Untitled

Grundlehren der
mathematischen Wissenschaften 348

A Series of Comprehensive Studies in Mathematics
Series editors
M. Berger P. de la Harpe N.J. Hitchin

A. Kupiainen G. Lebeau F.-H. Lin
S. Mori B.C. Ngô M. Ratner D. Serre
N.J.A. Sloane A.M. Vershik M. Waldschmidt
Editor-in-Chief
A. Chenciner J. Coates S.R.S. Varadhan
For further volumes:
www.springer.com/series/138
Dominique Bakry r Ivan Gentil r Michel Ledoux
Analysis and
Geometry of
Markov Diffusion
Operators
Dominique Bakry Michel Ledoux
Institut de Mathématiques de Toulouse Institut de Mathématiques de Toulouse
Université de Toulouse Université de Toulouse
and Institut Universitaire de France and Institut Universitaire de France
Toulouse, France Toulouse, France
Ivan Gentil
Institut Camille Jordan
Université Claude Bernard Lyon 1
Villeurbanne, France
ISSN 0072-7830 ISSN 2196-9701 (electronic)

Grundlehren der mathematischen Wissenschaften
ISBN 978-3-319-00226-2 ISBN 978-3-319-00227-9 (eBook)
DOI 10.1007/978-3-319-00227-9
Springer Cham Heidelberg New York Dordrecht London
Library of Congress Control Number: 2013952461
Mathematics Subject Classification: 39B62, 39B72, 47D07, 53C21
© Springer International Publishing Switzerland 2014

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed. Exempted from this legal reservation are brief excerpts in connection
with reviews or scholarly analysis or material supplied specifically for the purpose of being entered
and executed on a computer system, for exclusive use by the purchaser of the work. Duplication of
this publication or parts thereof is permitted only under the provisions of the Copyright Law of the
Publisher’s location, in its current version, and permission for use must always be obtained from Springer.
Permissions for use may be obtained through RightsLink at the Copyright Clearance Center. Violations
are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of pub-
lication, neither the authors nor the editors nor the publisher can accept any legal responsibility for any
errors or omissions that may be made. The publisher makes no warranty, express or implied, with respect
to the material contained herein.
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)

How far can you go with the Cauchy-Schwarz
inequality and integration by parts?
To Leonard Gross
Preface
Semigroups of operators on a Banach space provide very general models and tools
in the analysis of time evolution phenomena and dynamical systems. They have a
long history in mathematics and have been studied in a number of settings, from
functional analysis and mathematical physics to probability theory, Riemannian ge-
ometry, Lie groups, analysis of algorithms, and elsewhere.
The part of semigroup theory investigated in this book deals with Markov dif-
fusion semigroups and their infinitesimal generators, which naturally arise as solu-
tions of stochastic differential equations and partial differential equations. As such,
the topic covers a large body of mathematics ranging from probability theory and
partial differential equations to functional analysis and differential geometry for op-
erators or processes on manifolds. Within these frameworks, research interests have
grown over the years, now encompassing a wide variety of questions such as reg-
ularity and smoothing properties of differential operators, Sobolev-type estimates,
heat kernel bounds, non-explosion properties, convergence to equilibrium, existence
and regularity of solutions of stochastic differential equations, martingale problems,
stochastic calculus of variations and so on.
This book is more precisely focused on the concrete interplay between the ana-
lytic, probabilistic and geometric aspects of Markov diffusion semigroups and gen-
erators involved in convergence to equilibrium, spectral bounds, functional inequal-
ities and various bounds on solutions of evolution equations linked to geometric
properties of the underlying structure.
One prototypical example at this interface is simply the standard heat semigroup
(Pt )t≥0 on the Euclidean space Rn whose Gaussian kernel
1
e−|x| /4t ,
2
u = u(t, x) = pt (x) = t > 0, x ∈ Rn ,
(4πt)n/2
is a fundamental solution of the heat equation
∂t u = u, u(0, x) = δ0 ,
for the standard Laplace operator , thus characterized as the infinitesimal generator
of the semigroup (Pt )t≥0 .
vii
viii Preface
From the probabilistic viewpoint, the family of kernels pt (x), t > 0, x ∈ Rn ,

represents the transition probabilities of a standard Brownian motion (Bt )t≥0 as

E f (x + B2t ) = f (y) pt (x − y)dy = Pt f (x), t > 0, x ∈ Rn ,
Rn
for all bounded measurable functions f : Rn → R.

The third aspect investigated in this work is geometric, and perhaps less imme-
diately apparent than the analytic and probabilistic aspects. It aims to interpret, in
some sense, the commutation of derivation and action of the semigroup as a curva-
ture condition. For the standard Euclidean semigroup example above, the commu-
tation ∇Pt f = Pt (∇f ) will express a zero curvature, although this corresponds not
only to the curvature of Euclidean space as a Riemannian manifold but rather to the
curvature of Euclidean space equipped with the Lebesgue measure, invariant under
the heat flow (Pt )t≥0 , and the bilinear operator (f, g) = ∇f · ∇g.
In order to carry out the investigation along these lines, the exposition emphasizes
the basic structure of a Markov Triple1 (E, μ, ) consisting of a (measurable) state
space E, a carré du champ operator and a measure μ invariant under the dynamics
induced by . The notion of a carré du champ operator associated with a Markov
semigroup (Pt )t≥0 with infinitesimal generator L given (on a suitable algebra A of
functions on E) by
1
(f, g) = L(f g) − f Lg − gLf ,
2
will be a central tool of investigation, the associated -calculus providing, at least
at a formal level, a kind of algebraic framework encircling the relevant properties
and results.
These analytic, stochastic and geometric features form the basis of the investiga-
tion undertaken in this book, describing Markov semigroups via their infinitesimal
generators as solutions of second order differential operators and their probabilistic
representations as Markov processes, and analyzing them with respect to curvature
properties. The investigation is limited to symmetric (reversible in the Markovian
terminology) semigroups, although various ideas and techniques go beyond this
framework. We also restrict our attention to the diffusion setting, that is when the
carré du champ operator is a derivation operator in its two arguments, even in those
cases where the result could be extended to a more general setting. These restrictions
rule out many interesting fields of applications (discrete Markov chains, models of
statistical mechanics, most of the analysis of algorithms of interest in optimization
theory or approximations of partial differential equations, for example), but allow
us to concentrate on central features in the analysis of semigroups, in the same way
that ordinary differential equations are in general easier to handle than discrete se-
quences. Even within the field of symmetric diffusion semigroups, we have not tried
1 The terminology “Markov triple” should not, of course, be confused with solutions of the Markov
Diophantine equation x 2 + y 2 + z2 = 3xyz!

Preface ix
to cover all the possible interesting cases. In order to keep the monograph within a
reasonable size, we have had to omit, among other things, the specific analysis re-
lated to hypoelliptic diffusions, special features of diffusions on Lie groups, and
many interesting developments arising from infinite interacting particle systems.
In addition, although we have largely been motivated by the analysis of the be-
havior of diffusion processes (that is, solutions of time homogeneous stochastic dif-
ferential equations), rather than concentrating on the probabilistic aspects of the
subject, such as almost sure convergence of functionals of the trajectories of the un-
derlying Markov processes, recurrence or transience, we instead chose to translate
most of the features of interest into functional analytic properties of the Markov
structure (E, μ, ) under investigation.
Heat kernel bounds, functional inequalities and their applications to convergence
to equilibrium and geometric features of Markov operators are among the main
topics of interest developed in this monograph. A particular emphasis is placed on
families of inequalities relating, on a Markov Triple (E, μ, ), functionals of func-
tions f : E → R to the energy induced by the invariant measure μ and the carré du
champ operator ,

E(f, f ) = (f, f )dμ.
E
Typical functionals are the variance, entropy or Lp -norms leading to the main func-
tional inequalities of interest, the Poincaré or spectral gap inequality, the logarithmic
Sobolev inequality and the Sobolev inequality. A particular goal is to establish such
families of inequalities under suitable curvature conditions which may be described
by the carré du champ operator and its iterated 2 operator.
Similar inequalities are investigated at the level of the underlying semigroup
(Pt )t≥0 for the heat kernel measures, comparing Pt (ϕ(f )) (for some ϕ : R → R)
to Pt ((f, f )) or (Pt f, Pt f ), which give rise to heat kernel bounds. With this
task in mind, we will develop the main powerful tool of heat flow monotonicity, or
semigroup interpolation, with numerous illustrative applications and strong intuitive
content. To illustrate the principle, as a wink towards what is to come, let us briefly
present here a heat flow proof of the classical Hölder inequality which is very much
in the spirit of this book. In particular, the reduction to a quadratic bound is typi-
cal of the arguments developed in this work. Let f, g be suitable (strictly) positive
functions on Rn and θ ∈ (0, 1). For fixed t > 0, consider, at any point (omitted), the
interpolation

(s) = Ps eθ log Pt−s f +(1−θ) log Pt−s g , s ∈ [0, t],
where (Pt )t≥0 is the standard heat semigroup on Rn as recalled above. Together
with the heat equation ∂s Ps = Ps = Ps , the derivative in s of is given by

(s) = Ps eH − eH θ e−F eF + (1 − θ )e−G eG
x Preface
where F = log Pt−s f , G = log Pt−s g and H = θ F + (1 − θ )G. Now by standard

calculus,

e−H eH − θ e−F eF + (1 − θ )e−G eG
= |∇H |2 − θ |∇F |2 − (1 − θ )|∇G|2
which is negative by convexity of the square function. Hence (s), s ∈ [0, t], is
decreasing, and thus

(t) = Pt f θ g 1−θ ≤ (Pt f )θ (Pt g)1−θ = (0).
Normalizing by t n/2 and letting t tend to infinity yields Hölder’s inequality for the
Lebesgue measure. Actually, the same argument may be performed at the level of
a Markov semigroup with invariant finite discrete measure, thus yielding Hölder’s
inequality for arbitrary measures.
While functional inequalities and their related applications are an important focal
point, they also give us the opportunity to discuss a number of issues related to
examples and properties of Markov semigroups and operators. One objective of
this work is thus also to present the basic tools and ideas revolving around Markov
semigroups and to illustrate their usefulness in different contexts.
The monograph comprises three main parts.
The first part, covering Chaps. 1 to 3, presents some of the main features,
properties and examples of Markov diffusion semigroups and operators as con-
sidered in this work. In a somewhat informal but intuitive way, Chap. 1 intro-
duces Markov semigroups, their infinitesimal generators and associated Markov
processes, stochastic differential equations and diffusion semigroups. It also de-
scribes a few of the standard operations and techniques while working with semi-
groups. Chapter 2 develops in detail a number of central geometric models which
will serve as references for later developments, namely the heat semigroups and
Laplacians on the flat Euclidean space, the sphere and the hyperbolic space. Sturm-
Liouville operators on the line, and some of the most relevant examples (Ornstein-
Uhlenbeck, Laguerre and Jacobi), are also presented therein. On the basis of these
preliminary observations and examples, Chap. 3 then tries to describe a general
framework of investigation. While it would not be appropriate to try to cover in a
unique formal mould all the cases of interest, it is nevertheless useful to emphasize
the basic properties and tools in order to easily and suitably develop the -calculus.
In particular, it is necessary to describe with some care the various classes and al-
gebras of functions that we shall be dealing with and to show their relevance in
the classical smooth settings. Note that while infinite-dimensional models would re-
quire further care in this abstract formalism, the methods and principles emphasized
throughout this work are similarly relevant for them. Taking the more classical pic-
ture as granted, Chap. 3 may be skipped at first reading (or limited to the summary
Sect. 3.4).
Part II, forming the core of the text, includes Chaps. 4 to 6 and covers the three
main functional inequalities of interest, Poincaré or spectral gap inequalities, loga-
rithmic Sobolev inequalities and Sobolev inequalities. For each family, some basic
Preface xi
properties and tools are detailed, in tight connection with the reference examples of
Chap. 2 and their geometric properties. Stability, perturbation and comparison prop-
erties, characterization in dimension one, concentration bounds and convergence to
equilibrium are thus addressed for each family. The discussion then distinguishes
between inequalities for the heat kernel measures (local) and for the invariant mea-
sure (global) which are analyzed and established under curvature hypotheses. Chap-
ter 4 is thus devoted to Poincaré or spectral gap inequalities, closely related to
spectral decompositions. Chapter 5 deals with logarithmic Sobolev inequalities, em-
phasized as the natural substitute for classical Sobolev-type inequalities in infinite
dimension, and their equivalent hypercontractive smoothing properties. Sobolev in-
equalities form a main family of interest for which Chap. 6 provides a number of
equivalent descriptions (entropy-energy, Nash or Gagliardo-Nirenberg inequalities)
and associated heat kernel bounds. A significant proportion of this chapter is devoted
to the rich geometric content of Sobolev inequalities, their conformal invariance, and
the curvature-dimension conditions.
On the basis of the main functional inequalities of Part II, Part III, consisting of
Chaps. 7 to 9, addresses several variations, extensions and related topics of interest.
Chapter 7 deals with general families of functional inequalities, each of them hav-
ing their own interest and usefulness. The exposition mainly emphasizes entropy-
energy (on the model of logarithmic Sobolev inequalities) and Nash-type inequal-
ities. In addition, the tightness of functional inequalities is studied by employing
the tool of weak Poincaré inequalities. Chapter 8 is an equivalent description of the
various families of inequalities for functions presented so far in terms of sets and
capacities for which co-area formulas provide the suitable link. The second part of
this chapter is concerned with isoperimetric-type inequalities for which semigroup
tools again prove most useful. Chapter 9 briefly presents some of the recent impor-
tant developments in optimal transportation in connection with the semigroup and
-calculus, including in particular a discussion of the relationships between func-
tional and transportation cost inequalities (in a smooth Riemannian setting).
The last part of the monograph consists of three appendices, on semigroups of
operators on a Banach space, elements of stochastic calculus and the basics of dif-
ferential and Riemannian geometry. At the interface between analysis, probability
and geometry, these appendices aim to possibly supplement the reader’s knowledge
depending on his own background. They are not strictly necessary for the compre-
hension of the core of the text, but may serve as a support for the more special-
ized parts. It should be mentioned, however, that the last two sections of the third
appendix on the basics of Riemannian geometry actually contain material on the
-calculus (in a Riemannian context) which will be used in a critical way in some
parts of the book.
This book has been designed to be both an introduction to the subject, intended to
be accessible to non-specialists, and an exposition of both basic and more advanced
results of the theory of Markov diffusion semigroups and operators. Indeed we chose
to concentrate on those points where we felt that the techniques and ideas are central
and may be used in a wider context, even though we have not attempted to reach
the widest generality. Every chapter starts at a level which is elementary for the
xii Preface
notions developed in it, but may evolve to more specialized topics which in general
may be skipped at first reading. It should be stressed that the level of exposition
throughout the book fairly non-uniform, sometimes putting emphasis on facts or
results which may appear as obvious or classical for some readers while developing
at the same time more sophisticated issues. This choice is motivated by our desire
to make the text accessible to readers with different backgrounds, and also by our
aim to provide tools and methods to access more difficult parts of the theory or to
be applied in different contexts. This delicate balance is not always reached but we
nevertheless hope that the chosen style of exposition is helpful.
The monograph is intended for students and researchers interested in the mod-
ern aspects of Markov diffusion semigroups and operators and their connections
with analytic functional inequalities, probabilistic convergence to equilibrium and
geometric curvature. Selected chapters may be used for advanced courses on the
topic. Readers who wish to get a flavor of Markov semigroups and their applica-
tions should concentrate on Part I (with the exception of Chap. 3) and Part II. Via
an appropriate selection of topics, Part III tries to synthesise the developments of
the last decade. The book demands from the reader only a reasonable knowledge
of basic functional analysis, measure theory and probability theory. It is also ex-
pected that it may be read in a non-linear way, although the various chapters are not
completely independent. The reader not familiar with the main themes (analysis,
probability and geometry) will find some of the basic material collected together in
the appendices.
Each Chapter is divided into Sections, often themselves divided in Sub-Sections.
Section 1.8 is the eighth section in Chap. 1. Theorem 4.6.2 indicates a theorem in
Chap. 4, Sect. 4.6, and (3.2.2) is a formula in Sect. 3.2. An item of a given chapter
is also referred to in other chapters by the page on which it appears. There are no
references to articles or books within the exposition of a given chapter. The Sections
“Notes and References” at the end of each chapter briefly describe some historical
developments with pointers to the literature. The references are far from exhaustive
and in fact are rather limited. There is no claim for completeness and we apologize
for omissions and errors. For books and monographs, we have tried to present the
references in historical order with respect to original editions (although the links
point toward the latest editions).
This book began its life in the form of lectures presented by the first author
at Saint-Louis du Sénégal in April 2009. He thanks the organizers of this school
for the opportunity to give this course and the participants for their interest. This
work presents results and developments which have emerged during the last three
decades. Over the years, we have benefited from the vision, expertise and help of a
number of friends and colleagues, among them M. Arnaudon, F. Barthe, W. Beck-
ner, S. Bobkov, F. Bolley, C. Borell, E. Carlen, G. Carron, P. Cattiaux, D. Chafaï,
D. Cordero-Erausquin, T. Coulhon, J. Demange, J. Dolbeault, K. D. Elworthy,
M. Émery, A. Farina, P. Fougères, N. Gozlan, L. Gross, A. Guillin, E. Hebey,
B. Helffer, A. Joulin, C. Léonard, X. D. Li, P. Maheux, F. Malrieu, L. Miclo, E. Mil-
man, B. Nazaret, V. H. Nguyen, Z.-M. Qian, M.-K. von Renesse, C. Roberto, M. de
la Salle, L. Saloff-Coste, K.-T. Sturm, C. Villani, F.-Y. Wang, L. Wu and B. Ze-
garliński. We wish to thank them for their helpful remarks and constant support.
Preface xiii
F. Bolley, S. Campese and C. Léonard went through parts of the manuscript at sev-
eral stages of the preparation, and we warmly thank them for all their corrections
and comments that helped to improve the exposition.
We sincerely thank the Springer Editors C. Byrne and M. Reizakis and the pro-
duction staff for a great editing process.
We apologize for all the errors, and invite the readers to report any remarks,
mistakes and misprints. A list of errata and comments will be maintained online.
Lyon, Toulouse Dominique Bakry
June 2013 Ivan Gentil
Michel Ledoux
Basic Conventions
Here are some classical and basic conventions used throughout the book.
N is the set of integers {0, 1, 2, . . .}. The set of real numbers is denoted by R.
Functions (on some state space E) are always real-valued. Points in R are usually
denoted by x (if R is the underlying state space) or by r.
An element r ∈ R is positive if r ≥ 0, strictly positive if r > 0, negative if r ≤ 0
and strictly negative if r < 0. Moreover, R+ = [0, ∞) is the set of positive real
numbers while (0, ∞) denotes the set of strictly positive numbers. For r, s ∈ R,
r ∧ s = min(r, s) and r ∨ s = max(r, s). We agree that 0 log 0 = 0.
In the same way (and somewhat against the current), a positive (respectively
negative) function f (on E) is such that f (x) ≥ 0 (respectively f (x) ≤ 0) for every
x ∈ E. The function is strictly positive or strictly negative whenever the inequal-
ities are strict. Similarly, an increasing (respectively decreasing) function f on R
or some interval of R satisfies f (x) ≤ f (y) (respectively f (x) ≥ f (y)) for every
x ≤ y. The function f is said to be strictly increasing or strictly decreasing when-
ever the preceding inequalities are strict. A function is monotone if it is increasing
or decreasing.
Points in Rn are denoted by x = (x1 , . . . , xn ) = (xi )1≤i≤n (or sometimes
x = (x 1 , . . . , x n ) = (x i )1≤i≤n depending on the geometric context). The scalar prod-
uct and Euclidean norm in Rn are given by

n
n 1/2
x·y = xi yi , |x| = (x · x)
1/2
= xi2 .
i=1 i=1
The notation | · | is used throughout to denote the Euclidean norm of vectors and of
tensors.
The constant function equal to 1 on a state space E is denoted by 1. If A ⊂ E,
1A is the characteristic or indicator function of A.
All measures on a measurable space (E, F) considered here are positive mea-
sures. Positive (measurable) functions on (E, F) may take the value +∞. If μ is a
(positive) measure on (E, F), and if f is a function on E
which is
integrable with
respect to μ, its integral with respect to μ is denoted by E f dμ or E f (x)dμ(x),
xv
xvi Basic Conventions

or sometimes as E f (x)μ(dx). The Lebesgue measure on the Borel sets of Rn is

denoted by dx. If B is a Borel set in Rn , its Lebesgue measure is sometimes denoted
by voln (B).
The terminology “change of variables” is used in the broad sense of changing a
variable x into h(x) and a function f into ψ(f ). “Chain rule” is understood more
in an algebraic sense when h and ψ are polynomials.
The notations are supposed to be reasonably stable throughout the monograph.
Further definitions and conventions will be given in the text when they are needed.
A list of symbols and notations with the corresponding reference pages is given on
p. 523.
Contents
Part I Markov Semigroups, Basics and Examples

1 Markov Semigroups . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1 Markov Processes and Associated Semigroups . . . . . . . . . . 7
1.2 Markov Semigroups, Invariant Measures and Kernels . . . . . . 9
1.3 Chapman-Kolmogorov Equations . . . . . . . . . . . . . . . . . 16
1.4 Infinitesimal Generators and Carré du Champ Operators . . . . . 18
1.5 Fokker-Planck Equations . . . . . . . . . . . . . . . . . . . . . 23
1.6 Symmetric Markov Semigroups . . . . . . . . . . . . . . . . . . 24
1.7 Dirichlet Forms and Spectral Decompositions . . . . . . . . . . 29
1.8 Ergodicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
1.9 Markov Chains . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
1.10 Stochastic Differential Equations and Diffusion Processes . . . . 38
1.11 Diffusion Semigroups and Operators . . . . . . . . . . . . . . . 42
1.12 Ellipticity and Hypo-ellipticity . . . . . . . . . . . . . . . . . . 49
1.13 Domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
1.14 Summary of Hypotheses (Markov Semigroup) . . . . . . . . . . 53
1.15 Working with Markov Semigroups . . . . . . . . . . . . . . . . 56
1.16 Curvature-Dimension Condition . . . . . . . . . . . . . . . . . . 70
1.17 Notes and References . . . . . . . . . . . . . . . . . . . . . . . 74
2 Model Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
2.1 Euclidean Heat Semigroup . . . . . . . . . . . . . . . . . . . . 78
2.2 Spherical Heat Semigroup . . . . . . . . . . . . . . . . . . . . . 81
2.3 Hyperbolic Heat Semigroup . . . . . . . . . . . . . . . . . . . . 88
2.4 The Heat Semigroup on a Half-Line and the Bessel Semigroup . 92
2.5 The Heat Semigroup on the Circle and on a Bounded Interval . . 96
2.6 Sturm-Liouville Semigroups on an Interval . . . . . . . . . . . . 97
2.7 Diffusion Semigroups Associated with Orthogonal Polynomials . 102
xvii
xviii Contents
3 Symmetric Markov Diffusion Operators . . . . . . . . . . . . . . . 119

3.1 Markov Triples . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
3.2 Second Order Differential Operators on a Manifold . . . . . . . 137
3.3 Heart of Darkness . . . . . . . . . . . . . . . . . . . . . . . . . 151
3.4 Summary of Hypotheses (Markov Triple) . . . . . . . . . . . . . 168
Part II Three Model Functional Inequalities
4 Poincaré Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . 177
4.1 The Example of the Ornstein-Uhlenbeck Semigroup . . . . . . . 178
4.2 Poincaré Inequalities . . . . . . . . . . . . . . . . . . . . . . . . 181
4.3 Tensorization of Poincaré Inequalities . . . . . . . . . . . . . . . 185
4.4 The Example of the Exponential Measure, and Exponential
Integrability . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
4.5 Poincaré Inequalities on the Real Line . . . . . . . . . . . . . . 193
4.6 The Lyapunov Function Method . . . . . . . . . . . . . . . . . . 201
4.7 Local Poincaré Inequalities . . . . . . . . . . . . . . . . . . . . 206
4.8 Poincaré Inequalities Under a Curvature-Dimension Condition . 211
4.9 Brascamp-Lieb Inequalities . . . . . . . . . . . . . . . . . . . . 215
4.10 Further Spectral Inequalities . . . . . . . . . . . . . . . . . . . . 220
5 Logarithmic Sobolev Inequalities . . . . . . . . . . . . . . . . . . . 235
5.1 Logarithmic Sobolev Inequalities . . . . . . . . . . . . . . . . . 236
5.2 Entropy Decay and Hypercontractivity . . . . . . . . . . . . . . 243
5.3 Integrability of Eigenvectors . . . . . . . . . . . . . . . . . . . . 250
5.4 Logarithmic Sobolev Inequalities and Exponential Integrability . 252
5.5 Local Logarithmic Sobolev Inequalities . . . . . . . . . . . . . . 257
5.6 Infinite-Dimensional Harnack Inequalities . . . . . . . . . . . . 265
5.7 Logarithmic Sobolev Inequalities Under a Curvature-Dimension
Condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
6 Sobolev Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . 277
6.1 Sobolev Inequalities on the Model Spaces . . . . . . . . . . . . 278
6.2 Sobolev and Related Inequalities . . . . . . . . . . . . . . . . . 279
6.3 Ultracontractivity and Heat Kernel Bounds . . . . . . . . . . . . 286
6.4 Ultracontractivity and Compact Embeddings . . . . . . . . . . . 290
6.5 Tensorization of Sobolev Inequalities . . . . . . . . . . . . . . . 291
6.6 Sobolev Inequalities and Lipschitz Functions . . . . . . . . . . . 293
6.7 Local Sobolev Inequalities . . . . . . . . . . . . . . . . . . . . . 296
6.8 Sobolev Inequalities Under a Curvature-Dimension Condition . . 305
6.9 Conformal Invariance of Sobolev Inequalities . . . . . . . . . . . 313
6.10 Gagliardo-Nirenberg Inequalities . . . . . . . . . . . . . . . . . 323
6.11 Fast Diffusion Equations and Sobolev Inequalities . . . . . . . . 330
Contents xix
Part III Related Functional, Isoperimetric and Transportation

Inequalities
7 Generalized Functional Inequalities . . . . . . . . . . . . . . . . . 347
7.1 Inequalities Between Entropy and Energy . . . . . . . . . . . . . 348
7.2 Off-diagonal Heat Kernel Bounds . . . . . . . . . . . . . . . . . 355
7.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362
7.4 Beyond Nash Inequalities . . . . . . . . . . . . . . . . . . . . . 364
7.5 Weak Poincaré Inequalities . . . . . . . . . . . . . . . . . . . . 373
7.6 Further Families of Functional Inequalities . . . . . . . . . . . . 382
7.7 Summary for the Model Example μα . . . . . . . . . . . . . . . 386
8 Capacity and Isoperimetric-Type Inequalities . . . . . . . . . . . . 391
8.1 Capacity Inequalities and Co-area Formulas . . . . . . . . . . . 392
8.2 Capacity and Sobolev Inequalities . . . . . . . . . . . . . . . . . 396
8.3 Capacity and Poincaré and Logarithmic Sobolev Inequalities . . 399
8.4 Capacity and Further Functional Inequalities . . . . . . . . . . . 403
8.5 Gaussian Isoperimetric-Type Inequalities Under a Curvature
Condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411
8.6 Harnack Inequalities Revisited . . . . . . . . . . . . . . . . . . 421
8.7 From Concentration to Isoperimetry . . . . . . . . . . . . . . . . 425
9 Optimal Transportation and Functional Inequalities . . . . . . . . 433
9.1 Optimal Transportation . . . . . . . . . . . . . . . . . . . . . . 434
9.2 Transportation Cost Inequalities . . . . . . . . . . . . . . . . . . 438
9.3 Transportation Proofs of Functional Inequalities . . . . . . . . . 442
9.4 Hamilton-Jacobi Equations . . . . . . . . . . . . . . . . . . . . 451
9.5 Hypercontractivity of Solutions of Hamilton-Jacobi Equations . . 454
9.6 Transportation Cost and Logarithmic Sobolev Inequalities . . . . 458
9.7 Heat Flow Contraction in Wasserstein Space . . . . . . . . . . . 462
9.8 Curvature of Metric Measure Spaces . . . . . . . . . . . . . . . 464
Appendices
Appendix A Semigroups of Bounded Operators on a Banach Space . . 473
A.1 The Hille-Yosida Theory . . . . . . . . . . . . . . . . . . . . . . 473
A.2 Symmetric Operators . . . . . . . . . . . . . . . . . . . . . . . 475
A.3 Friedrichs Extension of Positive Operators . . . . . . . . . . . . 477
A.4 Spectral Decompositions . . . . . . . . . . . . . . . . . . . . . 478
A.5 Essentially Self-adjoint Operators . . . . . . . . . . . . . . . . . 481
A.6 Compact and Hilbert-Schmidt Operators . . . . . . . . . . . . . 483
A.7 Notes and References . . . . . . . . . . . . . . . . . . . . . . . 485
xx Contents
Appendix B Elements of Stochastic Calculus . . . . . . . . . . . . . . 487

B.1 Brownian Motion and Stochastic Integrals . . . . . . . . . . . . 487
B.2 The Itô Formula . . . . . . . . . . . . . . . . . . . . . . . . . . 491
B.3 Stochastic Differential Equations . . . . . . . . . . . . . . . . . 493
B.4 Diffusion Processes . . . . . . . . . . . . . . . . . . . . . . . . 495
B.5 Notes and References . . . . . . . . . . . . . . . . . . . . . . . 498
Appendix C Basic Notions in Differential and Riemannian Geometry 499
C.1 Differentiable Manifolds . . . . . . . . . . . . . . . . . . . . . . 500
C.2 Some Elementary Euclidean Geometry . . . . . . . . . . . . . . 502
C.3 Basic Notions in Riemannian Geometry . . . . . . . . . . . . . . 504
C.4 Riemannian Distance . . . . . . . . . . . . . . . . . . . . . . . 509
C.5 The Riemannian and 2 Operators . . . . . . . . . . . . . . . 511
C.6 Curvature-Dimension Conditions . . . . . . . . . . . . . . . . . 513
C.7 Notes and References . . . . . . . . . . . . . . . . . . . . . . . 518
Afterword . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 521
Chicken “Gaston Gérard” . . . . . . . . . . . . . . . . . . . . . . . . 521
Notation and List of Symbols . . . . . . . . . . . . . . . . . . . . . . . . 523
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 527
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 547
Part I
Markov Semigroups, Basics and Examples
Chapter 1
Markov Semigroups
This chapter is introductory and descriptive. It aims to introduce, somewhat infor-

mally, some of the basic ideas and concepts in the investigation of Markov semi-
groups, operators and processes, at the interface between analysis, partial differen-
tial equations, probability theory and geometry, loosely jumping from one area to
another. Readers only familiar with one, or two, or even none of these fields, will
find the necessary background in the various appendices.
The chapter is not intended to be read linearly, but should be used as a vade
mecum of ideas and tools on the topic of Markov semigroups. Each section in-
troduces and sheds light on one or several aspects of the investigation of Markov
semigroups, aiming at this point to develop intuition rather than giving precise defi-
nitions and hypotheses. Pointers to more precise descriptions and examples appear-
ing in subsequent chapters of the book are provided. In particular, some of the main
features of the analysis will be illustrated by model examples in Chap. 2, while
Chap. 3 will develop the complete formalism describing the suitable environment
in which the full theory may easily be developed. For the reader’s convenience, a
first set of precise assumptions concerning state spaces, measures, semigroups etc.,
is nevertheless put forward here in Sect. 1.14.
There are many ways to address Markov semigroups and generators. In this chap-
ter, we investigate their elementary properties starting with the assumption that we
already know the semigroup itself which then leads to the definition of the so-called
infinitesimal generator of the semigroup, which in return entirely describes the latter.
In Chap. 3, conversely, we begin with the generator, or rather its carré du champ op-
erator, given on some suitable class of functions, and use it to determine properties
of the semigroup. This setting then describes the formalism in which the monograph
will evolve.
Generally speaking, a semigroup P = (Pt )t≥0 is a family of operators acting on
some suitable function space with the semigroup property Pt ◦ Ps = Pt+s , t, s ≥ 0,
P0 = Id. Such families naturally arise in numerous settings, and describe evolu-
tion equations which may be investigated from various viewpoints. In particular,
these semigroups appear in the probabilistic context describing the family of laws
of Markov processes (Xt )t≥0 living on a measurable space E, the fundamental rela-
D. Bakry et al., Analysis and Geometry of Markov Diffusion Operators, 3

Grundlehren der mathematischen Wissenschaften 348,
DOI 10.1007/978-3-319-00227-9_1,
4 1 Markov Semigroups
tion being given by the conditional expectation representation

Pt f (x) = E f (Xt ) | X0 = x
for t ≥ 0, x ∈ E and f : E → R a suitable measurable function.

In order to illustrate, and introduce the reader to, these investigations, let us de-
scribe the simplest example, namely the heat or Brownian semigroup on the Eu-
clidean space Rn . This example may be introduced in many ways, depending on
the culture and purpose. According to the three main aspects of this monograph, we
briefly describe it from the analytic, probabilistic and geometric viewpoints.
First we start with the family of Gaussian kernels
1
e−|x| /4t ,
2
pt (x) = t > 0, x ∈ Rn .
(4πt)n/2
Formally, we may already agree that p0 (x) may be considered as the Dirac mass
at 0. The normalization is chosen so that pt (x) are
probability densities with respect
to the Lebesgue measure (denoted dx), that is Rn pt (x)dx = 1. It is a classical
result, and easy to see, that pt solves the (parabolic) heat equation
∂t pt = pt
where is the standard Laplacian on Rn .

From these kernels, one may define operators Pt , t ≥ 0, sending suitable func-
tions f : Rn → R to new functions Pt (f ) = Pt f defined by

Pt f (x) = f (y) pt (x, y)dy, t > 0, x ∈ Rn ,
Rn
where pt (x, y) = pt (x − y), (x, y) ∈ Rn × Rn . This definition applies for example

to any function f which is bounded and measurable, but the condition that f is
bounded may obviously be relaxed provided Pt f still makes sense. P0 is the identity
operator, P0 f = f . By standard convolution, these operators satisfy the semigroup
property Pt ◦ Ps = Pt+s , t, s ≥ 0, and thus form a so-called semigroup (moreover it
is Markov, or mass preserving, since Pt (1) = 1). On smooth functions f : Rn → R,
∂t Pt f = Pt f = Pt (f ).
In other words, u(x, t) = Pt f (x), t ≥ 0, x ∈ Rn , solves the partial differential (heat)

equation ∂t u = u with initial condition u(x, 0) = f (x). This initial condition
has however to be interpreted carefully. It is certainly true that, for any x ∈ Rn ,
limt→0 u(x, t) = f (x) provided f is continuous and bounded, but has to be under-
stood in a weaker sense when f is only measurable.
Accordingly, the Laplacian is called the infinitesimal generator of the (heat)
semigroup (Pt )t≥0 . This semigroup is one simple prototypical example which antic-
ipates several developments in this monograph. The heat equation ∂t u = u is the
basic parabolic partial differential equation linking the first order time derivative to
1 Markov Semigroups 5
second order space derivatives, highlighting the central importance of second order
differential operators of interest in this monograph.
On the probabilistic side, the densities pt describe the transition probabilities of
a random process (Bt )t≥0 in Rn called Brownian motion. More precisely, (Bt )t≥0 is
a (continuous) random process with values in Rn such that B0 = 0 (almost surely)
and the increments
Bt1 , Bt2 − Bt1 , . . . , Btk − Btk−1 , 0 ≤ t1 < · · · < t k ,
are independent centered Gaussian vectors in Rn with respective covariance matri-

ces t1 Id, (t2 − t1 ) Id, . . . , (tk − tk−1 ) Id where Id is the identity matrix. Note that
this n-dimensional Brownian motion may be written as Bt = (Bt1 , . . . , Btn ), t ≥ 0,
where (Bti )t≥0 , i = 1, . . . , n, are independent one-dimensional Brownian motions.
From this point of view, there is not much more to n-dimensional Brownian mo-
tion than in the one-dimensional case although dimension could (and will) play a
fundamental role in many functional inequalities developed later.
By construction, the distribution of Bt − Bs , t > s ≥ 0, is given by

y
E f (Bt − Bs ) = f √ pt−s (y)dy
Rn 2
(for f bounded and measurable, for example). In terms of the semigroup (Pt )t≥0 ,
the preceding may also be written as
√
E f (x + B2t ) = E f x + 2 Bt = Pt f (x), t ≥ 0, x ∈ Rn .
The process (x + Bt )t≥0 is (by definition) a Brownian motion starting from x ∈ Rn .

In this interpretation, the semigroup property is a consequence of the fact that
Bt+s (t, s ≥ 0) may be written as the sum of two independent Gaussian vectors
(Bt+s − Bs ) + Bs with respective covariance matrices t Id and s Id. This observa-
tion also illustrates the Markov property of the process (Bt )t≥0 , which states that
starting from Bs , the process (Bt+s )t≥0 is again a Brownian motion.
√
The somewhat annoying 2 factor in the preceding formulas justifies the proba-
bilistic convention of working with 12 instead of as the infinitesimal generator.
However we avoid the factor 12 throughout and mainly work with the generator .
Working then with stochastic calculus involving Brownian motion, we use from
time to time the process (B2t )t≥0 .
The geometric content of the heat or Brownian semigroup model is less immedi-
ately perceptible since it is developed on the flat Euclidean space Rn . However, anal-
ogous heat equations and Brownian processes may be considered on different state
spaces, such as the sphere or the hyperbolic space. The geometry of the underlying
state space, or rather of the Laplace operator acting on it, is then closely connected
with various properties of the corresponding Markov semigroups and processes. In
particular, curvature properties play an important role in this analysis, as we shall
see. On the flat space Rn , the (null) curvature is reflected on the heat semigroup
(Pt )t≥0 by the commutation property ∇(Pt f ) = Pt (∇f ) between the action of the
gradient operator ∇ and the semigroup Pt (on smooth functions f : Rn → R). By

Jensen’s inequality, this commutation produces the gradient bound

∇(Pt f ) ≤ Pt |∇f | .
In fact, as will be developed extensively, this commutation inequality corresponds

to a functional curvature property of the Laplace operator . The analysis of similar
relationships or gradient bounds in more general settings under certain curvature
conditions will reveal important information about the semigroups, and in turn will
shed light on one of our main objects of study, the functional inequalities.
The fundamental example of the heat or Brownian semigroup in Euclidean space
may therefore be analyzed in different ways, via partial differential equations, prob-
ability theory or geometry. These three modes of analysis will be central to our in-
vestigation of Markov semigroups and operators. The elementary properties of the
heat semigroup just presented may be generalized and deepened in many different
directions. As announced in the introduction, our investigation shall be restricted to
the family of diffusion (symmetric) Markov semigroups for which the generator (the
Laplace operator in the example of the heat semigroup) is a second order differential
operator satisfying the change of variables formula.
In this chapter we shall make some fairly informal remarks on various aspects of
the analysis of Markov semigroups and operators, using the Euclidean heat semi-
group as a focal example, in anticipation of future developments. We also introduce
some basic concepts, some of them being further developed in the appendices. Read-
ers are encouraged to complement this overview with relevant references from the
literature, a selection of which can be found in the Notes and References Sect. 1.17
at the end of the chapter.
Let us outline the contents of this chapter. The first sections describe some
of the basic elements in the general description of Markov semigroups, operators
and processes, dealing with state space, the Markov property, invariant measures,
infinitesimal generators and carré du champ operators. The associated Chapman-
Kolmogorov and Fokker-Plank equations are discussed next, connecting with prob-
abilistic Markov processes on one side and partial differential equations on the other.
Sections 1.6 and 1.7 are concerned with symmetric semigroups, Dirichlet forms and
spectral decompositions, while Sect. 1.8 introduces ergodicity. Markov chains on
finite or countable state spaces are then briefly presented as a simple illustration of
the previous concepts (although the study of Markov semigroups and operators will
be restricted in this book to the so-called diffusion semigroups, thus excluding such
discrete models). Probabilistic intuition is further developed on the basis of stochas-
tic differential equations and diffusion processes in Sect. 1.10 which illustrates the
main diffusion hypothesis central to the investigation. Diffusion semigroups and
their domains are discussed next, together with ellipticity and hypo-ellipticity. At
this stage, a first summary of the various hypotheses from the semigroup viewpoint
is outlined in Sect. 1.14 (Chap. 3 will provide a complete description of the formal-
ism and hypotheses). Section 1.15 then presents a variety of basic tools and oper-
ations on semigroups and their infinitesimal generators, which are of constant use
and help in applications. The last section introduces the concept of curvature and di-
1.1 Markov Processes and Associated Semigroups 7
mension of a Markov (diffusion) operator which will play a crucial role throughout,
in particular in the study of functional inequalities.
1.1 Markov Processes and Associated Semigroups

Although one main object of investigation will be the study of Markov semigroups
by themselves, the reader should keep in mind that their study is mainly motivated
by the description of laws of Markov processes. The close relationships between
Markov semigroups and Markov processes will be a constant theme throughout this
work. Indeed, some very important features of Markov semigroups become clear
from the probabilistic interpretation, and many aspects of the behavior of Markov
processes follow from a careful analysis of the associated semigroups and infinites-
imal generators.
In this interplay, the very first example of a Markov process, briefly discussed
above, is that of Brownian motion with associated heat semigroup and infinitesimal
generator the Laplacian . This example took place in the standard Euclidean space
Rn (equipped with its Borel σ -field). In general, a Markov process (Xt )t≥0 is a
family of random variables constructed on some probability space ( , , P) with
values in some set E, equipped with a σ -field F , thus defining a measurable space
(E, F). The set E need not be topological, however F cannot in general be any
σ -field.
Indeed, some measurable spaces are better than others, namely those for which
the measure decomposition theorem applies. This theorem states that whenever μ is
a probability measure on the product σ -field F ⊗ F on E × E, and if μ1 denotes
its projection on the first coordinate, then μ can be decomposed as μ(dx, dy) =
k(x, dy)μ1 (dx) for some kernel k(x, dy) (that is, a probability measure depending
in a measurable way on the variable x). This decomposition theorem is true for a
rather large class of measurable spaces, but nevertheless it is not always true. For
example, it holds when E is a so-called Polish space (that is a separable topological
space, the topology of which may be defined by a complete metric) equipped with
its Borel σ -field. This example of course includes the case of Rn , or any open set in
Rn , with the Borel σ -field. In the following, by a good measurable space we shall
mean a measurable space (E, F) for which the measure decomposition theorem
applies, and for which there exists a countable family generating the σ -algebra F .
Whenever there is a reference measure on (E, F), the latter requirement may be
relaxed to the weaker condition that the countable family generates the σ -algebra
only up to sets of measure 0. In particular, all Lp -spaces, 1 ≤ p < ∞, on (E, F)
(equipped with a σ -finite measure) are separable under this assumption. Functions
on a measure space are understood as equivalence classes of (real-valued) functions
with respect to the relation of almost everywhere equality, and as such equalities
and inequalities between functions always hold almost everywhere. (See Sect. 1.14
for precise definitions and properties.)
Throughout, we shall always work with good measurable spaces. This level of
generality is justified by the numerous examples of applications and illustrations
of the use of Markov evolutions in mathematical physics and biology, for example,
dealing in particular with processes on spaces with an infinite number of coordi-
nates. For such spaces, the setting of locally compact spaces is clearly not sufficient.
Although we do not really discuss such infinite-dimensional models in this book, it
is important to choose the correct framework to start with.
Markov processes are characterized by the basic Markov property. Let (Xtx )t≥0
be a measurable process on a probability space ( , , P) starting at time t = 0
from x ∈ E. When the initial point x is fixed or clear from the context, (Xtx )t≥0
is denoted more simply by (Xt )t≥0 . For a given starting point x ∈ E, denote by
Ft = σ (Xu ; u ≤ t), t ≥ 0, the natural filtration of (Xt )t≥0 . The Markov property
then indicates that for t > s, the law of Xt given Fs is the law of Xt given Xs ,
as well as the law of Xt−s given X0 , the latter property reflecting the fact that the
Markov process is time homogeneous, which is the unique case that will be consid-
ered here. The process (Xt )t≥0 (starting at x) is then said to be a Markov process.
One prototypical example is Brownian motion (with values in E = Rn ) for which
the independence of the increments immediately ensures the Markov property.
In a more analytical language, the Markov property may be described on
the finite-dimensional distributions of the process (Xt )t≥0 . Namely, if pt (x, dy)
is the probability kernel describing the distribution of Xt starting from x, the
law of the pair (Xt1 , Xt2 ), t1 ≤ t2 , is obtained after conditioning by Ft1 as
pt1 (x, dy1 ) pt2 −t1 (y1 , dy2 ). Iterating the procedure, the law of the sample
(Xt1 , . . . , Xtk ), 0 < t1 ≤ · · · ≤ tk , given that X0 = x, is
pt1 (x, dy1 ) pt2 −t1 (y1 , dy2 ) · · · ptk −tk−1 (yk−1 , dyk ).
The description of the laws of the various k-uples (Xt1 , . . . , Xtk ), 0 < t1 ≤ · · · ≤ tk ,
is certainly not enough all the properties of the Markov process itself. For example,
it does not allow the description of the law of the random variable XT where T
is the first (random) time when the process (Xt )t≥0 enters some set A. For such
tasks, it is in general necessary to request additional assumptions such as the fact
that (Xt )t≥0 lives in a topological space and has regular paths (continuous, or at
least right-continuous). For the purpose of our investigation, the study of finite-
dimensional marginal laws will suffice, so we shall mainly concentrate on them.
Given such a Markov process (or family of processes) {Xtx ; t ≥ 0, x ∈ E}, de-
fine the associated Markov semigroup (Pt )t≥0 on suitable measurable functions
f : E → R by the conditional formula

Pt f (x) = E f Xtx = E f (Xt ) | X0 = x , t ≥ 0, x ∈ E. (1.1.1)
In particular P0 f = f . The Markov property then indicates that, for every suitably
integrable function f : E → R, and every t, s ≥ 0,
x x
Pt+s f (x) = E f Xt+s = E E f Xt+s | Ft
x x
= E E f Xt+s | Xt
x
= E Ps f Xt = Pt (Ps f )(x).
1.2 Markov Semigroups, Invariant Measures and Kernels 9
In other words, Pt+s = Pt ◦ Ps , t, s ≥ 0, which is the fundamental property of

Markov semigroups (Pt )t≥0 .
Note that by duality, those semigroups (Pt )t≥0 also act on the set of measures ν
on F via the formula

Pt f dν = f d Pt∗ ν (1.1.2)
E E
(for every t ≥ 0 and suitable function f : E → R), introducing the dual semigroup
(Pt∗ )t≥0 . Formally speaking, and in relation with the probabilistic interpretation of
Markov semigroups in terms of Markov processes, when ν is the law of X0 , Pt∗ ν is
the law of the variable Xt .
As will be discussed at length in subsequent sections and chapters, such a Markov
semigroup (Pt )t≥0 may be entirely described by an operator L, called the infinites-
imal generator of (Pt )t≥0 , which occurs in many examples as a differential opera-
tor (such as the standard Laplacian in the case of Brownian motion). Furthermore,
although no topology on the set E is required, as mentioned earlier, it is quite im-
portant in practice that the trajectories t → Xt are not too irregular. This regularity
is seen through the action of good functions and there is usually a nice family A of
functions f : E → R which is rich enough (the domain or a core of it, as defined
later) such that, for any f ∈ A, the processes t → f (Xt ), t ≥ 0, are almost surely
right-continuous with left-limits (càdlàg in the French abbreviation). We shall not
need this type of abstract construction. In the best cases, these processes are actually
almost surely continuous, corresponding to the so-called diffusion case (presented
in Sect. 1.11 below). This is in particular the case when the Markov processes are
given as solutions of stochastic differential equations as described in Sect. 1.10.
1.2 Markov Semigroups, Invariant Measures and Kernels
This section introduces the basic notions of Markov semigroup and invariant mea-
sure and, still in an informal way, some of the general principles governing their
analysis. The framework will be made progressively more precise in subsequent
sections. The basics for the construction and properties of abstract semigroups are
presented in Appendix A.
1.2.1 Markov Semigroups, Invariant Measures
The fundamental object of investigation consists of a family P = (Pt )t≥0 of op-

erators defined on some set of real-valued measurable functions on (E, F). The
following list describes the first definitions and properties of such an object:
(i) For every t ≥ 0, Pt is a linear operator sending bounded measurable functions
on (E, F) to bounded measurable functions.
(ii) P0 = Id, the identity operator (initial condition).

(iii) Pt (1) = 1, where 1 is the constant function equal to 1 (mass conservation).
(iv) If f ≥ 0, then Pt f ≥ 0 (positivity preserving).
(v) For every t, s ≥ 0, Pt+s = Pt ◦ Ps (semigroup property).
If f : E → R is bounded measurable, we write Pt f = Pt (f ). Properties (iii)
and (iv) are often also described as “Pt is a Markov operator”, giving rise to the
terminology. Note that, as a consequence of the positivity and mass conservation
properties, by Jensen’s inequality, for every convex function φ : R → R, every t ≥ 0
and every, say bounded, measurable function f on E,

Pt φ(f ) ≥ φ(Pt f ). (1.2.1)
In particular, for the functions φ(r) = |r|p , r ∈ R, 1 ≤ p ≤ ∞, Pt (|f |p ) ≥ |Pt f |p .

The preceding five conditions actually (almost) characterize what will be called
a Markov semigroup (the semigroup property being described by the last prop-
erty (v)). What is missing is a continuity property at t = 0 which will turn out to
be important, but requires a new notion to be fully appreciated, namely the notion
of invariant (or stationary) measure.
Definition 1.2.1 (Invariant measure) Given a family P = (Pt )t≥0 on a measurable

space (E, F) as before, a (positive) σ -finite measure μ on (E, F) is said to be
invariant for P if for every bounded positive measurable function f : E → R and
every t ≥ 0,

Pt f dμ = f dμ. (1.2.2)
E E
In other words, according to (1.1.2), μ is invariant if Pt∗ μ = μ for every t ≥ 0.
Note that, by (1.2.1), if a measure μ is invariant for P, then, for every t ≥ 0, Pt is a

contraction on the bounded functions in Lp (μ) for any 1 ≤ p ≤ ∞. As is traditional,
Lp (μ) = Lp (E, μ) with associated norm · p (sometimes also denoted · Lp (μ) ),
1 ≤ p ≤ ∞, denote the Lebesgue spaces on the measure space (E, F, μ). By den-
sity, Pt may thus be extended to a bounded operator on Lp (μ) for all 1 ≤ p ≤ ∞,
and even a contraction in the sense that Pt p,p ≤ 1 where P p,p denotes the op-
erator norm of P in Lp (μ). In particular, the invariance definition (1.2.2) extends to
any f ∈ L1 (μ).
The existence of such an invariant measure is not always guaranteed, but most
semigroups of interest do have an invariant measure. The uniqueness question, or
whether the invariant measure is finite or not, should then be further investigated.
These questions already arise for Markov chains on countable spaces (see Sect. 1.9).
Very often, an invariant measure may be constructed by choosing any
t reasonable
initial probability measure μ0 and by considering any weak limit of 1t 0 Ps∗ μ0 ds as
t → ∞ (where we recall the dual semigroup (Pt∗ )t≥0 described in (1.1.2)). However,
this method only works in general when the invariant measure is finite, and must be
adapted in the general case. A typical example is the heat or Brownian semigroup
on Rn for which the invariant measure is given (up to multiplication by a constant)

by the Lebesgue measure (which may be obtained from the heat semigroup through
the normalization (4πt)n/2 Pt∗ μ0 as t → ∞ for any reasonable initial probability
measure μ0 ).
The invariant measure is in general only defined up to a multiplicative constant.
When it is finite, a canonical choice is to normalize it into a probability measure.
It is then usually unique. However, there is no canonical normalization when the
measure is infinite. The analysis of Markov semigroups with a finite (normalized)
invariant measure is usually easier than the general case. This issue is similar to
the one for Markov chains on countable state spaces, for which the existence of a
finite invariant measure automatically ensures that the associated Markov process is
recurrent (see Sect. 1.6 below).
The invariant measure, at least when it is a probability measure, has a clear prob-
abilistic interpretation on the associated Markov process {Xtx ; t ≥ 0, x ∈ E}. If the
process starts at time t = 0 with initial distribution μ (that is, the law of X0 is μ),
then it keeps this distribution at each time t since, by the Markov property, for any
(say, bounded) measurable function f : E → R,

E f (Xt ) = E E f (Xt ) | X0

(1.2.3)
= E Pt f (X0 ) = Pt f dμ = f dμ = E f (X0 ) .
E E
This property justifies the terminology of stationary measure for μ in a probabilistic

context. The same interpretation remains true in general if one accepts the idea of
random variables with distributions having infinite mass.
Summarizing the construction of a Markov semigroup at this point, provided
μ is a σ -finite invariant measure for P = (Pt )t≥0 , the latter consists of a family
of operators bounded in Lp (μ) satisfying the properties (i)–(v). The last condition
required to deal with a Markov semigroup is the following continuity property:
(vi) For every f ∈ L2 (μ), Pt f converges to f in L2 (μ) as t → 0 (continuity prop-
erty).
Since (Pt )t≥0 is a contraction and a semigroup, the continuity property (vi) ex-
presses that t → Pt f is continuous in L2 (μ) on R+ . We could have required con-
vergence in Lp (μ) for any 1 ≤ p < ∞, although for simplicity we stick to p = 2.
This property will almost always be satisfied in the applications (however not for
p = ∞).
The continuity property (vi) actually reflects the regularity properties of the as-
sociated Markov process {Xtx ; t ≥ 0, x ∈ E} through the representation (1.1.1). In-
deed, assume that A is a dense subset of bounded functions in L2 (μ) such that, for
any f ∈ A, the process t → f (Xt ) has right-continuous paths for any initial value
X0 = x and, for any t0 > 0, supt≤t0 |Pt f | ∈ L2 (μ). This is for example the case for
bounded functions as soon as μ is a probability measure, and for any f ∈ L2 (μ)
if μ is reversible—see Lemma 1.6.2. Then, for such functions, and for any x ∈ E,
limt→0 Pt f (x) = f (x) and limt→0 Pt f = f in L2 (μ) (a property extended then to

any f ∈ L2 (μ) since Pt is a contraction).
On the basis of the above properties (i)–(vi), the following statement summarizes
the definition of a Markov semigroup with invariant measure μ.
Definition 1.2.2 (Markov semigroup) A family of operators P = (Pt )t≥0 de-

fined on the bounded measurable functions on a state space (E, F) with invariant
σ -finite measure μ satisfying the properties (i)–(vi) is called a Markov semigroup
of operators.
Accordingly, the Markov semigroup P = (Pt )t≥0 defines a semigroup of contrac-

tions on L2 (μ) in the sense of Appendix A.
Very often property (iii) may be relaxed to Pt (1) ≤ 1. This condition appears
naturally when solving Schrödinger-type equations with negative potential (see
Sect. 1.15.6 below), but also when the underlying state space is not complete (see
Sect. 2.5, p. 96, or Sect. 3.2.2, p. 141). In such cases, P = (Pt )t≥0 will be called a
sub-Markov semigroup. Sub-Markov semigroups are only considered in this book
when the need arises.
1.2.2 Kernels
Markov operators Pt , t ≥ 0, as given in Definition 1.2.2, may be represented by

(probability) kernels corresponding to the transition probabilities of the associated
Markov process (and this is the way Markov semigroups are usually given). Namely,
for every bounded measurable function f : E → R,

Pt f (x) = f (y) pt (x, dy), t ≥ 0, x ∈ E, (1.2.4)
E
where pt (x, dy) is, for every t ≥ 0, a probability kernel (that is, for every x ∈ E,
pt (x, ·) is a probability measure, and for every measurable set A ∈ F , x → pt (x, A)
is measurable). Recall that such a representation formula is actually understood for
μ-almost every x in E, as are all identities between functions throughout this work.
The distribution at time t of the underlying Markov process Xtx starting at x is thus
given by the probability pt (x, ·).
From the representation (1.2.4), the operators Pt , t ≥ 0, can be extended to pos-
itive measurable functions, possibly with infinite values. Often, inequalities involv-
ing Pt f are accordingly stated for bounded or positive measurable functions f .
The representation (1.2.4) requires good measurable spaces and the measure
decomposition theorem and is established in the next proposition (more for self-
consistency, since again Markov semigroups associated with Markov processes are
usually given explicitly in terms of kernels).
Proposition 1.2.3 (Kernel representation) Given a good measure space (E, F, μ)

(for example a Polish space), with μ σ -finite, let P be a linear operator sending pos-
itive measurable functions to positive measurable functions. Then, if P (1) = 1 and
if P is bounded on L1 (μ), P may be represented by a probability kernel p(x, A),
x ∈ E, A ∈ F , in the sense that for every bounded or positive measurable function
f on E and (μ-almost) every x ∈ E,

Pf (x) = f (y) p(x, dy).
E
Proof We use a result from measure theory, holding on good measurable spaces,
known as the bi-measure theorem. This theorem states that whenever (E, F) is
a good measurable space, if there is a map (A, B) → ν(A, B) from F × F into
R+ such that for every A, B ∈ F , the maps C → ν(A, C) and C → ν(C, B)
are measures on F , then there exists a measure ν1 on the product space E × E
equipped with the product σ -field F ⊗ F such that for every (A, B) ∈ F × F ,
ν(A, B) = ν1 (A × B).
On the basis of this result, we establish the proposition when μ is a probability
measure (the case when μ is σ -finite then easily follows). Set, for (A, B) ∈ F × F ,

ν(A, B) = 1A P (1B )dμ,
E
where 1A (respectively 1B ) is the characteristic function of the set A (respectively

B). It is a bi-measure. To prove this, it is enough
to check that whenever (Ak )k∈N
is a decreasing sequence of sets in F with k Ak = ∅, then ν(Ak , B) converges to
0 as k → ∞, and similarly for the second argument with a corresponding sequence
(Bk )k∈N decreasing to the empty set. For the sequence (Ak )k∈N , it suffices to ap-
ply the dominated convergence Theorem to the sequence 1Ak P (1B ), k ∈ N. For the
sequence (Bk )k∈N , observe that since P is continuous in L1 (μ), P (1Bk ) converges
to 0 in L1 (μ), and hence, there exists a subsequence converging to 0 μ-almost ev-
erywhere. Again by the dominated convergence Theorem, ν(A, Bk ) → 0 along this
subsequence, and since (ν(A, Bk ))k∈N is decreasing, ν(A, Bk ) → 0.
The bi-measure theorem thus applies to ν, and therefore there exists a (probabil-
ity) measure ν1 on the σ -field F ⊗ F such that

ν1 (A × B) = 1A P (1B )dμ
E
for every A, B ∈ F . Since P (1) = 1, the first marginal of ν1 is μ and ν1 may be

decomposed by the measure decomposition theorem as
ν1 (dx, dy) = p(x, dy)μ(dx).
It is then easily checked that the kernel p(x, dy) fulfills the required conditions. The
proposition is established.
Very often, the family of kernels pt (x, dy) have densities with respect to a refer-
ence measure (often also the invariant measure).
Definition 1.2.4 (Density kernel) A Markov semigroup P = (Pt )t≥0 on (E, F) is

said to admit density kernels with respect to a reference σ -finite measure m on F
if there exists for every t > 0 a positive measurable function pt (x, y) defined on
E × E (up to a set of m ⊗ m-measure 0) such that, for every bounded or positive
measurable function f : E → R and (m-almost) every x ∈ E,

Pt f (x) = f (y) pt (x, y)dm(y).
E

In this case, E pt (x, y)dm(y) = 1 for (m-almost) every x ∈ E (reflecting the fact
that Pt (1) = 1).
In order for Pt f to make sense for any f ∈ L2 (m) in this definition, it

is
in general required that for all t > 0 and (m-almost) every x ∈ E,

2 dm(y) < ∞. In this case, (P )
t t≥0 is said to admit L -density kernels
p (x, y) 2
E t
with respect to m. As already mentioned, the reference measure will often be chosen
to be the invariant measure μ.
In the context of Definition 1.2.4, many of the properties of the Markov semi-
group (Pt )t≥0 may be translated into properties of the density kernels (for example
the Fokker-Planck equation described in Sect. 1.5 or reversibility in Sect. 1.6). How-
ever, it is not always easy to check that density kernels exist, which usually requires
additional knowledge on the semigroup. Beyond the case of Hilbert-Schmidt opera-
tors (see Appendix A, Sect. A.6, p. 483) where an explicit expression for this kernel
may be given (see, for example, (1.7.3) below), the main task would be to first prove
that the semigroup is a bounded operator from L1 (m) into L∞ (m). The following
proposition, which will illustrated in the study of heat kernel bounds in Chaps. 6
and 7, is a useful result in this regard.
Proposition 1.2.5 (Existence of density kernel) Let (E, F, m) be a good mea-

sure space with m σ -finite. Let P be a linear operator mapping L1 (m) into
L∞ (m) with operator norm P 1,∞ ≤ M. Then, there exists a measurable func-
tion (kernel density) p(x, y) on E × E with |p(x, y)| ≤ M for m ⊗ m-almost every
(x, y) ∈ E × E such that, for any f ∈ L1 (m) and (m-almost) every x ∈ E,

Pf (x) = f (y) p(x, y)dm(y).
E
Proof We only prove the result in the case when m is a probability measure, the gen-
eral case being easily adapted by σ -finiteness of m. For a good measurable space,
there exists a sequence (A )∈N of measurable sets which generates the σ -field
F up to sets of m-measure 0. For each ∈ N, let F be the σ -field generated by
A0 , . . . , A . It is also generated by a partition (B1 , . . . , Bj ) of E from which the
sets with m-measure 0 have been removed. For f ∈ L1 (m), define P f as the condi-
tional expectation E(Pf | F ) of Pf with respect to F . The operator P is bounded
from L1 (m) into L∞ (m) with norm M and may be represented as

j
P f (x) = Qj f 1B (x)
j
j =1
where

1
Qj f = Pf dm.
m(Bj ) Bj
The Qj ’s, j = 1, . . . , j , are linear operators, bounded on L1 (m), with norm
bounded above by M. By duality,

Qj f =

f (y) qj (y)dm(y)
E
for some bounded (by M) measurable functions qj , j = 1, . . . , j . Restricted to

functions which are F -measurable, the functions qj are linear combinations of
1B , j = 1, . . . , j . Thus, there exists a kernel p = p (x, y) which is F ⊗ F -
j
measurable, bounded by M, and such that for any integrable function f : E → R
measurable with respect to F and any x ∈ E,

P f (x) = f (y) p (x, y)dm(y).
E
In other words, p is the kernel of the operator E(P (E(f | F )) | F ). It is then not
hard to see that p , ∈ N, is an m ⊗ m-martingale with respect to the increasing
family of σ -fields F ⊗ F , ∈ N. Since it is bounded, by elementary martingale
theory, it converges (m ⊗ m-almost everywhere) to a measurable function p(x, y),
(x, y) ∈ E × E, bounded by M, which is easily seen to be a kernel of the operator
P . The proof is complete.
As a further general fact, it is worth mentioning that as soon as a Markov operator

P on a probability space (E, F, μ) is bounded from above by a constant C < 2, then
P 2 = P ◦ P is also bounded from below. This is the content of the following simple
proposition.
Proposition 1.2.6 Let P be a Markov operator (thus satisfying properties (iii)

and (iv) of the definition of a Markov semigroup) with density kernel p(x, y),
(x, y) ∈ E × E, with respect to an invariant probability measure μ on (E, F). Then,
as soon as P 1,∞ ≤ 1 + ε for some 0 ≤ ε < 1, for every positive function f in
L1 (μ),

√ 2
P f ≥ 1 − ε (1 + ε) f dμ.
2
E
In particular,
√ 2 the density kernel p 2 (x, y) of P 2 is bounded from below by
(1 − ε ) (1 + ε).

Proof Assume by homogeneity that E f dμ = 1. Setting g = Pf , g ≤ 1 + ε and
E gdμ = 1 since μ is invariant for P . For any 0 < a < 1 + ε, P f = P (g) ≥

2
a P (1{g≥a} ). But
P (1{g≥a} ) = 1 − P (1{g<a} ) ≥ 1 − (1 + ε)μ(g < a)
and by Markov’s inequality,
μ(g < a) = μ(1 + ε − g > 1 + ε − a)

1 ε
≤ (1 + ε − g)dμ = .
1+ε−a E 1+ε−a
Therefore,
a(1 − ε 2 − a)
P 2f ≥
1+ε−a
√
and it remains to choose the optimal a = (1 + ε)(1 − ε ). The proof is complete.
It is easy to see that the condition ε < 1 is necessary in the above statement:
consider for example on E = {0, 1} the Markov operator P given by Pf (0) = f (1)
and Pf (1) = f (0), invariant with respect to the uniform measure μ, for which ε = 1
and P 2 = Id. For this P , the kernel p 2 (x, y) with respect to μ is such that p 2 (0, 1) =
p 2 (1, 0) = 0.
1.3 Chapman-Kolmogorov Equations
Section 1.1 describes the basic principle used to construct a Markov semigroup from
a Markov process. Now conversely, given a Markov semigroup, the construction of a
Markov process associated to it relies on the Chapman-Kolmogorov equation which
expresses the semigroup property from a probabilistic point of view. As constructed
in the preceding sections, kernels and density kernels are only defined almost ev-
erywhere with respect to the underlying measure, and identities between them are
understood in this sense below.
Let P = (Pt )t≥0 be a Markov semigroup on L2 (μ) according to Definition 1.2.2.
The semigroup property Pt ◦ Ps = Pt+s translates to the kernels pt (x, dy) of the
representation (1.2.4) via the composition property, for all t, s ≥ 0, x ∈ E,

pt+s (x, dy) = pt (z, dy) ps (x, dz). (1.3.1)
z∈E
Such a relation is called a Chapman-Kolmogorov equation.

1.3 Chapman-Kolmogorov Equations 17
When the kernels admit densities pt (x, y), t > 0, (x, y) ∈ E × E, with respect to
a reference measure m on (E, F) as in Definition 1.2.4, the latter equation expresses
that, for all t, s > 0, (x, y) ∈ E × E,

pt+s (x, y) = pt (z, y) ps (x, z)dm(z). (1.3.2)
E
Now, as announced, from the Chapman-Kolmogorov equation (1.3.1), one may

construct, starting from any point x ∈ E, a Markov process (Xt )t≥0 on E by speci-
fying the distribution of (Xt1 , . . . , Xtk ), 0 ≤ t1 ≤ · · · ≤ tk , as

E f (Xt1 , . . . , Xtk ) = f (y1 , . . . , yk ) ptk −tk−1 (yk−1 , dyk )
E×···×E (1.3.3)
× ptk−1 −tk−2 (yk−2 , dyk−1 ) · · · pt1 (x, dy1 )
for every, say, positive or bounded measurable function f on the product space
E × · · · × E. When the distribution of such a k-dimensional vector (Xt1 , . . . , Xtk ) is
specified, it determines the distributions of each extracted lower dimensional vector
(for example, the law of (Xt1 , Xt3 ) can be deduced from that of (Xt1 , Xt2 , Xt3 )). It
is thus necessary that the system of finite-dimensional distributions be compatible,
which is precisely the content of the Chapman-Kolmogorov equation.
Equation (1.3.3) describes the law of the process (Xt )t≥0 starting at X0 = x. If
the initial distribution of X0 is given by some other measure ν, then the law of
(X0 , Xt1 , . . . , Xtk ) is given by

E f (X0 , Xt1 , . . . , Xtk ) = f (y0 , y1 , . . . , yk ) ptk −tk−1 (yk−1 , dyk ) · · ·
E×···×E (1.3.4)
× pt1 (y0 , dy1 ) ν(dy0 ).
Given a Markov semigroup P = (Pt )t≥0 , we therefore have candidates for the
distribution of a Markov process (Xt )t≥0 starting at a point x ∈ E. This entails a
correspondence between Markov processes and Markov semigroups (even without
the continuity condition (vi)). However, as already mentioned, the preceding finite-
dimensional description is not enough in general to characterize the full law of the
Markov process and its regularity properties. Usually, is does not even describe the
joint laws of the processes starting from different initial values x and y, as is the case
for example when solving stochastic differential equations. In fact, the process often
lives on a topological space, and one then looks for a process such that the trajec-
tories (the random maps t → Xt (ω) for ω ∈ ) are continuous, or right-continuous
with left limits (càdlàg). It requires extra work to construct these laws as probability
measures on the space of continuous (or càdlàg) maps on R+ with values in the state
space E. This question will not be addressed in this book, since most of the time the
concrete processes will be explicitly given, for example as solutions of stochastic
differential equations (see Sect. 1.10).
In the following, we will indistinctly speak of a Markov semigroup P = (Pt )t≥0

with state space E and of its associated Markov process, or family of Markov pro-
cesses, X = {Xtx ; t ≥ 0, x ∈ E}. When the starting point is implicit from the context
or fixed, we often simply denote the Markov process by (Xt )t≥0 .
1.4 Infinitesimal Generators and Carré du Champ Operators

This section introduces infinitesimal generators and carré du champ operators of
Markov semigroups which are central objects of interest encoding numerous fea-
tures and properties.
1.4.1 Infinitesimal Generators
A Markov semigroup P = (Pt )t≥0 as defined in Definition 1.2.2 is driven by an

operator called the infinitesimal generator of the Markov semigroup. According to
Appendix A, it may actually be introduced in a general framework.
A family of bounded linear operators (Pt )t≥0 on a Banach space B with the
semigroup (Pt ◦ Ps = Pt+s , t, s ≥ 0) and continuity (Pt f → f for every f ∈ B
as t → 0) properties is called a semigroup of (bounded) linear operators (on the
Banach space B). In the preceding setting of Markov (contraction) semigroups, the
Banach space B will always be the Hilbert space L2 (μ), but we could similarly work
(adapting (vi) of the definition of Markov semigroup) with Lp (μ), 1 ≤ p < ∞, or
the space of bounded continuous functions (provided E is topological), or any space
on which the family of operators (Pt )t≥0 acts naturally.
In this setting, the Hille-Yosida theory (cf. Sect. A.1, p. 473) indicates that there
is a dense linear subspace of B, called the domain D of the semigroup (Pt )t≥0 , on
which the derivative at t = 0 of Pt exists in B. The operator that maps f ∈ D to this
derivative Lf ∈ B of Pt f at t = 0 is a linear (usually unbounded) operator, called
the infinitesimal generator of the semigroup (Pt )t≥0 , denoted L. The domain D is
also called the domain of L, denoted D(L), and depends on the underlying Banach
space B. Actually, the generator L and its domain D(L) completely characterize the
semigroup (Pt )t≥0 acting on B.
Applied to a Markov semigroup P = (Pt )t≥0 as presented in Definition 1.2.2 with
B = L2 (μ) the preceding yields the following definition.
Definition 1.4.1 (Infinitesimal generator) Let P = (Pt )t≥0 be a Markov semigroup

with state space (E, F) and invariant measure μ. The infinitesimal generator L of
P in B = L2 (μ) is called the Markov generator of the semigroup P = (Pt )t≥0 with
L2 (μ)-domain D(L).
When dealing with B = Lp (μ), 1 ≤ p < ∞, in Definition 1.4.1, we sometimes

speak of the Lp (μ)-domain of L. In view of the connection with the Markov process
1.4 Infinitesimal Generators and Carré du Champ Operators 19
X associated with the semigroup P, the generator L will also be called the Markov
generator of X.
The linearity of the operators Pt , t ≥ 0, together with the semigroup property,
shows that L is the derivative of Pt at any time t > 0. Namely, for t, s > 0,

1 1 1
[Pt+s − Pt ] = Pt [Ps − Id] = [Ps − Id] Pt .
s s s
Letting s → 0 then yields

∂t Pt = LPt = Pt L. (1.4.1)
While traditionally reserved for Laplace operators , by extension (Pt )t≥0 will of-
ten be called the heat semigroup or heat flow with respect to the generator L (thus
solving the heat equation (1.4.1)).
Note that replacing L with c L for some c > 0 amounts to the time change t → ct.
This seemingly insignificant transformation actually allows us to freely normalize
generators. More sophisticated transformations, changing L to c(x)L where c(x) is
a strictly positive function, will be considered below in Sect. 1.15.2.
In the illustrations, it is usually difficult, but also often useless, to exactly describe
the domain D(L) of the generator L. It will be enough to determine a so-called core
D0 , that is a dense subspace with respect to the topology of the domain in the sense
that for every element f ∈ D(L), there is a sequence (fk )k∈N in the core D0 con-
verging to f and such that Lfk converges to Lf . The question of the description of
the domain, or a core of it, is a delicate question which we will return to later (see
Sect. 1.13). We refer the reader to Appendix A for further examples. In most exam-
ples of interest, and in particular in this monograph, the operator L is a differential
operator (for example the usual Laplace operator for the heat or Brownian semi-
group on Rn ). The statement that a function f is in the domain D(L) then hides a
number of properties. A first property would be that the function f should be smooth
enough so that the action of the generator (for example in the distributional sense)
gives rise to a function Lf in L2 (μ). As a second property, the function should be
well-behaved at infinity (or at the boundary of the state space) so that integration by
parts formulas (hidden in the symmetry property introduced in Sect. 1.6 and at the
heart of the analysis in this book) can be used.
The reader should also pay attention to the fact that in Definition 1.4.1 the semi-
group P = (Pt )t≥0 is given and that the infinitesimal generator L (and its domain)
is determined from it. One of the main difficulties in the analysis of semigroups is
that in general we would like to work in the other direction, that is, starting from the
knowledge of a Markov generator L on a set A0 of “nice” functions (for example,
when L is a differential operator, the set of smooth compactly supported functions).
This amounts to proving that this set of functions is a core in the domain, but for
a semigroup not yet described. Whether this knowledge is enough to determine the
semigroup (Pt )t≥0 is a delicate matter, related in the symmetric case to the question
of essential self-adjointness (see Appendix A, Sect. A.5, p. 481, and Chap. 3, Defi-
nition 3.1.10, p. 134). The question of determining when some unbounded operator
L given on such a class A0 may be extended to a larger domain in order to become

the infinitesimal generator of a Markov semigroup will be addressed in Chap. 3 to-
gether with the necessary tools to achieve this goal in practice. In the core of the
text, some efficient criteria to achieve the latter property in concrete examples will
be presented (see for instance Proposition 2.4.1, p. 95, Proposition 2.6.1, p. 102,
Proposition 3.2.1, p. 142, or Proposition A.5.4, p. 482.)
With respect to abstract semigroups of operators, the Markov semigroups
P = (Pt )t≥0 considered in this work share some further properties described by
Definition 1.2.2. In particular, they act on function spaces, satisfy the fundamental
stability property of preserving the unit function, and admit an invariant measure.
These properties may be translated (equivalently) on the generator L of the semi-
group P. If the constant 1 function is in L2 (μ), then it is in the domain D(L) and
L(1) = 0. This is the case when μ is finite. When the measure μ is infinite, 1
cannot be in the domain, and it is not enough in general to know that L(1) = 0
in order to conclude that conversely Pt (1) = 1 for every t ≥ 0. Further conditions
are necessary (such as boundary conditions), or the generator should itself satisfy
suitable
properties. Concerning

the invariant measure, differentiating at t = 0 the
identity E Pt f dμ = E f dμ (which is justified as soon as Lf is bounded for ex-
ample, or by Lemma 1.6.2 below in the reversible case) shows that if f is in the
L1 (μ)-domain, then

Lf dμ = 0.
E
Conversely, if this condition holds on a set of functions f which is dense in the
L1 (μ)-domain of L, then the measure μ will be invariant for P.
1.4.2 Carré du Champ Operators
To a Markov semigroup P = (Pt )t≥0 and its infinitesimal generator L, with (L2 (μ)-)
domain D(L), is naturally associated a bilinear form. Assume that we are given a
vector subspace A of the domain D(L) such that for every pair (f, g) of functions
in A, the product f g is in the domain D(L) (A is an algebra).
Definition 1.4.2 (Carré du champ operator) The bilinear map

1
(f, g) = L(f g) − f Lg − g Lf
2
defined for every (f, g) ∈ A×A is called the carré du champ operator of the Markov
generator L.
The definition of the carré du champ operator is clearly subordinate to the algebra
A. Such a class A will be mostly natural in a given context (for example smooth
functions on a manifold) and will be carefully described and discussed in Chap. 3.
1.4 Infinitesimal Generators and Carré du Champ Operators 21
The carré du champ operator (which is so named for reasons which will be made
clear below) will play a crucial role throughout this work. Numerous examples will
be made discussed in this monograph, but it may already be useful to mention the
simple example of the Laplacian L = on Rn giving rise to the standard carré du
champ operator (f, g) = ∇f · ∇g (the usual scalar product of the gradients of f
and g) for smooth functions f, g on Rn .
If f ∈ A, since Pt (f 2 ) ≥ (Pt f )2 for every t ≥ 0 by (1.2.1), in the limit as t → 0,
L(f 2 ) ≥ 2f Lf . It follows that the carré du champ operator is positive on A in the
sense that
(f, f ) ≥ 0, f ∈ A. (1.4.2)
This is a fundamental property of Markov semigroups. To somewhat lighten the
notation, we set (f ) = (f, f ), f ∈ A. Observe also that by bilinearity, (1.4.2)
immediately yields the Cauchy-Schwarz inequality
(f, g)2 ≤ (f ) (g), (f, g) ∈ A × A. (1.4.3)
As already mentioned, the operator L is usually only determined on a core of

the domain, however it completely characterizes the semigroup P = (Pt )t≥0 . One
of the main difficulties is to translate the properties of L to the semigroup P. A first
instance is positivity. As emphasized above, the positivity of the carré du champ
operator (1.4.2) is one of the main properties which will be carefully described
in examples of interest. Positivity may actually be addressed more generally. If
φ : R → R is a convex function, and f is a function in the domain D(L) such
that φ(f ) also belongs to D(L), then Jensen’s inequality (1.2.1) together with the
definition of L shows that
Lφ(f ) ≥ φ (f ) Lf. (1.4.4)
The fact that (f ) = (f, f ) ≥ 0 is precisely this property for the convex function
φ(r) = r 2 , r ∈ R. This property, together with the invariant measure μ, actually
establishes (at least formally) the positivity
preserving property of the semigroup
(Pt )t≥0 . Indeed, if f ≥ 0, setting (t) = E φ(Pt f )dμ, t ≥ 0, and differentiating
together with (1.4.4) yields that

(t) = φ (Pt f ) LPt f dμ ≤ Lφ(Pt f )dμ.
E E
But now, if φ(Pt f ) is in the domain D(L), then the integral on the right-hand side
of the preceding inequality is zero since μ is invariant. Therefore is decreasing
so that

φ(Pt f )dμ ≤ φ(f )dμ
E E
for every t ≥ 0. In particular, for the convex function φ(r) = |r|, and with f ≥ 0,

|Pt f |dμ ≤ |f |dμ = f dμ = Pt f dμ. (1.4.5)
E E E E
Therefore, Pt f ≥ 0 μ-almost everywhere whenever f ≥ 0, showing that Pt pre-

serves positivity for every t ≥ 0.
Now of course the preceding argument relies on (1.4.4) rather than only the pos-
itivity of . However, it is not hard to see that the positivity of implies (1.4.4)
when φ is a convex polynomial. Extending this to more general φ’s requires further
hypotheses. But there should be no surprise that in any given example, as soon as
is positive in the sense of (1.4.2), then (1.4.4) is also valid. In particular, for diffusion
semigroups as discussed later in Sect. 1.11, the diffusion property
Lφ(f ) = φ (f ) Lf + φ (f ) (f )
(for smooth φ : R → R) immediately yields (1.4.4) from the positivity prop-

erty (1.4.2).
1.4.3 Martingale Problem
To conclude this section, we illustrate the notion of infinitesimal generators from

the probabilistic viewpoint. Given a Markov semigroup P = (Pt )t≥0 , the gener-
ator L of P may be used to describe the law of the associated Markov process
X = {Xtx ; t ≥ 0, x ∈ E}. In the good cases (that is when there exists a countable
dense subset of the domain D(L) of L), the process (Xt )t≥0 may be chosen in such
a way that for any f in the domain D(L), the map t → f (Xt ) is right-continuous
and
t
f
Mt = f (Xt ) − Lf (Xs )ds, t ≥ 0, (1.4.6)
0
is a local martingale. This type of abstract construction will not be necessary in what
follows, but it might be useful to know that the Markov process X (as a probabil-
ity measure on the space of right-continuous functions with values in E) may be
obtained directly from the operator L. This construction amounts to solving the so-
called martingale problem associated with L, that is defining a probability law (or a
family of them) on the space of functions t → Xt from R+ with values in E such
f
that for any function f in the domain, the expression Mt given in (1.4.6) is a mar-
tingale. When the process is given as a solution of a stochastic differential equation
as described in Sect. 1.10, this martingale property will appear as a by-product of
Itô’s formula (see Appendix B, Sect. B.4, p. 495). Observe that, when the preceding
local martingale is a true martingale (for example when it is uniformly integrable,
f f
or better bounded), then taking expectation in (1.4.6) yields E(Mt ) = E(M0 ) and
therefore the identity
t
Pt f (x) = f (x) + Ps Lf (x)ds (1.4.7)
0
(for every t ≥ 0 and x ∈ E).

1.5 Fokker-Planck Equations 23
This picture may be further extended to smooth functions F (t, x) depending on

t ≥ 0 and x ∈ E under suitable conditions such that
t
F (t, Xt ) − (∂s + L)F (s, Xs ) ds, t ≥ 0, (1.4.8)
0
is also a (local) martingale. In particular, when choosing F (t, x) = PT −t f (x) for

f bounded and t ∈ [0, T ], F (t, Xt ) is a martingale on [0, T ]. Since
(∂s + L)F (s, Xs ) = 0, taking expectations at t = T , one recovers that

PT f (x) = E f (XT ) | X0 = x .
From the abstract semigroup viewpoint, the stochastic representation of the mar-
tingale problem (1.4.7) indicates that, for every t ≥ 0,
t
Pt = Id + Ps L ds. (1.4.9)
0
The second formulation (1.4.8) on the other hand is concerned with the heat equa-
tion ∂t G = LG (G(0, x) = f (x)). To formally solve it, it suffices to consider the
exponential of the operator L as
tk
Qt f = et L f = Lk f (1.4.10)
k!
k≥0
which clearly satisfies ∂t Qt = LQt , and thus Pt = Qt = et L . Hence, the first for-
mulation (1.4.7) expresses that ∂t Pt = Pt L while the second (1.4.8) indicates that
∂t Pt = LPt . The fact that the operator L commutes with Pt actually tells us that
Pt is in a sense a function of L. This statement will be given a precise meaning in
various examples below, but this basic observation can serve as a guide through-
out this monograph. Note, however, that it is quite impossible, given the represen-
tation (1.4.10) of Pt = et L , to analyze properties such as positivity preservation,
which is central to the analysis of Markov semigroups.
1.5 Fokker-Planck Equations

This short section describes a dual point of view on the Chapman-Kolmogorov equa-
tion, which is more commonly found in the partial differential equations literature.
Given a Markov semigroup P = (Pt )t≥0 with infinitesimal generator L as in the
preceding sections, by (1.2.4) the operators Pt , t ≥ 0, may be represented by Markov
kernels. Usually, they admit densities pt (x, y), t > 0, (x, y) ∈ E × E, with respect
to some reference measure m (for example the Lebesgue measure), not necessarily
the invariant measure, as described in Definition 1.2.4, so that

Pt f (x) = f (y) pt (x, y)dm(y), t > 0, x ∈ E,
E
on a suitable class of functions f on E. Since L is the generator of the semigroup

P = (Pt )t≥0 , the function pt (x, y), t > 0, x ∈ E, is a solution of the heat equation
with respect to L
∂t pt (x, y) = Lx pt (x, y), p0 (x, y) dm(y) = δx , (1.5.1)
where Lx denotes the operator L acting on the x variable. This expresses that
∂t Pt f = LPt f.
But one may similarly consider the dual equation, called the Fokker-Planck equation
(or Kolmogorov forward equation), for t > 0,
∂t pt (x, y) = L∗y pt (x, y), (1.5.2)
where L∗ is the adjoint of L with respect to the reference measure m in the sense
that

∗
f L g dm = g Lf dm
E E
for suitable functions f, g (see Sect. A.2, p. 475). The latter identity (1.5.2) then
expresses that
∂t Pt f = Pt Lf.
This dual Fokker-Planck equation may be obtained, at least formally, admitting that
derivation in t > 0 commutes with the integral in y, via the identities

∂t pt (x, y)f (y)dm(y) = pt (x, y) Lf (y)dm(y)
E E

= f (y) L∗y pt (x, y)dm(y)
E
which are valid on a suitable class of functions f , giving rise to (1.5.2).

Care has to be taken with the fact that the heat and Fokker-Planck equations do
not play the same role. When writing ∂t Pt f = Pt Lf , the function f is implicitly
in the domain of L (which usually requires that f is sufficiently smooth). However
the equation ∂t Pt f = LPt f can in general be applied for irregular functions (only
bounded and measurable, for example) since the operator Pt will usually be smooth-
ing, mapping irregular functions to regular ones (see Sect. 1.12). More details will
be provided in Chap. 3. One of the by-products of the tools developed in this work in
order to measure the convergence to equilibrium will be to quantify the smoothing
properties of the operators Pt .
1.6 Symmetric Markov Semigroups

One of the central themes of this work will be to detail the many useful properties of
the important class of symmetric Markov semigroups and their associated reversible
1.6 Symmetric Markov Semigroups 25
Markov processes. As before, denote by P = (Pt )t≥0 a Markov semigroup with state
space (E, F), invariant measure μ and infinitesimal generator L with L2 (μ)-domain
D(L). The associated Markov process is denoted by X = {Xtx ; t ≥ 0, x ∈ E}.
1.6.1 Symmetric Markov Semigroups, Reversible Measures
Symmetry of a Markov semigroup, or reversibility of the invariant measure, are de-

fined as follows.
Definition 1.6.1 (Symmetric Markov Semigroup) The Markov semigroup

P = (Pt )t≥0 is said to be symmetric with respect to the invariant measure μ, or
μ is reversible for P, if for all functions f, g ∈ L2 (μ) and all t ≥ 0,

f Pt g dμ = g Pt f dμ. (1.6.1)
E E
If the semigroup P = (Pt )t≥0 admits density kernels pt (x, y), t > 0,
(x, y) ∈ E × E, with respect to μ, then pt is a symmetric function on E × E.
If μ is a probability measure, the symmetry property for f = 1 shows that μ is
invariant. The converse is not true in general. Under reasonable minimal assump-
tions, an invariant measure is unique. Such a measure may or may not be reversible.
Hence Markov semigroups are divided into two classes, according to whether the in-
variant measure is reversible or not. In the fundamental examples studied in Chap. 2
we will learn how to distinguish between these two classes.
From a probabilistic point of view, the name reversible refers to reversibility
in time of the associated Markov process whenever the initial law is the invariant
measure. Indeed, from (1.2.3), we know that if the process starts from the invari-
ant distribution μ, then it keeps the same distribution for any time. Moreover, if
the measure is reversible, and if the initial distribution of X0 is μ, then for any
t > 0 and any partition 0 ≤ t1 ≤ · · · ≤ tk ≤ t of the time interval [0, t], the law of
(X0 , Xt1 , . . . , Xtk , Xt ) is the same as the law of (Xt , Xt−t1 , . . . , Xt−tk , X0 ). This is
easily seen from (1.3.4) when the initial measure ν is μ since in this case the mea-
sure pt (x, dy)μ(dx) is symmetric in (x, y), and then, by an immediate induction,
the measures
ptk −tk−1 (yk−1 , dyk ) ptk−1 −tk−2 (yk−2 , dyk−1 ) · · · pt1 (y0 , dy1 ) μ(dy0 )
are invariant under the change (y0 , . . . , yk ) → (yk , . . . , y0 ). Therefore, for any t > 0,
the law of the process (Xs )0≤s≤t is the same as the law of the process (Xt−s )0≤s≤t .
Hence, the law of the Markov process is “reversible in time”.
In terms of the generator L of the semigroup P, the reversibility property may be
expressed as

f Lg dμ = g Lf dμ (1.6.2)
E E
on suitable functions f, g in the domain D(L), as can be checked by differentiating

the identity (1.6.1) at t = 0. The operator L is thus a symmetric operator, however it
is unbounded on L2 (μ).
As discussed in Appendix A, there is a fundamental difference in infinite-
dimensional Hilbert spaces, namely that a symmetric operator is not necessarily
self-adjoint (see there for a summary of the fundamental properties of symmetric
and self-adjoint operators). For a symmetric semigroup (Pt )t≥0 , the generator L
defined on its domain D(L) is always self-adjoint. Nevertheless, as already em-
phasized and precisely described later in Chap. 3, we are in general given some
symmetric operator L on a class of functions (such as differential operators defined
on smooth compactly supported functions) for which the semigroup has yet to be
constructed. Those symmetric operators considered in this book always admit at
least one self-adjoint extension, that is the domain D(L) may always be extended
in such a way that L becomes a self-adjoint operator, which in turn entirely de-
scribes the semigroup (Pt )t≥0 (see Sect. 3.1, p. 120). Whether or not this exten-
sion is unique is related, as mentioned earlier, to the question of the uniqueness of
the semigroup (Pt )t≥0 associated to the description of L on D(L) and to the es-
sential self-adjointness property (Definition 3.1.10, p. 134). For example, on the
interval (−1, +1), there are many semigroups (Pt )t≥0 for which the generator co-
incides with the Laplace operator Lf = f acting on functions f which are smooth
and compactly supported in (−1, +1), which defines a symmetric operator for the
Lebesgue measure dx. In a probabilistic context, two such examples are the semi-
group associated with the Brownian motion killed at the boundary and the Brownian
motion reflected at the boundary. These examples correspond to two different self-
adjoint extensions of the operator, although the space of functions which are smooth
and compactly supported in (−1, +1) is dense in L2 (dx) (see Sects. 2.4, 2.5 and 2.6,
pp. 92, 96 and 97).
The following is a simple useful application of martingale theory to the analysis
of symmetric Markov semigroups which illustrates the interplay between probabil-
ity and analysis. The following lemma is in particular useful as mentioned earlier
in the study of the regularity properties of the associated Markov processes. The
proof relies on maximal inequalities for martingales. The result itself will not re-
ally be used later and should be considered more as an illustration of the power of
probabilistic techniques in this context.
Lemma 1.6.2 (Rota’s Lemma) Let P = (Pt )t≥0 be a Markov semigroup symmetric
with respect to μ. Then, for any 1 0 such that for any
measurable function f : E → R,

sup Ps f ≤ Cp f p .
s≥0 p
Proof We sketch the proof in the case where μ is a probability measure (the general
case has to be adapted with some care). It suffices to assume that f is positive
and, by approximation, bounded. Let (Xt )t≥0 be the Markov process with initial
1.6 Symmetric Markov Semigroups 27
distribution μ associated with (Pt )t≥0 . For T > 0, consider the (positive) martingale
Mt = PT −t (f )(Xt ), t ∈ [0, T ] (cf. Sect. 1.4.3). By Doob’s maximal inequality,

p p p
E sup Mt ≤ Cp E MT
0≤t≤T
for some constant Cp > 0. But MT = f (XT ) and since (by invariance) the law of
p p
XT is μ, E(MT ) = f p . Furthermore, conditional expectation is a contraction in
L (P), so that
p
p
p
E E sup Mt | XT ≤ Cp f p .
0≤t≤T
Now reversibility ensures that the law of (Xt , XT ) is the same as that of (XT −t , X0 )
so that E(Mt | XT ) = P2(T −t) f (XT ). It follows that

E sup Mt | XT ≥ sup E(Mt | XT ) = sup P2(T −t) f (XT ).
0≤t≤T 0≤t≤T 0≤t≤T
In conclusion

sup Pt f ≤ Cp f p
0≤t≤2T p
and it remains to let T go to infinity to get the result.
1.6.2 Markov Triple
When the measure μ is symmetric, the generator L and the semigroup P = (Pt )t≥0
are completely described by the measure μ and the carré du champ operator from
Definition 1.4.2. Indeed, for all functions f, g in the algebra A ⊂ D(L) on which
is defined,

(f, g)dμ = − f Lg dμ. (1.6.3)
E E
This basic formula is immediate by integration

of the definition (1.4.2) of (f, g)
together with the invariance property E Lf dμ = 0 and is actually an integration by
parts formula as will be seen explicitly in the case of diffusion processes. It will turn
out to be a fundamental tool in the analysis of Markov semigroups. (Note that in the
non-symmetric case, with μ invariant, the above integration by parts formula (1.6.3)
does not hold for all f and g.) Since is bilinear and positive ((1.4.2)), for every
function f for which Lf is well-defined,

f (−Lf )dμ ≥ 0.
E
The operator −L is therefore a positive operator, and for such positive operators,
the theory of symmetric and self-adjoint operators is much easier, as presented in
Appendix A.
As the carré du champ operator and the measure μ completely determine the
symmetric Markov generator L, we will mostly work throughout this monograph
with what will be called a Markov Triple (E, μ, ) consisting of a (σ -finite) measure
μ on a state space (E, F) and a carré du champ operator (on some suitable class
A of functions on E). This framework will be carefully presented in Chap. 3. The
adoption of this framework is particularly justified in our investigation of functional
inequalities in later chapters. As a standard example, consider the Lebesgue measure
on the Borel sets of Rn and the carré du champ operator (f, g) = ∇f · ∇g for
smooth functions f, g on Rn . Further basic examples will be discussed in this and
the next chapter.
1.6.3 Heat Equations
In the last part of this section, we collect a few observations on the Chapman-
Kolmogorov and Fokker-Planck heat equations in the symmetric context.
We first briefly revisit the Chapman-Kolmogorov and Fokker-Planck equations
for a symmetric Markov semigroup (Pt )t≥0 , and draw some useful conclusions.
Thus let μ be reversible for L. The operators Pt , t ≥ 0, are then symmetric in
L2 (μ) and often admit kernel densities with respect to μ, denoted pt (x, y), t > 0,
(x, y) ∈ E × E, which are symmetric in (x, y).
As discussed in Appendix A, Sect. A.2, p. 475, while symmetry and self-
adjointness might not coincide for unbounded operators, the generator L of a sym-
metric semigroup is self-adjoint, that is L = L∗ where the adjoint operator L∗ is
computed with respect to the invariant measure μ. Now, from the heat and Fokker-
Planck equations (1.5.1) and (1.5.2),
∂t pt (x, y) = Lx pt (x, y) = Ly pt (x, y), t > 0.
In particular
Lx pt (x, y) = Ly pt (x, y).
Recall that the notations Lx and Ly are to emphasize the action on the respective
variables x and y of the kernel pt (x, y). This is however not a simple consequence
of the symmetry property. For example, on R2 , the function h(x, y) = x 4 + y 4 is
symmetric in (x, y), but ∂x2 h = ∂y2 h. On the other hand, on Rn , all smooth functions
F of |x − y|2 satisfy x F = y F . This will be the case for the heat kernels of the
standard Laplacian.
The second illustration is related to the trace of the semigroup. Namely, the
Chapman-Kolmogorov equation (1.3.2) indicates that the scalar product in L2 (μ)
1.7 Dirichlet Forms and Spectral Decompositions 29
of the functions z → ps (x, z) and z → pt (z, y) is pt+s (x, y). In particular, by the
Cauchy-Schwarz inequality, for all t, s > 0 and (x, y) ∈ E × E,
pt+s (x, y)2 ≤ p2t (x, x) p2s (y, y).
Note that this inequality, when applied for t = s (and after the change t → 2t ),
indicates that, for every t > 0, the function pt (x, y) of (x, y) achieves its maximum
on the diagonal x = y. Moreover, in the symmetric case, and provided the following
expression makes sense, one has

p2t (x, x)dμ(x) = pt (x, y)2 dμ(x)dμ(y). (1.6.4)
E E E
When the latter is finite, the operator Pt is Hilbert-Schmidt (see Sect. A.6, p. 483,
for more information about Hilbert-Schmidt operators). In terms of the associated
Markov process X = {Xtx ; t ≥ 0, x ∈ E}, pt (x, x) is the probability that Xt returns
at time t > 0 to its initial point x (as for the Brownian bridge for example). Of
course, this naive representation is formal since in general this probability vanishes.
This basic formula however links the distribution of the so-called bridge process to
the trace of the semigroup.
1.7 Dirichlet Forms and Spectral Decompositions

In the context of symmetric Markov semigroups, the carré du champ operator imme-
diately gives rise to the notion of a Dirichlet form. Dirichlet forms play an important
role in the construction of semigroups (see Chap. 3), and the functional inequalities
studied throughout this work are also closely related to this object. Dirichlet forms
may be considered in non-symmetric frameworks but here we restrict the exposition
to the symmetric case.
1.7.1 Dirichlet Forms
For a symmetric Markov semigroup P = (Pt )t≥0 with infinitesimal generator L,

reversible measure μ and carré du champ operator on a class A of functions on
E, consider the (symmetric) bilinear operator

E(f, g) = (f, g)dμ, (f, g) ∈ A × A.
E
By the integration by parts formula (1.6.3),

E(f, g) = (f, g)dμ = − f Lg dμ (1.7.1)
E E
for any f, g ∈ A.
The bilinear form E may actually be defined on a class of functions larger than
A. To this end, observe first that from (1.7.1), E(f ) = E(f, f ) may be defined for
any function f ∈ D(L). Also (and this type of argument will be used constantly
throughout this book), for any f ∈ L2 (μ) and any t > 0,

∂t (Pt f )2 dμ = 2 Pt f LPt f dμ = −2 E(Pt f )
E E
and

∂t E(Pt f ) = −∂t Pt f LPt f dμ
E

=− (LPt ) dμ −
2
Pt f L2 Pt f dμ (1.7.2)
E E

= −2 (LPt f )2 dμ.
E
In particular, E(Pt f ) is a decreasing function of t and

t
f 2 dμ − (Pt f )2 dμ = 2 E(Ps f )ds ≥ 2t E(Pt f ).
E E 0
t
Changing t into 2, this inequality shows that the quantity

1
f 2 dμ − (Pt/2 f )2 dμ
t E E
is decreasing

in t > 0. But the latter may be rewritten by the reversibility property
as 1t E f (f − Pt f )dμ, and therefore, for f ∈ D(L), it converges to E(f ) when t
tends to 0.
This observation allows us to extend the definition of E(f ) to a wider class of
functions, defining D(E) as the set of functions f ∈ L2 (μ) for which the quantity

1 1
f 2 dμ − (Pt/2 f )2 dμ = f (f − Pt f )dμ
t E E t E
has a finite (decreasing) limit as t decreases to 0, and E(f ) to be this limit for
f ∈ D(E).
Definition 1.7.1 (Dirichlet form) For a symmetric Markov semigroup P = (Pt )t≥0
with reversible measure μ, the energy E(f ) is defined as the limit as t → 0 of

1
f (f − Pt f ) dμ
t E
for all functions f ∈ L2 (μ) for which this limit exists, defining in this way the
domain D(E). The Dirichlet form E(f, g) is defined by polarization for f and g in
the Dirichlet domain D(E), and E(f ) = E(f, f ).
1.7 Dirichlet Forms and Spectral Decompositions 31
If f, g ∈ D(L),

E(f, g) = f (−Lg)dμ = g(−Lf )dμ,
E E
while if f, g belong to the class A on which is defined,

E(f, g) = (f, g)dμ.
E
Hence D(L) ⊂ D(E) ⊂ L2 (μ). These objects will be further investigated in Chap. 3
where in particular D(E) will be constructed first and then D(L) will be extracted
from it.
1.7.2 Spectral Decompositions
For symmetric semigroups, the spectral decomposition of the generator L is a

very efficient tool for analyzing properties of functions and operators. We refer to
Sect. A.4, p. 478, for abstract definitions and to Proposition 3.1.6, p. 131, for con-
crete applications in the setting of Markov semigroups. In particular, for any func-
tion f ∈ L2 (μ) and any t > 0, the function Pt f always belongs to the domain D(L),
and also to D(Lk ) for any integer k ∈ N (by induction, f ∈ D(Lk ) if f ∈ D(Lk−1 )
and Lk−1 f ∈ D(L)—see Remark 3.1.7, p. 132). In particular, this would not be the
case for a non-symmetric semigroup.
We next briefly comment on the spectral decomposition of a Markov operator
L in the simple instance when there is a countable basis of eigenvectors for the
operator L. Assume therefore that (ek )k∈N is a Hilbertian basis of L2 (μ) consisting
of eigenfunctions of L with corresponding sequence of eigenvalues (λk )k∈N . Hence
−Lek = λk ek , k ∈ N.
In this case, by the integration by parts formula (1.6.3), for every integer k,

E(ek ) = (ek , ek )dμ = − ek Lek dμ = λk ek2 dμ = λk .
E E E
As a fundamental consequence, the eigenvalues of a Markov generator are always

negative (justifying the choice of sign in the definition −Lek = λk ek ). This expresses
in another way that −L is positive. The operator −L is sometimes called the oppo-
site operator (to L).
(formally) Pt = e , t ≥ 0, for a function f ∈ L (μ) decomposed as
Since tL 2
f = k∈N fk ek in the basis (ek )k∈N ,

Pt f = e−λk t fk ek .
k∈N
Then setting, for t > 0 and (x, y) ∈ E × E,

pt (x, y) = e−λk t ek (x) ek (y), (1.7.3)
k∈N
shows that

Pt f (x) = f (y) pt (x, y)dμ(y)
E
in accordance with the kernel representation (1.2.4). That the density kernels
pt (x, y) thus defined are positive is however not easy to see directly, even in simple
examples. From (1.6.4), the trace formula

pt (x, x)dμ(x) = e−tλk , t > 0, (1.7.4)
E k∈N
follows. Hence, information on the functions pt (x, x) yields estimates on the spec-
trum of L. Of course, there are numerous examples where the operator admits den-
sities without being Hilbert-Schmidt (the standard Euclidean heat semigroup being
the simplest example).
Contrary to the finite-dimensional case, in infinite dimension, symmetric or self-
adjoint operators L do not always admit an orthonormal basis of eigenvectors. This
is replaced in the general setting by a spectral decomposition. As described in
Sect. A.4, p. 478, the spectrum, which is the set of λ ∈ R such that L − λId is
not invertible, is actually divided into two parts, the discrete spectrum, which corre-
sponds to isolated eigenvalues associated with a finite-dimensional eigenspace, and
the essential spectrum. For example, the Laplace operator on Rn , Markov gener-
ator of the heat or Brownian semigroup, has an empty discrete spectrum. In ideal
settings, as described later in Chap. 4, the essential spectrum is entirely determined
by the behavior of L at infinity. This is in particular the case for Persson operators,
which are developed in Sect. 4.10.3, p. 227.
1.8 Ergodicity
In probability theory, ergodic properties usually relate to the long time behavior. In
the context of Markov processes {Xtx ; t ≥ 0, x ∈ E}, and when the invariant mea-
sure μ is a probability measure, it is in general expected that quantities such as

1 t x
f Xs ds
t 0

converge almost surely as t → ∞ to E f dμ, whatever the starting point x and for
suitable functions f . Although such results may in general be deduced from the
analysis developed in this monograph (see for example Sect. 1.15.9), we adopt here
a less ambitious approach. Ergodicity in our context will relate to the convergence
1.9 Markov Chains 33
of the law of Xtx to the invariant probability measure μ

as t → ∞. In other words,
for suitable functions f : E → R, Pt f (x) converges to E f dμ for any x ∈ E. (See
the precise Definition 3.1.11, p. 135.)
From both the point of view of probability and analysis, these ergodic properties
may be seen as “convergence to equilibrium”, the equilibrium being the invariant
measure (whenever finite). The analysis of convergence to equilibrium plays an im-
portant role in many fields, and many of the methods and functional inequalities
developed in this monograph will lead to (precise) quantitative bounds on the con-
vergence to equilibrium.
Determining the class of functions f such that Pt f (x) converges, and the sense
in which it converges, strongly depends on the context. According to the spec-
tral decomposition of self-adjoint operators as presented in Sect. A.4, p. 478,
(see also Sect. 1.7.2), the easiest way to understand ergodicity is to look at
L2 (μ)-convergence. In the context of symmetric Markov semigroups for which the
infinitesimal generators L are self-adjoint, the associated semigroups P = (Pt )t≥0
are actually L2 (μ)-ergodic. Indeed, the spectrum of L lies in (−∞, 0] and for
any f ∈ L2 (μ), Pt f converges to P∞ f in L2 (μ) as t → ∞ where P∞ is the
orthogonal projection onto the space associated with the eigenvalue 0, called the
invariant space. This eigenspace consists
of the solutions of the equation

Lf = 0
(f ∈ D(L) ⊂ L2 (μ)). In particular, since E (−Lf )f dμ = E(f ) = E (f )dμ, and
since (f ) ≥ 0, the invariant functions satisfy (f ) = 0 (μ-almost everywhere).
Now, in many cases, these invariant functions are constant. For example, in the
diffusion case (Sect. 1.11), (f ) stands for the square of the length of ∇f . A sim-
ilar description is available in the discrete Markov chain setting (Sect. 1.9). This
property is related to connexity of the state space as developed below in Sect. 3.3,
Definition 3.3.7, p. 151, the precise relationship between ergodicity and connexity
being described in Proposition 3.3.10, p. 157.
Once we know that invariant functions are constant, there are only two possi-
bilities:
either μ(E) < ∞, and assuming that μ is a probability measure, P∞ f =

E f dμ, or μ(E) = ∞, and since 0 is the only constant in L2 (μ), P∞ f = 0. These
two situations are represented by the Ornstein-Uhlenbeck semigroup and the stan-
dard heat semigroup in Rn (see Sect. 2.7.1, p. 103, and Sect. 2.1, p. 78).
Convergence to equilibrium therefore reflects the fact that Pt f converges in
L2 (μ) to P∞ (f ), which is the projection on the space of invariant functions. A sim-
ilar phenomenon can be found in other settings. For example, in differential geome-
try, one may also construct semigroups on p-forms (which are no longer functions,
or functions of a certain type on a larger space) for which the spectrum is still neg-
ative, and for which the invariant space may carry some non-constant or non-zero
terms, the so-called space of harmonic p-forms. The non-existence of non-vanishing
invariant forms in this context yields important topological information.
1.9 Markov Chains

Although this monograph is concerned with a continuous (diffusion) setting as pre-
sented in the next section (that is, diffusion semigroups and Markov processes with
continuous paths), it is of interest to briefly test and motivate the preceding abstract
definitions and properties of Markov semigroups and their infinitesimal generators
on the concrete example of a Markov chain on a finite or countable state space. Fur-
thermore, many of the difficulties that will be encountered in general may already be
described in the context of Markov chains on a countable state space. Indeed, in this
simple setting, generators appear as matrices L for which the associated semigroup
is nothing else than et L , that is the usual exponential of matrices. In particular, it is
natural in this context to start from the Markov generator and define from it the as-
sociated semigroup. Most of the properties of general Markov semigroups and gen-
erators have an analogue in this context, and this setting therefore provides a rich
source of ideas and basic counterexamples. Moreover, both the finite and count-
able models also have great importance when one considers numerical estimates.
For example, when solving a partial differential equation such as the heat equation,
discretization of space is often considered. When discretizing differential operators
like those described in Sect. 1.11 below, the resulting operators are generators of
Markov processes on finite or countable spaces.
In the following, we (briefly) review for the example of Markov chains on a finite
or countable state space some of the objects and properties considered in the abstract
framework of the previous sections. We refer the reader to general references on
Markov chains for a more complete picture.
1.9.1 Finite Markov Chains
We begin with one of the simplest cases, a finite state space. Thus given a finite set E
(equipped with the σ -field F of all its subsets), consider, as domain of any Markov
generator L, the vector space of all real-valued functions on E (with dimension the
cardinality of E, say N ). On the basis of the functions fx = 1x , x ∈ E, a Markov
generator L can be represented as a matrix (L(x, y))(x,y)∈E×E .
In this context, the (Markov) property L(1) = 0 expresses that, for every x ∈ E,
y∈E L(x, y) = 0. The carré du champ operator is given by
2
(f )(x) = L(x, y) f (x) − f (y) , x ∈ E.
y∈E
Hence, the carré du champ operator is positive if and only if all off-diagonal terms
are positive. A square matrix L = L(x, y), (x, y) ∈ E × E, with the
of the matrix L
two properties y∈E L(x, y) = 0 for all x ∈ E, and L(x, y) ≥ 0 for all x = y, is
called a Markov generator on the finite state space E.
The easiest way to produce such a Markov generator is to consider a stochas-
tic (or Markov) matrix P and to set L = λ(P − Id) for some λ > 0. The matrix
P = (P (x, y))(x,y)∈E×E actually represents the transition probabilities of a Markov
chain jumping at integer times from x to y with probability P (x, y). The continuous
time Markov process associated with P is obtained from the chain by replacing the
times between the jumps by independent exponential variables with parameter λ.
It is an elementary exercise to check that, for a square matrix L, the matrix et L ,

t ≥ 0, has all its entries strictly positive if and only if the off-diagonal entries of L
are positive. This exponential then describes the semigroup Pt = et L , t ≥ 0, with
infinitesimal generator L.
In this finite state space context, for a convex function φ : R → R, Jensen’s in-
equality Lφ(f ) ≥ φ (f ) Lf (1.4.4) expresses that, at any given point x ∈ E,

L(x, y) φ f (y) − φ f (x) ≥ L(x, y) φ f (x) (f (y) − f (x)
y∈E y∈E
which is a clear consequence of the convexity of φ.

