HugoAimar Book Distance and Measure Analysis and PDE

Hugo Aimar
Instituto de Matemática Aplicada del Litoral-IMAL
Distance and Measure in

Analysis and PDE
October 11, 2017
Springer
Use the template dedic.tex together with the
Springer document class SVMono for
monograph-type books or SVMult for
contributed volumes to style a quotation or a
dedication at the very beginning of your book
in the Springer layout
Foreword
Use the template foreword.tex together with the Springer document class SVMono
(monograph-type books) or SVMult (edited books) to style your foreword in the
Springer layout.
The foreword covers introductory remarks preceding the text of a book that are
written by a person other than the author or editor of the book. If applicable, the
foreword precedes the preface which is written by the author or editor of the book.
Place, month year Firstname Surname
vii
Preface
Use the template preface.tex together with the Springer document class SVMono
(monograph-type books) or SVMult (edited books) to style your preface in the
Springer layout.
A preface is a book’s preliminary statement, usually written by the author or ed-
itor of a work, which states its origin, scope, purpose, plan, and intended audience,
and which sometimes includes afterthoughts and acknowledgments of assistance.
When written by a person other than the author, it is called a foreword. The
preface or foreword is distinct from the introduction, which deals with the subject
of the work.
Customarily acknowledgments are included as last part of the preface.
Place(s), Firstname Surname

month year Firstname Surname
ix
Acknowledgements
Use the template acknow.tex together with the Springer document class SVMono
(monograph-type books) or SVMult (edited books) if you prefer to set your ac-
knowledgement section as a separate chapter instead of including it as last part of
your preface.
xi
Contents
1 Topological, uniform and quasi-metric spaces. . . . . . . . . . . . . . . . . . . . . . 7

1.1 Topology, neighborhoods and bases. . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2 The topology induced by the neighborhood system associated to a
proto-distance g. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3 Uniform spaces. Uniform topology. . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4 Metrizability of uniform spaces. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.5 Uniform structures induced by proto-distances. . . . . . . . . . . . . . . . . . . 16
1.6 Metrizability of the uniform structure induced by a proto-distance . . 18
1.7 Topology on quasi-metric spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.8 Metrization of quasi-metric spaces. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.9 Iteration of the triangle inequality for a quasi-distance. . . . . . . . . . . . 23
1.10 Second approach to the metrization of quasi-metric spaces.
Metrization of η-metric spaces. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.11 Quasi-distance of order α, open balls and Lipschitz functions on
quasi-metric spaces. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.12 Quasi-distance and balls. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
1.13 Quasi-distance and bands. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.14 Compact, totally bounded and complete sets in quasi-metric spaces. 37
1.15 Hausdorff measure and dimension of sets in quasi-metric spaces. . . . 38
1.16 Improving the smoothness of the boundaries of balls. . . . . . . . . . . . . 41
1.17 Dyadic type partitions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
1.18 Product of quasi-metric spaces. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
1.19 Problems and Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2 The weak homogeneity property on quasi-metric spaces. Assouad

metric dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2.1 Weak homogeneity property. Assouad dimension. . . . . . . . . . . . . . . . 57
2.2 Metric and topological properties of quasi-metric spaces satisfying
WHP. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.3 Hausdorff and Assouad dimensions. . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
xiii
xiv Contents
2.4 Wiener type covering lemmas. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

2.5 Whitney type covering lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
2.6 Lipschitz type partition of unity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
2.7 Whitney extension of continuous functions. . . . . . . . . . . . . . . . . . . . . . 75
2.8 Whitney extension of Lipschitz functions. . . . . . . . . . . . . . . . . . . . . . . 77
2.9 Lipschitz test functions for a fractional order theory of
distributions on quasi-metric spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
2.10 Problems and Comments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3 Spaces of homogeneous type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

3.1 Spaces of homogeneous type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.2 Boundedness and finite measure of a space of homogeneous type. . . 89
3.3 Atomic singletons are countable and isolated. . . . . . . . . . . . . . . . . . . . 91
3.4 Regularity of the measure. Approximation of indicators by
continuous functions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.5 Finite upper type and doubling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
3.6 More covering lemmas. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
3.7 Subspaces of a space of homogeneous type. . . . . . . . . . . . . . . . . . . . . 105
3.8 Normal spaces and normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
3.9 Some basic integrals of Newton - Riesz kernels on normal spaces. . . 121
3.10 Reverse doubling. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
3.11 The regularity of the functions ηx (r). . . . . . . . . . . . . . . . . . . . . . . . . . . 127
3.12 Hausdorff distance convergence of ε-nets. . . . . . . . . . . . . . . . . . . . . . . 132
3.13 A second view to the doubling property. Weak convergence and
doubling. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
3.14 Convergence of spaces of homogeneous type. . . . . . . . . . . . . . . . . . . . 134
3.15 On doubling set functions on a quasi metric space . . . . . . . . . . . . . . . 136
3.16 Some general geometric consequences of the weak homogeneity
property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
3.17 Spaces of homogeneous type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
4 Some examples of quasi-metric measure spaces. . . . . . . . . . . . . . . . . . . . 143

4.1 Some elementary and classical examples. . . . . . . . . . . . . . . . . . . . . . . 143
4.1.1 Some basic translation invariant structures on Rn . . . . . . . . . . 143
4.1.2 More translations invariant structures generalized
homogeneity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
4.1.3 Regular Riviére- Vitali families. . . . . . . . . . . . . . . . . . . . . . . . . 146
4.1.4 The Heinsenberg group. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
4.1.5 A dyadic distance on Euclidian Spaces. . . . . . . . . . . . . . . . . . . 149
4.2 Some examples from degenerate elliptic and parabolic equations. . . 150
4.2.1 Basic A p theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
4.2.2 A weighted parabolic space of homogeneous type and
degenerate heat equations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
4.2.3 A space of homogeneous type modeling the analytic
context associated to the Monge-Ampére equation. . . . . . . . . 158
Contents xv
4.3 Some basic examples from geometric measure theory. . . . . . . . . . . . . 161

4.3.1 The Cantor set. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
5 The Hardy-Littlewood maximal function and the differentiation

theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
5.1 The Hardy-Littlewood maximal operator . . . . . . . . . . . . . . . . . . . . . . . 167
5.2 Weak type (1, 1) for M . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
5.3 L p bounded of Mf . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
5.4 Lipschitz functions with bounded support are dense in L p
(0 < p < +∞) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
5.5 The Lebesgue differentiation theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 171
5.6 Dyadic tilings revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
5.7 Dyadic Calderón-Zygmund decomposition of the space associated
to a given integrable function f and to a given level λ > 0 . . . . . . . . 177
5.8 The dyadic maximal function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
6 Weighted norm inequalities for the Hardy-Littlewood maximal

operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
7 Some techniques for the L2 boundedness of linear operators . . . . . . . . 193

7.1 Young’s Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
7.2 Applications of Cotlar’s Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
7.3 An extension of Cotlar’s Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
7.4 Krein Lemma and Wittman’s method . . . . . . . . . . . . . . . . . . . . . . . . . . 199
7.5 An application of Cotlar’s lemma to regularization of Haar basis . . . 207
A Chapter Heading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217

A.1 Section Heading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
A.1.1 Subsection Heading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
Acronyms
Use the template acronym.tex together with the Springer document class SVMono
(monograph-type books) or SVMult (edited books) to style your list(s) of abbrevia-
tions or symbols in the Springer layout.
Lists of abbreviations, symbols and the like are easily formatted with the help of
the Springer-enhanced description environment.
ABC Spelled-out abbreviation and definition
BABI Spelled-out abbreviation and definition
CABR Spelled-out abbreviation and definition
xvii
Introduction
Electrostatic and gravitational potentials provide a natural starting point to introduce

singular integral type operators defined in different geometrical contexts. Green’s
representation formula for smooth functions states that any C 2 function on a do-
main Ω can be written as the sum of three potentials: (1) the Newtonian potential
generated by a space charge density, (2) the single layer potential generated by a
surface charge density, and (3) the double layer potential generated by a surface
dipole density. The underlying idea is to charge in some specific way a subset of the
3-d euclidian space, to model the system through a Schwartz distribution and then
to observe the specific mathematical properties of the potential and the electrostatic
field so generated.
Assume that q is the value of an electrical charge at the point x0 of the three
dimensional euclidian space R3 . The absolute value of the attracting or repelling
force on another charge q at the point x 6= x0 is given by Coulomb - Newton’s law
|q| |q|
|F(x, q)| = .
|x − x0 |2
The force is given by the vector field
qq x − x0
F(x, q) = .
|x − x0 |2 |x − x0 |
It is usual to say that the vector E(x) = F(x, 1) is the electric field generated by q
at x0 . In terms of Schwartz distributions we can write
E(x) = q [δx0 ∗ v](x) ,

where δx0 is the Dirac unit mass at x0 , v(x) = | x |−3 x, and
δx0 ∗ v = (δx0 ∗ v1 , δx0 ∗ v2 , δx0 ∗ v3 )

xi
with vi = |x|3
the components of v, i = 1, 2, 3.
1
2 Acronyms
The superposition principle allows us to say that a bounded system of electrical

charges is, in a very general sense, a compactly supported Schwartz distribution T
0
(T ∈ E , E = C ∞ ). Then, the electrical field generated by that system is given by
the convolution of T and v
E = T ∗ v = (T ∗ v1 , T ∗ v2 , T ∗ v3 ),
xi
where vi = |x|3
.
Of course the convolution defining E makes sense even for more
general distributions T.
The potential energy of the electric field E is any scalar field Φ defined in R3 for
which ∇Φ = E. So that, a potential for the electric field E = T ∗ v generated by a
0
distribution T ∈ E is given by
Φ = T∗N ,
1
with N(x) = |x| .
Three basic examples are in order.
F IRST E XAMPLE When the distribution of charges T is given by a C0∞ (R3 ) den-
sity f , then, both the potential and the electric field are functions,
f (y)
Z
Φ(x) = dy , and
R3 |x − y|
x−y
Z
E(x) = f (y) dy.
R3 |x − y|3
Let us point out that, in this case, both Φ and Ei , the i-th component of the electric
field are bounded by a fractional integral of the absolute value of f . In fact
|Φ(x)| ≤ I2 (| f |)(x) and |Ei (x)| ≤ I1 ( | f | )(x) ,

R
where Iα (g)(x) = R3 | x − y |α−3 g(y) dy is the Riesz potential of order α of g. On
the other hand, the Jacobian of
E(x)
or the second derivatives of Φ(x), can be explicitly computed by taking the cor-
responding derivatives of the kernel | x − y |−1 of Φ(x) when the point x does not
belong to the support of the charge density f . For example, if x 6∈ supp f , then
∂ 2Φ (x1 − y1 )(x2 − y2 )
Z
(x) = 3 f (y) dy.
∂ x1 ∂ x2 R3 | x − y |5
The same question when x belongs to the support of f can be considered as the
starting point of the Calderón-Zygmund theory of singular integral operators in the
euclidian space.
Let us observe that there is no better bound than |K12 (x − y)| ≤ | x − y |−3 for the
absolute value of the kernel K12 (x, y) = (x1 − y1 ) (x2 − y2 ) | x − y |−5 . So that, since
Acronyms 3
the dimension of the space is three, | x − y |−3 is the first not even locally integrable
power of | x − y |−1 , the reciprocal of the euclidian distance.
S ECOND E XAMPLE If a solid conductive body in R3 is charged, the repelling
forces inside the solid produce the migration of charges to the surface ∂ Ω of the
solid Ω . This physical situation can be modelled by a Borel measure µ supported
on ∂ Ω . A special case occurs when Ω is smooth and µ is given by a surface density
f with respect to the area measure dσ on ∂ Ω . In this case, we have dµ = f dσ with
f : ∂ Ω −→ R an appropriate function. In this way we get the single layer potential
induced by the surface density f on ∂ Ω , given by
f (y)
Z
Φ(x) = dσ (y),
∂Ω |x−y|
and the corresponding electric field
x−y
Z
E(x) = f (y) dσ (y).
∂Ω | x − y |3
Notice that both integrals are well defined and absolutely convergent if f is continu-
ous and x 6∈ ∂ Ω . Actually, since the dimension of ∂ Ω is two, the integral defining Φ
is absolutely convergent even for x ∈ ∂ Ω . For the integral defining the field E(x), let
us observe that if we take only into account the orders of magnitude of the kernel,
neglecting the fact that regularity of ∂ Ω would imply the almost orthogonality of
x − y and dσ (y), we have a kernel of size | x − y |−2 on a set of dimension two. So
that in this situation the value of the field precisely on the charged surface can not
be directly evaluated and has to be considered as a principal value giving rise to a
singular integral defined on the non-euclidian space ∂ Ω .
T HIRD E XAMPLE In the two previous examples electric potentials and electric
fields are nothing but scalar or vector convolutions of some special measures with
the basic kernels | x − y |−1 and (x − y)| x − y |−3 . Other mathematical objects with
more complex structure are also good models for some physical standard situations.
This is the case of the electrical dipole of size α and momentum u ∈ S2 . Let us
denote by δ0 the Dirac unit mass distribution located at the origin 0 of R3 . As a
Schwartz distribution, δ0 has all the partial derivatives ∂∂δx0i i = 1, 2, 3, well defined.
So that, the gradient ∇δ0 is well defined as a vector distribution and the directional
derivative
3
∂ δ0
Du δ 0 = ∑ u i
i=1 ∂ xi
is well defined as a scalar distribution. The distribution T = α Du δ0 is called the
dipole at the origin with momentum u ∈ S2 and size α ∈ R. It is not difficult to show
how this definition fits, as a limit, the standard physical idea of dipole as a system
of opposite, large and close charges supported on the line through the origin with
direction u. The potential of the field generated by the dipole T is of course given
by
4 Acronyms
Φ = N ∗ T = α(N ∗ Du δ0 ) = α (Du N ∗ δ0 ) = αu · ∇N.

The third type of potential entering in Green’s formula for C 2 functions can be
constructed using the dipoles as building blocks. Let Ω be a smooth domain in R3 :
bounded, connected with boundary of class C 1 . Given P ∈ ∂ Ω , let us denote by
ν(P) the unit outer normal at P.
We now construct a special charge system supported on ∂ Ω by attaching to each
point P ∈ ∂ Ω a dipole of size f (P) dσ (P) and momentum ν(P). From the superpo-
sition principle, we have a vector field given for x ∈
/ ∂ Ω by the electrostatic poten-
tial.
Z
Φ(x) = dΦP (x)
ZP∈∂ Ω
= f (P)Dν(P) N(x − P) dσ (P)
∂Ω
ν(P).(x − P)
Z
= f (P) dσ (P) .
∂Ω | x − P |3
R
Let us observe that with µi (E) = E∩∂ Ω f (P)νi (P) dσ (P), νi (P) the i-th compo-
nent of ν(P), i = 1, 2, 3; and µ = (µ1 , µ2 , µ3 ), we have for Φ(x) the expression
Φ = div µ ∗ N ,
3
∂ µi
where, of course, div µ = ∑ . All these formulas are absolutely convergent in-
i=1 ∂ xi
tegrals for x ∈
/ ∂ Ω . The immediate question about a possible meaning for Φ on ∂ Ω
posses again an existence problem for singular integrals on the non-euclidian setting
determined by ∂ Ω .
In a general sense, distance and measure are intimately related from this point
of view of classical potential theory which is also the source of the very deep links
between real, harmonic and Fourier analysis on one side and PDE’s on the other.
The mean value property harmonic functions defined in the n-dimensional euclidian
space
1
Z
µ(x) = µ(y) dy,
B(x, r) B(x,r)
is providing an elementary identification of the Laplace differential operator ∆ =

n
∂2
∑ ∂ x2 = div grad, with a geometrical structure on Rn given by a measure dx
i=1 i
(Lebesgue) and a distance (Euclidian) that defines the family of balls.
In the previous examples, the distance gives a control on the kernels of the opera-
tors which are defined by integration with respect to a measure. The classical results
of harmonic analysis give boundedness properties of the operators acting on Banach
spaces of functions in terms of the size, smoothness or cancellation properties of the
kernels. Several of these classical results have been extended to more general than
Acronyms 5
euclidian settings. In particular, to the context of spaces of homogeneous type where

metric and measure are related by the doubling property. The basic fact is given by
the success in extending the Calderón - Zygmund method for the study of singular
integral operators to spaces of homogeneous type. But more recent results and ap-
plications together with a better understanding of the geometric measure theoretical
issues involved are giving rise to a new and very active fields of mathematics.
Aside from the basic academic examples of metric measure spaces, we shall
concentrate on concrete models for non-homogeneous diffusions, non-isotropic set-
tings, the Monge - Ampère equation, etc, where the abstract results can be applied
to obtain regularity of solutions of PDE’s of elliptic or parabolic types.
The abstract approach has, from the point of view of the applications, the addi-
tional advantage given by the fact that discrete continuous at once and the discrete
approximations to continuous problems allow to have a new insight on the underly-
ing mathematical models.
Our approach is basic in the sense that no previous knowledge of the subject
is assumed and we aim to the exploration of a natural setting for the Calderón -
Zigmund theory, in a broad sense, as tool in analysis and PDE. We also wish to
provide an account of the structure as a useful model of interesting mathematical
and physical situations.
Even when we shall develop with some detail the basic aspects of quasi-metric
spaces, the elementary concepts of Assouad, Volberg - Konjagin and Hausdorff di-
mensions, this is not a book on geometric measure theory. The book [?] is an excel-
lent treatment of the interesting and challenging related geometric aspects.
The book is organized in three parts: point spaces, function spaces and operators.
Chapter 1
Topological, uniform and quasi-metric spaces.
Abstract Each chapter should be preceded by an abstract (10–15 lines long) that
summarizes the content. The abstract will appear online at www.SpringerLink.com
and be available with unrestricted access. This allows unregistered users to read the
abstract as a teaser for the complete chapter. As a general rule the abstracts will not
appear in the printed version of your book unless it is the style of your particular
book or that of the series to which your book belongs.
Please use the ’starred’ version of the new Springer abstract command for
typesetting the text of the online abstracts (cf. source file of this chapter template
abstract) and include them with the source files of your manuscript. Use the
plain abstract command if the abstract is also to appear in the printed version of
the book.
Some basic laws of nature, such as those concerning gravitational and coulom-
bian interactions, are given in terms of potentials depending generally on the dis-
tance r between the interacting particles. Since most of the classical Harmonic Anal-
ysis is based on the particular properties of the operators induced by the Newtonian
potential r−1 and its derivatives, we start by introducing the basic metric concepts.
“Metrics Everywhere” is the title that Mischa Gromov gives to the introduction of
his book “Metric Structures for Riemannian and Non-Riemannian Spaces” ([G]).
As we shall see in this chapter, quasi-metric spaces are not beyond the scope of
metric spaces. But as we shall appreciate in Chapter 4, the general concept of quasi-
distance is more suitable to produce mathematical models of a wide variety of phys-
ical and mathematical situations. So, paraphrasing M. Gromov, we could say that
quasi-metrics are even more than everywhere.
In this chapter we shall introduce the basic topological and metric concepts used
throughout the book. We are specially interested in the study of the topological and
metric structures generated on a set X by a general non-negative valued function g
defined on X×X. We start by searching sufficient conditions on the proto-distance g
in order to construct a topology τ on X. We explore here two possible approaches to
the construction of τ. First we consider the neighborhood system generated by the
7
8 1 Topological, uniform and quasi-metric spaces.
g-balls. Then we introduce and study the uniform structure generated by g on X×X.
Next we review the basic results on metrizability of the topology induced by uni-
form structures with countable bases. Then we apply them to prove the metrizability
of quasi-metric spaces in the quantitative form given by Macı́as and Segovia in [?].
Some elementary properties and generalizations of metric and quasi-metric spaces
are also included. In Sections 1.12 and 1.13 we give two useful approaches to the
construction of quasi-distances starting from the family of balls or bands. These re-
sults shall be used in Chapter 4 where we consider a model for the Monge - Ampère
setting.
Sections 1.14 and 1.15 are just reviews of some basic metric concepts such as
total boundedness and Hausdorff dimension in the quasi-metric context.
Sections 1.16 and 1.17 contain the metric aspects of two constructions based on
self-similarity arguments, the first due to Macı́as and Segovia [?] and the second due
to M. Christ [?] and G. David. Both of them are going to be useful tools in the next
chapters. The first will provide subspace structure to the balls. The second one shall
be the source of dyadic partitions and, thus, of wavelet type orthonormal basis for
the Lebesgue spaces.The last section deals with a special product quasi-metric space
that shall be used in Chapter 4 in order to produce a model for a non-homogeneous
space time setting for parabolic equations. The basic general references for this
chapter are Kelley’s book [?] and the above mentioned paper by R. Macı́as and C.
Segovia [?]. See also [?]. Specific references shall also be given in the corresponding
sections.
Let us finally point out that we shall only introduce in this chapter elementary
or academic examples used to illustrate or to disprove particular facts, more useful
examples and models are given in Chapter 4.
1.1 Topology, neighborhoods and bases.
This section aims only to fix notation and terminology.
Let X be a set. A given family τ, of subsets of X is a topology if it is closed

under finite intersections and arbitrary unions and both 0, / the empty set, and the
whole space X belong to τ. Each member of τ is called an open set. The couple
(X, τ) is called a topological space. Given a point x ∈ X, a set V containing an open
set A with x ∈ A is called a neighborhood of x.
Let (X, τ) be a topological space. A subset F of X is called closed if its comple-

ment, F c , is open. The intersection of all closed sets containing a given set A ⊂ X is
called the closure of A and denoted by Ā. A subset D of X is dense if D̄ = X. The
topological space (X, τ) is separable if there exists a countable dense subset in X.
1.2 The topology induced by the neighborhood system associated to a proto-distance g. 9
Given a topological space (X, τ), a family B of subsets of X is a basis for τ if

B ⊂ τ and each member of τ is an union of members of B.
On the other hand, given a family B of subsets of X with X = B, then B is

[
B∈B
a basis for some topology on X if and only if for every B1 and B2 in B and every
x ∈ B1 ∩ B2 there exists B3 ∈ B such that x ∈ B3 ⊂ B1 ∩ B2 .
Given a family S of subsets of X, with X =

[
S, then the family of all finite
S∈S
intersections of members of S is a basis for a topology τ on X. In this case we say
that S is a sub-basis for τ. Spaces with countable bases are separable.
1.2 The topology induced by the neighborhood system associated

to a proto-distance g.
In this section we give sufficient conditions on a general function
g : X×X → R+ +
0 = R ∪ {0}
in such a way that, for every x ∈ X, the family of sets containing g-balls centered at
x determines a neighborhood system and hence a topology on X.
Let X be a set. If a topology τ on X is given then, the function that to each x in X

assigns the system Nx of all neighborhoods of x satisfies the following properties:
(a) if U ∈ Nx , then x ∈ U;
(b) if U ∈ Nx and V ∈ Nx , then U ∩V ∈ Nx ;
(c) if U ∈ Nx and V ⊃ U, then V ∈ Nx ;
(d) if U ∈ Nx , then there exists V ∈ Nx such that V ⊂ U and V ∈ Ny for every y ∈ V .
The next result is an exercise in Kelley’s “ General Topology ”.
Proposition 1.1. Given a set X and a function N : x → Nx , assigning to each

x ∈ X a family Nx of parts of X satisfying (a), (b) and (c) then, the family
τ = {U : U ∈ Nx for every x ∈ U} is a topology on X.
Assume now that a set X is given and that a non-negative real function g is defined
on X×X. Given a point x ∈ X and a positive real number r, the g-ball with center x
and radius r is defined by
Bg (x, r) = { y ∈ X : g(x, y) < r}
In these first sections we shall emphasize the dependence on g, but when there
is no possible doubt about the underlying function g we use the standard notation
B (x, r).
Let us now consider the function N assigning to each x in X the family of subsets
of X given by
Nx = U : Bg (x, r) ⊂ U for some positive r .

Very mild conditions on g are enough to apply Proposition 1.1 to this function
N in order to get a topology on X induced by g.
Proposition 1.2. Let X be a set and let g be a non-negative function defined on

X×X which is identically zero on the diagonal ∆ of X×X. Then, the above defined
function N satisfies (a), (b), and (c). Hence
τ = τNg = U : for every x ∈ U there exists r > 0 with Bg (x, r) ⊂ U

is a topology on X.
Two extreme trivial examples of the situation are in order.
Example 1.1. X a set, g ≡ 0 on X×X. Since for every positive r the g-ball Bg (x, r)
is the whole space X, we get the trivial topology τ = { 0,
/ X}.
Example 1.2. X a set, g (x, y) = 1 if x 6= y and g (x, x) = 0 for every x ∈ X. Then the
g-ball centered at x ∈ X with radius r ∈ (0, 1) is the point { x}. So that τ = P ( X),
the set of all subsets of X.
Let us remark that the topology defined in 1.1 is not metrizable and that the
function g in 1.2 is actually a distance or metric, i.e: g (x, y) = 0 if and only if x = y
; g (x, y) = g (y, x) for all x and y in X, and g (x, z) ≤ g (x, y) + g (y, z) for every choice
of x, y and z in X. We shall say that g is a pseudo-distance or pseudo-metric if g
satisfies the above properties but g is not necessarily faithful: g (x, y) = 0 is possible
with x 6= y.
1.3 Uniform spaces. Uniform topology. 11
1.3 Uniform spaces. Uniform topology.
In this section we introduce the basic definitions and notation concerning uniform
structures on an abstract set.
All the basic concepts and results of this section are contained in Chapter 6 of
[?].
For a given set X, let us write P (X×X) for the family of all subsets of X×X.
Given U and V in P (X×X), define inverse and composition by
U −1 = { (x, y) : (y, x) ∈ U}
and
U ◦V = { (x, z) : (x, y) ∈ V and (y, z) ∈ U for some y ∈ X} .
We shall say that U is symmetric when U −1 = U. Let us denote by ∆ the identity

or diagonal {(x, x) : x ∈ X}.
Definition 1.1. Let X be set. A family U of subsets of X×X is an uniform structure

on X if
1. ∆ ⊂ U for every U ∈ U ,
2. U ∈ U if and only if U −1 ∈ U ,
3. for each U ∈ U there exists V ∈ U such that V ◦V ⊂ U,
4. for U ∈ U and V ∈ U , we have that U ∩V ∈ U ,
5. for U ∈ U and V ⊃ U, we have that V ∈ U .
If X is a set and U is an uniform structure on X, we say that (X, U ) is an uniform

space.
Non trivial examples of uniform spaces are given by metric spaces. Suppose that
(X, d) is a metric space, given r > 0 let us write
Bd (r) = { (x, y) ∈ X×X : d (x, y) < r} ,
the family Ud = {U ∈ P (X×X) : there exists r > 0 such that Bd (r) ⊂ U} defines
an uniform structure on X.
Definition 1.2. Given U , an uniform structure on X, a subfamily B of U is a basis

for U if and only if for each U ∈ U there exists V ∈ B with V ⊂ U.
For example, the family Bd of all metric bands Bd (r) , r > 0 is a basis for the
uniformity Ud generated by the distance d on X. Notice that the same uniformity
Ud has also a countable basis: Bd ( 1n ) : n ∈ N .
Given a subset U of X×X and x ∈ X we need to introduce notation for the section
of U at x :
U (x) = { y ∈ X : (x, y) ∈ U} .
Every uniform structure on a set X induces a topology on X.
Proposition 1.3. Let U be an uniform structure on the set X. Then, the family
τ = { S ⊂ X : ∀ x ∈ S ∃ U ∈ U such that U (x) ⊂ S}
is a topology on X. If B is a basis for U , then
Nx = {U (x) : U ∈ B}
is a neighborhood basis for τ: every neighborhood of x contains some member of

Nx .
The proofs of the above and the next propositions are both left as exercises.
Proposition 1.4. The non-empty family B of P (X× X) is a basis for an uniform

structure on X if and only if
1. if B ∈ B, then ∆ ⊂ B;
2. for every B ∈ B there exists B0 ∈ B such that B0 ⊂ B−1 ;
3. for every B ∈ B there exists B0 ∈ B with B0 ◦ B0 ⊂ B;
4. given B and B0 in B, there exists B00 ∈ B such that B00 ⊂ B ∩ B0 .
The uniform structure generated by B is
5. U = {U ∈ P (X×X ) : B ⊂ U for some B ∈ B}.

1.4 Metrizability of uniform spaces. 13
1.4 Metrizability of uniform spaces.
We shall say that an uniformity U on X is pseudo-metrizable (resp. metrizable)

if
there is a pseudo-distance (resp. distance) ρ on X such that the family of ρ-bands
Bρ (r) : r > 0 is a basis for U . Let us notice that if U is a pseudo-metrizable
uniformity for X, then U has a countable basis, namely Bρ ( n1 ) : n ∈ N . The main

result for our purposes is the converse: every uniformity with a countable basis is
pseudo-metrizable. We start by proving a slight modification for the metrization
Frink’s lemma given in [?]
Lemma 1.1. Let X be a set and let {Un : n ∈ Z} be a family of subsets of X×X
satisfying
[
1. Un = X×X;
n∈Z
2. ∆ ⊂ Un for each n ∈ Z;
3. Un+1 ◦Un+1 ◦Un+1 ⊂ Un for every n ∈ Z.
Then, there exists a non-negative function ρ defined on X×X such that
4. ρ (x, z) ≤ ρ (x, y) + ρ (y, z) for every x, y and z ∈ X;

5. Un ⊂ { (x, y) ∈ X×X : ρ (x, y) ≤ 2−n } ⊂ Un−1 for every n ∈ Z.
Moreover, if each Un is symmetric, then\the function ρ is a pseudo-distance on X

which becomes a distance if in addition Un = ∆
n
Proof. Let us start by noticing that 1.9.3 together with 1.9.2 implies that Un+1 ⊂ Un
for every n. Given a point (x,\ y) ∈ X×X, then either there exists n ∈ Z such that
(x, y) ∈ Un −Un+1 or (x, y) ∈ Un . So that, we have a well defined function f on
n∈Z
X×X given by

 2−n if (x, y) ∈ U\
n −Un+1 ,
f (x, y) = 0 if (x, y) ∈ Un .

n∈N
The function f has the following remarkable property
n
f (x0 , xn+1 ) ≤ 2 ∑ f (xi , xi+1 ) (1.1)
i=0
for every n ∈ Z and every choice of { x0 , x1 , ..., xn+1 } ⊂ X.
Let us observe that in particular, (1.1) is a triangle type inequality when n = 1.

But for n > 1, (1.1) is sharper than the iteration of the triangle type inequality with
constant K = 2. We shall say that { xr , xr+1 ..., xs } (s > r) is a chain of length s − r
joining xr with xs .
Let us prove (1.1) by induction on the length of the chain. The result is obvious
if the length of the chain is one. Assume that (1.1) holds for any chain of length less
than or equal to n. Let us prove (1.1) as stated for a chain of length n + 1. Let us
write
n
α = ∑ f (xi , xi+1 )
i=0
\
We can assume α > 0. Otherwise (xi , xi+1 ) ∈ Un for every
n∈Z
\
i = 0, 1, ..., n, so that, from 1.9.3, (x0 , xn+1 ) ∈ Un and f (x0 , xn+1 ) = 0 and (1.1)
n∈Z
follows. Let us first assume that there exists k = 1, · · · , n − 1 such that
k−1
α
∑ f (xi , xi+1 ) ≤ 2
i=0
but
k
α
∑ f (xi , xi+1 ) > 2
.
i=0
Observe that in the case k < n, we also have that
n
α
∑ f (xi , xi+1 ) ≤ .
i=k+1 2
Now, both chains, { xk+1 , ..., xk } and { x0 , ..., xn+1 } have length less than or equal
to n. Then
k−1
α
f (x0 , xk ) ≤ 2 ∑ f (xi , xi+1 ) ≤ 2 2 = α.
i=0
and
n
α
f (xk+1 , xn+1 ) ≤ 2 ∑ f (xi , xi+1 ) ≤ 2 = α.
i=k+1 2
1.4 Metrizability of uniform spaces. 15
Also, obviously, f (xk , xk+1 ) ≤ α. Let m be the first integer for which α ≥ 2−m .
Then 2−m+1 = 2−(m−1) > α ≥ 2−m . Since f takes only dyadic values, the above
inequalities imply that (x0 , xk ) ∈ Um , (xk+1 , xn+1 ) ∈ Um and (xk , xk+1 ) ∈ Um .
Finally, 1.9.3 implies that f (x0 , xn+1 ) ≤ 2−(m−1) = 2 × 2−m ≤ 2α, and (1.1) is
proved when such a k ∈ {1, 2, · · · , n − 1} exists. If otherwise, f (x0 , x1 ) > α2 we can
proceed in a similar way since Um ◦Um ⊂ Um ◦Um ◦Um .
Define ρ on X×X as the function that maps the point (x, y) to the nonnegative
real number
( )
n
inf ∑ f (xi , xi+1 ) : x0 = x, x1 , ..., xn , xn+1 = y; n ∈ N .
i=0
Of course ρ is non-negative and finite, since ρ (x, y) ≤ f (x, y) . The triangle in-
equality follows easily: let x, y and z be three points in X and let ε be a positive real
number, then there are two chains x0 = x, x1 , ..., xm = y and xm = y, xm+1 , ..., xm+k = z
such that
m−1
∑ f (xi , xi+1 ) < ρ (x, y) + ε
i=0
and
m+k−1
∑ f (xi , xi+1 ) < ρ (y, z) + ε.
i=m
Since of course x = x0 , x1 , ..., xm+k = z is a chain joining x and z, we have
m+k−1
ρ(x, z) ≤ ∑ f (xi , xi+1 ) < ρ(x, y) + ρ(y, z) + 2ε,
i=0
being ε arbitrary, we get 1.9.4. In order to prove 1.9.5 let us observe that (1.1)
implies the equivalence of ρ and f :
ρ (x, y) ≤ f (x, y) ≤ 2ρ (x, y) .
If (x, y) ∈ Un , then f (x, y) ≤ 2−n and also ρ (x, y) ≤ 2−n . On the other hand if
ρ (x, y) ≤ 2−n , we have that f (x, y) ≤ 2 × 2−n ≤ 2−(n−1) and this, in turn, implies
that (x, y) ∈ Un−1 .
If each Un is a symmetric set of X×X, f and ρ are both symmetric functions. Let
us finally\observe that if (x, y) satisfies ρ (x, y) = 0, then from 1.9.5 we conclude that
(x, y) ∈ Un . If the last intersection is the diagonal ∆ , then x = y and ρ becomes
n∈Z
faithful.
Theorem 1.1. The uniformity U on X is pseudo-metrizable if and only if has a

countable basis.
The proof of the Theorem will be an immediate consequence of Lemma 1.1 and
the next result.
Lemma 1.2. If B is a countable basis for the uniformity U on X, then there exists
a countable basis B̃ for U of symmetric sets satisfying 1.9.1, 1.9.2 and 1.9.3.
Proof. Let B = {Vk : k ∈ N} be a basis for U . The set U1 = V1 ∩V1−1 is symmetric

and belongs to U . Then there exists V in U such that V ◦V ⊂ U1 . Moreover, there
exists W ∈ U such that W ◦W ⊂ V ; then
W ◦W ◦W ⊂ W ◦W ◦W ◦W ⊂ U1 .
Now choose k1 ∈ N such that Vk1 ⊂ W . So that Vk1 ◦ Vk1 ◦ Vk1 ⊂ U1 . Take U2 =
Vk1 ∩Vk−1
1
. This argument can be iterated to obtain Un for n ≥ 3. For n ≤ 0 take for
example Un = X×X.
As a corollary, we easily see that the uniformity U on X is metrizable if and only

if U has a countable basis such that the intersection of all its members is exactly ∆ .
1.5 Uniform structures induced by proto-distances.
Let X be a set and let g be a non-negative function defined on X×X. Let us consider
the family B of all subsets of X×X defined by
Bg (r) = { (x, y) ∈ X×X : g(x, y) < r} ,

1.5 Uniform structures induced by proto-distances. 17
for some positive number r. In this section we search for sufficient conditions on
g to insure that the family B satisfies 1.8.1 to 1.8.4 of Proposition 1.4. For those
proto-distances g we will then have that the family
U g = {U ∈ P(X×X) : there exists Bg (r) ⊂ U}
shall become an uniform structure on X. Also, from Proposition 1.3, the sections
Nx = {U(x) : U ∈ B} shall constitute a neighborhood basis for the topology gen-
erated by the uniformity U g . On the other hand, let us observe that the section of
Bg (r) at x is nothing but the g-ball with center at x and radius r, i.e., Bg (r)(x) is the
g-ball Bg (x, r). So that, provided that g ≡ 0 on ∆ , the uniform topology will coincide
with the neighborhood topology described in Proposition 1.2.
In the next result sufficient conditions on g are given in order to get 1.8.1 to 1.8.4
for the family B of g-bands Bg (r), r > 0.
Theorem 1.2. Let g : X×X → R+

0 ∪ {0} be a function such that
1. g (x, x) = 0 for every x ∈ ∆ ;

2. there exists an increasing bijection ν of R+ such that
g (x, y) ≤ ν (g(y, x))
for every x ∈ X and every y ∈ X,

3. there exists an increasing bijection λ of R+ such that the inequality g (x, y) ≤
λ (max { g (z, y) , g (y, z)}) holds for every x, y and z ∈ X.
Then, the family B = { Bg (r) : r > 0} satisfies properties 1.8.1, 1.8.2, 1.8.3 and
1.8.4.
Proof. Property 1.8.1 follows from hypothesis 1.12.1 and 1.8.4 follows from the
fact that Bg (r) ⊂ Bg (s) when r < s. Let us prove 1.8.2. We shall prove that for every
r > 0 there exists s > 0 such that Bg (s) ⊂ (Bg (r))−1 . Since
(Bg (r))−1 = { (x, y) : (y, x) ∈ Bg (r)} = { (x, y) : g(y, x) < r} ,
taking s = ν −1 (r), for g (x, y) < s, we have that

ν (g (x, y)) < ν(s) = r, so that g(y, x) < r. In other words, (x, y) ∈ Bg (s) implies
(x, y) ∈ (Bg (r))−1 .
To prove 1.8.3 we shall show that for every r > 0 there exists s > 0 such that
Bg (s) ◦ Bg (s) ⊂ Bg (r). Since
Bg (s) ◦ Bg (s) = { (x, z) : ∃ y ∈ X / (x, y) ∈ Bg (s) and (y, z) ∈ Bg (s)}

= { (x, z) : ∃ y ∈ X / g (x, y) < s and g (y, z) < s} ,
taking s = λ −1 (r) we have that for (x, z) ∈ Bg (s) ◦ Bg (s),
g (x, z) ≤ λ (max { g (x, y) , g (y, z)}) < λ (s) = r,
so that (x, z) ∈ Bg (r).
Lemma 1.3. Given a function g satisfying 1.12.1, 1.12.2 and 1.12.3, the family
U g = {U ⊂ X×X : there exits r > 0 with Bg (r) ⊂ U}
is an uniform structure on X×X and the uniform topology on X coincides with the
neighborhood topology
τ g = A ⊂ X : for every x ∈ A there exists r > 0 with Bg (x, r) ⊂ A .

1.6 Metrizability of the uniform structure induced by a

proto-distance
As observed in the examples of Section 1.2, the neighborhood topology induced on

a set X by a proto-distance g satisfying 1.12.1 needs not be a metrizable, not even a
pseudometrizable, topology. Quasi-symmetric and quasi-triangular inequalities like
those contained in conditions 1.12.2 and 1.12.3 are sufficient to construct a count-
able basis for U g and to be able, then, to apply the results of Section 1.4. This can
be done in a qualitative way in order to deduce that U g is also induced by a distance
or, better, in a quantitative manner in order to produce a structure result for the class
of proto-distances g on X in terms of the functions ν, λ and some specific distance
ρ defined on X. When ν(t) = t and λ (t) = Kt for some constant K ≥ 1 we should
recover the result of Macı́as and Segovia in [?].
1.6 Metrizability of the uniform structure induced by a proto-distance 19
Theorem 1.3. Let g : X×X → R+ 0 be a function satisfying 1.12.1 and 1.12.3 which
is faithful: g (x, y) = 0 implies x = y and λ (t) ≥ (1 + ε)t for some ε > 0. Then, there
exists ρ : X×X → R+ 0 such that
1. ρ (x, y) = 0 if and only if x = y ;

2. ρ (x, z) ≤ ρ (x, y) + ρ (y, z), for every x, y and z in X;
3. with Λ = λ −2 = λ −1 ◦ λ −1 and [s] the integer part of the real number s, we have
that the inequality
h i h i
1
log ρ(x,y) 1
log2 ρ(x,y)
−1
Λ ◦Λ (1) ≤ g (x, y) ≤ Λ ◦Λ (1)
holds for every x and every y in X with x 6= y.
Moreover, if g is symmetric then ρ is a distance on X.
Proof. In order to apply Lemma 1.1 we have to construct a sequence {Un : n ∈ Z} of

subsets of X×X satisfying 1.9.1, 1.9.2, and 1.9.3. Define Un = { (x, y) ∈ X×X : g(x, y) < Cn },
where C0 = 1 and Cn = λ (−2n) (1) for n ∈ Z, with λ (k) = λ ◦ . . . ◦ λ , k times if k ∈ N
and λ (k) = λ −1 ◦ . . . ◦ λ −1
|k| times if −k ∈ N. Since Cn → ∞ for n → −∞, we have 1.9.1. Since g ≡ 0 on ∆ ,
then 1.9.2 is clear. Let us show 1.9.3. Notice that we can write Un+1 ◦Un+1 ◦Un+1 =
{ (x, w) ∈ X×X : ∃ y, z ∈ X / g(x, y) < Cn+1 ; g(y, z) < Cn+1 and g(z, w) < Cn+1 } =
3 . Then, from 1.12.3, given a point (x, w) in U (3) we certainly have that
Un+1 n+1
g(x, w) ≤ λ (max { g(x, y); g(y, w)})

≤ λ (max { g(x, y); λ (max {g(y, z); g(z, w)})})
≤ λ 2 (max { g(x, y); g(y, z); g(z, w)})

≤ λ 2 (Cn+1 ) = λ (2) λ −2(n+1) (1)
= λ (−2n) (1) = Cn .
(3)
In other words Un+1 ⊂ Un . Thus, from Lemma 1.1 there exists a non-negative
function ρ defined on X× X verifying 1.14.2 and 1.9.5. Since g is faithful then,
the intersection of the sequence {Un : n ∈ Z} is the diagonal ∆ of X×X. So that ρ
is also faithful which is 1.14.2. In order to prove 1.14.3 we shall use 1.9.5 : Un ⊂
{ (x, y) : ρ(x, y) ≤ 2−n } ⊂ Un−1 for every n ∈ Z. This means that if ρ (x, y) ≤ 2−n ,
then g (x, y) < Cn−1 and if ρ (x, y) > 2−(n+1) , then g (x, y) ≥ Cn+1 . So that, given
(x, y) ∈ X× X −∆ there exists a unique n = n (x, y) ∈ Z such that 2−(n+1) < ρ (x, y) ≤
2−n for this value of n we also have the inequalities
λ (−2) ◦ λ (−2n) (1) = Cn+1 ≤ g (x, y) < Cn−1 = λ (2) ◦ λ (−2n) (1).
Since n (x, y) = [− log2 ρ (x, y)] we have 1.14.3. The final remark in the statement
of Theorem 1.3 follows also from Lemma 1.1 since, being g symmetric so are the
sets Un , n ∈ Z.
Let us observe that if a generalized symmetric condition like 1.12.2 holds for g,
then the function G(x, y) = max { g (x, y) , g(y, x)} also satisfies 1.12.3 with perhaps
a different function λ . So that Theorem 1.3 can be applied to obtain a distance ρ
which is equivalent to G in the sense of 1.14.3.
Theorem 1.4. Let g be a non-negative symmetric function defined on X×X satisfy-

ing 1.12.1 and 1.12.3. Then, the uniform structure on X×X described in Corollary
1.3 and the uniform topology on X are pseudo-metrizablein the sense that there
exists a pseudo-distance ρ on X such that the family B ρ = Bρ (r) : r > 0 is a ba-
sis for the uniform structure, the family Nx = Bρ (x, r) : r > 0 is a basis for the
ρ
neighborhoods of any point x ∈ X, and Bρ = Bρ (x, r) : x ∈ X, r > 0 is a basis

for the topology τ g .
1.7 Topology on quasi-metric spaces
In this section we introduce the concept of quasi-metric space and some basic ex-
amples and definitions.
Let X be a set. A quasi-distance or quasi-metric on X is a function d : X×X →

R+
0 satisfying the following properties
(a) d (x, x) = 0 for every x ∈ X;

(b) d (x, y) = 0 implies x = y;
(c) d (x, y) = d (y, x) for every x and y;
(d) there exists a positive constant K such that the triangular type inequality
d (x, z) ≤ K (d (x, y) + d (y, z)) holds for every x, y and z ∈ X.
1.7 Topology on quasi-metric spaces 21
The couple (X, d) is a quasi-metric space. Of course generally K ≥ 1 and when

K = 1 we have a metric space. When only (a), (c) and (d) hold, we say that (X, d) is
pseudo-quasi-metric space. Let us observe that (d) is equivalent to
e max { d (x, y) ; d (y, z)}

d (x, z) ≤ K
e and every x, y and z ∈ X.

for some constant K
This is a particular case of the general situation considered in the previous sec-
tions. So that we have a topology on X generate by a quasi-distance d, given by
τ = {V ⊂ X : for every x ∈ V there exists r > 0 such that Bd (x, r) ⊂ V } .
Actually τ is the usual metric topology when d is a distance on X.
Recall that for the case of (X, d) metric space, d-balls are open sets. In fact, given
y ∈ Bd (x, r) we have that
Bd (y, r − d (x, y)) ⊂ Bd (x, r) .
For the general case of quasi-metric spaces this fact is no longer true.
Example 1.3. Let X = R, the set of real numbers. Let ϕ(t) be the function |t| for
|t| ≤ 1 and ϕ(t) = |t|+1 for |t| > 1. Observe that |t| ≤ ϕ(t) ≤ 2|t|. Then if we define
d on R×R by d (x, y) = ϕ (x − y), we have a faithful and symmetric d satisfying (d)
with K = 2; in fact
d(x, z) = ϕ (x − z) ≤ 2 |x − z| ≤ 2 ( |x − y| + |y − z| )
≤ 2 (ϕ (x − y) + ϕ (y − z))
≤ 2 (d (x, y) + d (y, z)) .
On the other hand, since for |x − y| < 1, d coincides with the usual distance on
R, the topology τ induced by d is the usual one. Let us finally remark that the d-ball
centered at the origin with r = 3/2.
Bd (0, 3/2) = { y ∈ R : ϕ(y) < 3/2} .

is nothing but the closed interval [−1, 1] which does not belong to τ.
Given d and d 0 two quasi-distances on the same set X we shall say that they are
equivalent if there exist constants c1 and c2 such that for every x and y ∈ X,
c1 d(x, y) ≤ d 0 (x, y) ≤ c2 d(x, y). (1.2)
In this case we write d ∼ d 0 . Let us also observe that if a symmetric function d 0

is equivalent to a quasi-distance d on X, then d 0 is also a quasi-distance d on X. Of
course, equivalent quasi-distances induce the same topology on X, since
Bd 0 (x, c1 r) ⊂ Bd (x, r) ⊂ Bd 0 (x, c2 r).
But the same topology on X can be obtained from non-equivalent quasi-distances

on X.
Let us observe that the quasi-distance d in Example 1.3 is equivalent to the usual
|x − y| for which, of course, open balls are open sets. The main result about quasi-
distances is that this fact is always true: given a quasi-distance d on a set X, then
there exists d 0 ∼ d such that d 0 -balls are open sets. We shall prove it in the next
sections.
1.8 Metrization of quasi-metric spaces.
As an application of Theorem 1.3 we obtain in this section the result of Macı́as and
Segovia [?] on the structure of quasi-distances.
Let X be a set and d be a quasi-distance on X, i.e. d is a non-negative function on

X×X satisfying (a), (b), (c) and (d) of Section 1.7. Then d satisfies all the hypotheses
in Theorem 1.3 with λ (t) = Kt. Thus, there exists a distance ρ on X verifying 1.14.3
with Λ (t) = λ −2 (t) = Kt2 . Then
1.9 Iteration of the triangle inequality for a quasi-distance. 23
−2 log2 K log2 ρ1
K −2 (ρ(x, y))2 log2 K = K −2 2

1
−2 1+log2 ρ(x,y)
= K j k
1 )
−2 1+ log2 ( ρ(x,y)
≤K
≤ d(x, y)
h i
1
−2 log2 ρ(x,y) −1
≤K

1 −2
−2 log2 ρ(x,y)
≤K
−2 log2 K log2 ρ1
= K4 2
= K 4 (ρ(x, y))2 log2 K .
Theorem 1.5. Let (X, d) be a quasi-metric space. Then there exists a distance ρ on
X and a positive number β depending only on K such that
d ∼ ρβ
in the sense of (1.2).
1.9 Iteration of the triangle inequality for a quasi-distance.
Given a quasi-distance d on a set X with constant K for the triangle inequality, by

iteration we have that
n−1
d(x0 , xn ) ≤ Kn−1 ∑ d(x j , x j+1 ),
j=0
for every choice of points x0 , x1 , .., xn in X and every n ∈ N. From the result of
Section 1.8, we can improve this exponential behavior on the length of the chain
joining x0 with xn .
Proposition 1.5. Given a quasi-metric space (X, d), there exist two constants β ≥ 1
and C > 0 depending only on K such that the inequality
n−1
d (x0 , xn ) ≤ C nβ −1

∑d x j , x j+1
j=0
holds for every finite sequence x0 , x1 , ..., xn of points of X.

Proof. From Theorem 1.5 we have a distance ρ on X, a number β ≥ 1 and two

positive constants C1 and C2 such that
C1 ρ β (x, y) ≤ d (x, y) ≤ C2 ρ β (x, y)
for every x and y ∈ X. Then, from Hölder’s inequality, we have
d(x0 , xn ) ≤ C2 ρ β (x0 , xn )
!β
n−1
≤ C2 ∑ ρ(x j , x j+1 )
j=0
 ! 1 β
n−1 β 
1− 1

≤ C2 n β ∑ ρ β (x j , x j+1 ) 
j=0

n−1
= C2 nβ −1 ∑ ρ β (x j , x j+1 )
j=0
C2 β −1 n−1
≤ n ∑ d(x j , x j+1 ).
C1 j=0
1.10 Second approach to the metrization of quasi-metric spaces.

Metrization of η-metric spaces.
The approach to metrization of quasi-metric spaces through the underlying uniform

structure on X, discussed in the above sections, does not directly exhibit the explicit
form of ρ in terms of d. Let us start this section by observing that if d is a distance
n
on X, then d (x, y) coincides with the infimum of the sums ∑ d (xi , xi+1 ) taken over
i=1
all chains joining x with y. On the other hand, if d is a distance on X so is d β for any
0 < β ≤ 1, since (a + b)β ≤ aβ + bβ for a and b non-negative real numbers. Notice
also that if d is a quasi-distance with constant K, then d β is a quasi-distance with
constant K β ≤ K for those values of β . These facts suggest that we should look for
the distance ρ induced by the quasi-distance d in the class of all functions of the
form
1.10 Second approach to the metrization of quasi-metric spaces. Metrization of η-metric spaces.
25
( )
n
ρ(x, y) = inf ∑ d β (xi , xi+1 ) : x = x1 , ..., xn+1 = y; n ∈ N
i=1
for some positive, perhaps small, β .
Theorem 1.6. Let X be a set and let d be a quasi-distance on X with constant K.

