] Kallunki, S., Shanmugalingam, N.: Modulus and continuous capacity, preprint.
[149] Kanai, M.: Rough isometries and the parabolicity of Riemannian manifolds. J.
Math. Soc. Japan 38 (1986), 227–238.
[150] Karidi, R.: Geometry of balls in nilpotent Lie groups. Duke Math. J. 74 (1994),
301–317.
[151] Kigami, J.: Harmonic calculus on p.c.f. self-similar sets. Trans. Amer. Math. Soc.
335 (1993), 721–755.
[152] Kigami, J.: Hausdorff dimension of self-similar sets and shortest path metrics. J.
Math. Soc. Japan 47 (1995), 381–404.
[153] Kigami, J.: Harmonic metric and Dirichlet form on the Sierpinski gasket. In:
Asymptotic problems in probability theory: stochastic models and diffusions on frac-
tals (Sanda/Kyoto, 1990), pp. 201–218, Pitman Res. Notes Math. Ser., 283, Long-
man Sci. Tech., Harlow, 1993.
[154] Kigami, J., Lapidus, M. L.: Weyl’s problem for the spectral distribution of Lapla-
cian on p.c.f. self-similar fractals. Comm. Math. Phys. 158 (1993), 93–125.
[155] Kilpeläinen, T.: Smooth approximation in weighted Sobolev spaces. Comment.
Math. Univ. Carolinae 38 (1995), 1–8.
[156] Kilpeläinen, T., Kinnunen, J., Martio.: Sobolev spaces with zero boundary
values on metric spaces. Potential Anal., to appear.
[157] Kinnunen, J., Martio, O.: The Sobolev capacity on metric spaces. Ann. Acad.
Sci. Fenn. Math. 21 (1996), 367–382.
[158] Kohn, R. V.: New integral estimates for deformations in terms of their nonlinear
strains. Arch. Rational Mech. Anal. 78 (1982), 131–172.
[159] Korányi, A, Reimann, H. M.: Foundations of the theory of quasiconformal map-
pings on the Heisenberg group. Adv. Math. 111 (1995), 1–87.
[160] Koskela, P.: Removable sets for Sobolev spaces Ark. Mat. 37 (1999), 291–304.
[161] Koskela, P. MacManus, P.: Sobolev classes and quasisymmetric maps. Studia
Math. 131 (1998), 1–17.
[162] Kozlov, S. M.: Harmonization and homogenization on fractals. Comm. Math.
Phys. 153 (1993), 339–357.
[163] Kusuoka, S., Stroock, D. W.: Applications of the Malliavin calculus. III. J. Fac.
Sci. Univ. Tokyo Sect. IA Math. 34 (1987), 391–442.
[164] Laakso, T.: Ahlfors Q-regular spaces with arbitrary Q admitting weak Poincaré
inequality. Geom. Funct. Anal., to appear.
 SOBOLEV MET POINCARÉ 97

[165] Lanconelli, E.: Stime sub-ellittiche e metriche riemanniane singolari. Seminario di