A (finite) measure μ is invariant with respect to L if and only if, for all y ∈ E,

μ(x) L(x, y) = 0.
y∈E
Such a measure always exists (it is an eigenvector of the transpose of L, with

eigenvalue 0 and positive coordinates). It is unique as soon as the matrix L is ir-
reducible, that is, for any pair of points (x, y) in E × E, there exists a path in E,
(x = x0 , x1 , . . . , x = y), such that for every i = 0, . . . , − 1, L(xi , xi+1 ) > 0. For
the measure μ to be reversible, the condition is that
μ(x) L(x, y) = μ(y) L(y, x) (1.9.1)
for all (x, y) ∈ E × E (also known as the detailed balance condition). Summing over
x this identity shows that μ is invariant, so that in this particular case, reversibility
implies invariance. Observe that when the matrix L is symmetric in the ordinary
sense, then the uniform measure on E is reversible for L. Moreover, when L is
irreducible, μ(x) > 0 for every x ∈ E, and the matrix
1
R(x, y) = μ(x) L(x, y) √ , (x, y) ∈ E × E,
μ(y)
is symmetric in the usual sense. It can then be diagonalized with respect to an
orthonormal basis, which means that there is a basis of eigenvectors of L which
is orthonormal in L2 (μ). The eigenvalues are therefore negative, and if ek , λk ,
k = 0, . . . , N , denote respectively the eigenvectors and eigenvalues of −L, then,
for every k,
et L (ek ) = e−λk t ek .

If the vectors ek are normalized in L2 (μ), that is, if x∈E ek (x)μ(x) = 1,
2 then

N
e−λk t ek (x) ek (y)
k=0
is positive for every pair (x, y) in E × E and every t ≥ 0 (this is not obvious).
The invariant measure μ is thus viewed as a positive eigenvector of the trans-

posed matrix of L. The fact that the transposed matrix of L admits a positive eigen-
vector is not specific to those Markov generators. It is a general property of matrices
with positive entries (see the above transformation L yielding a trivial change on
the spectrum without changing the eigenvectors). Such matrices always admit an
eigenvector with positive entries. It corresponds to the eigenvalue with maximal
modulus, and this eigenvalue is real and positive. Moreover, provided the matrix is
irreducible in the sense described above, this eigenvector is unique up to scaling with
strictly positive coordinates. These are known as the Perron-Frobenius eigenvector
and eigenvalue.
The matrices with positive entries (outside the diagonal) may serve as a toy
model for generators of positivity preserving semigroups (without the Markov prop-
erty Pt (1) = 1). The analogue of the Perron-Frobenius eigenvector is then a posi-
tive eigenvector (known as the ground state in the physics literature), and may be
used for example to remove a potential term from a generator (see for example
Sects. 1.15.8 and 1.15.6 below and Proposition 4.5.3, p. 198).
1.9.2 Countable Markov Chains
Beyond the simplest case of finite spaces, the next case to consider is when the space
is countable. It can thus be assumed that E = N but we tend to prefer the abstract
notation E below, which is particularly justified when one has to take into account
the geometric aspects of the state space E. For example, random walks on Z and
Zn do not have the same behavior concerning recurrence and transience (that is,
whether or not the corresponding Markov process tends to infinity with time).
In the context of a countable state space E, a Markov semigroup P = (Pt )t≥0
is described by an “infinite matrix” of positive kernels (pt (x, y))(x,y)∈E×E , t ≥ 0,
such that for all t ≥ 0 and x ∈ E, and any positive function f on E,

Pt f (x) = f (y) pt (x, y).
y∈E
As in the finite case, for any x ∈

E, pt (x, y) is a probability measure on E, that
is, pt (x, y) ≥ 0 for all y ∈ E and y∈E pt (x, y) = 1. Then the generator L is also
given by an infinite matrix (L(x, y))(x,y)∈E×E as
L(x, y) = ∂t pt (x, y)|t=0 , (x, y) ∈ E × E,
and, for any, say finitely supported, function f on E,

Lf (x) = L(x, y)f (y).
y∈E
The heat equation reads

∂t pt (x, y) = L(x, y) pt (x, y), p0 (x, y) = δx (y).
y∈E
Solving such an equation requires some properties of the matrix L. Among them,
note that whenever x = y, L(x, y) ≥ 0, and, for every x ∈ E, y∈E L(x, y) = 0.
As in the case of a finite space, the carré du champ operator is given by
2
(f )(x) = L(x, y) f (x) − f (y) , x ∈ E
y∈E
(on, again, say finitely supported functions), and an invariant measure μ for L is
characterized by

μ(y) = μ(x) L(x, y)
x∈E
for every y ∈ E. Again, finding an invariant measure μ requires the solution of an
infinite-dimensional eigenvector problem. The measure μ is reversible if
μ(x) L(x, y) = μ(y) L(y, x)
for all (x, y) ∈ E × E (detailed balance property). Hence, in order for μ to be re-
versible the ratios L(x,y) R(x)
L(y,x) must be expressible in the form R(y) for some function R.
A reversible measure may then be easily defined directly from the generator L, fix-
ing the value at one point x0 ∈ E, say μ(x0 ) = 1, and setting μ(x) = L(x 0 ,x)
L(x,x0 ) , x ∈ E.
If this measure is finite, it can be normalized into a probability measure. An invariant
measure will exist and is unique in the recurrent case. When the underlying Markov
chain is transient, there may exist many invariant measures.
It is also of interest to describe the paths of the Markov process (or family of
Markov processes) X = {Xtx ; t ≥ 0, x ∈ E} with generator L in these finite or count-
able contexts. Working directly in the countable state space setting, we introduce the
Markov matrix
L(x, y)
K(x, y) = − , x = y, K(x, x) = 0.
L(x, x)
We therefore exclude the possibility that L(x, x) = 0 for any x (such a point would
be a trap point, meaning that when the process arrives at this point, it stays there
forever). Look first at the discrete time Markov chain (X k )k∈N with matrix K as
Markov kernel, that is the sequence of random variables with X 0 = x such that, for
all k ∈ N and (x, y) ∈ E × E,

P X k = x = K(x, y).
k+1 = y | X
This process is known as the underlying Markov chain of the process X. Then the
associated Markov process X may described in the following terms. For each k ∈ N,
if Tk is the k-th jump time of (Xt )t≥0 , the sequence (XTk )k∈N is a Markov chain
which follows the same law as (X k )k∈N . The interval Tk+1 − Tk between the jump
follows an exponential law with parameter −L(XTk , XTk ), independently of what
happened before Tk . In other words, when the process arrives at some point x ∈ E,
it waits for an exponential time with parameter −L(x, x), and then jumps to y with
probability K(x, y).
This rough description hides some serious difficulties. It may happen that the
underlying Markov chain (X k )k∈N is transient (that is, goes to infinity when k → ∞)
and that the jumps are so quick (when L(x, x) goes to infinity) that the process so
described goes to infinity in finite time. The actual construction of the process given
the data contained in the matrix L requires further restrictions. We refer the reader
to the standard literature for a complete account of the topic.
These discrete models may often be thought of as approximations for diffusion
processes. As a basic example, consider the Markov process on Z/N ⊂ R, which
jumps from x to x ± N1 at exponential times with parameter N 2 . Its generator may
be described as

1 1
LN f (x) = N 2 f x + +f x − − 2f (x) ,
N N
which clearly converges to Lf (x) = f (x) when N goes to infinity (at least along
smooth functions). In a naive way, this reflects Donsker’s theorem which describes
convergence of a suitable renormalization of the random walk on Z to Brownian
motion.
1.10 Stochastic Differential Equations and Diffusion Processes
Diffusion semigroups are the main objects of investigation in this work, and will be
introduced in Sect. 1.11. They occur quite naturally in the probabilistic context of
stochastic differential equations and we present here some of the intuition behind
this description. The reader is referred to Appendix B for the necessary stochas-
tic calculus background and more precisely to Sect. B.4, p. 495, for the necessary
material on diffusion processes.
1.10.1 Diffusion Processes
In this framework, we begin with a stochastic differential equation, in R or Rn ,

dXtx = σ Xtx dBt + b Xtx dt, X0x = x, (1.10.1)
where (Bt )t≥0 is a standard Brownian motion (on R or Rn , respectively) and σ

and b are smooth functions with bounded derivatives. More precisely, say on Rn ,
1.10 Stochastic Differential Equations and Diffusion Processes 39
σ is an n × n matrix and b a vector. A process X = {Xtx ; t ≥ 0, x ∈ Rn } which is a

solution of (1.10.1) is usually called a diffusion process (more accurately, since the
initial point x is a variable, it is a family of diffusion processes). For a given smooth
function f : Rn → R, at least C 2 , Itô’s formula (Theorem B.2.1, p. 491) indicates
that
n
j
df Xtx = σi Xtx ∂j f Xtx dBti + Lf Xtx dt
i,j =1
f
= dMt + Lf Xtx dt, X0x = x,
f
where (Mt )t≥0 is a local martingale (given as a stochastic integral) such that
f
M0 = 0 and
1 i j 2
n n n
Lf = σk σk ∂ij f + bi ∂i f. (1.10.2)
2
i,j =1 k=1 i=1
f
In other words, (Mt )t≥0 solves the martingale problem (1.4.6) for L. Actually
f
(Mt )t≥0 is a true martingale as soon as f and Lf are bounded functions, and in
this case, taking expectations, for all t ≥ 0 and x ∈ Rn ,
t
x
E f Xt = f (x) + E Lf Xsx ds. (1.10.3)
0
In terms of the associated Markov semigroup

Pt f (x) = E f Xtx = E f Xtx | X0 = x , t ≥ 0, x ∈ Rn ,
so that (1.10.3) leads to the fundamental formula (1.4.9)

t
Pt f (x) = f (x) + Ps Lf (x)ds, t ≥ 0, x ∈ Rn .
0
We therefore recover here in a concrete way the characterization of the law of

the process X in terms of a martingale problem (1.4.6) discussed in Sect. 1.4.3.
The second description (1.4.8) of the martingale problem may similarly be recast
in terms of stochastic calculus. Namely, observe, again from Itô’s formula, that if
F (t, x) is a (smooth) function of the two variables t ≥ 0 and x ∈ Rn , then

dF t, Xtx = dMtF + (∂t + L)F t, Xtx dt, X0x = x
(where L is acting on the space variable x). In other words, if ∂t F = LF (and

F (0, x) = f (x)), then F (T − t, Xtx ) is a martingale on the time interval t ∈ [0, T ]
which is equal to F (T , x) at t = 0 and to f (XTx ) at time t = T . Taking expectations,

F (x, T ) = E F 0, XTx = E f XTx .
Hence if F (t, x) is a solution of the partial differential (heat) equation ∂t F = LF

with F (0, x) = f (x), then F (t, x) = Pt f (x), t ≥ 0, x ∈ Rn . Note that here the
operator L (that is, the coefficients of the original stochastic differential equa-
tion (1.10.1)) does not depend on t. The case of an equation with time dependence
would actually lead to similar consequences, however it is somewhat more difficult
to express.
Therefore understanding the distributions of Xtx , t ≥ 0, x ∈ Rn , reduces to the
study of the solutions of heat equations of the type ∂t F = LF . To this end, let us
have a closer look at the second order differential operator L of (1.10.2). It may be
written as
1 ij
n n
Lf = g (x)∂ij2 f + bi (x)∂i f (1.10.4)
2
i,j =1 i=1
where g = (g ij )1≤i,j ≤n = (g ij (x))1≤i,j ≤n denotes the symmetric matrix σ σ ∗ and

b = (bi )1≤i≤n = (bi (x))1≤i≤n . For a matrix g to be expressible in this form, it is
necessary that, at every point x, it is symmetric and positive, that is

n
g(V , V ) = g ij (x) Vi Vj ≥ 0 for all V = (Vi )1≤i≤n ∈ Rn .
i,j =1
We refer to the latter property of the operator L of (1.10.4) as semi-ellipticity.

Ellipticity on the other hand usually refers to semi-ellipticity together with non-
degeneracy of g, that is, it must be strictly positive or positive-definite. A further
discussion of these notions appears in Sect. 1.12 below. Every symmetric positive
matrix g can be diagonalized with respect to an orthonormal basis, and admits a
symmetric square root such that σ σ ∗ = σ 2 = g. The fact that this can be achieved
continuously on x requires some care, but in general σ may be chosen with at least
the same regularity as g as long as g is invertible (elliptic case). Even when g is de-
generate, one may choose a square root σ to be at least locally Lipschitz, provided
g is at least C 2 . We will generally not be concerned with such regularity issues in
a general setting since the operators L will usually be given explicitly in the form
g = σ σ ∗.
Once L is given, that is the matrix g = g(x) and the vector b = b(x), one can
construct the associated Markov process X = {Xtx ; t ≥ 0, x ∈ Rn } described above
by solving a stochastic differential equation, and then find the solution to the heat
equation ∂t F = LF with the help of the distribution of Xtx . There is therefore a
close connection between solutions of such heat equations and laws of diffusion
processes.
Note finally that the operator L does not fully describe the process
X = {Xtx ; t ≥ 0, x ∈ Rn } since there is always a choice of the square root of the
matrix g. However, it is possible to completely describe from L the distribution of
Xtx for every t as a martingale problem (1.4.6), the formulation of which only relies
on L (and not on the choice of σ ).
1.10 Stochastic Differential Equations and Diffusion Processes 41
1.10.2 Hörmander Form
Very often, the operator L is given in the so-called Hörmander form. Recall (cf. Ap-
pendix C, Sect. C.1, p. 500) that a vector field on Rn or on an open set of Rn is a
first order operator f → Zf which may be written as

n
Zf (x) = Z i (x) ∂i f (x).
i=1
The coefficients Z i (x) are the coordinates of Z in the base (∂i )1≤i≤n . If Z is a
vector field, then Z 2 = Z ◦ Z is a second order differential operator. Then, given a
collection (Zj )0≤j ≤d of vector fields (d does not need to be the dimension of the
space, it may be smaller or larger), the operator
1 2
d
L= Zj + Z0 (1.10.5)
2
j =1
is a second order differential operator, semi-elliptic with no 0-order term. For ex-
ample, the ordinary Laplacian on Rn is given in this form, with d = n and Zj = ∂j
for 1 ≤ j ≤ n and Z0 = 0. At least locally, an operator L given by (1.10.4) may
always be represented in this Hörmander form, so that there is no loss of generality
in restricting to operators of this form. This Hörmander form is particularly useful
when analyzing the hypoelliptic properties of L (see Sect. 1.12 below), however it
is not unique.
For such an operator L in Hörmander form (1.10.5), to construct the associated
Markov process, say on Rn , X = {Xtx ; t ≥ 0, x ∈ Rn } with generator L, it is enough
j
to choose d independent Brownian motions (Bt )t≥0 in Rn , 1 ≤ j ≤ d, and to solve
the stochastic differential equation

d
j
dXtx = Zj Xtx ◦ dBt + Z0 Xtx dt, X0x = x. (1.10.6)
j =1
(j )
Here, the notation Zj (Xtx ) ◦ dBt means that the standard Itô integral has been
replaced by the Stratonovich integral (Sect. B.4, p. 495, in Appendix B). The previ-
ous notation has to be understood component-wise, that is, for any coordinate Xtx,i ,
1 ≤ i ≤ n, of Xtx ,

d
j
dXtx,i = Zji Xtx ◦ dBt + Z0i Xtx dt.
j =1
This change from Itô’s integral to Stratonovich’s integral may be viewed as a mere
change of notation, but it has the distinct advantage of having a simpler chain rule
formula than the Itô integral, and therefore leads to a direct representation of L in
the desired form. The link between this stochastic differential equation and the as-
sociated Markov semigroup as a martingale problem (1.4.6) can be seen, as before,
f f
by checking that df (Xtx ) = dMt + Lf (Xtx )dt where Mt is a martingale, and this
is again the result of a direct application of Itô’s formula for Stratonovich integrals.
1.11 Diffusion Semigroups and Operators

Markov semigroups associated with solutions of stochastic differential equations as
presented in the previous section are prototypical examples of the so-called diffu-
sion semigroups. (Symmetric) Markov diffusion semigroups and operators are the
central objects of analysis in this monograph. In this section, we do not deal with the
most general family of diffusion operators and only present some basic examples of
interest. In particular, we only loosely describe the classes of functions on which
the various operators are acting. It should be pointed out that, with respect to the
previous probabilistic interpretation, we work here, and throughout the book, with
the generator L rather than 12 L.
1.11.1 Diffusion Operators
To introduce the family of diffusion semigroups and operators (discussed in further

detail in Chap. 3), we begin with E = Rn or an open subset of Rn as underlying
state space. Working on such a state space, smooth functions (at least C 2 but often
C ∞ ) with compact support in E are in the domains of the diffusion operators L
under investigation. To determine whether smooth functions form a core (in other
words whether it is enough to know L on these functions in order to describe the
associated Markov semigroup P = (Pt )t≥0 ) actually involves the intrinsic metric
and the boundary of E as will be discussed below (see Sect. 3.2.2, p. 141, and also
Sect. A.5, p. 481, for a general discussion when the operator L is symmetric).
On smooth functions f , thus on E = Rn or some open subset of Rn , a Markov
diffusion operator L is a second order differential operator of the form

n
n
Lf = g ij
∂ij2 f + bi ∂i f (1.11.1)
i,j =1 i=1
where g = g(x) = (g ij (x))1≤i,j ≤n and b = b(x) = (bi (x))1≤i≤n are smooth, respec-
tively n × n symmetric matrix-valued and Rn -valued, functions of x ∈ E. The fact
that L does not involve constant terms expresses the fact that L(1) = 0. The carré du
champ operator (cf. Definition 1.4.2) takes the form, on smooth functions f, g,

n
(f, g) = g ij ∂i f ∂j g. (1.11.2)
i,j =1
1.11 Diffusion Semigroups and Operators 43
That is positive expresses the semi-ellipticity property of the generator L. This

also explains why we limit ourselves to second order differential operators since
a differential operator of higher order can never have a positive carré du champ
operator.
When the operator L is given in the Hörmander form (1.10.5) (without the factor
1
2 the carré du champ operator is given by
),

d
(f, g) = Zj f Zj g,
j =1
from which positivity is immediate. It is also plain from this expression that L is
elliptic as soon as, at any point x, the vector space spanned by the vectors Zj (x),
1 ≤ j ≤ d, is Rn . This requires in particular d ≥ n.
The continuous setting of the previous examples allows for a crucial diffusion,
or change of variables, property emphasized in the following basic definition. At
this point, it is more a property of the differential operators (1.11.1) but it will
make sense in a general framework (emphasized in Chap. 3) for Markov genera-
tors L acting on some class A of (smooth) functions. For such an operator, recall
the carré du champ operator of Definition 1.4.2, as well as the shorthand notation
(f ) = (f, f ).
Definition 1.11.1 (Diffusion operator) An operator L, with carré du champ opera-

tor , is said to be a diffusion operator (or generator) if
L ψ(f ) = ψ (f ) Lf + ψ (f ) (f ) (1.11.3)
for every ψ : R → R of class at least C 2 and every suitably smooth function f .
The operator L is also said to satisfy the diffusion property or change of variables
formula. In the examples of second order differential operators L considered above,
one may for example choose sufficiently smooth functions f in this definition. Pre-
cise assumptions and descriptions of relevant classes of functions will be addressed
later in Chap. 3.
There is a similar, equivalent, change of variables formula for functions of several
variables,
k
L (f1 , . . . , fk ) = ∂i (f1 , . . . , fk ) Lfi
i=1
(1.11.4)

k
+ ∂ij2 (f1 , . . . , fk ) (fi , fj )
i,j =1
for : Rk → R of class C 2 (actually C ∞ will mainly be used in the core of the

book), and f1 , . . . , fk smooth. Formally, the diffusion property expresses that L is a
second order differential operator, or equivalently that is, in each argument, a first
order differential operator. For example,

ψ(f ), g = ψ (f ) (f, g) (1.11.5)
for smooth f and g, and similarly for families of functions.

From a purely algebraic point of view, if the operator L and its carré du champ
operator are defined on an algebra A of functions included in the domain D(L),
the formula (1.11.4) for a polynomial function boils down to the chain rule for-
mula
(f g, h) = f (g, h) + g (f, h)
for any triple (f, g, h) of elements of A, which precisely translates the fact that,
in each of its arguments, is a first order differential operator. Extending this to
the wider class of smooth functions generally requires further assumptions (see
Remark 3.1.4).
It may be observed that (1.11.4) applied for a system of coordinates (f1 , . . . , fn )
yields precisely the expression (1.11.1) of L in this system with g ij = (fi , fj ) and
bi = Lfi , 1 ≤ i, j ≤ n, written in those coordinates.
For diffusion operators, given a (smooth) function h, the operator f → (h, f )
is a first order differential operator, that is a vector field. In a smooth context, it is
denoted ∇h. With this notation,

n
∇h(f ) = g ij ∂j h ∂i f.
i,j =1

Alternatively, ∇h(f ) = ni=1 Z i ∂i f where Z i = nj=1 g ij ∂j h, 1 ≤ i ≤ n.
The diffusion property of a given Markov generator L is closely related to
the continuity properties of the sample paths of the associated Markov process
X = {Xtx ; t ≥ 0, x ∈ Rn }. As discussed in Sect. 1.1, given any Markov semigroup,
the associated Markov process may be constructed so that for every “nice” func-
tion f (in the domain of L), the real-valued process (f (Xt ))t≥0 is right-continuous
with left-limits. What is hidden behind the diffusion property is that these processes
(f (Xt ))t≥0 are continuous in this case. The proof of this claim requires some work
in stochastic calculus (with respect to non-continuous processes). At least, the dif-
fusion property implies a kind of continuity in measure for functions of the domain
as described in Theorem 3.1.16, p. 136.
Observe that the diffusion operators L for which the second order terms are all
zero (that is, for which the carré du champ operator is identically zero) are far
from being uninteresting. They correspond to vector fields, that is, first order dif-
ferential operators. For such an operator, the associated semigroup P = (Pt )t≥0
is of the form Pt f (x) = f (Xtx ) where Xtx is the path following the vector field
starting at x. It is the analogue of the Markov processes described by solutions of
stochastic differential equations, but deterministic. If the vector field is of the form
Zf (x) = ni=1 Z i (x)∂i f (x), then Xtx is a point in Rn whose coordinates are the
solution of the differential system X0 = x, dXti = Z i (Xt )dt. The law of the process
at time t is a Dirac mass at Xtx and randomness has disappeared. Randomness (or
diffusion) arises precisely from the second order terms in the differential operator.
1.11.2 Invariant Measures of Differential Operators
Given a diffusion operator L as in (1.11.1), let us now describe its invariant measure
when it admits a (smooth, strictly positive) density w with respect to the Lebesgue
measure dx (on E = Rn or an open subset of Rn ). This density w is given by the
equation L∗ w = 0 where L∗ is the adjoint of L with respect to the Lebesgue measure
which may be explicitly described, on smooth functions f , by

n
ij
n

∗
L f= ∂ij2 g f − ∂i bi f . (1.11.6)
i,j =1 i=1
For a different reference measure, it is necessary to take the adjoint with respect
to this new measure (for example, when the diffusion matrix is not the identity,
there is a better choice than the Lebesgue measure as will be seen below). The
invariant measure will usually be given in the further developments, and its existence
or finiteness will be clear from the context.
In general, the adjoint operator L∗ of (1.11.6) is a second order differential oper-
ator, but including a non-zero constant term (usually called the potential in quantum
mechanics and in the study of Schrödinger equations). In the simplest instances
L = + Z where is the usual Laplacian on Rn and Z = (Z i )1≤i≤n is a vec-
tor field, with the Lebesgue measure as reference measure. Hence, by (1.11.6), on
smooth functions f , the adjoint operator L∗ with respect to the Lebesgue measure
is
L∗ f = f − Zf − div(Z)f (1.11.7)
n
where div(Z) = i=1 ∂i Z i . Indeed, given smooth

compactly

supported functions
f, g (on E = Rn or an open subset of Rn ), E f g dx = E g f dx whereas, by
integration by parts, for every 1 ≤ i ≤ n,

f Z i ∂i g dx = − ∂i Z i f g dx = − g Z i ∂i f dx − g ∂i Z i f dx
E E E E
thus justifying the above form of L∗ .

Note that it is often easier to describe the evolution of the density with respect to
the (an) invariant measure since then there is no potential term in the adjoint of L.
Indeed, if μ is invariant for L and if L∗ denotes the adjoint of L with respect to μ,
for every compactly

function f in the domain D(L) and every function
supported
∗ g dμ =
g
in D(L), E f L E Lf g dμ. Therefore, if g = 1, by invariance of μ,

f L∗ (1)dμ = 0 so that L∗ (1) = 0 and L∗ does not include constant terms.
E
1.11.3 Reversible Measures of Differential Operators
In the context of second order differential operators L of the form (1.11.1), invariant
measures μ are not easy to identify in general. For example, on an open subset
of Rn , the density w of μ with respect to the Lebesgue measure is a solution of
L∗ w = 0. But when the measure μ is reversible, things are a lot easier. For example,
on an open set in Rn , or on a manifold given any local system of coordinates, an
operator given in the form
1 ij
n

Lf = ∂i wg ∂j f , (1.11.8)
w
i,j =1
where w is smooth and strictly positive, is clearly (by integration by parts) symmet-
ric in L2 (μ) where dμ = wdx is the measure with density w with respect to the
Lebesgue measure. This is indeed the general case when the reversible measure has
a non-vanishing density w, and therefore operators of the form

n n
n
Lf = g ij ∂ij2 f + ∂i g + g ∂i (log w) ∂j f
ij ij
i,j =1 j =1 i=1
are exactly those operators with second order terms given by the matrix (g ij ) and
reversible measure dμ = wdx. Therefore, the knowledge of (g ij ) and (bi ) in the
representation (1.11.1) allows for a direct identification of the invariant measure
dμ = wdx when it is reversible. In this case, for f, g smooth and compactly sup-
ported,

n
f Lg dμ = − g ij ∂i f ∂i g dμ = − (f, g)dμ. (1.11.9)
Rn Rn i,j =1 Rn
When the matrix g = (g ij ) is everywhere non-degenerate (elliptic case), there

is a more canonical way of decomposing the operator L of (1.11.1). Namely, in-
stead of using the Lebesgue measure as a reference measure to compute den-
sities, it is preferable in general to employ the Riemannian measure μg with
density wg = det(g)−1/2 with respect to the Lebesgue measure. In this case,
one introduces the Laplace-Beltrami operator which has the same second or-
der terms as L and μg as reversible measure: in a local system of coordi-
nates,
1
n

g = ∂i wg g ij ∂j
wg
i,j =1
and then L = g + Z where Z is a vector field. This operator g , called

the Laplace-Beltrami operator, or Laplacian, associated with the (co-)metric
g = g(x) = (g ij (x))1≤i,j ≤n is presented in Appendix C, Sect. C.3, p. 504, in a
Riemannian framework. With respect to the Riemannian framework, it is therefore
the co-metric g which naturally enters into the investigation of Markov diffusion
operators. We thus privilege the co-metric g in the notation. As explained there, the
decomposition L = g + Z has the advantage of being preserved under a change of
coordinates.
Under this decomposition
L = g + Z,
to a measure dμ = wdμg as soon as Zf = (log w, f ).

L is symmetric with respect
In other words, if Zf = ni=1 Z i ∂i f ,

n
Z =
i
g ij ∂j (log w), 1 ≤ i ≤ n.
j =1
The operator f → (log w, f ) is often identified with ∇ log w · ∇f , and in this case
Z is called a gradient vector field. The invariant reversible measure is then typically
written as
dμ = e−W dμg
with w = e−W and μg the Riemannian measure. The coordinate representa-
tion (1.11.8) takes the form
L = g − (W, ·) = g − ∇W · ∇ (1.11.10)
(see Sect. C.5, p. 511, for further details). Both the coordinate representation of
(1.11.2) and the integration by parts formula (1.11.9) are similar in this context. The
latter decomposition of L (the canonical decomposition) is most useful when com-
puting iterated carré du champ operators in view of curvature-dimension conditions
(see Sect. 1.16 below) since it is mainly with the language of Riemannian geometry
that computations are made understandable.
When dealing with an operator L = g − ∇W · ∇ symmetric with respect to
dμ = e−W dμg on a Riemannian manifold (M, g) with Riemannian measure μg ,
we sometimes speak of M as a weighted Riemannian manifold. The basic example
of a weighted measure dμ = e−W dx on Rn is already of significant interest.
In summary, it is easier to describe the invariant measure in the reversible case,
since it is then given locally, while in the general case, it is obtained from a differen-
tial equation. For symmetric diffusions, the carré du champ operator determines
the second order terms of the generator L while the invariant measure μ determines
the first order terms. (For comparison, note that such a characterization is not avail-
able in the discrete case, since for example on finite state spaces, the operator
already describes the generator L.)
1.11.4 Alternative Geometric Representations of Generators
In the last part of this section, we briefly discuss variations and a few operations on
the preceding representations.
Adding a new vector field Z to an operator L = g − ∇W · ∇ as above, it may

happen that the measure dμ = e−W dμg is still an invariant measure although no
longer reversible, that is the new operator L=L+Z is no longer symmetric in
L (μ). For
2

this to happen, it is enough that for smooth compactly supported func-
dμ = 0, or in other words that the adjoint (Z)
tions f , E Zf in L2 (μ) satis-
∗ of Z

∗ (1) = 0. In a local system of coordinates for which Zf
fies (Z) = ni=1 Z i ∂i f , an
easy computation shows that
n n n
∗
(Z) f = −
Z ∂i f −
i
∂i Z +
i i
Z ∂i (log ρ) f
i=1 i=1 i=1
where ρ is the density of μ with respect to the Lebesgue measure dx (while

w = e−W is the density with respect to the Riemannian measure μg ). The quan-
tity

n
n
n
=
divμ (Z) i +
∂i Z i ∂i (log ρ) =
Z ∂i Z
i + Z(log ρ)
i=1 i=1 i=1
then plays the role of the divergence (and actually is the usual divergence when μ
is the Lebesgue measure itself, that is ρ = 1). Then, the adjoint
L∗ of
L = L + Ẑ in
L2 (μ) may be written
− divμ (Z),
L∗ = g − ∇W − Z
a description similar to (1.11.7) in the Euclidean case. Hence, the measure μ is

invariant precisely when divμ (Z) = 0, and then the adjoint L∗ of
L=L+Z in
L (μ) is L − Z.
2
There is still another most useful representation of L = g − ∇W · ∇. We again
begin with the Euclidean case E = Rn where g is the usual Euclidean metric and
dμ = e−W dx. For any smooth compactly supported function f on Rn , consider
the gradient operator df = (∂i f )1≤i≤n = ∇f , which is a vector of smooth com-
pactly supported functions. Its adjoint ∇ ∗ is defined on smooth compactly supported
vector-valued functions V in such a way that, for any smooth compactly supported
function f ,

∗
∇ V f dμ = V · ∇f dμ.
E E
If V = (Vi )1≤i≤n , then

n
n
∇ ∗V = − ∂i Vi + Vi ∂i W
i=1 i=1
so that
L = −∇ ∗ ∇. (1.11.11)
This representation immediately extends to the general Riemannian case, where
∇f is replaced by the 1-form df given in a local system of coordinates by
1.12 Ellipticity and Hypo-ellipticity 49
(∂i f )1≤i≤n . The scalar product of 1-forms w and w is then given in this system

of coordinates by w · w = ni,j =1 g ij wi wj , so that again L = −d ∗ d, which may be
seen as an extension of (1.11.11).
When the operator is elliptic, df may be replaced by its gradient ∇f which is
now a vector field, where in a local system of coordinates,

n
∇i f = g ij ∂j f, 1 ≤ i ≤ n.
j =1

The scalar product on vectors V and V is defined as V · V
= ni,j =1 gij V i V
j where
∗
(gij ) is the inverse matrix of (g ). The adjoint ∇ is then defined by its action on a
ij
vector field Z as

∇ ∗ Zf dμ = Z · ∇f dμ
E E
for any smooth compactly supported function f , and we still have L = −∇ ∗ ∇,
which also extends (1.11.11) to the Riemannian case. We refer to Appendix C for
further details.
We conclude this section with a brief comment on integration by parts in the
context of diffusion operators. Namely, observe that when the measure μ is only
invariant, and only for diffusion operators, the integration by parts formula (1.6.3)
still holds when g = ψ(f ). Indeed, setting ψ = , invariance yields

0= L(f )dμ = ψ(f ) Lf dμ + ψ (f ) (f )dμ,
E E E
and thus, with g = ψ (f ),

(f, g)dμ = ψ (f ) (f )dμ = − g Lf dμ.
E E E
In particular, this observation explains why, in dimension one (for the so-called
Sturm-Liouville operators, Sect. 2.6, p. 97), the analysis will be easier and invariant
measures will automatically be reversible.
1.12 Ellipticity and Hypo-ellipticity
It has already been mentioned in Sect. 1.10 that ellipticity or hypo-ellipticity play
an important role in the theory. In particular, under these conditions, solving a dif-
ferential (heat) equation of the form ∂t u = Lu with respect to a differential operator
L with initial condition u(0, x) = f (x) has regularizing properties: even when f is
only measurable, the solution u(t, x) is smooth for t > 0.
Hypo-ellipticity is a notion intermediate between semi-ellipticity and ellipticity.
Recall first the notions of semi-ellipticity and ellipticity for a diffusion operator L
of the form (1.11.1)

n
n
Lf = g ij ∂ij2 f + bi ∂i f
i,j =1 i=1
where g = g(x) = (g ij (x))1≤i,j ≤n and b = b(x) = (bi (x))1≤i≤n . The operator L is

said to be semi-elliptic if the matrix g is, at every point x, symmetric and positive,

n
g(V , V ) = g ij (x) Vi Vj ≥ 0 for all V = (Vi )1≤i≤n ∈ Rn . (1.12.1)
i,j =1
Ellipticity refers to semi-ellipticity together with non-degeneracy of g, that is strictly

positive or positive-definite, in the sense that
if g(V , V ) = 0, then V = 0. (1.12.2)
We sometimes speak of uniform ellipticity whenever there exists a c > 0 such that
g(V , V ) ≥ c |V |2 for all V ∈ Rn . (1.12.3)
Semi-elliptic operators in particular include first order differential operators (vec-

tor fields) for which the matrix (g ij ) is zero everywhere (the associated carré du
champ operator (f ) vanishing for every f ). Such a first order differential opera-
tor would never have regularizing properties as may be seen, for instance, from the
example L = ∂x on E = R for which the solution of ∂t u = Lu with initial condition
u(0, x) = f (x) is just u(t, x) = f (x + t).
However, when the operator L is symmetric with respect to some measure μ
(or more precisely self-adjoint), the solution u(t, x) = Pt f (x) is always smooth in
some weak L2 (μ) sense, that is in the domain of Lk for any integer k ≥ 1 (p. 31).
This property may be deduced from spectral analysis (see Remark 3.1.7, p. 132).
But this global result is in general not enough to provide precise information on
smoothness properties of the solution in the usual sense. Moreover, when solving
any heat equation on u(t, x) of the form
∂t u = Lu, u(0, x) = u0 , (1.12.4)
boundary conditions have to be imposed to ensure some kind of uniqueness (see, for
example, in dimension one, Sects. 2.4, p. 92, 2.5, p. 96, and 2.6, p. 97). The simplest
conditions (which however produce in general only sub-Markov semigroups) are
Dirichlet boundary conditions which amount to assuming that u(t, x) is 0 at the
boundary of the domain for any t > 0, or in probabilistic terms to consider the
associated Markov process killed at the boundary. Again, such restrictions imply
that it is impossible to solve the heat equation (1.12.4) locally.
To be more precise, if L is defined, say, on an open set O ⊂ Rn , and if we consider
a solution of the equation in O1 ⊂ O with the initial condition u0 say compactly
supported in O1 , the solution is not unique: any kind of boundary condition on any
1.12 Ellipticity and Hypo-ellipticity 51
intermediate set O1 ⊂ O2 ⊂ O gives rise to a different solution in O1 . The solution

of (1.12.4) in O1 actually strongly depends on the behavior of L outside O1 .
Moreover, smoothness of u0 and of the coefficients g ij and bi in O1 does not nec-
essarily ensure smoothness of the solution u in O1 . In the degenerate case (g ij ) = 0,
it is easy to construct, in dimension two for example, smooth coefficients bi on
O1 = (−1, +1)2 and a solution u(t, x) in O1 with u(0, x) smooth and compactly
supported in O1 such that u(t, x) is not even continuous for t = 1 (the price to pay
is that the bi ’s are not continuous far from O1 ). However, when solving a stochastic
differential equation in O, if the coefficients are smooth and if the resulting pro-
cess does not explode in O, then the resulting law of Xtx smoothly depends on
x ∈ O, and as a result Pt u0 is smooth for any compactly supported function u0 in O
(see Sect. B.4, p. 495). But again this is a global result, requiring knowledge of the
operator L everywhere where the underlying process lives.
Therefore, there is very little hope of obtaining in a general setting local prop-
erties of the solution of (1.12.4). This is the place where a more precise analysis,
namely ellipticity or hypo-ellipticity, has to come into play.
As an introduction to this more refined analysis, we consider here only operators
L of the type (1.11.1) with smooth coefficients (g ij ) and (bi ) on an open set O ⊂ Rn
(or an open set in a manifold). Then, as already mentioned earlier (see (1.12.2)), we
say that L is elliptic in O if the matrix (g ij (x)) is positive at any point x ∈ O. In this
situation, any solution on [0, s] × O of the equation ∂t u = Lu is C ∞ on (0, s) × O,
whatever the initial condition u0 , which may even only be measurable.
In the theory of partial differential equations, a linear operator P defined on
some open set O is said to be hypo-elliptic if any u such that P u ∈ C ∞ is itself C ∞ .
Classical regularity results ensures that if L is elliptic on O, then it is hypo-elliptic.
Furthermore P = L − ∂t is hypo-elliptic on (0, ∞) × O (parabolic hypo-ellipticity).
Hörmander’s theory actually allows us to go beyond the elliptic case and to give
a precise criterion for hypo-ellipticity in terms of the operator. Assume that L is
defined on O in Hörmander form (1.10.5) as

d
L= Zj2 + Z0
j =1
where the Zj ’s, 0 ≤ j ≤ d, are smooth vector fields (which may always be achieved
at least locally). Consider then, at any point x ∈ O, the vector space Vp generated
by the vector fields Zj and their commutators up to some fixed order p ≥ 1, that is
V1 = span{Zj , 0 ≤ j ≤ d},

V2 = span Zj , [Zj , Zk ], 0 ≤ j, k ≤ d ,
...

Vp = span Vp−1 ∪ [Zj , V ], V ∈ Vp−1 , 0 ≤ j ≤ d .
Then, one main conclusion of the Hörmander theory is that if there exists an integer
p such that, for any x ∈ O, Vp = Rn , the operator L is hypo-elliptic on O.
When going from L to the parabolic form P = L − ∂t , Z0 has to be replaced by

Z0 − ∂t . Reserving Z0 to span the time direction, the following statement holds.
Proposition 1.12.1 For L in Hörmander form on O ⊂ Rn as above, set

1 = span{Zj , 1 ≤ j ≤ d}, and, for any p > 1,
V

p = span V
V p−1 ∪ [Zj , V ], V ∈ V
p−1 , 0 ≤ j ≤ d .
p = Rn , the operator
Then, if there exists an integer p such that for any x ∈ O, V
L − ∂t is hypo-elliptic on (0, ∞) × O. In other words, under this condition, any
solution of ∂t u = Lu on [0, s] × O is smooth on (0, s) × O, whatever the behavior
of L outside O.
The main difference with the “non-parabolic Hörmander Theorem” is that Z0

is no longer used to span V1 . For example, on R2 , L = ∂x2 + ∂y satisfies Hörman-
der’s condition but does not fulfill the conditions of Proposition 1.12.1 whereas
L = ∂x2 + x ∂y does. In probabilistic terms, under the Hörmander condition of Propo-
sition 1.12.1, the laws of the underlying process Xtx , t ≥ 0, have smooth densities
for all x ∈ O. We will come back to these notions and issues in Chap. 3, in particular
in Sect. 3.2, p. 137.
1.13 Domains
This section briefly discusses a few observations and hypotheses about domains of
Markov semigroups and their infinitesimal generators as they arose in the preceding
sections. It is only aimed at giving a flavor of the necessary framework and will
be completed in Chap. 3, Sect. 3.3, p. 151, in which precise hypotheses on the
various classes of functions necessary to develop the Markov semigroup calculus
are carefully discussed.
Throughout this chapter, we are dealing, on some state space (E, F), with a
symmetric Markov semigroup P = (Pt )t≥0 with infinitesimal generator L with
L2 (μ)-domain D(L) where μ is the invariant and reversible measure. As discussed
earlier in Sect. 1.4, to define conversely the semigroup P from its generator L, it is
necessary to know L on a core of its domain D(L). Usually however, the generator
is only given on some set A0 of functions. Some simple criteria to ensure that a
dense vector subspace of functions is a core (in particular in the symmetric case of
most interest) are actually available. In Sect. 3.2, p. 137, we show for example that
on a complete Riemannian manifold (in particular, on open sets in Euclidean spaces
with a suitably chosen metric), C ∞ functions with compact support always form
a core for operators of the form L = g − ∇W · ∇ for some smooth potential W
on the manifold. Further examples will be discussed later. In general, when dealing
with second order operators L with smooth coefficients on an open set in Rn or on a
manifold as given by (1.11.1), every solution of the (heat) equation ∂t f = Lf will
be smooth (as is the case when solving stochastic differential equations, as already
1.14 Summary of Hypotheses (Markov Semigroup) 53
mentioned in Sect. 1.12). The solution is actually as smooth as the initial data. If the
operator is smooth and elliptic (or only hypo-elliptic), then even the regularity on
the initial data is not required, as discussed in the previous section. We could then
work on the class of C ∞ functions, but unfortunately it is not included in L2 (μ),
unless the manifold is compact. These observations lead us to consider an extended
algebra A ⊃ A0 on which the operator L is defined as an extension together with
the associated carré du champ operator , regardless of integrability properties. In
the regular instances, this class A will be the class of smooth C ∞ functions.
Given a vector space A0 of real-valued functions f on E, a variety of hypotheses
may be considered in order to conveniently work with P and L on A0 . A general
set of conditions is that A0 is contained in all Lp (μ)-spaces, 1 ≤ p < ∞, is dense
in the domain D(L) of L (that is, is a core of the domain), and is stable under
products and compositions with C ∞ functions vanishing at 0. In particular, it is an
algebra. In addition, it may be imposed that whenever f, g ∈ A0 , then f Pt g ∈ A0
for every t ≥ 0, or even that A0 is stable under the action of Pt . Observe that under
the latter assumption, A0 is automatically a core of the domain. The first hypothesis
is formulated to include the example of the set A0 of smooth compactly supported
functions on a complete manifold, while the second covers compact manifolds (but
also non-compact instances such as, for example, the space of Schwartz functions
for the Laplacian on Rn ). The algebra A on which the generator and its carré du
champ operator may naturally be considered, corresponding to the class of smooth
functions, may be constructed as an extension of A0 .
In some situations, it is not necessary to assume stability by composition with
smooth functions (as for the Ornstein-Uhlenbeck semigroup if one wishes to work
only with polynomials, cf. Sect. 2.7.1, p. 103). Actually, there are numerous inter-
mediate hypotheses in order to suitably cover one example or another, and it would
be rather difficult to cover all the cases in this way. We often choose to work with the
maximal hypothesis (of stability by the semigroup), leaving to the reader to adapt
proofs and arguments to more specific or complicated cases. Section 3.3, p. 151,
supplies the necessary arguments to fully justify this reduction, and describes in
more detail the different technical conditions required on a set of functions on which
L is given in order to ensure that the described extension can be achieved.
1.14 Summary of Hypotheses (Markov Semigroup)

At this stage of the exposition, summarizing some of the objects and ideas put for-
ward so far, we collect a first set of hypotheses on the datum consisting of a Markov
semigroup and its infinitesimal generator which will be in force throughout the
book. Recall that Chap. 3 develops the complete framework in which the results
of this monograph will be presented, emphasizing the carré du champ operator and
the invariant measure, and demonstrating consistency between the two viewpoints.
H1. The state space is a measurable space (E, F). All (positive) measures on
(E, F) are assumed to be σ -finite. Given a (σ -finite) measure μ on (E, F),
(E, F, μ) is assumed to be a good measure space in the sense that there is a count-
able family of sets which generates F (up to sets of μ-measure 0), and that both the
measure decomposition theorem and the bi-measure theorem apply (cf. Sect. 1.2).
To make things simpler, and since it is unlikely that the reader would ever need
more, it may actually always be assumed that F is the completion with respect to
the measure μ of the Borel σ -field for a topology on E which makes E a com-
plete separable metric space (a so-called Polish space). However, the topology will
not be used. Every Lp (μ)-space (1 ≤ p < ∞) is therefore separable. Functions,
always real-valued, are classes of functions for the μ-almost everywhere equal-
ity, and equalities and inequalities such as f ≤ g are always understood to hold
μ-almost everywhere. (See p. 7.)
H2. On a good measure space (E, F, μ), a Markov semigroup P = (Pt )t≥0 is a
family of operators with the following properties (see pp. 9 and 11):
(i) For every t ≥ 0, Pt is a linear operator sending bounded measurable functions
on (E, F) into bounded measurable functions.
(ii) P0 = Id, the identity operator (initial condition).
(iii) Pt (1) = 1, where 1 is the constant function equal to 1 (mass conservation).
(iv) If f ≥ 0, then Pt f ≥ 0 (positivity preserving).
(v) For every t, s ≥ 0, Pt+s = Pt ◦ Ps (semigroup property).
(vi) For every f ∈ L2 (μ), Pt f converges to f in L2 (μ) as t → 0 (continuity
property).
(vii) For every p, 1 ≤ p < ∞, the operators Pt , t ≥ 0, extend to every Lp (μ) as
bounded (contraction) operators.

for the semigroup P = (Pt )t≥0 in the sense that for
H3. The measure
μ is invariant
any f ∈ L1 (μ), E Pt f dμ = E f dμ (Definition 1.2.1).
(vii) is then automatic and the operators Pt , t ≥ 0, are contractions on every
Lp (μ), 1 ≤ p ≤ ∞
H4. The operators Pt , t ≥ 0, are represented by Markov kernels

Pt f (x) = f (y)pt (x, dy), x ∈ E,
E
where pt (x, dy), t ≥ 0, is a family of probability kernels on E (that is, for fixed
x ∈ E, a probability measure on (E, F) such that for any A ∈ F , x → pt (x, A) is
measurable) (see p. 12).
H5. The Markov process (or family of processes) X = {Xtx ; t ≥ 0, x ∈ E} on a
probability space ( , , P) associated with the Markov semigroup P = (Pt )t≥0
satisfies, on bounded or positive measurable functions f : E → R, the conditional
formula

Pt f (x) = E f Xtx = E f (Xt ) | X0 = x , t ≥ 0, x ∈ E.
The semigroup determines the finite-dimensional distributions of the processes

{Xtx ; t ≥ 0}, x ∈ E, by the Chapman-Kolmogorov equations (Sect. 1.3).
1.14 Summary of Hypotheses (Markov Semigroup) 55
H6. The Markov semigroup P = (Pt )t≥0 is symmetric with respect to the measure
μ, or μ is reversible for P = (Pt )t≥0 , in the sense that for any pair (f, g) of func-
tions in L2 (μ),

f Pt g dμ = gPt f dμ
E E
(Definition 1.6.1).
H7. The domain D(L) of the infinitesimal generator L of the semigroup P = (Pt )t≥0
is the (linear) space of functions f ∈ L2 (μ) such that the limit
1
lim (Pt f − f )
t→0t
exists in L2 (μ). This limit, denoted by Lf , defines the infinitesimal (Markov) gen-
erator (or operator) L. The domain D(L) is always a dense subspace of L2 (μ)
(cf. Definition 1.4.1). On this domain, the semigroup (Pt )t≥0 solves the heat equa-
tion (with respect to L),
∂t Pt = LPt = Pt L.
For a symmetric semigroup, the generator L is self-adjoint on its domain.
H8. The domain D(E) of the Dirichlet form E is the space of functions f in L2 (μ)
such that the limit

1
lim f (f − Pt f )dμ
t→0 t E
exists. This limit, denoted by E(f, f ) = E(f ), defines the Dirichlet form E of the
semigroup P = (Pt )t≥0 with domain D(E). When f ∈ D(L),

E(f ) = f (−Lf )dμ
E
and D(L) ⊂ D(E) ⊂ L2 (μ) (cf. Sect. 1.7).

H9. The carré du champ operator associated with the Markov generator L is
a bilinear operator defined on an algebra A of functions included in the domain
D(L) by
1
(f, g) = L(f g) − f Lg − g Lf , (f, g) ∈ A × A
2
(Definition 1.4.2). One always has (f ) = (f, f ) ≥ 0 and

(f )dμ = − f Lf dμ = E(f ).
E E
Various hypotheses may hold on the vector space A of functions on which this
carré du champ operator is defined (see Sect. 1.13).
H10. The generator L satisfies the diffusion property which states that for any fam-
ily (f1 , . . . , fk ) of functions in A on which the carré du champ operator is de-
fined, and any smooth function : Rk → R such that (f1 , . . . , fk ) belongs to
D(L),
k
k
L (f1 , . . . , fk ) = ∂i (f ) Lfi + ∂ij2 (f ) (fi , fj ).
i=1 i,j =1
Equivalently, the carré du champ operator is, in each argument, a first order dif-
ferential operator. The diffusion property translates the fact that the trajectories of
the associated Markov process X = {Xtx ; t ≥ 0, x ∈ E} are in some sense continu-
ous (cf. Sect. 1.11).
1.15 Working with Markov Semigroups

This section reviews a variety of basic operations which may be performed with
Markov semigroups and generators. These operations are mostly presented on some
specific, but important, illustrative examples which may then be adapted to more
general settings if necessary. Depending on the operation, the various sections em-
phasize more the analytic, probabilistic or geometric aspect.
According to the preceding summary, we have a symmetric Markov semigroup
P = (Pt )t≥0 on a state space (E, F), with invariant and reversible measure μ and
infinitesimal generator L with L2 (μ)-domain D(L). The semigroup and its genera-
tor yield the associated carré du champ operator and Dirichlet form E. As usual,
X = {Xtx ; t ≥ 0, x ∈ E} denotes the associated Markov process for the probabilistic
interpretation. If necessary, we use some, or all, of the hypotheses alluded to at the
end of Sect. 1.13 on some suitable subset A0 ⊂ D(L) of “nice” functions, although
these will usually be clear from the given context (here we retain our somewhat
informal level of discussion).
1.15.1 Change of Coordinates
Given a Markov semigroup (Pt )t≥0 , or its associated Markov process (Xt )t≥0 , on
a state space E, a bi-measurable bijection B between E and another state space E
yields a Markov process (B(Xt ))t≥0 on E , with Markov semigroup

Pt f = Pt (f ◦ B) ◦ B −1 , t ≥ 0,
not really different from the original one. The properties of semigroups which we
would like to investigate should actually be invariant under such transformations.
There are indeed intrinsic quantities in a given Markov generator which are inde-
pendent of the choice of such representations. Properties of semigroups are often
1.15 Working with Markov Semigroups 57
easier to detect in a given system of coordinates, so that it is important to perform

such changes of coordinates. For example, when working on Rn , or some open set
in Rn , one may consider diffeomorphisms B (that is, bijective differentiable maps
with differentiable inverses). Such a diffeomorphism is classically called a change
of variables. Consider its action on a diffusion generator

n
n
Lf = g ij (x)∂ij2 f + bi (x)∂i f
i,j =1 i=1
as in (1.11.1). To describe the image generator L f = (L(f ◦ B)) ◦ B −1 , denote

by (y1 , . . . , yn ) = B(x1 , . . . , xn ) the coordinates of B. These functions are the new
coordinates with respect to which the operator will be given and

n
n

Lf =
g ij
(y)∂ij2 f +
bi (y)∂i f
i,j =1 i=1
where

n

Jki (y)J (y) g k B −1 (y)
j

g ij (y) =
k,=1
with Jji the Jacobian of the transformation

∂y i −1
Jji (y) = B (y) .
∂x j
The formula for b is however a bit more complicated since it involves both b and
the derivatives of the g ij ’s. There are expressions which are better behaved than oth-
ers (such as the matrix g = (g ij (x)) or the vector fields, and more generally tensors,
see below) under a change of variables.
It is important here to notice that, from this
point of view, the first order part ni=1 bi ∂i of L is not a vector field. However, in
the decomposition (1.11.10) L = g + Z, the term Z is a vector field. What does
not behave properly in this expression is the second order term ni,j =1 g ij (x)∂ij2 ,

which may appear as ni,j =1 ĝ ij (y)∂ij2 + ni=1 b̂i ∂i after a change of coordinates.
For example, if = ∂x2 + ∂y2 is the usual Laplacian on R2 for which g = Id and
b = 0, in polar coordinates it takes the form
1 2 1
∂r2 + ∂ + ∂r .
r2 θ r
In particular b = 0 in the first form
n buti b = 0 in the second, so that neither the vector
b, nor the first order operator i=1 b ∂i , are intrinsic.
As already mentioned, many ideas in the study of diffusion generators are
taken from Riemannian geometry, in which precisely invariants under the action
of changes of coordinates are investigated. This is in particular the case for ellip-
tic operators, for which the diffusion matrix g = g(x) = (g ij (x))1≤i,j ≤n is, at every
point x, symmetric positive-definite (see (1.12.2)). A Riemannian metric is thus nat-
urally associated with such an elliptic operator, and there is a way to work with and
to identify invariants under changes of coordinates in this geometry. The natural
framework for differential operators and invariants under changes of coordinates is
the setting of differentiable manifolds. The reader will find in Appendix C the basic
geometric notions necessary for the further developments.
Moreover, quantities invariant under a change of coordinates, called intrinsic,
will be of use in the analysis of semigroups. Among these quantities, some are
easier to manipulate, as mentioned above, namely tensors (cf. Sect. C.3, p. 504, for
a complete description of tensor quantities). In a change of coordinates, a tensor is
multiplied by the Jacobian matrix as much as there are top indices in the tensor and
by its inverse as much as there are bottom indices. In the definition of a diffusion
operator, g = (g ij )1≤i,j ≤n is a tensor (twice covariant) but b = (bi )1≤i≤n is not.
1.15.2 Time Change
The time change from t to ct (c > 0) on a Markov semigroup P = (Pt )t≥0 clearly
corresponds to a change of infinitesimal generator from L to c L. We investigate
here changes of L to c(x)L for some function c(x) that also correspond to time
changes, although of more complicated forms, but which have a natural probabilistic
interpretation.
Observe to start with that only positive functions c(x) have to be considered
since the carré du champ operator associated to c L is c , which has to be positive.
For the discussion here, we further restrict to strictly positive bounded functions
0 < a ≤ c(x) ≤ b < ∞ and to diffusion operators L (so that the associated Markov
processes X are sample continuous). While perhaps not the most interesting situa-
tion, once the basic principle is well understood in this case, one can easily adapt it
to the general case.
Now, given the Markov process (Xt )t≥0 (on some probability space ( , , P))
with generator L, starting at some (fixed) point x ∈ E, set for every t ≥ 0,
t 1
Ct = ds.
0 c(Xs )
This is an increasing continuous and differentiable process, with derivative

Ct = c(X1
t)
. For each fixed ω ∈ , Ct = Ct (ω) is a bijective map from R+ into
R+ . Its (generalized) inverse map At , t ≥ 0, is given by
At = inf {s > 0 ; Cs ≥ t}.

It is easily checked that, for every t ≥ 0,

1
At = = c(XAt ).
CA t
Since the process (Ct )t≥0 is adapted to the natural filtration (Ft )t≥0 of (Xt )t≥0 , At
t = XAt , t ≥ 0, is
is, for every t ≥ 0, a stopping time of this filtration. The process X
a Markov process (due to the strong Markov property applied to the stopping times
At ). For any function f in the domain of L, by Itô’s formula,
f
df (Xt ) = dMt + Lf (Xt )dt,
and, by composition,

df (XAt ) = At dMAt + Lf (XAt )dt .
f
The latter may be rewritten as
t ) = d M
df (X tf + c(X
t ) Lf (X
t )dt.
f tf )t≥0 is a martingale

Here (Mt )t≥0 is a martingale for the filtration (Ft )t≥0 , and (M
t )t≥0 is a Markov process for the generator
for the filtration (FAt )t≥0 . Hence, (X
c(x)L. Note that when c is constant, At = ct and we are back to the simple linear
change of time.
1.15.3 Products
This operation will often be referred to in this book as the tensorization proce-
dure. Given two independent families of Markov processes X = {Xtx ; t ≥ 0, x ∈ E}
y
and Y = {Yt ; t ≥ 0, y ∈ F } on possibly different state spaces E and F , the pair
y
(Xtx , Yt ), t ≥ 0, (x, y) ∈ E × F , is again a Markov process on the product space
E × F . If P = (Pt )t≥0 and Q = (Qt )t≥0 denote their respective semigroups, the
product semigroup R = P ⊗ Q is given by the tensor products Rt = Pt ⊗ Qt ,
t ≥ 0. More precisely, the product semigroup (Rt )t≥0 acts on products of functions
h(x, y) = f (x)g(y), f : E → R, g : F → R, as
Rt h(x, y) = Pt f (x) Qt g(y), t ≥ 0, x ∈ E, y ∈ F.
The semigroup (Rt )t≥0 may also be defined via the kernels of the representa-
tion (1.2.4) as

rt (x, y), (dz, du) = pt (x, dz) ⊗ qt (y, du)
for every t ≥ 0 and (x, y) ∈ E × F (with the obvious notation). In case the ker-
nels of (Pt )t≥0 and (Qt )t≥0 admit respective densities pt (x, z) and qt (y, u), t > 0,
(x, z) ∈ E × E, (y, u) ∈ F × F , with respect to measures m and n on E and F

following Definition 1.2.4, then the product semigroup (Rt )t≥0 has kernel densities
pt (x, z)qt (y, u) with respect to the product measure m(dz) ⊗ n(du).
The invariant measure of the product semigroup is the product of the invari-
ant measures. It is reversible as soon as each factor is. The generator of (Rt )t≥0
is the sum of the two generators. That is, if L and M are the respective genera-
tors of (Pt )t≥0 and (Qt )t≥0 , the generator L ⊕ M acts on a function h(x, y) as
Lx h + My h where Lx h denotes the action on the x variable, y being fixed, and con-
versely for My h. In particular, the Laplacian of Rn is the sum of the Laplacians on
the coordinates, and the (heat or Brownian) semigroup on Rn is the product of the
one-dimensional semigroups along each coordinate (and Brownian motion on Rn is
an n-tuple of independent one-dimensional Brownian motions).
This book will often be concerned with tensorization properties of various func-
tional inequalities, and will study how those functional inequalities behave under
this product procedure. This is an important issue when dealing with the behavior
of P⊗n for large n, which arises naturally in many areas such as statistical mechan-
ics, infinite particle systems, probability theory, statistics and non-linear evolution
equations.
1.15.4 Quotients, Commuting Vector Fields and Projections
Quite often, given a semigroup P = (Pt )t≥0 on a state space E with Markov gen-
erator L and associated Markov process X = {Xtx ; t ≥ 0, x ∈ E}, there is a map
h : E → F such that for every t ≥ 0 and every bounded measurable function
g : F → R, Pt (g ◦ h) is a function of h, call it (Qt g) ◦ h. In this case, the fam-
ily of operators Q = (Qt )t≥0 is again a Markov semigroup, with state space F , and
the push-forward process (h(Xt ))t≥0 is a Markov process with semigroup Q. In or-
der for this property to hold, it is enough that L(g ◦ h) = ( Lg) ◦ h for sufficiently
many suitable functions g : F → R where L is an operator on F . The semigroup
Q = (Qt )t≥0 then has Markov generator L. This new semigroup Q and generator

L may then be considered as the quotient semigroup and generator of P = (Pt )t≥0
and L respectively via the projection map h. In this setting, if the semigroup P has
invariant measure μ, then the image semigroup Q has invariant measure ν where ν
is the image measure of μ induced by the map h. This procedure may sometimes
be used to identify image measures, in particular when the operator L is easy to
identify.
Such a situation occurs in particular for diffusion operators when h is real-valued.
In this case, it suffices that Lh be a function of h, as well as (h). Indeed, if
Lh = b(h) and (h) = a(h), then by the diffusion property (1.11.3) and for a suffi-
ciently regular function g : R → R,
L(g ◦ h) = a(h)g (h) + b(h)g (h)

so that

Lg = a(x)g + b(x)g .
The same method applies when h is vector-valued via a multi-dimensional change
of variables formula.
Another situation which often occurs is where there are vector fields
Zj , 1 ≤ j ≤ k, commuting with L, and such that a function f of the domain D(L) is
of the form g ◦ h if and only if Zj f = 0, 1 ≤ j ≤ k. The space F is then viewed the
quotient of E by the action of these vector fields and the function h is the projection
on this quotient space. When there is just one such vector field Z, F is then simply a
parametrization of the integral curves of this vector field. If Zj = Z commutes with
L, then Z also commutes with Pt for every t ≥ 0, and if Zf = 0, then ZPt f = 0.
Indeed, taking the derivative in s ∈ [0, t] of Ps Z(Pt−s f ) yields
Ps (L Z − Z L)Pt−s f = 0.
Therefore, identifying the value at 0 with the value at t, Pt Zf = ZPt f . Hence the
semigroup (Pt )t≥0 preserves the functions vanishing on the vector field Z. More-
over, if Z is a vector field as above, let (Tt )t≥0 be the flow solution of the differen-
tial equation ∂t Tt f = ZTt f (called the exponential of the vector field Z). The flow
(Tt )t≥0 is simply the semigroup—in fact here a group, since there is no need for the
restriction t ≥ 0—with generator Z. Now, if Z with exponential (Tt )t≥0 commutes
with the generator L with semigroup (Pt )t≥0 , then Pt and Tt commute for every
t ≥ 0 (simply repeat the previous procedure by taking the derivative of Tu Pt Ts−u f
for u ∈ [0, s]).
As an illustration, a (smooth) function f in Rn is radial if and only if for all
infinitesimal rotations ij = x i ∂j − x j ∂i , 1 ≤ i, j ≤ n, ij f = 0. It is clear that
these vector fields ij commute with the (standard) Laplace operator on Rn .
Thus the heat semigroup (Pt )t≥0 preserves radial functions, and a new semigroup
may then be constructed by projection on the quotient space R+ , called the Bessel
semigroup (see Sect. 2.4.2, p. 94). Similarly, f only depends on (x1 , . . . , xn−1 ) if
and only if ∂n f = 0, and the image of the Brownian motion in Rn is again a Markov
process in Rn−1 (in this case simply Brownian motion in Rn−1 ). This principle
will be further illustrated in the next chapter in the case of the heat or Brownian
semigroup in Euclidean space, but also on the sphere or on the hyperbolic space in
which the Laplace operator admits a number of symmetries (see Sects. 2.2 and 2.3,
pp. 81 and 88).
Another way to form a semigroup on some image space F can be illustrated
using the carré du champ operator and the reversible measure μ of the semigroup
(Pt )t≥0 (that is, on the Markov Triple structure (E, μ, ), which will be emphasized
later). Namely, let ν be the image measure of μ by a measurable map h : E → F ,
and define a new carré du champ operator on functions g on F by

1 (g)(y) = E g(h) | h = y , y ∈ F,
the conditional expectation being taken under the law μ. (It is possible to give a
meaning to this expression even when μ is not a probability measure via a decom-
position of μ with respect to ν.) It is easily checked that 1 is indeed a carré du
champ operator, and if is of diffusion-type, then so is 1 . In contrast to the pre-
vious procedure, this construction does not require any special properties of the
function h with respect to L. However, the resulting generator is in general difficult
to compute explicitly. This approach will be used repeatedly throughout this work
when dealing with a Lipschitz map h : E → R to transfer functional inequalities
or concentration tail estimates of the space (E, μ) onto R (equipped with the im-
age measure), and will be the basis for most tail estimates for Lipschitz functions
discussed in the following chapters.
1.15.5 Adding a Vector Field
Adding a vector field is a simple analytic operation which however, from the prob-
abilistic point of view, is somewhat delicate. It takes the form of the so-called Gir-
sanov transformation. The basic principle is outlined here only for diffusion pro-
cesses, and even for simplicity for processes on E = Rn with generator given in the
Hörmander form (1.10.5)
1 2
d
L= Zj + Z0
2
j =1
for vector fields Z0 , Z1 , . . . , Zd (with the probabilistic convention). As we have

seen in Sect. 1.10.2, the associated Markov process X = {Xtx ; t ≥ 0, x ∈ Rn } may
be described in this case via the stochastic differential equation (1.10.6).
We would like to add to L a vector field Z, for simplicity of the form
1 j
d
Z= a (x)Zj
2
j =1
where the aj ’s, 1 ≤ j ≤ d, are smooth and bounded functions (beware that Z0 is
excluded in the previous sum). The new operator is L = L + Z, and the task is to
compare the semigroup (P t )t≥0 with generator
L with the semigroup (Pt )t≥0 with
generator L. From a probabilistic viewpoint, the idea is to consider the law of the
process (Xt )t≥0 as a measure on the space of continuous functions on R+ with
values in Rn . Then, the law of the process (X) t≥0 with generator
L has a density
with respect to the law of (Xt )t≥0 given, at each time t, by
t d
1 t j 2
d
j
Nt = exp a j (Xs )dBs − a (Xs )ds ,
0 2 0
j =1 j =1
where (B j )1≤j ≤d are the independent Brownian motions in Rn which appear in the
stochastic differential equation (1.10.6). For a (sketch of a) proof, observe that the
process (Nt )t≥0 thus defined is a martingale by Itô’s formula and that

f
d f (Xt )Nt = f (Xt ) dNt + Nt dMt + Lf (Xt )dt

1
d
j
+ i
∂j f (Xt ) a (Xt )Zi dt .
2
i,j =1
Therefore,
t
Ex Nt f (Xt ) −
Lf (Xt )dt = f (x)
0
where the notation Ex denotes the expectation conditional on X0 = x. This shows
that multiplying the law of Xt by the density Nt solves the martingale prob-
lem (1.4.6) associated to
L. In fact, defining for every suitable function f , every
t ≥ 0 and x ∈ Rn ,

t f (x) = Ex f (Xt )Nt ,
P
we have ∂t P t f = Pt t )t≥0 is indeed the semigroup with generator
Lf so that (P L.
There are of course many different ways to reach this conclusion. Observe that in
order for this conclusion to hold that the vector field Z be a linear combination of
the vector fields Zj , 1 ≤ j ≤ d. This condition is not restrictive when the operator
L is elliptic, but becomes a serious restriction when it is not.
1.15.6 Adding a Potential
It is often the case that one has to look at differential operators of the form

Lf = Lf + Vf,
where L is the generator of a Markov semigroup and Vf is multiplication by some

non-constant function V . This happens in particular when solving Schrödinger-type
equations. Such functions V are then called potentials. The heat equation ∂t f = Lf ,
and the associated semigroup (P t )t≥0 , are then similarly of interest. The latter semi-
group is still positivity preserving, but no longer satisfies P t (1) = 1. This is easy to
understand in the simplest example of finite spaces, since then the matrix L(x, y)
is replaced by the matrix L(x, y) which is equal to L(x, y) when x = y while

L(x, x) = L(x, x) + V (x).
As in the preceding sections, the aim is to compare the semigroup (P t )t≥0 with
generator L with the semigroup (Pt )t≥0 with generator L. When V is constant, we
have Pt = etV Pt , t ≥ 0. Hence Pt (1) ≤ 1 (that is (Pt )t≥0 is sub-Markov) as soon as
V ≤ 0 (and this will remain the case for every non-constant potential V ≤ 0). These
sub-Markov semigroups share a lot of common features with the semigroups killed
at the boundary such as those described later in Sect. 2.4, p. 92, and in Sect. 2.6,
p. 97. For a general (but bounded) potential V , the easiest way to represent (P t )t≥0
starting from (Pt )t≥0 is to use the same trick as for the Girsanov transformation, that
t )t≥0 from the law of the Markov process (Xt )t≥0 with generator L.
is to describe (P
This is done this time with the help of the Feynman-Kac formula. Working again for
simplicity on E = Rn , if
t
At = exp V (Xs )ds , t ≥ 0,
0
then, for t ≥ 0 and x ∈ Rn ,

t f (x) = Ex f (Xt )At .
P
This is much easier to see than for the Girsanov transformation, since in this case
f
d f (Xt )At = At dMt + Lf (Xt )dt + f (Xt )At V (Xt ) dt.
Therefore, for every t ≥ 0 and x ∈ Rn ,

t
t f (x) = f (x) + Ex
P As
L(Xs )ds ,
0
and hence ∂t P t
t f = P Lf .
If the starting operator L is symmetric with respect to some measure μ, so is
the new operator L. But this time, even for a probability measure μ, the constant
function 1 is no longer an eigenvector. If V ≤ 0, then the spectrum still lies in
(−∞, 0). Very often, it is in fact in (−∞, λ0 ) for some λ0 < 0. When this lower
bound of the spectrum of − L is in fact an eigenvalue, it corresponds to (at least) one
eigenvector U0 which is positive. This eigenvector is called the fundamental state (or
ground state) of the system (corresponding to the Perron-Frobenius eigenvector in
the finite Markov chain setting of Sect. 1.9). When it is strictly positive everywhere,
then the transformation of P t , t ≥ 0, into
1
Rt f = e−λ0 t Pt (U0 f ), t ≥ 0,
U0
satisfies Rt (1) = 1, and therefore defines a new Markov semigroup (often called
the ground state transform of the previous one). Hence, the study of such Markov
semigroups may be reduced to the study of Markov semigroups provided that the
ground state U0 may be identified, which is not an easy task in general (see also
Sect. 1.15.8 below for more details).
1.15.7 Removing a Gradient Field
Removing a gradient field (W, ·) (= ∇W · ∇) is not just adding the vector field
with opposite sign. The transformation described here is based on the same principle
as the h-transform and turns out to be very useful. Thus assume that L1 is a diffusion
generator written in the form
L1 f = Lf − (W, f )
on suitable functions f , where L is another generator with carré du champ operator

. It is easily seen that L1 also has as carré du champ operator. Applying L1 to
g = eW/2 f yields L1 g = eW/2 (Lf − W1 f ) with
LW (W )
W1 = − + .
2 4
In other words, denoting by M multiplication by eW/2 ,
M −1 L1 M = L − W1 .
The two operators L1 and L − W1 are conjugate to each other. In particular, for the
respective associated semigroups (Pt )t≥0 and (Qt )t≥0 ,

Qt f = e−W/2 Pt eW/2 f , Pt f = eW/2 Qt e−W/2 f , t ≥ 0.
Moreover, if Pt has kernel densities pt (x, y), t > 0, (x, y) ∈ E × E, with respect to
some measure m, that is

Pt f (x) = f (y) pt (x, y)dm(y),
E
then Qt has a kernel density with respect to m given by
e−(W (x)−W (y))/2 pt (x, y), (x, y) ∈ E × E.
Furthermore, if L is symmetric with respect to μ, then L1 is symmetric with respect

to dμ1 = e−W dμ, and the map f → eW/2 f is an isometry between L2 (μ1 ) and
L2 (μ). In this case, the operators L1 and L − W1 have the same spectral properties
and if the spectrum of one of them is discrete, so is the spectrum of the other, with
the same eigenvalues, the eigenvectors being in correspondence via the above map.
Very often, the latter may be applied in Rn with L1 = − ∇W · ∇. This operator
shares the same spectral properties as the Schrödinger operator − W1 . The anal-
ysis of such operators is quite well-known, in particular there is a lot of classical
analysis concerning their spectral properties, all of which carries over to the study
of the operators − ∇W · ∇.
1.15.8 h-Transform
This subsection deals with another current transformation of a Markov semi-

group known as the h-transform in potential theory. Given a Markov semigroup
P = (Pt )t≥0 on a state space E, let h : E → R be a strictly positive measurable
function. Under appropriate conditions on h, consider the new semigroup given by
1
Pth f = Pt (f h), t ≥ 0,
h
on a suitable class of functions f . If Pt is represented by a kernel pt (x, dy) as
in (1.2.4), then Pth is represented by the kernel
h(y)
pt (x, dy).
h(x)
If (Pt )t≥0 has generator L, then (Pth )t≥0 has generator
1
Lh f = L(hf ).
h
As an illustration, the h-transform can be used to describe the Fokker-Planck op-
erator (see Sect. 1.5) in the symmetric setting. For example, if L = g − ∇W · ∇
on a weighted Riemannian manifold with Laplace-Beltrami operator g and Rie-
mannian measure μg , then for h = eW , Lh = L∗ where L∗ is the adjoint of L in
L2 (μg ).
But this is not the main use of the h-transform. The aim is to continue working
with the generator of a Markov semigroup, and Lh is actually such a generator if
h is harmonic in the sense that Lh = 0. In this case, and provided L is a diffusion
operator according to Sect. 1.11,
Lh f = Lf + 2 (log h, f )
for every suitable function f .

There is a similar representation with strictly positive solutions of Lh = λh
(λ ∈ R) in which case Lh f = Lf + 2 (log h, f ) + λf corresponding to the Markov
semigroup
1
Pth = e−λt Pt (hf ), t ≥ 0.
h
In general, it is difficult to construct such strictly positive harmonic functions on the
whole space (for example on Rn , any strictly positive harmonic function is constant
when L = ), but it may be possible when restricted to suitable open sets (for ex-
ample on R, the function h(x) = x on (0, ∞)). Now, for some classes of harmonic
functions on a subset A of the state space (the minimal strictly positive ones), this
new semigroup corresponds to the Markov process conditioned to stay in A. The
associated process is known as the h-process associated to the original Markov pro-
cess with semigroup P.
It is beyond the scope of this short section to develop the full theory of
h-processes (which is indeed quite complicated and for which we refer the reader to
the appropriate literature). However, we may at least give a flavor of it by compar-
ison with Markov chains (cf. Sect. 1.9). Indeed, consider a Markov chain (Xk )k∈N
(on a finite set E, say) with transition matrix P = (P (x, y))(x,y)∈E×E , and assume
for simplicity that for any (x, y) ∈ E × E, P (x, y) > 0. The quantity P (x, y) rep-
resents the probability that the chain will jump from x to y. Now restrict the matrix
P (x, y) to (x, y) ∈ A × A where A ⊂ E, and denote by P A this restricted square
matrix. The (unique strictly positive) Perron-Frobenius eigenvector U of P A satis-
fies

P A (x, y) U (y) = λ U (x)
y∈A
for all x ∈ A, corresponding to the (real) eigenvalue λ with maximum modulus.

Then, consider the new Markov matrix on A × A defined by
1
QA (x, y) = P A (x, y) U (y), (x, y) ∈ A × A.
λU (x)
Given x ∈ A, for k < , consider the law of (X1 , . . . , Xk ) conditioned to the event
{Xi ∈ A, 1 ≤ i ≤ }. This is not the law of a Markov chain, but it becomes so if
→ ∞. In the limit, this law is the law of the Markov chain with transition ma-
trix QA . This appears as a consequence of some elementary (but rather tedious)
direct computations, which, in this case, can be made explicit from the matrix
P . The h-transform is then simply the transfer of this principle to the continuous
time case, in the larger setting of Markov semigroups. Notice that in this case the
h-transform is the same operation as the ground phase state transformation de-
scribed in Sect. 1.15.6.
1.15.9 Subordination
Subordination is another way to produce new semigroups from a given one, through
time averaging. Let (Pt )t≥0 be a Markov semigroup on E with generator L. For any
probability measure ν on R+ , define a new Markov operator by setting
∞
Pν = Pt dν(t).
0
The integral may be considered as an integral operator,

∞or more naively as acting on
bounded measurable functions f : E → R as Pν f = 0 Pt f dν(t).
From a probabilistic point of view, if (Xt )t≥0 is the Markov process associated
with the semigroup (Pt )t≥0 , consider a random variable T with values in R+ with
law ν, independent of the process (Xt )t≥0 . Then Pν is the law of the random variable
XT in the sense that for all bounded measurable functions f on E, and all x ∈ E,

Pν f (x) = E f (XT ) | X0 = x .
From the definition, it is clear that Pν is positivity preserving and satisfies

Pν (1) = 1 (ν is a probability measure). When Pt for t > 0 has a kernel density
pt (x,
y), (x, y) ∈ E × E, with respect to a measure m, then Pν has the kernel den-
∞
sity 0 pt (x, y)dν(t) as soon as this integral converges. Formally, Pν = Gν (−L)
where
∞
Gν (λ) = e−λt dν(t), λ ∈ R.
0
For example, if dν(t) = αe−αt dt for some α > 0, the resulting operator Gν is known
as the resolvent Gα = α−Lα
of L (see Sect. A.1, p. 473).
The preceding construction may actually be pushed further to define a new semi-
group (Qs )s≥0 of operators via a family νs , s ≥ 0, of probability measures. For this
to happen, it is enough that the family νs , s ≥ 0, satisfies, for any λ ∈ R+
∞
e−λt dνs (t) = e−sψ(λ)
0
(for some positive function ψ on R+ ). Such families of measures νs , s ≥ 0, on R+

are called subordinators. The semigroup (Qs )s≥0 then has generator −ψ(−L). In
general, the admissible functions ψ are described as
∞

ψ(λ) = dλ + 1 − e−λx d(x), λ ∈ R+ ,
0

∞
where d ∈ R+ and is a measure with support in R+ such that 0 (1 ∧ x) d(x)
< ∞.
The main example of a subordinator (known as the 12 -stable subordinator) is
described by the function ψ(λ) = λ1/2 giving rise to the family
1 s
dνs (t) = √ 3/2 e−s /4t dt,
1/2 2
s ≥ 0, (1.15.1)
2 π t
corresponding to d = 0 and d(x) = C x −3/2 dx. (In general, α-stable subordina-

tors satisfy√ψ(λ) = λα for α ∈ (0, 1) and d(x) = Cα x −1−α dx.) In this case, since
Qs = e−s −L , for a given f , the function H (s, x) = Qs f (x), s ≥ 0, x ∈ E, is a
solution of the equation
2
∂s + L H = 0, H (0, x) = f.
Therefore, while (Pt )t≥0 describes the solution of the parabolic heat equation as-
sociated to the generator L, the semigroup (Qs )s≥0 is the solution of the extended
elliptic equation ∂s2 + L = 0 on R+ × E (elliptic at least when L itself is elliptic).
Again, the preceding operation has a clear probabilistic interpretation. The mea-
1/2
sure νs arising in (1.15.1) is precisely the law of the hitting time
T = inf{t > 0 ; Bt = 0} for a one-dimensional Brownian motion (Bt )t≥0 starting
from B0 = s√> 0. Indeed, it is enough to observe that, for α > 0, the function
F (t, x) = e− α x−αt , t ≥ 0, x ∈ R, satisfies (∂t + ∂x2 )F = 0 and is bounded on R+ .
Therefore F (t, Bt ) is a martingale on [0, T ] such that E(F (T , BT )) = E(F (0, s))
(since B0 = s) from which it follows that
√
Es e−αT = e− α s
1/2
which is the known Laplace transform of the probability measure νs .
The preceding representation of the solutions of the elliptic equation
(∂s2 + L)H = 0 from the semigroup (Qs )s≥0 and the law of some random variable
on R+ may be easily extended to more general equations of the form
2
∂s + α(s)∂s + L H = 0, H (0, x) = f,
t = 0} where (B
with the help of the hitting time inf{t > 0, B t )t≥0 is the diffusion on
R+ with generator ∂t2 + α(t)∂t and initial value B 0 = s. At least for H bounded, one
may consider to this end the martingale H (B t , Xt ), t ≥ 0, with (Xt )t≥0 the Markov
process with generator L assumed to be independent of (B t )t≥0 . Then, by the mar-
tingale property, we use the fact that E(s,0) (H (B 0 , X0 )) = E(s,0) (H (BT , XT )). Of
course, this construction yields a new semigroup only when α is constant. For ex-
ample, later in Sect. 2.2, p. 81, a slight modification of the 12 -stable subordinator
α,1/2 2
will be considered, defined for α > 0 by dνs (t) = eαs−α t dνs (t) and satisfying
∞ √
e−λt dνs (t) = e−s( λ+α −α) , λ ∈ R+ .
α,1/2 2
(1.15.2)
0
The resulting semigroup is solution of

2
∂s − 2α∂s + L H = 0, H (0, x) = f.
To conclude this section, we note that the subordination procedure may be de-
veloped for lots of different representations of the operator L. These usually make
sense only in some specific situations or on some restricted classes of functions. For
example, the Riesz potentials (−L)−α may be represented for α > 0 as
∞
γα−1 t α−1 Pt dt
0

∞
where γα = 0 t α−1 e−t dt. This operator may be bounded from some Lp -space
into another Lq -space as is the case, for example, for the heat semigroup, but is only
defined on the space of mean-zero functions in the finite measure case in general.
Although these representations should be handled with care in general, they provide
useful norm bounds and positivity preserving properties of the associated operators.
Of particular interest is the value α = 1 corresponding to another useful repre-

sentation in this direction, which is the potential
∞
−1
(−L) = Pt dt.
0
This integral does not make sense in general although it may be defined under mini-
mal reasonable conditions. Furthermore, it has a clear meaning in specific instances.
For example, as will be discussed later in Chap. 4, as soon as the invariant measure
μ of (Pt )t≥0 is finite (and normalized into a probability measure) and satisfies a
Poincaré inequality, it holds that for some constant λ > 0,
Pt f 2 ≤ e−λt f 2 , t ≥ 0,

on the set L20 (μ) of functions f ∈ L2 (μ) with E f dμ = 0. By invariance of μ, Pt

maps L20 (μ) into itself, so that in this context (−L)−1 is well-defined as a bounded
operator on L20 (μ). It produces for any function f ∈ L20 (μ) a unique function
g ∈ L20 (μ) ∩ D(L) such that Lg = −f .
These observations are actually the key to ergodicity as outlined in Sect. 1.8.
Indeed, given f ∈ L20 (μ) with g such that Lg = −f , the martingale representa-
tion (1.4.6) shows that
t
g
g Xtx − g(X0 ) = Mt + f (Xs )ds, t ≥ 0.
0
g
With the help of martingale theory to handle 1t Mt , it may then be deduced that

1 t
t 0 f (Xs )ds converges to 0 (μ-almost surely) as t → ∞.
1.16 Curvature-Dimension Condition

This last section introduces the notion of a curvature-dimension condition through
the iterated carré du champ operator 2 which will play a central role in the inves-
tigation of functional inequalities associated to Markov semigroups and operators.
It describes the geometric aspect of this investigation. We only give here the main
flavor of this basic idea which will be developed in subsequent chapters. The geo-
metric Riemannian background at the origin of the concept is presented in Sects. C.5
and C.6, p. 511 and p. 513 of Appendix C.
1.16.1 2 Operator
Following the previous developments, consider a symmetric Markov semigroup

P = (Pt )t≥0 with invariant reversible measure μ and infinitesimal generator L. Re-
1.16 Curvature-Dimension Condition 71
call from Definition 1.4.2 the carré du champ operator defined on a suitable alge-
bra A of functions in the L2 (μ)-domain D(L) of L by
1
(f, g) = L(f g) − f Lg − g Lf , (f, g) ∈ A × A.
2
The idea (taken from a Riemannian viewpoint, as explained below) is to formally
repeat this definition, replacing the product operation by , to define a new operator
2 (iterated carré du champ operator) as
1
2 (f, g) = L (f, g) − (f, Lg) − (Lf, g) , (1.16.1)
2
for any pair (f, g) of functions such that the various terms on the right-hand side
are well-defined. As for the carré du champ operator , we often write more simply
2 (f ) = 2 (f, f ).
By symmetry of L with respect to the measure μ, and as long as f , Lf , g, Lg
and (f, g) are in the L2 (μ)-domain of L, and if moreover (f, g) is in the L1 (μ)-
domain, the integration by parts formula for the 2 operator reads as

2 (f, g)dμ = (Lf )(Lg)dμ. (1.16.2)
E E
It might also be useful to record at this stage that the diffusion property for L or
(Definition 1.11.1) leads to a change of variables formula for the 2 operator. For
example, using (1.11.3) and (1.11.5), if ψ : R → R is smooth enough, an elementary
calculation yields that

2 ψ(f ) = ψ (f )2 2 (f ) + ψ (f )ψ (f ) f, (f ) + ψ (f )2 (f )2 . (1.16.3)
Note that this formula is presented in (C.6.7), p. 516, as a consequence of the stan-
dard differential calculus rules in differentiable manifolds whereas the diffusion
property from Definition 1.11.1 emphasizes here a more intrinsic and efficient cal-
culus, called -calculus (or 2 -calculus). The power of this calculus will be demon-
strated at length throughout this work.
Some care has to be taken in the definition of and 2 concerning the choice of
the class A of functions. This issue will be discussed in Chap. 3. For the examples
considered in this short introduction, say on E = Rn or on a manifold, A may be
taken for simplicity to be the class of smooth functions with compact support. If
is the standard Laplacian on Rn , and if f : Rn → R is smooth, then
(f ) = |∇f |2 and 2 (f ) = |∇∇f |2
(where ∇∇f = (∂ij f )1≤i,j ≤n is the Hessian of f ). More generally, as developed in

Sect. C.6, p. 513, and in a Riemannian language, if L = g − ∇W · ∇ as described
in Sect. 1.11.3, then ((f ) = |∇f |2 and)
2 (f ) = |∇∇f |2 + Ric(L)(∇f, ∇f ) (1.16.4)

where Ric(L) is a symmetric tensor defined from the Ricci tensor Ricg of the Rie-
mannian (co-) metric g by Ric(L) = Ricg +∇∇W . On Rn with the flat Euclidean
metric, and for the usual Laplacian , this would simply be Ric(L) = ∇∇W and
thus
2 (f ) = |∇∇f |2 + ∇∇W (∇f, ∇f ) (1.16.5)
on smooth functions f : Rn → R.
1.16.2 Curvature-Dimension Inequalities
As is clear from these examples, unlike the carré du champ operator , the new oper-
ator 2 is not always positive. For Laplacians g on Riemannian manifolds (M, g),
the 2 operator is an expression of the Bochner-Lichnerowicz formula (1.16.4)
(cf. Theorem C.3.3, p. 509) and is positive if (and only if) the Ricci curvature Ricg
of the manifold is positive. More precisely, for an elliptic operator L, there is al-
ways a function ρ(x) such that for every smooth function f and at every point x
in the space, 2 (f ) ≥ ρ(x) (f ). For a Laplacian on a Riemannian manifold, the
best possible function ρ(x) in such an inequality is precisely the infimum of the
Ricci tensor (that is, at any point, the smallest eigenvalue of some symmetric matrix
evaluated from the coefficients of L). For non-elliptic operators, there is no such
function in general.
Many differential inequalities may be seen as consequences of an inequality of
the form
2 (f ) ≥ ρ (f ) (1.16.6)
for some ρ ∈ R and all f ’s in A (or in the respective domains of and 2 ). Such an
inequality will be called a curvature condition CD(ρ, ∞). To explain the meaning
of the second parameter ∞, we turn to a more general definition.
Definition 1.16.1 (Curvature-dimension condition) A diffusion operator L is said

to satisfy the curvature-dimension condition CD(ρ, n), for ρ ∈ R and n ∈ [1, ∞],
if for every function f : E → R in a sufficiently rich class A,
1
2 (f ) ≥ ρ (f ) + (Lf )2
n
(μ-almost everywhere).
The preceding definition is actually subordinate to the chosen class A of func-

tions. Detailed and useful conditions on the class A for which such a definition
should be considered will be deeply investigated in Chap. 3.
As discussed in Sect. C.6, p. 513 (to which we refer for a more detailed discus-
sion), for a Laplacian on a Riemannian manifold, the curvature-dimension condition
CD(ρ, n) holds if and only if the Ricci curvature is bounded from below by ρ and
1.16 Curvature-Dimension Condition 73
n is greater than or equal to the (topological) dimension of the manifold (that is, the
example from which the inequality actually takes its name). For example, the usual
Laplacian on Rn satisfies the condition CD(0, n) (since |∇∇f |2 ≥ n1 (f )2 ).
But this is not necessarily true for other operators. For example, the Ornstein-
Uhlenbeck operator investigated in Sect. 2.7.1, p. 103, satisfies CD(1, ∞) on any
(finite-dimensional) state space, but does not satisfy CD(ρ, n) for any finite n.
As presented in Sect. 1.11.3, a general elliptic differential operator L on a man-
ifold with dimension n is uniquely decomposed as L = g + Z where g is the
Laplacian associated to a Riemannian (co-) metric g and Z is a vector field. Then L
satisfies a curvature-dimension condition CD(ρ, m) if and only if m ≥ n and, set-
ting ∇S Z to be the symmetric covariant derivative of Z in the metric g (that is the
symmetrized tensor ∇Z),
1
Ricg −∇S Z ≥ ρ g + Z ⊗ Z. (1.16.7)
m−n
Note in particular that m = n only when Z = 0 so that, in this sense, Laplacians are
operators with minimal dimension among all elliptic operators. Note that here the
curvature condition CD(ρ, ∞) boils down to Ricg −∇S Z ≥ ρ g. In particular, when
Z = −∇W · ∇, as will normally be the case throughout this book, (1.16.7) reads as
Ricg +∇∇W ≥ ρ g. (1.16.8)
In particular, under this condition, a weighted Riemannian manifold (M, g)

equipped with the measure dμ = e−W dμg will sometimes be said to be of cur-
vature bounded from below by ρ. On the flat manifold Rn for which Ric = 0, the
latter simply amounts to a convexity condition on the potential W .
Curvature-dimension conditions behave nicely under the product procedure de-
scribed in Sect. 1.15.3. Indeed, if two operators Li on respective spaces Ei with
associated carré du champ operators i , i = 1, 2, satisfy the curvature-dimension
conditions CD(ρi , ni ), i = 1, 2, then the operator L1 ⊕ L2 on the product space
E1 × E2 with associated carré du champ operator 1 ⊕ 2 satisfies the CD(ρ, n)
condition with ρ = min(ρ1 , ρ2 ) and n = n1 + n2 . This may be seen directly from
the definition of the product carré du champ operator, or in a more naive way when
both spaces are manifolds in a local system of coordinates through the computations
developed in Sect. C.5, p. 511. Note that this tensorization property is not restricted
to the reversible case. This observation in particular fully justifies the name “dimen-
sion” for the parameter n.
Finally, note that the curvature-dimension condition CD(ρ, n) takes a simple
form in dimension one for operators defined on an interval by Lf = f − a(x)f .
In this case, CD(ρ, n) is equivalent to the differential inequality on a given by
a2
a ≥ ρ + . (1.16.9)
n−1
Such one-dimensional operators will be illustrated in model examples in Chap. 2.
Owing to their simplicity, they will prove to be useful when testing our intuition
regarding similar convexity properties for more general operators on more general
spaces.
1.17 Notes and References

This section collects some general references concerning various aspects of the
study of Markov semigroups, generators and processes, which was mostly devel-
oped in the second half of the 20th century. At this stage, only a few general refer-
ences (with subjective choices) are included. More precise references to specific re-
sults and properties will be pointed out within the main text. Since this chapter only
surveys and briefly presents a few specific aspects of the general theory of Markov
semigroups and processes, we refer the reader to these references, and more, for a
complete account of this vast subject. Recall that Chap. 3 will set up a consistent
framework in which the results of this monograph may be developed.
General references on semigroups of operators are [95, 96, 100, 143, 154, 168,
169, 241, 244, 264, 355, 356, 444]. Applications of semigroups of operators to par-
tial differential equations are developed in [174, 205, 344, 392, 393, 419]. The
standard main references [203, 247, 248] provide complete accounts on elliptic
and hypo-elliptic partial differential equations of the second order relevant to the
topics developed here. Elementary martingale theory may be found, for example,
in [152, 289, 350, 358, 363, 440] (cf. also Appendix B). The interplay between
potential theory, martingales and Markov processes is developed in the mono-
graphs [50, 72, 132, 142, 152, 153, 164, 170, 178, 252, 255, 307, 362, 363, 393,
418]. For a point of view related to Dirichlet forms, see [91, 189, 190, 294, 398].
Early developments of the theory in the spirit and at the root of this mono-
graph can be found in the lecture notes [26]. G. Royer’s introduction to logarith-
mic Sobolev inequalities [372] presents basic and relevant material on Markov
semigroups and processes along the lines of this investigation. The book [431] by
F.-Y. Wang is a comprehensive investigation of Markov operators and functional
inequalities with some overlap with the content of this book.
The outline on Markov processes and semigroups of Sects. 1.1, 1.2, 1.3 and 1.6
follows the preceding general references. Standard expositions on probabilities and
measures on general measurable spaces are [67, 83]. The measure decomposition
theorem in good measurable spaces, as well as Propositions 1.2.3 and 1.2.5, may be
found in [151] (see also [150]). (Actually, a statement such as for example Propo-
sition 1.2.5 does not necessarily require the good measurable space framework pro-
vided suitable martingale convergence theorems are available.)
The history of the carré du champ operator of the infinitesimal generator of
a Markov semigroup (Sect. 1.4) is discussed in [153], with a particular reference
to [366].
The Fokker-Planck point of view of Sect. 1.5 is emphasized mostly in the partial
differential equation literature with the Lebesgue measure as reference measure.
The basic notion of Dirichlet form presented in Sect. 1.7, gradually put forward
in various parts of mathematics, is taken here from the standard reference [189] by
1.17 Notes and References 75
M. Fukushima (see also [91, 190, 294, 431]). Rota’s Lemma (Lemma 1.6.2) may be
found in [196].
Ergodic properties refer to a variety of behaviors as time goes to infinity and only
a very specific feature is examined in Sect. 1.8 and throughout this book.
Elements on Markov chains (Sect. 1.9) on finite and countable spaces may be
found in [309, 333, 357].
Complete accounts on stochastic differential equations and diffusion processes
and semigroups, as outlined in Sects. 1.10 and 1.11, are [50, 152, 252, 255, 263,
350, 358]. Martingale problems and the interplay between analysis and probability
theory is exposed in [256, 393].
Hypo-ellipticity is one major topic of the Hörmander theory [247, 248] briefly
discussed in Sect. 1.12.
The relevant aspects on domains of infinitesimal operators are briefly presented in
Sect. 1.13 following the standard material and references on the subject (see above
and also Appendix A and Chap. 3).
Section 1.15 gathers a variety of tools in the investigation of Markov semigroups,
generators and processes. Some of them are part of the folklore and not always
explicitly stated in standard references. Girsanov (Sect. 1.15.5) and Feynman-Kac
(Sect. 1.15.6) formulas are presented in standard references on stochastic calculus
such as [252, 350, 358, 393]. Potential theory and h-transforms as mentioned in
Sect. 1.15.8 are deeply investigated in [164] (see also [153]) where the reader will
find a comprehensive account of the interplay between potential theory and proba-
bility theory. More on subordinators (Sect. 1.15.9) may be found in [15, 64, 65, 379]
and in the references therein. The basics on Riesz potentials may be found for ex-
ample in [164, 342, 388].
The 2 operator and the associated notion of a curvature-dimension condition
were introduced in the early contributions [24, 36] on the basis of the Bochner-
Lichnerowicz formula in Riemannian geometry towards the study of logarithmic
Sobolev inequalities (see Chap. 5) and Riesz transforms (cf. [26]). Curvature or
curvature-dimension conditions in terms of the 2 operator are sometimes referred
to as the “2 criterion”.
Chapter 2
Model Examples
This chapter is devoted to some basic model examples, which will serve as a guide
throughout this book. These model examples will also provide an opportunity to
illustrate some of the ideas, definitions and properties of Markov semigroups, op-
erators and processes introduced in the first chapter. Moreover, they will help to
set up the framework for the investigation of more general Markov semigroups and
generators as will be achieved in the next chapter.
The two simplest examples of diffusion semigroups and generators (at least
among those for which the semigroup is explicitly known) are the (Euclidean) heat
semigroup and the Ornstein-Uhlenbeck semigroup, with associated Brownian mo-
tion and Ornstein-Uhlenbeck process. There are of course other fundamental exam-
ples, as models or references for comparison, for which there is however in general
no explicit formulas for the semigroups so that they are only described in terms
of their generators. We present here some of these models, with a special focus on
the underlying Laplacian or diffusion generator. With respect to Chap. 1, most of the
Markov semigroups presented in this chapter will indeed be introduced by their gen-
erators (defined on classes of smooth functions). The examples considered here will
actually present an opportunity to discuss the existence and uniqueness of (symmet-
ric) semigroups with given generators (emphasized in the preceding chapter as the
essential self-adjointness issue). Complete justifications are developed in the next
chapter together with the description of the relevant classes of functions.
The chapter starts with the three geometric models of the heat semigroup on the
Euclidean space, the sphere and the hyperbolic space. For each model, we care-
fully describe the geometric framework and present the natural Laplacian giving
rise to the associated heat, or Brownian, semigroup. Sections 2.4 and 2.5 discuss
the heat semigroup with Neumann, Dirichlet or periodic conditions on R or on an
interval of R. More general Sturm-Liouville operators on an interval of the real line
are examined next, and are illustrated in Sect 2.7 by the examples of the Ornstein-
Uhlenbeck or Hermite, Laguerre and Jacobi operators. These important examples
are the three models of one-dimensional diffusion operators which may be diag-
onalized with respect to a basis of orthogonal polynomials, thus leading to deep