Then there exists a positive β less than or equal to one, depending only on K, such
that the function
( )
n
β
ρ(x, y) = inf ∑d (xi , xi+1 ) : x1 = x, ..., xn+1 = y; n ∈ N
i=1
1
is a distance on X with ρ β ∼ d.
We shall obtain Theorem 1.6 as a corollary of the more general result for η-
metric spaces stated in the following theorem.
Theorem 1.7. Let η be a non-negative, continuous, increasing and convex function

defined on the non-negative real numbers such that η(t) > 2t and η(0) = 0. Let
X be a set and let d be an η-metric on X: d is a non negative faithful symmetric
function such that d (x, x) = 0 and
d (x, z) ≤ η (max { d(x, y), d(y, z)})
for every x, y, z ∈ Z. Let ψ be a continuous, increasing and concave solution of the

inequality
ψ ◦ η ◦ η ≤ 2ψ
with ψ(1) = 1 and ψ(0) = 0. Then
( )
n
ρ(x, y) = inf ∑ ψ(d(xi , xi+1 )) : x1 = x, x2 , ..., xn+1 = y; n ∈ N
i=1
is a distance on X with
ψ −1 (ρ) ≤ d ≤ ψ −1 (2ρ).
Of course, the existence of such a ψ has to be proved. [See Lemma 1.5].
Let us shortly sketch now how Theorem 1.6 follows from Theorem 1.7. Notice
that we may always assume that the constant K of d is bigger than two. So that
η(t) = Kt satisfies the hypotheses of Theorem 1.7. Observe also that ψ(t) = t β is a
−1
solution of ψ o η o η = 2ψ when β = log2 K 2 .
In order to prove Theorem 1.7 we shall make use of the next two lemmas. The
proof of the first is left as an exercise, while the second is proved at the end of this
section.
Lemma 1.4. Let X be a set and let g be a non-negative symmetric function on X×X
vanishing on the diagonal ∆ . Then the function
n
ρ(x, y) = inf ∑ g(xi , xi+1 )
i=1
is a pseudo-distance bounded above by g(x, y) if the infimum is taken over all finite
chains x = x1 , ..., xn+1 = y joining x with y.
Lemma 1.5. Let η be a continuous, increasing and convex function defined on R+ 0

satisfying η(t) > 2t for every positive t and η(0) = 0, then the functional inequality
ψ ◦ η ◦ η ≤ 2ψ
has at least one solution ψ increasing, continuous and concave with ψ(1) = 1 and
ψ(0) = 0.
Proof of Theorem 1.7. Let us start by applying Lemma 1.4 to the function g(x, y) =
ψ (d(x, y)) for a ψ given by Lemma 1.5. So that ρ(x, y) is a pseudo-distance and
ρ(x, y) ≤ ψ (d(x, y)) for every (x, y) ∈ X×X. Since ψ is increasing we have that
ψ −1 (ρ) ≤ d. Notice that in order to finish the proof of the theorem we only need to
show that
ψ(d) ≤ 2ρ, (1.3)

since it gives us the remaining inequality for the equivalence and proves that ρ is
faithful and hence a distance. To prove (1.3) let us show that the inequality
1.10 Second approach to the metrization of quasi-metric spaces. Metrization of η-metric spaces.
27
k−1
ψ (d(x1 , xk )) ≤ 2 ∑ ψ (d(xi , xi+1 )) , (1.4)
i=1
holds for any choice of { x1 , ..., xk } ⊂ X and any k ∈ N. Of course (1.4) is true for
k = 2. Assume that (1.4) holds for any k ≤ n and take { x1 , ..., xn+1 } a set of n + 1
n
point in X with n ≥ 2. Let δ = ∑ ψ (d(xi , xi+1 )) . We can assume δ > 0. One and
i=1
only one of the following three situations is possible:
(a) ψ (d(x1 , x2 )) > δ /2;

n−1
(b) ∑ ψ (d(xi , xi+1 )) ≤ δ /2;
i=1
(c) there exists k = 1, 2, ..., n − 2 such that
k k+1
∑ ψ (d(xi , xi+1 )) ≤ δ /2 and ∑ ψ (d(xi , xi+1 )) > δ /2.
i=1 i=1
We leave the cases (a) and (b) as exercises and only consider here the generic
case (c). In the following inequalities we apply twice the triangular inequality, the
induction hypothesis and Lemma 1.5.
n
ψ (d(x1 , xn+1 )) ≤ ψ η max d(x1 , xk+1 );
!
o
η(max{d(xk+1 , xk+2 ), d(xk+2 , xn+1 )})
k
−1
≤ (ψ o η o η) max ψ 2 ∑ ψ(d(xi , xi+1 )) ; ψ −1 (δ );
i=1
n
ψ −1 2 ∑ ψ(d(xi , xi+1 ))
i=k+2
≤ (ψ o η o η) ψ −1 (δ )

≤ 2δ .
Proof of Lemma 1.5. Let us denote by η̃ the composition η o η. We have to solve the
inequality ψ (η̃(t)) ≤ 2ψ(t) with ψ(1) = 1. Since η̃(t) > 4t we have that η̃ (k) (1) is
an increasing sequence for k ∈ Z, with η̃ (k) = η̃ o η̃ o...o η̃,
k times; η̃ (−k) = η̃ (−1) o η̃ (−1) o...o η̃ (−1) and η̃ (0) the identity. Moreover
lim η̃ (k) (1) = +∞ and lim η̃ (k) (1) = 0 .

k−→∞ k−→−∞
Set tk for η̃ (k) (1); k ∈ Z. Let us define ψ on the sequence T = {tk ; k ∈ Z} by
ψ(tk ) = 2k ; k∈Z .
Let us first observe that on T the function ψ solves the equation ψ o η̃ = 2ψ. In
fact, for each k ∈ Z we have
ψ ◦ η̃(tk ) = ψ (η̃(tk )) = ψ(tk+1 ) = 2k+1

= 2 . 2k = 2ψ(tk ).
Let us show that the function ψ defined on R+ by piecewise linear interpolation

of the points (tk , 2k ); k ∈ Z satisfies the required properties. Of course ψ is increasing
since so is on T. Let us call mk the slope of the kth segment of the graph of ψ. In
other words
2k+1 − 2k 2k
mk = = .
tk+1 − tk tk+1 − tk
The concavity of ψ is equivalent to the inequality mk+1 ≤ mk which, in turn,

follows from η̃(t) ≥ 4t. In fact mk+1 ≤ mk if and only if 2(tk+1 − tk ) ≤
≤ tk+2 −tk+1 which is implied by 3tk+1 < 4tk+1 ≤ η̃(tk+1 ) = tk+2 . It remains only to
show that ψ is a solution for the inequality (ψ o η̃)(t) ≤ 2ψ(t) when t 6∈ T. Given
a positive t which does not belong to T there exist one and only one k ∈ Z such that
tk < t < tk+1 so that tk+1 < η̃(t) < tk+2 . The assumed convexity of η̃ implies the
inequality
η̃(t) − η̃(tk ) η̃(tk+1 ) − η̃(tk )

≤ ; (1.5)
t − tk tk+1 − tk
on the other hand, from the definition of ψ we have that
ψ(t) − 2k
= mk ; and (1.6)
t − tk
ψ(η̃(t)) − 2k+1
= mk+1 . (1.7)
η̃(t) − tk+1
1.11 Quasi-distance of order α, open balls and Lipschitz functions on quasi-metric spaces. 29
Applying (1.7), the definition of mk+1 , (1.5) and finally (1.6) we get
ψ(η̃(t)) = (η̃(t) − tk+1 ) mk+1 + 2k+1

η̃(t) − tk+1 k+1
= 2 + 2k+1
tk+2 − tk+1
tk+2 − tk+1 2k+1
≤ (t − tk ) + 2k+1
tk+1 − tk tk+2 − tk+1
= 2(t − tk ) mk + 2k+1
= 2(ψ(t) − 2k ) + 2k+1
= 2ψ(t).
1.11 Quasi-distance of order α, open balls and Lipschitz

functions on quasi-metric spaces.
In this section we introduce the main elementary analytical consequences of Theo-

rems 1.5 or 1.6. Following Macı́as and Segovia, we shall say that a quasi-distance δ
on X is of order α (0 < α ≤ 1) if there exists a constant C such that inequality
| δ (x, z) − δ (y, z)| ≤ Cr1−α δ α (x, y) (1.8)
holds for every x, y and z in X such that δ (x, z) < r and δ (y, z) < r.
Theorem 1.8. Given a quasi-metric space (X, d) with constant K, there exists a
quasi-distance δ of order α = α(K) on X equivalent to d.
Proof. Let us take δ = ρ 1/β where ρ and β are given in Theorem 1.6. Let r be
a positive number and x, y and z be three points in X satisfying δ (x, z) < r and
δ (y, z) < r. Then ρ(x, z) < rβ , ρ(y, z) < rβ and, obviously, any number ξ between
ρ(x, z) and ρ(y, z) satisfies the same estimate ξ < rβ . From the mean value theorem
we get
| δ (x, z) − δ (y, z)| = ρ 1/β (x, z) − ρ 1/β (y, z)

1 1/β −1
= ξ | ρ(x, z) − ρ(y, z)|
β
1
≤ r1−β δ β (x, y),
β
which gives the desired result with C = 1/β .
We leave as an exercise the proof of the following generalization of Theorem 1.8

to the case of η-metric spaces.
Theorem 1.9. Let (X, d) be an η-metric space with η as in Theorem 1.7. Then
δ = ψ −1 (ρ) satisfies the following property of order ψ
ψ(δ (x, y))

| δ (x, z) − δ (y, z)| ≤ ,
ψ+0 (r)
where r is any positive number larger than max{ δ (x, z); δ (y, z)} and ψ+0 is the right
derivative of ψ(r).
As a consequence of Theorem 1.8 we see that given any quasi-metric space (X, d)
and a fixed point z ∈ X, the function f : X → R defined by f (x) = δ (x, z) is contin-
uous. In fact, given a sequence {xn } in X such that xn → x, with n large enough, we
have that
| f (xn ) − f (x)| = | δ (xn , z) − δ (x, z)|

≤ C max { δ (xn , z); δ (x, z)}1−β δ β (xn , x)
≤ CK 1−β (1 + δ (x, z))1−β δ β (xn , x),
which tends to zero as n tends to infinity. Since for a given z ∈ X and a given positive
real number r, the δ -ball Bδ (z, r) is the inverse image by f of the interval (−∞, r),
we have that Bδ (z, r) is an open set in X.
Corollary 1.1. In any quasi-metric space (X, d) there exists a quasi-distance δ

equivalent to d for which δ -balls are open sets. Moreover the δ -balls form a ba-
sis for the topology induced on X by d.
1.11 Quasi-distance of order α, open balls and Lipschitz functions on quasi-metric spaces. 31
In a metric space (X, d) we have that for d(x, y) < r
B (x, r − d(x, y)) ⊂ B(y, r) ⊂ B (x, r + d(x, y)) .

In quasi-metric spaces of order β we have a generalization of this basic fact.
Lemma 1.6. Let δ be a quasi-distance of order β on X. Let K be the triangular

constant and C the constant in (1.8). Let r be a positive real number and x and y two
−1/β
points in X such that δ (x, y) < C(2K)1−β

r, then

Bδ x, r − γr1−β δ β (x, y) ⊂ Bδ (y, r) ⊂ Bδ x, r + γr1−β δ β (x, y)
β
r
for every γ such that C(2K)1−β ≤ γ < δ (x,y) .
−1/β
Proof. Since the inequality δ (x, y) < C(2K)1−β

r is equivalent to
β
C(2K)1−β < r
δ (x,y) ,
β
r
we can choose γ satisfying C(2K)1−β ≤ γ < . Now, for these values of γ
δ (x,y)
r − γ r1−β is positive and the set Bδ x, r − γ r1−β δ β (x, y) is

we have that δ β (x, y)
a non-empty δ -ball in X. Let z be a point in this ball, then
δ (z, y) ≤ | δ (z, y) − δ (z, x)| + δ (z, x)

< C (max{ δ (z, y); δ (z, x)})1−β δ β (x, y) + r − γ r1−β δ β (x, y)
n o
≤ C (K (r + δ (x, y)))1−β − γ r1−β δ β (x, y) + r
≤ r;
so that z ∈ Bδ (y, r). The second inclusion follows in a similar way.
Let us now show how quasi-distances of order β can be used to produce Lipschitz
functions on quasi-metric spaces. Assume that (X, d) is a quasi-metric space and that
δ is a quasi-distance of order β (0 < β ≤ 1) equivalent to d. Let ϕ : R → R be a
Lipschitz function with compact support. For x0 ∈ X a fixed point, define f : X → R
by f (x) = ϕ (δ (x0 , x)) . Since ϕ has compact support, then there exists R > 0 such
that f (x) = 0 for x ∈/ Bδ (x0 , R) or Bδ (x0 , R) = X. In order to estimate the absolute

value of f (x) − f (y) for any choice of x and y in X, let us first assume that δ (x, y) ≥
R. In this case we have the trivial estimates
2 kϕk∞ β kϕk∞
| f (x) − f (y)| ≤ 2 kϕk∞ ≤ δ (x, y) ≤ C β d β (x, y).
R β R
On the other hand, if δ (x, y) < R and nor x neither y belongs to the ball Bδ (x0 , R),
we obviously have | f (x) − f (y)| = 0. Assume that δ (x, y) < R and at least one of
them belongs to Bδ (x0 , R). Observe that then both x and y belong to Bδ (x0 , 2KR).
So that, with L the Lipschitz constant for ϕ, applying the property of order β of δ
we have the desired result
| f (x) − f (y)| ≤ L | δ (x, x0 ) − δ (y, x0 )|

≤ LC (2KR)1−β δ β (x, y)
≤ Ce L R1−β d β (x, y).
Corollary 1.2. In every quasi-metric space there exist non-trivial Lipschitz β func-
tions, for some positive β depending on K.
1.12 Quasi-distance and balls.
In this section we explore necessary and sufficient conditions on a family of subsets

of a given set in order to get a quasi-distance whose balls are equivalent to the given
family. This approach to quasi-metric spaces is useful in applications. See Chapter
4.
Let (X, d) be a quasi-metric space. Let K be the triangular constant for d. Let us
think about the family of all balls in (X, d) as the function
B : X×R+ → P(X)
mapping (x, r) to the d-ball B(x, r). It is easy to see that the function B satisfies the
following properties:
1.12 Quasi-distance and balls. 33
(a) for every x ∈ X fixed, B(x, ·) is a non-decreasing function of r with the usual
P(X);
orders in R+ and [
(b) for every x ∈ X, B(x, r) = X;
r>0
\
(c) for every x ∈ X, B(x, r) = { x};
r>0
(d) for every x ∈ B(y, r) and some constant K we have both
(d.1) B(y, r) ⊂ B(x, K r); and
(d.2) B(x, r) ⊂ B(y, K r).
The constant K can actually be chosen to be 2K.
The main result of this section is the converse of the above remark. Let us intro-
duce a definition. Given a set X and two functions Bi : X×R+ → P(X),
i = 1, 2; we shall say that B1 is equivalent to B2 , and we shall write B1 ∼ B2 if there
exist constants γ and Γ such that
B1 (x, γr) ⊂ B2 (x, r) ⊂ B1 (x,Γ r)
for every x ∈ X and every r > 0.
Theorem 1.10. Let X be a set and let B : X×R+ → P(X) be a function satisfying
(a), (b), (c) and (d) for some constant K , then there exists a quasi-distance d on X
such that the family Bd of d-balls is equivalent to the given B.
Proof. Let us start by noticing that given two points x and y in X, from (b) and (a),
we can get a positive r for which y ∈ B(x, r) and x ∈ B(y, r). So that d : X×X → R+
given by
d(x, y) = inf { r > 0 : x ∈ B(y, r) and y ∈ B(x, r)}
is a well defined symmetric function with d(x, x) = 0, since x ∈ B(x, r) for ev-
ery r > 0. On the other hand, if d(x, y) = 0 there exists a sequence {rn } of pos-
itive real numbers with rn → 0 for\ n → ∞ and x ∈ B(y, rn ) for every n. So that
from (a) and (c) we get that x ∈ B(y, r) = {y} and x = y. Let us show the
r>0
quasi-triangular inequality. Let x, y and z be three points in X. We shall prove that
d(x, z) ≤ K (d(x, y) + d(y, z)) . Given a positive ε, there exist r1 > 0 and r2 > 0
such that
r1 < d(x, y) + ε with (i) x ∈ B(y, r1 ) and

(ii) y ∈ B(x, r1 ),
r2 < d(y, z) + ε with (iii) y ∈ B(z, r2 ) and

(iv) z ∈ B(y, r2 ).
To prove that d(x, z) ≤ K (r1 + r2 ) we only need to check that

x ∈ B(z, K (r1 + r2 )) and z ∈ B(x, K (r1 + r2 )). From (iii)
y ∈ B(z, r2 ) ⊂ B(z, r1 + r2 ),
apply now (d.2) with y ∈ B(z, r1 + r2 ) in order to get that
B(y, r1 + r2 ) ⊂ B (z, K (r1 + r2 )) .
Since from (i) and (a) we have x ∈ B(y, r1 ) ⊂ B(y, r1 + r2 ), we conclude that
x ∈ B z, K (r1 + r2 ) . Applying (iv) and (ii) in a similar way we get that z ∈
B x, K (r1 + r2 ) . Let us show the equivalence of the families Bd of d-balls on X
and the given B. It follows from the definition of d that Bd (x, r) ⊂ B(x, r). Take now
a point y in the set B(x, r), from (d) we see that B(x, r) ⊂ B(y, K r) and B(y, r) ⊂
B(x, K r), then x ∈ B(y, K r) and y ∈ B(x, K r), so that
d(x, y) ≤ K r < (K + ε)r,
for every positive ε. Thus B(x, r) ⊆ Bd x, (K + ε)r for every positive ε.
Let us point out that properties (a), (c) and (d.1) imply property (d.2) with con-
stant K 2 . See 1.52 in Section 1.19.
1.13 Quasi-distance and bands.
Let (X, d) be a quasi-metric space. Let K be the triangular constant for d. Let us
think about the family of all d-bands in X×X as a function
1.13 Quasi-distance and bands. 35
Vd : R+ → P(X×X); r → Vd (r) = { (x, y) : d(x, y) < r}.
The symmetric sets Vd (r) satisfy the following properties:
d (r1 ) ⊆ Vd (r2 ) whenever r1 ≤ r2 ;

(a) V[
(b) Vd (r) = X×X;
r>0
\
(c) Vd (r) = ∆ ;
r>0
(d) there exists a constant K such that Vd (r)◦Vd (r) ⊂ Vd (K r) for every positive
r.
The constant K in (d) can be chosen to be 2K. The main result of this section is the
following.
Theorem 1.11. Let X be a set and let V : R+ → P(X×X) be a family of subsets of

X×X satisfying (a), (b), (c) and (d). Then there exists a quasi-distance d on X such
that for every positive r and every γ < 1 we have
V (γr) ∩V −1 (γr) ⊆ Vd (r) ⊂ V (r) ∩V −1 (r).
Given Vi : R+ → P(X×X) , i = 1, 2 two families of subsets of X×X we shall

say that V1 and V2 are equivalent and we shall write V1 ∼ V2 if there are positive
constants γ and Γ such that
V1 (γr) ⊆ V2 (r) ⊆ V1 (Γ r),

for every r > 0.
Corollary 1.3. Let X be a set and let V : R+ → P(X×X) be a family of subsets

of X×X satisfying (a), (b), (c) and (d). Let us denote by V −1 the family V −1 (r) =
(V (r))−1 . Then, if V ∼ V −1 we have a quasi-distance d on X for which Vd ∼ V.
Proof of Theorem 1.11. First of all, notice that
!−1
−1
= (X×X)−1 = X×X,
[ [
V (r) = V (r)
r>0 r>0
and that
!−1
−1
= ∆ −1 = ∆ .
\ \
V (r) = V (r)
r>0 r>0
Given a point (x, y) in X×X there exists r > 0 such that (x, y) ∈ Vr and (x, y) ∈ Vr−1 .
The function d(x, y) =inf{r : (x, y) ∈ Vr and (x, y) ∈ Vr−1 }, is then well defined on
X×X. It is easy to see that d is symmetric and faithful: d(x, y) = 0 if and only if
x = y. Let us check the triangular inequality. For x, y, z and ε > 0, choose r1 and r2
in such a way that
r1 < d(x, y) + ε , (x, y) ∈ Vr1 , (y, x) ∈ Vr1 ;
r2 < d(y, z) + ε , (y, z) ∈ Vr2 , (z, y) ∈ Vr2 .
Then, from (a), (b), (c) and (d) we get that
(x, z) ∈ Vr2 ◦Vr1 ⊆ Vr1 +r2 ◦Vr1 +r2 ⊆ VK (r1 +r2 ) .
so that
d(x, z) ≤ K (r1 + r2 ) < K (d(x, y) + d(y, z)) + 2K ε,
and the inequality follows since ε is arbitrary. Let us finally show that the family of
d-bands Vd is equivalent to V ∩V −1 . Notice first that
Vd (r) ⊂ (V ∩V −1 )(r).
In fact given (x, y) ∈ Vd (r), we have d(x, y) < r; then for some s < r, (x, y) ∈ V (s) ⊂ V (r)
and (x, y) ∈ V −1 (s) ⊂ V −1 (r). On the other hand if (x, y) ∈ (V ∩V −1 )(γr) for γ < 1,
then d(x, y) ≤ γr < r and so (x, y) ∈ Vd (r).
1.14 Compact, totally bounded and complete sets in quasi-metric spaces. 37
1.14 Compact, totally bounded and complete sets in quasi-metric

spaces.
In this section we review some classical metric concepts such as total boundedness,
which in a quantitative manner will become important in the next chapters. We shall
say that the family {Uα : α ∈ Γ } is an open covering of[ the set E ⊂ X, (X, d) is
a quasi-metric space, if each Uα is an open set and E ⊂ Uα . A set E is called
α∈Γ
compact if every open covering of E contains a finite subcovering, i.e. for every
{Uα : α ∈ Γ } open covering of E there exists a finite subset { α1 , α2 , ..., αm } ⊂ Γ
such that {Uαi : i = 1, ..., m} is an open covering of E.
Given a quasi-metric space (X, d), a set E is called bounded if there exists x0 ∈ X
and R > 0 such that E ⊂ B(x0 , R). The set E is called totally bounded if for every
positive ε there is a finite subset F of E such that d(x, F) < ε for every x ∈ E, where
d(x, F) = inf { d(x, y) : y ∈ F} . Observe that totally bounded sets are bounded sets
but the converse does not hold in general.
We shall say that a sequence { xn : n ∈ N} in a quasi-metric space (X, d) is a

Cauchy sequence if for every positive ε there is an N such that d(xn , xm ) < ε when-
ever n and m are larger than or equal to N. A subset E of X is called complete if
every Cauchy sequence converges in E to a point of E.
Let us state the result of this section which is well known for the case of metric
spaces.
Theorem 1.12. Let (X, d) be a quasi-metric space and let E be a subset of X. Then
E is compact if and only if E is complete and totally bounded.
Proof. Let ρ be a distance in X given by Theorem 1.1: d ∼ ρ γ ; γ ≥ 1. Since

the topology on X given by ρ is the topology given by d, ρ-compact sets are
d-compact sets. From the equivalence
d ∼ ργ
we easily see that d-completeness is ρ-completeness and that d-total boundedness
is the same as ρ-total boundedness. So we can apply the well know result for the
distance ρ to obtain it for the quasi-distance d.
1.15 Hausdorff measure and dimension of sets in quasi-metric

spaces.
A metric concept which can be extended to quasi-metric spaces is Hausdorff dimen-

sion. In this section we introduce the basic definitions of outer Hausdorff measure
and Hausdorff dimension in a quasi-metric space.
Let (X, d) be a given quasi-metric space. For a subset A of X define the diameter
of A by
|A|d = sup { d(x, y) : x, y ∈ A} .
For a given set E and a given positive δ , a countable family { Ai : i ∈ N} of

∞
[
subsets of X is called a δ -covering of E if E ⊂ Ai and |Ai |d ≤ δ for every i ∈ N.
i=1
Let s and δ be two positive numbers. Denote by R+ +
∗ the set R ∪ {0, +∞} and
consider the function Hδ with domain P(X) and values in R∗ given by
s +
( )
∞
Hδs (E) = inf ∑ |Ai |sd : {Ai } δ − covering of E .
i=1
Lemma 1.7. Hδs is an outer measure on P(X).
Proof. If E1 ⊆ E2 , then every δ -covering of E2 is a δ -covering of E1 , so that

Hδs (E1 ) ≤ Hδs (E2 ). Let us now show that
!
∞
[ ∞
Hδs Ej ≤ ∑ Hδs (E j ), (1.9)
j=1 j=1
for any sequence { E j : j ∈ N} in P(X). If the right hand side of the above inequality
equals +∞, there is nothing to prove. Assume that
∞
∑ Hδs (E j ) < +∞.
j=1
Let ε > 0, then, for each j ∈ N there exists a δ -covering { Aij : i ∈ N} of E j such
that
1.15 Hausdorff measure and dimension of sets in quasi-metric spaces. 39
ε
∑ |Aij |sd < Hδs (E j ) + 2 j .
i∈N
The family { Aij : i, j ∈ N; j ∈ N} is a δ -covering of E for which
ε
∑ |Aij |sd ≤ ∑ Hδs (E j ) + j
2
j,i∈N j∈N
= ∑ Hδs (E j ) + ε,
j∈N
since ε is arbitrary we have (1.9).
Lemma 1.8. For fixed E ⊂ X and δ > 0, we have that
Hδt (E) ≤ δ t−s Hδs (E)

whenever t > s.
Proof. Observe that if { A j : j ∈ N} is a δ -covering of E, then
∞ ∞ ∞
∑ |A j |td = ∑ |A j |t−s s
d |A j |d ≤ δ
t−s
∑ |A j |sd .
j=1 j=1 j=1
Looking at Hδs (E), for fixed E and s, as a function of δ , we see that it is non-
increasing. Then define the Hausdorff outer measure of order s on (X, d) by
H s (E) = lim Hδs (E) = sup Hδs (E).

δ →0 δ >0
The next lemma is basic for the definition of the Hausdorff dimension of a set in
a quasi-metric space.
Lemma 1.9. Let (X, d) be a quasi-metric space. Let E be a subset of X and set
ψ(s) = H s (E). Then ψ ≡ 0, ψ ≡ +∞ or there exists s0 ∈ R+ such that ψ(s) = +∞
for every s < s0 and ψ(s) = 0 for every s > s0 .
Proof. The result follows from the next two remarks
(a) if ψ(s) < ∞ and s0 > s, then ψ(s0 ) = 0;

(b) if ψ(s) > 0 and s0 < s, then ψ(s0 ) = +∞.
To prove (a) apply Lemma 1.8 to see that for each positive δ
0 0
Hδs (E) ≤ δ s −s Hδs (E),
0
now, since s0 > s, δ s −s → 0 as δ → 0 but, since ψ(s) < ∞, Hδs (E) is bounded when
0 0
δ → 0, so that H s (E) = limδ →0 Hδs (E) = 0. To show (b) use again Lemma 1.8 to
get that
0 0
δ s −s Hδs (E) ≤ Hδs (E).
Since ψ(s) > 0, the left hand side tends to infinity for δ → 0.
Hausdorff Dimension: When ψ ≡ 0 we say that the set E has Hausdorff dimen-
sion zero with respect to d, when ψ ≡ ∞, that the Hausdorff dimension of E with
respect to d is +∞, in the generic case we say that s0 is the Hausdorff dimension of
E with respect to the quasi-distance d and we write
s0 = dimH,d (E).
Let us observe that the dimension of a given set is invariant after changes of
equivalent quasi-distances. Let us also notice that for a given quasi-distance d, if ρ
is the distance in Theorem 1.5, for which d ∼ ρ γ for some γ ≥ 1, we have
dimH,ρ (E) = γ dimH,d (E)
Let us finally point out that H s is a quasi-metric outer measure on P(X) in the
sense that given two sets E1 and E2 such that d(E1 , E2 ) > 0, then H s (E1 ∪ E2 ) =
H s (E1 ) + H s (E2 ). This property suffices for the measurability of Borel sets with
respect to the Caratheodory σ −algebra defined by the outer measure H s .
1.16 Improving the smoothness of the boundaries of balls. 41
1.16 Improving the smoothness of the boundaries of balls.
In this section we introduce a procedure designed by Macı́as and Segovia in [?] in

order to get equivalent quasi-distances with better properties on the boundaries of
the balls defined by the new quasi-distance. The precise meaning of this smoothness
property will be clear later (Chapter 3). In this section we give the construction
procedure and we prove that the new quasi-distance δ is equivalent to the original
d. The basic technique resembles a self-similar fractal construction.
Let (X, d) be a quasi-metric space. Let us denote by
U(r) = { (x, y) : d(x, y) < r}, r > 0,

the d-band of width r around the diagonal ∆ in X × X.
We are going to construct a family {V (r) : r > 0} of bands containing ∆ by
joining “self similar copies of U(r)” an then we shall define δ on X×X as we did in
Section 1.13.
1
Assume that K is a triangular constant for d. Pick 0 < a < 2K . Given a positive
number r, we first construct a sequence U(r, n) of subsets of X×X in the following
way,
U(r, 0) = U(r)
U(r, 1) = U(a r) ◦U(r) ◦U(a r)
.. .. ..
. . .
U(r, n) = U(an r) ◦U(r, n − 1) ◦U(an r)
Since each U(s) contains the diagonal, we have that U(r) ⊂ U(r, n) for every n. But,
since a is small, the geometric growth of the sequence U(r, n), n ∈ N is controlled
by U(3K 2 r).
Lemma 1.10. For every r > 0 and every n ∈ N0 we have that
U(r) ⊂ U(r, n) ⊂ U(3K 2 r).
Proof. We only have to prove the second inclusion. Take (x, y) ∈ U(r, n). Then, there
exists a sequence x0 = x, x1 , ..., xn , yn , yn−1 , ..., y1 , y0 = y of points in X, satisfying
(xn , yn ) ∈ U(r) = U(r, 0), (1.10)

n− j
(x j , x j+1 ) ∈ U(a r),
(y j , y j+1 ) ∈ U(an− j r), j = 0, 1, .., n.
Let us now estimate d(x, y) by repeated use of the triangle inequality,
d(x, y) = d(x0 , y0 ) ≤ K 2 d(x0 , xn ) + d(xn , yn ) + d(yn , y0 )

≤ K 2 K d(x0 , xn−1 ) + d(xn−1 , xn ) + d(xn , yn ) + K(d(yn , yn−1 ) +

+ d(yn−1 , y0 )
h
≤ K 2 K n d(x0 , x1 ) + K n−1 d(x1 , x2 ) + ... + Kd(xn−1 , xn ) +
+ d(xn , yn )
i
+ K n d(y0 , y1 ) + K n−1 d(y1 , y2 ) + .. + Kd(yn−1 , yn )
< rK 2 2 (K n an + K n−1 an−1 + ... + Ka) + 1

" #
2 Ka
≤ rK 2 + 1 ≤ 3K 2 r.
1 − Ka
Let us now construct a new family {V (r) : r > 0} of subsets of X×X. Set V (r) =
∞
[
U(r, n) with r > 0. From Lemma (1.35), we have that
n=0
U(r) ⊂ V (r) ⊂ U(3K 2 r)
for every positive r. In other words, U ∼ V in the sense of Section 1.13. Let us also
notice that, for any positive number r we have
V −1 (r) = { (y, x) : (x, y) ∈ V (r)} = V (r),
since every U(r, n) is symmetric. It is clear that the family {V (r) : r > 0} satisfies
the properties (a), (b), (c) and (d) of Section 1.13. In fact, (a) follows from the
definition of V (r), (b) from U(r) ⊂ V (r) and (c) from both, U(r) ⊂ V (r) ⊂ U(3K 2 r).
To check (d), notice that
V (r) ◦V (r) ⊂ U(3K 2 r) ◦U(3K 2 r) ⊂ U(6K 3 r) ⊂ V (6K 3 r).
We can apply Corollary 1.3 to obtain a quasi-distance δ on X, defined by

1.17 Dyadic type partitions. 43
δ (x, y) = inf { r : (x, y) ∈ V (r)}
such that Vδ ∼ V.
Lemma 1.11. δ ∼ d.
Proof. Since Vδ ∼ V and V ∼ U we certainly have that Vδ ∼ U. This means that

there exist constants γ and Γ such that, for every positive r.
Vδ (γ r) ⊂ U(r) ⊂ Vδ (Γ r).
Take x 6= y and k ∈ Z such that 2k ≤ d(x, y) < 2k+1 . Since (x, y) ∈ U(2k+1 ), then
(x, y) ∈ Vδ (Γ 2k+1 ), hence δ (x, y) < Γ 2k+1 ≤ 2 Γ d(x, y). On the other hand,
since d(x, y) ≥ 2k , then (x, y) ∈
/ U(2k ), hence (x, y) ∈
/ Vδ (γ 2k ). This means that
k γ
δ (x, y) ≥ γ 2 > 2 d(x, y).
Lemma 1.12. For every x ∈ X and every r > 0 we have that
Bδ (x, r) = { y : (x, y) ∈ V (r)}.

In other words, the δ -balls are precisely the sections of the strips V (r).
Proof. Let us denote, as before, V (r)(x) for the section at x of V (r). Then, it is clear
from the definition of δ that Bδ (x, r) ⊂ V (r)(x). Take (x, y) ∈ V (r). Then, there is an
n ∈ N ∪ {0} such that (x, y) ∈ U(r, n). With the notation of Lemma 1.10 we have a
sequence joining x with y satisfying (1.10). Since the sequence is finite and in (1.10)
we have a finite number of strict inequalities, we see that (1.10) remains valid for
some positive s with s < r. Hence (x, y) ∈ U(s, n) ⊂ V (s) and δ (x, y) ≤ s < r.
1.17 Dyadic type partitions.
Let (X, d) be a quasi-metric space. A subset A of X is said to be bounded if there

exists x0 ∈ X and R > 0 such that A ⊂ B(x0 , R). Given a positive number ε and E ⊂ A
subsets of X, we say that E is ε-dense in A if for every x ∈ A there exists y ∈ E such
that d(x, y) < ε. A subset A ⊂ X is totally bounded if for every positive ε there
exist a finite subset F of A such that F is ε-dense in A.
In this section we shall assume that (X, d) is a quasi-metric space for which
bounded sets are totally bounded sets. The triangular constant of d shall be denoted
by K.
A subset E of X is ε-disperse if for every choice of x 6= y , x ∈ E and y ∈ E we

have that d(x, y) ≥ ε. A subset A of X is totally bounded if and only if for every
ε > 0 there exists a finite subset E of A such that E is a maximal ε-disperse subset
of A.
Since X can be written as increasing and countable union of d-balls and since
every d-ball is assumed to be totally bounded, we see that for every positive ε there
exists a subset Nε of X which is a maximal ε-disperse subset of X. We say that Nε
is an ε-net. Let us also observe that X is bounded if and only if for some positive ε
there exists a finite ε-net in X.
The aim of this section is to introduce the construction of dyadic type sets on
(X, d) given by M. Christ and G. David, (see [] and []). Let us first state the main
result of the section.
Theorem 1.13. Let (X, d) be a quasi-metric space such that bounded sets are totally
bounded. Then there exists a > 0, 0 < δ < 1, C > 0 and, for every j ∈ Z, an initial
interval K j = { 1, ..., k j } of N which could be all of N, such that to every ( j, k) ∈ A
({ j} × K j ) a subset Qkj of X and a point xkj are assigned, satisfying
[
=
j∈Z
1. Qkj is open set for every ( j, k) ∈ A ;

2. B( xkj , a δ j ) ⊂ Qkj for every ( j, k) ∈ A ;
3. Qkj ⊂ B(xkj ,Cδ j ) for every ( j, k) ∈ A ;
4. for every ( j, k) ∈ A and every i < j there exists a unique l ∈ Ki such that Qkj ⊆ Qil ,
5. if j ≥ i then either Qkj ⊆ Qil or jT i
/ k ∈ K j and l ∈ Ki ;
/ Qk Ql = 0,
Qkj is empty;
[
6. for every j, the interior of X
k∈K j
7. X is bounded if and only if there exists ( j, k) ∈ A such that X = Qkj .
Let us notice that, modifying eventually the constants a and C, the result is in-
variant if we change d to an equivalent quasi-distance. So that throughout the con-
struction we shall keep assuming that d-balls are opens sets.
To start with the construction of the family of dyadic sets, pick a number δ with
0 < δ < 1 and, for every j ∈ Z take a δ j -net N j in X,
N j = { xkj : k ∈ K j }
for some at most countable set K j . Notice that N j is δ j - dense in X and that its
intersection with every ball is finite. The next step is to introduce a tree structure on
A =
[
({ j} × K j ) .
j∈Z
Lemma 1.13. There exists a partial order on A satisfying the following tree
properties
1. ( j1 , k1 ) ( j2 , k2 ) implies j2 ≤ j1 ;
2. for every ( j1 , k1 ) ∈ A and every j2 ≤ j1 , there exists a unique k2 ∈ K j2 such that
( j1 , k1 ) ( j2 , k2 );
3. if ( j1 , k1 ) ( j1 − 1, k2 ), then d xkj11 , xkj12 −1 < δ j1 −1 ;
j1 −1
4. if d xkj11 , xkj12 −1 < δ 2K , then ( j1 , k1 ) ( j1 − 1, k2 ).
The partial order given on A by Lemma 1.39 can be seen as family tree in which
j ∈ Z is the generation, k ∈ K j is an individual in the jth generation, the relation
can be interpreted as “ descendant of”. So that 1.39.2 means that each individual in
the j1 generation has a unique ancestor in the j2 generation, provided that j2 j1 .
Properties 1.39.3 and 1.39.4 provide the link of the order with the (geo)metric
structure on X. The first one reads “ the offspring of the individual k of the j-th
generaton is not far away from k”. The second is a weak converse: “ if an individual
k1 of the generation j is ’very’ close to an individual k2 of the preceding generation
(j-1), then k1 is an offspring of k2 ”.
Assuming the existence of such an order on A , we define
Qkj =
[
B(xli , aδ i ),
(i,l)( j,k)
for ( j, k) ∈ A . We must show that for an appropriate choice of a and δ the sequence
{ Qkj : ( j, k) ∈ A } satisfies 1.38.1 to 1.38.7. Statement 1.38.1 follows directly from
the definition of Qkj since we are assuming that the d-balls are open sets. Also 1.38.2
∞
Ni is dense in X and contained in Qkj ,
[ [
is clear. Since for every j ∈ Z the set
i= j k∈ I j
1.38.6 easily follows. Property 1.38.7 follows from the fact that δ j → ∞ as j → −∞.
Proof of 1.38.3. Let z be a point in Qkj for a fixed ( j, k) ∈ A . There exists (i, l) ∈ A
with (i, l) ( j, k) such that z ∈ Bd (xli , a δ i ).
From 1.39.1, i ≥ j. If i = j, since is a partial order and 1.39.2 holds, then l = k,

so that z ∈ Bd (xkj , a δ i ) ⊂ Bd (xkj ,Cδ j ), provided that C ≥ a. Assume, then that i > j.
Let us apply 1.39.2 starting with (i, l) = ( j1 , k1 ) and j2 = i − 1. Then, there exists
a unique l1 ∈ Ii−1 such that (i, l) (i − 1, l1 ). If we keep applying 1.39.2 until we
reach the j-th generation, we get a finite sequence
(i, l) (i − 1, l1 ) (i − 2, l2 ) . . . ( j, li− j ).
Since is transitive, we have that (i, l) ( j, li− j ). Using now the uniqueness in
1.39.2 we conclude that li− j = k. From 1.39.3 we have the inequalities

d xli , xli−1
1
< δ i−1 ,

d xli−1
1
, xli−2
2
< δ i−2 ,
.. .. ..
. . .
j+1 j j+1 j i−(i− j) = δ j .
d xli− j−1 , xli− j = d xli− j−1 , xk < δ
Then, from the triangle inequality we have that

d(z, xkj ) ≤ K d z, xli + d xli , xkj

< K a δ i + K d xli , xlj+1i− j−1
+ d x j+1
,
li− j−1 k x j

≤ K a δ + K d xl , xli− j−2 + K d xlj+2
i 2 i j+2 2
i− j−2
, xlj+1
i− j−1
+

+ Kd xlj+1
i− j−1
, xkj
≤ K(a δ i + K i− j δ i−1 + K i− j−1 δ i−2 + ..... + Kδ j )
= Kδ j (a δ i− j + K i− j δ i− j−1 + K i− j−1 δ i− j−2 + .... + K).
∞
≤ K2δ j ∑ (Kδ )m .
m=0
1
Choosing δ ≤ 2K we have the desired result.
Properties 1.38.4 and 1.38.5 are consequences of the following lemma.
Lemma 1.14. For every j ∈ Z, Qkj ∩ Qlj 6= 0/ implies k = l.
Proof of 1.38.4. Given ( j, k) ∈ A and i < j , from 1.39.2 we have a unique l ∈ Ki

such that ( j, k) (i, l). Then
Qkj =
[ [
Bhm ⊂ Bhm = Qil ,
(h,m)( j,k) (h,m)(i,l)
with Bhm = Bd (xm

h , a δ h ). To show the uniqueness of l, assume that there exists
l ∈ Ii such that Qkj ⊂ Qli . Hence Qil ∩ Qil ⊃ Qkj 6= 0/ and Lemma 1.14 implies
l = l.
Proof of 1.38.5. Take j ≥ i and assume that for k ∈ I j and l ∈ Ii we have that Qkj ∩
/ From 1.39.2, there exists m ∈ Ii such that ( j, k) (i, m). Hence Qkj ⊂ Qim .
Qil 6= 0.
/ so that, from Lemma 1.14 we conclude that m = l and Qkj ⊂ Qil
Then Qim ∩ Qil 6= 0,
as desired.
Proof of Lemma 1.14. Assume that x is a point in the intersection of Qkj and Qlj .
l) ( j, l) such that x ∈ Bej1 ∩ Bej2 . Assume
k) ( j, k) and ( j2 , e
Then, there exist ( j1 , e
k l
that j2 ≥ j1 . Then
h i
d xej1 , xej2 ≤ K d xej1 , x + d x, xej2 < K[a δ j1 + a δ j2 ] ≤ 2aKδ j1 .
k l k l
1
If the number 2aK is less than one, i.e. if we choose a < 2K , then

d xej1 , xej2 < δ j1 .
k l

Since N j1 is δ1j - dense, either (i) j2 = j1 and d xej1 , xej2 = 0 or (ii) j2 > j1 .
k l

In this first case we have j2 , e l = j1 , e k , then from the uniqueness of ancestors
given by 1.39.2 we must have that ( j, k) = ( j, l), so that k = l as we wanted.
Let us consider now the case j2 > j1 .Since j1 + 1 ≤ j2 and ( j2 , e l) ∈ A , from

∗
1.39.2 there exists a unique k ∈ K j1 +1 such ∗
that ( j2 , l) ( j1 + 1, k ). Apply-
e
j1 j2 j
ing the triangle inequality, the inequality d xe , xe < 2aKδ 1 and the inequality
l l
d xkj1∗+1 , xej2 ≤ 2Kδ j1 +1 obtained in the proof of 1.38.3, we have that
l

d xkj1∗+1 , xej1 ≤ Kd xkj1∗+1 , xej2 + Kd xej2 , xej1
k l l k
< 2K 2 δ j1 +1 + 2aK 2 δ j1
= 2K 2 (a + δ )δ j1
δ j1
= ,
2K
choosing a and δ small enough. Applying 1.39.4 we get that ( j1 + 1, k∗ )

( j1 , e
k). Hence, from the transitivity of we have
l) ( j1 + 1, k∗ ) ( j1 , e
( j2 , e k) ( j, k).
l) ( j, l), applying again the uniqueness of the

Since we also know that ( j2 , e
ancestor given by 1.39.2 we get k = l as desired.

Proof of Lemma 1.13. Let us start by defining between members of the (j-1)-th
and j-th generations in A0 . Since the net
N j−1 = { xkj−1 : k ∈ K j−1 }
in X, we have that given ( j, k) ∈ A there exists at least one l ∈ K j−1

is δ j−1 - dense
such that d xkj , xlj−1 < δ j−1 and, since N j−1 is δ j−1 - disperse there exists at most
j−1
one m ∈ K j−1 such that d xkj , xmj−1 < δ2K . In fact, if xmj−1 and xnj−1 with n 6= m,
n ∈ K j−1 , m ∈ I j−1 were two points in N j−1 distant from xkj less than (2K)−1 δ j−1 ,
we would have that

d xnj−1 , xm
i−1
≤ Kd xnj−1 , xkj + Kd xkj , xmj−1 < δ j−1 ,

which is impossible, because N j−1 is δ j−1 -disperse.

δ j−1
Given ( j, k) ∈ A , if there exists m ∈ K j−1 such that d xkj , xmj−1 < 2K we say
that ( j, k) ( j − 1, m) and that ( j, k) is not related to any other element ( j − 1, l) for
l ∈ K j−1 .
an m ∈ K j−1 exists. Since there exists at least one

It may happen thatnone such
l ∈ K j−1 such that d xkj , xlj−1 < δ j−1 , we choose any one of those l and define
1.18 Product of quasi-metric spaces. 49
( j, k) ( j − 1, l) and we say that ( j, k) is not related to any other ( j − 1, m) for

m ∈ K j−1 .
Notice that with this definition, for every ( j, k) ∈ A there exists one and only
one l ∈ K j−1 such that ( j, k) ( j − 1, l). Since we are looking for a partial order in
A , reflexivity and transitivity complete the picture:
• ( j, k) ( j, k) for every ( j, k) ∈ A ;
• ( j, k) is not related to any other ( j, l), l 6= k ;
• ( j, k) (i, l) and j 6= i implies j > i and that there exist k1 ∈ K j−1 , k2 ∈
K j−2 , . . . , k j−i−1 ∈ Ki+1 such that ( j, k) ( j − 1, k1 )
( j − 2, k2 ) . . . (i, l).
It is easy to check 1.39.1 to 1.39.4.
1.18 Product of quasi-metric spaces.
Let (X, d) be a quasi-metric space with constant K and (Y, ρ) be a quasi-metric

space with constant k. Then X×Y becomes a quasi-metric space with
δ ((x1 , y1 ) ; (x2 , y2 )) = d(x1 , x2 ) + ρ(y1 , y2 )
and triangular constant given by max{ K, k}. This elementary situation can be ex-
tended to a more useful statement that shall be used in Chapter 4 when modelling a
degenerate parabolic equation.
Theorem 1.14. Let (X, d) be a quasi-metric space. Let Y be a given set. Assume
that for each x ∈ X there is a quasi-distance ρx on Y such that
1. there exists a constant k such that the inequalities
ρx (y1 , y3 ) ≤ k [ ρx (y1 , y2 ) + ρx (y2 , y3 )]
hold for every x ∈ X and every y1 , y2 , y3 in Y,

2. there exists a constant M such that
50
δ ((x1 , y1 ); (x2 , y2 )) = d(x1 ,1x2Topological,
) + ρx1 (y1 ,uniform
y2 ) + ρand
x2 (yquasi-metric
1 , y2 ) spaces.
is a quasi-distance on X×Y,
(b) The family of δ -ballsρis equivalent to the family of “squares” generated by d and
x1 (y1 , y2 ) ≤ M [ ρx2 (y1 , y2 ) + d(x1 , x2 )]
ρx in the sense that there exist constants C1 and C2 such that
x2 0∈, yX0 ),C

x1δ,((x
for every B and1 r) ⊆ Bdy1(x, 0y,2r)×B
every ∈ Y. ρx0 (y0 , r) ⊆ Bδ ((x0 , y0 ),C2 r) .
Then
Proof. It is easy to verify that δ is faithful and symmetric. Let us check that δ
(a) The function
satisfies a triangle inequality. Take three points (x1 , y1 ), (x2 , y2 ) and (x3 , y3 ) in X×Y,
then from the fact that d is a quasi-distance and from 1.41.1 and 1.41.2, we have
δ ((x1 , y1 ); (x3 , y3 )) = d(x1 , x3 ) + ρx1 (y1 , y3 ) + ρx3 (y1 , y3 )

≤ K [ d(x1 , x2 ) + d(x2 , x3 )] +
+ k [ ρx1 (y1 , y2 ) + ρx1 (y2 , y3 )] +

+ k ρx3 (y1 , y2 ) + ρx3 (y2 , y3 )
≤ Kd(x1 , x2 ) + Kd(x2 , x3 ) +
+ kρx1 (y1 , y2 ) + kM [ ρx2 (y2 , y3 ) + d(x1 , x2 )] +
+ kM [ ρx2 (y1 , y2 ) + d(x2 , x3 )] + kρx3 (y2 , y3 )
≤ (K + kM) d(x1 , x2 ) + kρx1 (y1 , y2 ) +
+ kMρx2 (y1 , y2 ) +
+ (K + kM) d(x2 , x3 ) + kMρx2 (y2 , y3 ) +
+ kρx3 (y2 , y3 )

≤ (K + kM) δ (x1 , y1 ); (x2 , y2 ) +

+ δ (x2 , y2 ); (x3 , y3 ) ,
which proves (a). Let us prove (b). It is enough to show that δ ((x0 , y0 ); (x, y))
is equivalent to the function max{ d(x0 , x); ρx0 (y, y0 )} with constants that are inde-
pendent of x, y, x0 and y0 . From the definition of δ we see that max{ d(x0 , x); ρx0 (y, y0 )} ≤
δ ((x0 , y0 ); (x, y)) . On the other hand, since from 1.41.2 we have the inequality

ρx (y, y0 ) ≤ M ρx0 (y, y0 ) + d(x0 , x) ,
then
1.19 Problems and Comments 51
δ ((x0 , y0 ); (x, y)) = d(x0 , x) + ρx0 (y, y0 ) +

+ ρx (y, y0 )
≤ d(x0 , x) + ρx0 (y, y0 ) +

+ M ρx0 (y, y0 ) + d(x0 , x)
≤ 2(M + 1) max{ d(x0 , x), ρx0 (y, y0 )}.
Hence
max{ d(x0 , x); ρx0 (y, y0 )} ≤ δ ((x0 , y0 ); (x, y))

≤ 2(M + 1){ d(x0 , x); ρx0 (y, y0 )}
So that
Bδ ((x0 , y0 ), r) ⊆ Bd (x0 , r)×Bρx0 (y0 , r)

⊆ Bδ ((x0 , y0 ), 2(M + 1)r) .
1.19 Problems and Comments
1.41. Prove Propositions 1.1 and 1.2. Is g ≡ 0 on ∆ necessary for the properties
(a),(b) and (c) for the function N defined in Section 1.2?
1.42. Prove Propositions 1.3 and 1.4.
1.43. Show that the uniformity U on X is metrizable if and only if U has a count-
able basis B such that the intersection of all the members of B is the diagonal
∆.
1.44. Explore the necessity of 1.12.1, 1.12.2 and 1.12.3 if we know that the family
B defined in Proposition 1.2 satisfies 1.8.1 to 1.8.4.
1.45. Give an example of a quasi-distance on R1 equivalent to the usual |x − y| such

that there are balls which are not even Borel subsets of R.
1.46. Show that the potential growth on the length of the chain in Proposition 1.5 is
generally unimprovable.
1.47. Consider the result of Proposition 1.5 when d is equivalent to a distance ρ on

X.
1.48. Prove Lemma 1.4.
1.49. Prove Theorem 1.9.
1.50. Prove the properties (a), (b), (c) and (d) in Section 1.12.
1.51. Give an example of a quasi-distance d on R for which the d-balls are open sets
but d is not of order α for any α > 0.
1.52. Prove that properties (a), (c) and (d.1) of Section 1.12 imply (d.2) with constant
K 2 . (Diego Maldonado)
1.53. Let (X, δ ) be a quasi-metric space. Let { Bδ (x, r) : x ∈ X, r > 0} be the family
of δ -balls on X. Apply Theorem 1.10 to this family in order to get a quasi-distance
d on X. Are δ and d related in some specific way?
1.54. Prove that Hausdorff dimension is invariant under changes of equivalent quasi-
distance.
1.55. Show that if d ∼ ρ γ , then dimH,ρ (E) = γ dimH,d (E), for any subset E of X.
1.56. Prove that for any subset E of X the infimum of dimH,d (E), taken over the set
of all quasi-distances d on X is zero. What if d ranges only on the set of all distances
on X?
1.57. Let (X, d) and (Y, δ ) be two quasi-metric spaces. Let 1 ≤ p < ∞, if
1
ρ p ((x1 , y1 ); (x2 , y2 )) = (d(x1 , x2 ) p + δ (y1 , y2 ) p ) p
and
ρ∞ ((x1 , y1 ); (x2 , y2 )) = max{ d(x1 , x2 ), ρ(y1 , y2 )} ,
are all equivalent quasi-distances on X×X.
1.58. Let X = R2 , E = Q = [0, 1] × [0, 1] the unit cube. Compute the Hausdorff di-
mension of Q with respect to
1
d((x1 , y1 ); (x2 , y2 )) = max{ |x1 − x2 |, |y1 − y2 | 2 }.
1.19 Problems and Comments 53
Compute also the dimension of an horizontal straight line and of any other straight
line in this setting.
1.59. Describe the topology generated on X = R by the function