Analysi Matematica. Istituto Matematico dell’Universita di Bologna. A.A. 1982-83.
[166] Lanconelli, E., Morbidelli, D.: On the Poincaré inequality for vector fields,
Ark. Mat., to appear.
[167] Lapidus, M. L.: Analysis on fractals, Laplacians on self-similar sets, noncommu-
tative geometry and spectral dimensions. Topol. Methods Nonlinear Anal. 4 (1994),
137–195.
[168] Lapidus, M. L.: Fractals and vibrations: can you hear the shape of a fractal drum?
In: Symposium in Honor of Benoit Mandelbrot (Curacao, 1995). Fractals 3 (1995),
725–736.
[169] Lawler, G. F.: Estimates for differences and Harnack inequality for difference
operators coming from random walks with symmetric, spatially inhomogeneous, in-
crements. Proc. London Math. Soc. 63 (1991), 552–568.
[170] Li, P., Schoen, R.: Lp and mean value properties of subharmonic functions on
Riemannian manifolds. Acta Math. 153 (1984), 279–301.
[171] Li, P., Wang, J.: Convex hull properties of harmonic maps. J. Diff. Geom. 48
(1988), 487–530.
[172] Li, P., Yau, S. T.: Estimates of eigenvalues of a compact Riemannian manifold.
In: Geometry of the Laplace operator (Proc. Sympos. Pure Math., Univ. Hawaii,
Honolulu, Hawaii, 1979), pp. 205–239, Proc. Sympos. Pure Math., 36, Amer. Math.
Soc., Providence, R.I., 1980.
[173] Long, R. L., Nie, F. S.: Weighted Sobolev inequality and eigenvalue estimates
of Schrödinger operators. In: Harmonic Analysis. (Tianjin, 1988). Lecture Notes in
Math. 1494, Springer, 1991.
[174] Liu, S., Sussman, H.: Shortest paths for sub-Riemannian metrics on rank-two
distributions. Mem. Amer. Math. Soc. 118 (1995), no. 564, 1–104.
[175] Lu, G.: Weighted Poincaré and Sobolev inequalities for vector fields satisfying
Hörmander’s condition and applications. Rev. Mat. Iberoamericana 8 (1992), 367–
439.
[176] Lu, G.: Existence and size estimates for the Green’s functions of differential opera-
tors constructed from degenerate vector fields. Comm. Partial Differential Equations
17 (1992), 1213–1251.
[177] Lu, G.: On Harnack’s inequality for a class of strongly degenerate Schrödinger
operators formed by vector fields. J. Diff. Int. Equations 7 (1994), 73–100.
[178] Lu, G.: The sharp Poincaré inequality for free vector fields: An endpoint result.
Rev. Mat. Iberoamericana 10 (1994), 453–466.
[179] Lu, G.: Embedding theorems on Campanato-Morrey spaces for vector fields and
applications. C. R. Acad. Sci. Paris Ser. I Math. 320 (1995), 429–434.
[180] Lu, G.: Embedding theorems into Lipschitz and BMO spaces and applications to
quasilinear subelliptic differential equations. Publ. Mat. 40 (1996), 301–329.
[181] Lu, G.: A note on a Poincaré type inequality for solutions to subelliptic equations.
Comm. Partial Differential Equations 21 (1996), 235–254.
[182] Luukkainen, J., Saksman, E.: Every complete doubling metric space carries a
doubling measure. Proc. Amer. Math. Soc. 126 (1998), 531–534.
[183] Macias, R. A., Segovia, C.: Lipschitz functions on spaces of homogeneous type.
Adv. Math. 33 (1979), 257–270.
[184] MacManus, P., Pérez, C.: Generalized Poincaré inequalities: sharp self-improving
properties. Internat. Math. Res. Notices 1998, No. 2, 101–116.
98 PIOTR HAJLASZ AND PEKKA KOSKELA