DOI 10.1007/978-3-319-00227-9_2,
78 2 Model Examples
analytic connections. Following the first chapter, the exposition often combines the
various analytic, probabilistic and geometric viewpoints.
2.1 Euclidean Heat Semigroup
From the operator-theoretic viewpoint developed in the first chapter, the heat or
Brownian semigroup in the Euclidean space Rn is the Markov diffusion semigroup
with infinitesimal generator the usual Laplacian = Rn on Rn defined on smooth
functions f : Rn → R by

n
f = Rn f = ∂i2 f
i=1
and with invariant and reversible measure the Lebesgue measure dx. The carré du
champ operator is simply given by

n
(f, g) = ∇f · ∇g = ∂i f ∂i g
i=1
for smooth functions f, g on Rn , and in particular (f ) = |∇f |2 is the standard

Euclidean length of the gradient of f . A convenient function algebra to use here is,
for example, the class of smooth (C ∞ ) compactly supported functions.
The associated Markov process is the standard Brownian motion B = (Bt )t≥0 in
Rn . It consists of n independent standard real Brownian motions Bt = (Bt1 , . . . , Btn ),
t ≥ 0, starting at x = (x1 , . . . , xn ) ∈ Rn . The distribution of Bt at time t is the law
of a Gaussian vector centered at x with covariance matrix t Id. B = (Bt )t≥0 ac-
tually corresponds to the more probabilistic normalization which works with 12 .
However, we mostly deal with later and therefore with (B t )t≥0 = (B2t )t≥0 as as-
sociated Markov process. Observe that this change from t to 2t in the formulas may
lead to some confusion when comparing with the standard literature on the subject.
The Laplace operator with the Lebesgue measure as invariant and symmetric
reference measure defines a Markov semigroup P = (Pt )t≥0 with kernel densities
1
e−|x−y| /4t ,
2
pt (x, y) = n/2
t > 0, (x, y) ∈ Rn × Rn . (2.1.1)
(4πt)
In other words, for every, say, bounded measurable function f : Rn → R,

Pt f (x) = f (y) pt (x, y)dy, t > 0, x ∈ Rn .
Rn
These kernels classically solve the (parabolic) heat equation ∂t pt = pt (where
acts either on x or y). The semigroup (Pt )t≥0 may via described in probabilistic
2.1 Euclidean Heat Semigroup 79
terms via the representation formula

√
Pt f (x) = Ex f (Bt ) = E f x + 2t G , t ≥ 0, x ∈ Rn , (2.1.2)
where G is a standard Gaussian vector in Rn with distribution the standard Gaussian

probability measure (2π)−n/2 e−|x| /2 dx.
2
As such, the semigroup (Pt )t≥0 is a Markov semigroup in the sense of (i)–(v)
of Definition 1.2.2, p. 12. To check the continuity assumption (vi), that is, for any
f ∈ L2 (dx), Pt f → f in L2 (dx) as t → 0, start with a smooth and compactly
supported function f and extend the result by density.
The semigroup P = (Pt )t≥0 is thus called the heat or Brownian semigroup on Rn .
It is one of the few examples for which an explicit description of the transition prob-
abilities is available. The Chapman-Kolmogorov equation (1.3.2), p. 17, appears as
a consequence of the fact that the sum of two independent Gaussian vector in Rn
with respective covariance matrices 2t Id and 2s Id is a Gaussian vector with covari-
ance 2(t + s) Id. Similarly, the dual Fokker-Planck equation (1.5.2), p. 24, is easily
checked with the observation that, for R = |x − y|2 , R = 2n and (R) = 4R (on
either x or y).
Note that the constant function 1 is not in the L2 (dx)-domain of since not
integrable with respect to the Lebesgue measure dx. As observed in Sect. 1.4, p. 18,
the equation Pt (1) = 1 is not a direct consequence of the fact that (1) = 0. We
come back to this issue later when dealing with completion and boundary values.
Among the further useful observations on the Euclidean heat semigroup (Pt )t≥0 ,
observe
that the operators Pt are not Hilbert-Schmidt. Indeed, for every t > 0 and
x ∈ Rn , Rn pt (x, y)2 dy = (8πt)
1
n/2 and hence

pt (x, y)2 dxdy = ∞
Rn Rn
(see Sect. A.6, p. 483). This is again a consequence of the fact that the reversible
measure (the Lebesgue measure) is infinite. Actually, there are no square integrable
eigenvectors of the Laplacian . Indeed, the eigenvectors are typically of the form
y → eix·y , which are never integrable.
On the geometric side, one can immediately verify that the 2 operator, (1.16.1),
p. 71, of the standard Laplacian on Rn is given on smooth functions f : Rn → R
by

n
2 2
2 (f ) = 2 (f, f ) = |∇∇f |2 = ∂ij f .
i,j =1
n
Since by the Cauchy-Schwarz inequality i,j =1 (∂ij f )
2 ≥ n1 ( ni=1 ∂i2 f )2 , for ev-
ery smooth f ,
1
2 (f ) ≥ (f )2
n
so that the Laplace operator on Rn satisfies the curvature-dimension condition
CD(0, n) of Definition 1.16.1, p. 72 (with optimal values).
80 2 Model Examples
Using this example, we next illustrate some of the semigroup tools and properties
described in Chap. 1. We start with the commutation relations with suitable vector
fields as described in Sect. 1.15.4, p. 60. Brownian motion B = (Bt )t≥0 is clearly
translation invariant. The process starting at x is the process starting at 0 trans-
lated from x. This property actually illustrates the commutation of the Laplacian
with translations. A translation in the direction u ∈ Rn is expressed on functions as
f (x) → Tt f (x) = f (x + tu), t ≥ 0, x ∈ Rn . As t ≥ 0 evolves, the family (Tt f )t≥0
is nothing else but the solution of the differential equation ∂t Tt f = ZTt f where Z
is the constant vector field Zf = ni=1 ui ∂i f . The vector field Z clearly commutes
with , and thus in the heat semigroup P = (Pt )t≥0 we have Pt (Ts f ) = Ts Pt f for
all choices of t and s (≥ 0).
A similar observation may be applied to rotations. The Laplacian indeed also
commutes with the operators ij = xi ∂j − xj ∂i , 1 ≤ i, j ≤ n. In particular, this
commutation justifies the invariance by rotation of the law of Brownian motion
B = (Bt )t≥0 when x = 0. Namely, if rijt denotes the rotation of angle t in the 2-plane
(ei , ej ), then for any f and x ∈ Rn , f (rijt x) = Rt f (x) where Rt = etij is the
(semi-) group generated by the vector field ij . Since ij vanishes at the origin,
and hence 0 is left-invariant under Rt , the commutation between and ij ensures
that for any bounded measurable function f , Ps Rt f (0) = Rt Ps f (0) = Ps f (0), jus-
tifying the rotational invariance of Brownian motion. Of course, there is no need to
use such a complicated argument to observe the obvious fact that the (Gaussian) law
of Bt is rotation invariant when the origin is 0. This presentation however empha-
sizes how to use similar arguments in more involved instances, when for example
the law of the associated Markov process is not explicitly known.
The law of the Brownian motion B = (Bt )t≥0 has of course several other remark-
able properties. In particular, for every t ≥ 0, the law of Bt starting from 0 is also the
law of t 1/2 B1 . This again is a commutation property, this time with the vector field
Df (x) = ni=1 xi ∂i f (x), the exponential of which being the dilation semigroup
f (x) → Dt f (x) = f (et x) for which
[, D] = D − D = 2.
By the same argument as above, Ps Dt = Dt Pe2t s for every s, t ≥ 0 which indeed

leads to the dilation property after noticing that Dt f (0) = f (0) for every t.
Finally, the heat kernels used to solve the heat equation lead through subordina-
tion to solutions to other equations (Sect. 1.15.9, p. 67). For example, the 12 -stable
subordinator given in (1.15.1), p. 68, leads to the Cauchy kernels qt (x, y), t > 0,
(x, y) ∈ Rn × Rn , given by
∞
1/2 t
qt (x, y) = ps (x, y)dνt (s) = Cn
0 (t 2 + |x − y|2 )(n+1)/2
(where Cn > 0 is the suitable normalizing constant). For any bounded measurable
function f on Rn , the Cauchy kernels yield the harmonic extension H (t, x) of f on
2.2 Spherical Heat Semigroup 81
R+ × Rn , that is the solution of

2
∂t + H = 0, H (0, x) = f (x),

via the representation H (t, x) = Rn f (y)qt (x, y)dy.
2.2 Spherical Heat Semigroup
This section is devoted to the analogue of the heat semigroup on the standard sphere
Sn (in Rn+1 ). This is one most useful model spaces of a compact manifold without
boundary, moreover of strictly positive (constant) curvature. It is closely related to
the Jacobi semigroups described below in Sect. 2.7.4.
To get a clear picture, it is important to first provide suitable geometric descrip-
tions of the state space in order to address the various objects under investigation,
such as the generator and carré du champ operator. The spherical Laplacian, de-
noted by Sn below, may be defined as the Laplace-Beltrami operator on the Rie-
mannian manifold Sn as introduced in Sect. C.3, p. 504. It is invariant and symmetric
with respect to the unique probability measure on Sn , denoted σn , which is invari-
ant under rotations in Rn . As an embedded (in Rn+1 ) manifold, representations of
the sphere Sn in explicit charts may actually be provided. We present two main
representations, the orthogonal projection representation and the stereographic pro-
jection representation. For each representation, we provide an explicit description
of the spherical Laplacian, the invariant measure and the associated carré du champ
operator. Below, we use the standard unit sphere Sn in Rn+1 , but simple scaling
arguments cover spheres of arbitrary radius. Recall also that, according to the expo-
sition in Appendix C, emphasis is placed on the (Riemannian) co-metric rather than
the usual metric.
2.2.1 Orthogonal Projection Representation
The orthogonal projection representation of a point in Sn ⊂ Rn+1 , n ≥ 1, amounts

to symmetrically cutting the sphere by the horizontal hyperplane Rn , and associating
to a given point x = (x1 , . . . , xn ) in the open unit ball B of this hyperplane the point

x1 , . . . , xn , 1 − |x|2

on the sphere Sn in Rn+1 . As usual, |x|2 = ni=1 xi2 is the square of the Euclidean
norm in Rn . The unit ball B of Rn is therefore a chart of the upper half-sphere (in the
language of Sect. C.1, p. 500). This chart is obtained by considering the orthogonal
projection from Rn+1 onto Rn . In Rn+1 , consider now a vector U tangent to the
of
sphere at a point x. Via this projection, the vector U is transformed into a vector U
82 2 Model Examples
i )1≤i≤n . Similarly, for every vector U

Rn with coordinates (U of Rn , and any x ∈ Sn ,

U may be seen as the projection of a vector tangent to the sphere at x. However,
the norm |U | of U in Rn+1 is not the norm of U in Rn , but a quantity related to

some Euclidean metric G = G(x) = (gij (x))1≤i,j ≤n by |U |2 = ni,j =1 gij (x)U i U
j
where
xi xj
gij (x) = + δij , 1 ≤ i, j ≤ n.
1 − |x|2
The dual metric (or co-metric, cf. Sect. C.3, p. 504) of interest here, given by the
inverse matrix g(x) of (gij )(x), is, at any point x ∈ Sn ⊂ Rn+1 in these local coor-
dinates,
g ij (x) = δ ij − x i x j , 1 ≤ i, j ≤ n.
At this level, we switch to upper indices in agreement with the notation of Rie-
mannian geometry (cf. Appendix C). (In the present setting, x i is at the same time
the coordinate and the differential 12 ∂i (|x|2 ). Therefore the choice of notation xi
or x i is irrelevant and depends on the context.) It is easily seen that the latter de-
fines a strictly positive metric on B. The Ricci tensor Ricij of this metric is equal to
(n − 1) g ij . This metric is thus of constant Ricci curvature (equal to n − 1). Actually,
the full Riemann curvature tensor is constant (cf. Sect. C.3, p. 504). At this point, the
Laplacian is only defined in the upper half-space. In order to extend the Laplacian
to the whole space one might consider a similar chart for the lower half-sphere, but
this would still exclude the equator (i.e., in this system of coordinates, the boundary
of the unit ball). However, it is not necessary to use charts to define the Laplacian:
any chart (in particular, projecting onto any hyperplane) gives the same operator via
a change of coordinates.
Let us now describe more precisely the Laplacian and its associated carré du
champ operator and invariant measure in the preceding system of coordinates
x = (x 1 , . . . , x n ) ∈ B. In this chart, the carré du champ operator of the spherical
Laplacian Sn of smooth functions f = f (x) and g = g(x) is given by

n
ij
(f, g) = δ − x i x j ∂i f ∂j g.
i,j =1
The invariant and reversible measure in this system of coordinates is

−1/2
dμ(x) = cn 1 − |x|2 dx = w(x)dx
(where dx is the Lebesgue measure on the unit ball B of Rn and cn > 0 is the
normalization constant which ensures that μ is a probability measure). This mea-
sure μ is therefore, up to some constant, the image of σn under the orthogonal
projection from Sn onto B. The Laplacian Sn itself in this chart takes the form
(compare (1.11.8), p. 46)
1 ij ij
n n n

S n = ∂i wg ∂j = δ − x i x j ∂ij2 − n x i ∂i . (2.2.1)
w
i,j =1 i,j =1 i=1
Having computed Sn in this local system of coordinates, observe that if we

consider x i and x j as the restrictions to the sphere Sn of the coordinates x i and x j
(1 ≤ i, j ≤ n) in the ambient space Rn+1 , that is the restriction of two linear forms
corresponding to two unit orthogonal vectors, then

Sn x i = −nx i , x i , x j = δ ij − x i x j .
Via the chain rule formula, if f is the restriction to the sphere of a smooth function
f (x 1 , . . . , x n+1 ) defined in Rn+1 , then Sn f is the restriction to the sphere of the
quantity

n+1
ij
n+1
δ − x i x j ∂ij2 f − n x i ∂i f. (2.2.2)
i,j =1 i=1
In other words, (2.2.1) is in fact valid in Rn+1 , and not only in the local system of
coordinates. Hence, for explicitcomputations, it is not necessary to replace one of

the coordinates, say x n+1 by ± 1 − ni=1 (x i )2 .
From the representation (2.2.2) of Sn as an operator on Rn+1 , it is not im-
mediate that the associated Markov process lives on the unit sphere. However, the
characterization of Sn may be changed slightly into an operator L in Rn+1 satisfy-
ing

L x i = −nx i , x i , x j = δ ij |x|2 − x i x j , 1 ≤ i, j ≤ n + 1.
This operator obviously coincides with Sn on the unit sphere. Moreover, it may
be observed that for this new operator L(|x|2 ) = 0 and (|x|2 ) = 0. Then, it is quite
immediate that for the diffusion process (Xtx )t≥0 with generator L, |Xtx |2 is constant,
and therefore that the process stays forever on the sphere it started from.
Both the carré du champ operator and the invariant measure μ of the spherical
Laplacian are invariant under the action of the rotations of Rn+1 . This is easily
seen for rotations with vertical axis. Consider the first order differential operators in
Rn+1 , ij = x i ∂j − x j ∂i , 1 ≤ i, j ≤ n + 1, already introduced in Sect. 2.1. Then
[Sn , ij ] = 0. Indeed, a smooth function f is rotationally invariant if and only if
ij f = 0 for every i, j . As already mentioned in the preceding section, the vector
field ij generates a (semi)-group of rotations in the plane (ei , ej ). Since every
rotation is a composition of planar rotations in orthogonal planes, the claim follows.
2.2.2 Stereographic Projection Representation
A second representation of interest is the so-called stereographic projection of Sn

in Rn+1 on the hyperplane Rn . Denote by N the north-pole (that is the point in
Rn+1 with coordinates (0, . . . , 0, 1)), and for every x ∈ Sn , x = N , consider the line
joining N to x. It cuts the horizontal hyperplane at T (x) ∈ Rn . The stereographic
84 2 Model Examples
map x → T (x) is a chart of Sn \ N on Rn . In this chart, the carré du champ operator

takes the form
1 2
(f ) = 1 + |x|2 |∇f |2
4
where |∇f |2 is the usual carré du champ operator associated with the Euclidean
Laplacian of a smooth function f . This formula can be proved by a direct compu-
tation of the Jacobian matrix of the stereographic projection. The invariant measure
in this system of coordinates is written as dμ(x) = cn (1 + |x|2 )−n dx where dx
is Lebesgue measure on Rn and cn > 0 is the normalization constant so that μ is
a probability measure. Once again, μ is the image of σn under the stereographic
projection T . The latter measure μ is sometimes called a Cauchy measure, in anal-
ogy with the one-dimensional case. In the stereographic system of coordinates, the
Laplace operator on Sn becomes
(1 + |x|2 )2 n−2 i n
Sn = − 1 + |x|2 x ∂i (2.2.3)
4 2
i=1
where is the Euclidean Laplacian in Rn . We refer the reader to Sect. 2.3 for
more details about the stereographic projection, viewed there as the restriction to
the sphere of an inversion in Rn+1 .
The preceding orthogonal (2.2.1) and stereographic (2.2.3) projection represen-
tations of the Laplacian on the sphere Sn , regarded as a sub-manifold of Rn+1 , do
keep the same metric and are thus equivalent. One of the interesting properties of
the stereographic projection is that, in this chart, the Euclidean and spherical met-
rics are proportional, in other words conformally equivalent. Generally speaking, a
conformal map from a Riemannian manifold M into itself is a map under which
the metric g of M is transformed to c(x)g for some strictly positive function c(x).
Two metrics g1 and g2 on a given manifold M are conformally equivalent when
g2 (x) = c(x)g1 (x) for some strictly positive function c(x). Conformal maps and
conformally equivalent metrics will play a crucial role in the study of Sobolev in-
equalities on Euclidean space and on the sphere (cf. Sect. 6.9, p. 313). Furthermore,
from a more analytical point of view, uniform ellipticity (that is, the existence of a
c > 0 such that (g ij ) ≥ c(δ ij ) in the sense of symmetric matrices, cf. (1.12.3), p. 50)
is not satisfied in the orthogonal projection representation at the boundary of the
unit ball (the equator of the sphere) whereas it is satisfied there in the stereographic
representation. Conversely, the reverse inequality (g ij ) ≤ c(δ ij ) holds in the projec-
tion representation in the neighborhood of the south-pole but does not hold in the
stereographic projection. Therefore, care must be taken over such properties which
are not invariant by changes of coordinates, and thus are not intrinsic.
There are still many other ways to consider the Laplace operator on the sphere
Sn and we briefly describe below some further examples (these, however, are not
really used later).
First, for Sn embedded into Rn+1 in the usual way, extend any smooth function
f on Sn to a function fdefined on a neighborhood of Sn in Rn+1 , which is indepen-
dent of the radius by setting f(x) = f ( |x|

x
). Then, if
h = ffor the usual Laplacian
in Rn+1 , h = Sn f where h is the restriction of h to Sn .
The connection between the usual Laplacian on Rn+1 and the spherical Laplacian
Sn may also be clarified in polar coordinates. Indeed, if a point x ∈ Rn+1 \ {0} is
parametrized by (r, θ ) where r = |x| is the Euclidean norm and θ = |x| x
∈ Sn , then
the Euclidean Laplacian = Rn+1 may be written as
n 1
Rn+1 = ∂r2 + ∂r + 2 Sn . (2.2.4)
r r
There is still another representation, more intrinsic in view of Lie group actions.
Consider the sphere Sn as the quotient space SO(n + 1)/SO(n), where SO(n + 1)
is the special orthogonal group in Rn+1 (SO(n) is then regarded as the subgroup of
SO(n + 1) which leaves the point (1, 0, . . . , 0) invariant). Recall the infinitesimal
rotations in Rn+1 , ij = x i ∂j − x j ∂i , 1 ≤ i, j ≤ n + 1. These vector fields preserve
functions
which are independent of the radius (since they commute with the oper-
ator n+1i=1 x i ∂ ). Now, for a function f on Sn given as the restriction to Sn of a
i
smooth function defined in a neighborhood of Sn in Rn+1 , the operator Sn may be
represented as

Sn f = 2ij f. (2.2.5)
1≤i<j ≤n+1
In this representation, observe that, as for the usual Laplace operator on Rn , Sn is
given as the sum of squares of vector fields which commute with it, although they
do not commute with each other. Moreover, many more vectors than the dimension
of the space have to be used. This is an example of a Casimir operator on a homo-
geneous space. Such operators play a fundamental role in the analysis of compact
Lie groups.
2.2.3 Spherical Heat Kernel
According to the general theory presented in Chap. 3, the spherical Laplacian, con-
sidered for example on the class of smooth (C ∞ ) functions on Sn , defines the gen-
erator of the so-called spherical heat or Brownian semigroup (Pt )t≥0 on the sphere
Sn . The Markov process associated with this operator is called the spherical Brow-
nian motion. The spherical heat semigroup admits kernel densities pt (x, y), t > 0,
(x, y) ∈ Sn × Sn , with respect to the invariant measure σn (cf. Definition 1.2.4,
p. 14). Since the semigroup commutes with rotations on the sphere, and therefore
with any rotation R, the heat kernels satisfy pt (x, y) = pt (Rx, Ry) (t > 0). But for
any two pairs (x, y) and (x , y ) of points on the sphere, there exists a rotation R
such that (x , y ) = (Rx, Ry) if and only if x · y = x · y . Using the intrinsic distance
on the sphere defined from the Laplacian (see Sect. C.4, p. 509, or (3.3.9), p. 166),
which in this case is d(x, y) = arccos(x · y), (x, y) ∈ Sn × Sn , pt (x, y) may be ex-
86 2 Model Examples
pressed as a function of the distance d(x, y) (and of t). Unfortunately, no explicit

value for this function is easy to handle, but it may be expressed in terms of the heat
kernels for Jacobi operators (see Sect. 2.7.4 below).
The harmonic analysis of the spherical Laplacian Sn is classical. The eigen-
vectors of Sn are the restrictions to the sphere Sn of the harmonic polyno-
mials (in Rn+1 ) homogeneous of degree k, and the eigenvalues, of −Sn , are
λk = k(k + n − 1), k ∈ N. In particular, the eigenvectors U associated with the
first non-trivial eigenvalue n are the restrictions to the sphere of the linear maps in
the ambient Euclidean space Rn+1 . Actually, they satisfy ∇∇U = −U Id (here and
in what follows Id = (g ij )) and the eigenvalue n is obtained after taking the trace.
The polar coordinate representation (2.2.4) of the Euclidean Laplace operator
from the spherical representation actually leads to a representation of harmonic
functions in the unit ball with given boundary value f on the sphere via subordina-
tion (cf. Sect. 1.15.9, p. 67). Indeed, setting r = e−t in (2.2.4), the Laplace operator
in Rn+1 is given by

e2t ∂t2 − (n − 1)∂t + Sn .
Using the subordinator given in (1.15.2), p. 69, with α = n−1
2 , the harmonic function
H in the ball with boundary value f in polar coordinates (r = e−t , x ∈ Sn ) takes the
form

H e−t , x = f (y)qt (x, y)dσn (y)
Sn
where
∞
(n−1)/2,1/2
qt (x, y) = ps (x, y) dνt (s), t > 0, (x, y) ∈ Sn × Sn .
0
While there are no simple expressions for heat kernels on spheres, the latter ker-
nel qt (x, y) is on the other hand quite simple and is given by the celebrated Poisson
formula. Indeed, it is classical that the harmonic extension H to the unit ball in Rn+1
of a function f : Sn → R may be represented as

1 − |z|2
H (z) = f (y)dσn (y). (2.2.6)
Sn |z − y|n+1
This formula is quite easy to check. Observe first that the map
1 − |z|2
z →
|z − y|n+1
is harmonic in the open unit ball B. Furthermore, for any z ∈ B, the measure
1−|z|2
ν(z, dy) = |z−y| n+1 dσn (y) is a probability measure on S . This is a consequence
n

of the harmonic property since the function m(z) = Sn ν(z, dy) is harmonic in B,
constant on any sphere of radius r < 1 (due to the rotation invariance of σn ) and
satisfies m(0) = 1, so that the constant value on spheres of radius r is 1. It remains

as z → z0 ∈ S , ν(z, dy) converges to the Dirac mass at z0 . Then,
to observe that n
the function Sn f (y)ν(z, dy) is harmonic on the ball, and for f continuous on Sn ,
converges to f (z0 ) when z converges to z0 ∈ Sn . This is exactly what is expected
from the Poisson kernel, proving (2.2.6).
According to the Poisson formula, in polar coordinates (r = e−t x, x ∈ Sn ), the
density kernel qt (x, y) with respect to σn of the measure ν(z, dy) may be written
1 − e−2t
qt (x, y) = .
|e−t x − y|n+1
Writing the harmonic extension of f as Qt f , the subordination representation ex-
presses that, as a semigroup,

(n − 1)2 n − 1
Qt = exp −t −Sn + − , t ≥ 0.
4 2
Applying Qt to an eigenvector Uk on the sphere with eigenvalue λk = k(k + n − 1),

we get Qt Uk = e−kt Uk since

(n − 1)2 n − 1
λk + − = k.
4 2
Hence Qt Uk = r k Uk (since r = e−t ) and we recover that Uk is indeed the restriction

to the sphere of a harmonic polynomial of degree k in Rn+1 .
The representation (2.2.6) of the Poisson kernel will be most useful for example
when dealing with Jacobi operators and polynomials (see Sect. 2.7.4 below). In
the next section, similar formulas will be considered for the hyperbolic Laplacian
(see (2.3.2)), in particular with recurrence identities in the dimension. Actually, a
similar recurrence (2.7.15) holds in the sphere case, as will be developed with the
tools of Jacobi operators.
2.2.4 Curvature-Dimension Condition
In the last part of this section, we describe the geometric and curvature features of
the sphere and its Laplacian. As the sphere Sn of dimension n is of constant cur-
vature n − 1, the Laplace operator Sn satisfies the curvature-dimension condition
CD(n − 1, n)
1
2 (f ) ≥ (n − 1) (f ) + (Sn f )2
n
(for all f , say, in the algebra A of smooth functions) from Definition 1.16.1, p. 72,
with optimal values (cf. Sect. C.6, p. 513). Such a curvature-dimension condition is
similar to those of the one-dimensional Jacobi operators investigated in Sect. 2.7.4
88 2 Model Examples
below which may actually be viewed as projections on a diameter of the spherical

Laplacian. As discussed there, curvature-dimension conditions for one-dimensional
generators are analyzed via (1.16.9), p. 73.
The status of n as a dimension in the curvature-dimension condition
CD(n − 1, n) may however be somewhat further analyzed. Starting from the pro-
jection representation (2.2.2) of Sn (defined on the unit ball), projecting it on
a vector subspace of dimension n0 < n (that is, letting Sn act on functions de-
pending only on the n0 first coordinates) yields the same operator but this time on
the unit ball of dimension n0 . In particular, this operator may thus be written as
n0 i
Sn0 − (n − n0 ) i=1 x ∂i where Sn0 is the spherical Laplacian of dimension n0
(in orthogonal projection). It is readily checked that, in these coordinates,

n0
− x i ∂i f = log 1 − |x|2 , f .
i=1

But now, the function U = 1 − |x|2 is nothing else than the restriction to the half-
upper sphere of the first coordinate. It is therefore an eigenfunction associated to the
first eigenvalue, so that the operator Sn0 can be decomposed as
Sn0 + (n − n0 )∇ log U
where ∇∇Sn0 U = −U Id. In particular, with Z = (n − n0 )∇ log U ,
1
Ric −∇S Z − Z ⊗ Z = (n − 1) Id .
n − n0
Comparing with (1.16.7), p. 73, observe that there is equality between tensors. Thus,
here we have found a fundamental example of operators satisfying the curvature-
dimension condition CD(n − 1, n) in an optimal way but for which n > n0 is not the
topological dimension of the state space. The latter operator will be further analyzed
in Sect. 6.9, p. 313, in connection with Sobolev-type inequalities.
2.3 Hyperbolic Heat Semigroup
The third model example is the heat semigroup on hyperbolic space. In the core of
this book, this model will not be used as often as the preceding models in Euclidean
and spherical spaces. So its description here will be somewhat more sketchy. We
start again with a geometric description of the underlying state space. Various rep-
resentations of the hyperbolic metric are available and here we emphasize two of
them.
2.3 Hyperbolic Heat Semigroup 89
2.3.1 Upper Half-Space Representation
The first one is the upper half-space representation given by

2
g ij = x n δ ij , 1 ≤ i, j ≤ n,
on E = Rn−1 × (0, ∞). As in the preceding section, we use here upper indices
in accordance with the Riemannian geometry convention, putting emphasis on the
co-metric. It is not necessary here to change charts since this one covers at once
the entire manifold. The resulting manifold, called the hyperbolic space Hn (in the
upper half-space representation), is not compact. In this representation, the carré du
champ operator may be written as
2
(f ) = x n |∇f |2
and the invariant reversible measure is dμ(x) = (x n )−n dx (beware of the notation
here, x n is the n-th coordinate and (x n )−n denotes its −n power). Here, as usual,
|∇f |2 is the carré du champ operator of a smooth function f on Rn for the usual
Laplacian Rn and dx is the Lebesgue measure. The associated hyperbolic Lapla-
cian, denoted Hn , is given in this representation by
2
Hn = x n Rn − (n − 2)x n ∂n . (2.3.1)
It should be pointed out that the hyperbolic metric degenerates at the bound-
ary {x n = 0} of the upper half-space E = Rn−1 × (0, ∞). However, the descrip-
tion (2.3.1) is enough to define a unique symmetric semigroup with infinitesimal
generator Hn . Following the developments in Chap. 3, for an elliptic operator on
a manifold, knowledge of its action on the class of smooth compactly supported
functions is enough to describe a unique symmetric semigroup with this operator as
infinitesimal generator (this is the issue of essential self-adjointness) as soon as the
manifold is complete (see Proposition 3.2.1, p. 142, and Corollary 3.2.2, p. 143).
Now the distance on E induced by Hn (see Sect. C.4, p. 509) is the standard hy-
perbolic Riemannian metric on E, and this metric is complete (the boundary is at an
infinite distance with respect to it). Therefore, the operator Hn is essentially self-
adjoint on E and is the infinitesimal generator of a unique symmetric semigroup
called the hyperbolic heat or Brownian semigroup. Moreover the operator satisfies
the curvature-dimension condition CD(−(n − 1), n) (see below), hence this unique
semigroup is indeed Markov.
From a probabilistic viewpoint, the heat semigroup on Hn is the Markov semi-
group (up to the probabilistic normalization of the Laplacian) associated with the
process solving the stochastic differential equation
dXti = Xtn ◦ dBti − (n − 1)Xtn dt, 1 ≤ i ≤ n,
(in Stratonovich form). It is easily seen that this process does not reach the bound-
ary {x n = 0} in finite time and therefore that the semigroup is indeed Markov. The
Markov process (Xt )t≥0 thus constructed is called hyperbolic Brownian motion.
90 2 Model Examples
The hyperbolic heat semigroup admits density kernels pt (x, y), t > 0,
(x, y) ∈ Hn × Hn , with respect to the invariant measure. They are expressed in terms
of the intrinsic distance associated with hyperbolic Laplacian Hn (see Sect. C.4,
p. 509 or (3.3.9), p. 166). In the upper half-space representation E = Rn−1 × (0, ∞),
the distance between (x1 , y1 ) and (x2 , y2 ) is given by

|x1 − x2 |2 + y12 + y22
cosh−1 .
2y1 y2
The density kernels pt (x, y) in dimension n are then expressed as functions

kn (t, d(x, y)) of the distance, where kn (t, d) may be defined by induction on n ≥ 2
by
√ ∞
se−s /4t
2
2 −t/4
k2 (t, d) = e ds
(4πt)3/2 d (cosh(s) − cosh(d))1/2

1 d d2
k3 (t, d) = exp −t −
4πt)3/2 sinh(d) 4t
and
e−nt
kn+2 (t, d) = ∂d kn (t, d). (2.3.2)
2π sinh(d)
The expressions take a different form according as n is even or odd (and are sim-
pler for odd n’s) while becoming increasingly more complicated as n increases.
A similar recurrence formula for the heat kernels on spheres will be given below
in (2.7.15) after the appropriate analysis of the corresponding Jacobi operators and
their expansions in Jacobi orthogonal polynomials.
As for the sphere, the hyperbolic metric is conformally equivalent to the Eu-
clidean metric. But now the invariant measure is infinite. The Riemann curvature
tensor of the metric is constant, as is the Ricci curvature, which is equal to −(n − 1).
(We have actually described in these three sections the only three metrics and spaces
with this property, the Euclidean space, the sphere and the hyperbolic space.) It may
be checked directly from the definition of Hn that it satisfies a curvature-dimension
condition CD(−(n − 1), n) in the sense of Definition 1.16.1, p. 72.
The Laplace operator Hn on hyperbolic space is invariant under rotations
around the axis of the last coordinate vector en , as well as under translations par-
allel to the hyperplane {x n = 0}. According to Sect. 1.15.4, p. 60, the associated
Markov semigroup leaves invariant the set of functions depending only on the last
coordinate, and gives rise along this coordinate to a one-dimensional semigroup
on (0, ∞) with generator x 2 ∂x2 − (n − 2)x∂x . After a change of variable setting
x = ey , the latter operator takes the simpler form ∂y2 − (n − 1)∂y . In particular
(y) = 1, the function y (which may be seen as the distance from infinity) hav-
ing gradient 1. By (1.16.9), p. 73, this one-dimensional operator still satisfies the
CD(−(n − 1), n) condition, and moreover there is equality in the differential in-
equality (1.16.9), which characterizes this condition in dimension one. Therefore,
2.3 Hyperbolic Heat Semigroup 91
the geometric properties of the Laplace operator of hyperbolic space may be recov-
ered from one-dimensional projections. Moreover, from the latter description, the
function h(x) = (x n )n−1 satisfies Hn h = 0 and therefore defines a harmonic posi-
tive function. The measure hdμ is invariant (although not reversible) for Hn , and
in particular there exists in this way at least one invariant measure different from the
reversible one (many such invariant measures exist, this is just one of them).
2.3.2 Open Ball Representation
Like the sphere Sn , the hyperbolic space Hn admits another representation on the
open unit ball B of Rn . In order to describe this, we return to the stereographic
projection, but this time from Sn−1 to the hyperplane {x n = 0} identified with Rn−1 .
This transformation is an inversion. In the Euclidean space Rn , the inversion with
center x0 ∈ Rn and radius r > 0 is the map which associates to each x = x0 its
inversion x ∈ Rn defined by the condition that x − x0 and x − x0 are proportional
and |x − x0 ||x − x0 | = r 2 . Analytically, it is given by
x − x0
ϕx0 ,r : x → x0 + r 2 .
|x − x0 |2
The sphere with center x0 and radius r is clearly stable under this transformation
(which is equal to the identity on it). This sphere is called the sphere of the inver-
sion ϕx0 ,r . It is only defined for x = x0 and it is an involution (ϕx20 ,r = Id). Actually,
the inversion ϕx0 ,r sends every sphere not containing x0 to sphere, and every sphere
containing x0 to a hyperplane not containing x0 . Similarly, it transforms every hy-
perplane containing x0 into a hyperplane containing x0 , and every hyperplane not
containing x0 into a sphere. Moreover, ϕx0 ,r preserves all the spheres orthogonal to
the sphere of inversion (two spheres are orthogonal if at any intersection point x,
the radii joining x to the centers of the spheres are orthogonal).
Now, the stereographic projection with pole N (north-pole) is actually√the re-
striction to the unit sphere Sn−1 of an inversion with center N and radius 2 (this
inversion clearly preserves the intersection of the unit sphere with the horizontal
hyperplane). But it also sends the upper half-space {xn > 0} onto the unit ball of
Rn . Via this transformation, the hyperbolic Laplacian Hn of (2.3.1) becomes an
operator on the open unit ball B with carré du champ operator
1 2
(f ) = 1 − |x|2 |∇f |2
4
and reversible measure cn (1 − |x|2 )−n dx. Actually, the inversions are conformal
maps of the Euclidean space, and it is not so surprising that the image under an in-
version of a metric conformally equivalent to the Euclidean metric is again confor-
mally equivalent to the Euclidean metric. In this system of coordinates, the operator
92 2 Model Examples
Hn becomes the operator defined on B as
1 2 n − 2 i n
Hn = 1 − |x|2 Rn + 1 − |x|2 x ∂i . (2.3.3)
4 2
i=1
This representation is quite similar to the stereographic representation (2.2.3) of the

spherical Laplacian Sn .
The latter representation (2.3.3) of the hyperbolic Laplacian gives rise to pro-
jections rather similar to the first representation on the upper half-space. Indeed,
in the previous form, the Laplacian Hn commutes with rotations centered at the
origin. Therefore, following Sect. 1.15.4, p. 60, it preserves radial functions. Setting
sinh(y/2)
|x| = cosh(y/2) , the operator acting on radial functions becomes the one-dimensional
operator
cosh(y)
∂y2 + (n − 1) ∂y .
sinh(y)
Again (y) = 1, and y is in fact the distance (in the hyperbolic metric) to the center
of the unit ball. The latter operator is therefore Hn acting on functions depending
only on this distance. Using once more (1.16.9), p. 73, this one-dimensional operator
satisfies the curvature-dimension condition CD(−(n − 1), n), with equality in the
differential inequality (1.16.9) characterizing this condition.
In the unit ball representation, it may be further observed that Hn is also in-
variant under rotations preserving the center of the unit ball. Going back and forth
between the upper half-space representation and the open unit ball representation
of Hn , we may therefore construct many such invariant transformations (indeed
the group of isometries of the hyperbolic space). Moreover, the harmonic function
h(x) = (x n )n−1 described in the upper half-space may be seen in the ball as a har-
monic function vanishing on the boundary except at one point. Using rotations, to
any point on the boundary of the ball may be associated a harmonic strictly positive
function, and therefore a new invariant measure. In this context, inversions assume
the role that is played by rotations in the case of the sphere. In the unit ball repre-
sentation of Hn , the inversions with respect to spheres orthogonal to the unit sphere
are transformations from the unit ball into itself. These inversions are isometries of
the hyperbolic space, and the Laplacian Hn is invariant under these inversions. It
is of interest to look for the vector fields on the ball corresponding to this invariance
which commute with Hn .
2.4 The Heat Semigroup on a Half-Line and the Bessel

Semigroup
Before addressing more general Sturm-Liouville operators and semigroups in

Sect. 2.6, we briefly discuss in this and the next section the heat semigroup on a
half-line and on a bounded interval of the real line. These examples will provide an
2.4 The Heat Semigroup on a Half-Line and the Bessel Semigroup 93
opportunity to introduce the Neumann and Dirichlet boundary conditions, which are
discussed more generally in Sect. 2.6, to address the issue of self-adjointness and to
provide intuitive probabilistic descriptions in terms of reflected and killed Brownian
motion.
2.4.1 The Heat Semigroup on a Half-Line
Consider the operator Lf = f on (0, ∞) acting on the set Cc∞ (0, ∞) of smooth
and compactly supported functions f on (0, ∞). As this operator is symmetric with
respect to the Lebesgue measure dx, we may look for a symmetric semigroup for
which Cc∞ (0, ∞) is included in the domain of its generator and for which this gen-
erator coincides with L on this class of functions. Note that while we considered
L on (0, ∞), we will see that the associated Markov process may in fact live in
R+ = [0, ∞) and that we will have to consider the natural state space on which the
semigroup lives to be [0, ∞) instead of (0, ∞). Whether we regard the space state to
be (0, ∞) or [0, ∞) is more a matter of taste, and in any case the boundary behavior
will have to be examined. This observation is relevant for most examples studied
here and in the next sections.
A semigroup with generator L in this setting is not unique. Indeed, a bounded
measurable function f defined on (0, ∞) may be extended in at least two dif-
ferent ways to the whole real line R. It may actually be extended to a symmet-
ric function fˆ (that is fˆ(−x) = fˆ(x)) or to an anti-symmetric function fˇ (that is
fˇ(−x) = −fˇ(x)), its value at 0 will not matter. Then, if (Pt )t≥0 is the heat semi-
group on R, Pt fˆ is symmetric, while Pt fˇ is anti-symmetric. Setting PtN f = Pt fˆ
and PtD f = Pt fˇ, taking the restriction to R+ = [0, ∞) yields two different semi-
groups (PtN )t≥0 and (PtD )t≥0 on R+ . It is easily seen that both semigroups are
symmetric with respect to the Lebesgue measure, and that they are positivity pre-
serving. (PtN )t≥0 and (PtD )t≥0 admit simple kernel densities via the standard heat
kernel (2.1.1) (on the line) given, for t > 0 and (x, y) ∈ (0, ∞) × (0, ∞), by
1
ptN (x, y) = pt (x, y) + pt (x, −y)
2
and
1
ptD (x, y) = pt (x, y) − pt (x, −y) .
2
The semigroup (PtN )t≥0 is Markov while (PtD )t≥0 is only sub-Markov (PtD (1) ≤ 1).
For any function f on (0, ∞), both semigroups u(t, x) = PtN,D f solve the heat
equation ∂t u = Lu on R+ × (0, ∞) and thus have L as infinitesimal generator on
Cc∞ (0, ∞). In conclusion, Cc∞ (0, ∞) is not a core of the domain of L, or equiva-
lently L is not essentially self-adjoint on this set (cf. Sect. A.5, p. 481, Sect. 1.6,
p. 24, and Sect. 1.12, p. 49).
94 2 Model Examples
For any bounded measurable function f on (0, ∞), and for any t > 0, PtN f is
a smooth function, symmetric, so that its derivative vanishes at x = 0. The semi-
group (PtN )t≥0 is called the semigroup with infinitesimal generator L and Neumann
boundary conditions (derivatives vanish at the boundary). On the other hand, PtD f
is also smooth, anti-symmetric and therefore vanishes at x = 0. The semigroup
(PtD )t≥0 is then called the semigroup with infinitesimal generator L and Dirichlet
boundary conditions (functions vanish at the boundary).
The Neumann and Dirichlet semigroups have clear probabilistic descriptions in
terms of the associated Brownian motion. Letting {B tx ; t ≥ 0, x ∈ R} be Brownian
motion with speed 2 on the line, it is easily shown that for every suitable function
f : (0, ∞) → R, and every t ≥ 0, x ∈ R+ ,
x
PtN f (x) = E f Bt .
The family (PtN )t≥0 is the semigroup of Brownian motion reflected at x = 0. On

the other hand,
x
t 1{t<T }
PtD f (x) = E f B
where T is the first time s ≥ 0 for which B s = −x). The latter
sx = 0 (that is B
assertion is actually not so immediate. In order to see why this is, observe that
F (t, x) = PtD f (x), t ≥ 0, x ∈ R, satisfies ∂t F = ∂x2 F and vanishes at x = 0. Then,
by Ito’s formula (Theorem B.2.1, p. 491), for any t > 0,

Ms = F t − s ∧ t, B s∧t
x
, 0 ≤ s ≤ t,
is a martingale, and hence E(MT ∧t ) = E(M0 ). Now, by definition of the stopping

tx ) = f (B
time T , if T ≤ t then F (t − T , B̃Tx ) = 0 while when T > t, F (0, B tx ) from
D
which the conclusion follows. Therefore (Pt )t≥0 is the semigroup of Brownian
motion killed at x = 0.
2.4.2 The Bessel Semigroup
In the probabilistic realm, another operation on standard Brownian motion is the

h-transform as described in Sect. 1.15.8, p. 66, which yields the law of Brownian
motion conditioned to stay ever positive. The positive harmonic function h on the
half-line R+ which vanishes at x = 0 is just h(x) = x. Performing the corresponding
h-transform turns the generator Lf = f into the new generator
2
Lf (x) = f (x) + f (x)
x
(for smooth functions f on (0, ∞)). This operator belongs to the family of so-called
Bessel semigroups, associated with Bessel processes with parameter 2. These pro-
cesses naturally appear when looking at radial parts of Brownian motion in dimen-
sion n. Indeed, since the standard Euclidean Laplacian commutes with rotations, it
2.4 The Heat Semigroup on a Half-Line and the Bessel Semigroup 95
preserves the class of radial functions. As described in Sect. 1.15.4, p. 60, if r de-
notes the function |x|, then (r 2 ) = 2n while (r 2 ) = 4r 2 . Hence, for any smooth
function f compactly supported in (0, ∞)
n−1
f (r) = f (r) + f (r).
r
(Compare with (2.2.4).) The image operator on (0, ∞) is thus Lf = f + n−1
x f .
The family of operators
κ
LBκ f = f + f
x
on (0, ∞) is known as the family of Bessel operators with parameter κ. Here we
restrict our attention to the case κ ≥ 0. The associated reversible measure is x κ dx
on (0, ∞). By means of the harmonic function hκ (x) = x 1−κ (= log x when κ = 1),
it is easily seen that the Markov process with generator LBκ starting from x > 0
never reaches the boundary 0 as soon as κ ≥ 1. A more precise explanation of why
this is the case will be given in Sect. 2.6 below.
Now, one may naively think that if a process does not reach the boundary, then
its generator is essentially self-adjoint, since different symmetric extensions cor-
respond to various boundary behaviors of the process. The above example of LB2
shows that this is not the case. Indeed, performing the h-transform procedure with
h(x) = x for the semigroups (PtN )t≥0 and (PtD )t≥0 of respectively reflected and
killed Brownian motions on (0, ∞) yields two semigroups with generator LB2 on
(0, ∞), while the process driven by the associated stochastic differential equation
does not reach the boundary. (Observe that the h-transform of the killed Brow-
nian motion produces a Markov semigroup, while the transform of the reflected
Brownian motion produces a semigroup (Pt )t≥0 which only satisfies Pt (1) ≥ 1.)
Moreover, the h-transform of LBκ yields the Bessel operator LB2−κ . Hence, follow-
ing Proposition 2.4.1 below, LBκ is essentially self-adjoint as soon as κ > 2. Since
h-transforms preserve such properties, LBκ is also self-adjoint in the range κ < 0.
In this setting, the associated process has a drift which pushes it to the boundary so
strongly that, once it has reached 0, there is no way to return to the open set (0, ∞).
This kind of duality between LBκ and LB2−κ will also be observed for Laguerre
semigroups (Sect. 2.7.3) with similar conclusions.
The value which appears as a limiting case for essential self-adjointness for κ ≥ 0
is thus κ = 2. The following, somewhat more general, Proposition 2.4.1 provides a
useful criterion to ensure that a given generator L on the half-line is essentially self-
adjoint, that is the space Cc∞ (0, ∞) of C ∞ compactly supported functions in (0, ∞)
is dense in the domain D(L) of L. As announced, it shows in particular that, as soon
as κ > 2, the Bessel generator LBκ defined on Cc∞ (0, ∞) is essentially self-adjoint.
Proposition 2.4.1 Let Lf = f + a(x)f be defined on Cc∞ (0, ∞), where a is a

smooth function on (0, ∞). Then the operator L is symmetric with respect to the
measure dμ = eA dx where A = a. Moreover, as soon as there exist two constants
2
c, C > 0 such that a (x) + a 2(x) ≥ xc2 − C, x > 0, then L is essentially self-adjoint.
96 2 Model Examples
Proof We briefly outline the arguments. The fact that L is symmetric with respect
to dμ = eA dx is immediate (see Sect. 2.6). Remove then the gradient in L accord-
ing to the technique described in Sect. 1.15.7, p. 65. The problem is reduced to
2
proving that if K = a2 + a4 then the operator L1 f = f − Kf is essentially self-
adjoint on (0, ∞) with respect to the Lebesgue measure. To this end, according
to Proposition A.5.3, p. 482, it is enough to show that for some λ ∈ R, the equa-
tion f = (λ + K)f (understood in the distributional sense) has no solution in
L2 (dx) = L2 ((0, ∞), dx) except 0. By the hypothesis, λ may be chosen so that
λ + K > K0 (x) for some ε > 0 where K0 (x) = xε2 . Any solution f on (0, ∞) of
f = (K + λ)f is as smooth as K. Assuming that f is not identically 0, up to a
sign change, let f (x0 ) > 0 for some x0 > 0. Now, if f (x0 ) > 0, it is easy to see
from f ≥ K0 (x)f that f is increasing on (x0 , ∞), and is therefore convex on this
interval. Being convex it grows at least linearly at infinity and therefore is not in
L2 (dx). On the other hand, if f (x0 ) < 0, from standard arguments, f is bounded
from below by the solution f0 of f0 = K0 f0 which has the same value and same
derivative at x0 . To verify that f is not in L2 (dx), it is therefore enough to show that
f02 is not integrable near 0. But the solutions of f0 = K0 f0 are linear combinations
of x α1 and x α2 where α1 and α2 are solutions of α(α − 1) = ε. Since f0 (x0 ) < 0,
√
f0 behaves like βx −α1 near the origin, with β > 0 and −2α1 = 1 + 1 + 4ε. The
conclusion then easily follows and the proposition is established.
2.5 The Heat Semigroup on the Circle and on a Bounded

Interval
Considering the unit circle S1 = {z ∈ C ; |z| = 1} via the parametrization z = e2iπx ,

a function on S1 is simply a periodic function on R with period 1. A function f˜
on S1 is thus identified with the 1-periodic function f (x) = f˜(e2iπx ). While peri-
odic functions are never integrable with respect to the Lebesgue measure on R, the

1
Lebesgue measure on S1 may be introduced by S1 f˜(z)dz = 0 f (x)dx. If f and
g are smooth and 1-periodic, then
1 1 1

f g dx = f g dx = − f g dx
0 0 0
since via integration by parts the boundary terms vanish.

On the real line, the standard heat semigroup (Pt )t≥0 commutes with translations.
Therefore, if f is periodic with period 1, so is Pt f for every t ≥ 0, and it is also
smooth when t > 0. Hence, the standard heat semigroup acting on periodic func-
p
tions induces a Markov semigroup (Pt )t≥0 on S1 . To describe it more precisely,
note first that for any periodic function f , one has by definition ∂t Pt f = (Pt f ) .
p p
Then, this semigroup has the Lebesgue measure on S as invariant measure. All
1
p
these considerations show that the above semigroup (Pt )t≥0 on S1 is just the heat
2.6 Sturm-Liouville Semigroups on an Interval 97
semigroup with infinitesimal generator Lf = f acting on (smooth) periodic func-

p
tions. Observe that (Pt )t≥0 commutes with rotations (which are nothing else but
translations on the real line via the preceding identification).
Any function f on [0, 1] may extended by periodicity to a new function fˆ on
the whole real line R. The restriction to [0, 1] of the action Pt fˆ of the standard heat
p
semigroup on the line is precisely Pt f . This construction yields a useful represen-
p p
tation of the kernel densities pt of the semigroup (Pt )t≥0 on S1 as
p

pt (x, y) = pt (x, y + k), t > 0, (x, y) ∈ [0, 1] × [0, 1].
k∈Z
Unfortunately, this representation as a sum of a series does not admit a closed form.
The preceding analysis may be pushed a bit further on the basis of the proper-
ties of the standard heat semigroup (Pt )t≥0 on the real line to further illustrate the
Neumann and Dirichlet boundary conditions emphasized in the previous section.
Indeed, any bounded measurable function f on [0, 1] may be extended by symme-
try to [−1, +1], and then by periodicity with period 2. The resulting function fˆ
is invariant under the symmetries about 0 and 1. For every t ≥ 0, Pt fˆ is a func-
tion on R which shares the same symmetries. Since a smooth function on R which
is symmetric under those symmetries has zero derivatives at 0 and 1, the result-
ing semigroup on [0, 1] corresponds to the Neumann boundary conditions. In the
same way, any function f defined on [0, 1] may be extended by anti-symmetry at
0 and then by periodicity. The resulting function fˇ is anti-symmetric about x = 1,
and so is Pt fˇ. The associated semigroup on [0, 1] then corresponds to the Dirichlet
boundary conditions.
2.6 Sturm-Liouville Semigroups on an Interval
In this section, we turn to the analysis of the family of diffusion operators on an

interval (bounded or unbounded) of the real line. This study extends, in the same
language, the previous examples of the heat semigroups corresponding to Neumann,
Dirichlet or periodic boundary conditions. The investigation of this family is quite
accessible, but nevertheless already indicates some of the main difficulties in the
study of general diffusion operators and semigroups.
2.6.1 Sturm-Liouville Operators
To fix the ideas, we begin with the case of a bounded interval of the real line R, say
[−1, +1]. Choose a generator L with smooth coefficients of the form
Lf = af + bf (2.6.1)
98 2 Model Examples
where a and b are smooth on [−1, +1]. A natural class of functions f on which L
acts is the family of smooth functions on [−1, +1] (equivalently the restrictions to
[−1, +1] of smooth functions defined on a neighborhood of [−1, +1]). Although it
is of interest and often necessary to consider coefficients which are unbounded at the
boundary of the interval (as for example in Sect. 2.4 or Sect. 2.7.4), we assume here
for simplicity of exposition that a and b are bounded on the whole closed interval
[−1, +1]. An immediate computation shows that the carré du champ operator
associated with the generator L of (2.6.1) is given, on smooth functions f and g, by
(f, g) = af g .
Since (f ) = (f, f ) = af 2 is always positive, we have to choose a ≥ 0. To sim-

plify the matter, we assume furthermore that a > 0 on [−1, +1], so that the operator
L is elliptic (Sect. 1.12, p. 49).
To investigate the operator L of (2.6.1), in order to simplify the expression we
first perform a change of variables. To this end, set a = α 2 where α > 0, and look
for y(x) such that dy = α(x)
dx d
. In other words, dy = α dx
d
. Hence, every function f (x)
g (y)
is of the form g(y(x)), and f (x) = α(x) . Therefore,
d2 d2 α d
2
= α2 2 − · .
dy dx α dx
In the new variable y, the operator L thus takes the form ∂y2 + c(y) ∂y and the carré
du champ operator is the usual (f ) = f 2 . Observe also that if α vanishes at one
of the boundaries of [−1, +1], the new interval on which varies may be unbounded
(although we assumed here for simplicity that this is not the case). For further pur-
poses, note that this change of variables is actually only possible in dimension one
(since there is only one possible metric up to a multiplicative function). Another
feature of dimension one is that the invariant measure is always reversible as well
as explicit (since every vector field is a gradient).
After this change of variable, we may again rescale the state space as to interval
[−1, +1] and switch back to the x variable, so that the operator now takes the form
Lf = f + c(x)f (2.6.2)
where c is a smooth function on [−1, +1]. In this form, the invariant measure is
given, up to a multiplicative constant, by its density
x
w(x) = exp c(y)dy ,
x0
the initial point x0 being a normalization constant for the underlying Markov pro-
cess. Denote by μ the probability measure with (normalized) density w with respect
to the Lebesgue measure (if the coefficients are all bounded, then the invariant mea-
sure is indeed finite).
2.6.2 Boundary Conditions
The description of the domain D(L) of the Sturm-Liouville operator L (2.6.2) is

a more delicate question. In contrast to what happens when the drift term c(x) is
infinite at the boundary, as in Proposition 2.4.1, the space of smooth compactly sup-
ported functions (in the interval [−1, +1]) is never dense in the domain when c(x)
is bounded, and knowing how the generator behaves in the interior of the interval is
not enough to fully describe the associated semigroup. As for the examples of the
two last sections, it is necessary to describe the boundary behavior. If the operator
has to be symmetric, for all functions f, g in some domain A0 to be made precise,
+1 +1
f Lg dμ = g Lf dμ
−1 −1
(where we recall that dμ = wdx). But for bounded (smooth) functions f and g, the
integration by parts formula indicates that
+1 +1 +1
f Lg dμ = − (f, g) dμ + f g w −1
−1 −1
and
+1 +1
0= (f Lg − g Lf )dμ = f g − f g w −1 .
−1
It is thus necessary to choose a domain A0 on which the boundary terms vanish

for functions f, g ∈ A0 . This determines the so-called “boundary conditions”. As
already observed for the heat semigroup on the half-line in Sect. 2.4.1, two natu-
ral choices arise, either the functions or their derivatives vanish. As also discussed
there, in probability theory, there are two classical boundary behaviors for diffusion
processes. Either the process is killed on the boundary, or it is reflected. (There are
of course other possibilities, such as jumping inside the interior following a law de-
pending on the exit point, as well as other options which shall not be considered
here.)
The so-called Dirichlet boundary conditions correspond to the case when the
semigroup is stable on functions vanishing at the boundary. The associated Markov
process is therefore killed on the boundary. Processes killed on the boundary are
rather easy to describe. Indeed, given the solution (Xt )t≥0 of the stochastic differ-
ential equation
√
dXt = 2 dBt + c(Xt )dt
inside some open domain O, starting from x ∈ O, consider the stopping time
T = inf{s ≥ 0 ; Xs ∈
/ O}. Then define the semigroup killed at the boundary by

PtD f (x) = Ex f (Xt )1{t<T } , t ≥ 0, x ∈ O. (2.6.3)
100 2 Model Examples
The reasoning which identified this semigroup with the semigroup with Dirichlet
boundary conditions on the half-line (Sect. 2.4.1) may be repeated here with no
change.
The Neumann boundary conditions correspond to the case when the semigroup
(PtN )t≥0 is stable on functions with derivative equal to 0 on the boundary (in higher
dimensions, with normal derivative equal to 0 on the boundary), with associated re-
flected Markov processes. In this case, it is not enough to solve an ordinary stochas-
tic differential equation in order to describe the semigroup and it is necessary to add
to the equation a drift term called local time, which is a measure on R+ supported
by the set of times where the process is at the boundary, and which is singular with
respect to the Lebesgue measure. We will not go into these considerations here.
It is not easy to connect Neumann or Dirichlet conditions and processes reflected
or killed at the boundary. It is not so difficult to see that for the process killed at the
boundary, starting from a continuous function f vanishing at the boundary of O,
and provided the coefficients of the stochastic differential equation are smooth and
bounded, then PtD f also vanishes at the boundary of O. However, for the semigroup
associated with the reflected process, it is in general much more difficult to make
sure that whenever f has a derivative vanishing at the boundary of O, then the same
holds for PtN f . We do not investigate this question here, which would lead to a
further study of local times.
In the case of Sturm-Liouville operators on an interval of the real line, we sys-
tematically consider below and throughout this work the Neumann boundary condi-
tions, and thus the Neumann semigroup which we denote by (Pt )t≥0 for simplicity.
These conditions will indeed make it possible for the constant function 1 to belong
to the domain and to satisfy Pt (1) = 1.
The semigroup (Pt )t≥0 associated with a Sturm-Liouville operator L on [−1, +1]
as in (2.6.2) with smooth and bounded coefficients is always Hilbert-Schmidt in
L2 (μ) for the invariant probability measure μ (which is also reversible in this one-
dimensional case). This will be studied later, since the operator actually satisfies a
Sobolev-type inequality, and in this case the density kernels pt (x, y) will actually be
bounded for t > 0. Its spectrum is therefore discrete. To understand the structure of
the eigenvectors, one has to solve the second order differential equation Lf = −λf .
But, for any λ, and for every choice of f (−1) and f (−1), there is a unique so-
lution in the open interval (−1, +1). For Neumann conditions, one has to impose
f (−1) = 0. Then, there exists an infinite sequence of values of λ for which the
solutions of this equation satisfy f (1) = 0 (the value of f (−1) does not matter and
can be fixed). This infinite sequence consists of the eigenvalues of −L with Neu-
mann boundary conditions, and the associated eigenfunctions, suitably normalized,
form an orthonormal basis of L2 (μ). In particular, the constant function (equal to
1) is an eigenfunction with eigenvalue 0.
Denote by (λk )k∈N the sequence of eigenvalues of −L as just described, and
by (fk )k∈N the sequence of the corresponding eigenvectors (normalized in L2 (μ)).
Then the kernel densities of (Pt )t≥0 with respect to the invariant measure μ are
given, following (1.7.3), p. 32, by

pt (x, y) = e−λk t fk (x)fk (y), t > 0, (x, y) ∈ [−1, +1] × [−1, +1]. (2.6.4)
k∈N
However, very few cases yield explicit expressions for λk and fk , and hence of this
heat kernel. The later three sections devoted to the Hermite, Laguerre and Jacobi
operators are such instances.
2.6.3 Essential Self-adjointness
For many interesting examples, it often happens that the drift coefficient is singular
at the boundary, with a boundary repulsion force on the process. There is a nice way
to determine whether the associated Markov process X = {Xtx ; t ≥ 0, x ∈ [−1, +1]}
reaches the boundary or not in terms of harmonic functions, that is, solutions of
Lh = 0. If L is of the form (2.6.2), so that Lh = h + c(x)h , then the harmonic

functions h are solutions of hh = −c(x), namely
x y
h(x) = exp − c(r)dr dy.
a b

Since the density w of the invariant measure μ satisfies ww = c, a harmonic function
h thus satisfies h = w1 . In terms of X, assuming that (Xt )t≥0 starts from an interior
point (say 0), then (h(Xt ))t≥0 is a local martingale. If Tu and Tv are the hitting
times of u, v ∈ (−1, +1) respectively, and if Tu,v = Tu ∧ Tv is the hitting time of the
boundary of [u, v] ⊂ (−1, 1), we get, since h is harmonic,

E h(XTu,v ) = h(0) = h(u) P(Tu < Tv ) + h(v) P(Tv < Tu ). (2.6.5)
Hence, if h(u) → ∞ as u → −1, the probability of reaching −1 before reaching any

other point in the interior is zero, and therefore the process never hits the boundary.
From a practical point of view, the critical case (for −1 say) occurs when c(x) x+1α
as x → −1. The process never hits the boundary, if α ≥ 1, and it reaches it when
α < 1.
As already pointed out earlier for Brownian motion on the half-line, not reaching
the boundary is not the same as being essentially self-adjoint. Indeed, when the drift
is singular, the definition of the generator L on the open interval may be enough
to fully describe the Markov semigroup and process. In analytical terms, the C ∞
compactly supported functions on the open interval (−1, +1) form a core of the
domain, and thus the operator is essentially self-adjoint. (In particular, whenever L
is self-adjoint, there is no need to care about the boundary conditions when solving
the heat equation.) However, this is usually not that easy to prove. The following
proposition provides a sufficient criterion. It is similar to Proposition 2.4.1 above.
Proposition 2.6.1 Let Lf = f + cf be a Sturm-Liouville operator on (−1, +1).

Assume that c is smooth in (−1, +1) and that there exist C1 , C2 > 0 such that for
every x ∈ (−1, +1),
c2 (x)
c (x) + ≥ C1 min (1 + x)−2 , (1 − x)−2 − C2 .
2
Then L is essentially self-adjoint.
α− α+
Observe that when c(x) 1+x at x = −1 and c(x) − 1−x at x = +1, then
the above condition requires that min(α− , α+ ) > 2, whereas the condition for the
process not to reach the boundary requires min(α− , α+ ) ≥ 1. Such arguments are of
course very specific to dimension one, since explicit computations can be performed
in this case. In higher dimensions, the idea is often to mimic the one-dimensional
case by using comparison arguments.
2.7 Diffusion Semigroups Associated with Orthogonal

Polynomials

real line R, a probability measure μ with exponential moments (that is, such
On the
that R eα|x| dμ(x) < ∞ for some α > 0) has a more or less canonical Hilbertian ba-
sis for the space L2 (μ) consisting of orthogonal polynomials, since polynomials are
then dense in L2 (μ). Such a basis is obtained by orthogonalization of the sequence
of polynomials x k , k ∈ N, in the scalar product of L2 (μ). Up to normalization and
change of sign, these orthogonal polynomials are unique, and in the following we al-
ways choose them so that they are normalized in L2 (μ) with strictly positive leading
coefficient.
There are only a few cases of orthogonal polynomials which are also eigenvectors
of diffusion operators. In dimension one, there are only, up to affine transformations,
the Hermite, Laguerre and Jacobi polynomials. This is why we give special atten-
tion to these polynomials in the following sub-sections. Moreover, these examples
belong to the few cases for which there is a complete description of the sequences
of eigenvalues and eigenvectors for a Sturm-Liouville operator. There is a huge lit-
erature on orthogonal polynomials, especially on these three main families. The
following only outlines what will be useful and relevant for the rest of the book.
As will become clear in the following, the Hermite, Laguerre and Jacobi opera-
tors analyzed here may be considered on various classes of functions, C ∞ functions,
C ∞ rapidly decreasing functions, or even polynomials. Again, Chap. 3 will present
the suitable framework and conditions covering such examples.
2.7 Diffusion Semigroups Associated with Orthogonal Polynomials 103
2.7.1 The Ornstein-Uhlenbeck Semigroup
The Ornstein-Uhlenbeck (or Hermite) semigroup is one of the simplest Sturm-

Liouville semigroups on the real line (which can then be extended, by product, to
Rn ) for which all the characteristic elements are explicit. It will serve as a main ref-
erence example in the investigation of many functional inequalities in later chapters.
The Ornstein-Uhlenbeck operator on the real line acts on smooth functions f on
R by
LOU f = f − xf .
According to the preceding section, or as can be immediately checked, its carré du
champ operator is (f ) = f 2 and its invariant and reversible probability measure is
the standard Gaussian probability distribution on the real line dμ(x) = e−x /2 √dx .
2
2π
The preceding definition may be extended to Rn by the product procedure out-
lined in Sect. 1.15.3, p. 59. The Ornstein-Uhlenbeck operator on E = Rn is then
given by LOU = − D where D is the dilation vector field acting on every smooth
function f by Df (x) = ni=1 xi ∂i f (x), x ∈ Rn . In other words, for every smooth
function f on Rn ,
LOU f = f − x · ∇f. (2.7.1)
To lighten the notation, we write LOU = L throughout this and the next sub-section.
The carré du champ operator is the usual gradient operator (f ) = |∇f |2 . Since
−D is the gradient of −|x|2 /2, the invariant and reversible probability measure is
given by
dx
dμ(x) = e−|x|
2 /2
(2π)n/2
that is μ is the standard Gaussian measure on Rn (with mean zero and covariance
the identity matrix). We may also characterize the Ornstein-Uhlenbeck operator L
as the unique operator such that
Lxi = −xi , (xi , xj ) = δij , 1 ≤ i, j ≤ n,
on the system of coordinates (xi )1≤i≤n .

The associated Markov semigroup and process may be explicitly described.
The Ornstein-Uhlenbeck semigroup P = (Pt )t≥0 with infinitesimal generator
L = LOU is conveniently presented by means of the associated Markov process
X = {Xtx ; t ≥ 0, x ∈ Rn } which admits the nice explicit representation

−t
√ t s
Xt = e
x
x+ 2 e dBs (2.7.2)
0
where (Bt )t≥0 is a standard Brownian motion (in Rn ) starting at the origin. This
process is indeed the solution of the stochastic differential equation
√
dXt = 2 dBt − Xt dt, X0 = x,
with the generator L = − x · ∇, defining the Ornstein-Uhlenbeck process.

The Ornstein-Uhlenbeck semigroup (Pt )t≥0 admits an explicit integral repre-
sentation since the distribution of each Xtx is Gaussian with covariance matrix
(1 − e−2t ) Id centered at e−t x so that, for suitable functions f : Rn → R,

Pt f (x) = E f e−t x + 1 − e−2t G
(2.7.3)

= f e−t x + 1 − e−2t y dμ(y), t ≥ 0, x ∈ Rn ,
Rn
where G is a standard Gaussian variable on Rn (with distribution μ). This formula

should be compared with (2.1.2) which describes a similar form for the Euclidean
Brownian semigroup. There are of course many connections between these two
semigroups. In particular, the Ornstein-Uhlenbeck semigroup (Pt )t≥0 admits ker-
nel densities (Mehler kernel) with respect to the Lebesgue measure given, for every
t > 0 and (x, y) ∈ Rn × Rn , by

1 |y − e−t x|2
exp − .
(2π(1 − e−2t ))n/2 2(1 − e−2t )
The densities with respect to the Gaussian invariant measure μ are on the other hand

1 |x|2 − 2et x · y + |y|2
pt (x, y) = exp − . (2.7.4)
(1 − e−2t )n/2 2(e2t − 1)
It can be verified that

pt2 (x, y)dμ(x)dμ(y) < ∞
Rn Rn
so that the operators Pt , t > 0, of the Ornstein-Uhlenbeck semigroup are Hilbert-

Schmidt (cf. Sect. A.6, p. 483).
The Ornstein-Uhlenbeck generator L of (2.7.1) satisfies the curvature condition
CD(1, ∞) of (1.16.6), p. 72. Indeed, one can easily check that for every smooth
enough function f on Rn ,
2 (f ) = |∇∇f |2 + |∇f |2 ≥ |∇f |2 = (f ).
It is not difficult to show that L cannot satisfy any CD(ρ, m) condition where
ρ ∈ R and m < ∞, expressing that in a sense the operator is intrinsically infinite-
dimensional (independently of the dimension of the state space). The curvature con-
dition CD(1, ∞) may also be described by the commutation property on the semi-
group
∇Pt f = e−t Pt (∇f ) (2.7.5)
for every t ≥ 0 and f sufficiently regular. This property may be verified directly
from the representation (2.7.3). Alternatively, starting with the commutation relation
[L, ∇] = L∇ − ∇L = ∇,
the vector-valued map (s) = Ps (∇Pt−s f ), 0 ≤ s ≤ t, satisfies (s) = (s).

Hence e−t (t) = (0) which amounts to (2.7.5). As a consequence of (2.7.5),

|∇Pt f | ≤ e−t Pt |∇f | ,
which is linked to the curvature condition CD(1, ∞) in Theorem 3.3.18, p. 163 (see
also Sect. 4.7, p. 206, and Sect. 5.5, p. 257, below).
As announced, of particular interest are the eigenvectors of the Ornstein-
Uhlenbeck generator which are described in terms of the Hermite polynomials
orthogonal with respect to the Gaussian measure. Here we begin with dimension
one, so
that μ denotes below the standard Gaussian measure on R (satisfying in par-
ticular R eα|x| dμ(x) < ∞ for some, or even every, α > 0). For every integer k, the
Ornstein-Uhlenbeck operator L sends the space Pk of polynomials of degree less
than or equal to k into itself. If the finite-dimensional vector space Pk is equipped
with the scalar product of L2 (μ), L defines a symmetric operator in a Euclidean
space. It may be diagonalized with respect to an orthonormal basis. By recurrence
over k, one then constructs a sequence of orthonormal polynomials in L2 (μ) which
are eigenvectors of L. Comparing the step from k to k + 1, the polynomial added
at k + 1 is none other than the polynomial of degree k + 1 orthogonal to Pk . The
sequence of orthonormal polynomials (Hk )k∈N obtained in this way is therefore
the orthonormal polynomials for the Gaussian measure μ known as the Hermite
polynomials. These polynomials are unique provided their signs are prescribed, and
by convention will be taken so that the coefficient of the term of highest degree is
strictly positive.
For each k ∈ N, Hk is of degree k and is an eigenvector of −L, that is
−L Hk = λk Hk
for some positive number λk . Inspecting the term of highest degree shows that
λk = k (H0 is a constant). Alternatively, the sequence (Hk )k∈N of Hermite poly-
nomials may be defined via the generating series
2 /2 sk
esx−s = √ Hk (x), x ∈ R, s ∈ R,
k∈N
k!
and it is then an exercise to check using this representation that −LHk = kHk for
every k. There are many other useful representations of the Hermite polynomials
Hk , k ∈ N, for instance

1 1
Hk (x) = √ E (x + iG) = √k
(x + iy)k dμ(y), x ∈ R
k! k! R

2
(observe that R es(x+iy) dμ(y) = esx−s /2 , s ∈ R). For example H0 (x) = 1,
H1 (x) = x, H2 (x) = √1 (x 2 − 1), and so forth. One may also deduce in this way the
2
transition densities (2.7.4) from the general expansion (2.6.4).
On the basis of this orthogonal decomposition of L2 (μ), the functions in the

L2 (μ)-domain D(L) of L are easily described. Any function in L2 (μ) can therefore
be developed as a series of Hermite polynomials

f= ak Hk (2.7.6)
k∈N

with ak = R f Hk dμ, k ∈ N, and f 22 = k∈N ak2 . In this case, the semigroup Pt
acts, for every t ≥ 0, as

Pt f = e−kt ak Hk .
k∈N
Similarly, it is expected that

−Lf = k ak Hk ,
k∈N
so that the functions in the L2 (μ)-domain of L are the functions f with de-
composition
(2.7.6) such that the series describing Lf is in L2 (μ), that is, such
that k∈N k ak2 < ∞. This condition therefore requires rapidly decreasing coeffi-
2
cients ak . Translating this condition into the regularity properties of a given func-
tion f ∈ L2 (μ) requires more care. The latter is actually equivalent to saying that
f ∈ C 1 and that its second derivative in the distributional sense is a function such
that f − xf is in L2 (μ).
For a function f decomposed in L2 (μ) as (2.7.6), the Dirichlet form of Defini-
tion 1.7.1, p. 30, is given by

E(f ) = k ak2
k∈N

(use that E(f ) = f (−Lf )dμ). Hence functions with E(f ) < ∞ are a priori less
regular than functions in the domain (if one agrees that the regularity of a function
is described in terms of the decay of its L2 (μ)-coefficients).
Finally, for any function f ∈ L2 (μ), Pt f is, for every t > 0, in the domain D(L)
and
C
LPt f 22 ≤ 2 f 22 .
t
To prove this assertion, note that the function re−r is bounded on r ∈ R+ , say by
C, so that for every k ∈ N and t > 0, k 2 e−2kt ≤ tC2 . Then compute the norm of
LPt f as above in its spectral decomposition. This property is actually not particu-
lar to the Ornstein-Uhlenbeck operator, and the same argument may be applied to
any symmetric semigroup provided the spectral representation may be used. Hence,
a symmetric semigroup is always smoothing in the L2 -sense (which is however
weak). Below, stronger regularizing properties will be investigated, depending on
the ellipticity of the generator. This comment emphasizes more the difference with
non-symmetric semigroups. For example, a dynamical system which pushes Dirac

masses onto Dirac masses can never be regularizing.
The spectral structure of the Ornstein-Uhlenbeck generator and semigroup in
dimension n is similar. Products of Hermite polynomials

n
Hki (xi ), (k1 , . . . , kn ) ∈ Nn , (x1 , . . . , xn ) ∈ Rn ,
i=1

are eigenfunctions, with eigenvalues − ni=1 ki of L, and define an orthonormal
basis of L2 (μ), where now μ is the standard Gaussian measure in Rn (the product
of the one-dimensional Gaussian measures on each coordinate). The spectrum is
again N, but this time the vector space associated with a given eigenvalue is no
longer one-dimensional.
It is worth mentioning that several useful integration by parts formulas for the
Gaussian measure μ may actually be established with the help of Hermite polyno-
mials. The simplest one is obtained using the fact that L(xi ) = −xi (1 ≤ i ≤ n) to
get that for f ∈ D(L),

xi f dμ = − L(xi )f dμ = (xi , f )dμ = ∂i f dμ. (2.7.7)
Rn Rn Rn Rn
Of course, such an identity also follows from a standard integration by parts on the
Gaussian density. While this standard argument may be iterated in various ways, the
Hermite polynomials as eigenvalues of L provide an efficient systematic tool. For
example, using similarly that L(|x|2 − n) = −2(|x|2 − n) yields that

|x| f dμ = n
2
f dμ + f dμ (2.7.8)
Rn Rn Rn
and so on.
The Ornstein-Uhlenbeck operator is related to another famous and well-studied
operator, the harmonic oscillator in Rn , given on smooth functions f by
Hf = f − |x|2 f. (2.7.9)
This is an operator with potential, symmetric with respect to the Lebesgue measure,
and corresponding to the simplest model of quantum mechanics. Denote by (Kt )t≥0
its associated semigroup. Observing that H (U0 ) = −n U0 where U0 = e−|x| /2 , the
2
ground state transformation described in Sect. 1.15.6, p. 63, may be performed. Thus
consider the semigroup
1
Rt f = ent Kt (U0 f ), t ≥ 0,
U0
with generator given on smooth functions f by
1
Lf = nf + H (U0 f ).
U0
Now
H (U0 f ) = −n U0 f + U0 f + 2 ∇U0 · ∇f
so that
Lf = f + 2∇ log U0 · ∇f
which is precisely the Ornstein-Uhlenbeck operator − x · ∇. The transformation
f → U0 f carries the analysis of H into the analysis of the Ornstein-Uhlenbeck
operator in terms of Hermite polynomials.
2.7.2 The Infinite-Dimensional Ornstein-Uhlenbeck Semigroup
Various functional inequalities in this book are concerned with the dependence of
constants with respect to dimension. Independence with respect to dimension then
usually allows for infinite-dimensional extensions. Actually, in most cases, infinite-
dimensional functional inequalities appear as limits of finite-dimensional inequal-
ities with constants controlled in terms of dimension. Since we will not discuss
many infinite-dimensional examples, it is of interest to describe at least one genuine
infinite-dimensional example, the infinite-dimensional Ornstein-Uhlenbeck opera-
tor. Nevertheless, the functional inequalities for Gaussian measures emphasized in
the future chapters will be restricted to their (dimension-free) finite-dimensional
statements. Infinite-dimensional versions may be reached either by a limiting argu-
ment on the finite-dimensional results, or by suitable adaptations of the formalism
and of the main principles of proofs as they can be drawn from the material dis-
cussed here.
As presented in the finite-dimensional case in the previous Section, the Ornstein-
Uhlenbeck operator in Rn may be described as the diffusion operator L such that
Lxi = −xi and (xi , xj ) = δij where xi , 1 ≤ i ≤ n, are linear forms corresponding
to an orthonormal basis. Moreover, these coordinates xi define independent standard
normal variables under the invariant measure. In a somewhat more abstract way,
L may be described by its action on linear forms (x) = · x, x ∈ Rn , as

L() = −, , = · ,
thus independent of the chosen orthonormal basis. This discussion provides an op-
portunity to mention that if μ is a centered Gaussian measure on Rn with non-
singular covariance matrix Q, then it is the invariant measure of the Ornstein-
Uhlenbeck generator

n
n

M ij ∂ij2 − MQ−1 x i ∂i
i,j =1 i=1
where M = (M ij )1≤i,j ≤n is any symmetric positive-definite matrix. For M the

identity matrix, the associated Ornstein-Uhlenbeck semigroup takes the form, for
f : Rn → R (cf. (2.7.3)),

−1 −1 1/2
Pt f (x) = f e−tQ x + Id −e−2tQ y dμ(y), t ≥ 0, x ∈ Rn .
Rn
From this description, the extension from Rn to infinite-dimensional (Hilbert or

Banach) spaces requires suitable systems of coordinates. However, some care has to
be taken, in particular with respect to the existence itself of an invariant measure. Re-
call first that a probability measure μ on the Borel sets of a (real separable) Banach
space B is said to be (centered) Gaussian if the linear functions x → (x) = , x,
∈ B ∗ , are, under μ, (centered) real Gaussian variables. Since weak limits of Gaus-
sian variables are also Gaussian, it is enough to have this property for a dense
subset of B ∗ . Now, for example, let H be a (real) separable Hilbert space with
scalar product ·, · equipped with a centered Gaussian measure μ. Given an or-
thonormal basis (ei )i≥1 of H, the corresponding functions Xi (x) = ei , x, i ≥ 1,
x ∈ H, should be, by orthogonality, independent standard normal variables under μ.
But then, the strong law of large numbers
would show that, μ-almost everywhere,
1 k 2 → 1 as k → ∞, so that 2 = ∞, contradicting the hypothesis
k X
i=1 i X
i≥1 i
that μ is supported on H. This suggests that the Gaussian measure μ should be
supported on a larger space.
In a Banach space setting, one must first address the issue of defining Hilber-
tian coordinates. These questions have been investigated in the context of abstract
Wiener spaces with the study of Gaussian measures in Banach spaces and their as-
sociated reproducing kernel Hilbert spaces (which yield the required orthonormal
system of coordinates with the scalar product induced by the covariance structure
of the Gaussian measure). We do not cover the general theory here, but describe
how to overcome these difficulties in the concrete and basic example of the Wiener
measure. One of the most used and classical examples of Gaussian measure in a
Banach space is the Wiener measure, the law of Brownian motion on the space
C0 ([0, 1]) of continuous functions w : [0, 1] → R such that w(0) = 0 (the so-called
canonical space). One crucial feature of the construction (not addressed here) is that
the Brownian paths may be shown to be almost surely continuous so that the law
of Brownian motion (Bt )t≥0 (restricted to t ∈ [0, 1]) indeed defines a probability
measure μ on C0 ([0, 1]) identifying w(t) with Bt (w), called the Wiener measure.
The dual space of C0 ([0, 1]) consists of bounded (signed) measures, via the duality

1 p
bracket ν, w = 0 w(s)dν(s), and for linear combinations ν = i=1 ai δti , ai ∈ R,
p
ti ∈ [0, 1], ν, w = i=1 ai Bti which defines a real centered Gaussian variable with
variance
2 1 2
p p
p
E
ai Bti = ai aj (ti ∧ tj ) = ai 1[0,ti ] dt.

i=1 i,j =1 0 i=1
By a density argument, this basic identity may be used to define a Hilbertian

subset of C0 ([0, 1]) with scalar product induced by L2 (μ) known as the repro-
ducing kernel Hilbert space H associated with the Wiener measure μ. Indeed, if
p
h(t) = i=1 ai (ti ∧ t), t ∈ [0, 1], h defines an absolutely continuous function in
C0 ([0, 1]) such that
2 2
1 p 1
|h|2H = ḣ2 (t)dt = E ai Bti = E ḣ(t)dBt .
0 i=1 0
The Hilbert space H is identified with the set of absolutely continuous elements

1
h of C0 ([0, 1]) such that 0 ḣ2 (t)dt < ∞, traditionally called the Cameron-Martin
Hilbert space (of the Wiener measure μ).
For any orthonormal sequence (ei )i≥1 in L2 ([0, 1], dt), the Wiener integrals (ran-
dom variables on the canonical space)
1
Xi (w) = ei (s)dBs (w), i ≥ 1,
0
are independent standard normal random variables (see Sect. B.1, p. 487). Note that
when the functions ei have one derivative in L2 ([0, 1], dt), Itô’s formula indicates
that
1
Xi (w) = ei (1)w(1) − ei (s)w(s)ds.
0
The sequence (Xi )i≥1 defines a set of orthonormal coordinates on the canoni-

t
such that, if vi (t) = 0 ei (s)ds, t ∈ [0, 1], i ≥ 1 (elements of H), the
cal space
series i≥1 vi Xi is almost everywhere convergent with distribution μ (although

i≥1 Xi = ∞ μ-almost everywhere).
2
On L2 (μ) with state space E = C0 ([0, 1]) and probability measure the Wiener
measure μ, consider the algebra A0 of functions f : C0 ([0, 1]) → R which are poly-
nomials in a finite number of such Xi ’s, that is f (w) = Pk (X1 (w), . . . , Xn (w)). The
Ornstein-Uhlenbeck operator L may then be defined as acting on such f ’s by

Lf = LPk X1 (w), . . . , Xn (w)
where on the right L is the usual n-dimensional Ornstein-Uhlenbeck operator. It

is easily seen that A0 is dense in L2 (μ), stable under L and that L is a diffusion
operator such that
LXi = −Xi , (Xi , Xj ) = δij , i, j ≥ 1.
The operator L generates a semigroup (Pt )t≥0 (infinite-dimensional Ornstein-

Uhlenbeck semigroup) described exactly as in (2.7.3)

Pt f (x) = f e−t x + 1 − e−2t y dμ(y), t ≥ 0, x ∈ E.
E
This representation may be verified directly for functions f ∈ A0 since it then boils
down to the corresponding representation in Rn . Observe that A0 is stable under
the action of (Pt )t≥0 , and therefore is a core for L. The representation is therefore
independent of the chosen base (ei )i≥1 of L2 ([0, 1], dt).
The CD(1, ∞) curvature property of the finite-dimensional Ornstein-Uhlenbeck
operator immediately extends to the infinite-dimensional setting. Furthermore, since
the operator is an infinite sum of one-dimensional Ornstein-Uhlenbeck operators,
its eigenvectors and eigenvalues are quite easy to describe. The eigenvalues of −L
are the positive integers. However, the corresponding eigenspaces are now infinite-
dimensional. For example, the eigenspace associated with the eigenvalue λ1 = 1 is
the space of linear functionals. The infinite-dimensional Ornstein-Uhlenbeck oper-
ator is thus an example of an operator with a purely point spectrum but which is not
discrete in the terminology of Sect. A.4, p. 478, in Appendix A.
2.7.3 The Laguerre Semigroup
The next example of a Sturm-Liouville operator which may be diagonalized in or-

thogonal polynomials is the Laguerre operator. This example has many similarities
with the Ornstein-Uhlenbeck example (so that not every detail is supplied here).
The Laguerre operator on E = R+ acts on smooth functions f on R+ (or re-
strictions to R+ of smooth functions on a neighborhood of R+ ) by
Lα f = xf + (α − x)f , (2.7.10)
where α > 0. Of course, any linear change x → ax (a > 0) would modify the co-
efficients a little, the chosen normalization being justified below. Its invariant and
reversible measure is the gamma distribution on R+
dμα (x) = γα−1 x α−1 e−x dx.
The carré du champ operator is given on smooth functions by (f ) = x f 2 .

It may be observed that the Laguerre operator for suitable parameters may be
constructed from the Ornstein-Uhlenbeck operator of the preceding section. More
precisely, 2 Ln/2 appears as the Ornstein-Uhlenbeck operator LOU = − x · ∇ on
Rn acting on radial functions of the form f ( |x|2 ) where f : R+ → R. In particular,
2
it is therefore of no surprise that the Laguerre operator Lα satisfies the curvature

condition CD( 12 , ∞), at least when 2α is an integer. It is easily checked that this
property remains true for any α ≥ 12 , and that the curvature 12 is the best possible.
For the other values, the nature of the boundary conditions has to be identified.
On the form (2.7.10) itself, it is however not immediately clear which conditions
should be imposed since (f ) = 0 at the boundary (the metric degenerates). The
2
change of variable x → y4 shows that we are dealing with Neumann boundary
conditions. Indeed, as explained in Sect. 2.6, this change of variable transforms Lα
2
into the operator f + 4α−2−y
2y f on R+ . In fact, following Sect. 2.6, the process
associated with this operator lives on (0, ∞) as soon as α ≥ 1, and is essentially
self-adjoint on (0, ∞) as soon as α > 32 . Comparing with the case of the hyperbolic
Laplacian on Rn−1 × (0, ∞) (Sect. 2.3) shows that this issue really depends on the
completeness of the space for the natural distance induced by the operator L as
described in (3.3.9), p. 166. In the Laguerre case, the distance from the boundary
{x = 0} to any point in (0, ∞) is finite, while in the case of the hyperbolic Laplacian
it is infinite.
As in the Ornstein-Uhlenbeck case, the Laguerre operator Lα may be diagonal-
ized with respect to a basis of orthogonal polynomials, namely the Laguerre or-
thogonal polynomials. Indeed, as before, Lα preserves the finite-dimensional space
of polynomial functions with degree less than k, and is symmetric on it when en-
dowed with the Euclidean structure inherited from L2 (μα ). Therefore there exists
an orthonormal basis (Lk )k∈N of L2 (μα ) consisting of polynomials Lk of degree k
which are eigenvectors of Lα . After normalization, this sequence is unique and de-
fines the Laguerre orthonormal polynomials associated with μα . As usual L0 = 1,
and the eigenvalues are computed just by looking at the highest degree term so that
−Lα Lk = k Lk
for every k ∈ N. Laguerre polynomials may be described by generating series.

Comparing with the Ornstein-Uhlenbeck operator from which it comes when
α = n2 , the spectrum of 2 Lα is identical to the even part of the Ornstein-Uhlenbeck
spectrum. That is, the Laguerre polynomial of degree k corresponds to a 2k-degree
Hermite polynomial in dimension n which is radial. Indeed, the 2k-degree Hermite
2
polynomial H2k is an even polynomial which may therefore be written as Qk ( x2 ) for
some degree k polynomial Qk . As the Ornstein-Uhlenbeck operator LOU on Rn is
2
invariant under rotations, for any unit vector e ∈ Sn−1 , the map x ∈ Rn → Qk ( (e·x)
2 )
is an eigenvector of LOU with eigenvalue 2k, and hence so is the map sending x to

(e · x)2
Qk dσn−1 (e)
Sn−1 2
where σn−1 is the uniform measure on the sphere Sn−1 ⊂ Rn . This function may be
written as Rk ( |x|2 ) where Rk is a further polynomial of degree k. Then Rk satisfies
2
2 Ln/2 Rk = LRk = −2kRk , and up to a multiplicative constant, it is precisely the

Laguerre polynomial with degree k and parameter α = n2 .
The image of the Laguerre operator as a projection of the n-dimensional
Ornstein-Uhlenbeck

operator leads to further relationships between the two √fam-
ilies. Since R+ xdμα = α, recentering and scaling the variable as x = α + α y
transforms the Laguerre operator Lα into the operator

y
1 + √ f (y) − yf (y)
α
√
on [− α, ∞). As α → ∞, this operator converges to the one-dimensional
Ornstein-Uhlenbeck generator. Indeed, under this change of variable, the measure
μα converges to the standard Gaussian measure, the orthogonal Laguerre poly-

nomials themselves converge to the corresponding Hermite polynomials, and the
convergence is furthermore valid at the level of the semigroups. For α = n2 , the
geometric picture of the preceding simply boils down to the central limit theorem
for sums ni=1 Xi2 where the Xi ’s are independent standard normal variables. The
same phenomenon will be observed for the Jacobi semigroups below (Sect. 2.7.4).
In the same way, many of the properties of the Hermite polynomials may be trans-
lated to Laguerre polynomials with integer parameter, and very often extended to
the general parameter.
The semigroup with generator Lα is called the Laguerre semigroup and is de-
noted by (Ptα )t≥0 . In this setting, it may actually be defined explicitly from its action
on the basis of Laguerre orthogonal polynomials, as was the case for the Ornstein-
Uhlenbeck semigroup. It coincides with the semigroup of the unique essentially
self-adjoint extension in the range α > 32 . The Laguerre semigroup admits kernel
densities with respect to the invariant measure.
Like the Ornstein-Uhlenbeck semigroup, the family of Laguerre semigroups
satisfies a number of remarkable identities. For example, the identity ∂x Lα =
(Lα+1 − Id)∂x on the generators Lα and Lα+1 leads to the commutation property
on the associated semigroups
∂x Ptα = e−t Ptα+1 ∂x
which may be compared to the corresponding formula (2.7.5) for the Ornstein-
Uhlenbeck semigroup. Similarly, the h-transform

x α−1 Lα x 1−α f = (L2−α + α − 1)f
translates into

x 1−α Ptα x α−1 f = et (α−1) Pt2−α f
which exhibits a duality between Lα and L2−α (0 < α < 2) very similar to the dual-
ity observed for Bessel semigroups or generators (see Sect. 2.4.2).
Products of Lα on Rn+ may be considered similarly.
2.7.4 The Jacobi Semigroup
The last family of Sturm-Liouville operators presented here is the one associated
with the so-called Jacobi polynomials. As mentioned earlier, it plays an important
role in the analysis of the Laplacian on spheres. It is, after the Ornstein-Uhlenbeck
and Laguerre operators, the third family of one-dimensional diffusion operators
which may be diagonalized with respect to a family of orthogonal polynomials.
It is a remarkable fact that these three operators are the only operators which have
this property.
The Jacobi operator on E = [−1, +1] acts on smooth functions f on [−1, +1]
(or restrictions to [−1, +1] of smooth functions on a neighborhood of [−1, +1]) by

Lα,β f = 1 − x 2 f − (α + β)x + α − β f , (2.7.11)
where α, β > 0. Its invariant and reversible measure is the beta distribution on
[−1, +1]
dμα,β (x) = Cα,β (1 − x)α−1 (1 + x)β−1 dx
where Cα,β > 0 is a suitable normalization constant. The carré du champ operator is
given on smooth functions f by (f ) = (1 − x 2 )f 2 . When α = β, we sometimes
speak of symmetric Jacobi operators (although all Jacobi operators are symmetric
with respect to the invariant measure).
While the Jacobi operators (2.7.11) may be considered for any values of α, β > 0,
the boundary issues are more easily described after a suitable change of variable.
Setting x = cos θ , we obtain in the new variable θ an operator on [0, π] which reads
(α + β − 1) cos θ + α − β
∂θ2 + ∂θ . (2.7.12)
sin θ
According to Proposition 2.6.1, this operator is self-adjoint on the interval (0, π)
(and hence the Jacobi operator Lα,β is self-adjoint on the interval (−1, +1)) as soon
as min(α, β) > 32 while the associated Markov process does not reach the boundary
as soon as min(α, β) ≥ 1.
As in the case of the Ornstein-Uhlenbeck operator, Lα,β sends the family of poly-
nomials of degree less than or equal to k into itself, and thus admits a basis of
eigenvectors consisting of polynomials. These polynomials are the orthogonal poly-
nomials associated with the measure μα,β , called the Jacobi polynomials (Jk )k∈N
(with implicit parameters α, β). Jacobi polynomials may be described by generating
series (see (2.7.14) below in the symmetric case).
The Ornstein-Uhlenbeck generator can be seen as a limit of Jacobi operators.
Indeed, for α = β = n2 , and after changing x into √xn , n1 Lα,β converges as n → ∞
to the (one-dimensional) Ornstein-Uhlenbeck operator LOU . (In fact, the conver-
gence is much more comprehensive, the measures, the eigenvectors and the eigen-
values converge, and the Markov semigroups themselves converge, together with
the finite-dimensional laws of the associated Markov processes.) In the same way,
if we translate [−1, +1] to [0, 2], and then change x into xn and α into n (with β
fixed), then n1 Ln,β converges to the Laguerre operator L2β with the same strength
of convergence as described in the remark above.
The eigenvalues of the Jacobi operators are easy to describe by considering the
action of Lα,β on the highest degree terms of the orthogonal polynomials. The eigen-
values of −Lα,β are given by the sequence
λk = k(k + α + β − 1), k ∈ N.
There is an analogy with the eigenvalues of the spherical Laplacian Sn , which
should come as no surprise. When n is an integer (n ≥ 1), and α = β = n2 , the Jacobi
operator Lα,β may indeed be seen as an image of Sn via its action on functions
depending only on one coordinate in Rn+1 , that is the functions which are invariant
under rotations leaving one point fixed (for example (1, 0, . . . , 0)). These functions
are called zonal. This may easily be seen via (2.2.2) acting on functions depending
only on x 1 . Hence, if Jk is a Jacobi polynomial of degree k, that is, an eigenvector
for Lα,α for α = n2 , then the restriction to Sn of the function Jk (x 1 ) defined on Rn+1
is an eigenvector for Sn .
This interpretation of the Jacobi operator Lα,α as the projection of Sn on
zonal functions leads to the fact that it satisfies the curvature-dimension condition
CD(2α − 1, 2α) for α = n2 (which may of course be checked directly from the rep-
resentation (2.7.12) and the general formula (1.16.9), p. 73). Observe furthermore
that (1.16.9) (which characterizes the curvature-dimension condition) is in fact an
equality, as it was for the radial part of the hyperbolic Laplace operator. Therefore
the symmetric Jacobi operator is a very good one-dimensional candidate for testing
properties related to curvature-dimension.
For non-symmetric Jacobi operators (α = β), there is also a representation
from spheres at least when α and β are half-integers α = n2 , β = p2 . The
Jacobi operator is then interpreted as a quotient of the sphere of dimension
n + p − 1 (up to a factor of 4). The eigenvalues of degree k correspond to
eigenfunctions on the sphere of dimension 2k as may be seen from the relation
4k(k + α + β − 1) = 2k(2k + n + p − 2). Of course, this observation in itself is
not enough to claim that the generator comes from the sphere models, but it is a
strong indication. To really understand the construction, let the spherical Lapla-
cian on Sn+p−1 ⊂ Rn+p act on functions of y = x12 + · · · + xp2 producing in this
way an operator defined on functions of y ∈ [0, 1]. After the change of variable
y = 2x − 1 towards the interval [−1, +1], one obtains (up to a factor of 4) the Ja-
cobi operator Lα,β with parameters α = n2 , β = p2 . It is then seen to satisfy the
curvature-dimension condition CD( n+p−2 4 , n + p − 1), this property being easily
verified even when n and p are no longer integers. However, in this dissymmetric
situation, there are many other curvature-dimension conditions which are not imme-
diately comparable with one another. For example, for n = p, the first interpretation
leads to an optimal condition CD(n − 1, n) for Ln/2,n/2 whereas the second one
leads to CD( n−12 , 2n − 1).
From this geometric interpretation, Ornstein-Uhlenbeck or Laguerre operators
may also be viewed as limits of spherical √ operators when the dimension n goes
to infinity on spheres scaled to have radius n. As a significant illustration, and a
source of inspiration for many results on the (infinite-dimensional) curvature con-
dition CD(ρ,
√ ∞), (normalized) uniform measures on spheres of dimension n and
radius n converge as n → ∞, when projected on k coordinates, to the standard
Gaussian measure on Rk . As mentioned above, a similar convergence holds at
the level of the operators and semigroups. This observation confirms the infinite-
dimensional character of the Ornstein-Uhlenbeck operator (as well as the Laguerre
operator) from the curvature-dimensional condition viewpoint.
The semigroup with generator Lα,β is called the Jacobi semigroup denoted by
α,β
(Pt )t≥0 . Like the Ornstein-Uhlenbeck and Laguerre semigroups, the Jacobi semi-
group may actually be defined explicitly from its action on the basis of Jacobi or-
thogonal polynomials (and coincides similarly with the semigroup of the unique es-
sentially self-adjoint extension in the range min(α, β) > 32 ). The Jacobi semigroup
α,β α,β
(Pt )t≥0 admits kernel densities pt (x, y), t > 0, (x, y) ∈ [−1, +1] × [−1, +1],
with respect to the invariant measure.
The family of Jacobi semigroups also satisfies some remarkable relations. Ob-
serve for example that

∂x Lα,β = Lα+1,β+1 − (α + β) Id ∂x .
α,β
This identity translates on the associated semigroups (Pt )t≥0 as
= e−(α+β)t Pt
α,β α+1,β+1
∂x Pt ∂x
α,β α+1,β+1
and therefore also leads to a simple relation between ∂x Jk and Jk−1 . In
α,β α+1,β+1
turn, these relations connect the kernels pt (x, y) and pt (x, y). Indeed, for
smooth functions f compactly supported in (−1, +1), writing

α,β
∂x pt (x, y)f (y)dμα,β (y)
[−1,+1]

= e−(α+β)t
α+1,β+1
pt (x, y)∂y f (y)dμα+1,β+1 (y)
[−1,+1]
yields after integration by parts

Cα,β α,β
e(α+β)t ∂x pt (x, y)
Cα+1,β+1
α+1,β+1 α+1,β+1
= − 1 − y 2 ∂y pt (x, y) + (α + β + 2)y + α − β pt (x, y)
where
Cα,β > 0 is the normalization

constant of the beta distribution μα,β . Now,
since [−1,+1] L(x)dμα,β = [−1,+1] L(x 2 )dμα,β = 0,

Cα,β 1 (α − β)2
= 1 − x 2 dμα,β = 1− ,
Cα+1,β+1 [−1,+1] α+β +1 α+β
α+1,β+1 α,β
providing an explicit formula linking pt to pt . This formula takes in par-
ticular a simpler form for y = ±1 and α = β, namely
e2αt
ptα+1,α+1 (x, ±1) = ± ∂x ptα,α (x, ±1). (2.7.13)
2(α + 1)(2α + 1)
This formula will be translated below in (2.7.15) into an identity for heat kernels on
the sphere.
The representation of Jacobi operators from the spherical Laplacian when the pa-
rameters are half-integers leads to nice formulas for the family of Jacobi orthogonal
polynomials via the Poisson formula (2.2.6). Let us for example illustrate the picture
when α = β = n2 , n ≥ 1. Any function f (x1 ) defined on [−1, +1] may be lifted to
a zonal function fˆ on the sphere by fˆ(x) = f (x · e1 ) where e1 = (1, 0, . . . , 0) ∈ Sn
(the variable x1 being with respect to as the first coordinate of a point x ∈ Sn ).
Then expand f ∈ L2 (μα,α ) with respect to the basis of Jacobi polynomials (Jk )k∈N ,

say f (x1 ) = k∈N ak Jk (x1 ). Since Jˆk is the restriction to Sn of a homogeneous
harmonic polynomial of degree k, Fk (X) = r k Jˆk (x) is harmonic in Rn+1 , where
X = rx, x ∈ S (polar coordinates). Then Qr f (x1 ) = k∈N ak r k Jk (x1 ), 0 ≤ r ≤ 1,
n

is such that Q r f is the harmonic extension to the unit ball of fˆ (read in polar coor-
dinates). This construction is represented by the kernel

Qr f (x1 ) = qr (x1 , y1 )dμα,α (y1 )
[−1,+1]

where qr (x1 , y1 ) = k∈N r
kJ
k (x1 )Jk (y1 ). On the other hand, by the Poisson for-
mula (2.2.6),

1 − r2

Qr f (x) = fˆ(y)dσn (y).
Sn |rx − y|n+1
Choosing x = e1 in the previous formula leads by identification to
1 − r2
r k Jk (1)Jk (y1 ) = (2.7.14)
(1 + r 2 − 2ry1 )(n+1)/2
k∈N
which is known as the generating function for the symmetric Jacobi polynomials.
The general formula for the Poisson kernel is however
a bit more complicated.
It requires the representation of a point on Sn as (x1 , 1 − x12 x̂), where x̂ ∈ Sn−1 ,
which then leads to

r k Jk (x1 )Jk (y1 )
k∈N

1 − r2
= n+1
dμα−1,α−1 (s).
[−1,+1] [1 + r 2 − 2rx1 y1 + 2rs (1 − x12 )(1 − y12 ) ] 2
Extending this formula to the general case where n is no longer an integer is then
easy and amounts to observing that both sides are solutions of the differential equa-
tion (∂r2 + nr ∂r + r12 Lα,α )F = 0, the remaining argument being similar.
The case of non-symmetric Jacobi operators may essentially be treated similarly
(at the expense of more complicated expressions), starting from the half-integer case
for the geometric picture and then extending to arbitrary parameters.
Conversely, the interpretation of the Jacobi operator as the spherical Laplacian
acting on functions which depend only on the first coordinate (that is, functions
depending only on the distance from (1, 0, . . . , 0) on the sphere) leads to useful
information about the sphere Sn itself. First, the image measure of the uniform mea-
sure σn on Sn via the map x = (x1 , . . . , x
n+1 ) → x1 is the measure μn/2,n/2 , and its
p
image measure via (x1 , . . . , xn+1 ) → 2 i=1 xi2 − 1 is μ(n+1−p)/2,p/2 . Moreover,
the spherical heat kernels pt (x, y), which, as we saw in Sect. 2.2, depend only on
the distance d(x, y) between x, y ∈ Sn , may be expressed in terms of the kernel
densities associated with the Jacobi operators. More precisely, if

pt (x, y) = sn t, d(x, y) = ŝn (t, x · y), t > 0, (x, y) ∈ Sn × Sn ,
n/2,n/2 n/2,n/2
then ŝn (t, u) = pt (1, u) where pt (x, y) is the kernel associated with the
Jacobi operator Ln/2,n/2 . Now, the recurrence formula (2.7.13) leads to a recurrence
formula for ŝn , namely
ent
ŝn+2 (t, u) = ∂u ŝn (t, u), (2.7.15)
(n + 1)(n + 2)
very similar to the corresponding formula for the hyperbolic space given in (2.3.2).