0 if y ≥ x,
g(x, y) =
1 if y < x.
in the sense of Section 1.2.
1.60. Let (X, d) be a quasi-metric space and let ∼ be an equivalence relation on X

such that the equivalence classes {x̄} are closed and “ almost - parallel” in the sense
that
d(u, ȳ) ' d(x̄, ȳ) ' d(x̄, v)
for every u ∈ x̄, every v ∈ ȳ, every x ∈ X and every y ∈ X, with constants independent
of x,y,µ and v. Here x̄ is the class of x, the distance d(A, B) from the set A to the set
B is given by the infimum of d(u, v), u ∈ A and v ∈ B and the distance from a point u
to a set B is just d({u}, B). Show that d( ¯ x̄, ȳ) = d(x̄, ȳ) is a quasi-distance on X/ ∼.
1.61. Consider the metrization of quasi-metric spaces when X = Rn and d is the

translation invariant quasi-metric given by a quasi-norm k k satisfying
kxk = 0 if and only if x = 0,

k x + y k ≤ K ( kxk + kyk ),
ktx k ≤ k x k if |t | ≤ 1.
1.62. Let (X, d) be a quasi-metric space. Notice that the set of those positive numbers
α such that there exists a distance ρ on X for which ρ α ' d is unbounded above. Is
the infimum of this set generally a minimum?
1.63. Chapters 1 and 6 of John Kelley’s “ General Topology ” cover largely all the
basic aspects of general topology and uniform structures needed, including Frink’s
metrization lemma for uniform structure with countable basis.
The central topic of this chapter, metrization of quasi-metric spaces, can be found
in the fundamental paper “ Lipschitz Functions on Spaces of Homogeneous Type ”
by Roberto Macı́as and Carlos Segovia. See also [?].
The results in Sections 1.16 and 1.17, originally designed with different purposes,
both rely on a self-similarity construction, resembling the iteration schemes used in
fractal geometry. The results in Section 1.16 are due to Macı́as and Segovia and are
contained in their 1981 paper “ A well-behaved quasi-distance for spaces of homo-
geneous type ”. The results in Section 1.17 are essentially those proved by Michel
Christ in 1990 in the paper “ A T(b) theorem with remarks on analytic capacity and
the Cauchy integral ”. See also [?].
References
In view of the parallel print and (chapter-wise) online publication of your book
at www.springerlink.com it has been decided that – as a genreral rule – references
should be sorted chapter-wise and placed at the end of the individual chapters. How-
ever, upon agreement with your contact at Springer you may list your references
in a single seperate chapter at the end of your book. Deactivate the class option
sectrefs and the thebibliography environment will be put out as a chapter
of its own.
References may be cited in the text either by number (preferred) or by au-
thor/year.1 The reference list should ideally be sorted in alphabetical order – even
if reference numbers are used for the their citation in the text. If there are several
works by the same author, the following order should be used:
1. all works by the author alone, ordered chronologically by year of publication
2. all works by the author with a coauthor, ordered alphabetically by coauthor
3. all works by the author with several coauthors, ordered chronologically by year
of publication.
The styling of references2 depends on the subject of your book:
• The two recommended styles for references in books on mathematical, physical,
statistical and computer sciences are depicted in [1, 2, 3, 4, 5] and [6, 7, 8, 9, 10].
• Examples of the most commonly used reference style in books on Psychology,
Social Sciences are [11, 12, 13, 14, 15].
• Examples for references in books on Humanities, Linguistics, Philosophy are [16,
17, 18, 19, 20].
• Examples of the basic Springer style used in publications on a wide range of
subjects such as Computer Science, Economics, Engineering, Geosciences, Life
Sciences, Medicine, Biomedicine are [21, 22, 24, 23, 25].
1. Broy, M.: Software engineering — from auxiliary to key technologies. In: Broy, M., Dener,
E. (eds.) Software Pioneers, pp. 10-13. Springer, Heidelberg (2002)
2. Dod, J.: Effective substances. In: The Dictionary of Substances and Their Effects. Royal
Society of Chemistry (1999) Available via DIALOG.
http://www.rsc.org/dose/title of subordinate document. Cited 15 Jan 1999
3. Geddes, K.O., Czapor, S.R., Labahn, G.: Algorithms for Computer Algebra. Kluwer, Boston
(1992)
4. Hamburger, C.: Quasimonotonicity, regularity and duality for nonlinear systems of partial
differential equations. Ann. Mat. Pura. Appl. 169, 321–354 (1995)
5. Slifka, M.K., Whitton, J.L.: Clinical implications of dysregulated cytokine production. J. Mol.
Med. (2000) doi: 10.1007/s001090000086
1 Make sure that all references from the list are cited in the text. Those not cited should be moved
to a separate Further Reading section or chapter.
2 Always use the standard abbreviation of a journal’s name according to the ISSN List of Title Word
Abbreviations, see http://www.issn.org/en/node/344

References 55
6. J. Dod, in The Dictionary of Substances and Their Effects, Royal Society of Chemistry.
(Available via DIALOG, 1999), http://www.rsc.org/dose/title of subordinate document. Cited
15 Jan 1999
7. H. Ibach, H. Lüth, Solid-State Physics, 2nd edn. (Springer, New York, 1996), pp. 45-56
8. S. Preuss, A. Demchuk Jr., M. Stuke, Appl. Phys. A 61
9. M.K. Slifka, J.L. Whitton, J. Mol. Med., doi: 10.1007/s001090000086
10. S.E. Smith, in Neuromuscular Junction, ed. by E. Zaimis. Handbook of Experimental Phar-
macology, vol 42 (Springer, Heidelberg, 1976), p. 593
11. Calfee, R. C., & Valencia, R. R. (1991). APA guide to preparing manuscripts for journal
publication. Washington, DC: American Psychological Association.
12. Dod, J. (1999). Effective substances. In: The dictionary of substances and their effects. Royal
Society of Chemistry. Available via DIALOG.
http://www.rsc.org/dose/Effective substances. Cited 15 Jan 1999.
13. Harris, M., Karper, E., Stacks, G., Hoffman, D., DeNiro, R., Cruz, P., et al. (2001). Writing
labs and the Hollywood connection. J Film Writing, 44(3), 213–245.
14. O’Neil, J. M., & Egan, J. (1992). Men’s and women’s gender role journeys: Metaphor for
healing, transition, and transformation. In B. R. Wainrig (Ed.), Gender issues across the life
cycle (pp. 107–123). New York: Springer.
15. Kreger, M., Brindis, C.D., Manuel, D.M., Sassoubre, L. (2007). Lessons learned in systems
change initiatives: benchmarks and indicators. American Journal of Community Psychology,
doi: 10.1007/s10464-007-9108-14.
16. Alber John, Daniel C. O’Connell, and Sabine Kowal. 2002. Personal perspective in TV inter-
views. Pragmatics 12:257–271
17. Cameron, Deborah. 1997. Theoretical debates in feminist linguistics: Questions of sex and
gender. In Gender and discourse, ed. Ruth Wodak, 99–119. London: Sage Publications.
18. Cameron, Deborah. 1985. Feminism and linguistic theory. New York: St. Martin’s Press.
19. Dod, Jake. 1999. Effective substances. In: The dictionary of substances and their effects.
Royal Society of Chemistry. Available via DIALOG.
20. Suleiman, Camelia, Daniel C. OConnell, and Sabine Kowal. 2002. ‘If you and I, if we, in this
later day, lose that sacred fire...’: Perspective in political interviews. Journal of Psycholin-
guistic Research. doi: 10.1023/A:1015592129296.
21. Brown B, Aaron M (2001) The politics of nature. In: Smith J (ed) The rise of modern ge-
nomics, 3rd edn. Wiley, New York
22. Dod J (1999) Effective Substances. In: The dictionary of substances and their effects. Royal
Society of Chemistry. Available via DIALOG.
23. Slifka MK, Whitton JL (2000) Clinical implications of dysregulated cytokine production. J
Mol Med, doi: 10.1007/s001090000086
24. Smith J, Jones M Jr, Houghton L et al (1999) Future of health insurance. N Engl J Med
965:325–329
25. South J, Blass B (2001) The future of modern genomics. Blackwell, London
Chapter 2
The weak homogeneity property on quasi-metric
spaces. Assouad metric dimension
The total boundedness of bounded subsets in a quasi-metric space is a basic in-

gredient to obtain Wiener type covering lemmas which, in a sense, are the starting
point for the solution of some analytical problems such as Withney’s extensions,
Hardy-Littlewood inequalities and Calderón - Zygmund decompositions.
A quantitative and more restrictive form of this total boundedness of d-balls, is

the weak homogeneity property first considered by Coiffman and de Guzmán in
[?, ?] and Coiffman and Weiss [?, ?] as a necessary condition for the existence of
doubling measures on a quasi-metric space. This condition leads to the concept of
metric Assouad dimension [?] (See also [?] and [?]).
In this chapter we introduce some basic analysis on finite dimensional quasi-

metric spaces. We also give the most elementary geometric and analytic conse-
quences of this property: Wiener and Whitney type covering lemmas, estimates for
the Hausdorff dimension, existence of partitions of unity in Lipschitz classes and
extensions theorems for continuous and Lipschitz functions.
2.1 Weak homogeneity property. Assouad dimension.
Let (X, d) be a quasi-metric space with constant K. As in Chapter 1, a subset A

of X is said to be ε-disperse , for some positive ε, if d(x, y) ≥ ε whenever x ∈ A,
y ∈ A and x 6= y. We shall say that the space (X, d) satisfies the weak homogeneity
property (WHP) if there exists a constant N ∈ N such that for every x ∈ X and every
r > 0, every 2r -disperse subset A of B(x, r) has at most N points. In other words, if
A ⊂ B(x, r) with d(y, z) ≥ 2r for every y ∈ A, z ∈ A, we have that ](A) ≤ N, where
](A) is the number of elements of the set A.
57
58 2 The weak homogeneity property on quasi-metric spaces. Assouad metric dimension
The euclidian space satisfies the WHP but the space of continuous functions on
[0, 1] with d( f , g) = k f − gk∞ does not.
Lemma 2.1. Let (X, d) be a quasi-metric space satisfying the WHP with constant
N. Then for every x ∈ X, every r > 0 and every m ∈ N, every 2rm -disperse subset of
B(x, r) contains at most N m points.
Proof. The case m = 1 is just the definition of WHP. Let us prove the result for m+1
r
assuming that it is true for m as stated. Take A a 2m+1 -disperse subset of B(x, r). Let
us construct another subset Ã of B(x, r) maximal such that Ã is 2rm -disperse. Then
[
B xi , 2rm . From the induction hypothesis we know that ](Ã) ≤ N m .

B(x, r) ⊂
xi ∈Ã
On the other hand, each set Ai = A ∩ B xi , 2rm for xi ∈ Ã, contains at most N points,

since it is 12 2rm -disperse subset of a ball of radius 2rm . Then
](A) ≤ ∑ ](Ai ) ≤ N ](Ã) ≤ N m+1 .

xi ∈Ã
Let us observe that for 0 < ε < r we have that if A ⊂ B(x, r) is ε-disperse, then
r n
](A) ≤ C ,
ε
r
where C is a finite constant and n = log2 N. To obtain this, pick m such that 2m+1 ≤
r
ε < 2m and apply Lemma 2.1. Let us look for the “smallest” number n satisfying
the above inequality. Let (X, d) be a quasi-metric space satisfying WHP. Let x be
a point in X, r be a positive number and let λ be a number larger than one. There
exists a unique k ∈ N such that 2k−1 < λ ≤ 2k . Assume that A is an r-disperse subset
of X. Then, from Lemma 2.1.
B(x, λ r) ∩ A ≤ ] B(x, 2k r) ∩ A

≤ N k ≤ Nλ log2 N .
Hence, the set S of the non-negative numbers α for which there exists a constant
C = C(α) such that the inequality

] B(x, λ r) ∩ A ≤ Cλ α ,
2.2 Metric and topological properties of quasi-metric spaces satisfying WHP. 59
holds for every x ∈ X, every positive r, every r-disperse subset A of X and every
λ > 1 is non-empty. The non-negative real number
dimA X = inf S,
is called the Assouad metric dimension of X. (See [?], [?]). Given a general quasi-
metric space (X, d), we define the Assouad uniform metric dimension of X as the
infimum of the set S if S 6= 0/ and dimA X = + ∞ if S = 0.
/
We conclude the section stating the results as a theorem.
Theorem 2.1. Let (X, d) be a quasi-metric space. Then dimA < ∞ if and only if
(X, d) satisfies WHP. Moreover, if N is a constant for WHP, then dimA ≤ log2 N.
2.2 Metric and topological properties of quasi-metric spaces

satisfying WHP.
The next result contains the main additional metric and topological properties of a
quasi-metric space with the WHP.
Lemma 2.2. Let (X, d) be a quasi-metric space with the WHP. Then
1. (X, d) is separable;
2. the topology τ induced by d has a countable basis;
3. bounded sets are totally bounded sets;
4. if (X, d) is complete, then, the Heine-Borel property holds: a subset A of X is
compact if and only if A is closed and bounded.
Proof. Notice that 2.3.2 follows from 2.3.1, 2.3.3 from Lemma 2.1 and 2.3.4 from
2.3.3 and Theorem 1.31 of Chapter 1. Let us prove 2.3.1. Take x0 a fixed point in X.
Let Dkl be a subset of B(x0 , k) maximal with dispersion 1l ; k ∈ N, l ∈ N. Since Dkl
[ [
is a finite set and X = B(x0 , k), we have that D = Dkl is a countable and dense
k∈N k,l
subset of X.
2.3 Hausdorff and Assouad dimensions.
In this section we prove that Hausdorff dimension is less than or equal to Assouad
dimension on a general quasi-metric space.
Theorem 2.2. Let (X, d) be a quasi-metric space, then dimH X ≤ dimA X.
Let us start by stating and proving an elementary covering lemma.
Lemma 2.3. Let (X, d) be a quasi-metric space with s = dimA X < ∞ then, for every
t > s there exists a constant C = C(t) such that for every x ∈ X and every R > r > 0,
t
the ball B(x, R) can be covered by less than C Rr balls, each of radius r.
Proof. Since dimA X = s and t > s, there exists C = C(t) such that the inequality
](B(x, λ r) ∩ A) ≤ Cλ t (2.1)
hold for every x ∈ X, every λ > 1, every r > 0, and every A r-disperse subset of X.
Let us now fix 0 < r < R and take A = { xi } a maximal r-disperse subset of B(x, R).
It is clear that the family { B(xi , r) : xi ∈ A} is a covering for B(x, R). Moreover,
applying (2.1) with λ = Rr > 1, we get that
] (A) = ] (B(x, R) ∩ A)
= ] B x, Rr r ∩ A

t
≤ C Rr ,
which gives the desired bound for the number of covering balls.
Proof of Theorem 2.4. We can assume that dimA X = s < ∞. Then, for every ball
B0 = B(x0 , R) we also have that dimA B0 ≤ s. Let us show that for every t > s we
have that H t (B0 ) < ∞, where H t is the t-dimensional Hausdorff outer measure.
t
Pick δ > 0. From Lemma 2.5, B0 can be covered by less than C δR balls with
[ t
radii K: Bi ⊃ B0 , ](I) ≤ C δR and diameter of Bi = |Bi | < δ . Hence
i∈I
2.4 Wiener type covering lemmas. 61
Hδt (B0 ) ≤ ∑ |Bi |t

i∈I
≤ ](I)δ t
t
≤ C δR δ t
= CRt ,
thus H t (B0 ) < ∞, for every t > s. Now, since H t (B0 ) < ∞, then H τ (B0 ) = 0 for
every τ > t and every ball B0 in X. Being H t an outer measure on the subsets of
X and being X a countable union of balls, we conclude that H τ (X) = 0 for every
τ > s, so that dimH (X) ≤ s.
The fact that the inequality in Theorem 2.4 can be strict can easily be proved by
constructing a sequence in R1 with positive Assouad dimension ( Section 2.14).
2.4 Wiener type covering lemmas.
Let (X, d) be a quasi-metric spaces. A given family F of d-balls in X can be seen

as a covering from several points of view. In some sense two extreme situations are
paradigmatic. The first one is to consider the set E = {C(B) : B ∈ F }, where C(B)
is the center of the given ball B, as the set being covered by F . The second, instead,
is to consider as the covered set the larger set given by the union of all members of
F . Between these two cases several variants are possible and relevant for particular
analytic problems. Sometimes the function C(B) has to be changed to some specific
function p(.) assigning to each ball B some point p(B) ∈ B.
When a given family of balls covers a bounded set of the euclidian space Rn , two
usual types of subcoverings are basic in real analysis: Wiener and Besicovitch. Only
the first type can actually be extended to our general setting.
Let us start by introducing the basic selection scheme for the Wiener type cover-
ing in a very simple finite situation.
Lemma 2.4. Let (X, d) be a quasi-metric space. Let E be a finite subset of X. As-
sume that for each point x ∈ E a ball centered at x is given or in other words, a
function r : E → R+ is given. Assume also that r is one to one. Then, there exists
F ⊂ E, F = {y1 , y2 , · · · , yk } such that with Bi = B(yi , r(yi )) and B0i = B(yi , 2K(yi ))
we have that
(a)Bi ∩ B j = 0/ if i 6= j;
k
B0i .
[
(b)E ⊂
i=1
Proof. Since r : E → R+ is one to one we can write E = {x j : j = 1, · · · , m = ](E)}

with r(x1 ) > r(x2 ) > · · · > r(xm ). To construct the subset F with the desired proper-
ties, let us pick its “first element” y1 = x1 . To choose “the second” y2 , let us first con-
/ B01 }, with B01 = B(y1 , 2K(y1 )). If this set is empty
sider the set { j = 1, · · · , m : x j ∈
take F = {y1 } and we certainly have (a) and (b). Otherwise, take j2 = min{ j =
/ B01 } and y2 = x j2 . Set B2 = B(y2 , r(y2 )) and B02 = B(y2 , 2K/y2 )).
1, · · · , m : x j ∈
Notice that B1 ∩ B2 = 0/ since, otherwise, with z ∈ B1 ∩ B2 we should have that
d(y1 , y2 ) ≤ K(d(y1 , z) + d(y2 , z)) < k(r(y1 ) + r(y2 )) < 2K(y1 ) which is in contradic-
tion with y2 ∈ / B01 . Assume that y1 , y2 , · · · , yl have been chosen in this way satisfying
l
B0i } is empty, we
[
Bi ∩ B j = 0/ for i 6= j, i, j = 1, · · · , l. If the set { j = 1, · · · , m : x j ∈
/
i=1
l
B0i } =
[
take F = {y1 , · · · , yl } and we are done. If, otherwise, { j = 1, · · · , l : x j ∈
/ 6 0,
/
i=1
denote its first element by jl+1 and define yl+1 = x jl+1 . Set Bl+1 = B(yl+1 , r(yl+1 ))
and B0l+1 its 2K-dilation. With the same proof as before we see that Bl+1 ∩ B j = 0/
for j = 1, · · · , l. Since E itself is finite, the process must stop and we obtain (a) and
(b) for the selected sequence.
Notice that no finite dimension of the space is required in the above lemma. But,
of course, in general quasi-metric spaces the finite sets are really far away from
“controlling”, in some sense, the general subsets or even the balls of the space.
The WHP or the total boundedness of bounded sets provide elementary tools to
recover that control. Another way of approaching continuous from discrete (finite)
is the compacity-continuity argument, but the standard uses of covering lemmas in
analysis, generally do not assume this form.
Let us observe that the hypothesis of injectivity on the function r(x) is only a
matter of simplicity and by no means essential.
In this more general situation the selection process can be done in the same way.
The only difference is that at each step one has to choose one of the several pos-
sible candidates and so, the algorithm is defined also by this option. Hence several
subcoverings satisfying (a) and (b) in Lemma 2.6 are possible.
Let us now introduce a general selection procedure that works in the non finite
case. Let (X, d) be a quasi-metric space, let E be a subset of X and let r : E → R+ be
a bounded function. Consider the covering of E by the family of balls {B(x, r(x)) :
x ∈ E}
Let ε be a positive number with ε < 1. Since r(.) is a bounded function, then
sup r(x) is finite and positive. Let us choose x1 ∈ E =: E1 in such a way that
x∈E
r(x1 ) > (1 − ε) sup r(x). Set B1 = B(x1 , r(x1 )) and B̃1 = B x1 , 2−ε

1−ε K(x1 ) . Let us
x∈E1
now consider the subset of E1 defined by E2 = E1 \ B̃1 . Pick x2 ∈ E2 such that
r(x2 ) > (1 − ε) sup r(x). Notice that B1 ∩ B2 = 0,
/ since otherwise if d(z, x1 ) < r(x1 )
x∈E2
2−ε
and d(z, x2 ) < r(x2 ) we would have that d(x1 , x2 ) < K(r(x1 ) + r(x2 )) < 1−ε K(x1 ),
which is a contradiction. The last inequality follows from the inequality r(x2 ) ≤
r(x1 )
sup r(x) < , since x2 ∈ E1 . In this way we can obtain a sequence {xi } of
x∈E1 1−ε
points in E and an associated sequence of disjoint balls {Bi = B(xi , r(xi ))} with
i)
r(x j ) < r(x
1−ε whenever j > i, which could be finite if some Ek = Ek−1 \ B̃k−1 = φ ,
2−ε

with B̃l = B xl , 1−ε K(xl ) . We shall say that these sequences have been obtained
from the original family {B(x, r(x)) : x ∈ E} by the Wiener selection process as-
sociated to ε. The usual ε is 21 , so the family of dilations of the disjoint balls is
B̃l = B(xl , 3K(xl )).
Lemma 2.5. Let (X, d) be a quasi-metric space with the WHP. Let E be a bounded
subset of X and r : E → R+ a given positive function. Let 0 < ε < 1. Then, there
exists a finite or countable sequence {xi } ⊂ E such that the balls Bi = B(xi , r(xi ))
2−ε
are pairwise disjoint and the balls B̃i = B(xi , 1−ε r(xi )) cover E;
[
E⊂ B̃i .
i
Proof. Since the set E is bounded, if the function r(x) is unbounded, there exists
one point x1 ∈ E such that E ⊂ B(x1 , r(x1 )). Assume, then, that r(x) is a bounded
function. So that we can apply the Wiener selection process associated to ε in or-
der to obtain a sequence {xi } ⊂ E such that the balls Bi = B(xi , r(xi )) are pairwise
disjoint.
[
To prove the lemma we only need to show that E ⊂ B̃i . If the selection process
i
stops at some level k, it happens only because Ek = 0,
/ and this fact is equivalent to
k−1
[
E⊂ B̃i . Let us assume that we actually have a sequence { xk : k ∈ N} ⊂ E and
i=1
the corresponding sequence of real numbers { rk = r(xk ) : k ∈ N}. Then rk → 0 as
k → ∞. Otherwise, for some positive δ and some subsequence rk j ; j ∈ N, we should
have rk j > δ . This fact would imply that the subsequence { xk j ; j ∈ N} is δ -disperse.
From the WHP we see that this is impossible since { xk j ; j ∈ N} ⊂ E is a bounded
[
set. To show that E ⊂ B˜k , take x ∈ E. Since the function r is positive on E, there
k
exists k ∈ N such that rk < (1 − ε)r(x). Then
(1 − ε)r(x) > rk > (1 − ε) sup r(z);

z∈Ek
so that x ∈
/ Ek . Then, for some l < k we necessarily have that x ∈ B̃l .
Let us notice that the argument used in the proof of Lemma 2.6 can be repeated
if instead of the weak homogeneity property we only have that d-balls are totally
1
bounded sets. Notice also that taking ε = 2 the covering balls become the usual
B(xi , 3Kr(xi )).
Lemma 2.6. Let (X, d) be a quasi-metric space such that d-balls are totally
bounded. Let E be a bounded subset of X and r : E → R+ a given function. Then
there exists a sequence {xi } ⊂ E such that the balls Bi = B(xi , r(xi )) are pairwise
disjoint and the sequence {B̃i } is a covering of E.
If the function[r(.) is bounded, not only the set E is covered by the sequence {B̃i }
but also the set B(x, r(x)). Moreover each B(x, r(x)) is contained in some fixed
x∈E
dilation of Bi for some i.
Lemma 2.7. Let (X, d) be a quasi-metric space with finite Assouad dimension. Let
E be a bounded subset of X and let r : E → R+ be any positive real valued bounded
function defined on E. Then, there exists a countable sequence {xi : i ∈ I}, which
can be finite, of points of E such that if Bi = B(xi , ri ); B̃i = B(xi , 3Kri ) and B̄i =
B(xi , 5K 2 ri ) with ri = r(xi ), we have
(i) Bi ∩ B
[j
= 0/ for i 6= j;
(ii)E ⊂ B̃i ;
i∈I
(iii)for every x ∈ E there exists i ∈ I such that B(x, r(x)) ⊂ B̄i ,
(iv)for every x ∈ E and every M > 0 there exists i ∈ I such that B(x, Mr(x)) ⊂ BM
i =
B(xi , K(2M + 3K)ri ).
Proof. Since the hypothesis of Lemma 2.8 are satisfied, we obtain a sequence {xi :
i ∈ I} ⊂ E such that the corresponding balls Bi and B̃i satisfy (i) and (ii) with ε = 12 .
[
Let us prove that (iii) is also satisfied. Since E ⊂ B̃i , given x ∈ E, the set {i ∈
i∈I
I : x ∈ B̃i } is non-empty. Recall that I is N or any initial interval of N. Hence the
set {i ∈ I : x ∈ B̃i } has a first element that we shall denote by i(x). If i(x) = 1 we
have that r(x) ≤ sup r < 2r1 , since we are assuming that r(.) is bounded. If i(x) ≥ 2
E
and i(x) is the first with the property x ∈ B̃i(x) , we necessarily have that x ∈
/ B̃k for
r(x)
k = 1, 2, · · · , i(x) − 1. Hence 2 ≤ 12 sup r < rk for k = 1, 2, · · · , i(x). So, in any case,
Ek
r(x) < 2ri(x) . Since x ∈ B̃i(x) we see that if z ∈ B(x, r(x)), then
d(z, xi ) ≤ K[d(z, x) + d(x, xi )]

< K[r(x) + 3Kri ]
< 5K 2 ri ,
which proves that B(x, r(x)) ⊂ B̄i . The result in (iv) follows in the same way since
now, given z ∈ B(x, Mr(x)) we have that
d(z, xi ) ≤ k[d(z, x) + d(x, xi )]

< k[Mr(x) + 3kri ]
< k(2M + 3k)ri .
Sometimes a non-centered version of the Wiener type covering lemma is useful.

Lemma 2.8. Let (X, d) be a quasi-metric space with finite Assouad dimension. Let
E be a bounded subset of X, let r : E → R+ be a bounded function and let y : E → X
be given in such a way that x ∈ B(y(x), r(x)) for every x ∈ E. Then there exists a
sequence {xi : i ∈ I} of points in E such that the balls Bi = B(yi , ri ) = B(y(xi ), r(xi ))
are pairwise disjoint and the set E is covered by the sequence {B∗i : i ∈ I} where
B∗i = B(yi , 10K 2 ri ).
Proof. Notice first that y(E), the image of E through y, is a bounded subset of
X. In fact, if E ⊂ B(x0 , R0 ) and x ∈ E, the d(y(x), x0 ) ≤ K(d(y(x), x) + d(x, x0 )) <
K/r(x) + R0 ), which is uniformly bounded since r(x) is assumed to be bounded.
For each z ∈ y(E) let us consider the set y−1 ({z}) = {x ∈ E : y(x) = z} and R(z) =
{r(x) : x ∈ y−1 ({z})}. Since r(x) is bounded we can choose x = x(z) ∈ y−1 ({z}) in
such a way that r(x(z)) = r0 (z) ≥ 21 sup R(z). Of course we have that r0 : y(E) → R+
is a bounded function on the bounded set y(E). So that, we can apply the previous
lemma with y(E) instead of E and r0 instead of r in order to get a sequence Bi =
B(zi , ri0 ) with ri0 = r0 (zi ) such that Bi ∩ B j = φ if i 6= j and for every z ∈ y(E) there
exists i ∈ I such that B(z, 2r0 (z)) ⊂ Bi = B∗ (zi , 10K 2 ri0 ). Take xi = x(zi ) ∈ y−1 ({zi }),
then Bi = B(y(xi ), r(xi )) = B(yi , ri ) and B∗i = B(yi , 10K 2 ri ). We only have to prove
that {B∗i : i ∈ I} is a covering for E. If x ∈ E, then x ∈ y−1 ({z}) for some z ∈ y(E).
For this z, there exists i ∈ I such that B(z, 2r0 (z)) ⊂ Bi . Also x ∈ B(y(x), r(x)) =
B(z, r(x)) ⊂ B(z, 2r0 (z)). Hence x ∈ Bi .
Let us observe that the sequence of intervals n1 , n is a sequence of balls in R1

covering the set R+ of positive real numbers such that no dilation of no disjoint sub-
family suffices to cover R+ . Notice also that applying the Wiener selection process
to the family of intervals n − 12 , n + 12 , n ∈ N in R1 , we could get for example the
sequence 2 j − 21 , 2 j + 12 , j ∈ N as a disjoint countable subfamily, but no fixed di-

lation of these interval suffice to cover the set of counters, N, of the given family. As
A.P. Calderón showed in [C], by improving the selection process, the boundedness
of E is irrelevant if the function r(B) is bounded on the given family of balls.
Lemma 2.9. Let F be a family of balls with bounded radii in the quasi-metric space
(X, d) which satisfies the WHP. Then there exists a disjoint sequence {B(xi , ri )} ⊂ F
such that for each ball B ∈ F there exists i for which B ⊂ B(xi , 4K 2 ri ).
Proof. Since r(B) is bounded above on F and we are assuming that F = 0, / we
have that M = supB∈F r(B) is finite and positive. Let 0 < λ < 1 to be fixed later. Set
Fk , k ∈ N, to denote the subfamily of those balls B in F such that λ k M < r(B) ≤
λ k−1 M. Let us take F̃1 to be any maximal disjoint subfamily of F1 . This means
that
1. (a1 ) F˜∞ ⊆ F∞ ;
2. (b1 ) B ∩ B0 = 0/ if B 6= B0 and both belong to F˜∞ ;
3. (c1 ) if B ∈ F∞ \ F˜∞ , then there exists B0 ∈ F˜∞ such that B ∩ B0 6= 0.
/
Since the balls in F˜∞ have radii bounded below by λ M and since the quasi-metric
space (X, d) has finite Assouad dimension, we conclude that F˜∞ is a countable
family of balls. Let us use the notation B1i = B(x11 , ri1 ); i ∈ I1 , for the balls in F˜∞ .
Here I1 is an initial interval in N which could be the whole N. By induction, for
k ∈ N, k ≥ 2, take F̃k such that
1. (ak ) F̃k ⊆ Fk ;
k
2. (bk ) B ∩ B0 = 0/ if B 6= B0 and both belong to F̃| ;
[
j=1
k
3. (ck ) if B ∈ Fk \ F̃k , then there exists B0 ∈ F̃| such that B ∩ B0 6= 0.
[
/
j=1
For the same reasons as in the case k = 1 we have that
F̃k = {Bki = B(xik , rik ) : i ∈ Ik }
for some finite or countable initial interval Ik of N. We claim that the family F˜ =
∞
F̃k has the desired properties. The balls in F˜ are pairwise disjoint from the
[
k=1
construction of each F̃k . We only have to check that choosing a suitable λ ∈ (0, 1),
each ball in F is contained in 4K 2 -dilation of some ball in F˜ . Take B ∈ F , hence
there exists only one k ∈ N such that B ∈ Fk . If B ∈ F̃k there is nothing to prove. If
B ∈ Fk \ F̃k , from (ck ) we have that there exists Bij = B(xij , rij ) ∈ F̃| for some j ≤ k
/ Let x = C(B), r = r(B) and y ∈ B ∩ Bij . Since
and some i ∈ I j such that B ∩ Bij 6= 0.
B ∈ Fk , we have that λ M < r ≤ λ M. Hence, given z ∈ B, applying the triangle
k k−1
inequality we have that
d(z, xij ) ≤ K(d(z, x) + d(x, xij ))

< K(r + K(d(x, y) + d(y, xij )))
< K(r + K(r + rij )).
j
ri
Since λ j M < rij ≤ λ j−1 M and k ≤ j, we have that r ≤ λ k−1 M ≤ λ1 λ j M < λ , we
finally obtain that
j K 2 1
d(z, xi ) < +K + 1 rij .
λ λ
Taking λ = K+1
3K we get that B ⊂ B(xij , 4K 2 rij ) as desired.
This dyadic type selection process introduced by A.P. Calderón can also be ap-
plied to obtain some kind of coverings with controlled overlapping instead of the
disjointness of the selected family. ( See Figure 1)
Notice that in Lemma 2.10 even when (X, d) is a metric space, i.e. K = 1, the
infimum of the dilation constants defining the covering family {B̃i } is 2. This fact
is irrelevant in general when dealing with doubling measures in analysis but it is
really an obstruction in non-doubling settings. In Euclidian settings Besicovitch type
lemmas are the natural substitutes since the covering sequence is {Bi } itself, no
enlarging is needed. We loose disjointness but we still have bounded overlapping.
The fact is that Besicovitch lemmas do not hold in general quasi-metric spaces with
finite Assouad dimension. (See [AF] and Chapter 4). The next result essentially due
to GuoghenLu ([L]), is in a sense half a way between Vitali-Wiener and Besicovitch.
winfig1-eps-converted-to.pdf
Fig. 2.1 COLOCAR ALGUNA REFERENCIA
Let G be the family of selected balls, G = {Bα1 , · · · , Bαm }. Let x0 be a fixed point
in X. Set G (x0 ) = {B ∈ G : x0 ∈ B},r(x0 ) = min{r(B) : B ∈ G (x0 )} and G k (x0 ) =
{B ∈ G (x0 ) : 2k r(x0 ) ≤ r(B) < 2k+1 r(x0 )}, k = 0, 1, 2, · · · . Notice that G k (x0 ) 6= 0/
only for a finite set of values of k. Let σ (x0 ) be the maximum of those values of k.
Hence, since
m σ (x0 )
∑ XBαi (x0 ) = ∑ XB (x0 ) = ∑ ∑ XB (x0 )
i=1 B∈G (x0 ) k=0 B∈G k (x0 )
the desired result shall be a consequence of the following estimates:

] G k (x0 ) ≤ Cδ −s ; (2.2)
1
σ (x0 ) ≤ C log . (2.3)
δ
Let us prove (2.2). Since for each B ∈ G k (x0 ), x0 ∈ B and r(B) < 2k+1 r(x0 ), we
have that Ck (x0 ) = {C(B) : B ∈ G k (x0 )} ⊂ B(x0 , 2k+1 r(x0 )}. Let us look for the
dispersion of the set Ck (x0 ), using that for B ∈ G k (x0 ) we also have r(B) > 2k r(x0 ).
k
0 )d
Notice that if d(C(B1 ),C(B2 )) < 2 r(x
K , with B1 and B2 in G k (x0 ) and, let us say,
B2 selected later than B1 , we have that B2 ⊂ B˜1 which is a contradiction. In fact,
take y ∈ B2 , then
d(y,C(B1 )) ≤ K(d(y,C(B2 )) + d(C(B2 ),C(B1 )))

2k r(x0 )δ

< K r(B2 ) +
K
≤ Kr(B1 ) + 2k r(x0 )δ
≤ (K + δ )r(B1 ).
2k r(x0 )δ
Hence Ck (x0 ) is a K -disperse subset is contained in a ball of radius 2k+1 r(x0 ) =
2K 2k r(x0 )δ
δ K . From the results of the previous section we conclude that

] G k (x0 ) = ] Ck (x0 ) ≤ Cδ −s
for s > dimA (X) and C a constant. Let us prove now the inequality (2.3). More
precisely we shall prove that for δ small and positive we have that δ 2σ (x0 ) ≤ 2K 2 .
Let B0 ∈ G (x0 ) such that r(B0 ) = r(x0 ). Hence B0 ∈ G 0 (x0 ). If σ (x0 ) = 0, the
desired inequality is immediate, since δ < 1. If, otherwise, σ (x0 ) ∈ N, let us take a
ball B ∈ G σ (x0 ) (x0 ) and prove that if the opposite inequality, δ 2σ (x0 ) (x0 ) and prove
that if the opposite inequality, δ 2σ (x0 ) > 2K 2 , is true we have the contradiction
B0 ⊂ B̃. In fact, for y ∈ B0 , we have
d(y,C(B)) ≤ K(d(y, x0 ) + d(x0 ,C(B)))

≤ K(K(d(y,C(B0 )) + d(C(B0 ),C(B))) + d(x0 ,C(B)))
< K(2Kr(x0 ) + r(B))
< δ 2r(x0 ) r(x0 ) + Kr(B)
≤ (δ + K)r(B).
the last inequality holds because B ∈ }σ (x0 ) (x0 ).
Lemma 2.10. Let (X, d) be a quasi-metric space with constant K of finite Assouad
dimension s. Then there exists δ0 > 0 such that for every δ0 > δ > 0 and every {Bα :
α ∈ A}, a finite collection of balls in X, there exists Γ ⊂ A, Γ = {α1 , α2 , · · · , αn }
such that
[ n
[
(a) Bα ⊂ B̃αi ;
α∈A i=1
n
1
(b) ∑ XBαi (x) ≤ Cδ −s log δ
i=1
with B̃αi the ball concentric with Bαi and K + δ times its radius and C a constant
that depends only on K and s.
2.5 Whitney type covering lemma 69
Notice that when (X, d) is a metric space, since K = 1, we have that the rate of
dilation of the covering balls is 1 + δ . So[that, the subsequence {Bαi } can be chosen
in such a way that they “almost cover” Bα , with overlapping controlled by (b).
α∈A
Proof of Lemma 2.11: Let δ be a given positive number. Let Bα1 be one of the
balls with largest radius in {Bα : α ∈ A}. Assuming that Bα1 , · · · , Bαm have been
m
[
selected, consider the family of those Bα , α ∈ A, such that Bα * (K + δ )Bαi . If
i=1
this family is empty we have (a) with n = m and the selection stops. If, otherwise,
some Bα is not covered by B̃α1 to B̃αm , we select one of these balls with largest
radius. Since {Bα : α ∈ A} itself is finite the selection process must stop and the
resulting family Bα1 , · · · , Bαn must satisfy (a).
Let us now show that δ can be chosen enough satisfying also (b).
2.5 Whitney type covering lemma
In the previous section we have proved the Wiener type covering lemma for bounded
subsets of a quasi-metric space for which the balls are totally bounded sets.
Applying this result to a special function r(x) we get in this section a Whitney
type covering lemma.
Theorem 2.3. Let (X, d) be a quasi-metric space with constant K. Let us assume
that every d-ball in X is a totally bounded set. Let Ω be an open and bounded
proper subset of X. Let λ be a given real number such that λ ≥ 1. Then, there exists
a sequence Bi = B(xi , ri ) of d-balls such that
1. B
[i
∩ B j = 0/ for i 6= j;
2. B̃i = Ω with B̃i = B(xi , 3Kri );
i
3. B∗i := B(xi , 3Kλ ri ) ⊂ Ω ;
4. for every x ∈ B∗i we have that 3Kλ ri ≤ d(x) ≤ 9K 3 λ ri , where d(x) = d(x,CΩ ) =
inf{ d(x, y) : y ∈ / Ω };
5. for every i there exists yi ∈/ Ω such that d(xi , yi ) < 9K 2 λ ri ,
6. for every n, the set { k : Bk ∩ B∗n 6= 0}
∗ / is finite.
Moreover, if dimA X < ∞, there exists a constant M such that
7. ]({ k : B∗k ∩ B∗n 6= φ }) ≤ M, for every n.
Proof. The function r : Ω → R+ given by

d(x)
r(x) = ,
6K 2 λ
with d(x,CΩ ) = d(x) can be used to apply Wiener[ type lemma. In this way we get
a sequence { Bi } = { B(xi , ri )} satisfying 2.8.1 and B̃i ⊃ Ω . To get 2.8.3 and then
i
2.8.2 we only have to observe that
d(xi ) d(xi )
3Kλ ri = 3Kλ = < d(xi ).
6K 2 λ 2K
Let us check the first inequality in 2.8.4. Since for every z ∈

/ Ω we have that d(xi ) ≤
d(xi , z) ≤ K(d(xi , x) + d(x, z)), then
d(xi ) − Kd(xi , x)
d(x, z) ≥
K
d(xi ) − K3Kλ ri
≥
K
d(xi )
d(xi ) − K3Kλ 6K 2λ
=
K
d(xi )
= = 3Kλ ri .
2K
Hence d(x) ≥ 3Kλ ri , as desired. To establish the upper bound for d(x) in 2.8.4,
given a positive number ε, take z ∈ / Ω in such a way that
d(xi , z) ≤ d(xi , z) ≤ d(xi ) + ε. Now, from the triangle inequality, we have that
d(x, z) ≤ K(d(x, xi ) + d(xi , z))

< K(3Kλ ri + d(xi ) + ε
≤ K(3Kλ ri + 6K 2 λ ri ) + εK
≤ 9K 3 λ ri + εK.
Let us now check 2.8.5. Since 9K 2 λ ri is larger than 6K 2 λ ri = d(xi ) and d(xi )
is the limit of a sequence d(xi , ym ), m ∈ N, ym ∈
/ Ω , we have the result. In order to
prove at once 2.8.6 and 2.8.7 we only have to observe that there exist constants N
and L such that for every k with B∗k ∩ B∗n 6= 0,
/ we have
2.6 Lipschitz type partition of unity. 71
(a) Bk ⊂ B(xn , Nrn );

(b) rk ≥ Lrn .
Both, (a) and (b) follow from 2.8.4, since for ξ ∈ B∗k ∩ B∗n we have that
3Kλ rk ≤ d(ξ ) ≤ 9K 3 λ rk
and
3Kλ rn ≤ d(ξ ) ≤ 9K 3 λ rn ,
hence rk and rn are comparable.
2.6 Lipschitz type partition of unity.
Whitney Type covering lemma is a basic tool for the construction of Whitney exten-
sion operators of regular functions defined on closed subsets of the whole space. A
second step in Whitney’s construction is to associate a regular partition of unity to
this particular covering of open sets.
Let us start by taking a C ∞ -function η defined in [0, ∞) such that η is supported

in [0, 2), η is identically one on [0, 1] and 0 ≤ η ≤ 1. Let (X, d) be a quasi-metric
space such that the d-balls are totally bounded sets. From the results in Chapter 1,
we know that there exist a distance ρ on X and a number β ≥ 1 such that δ = ρ β
is a quasi-distance equivalent to d. Assuming that K is the constant for δ and that
with B we denote the δ -balls in X, we can apply Whitney Lemma as stated in the
previous section for the space (X, d). In other words, if Ω is a bounded open subset
of X with X 6= Ω and λ ≥ 1, then, we get a sequence {Bi } of δ -balls satisfying 2.8.1
to 2.8.7 with δ instead of d everywhere. With that notation for the sequence {rn },
let us define
!
δ (x, xn )
ψn : X → R; ψn (x) = η .
3Krn
taking λ ≥ 2, we see that
B˜n = B(xn , 3Krn ) ⊂ supp ψn ⊂ B(xn , 6Krn ) ⊂ B(xn , 3Kλ rn ) = B∗n .

Hence, from 2.8.2, since ψn ≡ 1 on each B˜n , we have
XΩ ≤ ∑ XB̃n ≤ ∑ ψn .
n n
Notice that for fixed x ∈ Ω , from 2.8.6, the sum is finite.
On the other hand, if the space (X, d) has finite Assouad dimension, also (X, δ )
has finite dimension, then, there exists a constant M such that no point of Ω belongs
to more than M balls B∗n . Thus
∑ ψn ≤ ∑ XB∗n ≤ MXΩ .
n n
So that if (X, d) has the weak homogeneity property, we have both
XΩ ≤ ∑ ψn ≤ MXΩ .
Let us estimate the Lipschitz α norms of each ψn for α = β −1 in the space (X, d).
Assume that x and y are two points in B(xn ,Crn ) for some C larger than 3K. Then
! !
ρ β (x, xn ) ρ β (y, yn )
|ψn (x) − ψn (y)| = η −η
3Krn 3Krn
≤ k η 0 k∞ (3Krn )−1 ρ β (x, xn ) − ρ β (y, yn )
β −1
≤ k η 0 k∞ (3K)−1 (Crn )1 β rn−1 ρ(x, xn ) − ρ(y, yn )
−1/β
≤ Crn ρ(x, y)
−1/β 1/β
≤ C̃rn d (x, y)
−α α
= C̃rn d (x, y) .
Assume now that x ∈ B(xn , 6Krn ) and y ∈

/ B(xn ,Crn ) then, since for an appropri-
ate choice of C we have d(x, y) ≥ rn and
|ψn (x) − ψn (y)| = |ψn (x)| ≤ 1 ≤ rn−α d(x, y)α .

2.6 Lipschitz type partition of unity. 73
For a given function f defined on the quasi-metric space (X, d) with real values,
we define the Lipschitz α (0 < α ≤ 1) norm of f as the infimum of the constants A
for which the inequality
| f (x) − f (y)|
≤A
d(x, y)α
holds for every x,y ∈ X with x 6= y. Let us denote the norm be k f kα or k f kLip α .
Hence from the above consideration we have that
kψn kα ≤ Crn−α . (2.4)
Let us notice that since rn is comparable to the distance of the support of ψn to

CΩ in terms of ρ β , it is also comparable to the d-distance of supp ψn to CΩ and to
the d-diameter of the support of ψn . Let us now construct the partition of unity on
Ω . Define the sequence of functions φn : X → R given by
(
0 if x ∈
/ Ω,
φn (x) = ψn (x)
∑k ψk (x)
if x ∈ Ω .
We have that
supp φn ⊂ supp ψn ⊂ B(xn , 6Krn ) ⊂ B∗n . (2.5)
∑ φn = XΩ . (2.6)
n
These properties are true with the assumption that every ball in X is a totally
bounded set. For the next two properties of the sequence {φn } we shall assume that
the space (X, d) has finite Assouad dimension.
ψn
Since, in this case, ∑n ψn ≤ MXΩ , we have that φn ≥ M, hence
φn (x) ≥ M −1 for x ∈ Bn . (2.7)
Let us finally prove that for φn we have an estimate of type (2.2) for its Lipschitz
norm.
kφn kα ≤ Crn−α . (2.8)
So far we have used only λ ≥ 2. Let us now assume that λ ≥ 3K. In order to get
(2.6) let us start by considering the difference φn (x) − φn (y) for x ∈ B(xn , 6Krn ) and
y∈/ B(xn , 7K 2 rn ). Since in this case we have that δ (x, y) ≥ rn , we easily have the
desired estimate
!α
δ (x, y)
|φn (x) − φn (y)| ≤ 1 ≤
rn
−α
≤ Crn−α d(x, y)

.
The only nontrivial remaining case takes place when x ∈ B(xn , 6Krn ) and y ∈
B(xn , 7K 2 r − n). Since λ ≥ 3K, we have that both x and y belong to B∗n , hence
|ψk (x) − ψk (y)| = 0 for all k except perhaps for no more than M values of the index
k. Moreover, for these values of k we have that rn ∼ rk . With this remarks we have
that
ψn (x) ψn (y)
|φn (x) − φn (y)| = −
ψ
∑k k (x) ∑k ψk (y)
|ψn (x) − ψn (y)|
≤ +
∑k ψk (x)
∑k ψk (x) − ψk (y)
+ ψn (y)
∑k ψk (x) ∑k ψk (y)
≤ Crn−α
α
d(x, y) +
+ ∑ ψk (x) − ψk (y)
k
≤ C(1 + M)rn−α d(x, y)
α
which finishes the proof of (2.6).
As a summary of this section let us state the results in a theorem.
Theorem 2.4. Let (X, d) be a quasi-metric space with dimA X < ∞ and β ≥ 1 such
that d ' ρ β for some distance ρ on X. Then, associated to the sequence { Bi } given
2.7 Whitney extension of continuous functions. 75
in Theorem 2.8 we have a corresponding sequence { φi } of real functions defined on

X such that, with α = β −1 we have
1. B̃i ⊂ supp ϕi ⊂ B(xi , 6Kri ) ⊂ B∗i ;

2. 0 ≤ ϕi ≤ 1:
3. ∑ ϕi = XΩ ;
i
4. ϕi ≥ M −1 on Bi ;
5. kϕi kα ≤ CR−α
i .
2.7 Whitney extension of continuous functions.
The classical result due to Tietze, which in the context of Hausdorff topological
spaces gives a characterization of the separation property called normality, is ac-
tually an extension theorem preserving continuity. In the setting of general metric
and quasi-metric spaces Lipschitz-Hölder is the natural maximal regularity of real
valued functions. A classical result due to McShane, based essentially on the lat-
tice properties of the balls in the spaces of Lipschitz functions, gives an elementary
special extension theorem. (See [?]).
In this section and in the next we shall introduce a more universal extension
method based on the previous construction of Whitney type partitions of unity. The
original paper by Whitney is []. Chapter five of E. Stein [?] is the best source to learn
the method in the euclidian case. Whitney’s method is a basic tool for the extension
of general Besov spaces or even of doubling measures.
Assume that (X, d) is a bounded quasi-metric space with triangular constant K

satisfying the WHD. Let F be a nonempty closed proper subset of X. Then (F, d) is
itself a quasi-metric space satisfying the WHP. Let f : F → R be a given continuous
function. Set Ω = X\F and W a Whitney family for Ω given in Theorem 2.8. Let
{φn } be the partition of unity induced by W on Ω constructed in Section 2.6.
Let us now define the Whitney extension operator E that given a function defined
on F produces a function E f defined on the whole space X by
if x ∈ F;
(
f (x)
E f (x) = f
∑ k k (y )φ (x) if x ∈ Ω.
k
where { yk } is any sequence of points in F given 2.8.5. Notice that for each x ∈ Ω
the sum is finite and that different choices of the sequence { yk } ⊂ F satisfying 2.8.5
will generally produce different extension operators. We shall assume that once F
and W are fixed, we also fix the sequence { yk }, but our estimates will actually be
uniform in these variables.
Theorem 2.5. Let (X, d) be a bounded quasi-metric space satisfying the WHP. Let
F be a nonempty closed proper subset of X and let F be a continuous real func-
tion defined on F, then, the function E f defined as before is continuous on X and
Lipschitz α on every subset A of X with d(A, F) > 0.
In order to prove the theorem, we are going to use the extra property of the
Whitney covering that is contained in the next lemma.
Lemma 2.11. Under the conditions and notation of Theorem 2.8 with F c = Ω , there
exists a geometric constant C such that for every i and for any choice of yi satisfying
2.8.5, the inequality
d(y, yi ) ≤ Cd(y, x),
holds for every x ∈ B∗i and every y ∈ F.
Proof. For x ∈ B∗i we have that d(x, y) ≤ K(d(x, xi ) + d(xi , yi )) < a ri for some pos-
itive constant a. We also have from 2.8.4 that
d(B∗i , F) = d(B∗i ,CΩ ) ' ri .
Hence, for x ∈ B∗i and y ∈ F, applying the triangular inequality and the above esti-
mates we get the desired result
d(y, yi ) ≤ K(d(y, x) + d (yi , x))

< K(d(y, x) + bd (B∗i , F))
< Cd(y, x).
Proof of the Theorem 2.10. Let be
if x ∈ F;
(
f (x)
E f (x) = f
∑ k k (y )φ (x) if x ∈ Ω.
k
2.8 Whitney extension of Lipschitz functions. 77
Since on every subset A of X with d(A, F) > 0, only a finite number of terms of
the series defining E f are non-zero, from the results in Section 2.6 we have that E f
is Lipschitz α on A. Hence E f is continuous at every point belonging to X\F. Let
us prove that E f is continuous at each point in F. Let us fix a point y in F. Take
x∈/ F. Using the fact that ∑ φk ≡ 1 on Ω we can write
k
E f (y) − E f (x) = f (y) − ∑ f (yk )φk (x)

k

= ∑ f (y) − f (yk ) φk (x).
k
So that
E f (y) − E f (x) ≤ ∑ f (y) − f (yk ) φk (x).

k
Let ε be a given positive number. Since f is continuous at y. there exists

a positive δ such that | f (y) − f (z)| < ε provided that d(y, z) < δ , y ∈ F and
z ∈ F. Take η = δ /C with C ≥ 1 the constant in Lemma 2.11. Hence, for x ∈
B(y, η) we have that |E f (y) − E f (x)| ≤ Mε if x ∈/ F, with M the constant in
2.8.
2.8 Whitney extension of Lipschitz functions.
Let (X, d) be a quasi-metric space and let E be a subset of X. We shall say that a
bounded real function f defined on E belongs to the space Lip (γ, E) with γ > 0 if
| f (x) − f (y)|
| f |γ,E = sup
x,y∈E d(x, y)α
x6=y
is finite. The set Lip (γ, E) becomes a Banach space with norm defined by k f kγ,E =
k f k∞ + | f |γ,E , where k f k∞ = supx∈E | f (x)|. Let us state the main result of this
section.
Theorem 2.6. Let (X, d) be a bounded quasi-metric space of order α satisfying the
weak homogeneity property. Let F be a closed proper subset of X. Then the Whitney
extension operator E is a linear and bounded operator from Lip (γ, F) to Lip (γ, X)
for every 0 ≤ γ ≤ α. Moreover, the operator norm of E is bounded by a geometric

constant.
Proof. Let us start by observing that, since ∑ φk (x) = 1 for x ∈/ F, we have that
k
k E f k∞ ≤ k f k∞ . Let us estimate | E f |γ,X . We must get an upper bound of | E f (x)−
E f (y) | in terms of d(x, y)γ for every choice of x and y in X and 0 ≤ γ ≤ α. Of course
the case x ∈ F and y ∈ F is trivial since E f = f on F and | E f (x) − E f (y) | ≤
| f |γ,F d(x, y)γ . Let us now consider the case y ∈ F, x ∈ / F. Since ∑ φk (x) = 1, we
k
have, applying Lemma 2.11, that
| E f (y) − E f (x) | ≤ ∑ | f (y) − f (yk ) |φk (x)

k
≤ | f |γ,F ∑ d(y, yk )γ φk (x)
k
≤ C | f |γ,F ∑ d(y, x)γ φk (x)
k
= C | f |γ,F d(y, x)γ .
We have to consider the case x ∈

/ F and y ∈
/ F. Two essentially different situations
are still possible
(a) C d(x, y) ≥ min{ d(x), d(y)},

(b) C d(x, y) < min{ d(x), d(y)},
for some constant C to be specified.
(a) Assume that d(x) ≤ d(y), hence Cd(x, y) ≥ d(x). Then there exists z ∈ F such
that
(C + 1) d(x, y) ≥ d(x, z).
Notice also that
d(y, z) ≤ K [ d(y, x) + d(x, z)] ≤ C̃d(x, y).
With these estimates we can use the previous case to get