[185] MacManus, P., Pérez, C.: Trudinger inequalities without derivatives. Trans.
Amer. Math. Soc., to appear.
[186] Maheux, P., Saloff-Coste, L.: Analyse sur les boules d’un opérateur sous-
elliptique. Math. Ann. 303 (1995), 713–740.
[187] Malý, J., Ziemer, W.: Fine Regularity Properties of Solutions of Elliptic Par-
tial Differential Equations. Mathematical Surveys and Monographs, 51. Americal
Mathematical Society, 1997.
[188] Manfredini, M.: Compact embedding theorems for generalized Sobolev spaces.
Atti Accad. Naz. Lincei Cl. Sci. Fis. Mat. Natur. Rend. Lincei (9) Mat. Appl. 4
(1993), 251–263.
[189] Marchi, S.: Hölder continuity and Harnack inequality for De Giorgi classes related
to Hörmander vector fields. Ann. Mat. Pura Appl. 168 (1995), 171–188.
[190] Margulis, G. A., Mostow, G. D.: The differential of a quasiconformal mapping
of a Carnot–Carathéodory space. Geom. Funct. Anal. 5 (1995), 402–433.
[191] Martio, O.: John domains, bilipschitz balls and Poincaré inequality. Rev. Roumaine
Math. Pures Appl. 33 (1988), 107–112.
[192] Martio, O., Sarvas, J.: Injectivity theorems in plane and space. Ann. Acad. Sci.
Fenn., Ser. A I Math. 4 (1978/1979) 383–401.
[193] Mattila, P.: Geometry of Sets and Measures in Euclidean Spaces. Cambridge Stud-
ies in Advanced Mathematics, 44. Cambridge Univ. Press, 1995.
[194] Maz’ya, V. G.: A theorem on the multidimensional Schrödinger operator. Izv.
Akad. Nauk. 28 (1964), 1145–1172 (in Russian).
[195] Maz’ya, V. G.: Sobolev spaces (Springer 1985).
[196] Merkov, A. B.: Second order elliptic equations on graphs. Math. USSR-Sb. 55
(1986), 493–509.
[197] Metz, V., Sturm, K. T.: Gaussian and non-Gaussian estimates for heat kernels on
the Sierpinski gasket. In: Dirichlet forms and stochastic processes (Beijing, 1993),
pp. 283–289, de Gruyter, Berlin, 1995.
[198] Milnor, J.: A note on curvature and fundamental group. J. Diff. Geom. 2 (1968),
1–7.
[199] Mitchell, J.: On Carnot–Carathéodory metrics. J. Diff. Geom. 21 (1985), 35–45.
[200] Morbidelli, D.: Fractional Sobolev norms and structure of Carnot–Carathéodory
balls for Hörmander vector fields, Studia Math. to appear.
[201] Mosco, U.: Variational metrics on self-similar fractals. C. R. Acad. Sci. Paris Ser.
I Math. 321 (1995), 715–720.
[202] Moser, J.: On Harnack’s theorem for elliptic differential equations. Comm. Pure
Appl. Math. 14 (1961), 577–591.
[203] Mostow, G.: Quasiconformal mappings in n-space and the rigidity of hyperbolic
space forms. Publ. Math. I.H.E.S. 34 (1968), 53–104.
[204] Nagel, A., Stein, E. M. and Wainger, S.: Balls and metrics defined by vector
fields I: Basic properties, Acta Math. 55 (1985), 103–147.
[205] Näkki, R., Väisälä, J.: John disks, Exposition. Math. 9 (1991), 3–43.
[206] Nhieu, D. M.: Extensions of Sobolev spaces on the Heisenberg group. C. R. Acad.
Sci. Paris Ser. I Math. 321 (1995), 1559–1564.
[207] Nhieu, D. M.: An extension theorem for the Folland–Stein Sobolev spaces on strat-
ified nilpotent Lie groups, preprint.
 SOBOLEV MET POINCARÉ 99