Again, this chapter gathers in an informal way various models and results from
different areas and perspectives. The references below only include a few pointers
to the literature.
A standard and most valuable introduction to Euclidean and spherical harmonic
analysis covering most of Sects. 2.1 and 2.2 is the monograph [389]. The more geo-
metric (and Riemannian) aspects of the sphere and hyperbolic models are discussed
in e.g. [159, 194, 240, 260, 346, 354, 409]. The recent [187] is a good source on
hyperbolic space and Laplacian from the probabilistic viewpoint. The explicit ex-
pression for the hyperbolic heat kernels may be found in [144].
Dirichlet and Neumann boundary conditions are standard conditions in the anal-
ysis of partial differential equations on bounded open sets in Rn (cf. [96, 179, 203]).
The probabilistic aspects of Brownian motion killed or reflected at the boundary of
a half-line or an interval, briefly described in Sects. 2.4 and 2.5, may be found in the
standard references on stochastic calculus, including [255, 263, 350, 358].
A recent reference on Sturm-Liouville operators on the line is [447].
The examples of Sect. 2.7 are discussed in most textbooks on orthogonal poly-
nomials including [131, 402] and the recent [268] where many useful formulas are
displayed. The characterization of diffusion operators on the line which are diago-
nalized with respect to orthogonal polynomials, leading to the three examples of the
Hermite, Laguerre and Jacobi semigroups, was achieved by O. Mazet [301]. Recent
developments in higher dimensions are investigated in [41]. Infinite-dimensional
Ornstein-Uhlenbeck operators (Sect. 2.7.2) are presented in [26, 82, 142, 257, 308,
334, 417], and in the context of the analysis of Brownian paths on a manifold
in [251, 297, 391].
Chapter 3
Symmetric Markov Diffusion Operators
The aim of this chapter is to provide a general framework for the investigation of
symmetric Markov diffusion semigroups and operators. It is based on the early ob-
servations of Chap. 1 and on the investigation of the model examples in Chap. 2.
This chapter develops all the fundamental properties which will justify the com-
putations in the following chapters in the most convenient framework. It is not in-
tended to be read linearly, and could be skipped on first reading. The interested
reader should come back to this material when specific technical details have to be
understood (such as, for example, the use of curvature conditions towards gradient
bounds, which lies at the heart of much of the subsequent analysis). A summary of
the framework is presented at the end of the chapter.
The various questions addressed here are actually somewhat delicate. While the
chapter aims at describing a general abstract setup, it clearly cannot encompass all
the instances of interest. Moreover, each specific example often requires an inde-
pendent analysis with suitable assumptions and hypotheses. Therefore, we do not
try to be maximally general at each level, instead emphasizing the tools and meth-
ods which may then be suitably adapted to the cases of interest. These issues in
particular concern the classes of functions on which the various operators and semi-
groups are analyzed. Again, we present the main features of the classes of functions
required to develop the investigation rather than carefully describing each example.
In particular, the exposition freely emphasizes the formal and inspiring arguments
of the -calculus and heat flow monotonicity on which most of the following func-
tional inequalities will be based. Nevertheless, on the basis of the analysis of dif-
fusion operators on complete manifolds in Sect. 3.2, Sect. 3.3 develops a complete
and self-contained framework fully justifying the various results and conclusions
In Chap. 1, we investigated the analysis of Markov semigroups starting from the
knowledge of the semigroup itself (sometimes in a concrete instance), and intro-
duced the infinitesimal generator together with its domain as a tool to describe it.
In this chapter, we adopt the reverse point of view, analyzing the semigroup based
on the properties of an operator (a second order differential operator), or rather its
carré du champ operator , leading to the basic Markov Triple structure (E, μ, )
which will form the natural environment for future investigations.

DOI 10.1007/978-3-319-00227-9_3,
120 3 Symmetric Markov Diffusion Operators
The operator and carré du champ operator are initially defined on some set A0 of
functions (typically the set of smooth compactly supported functions). The analysis
of differential diffusion operators on smooth manifolds leads us to consider a larger
class A (playing the role of smooth functions without support conditions), and to
precisely analyze the relationships between the two classes A0 and A, in particu-
lar with regard to integration by parts formulas. Throughout, when working with
classes of (real-valued) functions (rather than points), it should be kept in mind that
equalities and inequalities between functions should always be understood to hold
almost everywhere (with respect to the underlying invariant measure).
The first section describes the main setting consisting of a triple (E, μ, ) with
state space E, σ -finite measure μ, carré du champ operator acting on an algebra
A0 of functions, and the diffusion property in force throughout this work. This al-
lows us to describe what we call a Standard Markov Triple, which is a convenient
setting allowing us to develop the formalism of Markov semigroups towards func-
tional inequalities and convergence to equilibrium. Section 3.2 presents, for the con-
crete example of second order differential operators on smooth complete connected
manifolds, some of the basic issues, properties and hypotheses needed to work with
diffusion operators and semigroups on natural classes of functions. This analysis
anticipates the further introduction of the extended function algebra A, which en-
ables the translation into the general framework of some central features such as
connexity, completeness, weak hypo-ellipticity etc. With the tool of essential self-
adjointness at the center of the construction, Sect. 3.3 then emphasizes the relevant
hypotheses and properties of Full Markov Triples, covering in particular the exam-
ples illustrated in Chaps. 1 and 2. Section 3.4 summarizes the various hypotheses
and definitions, introducing the Full Markov Triple as the convenient framework in
which most results of the book are presented. The reader may easily refer to it while
progressing through the subsequent chapters.
3.1 Markov Triples

This section presents the basic setting for the analysis of (symmetric diffusion)
Markov generators and semigroups. As already mentioned, when compared with
Chap. 1 and the various model examples in Chap. 2, the definitions put forward here
appear in the reverse order, emphasizing on the underlying invariant measure and
the carré du champ operator. These are actually the central objects of interest in the
investigation of functional inequalities and of their links with curvature conditions.
The exposition below makes clear how this framework encompasses the semigroup
and generator viewpoint.
3.1.1 Initial Structure
The following definition describes the initial setting of investigation. Recall the no-
tion of a good measurable space (p. 7).
3.1 Markov Triples 121
Definition 3.1.1 A triple (E, μ, ) consists of a good measurable space (E, F)

equipped with a σ -finite measure μ, a vector space A0 of (real-valued) bounded
functions dense in all Lp (μ)-spaces, 1 ≤ p < ∞, stable under products (that is A0
is an algebra), and a symmetric bilinear map : A0 × A0 → A0 (the carré du champ
operator) mapping two functions of A0 to an element of A0 , such that (f, f ) ≥ 0
for every f ∈ A0 . The fundamental identity

g, f dμ + 2 g (f, f )dμ = 2 (f g, f )dμ
2
(3.1.1)
E E E
holds for any f, g ∈ A0 . Moreover, there exists an increasing sequence of functions

in A0 converging μ-almost everywhere to the constant function 1.
Most often (f, f ) will be abbreviated as (f ).

The identity (3.1.1) of the previous definition, while curious at first glance, has
a very clear interpretation in the context of Markov chains on a finite set E as pre-
sented in Sect. 1.9, p. 33. Indeed, in this case a carré du champ operator has the
form
1
(f, g)(x) = a(x, y) f (y) − f (x) g(y) − g(x) , x ∈ E,
2
y∈E
for some matrix a(x, y), (x, y) ∈ E × E. Then, if μ(x) > 0 for any x ∈ E, (3.1.1)
is translated into the identity, valid for every pair (x, y) ∈ E × E,
μ(x)a(x, y) = μ(y)a(y, x),
which amounts to the reversibility (or detailed balance)

condition (1.9.1), p. 35, in
L2 (μ) of the associated Markov operator Lf (x) = y∈E a(x, y)[f (y) − f (x)]. If
a(x, y), (x, y) ∈ E × E, does not satisfy this balance condition, the symmetrized
matrix
μ(y)
a (x, y) = a(x, y) + a(y, x), (x, y) ∈ E × E,
μ(x)
does, and if
is then the
associated carré du champ operator, for any function
f : E → R, E (f )dμ = E (f )dμ so that everything is coherent.
It should be emphasized that Definition 3.1.1 involves an algebra A0 on which the
carré du champ operator is acting. This class is most often clear from the context,
although its properties should be carefully described, as will be developed in this
chapter. Typically, on a smooth manifold, A0 will be the set of smooth compactly
supported functions. On the basis of this example, and in order to develop a suitable
calculus on the carré du champ operator, it is necessary to deal with expressions
such as (fg, h) or (ψ(f ), g) for smooth functions ψ on the real line. When the
measure μ is infinite, constant functions are not in L2 (μ) and therefore it is neces-
sary to restrict the functions ψ to those satisfying ψ(0) = 0. It is therefore natural to
assume that A0 is an algebra stable under the action of smooth functions vanishing
at 0. Note that, on the other hand, it is also necessary to assert that (f, 1) = 0, so
that will have to be extended later to a wider class containing constant functions.
Remark 3.1.2 Working with classes of functions rather than points, it should be
emphasized that the preceding setting already ensures that the algebra

A0 is rich
enough in a measurable sense. For example, if f ∈ L1 (μ) and if E gf dμ ≥ 0 for
every positive g ∈ A0 , then f ≥ 0 (μ-almost everywhere). Indeed, given A ∈ F
with μ(A) < ∞, there exists a sequence (fk )k∈N in A0 converging to 1A in L1 (μ),
and therefore, up to the choice of a subsequence, μ-almost everywhere. By the sta-
bility by composition with functions vanishing at 0, gk = 2fk2 (1 + fk2 )−1 , k ∈ N,
forms a sequence of functions in A0 , uniformly bounded from
above and converg-

ing μ-almost everywhere to 1A . By dominated convergence, E gk f dμ → A f dμ

as k → ∞ and the conclusion follows.

In the same spirit, whenever μ(A) > 0, there
exists a positive g in A0 such that A gdμ > 0.
3.1.2 Diffusion Property
The central diffusion hypothesis will be in force in most of this book. While pre-
sented for the generator in Sect. 1.11, p. 42, it is defined here equivalently on the
carré du champ operator.
Definition 3.1.3 (Diffusion property) Let (E, μ, ) be a triple as in Definition 3.1.1

relative to an algebra A0 stable under the action of smooth (C ∞ ) functions vanishing
at 0. The carré du champ operator is said to have the diffusion property if for any
choice f1 , . . . , fk , g of functions in A0 , and any smooth (C ∞ ) function : Rk → R
vanishing at 0,
k
(f1 , . . . , fk ), g = ∂i (f1 , . . . , fk ) (fi , g). (3.1.2)
i=1
Recall that equalities between functions (in A0 here) are understood μ-almost
everywhere. In other words, (f, g) is a derivation on each of its arguments. Re-
stricted to polynomial functions , the preceding boils down to the chain rule
(f g, h) = f (g, h) + g (f, h) (3.1.3)
for any choice of f, g, h in A0 . Under the diffusion hypothesis, it is clear that the
fundamental identity (3.1.1) is satisfied.
3.1.3 Diffusion Operators
On the basis of some of the material of Chap. 1, we now describe how the preceding
framework actually includes the more usual description in terms of semigroups and
generators. This discussion will lead us to enrich the underlying algebra A0 with
several properties which, in addition to the diffusion property, will then form the
basic (diffusion) Markov Triple structure. The various properties of A0 presented
below will be summarized in Sect. 3.4.
Given a triple (E, μ, ) as in Definition 3.1.1 together with the diffusion property
from Definition 3.1.2, the associated symmetric operator L may then be defined, on
A0 , by the integration by parts formula,

g Lf dμ = − (f, g)dμ (3.1.4)
E E
for every f, g ∈ A0 (cf. (1.6.3), p. 27). In order to uniquely define the generator L
on A0 from (3.1.4), it is necessary to assume that for any f ∈ A0 , there exists a
finite constant C(f ) such that for any g ∈ A0 ,

(f, g)dμ ≤ C(f ) g2 (3.1.5)

E

so that g → E (f, g)dμ may be extended to a continuous linear form on L2 (μ).

(Note that Lf is then, as usual, only defined as a class of functions in L2 (μ).)
The next basic requirement on A0 is that for any f ∈ A0 , Lf ∈ A0 . Then, the
fundamental identity (3.1.1) of Definition 3.1.1 yields that, for every f ∈ A0 ,

L f 2 = 2f Lf + 2 (f ) (3.1.6)
(multiply (3.1.1) by g ∈ A0 and integrate against μ). By polarisation,
L(f g) = f Lg + g Lf + 2 (f, g)
for all f, g ∈ A0 .
It is also necessary to impose that

Lf dμ = 0
E
for any f ∈ A0 , what was identified as the invariance (or stationarity) of the mea-
sure μ in Chap. 1. Of course, if constant functions belong to A0 , this is automatically
true due to integration by parts (since then (f, 1) = 0 thanks to the chain rule). It
is however not wise in general to assume that constant functions belong to A0 . This
property will be automatic once the richer class A has been introduced in Sect. 3.3
below. By symmetry of , it is clear that the invariant measure μ is reversible for L
in the sense that

g Lf dμ = f Lg dμ (3.1.7)
E E
for every pair (f, g) ∈ A0 × A0 .
By the diffusion property of the carré du champ operator (Definition 3.1.3),

the operator L satisfies the analogous diffusion property (1.11.4), p. 43, that is
k
L (f1 , . . . , fk ) = ∂i (f1 , . . . , fk ) Lfi
i=1
(3.1.8)

k
+ ∂ij2 (f1 , . . . , fn ) (fi , fj )
i,j =1
for : Rk → R smooth (of class C ∞ ), (0) = 0, and f1 , . . . , fk in A0 . To ver-

ify this assertion, consider for simplicity the case where depends only on one
variable, so that the formula to be satisfied is

L (f ) = (f ) Lf + (f ) (f ) (3.1.9)
(for every f in A0 ). Now, by the integration by parts formula (3.1.4), this amounts
to

(f ), g dμ = f, (f )g dμ − g (f ) (f )dμ
E E E
for every g ∈ A0 , which appears as an immediate consequence of the diffusion
property for . Conversely, starting from a generator L, the change of variables
formula (3.1.9) for (r) = r 2 yields in return the carré du champ operator in the
form

2 (f ) = L f 2 − 2f Lf.
Everything is therefore coherent.
Remark 3.1.4 It is perhaps worth deciphering the relationship between the change
of variables formulas (3.1.2) and (3.1.8) for and L respectively, and the mere chain
rule (3.1.3), as given on p. 43.
On a commutative algebra such as A0 , a first order differential operator (with
no 0 order term), also often called a derivation or a vector field, is a linear op-
erator Z : A0 → A0 satisfying the chain rule Z(f g) = f Zg + gZf . An operator
X : A0 → A0 is a second order differential operator if its associated carré du champ
operator X (f, g) = 12 [X(f g) − f Xg − gXf ] is a first order operator in each ar-
gument. (One could define in this way differential operators of any order.) A first
order operator is then simply a second order operator with vanishing carré du champ
operator. Now, it is quite easy to see that,
for any first order operator Z, the change
of variables formula Z((f1 , . . . , fk )) = ki=1 ∂i (f1 , . . . , fk )Zfi is valid for any
polynomial function with (0) = 0 (extended without this restriction provided
A0 contains a unit 1 for which Z(1) = 0). In the same way, the change of vari-
able formula (3.1.8) is also valid for any second order differential operator X with
polynomial functions . Therefore, the change of variables formula is simply the
extension of the chain rule from polynomials to smooth functions. With these defi-
nitions, L is a second order differential operator, and a first order operator in each
of its arguments. Moreover, the 2 operator (Definition 3.3.12 below) is a second
order differential in each of its arguments, and the Hessian operator H (f )(g, h)
(Definition 3.3.13 below) is of second order in f and of first order in g and h.
3.1.4 Dirichlet Form and Domains
Given a triple (E, μ, ) as in Definition 3.1.1 with algebra A0 satisfying the dif-
fusion property and the preceding requirements, the next step is to understand on
which natural domain the (diffusion) operator L, given on A0 , may be extended.
The first task is to construct the Dirichlet form E and its domain D(E) by completion
from A0 , and then to build the domain D(L) of the generator L from this Dirichlet
domain. This procedure yields a self-adjoint operator L with L2 (μ)-domain D(L),
from which the semigroup P = (Pt )t≥0 with infinitesimal generator L and invari-
ant and reversible measure μ is then constructed following the lines presented in
Appendix A.
To start with, recall the Dirichlet form E introduced in Definition 1.7.1, p. 30, as

E(f, g) = (f, g)dμ = − f Lg dμ, (f, g) ∈ A0 × A0 . (3.1.10)
E E
Also write E(f ) for E(f, f ). Note that by the Cauchy-Schwarz inequality (both for
the quadratic form and in L2 (μ)),

E(f, g) ≤ E(f )1/2 E(g)1/2 .
The Dirichlet form E is a priori defined only on A0 × A0 but may actually be ex-
tended to a wider class, the so-called Dirichlet domain D(E). In fact, if A0 is en-
dowed with the Dirichlet norm
1/2
f E = f 22 + E(f ) , (3.1.11)
we may take the completion of A0 with respect to this norm to turn it into a Hilbert
space embedded in L2 (μ). For this to be possible, the form must be closable, that
is, if a sequence (fk )k∈N in A0 is such that fk 2 → 0 and if E(fk − f ) → 0
when k, → ∞, then E(fk ) → 0. But this is actually automatic with the help of the
operator L. Indeed, whenever k > ,

E(fk ) = − fk Lf dμ + E(fk , fk − f )
E
≤ fk 2 Lf 2 + E(fk )1/2 E(fk − f )1/2 .
Fixing and letting k go to infinity yields the conclusion.

Via this construction, the Dirichlet form E is now defined on a domain D(E)
which contains A0 , and in which A0 is dense (in the topology · E of the domain
as described above). Any function f in D(E) may thus be approximated in this

topology by a sequence (fk )k∈N in A0 .
The bilinear carré du champ operator may be extended to D(E). Indeed, for a
Cauchy sequence (fk )k∈N in D(E), since is positive, for every k, ∈ N,

(fk )1/2 − (f )1/2 2 ≤ (fk − f )
so that the sequence ((fk )1/2 )k∈N is a Cauchy sequence in L2 (μ). This allows us
to define in a proper way
(f ), and then (f, g) for f, g ∈ D(E), as an element of
L1 (μ) such that E(f ) = E (f )dμ.
At this stage, it is certainly worth making the following observation.
Proposition 3.1.5 If f ∈ D(E) and h ∈ A0 , then hf ∈ D(E) (in other words, D(E)
is an A0 -module). More precisely, whenever a sequence (fk )k∈N in A0 converges in
D(E) to f , then the sequence (hfk )k∈N converges in D(E) to hf . Moreover, for any
f, g in D(E) and any h ∈ A0 , (hf, g) = h (f, g) + f (h, g).
Proof From the diffusion property, for f, h ∈ A0 ,

2
E(hf ) = (hf )dμ ≤ 2 h (f ) + f 2 (h) dμ
E E

≤ 2 max h2∞ , (h) ∞ f E .
This inequality applied to an approximating sequence yields the result. The second
claim of the proposition is achieved in the same way.
As an illustration, if the underlying state space E is Rn (or a manifold), and

μ has a density with respect to the Lebesgue measure, D(E) corresponds to
the set of functions with one derivative in L2 (μ) (or more precisely for which
(f ) ∈ L1 (μ)). We shall be making strong use of the Dirichlet domain D(E)
throughout this work. Any formula involving only E(f ), as is the case for many
of the functional inequalities discussed in this book, will automatically extend from
functions in A0 to functions in this Dirichlet domain.
Once the Dirichlet form E and its domain D(E) have been constructed, the op-
erator L may be extended from A0 to its domain D(L) which is defined in the
following way. According to (3.1.5), say that a function f ∈ D(E) belongs to D(L)
if there exists a finite constant C(f ) such that
E(f, g) ≤ C(f ) g2
for every g ∈ D(E). The extension of L to D(L) (again denoted by L) is then

achieved as above via the integration by parts formula (3.1.4). It turns out that such
an operator L is then always self-adjoint and is the infinitesimal generator of a sym-
metric semigroup P = (Pt )t≥0 of contractions in L2 (μ) with invariant reversible
measure μ. Moreover D(L) coincides with the L2 (μ)-domain as the generator of
a semigroup of operators on a Banach space (cf. Sect. 1.4, p. 18, and Sect. A.1,
p. 473). On D(L), the semigroup (Pt )t≥0 with generator L solves the heat equation
∂t Pt = LPt = Pt L
(and accordingly is often called the heat semigroup or heat flow with respect to L).
It now remains to check the positivity preserving and Markov properties of
(Pt )t≥0 , which will be addressed in the next sub-sections.
3.1.5 Positivity Preserving Property
That the associated semigroup (Pt )t≥0 is positivity preserving (i.e. Pt f ≥ 0 when-
ever f ≥ 0) is not at all clear from the above construction. The proof given
via (1.4.5), p. 21, is purely formal and cannot be turned into a real proof without fur-
ther hypotheses on the triple (E, μ, ), thus we require another approach. Indeed,
in the general theory of Dirichlet forms, in order for (Pt )t≥0 (constructed from E)
to be positivity preserving it suffices that the Dirichlet domain D(E) is stable under
the maps f → ψ(f ), where ψ : R → R is a smooth function such that ψ(0) = 0
and |ψ | ≤ 1, and moreover that for such ψ ’s,

E ψ(f ) ≤ E(f ). (3.1.12)

In particular, E(f − ψ(f ), f ) ≥ 0, and whenever f ∈ D(L), E (f − ψ(f ))Lf dμ

≤ 0. Indeed, assuming (3.1.12), let us show that whenever f ∈ L2 (μ) is such
that f ≤ 1, then
∞Pt f ≤ 1. To this end, we introduce, for λ > 0, the resolvent
operator Rλ = 0 e−λs Ps ds, which is bounded on L2 by spectral decomposition
(cf. (A.1.2), p. 474). It is known that Rλ = (λ Id −L)−1 , and that for any f ∈ L2 (μ),
Rλ (f ) ∈ D(L). Moreover
tk k
Pt = lim e−λt λ 2 Rλ
λ→∞ k!
k∈N
which is clear from spectral decomposition (cf. Appendix A), and therefore, in
order to prove that Pt f ≤ 1, it is enough to show that g = λRλ (f ) ≤ 1. Now,
λg = Lg + λf , and from the preceding, for every smooth ψ with ψ(0) = 0 and
|ψ | ≤ 1,

λ g g − ψ(g) dμ = Lg g − ψ(g) dμ + λ f g − ψ(g) dμ
E E E

≤λ f g − ψ(g) dμ.
E
Thus

(f − g) g − ψ(g) dμ ≥ 0.
E
This property, established for any such smooth function ψ, easily extends to the
function ψ(r) = r ∧ 1, r ∈ R, from which

(f − g)(g − 1) dμ ≥ 0.
{g>1}
Since f ≤ 1, it follows that μ(g > 1) = 0. By homogeneity, for any a > 0, when-
ever f ≤ a, we have Pt f ≤ a, and therefore if f ≤ 0, then Pt f ≤ 0 which is the
announced positivity property. The argument ensures in the same way that the as-
sociated semigroup is sub-Markov (that is Pt (1) ≤ 1, where 1 is understood in the
sense of Definition 3.1.1).
Now, restricting to the case of diffusion operators, (3.1.12) easily holds. In-
deed, it is not hard to show that if ψ is smooth with ψ(0) = 0 and ψ bounded,
and if f ∈ D(E), then ψ(f ) ∈ D(E) and (ψ(f )) = ψ 2 (f )(f ). To see why
ψ(f ) ∈ D(E), for a sequence (fk )k∈N ∈ A0 converging to f ∈ D(E) in the Dirichlet
norm and μ-almost everywhere, and for k, ∈ N, we have

E ψ(fk ) − ψ(f ) ≤ ψ (fk )2 (fk − f ) + ψ (fk )2 − ψ (f )2 (fk )
E

+ 2ψ (fk )ψ (fk ) − ψ (f )(fk )1/2 (f )1/2 dμ.
Since the sequence ((fk ))k∈N is uniformly integrable, it follows that (ψ(fk ))k∈N
is a Cauchy sequence in D(E) and therefore convergent. Hence (3.1.12) is satisfied,
as announced.
Recall that the semigroup just constructed is moreover sub-Markov (that is, it
is positivity preserving and satisfies Pt (1) ≤ 1) rather than Markov (Pt (1) = 1).
Criteria for the latter to hold (Definition 3.1.12) will be presented later.
The preceding construction is the so-called Friedrichs self-adjoint extension of
the operator L initially given on A0 as described in Sect. A.3, p. 477, for which
A0 ⊂ D(L) ⊂ D(E).
Basically, the domain D(E) of the Dirichlet form is the L2 (μ)-domain of (−L)1/2 .
3.1.6 Semigroup and Spectral Decomposition
As already emphasized in Sect. 1.4, p. 18, and Sect. 1.6, p. 24, and above, the con-
struction of the Dirichlet form E with domain D(E) and of the self-adjoint operator
L with L2 (μ)-domain D(L) yields a Markov or sub-Markov semigroup P = (Pt )t≥0
on (E, F) with infinitesimal generator L and invariant reversible measure μ. The
semigroup viewpoint has several advantages and actually allows us to identify fur-
ther properties of interest.
Consider, for example, for a given function f in L2 (μ) and t > 0, the expression

1 1
f (f − Pt f )dμ = f 2 dμ − (Pt/2 f )2 dμ . (3.1.13)
t E t E E
By (1.7.2), p. 30, this expression is decreasing in t. The domain D(E) of the Dirich-
let form may then be described equivalently as the set of functions f ∈ L2 (μ) for
which the limit of this expression is finite when t → 0. If f is in the L2 (μ)-domain
D(L) of the generator L with semigroup P = (Pt )t≥0 , it is
also in the Dirichlet do-
main D(E) by the integration by parts formula E(f ) = − E f Lf dμ.
The expression (3.1.13) may moreover be rewritten using the kernels pt (x, dy)
of the representation (1.2.4), p. 12, as

1 2
f (x) − f (y) pt (x, dy)dμ(x). (3.1.14)
2t E E
In this form, we now observe that the expression is decreasing if f is replaced by

ψ(f ) where ψ : R → R is a contraction in the sense that |ψ(r) − ψ(s)| ≤ |r − s|,
r, s ∈ R. Therefore, the Dirichlet domain D(E) is stable under the action of such
contractions ψ and, for any f in D(E),

E ψ(f ) ≤ E(f ). (3.1.15)
With respect to (3.1.12), the latter therefore also covers contractions ψ which are not
necessarily smooth. Moreover, using Fatou’s lemma, whenever a sequence (fk )k∈N
in D(E) converges in L2 (μ) to f ∈ D(E), then, since the expression (3.1.14) is
decreasing in t,
E(f ) ≤ lim inf E(fk ). (3.1.16)
k→∞
In particular, for any a ∈ R, there is a sequence of smooth approximations

ψk (r), k ≥ 1, of the function r → r such that ψk vanishes on (a − k1 , a + k1 ).
Applying
(3.1.16) to fk = ψk (f ), k ≥ 1, then yields that, for any f ∈ D(E),
E(f ) ≤ {f =a} (f )dμ. Therefore, for every f ∈ D(E) and every a ∈ R,

(f )dμ = 0. (3.1.17)
{f =a}
This result captures the fact that (f ) = 0 as long as f remains constant. This will
be further developed in Chap. 8 as a weak co-area formula.
In this language, the contraction property (3.1.15) along the Dirichlet form may
be translated into a contraction property for itself (even without the diffusion
assumption). Indeed, for f ∈ A0 and tk → 0, μ-almost everywhere in x ∈ E,

1 2
(f )(x) = lim Ptk f (x) − Ptk (f )(x) .
2
k→∞ 2tk

But, if Pt f (x) = E f (y)pt (x, dy), this last expression may be written as

1 2
f (y1 ) − f (y2 ) ptk (x, dy1 ) ptk (x, dy2 ),
4tk E E
and is therefore clearly decreasing under the action of a contraction ψ . Hence, for
any ψ : R → R such that |ψ(r) − ψ(s)| ≤ |r − s|, r, s ∈ R, and every f ∈ A0 ,

ψ(f ) ≤ (f ), (3.1.18)
which then extends to functions in D(E).

We already observed that when ψ is smooth with bounded derivative, and un-
der the diffusion property, the formula (ψ(f )) = ψ (f )2 (f ) also immediately
extends from A0 to D(E). Furthermore, this point of view also yields immediate ac-
cess to the expression of in terms of the generator as expressed by (3.1.6). Indeed,
under the ESA property of Definition 3.1.10, any f ∈ D(L) is such that f 2 belongs
to the L1 (μ)-domain and

L f 2 = 2f Lf + 2 (f ) (3.1.19)
as is easily checked using any sequence in A0 which converges to f in the domain

topology. In particular, for any f ∈ L2 (μ) and any t > 0, (Pt f )2 belongs to the
L1 (μ)-domain.
The quadratic form (f ) may be thought of in different ways according as
f ∈ A0 or only in D(E), but the properties of which may be checked on A0
are in general easily extended to D(E), provided some care is taken. For example,
let ψ : R → R be smooth so that ψ is bounded. When A0 is dense in D(L), if
f ∈ D(L), then ψ (f ) ∈ L2 (μ) and ψ(f ) ∈ L1 (μ). In this case, ψ(f ) also belongs
to the L1 (μ)-domain of L (defined in Sect. 1.4.1, p. 18) and

L ψ(f ) = ψ (f ) Lf + ψ (f ) (f ).
Again, this identity is proved by approximation by functions in A0 exactly as

in (3.1.19) above (corresponding to ψ(r) = r 2 ). We thus recover along these lines
the diffusion property for L.
This semigroup language will feature in numerous basic proofs relying on in-
terpolation along (Pt )t≥0 via a formula sometimes known as Duhamel’s formula,
at the heart of heat flow monotonicity or semigroup interpolation. A typical exam-
ple, already used in (1.7.2), p. 30, of particular use and importance throughout this
monograph, is the following. Fix t > 0, and consider, for f ∈ L2 (μ),

(s) = Ps (Pt−s f )2 , s ∈ [0, t]
(at almost any point, omitted as usual). Then, from (3.1.19),

(s) = Ps L(Pt−s f )2 − 2Pt−s f LPt−s f = 2Ps (Pt−s f ) (3.1.20)
in L1 (μ). Taking integrals between 0 and t yields a useful variance representation

along the semigroup,
t

Pt f 2 − (Pt f )2 = 2 Ps (Pt−s f ) ds. (3.1.21)
0
Such representations and arguments will be developed at length in Chaps. 4, 5 and 6

in order to establish heat kernel inequalities.
The spectral decompositions of L and its associated semigroup P = (Pt )t≥0 are
further useful tools towards the description of the domains D(E) and D(L), as well
as of the semigroup (Pt )t≥0 itself. Several elementary properties in this regard are
summarized in the next proposition.
∞
Recall the spectral decomposition −L = 0 λ dEλ of the positive self-adjoint
operator −L on L2 (μ) as developed
in Sect. A.4, p. 478. Also recall from
∞
there that f ∈ L2 (μ) if and only if 0 dEλ f, f < ∞, and that whenever

∞
2
∞
g = 0 h(λ) dEλ (f ), we have E g dμ = 0 h2 (λ) dEλ f, f .

∞
Proposition 3.1.6 Let −L = 0 λ dEλ be the spectral decomposition of the oppo-
site of the self-adjoint operator L on L2 (μ). The following assertions hold.
(i) The domain D(L) of L is the set of functions f ∈ L2 (μ) such that
∞
(Lf )2 dμ = λ2 dEλ f, f < ∞
E 0

∞
and −Lf = 0 λ dEλ f .
(ii) The Dirichlet domain D(E) is the set of functions f ∈ L2 (μ) such that
∞
E(f ) = λ dEλ f, f < ∞.
0
(iii) For every t ≥ 0 and every f ∈ L2 (μ),

∞
Pt f = e−λt dEλ f.
0
In particular, for any t > 0, Pt f ∈ D(L) and

1
(LPt f ) dμ ≤ 2 2
2
f 2 dμ.
E e t E
(iv) For every t ≥ 0 and every f ∈ L2 (μ),

∞
E(Pt f ) = λe−2λt dEλ f, f .
0

Therefore E(Pt f ) ≤ 1
2et Ef
2 dμ for t > 0.
(v) For every f ∈ L2 (μ), the map t → E(Pt f ) is decreasing.
More or less, all items in this proposition are immediate from the definition of the
spectral decomposition. (iii) is a simple consequence of the bound r 2 e−r ≤ 4 e−2 on
R+ , (iv) follows from the bound re−r ≤ e−1 , while (v), which follows from the fact
that t → λe−2λt is decreasing, may be seen more directly using that

∂t E(Pt f ) = −∂t Pt f LPt f dμ = −2 (LPt f )2 dμ ≤ 0.
E E
It may finally be pointed out that Proposition 3.1.6 takes a much simpler form
in the case of a discrete spectrum. Indeed, in this casethere exists a sequence
(Ek )k∈N of orthogonal closed subspaces of L2 (μ) with k∈N Ek = L2 (μ) which
are eigenspaces of L associated with the eigenvalues
−λk , λk ≥ 0, k ∈ N. When a
function f ∈ L2 (μ) is decomposed into f = k∈N fk , then

−Lf = λk f k , Pt f = e−λk t fk , E(f ) = λk fk 22
k∈N k∈N k∈N
and so on.
Remark 3.1.7 Using the same arguments as in Proposition 3.1.6, it may be shown
that if f ∈ L2 (μ), then Pt f ∈ D(Lk ), t > 0, for any k ∈ N. Therefore, at the L2 (μ)
level, even in the absence of any hypo-ellipticity condition, the semigroup is smooth-
ing as an effect of symmetry only. Moreover, if f ∈D(E), then Pt f converges to f
in D(E) as t → 0. This is also true if f ∈ D(L) as k D(Lk ) is dense in D(Lp ) for
any p ∈ N, again by means of the smoothing properties of the semigroup.
3.1.7 Diffusion Markov Triple
At this point of the construction, we introduce the notion of a symmetric Diffusion

Markov Triple which summarizes the hypotheses on the triple (E, μ, ) and the un-
derlying algebra A0 presented so far. Section 3.4 below will summarize the various
hypotheses on the algebra A0 . Although the semigroup (Pt )t≥0 is not yet Markov
(mass conservative), the somewhat abusive terminology at this point is convenient.
Definition 3.1.8 (Diffusion Markov Triple) A Diffusion Markov Triple is a triple

(E, μ, ) as in Definition 3.1.1 with the diffusion hypothesis and linear operator
L : A0 → A0 defined by and satisfying the integration by parts formula

g Lf dμ = − (f, g)dμ
E E
for every f, g ∈ A0 . The measure μ is invariant and reversible with respect to L.

The associated Dirichlet form E is defined by

E(f, g) = (f, g)dμ
E
for every (f, g) ∈ A0 × A0 . The domain D(E) of the Dirichlet form E is the com-
pletion of A0 with respect to the norm f E = [f 22 + E(f )]1/2 , and E is extended
to D(E) by continuity together with the carré du champ operator . The symmetric
semigroup of contractions P = (Pt )t≥0 with infinitesimal generator L defined on its
L2 (μ)-domain D(L) is positivity preserving but not necessarily Markov (in general
it is only sub-Markov, i.e. Pt (1) ≤ 1).
3.1.8 Essential Self-adjointness
The following addresses the question of essential self-adjointness which will turn
out to be central in the abstract construction developed in Sect. 3.3 below, in partic-
ular concerning mass conservation. Besides the operator L constructed so far, it is
necessary to introduce an adjoint operator L∗ , where the duality refers both to the
measure μ and the algebra A0 .
Definition 3.1.9 (Adjoint operator) A function f ∈ L2 (μ) belongs to D(L∗ ) if

there exists a finite constant C(f ) such that for any g ∈ A0 ,

f Lg dμ ≤ C(f ) g2 .

E
In this situation, L∗ f is the unique element of L2 (μ) such that for any g ∈ A0 ,

f Lg dμ = g L∗ f dμ.
E E
By symmetry of L (cf. (3.1.7)), we always have that D(L) ⊂ D(L∗ ), and L∗ = L

on D(L). However, while L is self-adjoint on D(L), this does not mean that L = L∗
(L∗ being defined from A0 ). The self-adjointness property of L refers to the dual
operator constructed in the same way as L∗ , but where D(L) replaces A0 . In other
words, it is not true in general that the algebra A0 is dense in D(L) with respect to
the norm
1/2
f D(L) = f 22 + Lf 22 . (3.1.22)
This happens if and only if D(L∗ ) = D(L). In this case, the extension of L from
A0 to a larger domain as a self-adjoint operator is unique. This property will thus
play a central role our analysis, and is called essential self-adjointness, which there-
fore depends on the choice of A0 . The density of A0 in D(L) is equivalent to the
condition that the operator L defined on A0 has a unique self-adjoint extension, a

property also often referred to as uniqueness extension (see Appendix A, p. 476).
Definition 3.1.10 (Essential self-adjointness—ESA) The operator L is said to be

essentially self-adjoint if D(L) = D(L∗ ) (recall that D(L∗ ) is defined with respect
to A0 ). Equivalently, A0 is dense in D(L) in the D(L)-topology, that is A0 is a core
for L.
In order to ensure the ESA property, extra conditions are needed. The next sec-
tion will demonstrate how this property holds for second order elliptic differential
operators on smooth complete connected manifolds where A0 is the class of smooth
compactly supported functions. Section 3.3 will address this issue in the present ab-
stract framework.
Examples in Chap. 2 show that there may be many different self-adjoint exten-
sions L of a given operator L defined on A0 , and conditions for the ESA property
to hold may depend on the behavior of the measure at the boundary of the domain
(Sect. 2.4, p. 92, and Sect. 2.6, p. 97). In general, for various self-adjoint exten-
sions L, their domains D( L) are not comparable. To each self-adjoint extension L
corresponds a Dirichlet form with domain D(E) and an associated semigroup (not
necessarily Markov), using the procedure described by (3.1.13) below. In contrast
to the different domains D( L) however, the associated Dirichlet domains D(E) may
be compared. The domain D(E) of the Friedrichs extension that we constructed is
the minimal one, and corresponds in practice to operators with the Dirichlet bound-
ary conditions. There is also a maximal extension corresponding to the Neumann
boundary conditions. In the context developed here, namely if the operator is essen-
tially self-adjoint, the minimal and maximal extensions coincide and there is only
one Dirichlet domain (referred to as the unique extension).
3.1.9 Ergodicity and Mass Conservation
It is of great importance in many applications to understand the behavior of the

semigroup Pt as t → ∞. From the previous spectral decomposition, it is clear that
for any function f ∈ L2 (μ), Pt f converges in L2 (μ) to the projection of f on the
space of functions which satisfies Lf = 0 (see (iii) in Proposition 3.1.6). In many
natural situations, such a function will be constant, and therefore vanishes when the
measure is infinite. This property will be called ergodicity (see Definition 3.1.11
below). When the measure μ is finite (and therefore

under our convention a prob-
ability measure), this projection should be E f dμ. But for this to hold, it would
require the constant function 1 to be in the domain of the operator, which implies
that Pt (1) = 1. (Recall that according to Definition 3.1.1, 1 is understood as the in-
creasing limit of a sequence of elements in A0 .) This property is known as mass
conservation (or, depending on the context, Markov, stochastic completeness or
non-explosion). Even in the presence of the ESA property, it is not always satis-
fied. For instance, Sect. 2.6, p. 97, gives examples of operators which are essentially
self-adjoint but nevertheless not conservative.
These observations lead to the following two definitions in the setting of a
Diffusion Markov Triple (E, μ, ) with associated operator L and semigroup
P = (Pt )t≥0 .
Definition 3.1.11 (Ergodicity) The operator L is said to be ergodic if every

f ∈ D(L) such that Lf = 0 is constant (in particular f = 0 if μ has infinite mass).
Equivalently, any f ∈ D(E) such that (f ) = 0 is constant.
To verify the equivalence of the two possible definitions, note that if f ∈ D(L) is
such that Lf = 0, then E(f ) = 0 and therefore (f ) = 0. Conversely, if f ∈ D(E)
is such that (f ) = 0, then for any h ∈ D(E), E(f, h) = 0, and therefore f belongs
to D(L) with Lf = 0.
Definition 3.1.12 (Mass conservation) The semigroup P = (Pt )t≥0 is said to be

conservative if for every t ≥ 0, Pt (1) = 1.
Observe that since Pt is defined on any bounded measurable function, the def-
inition makes sense (and is in fact of great importance) even when μ(E) = ∞.
Section 3.2.4 below presents a useful criterion ensuring mass conservation for sec-
ond order differential operators on a complete Riemannian manifold. Section 3.3
further develops the abstract framework to deduce this property from essential self-
adjointness and curvature conditions. On the other hand, when solving a stochastic
differential equation, it is sometimes easy to directly check mass conservation. In
particular, we presented in the second chapter (Sect. 2.6, p. 97) criterions in dimen-
sion one which are almost necessary and sufficient conditions for this to happen.
From the preceding, the following proposition is immediate.
Proposition 3.1.13 Assume that ergodicity holds.

(i) If μ is infinite, for any f ∈ L2 (μ), limt→∞ Pt f = 0 in L2 (μ).
(ii) If μ is finite and the semigroup is conservative, for any f ∈ L2 (μ),

lim Pt f = f dμ in L2 (μ).
t→∞ E
This proposition will be related later to a connexity property (Sect. 3.2.1 and
Proposition 3.3.10).
Remark 3.1.14 It may be observed that under the ergodicity and the ESA prop-
erties, the set {Lh ; h ∈ A0 } is dense in the space L20 (μ) of functions in L2 (μ)
which are orthogonal to constants (that is L
2 (μ) itself when μ(E) = ∞). Indeed,
if f ∈ L2 (μ) is such that, for any h ∈ A0 , E f Lhdμ = 0, then f ∈ D(L∗ ) with
L∗ f = 0. From the ESA property, f ∈ D(L) and Lf = 0. Ergodicity allows us to
conclude that f is constant. It is then 0 when μ(E) = ∞. When μ(E) = 1, since
for any h ∈ A0 , Lh ∈ L20 (μ), the same conclusion follows working in L20 (μ) instead
of L2 (μ).
3.1.10 Standard Markov Triple
At this point of the construction, we emphasize the definition of a Standard Markov

Triple which is essentially a Diffusion Markov Triple with the ergodicity and mass
conservation properties. To lighten the terminology, “Diffusion” is not repeated. The
complete definition will be summarized in Sect. 3.4 below.
Definition 3.1.15 (Standard Markov Triple) A Standard Markov Triple is a Diffu-

sion Markov Triple (E, μ, ) with an algebra A0 (the properties of A0 being sum-
marized in Sect. 3.4) such that the associated operator L and semigroup P = (Pt )t≥0
are ergodic and Markov (conservative).
The structure of a Markov Triple induces on the probabilistic side correspond-

ing Markov processes. However, some care has to be taken in the description of
the laws of these processes and their regularity properties on classes of functions.
Indeed, as described in the first chapter, Sect. 1.3, p. 16, once the Markov semi-
group P = (Pt )t≥0 is constructed, one may consider the associated Markov process
X = {Xtx ; t ≥ 0, x ∈ E}, or more precisely the law of the process (Xtx )t≥0 starting
from any point x ∈ E. In the preceding context, recall that elements of A0 are not
functions, but classes of functions with respect to the relation of μ-almost every-
where equality. Therefore, properties such as continuity or regularity of the pro-
cesses (f (Xtx ))t≥0 for functions f ∈ A0 have no meaning per se. What is clearly
defined however is the law of the process when the initial distribution μ0 of X0 is
1A
absolutely continuous with respect to μ, for example dμ0 = μ(A) dμ for any A such
that 0 < μ(A) < ∞. The law of (Xt )t≥0 with initial condition X0 = x is then de-
fined up to sets of μ-measure 0 as conditional expectations. In practice, most of the
time the state space is actually a topological space, and moreover is often locally
compact. The conditional laws of (Xtx )t≥0 then have a version which is continu-
ous (in x) for the weak topology. In the abstract context, while it is meaningless
to talk about the law of (Xtx )t≥0 for any x ∈ E, when the measure decomposition
theorem applies, however, it makes sense for μ-almost every x. In particular, an
identity such that Pt f (x) = E(f (Xt ) | X0 = x) is understood μ-almost everywhere
as already emphasized.
3.1.11 Intermediate Value Theorem and Slicing
We close this section with two further properties which are relevant in this context.
Even though there is no topology on the state space E, functions in D(E) share
some minimal regularity as a consequence of the diffusion property. In particular
they satisfy a measurable form of the intermediate value theorem.
Theorem 3.1.16 (Intermediate value) Let (E, μ, ) be an ergodic Diffusion

Markov Triple. Let f ∈ D(E). If, for some a < b, μ(f ≤ a) > 0 and μ(f ≥ b) > 0,
then for any a < c < d < b, μ(f ∈ [c, d]) > 0.
3.2 Second Order Differential Operators on a Manifold 137
Proof Assume that μ(f ∈ [c, d]) = 0 and choose ψ : R → R smooth and constant
outside the interval [c, d] with ψ(a) = ψ(b), such that ψ (f ) = 0. Since by the
diffusion property

ψ(f ) dμ = ψ (f )2 (f )dμ = 0,
E E
ψ(f ) is constant μ-almost everywhere, contradicting the hypotheses μ(f ≤ a) > 0

and μ(f ≥ b) > 0. The theorem is established.
The slicing procedure described in the next result will prove to be particularly
useful.
Proposition 3.1.17 (Slicing) Let (E, μ, ) be an ergodic Diffusion Markov Triple.

For any strictly increasing sequence (ak )k∈N of strictly positive numbers, given a
function f : E → R+ , set fk = (f − ak )+ ∧ (ak+1 − ak ), k ∈ N. Then, if f ∈ D(E),

E(fk ) ≤ E(f ). (3.1.23)
k∈N
The proof of this result is an easy consequence of (3.1.16) and the change
of variable formula. Indeed, using some smooth approximation of the function
r → (r − ak )+ ∧ (ak+1 − ak ), and (3.1.16), it is easily seen that

E(fk ) ≤ (f )dμ.
{f ∈(ak ,ak+1 )}
An important feature used here is that (f ) = 0 on the set {f = a}, which is a
consequence of (3.1.17).
3.2 Second Order Differential Operators on a Manifold

Throughout this book, manifolds mean smooth connected manifolds (not neces-
sarily complete since, although we do not investigate boundary behaviors, specific
drifts as observed in Chap. 2 can make the second order differential operator under
investigation essentially self-adjoint).
In this section however, to illustrate the general setting presented in the previous
section, we concentrate on the example of a (smooth connected) complete Rieman-
nian manifold equipped with the class A0 of smooth (C ∞ ) compactly supported
functions in order to identify the precise issues that need to be addressed in the
definition of a Markov Triple (E, μ, ). (The next section will cover non-complete
instances at the expense of a deeper investigation.) The task will then be to translate
directly to the Markov generator L and its carré du champ operator most of the
required properties on the manifold, in order to ensure in particular properties such
that the density of A0 in the domain D(L) or the Markov property Pt (1) = 1. In
what follows, the manifold could of course be Rn , or some open set in Rn . Indeed,
as soon as the coefficients g ij (x) in (3.2.1) below are not constant, the relevant
quantities do not differ much in Rn or in a general manifold. This section makes
significant use of the material presented in Appendices A and C.
In the manifold context, here and throughout, the algebra A0 consists of the
smooth C ∞ compactly supported functions. The degree of smoothness may actu-
ally depend on the smoothness of the manifold, however it will be at least C 2 . For
simplicity, we only consider C ∞ manifolds and function algebras, for which the
generic terminology “smooth” will be used. More general instances may be adapted
accordingly.
We start with some general observations. As already pointed out in the last sec-
tion, when dealing with diffusion operators L and semigroups (Pt )t≥0 on manifolds,
classes of functions which are larger than the class A0 , such as the class of smooth
functions without special growth conditions, may be used. It is then necessary to be
cautious when using integration by parts. Indeed, as discussed in Sect. 1.6, p. 24,
for a symmetric diffusion operator L on a smooth manifold, the integration by parts
formula holds as soon as f and g are smooth and compactly supported, and it is
actually enough that one of the functions be compactly supported. If neither of the
smooth functions f and g is compactly supported, the validity of the formula re-
quires some knowledge on the behavior of the functions f and g at infinity depend-
ing on the underlying (reversible) measure μ. The statement that a function f is
in the domain D(L) of L hides two different aspects, a regularity assumption, and
a growth condition. Moreover, in the presence of boundaries, there are additional
requirements at the boundary (for example, normal derivatives at the boundary have
to vanish in the case of Neumann boundary conditions).
Similar comments hold for the associated semigroup P = (Pt )t≥0 . Indeed, when
the manifold is not compact, Pt f , t > 0, would never be compactly supported even
if f is. But the smoothness of Pt f for smooth compactly supported functions f is
in general a consequence of two types of criteria: when the semigroup is given as the
law of the solution of a stochastic differential equation with global properties such as
non-explosion (see Sect. B.4, p. 495), or from local considerations as a consequence
of ellipticity or hypo-ellipticity (see Sect. 1.12, p. 49).
It is therefore important to determine a setting in which the integration by parts
formula (3.1.4) and other properties hold beyond the preceding class A0 . For ex-
ample, in the core of the analysis, one is often led to consider quantities such as
(Pt f ). If we restrict ourselves to smooth functions f in the L2 (μ)-domain of L,
(Pt f ) is certainly not in general in L2 (μ). Furthermore, unless L may be defined
on any smooth function, the meaning of L(Pt f ) is not entirely clear. What would
then guarantee for a function such as g = (Pt f ) that ∂t Pt g = Pt (Lg)? Indeed,
coming back to the basic examples of symmetric diffusions with density kernels as
presented in the preceding chapters, so that for t > 0 and x ∈ E,

E
and Lx pt (x, y) = Ly pt (x, y), then, assuming that derivation and integration com-
mute,

∂t Pt f (x) = f (y) Lx pt (x, y)dμ(y)
E

= f (y) Ly pt (x, y)dμ(y) = Lf (y)pt (x, y)dμ(y).
E E
But the last step here is in fact integration by parts and requires some a priori knowl-
edge on the behavior with respect to μ of both pt (x, y) and f (y) at infinity.
As mentioned above, this discussion may also be illustrated in probabilistic terms
on the diffusion semigroups given by solutions of stochastic differential equations
as in Sect. 1.10, p. 38. From the representation (1.1.1), p. 8, of the semigroup in
terms of the expectation E(f (Xtx )), it is clear that as soon as the associated Markov
process X = {Xtx ; t ≥ 0, x ∈ E} is continuous and f and Lf are bounded and con-
tinuous, then ∂t Pt f = Pt Lf . Fortunately, with a minimum of assumptions on the
state space emphasized below, this property still holds for smooth functions f in
L2 (μ) as soon as Lf is also in L2 (μ).
Motivated by this discussion, the two main questions and properties which should
be addressed consistently are therefore the uniqueness of the symmetric extension
of L given the algebra A0 and the mass conservation (Markov) property Pt (1) = 1.
The uniqueness or essentially self-adjointness question (Definition 3.1.10), that is
whether A0 is a core for the operator L (cf. Sect. A.5, p. 481), is an important issue
since, when satisfied, any formula valid in A0 and continuous under the topology
of the domain D(L) would still hold on this domain. The Markov property, on the
other hand, allows for the correct probabilistic setting and interpretations.
The aim of this section is to establish these two properties in the context of sec-
ond order diffusion operators on a smooth complete connected manifold M, for the
class A0 of smooth (C ∞ ) compactly supported functions. Therefore, throughout this
section, we deal with a second order differential operator L acting in a local system
of coordinates on functions f : M → R in A0 as

n
n
Lf = g ij ∂ij2 f + bi ∂i f (3.2.1)
i,j =1 i=1
where (g ij (x))1≤i,j ≤n is a symmetric positive-definite matrix and (bi (x))1≤i≤n a

vector, both depending smoothly on x ∈ M. The (elliptic) operator L is furthermore
assumed to be symmetric with respect to a (σ -finite) measure μ on the Borel sets of
M. In Rn or a Riemannian manifold, μ will be assumed to have a smooth strictly
positive density with respect to the Lebegue or Riemannian measure. As announced,
this analysis deals simultaneously with another algebra of functions, namely the
class A of smooth (C ∞ ) functions on M, which is an extension of A0 , and to which
there is in general no difficulty in extending L and the associated carré du champ
operator (provided the coefficients defining L are smooth enough). Recall that, on
such functions, takes the form

n
(f, g) = g ij ∂i f ∂j g, (3.2.2)
i,j =1
and clearly L and satisfy the diffusion property of Definition 3.1.3. The associated
semigroup P = (Pt )t≥0 with generator L as constructed from the Dirichlet form
in Sect. 3.1 may already be assumed to be sub-Markov (Pt (1) ≤ 1).
The exposition here in the manifold case points towards the general setting de-
veloped in the next section. We try to identify the various properties used on the
manifold and the differential operator in order to express them directly in terms of
the operator and its carré du champ operator. The hypotheses put forward in Sect. 3.3
fully rely on this main example. The three basic properties required to achieve this
program and uniqueness of the symmetric extension are connexity, completeness
and (weak-) hypo-ellipticity, which are discussed in the following subsections.
3.2.1 Connexity
Recalling that (f ) stands for |∇f |2 (for Laplacians on manifolds, for example), a
minimum requirement is that if f : M → R is constant then (f ) = 0. This is an
easy consequence of the diffusion property (3.1.8). Conversely, it is also a natural
requirement that if a function f satisfies (f ) = 0, then it is constant. As illustrated
next, this property actually hides the connexity of the underlying state space together
with some intrinsic form of hypo-ellipticity.
In the context of finite Markov chains (see Sect. 1.9, p. 33), connexity is equiv-
alent to the fact that there is only one recurrence class, or in other words that the
Markov chain has a strictly positive probability of moving from any point to any
other point in a finite number of steps. In some abstract situation, connexity to-
gether with the diffusion property implies some minimal continuity for functions in
the Dirichlet space. It shows in particular that the diffusion property may never hold
on a discrete space.
If a diffusion operator L is elliptic in the sense of (1.12.2), p. 50, a function
f : M → R satisfying (f ) = 0 is locally constant. As announced, that locally
constant functions are constant entails connexity of the state space. For example,
if M is just two copies of R, that is M = R × {−1, +1} and if, for = ±1,
Lf (x, ) = ∂x2 f (x, ), then it is clear that any function f satisfying (f ) = 0 is
constant on each copy of R (each connected component of M) but may not be glob-
ally constant.
The picture may be even more complicated. Consider, on R2 , the operator
Lf = ∂x2 f for which (f ) = (∂x f )2 so that the connexity property does not hold.
Note that L is not hypo-elliptic in the sense described in Sect. 1.12, p. 49. How-
ever, in this context, the situation is rather simple. The variable y is just a parame-
ter, and the associated Markov process (Xt , Yt )t≥0 starting from (x, y) amounts to
(Xt , y)t≥0 where (Xt )t≥0 is a one-dimensional Brownian motion (with speed 2)
starting from x. The state space may thus be restricted to R × {y} and L may
be considered as an elliptic operator on R satisfying the connexity condition.
Let now Lf = ∂x2 f + x∂y f for which, still, (f ) = (∂x f )2 so that the opera-
tor is still
non-connected. But now, the associated Markov process is of the form
t
(Xt , y + 0 Xs ds)t≥0 and no longer preserves the second component. It is therefore
not possible to restrict here the analysis to R × {y}, and the connexity property does
not hold although the operator is now hypo-elliptic.
It may be thought that connexity is related to ellipticity. The following example
indicates that it is not the case. On R3 , consider the two vector fields
y x
Z1 = ∂x − ∂z , Z2 = ∂y + ∂z , (3.2.3)
2 2
and the operator L = Z12 + Z22 . This operator is not elliptic, but connected. Indeed,
(f ) = (Z1 f )2 + (Z2 f )2 , and a smooth function f which satisfies (f ) = 0 is
such that Z1 f = Z2 f = 0. Now the commutator [Z1 , Z2 ] = Z3 is Z3 f = ∂z f so
that f also satisfies Z3 f = 0 and is therefore constant. This model is classically
known as the Heisenberg (hypo-elliptic) operator.
Connexity will play a fundamental role in many functional inequalities below (re-
lated to tightness), and will be used to establish convergence to equilibrium (ergod-
icity). An abstract presentation of the connexity property will be further developed
in Sect. 3.3.
3.2.2 Completeness and Weak Hypo-ellipticity
Uniqueness of the self-adjoint extension (the ESA property of Definition 3.1.10) of

second order differential operators L of the form (3.2.1) on a manifold M will be
achieved by completeness of M and hypo-ellipticity of the operator. As illustrated
in Sects. 2.4 and 2.6, p. 92 and p. 97, there are situations where completeness is not
satisfied while ESA actually holds.
The standard completeness assumption on a Riemannian manifold (M, g) refers
to the associated Riemannian metric (cf. Sect. C.5, p. 511). To translate complete-
ness in our setting, we use an argument given in Sect. C.4, p. 509. Namely, according
to Proposition C.4.1, p. 511, on a complete Riemannian manifold (M, g), there ex-
ists an increasing sequence (ζk )k∈N of smooth compactly supported functions such
that limk→∞ ζk = 1 and |∇ζk | ≤ k1 for every k ≥ 1. Therefore, completeness in our
setting will refer to the existence of a sequence (ζk )k≥1 of positive functions in A0 ,
increasing to 1 and such that (ζk ) ≤ k1 for every k ≥ 1.
The second important property used here is hypo-ellipticity of the operator L as
presented in Sect. 1.12, p. 49. With respect to the general definition, we use this
local property here in the sense that any solution of L∗ f = λf is smooth. Recall
that L∗ is the adjoint of L on A0 with respect to μ as presented in Definition 3.1.9,
that is L∗ f = g if and only if for all h ∈ A0 ,

f Lh dμ = g h dμ.
M M
This weak form of hypo-ellipticity, again expressed in an intrinsic form, is actually

enough for later purposes (see Definition 3.3.8 below for an abstract formulation).
The following proposition links completeness and self-adjointness for hypo-
elliptic operators.
Proposition 3.2.1 Assume that L is a hypo-elliptic diffusion operator of the

form (3.2.1), symmetric with respect to μ, on a smooth complete connected man-
ifold M. Then, the operator L is essentially self-adjoint on the set A0 of smooth
compactly supported function. In other words, A0 is a core for the domain D(L) of
L, that is, the ESA property holds.
Proof In order to see that the operator L is essentially self-adjoint on A0 , we make

use of the criterion described in Proposition A.5.3, p. 482. (This criterion has already
been used in the proof of Proposition 2.4.1, p. 95.) Namely, the aim is to prove that
for some λ > 0 any L2 (μ)-solution f of L∗ f = λf is zero for the adjoint operator
L∗ of L. Assume for simplicity that λ = 1 (any λ > 0 would work similarly). By
hypo-ellipticity, f is smooth and L∗ f = Lf . Now, for ζ ∈ A0 ,

f ζ dμ =
2 2
(Lf )f ζ dμ = −
2
f, f ζ 2 dμ,
M M M
integration by parts being justified by the fact that ζ is compactly supported. To-
gether with the chain rule formula,

0≤ f ζ dμ = −
2 2
ζ (f )dμ − 2
2
f ζ (f, ζ )dμ.
M M M
As a consequence,

ζ (f )dμ ≤ 2
2
|f ζ | (f )1/2 (ζ )1/2 dμ.
M M
Now, if (ζ ) ≤ ε, ε > 0, by the Cauchy-Schwarz inequality,

1/2
√
ζ (f )dμ ≤ 2 ε f 2
2 2
ζ (f )dμ .
M M

Hence, M ζ 2 (f )dμ ≤ 4 εf 22 . Replacing

ζ with the functions ζk , k ≥ 1, of
the completeness property, it follows that M (f )dμ = 0, and therefore that f
is constant which in turn yields that f = 0 since Lf = f . The proposition is estab-
lished.
As an immediate consequence, we have

Corollary 3.2.2 For L = g − ∇W · ∇ on a smooth complete connected Rieman-

nian manifold (M, g) where W : M → R is a smooth potential, L is essentially
self-adjoint on the set of smooth compactly supported functions for the measure
e−W dμg , where dμg is the Riemannian measure on (M, g).
Recall that the setting of this corollary consisting of a Riemannian manifold

(M, g) with the measure dμ = e−W dμg is sometimes referred to as a weighted
Riemannian manifold. Any elliptic operator L symmetric with respect to μ may be
decomposed in the form L = g − ∇W · ∇ with g = (g ij ) a Riemannian (co-)metric
and g a Laplace-Beltrami operator (cf. Sect. 1.11.3, p. 46). The important aspect
is that the resulting properties only depend on the second order part of L and not
of W . In particular, in Rn , any operator of the form L = − ∇W · ∇ with smooth
W is essentially self-adjoint.
As already mentioned, the standard completeness assumption refers to the dis-
tance d on M defined from the Riemannian metric associated with any elliptic op-
erator of the form (3.2.1) as described in Sect. C.5, p. 511. In the dual formulation
put forward in Sect. C.4, p. 509,

d(x, y) = sup f (x) − f (y) , (x, y) ∈ M × M,
f ∈A,|∇f |≤1
where A is the set of smooth functions on M. The advantage of this representation

is that it will immediately lead below to a general definition replacing |∇f | ≤ 1 by
(f ) ≤ 1 and functions satisfying (f ) ≤ 1 will be called Lipschitz functions. The
diameter, when it exists, may be defined in a similar way as a supremum of this
distance function d(x, y) on E × E. Saying that the diameter is bounded therefore
amounts to placing bounds on Lipschitz functions.
3.2.3 Gradient Bounds
Proposition 3.2.1 above ensures that the hypo-elliptic diffusion operator L from
(3.2.1) on a smooth complete connected manifold M is essentially self-adjoint
and is therefore the infinitesimal generator in L2 (μ) of a symmetric semigroup
P = (Pt )t≥0 as discussed in Appendix A. While the Dirichlet form construction
of Sect. 3.1 already ensures that (Pt )t≥0 may be assumed sub-Markov, one central
question is to determine whether or not Pt (1) = 1 (for every t ≥ 0). As already
pointed out, depending on the context, this property is variously referred to as mass
conservation, Markov, stochastic completeness or non-explosion. Indeed, even in
the absence of boundary, it may happen that the associated Markov process (Xtx )t≥0
goes to infinity in finite time. Denoting by T this explosion time, if the manifold M
is complete, there is just one symmetric semigroup (Pt )t≥0 with generator L, and in
this case, Pt f (x) = E(f (Xtx )1{t<T } ) so that the Markov property Pt (1) = 1 does
not hold when T < ∞. As developed in this section, the validity of the mass conser-
vation or Markov property in a smooth manifold setting is closely related to gradient
bounds. The mass conservation property will require more information on the set A0
than the mere completeness of the manifold and makes use of the extended algebra
A of smooth functions on the manifold M.
A useful criterion towards mass conservation is the use of the curvature condi-
tion CD(ρ, ∞) of Definition 1.16.1, p. 72. This investigation will raise in addition
further issues concerning the algebras A0 and A. Recall to start with the 2 operator
from (1.16.1), p. 71, defined for f ∈ A0 by
1
2 (f ) = 2 (f, f ) = L (f ) − 2 (f, Lf ) .
2
The Riemannian content of the and 2 operators is emphasized in Sect. C.6,
p. 513, as (f ) = |∇f |2 and
2 (f ) = |∇∇f |2 + (Ricg −∇S Z)(∇f, ∇f )
where Ricg is the Ricci tensor and ∇S Z the symmetric part of ∇Z. As for L and
, the 2 operator may be extended in the manifold case to the class A of smooth
functions. Recall then the curvature condition CD(ρ, ∞), ρ ∈ R,
2 (f ) ≥ ρ (f )
(for all f ∈ A thus) which suitably extends the notion of Ricci curvature lower
bound on Riemannian manifolds.
One fundamental result with respect to the curvature condition CD(ρ, ∞) is the
following gradient bound or commutation between the actions of the semigroup
and the carré du champ operator. Simple examples are the Brownian and Ornstein-
Uhlenbeck semigroups in Rn with ρ = 0 or ρ = 1 respectively (cf. e.g. (2.7.5),
p. 104). The operator L (from (3.2.1)) is assumed to be elliptic in the sense
of (1.12.2), p. 50, since in general no curvature bounds may hold for non-elliptic
operators.
Theorem 3.2.3 (Gradient bound) Let L be an elliptic diffusion operator with semi-
group P = (Pt )t≥0 , symmetric with respect to μ, on a smooth complete connected
manifold M. If L satisfies the curvature condition CD(ρ, ∞) for some ρ ∈ R, then,
for every f ∈ A0 and every t ≥ 0,

(Pt f ) ≤ e−2ρt Pt (f ) .
It is not difficult to see that, conversely, the gradient bound of Theorem 3.2.3
for every t ≥ 0 implies in return the curvature condition CD(ρ, ∞) (see Corol-
lary 3.3.19 below or Theorem 4.7.2, p. 209, in the context of local Poincaré inequal-
ities). Below, we present a formal and easy proof of Theorem 3.2.3, and then we
address the question of how this formal argument can be justified.
The formal proof is very elementary, and relies on the basic interpolation argu-
ment along the semigroup extensively developed throughout this work. Consider,
for f ∈ A0 and fixed t > 0, the function defined by

(s) = Ps (Pt−s f ) , s ∈ [0, t]
(at any fixed point). Then,

(s) = LPs (Pt−s f ) − 2Ps (Pt−s f, LPt−s f ) .
The first term comes from the derivative of Ps and the heat equation, the second
one from the derivative of Pt−s f applied to the quadratic form . Rewriting this
identity with g = Pt−s f ,

(s) = Ps L(g) − 2 (g, Lg) = 2Ps 2 (g) .
It is worthwhile to observe that this identity is the analogue at the level of the 2
operator of the Duhamel formula (3.1.20) for . Now the CD(ρ, ∞) condition
simply reads ≥ 2ρ , from which the conclusion immediately follows.
Now, we have to wonder why and when such a proof could be justified, that is,
for which classes of functions f the preceding argument may be developed. The
operator Ps acts a priori on bounded functions, or on functions which are in some
Lp (μ)-space, but there is no guarantee that differentiation and commutation of L
and Ps are valid for those functions. More precisely, the first identities to be un-
derstood are LPs ((g)) = Ps (L(g)) as well as differentiation of (Pt−s f ) under
Ps (which consists, as already mentioned earlier, of an integration by parts together
with the derivation of an integral, and hence requires some justification). Of course,
working with an operator L with smooth coefficients on a compact smooth mani-
fold, the preceding is fully justified for sufficiently smooth functions. Indeed, in this
case, if f is smooth, so is Ps f , and so are g = Pt−s f and (g). They are more-
over bounded, and in the domain of L, so that the conclusion follows. However,
we would like to assert that this result is true for a generic operator L with smooth
coefficients on a (a priori non-compact) manifold. If f is smooth and compactly
supported, g = Pt−s f and (g) are still smooth, but no longer compactly supported
(it is not too hard to see that, under the CD(ρ, ∞) hypothesis, the kernel measures
pt (x, dy) have strictly positive densities everywhere for any x). Therefore care has
to be taken here when considering Ps ((g)).
Before moving on to a precise analysis of the arguments which justify the ingre-
dients used in the proof of Theorem 3.2.3, let us mention a reinforced form of it
which we shall also use extensively.
Theorem 3.2.4 (Strong gradient bound) In the setting of Theorem 3.2.3, if L satis-
fies the curvature condition CD(ρ, ∞) for some ρ ∈ R, then, for every f ∈ A0 and
every t ≥ 0,

(Pt f ) ≤ e−ρt Pt (f ) .
This is clearly an improvement upon Theorem 3.2.3 since, by Jensen’s inequality,

Pt (g 2 ) ≥ (Pt g)2 (Pt is sub-Markov). The formal proof of Theorem 3.2.4 is similar
to that of Theorem 3.2.3, except that we now consider

(s) = Ps (Pt−s f )1/2 , s ∈ [0, t].
The formal derivation yields in this case

(s) = Ps L (Pt−s f )1/2 − (Pt−s f )−1/2 (Pt−s f, LPt−s f ) .
Using the change of variables formula (diffusion property) for L, this may be rewrit-
ten as

−1/2 ((Pt−s f ))
Ps (Pt−s f ) 2 (Pt−s f ) − .
4 (Pt−s f )
From (C.6.4), p. 515, the CD(ρ, ∞) curvature hypothesis actually yields the rein-
forced inequality

4 (g) 2 (g) − ρ (g) ≥ (g) (3.2.4)
(for any g in A). Applying this inequality to g = Pt−s f then yields the differential
inequality ≥ ρ , from which the conclusion follows again.
The following now tries to fully justify the proof of the gradient bounds (The-
orems 3.2.3 and 3.2.4) in the smooth context of this section. According to the dis-
cussion in Sect. 1.11.3, p. 46, the elliptic operator L from (3.2.1) takes the form
L = g + Z where g is the Laplace-Beltrami operator on M associated with the
(co-) metric g = (g ij ) and Z is a smooth gradient field Zf = −∇W · ∇f . The
invariant measure of L is given by dμ = e−W dμg where dμg is the Riemannian
measure on (M, g). Recall the class A0 of smooth (C ∞ ) compactly supported func-
tions on M, and the class A of smooth (C ∞ ) functions. The carré du champ operator
is simply given by (f ) = |∇f |2 (on A0 or A). Recall that, according to Sect. C.6,
p. 513, the curvature condition CD(ρ, ∞) expresses, when Z = 0, a lower bound
on the Ricci curvature of the manifold.
Proof of Theorem 3.2.4 As a consequence of the completeness assumption, by

Proposition C.4.1, p. 511, Appendix C, there is a sequence (ζk )k≥1 of positive func-
√
tions in A0 , increasing to the constant function 1 and satisfying (ζk ) = |∇ζk | ≤ k1
for every k ≥ 1. Observe that this property only controls (ζk ), but not for example
Lζk . This drawback will make the analysis below rather delicate.
The CD(ρ, ∞) curvature assumption will
be used first to
extend to second order
derivatives the integration by parts formula M f Lg dμ = − M (f, g)dμ. Indeed,
from the very definition of the operator 2 and integration by parts,

2 (f )dμ = (Lf )2 dμ,
M M
which is valid for all functions f ∈ A0 . The first step is to extend this property to
smooth functions of the L2 (μ)-domain D(L) of L. In fact, the following consider-
ations will reveal that in our analysis

we then2 have to replace the Dirichlet form,
M |∇f | dμ, by quantities such as M |∇∇f | dμ involving Hessians. This aspect
2
increases the complexity of the integration by parts formulas. Some of the difficul-
ties are resolved by observing that
2 (f ) − ρ (f ) ≥ |∇∇f |2 (3.2.5)

which follows from the explicit description of the 2 operator in this Riemannian
context (cf. (C.5.3), p. 512).
To start with, note that for f ∈ A0 , from (3.2.5),

∇∇f 22 = |∇∇f |2 dμ
M

≤ −ρ (f )dμ + (Lf )2 dμ
M M
≤ |ρ| f 2 Lf 2 + Lf 22 (3.2.6)

|ρ| |ρ|
≤ f 2 + 1 +
2
Lf 22
2 2

|ρ|
≤ 1+ f 2D(L) .
2
Hence, if a sequence (fk )k∈N of functions in A0 converges as k → ∞ to f in the

domain norm · D(L) , the sequence ∇∇fk , k ∈ N, is a Cauchy sequence in the set
of symmetric tensors equipped with the norm
1/2
∇∇f 2 = |∇∇f | dμ
2
.
M
It therefore converges to some symmetric tensor K. The limit does not depend on
the approximating sequence, and is thus denoted Kf . If f ∈ A0 , then Kf = ∇∇f .
This remains true for any smooth function. To see why this is, it is enough to observe
that given h ∈ A0 ,
∇∇(hfk ) = h ∇∇fk + ∇h ⊗ ∇fk + ∇fk ⊗ ∇h + fk ∇∇h
for every k. In the limit as k → ∞,
∇∇(hf ) = h Kf + ∇h ⊗ ∇f + ∇f ⊗ ∇h + f ∇∇h,
from which the identification Kf = ∇∇f follows. Then the bound (3.2.6) extends
to functions f in the L2 (μ)-domain of L by density.
Having made these preliminary observations, we can now begin the proof of
Theorem 3.2.4. Choose a function f ∈ A0 and fix ε > 0. For t > 0, consider the
function in s ∈ [0, t] given by
1/2
(s) = e−2ρs (Pt−s f ) + ε .
By the reinforced inequality (3.2.4), it easily seen that
L + ∂s ≥ 0. (3.2.7)
We may also observe, from basic calculus on smooth functions, that

eρs ∇ (s) ≤ |∇∇Pt−s f | (3.2.8)
for every s ∈ [0, t]. From (3.2.6), |∇∇Pt−s f | is in L2 (μ) with L2 (μ)-norm bounded
from above by C1 f D(L) for some C1 > 0 depending only on ρ.
Choose now two auxiliary positive functions ξ, ζ : M → R, smooth and com-
pactly supported. Let, for s ≥ 0,

G(s) = ξ Ps ζ (s) dμ = ζ (s)Ps ξ dμ.
M M
Since ζ is compactly supported, there is no difficulty in differentiating this expres-

sion to get (with = (s) and = (s) for simplicity)

G (s) = ζ Ps ξ dμ + ζ LPs ξ dμ.
M M
From (3.2.7), for every s ≥ 0,

G (s) ≥ L(ζ ) − ζ L Ps ξ dμ = Lζ + 2 ( , ζ ) Ps ξ dμ.
M M
After integration by parts,

G (s) ≥ (ζ, )Ps ξ dμ − (ζ, Ps ξ )dμ.
M M
We therefore obtain the lower bound, for all s ≥ 0,

G (s) ≥ − (ζ ) ∞ ( ) 2 Ps ξ 2 + 2 (Ps ξ ) 2 ,
and together with (3.2.8) and (3.2.6),

G (s) ≥ −C2 (t) (ζ ) ∞ ξ D(L) f D(L)
where C2 (t) only depends on t. Therefore,

G(t) − G(0) ≥ C3 (t) (ζ ) ∞ ξ D(L) f D(L) .
Now replace ζ by the terms ζk , k ≥ 1 of the sequence which is used to characterize

completeness. In the limit as k → ∞,

ξ Pt (t) − (0) dμ ≥ 0.
M
This being true for any positive ξ ∈ A0 , it finally yields Pt ( (t)) ≥ (0). It re-
mains to let ε go to 0 to get the desired inequality. The strong gradient bound of
Theorem 3.2.4 is established.
There is therefore a huge difference between the formal proof of Theorem 3.2.4
given earlier and the actual, more honest proof, that justifies in practice the formal
computations. For example, the same result would also be true on a compact mani-
fold with convex boundary (with the Neumann conditions at the boundary), but yet
another proof would be required.
In Sect. 3.3, in the abstract framework of a Markov Triple, the preceding gradi-
ent bounds will be extended from the complete case to the essentially self adjoint
case. However, this extension requires a much more refined analysis of the relations
between the classes A0 and A.
Theorem 3.2.4 may be strengthened by letting ρ vary, i.e. by assuming the exis-
tence of a (smooth) function ρ : M → R such that 2 (f ) ≥ ρ (f ) at any point x
and for every f ∈ A. In a Riemannian setting with L = g − ∇W · ∇, the best func-
tion ρ = ρ(x) for which the preceding is satisfied is the minimal eigenvalue at the
point x of the symmetric tensor Ric +∇∇W . The next proposition is the announced
extension, together with its probabilistic interpretation.
Proposition 3.2.5 In the setting of Theorem 3.2.3 under the curvature condition
CD(ρ, ∞) for some ρ = ρ(x) bounded from below by c ∈ R, for any f ∈ A0 and
t ≥ 0,

t |∇f |
(Pt f ) = |∇Pt f | ≤ P
where (P t )t≥0 is the semigroup with generator L − ρ Id. In particular, if
{Xt ; t ≥ 0, x ∈ M} denotes the Markov process with generator L, for any t ≥ 0
x
and x ∈ M,

t
|∇Pt f |(x) ≤ Ex |∇f |(Xt ) e− 0 ρ(Xs )ds . (3.2.9)
The proof of this proposition closely follows the previous proof of Theorem 3.2.4
working now with

(s) = Ps (Pt−s f )1/2 , s ∈ [0, t].
√
√ ≥ c for every x ∈ M, we already have an a priori bound (Pt f ) ≤
Since ρ(x)
−ct
e Pt ( (f )) which justifies the various derivatives of . Then, it remains to
apply the Feynman-Kac formula of Sect. 1.15.6, p. 63.
3.2.4 Mass Conservation
On the basis of Theorems 3.2.3 and 3.2.4 of the last sub-section, we now consider
the mass conservation property Pt (1) = 1 under a CD(ρ, ∞) curvature condition
for some ρ ∈ R. Formally, conservation of mass reflects the fact that L(1) = 0.
However, when the measure μ is infinite, the constant function 1 cannot be in the
domain D(L).
Theorem 3.2.6 (Mass conservation) Let L be an elliptic diffusion operator with

semigroup P = (Pt )t≥0 , symmetric with respect to μ, on a smooth complete con-
nected manifold M. If the CD(ρ, ∞) curvature condition holds for some ρ ∈ R,
then for all t ≥ 0, Pt (1) = 1.
Proof To make sense of the statement, recall that Pt (1) is understood here as the
(increasing) limit of Pt ζk as k → ∞ where (ζk )k≥1 is any sequence of positive
functions in A0 increasing to 1.
Choose two positive functions ξ and ζ in A0 , and
consider, for t ≥ 0, the quantity M (Pt ζ − ζ )ξ dμ. By integration by parts,
t

(Pt ζ − ζ ) ξ dμ = ξ LP ζ dμ ds
s
M 0 M
t

≤ (ξ, Ps ζ )dμ ds (3.2.10)
0 M
t

≤
(ξ ) 1
(ζ ) ∞ e−ρs ds
0
where the last step uses the curvature condition via Theorem 3.2.4. Now replace ζ
by the terms ζk , k ≥ 1, of √
the sequence which is used to characterize completeness
of the manifold M. Since (ζk ) ≤ k1 , in the limit as k → ∞,

Pt (1) − 1 ξ dμ = 0,
M
which in turn yields Pt (1) = 1 since ξ ≥ 0 is arbitrary in A0 . The proof is com-

plete.
Another interesting consequence of the preceding tools is the following state-

ment, which ensures that the invariant measure is necessarily finite in spaces with
strictly positive curvature (such as the sphere, for example).
Theorem 3.2.7 Let L be an elliptic diffusion operator with Markov semigroup

P = (Pt )t≥0 , symmetric with respect to μ, on a smooth complete connected manifold
M. If the CD(ρ, ∞) curvature condition holds for some ρ > 0, then the invariant
measure μ is finite.
Proof When f is in L2 (μ) and t → ∞, Pt f converges in L2 (μ) to P∞ f which

is the projection of f on the space of invariant functions, that is, an element of the
domain D(L) satisfying LP∞ f = 0 (cf. Sect. 1.6, p. 24). Now, by hypo-ellipticity,
such
a function g = P∞ f is smooth, and being in L2 (μ) together with Lg, one
has M (g)dμ = 0 and thus g is constant by connexity. If μ(E) is infinite, then
this constant must be 0. Now, as in the proof of mass conservation Theorem 3.2.6,
choose two functions ξ and ζ in A0 . By (3.2.10) at t = ∞,
∞

(ξ − P∞ ξ ) ζ dμ ≤ (ξ ) (ζ ) e−ρs ds.
1 ∞
M 0
3.3 Heart of Darkness 151

smoothing functions ζk , as k → ∞, and whenever
Replacing as before ζ by the
μ(E) = ∞, it follows that M ξ dμ = 0 for any ξ ∈ A0 , yielding a contradiction.
Theorem 3.2.7 is proved.
3.3 Heart of Darkness
This section attempts to describe a general setting for the investigation of symmetric
Markov diffusion operators and semigroups, which will be convenient for the anal-
ysis of various questions related to functional inequalities, convergence to equilib-
rium, heat kernel bounds and so on. Although we do not want to restrict ourselves to
the case of smooth manifolds, in view of several important applications (in particular
in infinite dimension), we are now in a position to extract from the previous analysis
in manifolds the main features of the spaces of smooth compactly supported (A0 )
and smooth (A) functions used there. The abstract formalism put forward here is
a self-consistent framework in which the intuition behind the -calculus and semi-
group monotonicity may easily be developed, without having to take too much care
over the various families of functions involved in the analysis. In particular, this set-
ting is not intended for an investigation in itself. Applications will develop the ideas
emphasized through this formalism rather than the technical properties which have
to be investigated somewhat case by case.
Recall also that we do not want to impose a topology on the basic measure space
(E, F, μ) on which the Markov operators act, or on the space in which the Markov
processes live. Indeed, every notion should be invariant under measurable bijections
of the space (although such generality may in practice be quite useless). In partic-
ular, equalities and inequalities between functions have to be understood μ-almost
everywhere. The various classes of (real-valued) functions involved in the investi-
gation actually aim at replacing a pointwise analysis by a measurable one. In partic-
ular, the classes should be rich enough to develop a relevant calculus, leading below
to the necessary hypotheses in this regard (not always transparent). In applications
of interest, such as in a smooth manifold context, the various relevant inequalities
usually do hold pointwise.
The following thus describes a suitable framework enabling us to smoothly de-
velop the -calculus at the root of this investigation. The next Sect. 3.4 contains a
summary of the various definitions and hypotheses. The starting point for the analy-
sis is the underlying algebra A0 of a symmetric Diffusion Markov Triple (E, μ, )
(Definition 3.1.8). As already pointed out, this class A0 is in practise the class of
smooth (C ∞ ) functions with compact support. On A0 is given the carré du champ
operator . From the carré du champ operator is constructed the generator L to-
gether with the Dirichlet form E defined on its domain D(E). Many properties of
functions defined on A0 automatically extend to the Dirichlet space D(E). The semi-
group P = (Pt )t≥0 with infinitesimal generator L in L2 (μ) is symmetric with respect
to μ and sub-Markov (positivity preserving and such that Pt (1) ≤ 1 for every t ≥ 0).
A Diffusion Markov Triple together with the additional ergodicity and mass con-
servation properties yields the notion of a Standard Markov Triple defined in Defi-
nition 3.1.15.
But in the initial setting of a Diffusion Markov Triple (E, μ, ), many properties
(such as essential self-adjointness (ESA), ergodicity, mass conservation or gradient
bounds as developed in the smooth case) may or may not hold. Criteria for these
properties to hold actually require the introduction of a new larger class A of func-
tions, viewed as an extension of A0 , on which the operator L is defined as an exten-
sion together with the associated operator, regardless of integrability properties.
In the smooth manifold case, this class A is the class of smooth (C ∞ ) functions.
The following develops the abstract construction of such a class A. The ESA prop-
erty will appear as the cornerstone of the construction. Indeed, while positivity of
the carré du champ operator is the main ingredient of the extension results, at
the second order of the 2 operator, curvature conditions together with essential
self-adjointness allow for a parallel treatment.
3.3.1 Extended Algebra A
An extension of A0 is a class A of measurable functions f on E containing the

constants, stable under the action of smooth multivariate functions, and such that
hf ∈ A0 for all h ∈ A0 . In other words, A0 is an ideal in A. The extension A of A0
is not unique, and many algebras A suitably extending A0 may be considered. The
largest possible class A would be the set of functions f such that hf ∈ A0 for any
h ∈ A0 . It is however not always wise or necessary to work with this largest class
A. In many examples the ideal property is not suitable either, and may be relaxed in
some situations (at the price of some extra and tedious hypotheses on A).
Beyond the properties of A0 and described in Sect. 3.1 and summarized in
Sect. 3.4 below, the requirements on A are the following.
Definition 3.3.1 (Extended algebra A) A is an algebra of measurable functions

on E containing A0 , containing the constant functions and satisfying the following
requirements.
(i) Whenever f ∈ A and
h ∈ A0 , hf ∈ A0 (ideal property).
(ii) For any f ∈ A, if E hf dμ ≥ 0 for every positive h ∈ A0 , then f ≥ 0.
(iii) A is stable under composition with smooth (C ∞ ) functions : Rk → R.
(iv) The operator L : A → A is an extension of L on A0 . The carré du champ
operator is also defined on A × A by the formula, for every (f, g) ∈ A × A,
1
(f, g) = L(f g) − f L(g) − g L(f ) ∈ A.
2
(v) For every f ∈ A, (f ) ≥ 0.
(vi) The operators and L satisfy the change of variables formulas (3.1.2)
and (3.1.8) for any smooth (C ∞ ) function : Rk → R.
(vii) For every f ∈ A and every g ∈ A0 , the integration by parts formula

(f, g)dμ = − g Lf dμ = − f Lg dμ (3.3.1)
E E E
holds true.
(viii) For every f ∈ A0 and every t ≥ 0, Pt f ∈ A.
The ideal property (i) ensures in particular that whenever f ∈ A and h ∈ A0 ,

(f, h) ∈ A0 . In the same way, in many formulas (involving L, and later 2 ,
Hessians etc.) where some function h ∈ A0 appears, the result is in A0 . This will
allow us to freely use integration by parts arguments.
In view of Remark 3.1.2, (ii) is not a strong requirement, and is easy to check in
any practical situation. However, we do not want at this stage to impose on A0 any
kind of locality, and neither do we want to impose on A any kind of local integra-
bility. The ideal property ensures that (ii) makes sense without further assumptions.
The most important property in Definition 3.3.1 is (viii). In some sense, the al-
gebra A is the smallest class, according to each specific context, in which compu-
tations are easy to make and where functions in A0 are mapped through the action
of Pt . Note indeed that beyond the case of compact spaces, A0 is in general not
stable under the semigroup. The possible stability of A under (Pt )t≥0 reflects in
general smoothness properties of the coefficients of a given operator. Such opera-
tors with smooth coefficients should not be confused with hypo-elliptic operators
(cf. Definition 3.3.8 below), hypo-ellipticity being in general a much stronger re-
quirement.
Note finally that the validity of the integration by parts formula (3.3.1) may still
hold for some g ∈ / A0 but then usually requires some extra analysis (see for example
Proposition 3.3.16).
Remark 3.3.2 It might be worthwhile to mention again that the ideal and stability
properties of the extended algebra A provide an alternative understanding of the
density of A0 . For example, the non-negativity of on A0 may be extended to A
in the following way. Letting f ∈ A and g ∈ A0 , g ≥ 0, by the chain rule, for every
integer k ≥ 1,

0 ≤ f k g = k 2 f 2k−2 g 2 (f ) + 2kf 2k−1 g (f, g) + f 2k (g).
Changing f into f + 1, dividing by k 2 (f + 1)2k−2 and letting k → ∞ yields

g 2 (f ) ≥ 0, hence g (f ) ≥ 0, from which (f ) ≥ 0 by (ii) of Definition 3.3.1.
(See below in Remark 3.3.15 the corresponding extension for the curvature-
dimension condition with similar arguments.)
Before developing the analysis relating A0 , A, L and , it should be emphasized

that the preceding setting is not the only one that may be considered, and some sit-
uations could be handled with other environments. For instance, allowing constants
to belong to A0 will lead (through the ideal property) to A = A0 , and therefore to
the stability of A0 under (Pt )t≥0 . This is of course the most desirable situation (and
is valid, for example, on compact manifolds) and many properties (such as the ESA
property) are then automatic. It requires the measure μ to be finite. Relying on this
hypothesis actually allows us to favor intuition without taking too much care over
technical details.
There are also cases where A0 may be assumed to be stable under (Pt )t≥0 with-
out containing the constant functions. For instance, in Rn and for the standard
Laplace operator , or the Ornstein-Uhlenbeck operator, the Schwartz space of
rapidly decreasing functions together with all their derivatives is such an exam-
ple. With some extra work (depending on our knowledge of the model), this may
also be achieved in a fairly large setting. But it is unlikely that this will hold on a
non-compact manifold only under the curvature condition, which is at the heart of
the analysis developed in this monograph. For the Ornstein-Uhlenbeck, Laguerre
and Jacobi operators of Chap. 2, it is in general better to work with the algebra A0
of polynomials in order to take advantage of the particular structure of the associ-
ated semigroups. The functions in A0 are then no longer stable under composition
with smooth functions. The reader may easily adapt the various hypotheses to those
particular cases.
Remark 3.3.3 When the measure μ is finite and the semigroup (Pt )t≥0 is con-
servative, constant functions belong to D(L). Then A0 may be replaced by the
larger class Aconst
0 = {f + c ; f ∈ A0 , c ∈ R}, extending with (f, 1) = 0 and
L(1) = 0. In the same spirit, the change of variables formulas (3.1.2) and (3.1.8)
easily extend to functions f in Aconst
0 without the restriction that (0) = 0. How-
ever, care has to be taken not to apply the ideal property to Aconst 0 (since if it ap-
plies
we would be back
to A = A 0 ). Indeed, even the integration by parts prop-
erty E g Lf dμ = − E (f, g) dμ does not in general make sense for f ∈ A
when g is constant. It is also worth observing that the properties (f, 1) = 0 and
L(1) = 0 are direct consequences of the change of variables formula applied with
(r) = 1,
r ∈ R. In particular, integration by parts (3.3.1) entails the invariance
of μ as E Lf dμ = 0 for any f ∈ A0 . When the underlying invariant measure
is finite, it will also be convenient to deal in the later chapters with the family
Aconst+
0 = {f + c ; f ∈ A0 , f ≥ 0, c > 0}.
Remark 3.3.4 The ideal property (i) of Definition 3.3.1 is convenient for the analy-
sis on non-compact Riemannian manifolds, but would not be suitable in some other
settings. For example, in many models of statistical mechanics, where an infinite
number of variables is considered, A0 would be the space of (smooth compactly
supported) functions depending on a finite number of coordinates. Then, A would
be the space of smooth functions depending on an infinite number of variables
and the ideal property does not hold. Fortunately, in many infinite-dimensional in-
stances, this difficulty is easily overcome. For example, for the infinite-dimensional
Ornstein-Uhlenbeck semigroup described in Sect. 2.7.2, p. 108, the algebra A would
be the space of smooth functions depending on a finite number of coordinates, and
the stability under (Pt )t≥0 is valid due to the fact that the different components
do not interact. In general, however, when interaction is in action, a lot of analy-

sis has to be developed on approximating models with a finite number of particles,
and the major challenge is then to obtain functional inequalities where the control
on the constants does not depend on the dimension (in this context the number of
particles). While we do not address any specific infinite-dimensional models in this
work, the formalism put forward here, and more importantly some of the basic prin-
ciples of investigation, may be adapted to some extent to cover numerous examples
of interest.
3.3.2 Domains
At this point, there are two definitions of the operators L and , one on D(L) (or
D(E)) and one on the extended algebra A. The first task is to verify that they actually
coincide. Recall the adjoint L∗ of L (with respect to μ and A0 ) as presented in
Definition 3.1.9.
Proposition 3.3.5
(i) If f ∈ A ∩ D(L∗ ), then L∗ f = Lf .
(ii) If f ∈ A ∩ D(E), both definitions of , in D(E) and in A, coincide.
(iii) If f ∈ A ∩ D(E) and Lf ∈ L2 (μ), then f ∈ D(L).
Proof The first assertion

is immediate
as a consequence
of integration by parts. In-
deed, for any h ∈ A0 , E h L∗ f dμ = E f Lhdμ = E h Lf dμ. The result follows
from the density of A0 in L2 (μ).
The second assertion (ii) is a bit more delicate. For the moment, denote by A
the carré du champ operator on A, keeping the notation for the carré du champ
operator on D(E). We first check that for every
h ∈ A0 , (f, h) =
A (f, h). To this
end, it is enough to prove that for any k ∈ A0 , E k (f, h)dμ = E k A (f, h) dμ.
By Proposition 3.1.5, kf ∈ D(E) and (kf, h) = k (f, h) + f (k, h), while the
same identity holds for A from the diffusion property. Therefore, we need to check
that

(kf, h)dμ − f (k, h)dμ = A (kf, h)dμ − f A (k, h)dμ.
E E E E
But from the ideal property kf ∈ A0 , so that (kf, h) = A (kf, h), and similarly
f (h, k) = f A (h, k), hence equality holds. The same argument then allows us
to pass from (f, h) with f ∈ A ∩ D(E) and h ∈ A0 to (f, g) with f and g in
A ∩ D(E).
For the last point (iii), for every h ∈ A0 ,

E(h, f ) = h Lf dμ ≤ Lf 2 h2 ,

E
which extends to any h ∈ D(E) by density of A0 in D(E) with respect to the

D(E)-topology. This is precisely the definition of f ∈ D(L).
It is in general wrong to assert that f ∈ A ∩ L2 (μ) and (f ) ∈ L1 (μ) suffice

to assert that f ∈ D(E) as may be seen by the example of the Laplace operator on
[0, 1]. This implication actually holds under the essential self-adjointness assump-
tion ESA (Definition 3.1.10) as described by the next statement.
Proposition 3.3.6 Assume that the ESA property holds. Then

(i) If f ∈ A ∩ L2 (μ) and Lf ∈ L2 (μ), then f ∈ D(L).
(ii) If f ∈ A ∩ L2 (μ) and (f ) ∈ L1 (μ), then f ∈ D(E).
Proof (i) is immediate since under the hypotheses f ∈ D(L∗ ) (by the very definition
of D(L∗ )) and by the ESA property D(L) = D(L∗ ).
Turning to (ii), consider the linear map f : h → E h Lf dμ defined on A0 .

From the integration by parts formula and the Cauchy-Schwarz inequality,
1/2

f (h) ≤ E(h)1/2 (f )dμ .
E
The linear map f thus extends to a continuous linear form on the Hilbert space
D(E),
and may be represented as E(g, h) for some g ∈ D(E). Then for any h ∈ A0 ,
E (f − g)Lhdμ = 0, and therefore f − g ∈ D(L∗ ) (with L∗ (f − g) = 0). Since
∗
D(L ) = D(L) by the ESA assumption, f may be written as sum of an element of
D(L) and an element of D(E), hence in D(E). The proof is complete.
3.3.3 Connexity, Weak Hypo-ellipticity and Completeness
Having presented the extended algebra A and some of its properties, we next de-
scribe the connexity, weak hypo-ellipticity and completeness properties based on
the picture provided by the smooth manifold case of the previous section. Thus in
the following, the triple (E, μ, ) denotes a Diffusion Markov Triple and A the
extended algebra of Definition 3.3.1.
Definition 3.3.7 (Connexity) The Diffusion Markov Triple (E, μ, ) is said to be

connected if f ∈ A and (f ) = 0 imply that f is constant.
It should be mentioned here that with respect to the ergodicity property of Defi-
nition 3.1.11, connexity is a local property for functions in A whereas for ergodicity
the property holds for functions in D(E).
The second definition describes weak hypo-ellipticity in this context.
Definition 3.3.8 (Weak hypo-ellipticity) The Diffusion Markov Triple (E, μ, ) is

said to be weakly hypo-elliptic if for every λ ∈ R, any f ∈ D(L∗ ) which satisfies
L∗ f = λf belongs to A.
Very often (for example in the case of manifolds with hypo-elliptic operators,
as in Sect. 1.12, p. 49), it is also true that for any bounded measurable function
f and any t > 0, Pt f ∈ A. However, this property will not be used below, and
passing from weak hypo-ellipticity to this seemingly stronger form would require
extra assumptions on the algebra A.
The third definition is that of completeness in this context, following the char-
acterization of Proposition C.4.1, p. 511, in Appendix C, as extensively used in the
previous section on Riemannian manifolds.
Definition 3.3.9 (Completeness) The Diffusion Markov Triple (E, μ, ) is said to

be complete if there exists an increasing sequence (ζk )k≥1 of positive functions in
A0 such that limk→∞ ζk = 1 (μ-almost everywhere) and (ζk ) ≤ k1 , k ≥ 1.
Connexity and weak hypo-ellipticity yield ergodicity as defined in Defini-

tion 3.1.11.
Proposition 3.3.10 Under the connexity and weak hypo-ellipticity assumptions,

ergodicity occurs.
For a proof, let f ∈ D(L) be such that Lf = 0.