2.8 Whitney extension of Lipschitz functions. 79
E f (x) − E f (y) ≤ E f (x) − E f (z) + E f (z) − E f (y)

≤C f γ,F
[ d(x, z)γ + d(z, y)γ ]
≤ C̄ f γ,F
d(x, y)γ
(b) Let us write K(x) to denote the set { k : φk (x) 6= 0}. We know that there is a
constant M such that ]K(x) ≤ M. Since ∑ φk (y) = ∑ φk (x) = 1, we have that
k k

∑ φk (x) − φk (y) = ∑ φk (x) − φk (y) = 0.
k∈K(x)∪K(y) k
Then, for fixed l ∈ K(x) ∪ K(y), we can write
E f (x) − E f (y) = ∑ f (yk )φk (x) − ∑ f (yk )φk (y)

k k

= ∑ f (yk ) φk (x) − φk (y)
k∈K(x)∪K(y)

= ∑ f (yk ) − f (yl ) φk (x) − φk (y)
k∈K(x)∪K(y)
Hence from (2.6) and the Lipschitz character of f on F we have the estimates
!α
d(x, y)
|E f (x) − E f (y)| ≤ C | f |γ,F ∑ d(yk , yl )γ
.
k∈K(x)∪K(y)
rk
Now, for an appropriate choice of C in (b) the properties of the Whitney covering
imply that d(yk , yl ) ≤ Crl ' rk for every k in K(x) ∪ K(y), which has no more than
2M elements. Then, since 0 ≤ γ ≤ α and d(x, y) is bounded by a constant times rl ,
we get
!α
d(x, y)
| E f (x) − E f (y) | ≤ C | f
γ
|γ,F rl
rl
≤ C̄ | f |γ,F d(x, y)γ ,
with a purely geometric constant C̃.
2.9 Lipschitz test functions for a fractional order theory of

distributions on quasi-metric spaces
We have already defined the Banach spaces Lip (γ) = Lip (γ, X) on a quasi-metric
space (X, d) as the set of real functions f defined on X for which the norm k f kγ =
k f k∞ + | f |γ is finite. We know that Lip (γ) is non-trivial if 0 < γ ≤ α, where α =
β −1 for any β > 0 for which d ∼ ρ β , with ρ a distance on X. Once ρ and β are
fixed, α is the highest regularity order for functions defined on X.
Since we shall need a theory of Schwartz distributions in the abstract setting, we

start by considering the space of test functions. As in the classical case the two basic
ingredients for a test function are the boundedness of its support and its maximal
degree of regularity. This section is devoted to construct an inductive limit topology
on the space of Lip (γ)-functions (0 < γ < α) with bounded support. The basic
reference for the classical case is the book “Functional Analysis” by W. Rudin [?].
The theory of distributions on spaces of homogeneous type is introduced in [?].
Let (X, d) be a quasi-metric space. Take α = β −1 with β > 0 such that d ∼ ρ β

[ {Fn : n ∈ N} be a given sequence of closed
for some distance ρ defined on X. Let
subsets of X such that Fn+1 ⊇ Fn and Fn = X. For a given positive integer n define
n
Enα as the vector space of all real functions ϕ defined on X with support contained
in Fn which are of class Lipschitz(γ) for every γ ∈ (0, α). The set Enα is the space
of all test functions which are localized in Fn . A special case of this situation is
Fn = B(x0 , n) for every n ∈ N and some x0 ∈ X fixed.
A Frèchet structure on Enα is given by the distance
∞
δn (ϕ, ψ) = ∑ 2−i qi (ϕ − ψ)(1 + qi (ϕ − ψ))−1 ,
i=0
where q0 ( f ) = k f k∞ and qi ( f ) = | f |γi for some sequence γi increasing to α. With

this metric topology on Enα , a sequence {ϕk } converges to ϕ if and only if kϕk −
ϕk∞ → 0 as k → ∞ for every 0 < β < α. Moreover, the topology τn induced on Enα
by δn does not depend on the particular choice of the sequence {γi } approaching α.
Let us now consider the vector space E α of all real functions with support
contained in some Fn which are of class Lipschitz(β ) for every β < α. Hence
2.10 Problems and Comments. 81
[
Eα = Enα . Since for every n ∈ N we have that Enα ⊂ En+1
α and τn+1 restricted
n
Enα coincide with τn , we give to E α the strict inductive limit topology: a subset A of
E α belongs to τ if and only if for every n ∈ N the set A ∩ Enα belongs to τn .
The space (E α , τ) is a sequentially complete topological vector space with some

special and important extra properties:
(a) τ Enα = τn for every n ∈ N

(b) a sequence {ϕk } ⊂ E α converges to a function ϕ ∈ E α in the topology τ if and
only if there exists a positive integer n such that {ϕk } ⊂ Enα and kϕk − ϕk∞ +
|ϕk − ϕ|γ → 0 as k → ∞ for every 0 < γ < α,
(c) a subset A of E α is bounded if and only if there exists n ∈ N such that A ⊂
Enα , k ϕ k∞ ≤ M0 and | ϕ |γ ≤ Mγ for some constants M0 and Mγ , every ϕ ∈ A and
every γ ∈ (0, α).
Let us denote by D 0 the dual space of (E α , τ) i.e. T ∈ D 0 if T is a real or complex

linear and continuous functional defined on E α . As in the classical case, assuming
linearity, continuity is equivalent to sequential continuity at 0, to boundedness and
to continuity of every restriction τ Enα . The elements of D 0 are called distributions.
Of course evaluation at a fixed point x0 ∈ X is a basic example of a distribution

in D 0 = D 0 (X, d). In fact hτ, ϕi = ϕ(x0 ) is a linear and continuous operation on
(E α , τ). In this case we write T = δx0 using the standard notation for the Dirac unit
mass at x0 ∈ X.
2.10 Problems and Comments.
2.13. Give an example of a quasi-metric space for which every bounded set is totally
bounded but dimA X = +∞.
2.14. Give an example of a quasi-metric space with Assouad dimension strictly

larger than its Hausdorff dimension.
2.15. Prove the remark following Lemma 2.1.
2.16. Prove Lemma 2.7.
2.17. Consider the dyadic type partitions given in Theorem 1.38 of Chapter 1 under
the additional assumption that dimA X < +∞. Which extra properties of the dyadic
family can be obtained?
2.18. Let (X, d) be a quasi-metric space such that dimA X < +∞ and take 0 < ε <
δ < +∞. Construct Nδ δ -net and Nε ε-net in X such that Nδ ⊆ Nε .
2.19. Prove that if (X, d) satisfies the WHP then (X, δ β ) also satisfies the WHP for
every δ ∼ d and for every β > 0.
2.20
Chapter 3
Spaces of homogeneous type
In this chapter we shall introduce the basic definition and properties of space of
homogeneous type. There are several different approaches to the definition of space
of homogeneous type in the literature and they are generally not equivalent in a
strict sense. But the analytical problems solved in the structure have some kind of
robustness in the sense that the results are somehow independent of the approach.
The basic situation for this chapter occurs frequently in mathematics: two struc-
tures of different nature co-exist on a set and some a priori elementary interplay
between them produce amazing connections and properties, giving rise to a rich
setting. In our case the two structures are metric (quasi-metric) and measure (outer-
measure) and the basic interplay is given by what is known as the doubling property.
In Chapter 4 we shall introduce and discuss several useful examples of spaces of

homogeneous type.
3.1 Spaces of homogeneous type
Let (X, d) be a quasi-metric space. Let us denote, as we did before, P(X) for the
set of all subsets of X. Assume that on P(X) we have defined a non-negative, non-
increasing and finitely subadditive set function µ ∗ such that there exists a constant
A > 0 for which the inequality
µ ∗ (B(x, 2r)) ≤ Aµ ∗ (B(x, r)) (3.1)
83
84 3 Spaces of homogeneous type
holds for every x ∈ X and every r > 0. Inequality (3.1) is a doubling condition for
µ ∗ with respect to d in the sense that we are doubling the d-balls and their “volume”
in the sense of µ ∗ does not grow too much.
Notice that if there is an x in X and a positive number r such that µ ∗ (B(x, r)) = 0,
then for every bounded set E we have that µ ∗ = 0. In fact, given any subset E of X,
[k0
for some k0 ∈ N we can write E ⊂ B x, 2k r and
k=1
k0
!

∗ ∗ ∗
[
k
0 ≤ µ (E) ≤ µ (X) = µ B x, 2 r
k=1
k0
≤ ∑ µ∗ B x, 2k r
k=1
k0
≤ ∑ Ak µ ∗ (B(x, r))
k=1
= 0.
Observe also that if for some ball B(x, r) we have that µ ∗ (B(x, r)) = +∞, then
every ball and so every open set has infinite outer measure. To prove this fact we
only have to observe that given any other point y in X and any positive number s we
have that
B(x, r) ⊂ B(y, K(r + d(x, y)))

= B y, Ks (r + d(x, y)) s

⊂ B y, 2 j s ,

for j ≥ log2 Ks (r + d(x, y)) . Hence µ ∗ B(y, 2 j s) = +∞, then, from the doubling

property, A j µ ∗ (B(y, s)) = +∞.
These remarks show that aside from trivial situations (3.1) can be written as
0 < µ ∗ B(x, 2r) ≤ Aµ ∗ B(x, r) < ∞.

Let us now show that some mild additivity on µ ∗ is enough in order to have that
dimA X < ∞.
3.1 Spaces of homogeneous type 85
Lemma 3.1. Let (X, d) be a quasi-metric space. Let µ ∗ be a set functions as above
satisfying (3.1). Assume also that µ ∗ is a metric set function in the sense that if E1
and E2 are subsets for which d (E1 , E2 ) > 0, then µ ∗ (E1 ∪ E2 ) = µ ∗ (E1 ) + µ ∗ (E2 ).
Then (X, d) satisfies the weak homogeneity property or, in other words, dimA X < ∞.
Proof. Let ρ be a distance on X such that d ' ρ β for some β > 0. Let D be an
r
2 -disperse subset of Bρ (x0 , r) with respect to the distance ρ. In other words D ⊂
Bρ (x0 , r) and ρ(x, y) ≥ 2r for every x and y ∈ D with x 6= y. Then
r
ρ Bρ x, 8r , Bρ y, 8r ≥

and hence (3.2)
4
d Bρ x, 8r , Bρ y, 8r > 0

for every x and y ∈ D;
Bρ x, 8r ⊂ Bρ x0 , 89 r

for every x ∈ D; (3.3)
Bρ x0 , 98 r ⊂ Bρ x, 17

8 r for every x ∈ D; (3.4)
∗ 9 4 ∗ r

µ Bρ x0 , 8 r ≤ A µ Bρ x, 8 for every x ∈ D. (3.5)
Let us now take F a finite subset of D. Then, from (3.5), (3.2), the fact that µ ∗ is
a metric set function with respect to d and (3.3), we have that
µ ∗ Bρ x0 , 17

−4 ∗ 8 r
Bρ x0 , 17

A µ 8 r ](F) = ∑ 4
x∈F A
∗
≤ ∑ µ Bρ x, 8r

x∈F
!
∗
[
r

=µ Bρ x, 8
x∈F
≤ µ ∗ Bρ x0 , 98 r

.
These inequalities show that ](F) ≤ A4 for every finite subset F of D, hence
](D) ≤ A4 and (X, ρ) has finite Assouad dimension. As a consequence (X, ρ β ) and
of course (X, d) have finite Assouad dimension.
An example of the current situation is the following: Let X = Q be the set of all
rational numbers. Let d(x, y) = |x − y| be the restriction to Q of the usual euclidian
distance. Given a subset E of X let us denote by Ē its closure as a subset of R, the
set of all real numbers. Then, the set function µ ∗ (E) = λ (Ē), where λ is Lebesgue
measure, satisfies the doubling property and is a metric set function. Hence, from
Lemma 3.1, we have that dimA X < ∞.
However there is no way to obtain a non-trivial doubling measure on (X, d). In

fact, if µ is a Borel measure on (X, d) which is non-trivial, then there must be a

point x0 ∈ X such that µ {x0 } > 0. But if µ is doubling we have that µ x0 +

1
2n+1
, x0 + 21n , n ∈ N, is bounded below by a positive constant, since a fixed dilation

1
of x0 + 2n+1 , x0 + 21n contains the point x0 . Hence µ(x0 − 1, x0 + 1) = +∞ and,
again, we are in a trivial case.
The above remark shows both that the weak homogeneity property, i.e. dimA X <
∞, is not a sufficient condition to have a doubling measure on (X, d) and that having
a doubling metric outer measure is sufficient for the weak homogeneity.
Let us state the basic definition of space of homogeneous type as we shall use it
throughout.
Let X be a set and let F be a σ - algebra of parts of X. If µ is a measure on F and

d is a quasi-distance on X, we shall say that µ is a doubling measure with respect
to d or, briefly, that µ is doubling w.r.t.d, if the family {Bd (x, r) : x ∈ X, r > 0} of
all d-balls in X is contained in F and there exists a constant A > 0 such that the
inequalities

0 < µ Bd (x, 2r) ≤ A µ Bd (x, r) < ∞,
hold for every x ∈ X and every r > 0. Since given a quasi-distance d on X we have
a topology τd on X, the Borel σ - algebra with respect to d can be defined as the
intersection of all σ - algebras containing the topology τd and denoted by Bd or B
when there is no possible confusion.
Let us observe that if d and δ are two quasi-metrics on the set X, such that
d ' δ β for some β > 0, then since τd = τδ we have that Bd = Bδ and so family
of Borel sets is invariant under these types of changes of quasi-distances. However,
this invariance is not true for the doubling property as stated, since it can generally
happen that the d-balls belong to F but the δ -balls do not even when d ' δ . Recall
the examples and exercises of Chapter 1.
In order to give a concise statement of the next result we are going to use some
extra notation. Given a quasi-distance d on X we shall write R(d) to denote the set
{ ρ : ρ is a distance on X and ρ γ ' d for some positive γ} and
∆ (d) = { δ = ρ β : ρ ∈ R(d) and β > 0}
Lemma 3.2. Let (X, d) be a quasi-metric space. Let F be a σ - algebra of subsets

of X and let µ be a non-negative measure defined on F . Then
1. If ρ ∈ R(d) and µ is doubling w.r.t.ρ. then B := Bρ = Bd ⊂ F ;

2. the measure µ is doubling w.r.t.δ . for some δ ∈ ∆ (d) if and only if µ is doubling
w.r.t. δ . for every δ ∈ ∆ (d);
3. if µ is doubling w.r.t.d., then µ is doubling w.r.t.δ . for every δ ∈ ∆ (d);
4. if µ is doubling w.r.t.δ . for some δ ∈ ∆ (d) and the d-balls belong to F , then µ
is doubling w.r.t.d.
Proof. Let us start by proving 3.2.1. Since d ' ρ β for some γ > 0 we have that
Bρ = Bd and we denote them by B. In order to show that B ⊂ F we must prove
that the topology generated by d or ρ is contained in F . We know, from the doubling
property of µ w.r.t.ρ., that the ρ-balls are elements of F . Hence it is enough to
show that every open set can be written as a countable union of ρ-balls. Which is
in turn true if the space is separable, because the ρ-balls with rational radii and
center belonging to a countable dense subset of X, will perform a countable basis
for the topology. The separability is, as we have seen, a consequence of the weak
homogeneity property. We get that dimA X < ∞ from Lemma 3.1 applied to (X, ρ)
and the outer measure µ ∗ (E) = inf { µ(A) : A ⊃ E and A ∈ F } which is finitely
additive acting on separate ρ-balls, which is the only additivity actually used in the
proof of Lemma 3.1.
To prove 3.2.2 assume that µ is doubling w.r.t.δ . for δ = ρ γ , with ρ ∈ R(d) and
γ > 0. Take δ̄ in ∆ (d), δ̄ = ρ̄ γ̄ . Notice that
Bδ̄ (x, r) = { y ∈ X : ρ̄(x, y) < r1/γ̄ = Bρ̄ (x, r1/γ̄ )
is an open set in X, since ρ̄ is a distance generating the same topology as d. Hence

Bδ̄ (x, r) ∈ F for every x ∈ X and every r > 0. We only have to show that

0 < µ Bδ̄ (x, 2r) ≤ Ā µ Bδ̄ (x, r) < ∞,
for every x ∈ X and r > 0. Since δ̄ = ρ̄ γ̄ and δ = ρ γ with ρ β ' d and ρ̄ β̄ ' d we
necessarily have that
β̄ γ̄ γ̄
β γ̄
δ̄ = ρ̄ γ̄ = ρ̄ β̄ ' d β̄ ' ρ β̄
γ β γ̄ β γ̄
= ρ β̄ γ = δ β̄ γ
= δη.
In other words, there exist positive constants C1 ,C2 and η such that
C1 δ̄ ≤ δ n ≤ C2 δ̄ .
Thus, since

Bδ x, (2C1 r)1/η ⊂ Bδ̄ (x, 2r) ⊂ Bδ x, (2C2 r)1/η
and

Bδ x, (C1 r)1/η ⊂ Bδ̄ (x, r) ⊂ Bδ x, (C2 r)1/η
1/η
2C2
we have, with 2m ≥ C1 , that

0 < µ Bδ x, (2C1 r)1/η

≤ µ Bδ̄ (x, 2r)

1/η
≤ µ Bδ x, (2C2 r)1/η = µ Bδ x, 2C C1
2
(C1 r)1/η

≤ Am µ Bδ x, (C1 r)1/η

≤ Am µ (Bδ (x, r)) ≤ Am µ Bδ x, (C2 r)1/η < ∞,
which finishes the proof of 3.2.2. Statements 3.2.3 and 3.2.4 follow easily.
We shall say that (X, d, µ) is a space of homogeneous type if (X, d) is a quasi-

metric space and µ is a measure defined on a σ - algebra F of subsets of X such
that there exists ρ ∈ R(d) with µ doubling w.r.t.ρ. From Lemma 3.2 we deduce that
if (X, d, µ) is a space of homogeneous type, then the Borel sets belong to F and,
as a consequence, that continuous real or complex functions defined on (X, d) are
measurable functions. Of course also semi-continuous functions are measurable.
3.2 Boundedness and finite measure of a space of homogeneous type. 89
3.2 Boundedness and finite measure of a space of homogeneous

type.
Let us start by observing that if (X, d, µ) is a space of homogeneous type which is

bounded, i.e. X is a d-ball, then X is also a ρ-ball for ρ ∈ R(d), hence µ(X) < ∞.
The main result of this section is the converse: µ(X) < ∞ implies the bounded-
ness of X.
Theorem 3.1. Let (X, d, µ) be a space of homogeneous type. Then,

µ(X) < ∞ if and only if (X, d) is bounded.
[
Proof. Let us assume that µ(X) < ∞ but that X is unbounded. Since X = Bρ (x0 , k)
k∈N
for some distance ρ in R(d) and every x0 ∈ X, we have that

µ X\Bρ (x0 , k) = µ(X) − µ Bρ (x0 , k) (3.6)
tends to zero when k tends to infinity. Since we are assuming that X is unbounded,
for each k ∈ N there exists xk , a point in X, that does not belong to Bρ (x0 , 2k).
Hence, for each k ∈ N we have that

ρ(xk , x0 )
Bρ (x0 , k) ∩ Bρ xk , = 0,
/ and (3.7)
2
Bρ (x0 , k) ⊂ Bρ (xk , 2ρ(x0 , xk )) . (3.8)
Then, from (3.8), the doubling property for µ w.r.t.ρ. and (3.7), we have that

0 < µ(X) = lim µ Bρ (x0 , k)
k→∞

≤ lim sup µ Bρ (xk , 2ρ(x0 , xk ))
k→∞

ρ(xk , x0 )
≤ A2 lim sup µ Bρ xk ,
k→∞ 2
2

≤ A lim sup µ X\ Bρ (x0 , k) ,
k→∞
a contradiction since the right hand side is zero from (3.6).
The basic example of space of homogeneous type is indeed (Rn , |x − y|, λ ) where
|x − y| is the euclidian distance and λ is Lebesgue measure. In this example the
boundedness of the measures of a family of balls is equivalent to the boundedness
of the radii of the balls of the family. We want to point out that this is not the case
in a general space of homogeneous type. Taking X = R, d(x, y) = |x − y|, the σ -
1
algebra of Lebesgue measurable subsets of R1 and dµ = | x |− 2 dx, we have that
(X, d, µ) is a space of homogeneous type. Notice first that the d-balls are the open
intervals of real numbers, so that the balls belong to F . Since the function | x |−1/2
is locally integrable we see that the measure of each ball is finite and, certainly,
positive. Let us prove √ that µ is a doubling measure. We are going to show that,
actually, µ (B(x, r)) ' r in the sense that there exist positive constants C1 and C2
such that the inequalities
√ √
C1 r ≤ µ (B(x, r)) ≤ C2 r,
hold for every x ∈ R and every r > 0. From these inequalities the doubling property
follows at once. Since the weight function | x |−1/2 is even we only have to consider
the case x ≥ 0. Since for x = 0 we have that
Z r
√
µ (B(0, r)) = t −1/2 dt = 2 r
0
we shall assume that x > 0. Let us consider three cases for the relative position of x
with respect to the size of the ball given by r:
(a) 0 < r ≤ x ≤ 4r;

(b) 0 < x < r;
(c) x > 4r.
Case (a).
Z x+r √ √
t −1/2 dt = 2

µ (B(x, r)) = x+r− x−r
x−r
4r √
= √ √ ' r.
x+r+ x−r
3.3 Atomic singletons are countable and isolated. 91
Case (b).
√ √ Z x+r
2 r ≤ 2 x+r = t −1/2 dt
0
≤ µ (B(x, r))
Z x+r √ √ √
≤2 t −1/2 dt = 4 x + r ≤ 2 4 r.
0
Case (c).
√ 2r 2r 2 √
2 r≤ √ ≤ µ (B(x, r)) ≤ √ ≤ √ r.
x+r x−r 3
So that (X, d, µ) is a space homogeneous type. Notice now that the family of
intervals
n2 , (n + 1)2 : n ∈ N

is the family of balls
Bn := B n2 + n + 12 , n + 12 : n ∈ N

with unbounded radii. But µ(Bn ) = 2 for every n.
3.3 Atomic singletons are countable and isolated.
Let (X, d, µ) be a space of homogeneous type. We shall say that a point x0 ∈ X is an

atom if µ ({ x0 }) > 0. Notice that, being closed, the set {x0 } belongs to F .
Lemma 3.3. Let (X, d, µ) be a space of homogeneous type. Then the set A of all
atoms in X is countable.
Proof. Take x0 a fixed pointin X and ρ ∈ R(d). For n and m two positive integer
numbers let us write An,m = x ∈ Bρ (x0 , m) : µ ({ x}) > n1 . Since µ Bρ (x0 , m) <

∞, we have that each set An,m is finite. But A = An,m .

[
m,n
As usual, we shall say that a point x0 is isolated if there exists r > 0 such that
B(x0 , r) = { x0 } which is equivalent to Bρ (x0 , s) = { x0 } for some positive s, when
ρ ∈ R(d). Since ρ-balls have positive measure, then isolated points are atoms. Now
we give a result of Macı́as and Segovia [?] that shows the converse. We are going
to base our proof in the construction of a Whitney type family of balls that allows
some extensions to “higher dimensional” situations.
Theorem 3.2. A point x0 in a space of homogeneous type is an atom if and only if

x0 is isolated.
Proof. Let us assume that µ ({ x0 }) > 0. Take ρ ∈ R(d). If Bρ (x0 , 1) = { x0 } there is

nothing to prove, since Bd (x0 , c) = { x0 } for some positive constant. Assume, then,
that Bρ (x0 , 1) contains points x other than x0 . The set Ω = Bρ (x0 , 1) − { x0 } is a
bounded and open subset of X. Let r : Ω → R+ be the function defined by r(x) =
ρ(x, x0 )/6. Since Ω is bounded we can apply the Wiener type covering lemma in
order to get a finite or countable sequence { xi : i ∈ I} of points in Ω such that
the balls Bi = Bρ (xi , r(xi )) , i ∈ I are pairwise disjoint, (3.9)

and

[ ρ(xi , x0 )
we have that Ω ⊂ B̃i , with B̃i = Bρ xi , . (3.10)
i∈I
2
Since the ball Bρ (xi , 8r(xi )) = Bρ xi , 43 ρ(xi , x0 ) contains the point x0 for every

i ∈ I and, on the other hand, each Bi is a subset of Bρ x0 , 76 , for every finite subset

J of I we have that

](I)µ ({ x0 }) ≤ ∑µ Bρ (xi , 8r(xi ))
i∈J
≤ A3 ∑ µ(Bi )
i∈J
!
3
=A µ ∑ Bi
i∈J
≤ A3 µ Bρ (x0 , 7/6) .

3.3 Atomic singletons are countable and isolated. 93
Since the right hand side does not depend on J, we see that I is finite. Actually

3µ Bρ (x0 , 7/6)
](I) ≤ A .
µ({ x0 })
n o
So that, the set ri = ρ(x6i ,x0 ) : i ∈ I has a minimum which we shall call r > 0. Let
us prove that Bρ (x0 , 3r) = { x0 }. In fact, if x ∈ Ω then, for some i ∈ I we have that
x ∈ B̃i , which is equivalent to ρ(x, xi ) < ρ(x2i ,x0 ) . Hence
ρ(xi , x0 )
ρ(xi , x0 ) ≤ ρ(xi , x) + ρ(x, x0 ) < + ρ(x, x0 )
2
and
ρ(xi , x0 ) ρ(xi , x0 )
ρ(x, x0 ) > =3 = 3ri ≥ 3r > 0.
2 6
So that, for every point x ∈ Ω we have that ρ(x, x0 ) > 3r, which proves that
Bρ (x0 , 3r) = { x0 } as desired.
Notice that the boundary of X\ { x0 } can only be the empty set or the set { x0 }.
The non-trivial part of Theorem 3.5 can also be written as
∂ (X\ { x0 }) = { x0 } then µ ({ x0 }) = 0,
Let us explore with an example this form of Theorem 3.2 for the case of “higher
dimensional” boundaries. Observe that if X is the union of the upper half plane
H = R2+ of R2 and an horizontal line
L = { (x, y) ∈ R2 : y = c}.
with c < 0, the usual distance |x − y| and the Borel measure
µ(E) = area (E ∩ H) + length (E ∩ L),
then (X, |x − y|, µ) is a space of homogeneous type. The doubling constant A for µ
grows to infinity when c → 0− . Moreover, when c = 0 the measure µ is not doubling:
2
take r > 0, then µ (B ((0, r), 2r)) = πr2 , on the other hand µ (B ((0, r), 2r)) ≥ πr2 +r,
which makes the doubling impossible for r → 0+ .
The fact that (X, d, µ) is a space of homogeneous type when c < 0 can be directly
checked but it can also be obtained as a consequence of the results of a forthcoming
section.
The example above for the case c = 0 suggests that, being a one dimensional
set, the boundary of R2+ has to have zero measure if the measure is expected to
be doubling. Let us also recall that even in R1 with Lebesgue measure there are
bounded and open subsets whose boundaries have large measures. But in these cases
the open sets used have quite irregular boundaries.
Even when we have no differential structure on X, via Whitney type coverings of

an open set Ω we are able to introduce some basic idea of regularity of its boundary
∂ Ω , which in turn will be a sufficient condition for µ (∂ Ω ) = 0.
Let (X, d, µ) be a space of homogeneous type and let ρ ∈ R(d). We say that
a non-empty, open and bounded subset Ω of X is regular if given W = { Bi } a
Whitney family of pairwise disjoint balls for Ω , there exists a constant α such that
the dilations B̂i of Bi by C satisfy
[
∂Ω ⊂ B̂i , for every n.
i=n
In the problems at the end of the chapter we shall discuss how this regularity
condition is related to standard regularity in the euclidian context.
Lemma 3.4. Let (X, d; µ) be a space of homogeneous type and let Ω be a non-empty
open and bounded subset of X. Assume that Ω satisfies the regularity condition.
Then µ (∂ Ω ) = 0.
Proof. Since Ω is bounded, there exist x0 ∈ X and R > 0 such that Bi ⊂ Bρ (x0 , R)
for every i. Since the balls Bi are pairwise disjoint, we certainly have that
!
[
∑ µ(Bi ) = µ Bi ≤ µ Bρ (x0 , R) < ∞.
i=1 i=1
Hence ∑ µ(Bi ) → 0 as n → ∞. On the other hand, from the regularity of Ω , we

i=n
have that
3.4 Regularity of the measure. Approximation of indicators by continuous functions. 95
!
[
µ (∂ Ω ) ≤ µ B̂i
i=n

≤ ∑µ B̂i
i=n
≤ C (A, α) ∑ µ (Bi ) ,
i=n
which tends to zero if n tends to ∞.
3.4 Regularity of the measure. Approximation of indicators by

continuous functions.
Let (X, d, µ) be a space of homogeneous type. Let F be the σ - algebra of µ- mea-

surable sets. If B denotes the σ - algebra of all Borel sets, we have that B ⊆ F .
Let us write τ to denote the topology on X generated by d. Assume that ρ is a fixed
distance in R(d).
Given a set A ∈ F we shall say that A is regular or µ- regular if for every positive
number ε there exist G ⊂ τ and F closed (F c ∈ τ) such that F ⊂ A ⊂ G and µ(G −
F) < ε.
From the analytical point of view, the regularity of a large class of subsets of F
is an important fact that will be reflected as the density of continuous functions in
the usual Banach spaces of functions defined on X.
In this section we shall prove that bounded Borel subsets of a space of homoge-
neous type are regular.
Theorem 3.3. Let (X, d, µ) be a space of homogeneous type, then every bounded
Borel subset of X is regular.
Compactness of the inner approximation F can be obtained with the extra hy-
pothesis of completeness.
Theorem 3.4. Let (X, d, µ) be a complete space of homogeneous type. Then for
every bounded Borel subset E of X and every ε > 0 there exist a compact set K and
an open set G such that G ⊃ E ⊃ K and µ(G − K) < ε.
Proof. Since (X, d, µ) is a space of homogeneous type, then (X, d) has the weak
homogeneity property. Since (X, d) is complete we know that the space has the
Heine - Borel property: every closed and bounded subset of X is compact. But the
set F provided by Theorem 3.7 is closed and bounded.
As a consequence we obtain an approximation result by Lipschitz continuous

functions of the indicator functions of bounded Borel subsets of complete spaces of
homogeneous type.
Corollary 3.1. Let (X, d, µ) be a space of homogeneous type. For every bounded
Borel set E and every ε > 0 there exist a real function f defined on X of class
Lipschitz β with respect to d for some β > 0, a compact set K and an open set G
such that
1. K ⊂ E ⊂ G;
2. µ(G\K) < ε;
3. f ≡ XE on K ∪ Gc ;
4. | f (x) − XE (x)| ≤ 1 for every x ∈ G\K.
Proof. Take ρ ∈ R(d) and β such that ρ 1/β ' d. Since E is a bounded Borel set,
from Theorem 3.8, there exist K a compact set and a bounded open set G such that
3.9.1 and 3.9.2 hold. It is clear that the function
ρ (x, Gc )
f (x) =
ρ (x, Gc ) + ρ(x, K)
satisfies 3.9.3 and 3.9.4. The Lipschitz β character of f follows from the fact that
the product of two bounded Lipschitz functions is a Lipschitz function:
|g(x)h(x) − g(y)h(y)| ≤ |g(x) − g(y)||h(x)|

+ |g(y)||h(x) − h(y)|

≤ |g|β khk∞ + kgk∞ |h|β d(x, y)β
Notice that the function ρ(x, Gc ) + ρ(x, K) is bounded below since, otherwise,
ρ(Gc , K) = 0 and Gc ∩ K 6= 0,
/ which is impossible since K ⊂ G.
3.4 Regularity of the measure. Approximation of indicators by continuous functions. 97
Since every space of homogeneous type is σ - finite, a countable union of mea-

surable sets each with finite measure, Theorem 3.7 reduces to prove the next result.
Lemma 3.5. Let (X, ρ) be a metric space and let µ be a finite Borel measure on X.
Then, every Borel set is µ- regular.
Proof. It suffices to prove that every open set is µ- regular and that the class of all
µ- regular sets is a σ - algebra of subsets of X. Let V be an open set. Then
∞
[
c 1
V = x ∈ X : ρ (x,V ) ≥
n=1
n
[∞
= Fn ,
n=1
is an increasing and countable union of closed sets, i.e. V is a Fσ set. Since Fn ⊂ V

and µ(x) < ∞, we have that
µ(V \ Fn ) = µ(V ) − µ(Fn )
tends to zero as n tends to +∞. This finishes the proof of the fact that open sets are
µ- regular.
Let us write M = { A ⊂ X : A is µ −regular}; and show first that M is an algebra.

It is clear that the whole space X is µ- regular. For A ∈ M and ε > 0 there exist F
closed and G open such that
F ⊂A⊂G and µ(G\ F) < ε,
hence
X\ F ⊃ X\ A ⊃ X\ G
with
µ ((X\ F)\ (X\ G)) = µ(G\ F) < ε.

So that X\ A = Ac ∈ M . Take now two sets A1 and A2 in M and ε > 0. Then

for i= 1,2 we have Fi closed and Gi open with Gi ⊃ Ai ⊃ Fi and µ(Gi − Fi ) < ε2 .
But G = G1 ∪ G2 is open, F = F1 ∪ F2 is closed, G ⊃ A1 ∪ A2 ⊃ F and µ(G\ F) ≤
µ ((G1 \ F1 ) ∪ (G2 \ F2 )) < ε. This means that M is closed under finite unions.
Let us now prove that M is also closed under countable unions. Let { An : n ∈ N}
∞
be a given sequence of sets in M . We have to show that A = An ∈ M . Since we
[
n=1
already know that M is an algebra we may assume, without loosing generality, that
the sets An are pairwise disjoint: just take a new sequence { A0n }, given by A01 = A1
/ n−1
and A0n = An A0i which belong to M , are pairwise disjoint and its union is A.
[
i=1
Let ε be a given positive number. For each n ∈ N pick a closed set Fn and an open
set Gn in such a way that Fn ⊂ An ⊂ Gn and µ(Gn \ Fn ) < 2εn . For a finite n ∈ N, the
N
[ ∞
[
set F = Fn is closed, the set G = Gn is certainly open and F ⊂ A ⊂ G. Let us
n=1 n=1
estimate the measure of G\ F. Since
! !
∞
[ / N
[
G\ F = Gn Fn
n=1 n=1
" ! !# !
N
[ ∞
[ / N
[
= Gn ∪ Gn Fn ⊆
n=1 n=N+1 n=1
!
N
[ ∞
[
⊆ (Gn \ Fn ) ∪ Gn ,
n=1 n=N+1
we get
N ∞
µ(G\ F) ≤ ∑ µ(Gn \ Fn ) + ∑ µ(Gn )
n=1 n=N+1
N ∞
ε ε
≤ ∑ 2n + ∑ µ(Fn ) +
n=1 n=N+1 2n
∞
≤ ε+ ∑ µ(An ).
n=N+1
3.5 Finite upper type and doubling 99
Taking N large enough we have that µ(G\ F) < 2ε, since the series ∑ µ(An ) con-
n
verges to µ(A) ≤ µ(X) < ∞.
3.5 Finite upper type and doubling
Let (X, d, µ) be a space of homogeneous type. Let δ ∈ ∆ (d) be a power of a distance

on X which is equivalent to d. For each x ∈ X consider the function ηx : R+ →
R+ , given by ηx (r) = µ (Bδ (x, r)) . If the d-balls are measurable, we may consider
directly µ (Bd (x, r)) . The next lemma contains some basic facts regarding the family
of functions { ηx : x ∈ X}.
Lemma 3.6. Let (X, d, µ) be a space of homogeneous type and let be { ηx : x ∈ X}

defined as before. Then
1. each ηx is a non-decreasing function;

2. each ηx is left continuous;
3. for each x ∈ X we have that ηx (0+ ) = µ ({ x}) ;
4. there exists a positive constant A such that the inequalities
0 < ηx (2r) ≤ A ηx (r) < ∞

hold for every r ∈ R+ and every x ∈ X.
Proof. Property 3.11.1 follows from the monotonicity

[ of the measure. Property
3.11.2 is consequence of the fact that B(x, r) = B x, r − n1 . Since every ball
n∈N
has finite measure, property 3.11.3 follows from
∞
\
B x, 1n = { x}.

n=1
On the other hand, property 3.11.3 is just a rephrase of the doubling condition.
It is usual (see []) to say that a function η : R+ → R+ is of upper type β , with

0 < β < +∞ if there exists a constant C such that the inequality
η(ts) ≤ Csβ η(t) (3.11)
holds for every t > 0 and every s ≥ 1. The function η is said to be of lower type β if
there exists a constant C such that inequality (3.11) holds for every t > 0 and every
0 < s ≤ 1. It is clear the y η is of upper type β and γ > β then η is of upper type
γ. A family Λ of functions is said to be uniformly of upper type β if (3.11) holds,
with a fixed C, for every η ∈ Λ .
Lemma 3.7. Let (X, d) be a quasi-metric space and ρ ∈ R(d). Let µ be a given
measure on a σ - algebra F of parts of X containing the ρ- balls. Then (X, d, µ)is
a space of homogeneous type if and only if the family Λ = { ηx (r) = µ Bρ (x, r) :
x ∈ X} is uniformly of upper type β for some 0 < β < ∞.
Let us go back to the example of a space with mixed dimensions given in Section
3.3, where
X = { (x1 , x2 ) ∈ R2 : x2 > 0} ∪ { (x1 , c) : x1 ∈ R}

= H ∪L
for c < 0, d ((x1 , x2 ), (y1 , y2 )) = max{ |x1 − y1 |, |x2 − y2 |} and
µ(E) = area (E ∩ H) + length (E ∩ L).
Then for x = (x1 , x2 ) ∈ H we have that
 2
 4r if 0 < r ≤ x2 ,
ηx (r) = 2r(x2 + r) if x2 < r ≤ x2 + |c|,
2r(x2 + r) + 2r if x2 + | c | < r < ∞.

Let us observe that for c < 0 fixed and x ∈ X, the family ηx (r) is uniformly of upper
type β = 2. Let us also notice that this property loses uniformity when c → 0 for
x ∈ H and when c → −∞ for x ∈ L.
3.6 More covering lemmas. 101
3.6 More covering lemmas.
In the classical selection procedure for Wiener type lemmas the boundedness of the
covered set leads to a maximal radii criteria which in some sense is the most efficient
for this case. For the covering lemma considered in this section the assumption on
the covered set is, instead, given in terms of finiteness of the measure. Our procedure
has now two maximality criteria: first we pick balls with maximal measure and then,
among the balls which are close to these balls of maximal measure we pick those
with maximal radii.
The Wiener type covering lemma given in Chapter 2 works only for bounded
sets. In a space of homogeneous type we have sometimes that the measures of the
covering family of balls are known to be bounded. Since the boundedness of the
measures of balls does not imply the boundedness of the radii of the given balls, see
Section (SEE CORRESPONDENT SECTION), an essentially different selection
procedure is required. Let us state and prove the result of this section.
Lemma 3.8. Let (X, d, µ) be a space of homogeneous

[ type. Consider A = { Bα : α ∈ Γ },
a given family of d-balls such that E = Bα is contained in a measurable set with
α∈Γ
exists a sequence { Bi } = { B(xi , ri )} of pairwise disjoint
finite measure. Then, there[
balls in A such that E ⊂ B(xi ,Cri ) for some constant C that only depends on K.
i
Moreover, C can be chosen in such a way that each ball in A is contained in some
of the balls B(xi ,Cri ).
Proof. Let us start by noticing that if Λ is a subset of Γ and Bλ is given with λ ∈ Λ ,

then the family
G = { Bα : α ∈ Λ and B(xα , 2Krα ) ∩ Bλ 6= 0}

/
is nonempty and the set
R = { rα : Bα ∈ G }
is bounded. In fact, if the space X is bounded there is nothing to prove. If, otherwise,
the space is unbounded, then µ(X) = +∞. Assume that sup R = +∞. Then there is
a sequence { r j } ⊂ R with r j > rλ for every j, r j+1 > r j and r j → +∞ as j → +∞.
If X j is the center of one of the balls with radii equal to r j , we have that
B(xλ , r j ) ⊂ B x j , 4K 3 r j ,

since for z ∈ Bλ ∩ B(x j , 2Kr j ) if y ∈ B(xλ , r j ) the inequalities
d(y, x j ) ≤ K [ d(y, xλ ) + K (d(xλ , z) + d(z, x j ))] < 4K 3 r j
hold. Now since for ρ ∈ R(d) and some C > 0 we have that
r
i
Bρ xλ , ⊂ B(xλ , r j )
C
and
B x j , 4K 3 r j ⊂ Bρ x j ,C4K 3 r j .

r r
Hence µ Bρ (xλ , Cj ) ≤ µ Bρ (x j ,C4K 3 r j ) ≤ Ãµ Bρ x j , Cj . If M ∈ F with

µ(M) < +∞ and E ⊂ M we have that
r
j
Bρ x j , ⊂ B(x j , r j ) ⊂ E ⊂ M,
C
r r
then µ Bρ (xλ , Cj ) ≤ µ(M) < ∞ which is impossible since µ Bρ (xλ , cj ) tends to
+∞ for j → +∞.
Let us now give the inductive construction of the sequence { Bi }. Let A1 = A

be the given family of balls and B0,1 = B(x0,1 , r0,1 ) ∈ A1 be such that

2µ Bδ0,1 > sup{ µ Bδ : B ∈ A1 },
where, Bδ is the δ -ball with the same center and radius than B and δ = ρ γ ' d, ρ ∈ R(d).
Define now
A˜1 = { Bα = B(xα , rα ) ∈ A1 : B(xα , 2Krα ) ∩ B0,1 6= 0}.

/
From the remark above, the set R1 = { rα : Bα ∈ A˜1 } is bounded, hence we may
choose a ball B1 = B(x1 , r1 ) ∈ A˜1 such that 2r1 > sup R1 . Notice that if B = B(x, r) ∈
A1 and B ∩ B1 6= 0, / then B ⊂ B(x1 ,Cr1 ) for a constant C which depends only on K.
To prove this fact let us first observe that r ≤ K(1 + 2K) r1 , since otherwise, with
y ∈ B(x1 , 2Kr1 ) ∩ B0,1 and u ∈ B ∩ B1 , we would have
3.6 More covering lemmas. 103
d(x, y) ≤ K [ d(x, u) + K (d(u, x1 ) + d(x1 , y))]

< K [ r + K(r1 + 2Kr1 )]
< 2Kr,
which would imply that B = B(x, r) ∈ A˜1 , but r is too large
2r1 > sup R1 ≥ r > K(1 + 2K) r1 ≥ 3r1 .
Let us now show that assuming r ≤ K(1 + 2K)r1 we have that B ⊂ B(x1 ,Cr1 ) where
B = B(x, r) ∈ A1 and B ∩ B1 6= 0.
/ In fact, take z ∈ B and u ∈ B ∩ B1 , then
d(z, x1 ) ≤ K [ d(z, x) + K (d(x, u) + d(x, x1 ))]

< K [ r + K(r + r1 )]
≤ Cr1 .
Let us assume then, that for i ≥ 1 we have B j = B(x j , r j ) and

B0, j (x0, j , r0, j ), j = 1, 2, ..., i satisfying
j ∈ A
(a) B0, j = { B ∈ A : B ∩ [B1 ∪ ... ∪ B j−1 ] = 0};
/
(b) 2µ B0, j > sup { µ(B ) : B ∈ A j };
δ δ
(c) B j ∈ A˜j = { Bα = B(xα , 2Krα ) ∩ B0, j 6= 0};

/
(d) 2r j > sup R j , with R j = { rα : B(xα , rα ) ∈ A˜j };
(e) there exists a constant C > 0 such that for every B ∈ A j with B ∩ B j 6= 0/ we have
that B ⊂ B(x j ,Cr j ).
Let us now show that if there is some α ∈ Γ for which the corresponding ball Bα
is contained in none of the B(x j ,Cr j ), j = 1, 2, ..., i, then there exist two balls Bi+1 =
B(xi+1 , ri+1 ) and B0,i+1 satisfying (a) to (e) with j = i+1. Under our assumption, we
i
have that Ai+1 6= 0,
[
/ since the ball Bα is necessarily disjoint from B j , because,
j=1
otherwise, from (e), Bα ⊂ B(xh ,Crh ) for some h = 1, 2, ..., i. Pick, then, B0,i+1 ∈
Ai+1 such that

2µ Bδ0,i+1 > sup { µ(Bδ ) : B ∈ Ai+1 }.
Since Ri+1 is bounded we choose Bi+1 ∈ A˜i+1 with radius ri+1 such that 2ri+1 >
sup Ri+1 . With the same argument and the same constant C used for the case i = 0
we get (e) for j = i + 1.
If this selection process stops is just because for some i ∈ N every ball B ∈ A is
contained in some of the balls B(x j ,Cr j ); j = 1, 2, ..., i, and the sequence { B1 , ..., Bi }
satisfies the desired properties. Let us assume that we have an infinite sequence { Bi }
generated by our induction. Notice first that from the very definition of A j we have
that these balls are pairwise disjoint. Let us finally prove that for every B ∈ A there
exists j ∈ N such that B ⊂ B(x j ,Cr j ). Take B ∈ A . Let us assume that
∞
[
B∩ Bi = 0.
/
i=1
Then B ∈ Ai for every i ∈ N. Since B0,i ∈ A˜, property (d) implies that 2ri > r0,i .
Let y ∈ B(xi , 2Kri ) ∩ B0,i and z ∈ B0,i then
d(z, xi ) ≤ K [ d(z, x0,i ) + K ( d(x0,i , y) + d(y, xi ))]

< 6K 3 ri .
In other words, B0,i ⊂ B(xi , 6K 3 ri ). Since B ∈ Ai we have from the doubling property
that
r
i
0 < µ Bδ < 2µ Bδ0,i ≤ Cµ Bδ xi , ,
N
with N such that Bδ xi, Nri ⊂ B(xi , ri ) = Bi . Since the balls Bi are pairwise disjoint,

also the balls Bδ xi , Nri are disjoint. Hence
∞ r
i
µ(M) ≥ ∑ µ Bδ xi , = +∞,
i=1 N
∞
[
which is a contradiction. So that B ∩ Bi 6= 0.
/ Then, if we take
i=1
/ we clearly have that B ∈ A j and B ∩ B j 6= 0/ which from (e)
j = min{ i : Bi ∩ B 6= 0}
implies that B ⊂ B(x j ,Cr j ).
3.7 Subspaces of a space of homogeneous type. 105
3.7 Subspaces of a space of homogeneous type.
Let (X, d, µ) be a space of homogeneous type. A subset Y of X is said to be a sub-

space of homogeneous type of X, or (Y, d, µ) is said to be a subspace of (X, d, µ), if
Y ∈ F and (Y, d, µ) is itself a space of homogeneous type, considering the restric-
tion of d, of F and µ to Y.
Let us first notice that in the classical case of the euclidian space Rn with any of
its usual norms, this subspace property looks somehow like a regularity condition on
the set Y involved since open sets with inner cones preserve the doubling property
for the restrictions of the balls and Lebesgue measure.
In this section we shall consider a more specific situation: uniform families of

subspaces of a fixed space of homogeneous type (X, d, µ). Let us precise. If Y is
a subset of X, then the triangular constant of the restriction of d to X can only be
better than K and if ρ ∈ R(d) the same is true for the subset Y with, perhaps, better
equivalence constants. The doubling constant instead, depends strongly on the shape
of Y as elementary examples in the euclidian plane show. Assume that (X, d, µ) is
the euclidian space R2 with its usual structure but let us use, for simplicity, the norm
| x |∞ = max{ |x1 |, |x2 |} instead of the euclidian. For 0 < β ≤ 1 take
Yβ = { (x1 , x2 ) ∈ R2 : x1 ≥ 0 and x2 ≥ x1 }.
β
Consider the balls in (Yβ , d) centered at 0 with radii less than one:
B(0, r) = { (y1 , y2 ) ∈ Yβ : | y |∞ < r}.
ADD CORRESPONDENT FIGURE
1/β
Then B(0, r) = { (y1 , y2 ) ∈ R2 : 0 ≤ y2 < r, 0 ≤ y1 < y2 } so that
Z r
1/β β 1+ 1
|B(0, r)| = y2 dy2 = r β,
0 β +1
1+ 1
and the doubling constant is at least 2 β which tends to infinity for β → 0. Hence
{ (Yβ , d, µ) : 0 < β ≤ 1} is a family of subspaces of (X, d, µ) which is not uniform.
Several problems in real analysis, such as those involving Calderón-Zygmund

decompositions of a cube, can be solved in the euclidian spaces because of a subtle
property of the family of balls: the family of all euclidian balls in Rn is an uniform
family of subspaces of homogeneous type of Rn . In fact, there is a constant A > 0
such that, for every x ∈ Rn , every R > 0, every y ∈ B(x, R) and every r > 0 we have
the inequalities
0 < |B(y, 2r) ∩ B(x, R)| ≤ A |B(y, r) ∩ B(x, R)| < ∞.
Let us now take X = R2 , dµ = dx Lebesgue measure. We now proceed to define

a distance on X. Take a point x = (x1 , y1 ) ∈ R2 different from (0, 0). The function
2 2
x1 x2
F(x1 , x2 ; ρ) = +
ρ3 ρ
for (x1 , x2 ) fixed is a decreasing function of ρ > 0. Since F(x1 , x2 , ρ) tends to infinity
for ρ → +∞, there exists a unique ρ = ρ(x1 , x2 ) solving the equation
F(x1 , x2 ; ρ) = 1.
Let us notice that the function
d(x1 , x2 ) = max{ |x1 |1/3 , |x2 |}
defined on R2 satisfies the inequalities
√
d≤ρ ≤ 2 d.
Since d ((x1 , x2 ) − (y1 , y2 )) is an invariant distance on R2 we see that ρ ((x1 , x2 ) − (y1 , y2 ))

is at least a quasi-distance on R2 . From that equivalence we may also conclude that
the family of ρ-balls is equivalent to the family of d-balls. Hence, since the measure
of a d-ball of radius r > 0 is 4r4 , we have that (X, ρ, µ) with µ Lebesgue measure is
a space of homogeneous type. Of course also (X, d, µ) is a space of homogeneous
type, but let us concentrate in the former. Take r bigger than one. The ball Bρ (0, r)
is the open region bounded by the elliptic curve
x 2 x 2
1 2
+ = 1.
r3 r
Now, we are going to estimate, for r large the measure of the set
Er = Bρ ((0, 0), r) ∩ Bρ (r3 , 0), 1 .