[208] Oleinik, O. A., Radkevic, E. V.: Second order equations with nonnegative char-
acteristic form. Plenum Press, New York-London, 1973.
[209] Pansu, P.: Une inégalité isopérimétrique sur le groupe d’Heisenberg. C. R. Acad.
Sci. Paris Ser. I Math. 295 (1982), 127–131.
[210] Pansu, P.: Métriques de Carnot–Carathéodory et quasiisométries des espaces
symétriques de rans un. Ann. of Math. 129 (1989), 1–60.
[211] Peetre, J.: A remark on Sobolev spaces. The case 0 < p < 1. J. Approx. Theory
13 (1975), 218–228.
[212] Rao, M. M., Ren, Z. D.: Theory of Orlicz Spaces. Pure and Applied Math. 146.
Marcel Dekker, Inc. 1991.
[213] Rashevsky, P. K.: Any two points of a totally nonholonomic space may be con-
nected by an admissible line. Uch. Zap. Ped. Inst. im. Liebknechta, Ser. Phys. Math.
2 (1938), 83–94 (in Russian).
[214] Renardy, M., Rogers, R.: An Introduction To Partial Differential Equations.
Texts in Applied Mathematics, 13. Springer, 1993.
[215] Rigoli, M., Salvatori, M., Vignati, M.: A note on p-subharmonic functions on
complete manifolds. Manuscripta Math. 92 (1997), 339–359.
[216] Rigoli, M., Salvatori, M., Vignati, M.: Liouville properties on graphs. Mathe-
matika 44 (1997), 133–148.
[217] Rigoli, M., Salvatori, M., Vignati, M.: Subharmonic functions on graphs. Israel
J. Math. 99 (1997), 1–27.
[218] Rothschild, L. P., Stein, E. M.: Hypoelliptic differential operators and nilpotent
groups. Acta Math. 137 (1976), 247–320.
[219] Rudin, W.: Real and Complex Analysis, Third edition, Mc Graw–Hill, New York
1987.
[220] Saloff-Coste, L.: Uniformly elliptic operators on Riemannian manifolds. J. Diff.
Geom. 36 (1992), 417–450.
[221] Saloff-Coste, L.: A note on Poincaré, Sobolev, and Harnack inequalities. Inter-
nat. Math. Res. Notices (1992) no. 2, pp. 27–38.
[222] Saloff-Coste, L.: On global Sobolev inequalities. Forum Math. 6 (1994), 271–286.
[223] Saloff-Coste, L.: Parabolic Harnack inequality for divergence form second order
differential operators. Potential Anal. 4 (1995), 429–467.
[224] Sánchez-Calle, A.: Fundamental solutions and geometry of sums of squares of
vector fields. Invent. Math. 78 (1984), 143–160.
[225] Schinzel, A.: An analogue of Harnack’s inequality for discrete superharmonic func-
tions. Demonstratio Math. 11 (1978), 47–60.
[226] Semmes, S.: Finding curves on general spaces through quantitative topology, with
applications to Sobolev and Poincaré inequalities. Selecta Math. (N.S.) 2 (1996),
155–295.
[227] Semmes, S.: Some remarks about metric spaces, spherical mappings and their
derivatives. Publ. Mat. 40 (1996), 411–430.
[228] Shanmugalingam, N.: Newtonian spaces; a generalization of Sobolev spaces. Rev.
Mat. Iberoamericana, to appear.
[229] Simon, L.: Lectures on Geometric Measure Theory, Proc. of Centre for Math. Anal.
Austral. Nat. Univ. Vol. 3, 1983.
[230] Soardi, P. M.: Rough isometries and Dirichlet finite harmonic functions on graphs.
Proc. Amer. Math. Soc. 119 (1993), 1239–1248.
100 PIOTR HAJLASZ AND PEKKA KOSKELA

[231] Smith, W., Stegenga, D. A.: Hölder domains and Poincaré domains. Trans.
Amer. Math. Soc. 319 (1990), 67–100.
[232] Smith, W., Stegenga, D. A.: Sobolev imbeddings and integrability of harmonic
functions in Hölder domains. in: Potential theory (de Gruyter, Berlin 1992).
[233] Stein, E.: Singular integrals and differentiability properties of functions (Princeton
Univ. Press 1970).
[234] Stein, E.: Harmonic Analysis: Real-Variable Methods, Orthogonality and Oscilla-
tory Integrals (Princeton Univ. Press 1993).
[235] Strichartz, R. S.: Sub-Riemannian geometry, J. Diff. Geom. 24 (1986), 221–263.
Corrections: J. Diff. Geom. 30 (1989), 595–596.
[236] Sturm, T. K.: Analysis on local Dirichlet spaces. I. Recurrence, conservativeness
and Lp -Liouville properties. J. Reine Angew. Math. 456 (1994), 173–196.
[237] Sturm, T. K.: Analysis on local Dirichlet spaces. II. Upper Gaussian estimates
for the fundamental solutions of parabolic equations. Osaka J. Math. 32 (1995),
275–312.
[238] Sturm, K, T.: On the geometry defined by Dirichlet forms. In: Seminar on Stochas-
tic Analysis, Random Fields and Applications (Ascona, 1993), pp. 231–242 Progr.
Probab., 36, Birkhauser, Basel, 1995.
[239] Sturm, K. T.: Analysis on local Dirichlet spaces III. The parabolic Harnack in-
equality. J. Math. Pures Appl. 75 (1996), 273–297.
[240] Sub-Riemannian geometry. Progress in Mathematics, 144. Edited by Andre Bellaiche
and Jean-Jacques Risler. Birkhauser Verlag, Basel, 1996.
[241] Trudinger, N.: On imbedding into Orlicz spaces and some applications. J. Math.
Mech. 17 (1967), 473–483.
[242] Tam, L. F.: Liouville properties of harmonic maps. Math. Res. Lett. 2 (1995), 719–
735.
[243] Tartar, L.: Imbedding theorems in Lorentz spaces and applications. Boll. Unione
Mat. Ital. Sez. B Artic. Ric. Mat. (8) 1 (1998), 479–500.
[244] Triebel, H.: Fractals and spectra. Monographs in Mathematics 91, Birkhauser
1997.
[245] Tyson, J.: Quasiconformality and quasisymmetry in metric measure spaces. Ann.
Acad. Sci. Fenn. Math. 23 (1998), 525–548.
[246] Väisälä, J.: Lectures on n-dimensional quasiconformal mappings, Lecture Notes in
Math. 299, Springer 1971.
[247] Väisälä, J.: Uniform domains. Tohoku Math. J. 40 (1988), 101–118.
[248] Varopoulos, T.: Brownian motion and random walks on manifolds. Ann. Inst.
Fourier (Grenoble) 34 (1984), 243–269.
[249] Varopoulos, T.: Fonctions harmoniques sur les groupes Lie. C. R. Acad. Sci. Paris
Ser. I Math. 309 (1987), 519–521.
[250] Varopoulos, T.: Distances associees aux operateurs sous-elliptiques du second
ordre. C. R. Acad. Sci. Paris Ser. I Math. 309 (1989), 663–667.
[251] Varopoulos, T., Saloff-Coste, L., Coulhon, T.: Analysis and geometry on
groups (Cambridge Univ. Press 1992).
[252] Vodop’yanov, S. K.: Monotone functions on Carnot groups and their applications.
Sib. Math. Zh. 37 (1996), 1269–1295 (in Russian); English transl. Sib. Math. J. 37
(1996), 1113–1136.
 SOBOLEV MET POINCARÉ 101