Then f ∈ A by
the weak
hypo-ellipticity property. By integration by parts, 0 = E f Lf dμ = − E (f )dμ.
Hence (f ) = 0 and therefore f is constant from the connexity assumption.
As a further easy observation, the proof of Proposition 3.2.1 in the manifold case
immediately extends to a useful criterion for essential self-adjointness (the ESA
property).
Proposition 3.3.11 Under the connexity, weak hypo-ellipticity and completeness

assumptions, the ESA property holds.
3.3.4 Curvature-Dimension Conditions
To establish the gradient bounds described in Sect. 3.2 in the smooth Riemannian
manifold setting, we investigate curvature conditions in this abstract context. To
this end, we first define the 2 operator, already introduced in (1.16.1), p. 71, on the
algebra A, and define in the same way Hessians.
Definition 3.3.12 (2 operator) The 2 operator is the bilinear map

2 : A × A → A defined by
1
2 (f, g) = L (f, g) − (f, Lg) − (Lf, g)
2
for (f, g) ∈ A × A. As for the carré du champ operator , we often write more
simply 2 (f ) = 2 (f, f ).
The change of variables formula for 2 ((1.16.3), p. 71) is valid on A (being a

direct consequence of the change of variables formula for L)

2 ψ(f ) = ψ (f )2 2 (f ) + ψ (f )ψ (f ) f, (f ) + ψ (f )2 (f )2 (3.3.2)
(with f ∈ A and ψ : R → R smooth). Its multi-dimensional extension (C.6.5),

p. 516, is also similar.
The subsequent analysis will require us to consider, besides the 2 operator, ana-
logues of Hessians (on A). The following definition is the suitable abstract formu-
lation of the manifold case as described in Remark C.5.1, p. 512.
Definition 3.3.13 (Hessian) For a function f ∈ A, the Hessian of f is the bilinear

map H (f ) : A × A → A defined by
1
H (f )(g, h) = g, (f, h) + h, (f, g) − (f, (g, h) . (3.3.3)
2
Such Hessians appear in the chain rule formula for 2 ((C.6.6), p. 516) in the
form
2 (hf, g) = h 2 (f, g) + f 2 (h, g) + 2H (g)(h, f ). (3.3.4)
Many formulas involving the operator twice may also be seen as Hessians, such
as for example

h, (f ) = 2H (f )(f, h), f, (f, h) = H (h)(f, f ) − H (f )(f, h).
It is straightforward to observe that H (f )(g, h) is a first order operator in the

sense that, for k ∈ A, H (f )(kg, h) = kH (f )(g, h) + gH (f )(k, h). It is also a sec-
ond order differential operator with respect to the variable f , and satisfies the chain
rule
H (kf )(g, h) = kH (f )(g, h) + f H (k)(g, h) + (k, g)(f, h) + (f, g)(k, h).
Furthermore, the integration by parts formula, valid as soon as one of the functions
f, g, h belongs to A0 , reads

H (f )(g, h)dμ
E

1
= f L (g, h) + (h, Lg) + (g, Lh) + 2 Lg Lh dμ (3.3.5)
2 E

1
= g (f, Lh) − (h, Lf ) − L (f, h) dμ.
2 E
These chain rules extend to change of variables formulas for smooth functions as a
consequence of the formula for the operator (see Remark 3.1.4).
The curvature-dimension condition is defined using the 2 -operator, as in
Sect. 1.16, p. 70.
Definition 3.3.14 (Curvature-dimension condition) The curvature-dimension con-

dition CD(ρ, n), for ρ ∈ R and n ∈ [1, ∞], holds if for every f ∈ A,
1
2 (f ) ≥ ρ (f ) + (Lf )2 .
n
As presented in Sect. C.6, p. 513, this definition suitably extends Ricci curvature
lower bounds in dimensional Riemannian manifolds. Furthermore, the technique
developed for (C.6.4), p. 515, yields a reinforced version of the CD(ρ, ∞) curvature
condition of the form

4 (f ) 2 (f ) − ρ (f ) ≥ (f ) (3.3.6)
for every f ∈ A. Moreover, under the CD(ρ, ∞) condition, it also holds that
(cf. (C.6.8), p. 517)

H (f )(g, h)2 ≤ 2 (f ) − ρ (f ) (g)(h) (3.3.7)
for f, g, h ∈ A.
Remark 3.3.15 In the same way as the positivity of extends from A0 to A via the
ideal property, the change of variables formula and condition (ii) of Definition 3.3.1
(see Remark 3.3.2) may be used to extend the CD(ρ, n) inequality. The argument
is slightly more involved due to the second order terms in the change of variables
formula for 2 . Let us illustrate the principle on the simpler curvature inequality
CD(ρ, ∞). For a given function f ∈ A, it is enough to show that for any positive
function h ∈ A0 , hK(f ) ≥ 0 where K(f, f ) = 2 (f ) − ρ (f ). From the posi-
tivity of K on A0 and the Cauchy-Schwarz inequality (with respect to the positive
quadratic form K), K(hf, hf ) − 2f K(hf, h) + f 2 K(h, h) ≥ 0. Let then ψ : R → R
be smooth such that ψ(0) = 0 and apply the latter to ψ(h). By the change of vari-
ables formula, after some simplification,

ψ 2 (h)K(f ) + 4 ψ(h)ψ (h)H (f )(h, f ) + 2 ψ 2 (h) (f )(h) + (h, f )2 ≥ 0.
Now fix ε > 0 and choose ψ (smooth) such that ψ(0) = 0, and ψ(r) = 1 and
ψ (r) = 0 when r ≥ ε. It follows that K(f ) ≥ 0 on {h ≥ ε}, and it remains to let
ε → 0 to obtain the result.
3.3.5 Extensions
In order to develop the proof of the strong gradient bound (Theorem 3.2.4) in this
abstract setting, it is necessary to extend the 2 and the Hessian operators from
A0 to D(L), in the same way as extends from A0 to D(E). In a sense, the ap-
proach here parallels at the second order what has been developed in Sect. 3.1 at
the first order. At this second order level, positivity of has to be replaced by the
CD(ρ, ∞) curvature condition together with essential self-adjointness (ESA). This
goal is achieved in the following proposition.
Proposition 3.3.16 Assume the ESA assumption and the CD(ρ, ∞) condition for
some ρ ∈ R.
(i) The 2 operator extends to a continuous bilinear operator on D(L) satisfying

2 (f )dμ = (Lf )2 dμ (3.3.8)
E E
for all f ∈ D(L).

(ii) The CD(ρ, ∞) condition extends to D(L).
(iii) Whenever (f )∈N is a sequence in A0 converging to f ∈ D(L), then
(2 (f ))∈N converges in L1 (μ) to 2 (f ).
(iv) For any pair (g, h) ∈ A0 × A0 , the linear operator H (f )(g, h) defined for
f ∈ A0 extends to a bounded bilinear operator on D(L) × A0 × A0 , and may
be further extended to D(L) × D(E) × A0 . The Hessian operator H (f )(g, h)
defined for (f, g, h) ∈ A0 × A0 × A0 can be extended to D(L) × A0 × A0 ,
and further to D(L) × D(E) × A0 . It satisfies the extended inequality (3.3.7).
(v) For g and h in A0 , whenever (f )∈N is a sequence in A0 converging
to f ∈ D(L), then (H (f )(g, h))∈N converges to H (f )(g, h) in L2 (μ).
When h ∈ A0 and g ∈ D(E), then the convergence takes place in L1 (μ).
For fixed f ∈ D(L) and h ∈ A0 , if (g )∈N converges to g in D(E), then
(H (f )(g , h))∈N converges to H (f )(g, h) in L1 (μ).
(vi) For every f ∈ D(L), h, k ∈ A0 and g ∈ D(E),
H (f )(hg, k) = hH (f )(g, k) + gH (f )(h, k).
(vii) For every h ∈ A0 and f ∈ D(L), hf ∈ D(L) and, for any g ∈ D(L),
2 (hf, g) = h 2 (f, g) + f 2 (h, g) + 2H (g)(f, h).
(viii) For every f ∈ D(L), g ∈ D(E) and h ∈ A0 ,

H (f )(g, h)2 ≤ 2 (f ) − ρ (f ) (g)(h).

bilinear operator was extended from A0 to D(E) using
Proof In Sect. 3.1, the
that, on A0 , E(f ) = E (f )dμ, (f ) ≥ 0 and A0 is dense in D(E) with respect to
the D(E)-topology. Under the CD(ρ, ∞) and ESA conditions, the same procedure
may be performed on the bilinear positive map 2 (f ) −
ρ (f ) using
the density
of A0 in D(L). The basic ingredient here is the formula E 2 (f )dμ = E (Lf )2 dμ
which holds for any f ∈ A0 . The extension of the CD(ρ, ∞) condition to D(L) is
then straightforward. From (3.3.7), the extension of H (f )(g, h) follows the same
lines, and the change of variables formula (3.3.4) goes to the limit.
The next step, as was the case for the operator (Propositions 3.3.5 and 3.3.6),
is to identify the two definitions of 2 and the Hessian (on D(L) and on A), and
moreover to extend the integration by parts formula (3.3.8) to A. The following
proposition fulfills this task.
Proposition 3.3.17 Assume the ESA assumption and the CD(ρ, ∞) condition for
some ρ ∈ R. If f ∈ A ∩ D(L), the two definitions of 2 (f ) (on D(L) and on A)
coincide. The same is true for the Hessian H (f )(g, h) whenever f ∈ A ∩ D(L),
g ∈ A ∩ D(E) and h ∈ A0 .
Moreover, if f ∈ A ∩ D(L) then 2 (f ) ∈ L1 (μ) and (f ) ∈ L1 (μ), in which
case

2 (f )dμ = (Lf )2 dμ.
E E
Conversely, and under the additional ergodicity property, for f ∈ L2 (μ) ∩ A, if both
2 (f ) and (f ) belong to L1 (μ), then f ∈ D(L).
Proof We follow the same lines as the identification of on A and D(E)

(Proposition 3.3.5). Start with the identifications of the Hessians H (f )(g, h)
for f ∈ D(L) ∩ A and g, h ∈ A0 . To this end, denote by H (f ) the Hessian for
f ∈ D(L) and by HA (f ) the Hessian computed for f ∈ A. We aim to prove that
for every k ∈ A0 ,

kH (f )(g, h)dμ = kHA (f )(g, h)dμ.
E E
By the chain rule,

kH (f )(g, h)dμ = H (kf )(g, h)dμ − f H (k)(g, h) dμ
E E E

− (f, g)(h, k) + (f, h)(k, g) dμ.
E
By the integration
by parts formula (3.3.5), this expression is reduced to an inte-
gral of the form E f E(g, h, k)dμ where
E(g, h, k) (even if tedious

to write down
explicitly) belongs to A0 . Now similarly E kHA (f )(g, h)dμ = E f E(g, h, k)dμ
with the same expression E(g, h, k). Then, changing f into a sequence (f )∈N of
functions in A0 converging to f ∈ D(L), the identity of the two integrals follows
in the limit. Fixing now f ∈ D(L) and h ∈ A0 , and replacing g ∈ A0 by g ∈ D(E),
the same procedure (using the other chain rule for H and the other integration by
parts formula (3.3.5)) yields the announced coincidence of H and HA . The identity
of the 2 operators follow along the same lines.
We already know from Sect. 3.1 that if f ∈ D(L),
then Lf ∈
L2 (μ) (by def-
inition) and (f ) ∈ L (μ). Moreover, in D(L), E 2 (f )dμ = E (Lf )2 dμ and
1

− E f Lf dμ = E (f )dμ. Therefore both (f ) and 2 (f ) − ρ (f ) are in

L1 (μ), and so too is 2 (f ).
It remains to verify the last claim of the proposition, that is, under ergodic-
ity, when f ∈ A ∩ L2 (μ) with 2 (f ) and (f ) in L1 (μ), then f ∈ D(L). To-
wards this goal, following Proposition 3.3.6, it is enough to show that Lf ∈ L2 (μ).
But under the ESA and ergodicity properties {Lh ; h ∈ A0 } is dense in the space
L2 (μ) of L2 (μ)-functions orthogonal to constants. Restrict for simplicity to the
case μ(E) = ∞, in which case {Lh ; h ∈ A0 } is dense in the space L2 (μ) (the proof
is easily adapted to the case μ(E) = 1). Consider the bilinear form K(f, f ) =
2 (f, f ) − ρ (f, f ) and the linear form

f (h) = (Lf + ρf ) Lh dμ = K(f, h)dμ, h ∈ D(L).
E E
Since K is a positive bilinear form, it holds that K(f, h)2 ≤ K(f, f )K(h, h),
which implies for any h ∈ D(L),

|f (h)| ≤ C(f )hD(L) . Then, there exists some
g ∈ D(L) such that f (h) = E [Lg Lh + gh]dμ. But it is easily seen from the spec-
tral decomposition that in fact g = Lk, where k = (L2 + Id)−1 (L + ρ Id)f . Again
from the spectral decomposition, the operators
2 −1 −1 2
L + Id (L + ρ Id) and L2 + Id L +ρL
are bounded in L2 (μ), so that, for every h ∈ A0 ,

f (h) = (Lg + k) Lh dμ ≤ Lg2 + k2 Lh2 ≤ C (f ) Lh2 .

E
Thanks to the ergodicity and ESA properties, the linear form f can be extended to
L2 (μ), which shows that Lf + ρf ∈ L2 (μ), and therefore Lf ∈ L2 (μ). The proof
is complete.
Note from the preceding proof that whenever ρ ≥ 0, for a function f ∈ A ∩ L2 (μ)
such that 2 (f ) ∈ L1 (μ), the condition (f ) ∈ L1 is not needed to ensure that
Lf ∈ L2 (μ).
3.3.6 Gradient Bounds and Mass Conservation
On the basis of the previous curvature conditions, we now address the announced
gradient bounds (or commutation between Pt and ) which were developed in the
smooth manifold case in Sect. 3.2.3 under the hypotheses of connexity and com-
pleteness. The point here is that the result will be extended in the abstract frame-
work to the case when the underlying operator is essentially self-adjoint. As we
have seen in Proposition 3.3.11, essential self-adjointness holds under connexity,
weak hypo-ellipticity and completeness, so that the approach here goes further than
the manifold case.
Although the gradient bound (Pt f ) ≤ e−2ρt Pt ((f )) seems easier to obtain
than the strong bound, the main difficulty
√ is that we have no a priori control on the
L2 (μ)-norm of it, in contrast to that of (Pt f ). Unfortunately, this leads to some
extra complications. One main result is emphasized in the following theorem and its
corollary.
Theorem 3.3.18 (Strong gradient bound) Let (E, μ, ) be a Diffusion Markov

Triple satisfying the ESA property, with extended algebra A. Then, under the
CD(ρ, ∞) condition, for every f ∈ A0 and every t ≥ 0,

(Pt f ) ≤ e−ρt Pt (f ) .
Corollary 3.3.19 (Gradient bound) Let (E, μ, ) be a Diffusion Markov Triple

satisfying the ESA property, with extended algebra A. Then, under the CD(ρ, ∞)
condition, for every f ∈ A0 and every t ≥ 0,

(Pt f ) ≤ e−2ρt Pt (f ) .
Although the corollary seems a weaker result than the theorem (by Jensen’s in-
equality for Pt , which is sub-Markov), we will see later (cf. Theorem 4.7.2, p. 209)
that they are in fact both equivalent to the curvature CD(ρ, ∞) condition. Actually,
taking the derivative at t = 0 of (Pt f ) ≤ e−2ρt Pt ((f )) yields
L (f ) ≥ 2 (f, Lf ) + 2ρ (f )
which amounts to 2 (f ) ≥ ρ (f ).

The proof of Theorem 3.3.18 relies on the following lemmas.
Lemma 3.3.20 Under the ESA and CD(ρ, ∞) assumptions, for every f ∈ A0 , and
any t ≥ 0 and ε > 0, (Pt f ) + ε 2 − ε ∈ D(E).

Proof Set Gε = (Pt f ) + ε 2 − ε. Since Pt f ∈ A, by composition with a smooth
function, Gε ∈ A, and from the ESA property, it is enough to prove that Gε ∈ L2 (μ)
and (Gε ) ∈ L1√
√ (μ) (Proposition 3.3.6
(ii)). The first claim follows since we have
r + ε 2 − ε ≤ r for any r ≥ 0 and E (Pt f )dμ = E(Pt f ) ≤ E(f ) < ∞. The
second claim is a consequence of the bounds
((Pt f ))
(Gε ) = ≤ 2 (Pt f ) − ρ (Pt f ),
4((Pt f ) + ε 2 )
by the extended 2 inequality (3.2.4), and

2 (Pt f ) − ρ (Pt f ) dμ ≤ Lf 22 + |ρ| E(f ).
E
The lemma is therefore established.

Lemma 3.3.21 Under the ESA and CD(0, ∞) assumptions, for any
f ∈ A ∩ L2 (μ) such that Lf ∈ D(E) and (f ) + ε 2 − ε ∈ D(E) ∩ A for some
ε > 0, and for any h positive and bounded in D(E),
(f, Lf )
E h, (f ) + ε 2 − ε + h √ dμ ≤ 0.
E (f ) + ε

Proof Using the same notation Gε = (f ) + ε 2 − ε, for h ∈ A0

E(h, Gε ) = − h LGε dμ
E

42 (f )((f ) + ε 2 ) − ((f )) (f, Lf )
=− h + dμ.
E 4((f ) + ε 2 )3/2 (f ) + ε 2
Applying once again the extended 2 inequality (3.2.4) the announced inequality
follows in this case. It may then be extended to any h ∈ D(E) by density.
Proof of Theorem 3.3.18 We only prove the theorem when ρ = 0, the extension to
the general case being straightforward. Fix ε > 0, f ∈ A0 and h ≥ 0 in A0 . For
t > 0 fixed, consider

(s) = hPs Gε dμ = Gε Ps h dμ, s ∈ [0, t],
E E

with the notation Gε = (Pt−s f ) + ε 2 − ε as above. The task will be to show

From Lemma 3.3.20, Gε ∈ D(E). By spectral analysis, for
that is increasing.
every G ∈ D(E), ∂s E GPs hdμ = −E(Ps h, G). On the other hand,
(Pt−s f, Pt−s Lf )
∂s Gε = −
Gε + ε
while
(Pt−s Lf, Pt−s Lf ) + (Pt−s f, Pt−s L2 f ) (Pt−s f, Pt−s Lf )2

∂s2 Gε = − .
Gε + ε (Gε + ε)3
For any f ∈ A0 , (Pt f ) ∈ L1 (μ) showing that ∂s Gε 1 + ∂s2 Gε 1 ≤ C for some
constant C depending only on f ∈ A0 . Since Pt h is bounded, derivation of is
then justified and, by the usual change of variables formula,

(Pt−s f, Pt−s Lf )
(s) = −E(Ps h, Gε ) − Ps h dμ.
t−s f ) + ε)
((P 1/2
E
This is positive from Lemma 3.3.21. Hence (t) ≥ (0) for any positive h ∈ A0
and any ε > 0, leading to the conclusion. The theorem is established.
The gradient bounds of Theorem 3.3.18 immediately extend to functions f in

D(E) via an approximation argument. However, it is not clear that they apply to
functions in A without further hypotheses.
Under the CD(ρ, ∞) condition, further approximation procedures for functions
in the D(L)-domain by elements of A0 may be developed. Namely, from Corol-
lary 3.3.19 together with the Duhamel formula (3.1.21), for any f in A0 ,

t e2ρt − 1
Pt f 2 − (Pt f )2 ≥ 2 e2ρs ds (Pt f ) = (Pt f ).
0 ρ
(This type of local Poincaré inequality will be extensively examined in Sect. 4.7,
p. 206.) In particular, whenever (f )∈N (in A0 ) converges to f in Lp (μ) for some
p > 2, then (Pt f ) → (Pt f ) in Lq (μ) for any q < p2 and any t > 0.
Remark 3.3.22 The gradient bounds of Theorem 3.3.18 will play a major role in
the forthcoming chapters. In particular cases, similar results may be reached under
weaker hypotheses. For example, when solving stochastic differential equations, the
route proposed in Proposition 3.2.5 provides such bounds with stochastic calculus
tools. Similar results may be obtained for semigroups on bounded domains with
Neumann boundary conditions, provided there is enough information about the ge-
ometry of the boundary. Such an example would however never be covered by the
arguments presented here, due to the lack of the ESA property in this context. The
conclusion of Corollary 3.3.19 is still valid in non-diffusion instances (for example
for Markov chains on a finite space). In some examples, such as the Heisenberg
model (cf. (3.2.3)), the gradient bounds also hold up to a constant K > 1 as

(Pt f ) ≤ Ke−ρt Pt (f )
(while there is in general no CD(ρ, ∞) condition). In these examples, the weaker

gradient bound from Corollary 3.3.19 is usually far easier to obtain than the strong
version.
With the preceding gradient bounds, Theorem 3.2.6 of the previous section on
the mass conservation property extends to the present abstract setting. Recall that
the semigroup P = (Pt )t≥0 is conservative if Pt (1) = 1 for every t ≥ 0. As men-
tioned earlier, the mass conservation property is also known as the Markov, stochas-
tic completeness or non-explosion property. The following statement also covers
Theorem 3.2.7 in this abstract framework and is established similarly.
Theorem 3.3.23 (Mass conservation) Let (E, μ, ) be a Diffusion Markov Triple

with extended algebra A. Under the completeness, ESA and CD(ρ, ∞) as-
sumptions, the semigroup P = (Pt )t≥0 is conservative. Moreover, if ρ > 0, then
μ(E) < ∞.
3.3.7 Intrinsic Distance and Lipschitz Functions
According to the corresponding discussion in smooth manifolds, a distance function

associated with a Markov generator L and the underlying extended algebra A may
be defined as

d(x, y) = esssup f (x) − f (y) , (x, y) ∈ E × E, (3.3.9)
the esssup running over all bounded functions f in A such that (f ) ≤ 1. (Here,
the esssup is defined as the least measurable function on (E × E, F ⊗ F) which is
larger than f (x) − f (y) for μ ⊗ μ-almost every (x, y) ∈ E × E over the given class
of functions f .) The choice of an esssup (which thus rules out sets of μ-measure 0)
instead of a mere supremum is due to the fact that we are considering only classes
of functions. This distance, often called the intrinsic distance (although depending
on the underlying algebra A), coincides with the Riemannian distance in the case
of diffusions on a manifold with elliptic coefficients (see Sect. C.4, p. 509), and
also in some hypo-elliptic cases with the so-called Carnot-Carathéodory distance.
The diameter is defined as the L∞ (E × E, μ ⊗ μ)-norm of the distance function
d(x, y), (x, y) ∈ E × E. We make no claim that (3.3.9) effectively defines a distance
(beyond the triangle inequality which is obviously satisfied), and moreover that this
distance has anything to do with the completeness hypothesis. This would require
many more hypotheses on A and A0 (and would certainly be useless). In particular,
in some infinite-dimensional settings such as the Ornstein-Uhlenbeck example of
Sect. 2.7.2, p. 108, this distance is almost everywhere infinite. However, it will play
a key role in many functional inequalities or estimates on heat kernels in particular
by means of the associated Lipschitz functions. Distance to a measurable set A as
d(x, A) = esssup f (x), x ∈ E, where the esssup runs over all bounded functions f
in A such that (f ) ≤ 1 and f 1A = 0, may be considered similarly. In the above
mentioned example of the infinite-dimensional Ornstein-Uhlenbeck semigroup, this
distance is μ-almost everywhere finite as soon as μ(A) > 0. It will not be used
below but any result on Lipschitz function would also apply to these distances from
sets.
Definition 3.3.24 (Lipschitz function) A function f on a Diffusion Markov Triple

(E, μ, ) with extended algebra A is called Lipschitz if f ∈ A and (f ) ∈ L∞ (μ).
1/2
Denote by f Lip = (f )∞ its Lipschitz semi-norm. f is said to be 1-Lipschitz
if f Lip ≤ 1.
In this definition of Lipschitz function, the class A may be replaced alternatively

by D(E), leading in practice to similar results. Dealing with functions in the Dirich-
let domain is somewhat easier with respect to contraction properties (cf. (3.1.15)).
In the classical case of Rn , Rademacher’s Theorem tells us that Lipschitz functions
coincide with functions with almost everywhere bounded gradients.
The notion of the diameter of a Diffusion Markov Triple (E, μ, ) is defined
in the usual way where the set E is regarded as a metric space with respect to the
intrinsic distance d, and is described equivalently as
D(E, μ, ) = sup f˜∞ (3.3.10)

f Lip ≤1
where f˜(x, y) = f (x) − f (y), (x, y) ∈ E × E, the L∞ -norm being that of

L∞ (E × E, μ ⊗ μ). In particular, if the diameter is finite, every Lipschitz func-
tion which is finite μ-almost everywhere is bounded.
3.3.8 Full Markov Triple
To conclude this construction, starting from the (symmetric) Diffusion Markov

Triple Definition 3.1.8 given on some algebra A0 , in order to fully develop the
tools related to curvature conditions and gradient bounds, and to many other top-
ics such as integration by parts when functions do not belong to A0 , it has thus
been necessary to deal with an extended algebra A. The main assumptions on A0
and A are summarized in Sect. 3.4. This basis gives rise to the definition of a Full
Markov Triple which includes the connexity and ESA properties. Note that a Stan-
dard Markov Triple includes the mass conservation property while it was established
previously that a Diffusion Markov Triple with the ESA property satisfying a cur-
vature condition CD(ρ, ∞) for some ρ ∈ R is actually conservative.
As illustrated in Sect. 3.2, the Full Markov Triple structure covers the important
example of second order differential operators on smooth complete connected Rie-
mannian manifolds with A0 the class of smooth C ∞ compactly supported functions
and A the class of smooth C ∞ functions. Throughout this work, such examples
will always be considered for these algebras of functions under the generic name
of “smooth” functions. More general degrees of smoothness, at least C 2 , may be
investigated similarly.
Definition 3.3.25 (Full Markov Triple) A Full Markov Triple (E, μ, ) is a Stan-
dard Markov Triple (thus with algebra A0 and extended algebra A) satisfying more-
over the connexity and ESA properties.
Occasionally we shall also make the further assumptions of completeness or

weak hypo-ellipticity, however we shall always explicitly mention when these ad-
ditional assumptions have been made. Sometimes, to make things easier, it is also
convenient to assume that A = A0 (in which case the ESA property is automatic and
the gradient bounds are much easier to reach), corresponding to compact manifolds
(without boundary) in the smooth case.
Definition 3.3.26 (Compact Markov Triple) A Compact Markov Triple (E, μ, )

is a Full Markov Triple for which A = A0 .
For a Compact Markov Triple, A0 is stable under (Pt )t≥0 and therefore automat-
ically dense in D(L) with respect to its topology.
3.4 Summary of Hypotheses (Markov Triple)
In this last section, we summarize the previous investigation and present the typical
framework for the analysis of symmetric Markov diffusion operators in which we
will be working in most parts of this monograph. Although various results might
hold in broader settings, without in particular the symmetry or diffusion assump-
tions, we will always work with a Standard or Full Markov Triple (E, μ, ) as
presented in Definitions 3.1.15 and 3.3.25 with underlying algebra A0 and extended
algebra A respectively (the properties of which are recalled below). This

framework
includes the Dirichlet form E with domain D(E) defined by E(f ) = E (f )dμ, the
associated diffusion operator L with L2 (μ)-domain D(L), infinitesimal generator of
the Markov semigroup P = (Pt )t≥0 (with invariant and reversible measure μ), and
associated Markov process, or family of processes, X = {Xtx ; t ≥ 0, x ∈ E}. The
semigroup P = (Pt )t≥0 may be represented according to (1.2.4), p. 12, by probabil-
ity kernels as

Pt f (x) = f (y) pt (x, dy), t ≥ 0, x ∈ E.
E
The kernels pt (x, dy) describe the distribution at time t of the Markov process Xtx
starting at x. Functions are understood as classes of functions, and equalities and
inequalities between them hold μ-almost everywhere.
Throughout, we adopt the Full Markov Triple structure as a convenient frame-
work which allows us to describe the main results and freely develop the central
ideas and principles of the -calculus and heat flow monotonicity.
Standard Markov Triples, or even just Diffusion Markov Triples, suffice for many
statements. Similarly, some results only require a number of specific properties of
the Full Markov Triple definition, but for convenience we stick to the latter. If there
is any need to distinguish between Full and Standard Triples, generally speaking
Full Markov Triples are necessary as soon as gradient bounds, the 2 operator and
curvature-dimension conditions enter into play, otherwise Standard Markov Triples
may be used.
As mentioned earlier, it is sometimes appropriate to work in the Compact Markov
Triple setting (Definition 3.3.26), although the conclusions will actually hold in the
full case.
Finally, for simplicity of exposition, we mostly use the reduced terminology
“Markov Triple”, assuming that Full Markov Triple is meant (or only Standard
Markov Triple if this is enough for the purpose of the given results according to
the preceding remarks).
We present below a short synthesis of the different objects of interest and hy-
potheses described in this first part, to which the reader may refer while progressing
through the following chapters.
3.4 Summary of Hypotheses (Markov Triple) 169
3.4.1 Diffusion Markov Triple (Definition 3.1.8)
A Diffusion Markov Triple (E, μ, ) is composed of a measure space (E, F, μ), a

class A0 of real-valued measurable functions on E and a symmetric bilinear opera-
tor (the carré du champ operator) : A0 × A0 → A0 satisfying properties D1 to D9
below.
D1. The state space (E, F, μ) is a good measure space as defined in Sect. 1.14,
p. 53. The measure μ is σ -finite and, when finite, assumed to be a probability.
D2. A0 is a vector space of bounded measurable functions f : E → R, which is
dense in every Lp (μ), 1 ≤ p < ∞, stable under products (that is A0 is an alge-
bra) and stable under the action of smooth (C ∞ ) functions : Rk → R such that
(0) = 0.
D3. The carré du champ operator is a bilinear symmetric map A0 × A0 → A0
such that (f, f ) ≥ 0 for all f in A0 . (f, f ) is abbreviated as (f ). It satisfies
the diffusion hypothesis of Definition 3.1.3: for every smooth function : Rk → R
such that (0) = 0 and every f1 , . . . , fk , g ∈ A0 ,
k
(f1 , . . . , fk ), g = ∂i (f1 , . . . , fk )(fi , g). (3.4.1)
i=1
D4. For every f in A0 , there exists a finite constant C(f ) such that for every
g ∈ A0 ,

(f, g)dμ ≤ C(f ) g2 .

E
The Dirichlet form E is defined for every (f, g) ∈ A0 × A0 by

E(f, g) = (f, g)dμ.
E
E(f, f ) is abbreviated as E(f ). The domain D(E) of the Dirichlet form E is the
completion of A0 with respect to the norm f E = [f 22 + E(f )]1/2 . The Dirich-
let form E is extended to D(E) by continuity together with the carré du champ
operator .
D5. L is a linear operator on A0 defined by and satisfying the integration by parts
formula

g Lf dμ = − (f, g)dμ
E E
for all f, g ∈ A0 . The change of variables formula (3.1.8) for L is a consequence
of the change of variables formula (3.4.1) for .
D6. For the operator L defined in D5, L(A0 ) ⊂ A0 .
D7. The domain D(L) of the operator L is defined as the set of f ∈ D(E) for which
there exists a finite constant C(f ) such that for any g ∈ D(E)

E(f, g) ≤ C(f ) g2 .
On D(L), L is extended via the integration by parts formula for every g ∈ D(E).
L defined on D(L) is always self-adjoint.

D8. For every f ∈ A0 , E Lf dμ = 0.

D9. The semigroup P = (Pt )t≥0 is the symmetric semigroup with infinitesimal gen-
erator L defined on its domain D(L). It is positivity preserving but not neces-
sarily Markov (in general it is only sub-Markov, i.e. Pt (1) ≤ 1). Moreover, A0
is not necessarily dense in the domain D(L) with respect to the domain norm
f D(L) = [f 22 + Lf 22 ]1/2 . One may use the spectral decomposition to study
(Pt )t≥0 and various properties of functions in L2 (μ) (Proposition 3.1.6).
3.4.2 Standard Markov Triple (Definition 3.1.15)
A Standard Markov Triple is a Diffusion Markov Triple satisfying in addition the

ergodicity and mass conservation properties S10 and S11.
S10. The operator L is ergodic in the sense that any f ∈ D(L) such that Lf = 0
is constant. Equivalently, any f ∈ D(E) such that (f ) = 0 is constant (Defini-
tion 3.1.11).
S11. The semigroup P = (Pt )t≥0 is conservative in the sense that Pt (1) = 1 for
every t ≥ 0 (Definition 3.1.12).
In this context, further properties, not required in the Standard Markov Triple
structure, may be considered.
Adjoint Operator (Definition 3.1.9) The domain D(L∗ ) is the set of functions
f ∈ L2 (μ) such that there exists a finite constant C(f ) for which, for every g ∈ A0 ,

f Lg dμ ≤ C(f ) g2 .

E
On this domain, the adjoint operator L∗ is defined by integration by parts: for any
g ∈ A0 ,

∗
L (f ) = f Lg dμ = g L∗ f dμ.
E E
It holds that D(L) ⊂ D(L∗ ) and L∗ is an extension of L.
Essential Self-adjointness (ESA) (Definition 3.1.10) D(L) = D(L∗ ), or equiva-

lently A0 is dense in D(L) with respect to the D(L) topology (A0 is a core for L).
It is also equivalent to require that L defined on A0 as a symmetric operator has a
unique self-adjoint extension.
3.4 Summary of Hypotheses (Markov Triple) 171
3.4.3 Full Markov Triple (Definition 3.3.25)
A Full Markov Triple is a Standard Markov Triple for which there is an extended
algebra A ⊃ A0 of functions, with no requirements of integrability for elements of
A, satisfying the following requirements (Definition 3.3.1).
F1. Whenever f ∈ A and h ∈ A0 , hf ∈ A0 (ideal property).

F2. Whenever f ∈ A satisfies E hf dμ ≥ 0 for any positive h ∈ A0 , then f ≥ 0.

F3. A is stable under composition with smooth (C ∞ ) functions : Rk → R.
F4. The operator L : A → A is an extension of L on A0 . The carré du champ
operator is also defined on A × A by the formula
1
(f, g) = L(f g) − f L(g) − g L(f ) .
2
F5. For any f ∈ A, (f ) = (f, f ) ≥ 0.
F6. The operators and L satisfy the change of variables formulas (3.1.2)
and (3.1.8) respectively on A for any smooth (C ∞ ) : Rk → R.
F7. For every f ∈ A and g ∈ A0 , the integration by parts formula

(f, g)dμ = − g Lf dμ = − f Lg dμ
E E E
holds true.
F8. For every f ∈ A0 , and every t ≥ 0, Pt f ∈ A.
F9. The Markov Triple is connected in the sense that if f ∈ A, (f ) = 0 implies
that f is constant (Definition 3.3.7).
F10. The ESA property holds (Definition 3.1.10).
3.4.4 Compact Markov Triple (Definition 3.3.26)
A Compact Markov Triple is a Full Markov Triple such that A = A0 . The ESA
property is then automatic.
3.4.5 Miscellaneous
Algebra Aconst
0 (Remark 3.3.3) When μ is a probability measure, A0 may be
replaced by Aconst
0 = {f + c ; f ∈ A0 , c ∈ R}, extending by defining (f, 1) = 0
and L(1) = 0. The change of variables formulas (3.1.2) and (3.1.8) extend to func-
tions f ∈ Aconst
0 without the restriction that (0) = 0. The ideal property does not
apply to A0const (in which case A = A0 ). We also sometimes work with Aconst+
0 =
{f + c ; f ∈ A0 , f ≥ 0, c > 0}.
2 Operator (Definition 3.3.12) The 2 operator is the bilinear map A × A → A

defined by
1
2 (f, g) = L (f, g) − (f, Lg) − (Lf, g) .
2
2 (f ) denotes 2 (f, f ).
Hessian (Definition 3.3.13) The Hessian is a trilinear operator A × A × A → A

defined by
1
H (f )(g, h) = g, (f, h) + h, (f, g) − (f, (g, h) .
2
H (f )(g, h) is a first order operator in g and h and second order in f satisfying the
chain rule, for f, g, h, k ∈ A,
H (kf )(g, h) = kH (f )(g, h) + f H (k)(g, h) + (f, g) (k, h) + (f, h) (k, g).
The 2 operator satisfies the chain rule formula, for f, g, h ∈ A,
2 (hf, g) = f 2 (h, g) + h 2 (f, g) + 2H (g)(h, f ).
Curvature-Dimension Condition (Definition 3.3.14) A Standard Markov Triple

(E, μ, ) satisfies the curvature-dimension condition CD(ρ, n) for ρ ∈ R and
n ∈ [1, ∞] if for any f ∈ A0
1
2 (f ) ≥ ρ (f ) + (Lf )2 .
n
The curvature condition CD(ρ, ∞) amounts to 2 (f ) ≥ ρ (f ) for every f ∈ A0 .
The CD(ρ, n) condition automatically extend to A in the setting of a Full Markov
Triple (see Remark 3.3.15).
By the diffusion hypothesis, the CD(ρ, ∞) condition extends to Hessian in-
equalities in the form

H (f )(g, h)2 ≤ 2 (f ) − ρ (f ) (g)(h)
for f, g, h in A0 (or A). As a consequence, the reinforced curvature condition holds

(f ) ≤ 4 2 (f ) − ρ (f ) (f ).
Weak Hypo-ellipticity (Definition 3.3.8) The weak hypo-ellipticity property is

the fact that for every λ ∈ R, any f ∈ D(L∗ ) which is solution of L∗ f = λf belongs
to A.
Completeness (Definition 3.3.9) The completeness assumption is the existence

of a sequence (ζk )k≥1 of positive functions in A0 , increasing (μ-almost everywhere)
to the constant function 1 and such that (ζk ) ≤ 1k , k ≥ 1.
Intrinsic Distance (cf. (3.3.9)) The distance function on (E, μ, ) is defined by

d(x, y) = esssup f (x) − f (y) , (x, y) ∈ E × E,
where the essential supremum (with respect to the measure μ) is computed on

bounded functions f in A such that (f ) ≤ 1.
Lipschitz Function (Definition 3.3.24) A Lipschitz function is an element f of

1/2
A (or D(E)) such that (f )∞ < ∞. The Lipschitz constant f Lip is (f )∞ .
Diameter (cf. (3.3.10)) The diameter D = D(E, μ, ) of the Markov Triple

(E, μ, ) is defined as

D = esssup f (x) − f (y) ,
where the essential supremum (with respect to the measure μ) runs over all Lips-
chitz functions f with f Lip ≤ 1.
This chapter collects and formalizes a selection of definitions and properties ex-
tracted from Chaps. 1 and 2. In particular, several references relevant to this chapter
may already be found there.
The abstract framework of a Markov Triple (E, μ, ) emphasized in Sects. 3.1
and 3.3 is inspired by the Dirichlet form theory as developed in [91, 189, 190, 294].
It essentially summarizes and extends the early description put forward in the lecture
notes [26] (see also [27, 28]). The intermediate value Theorem 3.1.16 is recorded
in [39]. The slicing argument of Proposition 3.1.17 is part of the folklore on Dirichlet
forms and capacities (cf. [91, 303] etc.).
Section 3.2 emphasizes some fundamental regularity results in the manifold set-
ting. The uniqueness Proposition 3.2.1 and its Corollary 3.2.2 are suitable extrac-
tions from [247, 248]. In this manifold setting, completion and self-adjointness have
been deeply investigated by R. Strichartz [390]. Gradient bounds and commu-
tation properties between gradient and semigroup (Theorems 3.2.3 and 3.2.4), of
fundamental importance throughout this work, appeared in the probabilistic con-
text via the Bismut representation formula of Proposition 3.2.5 [69, 171, 172]
(see [251, 391]). The 2 approach to the gradient bounds originates in [36] and
has been promoted in [22, 24] (see also [25]) in connection with the analysis of the
boundedness of Riesz transforms. The mass conservation Theorem 3.2.6 under a
lower bound on the Ricci curvature essentially goes back to S.-T. Yau [443] (see ear-
lier [191]). It has been widely studied and extended under Ricci or volume growths
by numerous authors, and a detailed history is presented for example in [215, 218].
The monograph [217] by A. Grigor’yan is a comprehensive investigation of heat
kernel bounds on Riemannian manifolds and metric spaces in which the reader will
find complete references and historical developments.
The general setting put forward in Sect. 3.3 is largely inspired by the early con-
tribution [36] and already outlined in the lecture notes [26] (see also [27, 28]). The
complete and self-contained exposition developed here is new, and emphasizes in
particular essential self-adjointness as a critical tool in this context. The intrinsic
distance is also a classical feature of the theory of Dirichlet forms [91, 189, 190].
See also [243, 395, 396]. On gradient bounds for the hypo-elliptic Heisenberg model
(cf. Remark 3.3.22), see [30, 51, 282, 283, 425, 446].
Part II
Three Model Functional Inequalities
Chapter 4
Poincaré Inequalities
This chapter investigates the first important family of functional inequalities for
Markov semigroups, the Poincaré or spectral gap inequalities. These will provide
the first results towards convergence to equilibrium, and illustrate, at a mild and
accessible level, some of the basic ideas and techniques on Markov semigroups
and functional inequalities developed throughout this monograph, at the interplay
between analysis, probability theory and geometry.
Following the conclusions of Chap. 3 and the summary in Sect. 3.4, p. 168, it is
most convenient to present the results of this chapter, as in most parts of the book,
in the Full Markov Triple framework. Furthermore, according to the convention set
forth there, we use the terminology “Markov Triple” for Full Markov Triple. Results
with specific hypotheses will be clearly indicated.
More general settings may be considered for the analysis of spectral gap inequal-
ities. In order to state the Poincaré inequalities, it is actually enough to deal with the
minimal structure (E, μ, ) with finite invariant measure μ, even without the dif-
fusion property. Standard Markov Triples suffice for many of the results described
below, Full Markov Triples being necessary when dealing with inequalities involv-
ing gradient bounds and curvature conditions.
Recall that the Markov Triple setting includes a triple (E, μ, ) with state space
E, invariant reversible measure μ and carré du champ operator acting on an alge-
bra of bounded measurable functions A0 . The setting involves the Dirichlet form

E(f ) = (f )dμ
E
with domain D(E) and the associated diffusion operator L with domain D(L), gen-
erator of the Markov semigroup P = (Pt )t≥0 . Most inequalities are usually stated
and established for functions in D(E). In general, they are established in fact only
for the functions of the algebra A0 , which by construction is dense in the Dirichlet
domain D(E). The extended algebra A ⊃ A0 allows for the -calculus and gradient
bounds. In concrete examples, A0 plays the role of the set of smooth compactly sup-
ported functions while A represents smooth functions with no restriction on integra-
bility and support. The associated Markov semigroup P = (Pt )t≥0 with infinitesimal

DOI 10.1007/978-3-319-00227-9_4,
178 4 Poincaré Inequalities
generator L and invariant reversible measure μ admits a kernel representation

Pt f (x) = f (y) pt (x, dy), t ≥ 0, x ∈ E.
E
The Poincaré, or spectral gap, inequality is the simplest inequality which quantifies
ergodicity and controls convergence to equilibrium of the semigroup P = (Pt )t≥0
towards the invariant measure μ (in other words, the convergence of the kernels
pt (x, dy), x ∈ E, as t → ∞, towards dμ(y)). Since the kernels pt (x, dy) describe
the distribution at time t of the associated Markov process Xtx starting at x, this con-
vergence is translated equivalently as a convergence of Xtx towards the equilibrium.
On the basis of the example of the Ornstein-Uhlenbeck semigroup (Sect. 4.1),
Sect. 4.2 introduces the formal definition of Poincaré or spectral gap inequalities
in the context of a Markov Triple (E, μ, ) as above. Exponential decay in L2 (μ)
along the semigroup is part of the first main equivalent descriptions of Poincaré
inequalities. Tensorization properties are presented next. The example of the ex-
ponential measure on the line is discussed in Sect. 4.4, together with exponential
integrability properties of Lipschitz functions under a Poincaré inequality. The next
section studies Poincaré inequalities for measures on the real line or on an inter-
val of the real line. Section 4.5.1 first presents a characterization of measures on
the line satisfying a Poincaré inequality. Then, in Sect. 4.5.2, Poincaré inequalities
on an interval with respect to Neumann, Dirichlet and periodic boundary condi-
tions are discussed. Section 4.6 presents the Lyapunov function method of obtain-
ing Poincaré inequalities. Local Poincaré inequalities for heat kernel measures un-
der curvature conditions are investigated next in Sect. 4.7 via the basic semigroup
interpolation scheme. Global Poincaré inequalities for the invariant measure under
curvature-dimension conditions are developed in Sect. 4.8. Further inequalities of
Brascamp-Lieb-type are presented in Sect. 4.9. Finally, Sect. 4.10 is a somewhat
in depth investigation of the description of the bottom of the spectrum and essen-
tial spectrum of Markov generators together with criteria of general interest which
establish the discreteness of spectra via the notion of a Persson operator.
4.1 The Example of the Ornstein-Uhlenbeck Semigroup

The example of the Ornstein-Uhlenbeck semigroup will give the flavour of some of
the spectral properties and results that will be defined and investigated throughout
the chapter in greater generality. This elementary model will moreover provide an
opportunity to detail, step by step, some of the reasoning and tools which will be
developed later in more general cases.
The Ornstein-Uhlenbeck semigroup P = (Pt )t≥0 on the real line (for simplicity),
described in Sect. 2.7.1, p. 103, admits the standard Gaussian probability measure
dμ(x) = (2π)−1/2 e−x /2 dx as invariant and reversible measure, Lf = f − xf
2
as infinitesimal generator and (f ) = f 2 as carré du champ operator (acting

on smooth functions f : R → R). The Dirichlet form is given accordingly by
4.1 The Example of the Ornstein-Uhlenbeck Semigroup 179

E(f ) = R f 2 dμ. (While the set of smooth compactly supported functions is a

natural choice for the algebra A0 , it could be convenient here to work alternatively,
after some adaptations, with the algebra of polynomials.)
Convergence to equilibrium is best presented first on the spectral decomposition
of P and L which has been described in terms of Hermite polynomials. Recall (see
Sect. 2.7.1, p. 103) that a function f in L2 (μ) can be decomposed with respect to
the basis of Hermite orthonormal polynomials (Hk )k∈N as

f= ak Hk ,
k∈N

where, for every integer k, ak = Rf Hk dμ and

f 22 = f 2 dμ = ak2 < ∞.
R k∈N
In this case,

Pt f = e−kt ak Hk , t ≥ 0,
k∈N

so that Pt f converges in L2 (μ)towards a0 = R f H0 dμ = R f dμ as t → ∞ (re-

call that H0 = 1). This property expresses equivalently that the distribution of the
associated Ornstein-Uhlenbeck process Xtx converges as t → ∞ towards the invari-
ant measure μ (from any initial point x).
The task now is to control and quantify
this convergence.
On the basis of the
preceding decomposition, observe that f − R f dμ = k≥1 ak Hk so that, for every
t ≥ 0,
2 2

Pt f − f dμ = e −2kt 2
a ≤ e −2t
a 2
= e −2t
f − f dμ .
k k
R 2 k≥1 k≥1 R 2
This inequality describes an exponential convergence of Pt f towards the mean of

f ∈ L2 (μ) as t → ∞.
The latter family of inequalities may be analyzed infinitesimally. Assume for
simplicity that the mean of f is equal to 0 so that

(Pt f )2 dμ ≤ e−2t f 2 dμ
R R
for every t ≥ 0. Since both terms of this inequality are equal at t = 0, we can perform
a first order Taylor expansion at t = 0 of both sides, for any function f in the domain
D(L). More precisely, since Pt f = f + t Lf + o(t), we get that

(Pt f )2 dμ = f 2 dμ + 2t f Lf dμ + o(t)
R R R

= f 2 dμ − 2t E(f ) + o(t)
R
and, clearly,

−2t
e f dμ =
2
f dμ − 2t
2
f 2 dμ + o(t).
R R R

Hence R f 2 dμ ≤ E(f ). Since E(f + c) = E(f ) for any real c, it follows, after
centering, that for any function f ∈ D(L), and thus by extension for every f in the
Dirichlet domain D(E) of the Ornstein-Uhlenbeck operator,
2
f 2 dμ − f dμ ≤ E(f ). (4.1.1)
R R
This inequality (4.1.1) (or rather family of inequalities) may be read in two ways.
First, for smooth functions f , by construction of the carré du champ operator ,
2
f 2 dμ − f dμ ≤ E(f ) = (f )dμ = f 2 dμ.
R R R R
On the other hand, (4.1.1) expresses equivalently that the first non-trivial eigen-
value of (the opposite generator) −L is greater than or equal to
1. Indeed, if f
is non-constant such that −Lf = λf for some λ (≥ 0), then R f dμ = 0 and

2
R (f )dμ = R f (−Lf )dμ = λ R f dμ so that λ ≥ 1. Inequality (4.1.1) will
be called the Poincaré, or spectral gap, inequality for the Gaussian measure μ (with
respect to the carré du champ operator or the Dirichlet form E).
Note that (4.1.1) may be established directly on the spectral decomposition in
Hermite polynomials since
2
E(f ) = k ak2 ≥ ak2 = f dμ −
2
f dμ ,
k≥1 k≥1 R R
where, again, the last equality is a consequence of the fact that a0 = R f dμ. Note
also that (4.1.1) is optimal in the sense that functions of the form f = aH0 + bH1 =
a + bx for a, b ∈ R achieve equality in (4.1.1), thus describing the class of extremal
functions for this inequality.
Observe that similar arguments on multiple Hermite polynomials (or ten-
sorization tools as will be developed in Sect. 4.3 below) yield a Poincaré or
spectral gap inequality for the standard Gaussian probability measure dμ(x) =
(2π)−n/2 e−|x| /2
dx on Rn with carré du champ operator (f ) = |∇f |2 and Dirich-
2
let form E(f ) = Rn |∇f |2 dμ as described in the next proposition.
Proposition 4.1.1 (Poincaré inequality for the Gaussian measure) Let μ be the
standard Gaussian measure on the Borel sets of Rn . For every function f : Rn → R
in the Dirichlet domain D(E),
2
f dμ −
2
f dμ ≤ E(f ) = |∇f |2 dμ.
Rn Rn Rn
4.2 Poincaré Inequalities 181
If μ is a centered Gaussian measure on Rn with covariance matrix Q, by a simple

change of variables,
2
f 2 dμ − f dμ ≤ Q∇f · ∇f dμ
Rn Rn Rn
for every smooth function f : Rn → R.
4.2 Poincaré Inequalities

This section formally introduces the notion of Poincaré or spectral gap inequalities
for Markov operators and their invariant measures, and some of their first proper-
ties, in the context of Markov Triples (E, μ, ). While spectral properties may be
analyzed for general measures, the formal notion of a Poincaré inequality deals with
invariant probability measures.
4.2.1 Poincaré and Spectral Gap Inequalities
For a probability measure ν on a measurable space (E, F), we define the variance
of a function f in L2 (ν) as
2
Varν (f ) = f 2 dν − f dν . (4.2.1)
E E
Definition 4.2.1 (Poincaré inequality) A Markov Triple (E, μ, ) (with μ a prob-

ability measure) is said to satisfy a Poincaré, or spectral gap, inequality P (C) with
constant C > 0, if for all functions f : E → R in the Dirichlet domain D(E),
Varμ (f ) ≤ C E(f ).
The best constant C > 0 for which such an inequality holds is sometimes referred
to as the Poincaré constant (of the Markov Triple).
It is enough to state such a Poincaré inequality for a family of functions f which
is dense in the domain D(E) of the Dirichlet form E. This principle will be used
almost automatically when establishing Poincaré (and more general functional) in-
equalities, and typically functions from the underlying algebra A0 will be used to
this task.
The terminology between Poincaré and spectral gap inequality is somewhat fluc-
tuant. We often say that the probability measure μ satisfies a Poincaré inequality
P (C) with constant C (with respect to the carré du champ operator or the Dirich-
let form E). For example, by Proposition 4.1.1, the standard Gaussian measure μ on
Rn satisfies a Poincaré inequality P (1) with respect to the standard Dirichlet form

E(f ) = Rn |∇f |2 dμ. If f is an eigenfunction of −L with eigenvalue λ, applying
spectral gap inequality P (C) to f shows that, if λ = 0, then Cλ ≥ 1. Indeed,

the
E f dμ = 0 by invariance and

f 2 dμ = Varμ (f ) ≤ C E(f ) = C f (−Lf )dμ = Cλ f 2 dμ.
E E E
Hence, under P (C), every non-zero eigenvalue of −L is greater than or equal to

1
C . Indeed, even when it is not discrete, the spectrum of the symmetric positive
operator −L, as defined in Sect. A.4, p. 478, is included in {0} ∪ [ C1 , ∞), and P (C)
thus describes a gap in its spectrum.
We shall return to the following remark in several different contexts later.
Remark 4.2.2 A function with zero variance is constant. Therefore, one impor-
tant consequence of a spectral gap inequality for the invariant measure μ is that
if E(f ) = 0, then f is (μ-almost everywhere) constant. In particular, such a con-
clusion reinforces the ergodicity property of Definition 3.1.11, p. 135. It is also a
stronger form of connexity (Definition 3.3.7, p. 156). Functional inequalities with
this property will be called tight. A general principle, to be developed throughout
this book, is that tight inequalities have something to say about convergence to equi-
librium (cf. Sect. 1.8, p. 32).
Actually, as will be developed later for other functional inequalities, one can
consider non-tight Poincaré inequalities such as
2
f 2 dμ ≤ a f dμ + C E(f ), f ∈ D(E).
E E
This would not be of much help here, at least for probability measures, since this

f to the constant function f = 1 shows that a ≥ 1. Then,
inequality applied for
applied to g = f − E f dμ, we are back to the case a = 1. On the other hand, the
Nash-type inequalities investigated later in Sect. 7.4.1, p. 364,
2
f 2 dμ ≤ a |f |dμ + C E(f ), f ∈ D(E),
E E
will be of much greater interest.

In this spirit, the following elementary observation will prove useful in several
instances. We refer to Sect. 4.10.1 for probabilistic interpretations of these kind of
inequalities.
Lemma 4.2.3 Under the Poincaré inequality P (C), for a function f ∈ D(E) such
that f = 0 outside a set A ∈ F with μ(A) < 1, it holds that

C
f 2 dμ ≤ E(f ).
A 1 − μ(A)
4.2 Poincaré Inequalities 183
The proof is a simple application of the Cauchy-Schwarz inequality. Indeed,

2 2
f dμ = f dμ ≤ μ(A) f 2 dμ,
E A E

so that (1 − μ(A)) E f 2 dμ ≤ Varμ (f ) from which the claim follows by definition

of the Poincaré inequality P (C).
Remark 4.2.4 Inequalities of the form

f 2 dμ ≤ μ(A) E(f )
A
for functions f ∈ D(E) with support in A as in the previous lemma may hold in
various contexts. For example, the so-called Faber-Krahn inequalities, in the con-
text of Nash inequalities, belong to this family (see Remark 6.2.4, p. 284, and Re-
mark 8.2.2, p. 399).
4.2.2 Variance Decay
As illustrated earlier for on the Ornstein-Uhlenbeck semigroup, a Poincaré inequal-

ity for a probability measure μ describes (and in fact is equivalent to) exponential
convergence of the semigroup (Pt )t≥0 to equilibrium in L2 (μ). The same proof may
be used to derive the following main statement which formalizes this equivalence in
the general framework.
Theorem 4.2.5 (Exponential decay in variance) Given a Markov Triple (E, μ, )

with associated Markov semigroup P = (Pt )t≥0 , the following assertions are equiv-
alent:
(i) (E, μ, ) satisfies a Poincaré inequality P (C).
(ii) For every function f : E → R in L2 (μ), and every t ≥ 0,
Varμ (Pt f ) ≤ e−2t/C Varμ (f ).
(iii) For every function f ∈ L2 (μ), there exists a constant c(f ) > 0 (possibly de-
pending on f ) such that, for every t ≥ 0,
Varμ (Pt f ) ≤ c(f ) e−2t/C .
Proof Exactly as for the Ornstein-Uhlenbeck semigroup in Sect. 4.1, assertion

(i) follows from assertion (ii) by a first order Taylor expansion at t = 0. The converse
implication relies on the formula
d
Var(Pt f ) = −2 E(Pt f ) (4.2.2)
dt
for every f ∈ L2 (μ) and t > 0. Recall that according to Proposition 3.1.6 and Re-
mark 3.1.7, p. 131 and p. 132, Pt , t > 0, maps L2 (μ) into D(Lk ) for every in-
teger
k. Assume

then for simplicity that f has mean zero. Since by invariance,
E P t f dμ = E f dμ = 0,

(t) = Varμ (Pt f ) = (Pt f )2 dμ, t ≥ 0.
E
By the heat equation and integration by parts,

(t) = 2 Pt f LPt f dμ = −2 E(Pt f )
E
which yields (4.2.2). The Poincaré inequality P (C) therefore translates into the dif-
ferential inequality (t) ≤ − C2 (t), t > 0, which amounts to saying that the func-
tion e2t/C (t) is decreasing in t ≥ 0. The conclusion follows by comparing the
values of this function at t and at t = 0.
To prove the equivalence of the third statement (iii), it is convenient to record the
following lemma of independent interest.
Lemma 4.2.6 For any non-zero function f ∈ L2 (μ), the map t → log(Pt f 2 ) is
convex on R+ .

The proof is immediate. Indeed, if (t) = E (Pt f ) dμ, t > 0, as before, (t) =
2
2 E Pt f LPt f dμ and

(t) = 4 (LPt f )2 dμ.
E
Hence (t)2 ≤ (t) (t)

by the Cauchy-Schwarz inequality which amounts to
the convexity of log (t).
Returning to the proof of Theorem 4.2.5, in order to deduce (ii) from (iii), it
suffices from the preceding lemma to observe that if a convex function φ (on the
positive half-line) is bounded from above by β − αt, then it is bounded by φ(0) − αt.
The proof of the theorem is therefore complete.
For later purposes, it is useful to draw from the preceding arguments a general
inequality on the L2 (μ)-norm of f − Pt f (which sheds some further light on the
preceding equivalences). Namely, for every f ∈ D(E) and every t ≥ 0,

(f − Pt f )2 dμ ≤ 2 t E(f ). (4.2.3)
E
The elementary

proof follows the same pattern. For a function f ∈ D(E), the deriva-
tive of E (f − Pt f )2 dμ in t for t > 0 is equal, by symmetry and integration by
parts, to

−2 (f − Pt f ) LPt f dμ = 2 E(Pt/2 f ) − 2 E(Pt f ) ≤ 2 E(f ),
E
4.3 Tensorization of Poincaré Inequalities 185
since t → E(Pt f ) is decreasing (see Proposition 3.1.6, p. 131).

It is, in general, not so easy to establish a Poincaré inequality for a probability
measure μ with respect to a given carré du champ operator . One first approach
is to compare it to a simpler, or known, inequality. A first trivial observation in this
direction is that whenever μ satisfies a Poincaré inequality P (C) with respect to ,
and if 1 is another carré du champ operator such that ≤ a 1 for some a > 0,
then (μ, 1 ) satisfies P (aC).
The following proposition is a further simple, but important, observation con-
cerning stability by perturbation of Poincaré inequalities.
Proposition 4.2.7 (Bounded perturbation) Assume that the Markov Triple (E, μ, )
satisfies a Poincaré inequality P (C). Let μ1 be a probability measure with density
h with respect to μ such that b1 ≤ h ≤ b for some constant b > 0. Then μ1 satisfies
P (b3 C) (with respect to ).
The proof is elementary, since by duplication, for every probability measure ν,

1 2
Varν (f ) = f (x) − f (y) dν(x)dν(y). (4.2.4)
2 E E
With the somewhat more refined Lemma 5.1.7, p. 240, below, b3 may be replaced
by b2 in the preceding. An even more precise statement is that if h = ek , then b2 may
be replaced by eosc(k) where osc(k) = sup k − inf k is the oscillation of the function
k. (As usual, sup and inf stand here respectively for esssup and essinf with respect
to the measure μ.)
These first elementary observations already yield Poincaré inequalities for mea-
sures μ on R or Rn with a bounded density with respect to the
standard Gaussian
measure, and for Dirichlet forms, on the line say, of the form R a(x)f 2 (x) dμ(x)
where the function a(x) is bounded from below by a strictly positive constant.
4.3 Tensorization of Poincaré Inequalities
This section develops tensorization properties of Poincaré inequalities. One impor-

tant feature of Poincaré inequalities is their stability under products.
Proposition 4.3.1 (Stability under products) If (E1 , μ1 , 1 ) and (E2 , μ2 , 2 ) sat-

isfy Poincaré inequalities with respective constants C1 and C2 , then the product
Markov Triple (E1 × E2 , μ1 ⊗ μ2 , 1 ⊕ 2 ) satisfies a Poincaré inequality with
constant C = max(C1 , C2 ).
The proof will give a meaning to 1 ⊕ 2 , more precisely to the underlying

product Dirichlet form.
Proof As a first indication of the result, consider the elementary case when the cor-
responding (opposite) generators −L1 and −L2 have discrete spectra, (λ1k )k∈N and
(λ2 )∈N . The (opposite) generator −(L1 ⊕ L2 ) on the product space E1 × E2 has
spectrum (λ1k + λ2 )k,∈N . Its smallest strictly positive eigenvalue is thus the mini-
mum of the smallest strictly positive eigenvalues of (λ1k )k∈N and (λ2 )∈N . The con-
clusion follows in this case with C1 = 11 and C2 = 12 .
λ1 λ1
The general case is not much harder. The main point is to understand that the
Dirichlet form E on the product (E1 × E2 , μ1 ⊗ μ2 , 1 ⊕ 2 ) is given, on suitable
functions f : E1 × E2 → R, by

E(f ) = E1 (f )dμ2 + E2 (f )dμ1 .
E2 E1
Here E1 is the Dirichlet form on E1 applied to the first variable (with the second
one being fixed), and similarly E2 is the Dirichlet form on E2 applied to the second
variable (with the first one being fixed). First we need the following lemma.
Lemma 4.3.2 If f = f (x1 , x2 ) is a function of two variables (x1 , x2 ) ∈ E1 × E2 ,

measurable in x2 and in the Dirichlet space D(E1 ) for the first variable (that is, for
every x2 ∈ E2 , E1 (f (·, x2 )) is finite), then the function

g(x1 ) = f (x1 , x2 )dμ2 (x2 ), x1 ∈ E1 ,
E2
is in the Dirichlet space D(E1 ) and moreover

E1 (g) ≤ E1 f (·, x2 ) dμ2 (x2 ).
E2
Proof Rather than giving a formal proof, let us sketch the main idea. Dirichlet forms
are bilinear and positive, sending functions from the Dirichlet spaces into R+ . Such
quadratic forms are convex, and if φ : R → R is convex, then by Jensen’s inequality
(under suitable integrability properties),

φ F dμ2 ≤ φ(F )dμ2 .
E2 E2
This inequality applied to F (x2 ) = f (·, x2 ) then yields the announced claim.
We now conclude the proof of Proposition 4.3.1. Let f : E1 × E2 → R be in the

Dirichlet space of the product space, and let g(x1 ), x1 ∈ E1 , be as in Lemma 4.3.2.
The variance of f is easily decomposed as

2
Varμ1 ⊗μ2 (f ) = Varμ1 (g) + f (x1 , x2 ) − g(x1 ) dμ2 (x2 ) dμ1 (x1 ).
E1 E2
4.4 The Example of the Exponential Measure, and Exponential Integrability 187
By the Poincaré inequality for μ1 , the first term on the right-hand side
of this identity
is bounded from above by C1 E1 (g), and by Lemma 4.3.2, E1 (g) ≤ E2 E1 (f ) dμ2 .
By the Poincaré inequality for μ2 for every fixed
x1 ∈ E1 , the second term on the
right-hand side is bounded from above by C2 E1 E2 (f ) dμ1 . Therefore

Varμ1 ⊗μ2 (f ) ≤ C1 E1 (f ) dμ2 + C2 E2 (f ) dμ1 ≤ max(C1 , C2 ) E(f ).
E2 E1
Proposition 4.3.1 is established.
Remark 4.3.3 A somewhat different route may actually be used to establish Propo-
sition 4.3.1. On the product space E = E1 × E2 equipped with the product proba-
bility measure μ = μ1 ⊗ μ2 , consider, for every t ≥ 0 and x ∈ E, the probability
measure νt (dx, dy) = pt (x, dy)μ(dx) where pt (x, dy) denotes the kernel of the
semigroup (Pt )t≥0 (on E). As for the variance, it is easily checked that

2
f (x) − f (y) νt (dx, dy) = 2 f dμ − f Pt f dμ .
2
E E E E
When t = ∞, ν∞ (dx, dy) = μ(dx)μ(dy), which simply

corresponds

to the idea
that P∞ (x, dy) = μ(dy). Now, since by symmetry E f Pt f dμ = E (Pt/2 f )2 dμ,
the Poincaré inequality P (C) for μ is equivalent by Theorem 4.2.5 to

2
f (x) − f (y) νt (dx, dy)
E E

−t/C
2
≥ 1−e f (x) − f (y) ν∞ (dx, dy)
E E
(for every t ≥ 0). Since it is clear that νt = νt1 ⊗ νt2 corresponds to the tensor product
of the respective semigroups on E1 and E2 , the preceding inequality is immediately
stable under products.
The tensorization Proposition 4.3.1 may for example be used immediately to re-
cover the Poincaré inequality of Proposition 4.1.1 for the standard Gaussian measure
on Rn from its one-dimensional counterpart (4.1.1). Similarly, the Poincaré inequal-
ity for the exponential measure investigated in the next section may be tensorized to
multiple coordinates.
4.4 The Example of the Exponential Measure, and Exponential

Integrability
The preceding elementary tools are usually not sufficient to cover all the instances
of Poincaré inequalities. Another approach is to try, in a given context, to find
some particular functions with pleasant properties (called Lyapunov functions in
a Markov theory setting). We illustrate this strategy here with the example of the
exponential measure on the real line (and positive half-line), generalized later in
Sect. 4.6. On the basis of this example, we shall derive some integrability properties
of Lipschitz functions and measure concentration under a Poincaré inequality.
4.4.1 The Exponential Measure
The exponential measure dμ(x) = e−x dx on the positive half-line R+ , as a par-

ticular example of a gamma distribution with parameter 1, has been discussed in
Sect. 2.7.3, p. 111, on the Laguerre operator, in connection with spectral decompo-
sitions along Laguerre orthogonal polynomials. In particular, following the discus-
sion of Sect. 4.1 on the Gaussian measure, similar spectral decomposition arguments
may be used to show that for every function f : R+ → R in the Dirichlet domain of
the Laguerre operator,

Varμ (f ) ≤ xf 2 dμ (4.4.1)
R+
(the right-hand side describing the carré du champ operator and Dirichlet form as-
sociated to the Laguerre operator). The spectrum here is discrete (equal to N) and
the Poincaré inequality (4.4.1) expresses the gap between 0 and the first non-zero
eigenvalue 1.
However, one may also investigate the Poincaré inequality for the exponential
measure with respect to the standard carré du champ operator (f ) = f 2 and its
associated Dirichlet form E which is the content of the next proposition.
Proposition 4.4.1 Let dμ(x) = e−x dx on the positive half-line R+ . For every func-
tion f : R+ → R in the Dirichlet domain D(E) associated to the carré du champ
operator (f ) = f 2 ,

Varμ (f ) ≤ 4 f 2 dμ. (4.4.2)
R+
It should be pointed out that the Dirichlet domain in this case generates a sub-
Markov semigroup so that Proposition 4.4.1 is not really in the realm of Poincaré
inequalities. Moreover the essential self-adjointness property (Definition 3.1.10,
p. 134) is certainly not satisfied here (cf. Sect. 2.4, p. 92).
Proof It is enough to consider smooth compactly supported functions f : R+ → R.

Note first that

2
Varμ (f ) ≤ f − f (0) dμ
R+

since the variance minimizes the distance to constants in L2 (μ). Since R+ f 2 dμ

is not modified by adding a constant to f , it will be enough to establish the desired
inequality for functions f such that f (0) = 0. Then, by Fubini’s Theorem,

∞
f 2 dμ = f (x)2 e−x dx
R+ 0
∞ x
= 2 f (t)f (t)dt e−x dx
0 0
∞
=2 f (t)f (t) e−t dt = 2 ff dμ.
0 R+
The conclusion follows by applying the Cauchy-Schwarz inequality to the last inte-
gral.
We next somewhat broaden the spirit of the preceding proof. For this discus-
sion, let us consider the symmetric exponential measure on the whole real line,
dμ(x) = 12 e−|x| dx. The density of μ with respect to the Lebesgue measure is not
quite smooth and, if necessary, the potential x → |x| may be replaced by a smooth
function with the same behavior at infinity. However, we work with the given den-
sity since the argument which is developed is simpler in this case. The associated
Markov generator is Lf = f − sign(x)f , where sign(x) denotes the sign of x,
with carré du champ operator (f ) = f 2 . Its spectrum is not discrete. This is actu-
ally not due to the non-smooth coefficient, but rather to the behavior of |x| at infinity
so that the decay of μ is not strong enough. For example, it is easily seen that as
soon as λ ≥ 14 , there are no solutions to f − f = −λf which are square integrable
on (0, ∞).
Nevertheless, the operator L admits a Poincaré inequality (and thus a gap in the
spectrum) as can be seen from the following argument directly inspired by the proof
of Proposition 4.4.1. What follows is a bit formal, but may be easily justified. Con-
sider the function h(x) = |x|, x ∈ R. Then Lh = 2δ0 − 1 where δ0 is the Dirac mass
at 0, so that integration against it with a (smooth) function f such that f (0) = 0
gives 0. Let f be such a function, and write

f 2 dμ = − f 2 Lh dμ = f 2 , h dμ = 2 f (f, h)dμ.
R R R R
Now (h) = 1 so that |(f, h)| ≤ (f )1/2 . Hence, after an application of the
Cauchy-Schwarz inequality as in the proof of Proposition 4.4.1,
1/2 1/2
f 2 dμ ≤ 2 f 2 dμ (f )dμ ,
R R R
and therefore

Varμ (f ) ≤ 4 (f )dμ = 4 f 2 dμ (4.4.3)
R R
which is the announced Poincaré inequality for the symmetric exponential measure
μ on R. It is worth pointing out that the significant component of the argument is
that −Lh is bounded from below by a strictly positive constant away from 0 (outside
a compact set containing 0 would have been enough) and that (h) is bounded from
above. This principle will be amplified in Sect. 4.6 below, which deals with the
so-called Lyapunov functions.
4.4.2 Exponential Integrability
One miracle of the previous argument is that the constant 4 in Proposition 4.4.1
and (4.4.3) is optimal. To see why, we examine another consequence of Poincaré
inequalities, namely the integrability properties of Lipschitz functions. Recall the
notion of a Lipschitz function f from Definition 3.3.24, p. 166, with Lipschitz co-
1/2
efficient f Lip = (f )∞ . A function f is said to be 1-Lipschitz if f Lip ≤ 1.
Under a Poincaré inequality for a measure μ, Lipschitz functions are exponen-
tially integrable. This is the content of the following statement.
Proposition 4.4.2 (Exponential integrability) If μ satisfies a Poincaré inequality

P (C) for some C > 0, then for every 1-Lipschitz function f and every s < C4 ,

esf dμ < ∞.
E
Remark 4.4.3 Lipschitz functions belong either to A or to D(E). It is part of the

result here that Lipschitz functions are indeed in L2 (μ), and therefore also belong to
D(E), so that the statement may not be restricted to Lipschitz functions f ∈ D(E).
While the result is stated for f ∈ A, it applies with the same argument to Lipschitz
functions in D(E). In fact, in the core of the proof below, it is convenient to work
with functions in D(E) allowing for a simple truncation procedure, which however
does not preserve functions in A. The proof for functions in A is easily adapted
using a smooth truncation procedure.
Proof We first show that, under the Poincaré inequality

P (C), for every integrable
Lipschitz function f with f Lip ≤ 1, and every s < 4
C,

∞
−2
Cs 2
e dμ ≤ e
sf s E f dμ
1 − +1 . (4.4.4)
E 4
=0
It is enough to establish this inequality for a bounded function f (indeed one can
work with fk = (f ∧ k) ∨ (−k), k ∈ N, for which (fk ) ≤ 1 since (ψ(f )) ≤ (f )
for any contraction ψ : R → R, and let then k → ∞ by Fatou’s lemma). Apply the
Poincaré inequality P (C) to g = esf/2 , s ∈ R. Since

s2 s2
E(g) = e (f )dμ ≤
sf
esf dμ,
4 E 4 E

we get, setting Z(s) = Ee

sf dμ,
2
Cs 2 s
1− Z(s) ≤ Z .
4 2
Cs 2
If 1 − 4 > 0, then
2
Cs 2 −1 s
Z(s) ≤ 1 − Z .
4 2

Iterating the procedure by replacing s by s
yields the claim since lim→∞ Z( 2s )2 =

2
es E f dμ .
To reach the full conclusion of the statement, it remains to show that every
1-Lipschitz function f is indeed integrable (provided it is finite μ-almost every-
where). To this end, apply the Poincaré inequality to fk = |f | ∧ k for every k ∈ N.
In particular, since (fk ) ≤ (|f |) ≤ (f ) ≤ 1,
2
fk − fk dμ dμ ≤ c1
E E
for some constant c1 > 0 independent of k, and therefore

1

μ fk − fk dμ ≥ 2 c1 ≤ .
E 2
Let c2 > 0 be large enough so that μ(|f | ≥ c2 ) < 12 (since f is μ-almost every-
where finite). Hence, μ(|fk | ≥ c2 ) < 12 for every k. Now, if two measurable sets
S1 and S2 are such that μ(S1 ) ≥ 12 and μ(S2 ) > 12 , then, μ(S1 ∩ S2 ) > 0, and in
particular
S1 ∩ S2 = ∅.
√Therefore, for every k ∈ N, there exists an x ∈ E such that
|fk (x) − E fk dμ| < 2c1 and |fk (x)| < c2 , and hence

fk dμ < 2 c1 + c2 .
E
Since c1 and c2 are independent of k, it follows from Fatou’s lemma that

E |f |dμ < ∞. Hence
(4.4.4) applies, and since the infinite product is finite un-
der the condition s < C4 , the conclusion follows. The proof of Proposition 4.4.2 is
complete.
A careful examination of the infinite product

of (4.4.4) shows that (in the setting
4
described in Proposition 4.4.2), for every s < C,
√

2+s C
s(f − E f dμ)
e dμ ≤ √ . (4.4.5)
E 2−s C
Returning to the example of the exponential measure dμ(x) = 12 e−|x| dx on R, it

satisfies P (C) with C = 4 by (4.4.3). This corresponds to the critical exponent s = 1
in Proposition 4.4.2 achieved for the Lipschitz function f (x) = |x|. This indicates
that both the spectral gap constant in (4.4.3) as well as the exponent in Proposi-
tion 4.4.2 are optimal.
To conclude this sub-section, we briefly observe that new Poincaré inequalities
may be obtained from a given one by a contraction principle via Lipschitz functions.
Proposition 4.4.4 (Contraction principle) If (μ, ) satisfies a Poincaré inequality

P (C) on a state space E, and if F : E → R is 1-Lipschitz with respect to , then
the image measure μF of μ by F satisfies a Poincaré inequality P (C) with respect
to the usual carré du champ operator on the real line.
Proof The proof is immediate. For every suitable function g on R,
VarμF (g) = Varμ (g ◦ F ).
By the Poincaré inequality P (C) for μ and the change of variables formula,

Varμ (g ◦ F ) ≤ C (g ◦ F )dμ
E

g ◦ F (F )dμ
2
=C
E

2
g dμF ,
2
≤C g ◦ F dμ = C
E R
which is the announced claim.
4.4.3 Exponential Measure Concentration
Proposition 4.4.2 may be reformulated in the form of an exponential measure con-

centration property for the measure μ satisfying a Poincaré inequality P (C) for
some C > 0. Namely, for every Lipschitz function f and every r ≥ 0,
√
μ f≥ f dμ + r ≤ 3 e−r/ C f Lip . (4.4.6)
E
This is an immediate application of (4.4.5) with, for simplicity, s = √1 together

C
with Markov’s exponential inequality and homogeneity. The numerical constants
are not sharp. The inequality (4.4.6) is actually more a deviation inequality of f
from its mean. A concentration inequality around the mean is easily obtained to-
gether with the same inequality for −f yielding

√

μ f − f dμ ≥ r ≤ 6 e−r/ C f Lip , r ≥ 0.
E
4.5 Poincaré Inequalities on the Real Line 193
As a consequence of the measure concentration inequality (4.4.6), the image un-

der a Lipschitz map of a measure satisfying a spectral gap inequality is a measure
with tails at most exponential. This is in particular the case for the exponential mea-
sure itself. Together with the tensorization principle of Sect. 4.3, various applica-
tions of interest may be produced. For example, if f : R → R is Lipschitz with
respect to , then
1
F (x) = √ f (x1 ) + · · · + f (xn ) , x = (x1 , . . . , xn ) ∈ Rn ,
n
is Lipschitz on Rn with respect to the tensored on Rn with the same Lipschitz

parameter. Now, if μ is the probability distribution of some real-valued random
variable Z1 , μ⊗n is the distribution on Rn of the sample Z = (Z1 , . . . , Zn ) of inde-
pendent variables each with the same law as Z1 . Therefore, if μ satisfies a Poincaré
inequality P (C) with respect to the standard carré du champ operator on the line,
and if f Lip ≤ 1 for example, then
√
P F (Z) ≥ E F (Z) + r ≤ 3 e−r/ C
for every r ≥ 0. This is a consequence of the concentration inequality (4.4.6) applied

to the Lipschitz function F since μ⊗n satisfies P (C) on Rn by Proposition 4.3.1.
The preceding bound is independent of the dimension n. The same principle applies
to other Lipschitz functions of the sample Z, such as for example max1≤i≤n Zi .
Non-identically distributed samples may be considered similarly.
Remark 4.4.5 One may wonder how far the exponential integrability of Proposi-
tion 4.4.2, or the tail estimate (4.4.6), are from a Poincaré inequality. The Mucken-
houpt characterization presented in the next section clearly indicates that they are
not sufficient in general to entail a Poincaré inequality. However, under a curva-
ture condition CD(0, ∞), the concentration bound (4.4.6) surprisingly turns out to
imply in return a Poincaré inequality P (C ), which moreover has proportional con-
stants. This will be analyzed in Sect. 8.7, p. 425, on the basis of the local heat kernel
inequalities of Sect. 4.7 below (in particular their reverse forms).
4.5 Poincaré Inequalities on the Real Line
This section investigates Poincaré inequalities on the real line or on an interval of

the real line. We first discuss a useful criterion, known as Muckenhoupt’s crite-
rion, which characterizes Poincaré inequalities. Following that, we describe optimal
Poincaré inequalities on a bounded interval in terms of Neumann, Dirichlet and pe-
riodic boundary conditions.
4.5.1 The Muckenhoupt Criterion
This section describes a criterion for a measure μ on the real line R to satisfy a
Poincaré inequality with respect to the usual carré du champ operator (f ) = f 2
(recall that it is always possible to reduce to this case by a suitable change of vari-
ables, cf. Sect. 2.6, p. 97). This criterion, going back to the work of B. Muckenhoupt,
yields a useful necessary and sufficient condition for the Poincaré inequality to hold
(although it is not always completely straightforward to check). We only state it for
measures absolutely continuous with respect to the Lebesgue measure (see however
below). The Rn case will be treated in Chap. 8 in terms of more general measure-
capacity inequalities (which, however, are less useful in practice).
Theorem 4.5.1 (Muckenhoupt’s criterion) Let μ be a probability measure on the

Borel sets of R with density p with respect to the Lebesgue measure, and let m be a
median of μ (i.e. μ([m, +∞)) ≥ 12 and μ((−∞, m]) ≥ 12 ). Set

x 1
B+ = sup μ [x, +∞) dt
x>m m p(t)
and

m 1
B− = sup μ (−∞, x] dt.
x<m x p(t)
Then, a necessary and sufficient condition in order that μ satisfies a Poincaré in-
equality P (C) is that B = max(B+ , B− ) < ∞. Moreover, B ≤ 2C ≤ 8B where C is
the Poincaré constant of μ.
It should be mentioned that whenever B < ∞, the median is unique, since the
measure of any interval included in the support is strictly positive. A careful look
at the criterion indicates that in order for it to be satisfied, p1 has to be locally
integrable on the support of μ with respect to the Lebesgue measure. Hence the
measure μ should not have “holes” in its support. For example, it is immediately
checked that whenever p = 0 on some interval [x0 , x1 ] with m ≤ x0 < x1 such that
μ([x1 , +∞)) > 0, or when at some point x0 , p |x − x0 |, then the condition is not
satisfied. It is a good exercise to check that the measure on R with density cα e−|x|
α
with respect to the Lebesgue measure satisfies the Muckenhoupt criterion if and
only if α ≥ 1.
It is not necessary for a probability measure μ to have a density with respect
to the Lebesgue measure in order to satisfy a Poincaré inequality (use for example
the contraction Proposition 4.4.4 with a Lipschitz map which is constant on some
interval of positive measure). In fact, Theorem 4.5.1 may be extended and stated
without any change by taking p to be the density of the absolutely continuous part
of μ. For simplicity, and to make the exposition of the proof a little easier, we only
deal with measures with densities, and we assume that the suprema in B+ and B−
are taken over points x in the smallest closed interval which carries μ. The general
case is left to the reader.
We turn to the proof of Theorem 4.5.1

xunder these assumptions. We start with the
following simpler result. Set h+ (x) = m p(t)
1
dt, defined for x ≥ m, and similarly

m 1
h− (x) = x p(t) dt, defined for x ≤ m.
Proposition 4.5.2 If μ with density p is such that

∞ m
h+ (x) p(x)dx < ∞ and h− (x) p(x)dx < ∞
m −∞
where m is a median of μ, then it satisfies a Poincaré inequality.
Proof As in the proof of the Poincaré inequality for the exponential measure, it is
enough to bound from above
∞ m
2 2
f (x) − f (m) dμ and f (x) − f (m) dμ
m −∞

2 dμ.
by C E(f ) = C Rf By the Cauchy-Schwarz inequality, for every x > m,
2
2 x
f (x) − f (m) = f (t)dt
m
x 1
≤ f 2 (t) p(t)dt dt = E(f ) h+ (x).
R m p(t)
Therefore,
∞ ∞
2
f (x) − f (m) dμ ≤ E(f ) h+ (x) p(x)dx.
m m
Together with a similar treatment for the second integral, we immediately obtain a
Poincaré inequality with constant
∞ m
C = max h+ (x) p(x)dx, h− (x) p(x)dx .
m −∞
Next, we turn to the proof of Theorem 4.5.1 itself.
Proof of Theorem 4.5.1 First, the condition B < ∞ is sufficient for a Poincaré in-
equality P (C) to hold. To establish this claim, we proceed as in the proof of Propo-
sition 4.5.2 to bound [f (x) − f (m)] 2
for every x > m. However, we need a some-
what sharper bound. Set g(x) = h+ (x) and k(x) = g(x)p(x), x > m. Then, by
the Cauchy-Schwarz inequality,
2
2 x

x
2
x 1
f (x) − f (m) = f (t)dt ≤ f (t) k(t)dt dt.
m m m k(t)
Therefore, by Fubini’s Theorem,

∞ ∞ ∞ x
2 1
f (x) − f (m) p(x)dx ≤ f 2 (t)k(t) du p(x)dx dt.
m m t m k(u)
The choice of the function k is justified by the fact that g = 2k1

, and thus that
x 1
m k(u) du = 2g(x). The definition of B+ indicates that, for every x > m,

B+
g(x) ≤
μ([x, +∞))
so that, for t > m,

∞ x
∞
1 p(x)
du p(x)dx ≤ 2 B+ √ dx.
t m k(u) t μ([x, +∞))
√
Setting r(x) = μ([x, +∞)),
p(x)
√ = −2 r (x).
μ([x, +∞))
As a consequence of the preceding,

∞ ∞
2
f (x) − f (m) p(x)dx ≤ 4 B+ f 2 (t) k(t) r(t)dt.
m m
But

k(t) r(t) = p(t) μ [t, +∞) h+ (t) ≤ B+ p(t),
so that finally
∞ ∞
2
f (x) − f (m) p(x)dx ≤ 4B+ f 2 (t) p(t)dt.
m m

m
The term −∞ p(x)[f (x) − f (m)]2 dx is treated similarly with B− so that the mea-
sure μ indeed satisfies a Poincaré inequality with constant C = 4B.
Turning to the converse, the idea is to apply the Poincaré inequality P (C) to
suitable test functions f . We first replace the Poincaré inequality with the easiest
consequence of it given by Lemma 4.2.3 with A = [m, ∞) where m is the median
of μ (hence μ(A) ≤ 12 and 1−μ(A)1
≤ 2). Therefore, under P (C), for any function
f ∈ D(E) supported on [m, ∞),

f dμ ≤ 2C f 2 dμ.
2
R R

x
Then take a function g ∈ L2 (μ), and apply the latter to f (x) = m g(t)dt, x > m. It
follows that
∞ x 2 ∞
g(t)dt dμ(x) ≤ 2C g 2 (x)dμ(x).
m m m
Fix ε > 0 and r > m, and apply this inequality to the function
1
g(x) = 1[m,r] (x), x ∈ R.
p(x) + ε
The left-hand side of the previous inequality is therefore bounded from below by
∞ x 2 r 2
1
g(t) dt p(x)dx = dt μ [r, ∞) ,
r m m p(t) + ε
while the right-hand side is equal to

r r
p(x) 1
2C dx ≤ 2C dx.
m (p(x) + ε)2
m p(x) +ε
Comparing the two sides,

r 1
μ [r, ∞) dt ≤ 2C.
m p(t) + ε
As ε → 0, the first part of the conclusion, namely B+ ≤ 2C, follows. The second
part B− ≤ 2C is achieved similarly, yielding the announced bound B ≤ 2C. Theo-
rem 4.5.1 is established.
4.5.2 Poincaré Inequalities on an Interval
The Muckenhoupt criterion indicates that many probability measures on a bounded

interval of the real line satisfy a Poincaré inequality. This section is concerned with
optimal constants in certain instances. Dealing with intervals with boundaries, the
Full Markov Triple assumption (which requires essential self-adjointness) is not
assumed to hold in this sub-section, and instead we work with Diffusion Markov
Triples.
Before discussing the Poincaré inequalities themselves, it may be observed first
that, starting for example from the Poincaré inequality for the Gaussian measure on
the real line, the simple Lemma 4.2.3 yields inequalities of the form

f 2 dμ ≤ C E(f )
I
on any bounded interval I , for any smooth function f compactly supported in I

and for the usual carré du champ operator (f ) = f 2 , and for any measure μ with
strictly positive and bounded density. To understand the meaning of such inequali-
ties (which will be further explored in Sect. 4.10), consider for simplicity a bounded
interval I ⊂ R, with the usual carré du champ operator (f ) = f 2 and a reference
probability measure μ with a smooth density w bounded from below by some con-
stant a > 0. As described in Sect. 2.6, p. 97, this situation is related to the analysis

of the Sturm-Liouville operator Lf = f + ww f acting on smooth functions f on
an interval I ⊂ R. The following statement holds.

Proposition 4.5.3 For the Sturm-Liouville operator Lf = f + ww f acting on
smooth functions on a bounded interval I ⊂ R where w is smooth and bounded from
below by a constant a > 0, with invariant probability measure dμ = wdx, there is
a constant C > 0 such that, for any smooth function f compactly supported in the
interval I ,

f 2 dμ ≤ C f 2 dμ. (4.5.1)
I I
The best constant C in this inequality is where −λ0 is the lowest eigenvalue of
1
λ0 ,
the Sturm-Liouville operator L with Dirichlet boundary conditions (functions which
vanish at the boundary).
Proof From the analysis performed in Sect. 2.6, p. 97, the operator L defined on the
class A0 of smooth and compactly supported functions in the interval I extends to a
self-adjoint operator L (the Friedrichs extension) which is such that A0 is dense in
the Dirichlet domain D(E) of this extension, corresponding to the Dirichlet bound-
ary conditions. Moreover, the spectrum of (the opposite generator) −L is discrete,
consisting of an infinite increasing sequence of positive eigenvalues (λk )k∈N associ-
ated to eigenvectors (fk )k∈N which form an orthonormal basis of L2 (μ). It requires
a bit of further analysis (not developed here) to see that in fact λ0 is strictly pos-
itive, simple, and corresponds to a positive eigenvector f0 , strictly positive in the
interior of I (λ0 and f0 are the analogues of the Perron-Frobenius eigenvalue and
eigenvector in the context of finite Markov chains, see Sect. 1.9, p. 33).
applies. Namely, if f ∈ L (μ)
Now, the standard spectral decomposition picture 2
is developed in the basis (fk )k∈N in the form f = k∈N ak fk , then

−Lf = ak λk f k .
k∈N
Therefore,

E(f ) = − f Lf dμ = ak2 λk ≥ λ0 ak2 = λ0 f 2 dμ.
I k∈N k∈N I
Inequality (4.5.1) of Proposition 4.5.3 follows with C = λ10 , where λ10 appears to
be the best possible (since f0 saturates the inequality). On the other hand, if (4.5.1)
holds for all f ∈ A0 , then it extends to every function in the Dirichlet domain D(E)
by density. The proof of the proposition is therefore complete.
Of course, what is described in the previous proof extends similarly to bounded

domains on a manifold with any smooth elliptic operator and smooth density mea-
sure.
We next give an interpretation of the Poincaré inequality of Definition 4.2.1 for
the invariant measure of a Sturm-Liouville operator on an interval of the real line in
terms of Neumann boundary conditions. We again refer to Sect. 2.6, p. 97, for the
required background.

Proposition 4.5.4 For the Sturm-Liouville operator Lf = f + ww f acting on
smooth functions on an bounded interval I ⊂ R where w is smooth and bounded
from below by a constant a > 0, with invariant probability measure dμ = wdx,
there is a constant C > 0 such that, for every smooth function f on I (or restriction
to I of a smooth function defined on a neighborhood of I ),

Varμ (f ) ≤ C f 2 dμ. (4.5.2)
I
The best constant C in this inequality is λ11 , where −λ1 is the first non-zero eigen-
value of the Sturm-Liouville operator L with Neumann boundary conditions (func-
tions whose derivatives vanish at the boundary).
Proof This relies on the same analysis as that given in Proposition 4.5.3. The only
point to observe is that one may approximate a smooth function f defined in a
neighborhood of I by a sequence of smooth functions with derivatives which vanish
in a neighborhood of the boundaries, in such a way that the corresponding quantities
in the inequality (4.5.2) converge to the quantities for f . At a technical level, the idea
is to first approximate the derivatives. The details are left to the reader.
It is worth mentioning that there are cases when (4.5.2) is valid only for func-
tions which are compactly supported in I . As is illustrated by the next statement,
this situation corresponds to a spectral gap of the operator with periodic bound-
ary conditions. The next statement actually investigates more precisely the different
values of the constants in the preceding two propositions with the example of the
Lebesgue measure on the interval I = [0, 1].
Proposition 4.5.5 Let Lf = f be the Sturm-Liouville operator acting on smooth

functions f on [0, 1] with invariant measure the Lebesgue measure dx.
(i) For any smooth function f : [0, 1] → R such that f (0) = f (1) = 0,

1
f 2 dx ≤ 2 f 2 dx. (4.5.3)
[0,1] π [0,1]
The inequality is optimal and the sharp constant is attained by the function
f (x) = sin(πx).
(ii) For any smooth function f : [0, 1] → R,

1
Vardx (f ) ≤ 2 f 2 dx. (4.5.4)
π [0,1]
The inequality is optimal and the sharp constant is attained by the function
f (x) = cos(πx).
(iii) For any smooth function f : [0, 1] → R such that f (0) = f (1),

1
Vardx (f ) ≤ f 2 dx. (4.5.5)
4π 2 [0,1]
The inequality is optimal and the sharp constant is attained by functions of the
form f (x) = a cos(2πx) + b sin(2πx) + c where a, b, c are real constants.
Item (i) corresponds to the bottom of the spectrum with Dirichlet boundary con-
ditions (Proposition 4.5.3) while (ii) corresponds to the spectral gap for Neumann
boundary conditions (Proposition 4.5.4). Property (iii) corresponds to an estimate
on the spectral gap for periodic functions and is usually referred to as Wirtinger’s
inequality in the literature (see the Notes and References).
Proof The assertion (iii) is proved directly using a Fourier decomposition of a peri-
odic function similar to the Hermite expansion in Sect. 4.1. Namely if

f (x) = a0 + ak cos(2πkx) + bk sin(2πkx)
k≥1

(take if necessary a finite sum to start with), then assuming that a0 = [0,1] f dx = 0,

1 2
f dx =
2
ak + bk2 and f 2 dx = 2π 2 k 2 ak2 + bk2 .
[0,1] 2 [0,1]
k≥1 k≥1
The conclusion immediately follows. Note that any smooth function f on [0, 1]
such that f (0) = f (1) may be extended to a continuous 1-periodic function. While
it may happen that this extension is not C 1 , it may be approximated by a se-
quence of smooth periodic functions such that each term in the inequality con-
verges to the corresponding one for f . Therefore, the optimal (Poincaré) constant
in (4.5.5) is also the optimal constant in the Poincaré inequality on the interval [0, 1]
over the class of smooth 1-periodic functions. The second set (ii) of inequalities,
without any boundary condition, appears as a consequence of (i) by symmetriza-
tion and periodization (for f : [0, 1] → R arbitrary, define g : [−1, +1] → R by
g(x) = f (x) for x ∈ [0, +1], g(x) = f (−x) for x ∈ [−1, 0], and apply (i) to g on
the interval [−1, +1] after re-scaling). Finally (i) is a consequence of (iii) by anti-
symmetrization and periodization (for f : [0, 1] → R such that f (0) = f (1) = 0,
define g : [−1, +1] → R by g(x) = f (x) for x ∈ [0, +1], g(x) = −f (−x) for
x ∈ [−1, 0]).
4.6 The Lyapunov Function Method 201
4.6 The Lyapunov Function Method

This section investigates a method of establishing Poincaré inequalities close to the
method used for the exponential measure in Sect. 4.4.1. The form is however a bit
different.
4.6.1 The Lyapunov Function Method
The main principle is the following. In general, on Rn for example, it is not that
difficult to reach Poincaré inequalities on compact subsets, balls for example, by
comparison with known measures or gradients. Accordingly, we say that a Markov
Triple (E, μ, ), where μ is not necessarily a probability measure, with underlying
algebra A0 , satisfies a local Poincaré inequality on a measurable subset K with
0 < μ(K) < ∞ of the state space E if for some constant CK > 0 and every function
f in A0 ,

(f − mK ) dμ ≤ CK
2
(f )dμ (4.6.1)
K K

where mK = 1
μ(K) K f dμ. In other words, this is the Poincaré inequality with re-
spect to the carré du champ operator for functions restricted to K and with respect
to the measure μ restricted to K and normalized to be a probability measure. Actu-
ally, only a much weaker form of (4.6.1) will be used below, namely that for some
(measurable) set L ⊃ K and any f ∈ A0 ,

(f − mK )2 dμ ≤ CK,L (f )dμ (4.6.2)
K L
for some CK,L > 0.
Remark 4.6.1 On Rn equipped with the Lebesgue measure, or a manifold equipped

with a measure equivalent to the Lebesgue measure in any system of local coordi-
nates, whenever the set K is compact with smooth boundary, the local Poincaré in-
equality (4.6.1) on K is nothing else but the Poincaré inequality on K with Neumann
boundary conditions, as it is in dimension one (see Proposition 4.5.4). This may be
shown along the same lines, although with more technicalities. On the one hand, it
is always possible to approximate a smooth function f defined in a neighborhood
of K by a sequence (f )∈N of smooth
functions satisfying

the Neumann

bound-
ary conditions in such a way that K (f2 − μ(K) 1
K f dμ)2 dμ and
K (f )dμ
both converge to the corresponding quantities for f (this step is not so easy and
requires the use of normal coordinates in a neighborhood of the boundary). This
shows that the Poincaré inequality with Neumann boundary condition extends to
the inequality (4.6.1) valid for the restriction to K of smooth functions defined in a
neighborhood of K. On the other hand, it is quite easy to extend any smooth func-
tion defined on K and satisfying the Neumann boundary conditions into a smooth
function defined in a neighborhood of K, showing that the two inequalities are in

fact equivalent. If one is not concerned with optimal constants, it is easier to use the
simpler spectral inequality (4.10.1) in Sect. 4.10 below in a larger domain L ⊃ K,
for which the constant corresponds to the Dirichlet boundary conditions. This im-
mediately provides the weaker inequality (4.6.2).
We say that a function J : E → [1, ∞) (in the extended algebra A) is a Lyapunov

function (for the underlying generator L) if there are constants λ, b > 0 and a mea-
surable set K ⊂ E on which there is a local Poincaré inequality (4.6.2) (for some
L ⊃ K with constant CK,L ) such that J satisfies
LJ
1≤− + b 1K .
λJ
Theorem 4.6.2 (Lyapunov’s criterion) Let (E, μ, ) be a Markov Triple with μ a
probability measure satisfying the completeness property of Definition 3.3.9, p. 157.
If there exists a Lyapunov function J as above, then μ satisfies a Poincaré inequality
P (C) on the whole space E with constant C = λ1 + b CK,L .
Proof By density, it suffices to prove that for every f ∈ A0 ,

(f − m)2 dμ ≤ C (f )dμ
E E
for some m = m(f ) to be chosen later. Using of the Lyapunov function J , write

1 LJ
(f − m)2 dμ ≤ − (f − m)2 dμ + b (f − m)2 dμ. (4.6.3)
E λ E J K

Choose then m = μ(K) 1

K f dμ. The second term on the right-hand side of (4.6.3)

is then bounded from above thanks to (4.6.2) by b CK,L L (f )dμ. In order to deal
with the first term, we show that for every g ∈ A0 ,

LJ 2
− g dμ ≤ (g)dμ. (4.6.4)
E J E
To this end, integrate by parts to obtain

2
LJ 2 g
− g dμ = , J dμ. (4.6.5)
E J E J
Observe then that

g2 2g g2
,J = (g, J ) − 2 (J ) ≤ (g).
J J J
The first equality here follows from the diffusion property of the operator, while
the second is a direct consequence of the Cauchy-Schwarz inequality for , (1.4.3),
p. 21.
The proof is then completed provided (4.6.4) and (4.6.5) may be applied to
g = f − m, for which (g) = (f ). However, g is not in general in A0 even if
f is (recall that A0 stands in general for the set of smooth compactly supported
functions). To overcome this difficulty, use the completeness property and replace
g by g ζk , k ∈ N, where ζk are the functions used to define completeness and then
pass to the limit as k → ∞. In this way, Theorem 4.6.2 is established.
One may wonder whether the condition J ≥ 1 has indeed been used in the pre-
ceding proof. First, since the hypotheses in Theorem 4.6.2 are unchanged when J
is replaced by cJ , the preceding condition essentially amounts to the fact that J
is bounded from below by a strictly positive constant. This condition then ensures
that if we divide by J the resulting function g2J will be in A. Moreover, the com-
pleteness assumption is used only to justify the integration by parts formula (4.6.5).
This formula may

hold without it, for example as soon as L(log J ) and (log J ) are
in L1 (μ) and E L(log J )dμ = 0. Following Proposition 3.3.6, p. 156, it would be
enough for this purpose that L(log J ) ∈ L2 (μ).
The preceding Lyapunov criterion (Theorem 4.6.2) admits a number of vari-
ations, under somewhat different hypotheses and conclusions. Note for example
that it may be applied to the standard Gaussian measure on Rn with the choice of
J = 1 + |x|2 and K some Euclidean ball. The resulting Poincaré constant is how-
ever far from optimal (in particular it depends on the dimension). The criterion also
applies to the exponential measure with for example J = ec|x| for some sufficiently
small c > 0. Comparing with the previous (and optimal) proof of the Poincaré in-
equality for the exponential measure given in Proposition 4.4.1, setting J = ech , the
hypothesis on J may be deduced from a uniform bound on (h), an upper bound on
−Lh outside a compact set and a proper choice of the parameter c. The Lyapunov
hypothesis on J is then more general, although the proof given earlier, following
the same lines, was more precise.
4.6.2 Log-Concave Measures
A nice and powerful illustration of the Lyapunov function method is the following
statement for log-concave measures. A probability measure dμ = e−W dx on the
Borel sets of Rn such that W is a smooth convex function is called log-concave.
Theorem 4.6.3 (Kannan-Lovász-Simonovits-Bobkov Theorem) Let μ be a log-

concave probability measure on Rn . Then μ satisfies a Poincaré inequality with
respect the usual carré du champ operator (f ) = |∇f |2 .
Proof According to Theorem 4.6.2, it is enough to exhibit a suitable Lyapunov func-

tion. The idea is to choose c, R > 0 and J : Rn → R smooth with J (x) = ec|x| for
|x| ≥ R and J (x) ≥ 1 for |x| ≤ R. To verify that, for some c > 0 , J is a Lyapunov
function in

the preceding sense amounts to showing that for a convex function W
such that Rn e−W dx < ∞,
x · ∇W
lim inf > 0,
|x|→∞ |x|
which is left here as an exercise.
Note for further purposes that the dependence of the Poincaré constant on the
dimension n provided by the preceding proof is quite poor.
Theorem 4.6.3 may be applied to the uniform measure on a convex body in Rn .
Indeed let K be convex and compact with non-empty interior in Rn and consider the
convex function W (or rather a smooth approximation of it) defined by W (x) = 1 if
x ∈ K and W (x) = ∞ otherwise. The uniform normalized Lebesgue measure μK
on K thus satisfies a Poincaré inequality

VarμK (f ) ≤ C |∇f |2 dμK (4.6.6)
K
for some C > 0 and all smooth functions f : K → R. The classical Payne-
Weinberger Theorem, relying on a refined geometric analysis of the spectral de-
composition of the Laplace operator on K, describes the optimal estimate
D2
C≤ (4.6.7)
π2
in terms (only) of the diameter D of K. This inequality is optimal since it becomes
an equality when K is a ball, and includes in particular (4.5.4) in dimension one.
Even if (4.6.7) is optimal, the constant depends on the diameter and this estimate
cannot be used to prove a Poincaré inequality for general log-concave measures
(Theorem 4.6.3). Observe furthermore that (4.6.7) is false in general for non-convex
domains. A compact connected domain K in Rn still satisfies a Poincaré inequal-
ity (4.6.6) but the constant may depend on more than only the diameter (think for
example of a dumbbell).
4.6.3 A Glueing Property
According to the previous comments, the local Poincaré inequalities (4.6.1) on a

measurable subset K ⊂ E with 0 ≤ μ(K) < ∞ thus strongly depend on the shape
of the set K. In particular, the Poincaré constants CK do not have any kind of mono-
tonicity: a set K with small CK may contain a subset L with large CL . However,
they satisfy a useful glueing property illustrated by the following statement with
which we close this sub-section. Recall that μ is not necessarily a probability mea-
sure here.
Proposition 4.6.4 Let K and L be two measurable subsets of E such that

μ(K ∩ L) > 0 and μ(K ∪ L) < ∞. Then
μ(K ∪ L)
CK∪L ≤ max(CK , CL ).
μ(K ∩ L)
Here, the Poincaré inequalities (4.6.1) are assumed to hold on some suitable class
A1 of functions on E, not necessarily A0 . The proof relies on the following lemma,
which is of independent interest.
Lemma 4.6.5 Let K and L be two measurable subsets in a measure space

(E, F, μ) with μ(K ∩ L)
> 0 and μ(K ∪ L) < ∞.