Notice that the point (r3 , 0) is a boundary point of Bρ ((0, 0), r) . Claim: µ(Er ) ≤ Cr
for some constant C and every r large enough. Assume the claim and notice that
Bρ (r3 , 0), 1 is nothing but the euclidian ball centered at the point (r3 , 0). As r → ∞

the ball Bρ ((0, 0), r) becomes relatively narrow at the vertex (r3 , 0). The measure
of the unit ball in the space of homogeneous type Bρ ((0, 0), r) , ρ, µ centered at
points close to the vertex (r3 , 0) becomes smaller and smaller as r grows to infinity.
The proof of the claim follows from the fact that the points of intersection of
the boundary of Bρ ((0, 0), r) with the circle (x1 − r3 )2 + x22 = 1 have x2 coordinates
whose absolute value is of the order of 1r as r → ∞. In fact, the boundary of the ball
x2 x2
Bρ ((0, 0), r) is given by the points (x1 , x2 ) in R2 satisfying the equation r61 + r22 = 1
2 2
x
and those in the boundary of Bρ (r3 , 0), 1 are given by xr31 − 1 + r62 = r16 . Hence

x2 satisfies, at the intersection points
r !
6 x22
r 1− r2
− 1 + x22 = 1,
which, for r large, gives that | x | ' 1r .
Let us point out that if we consider the equivalent space (X, d, µ) with d(x1 , x2 ) =
max{ |x1 |1/3 , |x2 |}, we have that the family of all d-balls is an uniform family of
subspaces of homogeneous type of (X, d, µ) or (X, ρ, µ). So that, even when the
family of ρ-balls is not very well behaved we can find an equivalent quasi-distance
d on R2 for which the d-balls are well behaved.
The fact that this is always the case is given by a theorem proved by Macı́as and
Segovia in [?]: in every space of homogeneous type there is an equivalent quasi-
distance for which the family of balls is an uniform family of subspaces of the given
space.
Let us now finish this section proving that the construction given in Section 16
of Chapter 1, when applied to a space of homogeneous type provides a quasi-metric
δ whose balls constitute an uniform family of subspaces of homogeneous type of
the given space. We shall use the notation of Section 16 in Chapter 1, starting now
with a space of homogeneous type (X, d, µ). Hence we have that the quasi-distance
δ defined in that section is equivalent to d and that the δ -balls are the sections of
the corresponding bands V.
We can actually start by taking d 0 a quasi-distance on X such that d 0 ∼ d and

the d 0 -balls are open sets. Then we apply the construction given in Section 16 in
Chapter 1 to the space (X, d 0 , µ). In this way we also have that the δ -balls are open
sets, so that we do not have to take any special care for the measurability of the sets
involved in the next statement.
Theorem 3.5. Let (X, d, µ) be a space of homogeneous type and let δ be the quasi-
distance equivalent to d given in Section 16 of Chapter 1. Then, there exists a con-
stant C > 0 such that for every x ∈ X, every choice of r and R with 0 < r ≤ 2KR, the
inequality
Cµ (Bδ (y, r)) ≤ µ (Bδ (x, R) ∩ Bδ (y, r))
holds for every y ∈ Bδ (x, R).
As a consequence of the previous theorem, we have that the family of δ -balls is

an uniform family of subspaces.
Theorem 3.6. Let X, d, µ and δ be as before. Then, the family
{ Bδ (x, r) : x ∈ X, r > 0}
is an uniform family of subspaces of homogeneous type of the space (X, δ , µ).
Proof. Take x0 ∈ X and R a positive real number. Let us consider the quasi-
distance δ restricted to the set Y = Bδ (x0 , R) and the measure µ restricted to the
F -measurable parts of Y. Then the δ -ball with radius r in (Y, δ , µ) is given by
BY (x, r) = Bδ (y, r) ∩Y = Bδ (y, r) ∩ Bδ (x0 , R). From Theorem 3.14, for 0 < r ≤ 2KR
we have that
µ (BY (y, 2r)) = µ (Bδ (y, 2r) ∩ Bδ (x0 , R))

≤ µ (Bδ (y, 2r))
≤ A µ (Bδ (y, r))
A
≤ µ (Bδ (x0 , R) ∩ Bδ (x0 , R))
C
A
= µ (BY (y, r)) .
C
On the other hand, when r > 2KR, we have that Bδ (y, r) ⊃ Bδ (x0 , R), so that the
doubling inequality is trivial.
Corollary 3.2. Let X, d, µ and δ be as before. Then, the family of δ -balls in X is an

uniform family of subspaces of homogeneous type of the space (X, d, µ).
Let X be a set and let ρ be a distance on X. Write Uρ (r) to denote the set { (x, y) ∈
X×X : ρ(x, y) < r}. Let us also write, as in Section 16 of Chapter 1,
Uρ (r, 0) = Uρ (r)
Uρ (r, 1) = Uρ 4r ◦Uρ (r, 0) ◦Uρ 4r

.. .. ..
. . .
Uρ (r, n) = Uρ 4rn ◦Uρ (r, n − 1) ◦Uρ r

4n
for every n ∈ N. Let us fix such an n ∈ N. From Lemma 1.35, the family {Uρ (r, n) :
r > 0} satisfies Uρ (r) ⊂ Uρ (r, n) ⊂ Uρ (3r). Hence {Uρ (r, n) : r > 0} also satisfies
properties (a), (b), (c) and (d) in Section 1.13. Since, moreover
−1 r −1 r

Uρ (r, n) = Uρ−1 n ◦ Uρ (r, n − 1) ◦Uρ n = · · · = Uρ (r, n);
4 4
we have that
δn (x, y) = inf { r : (x, y) ∈ Uρ (r, n)}.
defines a quasi-distance on X. We can actually show that the triangular constant Kn

for δn is uniformly bounded above. In fact, for x, y and z ∈ X and ε > 0 let us pick
r1 and r2 such that
r1 < δn (x, y) + ε, (x, y) ∈ Uρ (r1 , n);

r2 < δn (y, z) + ε, (y, z) ∈ Uρ (r2 , n).
Then
(x, z) ∈ Uρ (r1 , n) ◦Uρ (r2 , n) ⊆ Uρ (r1 + r2 , n) ◦Uρ (r1 + r2 , n)

⊆ Uρ (3(r1 + r2 )) ◦Uρ (3(r1 + r2 ))
⊆ Uρ (6(r1 + r2 )) ,
so that
δn (x, z) ≤ 6(r1 + r2 ) < 6 (δn (x, y) + δn (y, z)) + 2ε.
This inequality implies that Kn = K(δn ) ≤ 6 for every n.
Let us also notice, by the way, that since
{ r : (x, y) ∈ Uρ (r)} ⊂ { r : (x, y) ∈ Uρ (r, n)} ⊂ { r : (x, y) ∈ Uρ (3r)},
we have that ρ(x, y) ≥ δn (x, y) ≥ 31 ρ(x, y). In other words the quasi-distances δn are
uniformly equivalent to ρ. Also, from the argument given in the proof of Lemma
1.37, we see that the δn -balls are the sections of the bands Uρ (r, n), i.e. Bδn (x, r) =
{ y : (x, y) ∈ Uρ (r, n)}.
Theorem 3.7. Let (X, ρ) be a metric space such that (X; ρ, µ) is a space of homo-
geneous type with doubling constant A. Then, there exists a constant c depending
only on A such that the inequality
µ Bδn (x, 1) ∩ Bδn (y, 4−i ) ≥ cµ Bδn (y, 4−i )

holds for every x and y in X satisfying δn (x, y) < 1 and every choice of integers with
1 ≤ i ≤ n.
Proof. Since δn (x, y) < 1, there exists 0 < r < 1 such that
(x, y) ∈ Uρ (r, n) ⊂ Uρ (1, n).

This fact means that there exists a chain of points of X joining x and y with expo-
nential decay on the ρ-distance of consecutive links of the chain. To be precise, we
have 2(n + 1) points in X denoted by
x0 , x1 , x2 , . . . , xn−1 , xn , yn , yn−1 , . . . , y2 , y1 , y0 ;
such that x0 = x, y0 = y, and
(x1 , x0 ) ∈ Uρ (4−n ), ρ(x1 , x0 ) < 4−n ;

.. ..
. .
(xn− j+1 , xn− j ) ∈ Uρ (4− j ), ρ(xn− j+1 , xn− j ) < 4− j ;
.. ..
. .
(xn , xn−1 ) ∈ Uρ (4−1 ), ρ(xn , xn−1 ) < 1/4;
(yn , xn ) ∈ Uρ (1), ρ(yn , xn ) < 1;
(yn , yn−1 ) ∈ Uρ (4−1 ), ρ(yn , yn−1 ) < 1/4;
.. ..
. .
(yn− j+1 , yn− j ) ∈ Uρ (4− j ), ρ(yn− j+1 , yn− j ) < 4− j ;
.. ..
. .
(y1 , y0 ) ∈ Uρ (4−n ), ρ(y1 , y0 ) < 4−n ;
with 1 ≤ j ≤ n. We shall prove that there exists an absolute constant k > 0 such that
for every integer i satisfying 1 ≤ i ≤ n, we have
(a) Bρ yn−i , 4−i−1 ⊂ Bδn y, 4−i ;

(b) Bρ yn−i , 4−i−1 ⊂ Bδn (x, 1);

(c) Bδn y, 4−i ⊂ Bρ yn−i , k 4−i−1 .

Assuming (a), (b) and (c), we easily get the desired result from the doubling
property of µ. In fact, with C = A[log2 k]+1 we have
µ Bδn (y, 4−i ) ≤ µ Bρ (yn−i , k 4−i−1 )

≤ Cµ Bρ (yn−i , 4−i−1 )

≤ Cµ Bδn (x, 1) ∩ Bδn (y, 4−i ) .

Let us prove (a). Since ρ ≥ δn , we only have to show that
Bρ yn−i , 4−i−1 ⊂ Bρ y, 4−i .

Take z ∈ Bρ yn−i , 4−i−1 . In order to estimate the ρ-distance from z to y we use the

links of the chain joining yn−i and y0 = y, hence
ρ(z, y) ≤ ρ(z, yn−i ) + ρ(yn−i , y0 )

n−1
< 4−i−1 + ∑ ρ(yn− j , yn− j−1 )
j=i
∞
< 4−i−1 + ∑ 4− j−1
j=i
−i
<4 ,
so that z ∈ Bρ y, 4−i as desired. Let us now prove (b). Take a point z ∈ Bρ yn−i , 4−i−1 ,

we are going to show that (z, x) ∈ Uρ (1, n) which implies δn (x, z) < 1 or, in other
words, z ∈ Bδn (x, 1). In order to prove that (z, x) ∈ Uρ (1, n) we only have to observe
that we can modify the given chain joining x and y to obtain a chain joining z and x,
in the following way:
y0 = z, y1
= z, . . . , yn−i−1
= z, yn−i , yn−i+1 , yn−i+2 , . . . , yn , xn , xn−1 , . . . , x1 , x0
= x.
Notice that ρ(yn− j+1 , yn− j ) < 4− j for 1 ≤ j ≤ n and since, of course, ρ(yn , xn ) <
1 and ρ(xn− j+1 , xn− j ) < 4− j , we have that (z, x) ∈ Uρ (1, n), as desired.
Let us, finally show that (c) holds. Take z ∈ Bδn y, 4−i . This means that (z, y) ∈

Uρ 4−i , n ⊂ Uρ (3 . 4−i ). Hence ρ(z, y) < 3 . 4−i . Using the links y = y0 , y1 , · · · , yn−i
of the chain and the above inequality we have that
ρ(z, yn−i ) ≤ ρ(z, y) + ρ(y0 , yn−i )

n−i−1
< 3 . 4−i + ∑ ρ(y j , y j+i )
j=0
n−i−1
< 3 . 4−i + ∑ 4 j−n
j=0
4
< 3 . 4−i + 4−i−1 .
3
40 −i−1
= 4 ,
3
which proves that z ∈ Bρ yn−i , k 4−i−1 , with k = 40

3 .
Let us notice that if (X, ρ, µ) is a metric space with a doubling measure with
doubling constant A, then the same is true for (X, λ ρ, µ) with λ any positive number
with the same constant A. Let us also observe that Uλ ρ (r, 0) = Uλ ρ (r) = Uρ λr ,

r r
Uλ ρ (r, 1) = Uλ ρ ◦Uλ ρ (r, 0) ◦Uλ ρ
r4 r r4
= Uρ ◦Uρ ◦Uρ
λr 4 λ λ4
= Uρ ,1 ,
λ
r

and, generally, Uλ ρ (r, n) = Uρ λ ,n . This means that
δnλ (x, y) = inf { r > 0 : (x, y) ∈ Uλ ρ (r, n)}

= λ δn1 (x, y)
= λ δn (x, y).
Since the constant c in Theorem 3.17 only depends on A and A is uniform in λ

we have that

µ Bδnλ (x, 1) ∩ Bδnλ (y, 4−i ) ≥ cµ Bδnλ (y, 4−i ) ,
for every x and y in X with δnλ (x, y) < 1, every 1 ≤ i ≤ n and every λ > 0. But this
fact can be rewritten in terms of δn -balls to produce the next result.
Corollary 3.3. Let (X, ρ, µ) be a space of homogeneous type with ρ a distance on

X and A as the doubling constant. Then, there exists a constant C depending only
on A such that the inequality

µ Bδn (x, R) ∩ Bδn (y, r) ≥ Cµ Bδn (y, r) .
holds for every x and y in X with δn (x, y) < R, every choice of r and R satisfying
4−n R < r < R and every n ∈ N.
The next step consists in taking the limit for n growing to infinity. Set, as in
Section 16 of Chapter 1,
∞
[
Vρ (r) = Uρ (r, n) and δ (x, y) = inf { r : (x, y) ∈ Vρ (r)}.
n=0
Since Uρ (r, n) ⊂ Uρ (r, n + 1) ⊂ Vρ (r) we have that ρ ≥ δn ≥ δn+1 ≥ δ ≥ ρ3 . Also

δ = lim δn . Then, if δ (x, y) < R, there exists n0 ∈ N such that δn (x, y) < R for every
n→∞
n ≥ n0 . On the other hand if 0 < r < R there exists n1 ∈ N such that 4−n R < r < R
for every n ≥ n1 . Since δn & δ we have that Bδn (z, s) % Bδ (z, s) and we have the
next result.
Corollary 3.4. Let (X, ρ, µ) be a space of homogeneous type with ρ a distance.

Then there exists a constant C depending only on the doubling constant such that
the inequality
µ (Bδ (x, R) ∩ Bδ (y, r)) ≥ Cµ (Bδ (y, r))
holds for every x and y in X with δ (x, y) < R and every R > r > 0.
Now we can go back to general spaces of homogeneous type. First of all let
us notice that equivalent quasi-distances d ∼ d 0 produce equivalent quasi-distances
δ ∼ δ 0 for an appropriate choice of the parameter a in Section 1.16. On the other
hand if ρ is a distance and β is a positive number we have that
1/β
Uραβ ((r, n) = Uρα (r1/β , n),
3.8 Normal spaces and normalization 115
where Udα (r, n) is defined as in Section 1.6 with a = α. This fact shows that
1/β β
δραβ (x, y) = δρα (x, y) , where δda (x, y) is the quasi-distance δ defined in Sec-
tion 1.16 taking explicit account of the original quasi-distance d and the parameter
a > 0 defining it. Since, of course

Bδρβ α (x, R) = B α 1/β
x, R1/β ,
δρ
taking ρ and β such that d ∼ ρ β we get Theorem 3.14.
3.8 Normal spaces and normalization
In some of the most basic examples of spaces of homogeneous type we have that
the measure of a ball of radius r is comparable to a power of r with constants that
do not depend on the center. We can actually normalize each space of homogeneous
type, by changing its quasi-distance, in such a way that for the resulting space the
measure of a ball is essentially comparable to its radius, with comparison constants
that do not depend on the center of the ball. Of course the new quasi-distance is
generally inequivalent to the original one in the strong sense, but they still generate
the same topology. The most basic example is given by the euclidian space Rn with
Lebesgue volume. The function |x − y|n defines a quasi-distance on Rn for which
the d-ball of radius r > 0 and any center has measure equal to a constant times r.
We shall say that a space of homogeneous type (X, δ , µ) is a normal space if

there exist constants C1 ,C2 , A1 and A2 such that for every r with A1 µ ({ x}) < r <
A2 µ(X), we have that the δ -ball B(x, r) is measurable and that
C1 r ≤ µ (B(x, r)) ≤ C2 r.
ADD CORRESPONDENT FIGURE
Assume that (X, d, µ) is a space of homogeneous type. Take d 0 = ρ β with ρ a

distance on X and β > 0 such that d ∼ d 0 . Recall that the topologies generated by
d, d 0 and ρ are all the same. Let us consider two functions defined on X×X by
δ (x, y) = inf { µ(B0 ) : x ∈ B0 , y ∈ B0 , B0 is a d 0 -ball}

and
∆ (x, y) = µ B0 x, d 0 (x, y) ,

for x 6= y and δ (x, x) = ∆ (x, x) = 0.
Lemma 3.9. The functions δ and ∆ are equivalent on X×X. In other words there
exist constants a and b such that
a δ (x, y) ≤ ∆ (x, y) ≤ b δ (x, y)
for every (x, y) ∈ X×X.
Proof. When x = y there is nothing to prove since both, δ and ∆ are zero. As-
sume, then x 6= y. Since d 0 (x, y) > 0 we have that both points x and y belong to
B0 (x, 2d 0 (x, y)) .
Hence
δ (x, y) ≤ µ B0 x, 2d 0 (x, y) ≤ A0 µ B0 x, d 0 (x, y)

= A0 ∆ (x, y),
with A0 the doubling constant for d 0 . On the other hand if B0 = B0 (z, r) is a d 0 -ball
containing both x and y, we have that
B0 x, d 0 (x, y) ⊂ B0 z, 3K 02 r

with K 0 the triangular constant for d 0 . In fact if µ ∈ B0 (x, d 0 (x, y)) , then
d 0 (u, z) ≤ K 0 d 0 (u, x) + d 0 (x, z)

< K 0 d 0 (x, y) + r

≤ K 0 K 0 d 0 (x, z) + d 0 (z, y) + r

≤ 3(K 0 )2 r
So that
µ B0 x, d 0 (x, y) ≤ µ B0 (z, 3K 02 r)

≤ Cµ(B0 )
for every ball B0 containing both x and y. Then C−1 ∆ (x, y) ≤ δ (x, y).
Notice that δ is symmetric but ∆ is generally not. The function
δ̄ (x, y) = ∆ (x, y) + ∆ (y, x)

is a symmetric function which is also equivalent to δ . Let us also observe that a

symmetric non negative function δ defined on X×X is a quasi-distance on X if
and only if δ is equivalent to a function ∆ such that ∆ (x, y) = 0 if and only if
x = y and for some constant K and every x, y and z ∈ X the inequality ∆ (x, z) ≤
K (∆ (x, y) + ∆ (y, z)) holds.
Lemma 3.10. The function δ is a quasi-distance on X.
Proof. From the remarks above we only have to check that ∆ is faithful and that
it satisfies a quasi-triangular inequality. Since for x 6= y two points in X we have
that d 0 (x, y) > 0, then the µ-measure of the ball B0 (x, d 0 (x, y)) is positive. So that
∆ (x, y) > 0. Let us now check that ∆ satisfies a quasi-triangular inequality. Take x, y
and z three points in X. From the triangular inequality for d 0 we get that d 0 (x, y) ≥
d 0 (x,z) 0 d 0 (x,z)
2K 0 or d (y, z) ≥ 2K 0 . Hence, at least one of the following set inclusions hold
B0 x, d 0 (x, z) ⊂ B0 x, 2K 0 d 0 (x, y) ,

or
B0 z, d 0 (x, z) ⊂ B z, 2K 0 d 0 (y, z) .

So that, for some constant C we have
∆ (x, z) ≤ C∆ (x, y) or ∆ (z, x) ≤ C∆ (z, y),

then
∆ (x, z) ≤ C (∆ (x, y) + ∆ (z, y))

≤ C̃ (∆ (x, y) + δ (y, z)) .
The last inequality follows from Lemma 3.20 since ∆ is equivalent to the sym-
metric function δ .
Theorem 3.8. Let (X, d, µ) be a space of homogeneous type such that the d-balls
are open sets. Then the function
δ (x, y) = inf { µ(B) : B is a d-ball containing x and y},
for x 6= y and δ (x, x) = 0 is a quasi-distance on X. The space (X, δ , µ) is a normal

space of homogeneous type and τd = τδ , i.e. δ and d induce the same topology on
X.
Let us have a picture for the δ -balls in X.
Lemma 3.11. Let X, d, µ and δ be as in the theorem. Then if for every x ∈ X and
every r > 0 we write F (x, r) to denote the, possibly empty, class of all the d-balls B
for which x ∈ B and µ < r, we have that
[
Bδ (x, r) = { x} ∪ B
B∈F (x,r)
is an open set (Bδ (x, r) ∈ τd ).
Proof. If y ∈ Bδ (x, r) and y 6= x, then there exists B a d-ball containing x and y such
that both x and y belong to B and moreover µ(B) < r. This shows that y belongs to
some ball in F (x, r). Since δ (x, x) =[ 0, we certainly have that x ∈ Bδ (x, r). Assum-
ing F (x, r) 6= 0,
/ take a point y ∈ B, then there exists a d-ball B containing
B∈F (x,r)
x and y with µ(B) < r. Hence δ (x, y < [ r and y ∈ Bδ (x, r). Since we are assum-
ing that the d-balls are open sets, then B is an open set. If F (x, r) 6= 0,
/ then
B∈F
B is open. If, otherwise, F (x, r) = 0/ we have
[ [
x∈ B and Bδ (x, r) =
B∈F (x,r) B∈F (x,r)
on one hand that Bδ (x, r) = { x} and on the other hand that every d-ball B containing
x has measure larger than or equal to r : µ(B) ≥ r for every d-ball for which x ∈ B.
This fact shows that { x} is an atom: µ ({ x}) > 0. So that { x} is isolated and hence
open.
Proof (Proof of Theorem 3.22). We know from Lemma 3.21 that δ is a quasi-
distance on X. Since from Lemma 3.23 we have that every δ -ball belongs to τd ,
then τδ ⊂ τd . In order to prove that τd ⊂ τδ we shall show that for every x ∈ X and
every r > 0 there exists a positive number s such that Bδ (x, s) ⊂ B(x, r) = Bd (x, r).
2
Let A be the doubling constant for (X, d, µ) and Ã ≥ A1+log2 3K . Take s =
1
Ã
µ (B(x, r)) . Let us prove that Bδ (x, s) ⊂ B(x, r). For y 6= x with y ∈ Bδ (x, s) there
exists a d-ball B = B(z,t) such that x ∈ B, y ∈ B and µ(B) < s = Ã1 µ (B(x, r)) . Notice
that y ∈ B (x, 2Kt) ⊂ B z, 3K 2t . Since, on the other hand

µ (B(x, 2Kt)) ≤ µ B(z, 3K 2t) ≤ Ã µ(B)

< µ (B(x, r)) ,
we have that 2Kt < r. Hence y ∈ B(x, r), as we wanted to show. Let us finally prove
that (X, δ , µ) is a normal space. Notice that, being open, every δ -ball is a measurable
subset of X. Take x ∈ X, observing that if µ(X) = µ ({ x}) we must have X = { x},
we can assume without loosing generality that µ(X) > µ ({ x}) . Then for some
b ∈ (1, 2] we have that µ(X) > bk > µ ({ x}) for some integer k. Take now a positive
real number r such that µ ({ x}) < r < µ(X). The sequence { µ (B(x, bn )) : n ∈ Z}
is non decreasing, tends to µ ({ x}) for n → −∞, and to µ(X) for n → +∞. Then,
there exists n ∈ Z such that
µ (B (x, bn )) < r ≤ µ B x, bn+1

.
Since µ (B(x, bn )) < r we have that B (x, bn ) ⊂ Bδ (x, r). Hence
µ (Bδ (x, r)) ≥ µ (B (x, bn )) = µ B x, b−1 bn+1

≥ C1 µ B(x, bn+1 )

≥ C1 r
for a geometric positive constant C1 .
Let us prove now that we also have the inequality
µ (Bδ (x, r)) ≤ C2 r,
for some geometric constant C2 , every x ∈ X and every r > 0. Since Bδ (x, r) is an
open set, it is measurable. If we show that Bδ (x, r) has finite measure, since it is the
union of the d-balls in F (x, r), we can apply Lemma 3.13 in Section 3.6 to obtain a
disjoint sequence { Bi } of balls in F (x, r) such that every ball in F (x, r) is contained
3.9 Some basic integrals of Newton - Riesz kernels on normal spaces. 121
in a fixed dilation B̃i of Bi . Since each Bi belongs to F (x, r), we necessarily have
that x ∈ Bi . Since the sequence Bi is pairwise disjoint we have that the sequence
{ Bi } consists of exactly one ball B1 ∈ F (x, r). Hence, for every B ∈ F (x, r) we
have that B ⊂ B˜1 . So that Bδ (x, r) ⊂ B˜1 and since B1 ∈ F (x, r)
µ (Bδ (x, r)) ≤ µ B˜1

≤ Ã µ(B1 )
< Ã r.
Let us show that µ (Bδ (x, r)) < ∞ by proving that actually Bδ (x, r) is a d-bounded
set. If µ(X) < ∞ there is nothing to prove. Assume that µ(X) = +∞. To prove the
claim it suffices to show that there exists a finite number M such that B ⊆ B(x, M) for
every B ∈ F (x, r). If, otherwise, for each n ∈ N there exists Bn = B(xn , rn ) ∈ F (x, r)
such that Bn 6⊆ B(x, n). Then, from the triangular inequality, rn > n/2K. So that
n
B x, ⊆ B(xn , 2Krn ).
2K
Hence
Ã r > Ã µ(Bn ) ≥ µ (B (xn , 2Krn ))

n
≥ µ B x, ,
2K
n

which is impossible since µ B x, 2K tends to µ(X) = +∞ for n → ∞.
3.9 Some basic integrals of Newton - Riesz kernels on normal

spaces.
Let (X, d, µ) be a normal space of homogeneous type: the d-balls are measurable
and there exist four constants A1 , A2 , K1 and K2 , with K2 ≤ 1 ≤
≤ K1 such that
• A1 r ≤ µ (B(x, r)) for r ≤ K1 µ(X);

• B(x, r) = X for r > K1 µ(X);
• A2 r ≥ µ (B(x, r)) for r ≥ K2 µ ({ x}) ;
• B(x, r) = { x} for r < K2 µ ({ x}) .
Notice that if the d-balls are measurable sets, then for x ∈ X fixed and γ ∈ R, since
the level sets of the function fR(y) = d γ (x, y) are balls, we have that f is a measurable
nonnegative function. Hence E f (y) dµ(y) is well defined for every E ∈ F . When,
for example, the space is (Rn , d, λ ) with d(x, y) = |x − y|n , λ is Lebesgue measure
and γ = αn − 1, 0 < α < n, the function d γ (x, y) is the kernel for the Riesz potential
of order α. The limit power between integrability and singularity in a normal space
is, generally speaking, the same as in the one dimensional euclidian case given by
γ = −1. The next result contains some basic local and global integrability properties
of these kernels.
We shall get upper and lower estimates of
Z
d(x, y)γ dµ(y)
r≤d(x,y)<R
as a function of r, R and γ, with 0 < r < R < ∞ and γ ∈ R, uniformly in x ∈ X, in a

normal space of homogeneous type (X, d, µ).
Let us start by assuming that K2 µ ({ x}) ≤ r < R ≤ K1 µ(X). Then, for any s > 0
with r ≤ s ≤ R we have the inequalities A1 s ≤ µ (B(x, s)) ≤ A2 s. Since τ = 2A A1 ≥ 2,
2
there exists a non-negative integer k such that rτ k ≤ R <

< rτ k+1 . So that, with τ = r≤d(x,y)<rτ k d(x, y)γ dµ(y) we have the inequalities
R
Z Z
J≤ d(x, y)γ dµ(y) ≤ J + d(x, y)γ dµ(y).
r≤d(x,y)<R rτ k ≤d(x,y)<R
Let us get an upper estimate for the last integral. Since in the domain of integration
τ −1 ≤ d(x,y)
R < 1, we have that
Z
d(x, y)γ dµ(y) ≤ CRγ µ (B(x, R))
rτ k ≤d(x,y)<R
≤ C̄Rγ+1 .
When k = 0 we generally have no better estimate than

3.9 Some basic integrals of Newton - Riesz kernels on normal spaces. 123
Z
0≤ d(x, y)γ dµ(y) ≤ C̄Rγ+1 .
r≤d(x,y)<R
To get estimates for I(γ) when k ≥ 1, we decompose its domain of integration in

τ-adic annuli in order to get
k−1 Z
J= ∑ d(x, y)γ dµ(y).
j j+1
j=0 rτ ≤d(x,y)<rτ
Since
Z
d(x, y)γ dµ(y) ' rγ τ jγ µ B(x, rτ j+1 ) − µ B(x, rτ j ) ,

rτ j ≤d(x,y)<rτ j+1
with constant depending only on the geometry of the space and since
A1 j+1 A2 j+1
rτ = A1 rτ j+1 − rτ
2 τ
≤ µ B(x, rτ j+1 ) − µ B(x, rτ j )

≤ A2 rτ j+1 ,
we have that
k−1
J ' rγ+1 ∑ τ j(γ+1) .
j=0
k−1
On the other hand, taking into account the behavior of ∑ τ j(γ+1) for k ≥ 1, when
j=1
γ = −1, γ > −1 and γ < −1, we obtain the results contained in the next statement.
Lemma 3.12. Let (X, d, µ) be a normal space. Let x be a point in X, let r < R be two
positive
R
real numbers and let γ ∈ R. Let us write I(x, r, R, γ) to denote the integral
γ
r≤d(x,y)<R d(x, y) dµ(y). Then
1. I(x, r, R, γ) ≤ CRγ+1 for 0 < r < R ≤ K1 µ(X);

2. I(x, r, R, −1) ' log Rr for K2 µ ({ x}) ≤ r < τr ≤ R ≤ K1 µ(X) and τ = 2A 2
A1 ;
3. I(x, r, R, γ) ' Rγ+1 for K2 µ ({ x}) ≤ r < τr ≤ R ≤ K1 µ(X) and γ > −1;
4. I(x, r, R, γ) ' rγ+1 for K2 µ ({ x}) ≤ r < τr ≤ R ≤ K1 µ(X) and γ < −1,
where all the equivalence constants do not depend on x, r and R.
Notice that from property 2 of the previous lemma we have that
Z
d(x, y)−1 dµ(y) = +∞,
B(x,r)
since if µ ({ x}) = 0 we can use 2 for r small enough and, otherwise, if µ ({ x}) > 0
the kernel d(x, y)−1 equals +∞ on a set of positive measure. We also have that
Z
d(x, y)−1 dµ(y) = +∞
d(x,y)≥r
for every r > 0 provided that µ(X) = +∞.

R
From property 4 of Lemma 3.24, we see that d(x,y)≥r d(x, y)
γ dµ(y) ' rγ+1 , pro-
K1
vided that µ(X) ≥ r ≥ K2 µ ({ x}) and γ < −1. From 3., for µ ({ x}) = 0, γ > −1
τ R
and r ≤ K1 µ(X) we get d(x,y)<r d(x, y)γ dµ(y) ' rγ+1 .
3.10 Reverse doubling.
Assume that µ is a Borel doubling measure on the usual euclidian metric space Rn ,
with doubling constant equal to A > 0. Take x∈ Rn and r > 0, then there exists y ∈
Rn suchthat | x−y | = 32 r.We have that B y, 2r ⊂ B(x, 2r), that B(x, r) ⊂ B y, 52 r ⊂
B y, 8 2r and that B y, 2r ∩ B(x, r) = 0./ So that
r
µ (B(x, r)) + µ B y, ≤ µ (B(x, 2r))
2
and r r
µ (B(x, r)) ≤ µ B y, 8 ≤ A3 µ B y, ,
2 2
hence
1
µ (B(x, 2r)) ≤ 1 + 3 µ (B(x, r)) .
A
3.10 Reverse doubling. 125
In other words, from the doubling property of µ on the euclidian space Rn we have
an effective growth of the measures of balls: there exists a constant γ > 1 such that
the inequality
µ (B(x, 2r)) ≥ γ µ (B(x, r)) ,
holds for every x ∈ Rn and every r > 0.
Notice that we have used a very sharp property of the euclidian distance: that
given x ∈ Rn and r positive there exists a point y ∈ Rn for which | y − x | = 32 r. When
a measure µ satisfies such a condition we shall say that µ is reverse doubling. We
have seen that on Rn , doubling implies reverse doubling. The converse does not
even hold true in R1 . In fact, take for example dµ(x) = ex dx.
Let us observe that since in a space of homogeneous type atoms are isolated
points, if µ ({ x}) > 0 we have that B(x, s) = { x} for every 0 < s < r0 for some
r0 > 0. Hence if 0 < 2r < r0 we have that
µ (B(x, 2r)) = µ ({ x}) = µ (B(x, r)) .
So that the reverse doubling is generally not a consequence of the doubling in

quasi-metric spaces. Notice also that a reverse doubling is also impossible in a
bounded setting, since for r large enough B(x, 2r) = B(x, r) for every x ∈ X. How-
ever if the space (X, d) is a “uniformly perfect”([?]) quasi-metric space we have that
doubling implies a reverse doubling.
Lemma 3.13. Let (X, d) be a quasi-metric space such that

the d-balls are Borel
sets and there exists b > 1 for which B x, bk+1 \ B x, bk 6= 0/ for every k ∈ Z and
every x ∈ X. Then, every doubling Borel measure on X satisfies a reverse doubling
property: there exist constants γ > 1 and C > 1 such that µ (B(x,Cr)) ≥ γ µ (B(x, r))
for every x ∈ X and r > 0.
in X and let k be a fixed integer number. Take y ∈

Proof. Let x be a point
B x, bk+2 \ B x, bk+1 . Then, assuming without loss of generality that b > 2K,
(a) B x, bk ∩ B y, bk = 0;

/
k k+3

(b) B y, b ⊂ B x, b ;
(c) B x, bk ⊂ B y, bk+3 .

Let us prove (a). Assume that there is a point z that belongs to both balls. Then

d(x, y) < K bk + bk = 2Kbk < bk+1 ,
which is impossible. To show (b) take z ∈ B y, bk , hence,

d(z, x) < K bk + bk+2 < 2Kbk+2 < bk+3 ,
so that z ∈ B x, bk+3 . Given z ∈ B x, bk , we have that

d(z, y) < K bk + bk+2 < bk+3 ,
and (c) is also proved. Then

µ B(x, bk ) + µ B y, bk ≤ µ B x, bk+3
and

µ B x, bk ≤ µ B y, bk+3 ≤ Ã µ B y, bk .
Hence

1
1+ µ B x, bk ≤ µ B x, bk+3 .
Ã
Given r > 0 pick k ∈ Z such that bk−1 < r ≤ bk . Then, with γ = 1 + Ã1 , we have that

γ µ (B(x, r)) ≤ µ B x, bk+3 ≤ µ B x, b4 r .

If (X, d, µ) is an unbounded non-atomic normal space, then a reverse doubling

with C > AA12 is a consequence of
A1 r ≤ µ (B(x, r)) ≤ A2 r.
3.11 The regularity of the functions ηx (r). 127
2A2
Notice that with τ = A1 , as in the previous section, we have that
A1 j+1 A2 j+1
τ = A1 τ j+1 − τ
2 τ
≤ µ B x, τ j+1 − µ B x, τ j

= µ B x, τ j+1 \ B x, τ j .

Hence B x, τ j+1 \ B x, τ j 6= 0.

/
There are non-atomic unbounded spaces of homogeneous type for which the
measure does not satisfy the reverse doubling property. Take for example X =
∞ h
[ i
kk , kk + 1 as a subspace of R with its standard structure of space of homo-
k=2
geneous type.
3.11 The regularity of the functions ηx (r).
The euclidian space Rn with the quasi-distance d(x, y) = | x−y |n and Lebesgue mea-
sure is a normal space for which the family of functions { ηx (r) = µ (B(x, r)) ; x ∈
Rn } has only one member which is a constant times the identity function, η(r) = cn r.
The fact that Rn is a group and the translation invariance of quasi-distance and mea-
sure are the reasons for which the space is so homogeneous that the class { ηx } has
only one member. Also Zn with the restriction of d and the counting measure has
this homogeneity property; the class { ξk (r) = ] (B(k, r) ∩ Zn ) ; k ∈ Zn } has only one
element. But η0 and ξ0 are very different at least from the point of view of regu-
larity. We shall say that (X, d, µ) satisfies property (P) if every d-ball is a Borel set
and { ηx : x ∈ X} is an uniform family of Lipschitz functions. In other words, there
exists a constant C such that
µ (B(x, r + s)) − µ (B(x, r)) ≤ Cs,
for every x ∈ X, every r > 0 and every s > 0. Let us observe that if (X, d, µ) satisfies
property (P), then we have the uniform upper estimate for normality from which
we conclude that (X, d, µ) has no atoms. But we do not have the lower bound in
general. This means that property (P) does not imply normality. Actually (X, d, µ)
may have property (P) without being a space of homogeneous type.
In contrast with the normalization procedure that can be applied to every space
of homogeneous type, in order to produce a new quasi-distance generating the same
topology, it is generally impossible to give a topology preserving quasi-distance
on a space of homogeneous type in order to produce a space satisfying property
(P). In fact, property (P) does not allow atoms, hence it does not allow points that
are at once closed and open sets. But this is a topological property that should be
preserved. For example, there are not quasi-distances d¯ on (Z, d, ]) (d the restriction
of the standard euclidian distance to Z, ] the counting measure) such that d¯ generates
the same topology as d on Z and (Z, d, ¯ ]) has property (P).
Sometimes a different kind of construction can be used in order to get a very

regular setting from a rough one. Just to illustrate a technique that can be useful, let
us keep considering the case of (Z, d, ]). In our further development of analytical
problems in this setting we shall deal with real valued functions defined on Z.
Say f : Z → R; k → f (k). Of course a “continuous” model for this situation is

(R, d, λ ), where d is the usual distance on the line, λ is Lebesgue measure, and f
changes to g(x) = ∑ f (k)X(k,k+1] (x). But while f is continuous, g is not. We could
k∈Z
instead build a semi-discrete model for the situation preserving a topological prop-
erty but improving another. Consider the set Z×(−1/2, 1/2) as a metric subspace of
the euclidian R2 . Let us write ρ to denote the restriction of the usual distance in the
plane to the set X. Let us consider the product measure µ = ]×λ on X, with ] the
counting measure on Z and λ the one dimensional Lebesgue measure on ( −1 2 , 12).
Now the space (X, ρ, µ) is more regular than the original (Z, d, ]) and the function
spaces on the latter can be regarded as corresponding subspaces on the former, when
we consider cylindrical functions that do not depend on the variable in (− 12 , 21 ).
Property (P) has, over normality, the extra advantage of allowing upper estimates
for some integrals of kernels without any a priori growth condition.
Let us start by introducing the basic context for some integral estimates which
are possible with the extra condition given by property (P). Radially decreasing
integrable kernels in a normal space define approximate identities which are well
behaved for pointwise convergence on the Lebesgue spaces associated to (X, d, µ).
In Rn there is an extension including more general kernels of these results due to Z
o ([?]) based on the classical Calderón-Zygmund technique for singular integrals.
In this context a non-negative integrable function k which is smooth in Rn − {0}
with | ∇k| ≤ C | x |−n−1 , defines a good approximate identity for the convolution.
To obtain a generalization of this situation to spaces of homogeneous type, let us
restrict our attention to the case in which k is also radial: k(x) = k̃(| x |). It is easy
to show that, k is radial, nonnegative, integrable and belongs to C 1 (Rn − {0}) with
| ∇k(x)| ≤ C| x |−n−1 for some constant C if and only if there exists a function ϕ
defined on R+ such that
ϕ(| x |n )
k(x) = , x 6= 0;
| x |n
and
1. ϕ ≥ 0;
2. ϕ ∈ C 1 (R+ );
3. ϕ is bounded;
4. |Rϕ 0 (t) | ≤ C t −1 , t > 0;
∞ dt
5. 0 ϕ(t) t < ∞.
Since the approximate identity is given by the family of kernels
1 x
= ϕ δ −n | x |n | x |−n ,

kδ (x) = n
k δ >0
δ δ
and since | x |n is the normalization of | x |, with ε = δ n > 0, we are load to consider

the integral operators with kernels

d(x, y) .
Kε (x, y) = ϕ d(x, y);
ε
on a normal space of homogeneous type for a function ϕ satisfying properties 1 to

5. The uniform integrability of the family of functions
{ Kε (x, .) : ε > 0, x ∈ X}
is one of the basic facts in order to fit the conditions for the family of kernels { Kε :
ε > 0} to behave as an approximate identity. In order to show how property (P) can
be used to estimate these integrals for general ϕ satisfying 1 to 5 we shall prove the
next result.
Lemma 3.14. Let (X, d, µ) be a normal space satisfying property (P). Let ϕ be a
function defined on R+ satisfying properties 1 to 5. Then, there exists a constant C
depending only on the geometric constants of the space and on the functions ϕ, such
that
R a/ε
ϕ(t) dtt for every ε ∈ (0, 1), every x ∈ X and ev-
R
(a) d(x,y)<a Kε (x, y) dµ(y) ≤C 0
ery a > 0, and
R R∞ dt
(b) X Kε (x, y) dµ(y) ≤ C 0 ϕ(t) t , for every ε ∈ (0, 1).
Proof. Let us first notice that (b) follows from (a) taking a → ∞. Let us prove (a).
For x ∈ X fixed, as a function of (ε, y) ∈ (0, 1)×X, Kε (x, y) is a measurable function.
Being X a σ -finite measure space and ϕ ≥ 0 we can apply Tonelli’s theorem in order
to guarantee that
Z 1 Z Z Z 1
−1 −1

Kε (x, y) dµ(y) dε = d(x, y) ϕ ε d(x, y) dε dµ(y)
0 X X 0
(3.12)
d(x,y)
After the change of variables given by ε → t = ε , the right hand side of (3.12)
becomes
Z Z ∞
dt
ϕ(t) 2 dµ(y).
X d(x,y) t
Interchanging again the order of integration, applying the normality and the fact
that Property (P) does not allow atoms, we have
Z 1 Z Z
∞ dt
Z
Kε (x, y) dµ(y) dε = ϕ(t) dµ(y) 2
0 X 0 d(x,y)<t t
dt
Z ∞
= ϕ(t) µ (B(x,t)) 2
0 t
dt
Z ∞
≤C ϕ(t) .
0 t
R
Hence X Kε (x, y) dµ(y) is finite for almost every ε ∈ (0, 1), but at first, the ex-
ceptional set could depend on ε. Property (P) guarantees that this is not the case.
From the previous estimates, we actually have that for fixed x ∈ X, the functions of
εR ∈ (0, 1) given by X Kε (x, y) dµ(y) belongs to L1 (0, 1). Then the same is true for
R
B(x,a) Kε (x, y) dµ(y) for a > 0 fixed. So that, from the one dimensional Lebesgue
theorem of differentiation of the integral of an integrable function, we have
Z ε0 +h Z
1
Z
Kε0 (x, y) dµ(y) = lim Kε (x, y) dµ(y) dε (3.13)
B(x,a) h→0 h ε0 B(x,a)
for almost every ε0 ∈ (0, 1). Let us now get on upper estimate for a general mean on
the right hand side of (3.13) using now the quantitative form of property (P):
Z ε0 +h Z Z ε +h
1 1
Z 0
Kε (x, y) dµ(y) dε = d(x, y)−1 ϕ(ε −1 d(x, y)) dε dµ(y)
h ε0 d(x,y)<a h d(x,y)<a ε0
Z d(x,y)/ε
1 dt
Z 0
= ϕ(t) 2 dµ(y)
h d(x,y)<a d(x,y)/ε0 +h t
Z a/ε0 Z
1 dt
= ϕ(t) dµ(y) 2
h a/ε0 +h tε0 ≤d(x,y)<a t
Z a/ε0 +h Z
1 dt
+ ϕ(t) dµ(y) 2
h 0 tε0 ≤d(x,y)<t(ε0 +h) t
Z a/ε0
≤C ϕ(t)t −1 dt.
0
Hence (a) holds for almost every ε ∈ (0, 1). Notice that so far we have not used
the smoothness of ϕ. Let ε be any number in (0, 1), n ∈ N and δ > ε such that (a)
holds for δ , hence, from properties 2 and 4 we have
Z Z
d(x, y)−1 |ϕ ε −1 d(x, y) − ϕ δ −1 d(x, y) | dµ(y)

Kε (x, y) dµ(y) ≤
N −1 ≤d(x,y)<a N −1 ≤d(x,y)<a
Z
d(x, y)−1 ϕ ε −1 d(x, y) dµ(y)

+
N −1 ≤d(x,y)<a
Z a/ε
δ −ε dt
≤ CaN δ +C ϕ(t) ,
δε 0 t
since a number δ satisfying the above conditions can be chosen as close as we want
to ε we get
Z a/ε
dt
Z
Kε (x, y) dµ(y) ≤ C ϕ(t) ,
N −1 ≤d(x,y)<a 0 t
uniformly in x and N.
Let us notice that if (X, d, µ) satisfies a reverse (P) of the type
µ (B(x, r + s)) − µ (B(x, r)) ≥ Cs,
we also have that
dt
Z Z ∞
Kε (x, y) dµ(y) ≥ C ϕ(t) ,
X 0 t
for every ε ∈ (0, 1) and every x ∈ X.
3.12 Hausdorff distance convergence of ε-nets.
Let (X, d) be a compact quasi-metric space with finite Assouad dimension. For every
positive number ε, every ε-net in X is a finite, hence compact, subset of X.
Let K be the family of all compact subsets K of X. For each K ∈ K and each
δ > 0 let us consider the set
K δ ,d = K δ = { x ∈ X : d(x, K) ≤ δ }.
The Hausdorff distance on K is given by
dH (K1 , K2 ) = inf{ δ > 0 : K1 ⊂ K2δ and K2 ⊂ K1 }.
Since d is equivalent to a power of a distance ρ on X, then dH is equivalent to a

power of the usual Hausdorff distance on the compact subsets of a metric space. So
that the space (K , dH ) is complete.
[ Let Y ⊂ X be an ε-net: Y is a maximal ε-disperse subset of X. Hence X =

Bd (y, ε) = Y ε . Since of course X ε = X ⊃ Y, we have that X ⊂ Y ε and Y ⊂ X ε , so
y∈Y
that dH (Y, X) ≤ ε. In particular if { N j : j ∈ Z} is any sequence of ε-nets in X, with
ε j → 0 as j → ∞, we have that
3.13 A second view to the doubling property. Weak convergence and doubling. 133
dH (N j , X) → 0, j → ∞.
In other words, sequences of nets with mesh tending to zero converge to the space
X in the Hausdorff sense.
3.13 A second view to the doubling property. Weak convergence

and doubling.
Let (X, d) be a quasi-metric space. Take δ = ρ β ∈ ∆ (d). Let F be a σ -algebra on X

containing the ρ-balls. Let η : R+ ∈ C (R ), η ≡
+
0 → [0, 2] be a function such that η
∞
δ (x,y)
1 on [0, 1] and η ≡ 0 on [2, +∞). Then the function ηr (x, y) = η r is contin-
uous as a function of x, y and r on X×X×R+ . The first elementary result of this
section is given in the next statement.
Lemma 3.15. With (X, d), F , δ and η as before and µ a positive measure on F
we have that (X, d, µ) is a space of homogeneous type if and only if there exists a
positive and finite constant A such that the inequalities
Z Z
0< η2r (x, y) dµ(y) ≤ A ηr (x, y) dµ(y) < ∞,
hold for every x ∈ X and every r > 0.
Proof. Follows from the elementary pointwise inequalities
XBδ (x,r) (y) ≤ ηr (x, y) ≤ XBδ (x,2r) (y).
Going back to the Schwartz distribution theory introduces in Section 2.9, we

have some other basic examples of distributions in D 0 induced by measures that
are finite on balls. Given a Borel measure ν such that for every x ∈ X and r > 0,
ν (Bδ (x, r)) < ∞, define
Z
hTµ , ϕi = ϕ(y) dµ(y),
y∈X
with ϕ ∈ E α . It is easy to prove that Tµ ∈ D 0 . A special class of such test functions

ϕ is given by the family { ηr (x, .) : x ∈ X, r > 0} considered in the previous lemma.
Since D 0 is the dual space of (E α , τ) considered in Section 2.9, the convergence in
D 0 of a sequence Tk of distributions to a distribution T is given by the individual
convergence of the numerical sequence { hTµ , ϕi : k ∈ N} to hT, ϕi for every test
function ϕ. So that if { µk : k ∈ N} is a sequence of doubling measures with uni-
formly bounded doubling constants on (X, d) such that Tµk → Tµ , in D 0 as k → ∞,
with µ a non trivial measure, then µ also satisfies the doubling property. In fact
Z
η2r (x, y) dµ(y) = hTµ , η2r (x, .)i
= lim hTµk , η2r (x, .)i
k→∞
Z
= lim η2r (x, y) dµk (y)
k→∞
Z
≤ A lim ηr (x, y) dµk (y)
k→∞
= A lim hTµk , ηr (x, .)i
k→∞
= A hTµ , ηr (x, .)i
Z
=A ηr (x, y) dµ(y).
Since all the measures involved in the preceding remark are doubling on (X, d),
then the support of each µk , and of µ, is the whole space X. The support here is
understood in the distributional sense: the complementary of the largest open set
where Tµ vanishes.
3.14 Convergence of spaces of homogeneous type.
In this section we consider a generalization of the situation described in the previous

section. The basic idea is to consider a space of homogeneous type as a limit of
spaces of homogeneous type with simple structure. For example the discretization
of a continuous setting.
Let (X, d) be a quasi-metric space and δ = ρ β ' d. Let us assume that X is

compact. Let { Kn } be a sequence of compact subsets of X, { Kn } ⊂ K with the
notation of Section 3.13. Assume that Kn → K∞ as n → ∞ in the sense of Hausdorff.
3.14 Convergence of spaces of homogeneous type. 135
Suppose now that for each n ∈ N there exists a Borel measure µn on X such
that supp µn ⊆ Kn , and (Kn , d, µn ) are all spaces of homogeneous type with a finite
uniform upper bound A for the doubling constants { An }. Assume that µn tends to a
nontrivial Borel measure µ∞ on X supported on K∞ , in the weak sense considered
in the previous section.
We are going to show that (K∞ , d, µ∞ ) is a space of homogeneous type.
For x ∈ K∞ and r > 0, we need to estimate
Z Z
η2r (x, y) dµ∞ (y) = η2r (x, y) dµ∞ (y).
K∞ X
Since µn → µ∞ in the sense of distributions on (X, d) and η2r (x, .) is a good test
function, we have that choosing n ≥ N(ε) the absolute values of the differences
Z Z
η2r (x, y) dµ∞ (y) − η2r dµn (y)
X X
and
Z Z
ηr (x, y) dµ∞ (y) − ηr (x, y) dµn (y)
X X
are both bounded by ε. But even when we know that (Kn , d, µn ) is a space of homo-
geneous type, we do not know whether or not x ∈ Kn . Since dH (Kn , K∞ ) → 0 as n →
∞, by choosing n even larger if necessary, we get aZ point xn ∈ Kn such that δ (xn , x) <
ε. Now, the absolute value of the difference [ η2r (x, y) − η2r (xn , y)] dµ(y) is
X
bounded above by
1
µn (Bδ (x, 4r)) k η 0 k∞ β (4r)β −1 ε 1/β .
2r
Since µ∞ is nontrivial, we must have a finite upper bound for µn (Bδ (x, 4r)) . Hence
given x ∈ K∞ and r > 0, for any ε > 0 we get that, for n large enough
Z Z
η2r (x, y) dµ∞ (y) ≤ η2r (xn , y) dµ(y)
K∞ Kn
+ ε +C(x, r, β , η) ε 1/β ,
with xn ∈ Kn . Since (Kn , d, µn ) is a space of homogeneous typeZwith doubling con-

stant An ≤ A < ∞, the first term on the right is bounded by A ηr (xn , y) dµn (y).
Kn
Arguing as before we can go back to ηr (x, y) adding another error of the order of
ε 1/β and finally to
Z Z
A η2r (x, y) dµ∞ (y) ≤ A ηr (x, y) dµ∞ (y) +Cε 1/β ,
K∞ K∞
which proves the doubling property for µ∞ on (K∞ , d).
3.15 On doubling set functions on a quasi metric space
We shall start by exploring some elementary but deep consequences of a doubling

property on a very general set function defined on a quasi-metric space.
Let (X, d) be a quasi-metric space.
We shall say that a set function µ ∗ , defined on P(X) with non-negative values,
is monotonic if for E ⊂ F we have µ ∗ (E) ≤ µ ∗ (F). A monotonic set function µ ∗ is
said to be trivial if µ ∗ (E) = 0 for every bounded set E or if µ ∗ (G) = +∞ for every
open set G in X. We say that a monotonic set function µ ∗ is doubling if there exists
a constant A such that the inequality
µ ∗ (B(x, 2r)) ≤ Aµ ∗ (B(x, r))

holds for every x ∈ X and every r > 0.
Finally, a set function µ ∗ is said to be a metric set function if µ ∗ (E ∪ F) =
µ (E) + µ ∗ (F) whenever d(E, F) > 0.
∗
Lemma 3.16. Let µ ∗ be a monotonic doubling set function in the quasi-metric space
(X, d). Then if µ ∗ (B) = 0 or µ ∗ (B) = +∞ for some ball B, we have that µ ∗ is trivial.
Proof. Assume first that for B = B(z, s) s > 0 we have that µ ∗ (B) = 0. If E is a
bounded set in (X, d), E ⊂ B(x, r) for some x ∈ X and r > 0. Thus E ⊂ B(x, r) ⊂
B(z, k(r + d(x, z))) ⊂ B(z, 2 j s), where j ∈ Z+ is chosen in such a way that K(r +
d(x, z)) ≤ 2 j s. Hence, by iteration of the doubling and monotonicity of µ ∗ we get
µ ∗ (E) ≤ µ ∗ (B(z, 2 j s)) ≤ A j µ ∗ (B) = 0.

3.15 On doubling set functions on a quasi metric space 137
which proves that µ ∗ is trivial.

Assume now that µ ∗ (B) = +∞ with B = B(z, s). If G is an open set in X, then
G ⊃ B(x, r) for some x ∈ G and some r > 0. We only have to show that µ ∗ (B(x, r)) =
+∞. Now B(z, s) ⊂ B(x, K(s + d(x, z))) ⊂ B(x, 2m r) for some m ∈ Z+ . Hence +∞ =
µ ∗ (B) ≤ Am µ ∗ (B(x, r)) and µ ∗ (B(x, r)) = +∞.
As a consequence of the above lemma we have that a nontrivial doubling and

monotonic set function µ ∗ satisfies the inequalities
0 < µ ∗ (B(x, 2r)) ≤ Aµ ∗ (B(x, r)) < ∞

for every x ∈ X and every r > 0.
In the next result we prove that the existence of a nontrivial, monotonic, doubling
and metric set function µ ∗ on X, d suffices for the weak homogeneity property or
finiteness of the Assouad metric dimension.
Theorem 3.9. Let (X, d) be a quasi-metric space. If P(X) supports a non-trivial,

monotonic, doubling and metric set function, then (X, d) has finite metric dimension.
Proof. Let ρ be a distance in X such that for some β > 0 we have that d ' ρ β .
Precisely, there exist two constants 0 < c1 ≤ c2 < ∞ such that
c1 d ≤ ρ β ≤ c2 d
on X × X.
From Lemma (2.?) in Chapter ??, in order to prove that (X, d) has the weak
homogeneity property, we only have to prove that (X, ρ) has that property. More-
over, Theorem 3.9 reduces to the metric case since also the hypotheses are in-
variant by changing d by ρ. In fact, notice first that d(E, F) > 0 if and only if
β
ρ(E, F) > 0. On the other hand it is easy to see that Bρ (x, 2r) ⊆ B(x, (2r)
c1 ) and that
β
B(x, (r)
c2 ) ⊆ Bρ (x, r). So that
(2)β c2 rβ
µ ∗ Bρ (x, 2r) ≤ µ ∗ B(x, · )
c1 c2
r β
≤ Am µ ∗ (B(x, ))
c2
≤ Am µ ∗ (Bρ (x, r)),
where m is the smallest integer for which 2β c1 ≤ c2 2m .