[253] Vodop’yanov, S. K., Greshnov, A. V.: Analytic properties of quasiconformal

mappings on Carnot groups. Sib. Math. Zh. 36 (1995), 1317–1327 (in Russian);
English transl. Siberian Math. J. 36 (1995), 1142–1151.
[254] Vodop’yanov, S. K., Markina, I. G.: Exceptional sets for solutions of subelliptic
equations. Sibirsk. Mat. Zh. 36 (1995), 805–818 (in Russian); English trasl. Siberian
Math. J. 36 (1995), 694–706.
[255] Volberg, A. L., Konyagin, S. V.: A homogeneous measure exists on any com-
pactum in IRn . Dokl. Akad. Nauk SSSR 278 (1984), 783–785 (in Russian); English
transl. Soviet Math. Dokl. 30 (1984), 453–456.
[256] Volberg, A. L., Konyagin, S. V.: On measures with the doubling condition. Izv.
Acad. Nauk SSSR 51 (1987), 666–675 (in Russian); English transl. Math. USSR Izv.
30 (1988), 629–638.
[257] Wu, J. M.: Hausdorff dimension and doubling measures on metric spaces. Proc.
Amer. Math. Soc. 126 (1998), 1453–1459.
[258] Xu, C. J.: Subelliptic variational problems. Bull. Soc. Math. France 118 (1990),
147–159.
[259] Xu, C. J.: Regularity for quasilinear second-order subelliptic equations. Comm.
Pure Appl. Math. 45 (1992), 77–96.
[260] Xu, C. J., Zuily, C.: Higher interior regularity for quasilinear subelliptic systems.
Calc. Var. 5 (1997), 323–343.
[261] Yau, S. T.: Harmonic functions on complete Riemannian manifolds. Comm. Pure
Appl. Math 28 (1975), 201–228.
[262] Zhou, X. Y.: Green function estimates and their applications to the intersections
of symmetric random walks. Stochastic Process. Appl. 48 (1993), 31–60.
[263] Ziemer, W.: Weakly differentiable functions, Graduate Texts in Mathematics 120
(Springer 1989).

Piotr Hajlasz: Institute of Mathematics, Warsaw University, ul. Banacha 2, 02–

097 Warszawa, Poland; E-mail: [email protected]

Pekka Koskela: University of Jyväskylä, Department of Mathematics, P.O. Box

35, FIN-40351 Jyväskylä, Finland; E-mail: [email protected]