Then, for any measurable func-
tion f : E → R such that K∪L f 2 dμ < ∞ and K∪L f dμ = 0,
2 2
1 1 μ(K ∩ L)
f dμ + f dμ ≤ 1− f 2 dμ.
μ(K) K μ(L) L μ(K ∪ L) K∪L
Proof Using homogeneity, we reduce to the case where μ(K ∪ L)

= 1 and
f 2 dμ = 1. Setting a = μ(K \ L), b = μ(L \ K), x =
K∪L

K\L f dμ,
y = L\K f dμ and z = K∩L f dμ, we have x + y + z = K∪L f dμ = 0 and, by
Cauchy-Schwarz,

x ≤a
2 2
f dμ, y ≤b
2 2
f dμ, z ≤ (1 − a − b)
2
f 2 dμ.
K\L L\K K∩L
Therefore, provided that a > 0 and b > 0,

x2 (x + y)2 y2 x2 z2 y2
+ + = + + ≤ f 2 dμ = 1.
a 1−a−b b a 1−a−b b K∪L
The left-hand side of the inequality to be established may be rewritten as
1 1 y2 x2
(x + z)2 + (y + z)2 = + .
1−b 1−a 1−b 1−a
y2 x2
After some tedious details, it appears that the supremum of 1−b + 1−a on the set
x2 (x+y)2 y2
a + 1−a−b + ≤ 1 is bounded from above by a + b = 1 − μ(K ∩ L) which yields
b
the claim. This bound is actually not optimal, but the optimal bound (given a and b)
has a less pleasant form. The argument is similar if either a = 0 or b = 0.
To complete the proof of Proposition 4.6.4, it is necessary to control the constant

CK∪L in the inequality
2
1
f dμ ≤ 2
f dμ + CK∪L (f )dμ
K∪L μ(K ∪ L) K∪L K∪L

for all functions f (in some class A1 ). To this end, assuming K∪L f dμ = 0, simply
write

f 2 dμ ≤ f 2 dμ + f 2 dμ
K∪L K L
and use the definition of CK and CL together with the lemma.
4.7 Local Poincaré Inequalities
In the framework of this chapter, this section deals with Poincaré inequalities not for
the invariant measure μ but for the heat kernel measures pt (x, dy), t ≥ 0, x ∈ E, of
the Markov semigroup P = (Pt )t≥0 (from the representation

Pt f (x) = f (y) pt (x, dy)
E
of (1.2.4), p. 12) with respect to the carré du champ operator . These inequalities
along the semigroup will be called local with reference to the (initial) point x and
the time variable t. (In particular, there is no need here to speak of an invariant or
reversible measure.) These are important illustrations of semigroup monotonicity
based on the Duhamel interpolation principle.
The inequalities investigated here are close in spirit to the local Poincaré inequal-
ities (4.6.1) where the measure restricted to a set K is replaced by the heat kernel
measure pt (x, dy) around a point x (highly concentrated around this point, at least
for small t’s), but play a different and fundamental role. Indeed, as will be shown
here, these inequalities are in general easier to reach, and provide a lot of important
information on the behavior of the semigroup. They are moreover a powerful tool
in the study of partial differential equations.
The local heat kernel inequalities will be presented and established for functions
in the algebra A0 of the underlying Markov Triple (E, μ, ). Depending on the
context, they may then be extended to any function in D(E) or D(L), or even to any
bounded measurable function. These extensions shall not be discussed below and
are left to the reader.
4.7.1 Poincaré Inequalities for Gaussian Kernels
To better introduce these ideas, we first start with the example of the Ornstein-
Uhlenbeck semigroup P = (Pt )t≥0 presented in Sect. 2.7.1, p. 103, whose generator,
on Rn , is given by Lf = f − x · ∇f with carré du champ operator (f ) = |∇f |2
(on an algebra A0 of smooth functions). In this simple and concrete example, both
the semigroup (Pt )t≥0 and the kernels pt (x, dy) are explicit. In particular, the rep-
resentation (2.7.3), p. 104, yields the commutation property ∇Pt f = e−t Pt (∇f )
4.7 Local Poincaré Inequalities 207
from which, for every t ≥ 0 and every f in A0 ,

(Pt f ) = |∇Pt f |2 ≤ e−2t Pt |∇f |2 = e−2t Pt (f ) (4.7.1)
(where we used the fact that, by convexity, (Pt h)2 ≤ Pt (h2 )).
Hence the gradient of Pt decays exponentially in t. This property has interest-
ing consequences via the Duhamel-type formula (3.1.21), p. 131. Let us recall the
principle. For fixed t > 0, and for a function f in A0 , consider

(s) = Ps (Pt−s f )2 , s ∈ [0, t]
(as usual at almost any point x ∈ E, omitted everywhere). Taking the derivative of
yields

(s) = Ps L (Pt−s f )2 − 2Pt−s f LPt−s f = 2Ps (Pt−s f )
as an immediate consequence of the basic chain rule

2 (g) = L g 2 − 2g Lg
applied to g = Pt−s f . Therefore, integrating the preceding in s ∈ [0, t],

t t
2

Pt f − (Pt f ) = (t) − (0) =
2
(s)ds = 2 Ps (Pt−s f ) ds.
0 0
On the basis of this interpolation formula, recall the commutation relation (4.7.1)

(Pt−s f ) ≤ e−2(t−s) Pt−s (f )
for every s ∈ [0, t]. Furthermore, by the semigroup property,

Ps (Pt−s f ) ≤ e−2(t−s) Pt (f ) .
As a consequence,

t
Pt f 2 − (Pt f )2 ≤ 2Pt (f ) e−2(t−s) ds = 1 − e−2t Pt (f ) . (4.7.2)
0
This inequality is a Poincaré inequality P (C), with constant C = 1 − e−2t , not for
the invariant (Gaussian) measure

μ, but for the heat kernel measure pt (x, dy) of the
representation Pt f (x) = Rn f (y) pt (x, dy), for any x ∈ E. However, as t → ∞,
we recover the Poincaré inequality of Proposition 4.1.1 since pt (x, dy) converges
to μ(dy). Recall that the initial point x is implicit in (4.7.2) (as is the case for all of
the local inequalities considered here).
The preceding argument via (4.7.1) may also be used to establish a converse
inequality. Indeed, by (4.7.1) again, for every s ≥ 0 and every g ∈ A0 ,

Ps (g) ≥ e2s (Ps g).
Therefore, with g = Pt−s f ,

Ps (Pt−s f ) ≥ e2s (Pt f ).
As a consequence this time,

t
Pt f 2 − (Pt f )2 ≥ 2 (Pt f ) e2s ds = e2t − 1 (Pt f ). (4.7.3)
0
The latter inequality, which may be considered as a reverse Poincaré inequality, is of

some particular interest. Indeed if for example f is bounded, then Pt f is Lipschitz
for every t > 0, with a Lipschitz norm improving as t → ∞, and of the order of
t −1/2 as t → 0. We know that if f ∈ L2 (μ), then Pt f belongs, for every t, to the
Dirichlet space D(E), but here more is available: whenever f is bounded, Pt f for
t > 0 has a bounded gradient with an explicit bound. The semigroup (Pt )t≥0 is thus
in some sense regularizing.
The preceding semigroup interpolation argument describes a heat flow mono-
tonicity proof of the Poincaré inequality for the Gaussian measure relying on the
commutation or gradient bound (4.7.1). This commutation has a curvature flavor
which is involved in geometric interpretations of the preceding heat kernel inequal-
ities. More precisely, the gradient bound (4.7.1) was already described in Chap. 3 as
being equivalent to the curvature condition CD(1, ∞) of Definition 3.3.14, p. 159
(cf. Corollary 3.3.19, p. 163). It is worth recalling the various steps in this con-
crete context of the Ornstein-Uhlenbeck operator. Namely, at t = 0, the two sides
of (4.7.1) are actually equal, and thus the asymptotics as t → 0 yields new informa-
tion of the form
L (f ) ≥ 2 (f, Lf ) + 2 (f ).
Recall now the 2 operator defined in (1.16.1), p. 71, by
1
2 (f ) = 2 (f, f ) = L (f ) − 2 (f, Lf ) .
2
The preceding inequality expresses that 2 (f ) ≥ (f ) for every f in A0 . In other
words, the generator L satisfies the curvature condition CD(1, ∞). In this Ornstein-
Uhlenbeck example, and as already presented in Sect. 2.7.1, p. 103, this curvature
condition may be checked directly, since for every sufficiently smooth function
f : Rn → R,
2 (f ) = |∇∇f |2 + |∇f |2 ≥ |∇f |2 = (f ).
In particular, the curvature condition CD(1, ∞) is independent of the dimension
n of the underlying state space Rn . The point here is that the commutation rela-
tion (4.7.1) is formally equivalent to the curvature condition CD(1, ∞).
The fact that the invariant measure μ is reversible does not play any particu-
lar role in the preceding developments, neither does the fact that it is a probability
measure. Integration by parts has not been used either. The calculus is only devel-
oped for the heat kernel measures pt (x, dy). In particular, the preceding applies
4.7 Local Poincaré Inequalities 209
in a completely similar way to the heat (Brownian) semigroup on Rn with the

classical Laplacian L = as infinitesimal generator (cf. Sect. 2.1, p. 78). In this
case Pt ∇ = ∇Pt , and hence, for every function f in the corresponding algebra A0
(smooth compactly supported functions for example), and every t ≥ 0,

(Pt f ) = |∇Pt f |2 ≤ Pt |∇f |2 = Pt (f ) .
As for (4.7.2), and its reverse form (4.7.3), we deduce that, for every t ≥ 0,

2t (Pt f ) ≤ Pt f 2 − (Pt f )2 ≤ 2t Pt (f ) . (4.7.4)
Differentiating at t = 0 yields here that 2 (f ) ≥ 0 (which is of course immedi-

ate since 2 (f ) = |∇∇f |2 ), expressing the curvature condition CD(0, ∞) of the
Brownian semigroup.
Remark 4.7.1 Since the kernels pt (x, dy) are Gaussian measures, the preceding
Poincaré inequalities (4.7.4) and (4.7.2) are just variations (with different means and
variances) of the Poincaré inequality for the standard Gaussian measure (Proposi-
tion 4.1.1) applied to functions y → f (x + ty). The approach is however completely
different, emphasizing interpolation along the semigroup, the commutation relations
and the connection with the 2 operator and curvature conditions.
4.7.2 Local Poincaré Inequalities
On the basis of the Brownian and Ornstein-Uhlenbeck semigroup examples, we next

present a general formulation of the preceding heat kernel inequalities following the
discussion on gradient bounds and curvature conditions in Corollary 3.3.19, p. 163
(in particular, the equivalence between (i) and (ii) below is part of the latter). Re-
call the curvature condition CD(ρ, ∞) expressing that 2 (f ) ≥ ρ (f ) for every
f ∈ A0 , or equivalently f ∈ A. All the local inequalities are stated for functions
in the algebra A0 , but actually extend to functions in A according to the extension
procedures described in Sect. 3.3, p. 151. Moreover, the reverse inequality (4.7.6)
(for t > 0) extends to bounded measurable functions.
Theorem 4.7.2 (Local Poincaré inequalities) Let (E, μ, ) be a Markov Triple

with semigroup P = (Pt )t≥0 . The following assertions are equivalent.
(i) The curvature condition CD(ρ, ∞) holds for some ρ ∈ R.
(ii) For every function f in A0 , and every t ≥ 0,

(Pt f ) ≤ e−2ρt Pt (f ) .
(iii) For every function f in A0 , and every t ≥ 0,

1 − e−2ρt
Pt f 2 − (Pt f )2 ≤ Pt (f ) . (4.7.5)
ρ
(iv) For every function f in A0 , and every t ≥ 0,
e2ρt − 1
Pt f 2 − (Pt f )2 ≥ (Pt f ). (4.7.6)
ρ
Further equivalent descriptions of the curvature condition CD(ρ, ∞) will be dis-

cussed throughout this monograph, in particular in the next chapter on logarithmic
−2ρt
and e ρ−1 have to be
2ρt
Sobolev inequalities (cf. Theorem 5.5.2, p. 259). Both 1−eρ
understood as 2t when ρ = 0.
Proof As mentioned before, the implication from (i) to (ii) is the content of the gra-
dient bound of Corollary 3.3.19, p. 163. The proof from the gradient bound (ii) to
the local Poincaré inequalities (iii) and (iv) has been discussed above for the Brow-
nian and Ornstein-Uhlenbeck semigroups and is similar here for any value of ρ. In
the same way, the local Poincaré inequality (iii) and its reverse form (iv) yield in
the limit as t → 0 the curvature condition CD(ρ, ∞). Indeed, using a second order
Taylor expansion
t2 2
Pt h = h + t Lh + L h + o t2
2
on (iii) for example, the left-hand side is given by
t2
t L f 2 + L2 f 2 − t 2 (Lf )2 − 2tf Lf − t 2 f L2 f + o t 2
2
while the right-hand side reads

2t (f ) − 2ρt 2 (f ) + 2t 2 L (f ) + o t 2 .
By definition of (f ), the first order terms vanish, yielding
1 2 2
L f − (Lf )2 − f L2 f ≤ −2ρ (f ) + 2 L (f ).
2
Developing further the chain rule, we obtain
2 (f, Lf ) + L (f ) ≤ −2ρ (f ) + 2 L (f )
which amounts to 2 (f ) ≥ ρ (f ) by definition of 2 (f ). The proof of the theorem

is complete.
Note that in general little is known about the heat kernel measures pt (x, dy),
t ≥ 0, x ∈ E. However, under a curvature condition, Theorem 4.7.2 ensures that
they satisfy Poincaré-type inequalities, and thus for example, by Proposition 4.4.2,
have exponential tails. On the other hand, as mentioned in the Ornstein-Uhlenbeck
4.8 Poincaré Inequalities Under a Curvature-Dimension Condition 211
example, the reverse Poincaré inequalities (4.7.6) of Theorem 4.7.2 provide a useful
quantitative regularization property in the form of
ρ
(Pt f ) ≤ (4.7.7)
e2ρt−1
for any bounded measurable function f such that |f | ≤ 1 and any t > 0 (cf. Proposi-
tion 8.6.1, p. 422, for a somewhat sharper bound). In particular, (Pt f ) = O(t −1/2 )
as t → 0.
Remark 4.7.3 It is worth mentioning that the implications from (ii) to (iii) and
(iv) of Theorem 4.7.2 work similarly if (ii) only holds up to some constant K ≥ 1,
that is, for every t ≥ 0 and all functions f in A0 ,

(Pt f ) ≤ K e−2ρt Pt (f ) .
The proof is entirely similar, and (iii), for example, then simply reads
K(1 − e−2ρt )
Pt f 2 − (Pt f )2 ≤ Pt (f ) .
ρ
This observation is of interest for some hypo-coercive examples as described in

Remark 3.3.22, p. 165. The point is that only K = 1 transfers back to the curvature
condition CD(ρ, ∞).
Remark 4.7.4 The preceding proof shows that each of the assertions of Theo-
rem 4.7.2 are equivalent to the statement that there exist a t0 > 0 and a function
c(t) = 2t − 2ρ t 2 + o(t 2 ) such that, for any t ∈ (0, t0 ) and any f in A0 ,

Pt f 2 − (Pt f )2 ≤ c(t)Pt (f )
(respectively

Pt f 2 − (Pt f )2 ≥ c(t) (Pt f ).
4.8 Poincaré Inequalities Under a Curvature-Dimension

Condition
The preceding section describes Poincaré inequalities under the curvature condition
CD(ρ, ∞) for the heat kernel measures pt (x, dy) of a Markov semigroup (Pt )t≥0 .
In the well behaved cases, typically when ρ > 0 (as for the Ornstein-Uhlenbeck
semigroup discussed above), these inequalities extend as t → ∞ to inequalities for
the invariant measure μ as a basic instance of ergodicity as emphasized earlier in
Sect. 1.8, p. 32. In particular, as discussed there, only the minimal connexity con-
dition (if (f ) = 0 then f is μ-almost everywhere constant) is required for this to
be the case. Moreover, it has been shown in Theorem 3.3.23, p. 165, that when-
ever ρ > 0, the invariant measure μ is always finite (and thus
a probability measure
by normalization) and ergodicity then ensures that P∞ f = E f dμ for every f in
L2 (μ). Under this assumption, the local Poincaré inequalities (4.7.5) for pt (x, dy)
of Theorem 4.7.2 extend as t → ∞ to a Poincaré inequality P ( ρ1 ) for the invariant
(probability) measure μ. The following statement summarizes this consequence. Al-
though it is formally contained in the stronger Theorem 4.8.4 below (corresponding
to n = ∞), it is worth stating independently.
Proposition 4.8.1 (Poincaré inequality under CD(ρ, ∞)) Under the curvature
condition CD(ρ, ∞) with ρ > 0, the Markov Triple (E, μ, ) satisfies a Poincaré
inequality P (C) with constant C = ρ1 . That is, for every function f ∈ D(E),
1
Varμ (f ) ≤ E(f ).
ρ
This proposition covers the example of the standard Gaussian measure on

Rn for which ρ = 1 thus yielding the optimal Poincaré inequality of Proposi-
tion 4.1.1. It actually covers more generally probability measures dμ = e−W dx on
Rn with respect to the standard carré du champ operator where W is a smooth
potential such that ∇∇W ≥ ρ Id (as symmetric matrices) for some ρ > 0 (equiv-
2
alently, W (x) − ρ|x|2 is convex). Indeed, according to Sect. 1.11.3, p. 46, the
Markov diffusion operator with invariant and symmetric measure μ is of the form
Lf = f − ∇W · ∇f with carré du champ operator (f ) = |∇f |2 on smooth
functions f . As a consequence of (1.16.5), p. 72, or an immediate verification in
this simple case, on every smooth function f : Rn → R,
2 (f ) = |∇∇f |2 + ∇∇W (∇f, ∇f )
where ∇∇W is the Hessian of W . By the convexity assumption,
∇∇W (∇f, ∇f ) ≥ ρ |∇f |2 .
In other words, the Markov Triple consisting of Rn with the measure dμ = e−W dx
for which ∇∇W ≥ ρ Id and the standard carré du champ operator therefore satisfies
the curvature condition CD(ρ, ∞). We present this conclusion as an independent
statement. It will also appear as a consequence of the stronger Brascamp-Lieb in-
equality below (Theorem 4.9.1).
Corollary 4.8.2 Let dμ = e−W dx be a probability measure on the Borel sets of Rn

where W : Rn → R is a smooth potential such that ∇∇W ≥ ρ Id for some ρ > 0.
Then μ satisfies the Poincaré inequality P ( ρ1 ).
A similar statement, with the same proof, holds on a weighted Riemannian man-
ifold (M, g) for the operator L = g − ∇W · ∇ with the reversible (probability)
4.8 Poincaré Inequalities Under a Curvature-Dimension Condition 213
measure μ having density e−W with respect to the Riemannian measure under the
curvature condition Ric(L) = Ricg +∇∇W ≥ ρ g with ρ > 0 (cf. Sect. 1.16, p. 70,
and Sect. C.6, p. 513).
The main purpose of this section is to investigate Poincaré inequalities for
a Markov Triple (E, μ, ) under the stronger curvature-dimension condition
CD(ρ, n) of Definition 3.3.14, p. 159, for some ρ > 0 and some finite dimension
n ≥ 1. Recall that this condition expresses that
1
2 (f ) ≥ ρ (f ) + (Lf )2
n
for all f ∈ A0 (or A). In particular, it will be observed how the (finite) dimension
will improve upon the preceding Poincaré constant in Proposition 4.8.1.
To this end, we start from the CD(ρ, n) condition for ρ > 0 and n > 1 integrated
with respect to μ to get that, for every f in A0 ,

1
2 (f )dμ ≥ ρ (f )dμ + (Lf )2 dμ.
E E n E
Now, by integration by parts and the construction of the operators and 2
(cf. Sect. 3.3, p. 151),

(f )dμ = f (−Lf )dμ and 2 (f )dμ = (Lf )2 dμ,
E E E E
so that the preceding inequality reads

ρn
(Lf )2 dμ ≥ f (−Lf )dμ.
E n−1 E
As a first intuitive approach, apply this inequality to a (non-zero) eigenfunction
f of −L with (positive) eigenvalue λ to see that

ρn
λ λ− ≥ 0.
n−1
ρn
Hence, if λ = 0, λ ≥ n−1 . By the correspondence between spectral gap and Poincaré
inequality, this bound improves upon Proposition 4.8.1 (note that the two Poincaré
constants agree when n = ∞).
As a consequence of the preceding, the first non-trivial eigenvalue λ1 of the
Laplace operator on an n-dimensional (compact) Riemannian manifold with Ricci
curvature bounded from below by ρ > 0 satisfies
ρn
λ1 ≥ , (4.8.1)
n−1
which is the essence of Lichnerowicz’ eigenvalue comparison Theorem. For the
Laplacian on the sphere Sn in Rn+1 (cf. Sect. 2.2, p. 81), or for the symmetric Ja-
cobi operator with dimension n (cf. Sect. 2.7.4, p. 113), the preceding lower bound
ρn
n−1 of the spectral gap is optimal since these two models both satisfy a curvature-
dimension condition CD(n − 1, n), yielding a Poincaré constant equal to n1 which
corresponds to the inverse of the first non-trivial eigenvalue. In particular, Lich-
nerowicz’s Theorem (4.8.1) compares the first non-trivial eigenvalue of the Lapla-
cian on a (compact) n-Riemannian manifold having a strictly positive lower bound
on the Ricci curvature by the one of the sphere with the same dimension and cur-
vature. Obata’s Theorem expresses conversely that when there is equality in (4.8.1),
the manifold is isometric to this sphere. Note that the result also holds on S1 (n = 1)
by Proposition 4.5.5 via a simple dilation from [0, 1] to [0, 2π]. But this may ap-
pear as a simple coincidence, since the condition CD(0, m), m ≥ 1, does not imply
a spectral gap inequality in general.
The preceding spectral argument implicitly assumes that the spectrum of L is dis-
crete. (We will actually establish later in Corollary 6.8.2, p. 306, that this is indeed
the case under these hypotheses.) We next provide a less formal approach relying
on semigroup tools and heat flow monotonicity. Along these lines, we start with a
useful dual description of the Poincaré inequality (also implicit from the previous
argument) which develops the heat flow argument via the second derivative of the
variance along the semigroup, or equivalently the first derivative of the Dirichlet
form, giving rise to the 2 operator.
Proposition 4.8.3 A Markov Triple (E, μ, ) satisfies a Poincaré inequality P (C)

for some C > 0 if and only if

E(f ) = (f )dμ ≤ C (Lf )2 dμ (4.8.2)
E E
for every function f in the domain D(L).
Proof
As already used above, for a given f ∈ A0 , consider the function (t) =

E (Pt f ) dμ, t ≥ 0, for which
2

(t) = −2 (Pt f )dμ and (t) = 4 (LPt f )2 dμ.
E E
Under (4.8.2), (t) ≥ − C2 (t), t ≥ 0, which integrates into the Poincaré inequal-
ity P (C) since
∞
C ∞ C
Varμ (f ) = − (t)dt ≤ (t)dt = − (0) = C (f )dμ.
0 2 0 2 E
Conversely, for every f ∈ D(L) with mean zero as it can be assumed,

1/2
(f )dμ = f (−Lf )dμ ≤ Varμ (f )1/2 (Lf )2 dμ
E E E
by the Cauchy-Schwarz inequality, and the conclusion follows from the application
of P (C).
4.9 Brascamp-Lieb Inequalities 215
Proposition 4.8.3 is actually a purely spectral statement only relying on ergodic-

ity. Under the ESA property of Definition 3.1.10, p. 134, (4.8.2) for every f ∈ A0
extends immediately to any f ∈ D(L). In this case, it is equivalent to

E(f ) ≤ C 2 (f )dμ (4.8.3)
E
which is valid for any f ∈ A0 . Moreover, under a curvature condition CD(ρ, ∞)

for some ρ ∈ R, the 2 operator and (4.8.3) extend to D(L).
The following main statement under the curvature-dimension condition
CD(ρ, n), which includes in particular the Lichnerowicz bound (4.8.1), follows.
Theorem 4.8.4 (Poincaré inequality under CD(ρ, n)) Under the curvature-dimen-
sion condition CD(ρ, n), ρ > 0, n > 1, the Markov Triple (E, μ, ) satisfies a
Poincaré inequality P (C) with constant C = n−1
ρn .
Proof Integration of the CD(ρ, n) condition yields as before

1
2 (f )dμ ≥ ρ (f )dμ + (Lf )2 dμ
E E n E

for any f in A0 . Since E (Lf )2 dμ = E 2 (f )dμ, (4.8.2) of Proposition 4.8.3

holds with C = n−1
ρ n . The theorem is established.
4.9 Brascamp-Lieb Inequalities
In this section, we deal with a stronger variant of Poincaré inequalities, known as

Brascamp-Lieb-type inequalities.
4.9.1 The Matrix Brascamp-Lieb Inequality
We first state and prove the classical version in Rn .
Theorem 4.9.1 (Brascamp-Lieb inequality) For a probability measure

dμ = e−W dx on Rn for which the smooth potential W : Rn → R is strictly con-
vex,

Varμ (f ) ≤ (∇∇W )−1 (∇f, ∇f )dμ
Rn
for every smooth compactly supported function f on Rn , where (∇∇W )−1 is the
inverse of the Hessian of W .
In this Euclidean context, for the operator L = − ∇W · ∇ with (f ) = |∇f |2 ,

the curvature condition CD(ρ, ∞) expresses that ∇∇W ≥ ρ Id in the sense of sym-
metric matrices. Hence, if ρ > 0,
1
(∇∇W )−1 (∇f, ∇f ) ≤ |∇f |2
ρ
so that the Brascamp-Lieb inequality improves upon Corollary 4.8.2 in this context.
To address the proof of Theorem 4.9.1, the following general lemma of indepen-
dent interest will be useful.
Lemma 4.9.2 Let A1 and A2 be symmetric operators defined on a common dense

domain D0 in a Hilbert space H. Assume that for some ε > 0 and any u ∈ D0 ,
εu, u ≤ A1 u, u ≤ A2 u, u.
Consider then the self-adjoint Friedrichs extensions of A1 and A2 (denoted in the

same way) with respective domains D(A1 ) and D(A2 ). The inverse operators A−1 1
and A−1
2 , defined by spectral decomposition, are bounded with norms bounded from
above by 1ε . Then, for any u ∈ H,
−1
A−1
2 u, u ≤ A1 u, u . (4.9.1)
Proof It is enough to prove (4.9.1) for u = A2 v with v ∈ D(A2 ). Then the in-
equality boils down to v, A2 v ≤ A−1
1 A2 v, A2 v. Now, it is easily checked that
D(A2 ) ⊂ D(A1 ), and setting B = A2 − A1 , the latter then reads A−1
1 Bv, A2 v ≥ 0.
In other words,
−1
A1 Bv, Bv + Bv, v ≥ 0
which is immediate since A1 and B are positive.
Proof of Theorem 4.9.1 The proof is only sketched. It uses in particular the ex-
tension of the operator L = − ∇W · ∇ and its associated semigroup (Pt )t≥0 to
vector-valued functions, and everything will ultimately rely on the commutation
property, easily checked on smooth functions f ,
∇Lf = (L − ∇∇W )∇f (4.9.2)
(where ∇f = (∂i f )1≤i≤n ).

To give a better sense to the latter, consider the Hilbert space H of vectors
u = (f1 , . . . , fn ) of functions in L2 (μ) with norm

n
u =
2
fi2 dμ.
Rn i=1
Denote by u, v the scalar product in H. For functions fi , i = 1, . . . , n, smooth and

compactly supported on Rn , introduce the operator

L(f1 , . . . , fn ) = (Lf1 , . . . , Lfn ).
The operator
L is symmetric in H, and for u = (u1 , . . . , un ) and v = (v1 , . . . , vn ) in
H,
n

−Lu, v = ∇u · ∇v dμ = ∇ui · ∇vi dμ
Rn Rn i=1
so that −L is positive.

Denote by M = M(x) multiplication by the matrix ∇∇W = ∇∇W (x) on el-
ements of H. The aim is to apply Lemma 4.9.2 with A1 = M and A2 = M − L,
or rather A1 = ε + M and A2 = ε + M − L for some ε > 0. The operators M
and M − L are symmetric on H. Since ∇∇W is strictly positive-definite, M is a
positive operator. In particular, M −
L is also positive. Therefore, by Lemma 4.9.2,
−1
ε + M −
L u, u ≤ (ε + M)−1 u, u (4.9.3)
for every u ∈ H.
Now let f ∈ A0 , where A0 is the set of smooth compactly supported functions
on Rn . Set
∞
gε = (ε − L)−1/2 f = c e−εs s −1/2 Ps f ds
0
for which

f 2 dμ = ε gε2 dμ + |∇gε |2 dμ.
Rn Rn Rn
(Here, the operator ∇ has to be properly extended from A0 to D(E) as an oper-
ator with values in H, but this is immediate.) Moreover, for f ∈ A0 , the identity
∇Lf = ( L − M)∇ from (4.9.2) extends to
1/2
∇gε = ∇ (ε − L)−1/2 f = ε + M −
L ∇f.
Indeed, if (Pt )t≥0 is the contraction semigroup on H with generator L − M, the

latter relies on the commutation P t ∇f = ∇Pt f , which is valid for f ∈ A0 and
from the fact that ∇Pt f ∈ D( L − M), which is easily checked from the definition
of D(M − L) and the techniques developed in Chap. 3. As a consequence, together
with (4.9.3),

−1
f 2 dμ = ε gε2 dμ + ε + M − L ∇f, ∇f
Rn Rn

≤ε gε2 dμ + (ε + M)−1 ∇f, ∇f .
Rn

It remains to let ε go to 0, with the further observation that whenever Rn f dμ = 0,

then limε→0 εgε 22 = 0. Indeed, lims→∞ Ps f 2 = 0 and
∞ ∞
1/2
ε g ε = c e −s −1/2
s P f ds ≤c e−s s −1/2 Ps/ε f 2 ds
2 s/ε
0 2 0
which tends to 0 with ε by dominated convergence. The proof is complete.
The Brascamp-Lieb inequality (Theorem 4.9.1) admits a simple extension to el-

liptic operators on a manifold (M, g) which we briefly describe now. Recall from
Sect. 1.11.3, p. 46, that in the elliptic case, L = g − ∇W · ∇, where the reversible
(probability) measure μ has density e−W with respect to the Riemannian measure.
Recall also (Sect. 1.16, p. 70) that in this case
2 (f ) = |∇∇f |2 + Ric(L)(∇f, ∇f ),
with Ric(L) = Ricg +∇∇W . The next statement is then the analogue of the
Brascamp-Lieb inequality in this context.
Theorem 4.9.3 In this Riemannian setting, if the tensor Ric(L) is strictly positive
everywhere, denoting by Ric(L)−1 its inverse tensor, then for every smooth com-
pactly supported function f on M,

Varμ (f ) ≤ Ric(L)−1 (∇f, ∇f ) dμ.
M
The proof is rather similar to the Euclidean case of Theorem 4.9.1, replacing all
vectors u by 1-forms w and the operator L by the operator
L given in a local system
of coordinates (and with Einstein’s summation notation) by

L wj = ∇ i ∇i wj − ∇ i (W ) · ∇i wj .
The same commutation formula holds

d Lf = L − Ric(L) df,
and the argument is then completed along the same lines.
4.9.2 Probabilistic Representations
A consequence of the Brascamp-Lieb inequality of Theorem 4.9.1, and of Theo-

rem 4.9.3 in the manifold setting M, is that if ρ(x) is the least eigenvalue, at the
point x, of the matrix ∇∇W (x) (in case of Theorem 4.9.1), or Ric(L) (in case of
Theorem 4.9.3), then, for every sufficiently smooth function f : M → R,

1
Varμ (f ) ≤ |∇f |2 dμ. (4.9.4)
M ρ(x)
This inequality may then be related to Proposition 3.2.5, p. 149, and the
semigroup
t )t≥0 with generator ∞
(P L = L−ρ Id. Indeed, by the development f = − 0 LPt f dt
of a smooth mean-zero function f and integration by parts,
∞
Varμ (f ) = ∇f · ∇Pt f dμ dt.
0 M
Now, by means of the gradient bound

t |∇f |
|∇Pt f | ≤ P (4.9.5)
of Proposition 3.2.5, p. 149, it holds that

∞

Varμ (f ) ≤
|∇f |Pt |∇f | dt dμ = |∇f |(−
L)−1 |∇f | dμ
M 0 M
from which (4.9.4) follows by an analysis similar to the one developed for Theo-
rem 4.9.1.
The gradient commutation bound (4.9.5) actually has an even more precise form
of a probabilistic nature, which may prove useful in a wide variety of contexts,
although it is not so easy to express. Namely, starting from the commutation for-
mula (4.9.2), the semigroup (Qt )t≥0 with generator
L − M acting on vector-valued
functions (or on 1-forms in the manifold case) can be rewritten as the (exact) for-
mula
∇Pt f = Qt ∇f, t ≥0
(on suitable functions f : M → R). Therefore, (4.9.5) amounts to |Qt u| ≤ P t (|u|)
for vector-valued functions u. A direct approach to this proceeds via the Feynman-
Kac-type probabilistic representations as described in Sect. 1.15.6, p. 63. Let us
briefly illustrate the principle in the flat Euclidean case M = Rn with
L = − ∇W · ∇, corresponding to Theorem 4.9.1, borrowing the notation from
its statement and proof.
Indeed, the semigroup (Qt )t≥0 admits a nice probabilistic interpretation in terms
of the Markov process {Xtx ; t ≥ 0, x ∈ Rn } with semigroup (Pt )t≥0 and generator L.
Consider its vector-valued extension (with generator L) given on a vector-valued
function u = (f1 , . . . , fn ) by

Pt u(x) = Ex u(Xt ) = Ex f1 (Xt ) , . . . , Ex fn (Xt ) , t ≥ 0, x ∈ Rn .
The vector-valued version of the Feynman-Kac formula of Sect. 1.15.6, p. 63, then
takes the form

Qt u(x) = Ex At u(Xt ) (4.9.6)
where now At , t ≥ 0, is a (random) matrix, being a solution of the ordinary differ-
ential equation

t ∂t At = −At M(Xt ), with A0 = Id. (Note that this is not in general
exp(− 0 M(Xs ) ds) since the matrices M(Xs ) do not commute, and the latter ex-
pression is not even symmetric even if the matrices M(Xs ) are.) By a standard
lemma from ordinary differential equation theory, if for any s ≥ 0, M(Xs )u, u ≥
ρ(Xs )|u|2 , then the matrix At satisfies

t
|At u| ≤ e− 0 ρ(Xs )ds
|u|.
As a consequence, the announced inequality |Qt u| ≤ P t (|u|) holds.

In the manifold setting (that is, as soon as the second order terms in L are
no longer constant), the preceding construction requires additional tools such as
stochastic parallel transport in order to give a reasonable meaning to the various
expressions. Indeed, the matrix operator Ric(L)(Xs ) acts naturally on the tangent
space above Xs , and one has to carry back these operations on the same vector
space (in general the tangent space above X0 ), causing extra trouble (and serious
technicalities) which will not be developed here.
4.10 Further Spectral Inequalities
In the last section of this chapter, we investigate some further spectral properties re-
lated to symmetric Markov semigroups. Essentially, Poincaré inequalities concern
Markov semigroups which are symmetric with respect to a given probability mea-
sure μ. In this context, the constant function 1 is always an eigenvector associated
with the 0 eigenvalue. Hence, Poincaré inequalities explore the gap between 0 and
the rest of the spectrum, and more precisely describe when the spectrum of −L is in-
cluded in {0} ∪ [λ1 , ∞) for some λ1 > 0. But it may happen that the spectrum itself
is included in [λ0 , ∞) for some λ0 > 0. This is in general the case for sub-Markov
semigroups, such as those related to Dirichlet boundary conditions on compact sets,
but it may also happen when the measure μ is infinite. These questions are addressed
in Sect. 4.10.1.
Furthermore, it is often of great interest to know when the spectrum is discrete,
that is when the essential spectrum is empty. Recall (cf. Appendix A) that this prop-
erty conceals several different facts. First that the spectrum is purely punctual, and
that to every point in the spectrum corresponds at least one non-zero eigenfunc-
tion (which is in L2 (μ) by definition). Second, that every point λ in the spec-
trum is isolated, that is, there exists an ε > 0 such that no point of the spectrum
lies in (λ − ε, λ + ε). Moreover, for any point in the spectrum, the correspond-
ing eigenspace is finite-dimensional. Section 4.10.2 investigates in this regard the
bottom of the essential spectrum, providing simple criteria related to Nash-type in-
equalities (further developed in Chap. 7).
In many situations, the essential spectrum depends only on “what happens at
infinity”. To give a precise meaning to this statement, we describe in Sect. 4.10.3 the
notion of a Persson operator, which leads once again to useful criteria for emptiness
of the essential spectrum.
We refer to Appendix A, and more precisely to Sect. A.4, p. 478, for the vari-
ous notions and definitions of the spectrum of a positive self-adjoint operator. The
4.10 Further Spectral Inequalities 221
first two sub-sections below are concerned with the minimal Markov triple struc-
ture (E, μ, ) as introduced in the first part of Sect. 3.1, p. 120, and focus on the
self-adjoint operator L associated with the triple.
4.10.1 Bottom of the Spectrum
As already mentioned, Poincaré inequalities make sense only when the refer-
ence measure μ is finite (in fact a probability measure), and when the semigroup
P = (Pt )t≥0 is Markov. It could nevertheless be the case, either for infinite mea-
sures (or for sub-Markov semigroups), that the following inequality

f 2 dμ ≤ C E(f ) (4.10.1)
E
holds for some C > 0 and some suitable class functions f on E. One such example
has already appeared in Lemma 4.2.3 above as a consequence of a Poincaré inequal-
ity on a larger set. When such an inequality holds for a class of functions f which
is dense in the Dirichlet domain D(E) of the given Dirichlet form E (and therefore
holds for every function f in it), it amounts to saying that the spectrum of the un-
derlying (opposite) generator −L lies in the interval [ C1 , ∞). Here, L is understood
as the minimal extension of the operator defined from the Dirichlet form E, that is,
its Friedrichs extension as described in Sect. 3.1, p. 120.
In the case of a bounded domain (in Rn or a manifold), this extension corre-
sponds to the Dirichlet boundary conditions. The connection between the spectrum
and (4.10.1) has been illustrated for Sturm-Liouville operators on an interval of the
real line in Sect. 4.5.2, giving rise to a useful connection between eigenvalues with
Dirichlet or Neumann boundary conditions. This section provides further material
on the spectrum of a diffusion operator.
From a spectral analysis point of view, the best constant C in (4.10.1) is λ10 where
λ0 is the bottom of the spectrum of −L (which is different from the spectral gap
related to Poincaré inequalities). This observation applies only when 1 is no longer
an eigenvector for −L which may indeed happen in a wide variety of situations:
either the measure μ is infinite, or the boundary conditions imposed on the domain
D(L) prevent 1 from belonging to it, in particular in the case of Dirichlet boundary
conditions.
Remark 4.10.1 For an elliptic operator L on an open domain K ⊂ Rn with regular

boundary, the natural probabilistic interpretation of the constant λ0 for the operator
L with Dirichlet boundary conditions in K may be described in terms of the expec-
tation of the hitting time of the boundary. To fix the ideas, assume that L is the usual
Laplace operator in some open relatively compact set O ⊂ Rn , and consider the
associated semigroup (PtD )t≥0 with Dirichlet boundary conditions on O. If (Xtx )t≥0
is Brownian motion starting from x ∈ O, and if
T = inf {t ≥ 0 ; Xt ∈
/ O},
then for all λ < λ0 , Ex (eλT ) < ∞ while for all λ ≥ λ0 , Ex (eλT ) = ∞. To understand
this result, the basic observation is that the bottom λ0 of the spectrum of − corre-
sponds to some strictly positive eigenvector f0 which vanishes at the boundary (this
strictly positive eigenvector corresponds to the ground state for Schrödinger oper-
ators and to the Perron-Frobenius eigenvector in the case of finite Markov chains).
Then eλ0 t f0 (Xt ), t ≥ 0, defines a martingale, and since f0 vanishes at the bound-
ary, the finiteness of Ex (eλ0 T ) would yield a contradiction. Hence Ex (eλT ) = ∞
for every λ ≥ λ0 . On the other hand, for any λ < λ0 , there exists a strictly posi-
tive eigenvector solution of −Lf = λf (not vanishing identically at the boundary),
which may be thought of as the bottom eigenvector on some enlarged set. The same
reasoning with the martingale eλt f (Xt ), t ≥ 0, leads to the fact that Ex (eλT ) must
be finite. This type of argument may be extended to more general settings, replacing
the Brownian motion by the Markov process with generator L.
One of the most important geometric examples of an inequality such as (4.10.1)

is given by the hyperbolic Laplace operator discussed in Sect. 2.3, p. 88. That is, the
state space E is the hyperbolic space Hn (n ≥ 2) which may be represented as the
upper half-space Rn−1 × (0, ∞) with reference measure dμ(x) = xn−n dx and carré
du champ operator (f ) = xn2 |∇f |2 . Here xn stands for the last coordinate (the one
which is strictly positive), dx for the standard Lebesgue measure and |∇f |2 for the
standard Euclidean carré du champ operator

of a smooth function f . Accordingly,
the Dirichlet form is given by E(f ) = Hn (f )dμ.
Proposition 4.10.2 For the hyperbolic space Hn of dimension n, and for any
smooth compactly supported function f on Hn ,

4
f 2 dμ ≤ E(f ).
Hn (n − 1)2
Proof The argument is similar to the proof of the Poincaré inequality for the ex-
ponential measure (Sect. 4.4.1). In the chosen upper half-space representation, the
hyperbolic Laplacian may be explicitly represented as
Lf = xn2 f − (n − 2)xn ∂n f.
Now, the coordinate function xn satisfies L(xn ) = −(n − 2)xn and (xn ) = xn2 .
Therefore, if f is a smooth and compactly supported function on Hn ,

1 −1 2 1
f dμ =
2
(−Lxn ) xn f dμ = xn , xn−1 f 2 dμ.
H n n−2 H n n−2 H n
Now,

xn , xn−1 f 2 = −f 2 xn−2 (xn ) + 2f xn−1 (xn , f )
= −f 2 + 2f xn−1 (xn , f )
≤ −f 2 + 2|f | (f )1/2
since (xn ) = xn2 . It follows that

1 2
1+ f dμ ≤
2
|f | (f )1/2 dμ
n − 2 Hn n − 2 Hn
1/2
2
≤ 2
f dμ E(f )1/2
n − 2 Hn
where we used the Cauchy-Schwarz inequality in the last step. The conclusion im-
mediately follows and the proposition is therefore established. One may check that
4
the constant (n−1) 2 is indeed optimal in this inequality.
4.10.2 Bottom of the Essential Spectrum
Here we return to the spectral information contained in a Poincaré inequality.

As pointed out earlier, the Poincaré inequality P (C) with constant C > 0 for a
Markov operator L indicates that the spectrum of (the opposite generator) −L lies
in {0} ∪ [ C1 , ∞). Now, recall (Sect. A.4, p. 478) that the spectrum of a positive self-
adjoint operator such as −L lies in R+ and may be divided into two parts, the dis-
crete spectrum, which corresponds to isolated eigenvalues with finite-dimensional
eigenspaces, and the rest which is called the essential spectrum and is denoted by
σess (−L). The spectrum is said to be discrete if it essential spectrum is empty (Def-
inition A.4.3, p. 480).
The following statement is a first description of some useful information on the
essential spectrum σess (−L) in terms of related families of spectral inequalities.
Proposition 4.10.3 Assume that σess (−L) ⊂ [σ0 , ∞) for some σ0 > 0. Then, for
every r > σ10 , there exists a finite-dimensional vector subspace F ⊂ L2 (μ) such
that, for any f ∈ D(E),

f dμ ≤ r E(f ) + (F f )2 dμ
2
(4.10.2)
E E
where F denotes the orthogonal projection onto F . Conversely, if there exists a

finite-dimensional vector subspace F ⊂ L2 (μ) such that (4.10.2) holds for some
r > 0 and all f ∈ D(E), then the essential spectrum σess (−L) lies in [ 1r , ∞).
1
Proof The first part is straightforward. Indeed, given r > σ0 , let F be the di-
rect sum of all the eigenspaces corresponding to eigenvalues λ < 1r . By defini-
tion of σess (−L), there is only a finite number of such eigenvalues and the sub-
space F is finite-dimensional. Let f ∈ D(E), and set for simplicity g = F (f ),
h = f − F (f ). Clearly E(g, h)
= 0 and E(f ) = E(g) + E(h) ≥ E(h). On the other
hand, by construction, E(h) ≥ 1r E h2 dμ. It follows that

f dμ =
2
g dμ + h dμ ≤
2 2
g dμ + rE(h) ≤
2
g 2 dμ + r E(f )
E E E E E
which is (4.10.2). Conversely, under (4.10.2) for some r > 0 and some subspace F ,
assume that there exists a λ ∈ σess (−L) such that λ < 1r . Recall from Weyl’s criterion
(Theorem A.4.4, p. 481) that for any ε > 0, there exists an orthonormal sequence
(fk )k∈N such that λfk + Lfk 2 ≤ ε for
every k. Since fk has norm 1 in L (μ),
2
by the Cauchy-Schwarz inequality − E (λfk + Lfk )fk dμ ≤ ε which amounts to

E(fk ) ≤ λ + ε, k ∈ N. Choose then ε such that r(λ + ε) < 1. Applying (4.10.2) to
fk for every k yields
1 ≤ r(λ + ε) + E fk 22 .
But an orthonormal sequence in a Hilbert space is always weakly converging to 0,
so that its projection on any finite-dimensional subspace converges (strongly) to 0.
Therefore E fk 2 → 0 which produces a contradiction. The proof of Proposi-
tion 4.10.3 is complete.
A more precise look at the preceding result yields an interesting consequence of

the fact that σess (−L) = ∅.
Proposition 4.10.4 If the spectrum of −L is discrete (i.e. σess (−L) = ∅), there exist
a positive function w ∈ L2 (μ) and a map β : (0, ∞) → (0, ∞) such that, for any
r ∈ (0, ∞) and any f ∈ D(E),
2
f 2 dμ ≤ r E(f ) + β(r) |f |w dμ . (4.10.3)
E E
Equivalently, there exists a positive function w ∈ L2 (μ) and an increasing concave

(r)
function : (0, ∞) →
(0, ∞) satisfying limr→∞ r = 0 such that for any func-
tion f in D(E) with E |f |wdμ = 1,

f 2 dμ ≤ E(f ) . (4.10.4)
E
Converses of this proposition will be studied below when the semigroup with
Markov generator L has a density kernel (Theorem 4.10.5) or for Persson operators
(Theorem 4.10.8).
Proof For any r > 0, denote by Fr the vector space spanned by all the eigenvectors
with eigenvalues less than 1r and by n(r) its dimension. Let (fk )k∈N be a sequence
of eigenvectors such that, for any r, (f0 , . . . , fn(r) ) defines an orthonormal basis of
r . Let (αk )k∈N be a decreasing
F sequence of strictly positive real numbers such that
α
k∈N k < ∞. Set w = k∈N k |fk |. The function
α
w belongs to L (μ) and, for
2
any 0 ≤ k ≤ n(r) and any f ∈ L (μ), setting ck = E ffk dμ, k ∈ N,

2

1
|ck | ≤ |f |w dμ.
αn(r) E
n(r) −2
Since Fr (f )22 = 2
k=0 ck , it follows that Er (f )22 ≤ n(r)αn(r) . It then suf-
−2
fices to choose β(r) = n(r)αn(r) , r > 0, and to follow the proof of Proposi-
tion 4.10.3 above. Inequality (4.10.4) is obtained with (r) = infs>0 [rs + β(s)],
which is clearly increasing and concave and satisfies limr→∞ (r)
r = 0. Conversely,
if (4.10.4) holds for some , it may be assumed to
be C 1 , therefore satisfying
limr→∞ (r) = 0 (namely, replace by (r) = 2 r (s)ds which is still in-
r 0
1 (r)). The version (4.10.3)
creasing and concave and satisfies 2 (r) ≤ (r) ≤
then follows from the family of inequalities (concavity)
(s) ≤ (s0 )s + (s0 ) − s0 (s0 ), s > 0,
setting r = (s0 ). The proposition is established.
The family of inequalities described in Proposition 4.10.4 will be further investi-

gated in Chap. 7 (Sect. 7.4.3, p. 370) under the name of generalized weighted Nash
inequalities associated to the growth functions β or .
As announced, we next examine some converses to Proposition 4.10.4. It is not
true that the spectrum of −L is always discrete under a generalized weighted Nash
inequality (4.10.3) or (4.10.4). For example, any infinite product of Gaussian mea-
sures provides a case where a logarithmic Sobolev inequality holds (see Chap. 5),
and thus a generalized weighted Nash inequality, whereas the eigenspace associated
with the first non-zero eigenvalue is infinite-dimensional. However, imposing an ex-
tra assumption on the semigroup will be enough. Recall (see Definition 1.2.4, p. 14)
that a symmetric operator P has an L2 -density kernel with respect to μ if there ex-
ists a (symmetric, positive) measurable
function p(x, y) on the product space E × E
such that, for μ-almost every x ∈ E, E p(x, y)2 dμ(y) < ∞ and

Pf (x) = f (y) p(x, y)dμ(y), x ∈ E.
E

Note that it is not required here that E E p 2 (x, y)dμ(x)dμ(y) < ∞, which would
imply that P is Hilbert-Schmidt and therefore compact (cf. Sect. A.6, p. 483).
Theorem 4.10.5 Assume that for some r > 0 there exists a positive function
w ∈ L2 (μ) such that for any f ∈ D(E),
2
f 2 dμ ≤ r E(f ) + |f |w dμ . (4.10.5)
E E
Assume moreover that there exists a t > 0 for which Pt has an L2 -density kernel
with respect to μ. Then the essential spectrum σess (−L) of −L lies in [ 1r , ∞).
In particular, if there exists a positive function w ∈ L2 (μ) such that for every
r > 0, there exists a β(r) > 0 such that for all f ∈ D(E),
2
f dμ ≤ r E(f ) + β(r)
2
|f |w dμ ,
E E
and if Pt has an L2 -density kernel with respect to μ for some t > 0, then the spec-
trum of −L is discrete.
Proof The second part of the statement follows from the first part by applying it
to the weight β(r)1/2 w (for any r > 0). We thus concentrate on the first assertion.
Let Gw = {g ∈ L2 (μ) ; |g| ≤ w}. First observe that for an operator P with a pos-
itive density kernel in L2 (μ), the closure P (Gw ) of P (Gw ) is compact. Indeed,
given any sequence (gk )k∈N in Gw , thus bounded in L2 (μ), there exists a subse-
quence (gk )∈N which converges weakly, say to some function g ∈ L2 (μ). Then,
P having a density kernel, P gk converges μ-almost everywhere to P g, and more-
over |P gk | ≤ P w ∈ L2 (μ). From the dominated convergence Theorem, P gk then
converges to P g in L2 (μ) as → ∞.
Now (4.10.5) applied to Pt f yields, with K = Pt (Gw ),
2
Pt f 22 ≤ r E(Pt f ) + sup f g dμ . (4.10.6)
g∈K E

Indeed, with h = sign(Pt f )w, E |Pt f |wdμ = E Pt f hdμ = E f Pt hdμ, so that

|Pt f |w dμ ≤ sup f g dμ.
E g∈K E
From this point, we may proceed as in the proof of Proposition 4.10.3 and choose
by contradiction λ < 1r such that λ ∈ σess (−L). For any ε > 0, there exists an or-
thonormal sequence (fk )k∈N ⊂ D(L) such that λfk + Lfk 2 ≤ ε for
every k ∈ N.
Looking at the derivative in s of e−2λ(t−s) Ps fk 22 , and recalling that E fk2 dμ = 1,
it follows that for every k ∈ N,
√
Pt fk 2 − e−2λt ≤ 2 t ε.
2
√
Combined with |E(Pt fk ) − λPt fk 22 | ≤ ε, it follows that
√
E(Pt fk ) − λe−2λt ≤ (1 + 2λt) ε.
Now, since (fk )k∈N converges weakly to 0 as k → ∞, for any compact set
K ⊂ L2 (μ),

lim sup fk g dμ = 0.
k→∞ g∈K E
Applying (4.10.6) to fk , letting k → ∞, and choosing sufficiently small ε, we obtain

a contradiction. Theorem 4.10.5 is established.
In Theorem 4.10.5, the hypothesis that Pt has a density kernel may be replaced by
the fact that there exists a sequence (P )∈N of operators with density kernels (not
necessarily positive) such that, in operator norm, lim→∞ Pt − P = 0. Indeed,
this assumption still ensures that Pt (Gw ) is relatively compact.
4.10.3 Persson Operators and the Essential Spectrum
In practice, Theorem 4.10.5 is not that useful since to establish that Pt has an L2 (μ)
density kernel, it is often necessary to prove first that it is Hilbert-Schmidt, and
therefore compact, which already implies that −L has a discrete spectrum. In the
following we develop some further techniques which will enable us to reach the
same conclusion. Actually, to complete the picture, the reader should be aware of
the huge difference that may exist, in the analysis of Markov semigroups, between
finite and infinite dimension. In particular, in reasonable situations, the essential
spectrum only depends on what happens at infinity. To allow such an analysis, we
introduce the notion of a Persson operator.
Definition 4.10.6 (Persson operator) Let (E, μ, ) be a Markov Triple. ! Let

(Ak )k∈N be an increasing sequence of measurable subsets of E with E = k∈N Ak
and μ(Ak ) < ∞ for every k ∈ N. Let λess be the infimum of the essential spec-
trum σess (−L) of −L. We say that L is Persson with respect to (Ak )k∈N if
λess = supk∈N λk where λk = inf E(f ), the
infimum being taken over all functions
f ∈ D(E) which are 0 on Ak and such that E f 2 dμ = 1.
In concrete instances, (Ak )k∈N will typically be any increasing sequence of com-
pact subsets of E (in finite dimension), or even just balls. Observe that since the
sequence (Ak )k∈N is increasing, the sequence (λk )k∈N is increasing too. As a conse-
quence, for a Persson operator with respect to a sequence (Ak )k∈N , to assert that the
spectrum is discrete it is enough to prove that the corresponding increasing sequence
(λk )k∈N converges to ∞.
Given a sequence (Ak )k∈N as in Definition 4.10.6, introduce for
each k ∈ N the
space L2 (Ak , μ) of all measurable functions f : E → R such that Ak f 2 dμ < ∞
with associated norm f 2,Ak . Define furthermore
" #
H1 (Ak ) = f ∈ D(E) ∩ L2 (Ak , μ) ; (f )dμ < ∞ .
Ak
One main result concerning Persson operators is the following.
Proposition 4.10.7 For a sequence (Ak )k∈N as above, assume that for any k ∈ N,
there exists a function ψk ∈ D(E) with values in [0, 1] such that ψk = 0 on Ak ,
ψk = 1 on Ack+1 , and such that (ψk ) is bounded. Assume furthermore that the
embedding from H1 (Ak ) into L2 (Ak , μ) is compact for each k. Then the operator
L is Persson with respect to the sequence (Ak )k∈N .
Proof Let σ = supk∈N λk in Definition 4.10.6. The inequality

! λess ≤ σ is univer-
sal and requires nothing more than the fact that E = k∈N Ak . Indeed, choose
any λ < λess . By definition of the essential spectrum (Definition A.4.3, p. 480), the
spectral space Hλ (that is the image of L2 (μ) under the spectral projection Eλ ) is
finite-dimensional, thus its unit ball is compact. Now, if K is any compact subset

of L2 (μ), limk→∞ supg∈K g 2 dμ = 0. Therefore, given any ε > 0, there exists a

Ack

k large enough so that, for any g ∈ Hλ with E g 2 dμ = 1, Ac g 2 dμ ≤ ε. Let then

k
f ∈ D(E) be supported in Ack with E f 2 dμ = 1. Denote by g the projection of f
on Hλ , and set h = f − g. The functions h and g are orthogonal in L2 (μ) with norm
less than or equal to 1. Moreover, since g + h = 0 on Ak ,

g 2 dμ = − gh dμ = gh dμ ≤ ε.
Ak Ak Ack

2 dμ ≤ 2ε,

E h dμ ≥ 1 − 2ε.
Therefore so that 2 Now,
Eg

E(f ) = E(g) + E(h) ≥ E(h) ≥ λ h2 dμ ≥ λ(1 − 2ε).
E
This being valid for any f ∈ D(E) with support in Ack , we get λ(1 − 2ε) ≤ λk , and
the announced inequality λess ≤ σ follows.
Turning to the converse inequality λess ≥ σ , let λ belong to the essential spec-
trum σess . Applying again Weyl’s criterion (Theorem A.4.4, p. 481), for any ε > 0,
there exists an orthonormal sequence (f )∈N in D(L) such that λf + Lf 2 ≤ ε
for every . As in previous similar proofs, |E(f ) − λ| ≤ ε, and therefore the se-
quence (E(f ))∈N is bounded together with the sequence (f 2 )∈N . Hence, for
any k ∈ N, the sequence (f )∈N is also bounded in H1 (Ak ), so that by the compact-
ness assumption, there exists a subsequence which converges in L2 (Ak , μ). As the
sequence is orthonormal in L2 (μ), it converges weakly to 0 and therefore the limit
of this subsequence is 0. This may be done for any subsequence of (f )∈N , which
shows that (f )∈N converges to 0 in L2 (Ak , μ) for any k.
Now fix k ∈ N and consider, according to the hypotheses, ψk with values in [0, 1]
vanishing on Ak , equal to 1 on Ack+1 , and with bounded gradient. For every , let
g = ψk f . Then

E(g ) = ψk (f )dμ + 2 ψk f (f , ψk )dμ + f2 (ψk )dμ.
2
E E E
Since (ψk ) = 0 on Ack+1 , E(g ) ≤ E(f ) + ε with lim→∞ ε = 0. Moreover

E(f ) ≤ λ E f2 dμ + ε = λ + ε. It follows that
E(g ) ≤ λ + ε + ε .

On the other hand, E (1−ψk )2 f2 dμ ≤ Ak+1 f2 dμ and thus lim→∞ E g2 dμ = 1.
Therefore, as → ∞, λk ≤ λ + ε and the announced inequality follows. The proof
of Proposition 4.10.7 is complete.
The use of the Persson hypothesis may be combined with the weighted Nash
inequalities of Theorem 4.10.5 to reach the following main characterization.
Theorem 4.10.8 Assume that L is Persson. Then, the spectrum of −L is discrete if

and only if there exist a positive function w ∈ L2 (μ) such that for every r > 0, there
exists a β(r) > 0 such that for all f ∈ D(E),
2
f dμ ≤ r E(f ) + β(r)
2
|f |w dμ .
E E
Proof Denote by (Ak )k∈N a sequence for which L is Persson. The task is to bound
from below the corresponding values λk , k ∈ N, defined in Definition

4.10.6. To this
end, let r > 0 be fixed, and choose k = k(r) such that β(r) Ac w 2 dμ ≤ 12 . Then, by
k
the hypothesis applied to a function f with support in Ack , and the Cauchy-Schwarz
inequality,
2
1
f 2 dμ ≤ r E(f ) + β(r) |f |w dμ ≤ r E(f ) + f 2 dμ,
E Ack 2 E
whence λk(r) ≥ 2r1 . Since r > 0 is arbitrary, the increasing sequence (λk )k∈N con-
verges to ∞. The converse is achieved via Proposition 4.10.4. The proof is therefore
complete.
The compactness hypothesis in Proposition 4.10.7

ensures that, for fixed k ∈ N,
given any sequence (f )∈N in D(E) such that Ak [(f ) + f2 ]dμ is uniformly
bounded, there exists a subsequence which converges in L2 (Ak , μ). In practice
when dealing with elliptic operators on Rn or on open subsets O ⊂ Rn with smooth
coefficients, (Ak )k∈N will be any increasing sequence of “nice” sets such as balls.
Then the embedding is nothing else than the classical Rellich-Kondrachov Theo-
rem (Theorem 6.4.3, p. 291). When the space Cc∞ (O) of C ∞ compactly supported
functions in O is dense in the domain D(E), it is easily seen, playing with the func-
tions ψk between Ak and Ak+1 , that the
infimum in the definition of λk may be
restricted to functions f ∈ Cc∞ (Ack ) (with E f 2 dμ = 1). As a consequence, Propo-
sition 4.10.7 yields the following useful corollary.
Corollary 4.10.9 Let L = − ∇W · ∇ in Rn be such that

1
lim −W (x) + |∇W |2 (x) = ∞.
x→∞ 2
Then, the spectrum of −L is discrete. The same is true if Rn is replaced by an open

subset O ⊂ Rn provided that L is essentially self-adjoint for Cc∞ (O), under the
preceding limit at the boundary of O.
Proof We first deal with the case of Rn , for which we already know that L is essen-
tially self-adjoint thanks to Corollary 3.2.2, p. 143. Recall that here dμ = e−W dx.
Choose as the sequence (Ak )k∈N the sequence of balls centered at the origin with
radius k which clearly satisfies the hypotheses of Proposition 4.10.7. It is therefore

enough to show that for any smooth function f compactly supported in Ack ,

|∇f |2 dμ ≥ ck f 2 dμ
Rn Rn
for some ck → ∞. Changing f into geW/2 , and after integration by parts (justified
since g is compactly supported), this amounts to proving that for any g smooth and
compactly supported in Ack ,

|∇g|2 + Rg 2 dx ≥ ck g 2 dx
Rn Rn
where R = − 12 W + 14 |∇W |2 . This is however a direct consequence of the hypoth-

esis.
For an open set O ⊂ Rn , let (Ak )k∈N consisting of the sets of Ak = {J ≤ k} satis-
fying the
hypotheses of
Proposition 4.10.7. The argument then reduces to examining
the ratio O |∇f |2 dμ/ O f 2 dμ for functions which are compactly supported in Ack
and the same transform of f → f eW/2 applies. The proof is complete.
Similar proofs may be adapted to more general settings, in particular to

manifolds. As an example, for every α > 1, the probability measure in Rn ,
dμα (x) = cα e−|x| , has a discrete spectrum. For α = 1, the discrete spectrum is
α
empty as already observed on p. 189.
Due to their interplay with spectral properties, Poincaré inequalities form a vast
subject. They have been investigated and studied in a wide variety of settings, in
particular in the analysis of partial differential equations, and under various names
in the literature. This chapter only focuses on the aspects of Poincaré inequalities
emphasized in this book, namely functional inequalities, convergence to equilibrium
and heat kernel bounds. In particular, we do not address Poincaré inequalities on
balls (local Poincaré inequalities), doubling properties and the related analysis on
metric spaces. For these topics, we refer, for example, to [9, 13, 125, 217, 230, 236,
376, 397, 422] and to the references therein.
Poincaré inequalities also sometimes refer to the comparison of the L2 -norm of
a function with its gradient (in the form (4.10.1)) for functions vanishing at the
boundary of some domain. This form is related to the Dirichlet eigenvalue problem
in partial differential equations.
The Poincaré inequalities investigated in this chapter, and throughout this book,
mainly deal with Neumann boundary conditions comparing the L2 -norm of a func-
tion with mean zero with the one of its gradient. In this form, they may probably
be traced back to the contributions [348, 349] where H. Poincaré used duplication
((4.2.4)) and a clever change of variables to establish the inequality on bounded
convex domains in Rn . They are sometimes called Poincaré-Wirtinger inequalities
in reference to W. Wirtinger who seemingly analyzed functions on the circle (corre-
sponding to (4.5.5)) using a spectral decomposition along the trigonometric system.
W. Wirtinger, who is mentioned in the book [70], also considered the extension on
the sphere using spherical harmonics. The term “Poincaré inequality” first appears
in the book by R. Courant and D. Hilbert [140]. See [8, 300] for a short historical
introduction to Poincaré inequalities in the context of partial differential equations.
We refer in addition to [317] for a comprehensive survey of inequalities that in-
volve a relationship between a function and its derivatives or integrals, including in
particular Poincaré inequalities.
The name of Poincaré inequality has been used for numerous inequalities of the
same form or which are obtained similarly. In particular, the Poincaré inequality for
the Gaussian measure (Proposition 4.1.1) may be traced back to the early 30s in the
physics literature along with the Hermite eigenfunction expansion. It appears later
in papers of J. Nash [323], H. Chernoff [130], L. Chen [127] and in many places
today.
The main classical exponential decay Theorem 4.2.5 in Sect. 4.2.2 is a standard
result in Fourier-type analysis, and a fundamental part of the theory of Poincaré
inequalities. Lemma 4.2.6 may be found in [139].
The stability by bounded perturbations and tensorization (Sect. 4.3) properties of
Poincaré inequalities form part of the folklore of the subject.
The Poincaré inequality for the exponential measure of Proposition 4.4.1 is
pointed out in [403]. Its interpretation as a Lyapunov criterion has been empha-
sized in, for example, [29] (see below). Exponential integrability under a Poincaré
inequality goes back to the works [89] and [222]. It was revived and improved in
the corresponding investigation under logarithmic Sobolev inequalities (see the next
chapter) by S. Aida and D. Stroock [7], L. Gross and O. Rothaus [227] and others
(cf. [276]). The proof presented here is extracted from [79]. More on applications to
measure concentration may be found in [90, 276, 278].
The Muckenhoupt criterion presented in Sect. 4.5.1 also has a long history start-
ing with early papers by G. Hardy [231, 232], later developed in [321, 405, 410].
In particular, B. Muckenhoupt presented in [321] a characterization with respect to
two measures on R+ (see also Chap. 8 in this regard). More details on those re-
sults can be found in [269]. In the form of Theorem 4.5.1, the criterion is recalled
in [77] by comparison with the criterion for logarithmic Sobolev inequalities (see
the next chapter). See also [14] for further details. More precise comparisons be-
tween the Muckenhoupt functional and the optimal Poincaré constant are developed
in [310]. The Poincaré inequalities on a bounded interval of the real line presented
in Sect. 4.5.2 are more or less classical and may be found in various places in the
literature, including [203].
The Lyapunov method discussed in Sect. 4.6 has its origin in the Markov chain
context (cf. e.g. [309]) where it has been used as a tool towards convergence to
equilibrium. It has also been developed by L. Hörmander in partial differential equa-

tions, already with the aim of establishing functional inequalities. The method has
been quite successful in the context of functional inequalities in recent years, see
e.g. [29, 33, 119] and the forthcoming monograph [118] for further developments.
Theorem 4.6.3 on Poincaré inequalities for log-concave measures is due to R. Kan-
nan, L. Lovász and M. Simonovits [262] in the case of the uniform measure on a
convex body and to S. Bobkov [74] in the general case. The Poincaré constant from
these works, based on the variance of the measure, improves upon the mere diameter
bound (as in (4.6.6)) but is still far from the famous KLS conjecture from [262] as-
serting that the sharp constant should be achieved on affine functions (see [228] for
a recent short survey). In dimension one, an another estimate in terms of the mean
value of the distance to the median has been pointed out in [186]. The Lyapunov
proof of Theorem 4.6.3 is given in [29]. The optimal dependence on the diame-
ter (4.6.6) for a convex domain is due to L. Payne and H. Weinberger [53, 343]. The
glueing Proposition 4.6.4 is taken from [187].
The local Poincaré inequalities for heat kernel measures of Sect. 4.7 were put
forward in [27] in parallel with their counterparts for logarithmic Sobolev inequal-
ities discussed in the next chapter (see also [28, 277, 394]). The basic semigroup
interpolation argument goes back to the seminal work [36] in the context of logarith-
mic Sobolev inequalities (see next chapter). Illustrations of the underlying Duhamel
principle from the 19th century may be found, for example, in [258]. Local inequal-
ities (here and in the next chapters) associated to operators of the form L = g + Z
for a general vector field Z may be investigated similarly, although the non-compact
case raises several delicate technical issues (cf. [431, 434, 435]). The local Poincaré
inequalities for heat kernel measures may be seen as counterparts of local Poincaré
inequalities on balls. The latter have been investigated deeply in the analysis of
spaces endowed with measures with doubling properties (without curvature assump-
tions). In particular, the works of A. Grigor’yan [214] and L. Saloff-Coste [374]
connect them with Sobolev-type and Harnack inequalities (cf. [217, 376]).
Poincaré inequalities under a curvature-dimension condition in Sect. 4.8 have
their origin in the Lichnerowicz eigenvalue comparison Theorem (4.8.1) [286] for
manifolds with a strictly positive lower bound on the Ricci curvature (see, for exam-
ple [121, 123, 126, 194, 346]). See also [246, 249]. Obata’s Theorem, stating that
whenever there is equality in the Lichnerowiz lower bound (4.8.1) then the mani-
fold is isometric to a sphere, goes back to [335]. In the context of the -calculus,
Poincaré inequalities under a curvature-dimension condition first appeared in [36]
(cf. [26] for a first synthesis). In particular, Theorem 4.8.4 includes the Wirtinger
inequality for the standard n-sphere. Sharper spectral comparison theorems in this
context are developed in [42] (see the references therein).
The Brascamp-Lieb inequality (Theorem 4.9.1) goes back to [93] (see [288]).
The proof here is inspired by [385] and was quite successful under the name
of the Witten Laplacian method in the study of models in statistical physics
(cf. [237–239]). The probabilistic representation (4.9.6) of the gradient of Pt is
known as Bismut’s formula [69] and is detailed in a manifold setting in, for example,
[251, 391, 394].
Section 4.10 collects a variety of results and observations on spectral properties

scattered all over the literature. Details on the positivity of eigenvectors may be
found in [355]. In the regular case, results and history on the existence of strictly
positive harmonic functions are discussed in [347]. The Poincaré inequality on the
hyperbolic space (Proposition 4.10.2) is proved in a general context in [306] (see
also [121]). Theorem 4.10.5 is due to F.-Y. Wang [429] and expands on earlier work
of A. Persson [345] on Schrödinger operators (see also [163] and [219]).
Chapter 5
Logarithmic Sobolev Inequalities
After Poincaré inequalities, logarithmic Sobolev inequalities are amongst the most
studied functional inequalities for semigroups. Indeed, they contain much more in-
formation than Poincaré inequalities, and are at the same time sufficiently general
to be available in numerous cases of interest, in particular in infinite dimension (as
limits of Sobolev inequalities on finite-dimensional spaces). Moreover, they entail
remarkable smoothing properties of the semigroup in the form of hypercontractivity.
The structure of this chapter is quite similar to the preceding one on Poincaré
inequalities. In particular, the setting is that of a Full Markov Triple, abbreviated
as “Markov Triple”, (E, μ, ) with associated Dirichlet form E, infinitesimal gen-
erator L, Markov semigroup P = (Pt )t≥0 and underlying function algebras A0 and
A (cf. Sect. 3.4, p. 168). Logarithmic Sobolev inequalities under the invariant mea-
sure only concern finite (normalized) measures μ. It should be mentioned that log-
arithmic Sobolev inequalities involve entropy and Fisher information, which deal
with strictly positive functions. It will therefore be convenient to deal with the class
Aconst+
0 of Remark 3.3.3, p. 154, consisting of functions which are sums of a positive
function in A0 and a strictly positive constant.
Once again, several definitions and properties make sense for more general triples
(E, μ, ), and the Standard Markov Triple assumption suffices for a number of
results. Contrary to the preceding chapter on Poincaré inequalities, the diffusion
property can however not be discarded.
The first section introduces the basic definition of a logarithmic Sobolev inequal-
ity together with its first properties. Section 5.2 presents the exponential decay in en-
tropy and the fundamental equivalence between logarithmic Sobolev inequality and
hypercontractivity. The next sections discuss integrability properties of eigenvectors
and of Lipschitz functions under a logarithmic Sobolev inequality and present, as in
the case of Poincaré inequalities, a criterion for measures on the real line to satisfy
a logarithmic Sobolev inequality (for the usual gradient). Sections 5.5 and 5.7 deal
with curvature conditions, first for the local logarithmic Sobolev inequalities for
heat kernel measures, then for the invariant measure with additional dimensional in-
formation. Local hypercontractivity and some applications of the local logarithmic
Sobolev inequalities towards heat kernel bounds are further presented. Section 5.6

DOI 10.1007/978-3-319-00227-9_5,
236 5 Logarithmic Sobolev Inequalities
develops Harnack-type inequalities under the infinite-dimensional curvature con-

ditions CD(ρ, ∞), and describes links with reverse local logarithmic Sobolev in-
equalities.
5.1 Logarithmic Sobolev Inequalities

This section formally introduces the notion of a logarithmic Sobolev inequality for
a Markov Triple (E, μ, ) and presents some of its first properties and equivalent
formulations.
5.1.1 Logarithmic Sobolev Inequalities
For a measure ν (not necessarily finite) on

a measurable space (E, F), define for all
positive integrable functions f such that E f | log f |dν < ∞,

Entν (f ) = f log f dν − f dν log f dν . (5.1.1)
E E E
(Recall that 0 log 0 is interpreted as 0 in this definition.) When ν is a probability

measure, since the function φ(r) = r log r is strictly convex on R+ , by Jensen’s
inequality, Entν (f ) ≥ 0 and is equal to 0 only if f is constant. Note furthermore
that Entν (cf ) = c Entν (f ) for every c ≥ 0. For future purposes, let us also mention
the classical entropic inequality for a probability measure ν,

f g dν ≤ Entν (f ) + f dν log g
e dν (5.1.2)
E E E
whenever
f ≥ 0 and g are suitably integrable.

g For
a sketch of the proof, assume that
f dν = 1 and set d ν̃ = f dν so that e dν = e g f −1 d ν̃. Then apply Jensen’s
E E E
inequality with respect to the probability measure ν̃ and the concave function log.
The entropic inequality (5.1.2) implies the useful variational formula, for ν a prob-
ability measure, and every positive measurable function f : E → R,

Entν (f ) = sup f g dν, (5.1.3)
E

where the supremum runs over all functions g such that E eg dν ≤ 1.

The following definition introduces the notion of a logarithmic Sobolev inequal-
ity.
Definition 5.1.1 (Logarithmic Sobolev inequality) A Markov Triple (E, μ, )

(with μ a probability measure) is said to satisfy a logarithmic Sobolev inequal-
ity LS(C, D) with constants C > 0, D ≥ 0, if for all functions f : E → R in the
5.1 Logarithmic Sobolev Inequalities 237
Dirichlet domain D(E),

Entμ f 2 ≤ 2C E(f ) + D f 2 dμ.
E
When D = 0, the logarithmic Sobolev inequality will be called tight, and will be
denoted by LS(C). When D > 0, the logarithmic Sobolev inequality LS(C, D) is
called defective.
The best constant C > 0 for which such an inequality LS(C) holds is sometimes
referred to as the logarithmic Sobolev constant (of the Markov Triple). Note that
it is part of the information

contained in the logarithmic Sobolev inequality that
provided f ∈ D(E), then E f 2 log(1 + f 2 )dμ < ∞, that is f belongs to the Orlicz
space L2 log L(μ) (observe further that L2 log L(μ) ⊃ Lp (μ) for every p > 2 for
the comparison with Sobolev inequalities in the next chapter). As usual, it is enough
to state (and prove) such inequalities for a family of functions f which is dense in
the Dirichlet domain D(E) (typically the algebra A0 ).
We often simply say that the probability measure μ satisfies a logarithmic
Sobolev inequality LS(C, D) or LS(C) (with respect to the underlying carré du
champ operator or Dirichlet form E). The normalization constant 2 is chosen
for further comparisons. By the homogeneity property of entropy, the logarithmic
Sobolev inequality of Definition 5.1.1 is homogeneous (invariant under a change of
f to cf ). The tight logarithmic Sobolev inequality LS(C) implies that if E(f ) = 0,
then f is constant. This observation is in agreement with the corresponding prop-
erty for Poincaré inequalities (cf. Remark 4.2.2, p. 182, and the link with connexity
in Sect. 3.2.1, p. 140).
The
√ logarithmic Sobolev inequality LS(C, D) is often presented equivalently
for f , f ≥ 0 (see Remark 5.1.2 below), and then takes the form, by the chain rule
formula,

C (f )
Entμ (f ) ≤ dμ + D f dμ. (5.1.4)
2 E f E
Since
f may be changed by homogeneity into cf , it can be assumed moreover that

E f dμ = 1. Setting dν = f dμ then defines a probability measure on (E, F). The
quantity

H(ν | μ) = Entμ (f ) = log f dν (5.1.5)
E
is then the relative entropy of the probability measure ν with respect to μ whereas

(f )
I(ν | μ) = Iμ (f ) = dμ = 4 f dμ (5.1.6)
E f E
is called the Fisher information of ν with respect to μ. A logarithmic Sobolev in-

equality LS(C) thus compares the entropy to the information
C
H(ν | μ) ≤ I(ν | μ) (5.1.7)
2
for every probability measure ν absolutely continuous with respect to μ, a language

often used in the context of information theory (where logarithmic Sobolev inequal-
ities were first considered).
Remark 5.1.2 The meaning of Iμ (f ) where the function f may vanish should be

sense when f = g for any g ∈ A0 , or even
clarified. It actually makes perfect 2
g ∈ D(E), since then Iμ (f ) = 4 E (g) dμ. In a concrete setting, it also makes

sense for any smooth strictly positive function f as the integral E (f )
f dμ, where
the latter integral may or may not be convergent. When f is only positive, observe
that Iμ (f ) = limε0 Iμ (f +ε). Therefore when dealing with such quantities, it is of-
ten enough to consider Iμ (f ) for f in Aconst+
0 . Throughout this section, inequalities
involving a Fisher-type expression will usually be established first for functions in
this class and then extended to more general functions taking limits. This procedure
will often be implicit.
The prototypical logarithmic Sobolev inequality is that for the standard Gaus-
sian measure on R or Rn for which LS(C) holds with C = 1, independently of
the dimension n of the underlying state space. (The value C = 1 justifies the nor-
malization chosen in Definition 5.1.1. This normalization will also be convenient
for the comparison with Poincaré inequalities below.) However, while the Poincaré
inequality for the Gaussian measure may be established by a simple spectral expan-
sion along the Hermite polynomials, the proof of the logarithmic Sobolev inequality
is more delicate, and is thus postponed until later in the chapter (Proposition 5.5.1).
The name logarithmic Sobolev inequality describes a weak form of the Sobolev
inequalities studied in Chap. 6. However, while Sobolev inequalities will typically
hold only in finite-dimensional spaces, logarithmic Sobolev inequalities may be
investigated in infinite-dimensional contexts (in probability theory, statistical me-
chanics, statistics, and in numerous settings dealing with an infinite number of co-
ordinates), where they provide useful bounds and controls. It will be shown later
how they really appear as limits of Sobolev inequalities as the dimension tends to
infinity (see Remark 6.2.6, p. 285, and Sect. 6.5, p. 291). Logarithmic Sobolev in-
equalities are also of interest in finite dimension as will be discussed in the chapter
and throughout the book.
5.1.2 Tightening and Perturbation
The tight logarithmic Sobolev inequality LS(C) is stronger than the Poincaré in-
equality P (C) (Definition 4.2.1, p. 181). This is the content of the following propo-
sition, which furthermore shows that a logarithmic Sobolev inequality LS(C, D)
together with a Poincaré inequality actually yield a tight logarithmic Sobolev in-
equality.
Proposition 5.1.3 (Tightening with a Poincaré inequality) A tight logarithmic

Sobolev inequality LS(C) implies a Poincaré inequality P (C). Furthermore, a log-
arithmic Sobolev inequality LS(C, D) together with a Poincaré inequality P (C )

implies a tight logarithmic Sobolev inequality LS(C + C ( D2 + 1)).
Proof For the first assertion, apply LS(C) to f = 1 + εg where g ∈ D(E) with
E g dμ = 0. As ε → 0, it is not difficult to check by a Taylor expansion that

Entμ f 2 = 2ε 2 g 2 dμ + o ε 2
E
while, clearly, E(f ) = ε 2 E(g). It follows that

2 g 2 dμ ≤ 2C E(g)
E
which amounts to the Poincaré inequality P (C).
Turning to the second claim, the following lemma will be useful.
Lemma
2 5.1.4 (Rothaus’ Lemma) Let f : E → R be measurable and such that

E f log(1 + f 2 ) dμ < ∞. For every a ∈ R,

Entμ (f + a)2 ≤ Entμ f 2 + 2 f 2 dμ.
E
Proof With f and a fixed, the aim is to verify that the function

ψ(r) = Entμ (rf + a)2 − Entμ (rf )2 − 2 (rf )2 dμ, r ∈ R,
E
is negative. To this end, observe that ψ is concave. Namely,

ψ (r)
= f log v dμ −
2 2 2
f dμ log 2
v dμ
2 E E E
−1 2

− Entμ f 2 − 2 2
v dμ f v dμ
E E
where v = rf + a. Now, by concavity and Jensen’s inequality,

Entμ f 2 ≥ f 2 log v 2 dμ − f 2 dμ log v 2 dμ .
E E E
Hence ψ is concave, and since, as is easily checked,

ψ(0) = ψ (0) = 0, the an-
nounced claim follows. Note that whenever E v 2 dμ = 0 in this argument, then
rf + a = 0 μ-almost everywhere, and the conclusion is immediate in this case. The
lemma is proved.

Applied to fˆ = f − E f dμ with a = E f dμ, the previous lemma yields

Entμ f 2 ≤ Entμ fˆ2 + 2 Varμ (f ). (5.1.8)
On this basis, the proof of the second assertion of Proposition 5.1.3 is immediate.
Indeed, by the logarithmic Sobolev inequality LS(C, D) applied to fˆ, and (5.1.8),

2
Entμ f ≤ 2C E(f ) + D fˆ2 dμ + 2 Varμ (f ),
E

and since E fˆ2 dμ = Varμ (f ), it remains to use the Poincaré inequality P (C ).

Proposition 5.1.3 is therefore established.
Remark 5.1.5 Examples will be provided below for which the logarithmic Sobolev
constant is strictly larger than the (finite) Poincaré constant (or is even infinite). In
this respect, observe that whenever the smallest (non-trivial)

eigenvalue λ1 of −L
(provided it exists) admits an eigenfunction h such that E h3 dμ = 0, then the Tay-
lor expansion to the next term around h used in the proof of Proposition 5.1.3 yields
such an instance for which the logarithmic Sobolev constant of μ is necessarily
strictly larger than the spectral gap constant (equal to λ11 ).
As for Poincaré inequalities, logarithmic Sobolev inequalities for a probability

measure μ with respect to a given carré du champ operator are often established
by comparison with known ones. For example, if μ satisfies a logarithmic Sobolev
inequality LS(C, D) with respect to , and if 1 is another carré du champ operator
such that ≤ a 1 for some a > 0, then (μ, 1 ) satisfies LS(aC, D).
The following perturbation result is an analogue of Proposition 4.2.7, p. 185.
Proposition 5.1.6 (Bounded perturbation) Assume that the Markov Triple (E, μ, )
satisfies a logarithmic Sobolev inequality LS(C, D). Let μ1 be a probability mea-
sure with density h with respect to μ such that b1 ≤ h ≤ b for some constant b > 0.
Then μ1 satisfies LS(b2 C, b2 D) (with respect to ).
The proposition is a simple consequence of the following variational formula.

As for the Poincaré inequality, if h = ek , b2 may be replaced by eosc(k) where again
osc(k) = sup k − inf k.
Lemma 5.1.7 Let φ : I → R on some open interval I ⊂ R be convex of class C 2 .

For every (bounded or suitably integrable) measurable function f : E → R with
values in I ,

φ(f ) dμ − φ f dμ = inf φ(f ) − φ(r) − φ (r)(f − r) dμ.
E E r∈I E
The important issue here, for both the proof of the lemma and its application,

is
that φ(s) − φ(r) − φ (r)(s − r) ≥ 0, r, s ∈ I , by convexity. With s = E f dμ, the
left-hand side of the equality of the lemma is less

than or equal to its right-hand
side, while equality is achieved precisely for r = E f dμ. Now for the function
φ(r) = r log r on I = (0, ∞), Proposition 5.1.6 is easily established since

Entμ1 (f ) = inf φ(f ) − φ(r) − φ (r)(f − r) dμ1 ≤ b Entμ (f )
r∈I E

(while obviously E (f )dμ ≤ b E (f )dμ1 ).
5.1.3 Slicing of Logarithmic Sobolev Inequalities
To conclude this section, we present some alternative forms of logarithmic Sobolev

inequalities. Like many of the functional inequalities encountered throughout this
book, logarithmic Sobolev inequalities may appear in a number of different forms,
and it is not always easy to recognize them in disguise. Below we present one of
many examples, included in the so-called Nash family which we shall meet later in
Chaps. 6 and 7.
Proposition 5.1.8 If (E, μ, ) satisfies a logarithmic Sobolev inequality LS(C, D),

then, for any p ∈ [1, 2) and any f ∈ D(E),

f 22
f 22 log ≤ C1 E(f ) + D1 f 22 , (5.1.9)
f 2p
with C1 = (2−p)C
p and D1 = (2−p)D
p . Conversely, if (5.1.9) holds for some p ∈ [1, 2)
and every f ∈ D(E), then, for constants c(p) > 0 and d(p) > 0 depending only on
p, LS(C, D) holds with C = c(p)C1 and D = c(p)D1 + d(p).
Proof We first show that LS(C, D) implies (5.1.9). The argument only relies
on the rephrasing of Hölder’s inequality as the convexity of the map
r → φ(r) = log(f 1/r ), r ∈ (0, 1] (for f = 0). Therefore, for 1 ≤ p < 2,
φ( p1 ) − φ( 12 )
1
≥ φ .
1
p − 1
2
2
Now, derivatives of Lp (μ)-norms give rise to entropy as

1 Entμ (f 2 )
φ =− ,
2 f 22
(2−p)C (2−p)D
from which (5.1.9) follows with C1 = p and D1 = p . Note that D1 = 0
for a tight logarithmic Sobolev inequality LS(C).
To get the reverse inequality, we introduce a slicing technique, already men-

tioned in Proposition 3.1.17, p. 137, and which will be used later in the study of
Sobolev inequalities and formalized in a more efficient way in Chaps. 6 and 8. Let
f ∈ D(E), which may be assumed positive as E(|f |) ≤ E(f ). By homogeneity, it
can also be assumed that f 2 = 1. The idea is to apply (5.1.9) to the family of
functions fk = (f − 2k )+ ∧ 2k , k ∈ Z (which still belong to the Dirichlet domain
D(E)). To this end, let Nk = {f > 2k }, k ∈ Z. Since f = k∈Z 1Nk \Nk+1 f , observe
first that

3 2k
2 μ(Nk ) ≤ f 2 dμ ≤ 3 22k μ(Nk ).
4 E
k∈Z k∈Z

2
In other words, since E f dμ = 1, setting αk = 2 μ(Nk ), k ∈ Z,
2k
1 4
≤ αk ≤ .
3 3
k∈Z
In particular αk ≤ 4
3 for every k. Now, for every k ∈ Z,
2k 1Nk+1 ≤ fk ≤ 2k 1Nk ,
αk+1
so that, letting βk = fk 22 , 4 ≤ βk ≤ αk and, for 1 ≤ p < 2,
fk 2p ≤ 22k μ(Nk )2/p = 2−2k((2/p)−1) αk

2/p
for every k ∈ Z. Therefore,

(2/p)−1
fk 22 3 βk
≥ 22k((2/p)−1)
fk 2p 4 αk
2/p
where we used that αk ≤ ( 34 )1−(2/p) αk . Now apply (5.1.9) to fk for every k ∈ Z
to get that

αk
k βk ≤ c(p) βk log + C1 E(fk ) + D1 + d(p) βk
βk
for constants
c(p), d(p) > 0 only depending on p. Decomposing f again as
f = k∈Z 1Nk \Nk+1 f shows similarly (after some details) that

f log f dμ ≤ 3 log 4
2 2
k αk + 8 .
E k∈Z
(The numerical constant 8 is far from optimal in this last inequality.)

It thus follows from the preceding that, for possibly new values of c(p) and d(p),

αk
f log f dμ ≤ c(p)
2 2
βk log + C1 E(fk ) + D1 + d(p) .
E bk
k∈Z k∈Z
5.2 Entropy Decay and Hypercontractivity 243
By Jensen’s inequality for the concave function log and the fact that αk+1
4 ≤ βk ≤ αk
for every k,

αk k∈Z αk 4
βk log ≤ βk log ≤ log 4
βk k∈Z βk 3
k∈Z k∈Z
from which it follows that

f 2 log f 2 dμ ≤ c(p) C1 E(fk ) + d(p) + D1 .
E k∈Z

With the help of Proposition (3.1.17), p. 137, k∈Z E(fk ) ≤ E(f ), so that

f 2 log f 2 dμ ≤ c(p)C1 E(f ) + c(p)D1 + d(p),
E
which amounts by homogeneity to the announced logarithmic Sobolev inequality.