Let us, then, prove the theorem assuming that d is a metric on X. Let D be an
r
2 -disperse subset of B(x0 , r). In other words D ⊂ B(x0 , r) and d(x, y) ≥ 2r for every
choice of x 6= y with x and y in D. The following statements are easy consequences
of the triangle inequality (K = 1),
(a) d(B(x, 8r ), B(y, 8r )) ≥ r
4 for every x and y ∈ D with x 6= y;
(b) B(x, 8r ) ⊆ B(x0 , 89 r), x ∈ D;

(c) B(x0 , 98 r) ⊆ B(x, 178 r), x ∈ D.
Let F be a finite subset of D. Using (c), the doubling property of µ ∗ , (a), the
metric property of µ ∗ and (b) we get
9 9
](F)A−s µ ∗ (B(x0 , r)) = ∑ A−s µ ∗ (B(x0 , r))
8 x∈F 8
17
≤ ∑ A−s µ ∗ (B(x, 8
r))
x∈F
r
≤ ∑ µ ∗ (B(x, 8 ))
x∈F
r
= µ ∗ ∪x∈F B(x, )
8
∗ 9
≤ µ (B(x0 , r)).
8
Being µ ∗ non-trivial, µ ∗ (B(x0 , 98 r)) is a positive real number, hence ](F) ≤ A5

for every finite subset F of D. Then ](D) ≤ A5 , as desired.
The most standard and classical case of the above general situation is provided
by the Euclidean space Rn equipped with the usual distance d(x, y) = |x − y| and the
outer Lebesgue measure µ ∗ . In this case A = 2n is the best possible constant for the
doubling property. Actualy, since the balls are open sets the doubling property can
be started in terms of the Lebesgue measure µ (the restriction of µ ∗ to the Lebesgue
measurable sets) as
µ(B(x, 2r)) = 2n µ(B(x, r)),

for every x ∈ Rn and every r > 0.
As a second example let us consider the space (X, d) where X = Q is the set
of rational numbers and d is the usual distance; d(q1 , q2 ) = |q1 − q2 |. Let us con-
sider the following set function defined on the family P(Q) of all the subsets of Q
by µ ∗ (E) = |E|, where E is the closure in R of E and |A| is the one dimensional
Lebesgue measure of the measurable set A. It is easy to check that µ ∗ is a non-
trivial, monotonic, doubling and metric set function on (Q, d). But in (Q, d) it is
impossible to obtain a Borel, non-trivial and doubling measure. In fact, since Q is
the disjoint union of its countable singletons {qn : n ∈ Z+ } and we are looking for
a non-trivial measure µ, some {qn } has to have positive measure. In other words
1
α = µ({qn }) > 0. But since qn ∈ 4(qn + 2k+1 , qn + 21k ) ∩ Q = B(qn + 2k+2
3 4
, 2k+2 ), we
3 4
have that µ(B(qn + 2k+2 , 2k+2 )) ≥ α > 0. Hence if such a doubling measure exists
we should have that
1 1
µ((qn + , qn + k ) ∩ Q) ≥ A−2 α,
2k+1 2
3.16 Some general geometric consequences of the weak homogeneity property 139
for every k ∈ Z+ . But
1
µ(B(qn + )) = µ((qn , qn + 1) ∩ Q)
2
∞
1 1
≥ ∑ µ((qn + k+1 , qn + k ) ∩ Q) = +∞.
k=0 2 2
Which is a contradiction.
3.16 Some general geometric consequences of the weak

homogeneity property
We shall say that a quasi-metric space (X, d) satisfies the weak homogeneity prop-
erty if on P(X) there exists a non-negative, non-trivial, monotonic, doubling and
metric set function µ ∗ . A point x0 in X is called an atom if µ ∗ ({x0 }) > 0. As usual
in a metric setting, a point x is said to be isolated if for some ε > 0 we have that
x = B(x, ε). Since µ ∗ in non-trivial and isolated points are balls, it is easy to see that
isolated points are atoms. The next result proves the converse.
Theorem 3.10. Let (X, d) be a quasi-metric space with the w.k.p.. Then atoms are
isolated.
Proof. Assume that µ ∗ ({x0 }) > 0. If X = {x0 }, then there is nothing to proof. As-
sume that X 6= {x0 } and that x0 is not isolated. Fix ρ ∈ R(d). We shall inductively
build a sequence of points xk and two sequences of ρ-balls {Bk0 } and {Bk1 } satisfying
(1) Bk0 is centered at x0 and Bk1 at xk ;
(2) Bk1 ∩ Bk0 = 0;/
k+1 k
(3) B1 ⊆ B0 ;
(4) x0 ∈ 3Bk1 ;
(5) Bk1 ⊂ Bρ (x0 , 2ρ(x1 , x0 ));
(6) ρ(Bk+1 k
1 , B1 ) > 0.
In fact, choose any x1 6= x0 , B11 = Bρ (x1 , ρ(x12,x0 ) ) and B10 = Bρ (x0 , ρ(x12,x0 ) ). It is
clear that these balls satisfy (1), (2), (4) and (5).
Since x0 is not isolated we can choose x2 6= x0 with x2 ∈ Bρ (x0 , ρ(x14,x0 ) ). It is
now easy to check that (3) and (6) with k = 1 are satisfied. Inductively we get the
sequences {xk }, {Bko } and {Bk1 } satisfying properties (1) to (6).
From the doubling property, the monotonicity of µ ∗ and (4) we have that
µ ∗ (Bk1 ) ≥ A−2 µ ∗ (3Bk1 ) ≥ A−2 µ ∗ ({x0 }).

Now, from (6) and the metric property of µ ∗ we have that for each M ∈ N
M
µ ∗ ∪M k
∑ µ ∗ (Bk1 )

k=1 (B1 ) =
k=1
−2 ∗
≥A µ ({x0 })M.
Hence, from monotonicity of µ ∗ and property (5) of the sequence Bk1 ,
µ ∗ Bρ (x0 , 2ρ(x1 , x0 )) ≥ A−2 µ ∗ ({x0 })M,

for every M ∈ N. So that µ ∗ is trivial, a contradiction.
Proposition 3.1. Let (X, d) be a quasi-metric space satisfying the weak homogene-
ity property. Then the set A of all atoms is at most countable.
Proof. From Theorem 3.10 we know that (X, d) has finite metric Assouad dimen-
sion. Hence (X, d) is separable. In other words, there exists in X a subset D which
is dense and contable. Since each point in A is isolated, necessarily D ⊇ A. Hence A
is a most countable.
There is still another metric property in the weak homogeneity property. A subset
E of a quasi-metric space (X, d) is said to be bounded if for some x0 ∈ X and some
R > 0 we have that E ⊆ B(x0 , R). If µ ∗ is a set function for the weak homogeneity
we have that µ ∗ (E) < ∞ for every bounded set E. In particular if the whole space X
is bounded, then µ ∗ (X) < ∞. The next result shows how a converse of this property
follows from the doubling of µ ∗ .
Theorem 3.11. Let (X, d) be a quasi-metric space with the weak homogeneity prop-
erty. Let µ ∗ be any set function associate to the w.h.p. for (X, d). If µ ∗ (X) < ∞, then
X is bounded.
Proof. It is easy to see that the problem reduces to the metric case. Assume then
that d is a distance on X. If X is unbounded then for each ball B0 in X there exists
B1 such that:
(a) B1 ∩ B0 = 0;
/
(b) B0 ⊆ 2B1 = B(c(B1 ); 2r(B1 ));
(c) B1 ⊆ B(C(B0 ), 3r(B1 )) =: B2 ,
where C(B) is the center of B and r(B) is its radious. In fact, take C(B1 ) 6∈ 2B0 and
r(B1 ) = 21 d(C(B1 ),C(B0 )). By iteration of the previous basic construction we shall
prove that if X is unbounded, then µ ∗ (X) = +∞. Let us take a fixed ball B0 in X,
say B0 = B(x0 , 1). Set B1 and B2 to denote the balls provided by the basic algorithm
when we start with B0 = B(x0 , 1). Now let us write B3 and B4 to denote the balls
provided by that basic .....???? we take B2 in placed of B0 . In other words
(a’) B3 ∩ B2 = 0;
/
(b’) B2 ⊆ 2B3 ;
(c’) B1 ⊆ B(C(B0 ), 3r(B3 )) =: B4 .
Iteration gives rice to a sequence of balls
B0 , B1 , B2 , B3 , B4 , . . .
such that the add indexed balls are pairwise disjoint, moreover they are separated in
the metric sense and 2B2k+1 ⊃ B0 for each integer k ≥ 0. Hence, for each M ∈ N we
have that
M
µ ∗ (X) ≥ µ ∗ ∪M ∑ µ ∗ (B2k+1 )

k=0 B2k+1 =
k=0
−1 ∗
≥A µ (B0 )M.
So that µ ∗ (X) = +∞.
3.17 Spaces of homogeneous type
Given a quasi-metric space (X, d), we say that a Borel measure µ on X is doubling
if for some distance ρ for which d ∼ ρ β for some β > 0, and for some constant
A > 0 we have the inequalities
0 < µ(Bρ (x, 2r)) ≤ Aµ(Bρ (x, r)) < ∞ (3.14)

for every x ∈ X and r > 0.
Let us write R(d) to denote the class of all distances on X for which ρ β ∼ d for
some β positive.
Lemma 3.17. A Borel measure µ on the quasi-metric space (X, d) is doubling if
and only if for each ρ ∈ R(d) and each m > 0 there exists a constant A(ρ, m) such
that (∗, ∗) holds with ρ m instead of ρ and A(ρ, m) instead of A.
An elementary example is again provided by the Euclidean case. If in Rn we
take ρ(x, y) = |x − y| and m = n, we get that the Lebesgue measure of any ρ m ball
of radius r > 0 is given by wn r, where wn is the volume of the unit ball. hence the
doubling property for the ρ m balls holds with A = 2.
We say that (X, d, µ) is a space of homogeneous type if (X, d) is a quasi-metric
space and µ is a doubling measure on a σ -algebra Σ of subsets of X containing
the topology τd induced by d. Notice that since Σ ⊃ τd , every Borel subset of X is
measurable.
As a result of the consideration of the two previous sections we have the follow-
ing basic properties of spaces of homogeneous type.
Theorem 3.12. Let (X, d, µ) be a space of homogeneous type. Then
(1) (X, d) has the weak homogeneity property;

(2) (X, d) has finite metric Assouad dimension;
(3) (X, d) has finite metric Hausdorff dimension;
(4) X is bounded if and only if µ(X) < ∞;
(5) X is isolated if and only if µ({x}) > 0.
Proof. All we have to do is to show that there exists on P(X) a positive, non-trivial,
monotonic, doubling and metric set function µ ∗ . Define
µ ∗ (E) = inf{µ(A) : A ∈ Σ and A ⊃ E}

for E ⊆ X. Notice that µ ∗ when restricted to Σ coincides with µ. It is clear that µ ∗
is positive and non-decreasing. Since for ρ ∈ R(d) any ρ-ball has finite and positive
measure we see that µ ∗ (E) < ∞ for every bounded set E and µ ∗ (G) > 0 for every
open set G. In other words, µ ∗ is non-trivial. The doubling for µ ∗ follows from the
doubling for µ on ρ-balls with ρ ∈ R(d).
Let us observe that the inequality µ ∗ (E ∪ F) ≤ µ ∗ (E) + µ ∗ ( f ) is immediate.
There is nothing to prove if µ ∗ (E) = +∞ or µ ∗ (F) = +∞. Assume then that both
are finite. Then, given ε > 0, there exist two measurable sets A1 and A2 such that
µ(A1 ) < µ ∗ (E) + ε

and
µ(A2 ) < µ ∗ (F) + ε

with E ⊆ A1 and F ⊆ A2 . Hence E ∪ F ⊆ A1 ∪ A2 , so that
µ ∗ (E ∪ F) ≤ µ(A1 ∪ A2 ) ≤ µ(A1 ) + µ(A2 )

≤ µ ∗ (E) + µ ∗ (F) + 2ε.
It only remains to be proved that if d(E, F) > 0, then µ ∗ (E ∪ F) ≥ µ ∗ (E) +

µ ∗ (F). Notice first that d(E, F) > 0 is equivalent to ρ(E, F) > 0 for any ρ ∈ R(d).
It is enough to prove that for every A ∈ Σ such that A ⊃ E ∪ F we have
µ(A) ≥ µ ∗ (E) + µ ∗ (F).

Notice that E ⊂ {y : ρ(E, y) < ρ(E,F)
2 } = G1 and that F ⊂ {y : ρ(E, y) >
ρ(E,F)
2 }=
G2 . Both, G1 and G2 are disjoint open set. So that A ∩ G1 and A ∩ G2 belong to Σ .
Moreover A ∩ G1 ⊃ E and A ∩ G2 ⊃ F. Then
µ(A) ≥ µ(A ∩ G1 ) + µ(A ∩ G2 )

≥ µ ∗ (E) + µ ∗ (F).
Chapter 4
Some examples of quasi-metric measure spaces.
In this chapter we shall consider examples of spaces of homogeneous type and some
related quasi-metric measure space. Aside from the very basic examples already
introduced in the previous chapters, we shall try here to emphasize the point of view
of mathematical modelling of somehow concrete situations through the structure of
quasi-metric measure spaces.
In the first section of the chapter we shall review some basic, well known or clas-
sical cases in which doubling measures on quasi-metric spaces provide structures of
spaces of homogeneous type. Some of them shall share as the underlying set X the
euclidian space. Sometimes the metric structure happens to be invariant under the
group translations, sometimes the measure is invariant but generally none of these
structures will be strongly related to the algebraic. In the second section we shall
introduce some models associated to elliptic and parabolic type partial differential
equations from the point of view of Harnack inequalities and regularity theory for
weak solutions. The third section is devoted to consider a model for the real analysis
associated to the Monge-Ampere equation.
In section 4 we shall consider examples coming from geometric measure theory
such as surfaces, curves, fractals and more general sets.
4.1 Some elementary and classical examples.
4.1.1 Some basic translation invariant structures on Rn .
Let X = Rn = {x = (x1 , · · · , xn ) : xi ∈ R} be the Euclidean space. Let us write |x|∞

to denote the max norm on X, i.e., |x|∞ = max |xi |. The ball B∞ (x, r) with center at
1≤i≤n
x ∈ Rn and radius r > 0 with respect to the distance d(x, y) = |x − y|∞ is the cube
B∞ (x, r) = {y ∈ Rn : |xi − yi | < r}. If µ is used to denote Lebesgue measure on the
Lebesgue measurable subsets of Rn , then we have that µ(B∞ (x, r)) = 2n rn . Since on
Rn all the norms are equivalent we have that (Rn , d, µ) is a space of homogeneous
143
144 4 Some examples of quasi-metric measure spaces.
type for d(x, y) = kx − yk with k.k a norm on Rn . Moreover, the measure of any
ball with respect to any norm on Rn is of the order of rn with constants depending
only on n and on the norm defining the family of balls. Hence (Rn , d, µ) becomes a
normal space if we take d(x, y) = kx − ykn , with k.k a norm on Rn . From the point of
view of the distribution of the measure of balls, this normalization has even stronger
properties, since µ(B(x, r)) is actually a linear function
! of r independent of x ∈ Rn .
n
A particular case is given by Rn , ∑ |xi − yi |n , µ .
i=1
Notice that the topology generated by any of these quasi-distances is the usual
one which certainly is contained in the σ -algebra of µ-measurable subsets of Rn .
Actually the convexity of the balls provided by the norms is not needed and we can
change the norm by a distance of the form d p (x, y) =
n
= ∑ |xi − yi | p for p > 0. Since neither isotropy is needed, another family of quasi-
i=1
distances providing space of homogeneous type structure to Rn with Lebesgue mea-
n
sure is given by d(x, y) = ∑ |x j − y j |α j . Even when all these quasi-distances are
j=1
equivalent form the topological point of view, they are generally not equivalent from
the analytical point of view in the sense that dd12 need not to be a bounded function
when d1 and d2 are two such quasi-distances on Rn .
4.1.2 More translations invariant structures generalized

homogeneity.
The natural homogeneity associated to a parabolic partial differential equation such

2
as the heat equation, ∂∂tu = ∂∂ xu2 in one space dimension, is reflected in the invariance
of the equation changing the variables u(x,t) −→ u(λ x, λ 2t) or in the restrictions for
the convergence of finite diference solutions to exact solutions: ∆∆tx2 ≤ 12 . The func-
p
tion d((x,t), (y, s)) = max{|x − a|, |t − s|} defined in R2 is a traslation invariant
distance that is not obtained from a norm. Instead d satisfies
d(Tλ (x,t), Tλ (y, s)) = λ d((x,t), (y, s)),

where
2 λ 0
Tλ (x,t) = (λ x, λ t) =
0 λ2
Since the d-ball with center (0, 0) and radius r > 0 is the rectangle
(−r, r) × (−r2 , r2 ), we have that its Lebesgue measure is given by 4r3 . Hence
(R2 , d, µ) is a space of homogeneous type if µ is Lebesgue measure. It is worth
noticing that the Hausdorff dimension of (R2 , d) is three.
4.1 Some elementary and classical examples. 145
Generalized homogeneity can take a more general form in the n-dimensional eu-
clidean space through the nonisotropic dilations induced by an n × n matrix Agiven
A log λ 10
by Tλ = e for λ > 0. For the example above we have n = 2, A = and
02

λ 0
Tλ = , λ > 0 for the associated non- isotropic dilations of the plane. The
0 λ2
set R − {0} can be covered by the parabolic parametric curves Tλ x = (λ x1 , λ 2 x2 )
2
with λ as the parameter and x = (x1 , x2 ) ∈ R2 − {0}. Each of these curves intersects
the boundary of the square (−1, 1) × (−1, 1) = Q at one and only one point. So that,
for each x = (x1 , x2 ) 6= 0 there exists one and only one positive number, that we call
ρ∞ (x), such that T1/ρ∞ (x) x is the only point in the curve {Tλ x : λ > 0} that belongs
to ∂ Q. Since to find ρ∞ (x) we have to solve the equation |T1/λ (x)|∞ = 1, it is easy
p
to check that ρ∞ (x) = max{|x1 |, |x2 |}, which gives the distance d considered as a
model for the heat equation.
In general given an n × n matrix A and a balanced, open, bounded and convex set
C in Rn such that each curve of the family {Tλ x = eA log λ x : x ∈ Rn \{0}} intersects
∂C at one and only one a function ρC (x) can be constructed exactly as we did
point,
10
for C = Q and A = . Actually the properties on C are used to think of this
02
intersection as the solution of kT 1 xk = 1 for the norm k k whose unit ball is C.
λ
A frecuently used special case is to consider a positive definite diagonalizable

matrix A with positive eigenvalues A1 , a2 , · · · , an and C to be the unit euclidean ball.
In this situation the generalized dilations are given by
 a 
λ 1 0 ··· 0
 0 λ a2 · · · 0 
Tλ =  ··· ··· ··· ··· 

0 0 · · · λ an
and the distance structure is induced by the only λ > 0 that solves the equation
x12 x22 xn2

+ + · · · + = 1.
λ 2a1 λ 2a2 λ 2an
This distance turns out to be equivalent to the one obtained by substituting the unit
standard ball by the unit cube, that takes the explicit form.
d(x, y) = max |xi − yi |1/ai .

i=1,2,··· ,n
Since any d-ball of radius r > 0 is just a parallelepiped with sides of lengths
n
∑ ai
2ra1 , 2ra2 , · · · , 2ran , the Lebesgue measure of Bd (x, r) is just 2n ri=1 = 2n rtrace(A) .
Hence (Rn , d, µ) is a space of homogeneous type with µ being Lebesgue measure.

Moreover the Hausdorff dimension and the Assouad dimension of (Rn , d) is the
trace of the matrix A.
Notice also that d is a distance if and only if ai ≥ 1 for every i = 1, 2, · · · , n.

Generally d is a quasi-distance with triangular constant depending on the size of the
n
minimum ai . Observe also that d is equivalent to the quasi-distance ∑ |x j − y j |αj
j=1
1
introduced in 4.1.1 with α j = aj .
4.1.3 Regular Riviére- Vitali families.
In [?] N. Riviére introduced families of neighborhoods of the identity of a group,

that we shall take to be Rn , satisfying certain properties that suffice for the analysis
of singular integral operators. In this section we shall show that such regular Vitali
families of neighborhoods are essentially the d-balls for an appropiate translation
invariant quasi-distance on Rn . A Vitali family in Rn is a family {Uα : α > 0} of
subsets of Rn satisfying
i) each Uα is an open set with Uα compact,
ii) Uα ⊂ Uβ for α < β ;
\ \
iii) Uα = Uα = {0}; for each α > 0,
α>0 α>0
iv) |Uα −Uα | ≤ A|Uα | for some constant A, where U −V = {x − y : x ∈ U and y ∈ V }
and |E| denotes the Lebesgue measure of E;
v) |Uα | is a left continuous function of α > 0;
vi) Uα = −Uα for every α > 0.
A Vitali family {Uα : α > 0} is said to be regular if there exists a continuous function
φ : R+ → R+ satisfying
vii) Uα −Uα ⊆ Uφ (α) for every α > 0;
viii) |Uφ (α) | ≤ A|Uα |, α > 0.
For a given regular Vitali family {Uα : α > 0} in Rn , let us define ρ(x) = inf{|Uα | : x ∈ Uα }
and d(x, y) = ρ(x − y). Notice that if B(0, ε) is an euclidian ball contained in Uα ,
from (vii) we have that B(0, 2ε) ⊆ Uα −Uα ⊆ ⊆ UΦ (α). So that, by iteration, we
obtain B(0,
[
2k ε) ⊆ UΦk (α), where Φ k (α) = (Φ[◦ · · · Φ)(α) is the composition of Φ k times.
Since B(0, 2k ε) = Rn , we also have that Uα = Rn . This remark shows that
k∈N α>0
ρ(x) is a non-negative real number for each x ∈ Rn .
From property (vi) of the family {Uα : α > 0} we see that d is symmetric.
\
On the other hand, U 1/n = {0} and, since each U α has finite measure, we have
n∈N
that
lim |Uα | = lim |Un | = 0.

α→0 n→∞
This observation shows us that ρ(0) = 0 and that d(x, x) = 0. On the other hand,
if ρ(x) = 0, hence x = 0 and d is faithfull. Let us show that (viii) implies a quasi-
triangular inequality for d. In fact, take x and y to be two points in Rn . Let ε > 0 be
given. Then there exists α > 0 and β > 0 such that x ∈ Uα , y ∈ Uβ , |Uα | < ρ(x) + ε
and |Uβ | < ρ(y) + ε. So that x + y ∈ Uα +Uβ = Uα −Uβ ⊆ Uγ −Uγ ⊆ UΦ(γ) , where
γ = max{α, β }. This proves that

ρ(x + y) ≤ |UΦ(γ) | ≤ A|Uγ | ≤ A |Uα | +Uβ ≤ A (ρ(x) + ρ(y)) + 2εA,
from which the quasi-triangle inequality holds with K = A for d. This fact shows us
that (Rn , d) is a quasi-metric space.
Since d(x, y) = ρ(x − y), to obtain the d-ball with center x ∈ Rn and radius r > 0,
we only have to translate to x the d-ball
[ with center 0 and radius r. Notice that
Bd (0, r) = {y ∈ Rn : ρ(y) < r} = Uα and that |Bd (0, r)| = lim |Uα |,
{α:|Uα |<r} α→α0− (r)
where α0 (r) = sup{α : |Uα | < r}. Hence |Bd (0, r)| ≤ r. Let us now obtain a lower
bound for the measure of Bd (0, r). Notice that Φ(α) > α for every α > 0, since
Uα − Uα ⊆ UΦ(α) and Uα − Uα is strictly larger than Uα itself. Since Φ is as-
sumed to be continuous, for some positive ε we have Φ(α0 − ε) > α0 , then
|Bd (0, r)| ≥ |Uα0 − ε| ≥ A−1 |UΦ(α0 −ε) |. Now, since Φ(α0 − ε) > α0 , we have that
Φ(α0 − ε) ∈ / {α : |Uα | < r} and |UΦ(α0 −ε) | ≥ r. Hence (Rn , d, Lebesgue) is a nor-
mal space of homogeneous type.
4.1.4 The Heinsenberg group.
The basic structure of the Heinsenberg group has tis origin in the theory of oper-
ators in quantum mechanics. Perhaps the most elementary way of introducing the
Heinsenberg group is to consider a special subset
  
 1 x1 x3 
H =  0 1 x2  : xi ∈ R, i = 1, 2, 3
0 0 1
 
 real 3 ×3 matrixes

of  the standard matrix product. Notice that if A =
equiped with
1 x1 x3 1 y1 y3
 0 1 x2  and B =  0 1 y2 , then AB belongs to H since
0 0 1 0 0 1
 
1 x1 + y1 x3 + y3 + x1 y2
AB =  0 1 x2 + y2 .
0 0 1
Since the objects of H have actually three degrees of freedom, we can and will
identify H with R3 under the application T : R3 → H , defined by
1 x1 x3 + x12x2
 
T (x1 , x2 , x3 ) =  0 1 x2 .
0 0 1
Its inverse T −1 : H → R3 is given by
 
1 y1 y3 y1 y2
T −1 =  0 1 y2  = y1 , y2 , y3 − .
2
0 0 1
Now, we define the Heisenberg product ∗ on R3 in such a way that T becomes
a group isomorphism between (R3 , ∗) and H with the usual matrix product. This
leads to

x1 y2 − y1 x2
x ∗ y = (x1 , x2 , x3 ) ∗ (y1 , y2 , y3 ) = x1 + y1 , x2 + y2 , x3 + y3 +
2
The basic “homogeneity” of this group structure on R3 is quite different from the
standard one. In fact, if δλ (x) = (λ x1 , λ x2 , λ 2 x3 ) for λ > 0, we have that δλ (x ∗ y) =
δλ (x) ∗ δλ (y). As in the parabolic case, this fact suggests that a good distance for
H should be given, after ∗ invariance, by
p
ρ(x) = ρ(x1 , x2 , x3 ) = |x1 | + |x2 | + |x3 |.
In other words, we consider in R3 the distance
d(x, y) = ρ(x ∗ y−1 ),

where y−1 = −y is the inverse of y with respect to ∗.
Let y be a fixed point in R3 and let E be a Borel measurable subset of R3 . The

set
E ∗ y = {x ∗ y : x ∈ E}

x1 y2 − y1 x2
= Sy (x) = x1 + y1 , x2 + y2 , x3 + y3 + :x∈E
2
is also a Borel set in R3 , and

Z
|E ∗ y| = |Sy (E)| = |J Sy (x)| dx,
E
where J Sy is the Jacobian of Sy . But
 
1 0 0
J Sy =  0 1 0 ,
y2
2 − y21 1
hence |E ∗ y| = |E|. Also |y ∗ E| = |E|. In other words, for (H , ∗) Lebesgue mea-
sure is Haar measure. Let us estimate the measure of a d-ball. From invariance
we can assume that the p given ball is centered at 0. We can also change ρ to
ρ∞ (x) = max{|x1 |, |x2 |, |x3 |} since they are equivalent but the ρ∞ balls have a sim-
pler description in rectangular coordinates: B∞ (0, r) = (−r, r) × (−r, r) × (−r2 , r2 ).
Hence |B∞ (0, r)| = 8r4 , and (H , d, |.|) is a space of homogeneous type which can
be normalized by changing the distance d by the quasi-distance d 4 .
4.1.5 A dyadic distance on Euclidian Spaces.
n 1) be the unit interval in R. For oj = 0, 1, 2, · · · , let us consider the fam-

Let X = [0,
ily D j = Ikj = 2kj , k+1
2j
: k = 0, · · · , 2 j − 1 of all dyadic intervals of j-level. Set
∞
D= D j to denote the family of all dyadic intervals of [0, 1). Given two points
[
j=0
x and y in [0, 1), define d(x, y) as the length of the smallest dyadic interval contain-
ing both, x andy. In other words d(x, y) = inf{|I| : I ∈ D and x, y ∈ I}. Notice that
d 21 − ε, 12 + ε = 1 for every ε > 0, hence d is far away from being equivalent to
the usual distance in [0, 1). Notice that d(x, y) = d(y, x). Since every point x ∈ X
belongs to one of the intervals of each level, d(x, x) = 0. On the other hand from the
nesting property of dyadic intervals we have that d(x, y) = 0 implies x = y. Take now
x, y, z three different points in X. Suppose that I ∈ D is the smallest interval con-
taining x and y and that J is the smallest dyadic interval containing y and z. Hence
d(x, y) = |I| and d(y, z) = |J|. Since y ∈ I ∩J, then I ⊆ J or J ⊆ I, from the basic prop-
erties of the family D. In the first case x also belongs to J and d(x, z) ≤ |J| = d(y, z).
In the second case z also belongs to I and d(x, z) ≤ |I| = d(x, y). In other words,
d(x, z) ≤ max{d(x, y), d(y, z)}.
Given x ∈ X and r > 0, the d-ball centered at x with radius r is given by
Bd (x, r) = {y ∈ X : min{|I| : I ∈ D, x, y ∈ I} < r}

[
= I = I(x, r),
|I|<r, x∈I
where I(x, r) belongs to D and is the largest possible dyadic interval containing x
with length < r. If j ∈ N ∪ {0} is such that 2− j < r ≤ 2− j+1 , then, the length of
I(x, r) = Bd (x, r) is precisely 2− j . Hence (X, d, µ) is a space of homogeneous type
with µ equal Lebesgue measure, moreover it is normal, since
r
≤ µ(Bd (x, r)) = 2− j < r.
2
The basic scheme of this example can be extended to more general situations.
See................
4.2 Some examples from degenerate elliptic and parabolic

equations.
4.2.1 Basic A p theory
In this subsection we sketch those properties of Muckenhoupt weights which are

relevant for our purpose of modelling degenerate elliptic and parabolic equations
from the point of view of Harmack’s inequalities.
Let (X, d, µ) be a space of homogeneous type. Let 1 < p < ∞. A non-negative

locally integrable function w defined on X is said to belong to A p = A p (X) =
A p (X, d, µ) if there is a constant C for which the inequality
Z Z p−1
1 p
− p−1
w dµ w dµ ≤ C(µ(B)) p ,
B B
holds for every d-ball B of X. Let us notice that, from Hölder inequality, the opposite
inequality
Z 1 Z p−1
p p
Z
1 − 1p −1p−1
µ(B) = w w p dµ ≤ w dµ w dµ
B B B
is always true. So that the weight w satisfies A p (X) if and only if

Z Z p−1
1
w dµ w− p−1 dµ ∼
= (µ(B)) p ,
B B
where the equivalence constant may depend on w but not on B.
The fundamental results concerning A p classes are those related the boundedness
of the classical operators of harmonic analysis. We shall deal with these problems
4.2 Some examples from degenerate elliptic and parabolic equations. 151
later on. Now we are only

Z interested in somer very basic results of the theory. Let
us write w(E) to denote w dµ for any Borel set E.
E
Lemma 4.1. Let 1 < p < ∞ and w ∈ A p (X). Then, there exists a constant C such
that the inequality
w(E) 1/p

µ(E)
≤
µ(B) w(B)
holds for every d-ball B of X and every measurable subset E of B.
Proof. The desired inequality follows from Hölder inequality and Muckenhoupt
condition in the following way
Z p−1
p
Z
1 1 1
µ(E) = w p w− p dµ ≤ (w(E))1/p w− p−1 dµ
E B
1/p
Cµ(B)
≤ (w(E))1/p R .
B w dµ
As a corollary of the above lemma, we have that (X, d, w dµ) is also a space of
homogeneous type. In fact, we have only to check that the d-balls of X satisfy the
doubling property with respect to the new weighted measure dν = wdµ. Let x ∈ X
and r > 0 be given. Apply the previous lemma with B = B(x, 2r) and E = B(x, r),
hence
µ(B(x, r)) p

ν(B(x, r))
≤ Cp .
µ(B(x, 2r)) ν(B(x, 2r))
Since µ itself is doubling with doubling constant A, we get that ν is doubling with
constant A pC p .
A particular case of the above situation is obtained by taking (X, d, µ) = (Rn , |. −

.|, Lebesgue). Hence, if w is any A p (Rn ) weight, we have that (Rn , |. − .|, wdx) is
a space of homogeneous type with a generally nontranslation invariant measure
dν = wdx. In this special case, in which the underlying space is the Euclidian when
(X, d, µ) = (Rn , |. − .|, Lebesgue ). Hence, if w is any A p (Rn ) weight, we have that
(Rn , |. − .|, w dx) is a space of homogeneous type with a generally nontranslation in-
variant measure dν = w dx. In this special case, in which the underlying space is the
Euclidian one, we have an actual growth of the w dx measure of balls as functions
of radius (See also section 3.10).
Lemma 4.2. If µ is a doubling Borel measure on Rn , then there exists γ > 1 such
that
µ(B(x, 2r)) ≥ γ µ(B(x, r)

for every x ∈ Rn and every r > 0
Proof. Take y ∈ Rn in such a way that |x − y| = 32 r. Then

(i) B(y, 2r ) ⊂ B(x, 2r);
(ii) B(x, r) ⊂ B(y, 52 r) ⊂ B(y, 8 2r );
(iii) B(x, r) ∩ B(y, 2r ) = 0;
/
Hence
r
µ(B(x, 2r)) ≥ µ(B(x, r)) + µ B y,
2 r
−3
≥ µ(B(x, r)) + A µ B y, 8
2
≥ (1 + A−3 )µ(B(x, r)),
and the result follows with γ = 1 + A−3 > 1.

[
The first class of non-trivial weights belonging to A∞ (Rn ) = A p (Rn ) is
p>1
ωα (x) = |x|α with α > −n. Actually ωα ∈ A p (Rn if and only if −n < α < n(p − 1).
This fact can be generalized to spaces of homogeneous type.
Lemma 4.3. Let (X, d, µ) be a normal space of homogeneous type with no atoms.
Then w(x) = d α (x, x0 ) belongs to A p (X, d, µ) for −1 < α < p − 1 and every x0 ∈ X.
Proof. Under the current assumptions on (X, d, µ) we have that A1 r ≤

≤ µ(B(x, r)) ≤ A2 r for 0 < r ≤ K1 µ(x), where A1 , A2 and K1 are constants. Let B =
B(x, r) be a given ball. Let us assume first that x, x0 and r are related by d(x, x0 ) ≤
Mr. Then B ⊂ B(x0 , K(M + 1)r) and
Z Z
w dµ ≤ d α (x, x0 )dµ(x)
B B(x0 ,K(M+1)r)
∞ Z
= ∑ d α (x, x0 )dµ(x)
−j − j−1 r)
j=0 B(x0 ,K(M+1)2 r)−B(x0 ,K(M+1)2
≤ Crα ∑ 2α j µ(B(x0 , 2− j r))

j=0
≤ Crα ∑ 2−α j 2− j r
j=0
1+α
≤ Cr ,
since α > −1. With the same type of estimates we obtain that
Z Z
1 α
w− p−1 dµ ≤ d − p−1 (x, x0 )dµ(x)
B B(x0 ,K(M+1)r)
α ∞ αj
− p−1 p−1 − j
≤ Cr ∑2 2 r
j=0
p−1−α
≤ Cr p−1 ,
since α < p − 1. So that
Z Z p−1
1
− p−1
wdµ w dµ ≤ Cr1+α r p−1−α
B B
= Cr p
≤ Cµ(B) p ,
if 0 < r ≤ K1 µ(X). Notice that if, on the other hand r > K1 µ(X), then B = X and the
1
A p condition on the whole space is nothing but the integrability of w and w− p−1 . Let
us assume now that d(x, x0 ) > Mr. In this case, every point y of the ball B satisfies
that d(y, x0 ) ' d(x, x0 ), so that, on B, we have that
1 α
ω(y) ' d α (x, x0 ) ω − p−1 (y) ' d − p−1 (x, x0 ).
and
In fact, d(y, x0 ) ≤ K(d(y, x)+d(x, x0 )) < K(r +d(x, x0 ) < K 1 + M1 d(x, x0 ). On the

M = 2K, we have
other hand, taking that d(x, x0 ) ≤ K(r + d(y, x0 )) for y ∈ B. Hence
d(x,x0 )
d(cx, x0 ) < K 2K + d(y, x0 ) . With the above remarks the desired A p estimates
follow now easily
Z Z p−1 ip
1
h α
− p−1
w dµ w dµ ' d α (x, x0 )µ(B) d − p−1 (x, x0 )µ(B)
B B
' µ(B).
When X = Rn ; d(x, y) = |x − y| and dµ = w dx and w ∈ A2 (Rn , d, dx), the space

(X, d, µ) or its normalization, become the basic models for the study of regularity of
weak solutions of degenerate elliptic equations in divergence form considered first
by G. Fabes, C. Kenig and R. Serapioni in [...]. If A = (ai j (x)) is a n × n symmetric
matrix satisfying a degenerate ellipticity condition of the form
λ w(x)|ψ|2 ≤ ψAψ T ≤ Λ w(x)|ψ|2 ,

a generalization of de Giorgi-Nash-Moser regularity theory is obtained for weak

solutions of the “elliptic” operator
¯ ∂ ∂u
Lu = ∇.A∇u =∑ ai j
i, j ∂ x j ∂ xj
when w ∈ A2 (Rn ).
As we have already observed, an A2 weight can actually vanish at some points,

hence the ellipticity of L is lost at those points. On the other hand, from Lemma 4.1
we have that w can not vanish on a set of positive Lebesgue measure, so that the
ellipticity of L is lost only on a “small” set.
4.2.2 A weighted parabolic space of homogeneous type and

degenerate heat equations.
In this model we shall use the results of the previous section and Theorem 1.41
in Section 1.18 of Chapter 1. Let X = Rn and
d(x1 , x2 ) = |x1 − x2 |∞ = max |x1i − x2i |,

i=1,··· ,n
where is the ith component of x j . Let Y = Rn . We shall introduce a family {ρx :

xij
x ∈ Rn } of quasi-distances on Y satisfying the conditions 1 and 2 of Theorem 1.41.
Let us start by a thecnical lemma. We shall use the notation Q(x, r) for the d-ball
with center x and radious r.
1
Lemma 4.4. Let 1 < p < ∞ and w ∈ A p (Rn ). With v = w− p−1 and x ∈ Rn , the func-
tion of r > 0 defined by
Z p−1
hx (r) = v(y)dy
Q(x,r)
satisfies the following properties:

1. hx is strictly increasing;
2. hx is continuous;
3. hx (0+ ) = 0;
4. hx is onto R+ .
Proof. Let us first notice that if p0 is the conjugate Hölder exponent of p, since
1
w ∈ A p (Rn ), from Hölder inequality, we also have that v = w− p−1 ∈ A p (Rn ). Hence
v itself satisfies, with p0 instead of p, all the results of the previous sub-section. In
particular, Lemma 4.1 implies that v can not vanish in a set of positive Lebesgue
measure. Since for r2 > r1 > 0 we obviously have |Q(x, r2 ) \ Q(x, r1 )| > 0, we have
that hx (r2 ) > hx (r1 ), which proves (1). Properties (2) and (3) follow from the ab-
solute continuity of the Lebesgue integral in Rn . Property (4) follows from Lemma
4.2 applied to vdx since, by iteration,
Z Z
v(y)dy ≥ γ k v(y)dy,
Q(x,2k ) Q(x,1)
so that hx (r) → ∞ as r → ∞.
From the above lemma, for each x ∈ Rn , the function hx is invertible on R+ . This
allows us to define the family of functions
ρ x : R × R → R+
0
by
ρx (s,t) = h−1
x (|t − s|),
for every x ∈ Rn . The basic fact is that the setting (Rn , d), (R, ρx ) for x ∈ Rn , satisfies
hipotheses (1) and (2) of Theorem 1.41.
Lemma 4.5. The function
δ : Rn+1 × Rn+1 → R+
0,
defined by
δ ((x,t); (y, s)) = |x − y|∞ + h−1 −1

x (|t − s|) + hy (|t − s|),
defines a quasi-distance in space-time Rn+1 . Moreover, the family of δ -balls is

equivalent to the family of rectangles
R((x,t), r) = Q(x, r) × (t − hx (r),t + hx (r)),
and each δ -ball is an open set.
Proof. Let us start by showing that for some fixed k we have the uniform quasi-
triangular inequality
ρx (s, u) ≤ k(ρx (s, y) + ρx (t, u))
for every s, t, u ∈ R and every x ∈ Rn . To prove this inequality it is enough to show
that h−1 −1 −1
x (α + β ) ≤ k[hx (α) + hx (β )], for some constant k, for every α > 0, β > 0
n
and x ∈ R . From the above lemma, this inequality is actually equivalent to
hx (a) + hx (b) ≤ hx (k(a + b)),

for every a > 0 and b > 0. In terms of the definition of hx we have to find a constant
k for which
Z p−1 Z p−1 Z p−1
v(y)dy + v(y)dy ≤ v(y)dy
Q(x,a) Q(x,b) Q(x,k(a+b))
for every a > 0, b > 0, x ∈ Rn . Let γ be the parameter provided by Lemma 4.2 when
dµ = v dx which is certainly doubling. Since γ > 1 there exists m ∈ N such that
1
γ m−1 < 2 p−1 ≤ γ m . Hence
Z p−1 Z p−1 Z p−1

v(y)dy + v(y)dy ≤2 v(y)dy
Q(x,a) Q(x,b) Q(x,a+b)
1 p−1
≤ 2 p−1 v(Q(x, a + b))
≤ (γ m v(Q(x, a + b))) p−1

≤ v(Q(x, 2m (a + b))) p−1
which is the desired inequality with k = 2m . Notice that from Lemma 4.4 we have
that each ρx is a quasi-distance on R. Next we show that
h−1 −1

x1 (r) ≤ 2 hx2 (r) + |x1 + x2 |∞
for every x1 and x2 ∈ Rn and every r > 0. This inequality shows that the family ρx
satisfies, with respect to the distance d(x1 , x2 ) = |x1 − x2 |∞ , the second hypothesis
of Theorem 1.41 and, as a consequence, δ becomes a quasi-distance in space-time
Rn+1 .
To check the above inequality, notice first that there is nothing to prove if
h−1 −1
x1 (r) ≤ 2|x1 − x2 |∞ . Assume then that hx1 (r) > 2|x1 − x2 |∞ and let us prove that,
in this case, h−1 −1 1 −1 −1
x1 (r) ≤ 2hx2 (r). In fact, since Q(x2 , 2 hx1 (r)) ⊂ Q(x1 , hx1 (r)), we
have that
! p−1
1 −1
Z
hx2 hx1 (r) = v(y) dy
2 Q(x2 , 21 h−1
x1 (r))
Z
! p−1
≤ v(y) dy
Q(x1 ,h−1
x1 (r))
= hx1 (h−1
x1 (r))
= r.
hence 21 h−1 −1
x1 (r) ≤ hx2 (r).
In order to prove the statements about the family of δ -balls on Rn+1 , let us notice
that from the last inequality proved we have that
δ ((x,t), (y, s)) ' |x − y|∞ + hx −1(|t − s|)
which is actually equivalent to the largest of the two terms on the right hand side. In
other words, for some constants 0 < c1 < c2 < ∞, we have that
c1 δ ((x,t), (y, s)) ≤ max{|x − y|∞ + h−1

x (|t − s|)}
≤ c2 δ ((x,t), (y, s)).
From these inequalities we see that the δ -balls with center (x,t) and radius r >
0 contains a rectangle R((x,t), ar) and is contained in a rectangle R((x,t), br). In
particular this fact shows that the δ -topology (see Chapter 1) induced in Rn+1 is the
usual topology. Finally notice that with the usual topology δ is continuous, so that
the open δ -balls in space-time are open sets
Let us try to give a physical meaning to this distance in space-time. Assume

that Rn is a “non-homogeneous” media in which a diffussion is taking place. Non
homogeneous here means only that the physical properties of the material can be
changing from point to point. We may also assume that the media is non-isotropic
which means that at each point the physical properties may depend on direction.
The general model for this situation is given by a conductivity matrix A(x) which
provides a generalized Fourier law for the thermal energy giving rise to the parabolic
type differential equation
∂u
= div A grad u
∂t
After the celebrated work by De Giorgi, Nash and Moser and the remarkable exten-
sions of the elliptic situation to degenerate ellipticity started by Fabes, Kenig and
Serapioni in [FKS], F. Chiarenga and R. Serapioni n [CS], consider the problem of
Harnack inequality for positive weak solutions of the above equation when
n
λ w(x)|ψ|2 ≤ ∑ ai j (x,t)ψi ψ j ≤ Λ w(x)|ψ|2
i, j=1
and w is a Muckenhoupt weight of the class A1+ 2 (Rn .

n
Let us next introduce a measure on (Rn+1 , δ ) where δ is defined as in Lemma 4.2

by a weight w ∈ A p (Rn ) and let us observe the result when p = 1 + 2n is the specific
index considered by Chiarenga and Serapioni.
Let E be a Borel subset of Rn+1 . Define

Z Z
µ(E) = w(x)dxdt
E
Notice that
µ(R((x,t); r)) = w(Q(x, r))2hx (r)

Z Z p−1
1
=2 w(y)dy w(y)− p−1 dy
Q(x,r) Q(x,r)
∼
= |Q(x, r)| p
' rnp
Since δ -balls are open sets and equivalent to the family of rectangles, we have
that
µ(Bδ ((x,t), r)) ' rnp ,

where the equivalence constants do not depend on x, t and r. This fact proves that
(Rn+1 , δ , µ) is a space of homogeneous type and moreover, the Hausdorff dimen-
sion of Rn+1 with respect to δ is np. In particular, when p = 1 + 2n we see that the
dimension of space-time is np = n + 2 which ispthe actual parabolic dimension of
space-time when d((x,t), (y, s)) = max{|x − y|, |t − s|}. Let us also observe that
(Rn+1 , δ n+1 , µ) is a normal space of homogeneous type with no atoms and with no
atoms and with µ(Rn+1 ) = +∞.
4.2.3 A space of homogeneous type modeling the analytic context

associated to the Monge-Ampére equation.
Given a potential (scalar field) u in Rn , the fundamental physical properties of the

vector field V = ∇u, the gradient of u, is contained in the second order tensor D2 u =
∂ 2u
∂ xi ∂ x j ; i = 1, 2, · · · , n, j = 1, 2, · · · , n, which is the Hessian of u and the Jacobian
of V = ∇u. In particular the trace of D2 u is the divergence of the vector field and
the Laplacian of u which is a local measure of the compressibility of V = ∇u. On
the other hand, if we are interested in measuring the volume changes produced by
∇u = V as a change of variables in Rn , we need to consider the Jacobian determinant
of ∇u which is
Lu = detD2 u,
the Monge-Ampére operator. In other words, the two basic invariants: trace and
determinant are inducing the two basic operators Laplace and Monge-Ampére. The
big mathematical difference among them is given by the non-linearity inherent to
the determinant.
Actually the Laplacian can be recovered from the Monge-Ampére operator by a

special linearization.
For the sake of computational simplicity, let us assume for a while, that the di-
mension of the space is n = 2. Let us compute the directional derivative of L in the
2
direction of φ (x) = |x|2 at u. The duality between the Lagrangian and the Hamilto-
nian approaches to classical mechanics gives a reasonable interpretation in terms of
kinetics energy of the fact that this particular quadratic form given by φ is specially
interesting. We have to find a limit as t → 0 of
L(u + tφ ) − L(u)
t
Notice that ∇φ = x and hence, D2 φ = I is the identity matrix. So that
L(u + tφ ) = detD2 (u + tφ ) = det(D2 u + tD2 φ )

= det(D2 + tI).
but as it is easy to show, det(A + tI) = det A + t traceA + t 2 . Hence
L(u + tφ ) − L(u) = t traceD2 u + t 2

= t4u + t 2
and,
L(u + tφ ) − L(u)
→ 4u,
t
as t → 0. In other words 4u is the directional derivative of L in the direction of
2
φ (x) = |x|2 at u,
D |x|2 L = 4
2
Of course there is nothing special with this particular φ in order to try to get Dφ L,
but when φ is a convex function Dφ L becomes an elliptic type operator
Dφ L(u) = Lφ (u) = trace(ΦD2 u),

where Φ is the cofactors matrix associated to D2 φ which is positive semi-definite.
The linearized operator Lφ is in non-divergence form and the degeneracy and non-
degeneracy of the ellipticity shall strongly depend on the convexity regularity prop-
erties of φ . The function φ gives rise to a particular geometric setting on Rn where
Euclidian balls are substituted by “sections” induced by φ and Lebesgue measure is
changed by the volume of the image through ∇φ . In this section our aim is to give
a structure of space of homogeneous type modeling this geometric setting when φ
satisfies the basic properties introduced by Caffarelli and Gutiérrez in [ ] and [ ].
Let us start by the basic observation of the fact that the usual metric-measure
structure on Rn given by Euclides-Lebesgue can be obtained from the Monge-
Ampére approach applied to the usual kinetic energy quadratic form given by
2
φ (x) = |x|2 , x ∈ Rn . For fixed x0 ∈ Rn , the tangent plane to the graph of φ at
(x0 , φ (x0 )) is given by
lx0 = φ (x0 ) + 4φ (x0 ).(x − x0 )

|x0 |2 |x0 |2
= + x0 .(x − x0 ) = x.x0 −
2 2
If we now lift this plane by ε > 0, the new plane intersects the graph of φ and the
section of φ at x0 with height ε is defined by
Sφ (x0 , ε) = {x ∈ Rn : φ (x) < lx0 (x) + ε}.

In our particular case, the inequality defining Sφ (x0 , ε) is
|x|2 |x0 |2
< x.x0 − + ε,
2 2
which is nothing but |x − x0 |2 < 2ε. In other words,
√
Sφ (x0 , ε) = B(x0 , 2ε),
the family of sections is the family of Euclidian balls. Moreover the Monge-Ampére
measure is given by
du = detD2 φ dx,
which coincides with Lebesgue measure since D2 φ = I, the identity matrix in Rn .
Hence, the standard Euclidian-Lebesgue structure of homogeneous type for Rn can
be regarded as a particular realization of the Monge-Ampére setting when the con-
2
vex function is the standard kinetic energy |x|2 .
A general convex function φ defined on Rn satisfies the inequality
φ (λ x + (1 − λ )y) ≤ λ φ (x) + (1 − λ )φ (y).
for every x, y ∈ Rn and every λ ∈ (0, 1). Even when convexity implies enough reg-
ularity, for the sake of simplicity we shall keep assuming that φ is a C ∈ convex
function. In particular for each point the supporting hyperplane is the tangent one:
Πx0 = {(x, y) ∈ Rn+1 : y = lx0 (x), x ∈ Rn }
where lx0 (x) = ϕ(x0 ) + ∇ϕ(x0 ).(x − x0 ).

Given a point x0 ∈ Rn and t > 0 the section of φ at x0 and height t is the set
Sφ (x0 ,t) = {x ∈ Rn : φ (x) < lx0 (x) + t}

4.3 Some basic examples from geometric measure theory. 161
4.3 Some basic examples from geometric measure theory.
4.3.1 The Cantor set.
Let F be the standard Cantor function which can be considered as a probability

distribution after extending it as zero for x ≤ 0 and one for x ≥ 1. Let µ = µF
be the Borel measure on R for which µ((a, b]) = F(b) − F(a) for every a < b.
Notice that since F is continuous the µ-measure of each single point vanishes and
µ((a, b]) = µ((a, b)). For the standard construction of the Cantor set on [0, 1] we
introduce the notation
n
Cn = 2j=1 Inj , Inj = [anj , bnj ]
and
C = ∩∞
n=1Cn ,
where Cn is the n-th approximation of the Cantor set and each one of the intervals Inj
has length 3−n . For each n ∈ N, let us denote by Ln the set of left points of intervals
n
Inj : j = 1, · · · , 2n . In other words Ln = {a1n , a2n , · · · , a2n }.
We shall give a special structure of space of homogeneous type on Ln , for each

n ∈ N. Let us start by defining a distance dn on Ln . Notice that x ∈ Ln if and only
if 3n x is a non-negative integer which can be written in basis 3 with at most n dig-
its, all of them in {0, 2}. In other words x can be obtained by dividing by 3n an
integer which in basis 3 has at most n digits and none of them is one. Hence Ln
can be identified with {0, 2}n , the set of all ordered lists (x1 , x2 , · · · , xn ) with xi = 0
or xi = 2. Given x, y ∈ Ln , let (x1 , x2 , · · · , xn ) and (y1 , y2 , · · · , yn ) be their respective
representations in {0, 2}n . Set

0 if x = y,
dn (x, y) =
3− j if xi = yi for i < j and x j 6= y j .
for each n ∈ N. Notice that each dn is actually a distance on Ln . In fact, if x =
(x1 , · · · , xn ),(y1 , · · · , yn ) and z = (z1 , · · · , zn ) are given, then if xi = yi for i < j1 and
yi = zi for i < j2 , we also have that xi = zi for i < min{ j1 , j2 }.
Next we define an uniformly distributed probability measure on each Ln by

µn ({x}) = 2−n for each xn ∈ Ln .
Lemma 4.6. {(Ln , dn , µn ) : n ∈ N} is an uniform sequence of spaces of homoge-

neous type in the sense that we can pick an uniform doubling constant for all the
members of the sequence
Proof. Let us denote by Bn (x, r) the dn -ball with center x ∈ Ln and radious r in the
metric space (Ln , dn ). Notice that, for j ∈ N we have that
Bn (x, 3− j ) = {y ∈ Ln : dn (x, y) < 3− j }

= {y ∈ Ln : yi = xi , i = 1, 2, · · · , j},
which reduces to the point {x} if j ≥ n. Then the number of elements in Bn (x, 3− j )
is 2n− j when j ≤ n and is
COMPLETAR
µn (Bn (x, 3− j )) = 2−n ](Bn (x, 3− j ))

−j
2 for j ≤ n,
=
2−n for j ≥ n.
Notice that since the dn diameter of Ln is bounded, it is enough to check the doubling
property for 0 < r < 1. Pick j ∈ N such that 3− j < r ≤ 31− j , then
µn (Bn (x, 2r)) ≤ µn (Bn (x, 32− j ))

−j
4.2 for j ≤ n + 2,
=
2−n for j ≥ n + 2.
On the other hand
µn (Bn (x, r)) ≥ µn (Bn (x, 3− j ))

−j
2 for j ≤ n,
=
2−n for j ≥ n.
So that
0 < µn (Bn (x, 2r)) ≤ 4µn (Bn (x, r)) < ∞
for every x ∈ Ln and every r > 0.
Let us now show that we can change dn by the restriction to Ln of the usual
distance on R.
Lemma 4.7. For every n ∈ N we have the inequalities
dn (x, y) ≤ |x − y| ≤ 3dn (x, y),
for every x and y in Ln .

Proof. Take x and y in Ln . Assume that x 6= Y. Hence dn (x, y) = 3− j for some j < n.
So that, with
n n
xi yi
x=∑ i y=∑ i
i=1 3 i=1 3
we have
n
|x − y| = ∑ 3−i (xi − yi )
i= j
n
≤ ∑ 3−i |xi − yi |
i= j
n
≤ 2 ∑ 3−i
i= j
< 3.3− j
= 3.dn (x, y).
On the other hand,
n
|x − y| ≥ 3− j |x j − y j | − ∑ 3−i (xi − yi )
i= j+1
2 1
≥ −j − j
3 3
= dn (x, y)
From the two lemmas we may conclude that also (Ln , ρ, µn ) with ρ(x, y) = |x −y|
is an uniform family of spaces of homogeneous type. In fact, for x ∈ Ln and r > 0,
µn (Bρ (x, 2r)) ≤ µn (Bn (x, 2r))

≤ 4µn (Bn (x, r))
r
≤ 43 µn Bn x,
3
≤ 43 µn (Bρ (x, r))
Let us next show that µn as a sequence of probability measures on the Borel

sets of R converges in the weak ∗ sense to µ. we have to prove that for every ϕ ∈
C ([0, 1]) we have that
Z Z
ϕ(x) d µn (x) → ϕ(x)dF(x)
[0,1] [0,1]
when n → ∞.
Since µn is uniformly distributed on Ln , each integral in the left is easy to com-

pute. In fact
n
1 2
Z
ϕ(x)dµn (x) = n ∑ ϕ(anj ).
[0,1] 2 j=1
Z On the other hand, being ϕ continuous and F of bounded variation, the integral
ϕdF exists and can be computed as the limit of the Riemann -Stieltjes sums of
[0,1]
a sequence of nested partition of [0, 1] with “norm” tending to zero. Let us consider
the sequence of partitions of [0, 1] given by
l
Πn = { : l = 0, 1, 2, · · · , 3n }, n ∈ N.
3n
Notice that ](Πn ) = 3n + 1 and that Ln is a subset of Πn with ](Ln ) = 2n . Since
the approximate Cantor function at level k + 1, Fk+1 is obtained by modifying Fk
only in the inner points of each component interval Ikj of Ck , we have that Fk+1 ≡ Fk
◦ ◦
in the complement of C. Hence F(x) = Fk (x) for every x ∈/C and every k ∈ N. Notice
◦
now that Πn ∩ C= 0/ if k ≥ n. So that F 3ln = Fn 3ln for every l = 0, · · · , 3n , and a

Riemann- Stieltjes sum associated to the partition Πn for the function ϕ with respect
to the measure dµ = dF, is given by
3n −1

l +1
l l
SΠn = ∑ ϕ F − F
l=0 3n
3 n 3 n
3n −1
l l +1 l
= ∑ ϕ n
Fn n
− F n n
l=0 3 3 3
l+1 l 1 l

Since, by construction of Cn and Fn , Fn 3n = Fn 3n + 2n if 3n ∈ Ln and
Fn l+1 l
l
3n = Fn 3n if 3n ∈ / Ln , we have that
2n
1
Z
SΠn = n ∑ ϕ(anj ) = ϕdµn .
2 j=1 [0,1]
Theorem 4.1. If C is the Cantor set, ρ is the restriction to C of the usual euclidian
distance and µ is the Borel measure induced by the Cantor function F, we have that
(C; ρ, µ) is a space of homogeneous type.
Proof. Take x ∈ C and r > 0, Let ψ be the continuous function defined on R+ taking
the value one in [0, 1], vanishing otuside [0, 2] and linear in [1, 2]. Let xn ∈ Cn be a
j
sequence of points for which xn → x as n → ∞. Pick j ∈ N such that 2 > 20. Since
|x−y|
the functions ϕx,α (y) = ψ α are continuous, from the weak-* convergence of
µn to µ, and the trivial estimates X[0,1] (t) ≤ ψ(t) ≤ X[0,2] (t), we have the following
inequalities proving the theorem

|x − y|
Z
µ(B(x, 2r)) = X[0,1] dµ(y)
C 2r

|x − y|
Z
≤ ψ dµ(y)
C 2r

|x − y|
Z
= lim ψ dµn (y)
n→∞ Cn 2r

|x − y|
Z
≤ lim inf X[0,2] dµn (y)
n→∞ Cn 2r
= lim inf µn (Cn ∩ B(x, 4r))
n→∞
≤ lim inf µn (Cn ∩ B(xn , 5r))
n→∞
r
3j
≤ 4 . lim inf µn Cn ∩ B xn ,
r4
n→∞

3j
≤ 4 . lim inf µn Cn ∩ B x,
n→∞
2
2|x − y|
Z
= 43 j lim inf X[0,1] dµn (y)
n→∞ Cn r

2|x − y|
Z
≤ 43 j lim ψ dµn (y)
n→∞ Cn r

2|x − y|
Z
= 43 j ψ dµ(y)
C r

2|x − y|
Z
3j
≤4 X[0,2] dµ(y)
C r
= 43 j µ(B(x, r))
Let us observe that the formula that we obtained for the µn measure of Bn (x, 3− j )
when x ∈ Cn , i.e.
−j
−j 2 for j ≤ n,
µn (Bn (x, 3 )) =
2−n for j > n.
allows, following the same pattern, to prove that locally, for small r, we have that
µ(B(x, r)) ' rα with α = log 2
log 3 . As by product we obtain that the Hausdorff dimen-
sion of C is α.
Chapter 5
The Hardy-Littlewood maximal function and the
differentiation theorem
This chapter is devoted to obtain Lebesgue type differentiation theorems on metric

measure spaces through the Hardy-Littlewood maximal function method. We start
by obtaining weak and strong type estimates for the Hardy-Littlewood maximal
operator on Lebesgue spaces using the basic covering and interpolation methods.
We shall introduce a dyadic version of the Hardy-Littlewood Maximal Operator
built on the dyadic sets given in §1.17. In the last section we prove the differentiation
theorem.
5.1 The Hardy-Littlewood maximal operator
Let (X, d) be a quasi-metric space and let ρ and β be a distance on X and a positive
number such that d ∼ ρ β = d 0 . The existence of ρ and β is given by Theorem 1.17.
Each d 0 -ball is also a ρ-ball, hence an open set. If F is some σ -algebra of subsets of
X that contains the ρ-balls and µ is a positive measure on F such that each ρ-ball
has positive and finite measure, we can consider mean values of locally integrable
functions f in different regions of the space and scales by taking the ??????
1
Z
mB ( f ) = f dµ.
µ(B) B
Hence several versions of the Hardy-Littlewood maximal function can be defined

in this setting. The non-centered maximal function
1
Z
M f (x) = sup | f (y)|dµ(y),
x∈B µ(B) B
where the supremum is taken over the family of all ρ (or d 0 )-balls containing x,
where f is a measurable function which is locally integrable (i.e. integrable on each
ρ-ball). The centered maximal function
167
168 5 The Hardy-Littlewood maximal function and the differentiation theorem
1
Z
Mc f (x) = sup | f (y)|dµ(y),
r>0 µ(B(x, r)) B(x,r)
where the supremum is taken over the family of the ρ-ball (or d 0 -balls) centered at
x. If the original d-balls Bd belong to F we have similar versions M d and Mcd of the
operators, taking d-balls instead of ρ-balls. Since for every λ > 0 the sets
Ωλ = {x ∈ X : M f (x) > λ } = ∪m B
B (| f |)>λ
is open, we have that M f is lower semicontinuous. It is also clear that Mc f (x) ≤

M f (x) for every x and every f .
Lemma 5.1. (5.1) Let X, d, F , ρ and β as before. If (X, dµ) is a space of homo-
geneous type in the sense that µ is doubling for ρ-balls, then
(5.1.1) the maximal functions Mc and M are equivalent,
(5.1.2) provided that all d-balls belong to F we also have that M d and Mcd are equiva-
lent to M and Mc ,
(5.1.3) the maximal function M f (x) is measurable for each f locally integrable.
Proof. We have already observed that Mc ( f ) ≤ M f (x), for every x. On the other
hand if B = B(y, r) is a ρ-ball containing x, we have that B ⊂ B(x, 2r) and, since
B(x, 2r) ⊂ B(y, 3r) we also obtain an estimate for mB (| f |)
µ(B(x, 2r)) µ(B(y, 3r))

mB (| f |) ≤ mB(x,2r) (| f |) ≤ m (| f |).
µ(B) µ(B(y, r)) B(x,2r)
So that, from the doubling property (5.1.1) follows immediately. The statement
(5.1.2) can be proved similarly. To prove (5.1.3) we only have to use that M f is lower
semicontinuous and that B, the σ -algebra of Borel measurable sets, is contained in
F . The last claim follows from the facts that ρ-balls belong to F and that (X, ρ) is
separable since, supporting the double measure µ, (X, ρ) has the weak homogeneity
property.
We shall keep assuming along this chapter that (X, dµ) is space of homogeneous
type such that d-balls are open and that M is taken with respect to the d-balls. Let
us observe that if f ∈ L∞ , then M f is bounded by || f ||∞ . If X is unbounded and f is
bounded and has support in the ball B(x0 , r), then, the Hardy-Littlewood maximal
function for f , M f (x), behaves like µ(B(x0 , d(x, x0 ))) for d(x, x0 ) large.
5.2 Weak type (1, 1) for M
In this section we shall use the Wiener type covering lemmas proved in Chapter
2 to prove the weak type estimate stated in the next theorem.
5.2 Weak type (1, 1) for M 169
Theorem 5.1. Let (X, d, µ) be a space of homogeneous type such that d-balls are
open sets. Then, there exists a constant C depending only on the geometric constants
k and A of (X, d, µ) such that the inequality
C
µ({x ∈ X : M f (x) > λ }) ≤ || f ||L1 ,
λ
holds for every locally integrable function f and every λ > 0
Proof. We only need to prove the desired inequality when f ∈ L1 (X, µ). Let R be a
fixed positive real number and x0 a fixed a fixed point in X. Consider the subset of
X defined by
ER = {x ∈ B(x0 , R) : M f (x) > λ }.

For each x in the bounded set ER , there exists a d-ball B(y(x), r(x)) containing x
such that
1
Z
| f (x)|dµ(y) > λ .
µ(B(y(x), r(x))) B(y(x),r(x))
If the space (X, d) is bounded we have that X = B(y, R0 ) for some R0 > 0 and
every y ∈ X. Hence, in this case, we can take r(x) to be bounded by R0 . If, on the
other hand, X is unbounded we have that µ(X) = +∞, so that
1 || f ||1
Z
| f (x)|dµ(y) ≤ .
µ(B(y(x), r(x))) B(y(x),r(x)) µ(B(y(x), r(x)))
Now, since B(x, r(x)) ⊂ B(y(x), 2Kr(x)) and x ∈ ER ⊂ B(x0 , R), if for some sequence
{xm } of points in ER we have r(m) −→ ∞, we also have for m large enough that
1
B(x0 , r(xm )) ⊂ B(xm , r(xm )) ⊂ B(y(xm ), 2Kr(xm )).
2K
1
Since µ(B(x0 , 2K r(xm ))) tends to µ(X) = +∞ as m tends to infinity, from the above
1
inequality and the fact that µ(B(x0 , 2K r(xm ))) ≤ Ãµ(B(y(xm ), r(xm ))) we have the
contradiction
Ã|| f ||1
0 < λ < mB(y(xm ),r(xm )) (| f |) ≤ 1
−→ 0 as m −→ ∞.
µ(B(x0 , 2K r(xm )))
so that the function r : X −→ R+ , x −→ r(x) is bounded above and we can apply

Lemma....(??Wiener 3??) in order to obtain a finite or countable sequence {xi : i ∈ I}
of points in ER such that the balls Bi = B(y(xi ), r(xi )) are pairwise disjoint and the
set ER is covered by the sequence {B?i : i ∈ I} with B?i = B(y(xi ), 10K 2 r(xi )). Then,
using again the doubling condition for µ, we obtain
µ(ER ) ≤ ∑ µ(B?i ) ≤ A ∑ µ(Bi )

i∈I i∈I
A A A
Z Z
< | f (y)|dµ(y) = | f (y)|dµ(y) ≤ || f ||1 .
λ∑i∈I Bi λ ∪i∈I Bi λ
Notice that, since the last term does not depend on R???? and the first increases
to µ({x ∈ X : M f (x) > λ }), we get the desired inequality with a constant A that
depends on A and K.
5.3 L p bounded of Mf
Since we have observed that ||M f ||∞ ≤ || f ||∞ and we have proved the weak type
inequality in Section 5.2, the L p -boundedness of M for 1 < p < ∞ follows from
the standard Marcinckiewiez interpolation technique which does not depend on the
geometric nature of the domain. In this section we shall briefly outline the proof of
the L p -boundedness of M.
Theorem 5.2. Let (X, dµ) be a space of homogeneous type such that the d-balls
are open sets, then there exists a geometric constant C such that the inequalities
1/p
1/p p
||M f || p ≤ 2 C || f || p ,
p−1
hold for every f ∈ L p (X, µ) and every p > 1.
Proof. Given a function f ∈ L p (X, µ) with 1 < p < ∞, we may compute the L p -
norm of M f by using its distribution function
Z ∞
||M f || pp = p λ p−1 µ({M f > λ })dλ .
0
For a given positive λ , decompose f as f = f λ + fλ , where f λ equals f if | f | ≥ λ2

and f λ = 0 if | f | < λ2 . Since | fλ | ≤ λ2 everywhere and M f ≤ M f λ + M fλ , we
see that {M f > λ } ⊂ { f λ > λ2 }, so that µ({M f > λ }) ≤ µ({M f λ > λ2 }). Hence,
applying the weak type (1, 1) inequality with f λ instead of f , λ2 instead of λ and
changing the order of integration, we get
5.5 The Lebesgue differentiation theorem 171
Z ∞
||M f || pp ≤ p λ p−1 µ({M f λ > λ /2})dλ
0
Z ∞ Z
≤ 2p C λ p−2 | f λ |(x)dµ(x) dλ
0 x∈X
Z ∞ Z
= 2p C λ p−2 | f (x)|dµ(x) dλ
0 x∈X:| f (x)|≥λ /2
2p p
Z Z 2| f (x)| Z
= 2p C | f (x)| λ p−2 dλ dµ = | f (x)| p dµ,
x∈X 0 p−1 X
which is the desired inequality.
5.4 Lipschitz functions with bounded support are dense in L p

(0 < p < +∞)
A basic point in the proof of the Lebesgue differentiation theorem is the fact that
it holds true for every function of a dense subspace of the Lebesgue spaces.
In this section we apply Corollary 3.9 in § 3.4 of Chapter 3 in order to show that
for some positive β the functions of class Lipschitz β with bounded support are
dense in L p (X, B, µ) (0 < p < +∞), when (X, d) is complete inRthe Cauchy sense
if L p (X, B, µ) denotes the Borel measurable functions such that X | f | p dµ < ∞.
Theorem 5.3. Let (X, d, µ) be a complete space of homogeneous type. Then there
exists β > 0 such that the space Λβ of Lipschitz β functions with bounded support
is dense in L p (X, B, µ)for every positive and finite p.
Proof. Since given f ∈ L p (X, B, µ) we have that x∈B(x p

R
/ 0 ,R) | f (x)| dµ(x) tends to
zero as R tends to infinity for x0 fixed, it is enough to prove that every L p −function
with bounded support can be approximated in the L p sense by a function in Λβ , for
some β > 0. Since every L p −function with bounded support can be approximated
in the L p sense by simple functions built on the characteristic functions of bounded
measurable sets, we only have to prove that if E is a bounded Borel subset of X and
ε is a positive number, then there exists f ∈ Λβ such that X |χE − f | p dµ < ε. This
R
can be achieved applying Corollary 3.9 in § 3.4 of Chapter 3.
5.5 The Lebesgue differentiation theorem

Let f be a real continuous function defined on the space of homogeneous type

(X, d, µ). Let ρ be a distance on X such that ρ α ∼ d for some α > 0. Then
mBρ (x,r) ( f ) tends to f (x) as r tends to zero for every x ∈ X. The same is true for
the mean values mB(x,r) ( f ) taken over the family of ρ α −balls or every other quasi-
distance for which the balls are measurable sets. Actually, for continuous functions
f , we have that
1
Z
| f (y) − f (x)|dµ(y) −→ 0
µ(B(x, r)) B(x,r)
as r −→ 0, provided that the balls are measurable sets.

Given a locally integrable function f defined on the space of homogeneous type
(X, d, µ), we say that x ∈ X is a Lebesgue point for f if there is a real number
f˜(x) such that mB(x,r) (| f − f˜(x)|) converges to zero when r −→ 0, with B(x, r) =
Bδ (x, r) for some δ ∼ d such that δ −balls are measurable. Notice that the Lebesgue
character of a point x ∈ X for a given function f does not depend on the particular
δ equivalent to d.
Theorem 5.4. Let (X, d, µ) be a space of homogeneous type such that continuous
functions are dense in L1 . Let f be a locally integrable function defined on X. Then
almost every point x ∈ X is a Lebesgue point for f . Moreover f˜(x) = f (x) almost
everywhere and mBδ (x,r) ( f ) −→ f (x), r −→ 0 almost everywhere for every δ ∼ d
for which δ −balls belong to F .
Proof. Let x0 ∈ X be a fixed point. Since the given f is locally integrable, we

have that f χBδ (x0 ,M) ∈ L1 for every M ∈ N. So that it suffices to prove the theo-
rem for f ∈ L1 . For f in L1 , let us consider the set E of those x ∈ X for which
lim supr−→0 mBδ (x,r) (| f − f (x)|) > 0. Hence
1
E = ∪n∈N En where En = {x ∈ X : lim sup mBδ (x,r) (| f − f (x)|) > }.
r−→0 n
Let g be any integrable and continuous function defined on X, then
1 1
En ⊂ {x ∈ X : Mc ( f − g)(x) > } ∪ {x ∈ X : | f (x) − g(x)| > }.
2n 2n
The weak type (1, 1)of the Hardy-Littlewood maximal function and Tchebychev
inequality, show that µ(En ) ≤ 4nC|| f − g||1 , for every continuous function g. Hence
En and so E are null sets.
As a by product of the results of the above theorem and the approximation result
given in § 5.4 we get the next statement
5.6 Dyadic tilings revisited 173
Corollary 5.1. Let (X, d, µ) be a complete space of homogeneous type and let f be
a Borel locally integrable function on X, then almost every point x ∈ X is a Lebesgue
point for f .
5.6 Dyadic tilings revisited
In this section we go back to the dyadic partitions constructed by M. Christ and

already introduced in a more general environment in § 1.17 of Chapter 1. There the
partitions were built under the metric hypothesis of equivalence of boundedness and
total boundedness. The quantitative additional information given by the finiteness
of the metric Assouad dimension allows us to prove some extra properties of the
dyadic sets. And the existence of a doubling measure such that continuous functions
are dense in the space of integrable functions will allow us to give an improvement
of the “almost covering property” 1.38.6: the boundaries of the dyadic sets have
zero measure.
Theorem 5.5. Let (X, d, µ) be a space of homogeneous type such that continuous
functions are dense in L1 . Let D = ∪ j∈Z D j be any of the dyadic families of subsets
of X constructed by M. Christ as in § 1.17. Then
µ(X\ ∪Q∈D j Q) = 0,
for every j ∈ Z.
Proof. Notice that for each i ∈ Z the set Ei = ∪Q∈Di Q is open. The indicator func-
tion χ j of E j is certainly a locally integrable function. So that from the Lebesgue
differentiation theorem proved in § 5.5 we have that mB4 (x,r) (χ j )(x) converges to
χ j (x) = 0 when r −→ 0 for µ−almost every point x ∈ X \ E j and every 4 ∼ d for
which the balls are open sets. This implies that the set
µ(B4 (x, r) ∩ E j )
N = {x ∈ X \ E j : lim sup > 0},
r−→0 µ(B4 (x, r))
4 µ(B (x,r)∩E )
j
is a µ−null set: µ(N) = 0. If we prove that the inequality lim supr−→0 µ(B >
4 (x,r))
0 is true for every x ∈ X, then N = X \ E j so that µ(X \ E j ) = 0. Since for every
i ∈ Z, Ni is δ i −dense in X, we have that for any given x ∈ X there exists k ∈ K (i)
such that d(xki , x) < δ i . Then for some constants C1 and C2 (C2 > C1 ) we have
B4 (xki , aδ i ) ⊂ B4 (x,C1 δ i ) ⊂ B4 (xki ,C2 δ i ). If 4 is the quasi-distance used in the

construction of the sets Qik in § 1.17, we have that
B4 (xki , aδ i ) ⊂ Qik ⊂ E i ⊂ E j ,
for every i ≥ j. Hence, for these values of i we obtain the inequalities
µ(B4 (x,C1 δ i ) ∩ E j ) µ(B4 (xki , aδ i ))

≥ > C3 > 0
µ(B4 (x,C1 δ ))i µ(B4 (xki ,C2 δ i ))
for some positive constant C3 . Since 0 < δ < 1 we see that
µ(B4 (x, r) ∩ E j )
lim sup > 0.
r−→0 µ(B4 (x, r))
Let us now state and prove some additional properties of the family D under the
assumption of finiteness of the metric Assouad dimension which, of course, hold
if (X, d, µ) is a space of homogeneous type. For a given positive number R, we
shall say that two dyadic sets Qkj and Qlj of the same level j ∈ Z are R − neighbors
if the inequality d(xkj , xlj ) ≤ Rδ j holds. Let us write NR (Q) to denote the set of
R−neighbors of Q ∈ D. Given a dyadic set Q ∈ D we define quadrant of Q as the
subset of X given by
C (Q) =
[
Q.
Q⊇Q, Q∈D
Lemma 5.2. Let (X, d) be a quasi-metric space with finite metric Assouad dimen-
sion. Let D be a dyadic family on (X, d). Then there exist constants M1 , depending
only on the geometric constants of the space, and M2 depending also on R > 0 such
that
n o
(a)] k ∈ K ( j) : Qkj ⊆ Qlj−1 ≤ M1 , for every l ∈ K ( j − 1) and every j ∈ Z;
(b)if C (Q) ∩ C (Q0 ) 6= 0,

/ then C (Q) = C (Q0 ) and ] {C (Q) : Q ∈ D} ≤ M1 ,
(c)] NR (Q) ≤ M2 .
Proof. (a)If k ∈ K ( j) and Qkj ⊆ Qkj−1 for some j ∈ Z and some l ∈ K ( j − 1),
then xkj ∈ Qkj ⊆ Qlj−1 ⊆ B(xlj−1 ,Cδ j−1 ) = B(xlj−1 , (δ −1C)δ j ). So that the set of
those xkj is δ j −disperse and is contained in a ball with radius a constant times
δ j , hence the number of offspring of Qlj−1 is bounded by a geometric constant.
(b)Notice first that if Q ⊆ Q̃ are two elements of D, then C (Q) = C (Q̃). Assume that
Q and Q0 are two dyadic sets, Q ∈ D and Q0 ∈ D, such that C (Q) ∩ C (Q0 ) 6= 0. /
5.6 Dyadic tilings revisited 175
Then there exist two dyadic sets Q̄ and Q̃ with Q̄ ⊇ Q and Q̃ ⊇ Q0 such that
Q̄ ∩ Q̃ 6= 0,
/ so that Q̄ ⊇ Q̃ or Q̃ ⊇ Q̄. In each of these cases we have that Q
and Q0 have a common ancestor Q∗ . From the remark above we see that C (Q) =
C (Q∗ ) = C (Q0 ), as desired. To prove that #({C (Q) : Q ∈ D}) is bounded above,
we only have to prove that any ball B0 = B(x0 , R0 ) we have that #({C (Q) : Q ∈
D and C (Q) ∩ B0 6= 0}) / is bounded by a geometric constant which does not
depend on x0 and R0 . For R0 > 0 given, take j ∈ Z such that R0 ∼ δ j . Since N j
is δ j −disperse and each Qkj is contained in a ball with radios of the order δ j , we
see that the set S of those dyadic sets Q ∈ D j for which Q ∩ B0 6= 0/ is bounded
above by a geometric constant that does not depend on x0 and R0 . Since, on the
other hand
{C (Q) : Q ∈ D and C (Q) ∩ B0 6= 0}

/ = {C (Q) : Q ∈ S}.
we have the desired bound.

(c)The proof of (c) is also simple consequence of the finiteness of the metric di-
mension.
In the next theorem we prove an interesting additional feature of these dyadic

families: {(Q, d, µ) : Q ∈ D} is a uniform family of sub-spaces of homogeneous
type of (X, d, µ).
Theorem 5.6. Let (X, d, µ) be a space of homogeneous type with geometric con-
stants K and A. Let D be a dyadic family on (X, d) constructed as in § 1.17, with
constants a, δ and C. Then, there exists a constant Ā, depending only on K, A, a, δ
and C such that for every ( j, k) ∈ A , (Qkj , d, µ) is a space of homogeneous type with
geometric constants K and Ā. In other words {(Qkj , d, µ) : ( j, k) ∈ A } is a uniform
family of spaces of homogeneous type.
Proof. In Problem (????) of Chapter 3 we have proved that if a measure ν is defined

on a σ −algebra containing the Borel sets of a quasi-metric space (Y, ρ) with open
balls and ν satisfies the doubling property, with a doubling constant A0 , for the balls
centered on a dense subset of Y , then (Y, ρ, ν) is a space of homogeneous type with
doubling constant Ã0 that depends only on A0 and on the triangle constant for ρ.
So that, for a fixed Q ∈ D we only have to prove the uniform doubling property
for balls centered at points x ∈ Q \ ∪Q0 ∈D ∂ Q0 . Let us assume that Q ∈ D j0 , so Q =
j
Qk00 for some k0 ∈ K ( j0 ). With BQ (x, r) we shall denote the 4−balls of the space
(Q, 4, µ)if 4 ∼ d is the quasi-distance on X used to obtain the dyadic sets as unions
of 4−balls: BQ (x, r) = Q ∩ B(x, r), x ∈ Q, x ∈ / ∪Q0 ∈D ∂ Q0 , r > 0 and B(x, r) =
{y ∈ X : 4(x, y) < 1}. Notice first that if r ≥ 2KCδ j0 , we have that B(x, 2r) ⊇
j
B(x, r) ⊇ B(xk00 ,Cδ j0 ) ⊇ Q, hence BQ (x, 2r) = BQ (x, r) = Q. So that, in this case,
the doubling property trivially holds with constant equal to one. Let us then assume
that 0 < r < 2KCδ j0 and let us pick j1 ≥ j0 such that 2KCδ j1 +1 ≤ r < 2KCδ j1 .
Since x does not belong to the boundaries of sets in D, there exists k1 ∈ K ( j1 + 1)
such that x ∈ Qkj11 +1 ⊆ Qk00 = Q. Then
j
B(xkj11 +1 , aδ j1 +1 ) ⊆ Qkj11 +1 = Qkj11 +1 ∩ Q

⊆ B(xkj11 +1 ,Cδ j1 +1 ) ∩ Q
⊆ B(x, 2KCδ j1 +1 ) ∩ Q
⊆ BQ (x, r).
On the other hand, since 2r < 4KC j1 +1 and x ∈ Q j1 +1 ⊆ B(x j1 +1 ,Cδ j1 +1 ), we

δ δ k1 k1
have
4KC
BQ (x, 2r) ⊆ B xkj11 +1 ,CK( + 1)δ j1 +1 .
δ
Thus

j1 +1 4KC j1 +1
µ(BQ (x, 2r)) ≤ µ B xk1 ,CK[ + 1]δ
δ

≤ Ãµ B(xkj11 +1 , aδ j1 +1 )
≤ Ãµ (BQ (x, r)) .
A picture for the case 0 < r < 2KCδ j0

Dibu
The next lemma shows in particular that the quadrants are subspace of a space of
homogeneous type for a given dyadic family D.
Lemma 5.3. Let (X, d, µ) be a space of homogeneous type. Let Yk (k ∈ N) be an

increasing sequence of subsets of X such that {(Yk , d, µ) : k ∈ N} is a uniform family
of subspaces of homogeneous type of (X, d, µ). Then if Y = ∪k∈NYn we have (Y, d, µ)
is a space of homogeneous type.
Proof. Let y be a point in Y and r be a positive real number. Then there exists k0 ∈ N
such that y ∈ Yk for every k ≥ k0 . Hence there exists a constant A which does not
depend on x, r or k ≥ k0 such that
µ (B(x, 2r) ∩Yk ) ≤ Aµ (B(x, r) ∩Yk ) ,

moreover, the left hand side is bounded below by µ B(x, 2r) ∩Yk0 > 0 and the right
hand side is bounded above by µ(B(x, r)) < ∞. Taking the limit for k −→ ∞ in the
above inequality we obtain the doubling property on (Y, d, µ).
>0
5.7 Dyadic Calderón-Zygmund decomposition of the space associated to a given integrable function f and to a given level λ 177
Corollary 5.2. Let (X, d, µ) be a space of homogeneous type and let D be a Christ
type dyadic family, then the quadrants are subspaces: (C , d, µ) is a space of homo-
geneous type for every quadrant C defined by D.
From the results of Chapter 3, we know that for a space of homogeneous type
boundedness and finiteness of the measure of the whole space are equivalent. So
that a quadrant C is unbounded if and only if µ(C ) = ∞. But the only case in which
a quadrant is bounded is when the space itself is bounded and then is only one
quadrant which is also a dyadic set.
Lemma 5.4. Let (X, d, µ) be an unbounded space of homogeneous type. Let D be a

Christ dyadic family in X. Then every quadrant C is unbounded.
Corollary 5.3. Let (X, d, µ) be a space of homogeneous type and D as before. Then
• a) If (X, d) is bounded then X = Q for some Q ∈ D and there is only one quadrant
with finite measure,
• b) if (X, d) is unbounded then there is a finite number of quadrants each with
infinite measure.
5.7 Dyadic Calderón-Zygmund decomposition of the space

associated to a given integrable function f and to a given
level λ > 0
1 R
We shall keep using the standard notation mE ( f ) = µ(E) E f dµ for a measurable
set E with positive measure and an integrable function f . When µ(E) = +∞ we
understand that mE ( f ) = 0
Theorem 5.7. Let (X, d, µ) be a space of homogeneous type and let D be a dyadic
family of the type considered in § 5.6. Let f be a non-negative µ−integrable function
defined on X and let λ be a positive number with λ ≥ mX ( f ). Then there exists a
family H = H ( f , λ ) ⊆ D such that
(a)Q ∩ Q0 = 0/ for Q ∈ H , Q0 ∈ H and Q 6= Q0 ;

(b)mQ ( f ) > λ for every Q ∈ H ;
(c)mQ̃ ( f ) ≤ λ for every Q̃0 ⊇ Q for some Q ∈ H ;
(d)mQ0 ( f ) ≤ λ for every Q ∈ D such that Q0 ∩ ∪Q∈H Q = 0;

/
(e)assuming that continuous functions are dense in L1 , f (x) ≤ λ for µ−almost

every x ∈
/ ∪Q∈H Q.
Proof. Let H1 be the family of all the dyadic sets Q ∈ D for which the mean value
of f over Q is larger than λ :
H1 = {Q ∈ D : mQ ( f ) > λ }.
If H1 = 0/ take H = H1 = 0. / Properties a), b), c) and d) are trivial and e) fol-

lows from d) by the Lebesgue differentiation theorem through the dyadic sets (see
Problem (???)). Let us assume that H1 6= 0.
/ Notice now that for each Q ∈ H1 the
class
{Q̄ : Q̄ ∈ H1 and Q̄ ⊃ Q}
is bounded above with the order given by inclusion. If X is bounded, X itself is a
dyadic set and mX ( f ) ≤ λ . In other words X itself is an upper bound for that family.
|| f ||1
If on the other hand X is unbounded, since f ∈ L1 , mQ0 ( f ) ≤ µ(Q 0 ) tends to zero if
0 0
the level of the dyadic set Q tends to infinity while Q ⊃ Q. Hence the inequality
µQ0 ( f ) > λ > 0 can not be true for large scales for Q0 if Q0 ⊃ Q. So that for each
Q ∈ H1 there exists a unique Q̃ ∈ D which is maximal with the properties Q̃ ∈ H∞
and Q̃ ⊃ Q. Let H be the class of those Q̃. In other words
H = {Q̃ : Q̃ is maximal with mQ̃ ( f ) > λ }.
It is easy to check that H satisfies a) to e).
5.8 The dyadic maximal function
Let (X, d, µ) be a space of homogeneous type and let D be a dyadic family of Christ
on (X, d). For a locally integrable function f defined on X, its dyadic maximal
function is given by
1
Z
M dy f (x) = sup | f (y)|dµ(y),
x∈Q∈D µ(Q) Q
provided that x ∈ ∪Q∈D Q and M dy f (x) = 0 otherwise. Let us observe that the excep-
tional set where M dy is defined to be zero is the set of the boundaries of quadrants,
which is always contained in the null set ∪Q∈D ∂ Q. Notice that for a given real num-
ber α the set E = {M dy f (x) > α} is open. In fact, if α < 0 the set E is the whole
space and if α ≥ 0, then the set E is a union of elements of D, which are all open
5.8 The dyadic maximal function 179
sets. In other words M dy f is lower semicontinuous and as a consequence, Borel

measurable.
It is easy to show that M dy f is pointwise dominated by M f . More precisely, there
exists a constant C, depending only on the geometric constant of the space such that
the inequality M dy f (x) ≤ CM f (x) holds for every x ∈ X. This inequality follows
from the doubling property of µ and from the metric control of the dyadic family
given by (1.38.2) and (1.38.3).
Let us also observe that if the function f has support contained in one of the
quadrants, then M dy f vanishes in every point of any other quadrant.
This fact shows that a pointwise estimate of M by M dy is generally impossible.
Nevertheless, as we shall show, we can get some control on the level sets of M f in
terms of the level sets of M dy f which can be useful.
Theorem 5.8. Let (X, d, µ) be a space of homogeneous type and let D be a given
dyadic family of Christ type in (X, d). Then there exists geometric constants R and
L such that for every integrable function f and every λ > 0 there exists a disjoint
family H ⊂ D such that except for sets of measure zero we have
(a){x ∈ X : M dy f (x) > λ } = ∪Q∈H ;

(b){x ∈ X : M f (x) > Lλ } ⊂ ∪Q∈H ∪Q0 ∈NR (Q) Q0 },

where M f is the non-centered Hardy-Littlewood maximal function and NR (Q) the

family of R−neighbors of Q as defined in (???).
Observe that from (a) we can obtain a version of the weak type inequality (1, 1)
for the dyadic maximal function which is sharper than the one obtained from esti-
mate M dy f ≤ CM f . Notice also that the control that H , hence M dy f , has on the
level sets of the Hardy-Littlewood maximal function M f , is given in terms of neigh-
bors of elements of H which generally do no “preserve quadrants”.
????PROOF OF THEOREM (?.?):

Take f ∈ L1 (X, µ) and λ > 0. If λ ≥ mX (| f |) we can apply the Calderón-
Zygmund decomposition given in the previous section to | f | and λ , in order to
get a family H ⊂ D satisfying (a) to (e) of Theorem (?.?). If, on the other hand,
λ < mX (| f |), this is because µ(X) < ∞. So that X = Qkj for some ( j, k) ∈ A . In
this case we take H to be the family whose only element is the dyadic set Qkj .
Let us start by proving that with this definition of H we have (a). Notice first that
H is empty if and only if mQ (| f |) ≤ λ for every Q ∈ D which is equivalent to
M dy f (x) ≤ λ for every x. Hence both sides of (a) are empty in this situation. If H
has X as its only element, then mX (| f |) > λ implies M dy f (x) > λ for every x and
both sides of (a) become X. For the generic case, H 6= 0/ and H 6= {X}, using
(.?, .?, b) of Theorem (.?.?) we see that if x ∈ Q ∈ H , then M dy f (x) ≥ mQ (| f |) > λ .
Hence the right hand side of (a) is contained in the left hand side. Given now a
point x for which M dy f (x) > λ , we have that, for some Q ∈ D, mQ (| f |) > λ and
x ∈ Q. From the construction of H there must exist Q0 ∈ H with Q0 ⊃ Q, hence

x ∈ ∪Q∈H Q.
In order to prove (b) let us start by noticing that, by taking L large enough if
needed, it suffices to prove it for the centered maximal function M c f and any equiv-
alent quasi-distance Notice that if H = 0/ the inclusion in (b) holds if we prove that
for some geometric constant L we have that M f (x) ≤ Lλ for every x. We know that
H = 0/ if and only if mQ (| f |) ≤ λ for every Q ∈ D. Let us show that, in this case,
we can find a geometric constant L such that mB (| f |) ≤ Lλ for every ball. Assume
that B = B(x, r). Take j ∈ Z in such a way that δ j+1 ≤ r < δ j . Let us consider the
family
G (x, r) = {Q ∈ D j : Q ∩ B(x, r) 6= 0},
/
of dyadic sets of level j. From the homogeneity property we see that ] (G (x, r)) ≤ M
for some geometric constant M. From the doubling property of µ and the metric
control D we see that µ(Q) ≤ Cµ(B) for some geometric constant C and every
Q ∈ G (x, r).
1
Z
mB (| f |) ≤ ∑ µ(B) | f |dµ
Q∈G (x,r) Q
≤ Ã ∑ mQ (| f |)
Q∈G (x,r)
≤ ÃMλ ,
as wanted.
We assume that H 6= 0/ and that H 6= {X} since the last case is trivial. Recall
that the current geometric constants involved are K, A, δ , a and C. With R = K(C +
K(1 +C)) we have the following statement.
/ ∪Q∈D ∂ Q, j ∈ Z and G (x, r) be as before: δ j+1 ≤

Claim: Let B = B(x, r), with x ∈
r < δ and G (x, r) = {Q ∈ D j : Q ∩ B 6= 0}.
j / Then if Q ∈ G (x, r) and Q ⊆ Q̃, for
some Q̃ ∈ D, we have that x ∈ ∪Q0 ∈NR (Q̃) Q0 .
In order to prove the Claim, let us take a point y ∈ Q ∩ B and let us write Q̃ = Qil
for some l ∈ K (i) and i ≤ j and Q = Qkj for some k ∈ K ( j). Then
d(x, xli ) ≤ K d(x, y) + d(y, xli ) < K(r +Cδ i ) < K(δ i +Cδ i ) ≤ K(1 +C)δ i .

In other words x is a point of the ball B(xli , K(1 +C)δ i ). Since does not belong to the
residual boundaries Z = ∪Q∈D ∂ Q, then for some m ∈ K (i) we have that x ∈ Qim .
Hence
i
, xli ) ≤ K d(xm
i
, x) + d(xli , x) < K(Cδ i + K(1 +C)δ i ) = Rδ i .

d(xm
5.8 The dyadic maximal function 181
The last inequality shows that Qim is an R neighbor of Qil = Q̃. Since we know that
x ∈ Qim , we have the desired claim: x ∈ Qim ∈ NR (Q̃).
From the claim we conclude that if x is any point of X which does not belong
to Z ∪ [∪Q̃∈H (∪Q0 ∈NR (Q̃) )], then: NO Q ∈ G (x, r) is contained in a Q̃ ∈ H . Let us
prove that the mean value of | f | over the ball B(x, r) is bounded above by a constant
times λ . Since for every Q ∈ G (x, r) we have that Q ∩ ∪Q̃∈H Q̃ = 0, / from (d) of
Theorem (?.?) above, we have that mQ (| f |) ≤ λ and, again,
1
Z
µ(Q)
| f |dµ ≤ ∑ mQ (| f |) ≤ M Ãλ .
µ(B(x, r)) B(x,r) Q∈G (x,r)
µ(B(x, r))
Chapter 6
Weighted norm inequalities for the
Hardy-Littlewood maximal operator
Let (X, d, µ) be a space of homogeneous type. Any non-negative, non-trivial locally

integrable function w defined on X shall be called a weight. Let 1 < p < ∞, we shall
say that a weight w satisfies the A p (X)−Muckenhoupt ?? condition on the space of
homogeneous type (X, dµ) if and only if there exists a positive constant C such that
the inequality
p−1
1 1
Z Z
1
w dµ w− p−1 dµ ≤ C,
µ(B) B µ(B) B
holds for every d−ball B in (X, d).

1
Let us observe that the A p (X) condition on w implies
R
that σ = w− p−1 is also
locally integrable.
R
In fact, if for some ball B0 we have B0 σ dµ = +∞,
R
we necessar-
illy have that B σ dµR= +∞ for every ball B containing B0 , so that B w dµ = 0 for
every such ball, then X w dµ = 0 and w = 0.
The next lemma collects some basic properties of A p weights that will be used in
the proof of the main result.
Lemma 6.1. Let 1 < p < ∞. We have

1
(2.1.a) If w ∈ A p (X), then σ = w− p−1 ∈ Aq (X) with p + q = pq.
(2.1.b) Given w ∈ A p (X), then there exist positive constants C and δ such that
the inequality
w(E) δ

µ(E)
≤ ,
µ(B) w(B)
R
holds for every d−ball B and every measurable subset E of B. Here w(E) = E w dµ.
(2.1.c) (X, d, wdµ) is also a space of homogeneous type for w ∈ A p (X).
183
184 6 Weighted norm inequalities for the Hardy-Littlewood maximal operator
Proof. (2.1.a) From the remark following the definition of A p we know that if w ∈
A p , then σ is locally integrable. On the other hand σ is non-trivia. since otherwise
w would not be locally integrable. In other words, σ is a weight on X. Let us show
that σ belongs to Aq . We only have to apply Hlder inequality to obtain
q−1
1 1
Z Z
1
− q−1
σ dµ σ dµ =
µ(B) B µ(B) B
q−1
1 1
Z Z 1
1
− p−1
= w dµ w (q−1)(p−1) dµ
µ(B) B µ(B) B
( 1
p−1 ) p−1
1 1
Z Z
1
− p−1 1
= w dµ w dµ ≤ C p−1
µ(B) B µ(B) B
where C is the A p (X) constant for w.
(2.1.b) Let w be an A p −weight and let E be a measurable subset of the ball B.

Then Z Z
1 1
µ(E) = dµ = w p w− p dµ
E E
Z 1 Z 1 Z p−1
p q q 1 1 p
≤ w dµ w− p dµ ≤ w(E) p w− p−1 dµ .
e E B
Now, from the A p condition for w we see that

p−1
µ(B) p
Z
1
− p−1
w dµ ≤CR .
B B w dµ
Then
1 µ(B)
µ(E) ≤ Cw(E) p 1 ,
w(B) p
1
which proves (b) with δ = p and C the A p constant for w.
(2.1.c) Take B = B(x, 2r) and E = B(x, r) in (2.1.b).
Let us notice that as consequence of (2.1.b) we have that for every set E of posi-
tive µ−measure we have that w(E) is also positive- On the other hand we certainly
have if µ(E) = 0, then w(E) = 0. It follows that the space of µ−essentially bounded
functions L∞ (X, dµ) is the same as L∞ (X, wdµ). The space L p (w) (0 < p < ∞) is
the class real valued measurable functions defined on X for which
Z
|| f || pp,w = | f (x)| p dµ(x)
X
6 Weighted norm inequalities for the Hardy-Littlewood maximal operator 185
is finite.
Theorem 6.1. Let 1 < p < ∞ and w be a weight on (X, d, µ) such that the Hardy-
Littlewood maximal operator satisfies the inequality
Z Z
|Mg| p w dµ ≤ C |g| p w dµ
X X
for some constant C and every measurable function g, then w belongs to the Muck-
enhoupt A p (X) class on the space of homogeneous type (X, d, µ).
Proof. Let us first take f ≥ 0 a locally integrable non-negative real function defined
on X and B a fixed d−ball in X. Since for every x ∈ B, mB ( f ) = mB (χB f ) is bounded
by M(χB f )(x), we have the pointwise inequlity
mB ( f )χB (x) ≤ M(χB f )(x).

Taking g = χB f in the inequality of the hypothesis, we see that
Z
w(B) mB ( f ) ≤ C f p w dµ.
B
Notice now that if E is any measurable subset of X with µ(E) > 0, then by taking
g = χE , since Mg > 0 almost everywhere, we have that w(E) > 0. In particular, for
every ball B we have that w(B) > 0. So that if f is bounded above and below by
positive constants we have that
Z Z p Z −1
p
w dµ f dµ f w dµ ≤ C µ(B) p .
B B B
1
Take now a sequence fn increasing to w− p−1 in order to obtain the desired result
Z Z p−1
1
w dµ w− p−1 dµ ≤ C µ(B) p .
B B
The main result of this section is the converse of the preceding theorem. We shall
give the proof given in [AM] which is an extension to spaces of homogeneous type
of the remarkably simple proof of Christ an Fefferman [CF].
Theorem 6.2. Let (X, d, µ) be a space of homogeneous type and let w be an

A p −weight (1 < p < ∞). The M is a bounded operator on L p (w), in other words
Z Z
|M f | p w dµ ≤ C | f | p w dµ,
X X
where C depends only on the geometric constants on p and on the A p −constant of
w.
The proof of Theorem (2.3) depends on an adequate Calderón-Zygmund type

lemma given in the next statement.
Lemma 6.2. Let (X, d, µ) be a space of homogeneous type with µ(X) = +∞. For
any ball B we shall denote by B̃ the ball concentric to B with 5K 2 times its radius
and B̂ the ball concentric to B and 15k5 its radius. Let A be a constant such that
µ(B̂) ≤ µ(B). For any non-negative integrable function f with bounded support,
b ≥ 2A3 > 1 and any k ∈ Z such that the set Ωk = y ∈ X : bk+1 ≥ M f (y) > bk

is non-empty, there exists a sequence of balls {Bki }i∈N satisfying
(2.4.1) Ωk ⊂ ∪∞ k
i B̃i ;
(2.4.2) Bki ∩ Bkj = 0/ if i 6= j;
(2.4.3) for every Bki , there exists xik ∈ Bki such that if rik is the radius of Bki , r ≥
5K 2 rik and xik ∈ B = B(y, r), then
bk+1 ≥ M f (xik ) ≥ mBk f > bk ≥ mB f ;

i
j k
(2.4.4) if x ∈
/ ∪∞j=k ∪∞
i=1 B̃i and M f (x) < ∞, then M f (x) ≤ b ;
/ and let Akj = ∪(l,n)∈I k B̃ln ,

(2.4.5) let I kj = {(l, n) ∈ Z × N : l ≥ k + 2, B̃ln ∩ B̃kj 6= 0}
j
then 2µ(Akj ) ≤ µ(Bkj );
(2.4.6) let E kj = B̃kj − Akj , then 2µ(E kj ) ≥ µ(B̃kj ) and µ(X − ∪k, j E kj ) = 0. If x ∈ E kj
and M f (x) < ∞, then M f (x) ≤ bk+2 ;
µ(B̃kj )
(2.4.7) if Fjk = Bkj − Akj , then µ(Fjk ) ≥ 2A and ∑+∞ ∞
−∞ ∑ j=1 χ k (x) ≤ 3, where χE
Fj
denotes the characteristic function of E.
Proof of Theorem (2.3) Let us first assume that µ(X) = +∞. Let f be as in Lemm
(2.4), then from (2.4.6) and (2.4.3) we have, with w ∈ A p (X) (1 < p < ∞),
Z Z
(M f ) p w dµ ≤ ∑ (M f ) p w dµ
X k
k, j E j
≤ b2p ∑ bkp w(E kj ) ≤ b2p ∑(mBk f ) p w(E kj )

j
k, j k, j
" #p !p
2p 1
Z
−1
σ (Bkj )
=b ∑ (f σ ) σ dµ w(E kj )
k, j σ (Bkj ) Bkj µ(Bkj )
" #p
1
Z σ (B̃kj )
≤b ∑ 2p
k)
( f σ −1 ) σ dµ σ (E kj )
k, j σ (B j B k
j σ (E kj )
" # p−1 
 σ (Bk ) w(B̃ki ) 
j
· ·
 µ(Bkj ) µ(Bkj ) 
From Lemma (2.1) and the very definition of w ∈ A p (X) we have that
" 
 σ (Bk ) p−1 w(B̃k ) 
#
j i
·
 µ(Bkj ) µ(Bkj ) 
is bounded by a constant depending on the A p constant for w. On the other hand,

from (2.1.a) and (2.1.b) we see that
!γ
σ (B̃kj ) µ(B̃kj )
≤C ,
σ (E kj ) µ(E kj )
for some positive and finite constants C and γ. The last term is bounded by a constant
µ(B̃kj )
since from (2.4.b) µ(E kj )
≤ 2. Then, the L p norm of M f is bounded by a constant
times the quantity
( " #p ) 1p
1
Z
∑ ( f σ −1 ) σ dµ σ (E kj ) .
k, j σ (Bkj ) Bkj
But σ (E kj ) ≤ σ (B̃kj ), which from (2.4.7) is bounded by 2A µ(Fjk ). Since Fjk ⊂ Bkj
and the overlapping of the E kj is bounded, we have that
Z 1
p
−1 p
||M f ||L p (wdµ) ≤ C |Mσ ( f σ )| σ dµ ,
X
1 R
where Mσ g(x) = supx∈B σ (B) B g σ dµ. Since (X, d, σ dµ) is also a space of homo-
geneous type and Mσ is the natural maximal operator on this space we certainly
have its L p (σ µ) boundedness. In other words
Z 1 Z 1
p p
p p
|Mσ g| σ dµ ≤C |g| σ dµ .
X X
Then
Z 1
p
||M f ||L p (wdµ) ≤ C | f σ −1 | p σ dµ
X
Z 1
p
≤C | f | p σ 1−p dµ
X
Z 1
p
=C | f | p w dµ
X
= C || f ||L p (wdµ) ,
which is the desired inequality for the case µ(X) = ∞.
Let us now considerer the case µ(X) < ∞. Set Y = X × R, δ : Y −→ R+ ∪ {0}

defined by δ ((x1 ,t1 ), (x2 ,t2 )) = max{d(x, x2 ), |t1 − t2 |} and ν = µ × λ with lambda
the one-dimensional Lebesgue measure. Now (Y, δ , ν)is a space of homogeneous
type with ν(Y ) = +∞. Since µ(X) is finite, then X is bounded, so that for some
positive R and every x ∈ X we have that Bd (x, R) = X. Take now a weight w ∈
A p (X) and f a measurable function on X, the W (x,t) = w(x) is an A p (Y )−weight.
Moreover F(x,t) = f (x)χ−2R,2R (t) is a measurable function on Y . Then we have the
inequality
M f (x) = MX f (x) ≤ MY F(x,t)
for every t ∈ (−R, R). Since ν(Y ) = +∞ we may use the case just proved to show
that
Z RZ
|MY F(x,t)| p W (x,t) dµ(x) dt
−R X
Z Z
≤ |MY F(x,t)| pW (x,t) dµ(x) dt
R X
Z Z Z 2R Z
≤ |F(x,t)| pW (x,t) dµ(x) dt ≤ C | f (x)| p w(x) dµ(x) dt
R X −2R X
= C 4R || f ||Lp p (wdµ) .
Notice finally that the left hand side of this inequality is an upper bound for
Z
2R |M f (x)| p w(x) dµ(x) = 2R||M f ||Lp p (wdµ) .
X
Proof of Lemma (2.4) We shall make use of the following version of Wiener type
covering lemma.
“Let E be a bounded subset of X and assume that for each x ∈ E there exist
y(x) ∈ X and r(x) > 0 such that x ∈ B(y(x), r(x)). Then, there exists a sequence of
disjoint balls {B(y(xi ), r(xi ))} such that E ⊂ ∪i B(y(xi ), 5K 2 r(xi )).”
Recall that given a ball B we write B̃ for 5K 2 B, B̂ = 15K 5 B and A satisfies µ(B̂) ≥
Aµ(B). Assume that the set Ωk = {y ∈ X : bk+1 ≥ M f (y) > bk } = 6 0/ with k ∈ Z,
b ≥ 2A3 and M f is the non-centered maximal function for f which is integrable and
has bounded support. Assume that f ≥ 0. Take x ∈ Ωk . Consider the set
Rk (x) = {r > 0 : mB f > bk for some B = B(y, r) such that x ∈ B}.