Proposition 5.1.8 is established.
In the preceding proof, going from one form of the inequality to the other and
back does not preserve the constants. While the proof shows that the tight loga-
rithmic Sobolev inequality LS(C) implies D1 = 0, conversely (5.1.9) with D1 = 0
in the preceding proof does not yield back LS(C) for some C > 0. However, a
Taylor expansion f = 1 + εg in (5.1.9) with D1 = 0 yields a Poincaré inequality
C1
P ( 2−p ), so that, together with Proposition 5.1.3, we actually reach a tight logarith-
mic Sobolev inequality in this case.
5.2 Entropy Decay and Hypercontractivity
Like Poincaré inequalities (cf. Sect. 4.2.2, p. 183), logarithmic Sobolev inequal-
ities have something to tell about convergence to equilibrium of the semigroup
P = (Pt )t≥0 towards the invariant (probability) measure μ. In addition, logarith-
mic Sobolev inequalities ensure smoothing properties of the semigroup in the form
of hypercontractivity.
5.2.1 Exponential Decay in Entropy
The following statement is the analogue for entropy of the exponential decay in
L2 (μ) produced by a Poincaré inequality (Theorem 4.2.5, p. 183). Furthermore, the
proof of this result exhibits the fundamental relation between decay of the entropy
along the semigroup and the Fisher information (5.1.6).
Theorem 5.2.1 (Exponential decay in entropy) The logarithmic Sobolev inequality

LS(C) for the probability measure μ is equivalent to saying that for every positive
function f in L1 (μ) (with finite entropy),
Entμ (Pt f ) ≤ e−2t/C Entμ (f )
for every t ≥ 0.
It is not obvious a priori that the convergence in this statement is stronger than the
convergence in variance as produced by a Poincaré inequality. One point is that the
convergence holds for a larger class of functions. For example, if μ is the standard
Gaussian distribution on the real line (for which LS(C) holds with C = 1) and if
2
f (x) = ecx /2 , x ∈ R, c ∈ (0, 1), then Varμ (f ) = ∞ if c ≥ 12 whereas Entμ (f ) =
−3/2 + (1 − c)−1/2 log(1 − c)) for every c ∈ (0, 1).
2 (c(1 − c)
1
Furthermore, convergence in entropy yields convergence in total variation ac-

cording to the following scheme. For probability measures μ, ν on (E, F), intro-
duce, according to (5.1.5), the relative entropy H(ν | μ) of ν with respect to μ as

dν
H(ν | μ) = log dν (5.2.1)
E dμ
whenever ν is absolutely continuous with respect to μ with Radon-Nikodym deriva-
tive f = dμ
dν
, and H(ν | μ) = ∞ if not. In other words, if ν μ with Radon-
Nikodym derivative f , H(ν | μ) = Entμ (f ). Then Theorem 5.2.1 may be reinter-
preted as
H(νt | μ) ≤ e−2t/C H(ν0 | μ)
where dνt = Pt f dμ, t ≥ 0.
Now, the celebrated Pinsker-Csizsár-Kullback inequality indicates that for μ, ν
probability measures on (E, F),
1
μ − ν2TV ≤ H(ν | μ) (5.2.2)
2
where μ − νTV is the total variation distance

μ − νTV = sup μ(A) − ν(A).
A∈F
Therefore, under a control of H(ν0 | μ), Theorem 5.2.1 implies the stronger conver-
gence of νt towards μ in total variation.
For the sake of completeness, here is a brief proof of (5.2.2) (which shares
some features with the proof of Lemma 5.1.4 above). If f
≥ 0 denotes the
Radon-Nikodym derivative dμ dν
, then, thanks to the identity E |1 − dμ dν
|dμ =
2μ − νTV , (5.2.2) translates into
2
|1 − f |dμ ≤ 2 Entμ (f ).
E
Set fs = sf + 1 − s, s ∈ [0, 1], and consider

2 2
(s) = 2 Entμ (fs ) − |1 − fs |dμ = 2 Entμ (fs ) − s 2
|1 − f |dμ .
E E
Then, (0) = (0) = 0 and

2
(1 − f )2
(s) = 2 dμ − 2 |1 − f |dμ , s ∈ [0, 1).
E fs E
By the Cauchy-Schwarz inequality, ≥ 0 so that is positive, which completes

the proof.
We next address the proof of Theorem 5.2.1. The principle of proof is the same as
that for Theorem 4.2.5, p. 183, this time differentiating entropy (t) = Entμ (Pt f ),
t ≥ 0, rather than variance. The argument relies on the following main identity
which is an analogue of (4.2.2), p. 183, for entropy. Recall the Fisher information
Iμ from (5.1.6).
Proposition 5.2.2 (de Bruijn’s identity) For every positive f in D(E),
d
Entμ (Pt f ) = −Iμ (Pt f ). (5.2.3)
dt
The proof is immediate. Indeed, by the heat equation and integration by parts,
for any f ∈ Aconst+
0 to start with, and then extending to f ∈ D(E),

d
Entμ (Pt f ) = (1 + log Pt f ) LPt f dμ
dt E

=− (Pt f, 1 + log Pt f )dμ
E

(Pt f )
=− dμ = −Iμ (Pt f ).
E Pt f
This basic relation between the time derivative of entropy and the Fisher informa-
tion in particular tells us that entropy is decreasing along the flow of the semigroup,
a property often referred to as the Boltzmann (H -) Theorem.
On the basis of de Bruijn’s identity, the logarithmic Sobolev inequality LS(C) in
its Fisher information formulation (5.1.4)
C
Entμ (g) ≤ Iμ (g)
2
translates into the differential inequality (t) ≤ − C2 (t), t > 0, from which the in-
equality of Theorem 5.2.1 easily follows. Conversely, differentiating this inequality
at t = 0 yields the logarithmic Sobolev inequality by the same argument.
5.2.2 Hypercontractivity
The evolution of the semigroup P = (Pt )t≥0 under a logarithmic Sobolev inequal-
ity may be described in another, new, formulation, at the origin of the notion of
logarithmic Sobolev inequality, known as the hypercontractivity property. This is a
smoothing property that exactly translates the logarithmic Sobolev inequality. The
next theorem presents this fundamental equivalence.
Theorem 5.2.3 (Gross’ hypercontractivity Theorem) For parameters C > 0,

D ≥ 0, 1 < p < ∞, let q(t), 1 < q(t) < ∞, and M(t) ≥ 0 be defined for every
t ≥ 0 by

q(t) − 1 1 1
= e2t/C , M(t) = D − .
p−1 p q(t)
For a Markov Triple (E, μ, ) with semigroup P = (Pt )t≥0 , the following assertions
are equivalent.
(i) The logarithmic Sobolev inequality LS(C, D) holds.
(ii) For some (every) 1 < p < ∞, and for every t ≥ 0 and every f ∈ Lp (μ),
Pt f q(t) ≤ eM(t) f p .
The second assertion of the theorem is thus called hypercontractivity of the semi-
group (Pt )t≥0 . Note that M(t) = 0 under a tight logarithmic Sobolev inequality
(D = 0) so that LS(C) holds if and only if for every t > 0 and f ∈ Lp (μ),
Pt f q ≤ f p
q−1
for (some or any) 1 < p < q < ∞ such that e2t/C ≥ p−1 . What Theorem 5.2.3
tells us is that, under a logarithmic Sobolev inequality, not only are the operators
Pt contractions in all Lp (μ)-spaces, but they improve integrability for large t (in
particular they are contractions from Lp (μ) to Lq (μ) when D = 0). This may be
shown to be optimal for some examples, such as the Ornstein-Uhlenbeck semigroup
(see below). Note that the theorem does not tell us anything in the case where p = 1.
Proof Once the correct starting point has been identified, the proof is rather simple.
To show that LS(C, D) implies hypercontractivity, we use the traditional strategy
of differentiating a suitable functional. Below, f is a function in Aconst+
0 (the result
being extended to functions in Lp (μ) by standard density arguments). A first obser-
vation, already used earlier in this chapter, is that the derivative of the norm · q
along its parameter q gives rise to the entropy. Namely,

q
∂q f q = ∂q f q dμ
E

1 q
= f log f dμ =
q
Entμ f + f dμ log f dμ .
q q
E q E E
On the other hand, for every fixed q > 1, by integration by parts and the diffusion
property,

∂t (Pt f )q dμ = q (Pt f )q−1 LPt f dμ
E E

= −q(q − 1) (Pt f )q−2 (Pt f )dμ.
E
Finally, the logarithmic Sobolev inequality LS(C, D) applied to f q/2 (q > 0)

yields, again by the change of variables formula,

Cq 2
Entμ f q ≤ f q−2 (f )dμ + D f q dμ. (5.2.4)
2 E E
On the basis of these three observations, the proof may easily be completed. For
f in Aconst+
0 as before, consider

= (t, q) = (Pt f )q dμ, t ≥ 0, q > 1.
E
By the preceding, the LS(C, D) inequality tells us that
C D 1
∂q ≤ − ∂t + + log .
2(q − 1) q q
For the given expressions of q(t) and M(t), it follows immediately from this in-
equality that H (t) = q(t)
1
log( (t, q(t))) − M(t) is decreasing in t ≥ 0. Hypercon-
tractivity is then simply the inequality H (t) ≤ H (0).
This shows that the logarithmic Sobolev inequality LS(C, D) implies hypercon-
tractivity for any p > 1. Conversely, start from the hypercontractivity inequality
Pt f q(t) ≤ eM(t) f p
for some p > 1. Taking the derivative at t = 0 gives back (5.2.4) with q = p which
amounts to the logarithmic Sobolev inequality LS(C, D) after changing of f into
f 2/p . The proof is complete.
Remark 5.2.4 The above proof may easily be modified to yield a reverse hypercon-
tractivity property. Indeed, choosing a tight logarithmic Sobolev inequality LS(C)
for simplicity, LS(C) holds if and only if for every t > 0 and every strictly positive
measurable function f : E → R,
Pt f q ≥ f p
q−1
for (some or any) −∞ < q 0 and not for any t > 0.
This is the content of the following statement. Of course, the exact correspondence
between constants is then lost.
Theorem 5.2.5 Given a Markov Triple (E, μ, ) with Markov semigroup

P = (Pt )t≥0 , if for some t0 > 0 and some q > 2, there exists a constant M > 0
such that
Pt0 f q ≤ M f 2
for all f ∈ L2 (μ), then a logarithmic Sobolev inequality LS(C, D) holds (with
2q log M
C = 2q t0
q−2 and D = q−2 ).
This theorem may be shown to follow from a general relationship between en-
tropy and energy along the semigroup, which is an analogue of (4.2.3), p. 184.
Proposition 5.2.6 (Cattiaux’s inequality) For every function f in D(E) and every
t ≥ 0,

f 2 log f 2 dμ ≤ 2t E(f ) + f 2 log(Pt f )2 dμ.
E E
More generally, for all functions f, g in D(E) and every t ≥ 0,

f 2 log g 2 dμ ≤ 2t E(f ) + f 2 log(Pt g)2 dμ.
E E
Proof We only establish the first assertion, the second one being proved in the same
way. As usual, it is convenient to work with f ∈ Aconst+
0 and to extend to D(E) by
density. As in many of
the earlier proofs, consider a functional along the semigroup
given here by (t) = E f 2 log Pt f dμ, t ≥ 0. Then, by the integration by parts and
diffusion properties,

LPt f 2f f2
(t) = f 2
dμ = − (f, Pt f ) + (Pt f ) dμ.
E Pt f E Pt f (Pt f )2
Now, by the Cauchy-Schwarz inequality for the quadratic form ,
2f f2
− (f, Pt f ) + (Pt f ) ≥ − (f )
Pt f (Pt f )2

so that (t) ≥ − E (f )dμ = − E(f ) for every t. Integrating from 0 to t yields
the result.
Proof of Theorem 5.2.5 Let q > 2 and f ∈ D(E) with f 2 = 1. By Proposi-

tion 5.2.6, for every t ≥ 0,

f log f dμ ≤ 2t E(f ) + f 2 log(Pt f )2 dμ
2 2
E E

2
= 2t E(f ) + f 2 log |Pt f |q dμ.
q E
By Jensen’s inequality with respect to the probability measure f 2 dμ,

f log |Pt f | dμ ≤
2 q
f log f dμ + log |Pt f |q dμ.
2 2
E E E
Therefore

2
1− f 2 log f 2 dμ ≤ 2t E(f ) + log Pt f 2q ,
q E
from which Theorem 5.2.5 immediately follows.
Note that Theorem 5.2.5 (with possibly different bounds) may be shown to follow
alternatively from Lemma 4.2.6, p. 184, and Proposition 5.1.8 (use by duality that
Pt f 2 ≤ Mf q ∗ where q ∗ is the conjugate exponent of q).
5.2.3 Tensorization of Logarithmic Sobolev Inequalities
To conclude this section, we observe that, like Poincaré inequalities (Sect. 4.3,
p. 185), logarithmic Sobolev inequalities are stable under products. This dimension-
free feature is one fundamental aspect justifying the importance of logarithmic
Sobolev inequalities in infinite-dimensional contexts. It technically allows for the
extension of logarithmic Sobolev bounds on finite-dimensional space to infinite di-
mension (such as, for example, the extension from finite-dimensional to infinite
dimensional Gaussian measures).
Proposition 5.2.7 (Stability under product) If (E1 , μ1 , 1 ) and (E2 , μ2 , 2 ) sat-

isfy logarithmic Sobolev inequalities LS(C1 , D1 ) and LS(C2 , D2 ) respectively, then
the product (E1 × E2 , μ1 ⊗ μ2 , 1 ⊕ 2 ) satisfies a logarithmic Sobolev inequality
LS(max(C1 , C2 ), D1 + D2 ).
Proof A quick proof may be provided via the equivalence with hypercontractiv-
ity (Theorem 5.2.3). Indeed, if (Pt1 )t≥0 and (Pt2 )t≥0 denote the respective Markov
semigroups, and if Pti is bounded from Lp (μ) into Lq (μ) with norm eMi for t
q−1
such that e2t/Ci ≥ p−1 , i = 1, 2, it is immediate that Pt1 ⊗ Pt2 is bounded from
Lp (μ1 ⊗ μ2 ) into Lq (μ1 ⊗ μ2 ) with norm bounded from above by eM1 +M2 for
q−1
e2t/ max(C1 ,C2 ) ≥ p−1 .
The proposition may also be proved in the same way as Proposition 4.3.1, p. 185,
with a traditional scheme which is often used when dealing with correlations in
statistical mechanics models for example. Indeed, using the notation therein,

Entμ1 ⊗μ2 f 2 = Entμ1 g 2

+ f 2 (x1 , x2 ) log f 2 (x1 , x2 )dμ2 (x2 )
E1 E2

− f 2 (x1 , x2 )dμ2 (x2 ) log f 2 (x1 , x2 )dμ2 (x2 ) dμ1 (x1 )
E2 E2

where now g ≥ 0 is defined by g 2 (x1 ) = E2 f 2 (x1 , x2 )dμ2 (x2 ), x1 ∈ E1 . Using

that E1 (g) ≤ E2 E1 (f )dμ2 , the conclusion easily follows. However, the latter re-
quires more here in the form of the change of variables formula for carré du champ
operators and the Cauchy-Schwarz inequality yielding that

1 (g)(x1 ) ≤ 1 f (x1 , x2 ) dμ2 (x2 ).
E2
5.3 Integrability of Eigenvectors

While the spectral content of Poincaré inequalities is quite clear, this is not so obvi-
ous for logarithmic Sobolev inequalities. In this short section, we draw integrability
properties of the eigenvectors of a diffusion operator L with Markov semigroup
(Pt )t≥0 from the hypercontractive Theorem 5.2.3. Provided that the spectrum is dis-
crete and that the eigenvalues go to infinity fast enough, these integrability properties
of eigenvectors are actually equivalent to hypercontractive bounds, and therefore to
logarithmic Sobolev inequalities.
We illustrate the picture in the Gaussian case first. As mentioned above, the stan-
dard Gaussian measure dμ(x) = (2π)−n/2 e−|x| /2 dx on the Borel sets of Rn will
2
be shown below (Proposition 5.5.1) to satisfy LS(1) (with respect to the standard
carré du champ operator). Therefore, by Theorem 5.2.3, the associated Ornstein-
Uhlenbeck semigroup P = (Pt )t≥0 (as presented in Sect. 2.7.1, p. 103) is hypercon-
q−1
tractive in the sense that for every 1 0 such that e2t ≥ p−1 ,
and every f ∈ L (μ),
p
Pt f q ≤ f p . (5.3.1)
Eigenfunctions of the Ornstein-Uhlenbeck operator are described by Hermite
polynomials. Recall from Sect. 2.7.1, p. 103, that the Hermite polynomials (Hk )k∈N
form an orthogonal basis of L2 (μ) on the real line. Similarly, the polynomials
H (x) = Hk1 (x1 ) · · · Hkn (xn ), x = (x1 , . . . , xn ) ∈ Rn , k1 , . . . , kn ∈ N,

5.3 Integrability of Eigenvectors 251
form an orthogonal basis of L2 (μ) on Rn . Now let (H )∈Lk be a given family of
such polynomials with the property that k1 + · · · + kn = k for some fixed k ∈ N, and
set

Q= a H .
∈Lk
Such a polynomial Q is an eigenvector of the Ornstein-Uhlenbeck operator with

eigenvalue k, that is −LQ = kQ. Hence, the action of the Ornstein-Uhlenbeck semi-
group (Pt )t≥0 on Q is given by Pt Q = e−kt Q, t ≥ 0. In particular, by (5.3.1), for
every t > 0,
e−kt Qq = Pt Qq ≤ Q2
where q = 1 + e2t . Hence, for every q > 2,
Qq ≤ (q − 1)k Q2 . (5.3.2)

This moment inequality furthermore provides sharp exponential tails on the dis-
tribution of Q (μ(|Q| ≥ r) is bounded from above by e−c r as r → ∞). When
2/k
considering infinitely many coordinates xi , this inequality extends to convergent (in

L2 (μ)) sums Q, and the resulting tail inequalities are not so easy to obtain by other
means.
These integrability properties of eigenvectors are not particularly tied to the ex-
ample of the Ornstein-Uhlenbeck semigroup and the Hermite polynomials. They
hold for any eigenvector of the generator under the hypercontractive estimate of
Theorem 5.2.3. In fact, in contrast to the family of Sobolev inequalities studied in
the next chapter, logarithmic Sobolev inequalities do not in general say anything
about the existence of eigenvectors (or the existence of any point in the spectrum),
unless extra properties such as that of Persson are imposed (cf. Sect. 4.10.3, p. 227).
There are indeed examples, in infinite dimension, where hypercontractivity holds
but the spectrum of L is not discrete.
However, when there are eigenvalues with associated eigenvectors, logarithmic
Sobolev inequalities do produce Lp -bounds. Indeed, as in the Gaussian example,
if h is an eigenvector of −L associated with eigenvalue λ, under the logarithmic
Sobolev inequality LS(C, D), and thus hypercontractivity (Theorem 5.2.3), for any
q >2
D( 21 − q1 )
hq ≤ e (q − 1)λC/2 h2 .
We simply apply Theorem 5.2.3 with p = 2 to h, for which Pt h = e−λt h, choosing
t according to the rule q − 1 = e2t/C . This shows in particular that any eigenvector
is in every Lq (μ)-space.
The preceding admits a somewhat surprising converse. Assume that the operator
L admits a point spectral decomposition in the sense that any f ∈ L2 (μ) may be
decomposed into an orthogonal sequence

f= k (f )
k∈N
where −L k (f ) = λk k (f ) for every k ∈ N. (Comparing with the discrete spec-

trum hypothesis of Definition A.4.3, p. 480, it should be mentioned that this point
spectral decomposition does not assume that the corresponding eigenspaces are
finite-dimensional.) Assume moreover that the spectral
values λk , k ∈ N, increase
fast enough to infinity such that, for some ε > 0, k∈N e−ελk < ∞. In this context,
as soon as, for some q > 2 and some a, C > 0,

k (f ) ≤ C eaλk f 2 (5.3.3)
q
for every k ∈ N, then, by the triangle inequality,

−λ t

Pt f q = k (f )
≤ C f 2 e(a−t)λk ≤ C1 f 2
e k
k∈N q k∈N
for every t ≥ a + ε. Therefore, for such a large t, the operator Pt is bounded

from L2 (μ) into Lq (μ), and, by Theorem 5.2.5, a logarithmic Sobolev inequal-
ity LS(C, D) holds, and hence we have full hypercontractivity. Notice that (5.3.3)
is nothing else than the similar upper bound on the eigenvectors in the associated
eigenspaces. In this setting, hypercontractivity appears somewhat surprisingly as a
consequence of spectral analytic bounds on the generator (via information on the
size of the eigenvectors).
5.4 Logarithmic Sobolev Inequalities and Exponential

Integrability
As shown in Sect. 4.4.2, p. 190, Lipschitz functions are exponentially integrable
whenever μ satisfies a Poincaré inequality P (C). We investigate in this section the
corresponding property under a logarithmic Sobolev inequality and the associated
Gaussian measure concentration.
5.4.1 Exponential Integrability
Exponential integrability under a logarithmic Sobolev inequality is achieved

through a method known as the Herbst argument. For a given Lipschitz function
1/2
f ∈ D(E) ∪ A, recall its Lipschitz semi-norm f Lip = (f )∞ . A function
f ∈ D(E) ∪ A is said to be 1-Lipschitz if f Lip ≤ 1.
Proposition 5.4.1 (Herbst’s argument) If μ satisfies a logarithmic Sobolev in-

equality LS(C) for some C > 0, then for every 1-Lipschitz function f and every
σ 2 < C1 ,

2 2
eσ f /2 dμ < ∞.
E
5.4 Logarithmic Sobolev Inequalities and Exponential Integrability 253
More precisely, any 1-Lipschitz function f is integrable and for every s ∈ R,

2
esf dμ ≤ es E f dμ+Cs /2 . (5.4.1)
E
At the expense of further technical arguments not developed here, the exponen-

2 2
tial integrability result E eσ f /2 dμ < ∞ for every σ 2 < C1 still holds under a
defective logarithmic Sobolev inequality LS(C, D).
With respect to Poincaré inequalities (Proposition 4.4.2, p. 190), squares of Lip-
schitz functions are here exponentially integrable (and not just the Lipschitz func-
tions themselves). In particular, the exponential measure, which satisfies a Poincaré
inequality, cannot satisfy a logarithmic Sobolev inequality (for the standard carré
du champ operator). Note that the integrability level in Proposition 5.4.1 is optimal
since the standard Gaussian measure satisfies LS(1) (cf. Proposition 5.5.1 below),
and (5.4.1) is sharp in this example with f linear.
Proof The proof is perhaps even simpler than the corresponding one for Poincaré
inequalities. We start with (5.4.1), and similarly restrict ourselves to bounded Lips-
chitz functions in D(E) (or A) the general case being reached through truncation and
approximation (cf. Remark 4.4.3, p. 190).
Therefore, given a bounded 1-Lipschitz
function f ∈ D(E), set for s ∈ R, Z(s) = E esf dμ. The aim is to apply LS(C) to
esf/2 for every s. Towards this goal, observe that

Z (s) = f esf dμ
E
while

Entμ esf = s f esf dμ − Z(s) log Z(s) = sZ (s) − Z(s) log Z(s).
E
On the other hand, since (f ) ≤ 1,

s2 s2
E esf/2 = esf (f )dμ ≤ Z(s).
4 E 4
The logarithmic Sobolev inequality LS(C) then expresses the differential inequality
in s ∈ R,
Cs 2
s Z ≤ Z log Z + Z.
2
It remains to integrate this inequality. To this end, set F (s) = 1
log Z(s) (with

s
F (0) = E f dμ) so that F (s) ≤ C2 for every s. It follows that

Cs
F (s) ≤ f dμ +
E 2
which amounts to the control (5.4.1) of the Laplace transform.
Integrating (5.4.1) along the centered Gaussian measure with variance σ 2 in the
s variable yields, by Fubini’s Theorem, that
2
2 f 2 /2 1 σ2
eσ dμ ≤ √ exp f dμ
E 1 − Cσ 2 2(1 − Cσ 2 ) E
1
for every σ 2 < C. Hence the first claim of the proposition holds and the proof is
complete.
Proposition 5.4.1 admits a simple and useful variation in the form of mo-
ment estimates. Namely, if R(p) = f 2p , p ≥ 2, for a function f ∈ Lp (μ) with
p/2 dμ < ∞, the derivative R (p) of R in p is given by

E (f )
2
2
R(p)1−(p/2) Entμ f p .
p
By the logarithmic Sobolev inequality LS(C) and Hölder’s inequality,

2/p
R (p) ≤ C (f )p/2 dμ .
E
By integration and monotonicity, the following statement holds.
Proposition 5.4.2 (Moment bounds) If μ satisfies a logarithmic Sobolev inequality

LS(C) for some C > 0, then for every p ≥ 2 and every f ∈ Lp (μ),
2/p
f 2p ≤ f 22 + C(p − 2) p/2
(f ) dμ .
E
If f is Lipschitz, then
f 2p ≤ f 22 + C(p − 2)f 2Lip
for every p ≥ 2. By a series expansion, the growth in p amounts to the same level
of integrability as that in Proposition 5.4.1.
The next statement is an analogue of Proposition 4.4.4, p. 192, in the setting of
logarithmic Sobolev inequalities. The proof is entirely similar.
Proposition 5.4.3 (Contraction principle) If (μ, ) satisfies a logarithmic Sobolev

inequality LS(C, D) on a state space E, and if F : E → R is Lipschitz with respect
to with F Lip ≤ 1, then the image measure μF of μ under F satisfies a logarith-
mic Sobolev inequality LS(C, D) with respect to the usual carré du champ operator
on the real line.
5.4 Logarithmic Sobolev Inequalities and Exponential Integrability 255
5.4.2 Gaussian Measure Concentration
As in Sect. 4.4.3, p. 192, for Poincaré inequalities, Proposition 5.4.1 may be refor-
mulated in the form of a Gaussian measure concentration property for the measure
μ satisfying a logarithmic Sobolev inequality LS(C) for some C > 0. Indeed, for
every Lipschitz function f and every r ≥ 0,

−r 2 /2Cf 2Lip
μ f≥ f dμ + r ≤ e . (5.4.2)
E
This is an immediate application of (5.4.1) together with Markov’s exponential in-

equality and homogeneity. The numerical constant 2 is sharp. A concentration in-
equality around the mean is easily obtained together with the same inequality for
−f yielding

−r 2 /2Cf 2Lip

μ f − f dμ ≥ r ≤ 2 e , r ≥ 0.
E
These Gaussian concentration inequalities have of course to be compared to the

corresponding exponential concentration inequalities (4.4.6), p. 192, under Poincaré
inequalities.
Again as in Sect. 4.4.3, p. 192, as a consequence of the tensorization Proposi-
tion 5.2.7, if μ on E satisfies a logarithmic Sobolev inequality LS(C) with respect
to , then the product measure μ⊗n on E n satisfies LS(C) with respect to the prod-
uct carré du champ operator, independently of the dimension n of the state space
E n . Tail inequalities for Lipschitz functions of a sample Z = (Z1 , . . . , Zn ) of inde-
pendent random variables with common distribution μ on R satisfying LS(C) may
then be deduced from Proposition 5.4.1. For example, if f : R → R, f Lip ≤ 1,
and
1
F (x) = √ f (x1 ) + · · · + f (xn ) , x = (x1 , . . . , xn ) ∈ Rn ,
n
then for every r ≥ 0,

P F (Z) ≥ E F (Z) + r ≤ e−r /2C .
2
In particular, these dimension-free Gaussian tail inequalities suitably describe the

central limit theorem. More general Lipschitz functions as well as non-identically
distributed samples may be considered similarly.
Remark 5.4.4 In analogy with Remark 4.4.5, p. 193, in the context of Poincaré
inequalities, one may wonder how far the concentration bounds of the Herbst ar-
gument (Proposition 5.4.1 and (5.4.2)) are from the logarithmic Sobolev inequal-
ity LS(C) (or LS(C, D)). Again, the Muckenhoupt characterization below (The-
orem 5.4.5) indicates that they are not sufficient in general to entail a logarithmic
Sobolev inequality. However, under a curvature condition CD(0, ∞), they do imply
a logarithmic Sobolev inequality with moreover proportional constants. This result
will be presented in Sect. 8.7, p. 425, together with the corresponding result for
Poincaré inequalities on the basis of the local heat kernel inequalities of Sect. 5.5
below (in particular the reverse forms).
5.4.3 Muckenhoupt’s Criterion
We conclude this section with an analogue of the Muckenhoupt criterion for log-
arithmic Sobolev inequalities. As for Poincaré inequalities (Sect. 4.5.1, p. 194),
it is possible to characterize measures μ on the real line R satisfying a logarith-
mic Sobolev inequality LS(C) with respect to the usual carré du champ operator
(f ) = f 2 . The next statement is an analogue of Theorem 4.5.1, p. 194, with a
similar although more involved proof not presented here. As for the correspond-
ing equivalence for Poincaré inequalities, this result appears as the one-dimensional
version of more general relationships between capacities and measures discussed in
Chap. 8.
Theorem 5.4.5 (Muckenhoupt’s criterion) Let μ be a probability measure on the

Borel sets of R with density p with respect to the Lebesgue measure, and let m be a
median for μ. Set
x
1 1
B+ = sup μ [x, +∞) log dt
x>m μ([x, +∞)) m p(t)
and

1 m 1
B− = sup μ (−∞, x] log dt.
x<m μ((−∞, x]) x p(t)
Then, a necessary and sufficient condition in order that μ satisfies a loga-
rithmic Sobolev inequality LS(C) is that B = max(B+ , B− ) < ∞. Moreover,
cB ≤ C ≤ c B for numerical c, c > 0 where C is the logarithmic Sobolev constant
of μ.
It is a good exercise to verify that the probability measure on R with density

cα e−|x| with respect to the Lebesgue measure satisfies the Muckenhoupt criterion
α
for the logarithmic Sobolev inequality if and only if α ≥ 2. Recall that the corre-
sponding criterion for the Poincaré inequality yields in this case α ≥ 1, thus distin-
guishing clearly the two families of inequalities. As a further example of interest, the
spectral decomposition of the Laguerre operator in Sect. 2.7.3, p. 111, shows that
the exponential measure μ on R+ has spectral gap 1 for the carré du champ operator
(f ) = xf 2 . Since the Laguerre operator is of curvature CD( 12 , ∞), according to
Proposition 5.7.1 below, μ satisfies a logarithmic Sobolev inequality with
√ constant
C = 2. This constant is optimal since the 1-Lipschitz map f (x) = 2 x (with re-

2 2
spect to ) is such that R+ eσ f /2 dμ < ∞ for every σ 2 < 12 and nothing better
(Proposition 5.4.1).
5.5 Local Logarithmic Sobolev Inequalities 257
5.5 Local Logarithmic Sobolev Inequalities
Following the treatment of Poincaré inequalities in Sect. 4.7, p. 206, of the pre-
vious chapter, this section investigates logarithmic Sobolev inequalities under the
semigroup P = (Pt )t≥0 , that is with respect to the heat kernel measures of the rep-
resentation (1.2.4), p. 12,

Pt f (x) = f (y) pt (x, dy), t ≥ 0, x ∈ E.
E
The approach further develops the main principle of heat flow monotonicity using
the Duhamel interpolation formula, here in the context of entropy.
Before addressing the general framework, we first consider the Ornstein-
Uhlenbeck example leading to the logarithmic Sobolev inequalities for Gaussian
measure. The section closes with a version of hypercontactivity for heat kernel
measures and some heat kernel bounds. The local heat kernel inequalities will be
presented and established for functions in the algebra A0 of the Markov Triple
(E, μ, ), and we technically work with the class Aconst+
0 .
5.5.1 Logarithmic Sobolev Inequalities for Gaussian Kernels
We commence our analysis of local logarithmic Sobolev inequalities, again by start-

ing with the example of the Ornstein-Uhlenbeck semigroup P = (Pt )t≥0 on Rn with
generator Lf = f − x · ∇f and carré du champ operator (f ) = |∇f |2 on an
algebra A0 of smooth functions (cf. Sect. 2.7.1, p. 103). Recall from (2.7.5), p. 104,
that ∇Pt f = e−t Pt (∇f ) so that, for every f ∈ A0 , and every t ≥ 0,

|∇Pt f | ≤ e−t Pt |∇f | . (5.5.1)
As for the local Poincaré inequalities in Sect. 4.7, p. 206, consider now, for t > 0
and f ∈ Aconst+
0 ,

(s) = Ps ψ(Pt−s f ) , s ∈ [0, t],
where ψ(r) = r log r, r ∈ R+ . Then, setting g = Pt−s f , by the heat equation and
the change of variables formula,

(g)
(s) = Ps L ψ(g) − ψ (g) Lg = Ps ψ (g) (g) = Ps .
g
Now, by (5.5.1),
2
(g) = |∇Pt−s f |2 ≤ e−2(t−s) Pt−s |∇f | .
Since pt−s (x, dy) is a probability measure, by the Cauchy-Schwarz inequality, for
every h,
2
h
(Pt−s h)2 ≤ Pt−s f Pt−s .
f
With h = |∇f |, this amounts to

(Pt−s (|∇f |))2 |∇f |2 (f )
≤ Pt−s = Pt−s .
Pt−s f f f
As a consequence,

−2(t−s) (f ) −2(t−s) (f )
(s) ≤ e Ps Pt−s =e Pt .
f f
By definition of (s), s ∈ [0, t], it follows that, for every t ≥ 0,

1 − e−2t (f )
Pt (f log f ) − Pt f log Pt f ≤ Pt . (5.5.2)
2 f
It should be emphasized that with respect to the corresponding Poincaré inequal-
ity (4.7.2), p. 207, the logarithmic Sobolev inequality (5.5.2) makes use of the com-
mutation (5.5.1) rather than only (4.7.1), p. 207 (strong gradient bound versus gra-
dient bound in the terminology of Chap. 3). This aspect will be further developed in
Theorem 5.5.2 below.
The preceding inequality is thus a (tight) logarithmic Sobolev inequality
LS(1 − e−2t ) for the heat kernel measures pt (x, dy), with a constant independent
of the initial point x (implicit throughout the argument), which converges to 1 as
t → ∞. Since the kernels pt (x, dy) converge to the standard Gaussian measure μ,
as seen from (2.7.3), p. 104, we therefore conclude that the latter satisfies LS(1).
Proposition 5.5.1 (Logarithmic Sobolev inequality for the Gaussian measure) The
standard Gaussian measure μ on the Borel sets of Rn satisfies LS(1). In other
words, for every function f : Rn → R in the Dirichlet space of the standard carré
du champ operator (f ) = |∇f |2 ,

Entμ f 2 ≤ 2 E(f ) = 2 |∇f |2 dμ.
Rn
This proposition states the logarithmic Sobolev inequality for the centered Gaus-
sian measure μ on Rn with identity as covariance matrix (sometimes called the
Gaussian logarithmic Sobolev inequality). If μ is a centered Gaussian measure on
Rn with covariance matrix Q, a simple change of variables shows that for every
smooth function f on Rn ,

2
Entμ f ≤ 2 Q∇f · ∇f dμ.
Rn
The constant in the Gaussian logarithmic Sobolev inequality of Proposition 5.5.1

is sharp. This may be shown in various ways. For example, by comparison with
the Poincaré inequality (Proposition 5.1.3) and the first eigenvector (with eigen-
value 1). It may also be shown that functions f (x) = eax (on the line) satisfy equal-
ity in LS(1) (and are actually the only extremal functions). Alternatively, Proposi-
tion 5.4.1 on the exponential integrability of Lipschitz functions implies that C = 1
cannot be improved by the simple example of f (x) = x.
Another approach to optimality would be via the equivalence with hypercontrac-
tivity of Theorem 5.2.3, namely that for every t > 0, Pt is a contraction from L2 (μ)
into Lq(t) (μ) for q(t) = 1 + e2t . Since the Ornstein-Uhlenbeck semigroup is one
of the rare examples for which the semigroup and its kernel are explicitly known
(cf. Sect. 2.7.1, p. 103), the optimality of this embedding may be established di-
rectly. It is however a bit delicate. Much easier is to observe that if q > q(t), then
Pt is unbounded from L2 (μ) into Lq (μ), so that the norm of Pt as an operator from
L2 (μ) into Lq (μ) is equal to 1 for q ∈ [2, q(t)] and to +∞ for q > q(t).
As for local Poincaré inequalities, the preceding logarithmic Sobolev inequal-
ity (5.5.2) for Pt may be reversed in the form of
e2t − 1 (Pt f )
Pt (f log f ) − Pt f log Pt f ≥ · (5.5.3)
2 Pt f
for every f positive in A0 (where the right-hand side is understood, according to
Remark 5.1.2, as limε0 (P tf)
Pt f +ε ).
In a further similarity with the chapter on Poincaré inequalities, the same families
of inequalities may be established for the Euclidean (Brownian) heat semigroup
P = (Pt )t≥0 on Rn . Recall that in this case |∇Pt f | ≤ Pt (|∇f |). The corresponding
local inequalities for the heat kernel measures then take the form

(Pt f ) (f )
t ≤ Pt (f log f ) − Pt log Pt f ≤ t Pt .
Pt f f
Since at time t = 12 , the distribution of the heat semigroup Pt is the standard Gaus-
sian measure, the first inequality recovers Proposition 5.5.1.
5.5.2 Local Logarithmic Sobolev Inequalities
The preceding analysis may be performed for any diffusion Markov semigroup
P = (Pt )t≥0 under the curvature condition CD(ρ, ∞) of Definition 3.3.14, p. 159.
With respect to Theorem 4.7.2, p. 209, dealing with local Poincaré inequalities, the
condition CD(ρ, ∞) is enriched by several new (stronger) equivalences, in partic-
ular the reinforced curvature condition (ii) and the strong gradient bound (iii).
Theorem 5.5.2 (Local logarithmic Sobolev inequalities) Let (E, μ, ) be a

Markov Triple with semigroup P = (Pt )t≥0 . The following assertions are equiva-
lent.

(ii) For every function f in A0 ,

4 (f ) 2 (f ) − ρ (f ) ≥ (f ) .
(iii) For every function f in A0 , and every t ≥ 0,

(Pt f ) ≤ e−ρt Pt (f ) . (5.5.4)
(iv) For every positive function f in A0 , and every t ≥ 0,

1 − e−2ρt (f )
Pt (f log f ) − Pt f log Pt f ≤ Pt . (5.5.5)
2ρ f
(v) For every positive function f in A0 , and every t ≥ 0,
e2ρt − 1 (Pt f )
Pt (f log f ) − Pt f log Pt f ≥ . (5.5.6)
2ρ Pt f
−2ρt
When ρ = 0, the quantities 1−e2ρ and e 2ρ−1 have to be replaced by t (as is
2ρt
the case for the heat semigroup on Rn of curvature CD(0, ∞)). Also recall from
Remark 5.1.2 that the Fisher-type expressions such as Pt ( (f )
f ) are understood as
limε0 Pt ( (f )
f +ε ). The inequalities in Theorem 5.5.2 extend to positive functions
f in A following the various extension procedures described in Sect. 3.3, p. 151.
Moreover, as for the corresponding Poincaré inequality (4.7.6), p. 210, the reverse
logarithmic Sobolev inequality (5.5.6) (for t > 0) extends to measurable positive
functions.
Proof We refer to Sect. C.6, p. 513, in Appendix C for the equivalence between
the CD(ρ, ∞) condition and (ii) already used in Theorem 3.3.18, p. 163, and at the
heart of the (strong) gradient bound (iii). As explained there, this apparent reinforce-
ment of the curvature condition only relies on the chain rule formulas for L and ,
that is, the diffusion property. The proof from the gradient bound (iii) to the local
logarithmic Sobolev inequality (iv) and its reverse form (v) is exactly the same as the
one presented above for the Ornstein-Uhlenbeck and Brownian semigroups (with
the only modification being the constant ρ). To complete the circle of equivalences,
proceed as usual to the asymptotics as t = 0 of the local logarithmic Sobolev in-
equality or the reverse inequality yielding the curvature condition CD(ρ, ∞). More
simply, use that these (local) logarithmic Sobolev inequalities imply the correspond-
ing (local) Poincaré inequalities (by Theorem 5.1.3) and the corresponding result for
Poincaré inequalities (Theorem 4.7.2, p. 209). The proof is therefore complete.
As in the corresponding Remark 4.7.3, p. 211, on local Poincaré inequalities, the

implications from (iii) to (iv) and (v) of Theorem 5.5.2 work similarly if (iii) only
holds up to some constant K ≥ 1, that is, for every t ≥ 0, and all functions f in A0 ,
√
(Pt f ) ≤ K e−ρt Pt (f ) .
The proof is entirely similar, and (5.5.5) for example, then simply reads

K(1 − e−2ρt ) (f )
Pt (f log f ) − Pt f log Pt f ≤ Pt .
2ρ f
Similarly, the analogue of Remark 4.7.4, p. 211, indicates that all the assertions
of Theorem 5.5.2 are also equivalent to saying that there exist t0 > 0 and a function
c(t) = t − ρ t 2 + o(t 2 ) such that, for any t ∈ (0, t0 ) and any positive f in A0 ,

(f )
Pt (f log f ) − Pt f log Pt f ≤ c(t) Pt ,
f
respectively
(Pt f )
Pt (f log f ) − Pt f log Pt f ≥ c(t) .
Pt f
While the proof of Theorem 5.5.2 is based on the √ reinforced curvature condi-
tion in terms of the commutation (5.5.4) of Pt and , the standard definition of
CD(ρ, ∞) is actually enough to reach the local logarithmic Sobolev inequalities of
Theorem 5.5.2 (as for Poincaré inequalities). It is worth emphasizing the argument
here since the principle will be used again in subsequent developments. The follow-
ing proposition is the key technical step towards this goal, illustrating the role of the
2 operator in the heat flow monotonicity principle. It will allow us later to reach
stronger conclusions such as Sobolev-type inequalities as well as Harnack-type in-
equalities.
Proposition 5.5.3 Let f be positive in A0 , and let for t > 0,
(s) = Ps (Pt−s f log Pt−s f ), s ∈ [0, t].
Then,

(s) = Ps Pt−s f (log Pt−s f )
and

(s) = 2Ps Pt−s f 2 (log Pt−s f ) .
Now, on the basis of these identities, under the curvature condition CD(ρ, ∞),
that is 2 (f ) ≥ ρ (f ) for every f ∈ A0 (or A), it follows that ≥ 2ρ . The
local inequalities of Theorem 5.5.2 are then easily derived in this alternative way.
Proof As usual, we work with f ∈ Aconst+

0 . The first derivative is a fairly general
property, already used in the proof of Theorem 5.5.2. Indeed, whenever ψ : R → R
is smooth enough, setting (s) = Ps (ψ(Pt−s f )) and g = Pt−s f , s ∈ [0, t], then

(s) = Ps Lψ(g) − ψ (g) Lg = Ps ψ (g) (g)
by the diffusion property. For ψ(r) = r log r, r > 0, ψ = 1

r and 1
g (g) =
g (log g), which yields the claim.
The second derivative is more specific to the function ψ(r) = r log r. We have,
with = ψ , again by the heat equation,

(s) = Ps L (g) (g) − (g) Lg (g) − 2(g) (g, Lg) .
It remains to consider the quantity

E(g) = L (g) (g) − (g) Lg (g) − 2(g) (g, Lg)
which by the change of variables formula (3.1.9), p. 124, for L may be written as

E(g) = 2(g) 2 (g) + 2 (g) g, (g) + (g) (g)2 .
For the choice of ψ(r) = r log r, and thus (r) = 1r , this expression may be directly
compared to the change of variables formula (3.3.2), p. 158, for the 2 operator
(consequence of the diffusion property of L), that is

2 ψ1 (g) = ψ1 (g) 2 (g) + ψ1 (g)ψ1 (g) g, (g) + ψ1 (g) (g)2 .
2 2
With ψ1 (r) = log r, E(g) = 2g 2 (log g) from which the conclusion follows.
Remark 5.5.4 The functions ψ(r) = r 2 (r ∈ R) and ψ(r) = r log r (r ∈ R+ ) are the
only ones for which the second derivative of (s) = Ps (ψ(Pt−s f )), s ∈ [0, t], takes
such a simple form. For ψ(r) = r 2 , even the subsequent derivatives have nice ex-
pressions yielding the iterated Gamma operators 3 , 4 . . . (constructed by the same
rule as 2 ). This is no longer the case for the entropy ψ(r) = r log r where the third
derivative already has a complicated form. Nevertheless, analogues of the preceding
local inequalities may be obtained by replacing r 2 or r log r by r α , 1 < α < 2 (or
even more general functions as in Proposition 7.6.1, p. 383, below). Such inequal-
ities (modified a little) are of interest for heat kernel decays as well as tail bounds
between exponential and Gaussian behaviour (cf. Chap. 7).
5.5.3 Local Hypercontractivity
Under a logarithmic Sobolev inequality, the Markov semigroup P = (Pt )t≥0 is hy-
percontractive (Theorem 5.2.3). The same argument may be developed under a lo-
cal logarithmic Sobolev inequality yielding hypercontractive characterizations of
the curvature bound CD(ρ, ∞). One advantage of this description is that it ex-
tends to positive measurable functions, and is therefore stable under various kinds
of semigroup convergences. Furthermore, as shown by its proof, Theorem 5.5.5 pro-
vides a semigroup characterization of the curvature condition CD(ρ, ∞) without
any derivation argument.
Theorem 5.5.5 (Local hypercontractivity) Let (E, μ, ) be a Markov Triple with

semigroup P = (Pt )t≥0 . The following assertions are equivalent.

(ii) For all t > 0, 0 < s ≤ t and 1 < p < q < ∞ such that
q −1 e2ρt − 1
= 2ρs , (5.5.7)
p−1 e −1
1/q p 1/p
Ps (Pt−s f )q ≤ Pt f (5.5.8)
for all positive measurable functions f .
1−e−2ρt e2ρt −1
As usual, when ρ = 0, then 2ρ has to be replaced by t and e2ρs −1
by st .
Proof Theorem 5.5.2 indicates that (i) is equivalent to the local logarithmic Sobolev
inequality (5.5.5). Assume then that (5.5.5) holds. Let f be in Aconst+
0 and consider
1/q(s)
(s) = Ps (Pt−s f )q(s) , s ∈ [0, t],
where q : [0, t] → (1, ∞). If g = Pt−s f ,
q 2 (q − 1) q−2
q−1 (s) = Entps g q +
Ps g (g)
q

1 − e−2ρs q −1
≤ q2 +
Ps g q−2 (g)
2ρ q
by the local logarithmic Sobolev inequality (5.5.5). Now, if q > 1 satisfies
1−e−2ρs
2ρ + q−1
q = 0, that is,
q(s) − 1 e2ρt − 1
= ,
q(t) − 1 e2ρs − 1
then is increasing which amounts to (5.5.8) (with q(s) = q and q(t) = p).
Conversely, assuming (ii), a first order Taylor expansion (with t fixed) as p = 2,
−2ρt
q = 2(1+ε) and s = t (1−αε)+o(ε) with α = (1−eρt ) as ε → 0 (for which (5.5.7)
is satisfied in the limit), yields the local logarithmic Sobolev inequality (5.5.5) which
is equivalent to CD(ρ, ∞). Inequality (5.5.8) for functions in Aconst+ 0 extends by
density to positive measurable functions. The proof of Theorem 5.5.5 is complete.
The same proof shows that

p 1/p 1/q
Pt f ≤ Ps (Pt−s f )q ,
t > 0, 0 < s ≤ t, −∞ < q < p < 1 satisfying (5.5.7) and f strictly positive, may be
added to the equivalences of Theorem 5.5.5.
It should be observed that Theorem 5.5.5 is not a direct consequence of the hy-
percontractivity Theorem 5.2.3 applied to the local logarithmic Sobolev inequalities,
which would correspond to the semigroup with same carré du champ operator
and reversible measure pt (x, dy). On the other hand, Theorem 5.5.5 applied with
s = t − u yields as t → ∞ with u fixed the hypercontractive bound Pu f q ≤ f p
of Theorem 5.2.3.
This comment implies that (5.5.8) of Theorem 5.5.5 is optimal for the Ornstein-
Uhlenbeck semigroup in the sense that ρ = 1 is the optimal parameter such
that (5.5.8) holds. For the Brownian semigroup (ρ = 0), Theorem 6.7.7, p. 304,
will indicate how this inequality becomes optimal with Gaussian extremal functions
and parameters depending upon the dimension.
5.5.4 Some Heat Kernel Bounds
The application discussed here concerns some cheap but rough heat kernel bounds
which may be produced from the exponential integrability of heat kernel measures.
Assume that the Markov semigroup P = (Pt )t≥0 is such that Pt , t > 0, admits a
density kernel pt (x, y) with respect to the invariant measure μ as

Pt f (x) = f (y) pt (x, y)dμ(y), t > 0, x ∈ E,
E
where pt (x, y) is symmetric since the measure μ is reversible (Definition 1.2.4,

p. 14).
By the Chapman-Kolmogorov equation (1.3.2), p. 17, for every t > 0 and
(x, y) ∈ E × E,

p2t (x, y) = pt (x, z) pt (z, y)dμ(z).
E
Then, for c(t) > 0 to be specified below, and with d the intrinsic distance on
(E, μ, ) (Sect. 3.3.7, p. 166),

−d(x,y)2 /c(t) 2 2
p2t (x, y) ≤ e pt (x, z) e2d(x,z) /c(t) pt (z, y)e2d(z,y) /c(t) dμ(z).
E
By the Cauchy-Schwarz inequality, and symmetry,

p2t (x, y) ≤ e−d(x,y) /c(t) pt (x, z)2 e4d(x,z) /c(t) dμ(z).
2 2
Assume now a uniform bound on the kernel pt (x, z) ≤ K(t) (such bounds will
be discussed later in Chaps. 6 and 7). Furthermore, assume that we are in a setting in
which the local logarithmic Sobolev inequalities (5.5.5) of Theorem 5.5.2 hold for
the measures pt (·, z)dμ(z). Therefore, together with the exponential integrability
result for the Lipschitz distance function (Proposition 5.4.1), there exists a c(t) > 0
such that

2
pt (x, z) e4d(x,z) /c(t) dμ(z) < ∞
E
5.6 Infinite-Dimensional Harnack Inequalities 265
and the bound is uniform in x. As a consequence of this discussion,
p2t (x, y) ≤ K (t) e−d(x,y)

2 /c(t)
(5.5.9)
where K (t) > 0 only depends on t (and the choice of c(t) in accordance with
geometric features). This bound emphasizes, at a mild level, the standard Gaussian
bounds which may be expected on heat kernels. Such bounds will be addressed later
in Chap. 7. It should be added that the only features used in the previous argument
are the triangle inequality for the distance function d and the fact that z → d(x, z)
is 1-Lipschitz.
5.6 Infinite-Dimensional Harnack Inequalities
This section is concerned with Harnack-type inequalities under the curvature condi-
tion CD(ρ, ∞) and with various applications of the heat kernel logarithmic Sobolev
inequalities of the previous section.
The main result is a Harnack-type inequality, a consequence of the gradient
bounds under curvature conditions (cf. Theorem 3.3.18, p. 163, or (iii) of Theo-
rem 5.5.2). We state it for convenience in the setting of smooth complete connected
Riemannian manifolds (in order to freely speak of geodesics joining two points)
although the Markov Triple version may also be reached by different means (see
below). The algebra A0 then stands for the family of smooth compactly supported
functions.
Theorem 5.6.1 (Wang’s Harnack inequality) Let P = (Pt )t≥0 be a Markov semi-
group with infinitesimal generator L = g − ∇W · ∇ on a complete connected
Riemannian manifold (M, g). The following are equivalent.
(ii) For every positive measurable function f on M, every t > 0, every x, y ∈ M
and every α > 1,

α αρ d(x, y)2
(Pt f ) (x) ≤ Pt f (y) exp
α
(5.6.1)
2(α − 1)(e2ρt − 1)
where d(x, y) is the Riemannian distance from x to y.
Inequality (5.6.1) is a Harnack-type inequality for infinite-dimensional diffusion

operators (compare with the dimensional Harnack inequality (6.7.6), p. 302).
Proof We prove the implication from (i) to (ii) for α = 2 and ρ = 0. Hence, the
inequality to establish is, for a positive smooth function f ,
2
(Pt f )2 (x) ≤ Pt f 2 (y) ed /2t
where d = d(x, y). The idea of the proof relies on a variation of the usual interpo-
lation scheme by joining x and y by a constant speed geodesic (xs )s∈[0,t] and then
analyzing the function

(s) = Ps (Pt−s f )2 (xs ), s ∈ [0, t].
Setting g = Pt−s f as usual,

(s) = 2Ps |∇g|2 (xs ) + xs · ∇Ps g 2 (xs )
d
≥ 2Ps |∇g|2 − ∇Ps g 2 .
t
Under the curvature condition CD(0, ∞), by the gradient bound (iii) of Theo-
rem 5.5.2,

∇Ps g 2 ≤ Ps ∇ g 2 = 2Ps g|∇g| .
Therefore, for every s ∈ [0, t],

d d2 d2
(s) ≥ 2Ps |∇g|2 − g|∇g| ≥ − 2 Ps g 2 = − 2 (s).
t 2t 2t
Integrating this differential inequality immediately yields the conclusion. The gen-
eral case, α > 1 and ρ = 0, is similar, the only difference is that we have to choose
a geodesic (xs )s∈[0,1] with a non-constant speed.
For the converse implication from (ii) to (i), assume again for simplicity that
ρ = 0. Let α = 1 + ε and let yε be the exponential map starting from x with initial
tangent vector v. A Taylor expansion (as ε goes to 0) of (5.6.1) then yields at the
point x ∈ M,
|v|2
Pt (f log f ) − Pt f log Pt f ≥ −v · ∇Pt (f ) − Pt f .
4t
Choosing v = −2t ∇P tf
Pt f yields the logarithmic Sobolev inequality in its reverse
form (5.5.6) which is equivalent to the curvature condition CD(ρ, ∞). The proof is
therefore complete.
Remark 5.6.2 (Log-Harnack inequality) The same method applied to (s) =

Ps (exp(Pt−s log f ))(xs ) where (xs )s∈[0,t] is a geodesic between x and y, provides
a log-Harnack inequality under the condition CD(ρ, ∞),
ρ d(x, y)2
Pt (log f )(x) ≤ log Pt f (y) + (5.6.2)
2(e2ρt − 1)
for any strictly positive bounded measurable function f on M, any t > 0 and any
x, y ∈ M. This inequality may be proved alternatively by letting α → ∞ in (5.6.1)
(after the changing f into f 1/α ).
5.6 Infinite-Dimensional Harnack Inequalities 267
The latter log-Harnack inequality has a simple interesting consequence for lower
bounds on the kernel density pt (x, y) of Pt with respect to μ, which we assume
to be a probability density. Namely, applying it to f (z) = pt (x, z) + ε, z ∈ M, and
letting ε → 0, it implies that for any x, y ∈ M,

ρ d(x, y)2
pt (x, z) log pt (x, z)dμ(z) ≤ log p2t (x, y) + .
M 2(e2ρt − 1)
By Jensen’s inequality for the convex function r → r log r, r ∈ R+ , with respect
to μ, since pt (x, ·) is a probability density, the left-hand side of this inequality is
positive. Replacing t by 2t , it follows that, for all t > 0 and (x, y) ∈ M × M,

ρ d(x, y)2
pt (x, y) ≥ exp − ρt . (5.6.3)
2(e − 1)
We conclude this section with a useful observation concerning the local logarith-
mic Sobolev inequalities of Theorem 5.5.2 and their relationships with the previ-
ous Harnack inequalities. It appears that the reverse logarithmic Sobolev inequali-
ties (5.5.6) of Theorem 5.5.2 provide rather precise information on the kernels as
t → 0, which in turn get close to the Harnack inequality (5.6.1) of Theorem 5.6.1.
Consider a function f ∈ Aconst+
0 such that ε ≤ f ≤ 1 for some ε > 0. Then
Pt (f log f ) ≤ 0 and thus (5.5.6) ensures that
e2ρt − 1 (Pt f )
−Pt f log Pt f ≥ .
2ρ Pt f
It then follows by elementary means that

1 ρ
log ≤ . (5.6.4)
Pt f 2(e2ρt − 1)
This gradient bound is an analogue of (4.7.7), p. 211, in the context of local Poincaré
inequalities.
With the aim of applying (5.6.4) to Harnack inequalities, we assume for simplic-
ity that ρ = 0. Then, by definition of the intrinsic distance (3.3.9), p. 166, for every
t > 0 and x, y ∈ E,

1 1 d(x, y)
log ≤ log + .
Pt f (x) Pt f (y) 2t
After some work, it may then be shown that for each ε > 0, there exists a C(ε) > 0
such that
2
(Pt f )2 (x) ≤ C(ε) Pt f 2 (y) ed(x,y) /(2+ε)t ,
which is as close as possible to (5.6.1) (for α = 2). Employing more refined ar-
guments of an isoperimetric nature, the optimal bound will be achieved using this
principle in Sect. 8.6, p. 421.
5.7 Logarithmic Sobolev Inequalities Under

a Curvature-Dimension Condition
In this section we present an analysis corresponding to that of the Poincaré inequal-

ity of Sect. 4.8, p. 211, and investigate logarithmic Sobolev inequalities for the
invariant probability measure μ under a curvature-dimension condition CD(ρ, n)
(Definition 3.3.14, p. 159).
To start with however, observe that under the condition CD(ρ, ∞) with ρ > 0,
one may let t → ∞ in the local inequalities of Theorem 5.5.2 to get in the limit that
for every positive function f in A0 ,

1 (f )
Entμ (f ) ≤ dμ. (5.7.1)
2ρ E f
In other words (changing f into f 2 and extending to the domain D(E)), μ sat-
isfies LS( ρ1 ). This limiting procedure requires the ergodicity properties described
in Sect. 1.8, p. 32, which hold in particular under connexity (cf. e.g. Sect. 3.2.1,
p. 140).
Although this logarithmic Sobolev inequality is contained in Theorem 5.7.4 be-
low (corresponding to n = ∞), it is worth stating it independently.
Proposition 5.7.1 (Logarithmic Sobolev inequality under CD(ρ, ∞)) Under the
curvature condition CD(ρ, ∞), ρ > 0, the Markov Triple (E, μ, ) satisfies a log-
arithmic Sobolev inequality LS(C) with constant C = ρ1 . That is, for every function
f ∈ D(E),
2
Entμ f 2 ≤ E(f ).
ρ
The constant is optimal for the example of the Ornstein-Uhlenbeck semi-

group with invariant measure the standard Gaussian measure corresponding to
the curvature condition CD(1, ∞). The proposition covers probability measures
dμ = e−W dx with respect to the standard carré du champ operator on Rn , where
W satisfies ∇∇W ≥ ρ Id (as symmetric matrices) for some ρ > 0 (equivalently,
2
W (x) − ρ|x|
2 is convex).

Then μ satisfies the logarithmic Sobolev inequality LS( ρ1 ).
As for the corresponding Poincaré inequality (Proposition 4.8.2, p. 212), with

respect to which it is actually a strengthening, a similar statement, with the
same proof, holds on a weighted Riemannian manifold (M, g) for the operator
L = g − ∇W · ∇ with the reversible (probability) measure μ having density e−W
5.7 Logarithmic Sobolev Inequalities Under a Curvature-Dimension Condition 269
with respect to the Riemannian measure under the curvature condition
Ric(L) = Ricg +∇∇W ≥ ρ g
(cf. Sect. 1.16, p. 70, and Sect. C.6, p. 513).

Towards further developments under a curvature-dimension condition, it is
worthwhile to formulate an analogue of Proposition 4.8.3, p. 214, which amounts to
the second derivative of entropy along the semigroup, that is according to Bruijn’s
identity (Proposition 5.2.2), to the derivative of the Fisher information. This further
emphasizes the semigroup interpolation principle and its connection with the 2
const+
operator, now with
respect to the invariant measure. Indeed, for f ∈ A0 say,
consider (t) = E Pt f log Pt f dμ, t ≥ 0. As in Proposition 5.5.3,

(t) = − Pt f (log Pt f )dμ = −Iμ (Pt f )
E
and

(t) = 2 Pt f 2 (log Pt f )dμ
E
where we recall the Fisher information Iμ from (5.1.6). Now, if (t) ≥ − C2 (t)
for some C > 0, then e2t/C (t) is increasing in t ≥ 0, and therefore
Iμ (Pt f ) = − (t) ≤ −e−2t/C (0) = e−2t/C Iμ (f ) (5.7.2)
for every t ≥ 0. Writing then from de Bruijn’s identity, for f positive in A0 ,

∞
Entμ (f ) = Iμ (Pt f )dt, (5.7.3)
0
the logarithmic Sobolev inequality LS(C) follows.

Note that integrating the gradient bound (iii) of Theorem 5.5.2 with respect to the
invariant measure μ together with the Cauchy-Schwarz inequality yields that, under
the curvature condition CD(ρ, ∞) for some ρ ∈ R,
Iμ (Pt f ) ≤ e−2ρt Iμ (f ) (5.7.4)
for every positive function f in A0 and every t ≥ 0. We therefore recover in this

way Proposition 5.7.1 whenever ρ > 0.
As announced, from the expressions of and above, an analogue of Propo-
sition 4.8.3, p. 214, may be given at this stage.
Proposition 5.7.3 A Markov Triple (E, μ, ) satisfies a logarithmic Sobolev in-

equality LS(C) for some C > 0 if

f (log f )dμ ≤ C f 2 (log f )dμ (5.7.5)
E E
for every positive function f in A which is bounded together with (f ) and Lf .
At a technical level, (5.7.5) should actually be used for Pt f whenever f ∈ Aconst+

0 .
Now, under a curvature condition CD(ρ, ∞) for some ρ ∈ R, the analysis devel-
oped in Sect. 3.3, p. 151, indicates that for t > 0, Pt f is indeed in A and moreover
bounded together with (Pt f ) and LPt f . Proposition 5.7.5 will only be used below
under curvature bounds.
Inequality (5.7.5) is very similar indeed to the dual inequality

(f )dμ ≤ C 2 (f )dμ
E E
of Proposition 4.8.3, p. 214, used for Poincaré inequalities. But whereas the latter
expresses a dual equivalent formulation of the Poincaré inequality, (5.7.5) turns out
to be strictly stronger in general than the logarithmic Sobolev inequality LS(C).
On the basis of the previous semigroup tools, we next investigate logarithmic
Sobolev inequalities under curvature-dimension conditions. As for Poincaré in-
equalities (cf. Theorem 4.8.4, p. 215), the constant C = ρ1 in Proposition 5.7.1 may
be improved under a curvature-dimension hypothesis CD(ρ, n) with n < ∞. Again
by Proposition 5.1.3, the next statement strengthens the Poincaré inequality from
Theorem 4.8.4, p. 215. The proof is a prototypical example of the power of the
change of variables formula for the 2 operator (and several integral identities in-
volving 2 in the proof below will turn out to be useful later in this monograph).
Theorem 5.7.4 (Logarithmic Sobolev inequality under CD(ρ, n)) Under the
curvature-dimension condition CD(ρ, n), ρ > 0, n > 1, the Markov Triple (E, μ, )
satisfies a logarithmic Sobolev inequality LS(C) with constant C = n−1
ρn .
Proof The objective is to reach the criterion of Proposition 5.7.3. The main argu-
ment relies on the chain rule formula (3.3.2), p. 158, for the 2 operator used here
with ψ(r) = ear for a real parameter a so to get, for g ∈ A0 ,

2 eag = a 2 e2ag 2 (g) + a g, (g) + a 2 (g)2 . (5.7.6)
Note that 2 (eag ) actually takes place in the extended algebra A. Now on the other
hand, by Proposition 3.3.17, p. 161, which applies to eag with g ∈ A0 , and the
change of variables for L,

ag 2
2 eag dμ = L e dμ
E E

2
= a2 e2ag Lg + a (g) dμ
E

= a2 e2ag (Lg)2 + 2a Lg (g) + a 2 (g)2 dμ.
E
5.7 Logarithmic Sobolev Inequalities Under a Curvature-Dimension Condition 271
Observe that since g ∈ A0 , all the terms in the latter integral are bounded and there-
fore integrable with respect to the finite measure μ. Integrating by parts,

e 2ag
Lg (g)dμ = − e 2ag
g, (g) dμ − 2a e2ag (g)2 dμ. (5.7.7)
E E E
After some elementary algebra, it follows from the preceding identities that

e 2ag
(Lg) dμ =
2
e2ag 2 (g) + 3a g, (g) + 4a 2 (g)2 dμ. (5.7.8)
E E
This identity turns out to be most useful, providing a way to classify many kinds of
quantities which are related to each other after integration by parts.
Now, by (5.7.6) again, the curvature-dimension condition CD(ρ, n) applied to
eag yields that
1 2
2 (g) + a g, (g) + a 2 (g)2 ≥ ρ (g) + Lg + a (g) .
n
After multiplication by eg and integration with respect to μ,

eg 2 (g) + a g, (g) + a 2 (g)2 dμ
E

≥ρ eg (g)dμ
E

1
+ eg (Lg)2 + 2a Lg (g) + a 2 (g)2 dμ (5.7.9)
n E

=ρ eg (g)dμ
E

1
+ eg (Lg)2 − 2a g, (g) + a 2 − 2a (g)2 dμ
n E
where in the last step, (5.7.7) (with a = 12 ) was used again. In the latter, replace now

the term E eg (Lg)2 dμ by (5.7.8) (with a = 12 ) to get that

ρn
eg 2 (g)dμ ≥ eg (g)dμ
E n−1 E

3 − 2(n + 2)a
+ eg g, (g) dμ (5.7.10)
2(n − 1) E

(a − 1)2 − na 2
+ eg (g)2 dμ.
n−1 E

Choosing a = 2(n+2)3
(for which the term E eg (g, (g))dμ disappears) yields
that

ρn 4n − 1
eg 2 (g) dμ ≥ eg (g)dμ + eg (g)2 dμ (5.7.11)
E n − 1 E 4(n + 2)2
E
and in particular

ρn
eg 2 (g)dμ ≥ eg (g)dμ. (5.7.12)
E n−1 E
This is almost the required (5.7.5) although for f = eg where g ∈ A0 . Now, by
Proposition 3.3.6, p. 156, and Lemma 3.3.16, p. 160, under the curvature condi-
tion (5.7.12) extends to functions f ∈ A which are bounded, together with (f ) and
Lf (such functions belong to D(L)). The proof of Theorem 5.7.4 is complete.
Theorem 5.7.4 applies to the sphere Sn in Rn+1 (Sect. 2.2, p. 81), and for the
symmetric Jacobi operator with dimension n (Sect. 2.7.4, p. 113), the lower bound
ρn
n−1 of the logarithmic Sobolev constant being optimal since these two models
both satisfy a curvature-dimension condition CD(n − 1, n) and their Poincaré con-
stant (equal to n1 ) is already optimal. The case n = 1 corresponding to the one-
dimensional torus, for which ρ = 0, may easily be integrated into the picture. In-
deed, by Proposition 4.5.5, p. 199, the Poincaré inequality P (1) holds in this case,
and by the dual formulation of Proposition 4.8.3, p. 214, for every f in D(L),

2 (f )dμ ≥ (f )dμ.
E E
This inequality applied to f = eg/2 (for g ∈ A0 ) yields by (5.7.6) that

1 1
eg 2 (g) + g, (g) + (g)2 dμ ≥ eg (g)dμ. (5.7.13)
E 2 4 E
Now, under CD(0, 1), (5.7.9) with a = 1

3 shows that

2
eg 2 (g)dμ ≥ eg (Lg)2 − g, (g) − (g)2 dμ.
E E 3
Comparing with (5.7.8) with a = 1

2 indicates that

2
eg g, (g) dμ + eg (g)2 dμ ≤ 0.
E 3 E
Therefore, by (5.7.13),

e 2 (g)dμ ≥
g
eg (g)dμ,
E E
ρn
that is (5.7.12) with the convention that n−1 = 1. The logarithmic Sobolev inequal-
ity with constant 1 (equal to the Poincaré constant) follows similarly. Of course,
simple integrations by parts on the torus may be used to provide a direct proof. The
constant is optimal since it is already optimal for the Poincaré inequality.
By analogy with the Poincaré inequalities of Proposition 4.5.5, p. 199, we may
emphasize the preceding conclusion for n = 1 on the unit interval (after scaling).
Proposition 5.7.5 Let Lf = f be the Sturm-Liouville operator acting on smooth

functions f on [0, 1] with invariant measure the Lebesgue measure dx.
(i) For any smooth function f : [0, 1] → R,

2
f dx
2
Entdx f 2 ≤ 2
π [0,1]
and the constant is optimal.

(ii) For any smooth function f : [0, 1] → R such that f (0) = f (1),

1
f dx
2
Entdx f 2 ≤
2π 2 [0,1]
and the constant is optimal.
These two results correspond respectively to Neumann and periodic boundary

conditions on the unit interval. The first assertion follows from the second one by
symmetrization and periodization. Sharpness of the constants is a consequence of
the corresponding result for Poincaré inequalities.

The notion of entropy has a long run in mathematics and information theory. Land-
mark contributions involving entropy are due to L. Boltzmann in mathematical
physics back in the 19th century, and to C. Shannon in information theory [383].
Logarithmic Sobolev inequalities have emerged under different forms and names
in the second half of the 20th century. Although not identified as such, one of the
first occurrences of logarithmic Sobolev inequalities (in their Euclidean version dis-
cussed in the next chapter) may be found in the work [387] by A. Stam in the con-
text of information theory (see [106, 141, 157, 424, 426]). A further early obser-
vation is due to P. Federbush [183] in mathematical physics. However, it is really
with the seminal paper of L. Gross [224] that logarithmic Sobolev inequalities were
clearly identified and emphasized as a central object of interest. One main result
of [224] is the equivalence (Theorem 5.2.3) with hypercontractivity (in a rather gen-
eral context). Hypercontractivity for the Ornstein-Uhlenbeck semigroup had been
put forward earlier by E. Nelson [325, 326] (see also later [327, 328]), while a first
hypercontractive result goes back to J. Glimm [204] and to B. Simon and R. Høegh-
Krohn [384] (in the form of Theorem 5.2.5). Hyperboundedness in the sense of the
hypothesis of Theorem 5.2.5 implies a spectral gap, as conjectured in [384]. This
was recently established in [311] (see [432] for a prior weaker result), and consid-
erably improves upon Proposition 5.1.3 (although not quantitatively). The existence
and uniqueness of ground states for Schrödinger operators [223] was a further im-
portant step in the theory. The interplay between logarithmic Sobolev inequalities
and hypercontractivity has actually been quite rich and fruitful. The reference [145]
gives an account of some of the main aspects of this connection, while the re-
views [225] and [226] survey the developments of logarithmic Sobolev inequalities
over the last decades and collect references. Further general references on logarith-
mic Sobolev inequalities are [372] and [14] as an introduction, and [229, 238] with
an emphasis on statistical physics models (for bounded and unbounded spins).
The tightening Proposition 5.1.3 has been observed in [158] (on the basis of an
independent proof of (5.1.8)). Lemma 5.1.4 is due to O. Rothaus [370] as is Re-
mark 5.1.5 [367–369]. The stability by perturbation property, Proposition 5.1.6, is
due to R. Holley and D. Stroock [245]. The general proof presented here on the
basis of Lemma 5.1.7 is taken from [14] (see also [372]). Proposition 5.1.8 is part of
the slicing technique further developed in the context of Sobolev-type inequalities
in Chap. 6.
Theorem 5.2.1 goes back to the origin of entropy and to the Boltzmann
H -Theorem (see [86, 424]). The Pinsker-Csizsár-Kullback inequality (5.2.2) can
be found in numerous references including e.g. [14, 141, 426]. de Bruijn’s iden-
tity of Proposition 5.2.2 is recorded in [387] (see also [157]) for the Euclidean
heat semigroup and extensively developed in a more general framework in [36].
Exponential decays in uniform norms for models from statistical mechanics may
be developed with the same principle (see e.g. [229]). The main equivalence Theo-
rem 5.2.3 with hypercontractivity is due to L. Gross [224]. Reverse hypercontractiv-
ity (Remark 5.2.4) was first observed by C. Borell and S. Janson [88]. Theorem 5.2.5
goes back to [384]. Proposition 5.2.6 is due to P. Cattiaux [115], who gave a dif-
ferent proof. A main feature of logarithmic Sobolev inequalities, stability under
products, had already been emphasized and used in a critical way in [224]. The
dimension-free property is in particular a powerful argument in statistical mechan-
ics (cf. [229, 372]).
Integrability of Wiener chaos via hypercontractivity was emphasized by C. Borell
[88]. The discussion at the end of Sect. 5.3 on spectral features of hypercontractive
bounds is taken from [81].
The so-called Herbst argument from Proposition 5.4.1 in Sect. 5.4 goes back
to an unpublished letter (1975) of I. Herbst to L. Gross. The argument is presented
in [146], and emphasized later in the paper [6] which revived interest in the question
and gave rise to several subsequent contributions, in particular related to measure
concentration, synthesized in the notes [276] (to which we refer for more details).
See also [90] for applications to concentration inequalities. Proposition 5.4.2 is
taken from [7]. The Muckenhoupt criterion of Theorem 5.4.5 is due to S. Bobkov
and F. Götze [77] and is based on the corresponding result for Poincaré inequalities
(see the preceding chapter).
There are at least fifteen different proofs of the Gaussian logarithmic Sobolev
inequality (Proposition 5.5.1), some of them mentioned in this book, including the
proof of L. Gross [224] relying on the two-point space inequality, tensorization and
the central limit Theorem. The proof given here using heat flow monotonicity along
the Ornstein-Uhlenbeck semigroup goes back to the seminal paper [36] and is par-
tially inspired by the stochastic calculus proof of Nelson’s hypercontractivity by
J. Neveu [329]. A careful and detailed exposition of the argument in the language of
partial differential equations is given in [412] in the Gaussian case, and later in [16]
for strictly convex potentials in Rn . The paper [412] actually emphasized the impor-
tance of logarithmic Sobolev inequalities towards convergence to equilibrium in this
context. The logarithmic Sobolev inequality for Gaussian measures (as well as the
corresponding Poincaré inequality) has been extended to the infinite-dimensional
setting of the Wiener measure (cf. Sect. 2.7.2, p. 108), and further to Brownian
paths on a Riemannian manifold. With this task in mind, finite-dimensional approx-
imations may be used. Alternatively, the semigroup scheme may also be developed
directly in the infinite-dimensional context, in particular by means of the Clark-
Ocone probabilistic interpolation formula for functionals of the Brownian paths
(cf. [5, 105, 182, 250, 417] and [251] for a synthesis). The local heat kernel inequal-
ities of Sect. 5.5 have been analyzed in [27] (see also [28, 277, 394]) on the basis of
the ideas in [36]. The latter paper, which introduces the basic interpolation scheme
along the semigroup and the importance of the 2 operator, also describes the log-
arithmic Sobolev inequalities under the curvature-dimension condition of Sect. 5.7.
See [26] for a first synthesis. Remark 5.5.4 is developed in [274] for entropy. Local
hypercontractivity (Theorem 5.5.5) is taken from [31].
The Harnack inequalities for an infinite-dimensional diffusion operator from
Theorem 5.6.1 are due to F.-Y. Wang [428]. The log-Harnack inequality (5.6.2) can
be found in [76] as well as in [434] as a limit case of the previous Harnack inequal-
ities. The lower bound (5.6.3) is pointed out in [435]. Discussions and extensions
are developed in a series of papers by this author [431, 433, 435]. The useful ob-
servation (5.6.4) is due to M. Hino [242] and will be employed in a critical way in
Sect. 8.7, p. 425. Heat kernel bounds of the type (5.5.9) are surveyed in [217].
Logarithmic Sobolev inequalities under curvature-dimension condition
(Sect. 5.7) go back to [26, 36] (see also [371]). That the integral criterion from
Proposition 5.7.3 is not equivalent to the corresponding logarithmic Sobolev in-
equality has been pointed out by B. Helffer ([14, 238]). Proposition 5.7.5 on
logarithmic Sobolev inequalities on an interval may be found in [367, 368, 439]
and [173]. Hypercontractivity on the n-sphere via ultraspherical polynomials is
studied in [322].
Chapter 6
Sobolev Inequalities
Following our study of Poincaré and logarithmic Sobolev inequalities, this chap-
ter is devoted to the investigation of Sobolev inequalities. Sobolev inequalities play
a central role in analysis, providing in particular compact embeddings and tight
connections with heat kernels bounds. They are also deeply linked with the geo-
metric structure of the underlying state space through conformal invariance. Here
our study only covers a small fraction of the vast subject of Sobolev inequalities,
with a specific focus on the main theme of Markov diffusion operators and semi-
groups. As in the preceding chapters, we work here in the context of a Full Markov
Triple (“Markov Triple”) (E, μ, ) with Dirichlet form E, infinitesimal generator L,
Markov semigroup P = (Pt )t≥0 and underlying function algebras A0 and A as sum-
marized in Sect. 3.4, p. 168. Once again, several properties and results described in
this chapter remain valid in more general settings (in particular for Standard Markov
Triples), and the reader may adapt the statements and proofs if necessary.
The chapter starts with a brief exposition of the classical Sobolev inequalities on
the model spaces, namely the Euclidean, spherical and hyperbolic spaces. These
examples will both give hints and provide a better understanding of further de-
velopments. The next section investigates the various definitions of Sobolev-type
inequalities in the Markov Triple context, emphasizing in particular (logarithmic)
entropy-energy and Nash-type inequalities. These inequalities will be further inves-
tigated in Chap. 7. Section 6.3 presents the basic equivalence between Sobolev in-
equalities and (uniform) heat kernel bounds (ultracontractivity), and Sect. 6.4 high-
lights applications to compact embeddings. Sections 6.5 and 6.6 address issues on
tensorization properties of Sobolev-type inequalities and diameter, Lipschitz func-
tions and volume estimates. In particular, with respect to Poincaré and logarithmic
Sobolev inequalities, tensorization of Sobolev inequalities have to take into account
a dimensional parameter. Local inequalities under the semigroup (Pt )t≥0 are inves-
tigated next, providing in particular a heat flow approach to the celebrated Li-Yau
parabolic inequality. Section 6.8 establishes the sharp Sobolev inequality under the
curvature-dimension condition CD(ρ, n) (ρ > 0, n < ∞), covering the example
of the standard sphere. The subsequent section describes the conformal invariance
properties of Sobolev inequalities, and, as a consequence, the sharp Sobolev in-

DOI 10.1007/978-3-319-00227-9_6,
278 6 Sobolev Inequalities
equalities in the Euclidean and hyperbolic spaces on the basis of the inequality on
the sphere. Gagliardo-Nirenberg inequalities form a further family of equivalent
Sobolev-type inequalities, well-suited to non-linear porous medium and fast diffu-
sion equations. These topics, and their geometric counterparts, are presented and
detailed in Sects. 6.10 and 6.11, providing in particular a fast diffusion approach to
several Sobolev inequalities of interest.
6.1 Sobolev Inequalities on the Model Spaces
The fundamental example of a Sobolev inequality takes place in the Euclidean space
Rn equipped with the Lebesgue measure dx and the standard carré du champ op-
erator (f ) = |∇f |2 . The basic Sobolev inequality expresses here that for every
smooth compactly supported function f on Rn , n > 2 (and thus every f in the
Dirichlet domain),

2
2
f p ≤ Cn |∇f | 2 = Cn |∇f |2 dx (6.1.1)
Rn
where p = n−2 2n
and Cn > 0 is an explicit constant whose optimal value is known
(and will be explicitly calculated below in Theorem 6.9.4). The norms are of course
understood here with respect to the Lebesgue measure. For simplicity |∇f |2 is
denoted ∇f 2 below.
2n
The exponent p takes the value n−2 (n > 2) for a very good reason. It is only for
this value that the Sobolev inequality (6.1.1) is invariant under dilation, that is by
the change of f (x) into fs (x) = f (sx), s > 0. Indeed, fs p = s −n/p f p while
∇fs 22 = s 2−n ∇f 22 , and it is only for the exponent p = n−2
2n
that the inequality
can hold, any other value leading to a contradiction as s → 0 or ∞. Dilations actu-
ally play a central role in the analysis of Sobolev inequalities in Euclidean space as
will be illustrated throughout this chapter.
On the unit sphere Sn in Rn+1 , with the normalized uniform measure μ and the
spherical carré du champ operator as defined in Sect. 2.2, p. 81, there is also
a Sobolev inequality which takes a somewhat different form, but with the same
exponent p = n−2 2n
. Namely, for every smooth function f on Sn (in the Dirichlet
domain)

4
f 2p ≤ f 22 + (f )dμ. (6.1.2)
n(n − 2) Sn
The norms denote here the norms in Lr (μ), r ≥ 1. In this form, the Sobolev inequal-
ity is closer to the Poincaré and logarithmic Sobolev inequalities discussed in the
preceding chapters, and may be established for large families of Markov Triples. In
the absence of dilations, the value of p = n−2
2n
is rather mysterious. By monotonic-
ity (Jensen’s inequality), different values of 2 ≤ p ≤ n−22n
may actually be consid-
2n
ered, however not exceeding the critical value n−2 . Note also that (6.1.2) is tight
6.2 Sobolev and Related Inequalities 279
in the sense of Remark 4.2.2, p. 182, implying that the only functions f such that
(f ) = 0 are constant.
Finally, there is also a Sobolev inequality on the hyperbolic space Hn
(cf. Sect. 2.3, p. 88) which takes the form

f p ≤ −A f 2 + Cn
2 2
(f )dμ (6.1.3)
Hn
for some (explicit) A > 0, Cn > 0, and every smooth function f : Hn → R.

These inequalities will be established as the chapter unfolds (cf. Theorem 6.9.4
below, where the various constants will be made explicit). Actually, the (deep) con-
nections between the three Sobolev inequalities, on the three model spaces of ge-
ometry, Euclidean, spherical and hyperbolic, as well as the explicit values of the
constants and of the extremizers, will be discussed later in Sect. 6.9 by means of
conformal invariance, and form the basis of the geometric analysis of Sobolev in-
equalities.
6.2 Sobolev and Related Inequalities
Following the preceding introductory section, we present families of Sobolev in-

equalities, and some of their equivalent forms, in the setting of a Markov Triple
(E, μ, ). Recall the norm · p in Lp (μ), 1 ≤ p ≤ ∞.
6.2.1 Sobolev Inequalities
The next definition introduces the notion of a Sobolev inequality in a general context.
Definition 6.2.1 (Sobolev inequality) A Markov Triple (E, μ, ) is said to satisfy

a Sobolev inequality S(p ; A, C) with exponent p > 2 and constants A ∈ R, C > 0,
if for all functions f in the Dirichlet domain D(E),
f 2p ≤ A f 22 + C E(f ).
On the basis of the model examples, the exponent p will often take the form
p = n−22n
for some n > 2 (not always an integer), and we then speak of a Sobolev
inequality Sn (A, C) of dimension n (> 2) and constants A ∈ R, C > 0. Under a
Sobolev inequality S(p ; A, C), if f ∈ D(E) then f ∈ Lp (μ). As usual, it is enough
to state (and prove) such inequalities for a family of functions f which is dense in
the Dirichlet domain D(E) (typically the algebra A0 ).
When the measure μ is finite and normalized into a probability measure, apply-
ing the Sobolev inequality from Definition 6.2.1 to f = 1 shows that A ≥ 1. (In
particular, a Sobolev inequality S(p ; A, C) with A = 0 (or A ≤ 0) can only hold if
the measure μ has infinite mass.) The inequality is called tight if A = 1, and is then
denoted by S(p ; C) (respectively Sn (C)) for simplicity. The best constant C > 0
for which such an inequality S(p ; C) or Sn (C) holds is sometimes referred to as
the Sobolev constant (of the Markov Triple).
As for logarithmic Sobolev inequalities, tightness only holds if there is a Poincaré
inequality (Definition 4.2.1, p. 181).
Proposition 6.2.2 (Tightening with a Poincaré inequality) Let (E, μ, ) be a

Markov Triple with μ a probability measure, and let p > 2. A tight Sobolev inequal-
ity S(p ; C) implies a Poincaré inequality P ( p−2
C
). Furthermore, a Sobolev inequal-
ity S(p ; A, C) together with a Poincaré inequality P (C ) imply a tight Sobolev in-
equality S(p ; (p − 1)(AC + C)).
It should be noted that the application of the first assertion to the Sobolev in-
equality (6.1.2) on the sphere Sn yields the Poincaré inequality P ( n1 ) with its sharp
4
constant (cf. Theorem 4.8.4, p. 215). In particular, both constants (1 and n(n−2) ) in
the Sobolev inequality (6.1.2) are optimal.
Proof The proof is similar to Proposition 5.1.3, p.

238. For the first assertion, apply
S(p ; C) to f = 1 + εg where g ∈ D(E) with E gdμ = 0. As ε → 0, it is not
difficult to verify by a Taylor expansion that

f p − f 2 = (p − 2) ε
2 2 2
g 2 dμ + o ε 2
E
while, clearly, E(f ) = ε 2 E(g). It follows that

(p − 2) g 2 dμ ≤ C E(g)
E
C
which amounts to the Poincaré inequality P ( p−2 ).
The second assertion relies on the analogue of (5.1.8), p. 240, stating that, for
p > 2 and every f ∈ Lp (μ),
2
2
f 2p ≤ f dμ + (p − 1) fˆp (6.2.1)
E

where fˆ = f − E f dμ. (This inequality actually covers

(5.1.8) as p → 2.) For
a proof, consider a bounded function f such that E f dμ = 1 represented as

f = 1 + rg, r ∈ R, with E gdμ = 0 and E g 2 dμ = 1. The inequality to be es-

tablished is then
ψ(r) = 1 + rg2p ≤ 1 + (p − 1) r 2 g2p
which may be shown by elementary calculus, proving that ψ (r) ≤ 2(p − 1)g2p .
Using (6.2.1), apply S(p ; A, C) to fˆ (in D(E)) and then the Poincaré inequality
P (C ) to get
2
2
f 2p ≤ f dμ + (p − 1) A fˆ2 + C E(f )
E
2

≤ f dμ + (p − 1) A C + C E(f ).
E
By Jensen’s inequality, this inequality is even better than the expected Sobolev in-
equality S(p ; (p − 1)(AC + C)). The proof of the proposition is complete.
6.2.2 Logarithmic Entropy-Energy and Nash Inequalities
We next turn to an important feature of Sobolev-type inequalities, namely that they

may be presented equivalently (up to constants) in various forms, such as families
of logarithmic Sobolev inequalities or other functional inequalities. Further equiva-
lent formulations will appear throughout this chapter as well as in Chaps. 7 and 8.
While these various formulations mostly originate from inequalities in Euclidean
space, they may be addressed in an abstract framework. For the matter of com-
parison with the most familiar inequalities in Euclidean space, we use the notation
Sn (A, C) where n > 2 reflects a dimension. Recall the entropy notation Entμ intro-
duced in (5.1.1), p. 236.
Proposition 6.2.3 For a Markov Triple (E, μ, ), the following implications hold.

(i) Under Sn (A, C), A ≥ 0, C > 0, for every function f in D(E) with E f 2 dμ = 1,
n
Entμ f 2 ≤ log A + C E(f ) . (6.2.2)
2
(ii) Under (6.2.2), for every function f in D(E),
n/2
f n+2
2 ≤ A f 22 + C E(f ) f 21 . (6.2.3)
(iii) Conversely, under (6.2.3), a Sobolev inequality Sn (A1 , C1 ) holds with con-
stants A1 ≥ 0 and C1 > 0 only depending on n, A and C. Moreover A1 = 0
whenever A = 0.
This proposition introduces in particular the logarithmic entropy-energy inequal-

ity (6.2.2) and the Nash inequality (6.2.3) on the basis of their classical counterparts
in Rn . The classical Nash inequality in Rn says that for every smooth compactly
supported function f : Rn → R,
f n+2
2 ≤ Cn ∇f n2 f 21 (6.2.4)
where the sharp constant Cn is known (see the Notes and References). While Nash
inequalities with A < 0 may be considered similarly, we stick for simplicity to the
case A ≥ 0. By concavity, the logarithmic entropy-energy inequality (6.2.2) is equiv-
alent to the family of (defective) logarithmic Sobolev inequalities

Entμ f 2 ≤ (r) E(f ) + (r), r ∈ (0, ∞), (6.2.5)

2
for every f in D(E) with E f dμ = 1, where (r) = n2 log(A + Cr) and
(r) = (r) − r (r), r ∈ (0, ∞). The classical logarithmic entropy-energy in-
equality in Rn is discussed below in Proposition 6.2.5 as an equivalent form of the
logarithmic Sobolev inequality for Gaussian measures. Generalized entropy-energy
and Nash-type inequalities will be thoroughly studied in Chap. 7.
When μ is a probability, the logarithmic entropy-energy and Nash inequali-
ties (6.2.2) and (6.2.3) are called tight if A = 1. A tight logarithmic entropy-energy
inequality implies a logarithmic Sobolev inequality LS( Cn 4 ). Applied to f = 1 + εg
with ε → 0, the tight Nash inequality implies a Poincaré inequality P ( Cn 2 ).
Proposition 6.2.3 thus indicates that, up to constants (but we will however see
that optimal constants play a crucial role in a variety of problems of interest), the
preceding three inequalities are equivalent. Many further formulations may be con-
sidered here, as is clear from the proof, such as for example the Gagliardo-Nirenberg
inequalities presented in Sect. 6.10, the preceding being however amongst the most
used ones. Note that both the logarithmic entropy-energy and Nash inequalities
make sense for n ≥ 1, or even n > 0, while the Sobolev inequality requires n > 2.
Proof of Proposition 6.2.3 We begin with the first implication, proceeding as

in Proposition 5.1.8, p. 241, with the convex function φ : r → log(f 1/r ),
r ∈ (0, 1], for which φ ( 12 ) = − Entμ (f 2 ) (under f 2 = 1). Thus, by convexity
with p = n−22n
> 2,

1 1 1
n φ −φ ≤φ ,
2 p 2
which therefore amounts to
n
Entμ f 2 ≤ log f 2p . (6.2.6)
2
It remains to apply the Sobolev inequality Sn (A, C) to obtain the conclusion.
We next turn to the second claim, using the same method. Write here

1 1
2 φ(1) − φ ≤φ ,
2 2
that is, again for a function f such that f 2 = 1,

1
log ≤ Entμ f 2 .
f 1
2
An application of (6.2.2) immediately yields (6.2.3).

We are left with the last implication, which is more involved and which re-
lies on the slicing technique already developed in Proposition 5.1.8, p. 241. It is
enough to establish the Sobolev inequality for a positive bounded function f in
D(E). The idea is to apply the Nash inequality (6.2.3) to the sequence of functions
fk = (f − 2k )+ ∧ 2k , k ∈ Z. Let Nk = {f > 2k }, k ∈ Z. Since
2k 1Nk+1 ≤ fk ≤ 2k 1Nk ,

2 μ(Nk+1 ) ≤
2k
fk dμ ≤ 2 μ(Nk ) and
2 2k
fk dμ ≤ 2k μ(Nk )
E E
for every k ∈ Z. Therefore, (6.2.3) applied to each fk yields that

2k 1+(n/2) n/2
2 μ(Nk+1 ) ≤ βk 22k μ(Nk )2 (6.2.7)
where βk = A 22k μ(Nk ) + C E(fk ), k ∈ Z. In other words, with αk = 2pk μ(Nk )

(p = n−2
2n
), for every k ∈ Z,
n/(n+2) 4/(n+2)
αk+1 ≤ 2p βk αk .
n+2 n+2
By Hölder’s inequality with exponents n and 2 , it follows that
n/(n+2) 2/(n+2)
αk+1 ≤ 2p βk αk2 .
k∈Z k∈Z k∈Z

By the upper bound 2
k∈Z αk ≤ ( k∈Z αk )2 , it follows that
(n−2)/n
αk ≤ 22(n+2)/(n−2) βk .
k∈Z k∈Z
Now, according to Proposition 3.1.17, p. 137,

E(fk ) ≤ E(f ),
k∈Z

while on the other hand, since f = k∈Z 1Nk \Nk+1 f ,
3 2k
f 2 dμ ≥ 22k μ(Nk ) − μ(Nk+1 ) = 2 μ(Nk ).
E 4
k∈Z k∈Z
Hence, by definition of the βk ’s,

4A
βk ≤ f 2 dμ + C E(f ).
3 E
k∈Z

It remains to observe that E f p dμ ≤ 2p k∈Z αk to conclude the Sobolev inequal-
ity Sn (A1 , C1 ) for some A1 and C1 only depending on n and the Nash constants
A, C. Clearly A1 = 0 whenever A = 0. The proof of Proposition 6.2.3 is thus com-

plete.
Remark 6.2.4 As for Poincaré inequalities (Lemma 4.2.3, p. 182), the Cauchy-
Schwarz inequality may be used in the Nash inequality (6.2.3) to reach spectral
inequalities on functions in D(E) with support in some (measurable) subset B of E.
For example, whenever μ(B) ≤ (2A)−n/2 , for any function f ∈ D(E) with support
in B,

f 2 dμ ≤ 2Cμ(B)2/n E(f ).
B
In the classical Euclidean case, corresponding to A = 0 (and μ(E) infinite), this
inequality is known as a Faber-Krahn inequality. It describes a lower bound on the
spectrum of the restriction of the operator L to B with Dirichlet boundary condi-
tions. (See also Remark 8.2.2, p. 399.) Using similar slicing techniques, this in-
equality, valid for any B ⊂ E, implies in turn the corresponding Sobolev inequality
with the same dimension (but different constants).
As mentioned previously, in Rn , the logarithmic entropy-energy inequal-

ity (6.2.2) for the Lebesgue measure is actually equivalent to the logarithmic
Sobolev inequality for Gaussian measures (Proposition 5.5.1, p. 258).
Proposition 6.2.5 (Euclidean logarithmic Sobolev inequality) In Rn , for the Lebes-

gue measure

dx and the usual carré du champ operator and Dirichlet form
E(f ) = Rn |∇f |2 dx,

n 2
Entdx f 2 ≤ log E(f ) (6.2.8)
2 nπe

for every function f : Rn → R in the Dirichlet domain D(E) with Rn f 2 dx = 1.

2
The constant nπe is optimal.
Proof Start from the logarithmic Sobolev inequality for the standard Gaussian mea-
sure dμ(x) = (2π)−n/2 e−|x| /2 dx in Rn (Proposition 5.5.1, p. 258),
2

2
Entμ g ≤ 2 |∇g|2 dμ. (6.2.9)
Rn
Apply this inequality to g(x) =

(2π)n/4 e|x| /4 f (x),
2

x ∈2 R , where f is smooth with
n
compact support and such that Rn f dx = 1 (= Rn g dμ), to get that

2

1 n
Entμ g 2 = Entdx f 2 + f 2 |x|2 dx + log(2π).
2 Rn 2
On the other hand,
2
|x|2 2
|∇g| dμ =
2 ∇f + x f dx = |∇f | + x · f ∇f +
2
f dx.
2 4
Rn Rn Rn
Integrate by parts the middle term of the integral on the right-hand side of the pre-
ceding equality to get

1 1
x · f ∇f dx = ∇ |x|2 · ∇ f 2 dx = − f 2 |x|2 dx.
R n 4 R n 4 Rn
Since (|x|2 ) = 2n, after (a somewhat miraculous) simplification, the logarithmic

Sobolev inequality (6.2.9) for μ implies that
n
Entdx f 2 ≤ 2 E(f ) − log 2πe2 .
2

Now change f (x) into fs (x) = s n/2 f (sx), x ∈ Rn , s > 0. Since fs2 dx =
Rn
Rn f dx = 1, E(fs ) = s E(f ) and
2 2

Entdx fs2 = Entdx f 2 + n log s,
applying the preceding inequality to fs for every s > 0, and optimizing (s 2 = 4En(f ) )
yields the conclusion. The constant is optimal since this proof clearly indicates
that (6.2.8) for the Lebesgue measure is actually equivalent to the logarithmic
Sobolev inequality (6.2.9) for the Gaussian measure μ, which is sharp. Since the
exponential functions ea·x , a ∈ Rn , x ∈ Rn , are the extremal functions of the Gaus-
sian logarithmic Sobolev inequality, extremal functions of the Euclidean logarith-
2 2
mic Sobolev inequality are given by Gaussian kernels ea·x−|x| /2σ , a ∈ Rn , σ > 0,
x ∈ Rn . Proposition 6.2.5 is established.
Remark 6.2.6 The Euclidean logarithmic Sobolev inequality of Proposition 6.2.5

may actually be shown to follow from the Sobolev inequality in Rn with its sharp
constant. Indeed, apply (6.1.1) to f ⊗k on Rnk for a given (smooth compactly sup-
ported) function f : Rn → R. With the help of the sharp constant of the Sobolev
inequality in Euclidean space (cf. Theorem 6.9.4), (6.2.8) follows in the limit as
k → ∞.
The fact that, after the action of dilations, an entropy-energy or logarithmic

Sobolev inequality is equivalent to a Sobolev inequality is actually a common phe-
nomenon. It is an illustration of the fact that, for the optimal exponent p = n−2 2n
,
the Sobolev inequality is invariant under dilations in the Euclidean space. The same
operation may be performed for any functional inequality on the Euclidean space.
After optimization under dilations, it produces in general some new functional in-
equality which is equivalent to a Sobolev inequality. Further illustrations of this
phenomenon will be presented in Sect. 6.10 which deals with Gagliardo-Nirenberg
inequalities. Unfortunately, there are very few such functional inequalities for which
this operation is effectively tractable and produces useful inequalities.
6.3 Ultracontractivity and Heat Kernel Bounds

Ultracontractivity is for Sobolev inequalities what hypercontractivity is for logarith-
mic Sobolev inequalities. However, historically ultracontractivity was investigated
earlier and expresses a stronger property, namely that under a Sobolev inequality
for a Markov Triple (E, μ, ), the associated semigroup (Pt )t≥0 maps L1 (μ) into
L∞ (μ). By interpolation it then also maps Lp (μ) into Lq (μ), 1 ≤ p < q ≤ ∞.
In the following, it will sometimes be convenient to state a number of embedding
results with the operator norm
Pt p,q = sup Pt f q (6.3.1)

f p ≤1
of Pt from Lp (μ) into Lq (μ), 1 ≤ p, q ≤ ∞.

Under a logarithmic Sobolev inequality, hypercontractivity thus indicates that
Pt p,q ≤ 1 for a suitable relation between 1 0 (Theo-
rem 5.2.3, p. 246). Ultracontractivity formally expresses that Pt 1,∞ < ∞, and
more precisely investigates the decay of Pt 1,∞ as a function of t > 0. Theo-
rem 6.3.1 below describes the equivalence between this decay and Sobolev in-
equalities. To get a better feeling for the quantity Pt 1,∞ , recall from Proposi-
tion 1.2.5, p. 14, that whenever Pt 1,∞ is finite, the operators Pt , t > 0, are repre-
sented by a density kernel in L∞ (μ ⊗ μ) (thus with respect to the invariant measure
μ) in the sense that there exists, for every t > 0, a (symmetric) measurable func-
tion
pt (x, y) on the product space E × E such that, for μ-almost every x ∈ E,

p (x, y)2 dμ(y) < ∞,
E t

E
and Pt 1,∞ = pt (·, ·)∞ in L∞ (E × E, μ ⊗ μ).

The next statement presents the equivalence between Sobolev inequalities and
uniform heat kernel bounds. With respect to the corresponding hypercontractivity
result, exact constants are not really preserved in the equivalence, and thus C below
varies from line to line (see however Remark 6.3.2 after the proof).
Theorem 6.3.1 (Ultracontractivity) Let (E, μ, ) be a Markov Triple with semi-

group P = (Pt )t≥0 . Let further n > 2 be a (Sobolev) dimension. The following are
equivalent.
(i) The Sobolev inequality Sn (A, C) holds for some constants A ≥ 0 and C > 0.
(ii) There is a constant C > 0 such that for every 0 < t ≤ 1,
C
Pt 1,2 ≤ .
t n/4
(iii) There is a constant C > 0 such that for every 0 < t ≤ 1,
C
Pt 1,∞ ≤ .
t n/2
6.3 Ultracontractivity and Heat Kernel Bounds 287
If A = 0, Sn (C) is equivalent to (ii) and (iii) holds for every t > 0.
Note that the requirement n > 2 is only necessary for the Sobolev inequality
Sn (A, C) in (i) and not for the ultracontractive bounds (ii) and (iii). Actually, the
proof below will transit through Nash inequalities (6.2.3) justifying the extension to
any n > 0.
Proof Start with the equivalence between (ii) and (iii). Note first that if
Pt 1,2 ≤ K(t), then, by duality, Pt 2,∞ ≤ K(t). Since Pt = Pt/2 ◦ Pt/2 , it is
therefore bounded from L1 (μ) into L∞ (μ) with norm K 2 ( 2t ). This shows that
(ii) ⇒ (iii). The converse implication is obtained by the classical Riesz-Thorin in-
terpolation Theorem which asserts that, for an operator P , whenever P p1 ,q1 ≤ K1
and P p2 ,q2 ≤ K2 , then for every θ ∈ [0, 1],
P pθ ,qθ ≤ K1θ K21−θ
where
1 θ 1−θ 1 θ 1−θ
= + , = + .
pθ p1 p2 qθ q1 q2
Applying this result with p1 = p2 = 1, q1 = ∞, q2 = 1, K1 = C t −n/2 , K2 = 1
(since Pt is a contraction) and θ = 12 then yields the claim since pθ = 1, qθ = 2 and
(1−1)
Kθ = C 1/2 t −n/4 . The same argument shows that Pt p,q ≤ (Ct − 2 ) p q , t > 0,
n
1 ≤ p < q ≤ ∞, described in Corollary 6.3.3 below.

Consider now the equivalence between the first two assertions of Theorem 6.3.1.
There are numerous ways to deduce ultracontractive bounds from a Sobolev in-
equality. The simplest is perhaps to use the Nash inequality (6.2.3) (which is a con-
sequence of the Sobolev inequality Sn (A, C) by Proposition 6.2.3). To this end,
we make use of the method developed in Theorem 4.2.5, p. 183, to prove the de-
cay of the semigroup in L2 (μ) under a Poincaré inequality.

That is, given a pos-
itive function f in D(E) with integral 1, set
(t) = E (Pt f )2 dμ, t ≥ 0. Then
(t) = −2 E(Pt f ), and by invariance of μ, E Pt f dμ = 1 for every t ≥ 0. The
Nash inequality (6.2.3) applied to Pt f then reads
1−θ
C
(t) ≤ A (t) − (t) , t ≥ 0,
2
where we set θ = 2
n+2 . It follows that the function eλt (1 − A −r (t)), t ≥ 0, is
decreasing, where
2A r θ 2
λ= and r = = .
C 1−θ n
Therefore, for every t > 0,
1/r n/2
A A
(t) ≤ ≤ . (6.3.2)
1−e −λt + A e−λt 1 − e−λt
(0)r
If A > 0, the latter is bounded from above by C t −n/2 for some constant C > 0 for
every 0 < t ≤ 1. If A = 0, 1−eA−λt is replaced by its limit 2rt
C
= Cn
4t (for all t > 0).
Next turn to the converse implication, that is from the decay (ii) to the Nash
inequality (6.2.3) equivalent to the Sobolev inequality (i). Using the same nota-
tion, we only give the argument for A > 0 (the
case A = 0 being similar). It
may be assumed, again for f ≥ 0 such that E f dμ = 1, that for every t > 0,
(t) ≤ C (1 + t −n/2 ). By Lemma 4.2.6, p. 184, the function log is convex. There-
fore, for every θ ∈ [0, 1], and every t > 0,
−n/2 θ
t
(t) ≤ (0) 1−θ
C 1+ .
θ
Choose then θ = α t for some fixed α to be specified below, which can be achieved
as soon as t is small enough. At t = 0, the two sides of the preceding inequality are
equal. A Taylor expansion at t = 0 then yields

−2 E(f ) ≤ (0) −α log (0) + α log C 1 + α n/2 .
Choose α = E (0)
(f )
so that

E(f ) n/2
log (0) ≤ 2 + log C 1 +
(0)
and hence
n/2
(0)1+(n/2) ≤ K (0)n/2 + E(f )n/2 ≤ K1 (0) + E(f ) ,

for some K, K1 > 0. Since (0) = E f 2 dμ and E f dμ = 1, the announced Nash

inequality follows by homogeneity.
The proof of Theorem 6.3.1 is complete.
Remark 6.3.2 The preceding proof does not attempt to make the constants sharp,
and of course the arguments may be tightened at some point. It is nevertheless pos-
sible to draw some (non-optimal) information concerning the relationships between
the various constants C in Theorem 6.3.1. For example, under the Sobolev inequal-
ity Sn (A, C) for some A ≥ 0, C > 0,
C
Pt 1,∞ ≤
t n/2
for every 0 < t ≤ 1 (every t > 0 if A = 0) where C = ( Cn

2 (1 +
4A n/2
Cn )) .
In the last part of this section, we make a few observations and describe some
consequences of Theorem 6.3.1. In particular, we assume in the following that A ≥ 0
in a given Sobolev inequality Sn (A, C). As a consequence of the Sobolev inequal-
ity (6.1.1), one recovers the fact that the standard heat kernels on Rn are uniformly
6.3 Ultracontractivity and Heat Kernel Bounds 289
bounded by C t −n/2 , t > 0, which is of course obvious from the explicit representa-
tion (2.1.1), p. 78. A more careful analysis will show in Corollary 7.1.6, p. 354, that
the optimal bound (4πt)−n/2 may actually be deduced from the Euclidean logarith-
mic Sobolev inequality (6.2.8) with its sharp constant. In Chap. 7, further methods
producing ultracontractive bounds from Sobolev inequalities will be developed.
The first corollary describes other forms of ultracontractive bounds.
Corollary 6.3.3 Under Sn (A, C), for any 1 ≤ p < q ≤ ∞,
C
Pt p,q ≤ n 1 1
t 2(p−q )
for every 0 < t ≤ 1 or t > 0 depending on whether A > 0 or A = 0, where C > 0

depends on the Sobolev constants A, C and n, p, q. Moreover, for any λ > 0, the
resolvent operator Rλ = (λ Id −L)−1 is bounded from Lp (μ) into L∞ (μ) as soon
pn
as p > n2 . When p ≤ n2 , it is bounded from Lp (μ) into Lq (μ) for q < n−2p .
Proof The first assertion was mentioned in the proof of Theorem 6.3.1 (as a
consequence of the Riesz-Thorin Theorem and the fact that Pt is a contrac-

∞all L −λt
tion on p (μ)-spaces). For the resolvent R , use the integral representation
λ
Rλ = 0 Pt e dt from (A.1.2), p. 474. By the preceding, Pt p,∞ ≤ Ct −n/2p ,
0 < t ≤ 1. In addition, since the operators Pt , t ≥ 0, are contractions in all
Lp (μ)-spaces, the norm Pt p,∞ is decreasing in t, and is in particular bounded
for t ∈ (1, ∞). Consequently, for some constant C > 0 only depending on p and the
Sobolev constants,
∞ 1 ∞
−λt −n/2p
Rλ p,∞ ≤ Pt p,∞ e dt ≤ C t dt + C e−λt dt.
0 0 1
The claim follows for p > n2 . Together with a further interpolation, the same argu-
ment leads to the corresponding assertion for p < n2 . Corollary 6.3.3 is proved.
When the measure μ is a probability measure and when the Sobolev inequality
is tight (that is Sn (1, C) holds), more precise bounds as t → ∞ may be obtained.
Recall that in this setting tightness is equivalent to a Poincaré inequality (Proposi-
tion 6.2.2).
Proposition 6.3.4 Assume that μ is a probability measure satisfying a Sobolev

inequality Sn (A, C1 ) and a Poincaré inequality P (C2 ). Then there are constants
C > 0 and T > 0, depending only on A, C1 , C2 and n, such that the density kernels
pt (x, y), t > 0, (x, y) ∈ E × E, of Pt with respect to the measure μ satisfy

pt (x, y) − 1 ≤ C e−t/C2
for every t ≥ T and (μ ⊗ μ-almost) every (x, y) ∈ E × E.

Proof We prove
the result with T = 2. For t > 0, consider the operator
Pt0 f = Pt f − E f dμ whose kernel is pt (x, y) − 1. From the hypotheses,
√ √
P10 1,∞ ≤ C so that P10 1,2 ≤ C and P10 2,∞ ≤ C . On the other hand,
the Poincaré inequality indicates by Theorem 4.2.5, p. 183, that
0
P f ≤ e−t/C2 f 2 .
t 2
By composition P2+t0 = P 0 ◦ P 0 ◦ P 0 is then bounded from L1 (μ) into L∞ (μ) with

1 t 1
norm Ce −t/2C2 , from which the upper bound on |p (x, y) − 1| follows.
t
It is worth mentioning that the general Proposition 1.2.6, p. 15, applied to Pt

combined with an upper bound of the form pt (x, y) ≤ 1 + Ce−t/C2 yields that
pt (x, y) ≥ 1 − 2Ce−t/4C2 , which is therefore somewhat worse than the bound of
the previous proposition.
6.4 Ultracontractivity and Compact Embeddings
Theorem 6.3.1 indicates that as soon as a Sobolev inequality holds, the semigroup
(Pt )t≥0 is ultracontractive, and may therefore be represented by a bounded density
kernel pt (x, y), t > 0, (x, y) ∈ E × E, as

Pt f (x) = f (y)pt (x, y)dμ(y) t > 0, x ∈ E,
E
with respect to the invariant measure μ. When the measure μ is finite, the fact that Pt
may be represented by a bounded kernel pt (x, y) implies that it is Hilbert-Schmidt,
since then

pt2 (x, y)dμ(x)dμ(y) < ∞.
E E
Therefore, the operator Pt , t > 0, is compact (cf. Sect. A.6, p. 483). We summarize
these conclusions in a statement.
Corollary 6.4.1 Under a Sobolev inequality Sn (A, C) for a finite measure μ, the
operators Pt , t > 0, are Hilbert-Schmidt and the spectrum of the Markov generator
L is discrete.
It is a specific consequence of this result that the embedding from the Dirichlet
space D(E) into L2 (μ) is compact. Recall that D(E) is endowed with the norm
f E = [f 22 + E(f )]1/2 .
Theorem 6.4.2 Under a Sobolev inequality Sn (A, C) for a finite measure μ, for
any sequence (fk )k∈N in D(E) bounded with respect to the norm · E , there is a
subsequence which converges in L2 (μ).
6.5 Tensorization of Sobolev Inequalities 291
Proof If necessary, normalize μ into a probability measure. Since (fk )k∈N is

bounded in L2 (μ), extract first a subsequence (fk )∈N which converges weakly to
f in L2 (μ). Since Pt (t > 0) is compact, Pt fk → Pt f (strongly) in L2 (μ). Then,
for every , write
fk − f 2 ≤ fk − Pt fk 2 + Pt fk − Pt f 2 + Pt f − f 2 .
From (4.2.3), p. 184,

fk − Pt fk 22 ≤ 2t E(fk ).
Since sup∈N E(fk ) < ∞, first letting → ∞ and then t → 0, it follows that
fk → f in L2 (μ), proving the claim.
Note that under the Sobolev inequality Sn (A, C), the convergence of the subse-
quence (fk )∈N in the preceding proof also takes place in Lq (μ), 2 ≤ q < n−2
2n
.
While Theorem 6.4.2 applies to Sobolev inequalities on spheres S or compact
n
manifolds, it does not apply directly on Rn since the invariant Lebesgue measure dx
is not finite. The same arguments do however apply on bounded sets in Rn , leading
to the classical Rellich-Kondrachov Theorem. Given a bounded open set O in Rn ,
set, for a smooth and compactly supported function f on O,

f 2H1 (O) = f 2 dx + |∇f |2 dx.
O O
Let H1 (O) be the Hilbert completion of the space of smooth compactly supported
functions in O with respect to this norm, considered as a subspace of L2 (O, dx).
Theorem 6.4.3 (Rellich-Kondrachov Theorem) In the preceding setting, the em-

bedding from H1 (O) into L2 (O, dx) is compact.
Proof Via the stereographic projection from the sphere Sn onto Rn described in
Sect. 2.2.2, p. 83, compactly supported functions in O may be considered as func-
tions defined on Sn . Moreover, O being bounded, the norm f H1 (O) is equiva-
lent to the Dirichlet norm f E on the sphere since on a bounded set O, both the
measures and the carré du champ operators are comparable for the usual Laplacian
and the Laplacian on the sphere. It then remains to apply Theorem 6.4.2 on the
sphere to extract, from any bounded sequence in H1 (O), a subsequence convergent
in L2 (O, dx).
6.5 Tensorization of Sobolev Inequalities

In contrast to Poincaré or logarithmic Sobolev inequalities, Sobolev inequalities do
not tensorize so well. Under Sobolev inequalities Sn1 (A1 , C1 ) and Sn2 (A2 , C2 ) on
respective spaces E1 and E2 , the product space E1 × E2 carries a Sobolev inequality
Sn1 +n2 (A, C) with constants A ≥ 0, C > 0 (depending on Ai ≥ 0, Ci > 0, ni > 2,
i = 1, 2) which are not so easy to describe if one is interested in sharp bounds.