Let us observe that Rk (x) is a non-empty bounded subset of R+ . In fact, since for
x ∈ B(y, r) we have that B(y, r) ⊂ B(x, 2Kr) ⊂ B(y, 4K 2 r), we have that
1 || f ||1
Z
mB f = f ≤C
µ(B) B µ(B(x, r))
which tends to zero as r −→ ∞, since f ∈ L1 and µ(X) = +∞. So that the condition
mB > bk can not hold for r large.
Now, since Ωk is bounded let us choose r(x) ∈ Rk (x) in such a way that if r is any
number bigger than 5K 2 r(x), then r ∈
/ Rk (x). Thus, there is a point y(x) ∈ X such
that
(?) bk+1 ≥ M f (x) ≥ mB(y(x),r(x)) f > bk ≥ mB(y,r) f ,
whenever r ≥ 5K 2 r(x) and x ∈ B(y, r).

The boundedness of the support of f implies the boundedness of Ωk for each
k ∈ Z. Then Ωk , rk and y(x) are given in the conditions of the Wiener type covering
lemma stated above. Let us use it. In this way we obtain a sequence {Bki } with i
belonging to a subset of N and k varying over all the integers of Z for which Ωk is
non-empty. Properties (2.4.1) and (2.4.2) are given by the Wiener lemma. Property
(2.4.3) is nothing but (?) for the particular xik provided by Wiener. Since the set
where M f < ∞ can be written as the union of the sets Ωk , for k ∈ Z, if f is non-
trivial, then (2.4.4) follows from (2.4.1).
Proof of (2.4.5) Let us first show that if l ≥ k + 2, n ∈ N and B̃ln ∩ B̃kj 6= 0, / then B̃ln ⊂
B̂kj , even more rnl ≤ rkj . Indeed, assume rnl > rkj , then B̃kj ⊂ B̃ln . The last inequality in
(2.4.3) applied to B(y, r) = B̂ln gives
µ(Bln ) 1
bk ≥ m = ·m l (f) ≥ m l f,
B̂ln l
µ(B̂n ) Bn A Bn
now, by the third inequality in (2.4.3) applied to the pair (l, n) we have that
1 1 1 k+2
m f > bl ≥ b .
A Bln A A
In other words, we would have bk ≥ A1 bk+2 which contradicts our choice of b.

So that, with l ≥ k + 2, n ∈ N and B̃ln ∩ B̃kj 6= 0/ we certainly have B̃ln ⊂ B̃kj . This fact
together with (2.4.3) and (2.4.2) gives us (2.4.5) in the following way
µ(Akj ) ≤ ∑ µ(B̃ln ) ≤ A ∑ µ(Bln )

I kj I kj
Z
! Z !
−l −l
≤ A∑b f dµ ≤ A ∑ b f dµ
Bln B̃kj
I kj l≥k+2
b−k−2 b−k−1
≤ A· · µ(B̃kj ) m k ( f ) ≤ A2 m ( f ) · µ(Bkj )
1 − 1/b B̃ j b − 1 B̃kj
A2 1 µ(Bkj )
≤ · · µ(Bkj ) ≤ .
b b−1 2
Proof of (2.4.6) DIBU??????
µ(E kj ) = µ(B̃kj ) − µ(Akj ∩ B̃kj ) ≥ µ(B̃kj ) − µ(Akj )

1 1
≥ µ(B̃kj ) − µ(Bkj ) ≥ µ(B̃kj ).
2 2
Take now x such that M f (x) < ∞. Then x ∈ Ωk for some k ∈ Z. By (2.4.1) x ∈ B̃kj
for some j ∈ N. Assume that x ∈ Akj , then there exists (l, n) ∈ I kj such that x ∈ B̃ln ,
and from (2.4.3) we obtain
1 bk+2
M f (x) ≥ mB̃n ≥ ml > > bk+1 ,
l A Bn A
which is a contradiction. Thus {E kj }
is a covering of the set {x : M f (x) > ∞}. On
the other hand, on account of the weak type (1, 1) boundedness of the H-L maximal
operator, the set {x : M f = +∞} is of measure zero and (2.4.6) is proved.
Proof of (2.4.3): From (2.4.2) we have

∞
∑ χFjk (x) = χ∪ j Fjk (x) ≤ χ∪ j E kj (x).
j=1
By definition of E kj it follows that no point of X belongs to more than three of the

sets ∪E kj . Then
∑ ∑ χ k (x) ≤ ∑ χ k (x) ≤ 3.
Fj ∪ jE j
k j k
Chapter 7
Some techniques for the L2 boundedness of
linear operators
The classical methods for the analysis of integral convolution operators acting on
L2 spaces, range from Young inequality when the kernels are integrable functions to
the Fourier transform method when dealing with singular kernels. While Young in-
equalities can easily be extended to general non Euclidean settings, the more subtle
and hence useful tool provided by the Fourier Transform is not available in such a
general framework.
Several interesting questions arise from the above observation. We shall be con-
cerned with at least two of them. First, we look for adequate substitutes of the
Fourier analysis as a tool for the study of boundedness of operators. Two well known
functional analysis based criteria are used in several contexts and situations: Cotlar’s
Lemma and Krein’s Lemma. Second, we may try to build a Fourier analysis in L2
sharing the basic features of the standard one. Even when every separable Hilbert
space has an orthonormal basis, in order to get some control on integral type op-
erators, some more concrete Fourier bases are needed. In particular we would like
to have a metric control on the basic function. It turns out that some “time-scale”
localization is possible on general setting allowing the construction of good enough
wavelet type bases.
The aim of this chapter is to introduce and briefly illustrate these L2 methods.
We start by an extension of Young’s inequality which in the discrete case is usually
called Schur Lemma. Then we prove the remarkable result contained in Cotlar’s
Lemma. Next we state and prove Krein’s Lemma. Finally we introduce a wavelet
type Fourier analysis on some metric measure spaces.
7.1 Young’s Inequality
Along this section we shall assume that (X, µ) is a σ -finite measure space. An inte-
gral operator acting on real functions defined on X is induced by a kernel K = K(x, y)
defined on X × X in the standard way
193
194 7 Some techniques for the L2 boundedness of linear operators
Z
T f (x) = K(x, y) f (y) dµ(y), (7.1)
y∈X
provided that the integral exists in the Lebesgue sense. In other words, conditions
on K and f should guarantee the absolute convergence of the above integral at least
for almost every x ∈ X. The extension of Young’s inequality is contained in the next
result
Theorem 7.1.R
Let K be a measurable Rreal valued function on X × X such that the
inequalities |K(x, y)| dµ(y) ≤ C and |K(y, x)| dµ(y) ≤ C hold for some constant
C and every x ∈ X. Then for every f ∈ L p ; 1 ≤ p ≤ ∞ the integral in 7.1 is absolutely
convergent for almost every x ∈ X and the operator T is bounded on L p , with kT k ≤
C.
Proof. Assume first that 1 < p < ∞. Let q denote the Hölder conjugate of p : q = 1p .
1 1
The function |K(x, y)|| f (y)| = |K(x, y)| q · |K(x, y)| p | f (y)| is measurable as a func-
tion on the product measure space X × X. From Hölder’s inequality we see that
Z 1 Z 1
q p
Z
p
|K(x, y)|| f (y)| dµ(y) ≤ |K(x, y)| dµ(y) |K(x, y)|| f (y)| dµ(y)
Z 1
1 p
p
≤C q |K(x, y)|| f (y)| dµ(y) .
Hence, integrating with respect to x the above inequality rised to the power p, we
obtain
Z Z p Z Z
p
p
|K(x, y)|| f (y)| p dµ(y) dµ(x)

|K(x, y)|| f (y)| dµ(y) dµ(x) ≤ C q
Interchanging integrals and using again the uniform boundedness p

for the inte-
grals of |K| we get that the last integral is bounded above by C1+ a k f k pp . Since this
quantity is finite we are proving that the integral in 7.1 is absolutely convergent for
almost every x. On the other hand, the same inequality proves that kT f k p ≤ Ck f k p .
as desired. The cases p = 1 and p = ∞ are even easier.
Corollary 7.1. Let (a − i j : i ∈ Z; j ∈ Z) be a given infinite real valued matrix such
that the inequalities ∑i∈Z |ai j | ≤ C and ∑i∈Z |a ji | ≤ C hold for some constant C and
every j ∈ Z. Then the series
(Av)i = ∑ ai j v j
j∈Z
is absolutely convergent for every i ∈ Z and for every v = (v j : j ∈ Z) ∈ l p = L p (Z, ]),

A
where ] is the countiny measure and 1 ≤ p ≤ ∞. Moreover the operator v = (v j ) →
w = ((Av )i ) is bounded in l p and kAv kl p ≤ Ckvkl p .
Proof. Apply Theorem 7.1 with X = Z and µ = ], the counting measure on Z
7.2 Applications of Cotlar’s Lemma 195
7.2 Applications of Cotlar’s Lemma
Let us start this section with the original example of the Hilbert transform in a
somehow new point of view which reflects the role of the scaling group as the space
(Y, ν) in the general approach given in the previous section.
Let H be the space L2 (R) with the standard Lebesgue measure. Let us write k · k2
to denote the norm in L2 (R) and k · k to denote operator norms. Let (Y, ν) be the
multiplicative group R+ with its standard Haar measure, in other words Y = R+ and
dν = dtt .
Let τ : R+ → B(L‘2, L2 ) be given by
Z
(τt f )(x) = kt (x − y) f (y) dy
R
(kt ∗ f )(x),
where
1 z
kt (z) = k1
t t
and
X[1,2) (z)
k1 (z) = .
z
X (z)
In other words, kt (z) = [t,2t)z . Notice that each kt is in L1 (R) and that kkt k1 =
2 log 2 for every t > 0. So that kt ∗ f belongs to L2 (R) for each f ∈ L2 (R). Hence
τt f is well defined for each t > 0 as a bounded linear operator from L2 (R) to L2 (R).
In orther to check the basic properties of the function τ, required to apply the gener-
alized Cotlar’s Lemma, let us start by noticing that the family EN = ( N1 , N); N ∈ N
of subsets of R+ is a sequence of sets of finite measure which cover R+ : ν(EN ) =
R N dt 2
1/N t = 2 log N. Let f and g be two L (R) functions, then the function
Z
t → hτt ( f ), gi2 = (kt ∗ f )(x)g(x) dx
R
not only is measurable in (R+ , dtt ) but also continuous of t > 0. In fact
Z
hτt ( f ), gi2 − hτs ( f ), gi2 = (kt − ks ) ∗ f (x)g(x) dx
R
≤ k(kt − ks ) ∗ f k2 kgk2 ≤ kkt − ks k1 k f k2 kgk2
On the other hand

1 z 1 z
Z
kkt − ks k1 = k1 ( ) − k1 ( )| dz
R t t s s
1 1 1 z z
Z Z
≤ t| − ||kt (z)| dz + |k1 ( ) − k1 ( )| dz
R t s s R t s
t 1 dz
Z
≤ |1 − | + X (z) − X[s,2s) (z) ,
s s R [t,2t) z
which clearly tends to zero when s tends to t.
On the other hand kτt k ≤ kkt k1 = 2 log 2, so that kτt k is in L2 (EN , dtt ) for each
N ∈ N. So far, we are done with the general hypotheses of Cotlar Lemma. Let us
check that there exists a function γ and a constant A satisfying the quantitative esti-
mates required to apply the lemma. The conclusion shall finally be the uniform (in
N) boundedness of the operator norm of EN τt dtt . After that we shall compare this
R
integral with the Hilbert transform.
Lemma 7.1.
(a) τt∗ is the integral operator with kernel kt∗ (z) = kt (z);
(b) kτt∗ τs k ≤ kkt∗ ∗ ks k1 andpkτt τs∗ k ≤ kkt ∗ ks∗ k1 = kkT∗ ∗ ks k1 ;
(c) the function γ(t, s) = kkt∗ ∗ ks k1 satisfies the summability required, more
over
R dt R ds R dµ
R+ γ(t, s) t = R+ γ(t, s) s = R+ γ(µ, 1) µ are finite.
7.3 An extension of Cotlar’s Lemma
Sometimes the family of operators to be added is indexed an a more general set than
the integers. When the index set is a measure space the method still works to provide
the boundedness of integral of families of operators instead of sums.
Let (Y, ν) be a σ -finite positive measure space. Let H be a separable Hilbert
space with scalar product h · iH and norm k kH . Set B(H, H) to denote the space of
linear and continuous operators on H. Let
τ : Y → B(H, H)
be the given family of operators on H indexed by Y . In other words, for each y ∈

Y, τ(y) is a linear and bounded operator on H. The norm in B(H, H) will be denoted
by k · k.
We shall assume that τ is measurable and that τ belongs to L1 (E, dν) for every
ν measurable subset E of Y with ν(E) < ∞. Precisely
(a) for each choice of f and g ∈ H we require the measurability of the scalar
valued function
y → hτ(y) f , giH ;
7.3 An extension of Cotlar’s Lemma 197
(b) kτ(y)k ∈ L1 (E, ν) for each measurable E ⊂ Y with ν(E) < ∞.

R
It is easy to see that (a) implies the measurability of kτ(y)k, so that E kτ(y)k dν(y)
is well defined.
Under the above
R
assumptions, for each ν-measurable E ⊂ Y with ν(E) < ∞, the
operator TE = E τ(y) dν(y)R
is well defined as a member of B(H, H). In fact, the
bilinear form b( f , g) = E hτ(y) f , giH dν(y) is continuous, since
hτ(y) f , giH ≤ kτ(y) f kH kgkH

≤ kτ(y)kk f kH kgkH ,
so that
Z Z
hτ(y) f , giH dν(y) ≤ kτ(y)k dν(y) k f kH kgkH
E E
Hence, for f ∈ H fixed, the linear form b( f , ·) is continuous on H and Riesz
representation theorem gives a unique element h = h( f ) ∈ H such that
b( f , g) = h( f ), giH
R
for every g ∈ H and kh( f )kH = kb( f , ·)k ≤ E kτ(y)k dν(y) k f kH .
Since it is also clear that the application
R
f → h( f ) is linear, we have that it be-
long
R
to B(H, H). We
R
shall denote it by E τ(y) dν(y) or by TE . Notice also that
k E τ(y) dν(y)k ≤ E kτ(y)k dν(y) as it could be expected.
R
Of course, when Y = N and ν is the combining measure, {1,...,N} τ(y) dν(y) cab
be written as ∑Ni=1 Ti where Ti = τ(i).
Theorem 7.2. Let H be a separableSHilbert space. Let (Y, ν) be a ν-finite positive

measure space. Let γ : Y ×Y → R+ {0} be a measurable function such that
nZ Z o
A = sup max γ(z, y) dν(z), γ(y, z) dν(z) < ∞.
y∈Y Y Y
Assume that τ : Y → B(H, H) is a given operator valued function on Y which is

integrable on each subset E of Y with finite measure. If kτ ∗ (y)τ(z)k ≤ γ 2 (y, z) and
kτ(y)τ ∗ (z)k ≤ γ 2 (y, z), then the inequality
Z
k τ(y) dµ(y)k ≤ A
E
holds uniformly on the family of all measurable subsets E of Y with ν(E) < ∞.
Proof. Once the formal aspects regarding the composition and basic algebra of op-
erators defined as integrals are solved, the proof follows the lines of Theorem 7.1.
Set TE = E τ(y) dν(y). Notice that TE∗ = E τ ∗ (y) dν(y) and that if S (y) is an-
R R
other operator valued function with the same properties as τ(y), we have that
Z Z
τ(y) dµ(y) S (y) dµ(y)
E E
as a composition of operators coincides with
Z Z
τ(y)S (z) d(ν × ν)(y, z)
E×E
as an operator valued integral on E × E.
Both properties follow from the uniqueness in Riesz representation. Hence, since
1 1
kTE k = (TE TE∗ ) 2k 2k+1 for every k ∈ N, we write
Z Z
(TE TE∗ )M = ··· τ(y1 )τ ∗ (y2 )τ(y3 ) . . . τ ∗ (y2M ) dν 2M (y)
y∈E 2M
where, for simplicity we wrote y = (y1 , y2 , . . . , y2M ); E 2M = E ×. . .×E 2M times and

ν 2M = ν ⊗. . .⊗ν 2M times as well. The norm of the operator τ(y1 )τ ∗ (y2) . . . τ ∗ (y2M )

can be estimated, for fixed y, in two ways associating the composition as τ(y1 )τ ∗ (y2 ) . . . τ(y2M−1 )τ ∗ (yM )

first and then

τ(y1 ) τ ∗ (y2 )τ(y3 ) . . . τ(y2M−2 )τ(y2M−1 ) τ ∗ (y2M )

Taking the geometric mean of the two estimates and using the quasi-orthogenality
hypothesis we get that
M Z Z
k TE TE∗ k ≤ · · · Aγ(y1 , y2 )γ(y2 , y3 ) . . . γ(y2M−1 , y2M ) dν 2M (y)
y∈E 2M
Since the function in the above integral is nonnegative and measurable an (Y, ν)
is ν-finite, Tonelli’s theorem allow us repeated integration.
If we integrate first with respect to y2M and use the uniform integrability proper-
ties of γ, we get
Z Z
k(TE TE∗ )M k ≤ A2 ... γ(y1 , y2 ) . . . γ(y2M−2 , y2M−1 ) dν(y2M−1 ) . . . d(y1 ).
y1 ∈E y2M−1 ∈E
Repeating this argument with y2M−1 y2M−2 up to y2 we get

Z
k(TE TE∗ )M k ≤ A1+2M−1 dν(y1 ) = A2M ν(E)
E
So that
7.4 Krein Lemma and Wittman’s method 199
k 1 k
1
k+1
kTE k = k(TE TE∗ )2 k 2k+1 ≤ A2·2 ν(E) 2
1
k+1
= A ν(E) 2 ,
for every k ∈ N. Hence kTE k ≤ A.
Notice that the above result contain an extension of Cotlar’s lemma even for the
discrete case since the bounding sequence γ needs not to be in “convolution” form.
7.4 Krein Lemma and Wittman’s method
Let us start by introducing the abstract setting. Let H be a real Hilbert space with
inner product h·iH and norm k kH . Let (B, k · kB) be a Banach space satisfying the
following properties;
i) B ⊂ H,
ii) B̄H = H (B is dense in H),
iii) B ,→ H, kX kH ≤ CkX kB , x ∈ B.
Theorem 7.3. (Krein’s Lemma). Let H and B as before. Let T : B → B be a bounded

linear operator on B such that hT x, yiH = hx, TyiH for every x and every y in B. Then
T has a unique bounded and linear extension T̃ to H with
kT̃ xkH ≤ kT kB kX |H ,
for every x ∈ H, where for simplicity we use the notation kT kB . With this notation
the inequality reads kT̃ kH ≤ kT kB .k
Proof. Since when T = 0 there is nothing to prove, we may assume by substituting

T by kTTkB , that kT kB = 1. Hence, we are assuming that kT xkB ≤ kxkB . We have to
prove that for x ∈ B with kX kH = 1 we have that kT xkH ≤ 1. Once this is proved
T̃ can be defined as the unique extension of T to H by density of B.
Take, then x ∈ B with kX kH = 1 amd defome vn = hT n x, T n xiH , n ∈ N.
Let us now consider the quadratic form kT n−1 x + λ T n+1 xk2H in λ ∈ R. Then
0 ≤ kT n−1 x + λ T n+1 xk2H = hT n−1 x + λ T n+1 x, T n−1 x + λ T n+1 xiH

= hT n−1 x, T n−1 xiH + 2λ hT n−1 x, T n+1 xiH + λ 2 hT n+1 x, T n+1 xiH
= vn−1 + 2λ hT n x, T n xiH + λ 2 vn+1
= vn−1 + 2λ vn + λ 2 vn+1 .
Hence Hv2n − 4vn+1 vn−1 ≤ 0 or v2n ≤ vn+1 vn−1 .

Since T 0 is the identity we have that v0 = hT 0 x, T 0 xiH = hx, xiH = kxk2H = 1.

Then
v2
v21 ≤ v0 · v2 ; v1 ≤ (v1 6= 0)
v1
Now
v2 v3 v2 v3
v22 ≤ v1 · v3 , then ≤ (v1 6= 0) and v1 ≤ ≤ .
v1 v2 v1 v2
Notice that if v2 = 0, then v1 = 0. So that, under the assumption v1 6= 0 we get
the following chain of inequalities
v2 v3 v4 vn
v1 ≤ ≤ ≤ ≤ ... ≤ .
v1 v2 v3 vn − 1
And
vn−1
1 = v1 · v1 · · · · · v1
v2 v3 vn
≤ · ·····
v1 v2 vn−1
vn
= ,
v1
or vn1 ≤ vn , which certainly implies v ≤ lim infn→∞ (vn )1/n .

Now, since the inclusion of B in H is continuous, we have that
vn = kT n xk2H
≤ C2 kT n xk2B
≤ C2 kT k2B nkxk2B ,
≤ C2 kxk2B
then
2/n
(vn )1/n ≤ C2/n · kxkB .
Hence v1 ≤ lim inf(vn )1/n ≤ 1.
Since v1 = kT xk2H we are done.
Corollary 7.2. Let H and B as before. Assume that Ti , i = 1, 2, are two linear and
bounded operators on B with kTi xkB ≤ ci kxkB , i = 1, 2 and
hT1 x, yiH = hx, T2 yiH , x, y ∈ B.

√
Then kTi xkH ≤ c1 c2 kxkH , x ∈ B, i = 1, 2.
Proof. We shall apply Fhm ( . ) to the operator T = T1 T2 . Notice that T : B → B is

linear and bounded with kT kB ≤ c1 c2 . Also
hT x, yi = hT2 x, T2 yiH = hT2 y, T2 xiH

= hT1 T2 y, xiH = hx, TyiH .
Hence, from Theorem 7.1 we have that kT xkH ≤ c1 c2 kxkH , x ∈ B. In a similar

way we obtain that
kSxkH ≤ c1 c2 kxkH , x ∈ B,
where S = T2 T1 .
Then
kT1 xk2H = hT1 x, T1 xiH

= hT2 T1 x, xiH
≤ kT2 T1 xkH · kxkH
≤ c1 c2 kxk2H .
We can proceed similarly with T2 .
As an application of the above results to the L2 boundedness of singular integrals

we shall state and prove the next result which is due to R. Wittmann [?].
For 0 < α ≤ 1 we shall denote by Λ α the space of all bounded real functions on
Rn that satisfy the condition
| f (x) − f (y)| ≤ C|x − y|α

for every x and y ∈ Rn and some constant C. The norm in Λ α is k f kΛ α = k f k∞ +
k f kα where k f kα = supx6=y | f (x)− f (y)| α
|x−y|α . By Λ0 we shall denote the class of all the
functions in Λ α with compact support.
Theorem 7.4. (R. Wittmann) Assume that for some 0 < α ≤ 1 we have two linear
operators S and T from Λ0α to Λ α satisfying the following conditions:
R R
(a) f(y)Sg(y) dy = Rn T f (y)g(y) dy for every f and g ∈ Λ0α ;
Rn
(b) max kS f kα , kT f kα ≤ Ck f kα for every f and f ∈ Λ0α ;
α
(c) max kS f k∞ , kT f k∞ ≤ C diam(supp f ) k f kα for every f and f ∈ Λ0α ;
then
kS f k2 ≤ Ck f k2 and
kT f k2 ≤ Ck f k2 for every f ∈ Λ0α .
The proof of Theorem 7.4 is based on the next construction which gives a basic
tool for localization in Lipschitz classes.
Lemma 7.2. Let η : R+ → R be a non-increasing C ∞ function with

(
1 for t ≤ 12
η(t) =
0 for t ≥ 1
Set ϕ : Rn → R by ϕ(x) = η(|x|) and, for r > 0, ϕ r (x) = ϕ( xr ). Then
(1) 0 ≤ ϕ r ≤ 1;
(2) ϕ r ≤ ϕ s for r ≤ s;
(3) {ϕr 6= 0} ⊆ {ϕ2r = 1};
n
r>0 {ϕr = 1} = R ;
S
(4)
(5) r
|sopϕ | ≤ cn r ;n
(6) ϕr ∈ Λ0α with kϕr kα ≤ cr−α .
Proof. Fix r > 0. Define Br = {g = ϕ r f : f ∈ Λ α }. With the norm kgkBr = kgk∞ +

rα kgkα , Br becomes a Banach space. Set Hr to denote the L2 (Rn ) closure of Br .
Equipped with the restriction of the scalar product and norm of L2 (Rn ) the space Hr
is a Hilbert space that contains Br . Moreover the inclusion of Br in Hr is continuous.
In fact, for g ∈ Br
Z Z
kgk22 = |g(y)|2 dy = |g(y)|2 dy
Rn B(0,r)
≤ kgk2∞ |B(0, r)|

≤ crn kgk2Br .
For the given operators S and T and for the given r > 0 define
Sr , Tr : Br → Br by
Sr (g) = ϕr (S(ϕr g)) and Tr (g) = ϕr (T (ϕr g)). Since all the function involved are
continuos and with compact support, from (a) we obtain for Sr and Tr the duality
required in Corollary 7.2. In fact
Z Z
Sr (g)(x)h(x) dx = S(ϕr g)(x)ϕr (x)h(x) dx
Rn n
ZR
= ϕr (y)g(y)T (ϕr h)(y) dy
n
ZR
= g(y)Tr (h)(y) dy
R
Sr and T are linear and bounded on Br . We shall only deal with Sr . Of course,
R r
for a, b R and g1 and g2 ∈ Br we have that
Sr (ag1 + bg2 ) = ϕr S(ϕr (ag1 + bg2 ))

= ϕr S(aϕr g1 + bϕr g2 )

= ϕr aS(ϕr g1 ) + bS(ϕr g2 )
= aSr (g1 ) + bSr (g2 ).
Let us check the boundedness of Sr as an operator on Br .

Let us start by estimating the L∞ norm of Sr (g) using (c):
kSr (g)k∞ = kϕr S(ϕr g)k∞ ≤ kS(ϕr g)k∞

≤ Crα kϕr gkα

≤ Crα kϕr k∞ gkα + kϕr kα kgk∞

≤ C rα kgkα + kgk∞
= CkgkBr
Let us now estimate the Lipα norm of Sr (g) using (b) and (c);
kSr (g)kα = kϕr S(ϕr g)kα

≤ kϕr kα kS(ϕr g)k∞ + kϕr k∞ kS(ϕr g)kα
≤ Cr−α kϕr gkα rα +Ckϕr gkα
≤ Ckϕr gkα

≤ C kϕr k∞ kgkα + kϕr kα kgk∞
≤ C kkgkα + r−α kgk∞ .

Then
kSr (g)kBr = kSr (g)k∞ + rα kSr (g)kα

≤ CkgkBr +C r[αkgkα + kgk∞

≤ C rα kgkα + kgk∞
≤ CkgkBr .
Since analogous estimates hold true for T instead of S, we are in position to apply
the corollary to the theorem of Krein in order to get that there exists a constant C > 0
such that
kSr gkHr ≤ CkgkHr , and
kTr gkHr ≤ CkgkHr ,

for every g ∈ Br and every r > 0. By property (4) of the family {ϕ r } we obtain
the desired inequality by leting r → +∞ when g is a smooth function with compact
support.
Application: the case of the Hilbert transform. It is easy to see that the Schwartz
distribution defined by
1
Z
ϕ(y)
hv· · , ϕi = lim dy
x ε→0 ε≤|y| y
extends from ϕ ∈ S (R) to ϕ ∈ Λ0α for any 1 ≥ α > 0. Hence the Hilbert transform
H f (x) = lim Hε f (x)

ε→0
with
f (y)
Z
Hε f (x) = dy,
|x−y|≥ε x−y
1
is well defined on Λ0α (R) for every x ∈ R. In other words, the convolution of v.p. x
with a function of Λ0α is well defined and H f = v.p. 1x ∗ f .
We shall prove that S = H and T = −H satisfy the conditions of Theorem ( . )
for some α. Notice that (b) and (c) imply that H f ∈ Λ α for f ∈ Λ0α . Let us start by
(c). Of course we only have to get the estimate for S = H. Since H is of convolution
type and since the L∞ (R) norm is translation invariant we can assume that f ∈ Λ0α
and that O ∈ ∂ (supp f ), the boundary of the support of f . Set R = diam supp f .
Then
supp f ⊂ [−R, R] = I(0, R)

and
f (0) = 0.
Suppose first that x ∈
/ I(0, 2R) = (−2R, 2R).
Then
f (y)
Z
S f (x) = H f (x) = dy,
y∈R − y
x
because f vanishes around x and there is not singularity in the integral.
In this situation we have that
f (y)
Z R
|S f (x)| ≤ dy
−R x−y
R dy
Z
≤ k f k∞
−R x − y

α
≤ C r kgkα + kgk∞
≤ Ckgk∞ .
But for z ∈ supp f | f (z)| := | f (z)− f (0)| ≤ k f kα |z|α ≤ C(diam supp f )α k f kα

which has the desired form for (c) in Theorem ( . ). Assume now that x ∈ I(0, 2R) =
(−2R, 2R). Then
f (y) − f (x)
Z
H f (x) = dy
4−R>|x−y| x−y
(gráfico)
Now
Z
|H f (x)| ≤ k f kα |x − y|α−1 dy
|x−y|<4R
= C(diam supp f )α k f kα ,
as desired.
To check (b) we only need to obtain an estimate of the form
kH f kα ≤ Ck f kα for f ∈ Λ0α .
Let x 6= y be two given points in R. Take ε > 0 small enough to have ε < 21 |x − y|
and ε1 > 2|x − y|. If we write H ε to denote the truncated Hilbert transform given by
f (z)
Z
H ε f (x) = dz
ε≤|x−z|<ε −1 x−z
from the basic properties of the Hilbert kernel, we have that
f (z) f (z)
Z Z
H ε f (x) − Hε f (y) = dz − dz
ε≤|x−z|<ε −1 x − z ε≤|y−z|<ε −1 y − z
f (z) − f (x) f (z) − f (y)
Z Z
= dz − dz
ε≤|x−z|<ε −1 x−z ε≤|y−z|<ε −1 y−z
f (z) − f (x) f (z) − f (y)
Z Z
= dz − dz+
ε≤|z−x|<2|x−y| x−z ε≤|z−y|<2|x−y| y−z
Z 1 1
+ − ( f (z) − f (x)) dz+
2|x−y|≤|z−x|<ε −1 x − z y−z
Z 1 1
Z
+ ( f (z) − f (x)) dz − ( f (z) − f (x)) dz
2|x−y|≤|z−x|<ε −1 y − z 2|x−y|≤|y−z|<ε −1 y − z
1
Z
+ ( f (x) − f (y)) dz = I + II + III + IV +V.
2|x−y|≤|y−z|<ε −1 y − z
Notice that V = 0 and that I and II share the same pattern. Let us estimate the
absolute values of I, III and IV .
Z
|I| ≤ k f kα |x − z|α−1 dz ≤ Ck f kα |x − y|α .
ε≤|z−x|<2|x−y|
For III we shall use that if 2|x − y| ≤ |z − x|, as in the domain of III, then
1
|x − z| ≤ |x − y| + |y − z| ≤ |x − z| + |y − z|,
2
hence |x − z| ≤ 2|y − z|. So that
y−x
Z
|III| ≤ k f kα |x − z|α dz
2|x−y|≤|z−x|<ε −1(x − z)(y − z)
|x − z|α
Z
≤ 2k f kα |x − y| 2
dz
2|x−y|≤|z−x| |x − z|
For 0 < α < 1 we have that
(III) ≤ Ck f kα · |x − y|α .
Let us, finally, get a bound for (IV ).
Notice that
1
Z
|IV | ≤ | f (z) − f (x)| dz,
∆ |y − z|
where ∆ is the symmetric difference of the sets {z : 2|x − y| ≤ |z − x| < ε −1 } and
{z : 2|x − y| ≤ |z − y| < ε −1 }. Assume that y > x. Then
7.5 An application of Cotlar’s lemma to regularization of Haar basis 207
Z y− 1 Z x−|x−y| Z y+2|x−y| Z y+ 1
ε ε
|IV | ≤ + + +
x− ε1 x−2|x−y| y+|x−y| x+ ε1
= Aε + B +C + Dε
Let us estimate C, B follows · · · the same way.
Z y+2|x−y| Z y+2|x−y|
1 |z − xα |
| f (z) − f (x)| dz ≤ k f kα dz
y+|x−y| |y − z| y+|x−y| |y − z|
Z y+2|x−y|
α dz
≤ Ck f kα |x − y|
y+|x−y| |y − z|
Z 2|x−y|
α dµ
= Ck f kα |x − y| = C|x − y|α k f kα .
|x−y| µ
which is the desired estimate. For Dε and Aε the direct estimate seems to be worse
Z y+ 1
ε 1
Dε = | f (z) − f (x)| dz
x+ ε1 |y − z|
Z y+ 1
ε 1
≤ 2k f k∞ dz
y−|x−y|+ ε1 |y − z|
1/ε 1
= 2k f k∞ log = 2k f k∞ log
1/ε − |x − y| 1 − ε|x − y|
which tends to zero as ε → 0. Hence
|H f (x) − H f (y)| = lim |H ε f (x) − H ε f (y)|

ε→0
≤ Ck f kα |x − y|α .
7.5 An application of Cotlar’s lemma to regularization of Haar

basis
Let us recall briefly some basic concepts of bases in Hilbert spaces. If B is a Banach
space, a sequence e : N → B is said to be SCHANDER BASIS for B if for every
f ∈ B there exists a unique sequence of scalars C : N → R such that f = ∑∞ k=1 Ck ek ,
where the convergence occurs in the norm of B.
If H is a Hilbert space, a sequence b : N → H is said to be a BESSEL SEQUENCE
WITH BOUND B if the inequality
∞
2
∑ h f , bk i ≤ Bk f k2H (7.2)
k=1
holds for every f ∈ H.

A sequence e : N → H is said to be an ORTHONORMAL BASIS FOR H if and
only if e is a Schauder basis for H and e is ORTHONORMAL:
hek , e j i = δ jk
A sequence r : N → H is a RIES BASIS FOR H if there exist U : H → 1 linear,
bounded,
S
one-to-one and onto and e : N → H an orthonormal basis for H such that
r = e.
A system g : N → H is COMPLET if the linear span of {gk : k ∈ N} is dense in
H.
Let us state as a theorem that can be found in [C] the following two basic results
regarding Riesz bases which shall be useful in our further application.
Theorem 7.5. Let H be a separable Hilbert space.

(A) r : N → H is a Riesz basis for H if and only if r is complete and there exist two
positive constants A y B such that for every finite sequence of scalers c : N → R
with c j = 0 for every j ≥ N(c), the inequalities
A ∑ |ck |2 ≤ k ∑ ck rk k2 ≤ B ∑ |ck |2
hold. A and B are said to be Riesz bounds for r.

(B) If r : N → H is a Riesz basis for H with Riesz bounds A and B and b : N → H
is a Bessel sequence
q 2 with bound R < A, then r + b is a Riesz basis with Riesz
q 2
bounds A 1 − A and B 1 − RB
R
As a corollary of (B) we have that if e : N → H is a orthonormal basis for H and

b : N → H is a Bessel
√ sequence with √ bound R < 1 then e ± b are Riesz bases with
Riesz bounds (1 − R)2 and (1 + R)2 .
In this section we are interested in the construction of Riesz bases for L2 (R),
with Riesz bounds as close to one as desired, by regularization of the Haar basis.
This result follows from the next theorem by choosing ε small enough.
Theorem 7.6. Let H = L2 (R) and h(x) = X[0,1/2) (x) − X[1/2,1] (x)R the basic Haar
function in R. For ϕ ∈ C0∞ (R), even discreasing in (0, ∞) with R ϕ dx = 1, set
hε = h ∗ ϕε , ε > 0 and small. Then there exist positive constants α and β such that
bεjk = 2 j/2 b(2 j x − k), with bε (x) = h(x) − hε (x), is a Bessel sequence with bound
αε β . n o
Hence (hε ) jk : ( j, k) ∈ Z2 is smooth Riesz basis for L2 (R) with Riesz bounds
as close to one as desired.
Proof. We have to prove that for each f ∈ L2 (R)

D E 2
∑ f , b2jk ≤ αε β k f k22 .
( j,k)∈Z2
It is enough to show that for each “rectangle” of the form A = A1 × A2 ⊂ Z × Z

we have the inequality
D E 2
∑ f , bεjk ≤ αε β k f k22 ,
( j,k)∈A
with α and β independent of A, f and ε. Let T be the operator on L2 (R) defined by

D E
T f = ∑ f , bεjk hkj .
A
It is clear, since A is finite, that T is linear and bounded as an operator on L2 (R).

2)
Since h(hkj ) is an orthonormal
D
2 2
Ebasis for L (R), the L (R) norm of T f is the `
2(Z
norm of the sequence f , bεjk . In fact

D E 2
kT f k22 = ∑ f , bεjk .
( j,k)∈A
Hence, in order to get the result, it is enough to prove that
kT f k22 ≤ α 2 ε 2β k f k22 .
The operator T can be composed as a sum of operators T j acting at the same scale
2− j . In fact
T= ∑ Tj ,
j∈A1
with
Tj f = ∑ < f , bcjk > hkj for j ∈ A1 .

k∈A2
We shall look at the sequence T j with (under?) the light of Cotlar’s Lemma. It is
easy to check that the adjoint T j∗ of T j is given by
T j∗ g = ∑ < g, hkj > bεjk .

k
We have to estimate the norms of the operators Ti T j∗ and Ti∗ T j . It is clear from
the cualitative properties of hkj and bεjk that now the operators T − j and T j∗ are quite
different, so the operators T j and T do not look like standard singular integrals.
Some simplifications and reduction due to the self similarity √of the context are
in order. Let ∆t denote the isometry in L2 (R) given by ∆t f (x) = tg(tx)t > 0. The
following properties are easy to check.
(1) (∆t )∗ = ∆t −1
(2) T j = ∆2 j T0 ∆2− j
(3) T j∗ = ∆2 j T0∗ ∆2− j

(4) Ti T j∗ = ∆2i T0 T j−i
∗ ∆ ∗
2−i = ∆ 2 j Ti− j T0 ∆ 2− j
(5) ∗ ∗ ∗
Ti T j = ∆2 j Ti− j T0 ∆2− j = ∆2i T0 T j−i ∆2−i
(6) kTi T j∗ k = kT0 T j−i
∗ k = kT T ∗ k
i− j 0
(7) ∗ ∗
kTi T j k = kT0 T j−i k = kTi− ∗ T k
j 0
(8) ∗ ∗
kT0 T j k = k(T0 T j ) k = kT j T0∗ k = kT0 T−∗ j k.
∗
Hence to apply Cotlar’s Lemma to our current situation we have to show that
(a) kT0∗ T j k ≤ aε b γ( j), j ∈ Z
(b) kT0 T j∗ k ≤ aε b γ( j), j ∈ Z
p positive constants a and b and for some sequence γ( j) such that A =

for some
∑ j∈Z γ( j) < ∞.
The estimate in (a) is easy. In fact
T0∗ T j ( f ) = ∑ T j f , h0k bε0k

k∈A2
= ∑ ∑ f , bεj` h`j , h0k bε0k

k∈A2 `∈A2
Since (h`j ) is an orthonormal system then T0∗ T j ( f ) = 0 if j 6= 0. If j = 0, then
T0∗ T0 ( f ) = ∑ f , bε0k bε0k .

k∈A2
Since kT0∗ T0 k = kT0 k2 we have to prove that kT0 k tends to 0 for ε → 0. But T0 f =
2
∑k∈A2 f , bε0k h0k , so that kT0 f k2 = ∑k∈A2 f , bε0k .
2
Hence kT0 f k2 ≤ ∑k∈Z f , bε0k . Now
Z
f , bε0k = f (x)(h − hε )(x − k) dx
R
Z
= f (x − k)(h − hε )(x) dx
R
Z ε
= f (x − k)(h − hε )(x) dx+
−ε
Z 1/2+ε
+ f (x − k)(h − hε )(x) dx+
1/2−ε
Z 1+ε
+ f (x − k)(h − hε )(x) dx
1−ε
and h − hε is bounded by 1. Each one of the above three terms on the right hand side
can be handled in the same way. Let us consider only the first one:
Z ε Z ε
f (x − k)(h − hε )(x) dx ≤ f (x − k) dx
−ε −ε
Z ε 1/2
2
≤ f (x − k) dx (2ε)1/2
−ε
Z k+ε 1/2
2
= f (x) dx (2ε)1/2 .
k−ε
So that
Z k+ε
2 2
f , bε0k ≤ Cε f (x) dx
k−ε
and
Z k+ε
2
kT0 f k2 ≤ C.ε ∑ f (x) dx
k k−ε
≤ Cεk f k22 ,
as desired.
The estimate in (b) is not so simple. From the above considerations we only have
to deal with the case j ≥ 1. For a function g ∈ L2 (R) we shall estime kT0 T j∗ gk22 .
From the definitions of T0 and T j∗ we have that
T0 T j∗ g = T0 (T j∗ g) = ∑ T j∗ g, bε0k h0k
k∈A2
= ∑ ∑ g, h`j bεj` , bε0` h0k

k∈A2 `∈A2
= ∑ ak h0k ,
k∈A2
where
ak = ∑ g, h`j bεj` , bε0k .

`∈A2
Since {h0k } is orthonormal,
kT0 T j∗ gk22 = ∑ |ak |2

k∈A2
2
= ∑ ∑ g, h`j bεj` , bε0k
k∈A2 `∈A2
2
≤ ∑ ∑ g, h`j bεj` , bε0k
k∈Z `∈Z
Notice that
Z
bεj` , bε0k = bεj` (x)bε0k dx
R
Z
= 2 j/2 bε (2 j x − `)bε (x − k) dx
R
Z
= 2 j/2 bε (2 j y − (` − 2 j ))bε (y) dy
R
= bεj(`−2 j k) , bε00 .
So that, since j ≥ 1, 2 j k ∈ Z and
kT0 T j∗ gk22 ≤ g, h`j

∑ ∑ bε j , bε00
j(`−2k )
k∈Z `∈Z

∑ ∑ g, h`+2 j k bεj` , bε00
k∈Z `∈Z
In order to deal with the inner sum, we shall partition Z in three sets
A = {` ∈ Z : supp bε00 ∩ supp bεj` =},
B = {` ∈ Z :supp bε00 ∩ supp bεj` 6=},

1
and supp bεj` ∩ {0, , 1} =
2
and
C = Z(A ∪ B).
A picture of bε00 can be of some help.
f
−8 −6 −4 −2 0 2 4 6 8
With the above notation we have
2
kT0 T j∗ gk22 ≤ ∑ ∑+∑+∑
k `∈A `∈B `∈C
2
=∑ ∑+∑
k `∈B `∈C
2 2
≤ 2∑ ∑ +2∑ ∑
k `∈B `∈C
= I + II.
Estimate for I: In this case ` ∈ B. We shall estimate the cardinality of B and we

shall get a sharp estimate for | < bεj` , bε00 > | which is given by the facts that bε00 is
smooth on the support of bεj` and that the mean value of bεj` vanishes. Since j ≥ 1 is
/ has at most Cε2 j elements.
fixed and ε is small then the set {` : supp bε00 ∩ bεj` 6= 0}
In order to estimate | < bεj` , bε00 > | we observe that the support of bεj` has three
connected components I εj` , J εj` and K εj` , each one of them of length 2ε2− j . We shall
only deal with I εj` , the other two J εj` and K εj` can be handled in a similar way. An im-
portant additional property of the sequence bεj` is that not only they have vanishing
integrals, but also the integral of bεj` on I εj` , on J εj` and on K εj` are all of them equal
to zero. Set x j` to denote the center of I εj` . Then
Z
| bεj` (x)bε00 (x) dx| =
I εj`
Z
=| (bε00 (x) − bε00 (x j` ))bεj` (x) dx| ≤
I εj`
Z
≤ |bε00 (x) − bε00 (x j` )||bεj` (x)| dx|
I εj`
1
Z
≤ |x − x j` ||bεj` (x)| dx
ε I εj`
2ε2− j
Z
≤ |bεj` (x)| dx
ε I εj`
≤ 2 · 2− j · 2 j/2 |I εj` |
3j
= C · ε · 2− 2 .
Hence, for ` ∈ B we have that
j 3j 2
−2
I ≤ 2∑ ∑ | < g, h`+2 j k > |C.ε.2
k `∈B
j 2
≤ C.ε .2−3 j ∑ 2
∑ | < g, h`+2 jk > |
k `∈B
j
≤ C.ε 2 .2−3 j ∑ 2

∑ | < g, h`+2 jk > | ](B)
k `∈B
Z j
≤ C.ε 2 .2−2 j ∑ ∑ |g|2 dx kh`+2 j k k22

k `∈B I`+2 j k
Z
= C.ε 3 .2−2 j ∑ ∑ XI`+2 |g(x)|2 dx

j (x)
R jk
`∈B k
Z
≤ C.ε 3 .2−2 j .](B) |g(x)|2 dx
R
≤ C.ε 4 .2− j kgk22 ,
which has the desired structure.

Estimate for II: The basic fact here is that the number of elements of the set C
has a uniform upper bound. In fact, since
1
C ⊆ {` ∈ Z : {0, , 1} ∩ supp bεj` 6= 0}
/
2
1
⊆ {` ∈ Z : 0 ∈ supp bεj` } ∪ {` ∈ Z : ∈ supp bεj` } ∪ {` ∈ Z : 1 ∈ supp bεj` }
2
and since ∑` Xsupp mε (x) ≤ 2 for every x ∈ R, none of the last three sets contains
j`
more than two elements, hence ](C) ≤ 6.
On the other hand since the discontinuity points of bε00 may now belong to the
support of bεj` , the smoothness argument for the estimate for I is no longer true.
But the uniform boundedness of ](C) allows now the use of a worse estimate for
| < bεj` , bε00 > |.
By Schwartz inequality we have that
Z
| < bε00 , bεj` > | ≤ |bε00 (x)||bεj` (x)| dx
R
Z
|bεj` (x)| dx
R
Z 1/2
|bεj` (x)|2 dx |supp bεj` |1/2
R
√
C ε2− j/2 .
So that, by Schwartz for sums and Bessel inequality for subsequences of the Haar
system,
j ε ε
2
II = 2 ∑ ∑ | < g, h`+2 j k > || < b j` , b00 > |
k `∈C
−j j 2
≤ Cε2 ∑ ∑ | < g, h`+2 jk > |
k `∈C
−j j 2

≤ Cε2 ∑ ∑ | < g, h`+2 jk > | ](C)
k `∈C
j
≤ Cε2− j 2

∑ ∑ | < g, h`+2 jk > |
`∈C k
≤ Cε2− j ](C)kgk22
≤ Cε2− j kgk22 ,
which has also the required structure.

Appendix A
Chapter Heading
All’s well that ends well
Use the template appendix.tex together with the Springer document class SVMono
(monograph-type books) or SVMult (edited books) to style appendix of your book
in the Springer layout.
A.1 Section Heading
Instead of simply listing headings of different levels we recommend to let every

heading be followed by at least a short passage of text. Furtheron please use the
LATEX automatism for all your cross-references and citations.
A.1.1 Subsection Heading

LATEX automatism for all your cross-references and citations as has already been
described in Sect. A.1.
For multiline equations we recommend to use the eqnarray environment.
a×b = c
a×b = c (A.1)
A.1.1.1 Subsubsection Heading

217
218 A Chapter Heading
Fig. A.1 Please write your

figure caption here
LATEX automatism for all your cross-references and citations as has already been
described in Sect. A.1.1.
Please note that the first line of text that follows a heading is not indented,
whereas the first lines of all subsequent paragraphs are.
Table A.1 Please write your table caption here

Classes Subclass Length Action Mechanism
Translation mRNAa 22 (19–25) Translation repression, mRNA cleavage
Translation mRNA cleavage 21 mRNA cleavage
Translation mRNA 21–22 mRNA cleavage
Translation mRNA 24–26 Histone and DNA Modification
a Table foot note (with superscript)
Glossary
Use the template glossary.tex together with the Springer document class SVMono
(monograph-type books) or SVMult (edited books) to style your glossary in the
Springer layout.
glossary term Write here the description of the glossary term. Write here the de-
scription of the glossary term. Write here the description of the glossary term.
219

HugoAimar Book Distance and Measure Analysis and PDE

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

HugoAimar Book Distance and Measure Analysis and PDE

Uploaded by

Copyright:

Available Formats

Hugo Aimar

Instituto de Matemática Aplicada del Litoral-IMAL

Distance and Measure in

Place, month year Firstname Surname

Place(s), Firstname Surname

1 Topological, uniform and quasi-metric spaces. . . . . . . . . . . . . . . . . . . . . . 7

2 The weak homogeneity property on quasi-metric spaces. Assouad

2.4 Wiener type covering lemmas. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

3 Spaces of homogeneous type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

4 Some examples of quasi-metric measure spaces. . . . . . . . . . . . . . . . . . . . 143

4.3 Some basic examples from geometric measure theory. . . . . . . . . . . . . 161

5 The Hardy-Littlewood maximal function and the differentiation

6 Weighted norm inequalities for the Hardy-Littlewood maximal

7 Some techniques for the L2 boundedness of linear operators . . . . . . . . 193

A Chapter Heading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217

Electrostatic and gravitational potentials provide a natural starting point to introduce

E(x) = q [δx0 ∗ v](x) ,

δx0 ∗ v = (δx0 ∗ v1 , δx0 ∗ v2 , δx0 ∗ v3 )

The superposition principle allows us to say that a bounded system of electrical

|Φ(x)| ≤ I2 (| f |)(x) and |Ei (x)| ≤ I1 ( | f | )(x) ,

Φ = N ∗ T = α(N ∗ Du δ0 ) = α (Du N ∗ δ0 ) = αu · ∇N.

is providing an elementary identification of the Laplace differential operator ∆ =

euclidian settings. In particular, to the context of spaces of homogeneous type where

1.1 Topology, neighborhoods and bases.

This section aims only to fix notation and terminology.

Let X be a set. A given family τ, of subsets of X is a topology if it is closed

Let (X, τ) be a topological space. A subset F of X is called closed if its comple-

Given a topological space (X, τ), a family B of subsets of X is a basis for τ if

On the other hand, given a family B of subsets of X with X = B, then B is

Given a family S of subsets of X, with X =

1.2 The topology induced by the neighborhood system associated

In this section we give sufficient conditions on a general function

Let X be a set. If a topology τ on X is given then, the function that to each x in X

The next result is an exercise in Kelley’s “ General Topology ”.

Proposition 1.1. Given a set X and a function N : x → Nx , assigning to each

Bg (x, r) = { y ∈ X : g(x, y) < r}

Nx = U : Bg (x, r) ⊂ U for some positive r .

Proposition 1.2. Let X be a set and let g be a non-negative function defined on

τ = τNg = U : for every x ∈ U there exists r > 0 with Bg (x, r) ⊂ U

Two extreme trivial examples of the situation are in order.

1.3 Uniform spaces. Uniform topology.

We shall say that U is symmetric when U −1 = U. Let us denote by ∆ the identity

Definition 1.1. Let X be set. A family U of subsets of X×X is an uniform structure

If X is a set and U is an uniform structure on X, we say that (X, U ) is an uniform

Bd (r) = { (x, y) ∈ X×X : d (x, y) < r} ,

Definition 1.2. Given U , an uniform structure on X, a subfamily B of U is a basis

Every uniform structure on a set X induces a topology on X.

τ = { S ⊂ X : ∀ x ∈ S ∃ U ∈ U such that U (x) ⊂ S}

is a topology on X. If B is a basis for U , then

is a neighborhood basis for τ: every neighborhood of x contains some member of

Proposition 1.4. The non-empty family B of P (X× X) is a basis for an uniform

The uniform structure generated by B is

5. U = {U ∈ P (X×X ) : B ⊂ U for some B ∈ B}.

1.4 Metrizability of uniform spaces.

We shall say that an uniformity U on X is pseudo-metrizable (resp. metrizable)

Then, there exists a non-negative function ρ defined on X×X such that

4. ρ (x, z) ≤ ρ (x, y) + ρ (y, z) for every x, y and z ∈ X;

Moreover, if each Un is symmetric, then\the function ρ is a pseudo-distance on X

The function f has the following remarkable property

for every n ∈ Z and every choice of { x0 , x1 , ..., xn+1 } ⊂ X.

Let us observe that in particular, (1.1) is a triangle type inequality when n = 1.

Observe that in the case k < n, we also have that

Since of course x = x0 , x1 , ..., xm+k = z is a chain joining x and z, we have