However, the parameter n behaves as an expected dimensional parameter (which is
an important issue in this product procedure).
There are several ways to reach the preceding tensorization property. The sim-
plest is perhaps to transit through ultracontractivity and Theorem 6.3.1. Indeed, if
two operators Pi on Ei are bounded from L1 (μi ) into L∞ (μi ) with norms Ki ,
i = 1, 2, then their product P1 ⊗ P2 is bounded from L1 (μ1 ⊗ μ2 ) into L∞ (μ1 ⊗ μ2 )
with norm K1 K2 . The claim then easily follows with Ki = Ci t −ni /2 .
Another approach may be developed on the basis of the logarithmic entropy-
energy inequality (6.2.2), equivalent by Proposition 6.2.3 to the Sobolev inequality
Sn (A, C) with the same dimension n. A particularly meaningful statement is ob-
tained when dealing with entropy-energy inequalities with the same constant A ≥ 0
(for example A = 0). Recall from Proposition 4.3.1, p. 185, the Dirichlet form E
with domain D(E) on a product space E1 × E2 .
Proposition 6.5.1 On Markov Triples (Ei , μi , i ), i = 1, 2, assume that the log-

arithmic entropy-energy inequalities (6.2.2) hold with respective dimension ni > 0
and constants Ai , Ci with Ai = A and Ci = nCi for some A ≥ 0 and C > 0. Then,
on the product space (E1 × E2 , μ1 ⊗ μ2 , 1 ⊗ 2 ), the logarithmic entropy-energy
inequality

n C
Entμ f 2 ≤ log A + E(f )
2 n

2
holds for every function f such that E f dμ = 1 in the Dirichlet space D(E) of
the product space E = E1 × E2 , where μ = μ1 ⊗ μ2 and n = n1 + n2 .
Proof According to (6.2.5), the respective inequalities (6.2.2) are equivalent to the
families of (defective) logarithmic Sobolev inequalities

Entμi f 2 ≤ i (ri ) Ei (f ) + (ri ), ri ∈ (0, ∞),

for every f in D(Ei ) with Ei f 2 dμi = 1, i = 1, 2, where
C
i (r) =
2(A + Cr
ni )
and
Cr
ni Cr
i (r) = i (r) − r i (r) =
ni
log A + − .
2 ni A + Cr
ni
ri
To add dimensions, choose ni = r
n for r > 0 so that
C
1 (r1 ) = 2 (r2 ) =
2(A + Cr
n )
6.6 Sobolev Inequalities and Lipschitz Functions 293
and
Cr
n Cr
1 (r1 ) + 2 (r2 ) = log A + − n
.
2 n A + Cr
n
The conclusion then follows from the tensorization of logarithmic Sobolev inequal-
ities in Proposition 5.2.7, p. 249, and retrieves the logarithmic entropy-energy in-
equality on the product space. Proposition 6.5.1 is established.
The previous proposition is of particular interest when A = 0, which shows

that the Euclidean logarithmic Sobolev inequality (6.2.8) is stable under products.
Proposition 6.5.1 may also be applied to tensorize the same logarithmic entropy-
energy inequality
n
Entμ f 2 ≤ log A + C E(f )
2
k-times, k ≥ 1, yielding

kn C
Entμ⊗k f 2 ≤ log A + E(f )
2 k

for every function f on the k-fold product space E k such that E k f 2 dμ⊗k = 1. This
is the procedure used in Remark 6.2.6 to reach the optimal Euclidean logarithmic
Sobolev inequality from the Sobolev inequality with sharp constants. If A = 1, as
k → ∞, the latter implies (formally) a logarithmic Sobolev inequality LS( Cn 4 ) on
the infinite-dimensional product space. This is another illustration of the importance
of logarithmic Sobolev inequalities in infinite dimension (as a form of limiting or
dimension-free Sobolev inequalities).
6.6 Sobolev Inequalities and Lipschitz Functions

In the corresponding sections in Chaps. 4 and 5, integrability properties of Lipschitz
functions with respect to a measure satisfying either a Poincaré or a logarithmic
Sobolev inequality were described. The picture here, under a Sobolev inequality in
the finite measure case, is that Lipschitz functions are actually (uniformly) bounded.
Recall from Sect. 3.3.7, p. 166, that the operator induces a natural distance d(x, y)
on points (x, y) ∈ E × E (with respect to the extended algebra A). (For example,
for the classical carré du champ operator on Rn , d gives rise to the Euclidean dis-
tance.) According to (3.3.10), p. 167, to control the diameter D = D(E, μ, ) of
the Markov Triple (E, μ, ), it is convenient to bound Lipschitz functions as
D = D(E, μ, ) = sup f˜∞

f Lip ≤1
where f˜(x, y) = f (x) − f (y), (x, y) ∈ E × E, and the norm is taken in

L∞ (E × E, μ ⊗ μ).
That the diameter D is finite under a Sobolev inequality with respect to a prob-
ability measure μ may, for example, be achieved via logarithmic entropy-energy
inequalities (equivalent to Sobolev inequalities by Proposition 6.2.3), much in the
spirit of the Herbst argument in Sect. 5.4.1, p. 252.
Proposition 6.6.1 Let (E, μ, ) be a Markov Triple with μ a probability measure.

Under the tight logarithmic entropy-energy inequality

2 n
Entμ f ≤ log 1 + C E(f ) , f ∈ D(E), f 2 dμ = 1,
2 E
the diameter D = D(E, μ, ) is finite.
Proof Set (r) = n2 log(1 + Cr), r ≥ 0. We proceed as in the proof of Proposi-

tion 5.4.1, p. 252,

using the same notation. For a (bounded) 1-Lipschitz function f ,
setting Z(s) = E esf dμ, s ∈ R, we reach in the same way the differential inequality
2
1 s
F (s) ≤ 2 , s = 0
s 4

on F (s) = 1
s log Z(s), with F (0) = E f dμ. Integrating, it follows that for every
s ∈ R,

s 2
1 u
es(f − E f dμ)
dμ ≤ exp s 2
du .
E 0 u 4

∞ 2
Setting C = 0 u12 ( u4 )du, which is finite by definition of , it easily follows as

s → ±∞ that f − E f dμ∞ ≤ C. The conclusion follows by the very definition

of the diameter D.
The bound produced by Proposition 6.6.1 is not sharp. For example, starting
from the Sobolev inequality (6.1.2) on the sphere Sn (with diameter π ), and then the
associated logarithmic entropy-energy inequality from Proposition 6.2.3, it may be
shown that D ≤ C for some (large) C > π . More refined methods not discussed in
this book (see the Notes and References) allow us to reach directly from the Sobolev
inequality (6.1.2) the optimal bound D ≤ π . This (more refined) argument actually
relies on the application of the Sobolev inequality to some suitable transformation
(related to extremal functions of the Sobolev inequality on the sphere) of a Lipschitz
function (and not just the exponential function as for entropy-energy inequalities).
Next, we concentrate on further aspects concerning Lipschitz functions under
Sobolev-type inequalities. One such aspect deals with volumes of balls.
Proposition 6.6.2 Given a positive 1-Lipschitz function f on (E, μ, ), set

V (r) = μ(f ≤ r), r ≥ 0. Assume that the Sobolev inequality Sn (A, C) with A ≥ 0
6.6 Sobolev Inequalities and Lipschitz Functions 295
holds and that V (r0 ) < ∞ for some r0 > 0. Then, provided that
log V (r)
lim sup < ∞, (6.6.1)
r→0 log r
there is a constant c > 0 such that V (r) ≥ c r n for every 0 ≤ r ≤ 1 (every r ≥ 0 if

A = 0).
Choosing for f the distance function to a given point, V (r) corresponds to the
volume of a ball in the metric space (E, d). The proposition thus indicates that,
provided the volume of the balls does not decay too rapidly to 0 as r → 0, the
volume of small balls is bounded from below by c r n where n is the dimension
of the Sobolev inequality. This result is already coherent on Rn . In particular, the
exponent n is minimal and determined by the volume of (small) balls. This property
explains why, for a diffusion operator with smooth coefficients on a Riemannian
manifold with dimension n, the Sobolev exponent is at least n.
Proof Assuming f is in A, apply the Sobolev inequality Sn (A, C) to smooth ap-

proximations of the functions g = (1 − fr )+ , r > 0 (which belong, for r small
enough, to D(E) under the condition V (r0 ) < ∞). Since
1 1
1 r ≤ g ≤ 1{f ≤r} and (g) ≤ 1{f ≤r} ,
2 {f ≤ 2 } r2
the inequality Sn (A, C) yields, for some C1 > 0,
2/p
r 1 2C1
V ≤ C1 1 + 2 V (r) ≤ 2 V (r)
2 r r
provided r ≤ 1 (where, as usual, p = 2n

n−2 ). After some rewriting, the latter amounts
to

V ( 2r ) V (r) p/2
C2 r n ≤ C 2 n (6.6.2)
(2) r
for some suitable constant C2 > 0. Assume now that there exists an r1 ≤ 1 such that
C2 V r(rn1 ) = a < 1. Then, iterating (6.6.2), for every k ≥ 1,
1
k
r1 a (p/2) r1n
V ≤
2k C2 2kn
which yields a contradiction with the hypothesis (6.6.1) as k → ∞ since p > 2.

Therefore V (r) ≥ c r n for some c > 0 and every 0 ≤ r ≤ 1. The proof is easily
modified to see that the preceding holds for every r ≥ 0 whenever A = 0 in the
Sobolev inequality. Proposition 6.6.2 is established.
The proof of the previous proposition actually shows more. Namely, if the vol-
ume of small balls is not at least of the order r n , then it has to decay at least expo-
−b
nentially as e−cr as r → 0 for some b, c > 0.
6.7 Local Sobolev Inequalities

The local inequalities for heat kernel measures which will be described here are the
analogues of the local Poincaré and logarithmic Sobolev inequalities of Sect. 4.7,
p. 206, and Sect. 5.5, p. 257, but at the same time are more subtle since they in-
volve a dimensional parameter, as is the case for Sobolev inequalities. While the
curvature-dimension condition CD(ρ, n) is closely related to Sobolev-type inequal-
ities as will be illustrated in the following sections, there is however no hope of
characterizing it via local Sobolev inequalities on the associated (linear) heat semi-
group. This difficulty is due to the fact that the heat semigroup essentially behaves
as a Gaussian kernel (and is exactly this in the Euclidean space), and such kernels
cannot satisfy Sobolev-type inequalities (the best functional inequalities for them
being logarithmic Sobolev inequalities). Furthermore, the methods developed be-
low to reach Sobolev inequalities are non-linear. Entropy-energy inequalities will
be established later in Sect. 6.8 by the heat flow method under curvature-dimension
conditions, but by means of integration by parts formulas which are not available for
local inequalities. The alternative geometric methods of Sects. 6.9 and 6.11 (confor-
mal transformations, porous medium or fast diffusion equations, etc.) do not yield
any information on the heat semigroup.
The project pursued here will actually deal with sharpened forms of the local
logarithmic Sobolev inequalities of Sect. 5.5, p. 257, introducing the dimensional
parameter of the curvature-dimension hypothesis. In particular, the conclusions will
be shown to cover sharp Harnack inequalities. Hypercontractivity for heat kernel
measures may be investigated similarly.
6.7.1 Dimensional Logarithmic Sobolev Inequality

for the Gaussian Measure
As a first step, we are looking for inequalities which hold for Gaussian mea-
sures in finite dimension, precisely capturing the dimensional feature. To this
end, we begin with the Euclidean logarithmic Sobolev inequality of Proposi-
tion 6.2.5. Recall also the standard Gaussian measure on the Borel sets of Rn ,
dμ(x) = (2π)−n/2 e−|x| /2 dx.
2
Proposition 6.7.1 Let μ be the standard Gaussian measure on Rn . Then

1 n 1
Entμ (g) ≤ g dμ + log 1 − g (log g)dμ (6.7.1)
2 Rn 2 n Rn
6.7 Local Sobolev Inequalities 297

for every smooth positive function g on Rn with Rn gdμ = 1 and such that g and
g (log g) are well-defined and integrable with respect to μ.
Comparing with the general setting, if A0 denotes the set of smooth compactly
supported functions on Rn , Proposition 6.7.1 is established for functions in Aconst+
0
(cf. Remark 3.3.3, p. 154), possibly extended then to larger classes of functions.
Note that (6.7.1) improves upon the standard logarithmic Sobolev inequal-
ity (6.2.9) for the Gaussian measure μ since log(1 + r) ≤ r, r ≥ 0, and
2 |∇g|2
g (log g) = g − g ∇(log g) = g − .
g
Turning to the proof, which is elementary, set g = ĝ 2 where ĝ ∈ Aconst+ and
0
2 dμ = 1. (Actually, in this concrete setting, the smooth approximations can
R n ĝ
be made easily.) Then use the Euclidean logarithmic Sobolev inequality (6.2.8) in
which
we change f back to (2π)−n/4 ĝ e−|x| /4 . After integration by parts for the
2
term Rn |∇f |2 dx (as in the proof of Proposition 6.2.5), we get

1
Entμ ĝ 2 ≤ ĝ 2 |x|2 + n log(2π) dμ
2 Rn

n 2 2 n |x|2
+ log |∇ ĝ| + ĝ
2
− dμ .
2 nπe Rn 2 4
By a further integration by parts of the form (2.7.8), p. 107, the conclusion easily
follows.
This proof does indeed retrieve the logarithmic Sobolev inequality for μ from the
Euclidean logarithmic Sobolev inequality. This self-improving property might ap-
pear as surprising unless the use of dilations to optimize the Euclidean logarithmic
Sobolev inequality is recalled (cf. the proof of Proposition 6.2.5). Extremal func-
tions in (6.7.1) are Gaussian kernels. Note furthermore that (6.7.1) is a second order
inequality and that the second order operator which appears is the standard Lapla-
cian on Rn and not the Ornstein-Uhlenbeck operator L = − x · ∇ (which is
naturally associated with the Gaussian measure μ). Finally, (6.7.1) implicitly con-
tains the fact that for any g ∈ Aconst+
0 (say),

g (log g)dμ < n g dμ = n,
Rn Rn
in other words

2
g dμ − g ∇(log g) dμ < n g dμ.
Rn Rn Rn
But, from the integration by parts formula (2.7.8), p. 107, which has already been
used earlier,

g dμ = g |x|2 − n dμ,
Rn Rn
and the requested inequality therefore amounts to the obvious

2

g ∇(log g) dμ + g |x|2 dμ > 0.
Rn Rn
Remark 6.7.2 Applying (6.7.1) of Proposition 6.7.1 to 1 + εg and letting ε → 0

yields
2
1
Varμ (g)2 ≤ E(g) − g dμ . (6.7.2)
2n Rn
One may wonder at the nature
of this inequality,

and ask for a direct proof. It was
already mentioned above that Rn g dμ = Rn g [|x|2 − n]dμ. Since u2 = |x|2 − n
is an eigenfunction of −L with eigenvalue 2,

1 1
g dμ = − g Lu2 dμ = ∇g · ∇u2 dμ.
Rn 2 Rn 2 Rn

Furthermore, it is immediately checked that Rn u22 dμ = 2n (using for example

that |∇u2 |2 = 4|x|2 ). Write then a given function g as
g = α + β u2 + ĝ where
ĝ is orthogonal both to the constants and to u2 . Since Rn ∇ ĝ · ∇u2 dμ = 0, the in-

equality (6.7.2) reduces to Rn ĝ 2 dμ ≤ Rn |∇ ĝ|2 dμ which is nothing else but the

Poincaré inequality for μ applied to ĝ.
6.7.2 Local Sobolev Inequalities
The dimensional logarithmic Sobolev inequality of Proposition 6.7.1 may be ana-

lyzed further after translation and dilation, leading to abstract versions of heat kernel
inequalities. Indeed, applying (6.7.1) to g(y) = f (x + ty), x, y ∈ Rn , t ≥ 0, yields,
for the heat semigroup (Pt )t≥0 in the Euclidean space, for any t > 0, and all func-
tions f ∈ Aconst+
0 ,

n 2t Pt (f (log f ))
Pt (f log f ) − Pt f log Pt f ≤ t Pt f + log 1 − · .
2 n Pt f
2
Now let t → 0 using the Taylor expansion Pt f = f + t f + t2 2 f . After a some-
what tedious calculation, one is left with the inequality
1 2
2 (log f ) ≥ (log f )
n
which is precisely the curvature-dimension condition CD(0, n) for the Laplace op-
erator on Rn (applied to log f ). In other words, while the standard Gaussian loga-
rithmic Sobolev inequality (6.2.9) only captures by dilation the CD(0, ∞) curvature
condition, the improved inequality (6.7.1) allows us to reach the dimension n. There
are few inequalities in Euclidean space with such a property. However, as soon as
one is found, it may usually be extended to a general framework.
The following theorem is one such illustration, providing a further example of
the semigroup interpolation scheme.
Theorem 6.7.3 (Local Dimensional Logarithmic Sobolev Inequalities) Let

(E, μ, ) be a Markov Triple with semigroup P = (Pt )t≥0 and generator L. The
following assertions are equivalent.
(i) The curvature-dimension condition CD(0, n) holds.
(ii) For every function f in Aconst+
0 , and every t ≥ 0,

2t
Pt f L(log Pt f ) ≥ Pt f L(log f ) 1 + L(log Pt f ) . (6.7.3)
n
(iii) For every function f in Aconst+

Pt (f log f ) − Pt f log Pt f
(6.7.4)
n 2t Pt (f L(log f ))
≤ t LPt f + Pt f log 1 − .
2 n Pt f
(iv) For every function f in Aconst+

Pt (f log f ) − Pt f log Pt f
(6.7.5)
n 2t
≥ t LPt f − Pt f log 1 + L(log Pt f ) .
2 n
As usual, restricting the inequalities to functions in Aconst+

0 is merely a matter of
comfort. In reasonable settings, the various inequalities in Theorem 6.7.3 immedi-
ately extend to any positive bounded function f in A for which every term makes
sense, following the various extension procedures described in Sect. 3.3, p. 151.
Note that as n → ∞, the theorem covers both the logarithmic Sobolev inequality
and the reverse form described in Theorem 5.5.2, p. 259, under the curvature condi-
tion CD(0, ∞).
Proof Start with the most important step, namely the proof of (ii) under the
curvature-dimension CD(0, n) hypothesis (which will then imply (iii) and (iv)).
As usual, set (s) = Ps (φ(Pt−s f )), s ∈ [0, t], t > 0, where φ(r) = r log r, r ∈ R+ ,
and hence f ∈ Aconst+
0 . As already observed earlier in Proposition 5.5.3, p. 261,

(Pt−s f )
(s) = Ps = Ps Pt−s f (log Pt−s f )
Pt−s f
and

(s) = 2Ps Pt−s f 2 (log Pt−s f ) .
By the CD(0, n) hypothesis applied to log Pt−s f ,
2 2 [Ps (Pt−s f L(log Pt−s f ))]2

n (s) ≥ 2Ps Pt−s f L(log Pt−s f ) ≥
Ps Pt−s f
where the last inequality follows from the Cauchy-Schwarz inequality for the mea-
sure Ps with the functions (Pt−s f )1/2 L(log Pt−s f ) and (Pt−s f )1/2 . The denomi-
nator of the expression on the right-hand side of the previous inequality is actually
Pt f by the semigroup property (and is thus independent of s). Furthermore, by the
change of variables formula,
Pt−s f L(log Pt−s f ) = LPt−s f − Pt−s f (log Pt−s f ),
so that the numerator of the same expression takes the form [LPt f − (s)]2 . This
results in the differential inequality
2 [LPt f − (s)]2
(s) ≥ , s ∈ [0, t]. (6.7.6)
nPt f
In this differential inequality, the terms LPt f and Pt f behave as constants, so that
we are actually dealing with an inequality of the form u ≥ (a + bu)2 , with u = .
More precisely, setting α = nP2t f , the differential inequality (6.7.6) leads to the in-
equality, valid for any 0 ≤ u ≤ v ≤ t,

(v) − LPt f − (u) − LPt f
(6.7.7)
≥ α(v − u) (0) − LPt f (t) − LPt f ,
which holds regardless of the signs of (v) − LPt f and (u) − LPt f . The in-
equality (6.7.3) of (ii) is then a direct application of (6.7.7) with u = 0 and v = t.
As already obtained in the proof of Theorem 5.5.2, p. 259, under the CD(0, ∞)
condition,

(Pt f ) (f )
≤ Pt ,
Pt f f
that is Pt f L(log Pt f ) ≥ Pt (f L(log f )). Combined with (6.7.3), this immediately
yields that
2t
1+ L(log Pt f ) > 0 (6.7.8)
n
(if r ≥ s(1 + αtr) and r ≥ s, then necessarily 1 + αtr > 0).
Furthermore, we may deduce from (6.7.7) that
1 1
≤ (s) − LPt f ≤ ,
αs − ( (0) − LPt f )−1 α(t − s) − ( (t) − LPt f )−1
from which the bounds (6.7.4) and (6.7.5) immediately follow by integration be-
tween 0 and t.
Finally, any of (ii), (iii), (iv) yields CD(0, n) by a Taylor expansion at t = 0. The
proof is complete.
As for Proposition 5.5.3, p. 261, it is worth observing that (6.7.4) and (6.7.5) are
obtained directly from the curvature-dimension condition without the explicit use of
the gradient bound (3.3.18), p. 163, contrary to their weaker forms in Theorem 5.5.2,
p. 259. As such, this strategy of proof may be used in settings where gradient bounds
are lacking, for example in hypoelliptic settings where on the other hand extensions
of the curvature-dimension condition may be available.
Remark 6.7.4 If the condition CD(0, n) is replaced by the more general CD(ρ, n),
ρ ∈ R, the differential inequality on becomes
2[LPt f − (s)]2
(s) ≥ + ρ (s), s ∈ [0, t].
nPt f
This is not harder to solve, except that the corresponding result leads to expres-
sions involving trigonometric and logarithmic functions (depending on the sign
of ρ and the values of LPt f and Pt f ), and the final result is far less pleasant to
express. Fundamentally, such a differential equation (especially when ρ > 0) may
have no bounded solutions on [0, t] whenever the coefficients (here LPt f and Pt f )
are not controlled in some proper way with respect to the initial or final values
(0) and (t). When ρ = 0, the Li-Yau inequality expresses precisely this non-
explosion property, and the equivalent form under ρ = 0 may similarly be estab-
lished.
The proof of Theorem 6.7.3, in particular, (6.7.8), includes the celebrated Li-
Yau parabolic inequality. According to Remark 6.7.4 above, versions under the
curvature-dimension CD(ρ, n) condition are also available.
Corollary 6.7.5 (Li-Yau inequality) Under the curvature-dimension condition

CD(0, n), for every function f in Aconst+
0 , and every t > 0,
n
L(log Pt f ) > − .
2t
Alternatively,
(Pt f ) LPt f n
2
− < . (6.7.9)
(Pt f ) Pt f 2t
As above, in reasonable situations, the Li-Yau inequality extends to any strictly

positive bounded measurable function f . The Li-Yau inequality is actually an iden-
tity for the heat semigroup on Rn when f is a Dirac mass at x. The inequality is
thus optimal and expresses a comparison with the Euclidean model.
6.7.3 Harnack Inequalities
The Li-Yau inequality is one of the main tools in establishing Harnack-type in-
equalities for solutions of heat equations and Gaussian heat kernel bounds with
the correct dependence on dimension (on say Riemannian manifolds). As part of
the family of logarithmic Sobolev inequalities for heat kernel measures, it may be
qualitatively compared to Wang’s (infinite-dimensional) Harnack inequality (The-
orem 5.6.1, p. 265). For simplicity, we state the Harnack inequality in the setting
of a diffusion operator L = g − ∇W · ∇ on a complete connected Riemannian
manifold (M, g).
Corollary 6.7.6 (Harnack inequality under CD(0, n)) Under the curvature-
dimension condition CD(0, n), for every positive measurable function f on M,
every x, y ∈ M and every 0 < t < t + s,

t + s n/2 d(x,y)2 /4s
Pt f (x) ≤ Pt+s f (y) e , (6.7.10)
t
where d(x, y) is the Riemannian distance from x to y. Conversely, the Harnack
inequality (6.7.10), which holds for every x, y ∈ M and every 0 < t < t + s, implies
in return the Li-Yau inequality (6.7.9).
For the proof, as in Theorem 5.6.1, p. 265, let (xu )u∈[t,t+s] be a geodesic with
constant speed joining x to y and consider
2 /4s 2
(u) = un/2 eud(x,y) Pu f (xu ), u ∈ [t, t + s].
Then the Li-Yau inequality tells us precisely that the derivative of log is positive
on the interval [t, t + s] from which the claim follows. The converse statement is
established as the corresponding assertion in Theorem 5.6.1, p. 265.
Harnack inequalities as in Corollary 6.7.6 are traditionally used to produce sharp
off-diagonal Gaussian bounds on heat kernels of the form
pt (x, y) ≤ C(n, t, ε) e−d(x,y)

2 /(4+ε)t
for t > 0, ε > 0, (x, y) ∈ M × M, where C(n, t, ε) is of polynomial decay in time

depending on geometric features such as volumes of balls. We briefly mention here
some results of this type for the Laplacian heat kernels on a Riemannian manifold
(M, g) with dimension n and positive Ricci curvature (referring to the Notes and
References for suitable background). A first step in the investigation is that by inte-
gration of (6.7.10) over a ball,
Cn
pt (x, x) ≤ √
V (x, t )
for all t > 0 and x ∈ M, where V (x, r) is the volume of the ball with center x ∈ M
and radius r > 0, and Cn > 0 only depends upon n. Together with geometric volume
comparison theorems in manifolds with positive Ricci curvature, the latter may then
be shown to imply uniform heat kernel bounds in the form of ultracontractivity,
and thus functional inequalities of Sobolev-type (cf. Sect. 6.3). Off-diagonal upper
bounds are achieved using the techniques presented in Sect. 7.2, p. 355, in the next
chapter. A typical result in this context is that
Cn,ε
√ 1/2 e−d(x,y) /(4+ε)t
2
pt (x, y) ≤ √ 1/2 (6.7.11)
V (x, t ) V (y, t )
for all t > 0, ε > 0, (x, y) ∈ M × M.
Lower bounds may be investigated similarly, and improve with a polynomial fac-
tor the lower bound (5.6.3), p. 267. For example, Corollary 6.7.6 (for f approaching
Dirac mass at x) yields that

t + s n/2 d(x,y)2 /4s
ps (x, x) ≤ pt+s (x, y) e
t
for all s > 0 and t > 0. Provided the local asymptotics lims→0 (4πs)n/2 ps (x, x) = 1
holds, it follows that
1
e−d(x,y) /4t
2
pt (x, y) ≥ (6.7.12)
(4πt)n/2
for all t > 0 and (x, y) ∈ M × M.
These precise upper and lower bounds on heat kernels actually significantly im-
prove milder small time asymptotics at a logarithmic scale, which on the other
hand may be considered in greater generality. For example, the classical Laplace-
Varadhan small time asymptotics
lim 4t log pt (x, y) = −d 2 (x, y) (6.7.13)

t→0
holds for solutions of large classes of second-order differential operators on a Rie-

mannian manifold.
We refer to Sect. 6.12 for more on geometric and non-geometric heat kernel
bounds. In the context of this book, such heat kernel bounds will be investigated
under functional inequalities in Chap. 7.
6.7.4 Local Dimensional Hypercontractivity
The curvature-dimension condition CD(0, n) may also be used to reinforce, with

the dimensional factor n, Theorem 5.5.5, p. 262, on local hypercontractivity by
means of the dimensional logarithmic Sobolev inequality (6.7.4). Again, one ad-
vantage of local hypercontractivity is that it produces a description of the curvature-
dimension CD(0, n) condition which applies to any measurable function. An equiv-
alent form of CD(ρ, n), ρ = 0, could be established along the same lines, although
the result (as for Theorem 6.7.3) is far less simple to state.
Theorem 6.7.7 (Local dimensional hypercontractivity) Let (E, μ, ) be a Markov

Triple with semigroup P = (Pt )t≥0 . The following assertions are equivalent.
(i) The curvature-dimension condition CD(0, n) holds.
(ii) For all 1 < q1 < q2 < ∞, u1 , u2 > 0 and σ = q1 u2 − q2 u1 ≥ 0, and for all
positive measurable functions f on E,
1/q2 1/q1
Pu1 (Pσ f )q2 ≤ M n/2 Pu2 f q1 (6.7.14)
where
1−(1/q1 ) (1/q2 )−1 (1/q2 )−(1/q1 )
q1 − 1 q2 − 1 q1 u2 − q2 u1
M= .
u1 u2 q2 − q1
Proof We proceed as in the proof of Theorem 5.5.5, p. 262, and only outline the
main steps. Starting from the curvature condition (i), let f ∈ Aconst+
0 as usual, and
consider
1/q
(s) = Pu (Pt−s f )q , s ∈ [0, t], t > 0,
where now q : [0, t] → (1, ∞) and u : [0, t] → [0, ∞) are functions of s. Then, with
g = Pt−s f ,
q 2 q−1 q q2
= Ent p g + u − 1 Pu g q−1 Lg
q u
q
(6.7.15)
q2
+ u (q − 1)Pu g q−2 (g) .
q
Theorem 6.7.3 indicates that (i) is equivalent to the local logarithmic Sobolev in-
equality (6.7.4). By linearization, the latter amounts to the family of inequalities
with parameter κ > 0,
n
Entpu (f ) ≤ u LPu f + (κ − 1 − log κ)Pu f − u κPu f L(log f ) .
2
Now apply these inequalities to g q and compare with (6.7.15). Choosing
q = q(s), u = u(s) and κ = κ(s) solving the system
⎧q
⎪
⎨ (u − 1) + u(1 − κ) = 0
q
q (6.7.16)
⎪
⎩u(q − 1 + κ) + u (q − 1) = 0
q
2
we get that qq −1 ≤ n2 A(κ) with A(κ) = κ − 1 − log κ, κ > 0. Assuming that q
is decreasing, it follows that (s) ≤ enMs,t /2 (t) where
t
q (r)
Ms,t = − 2
A κ(r) dr.
s q(r)
6.8 Sobolev Inequalities Under a Curvature-Dimension Condition 305
In other words,
1/q(s) 1/q(t)
Pu(s) (Pt−s f )q(s) ≤ enMs,t /2 Pu(t) f q(t) .
In order to optimize the value of Ms,t , we choose q(r), u(r) and κ(r) on [s, t] of
the form
r +γ 1−q
q(r) = , u(r) = αr + β, κ= ,
αr + β 1 − αq
where α, β, γ are constants that we adjust to fit the values of q(0) = q2 and
q(t) = q1 (and for which u(0) = u1 and u(t) = u2 ) (and therefore q is decreas-
ing since it is monotonic). It is easily checked that those functions solve the sys-
tem (6.7.16). The conclusion then follows after tedious computations. As for The-
orem 6.7.3, (ii) yields CD(0, n) by a Taylor expansion as u1 goes to u2 and q1
to q2 .
The inequalities (6.7.14) in Theorem 6.7.7 are optimal in the sense that if L is the
Laplacian in Rn (therefore satisfying CD(0, n)), these inequalities are equalities for
2
fα (x) = eα|x| , x ∈ Rn , α ∈ R. There is also a version of Theorem 6.7.7 in the range
−∞ < q2 < q1 < 0 or 0 < q2 < q1 < 1.
6.8 Sobolev Inequalities Under a Curvature-Dimension

Condition
This section is devoted to Sobolev inequalities for the invariant measure under
curvature-dimension hypotheses. As a curvature condition CD(ρ, ∞) with ρ > 0
implies a logarithmic Sobolev inequality, a stronger curvature-dimension condition
CD(ρ, n) with ρ > 0 and n < ∞ is expected to ensure the validity of a Sobolev
inequality with exponent n.
6.8.1 Logarithmic Entropy-Energy Inequality Under CD(ρ, n)
It is not so easy to prove the Sobolev inequality directly. Again, a convenient tool is
to establish instead a logarithmic entropy-energy inequality (6.2.2). The following
theorem is a first main result in this direction. As usual, we consider a Markov Triple
(E, μ, ) under the curvature-dimension condition CD(ρ, n), ρ ∈ R, n ≥ 1,
1
2 (f ) ≥ ρ (f ) + (Lf )2
n
for every f ∈ A0 (or A) from Definition 3.3.14, p. 159. Recall that since ρ > 0 in
the statement below, the invariant measure μ is assumed to be finite and normalized
into a probability measure (cf. Theorem 3.3.23, p. 165).
Theorem 6.8.1 Under the curvature-dimension condition CD(ρ, n), ρ > 0,

1 ≤ n < ∞, the Markov Triple (E, μ, ), with μ a probability measure, satisfies
the tight logarithmic entropy-energy inequality

n 4
Entμ f 2 ≤ log 1 + E(f )
2 ρn

2
for every f ∈ D(E) with E f dμ = 1.
As announced, by Proposition 6.2.3, a Sobolev inequality Sn (A, C) of dimension

n holds under the curvature-dimension condition CD(ρ, n) with ρ > 0 and n < ∞.
Moreover, together with Proposition 6.6.1, it follows from Theorem 6.8.1 that under
the CD(ρ, n) condition with ρ > 0 and n < ∞, the diameter of (E, μ, ) is finite.
Finally, thanks to Corollary 6.4.1,
Corollary 6.8.2 Under the curvature-dimension condition CD(ρ, n) with ρ > 0

and n < ∞, the spectrum of the generator L is discrete and the Markov operators
Pt , t > 0, are Hilbert-Schmidt.
We turn to the proof of Theorem 6.8.1 which is similar to the proof of the loga-
rithmic Sobolev inequality under the CD(ρ, ∞) condition (Theorem 5.7.4, p. 270).
√
Proof of Theorem 6.8.1 Working with f = f0 , f0 ∈ Aconst+ 0 , the inequality to be
proved is

n 1
Entμ (f ) ≤ log 1 + f (log f )dμ (6.8.1)
2 ρn E

provided E f dμ = 1. Set then, as usual, (t) = Entμ (Pt f ), t ≥ 0. We already

know (cf. the proof of Proposition 5.7.3, p. 269) that for every t ≥ 0,

(t) = − Pt f (log Pt f )dμ and (t) = 2 Pt f 2 (log Pt f )dμ.
E E
By the curvature-dimension condition CD(ρ, n) applied to log Pt f , for every t ≥ 0,

2 2
(t) ≥ −2 ρ (t) + Pt f L(log Pt f ) dμ.
n E

Now, since E Pt f dμ = 1, by the Cauchy-Schwarz inequality, with g = Pt f to ease

the notation,
2 2
2
g L(log g) dμ ≥ g L(log g)dμ = g (log g)dμ .
E E E
The preceding leads to the differential inequality

2 2
≥ −2 ρ + .
n
In other words, the function

−2ρt 1 ρ
e −
n (t)
is increasing in t ≥ 0. Recalling that ≤ 0, it follows that for every t ≥ 0,

1 ρ 1 −1
− (t) ≤ ρ e 2ρt
− − .
n (0) n
Finally,

t n (0)
(0) − (t) = − (s)ds ≤ log 1 − 1 − e−2ρt .
0 2 ρn
As t → ∞, the announced tight logarithmic entropy-energy inequality (6.8.1) fol-
lows. Theorem 6.8.1 is established.
We next discuss some features of the logarithmic entropy-energy inequality of

Theorem 6.8.1. First, using the inequality log(1 + r) ≤ r, r ≥ 0, it yields the log-
arithmic Sobolev inequality LS( ρ2 ), or alternatively let n → ∞ (corresponding to
the curvature condition CD(ρ, ∞)). This is however not the optimal logarithmic
Sobolev inequality as demonstrated by Theorem 5.7.4, p. 270. On the other hand,
the inequality put forward in Theorem 6.8.1 is not the only one of that kind which
may be produced by a similar proof. For example, recall (5.7.11), p. 272,

ρn 4n − 1
eg 2 (g) dμ ≥ eg (g)dμ + eg (g)2 dμ.
E n−1 E 4(n + 2)2 E

If E eg dμ = 1, by Jensen’s inequality,
2
eg (g)2 dμ ≥ eg (g)dμ ,
E E
and the method of proof of Theorem 6.8.1 can then be used in the same way to get
this time the differential inequality
2ρ n (4n − 1) 2
≥ − + .
n−1 2(n + 2)2
In this way, we end up with the logarithmic entropy-energy inequality

2 p 4(n − 1)
Entμ f ≤ log 1 + E(f )
2 pρ n

2
for every f ∈ D(E) such that E f 2 dμ = 1, where p = 4(n+2)4n−1 . This inequality
is of weaker Sobolev dimension but yields this time the sharp logarithmic Sobolev
inequality.
More generally, every inequality of the form

eg 2 (g)dμ ≥ eg (g)dμ
E E

for every g ∈ A0 with E eg dμ = 1 would give, by the same principle, an entropy-

energy inequality

Entμ f 2 ≤ E(f ) , f ∈ D(E), f 2 dμ = 1,
E
for some growth function : R+ → R+ of the type studied in Chap. 7. However,

the best inequality of this type under the CD(ρ, n) condition does not seem to be
known.
6.8.2 Sobolev Inequality Under CD(ρ, n)
Via Proposition 6.2.3, Theorem 6.8.1 thus shows how the curvature-dimension con-
dition CD(ρ, n) with ρ > 0 and n < ∞ produces a Sobolev inequality Sn (A, C)
of dimension n. This Sobolev inequality may furthermore be tightened. However,
the constants given by this process are not sharp. Actually, the CD(ρ, n) condition,
ρ > 0, n < ∞, may indeed be shown to reach optimal Sobolev inequalities but using
a completely different technique of proof.
Theorem 6.8.3 (Sobolev inequality under CD(ρ, n)) Under the curvature-dimen-
sion condition CD(ρ, n), ρ > 0, 2 < n < ∞, the Markov Triple (E, μ, ),
with μ a probability measure, satisfies a Sobolev inequality Sn (C) with constant
4(n−1)
C = ρn(n−2) . That is,
4 n−1
f 2p ≤ f 22 + · E(f ) (6.8.2)
n(n − 2) ρ
for every f ∈ D(E) where p = 2n

n−2 .
It is not known how to prove this theorem using (linear) heat flow monotonic-
ity. In fact, the semigroup proof of Theorem 6.8.1 (for example) may be adapted
2n
to Sobolev inequalities up to some (rather mysterious) exponent 2 < p < n−2 . In
Sect. 6.11, alternative evolution equations (fast diffusion equations) will be devel-
oped in order to obtain Sobolev inequalities in the form of Gagliardo-Nirenberg
inequalities. Here, we use instead non-linear methods on extremal functions. We
first sketch the formal argument, and then provide the necessary technical details
needed to complete the proof.
Proof We begin with the inequality
f 2p − f 22 ≤ C E(f )

for all f ∈ D(E) with C > 0 the optimal constant. If it exists, let f
be a positive non-
constant extremal function for this inequality, normalized so that E f p dμ = 1, that
is so that
f 2p − f 22 = C E(f )
and for every u ∈ D(E),
f + u2p − f + u22 ≤ C E(f + u).
Changing u into εu and letting ε → 0 shows that

f p−1 u dμ − f u dμ = 0
E E

provided E(f, u) = E (−Lf )u dμ = 0. Therefore, f ∈ D(L) and must satisfy the

extremal function equation
f p−1 − f = −C Lf. (6.8.3)
Later, in the true proof, we shall show that such an extremal function is bounded
from above and below, justifying all the integrations below. Setting f = eg , the
equation becomes

e(p−2)g = 1 − C Lg + (g) . (6.8.4)
Then multiply both terms of the latter identity by ebg Lg, b ∈ R, and integrate with
respect to μ. By integration by parts, it follows that

(p − 2 + b) e(p−2)g ebg (g)dμ
E

=b ebg (g)dμ + C ebg (Lg)2 dμ
E E

−C ebg g, (g) dμ − C b ebg (g)2 dμ.
E E
The following makes use of various integral formulas on the 2 operator devel-
oped in the proof of Theorem 5.7.4, p. 270. For example, by (5.7.8), p. 271, with
2a = b,

3b
ebg (Lg)2 dμ = ebg 2 (g)dμ + ebg g, (g) dμ
E E 2 E

+ b2 ebg (g)2 dμ.
E
Therefore the preceding equation becomes

(p − 2 + b) e(p−2)g ebg (g)dμ = b ebg (g)dμ + C ebg 2 (g)dμ
E E E

3b
+C −1 ebg g, (g) dμ
2 E

+ C b(b − 1) ebg (g)2 dμ.
E
Now replace e(p−2)g by its expression from (6.8.4) and integrate by parts again to
get, after some algebra,

p−2
ebg 2 (g)dμ = ebg (g)dμ
E C E

b
+ p−1− ebg g, (g) dμ (6.8.5)
2 E

+ (p − 2)(b − 1) ebg (g)2 dμ.
E
On the other hand, arguing as for (5.7.10), p. 271, the CD(ρ, n) condition yields,
for every a ∈ R,

ρn
e 2 (g)dμ ≥
bg
ebg (g)dμ
E n−1 E

3b − 2(n + 2)a
+ ebg g, (g) dμ (6.8.6)
2(n − 1) E

b2 − 2ab − (n − 1)a 2
+ ebg (g)2 dμ.
n−1 E
It remains to compare (6.8.5) and (6.8.6), and to adjust

the parameters a, b ∈ R
in such a way that the coefficients in front of the terms E ebg (g, (g))dμ and
bg
E e (g) dμ coincide. Recalling that p = n−2 , this is achieved by choosing
2 2n
b n−1 2(n − 3)
a= − and b = .
2 n−2 n−2
As a consequence, it follows that

p−2 ρn
ebg (g)dμ ≥ ebg (g)dμ.
C E n − 1 E

bg p−2 ρn
Since E e (g) dμ > 0 (because f is assumed to be non-constant), C ≥ n−1 ,
or in other words,
4(n − 1)
C≤ .
ρ n(n − 2)
This bound concludes the announced Sobolev inequality.

This scheme of proof is rather simple, but unfortunately non-constant extremal
functions of Sobolev inequalities do not always exist. In order for the preceding
formal proof to work, it is necessary to deal with almost extremal functions at the
almost optimal exponent. To justify the procedure, we make use of the logarithmic
entropy-energy inequality of Theorem 6.8.1 (and thus of the Sobolev inequality of
dimension n but with worse constants).
2n
Consider 2 < q < n−2 (which will later converge to p = n−2
2n
) and ε > 0, and
denote by C(q, ε) > 0 the best constant in the Sobolev inequality
f 2q ≤ (1 + ε)f 22 + C(q, ε) E(f )
for every f ∈ D(E). As explained below, choosing q < n−2 2n

will ensure the ex-
istence of a function saturating this inequality, while ε > 0 will ensure that it
is not constant. Moreover, it will be shown that such an extremal function is
bounded from above and below, justifying all the integration by parts arguments
of the formal proof. The extremal function f , provided it exists, will then satisfy
f q−1 = (1 + ε)f − C(q, ε) Lf , and with the same arguments as above, one ends up
with the upper bound
(n − 1)(q − 2)
C(q, ε) ≤ .
nρ
The conclusion follows as q → p = n−2 2n
and ε → 0.
We are thus left with the question of the existence of such an extremal function.
First, recall from Sect. 6.4 that the embedding of D(E) into Lq (μ) is compact for
every q ∈ [2, p). Let then (fk )k∈N be a bounded sequence of functions in D(E)
which are almost optimal for the Sobolev inequality, that is such that, for every
k ≥ 1,
1
fk 2q ≥ (1 + ε)fk 22 + C(q, ε) E(fk ) −
k
(and fk 22 + E(fk ) = 1). Since E(|fk |) ≤ E(fk ), it may be assumed that fk ≥ 0
for every k. By the compactness property, there is a subsequence, still denoted by
(fk )k∈N , converging almost surely and in Lq (μ), and weakly in D(E), towards a
positive function f ∈ D(E). For this function, for any u ∈ D(E),

q−1
f − (1 + ε)f u dμ = C(q, ε) E(f, u).
This function f is therefore in the domain D(L) and satisfies the extremal function
equation
f q−1 = (1 + ε)f − C(q, ε) Lf.
Next, we make sure that f is bounded from above and below (by a strictly posi-
tive constant). To this end, observe that the equation for extremal functions may be
rewritten as
−1
−1 1+ε q−1
f = C(q, ε) Id −L f = C(q, ε)−1 Rλ f q−1 ,
C(q, ε)

∞
where λ = C(q,ε)
1+ε
and Rλ = (λ Id −L)−1 = 0 e−λt Pt dt is the resolvent operator.
For the upper bound, Corollary 6.3.3 already indicates that Rλ (f q−1 ) is bounded
as soon as f ∈ Lr (μ) for some r > (q − 1) n2 . But from the same argument, as

soon as f ∈ Lr (μ), then f ∈ Lr (μ) for any r < ar where a = (q − 1) − 2r n . Since
a < 1 as soon as r > (q − 2) 2 , by a simple iterative procedure, it follows that f is
n
bounded as soon as it belongs to Lq0 (μ) for some q0 > (q − 2) n2 . Since in our case
f ∈ Lq (μ) and q > (q − 2) n2 , it is indeed bounded. For the (strictly positive) lower
bound, Proposition
6.3.4 provides a lower bound on the kernel of Pt , and thus, say
for t ≥ 1, Pt g ≥ c E gdμ for some c > 0 and every positive function g. This in turn
implies that
∞
−λt Ce−λ
Rλ f q−1 ≥ Pt f q−1
e dt ≥ f q−1 dμ
1 λ E
from which the claim easily follows. Putting together the various pieces, the proof
of Theorem 6.8.3 is complete.
Remark 6.8.4 The preceding proof similarly shows that, under the curvature-
dimension condition CD(ρ, n) with ρ > 0 and 2 < n < ∞, for every f ∈ D(E),
f 2q − f 22 (n − 1)
≤ E(f ) (6.8.7)
q −2 ρn
for every 2 ≤ q ≤ n−2
2n
. Such an inequality will be called later in Sect. 7.6, p. 382,
a Beckner-type inequality. In this form, it actually holds for every 1 ≤ q ≤ n−2 2n
,
the value q = 1 corresponding to the Poincaré inequality, the value q = 2 to the
logarithmic Sobolev inequality (in the limit), and q = n−2
2n
to the Sobolev inequality.
By Proposition 6.2.2, any inequality in this family implies the Poincaré inequality.
Since on the sphere Sn , ρ = n − 1, the constant in (6.8.7) is optimal for any q, and
in particular Theorem 6.8.3 leads to the optimal Sobolev inequality (6.1.2) in this
case. When 1 ≤ n ≤ 2, inequality (6.8.7) may actually be considered for any q ≥ 1
and implies, as q → ∞, Moser-Trudinger-type inequalities of the form

(n − 1)
log e dμ − f dμ ≤
f
E(f ), f ∈ D(E). (6.8.8)
E E 2ρ n
Remark 6.8.5 Under the condition CD(ρ, n), for ρ > 0, 2 < n < ∞, the family of
inequalities (6.8.7) in the previous remark describes optimal Poincaré, logarithmic
Sobolev and Sobolev inequalities. The constant in (6.8.7) may actually be slightly
2n
improved together with the spectral gap whenever q < n−2 . Indeed denoting by λ1
6.9 Conformal Invariance of Sobolev Inequalities 313
the first non-trivial eigenvalue of −L (the spectrum of −L is discrete in this case),

(n−1) 1
ρ n in (6.8.7) may be replaced by κ where
ρn
κ =θ + (1 − θ )λ1
n−1
and
(q − 1)(n − 1)2
θ = θ (q) = ∈ [0, 1].
(q − 2) + (n + 1)2
Note that θ (p) = 1 for the critical exponent p = n−2
2n
while θ (1) = 0. In particular,
when q = 2 this corresponds to the logarithmic Sobolev inequality,
(n − 1)2 ρ n 4n
κ= · + λ1 .
(n + 1) n − 1 (n + 1)2
2
ρn
Since λ1 ≥ n−1 by (4.8.1), p. 213, these bounds improve upon (6.8.7). They are ob-
tained by a simple

modification

at the end of the (formal) proof of Theorem 6.8.3,
using that λ1 E (f )dμ ≤ E (Lf )2 dμ according to Proposition 4.8.3, p. 214. As a
geometric consequence, together with Obata’s Theorem, an n-dimensional Rieman-
nian manifold with Ricci curvature bounded from below by n − 1 for which κ = n
2n
for some q < n−2 (in particular if the logarithmic Sobolev constant is equal to n1 ) is
isometric to the n-sphere Sn (Obata’s Theorem tells us this is the case when λ1 = n).
6.9 Conformal Invariance of Sobolev Inequalities

In this section, we will learn how the preceding optimal Sobolev inequality on
the sphere (cf. (6.1.2) and Remark 6.8.4), established by means of the curvature-
dimension criterion in the abstract framework of a Markov diffusion operator in
Theorem 6.8.3, actually yields, and is equivalent to, the corresponding optimal
Sobolev inequalities on the Euclidean and hyperbolic spaces. The crucial property
used to reach this equivalence is conformal invariance, which is linked to the exis-
tence of extremal functions. As already mentioned in the previous section, extremal
(non-constant) functions for Sobolev inequalities at the critical exponent do not al-
ways exist (actually even for non-critical exponents). However, in the Euclidean
space or on the sphere, such extremal functions at the critical exponent do exist.
As discussed in this section, this property relies on particular transformations of
the underlying state space which preserve the metric (up to dilation). These are the
conformal maps, and (suitably modified) Sobolev inequalities are invariant under
such conformal maps.
6.9.1 Conformal Invariant
Conformal invariance may be presented in various ways. On the one hand, it may
be seen simply as the change on a Riemannian manifold (M, g) with dimension n
of the co-metric g into c2 g, or of the operator into c2 , where c is some given

(strictly positive) function. At the same time, the measure μ associated with the
model has to be changed into c−n μ (which is the case when dealing with Laplace-
Beltrami operators in dimension n and the Riemannian measure defined from the
metric). There is actually a way (using the scalar curvature) to present the classical
Sobolev inequalities on the three model spaces as invariant under such a family of
transformations. On the other hand, such transformations may sometimes be pro-
duced by maps ψ : M → M such that (f ◦ ψ) = (c2 (f )) ◦ ψ. In this case, the
new metric is just the former one viewed through a change of coordinates. Such
a map ψ is called a conformal map, and usually induces extra information on the
conformal transformation itself.
This section presents the basic material on conformal transformations and
Sobolev inequalities, from which extremal functions (and further properties) will
be extracted. This will be achieved in the context of a smooth (at least C 2 ) Rie-
mannian manifold (M, g) (or some open set in M) of dimension n. The reader may
however keep in mind the three model spaces Rn , Sn and Hn . Sobolev inequalities
will always be understood to hold for the class A0 of smooth (C ∞ ) compactly sup-
ported functions on M. Recall in addition the class A of smooth (C ∞ ) functions
on M.
In this setting, we begin by recalling the notion of scalar curvature scg on (M, g)
(see (C.3.2), p. 508).
Definition 6.9.1 (Scalar curvature) The scalar curvature scg (x) at a point x ∈ M is
the trace of the Ricci tensor. In a local coordinate system,

n
scg (x) = g ij (x) Ricij (x)
i,j =1
where (Ricij (x))1≤i,j ≤n is the Ricci tensor at the point x ∈ M.
On the sphere Sn ⊂ Rn+1 with dimension n and constant Ricci curvature

Ricij = (n − 1)gij , scg (x) = n(n − 1) at any point. Similarly scg (x) = 0 on Rn
and scg (x) = −n(n − 1) on the hyperbolic space Hn . In particular, the (optimal)
Sobolev inequality on the sphere (6.1.2) at the critical exponent p = n−2
2n
with re-
spect to the uniform probability measure μ and the associated Dirichlet form E may
therefore be rewritten as

4 (n − 2)
f 2p ≤ scg f 2 dμ + E(f )
n(n − 2) 4(n − 1) Sn
for every f ∈ D(E). Such a formulation of the Sobolev inequality is actually
generic, and emphasizes the scalar curvature as a so-called conformal invariant.
This conformal transformation may be presented in a general context, of indepen-
dent interest. Consider the framework

of a Markov Triple (E, μ, ) with generator
L and Dirichlet form E(f ) = E (f )dμ. The n-conformal class, n ≥ 2 (it is possi-
ble to restrict to n > 2, but n = 2 is a special important case, even if many formulas
have to be adjusted), of (E, μ, ) is the set of all Markov Triples (E, c−n μ, c2 ) for
all strictly positive functions c ∈ A. An n-conformal invariant of the n-conformal
class of (E, μ, ) is a map S = S(μ, ) : E → R depending only on μ and such
that, for any function c = eτ ∈ A,

−n n−2 n−2
S c μ, c = c S(μ, ) +
2 2
Lτ − (τ ) . (6.9.1)
2 2
With this definition, the following statement holds.
Proposition 6.9.2 Setting p = n−2

2n
as usual, n > 2, let S(μ, ) be an n-conformal
invariant map. Then, the Sobolev inequality

f p ≤ C
2
S(μ, )f dμ + E(f ) , f ∈ D(E),
2
E
for some C > 0, is invariant in the n-conformal class of (E, μ, ). That is, if it holds
for the pair (μ, ), then it holds with the same constant C for the pair (c−n μ, c2 )
for every strictly positive function c ∈ A.
Proof The proof is immediate. It suffices to change f into c(2−n)/2 f in the Sobolev
inequality and, after the use of (6.9.1), to integrate by parts the mixed term in the
change of variables formula for the operator

n − 2 2 (n − 2)2 2
c(2−n)/2 f = c2−n (f ) − f ,τ + f (τ ) .
2 4
Note that by the change f → c(2−n)/2 f used in Proposition 6.9.2, the generator
L associated with (E, μ, ) is transformed for (E, c−n μ, c2 ) with c = eτ into

L = c2 Lf − (n − 2) (τ, f ) . (6.9.2)
The formula assumes a particularly nice form when n = 2.
6.9.2 Geometric Conformal Invariant
According to Proposition 6.9.2 above, the aim here will be to show that, in a Rie-
mannian setting, the scalar curvature is such an n-conformal invariant (thus satis-
fying (6.9.1)), where n is the dimension of the manifold. We refer to Appendix C
for the necessary Riemannian background. In Riemannian geometry, when g is the
(co-) metric and dμg the Riemannian volume element, the change of into c2
and dμg into c−n dμg corresponds to a conformal change of the metric g into c2 g.
n g =ij(g ), that is when

Recall that, for a Laplacian associated to the (co-) metric ij
the carré du champ operator takes the form (f ) = i,j =1 g ∂i f ∂j f , the Rie-
mannian measure is given in local coordinates by dμg = det(g)−1/2 dx, where dx
is the Lebesgue measure in the coordinate system. (Observe that, contrary to our
usual convention, it is not normalized whenever it is finite.) The aim is to examine
how the scalar curvature is modified by such a conformal transformation. To this
end we first study the Ricci curvature. Two approaches are available. One via the
2 -calculus for the new operator, and another in local coordinates. Both are heavy
and somewhat cumbersome, but it might be easier to work with the 2 operator,
which we do below.
While the case of a Laplacian will be our main illustrative example, the 2 for-
mulas which need to be developed are just as easy to describe in the general case of
a generator of the form L = g − ∇W · ∇ for some smooth potential W on a Rie-
mannian manifold (M, g) with symmetric invariant measure dμ = e−W dμg . This
setting will furthermore be suited to our later investigation of non-geometric exam-
ples. We thus discuss the 2 operations below in this more general setting. In this
instance, recall from (1.16.4), p. 71 (cf. also Sect. C.6, p. 513), that the 2 operator
is given on smooth functions by
2 (f ) = |∇∇f |2 + Ric(L)(∇f, ∇f )
where Ric(L) is a symmetric tensor defined from the Ricci tensor Ricg of the Rie-
mannian manifold (M, g) by Ric(L) = Ricg +∇∇W .
In this setting, the first step will be to compute, for any n and any smooth function
c, the Ricci tensor associated with (c−n μ, c2 ) from the Ricci tensor associated to
(μ, ). Although most interesting formulas will be achieved when n is the dimen-
sion of M, assume for the moment that it is distinct from this dimension, denoted
here by n0 . The computations below are performed as usual on the class A0 of
smooth compactly supported functions on M.
Given c written in the form c = eτ , the 2 operator associated with c2 L may be
expressed in terms of the 2 operator of L as

c4 2 (f ) + 2 τ, (f ) + L τ + 2 (τ ) (f ) − 2 Lf (f, τ ) .
According to Proposition 6.9.2, and for a generic n, the aim is to modify this formula
in terms of the operator L of (6.9.2), which thus requires us to replace 2 (f ) by
2 (f ) + (n − 2)∇∇τ (∇f, ∇f ) and L by L − (n − 2)(τ, ·) in the preceding. Hence,
denoting by 2 the 2 operator of
L,
(

2 (f ) = c4 2 (f ) + (n − 2)∇∇τ (∇f, ∇f ) + 2 τ, (f )
)
+ L τ − (n − 4) (τ ) (f ) − 2 Lf (f, τ ) + 2(n − 2) (f, τ )2 .
The second order terms are gathered in the expression

= ∇∇f + 2∇f # ∇τ − (f, τ ) g
∇∇f
where ∇f # ∇τ denotes the symmetric tensor product, that is

1
(∇f # ∇τ )(∇g, ∇h) = (f, g) (τ, h) + (f, h) (τ, g) .
2
Hence
(
|2 = c4 |∇∇f |2 + 2 τ, (f ) + 2 (f ) (τ )
|∇∇f
)
+ (n0 − 2) (f, τ )2 − 2g f (f, τ ) .
In this formula, n0 (the effective dimension of the manifold) appears as the norm,
in the space of symmetric matrices, of the identity matrix in the form of the term
(f, τ ) g. The term g f appears as the scalar product, still in the space of symmet-
ric matrices, of ∇∇f and g, that is the trace of ∇∇f . Lastly, comparing the two
formulas,
(
Ric(L)(∇f, ∇f ) = c4 Ric(L)(∇f, ∇f ) + (n − 2)∇∇τ (∇f, ∇f )

+ L τ − (n − 2)(τ ) (f ) + 2 ∇W (f ) (f, τ )
)
+ (2n − n0 − 2) (f, τ )2 .
In other words, in terms of Ricci tensors with lower indices, that is as symmetric
operators acting on the tangent space,

Ric(
L) = Ric(L) + (n − 2) ∇∇τ + L τ − (n − 2) (τ ) g
+ 2 ∇W # ∇τ + (2n − n0 − 2)∇τ # ∇τ.
This formula for the Ricci curvature of

L is rather heavy, but considerably sim-
plifies when n = n0 is the dimension of the manifold and L = g is a Laplacian
(W = 0), and it is in this setting that we shall mainly apply the formula. Indeed,
after identification,

Ric( g ) = Ricg +g g τ + (n − 2) ∇∇τ + ∇τ # ∇τ − (τ )g .
Taking the trace to reach scalar curvature, we get the following formula for the scalar
curvature scg of
L in terms of scg ,

scg = c2 scg + (n − 1) 2g τ − (n − 2) (τ ) .
The latter may also be rewritten, recalling c = eτ , as

4(n − 1) (n−2)τ/2 −(n−2)τ/2
scg = e scg −
2τ
e g e . (6.9.3)
n−2
n−2
This equation (6.9.3) shows that the function 4(n−1) scg satisfies the defining condi-
tion (6.9.1) of an n-conformal invariant, which is the announced claim.
Observe that when n is the dimension of the manifold (which is the case here), the
n-conformal class of a given Laplacian only contains Laplacians. Note furthermore
that, setting f = e−(n−2)τ/2 and as usual p = n−2
2n
, n > 2, (6.9.3) can be rewritten as
4(n − 1)
f p−1 scg = f scg − g f. (6.9.4)
n−2
As a consequence of these developments and of Proposition 6.9.2, we may now
state the following main conclusion.
Theorem 6.9.3 (Conformal invariant Sobolev inequality) Let g be the Laplace-

Beltrami operator on a Riemannian manifold (M, g) with dimension n (> 2)
equipped with the Riemannian measure μg , with associated carré du champ
operator
(f ) = |∇f |2 (on smooth functions f : M → R) and Dirichlet form
E(f ) = M (f )dμg . Then the scalar curvature scg is an n-conformal invariant
of the Markov Triple (M, μg , ). That is, if the Sobolev inequality

n−2
f p ≤ C
2
scg f dμ + Eg (f ) , f ∈ D(E),
2
4(n − 1) M
holds, where p = n−2

2n
, then the same inequality holds when μg is changed into
−n 2
c μg and into c where c is a smooth strictly positive function.
6.9.3 Conformal Invariance and Sharp Sobolev Inequalities
The preceding Theorem 6.9.3 has important consequences, in particular, as an-

nounced, it leads to optimal constants and extremal functions in the Sobolev inequal-
ities on the three model spaces. Denote below by ωn the volume (for the uniform
measure) of the standard sphere Sn in Rn+1 .
Theorem 6.9.4 (Optimal Sobolev inequalities on the model spaces) For the Rie-
mannian measures μ on the three model spaces Rn , Sn , Hn (not normalized on
Sn ), n > 2, the following optimal Sobolev inequalities with p = n−2
2n
hold (for every
function f in the respective Dirichlet domains D(E)).
(i) On the Euclidean space Rn ,
4
f 2p ≤ 2/n
E(f ).
n(n − 2) ωn
(ii) On the sphere Sn ,

1 4
f 2p ≤ 2/n
f 2 dμ + 2/n
E(f ).
ωn Sn n(n − 2) ωn
(iii) On the hyperbolic space Hn ,

1 4
f 2p ≤ − 2/n
f 2 dμ + 2/n
E(f ).
ωn Hn n(n − 2) ωn
Remark 6.9.5 For the Sobolev inequality in Rn (case (i)), the L2 (dx) norm of the
function f does not appear in the statement. However, the fact that f ∈ D(E) (and
therefore that f ∈ L2 (dx)) may not be removed without care since, for example,
the inequality cannot hold for a non-zero constant function. However, through stan-
dard localization and symmetrization procedures in Euclidean space, the Sobolev
inequality (i) may be extended to any smooth function f such that |∇f | ∈ L2 (dx)
and such that voln (|f | ≥ ε) < ∞ for any ε > 0.
Proof We start from the optimal Sobolev inequality on the sphere Sn established in
Theorem 6.8.3, rewritten here under the un-normalized uniform measure. Now, it
was observed in Sect. 2.2.2, p. 83, that in the stereographic projection, the metric of
the sphere is a conformal transformation of the metric of the Euclidean space Rn .
The scalar curvature of Rn is 0, and the resulting optimal Sobolev inequality on Rn
is thus obtained from Theorem 6.9.3. By the conformal invariance principle, these
two Sobolev inequalities are therefore equivalent. One can add to the equivalence
the Sobolev inequality on the hyperbolic space Hn since Hn may be identified with
Rn−1 × (0, ∞) with a metric given as a conformal transformation of the Euclidean
metric with constant scalar curvature −n(n − 1). The proof is complete.
A further important property and application drawn from Theorem 6.9.3 con-
cerns extremal functions. Assume that the two metrics given by and c2 have
constant scalar curvature equal to n(n − 1). Then, the equation (6.9.4) of conformal
transformation of scalar curvature indicates that the function f = c−(n−2)/2 satisfies
the equation
4 Lf
f p−2 = 1 − ,
n(n − 2) f
which is precisely Eq. (6.8.3) of extremal functions for the Sobolev inequality on the
sphere with constant C = n(n−2)4
. Now, there exist conformal maps on the sphere,
that is diffeomorphisms from Sn onto itself, such that the image carré du champ
operators are of the form c2 . But under such a conformal map, the metric c2 is
nothing else but the metric viewed under a change of coordinates. The new metric
has constant scalar curvature (since curvature is independent of the coordinates and
invariant under diffeomorphisms). These observations thus explain the existence of
non-constant extremal functions for the Sobolev inequality at the critical exponent
on the sphere. (It may be shown that there are no non-constant extremal functions
for the Sobolev inequality (6.8.7) below the critical exponent.)
It might be useful to briefly describe the conformal maps from Sn into itself. Let
Q be a point in Rn+1 exterior to the sphere Sn . To each I ∈ Sn associate the other
intersection point J between Sn and the line QI . The map I → J is then such a
conformal map as a consequence of the following three observations:
(i) The inversions of Rn+1 are conformal transformations of Rn into itself.

(ii) The map I → J is the restriction to Sn of the inversion (with center Q and
inversion sphere orthogonal to Sn ).
(iii) The metric of Sn is that of Rn+1 as an embedded manifold.
These maps may also be interpreted as dilations under stereographic projec-
tion.
The same reasoning may be applied to the passage from Rn to Sn . By stereo-
graphic projection (cf. Sect. 2.2.2, p. 83), the spherical carré du champ operator

is given in terms of the (standard) Euclidean carré du champ operator by

1 + |x|2 2

= .
2
2
Comparing the scalar curvatures, setting c(x) = 1+|x|
2 and f = c−(n−2)/2 , it follows
that
4
f p−1 = − f
n(n − 2)
which is the equation of the extremal functions for the optimal Sobolev inequality on
Rn . By translation, dilation, and homogeneity, all extremal functions of the Sobolev
inequality in Rn are of the form
−(n−2)/2
fσ,b,x0 (x) = σ 2 + b |x − x0 |2 , σ > 0, b > 0, x0 ∈ Rn . (6.9.5)
It should be observed that such functions are not in L2 (dx) when n = 3, 4. However,
they satisfy the equation for extremal functions and the Euclidean Sobolev inequal-
ity (i) applies to them thanks to Remark 6.9.5. These functions will appear naturally
when solving fast diffusion equations in Rn (see Sect. 6.11 below) justifying the
role played by those equations in connection with Sobolev inequalities.
Similarly, on the sphere Sn , extremal functions for the optimal Sobolev inequality
((ii) in Theorem 6.9.4) may be written as
−(n−2)/2
fb,e (x) = σ + b (e · x) , (6.9.6)
where x, e ∈ Sn , σ > 1, b ∈ R and σ 2 = 1 + b2 .

In the same way, extremal functions for the symmetric Jacobi operator on
[−1, +1] with invariant reversible measure dμn (x) = Cn (1 − x 2 )(n/2)−1 dx
(Sect. 2.7.4, p. 113) for the Sobolev inequality with exponent p = n−2 2n
are
fb (x) = (σ + bx) −(n−2)/2 , where σ = 1 + b . Although it is quite easy in this
2 2
latter case to check that the extremal function equation

(6.8.3) pis satisfied with con-
stant C = n(n−2)
4
, it is not completely obvious that [−1,+1] fb dμn = 1 (to ensure
that it is indeed an extremal function). But this is easily obtained using the change
of variable y = 1+αx
α+x
, |α| < 1, which leads to
+1 +1

2 (n/2)−1
n/2 (n/2)−2 dy
1−x dx = 1 − α 2 1 − y2
−1 −1 (1 − αy)n
from which the conclusion follows with α = − σb . Observe that, for n integer, n > 2,
this change of variable from x to y is the trace on zonal functions of the conformal
transformation of the sphere.
6.9.4 Non-geometric Conformal Invariant
The conformal invariance result of Theorem 6.9.3 is not restricted to the Laplacian
case and to the case where n is the dimension of the manifold. In fact, other general
(and sometimes useful) forms of conformal invariance may be considered once a
good candidate to replace the scalar curvature is determined. In the last (and some-
what technical) part of this section, we discuss a few relevant instances of interest
on the basis of the operations on 2 already developed for Theorem 6.9.3.
In the following examples, we restrict ourselves to the ncase of elliptic opera-
tors on some n0 -dimensional manifold M. Set (f ) = i,j0 =1 g ij ∂i f ∂j f where
g = (g ij ) is elliptic, and as usual denote respectively by g , μg and scg the Laplace
operator, Riemannian measure and scalar curvature associated with the (co-) metric
g. Using this notation we look at measures μ which have a density e−W with respect
to μg . The following non-geometric conformal invariance result illustrates a variety
of examples.
Proposition 6.9.6 In the preceding Riemannian context, for n > n0 ,

dμ = e−W dμg and any α ∈ R, the map

Sα (μ, ) = γn (α) scg − αW + βn (α) (W ) (6.9.7)
is n-conformal invariant where

α(n − 2n0 + 2) − 2(n0 − 1)
βn (α) =
2(n − n0 )
and
n−2
γn (α) = .
4(n0 − 1) − α(n − n0 )
Proof It is enough to check that Sα (μ, ) satisfies (6.9.1). The transformation rules
for scg under the change (μg , ) → (c−n0 μg , c2 ) was already considered previ-
ously. It remains to observe that if the density of μ is e−W for (μg , ), the density
of c−n μ with respect to c−n0 μg is equal to e−W cn0 −n = e−W −(n−n0 )τ . The compu-
tations are then straightforward (although somewhat tedious).
In fact Proposition 6.9.6 applies for any n = n0 , but in view of Sobolev in-
equalities, it is only interesting for n > n0 . As alluded to above, restricting to
Laplace operators leads to the restriction of conformal transformations to n = n0 .
Indeed, the change of μ into c−n μ changes a Riemannian measure into a measure
which is not the Riemannian measure associated to c2 as soon as n = n0 . When

dμ = e−W dμg , for a conformal change c where c is a function of W (which is often
the case), then there are wider classes of conformal invariant maps. For non-elliptic
operators, neither the Laplace operator nor the scalar curvature are defined. There is
no known general conformal invariant map in this setting.
The non-geometric conformal invariance result of Proposition 6.9.6 may also be
used to obtain optimal Sobolev inequalities in situations where the n-dimensional
parameter of the Sobolev inequality does not correspond to the dimension of the un-
derlying manifold. The method may actually be used to extend the correspondence
between spheres and Euclidean spaces for non-integer values of the dimensional
parameter.
Consider, for example, the Sobolev inequality in the Euclidean space Rn . Writing
R = Rn0 −1 × Rn−n0 +1 , this inequality may be considered on functions which are
n
radial in the variable y ∈ Rn−n0 +1 . We then get a Sobolev inequality on Rn0 −1 × R+

for the Euclidean carré du champ operator in Rn0 −1 × R+ , but for the weighted
measure ωn−n0 +1 r n−n0 dxdr. This Sobolev inequality is still valid, with optimal
constants, when n is no longer an integer provided the definition of ωn is suitably
adapted.
To reach this claim, recall that the Sobolev inequality in Rn may be deduced
from the Sobolev inequality on the sphere via stereographic projection, and that the
Sobolev inequality on the sphere may be established from the abstract curvature-
dimension condition CD(n − 1, n). We may therefore try to follow the same
path to get this new inequality when n ∈ / N. To start with, as was mentioned in
Sect. 1.16.2, p. 72, on the unit n0 -dimensional sphere x12 + · · · + xn20 +1 = 1, the
coordinate function U (x) = x1 satisfies ∇∇U = −U Id, and hence the operator
Sn0 + (n − n0 )∇ log U satisfies the CD(n − 1, n) condition on the half-sphere
{x1 > 0}. Therefore, a tight Sobolev inequality holds on this half-sphere, with
4
dimension n and constant n(n−2) . The measure on this half-sphere has density
n−n
cn,n0 x1 0 with respect to the Riemannian measure of the sphere (cn,n0 being the
normalizing constant). Next, consider the n-conformal map Sα (μ, ) of Proposi-
0 −1)
tion 6.9.6 with the choice of α = 2(n
n−2n0 (corresponding to (n − n0 )β(α) = α).
An easy computation shows that Sα (μ, ) = n(n−2) 4 (in other words, for this value
of α, the corresponding scalar curvature is constant for this weighted manifold).
Therefore, with this weighted measure, we are in a situation similar to that for the
Riemannian measure, except that n is no longer the dimension of the sphere, but the
dimensional parameter of the Sobolev inequality.
To transfer this conclusion from the spherical model to the Euclidean model,
apply the stereographic projection of the sphere on Rn0 from the north pole
(0, . . . , 0, 1). The image of the previous half-sphere becomes the half space
{x1 > 0}. Under the stereographic projection, the function U = x1 takes the form
2x1
U (x) = 1+|x| 2 . Hence, the sharp Sobolev inequality on the half-sphere is trans-
formed into a sharp n-dimensional Sobolev inequality on Rn0 −1 × R+ , with carré

6.10 Gagliardo-Nirenberg Inequalities 323
du champ operator
2
1 + |x|2
(f ) = |∇f |2
2
and measure
n−n
2n x1 0
dν(x) = · dx
ωn,n0 (1 + |x|2 )n
where ωn,n0 is the normalizing constant (so that ν is a probability measure). The
resulting inequality takes the form, with p = n−2
2n
,
2/p
|f |p dν
Rn0 −1 ×R+

4
≤ Sα (μ, ) f 2 dν + (f )dν .
n(n − 2) Rn0 −1 ×R+ Rn0 −1 ×R+
2
Apply the n-conformal invariance result to the function c(x) = 1+|x|2 to obtain a
new n-dimensional inequality on Rn0 −1 × R+ , but with the usual carré du champ
n−n
operator |∇f |2 and the measure x1 0 dx. For this pair, and for the same α, Sα = 0
so that, finally
n−n
x1 0 dx 2/p
|f | p
Rn0 −1 ×R+ ωn,n0
(6.9.8)
n−n
4 x 0 dx
≤ |∇f |2 1 .
n(n − 2) Rn0 −1 ×R+ ωn,n0
This Sobolev inequality is precisely the announced one. In this non-geometric pic-
ture, the quantity Sα behaves in the same way as the scalar curvature for Laplace
operators as it is constant strictly positive on the sphere and vanishes on Rn .
Further connections between Sobolev inequalities, extremal functions and met-
rics with constant scalar curvature may be developed along these lines. For example,
as already mentioned, there is a strong connection, through (6.9.4), between confor-
mal transformations of a space into itself which preserve constant scalar curvature
and extremal solutions of Sobolev inequalities. Furthermore, Sobolev inequalities
may be used to control functions c for which the metric associated with c2 has
constant scalar curvature.
6.10 Gagliardo-Nirenberg Inequalities
Up to now, Sobolev inequalities on spheres under the curvature-dimension condition

CD(ρ, n) were established by means of the heat semigroup (Pt )t≥0 , that is, by solv-
ing the differential equation ∂t u = Lu. The sharp constants are however obtained by
some variant, nevertheless still dealing with the infinitesimal Markov generator L
of (Pt )t≥0 . As discussed in the preceding Sect. 6.9, together with conformal invari-
ance, the sharp Sobolev inequality on the sphere then yields the optimal Sobolev
inequalities in Euclidean and hyperbolic spaces. In this picture, the spherical case,
with its sharp constants, and the curvature-dimension condition, appear as critical.
However, this methodology has not yet produced any type of Sobolev inequal-
ity comparable to the Euclidean inequality directly from the curvature-dimension
condition CD(0, n), and similarly for the hyperbolic space under CD(ρ, n) with
ρ < 0. One explanation for this is the following. When ρ > 0, the invariant mea-
sure is finite (cf. Theorem 3.3.23, p. 165) and may thus naturally be normalized into
a probability measure. It is actually from this normalization that optimal constants
are produced. By conformal invariance, this normalization is also reflected in the
sharp Sobolev inequalities on Rn and Hn . However, when the invariant measure
has infinite mass, there is no natural normalization, and hence when the measure is
multiplied by a (strictly positive) constant, the constants in the Sobolev inequality
are modified. Therefore, there is no hope of directly obtaining Sobolev inequali-
ties under curvature-dimension conditions without any further information on the
measure.
In this section, we examine how to suitably modify Sobolev inequalities in Eu-
clidean space in order to obtain new families of inequalities which will be shown in
the next section to be accessible under the curvature-dimension condition CD(ρ, n)
for ρ ∈ R, via non-linear evolution equations. With this task in mind, we first ex-
tend the Sobolev and Nash inequalities to the more general family of Gagliardo-
Nirenberg inequalities, and then consider entropic formulations of the latter.
6.10.1 Gagliardo-Nirenberg Inequalities
Classical Gagliardo-Nirenberg inequalities in Euclidean space may be defined in

the abstract context of Markov Triples (E, μ, ). The definition below may be con-
sidered for any p > 2 but as for Sobolev inequalities we emphasize a dimensional
parameter n (> 2) as p = n−2
2n
.
Definition 6.10.1 (Gagliardo-Nirenberg Inequality) Let n > 2, p = n−2 2n

, and q, s
be such that 1 ≤ s ≤ q ≤ p. The Markov Triple (E, μ, ) is said to satisfy a
Gagliardo-Nirenberg inequality GNn (q, s ; A, C) with dimension n, parameters
q, s, and constants A ≥ 0, C > 0, if for all functions f in D(E) ∩ Ls (μ),
θ/2
f q ≤ A f 22 + C E(f ) f s1−θ
where θ ∈ [0, 1] is such that

1 θ 1−θ
= + .
q p s
For simplicity, only A ≥ 0 is considered here although the formal definition in-
cludes A ∈ R. The Sobolev inequality Sn (A, C) of Definition 6.2.1 belongs to the
Gagliardo-Nirenberg family for θ = 1 or q = s, while the Nash inequality (6.2.3)
is achieved for q = 2, s = 1. As for the Nash inequality (cf. Proposition 6.2.3), any
inequality GNn (q, s ; A, C) is equivalent, up to constants, to the Sobolev inequality
Sn (A, C) with the same dimension n.
Proposition 6.10.2 In the preceding setting, the Sobolev inequality Sn (A, C)

implies the Gagliardo-Nirenberg inequality GNn (q, s ; A, C). Conversely, the
Gagliardo-Nirenberg inequality GNn (q, s ; A, C) implies a Sobolev inequality
Sn (A , C ).
Proof The proof is very similar to the proof of Proposition 6.2.3. For the first im-
plication, use again the concavity of the function r → log(f 1/r ), but this time
between the points s ≤ q ≤ p. Conversely, use the slicing procedure, applying
GNn (q, s ; A, C) to (f − 2k )+ ∧ 2k , k ∈ Z. The steps of the proof are similar, al-
though a bit heavier.
By concavity of the logarithm, a Gagliardo-Nirenberg inequality
GNn (q, s ; A, C) implies the inequality

f 2q ≤ θ A f 22 + C E(f ) + (1 − θ )f 2s , f ∈ D(E) ∩ Ls (μ). (6.10.1)
In Rn , there exist Gagliardo-Nirenberg inequalities GNn (q, s ; 0, C) (due for ex-
ample to Proposition 6.10.2), and we will be interested in sharp constants for some
values of the parameters q, s. But observe first that in this Euclidean (homogeneous)
case, the weak form (6.10.1) is as good as the original one. Indeed, changing f (x)
into fr (x) = f (rx), x ∈ Rn , r > 0, in (6.10.1) with A = 0, and after a change of
variables, we get
2n( q1 − p1 ) 2n( q1 − 1s )
f 2q ≤ C θ r E(f ) + (1 − θ ) f 2s r .
Optimizing in r > 0 returns the inequality GNn (q, s ; 0, C) with the same con-
stant C.
There is another, more subtle, way to linearize Gagliardo-Nirenberg inequalities
which will turn out to be most useful below.
Proposition 6.10.3 For any α > 0, α = 1, such that θ α < 1, the Gagliardo-
Nirenberg inequality GNn (q, s ; A, C) is equivalent to the inequality

q ≤ αθ A f 2 + C E(f ) + (1 − αθ ) f s , f ∈ D(E) ∩ L (μ), (6.10.2)
f 2α 2 2β s
α(1−θ)
where β = 1−θα .
Proof Setting Q = A f 22 + C E(f ), the Gagliardo-Nirenberg inequality

GNn (q, s ; A, C) takes the form

f 2q α Q αθ
≤ .
f 2s f 2s
As θ α < 1, the function r αθ , r ≥ 0, is concave and thus, for any r0 > 0,
r αθ ≤ α θ r0αθ−1 r + (1 − αθ )r0αθ .
2γ
Choose then r0 = f s with γ = α−1
1−αθ . Conversely, replace f by cf and optimize
in c > 0.
It should be observed that constants are preserved in the equivalences of Propo-

sition 6.10.3, and thus so are extremal functions, when they exist.
It is quite remarkable that there is a sub-family of Gagliardo-Nirenberg inequal-
ities in Rn (for the Lebesgue measure and the standard carré du champ operator
and Dirichlet form E) with sharp constants. This is the content of the next statement.
Extremal functions will be determined at the same time.
Theorem 6.10.4 (Del Pino-Dolbeault Theorem) Let ν > n > 2, and set
2ν 2(ν − 1)
q= , s=
ν −2 ν −2
so that 2 < s < q < p = n−2

2n
and q = 2(s − 1). Then, on Rn , the optimal Gagliardo-
Nirenberg inequality GNn (q, s ; 0, C) holds. That is, for every f ∈ D(E) ∩ Ls (μ),
f 2q ≤ C E(f )θ f s2(1−θ)
where q1 = pθ + 1−θ s and where the (optimal) constant C > 0 is the one for which
there is equality for the function
−(ν−2)/2
f (x) = 1 + |x|2 , x ∈ Rn .
The limiting case ν → n, θ → 1, corresponds to the standard Sobolev inequality

in Rn with its sharp constant and extremal functions (6.9.5) of the same form. As
for the case of the Sobolev inequality, these extremal functions are not in L2 (dx)
when 2ν ≤ n + 4, but the same kind of extension to various classes of functions as
the one described in Remark 6.9.5 may be developed similarly.
Proof Although the result is true as stated for ν ≥ n > 2, the method of proof devel-
oped here only works for ν ≥ n + 12 , n > 2, to which we restrict below. It is however
a quite general strategy to deduce sharp inequalities from other optimal inequalities
provided there exist extremal functions. We first deal with the case ν = n + m2 for
some integer m ≥ 1 and start from the optimal Sobolev inequality on Rn+m ,
g2p ≤ Cn+m E(g)

2(n+m)
where p = n+m−2 . Given a smooth compactly supported function f : Rn → R+ ,
the idea is to apply this Sobolev inequality to
−(n+m−2)/2
g(x, y) = f (x) + |y|2 , (x, y) ∈ Rn × Rm ,
and to integrate it with respect to the Lebesgue measure in y ∈ Rm . To this end,

observe that
(n + m − 2)2 −(n+m)
|∇g|2 = f + |y|2 |∇f |2 + 4 |y|2 .
4
Now, for a > 0,

−(n+m)
a + |y|2 dy = cn,m a −n−(m/2)
Rm
and

−(n+m)
|y|2 a + |y|2 dy = dn,m a −n−(m/2)+1
Rm
where cn,m , dn,m > 0 only depend on n, m. In this picture, the measure
(1 + |x|2 )−(n+m) dx is simply the image of the spherical measure of Rn+m (viewed
in stereographic projection from Sn+m onto Rn+m , cf. Sect. 2.2.2, p. 83) on Rn by
orthogonal projection.
The integrated Sobolev inequality therefore takes the form
2/p
f −n−(m/2) dx
Rn

−n−(m/2)
≤ C1 f |∇f | dx + C2
2
f −n−(m/2)+1 dx
Rn Rn
for constants C1 , C2 > 0 only depending on n, m. The construction ensures that this
inequality is an equality for f = 1 + |x|2 since in this case the function g is an ex-
tremal function of the Sobolev inequality in Rn+m . Changing f into f −4/(2n+m−4)
yields the inequality, with further constants C3 , C4 > 0,
2/p
f q dx ≤ C3 |∇f |2 dx + C4 f s dx
Rn Rn Rn
for the values q, s specified in the statement (recall that ν = n + m2 ), and for which
there is now equality for f = (1 + |x|2 )−(ν−2)/2 . Changing f to cf and optimizing
in c > 0 yields the announced Gagliardo-Nirenberg inequality. The inequality is
optimal since it admits non-constant extremal functions.
The preceding argument however only works when the parameter m is an integer.
To cover the general case, it should be observed that we only used functions of
r = |y| in Rm and the Sobolev inequality of dimension n + m in Rn × R+ for
the standard carré du champ operator and the measure dμ(x, r) = r m−1 dxdr. As
presented in Sect. 6.9, this Sobolev inequality may be deduced from the spherical
case for m ≥ 1 via conformal transformations. The proof of Theorem 6.10.4 may
therefore be completed in this way.
Remark 6.10.5 There is a counterpart to Theorem 6.10.4 in the range q, s < 2

with s = 2(q − 1) for which the extremal functions are of the form f (x) =
[(σ − |x|2 )+ ]1/(2−q) , thus being compactly supported. The proof is very similar.
Interestingly enough, letting q, s → 2 (from both sides) gives rise to the Euclidean
logarithmic Sobolev inequality (Proposition 6.2.5) with the Gaussian kernels as ex-
tremals.
6.10.2 Entropic Form of Gagliardo-Nirenberg Inequalities
The family of Gagliardo-Nirenberg inequalities of Theorem 6.10.4, including

the classical Sobolev inequality, shares extremal functions of the same form
(σ 2 + b |x|2 )α (from which optimal constants are determined). The existence of ex-
plicit extremal functions for this sub-family of Gagliardo-Nirenberg inequalities ac-
tually leads to a new, equivalent, entropic formulation which we present next. This
formulation will in turn allow for the control of non-linear evolution equations (such
as porous medium or fast diffusion equations) as developed in the following section.
Moreover, it is suitable for an extension to more general settings under curvature-
dimension conditions.
Proposition 6.10.6 For ν ≥ n > 2, set
ν − 1 −1/ν
H (r) = − r 1−1/ν and (r) = H (r) = − r , r ∈ (0, ∞).
ν
Fix b > 0 and let vσ = h−ν σ,b , σ > 0, where hσ,b = hσ,b (x) = σ + b|x| , x ∈ R .
2 2 n
Let f : R →
R+ be smooth and compactly supported and choose σ > 0 such that
n
R n f dx = Rn vσ dx. Then, the Gagliardo-Nirenberg inequality of Theorem 6.10.4

for the parameter ν is equivalent to

H (f ) − H (vσ ) − (f − vσ )(vσ ) dx
Rn
(6.10.3)
b2 ν 2
≤ f ∇ (f ) − (vσ ) dx
4(ν − 1) Rn
(holding for every such function f ). When ν = n, the latter amounts to the optimal
Sobolev inequality on Rn .
Since the function H is (strictly) convex, the left-hand side of (6.10.3) is always
positive. It is zero only for f = vσ , and the inequality measures, in a certain entropic
sense, the distance from f to vσ .
Proof Fix b = 1 (the general case being similar), and to simplify the notation, set
h = hσ,1 . Using integration by parts on the right-hand side, (6.10.3) may be rewritten
as

1 1
f h − |∇h|2 dx + h1−ν dx
Rn 4 ν − 1 Rn

1 1 1 (ν−2)/2ν 2
≤ f 1−1/ν
ν − h dx + ∇ f dx.
ν − 1 Rn 2 (ν − 2) Rn
2
The function h = σ 2 + |x|2 satisfies h − 1

4 |∇h|2 = σ 2 and h = 2n so that the
preceding inequality is equivalent to

1
σ 2
f dx + h1−ν dx
Rn ν − 1 Rn

ν−n 1 (ν−2)/2ν 2
≤ f 1−1/ν dx + ∇ f dx.
ν − 1 Rn (ν − 2) Rn
2
Choosing f = vσ = h−ν yields equality (while not compactly supported, it is easily

checked that the previous computation is still valid for it). But for this particular
choice of f ,
(ν−2)/2ν 2
∇ f = (ν − 2)2 h−ν h − σ 2 .

Hence Rn h−ν dx and Rn h1−ν dx are related as

2σ 2 (ν − 1)
h 1−ν
dx = h−ν dx.
Rn 2ν − n − 2 Rn

On the other hand, Rn f dx = Rn h−ν dx = Cn,ν σ n−2ν where Cn,ν > 0 only de-

pends on
n and ν. Replacing the values of Rn h1−ν dx and σ by their values in
terms of Rn f dx, we are left with
(2ν−n−2)/(2ν−n)

Cn,ν f dx
Rn

ν−n 1 (ν−2)/2ν 2
≤ f 1−1/ν dx + ∇ f dx,
ν −1 Rn (ν − 2)2 Rn
where Cn,ν > 0 is a another constant. Finally, changing f into f 2ν/(ν−2) yields the
non-homogeneous Gagliardo-Nirenberg inequality of Proposition 6.10.3 with β = 2s

and q and s as in Theorem 6.10.4. The resulting inequality is sharp since by con-
struction it admits non-constant extremal functions. For the converse implication,
it suffices to translate the Gagliardo-Nirenberg inequality of Theorem 6.10.4 in its
non-homogeneous form and to follow the path back. The proof of Proposition 6.10.6
is therefore complete.
6.11 Fast Diffusion Equations and Sobolev Inequalities
This section will be a bit formal (and at the same time somewhat technical). While
we observed that linear heat equations do not allow us to reach sharp Sobolev in-
equalities, the aim here is to show that possibly non-linear evolution equations may
be used to establish optimal Sobolev inequalities under curvature-dimension condi-
tions CD(0, n) for example.
The method may be applied to more general situations and equations, but we only
concentrate here on the porous medium and fast diffusion equations, with a partic-
ular emphasis on the latter which is well-suited to both curvature-dimension con-
ditions and Sobolev-type inequalities. In addition, this fast diffusion equation plays
the same role with respect to Sobolev and curvature-dimension CD(ρ, n) inequal-
ities as the usual heat equation with respect to logarithmic Sobolev and curvature
CD(ρ, ∞) inequalities. Deeper relationships are actually underlying the picture.
The Ornstein-Uhlenbeck semigroup is a model case to test logarithmic Sobolev in-
equalities and curvature conditions. Indeed, under CD(ρ, ∞), the heat flow mono-
tonicity method produces in this case logarithmic Sobolev inequalities, both for the
heat semigroup (Pt )t≥0 and for the invariant measure μ (when ρ > 0) which are
optimal in the example of the Ornstein-Uhlenbeck model on Rn . As already ob-
served in Sect. 5.5, p. 257, this is fully coherent since for this model, on the one
side the heat kernels (that is the solutions of the heat equation starting from Dirac
masses) are Gaussian measures, for which the logarithmic Sobolev inequality is
optimal, and on the other, the logarithmic Sobolev inequality for any Gaussian mea-
sure after dilations and translations yields the logarithmic Sobolev inequality for
the Ornstein-Uhlenbeck semigroup, which in turn gives in the limit the curvature
condition CD(1, ∞). The same picture holds with the Euclidean heat semigroup in
place of the Ornstein-Uhlenbeck semigroup with respect to the curvature condition
CD(0, ∞). The question now is whether this full set of equivalences and mod-
els may be extended to Sobolev and curvature-dimension CD(ρ, n) inequalities
(for some finite dimension n). The models for these Sobolev inequalities are now
spheres, which may be seen via stereographic projections on the Euclidean space,
and for which the references measures are Cauchy measures (cf. Sect. 2.2, p. 81).
Although the game with dilations and translations may no longer be played in this
context (and should be replaced by conformal transformations on the sphere), the
heat equation may still be replaced by the fast diffusion equation described below,
which, although non-linear, produces in Rn the required Cauchy measures when the
initial data are Dirac masses. While this method will not be entirely satisfactory, it
certainly shows that non-linear fast diffusion equations could play a similar role for
Sobolev inequalities as linear heat equations for logarithmic Sobolev inequalities
with respect to curvature-dimension CD(ρ, n) inequalities.
The main purpose of this section will be to establish the entropic version of the
Gagliardo-Nirenberg inequality of Proposition 6.10.6 (which requires some addi-
tional stabilizing function v) under a curvature-dimension condition CD(0, n). As
discussed in the previous section, this Gagliardo-Nirenberg inequality is equivalent
to the sharp Sobolev inequality in Rn . Recall that stating a Sobolev inequality for
6.11 Fast Diffusion Equations and Sobolev Inequalities 331
an infinite measure requires a given normalization. Here, the choice of the func-
tion v in the general context will play the role of this normalization. In some sense,
the fast diffusion equation described below is a model case illustrating the inter-
play between entropy methods for evolution equations, functional inequalities and
curvature-dimension conditions. The case of CD(ρ, n) with ρ > 0 is briefly ad-
dressed next.
6.11.1 Porous Medium and Fast Diffusion Equations
The porous medium and fast diffusion equations are evolution equations of the form,
say on Rn ,

∂t u = um
for u = u(t, x) = ut (x), t ≥ 0, x ∈ Rn . Porous medium corresponds to m > 1 and
fast diffusion to m < 1. It is not obvious that there exist solutions to such equa-
tions at any time, in particular when m < 1. Actually, in Rn and for small m, the
solution starting from an initial condition u0 > 0 vanishes in finite time. We do
not discuss here existence (or uniqueness) issues for these equations. Our aim is
rather to explain how they may be used to reach Sobolev or Gagliardo-Nirenberg
inequalities under curvature-dimension CD(ρ, n) conditions, exactly as the heat
equation is used to reach Poincaré or logarithmic Sobolev inequalities under curva-
ture conditions CD(ρ, ∞). Namely, via Proposition 6.10.6, functional inequalities
of Sobolev-type are related to evolution equations through the control of entropic
quantities. As was extensively developed in the preceding chapters, on the one hand,
functional inequalities may be used to prove exponential decay of entropy related
to the evolution, where the functional inequality is translated into a differential in-
equality linking the entropy and its time derivative along the evolution equation. On
the other hand, functional inequalities may be reached by deriving entropy twice,
and using curvature-dimension conditions to directly reach this exponential decay.
In what follows, we present a general framework to control entropy of general
evolution equations with stabilizing term v. The general form of the second differen-
tial of the entropy is quite heavy, and the comparison with the curvature-dimension
CD(ρ, n) condition leads to the rather complicated form of the general Proposi-
tion 6.11.6 below. However, a much simplified and surprising form occurs for the
fast diffusion equation with parameter m = 1 − n1 , leading to the optimal form of the
Gagliardo-Nirenberg inequality under the CD(0, n) hypothesis. A slightly modified
version also leads directly to the optimal Sobolev inequality under the CD(ρ, n)
condition with ρ > 0, where no stabilizing term v is required.
We work below with a state space E that will either be Rn or a smooth manifold,
although most of what follows may be extended to more general settings at the
expense of several notational complications. Moreover, the various integration by
parts performed below, typically for a solution at time t > 0 of the associated fast
diffusion equation, should be carefully justified to turn this scheme into an actual
proof of functional inequalities in the context of general Full Markov Triples. Once
again, we will keep the exposition at a formal level in order to better emphasize the
main principle of proof.
Thus given a Markov Triple (E, μ, ), where E = Rn or a smooth (weighted)
Riemannian manifold, with associated diffusion operator L, and a function at
least C 2 on R+ , a first set of conclusions may be developed for the extended porous
medium and fast diffusion equations in the form of the non-linear equation
∂t u = L(u), u0 = f, (6.11.1)
where u = u(t, x) = ut (x), t ≥ 0, x ∈ E, with initial condition f .

In Rn or a Riemannian manifold, the generator L can be written as −∇ ∗ ∇ where
∗
∇ is the adjoint of ∇ on L2 (μ) (see (1.11.11), p. 48). Choosing a function such
that r (r) = (r), the extended equation (6.11.1) may then be rewritten as

∂t u = −∇ ∗ u∇(u) .
In this form, the heat equation corresponds to (r) = log r, r ∈ (0, ∞). A priori, the
behavior of u as t → ∞ is unknown, but if the interchange between differentiation
and integration is justified, we should at least have

∂t u dμ = ∂t u dμ = 0.
E E
On a compact Riemannian

manifold, it may thus be expected that ut converges
at infinity towards E udμ. On the other hand, in Rn , the behavior of u will be
forced to a given asymptotic behavior by a suitable choice of function v (to be made
explicit below) by modifying the equation as

∂t u = −∇ ∗ u ∇ (u) − (v) . (6.11.2)
This function v may be considered as a stabilizing term and plays an important role
in the associated functional inequality. For example, for the standard heat equation
on Rn , the choice for (r) = log r would be v = e−|x| /2 and the resulting equation
2
is the Fokker-Planck equation (cf. Sect. 1.5, p. 23),
∂t u = u + x · ∇u + nu = u + ∇ · (ux),
describing the evolution equation of the density of the Ornstein-Uhlenbeck semi-

group with respect to the Lebesgue measure, which produces a different form than
the usual evolution with respect to the reversible (here Gaussian) measure. Setting
u = û e−|x| /2 , it boils down to ∂t û = û − x · ∇ û, that is, the heat equation associ-
2
ated with the Ornstein-Uhlenbeck operator.

After this transformation, it may be expected that, as t → ∞, ut will converge
towards v̂ where v̂ is such that (v) = (v̂) + c for a certain constant
c, since

these
functions are obviously stable solutions of (6.11.2). Now, since E v̂dμ = E udμ,
these two conditions will

in general
determine v̂ uniquely. Replacing in (6.11.2) v
by v̂ so that in addition E v̂dμ = E f dμ where f = u0 is the initial condition,
ut is expected to
converge towards v̂, a stationary solution of the equation with
conserved mass E f dμ.
In the following, u = u(t, x) = ut (x) denotes the solution of (6.11.1) and v the
stabilizer as constructed above. Moreover, these functions and the functions con-
structed with them are implicitly assumed to be sufficiently smooth in order for the
various expressions and arguments to make sense, and in particular for the validity
of the various integrations by parts.
6.11.2 Entropy Decay
In order to quantify the previous convergence along the same lines as the corre-
sponding development for logarithmic Sobolev inequalities in Chap. 5, an analogue
of the entropy functional needs to be considered. The quantity that will replace en-
tropy in this case is given by

F (u) = H (u) − u (v) dμ
E
where H on (0, ∞) is such that H = . Note that for the heat equation, H (r)
is indeed r log r − r, close to the standard entropy. When H is convex (which
will be the case below for the fast diffusion equation with m = 1 − n1 where
H (r) = −nr 1−1/n ), then F (u) − F (v) ≥ 0, since

F (u) − F (v) = H (u) − H (v) − (u − v)H (v) dμ.
E
Moreover, F (u) − F (v) vanishes only when u = v provided H is strictly convex.

The quantity F (u) − F (v) may thus be considered as a distance from u to v.
The first operation to perform in this context is to compute the time derivative of
the functional F as an analogue of de Bruijn’s identity (Proposition 5.2.2, p. 245) in
case of the standard entropy with the suitable (modified) Fisher information denoted
by Iμ,F .
Proposition 6.11.1 In the preceding notation, setting ξ = (u) − (v),

d
F (u) = − Iμ,F (u) = − u (ξ )dμ.
dt E
Proof We make use of the following simple integration by parts formula, which will
be used repeatedly below. Namely, for each smooth function h on E (for example
in the algebra A0 ),

∂t u h dμ = − u (ξ, h)dμ. (6.11.3)
E E
To prove this formula, use the identity ∂t u = ∇ ∗ u · ∇ξ and, assuming everything is

sufficiently smooth, integrate by parts to get

(∂t u) h dμ = − u ∇ξ · ∇h dμ = − u (ξ, h)dμ
E E E
which is the claim. On the basis of (6.11.3), it then suffices to observe that

d
F (u) = ∂t u (u) − (v) dμ = (∂t u) ξ dμ.
dt E E
The proposition is established.
By comparison with the entropic version of the Gagliardo-Nirenberg inequalities

(Proposition 6.10.6), the following statement describes exponential convergence in
the entropy F along the fast diffusion equation towards the steady states whose form
is closely related to the extremals of the Gagliardo-Nirenberg inequalities.
Proposition 6.11.2 In Rn , if H (r) = −n r 1−1/n , r ∈ (0, ∞), that is, for the equa-
tion

∂t u = −(n − 1)∇ ∗ u ∇ u−1/n − vσ−1/n
2 −n

where vσ = (σ + b |x| ) , provided σ > 0 is chosen so that Rn vσ dx =

2
Rn u0 dx, then for every t ≥ 0,

0 ≤ F (ut ) − F (vσ ) ≤ e−ct F (u0 ) − F (vσ )
with c = 4(n−1)
b2
. In other words, ut converges exponentially towards vσ in the en-
tropic sense as t → ∞.
For the proof, it suffices to observe that Proposition 6.10.6 actually yields that
d 4(n − 1)
F (ut ) − F (vσ ) ≤ − F (ut ) − F (vσ ) .
dt b2
As for logarithmic Sobolev inequalities (cf. Theorem 5.2.1, p. 244), the entropy
decay of Proposition 6.11.2 implies in return the Gagliardo-Nirenberg inequalities
of Proposition 6.10.6 with ν = n in the limit t → 0.
We next turn to the second step of the analysis, namely computing, as for classical
entropy, the second order derivative of the functional F (that is, the derivative of the
generalized Fisher information Iμ,F ) along the fast diffusion evolution with the help
of the 2 operator. The curvature-dimension condition will then enter into play when
establishing functional inequalities along the lines developed for Poincaré and loga-
rithmic Sobolev inequalities in Sect. 4.8, p. 211, and Sect. 5.7, p. 268, respectively.
This second order differentation will be performed for some general function
along the evolution (6.11.2) assuming however that the state space E is a smooth
manifold M and that the operator L is elliptic (allowing for the use of the differential
calculus described in Sect. C.5, p. 511). Questions of existence, regularity and do-
mains for the solutions of the non-linear equations (6.11.2) are not discussed here. In
practice, for any such evolution equation related to and v, with initial data u0 = f ,
it may be quite hard to justify the formal computations described below. In particu-
lar, for fast diffusion equations, which is the central point of interest below, (r) is
singular at r = 0 which causes serious difficulties. Since the point in what follows
is to use those non-linear evolution equations to obtain functional inequalities from
curvature-dimension conditions CD(ρ, n) together with geometric properties of the
function v involved in the equation, it is in general necessary to first approximate
the singular map by some smooth approximation ε . For this approximation,
all the computations and integrations by parts may be justified by some previous a
priori analysis so to yield an approximated functional inequality which converges
in the limit ε → 0 to the expected result. We however do not enter into these quite
technical and tedious considerations here, assuming as mentioned above the regular-
ity of u and v and of related expressions justifying the various integration by parts
formulas. The aim of this study is rather to show that the modified fast diffusion
equation leads, through these formal computations, precisely to the entropic form
of the Sobolev or Gagliardo-Nirenberg inequalities described in Proposition 6.10.6.
The next lemma describes the announced second derivative operation on the
functional F .

Lemma 6.11.3 Let R be the function R(r) = r (r)
(r) , r ∈ (0, ∞). Set S = −(u),
ξ = (u) − (v), which are functions (depending on t) from E to R. Then, with
the generalized Fisher information Iμ,F of Proposition 6.11.1,

d
Iμ,F (u) = − u K dμ
dt E
where

K = 2u (u) 2 (ξ ) + ξ, (ξ ) − R(u) + 2 S, (ξ )
R(u) + 1 + uR (u)
− 2R(u) ξ, (ξ, S) + 2 (ξ, S)2 .
u (u)
In the examples of interest, ≤ 0 so that S = −(u) ≥ 0.
Proof Again, the exchanges between differentiation and integration will not al-
ways be fully justified (although care may be developed to this end). Since
∂t ξ = (u)∂t u,

d
Iμ,F (u) = ∂t u (ξ ) + 2u (u)∂t u, ξ dμ.
dt E
By integration by parts with respect to the carré du champ operator , for suitable
functions f, g (say in the algebra A0 ),

u (f, g)dμ = − f u Lg + (u, g) dμ.
E E
Therefore, the second term on the right-hand side of the preceding identity may be
rewritten as

−2 ∂t u (u) u Lξ + (u, ξ ) dμ.
E
From the evolution of u, we have with (6.11.3) that

d
Iμ,F (u) = − u R1 dμ
dt E
where

R1 = ξ, (ξ ) + 2 ξ, (S, ξ ) − 2 ξ, u (u) Lξ .
By the change of variables formula and the definition of the 2 operator,

ξ, u (u) Lξ = u (u) (ξ, Lξ ) + u (u) + (u) Lξ (ξ, u)

= u (u) (ξ, Lξ ) − R(u) + 1 Lξ (ξ, S)

1
= u (u) −2 (ξ ) + L(ξ ) − R(u) + 1 Lξ (ξ, S).
2
Furthermore,

u2 (u)L (ξ )dμ = − u2 (u), (ξ ) dμ
E E

= u R(u) + 2 S, (ξ ) dμ.
E
Moreover, via integration by parts,

− u R(u) + 1 Lξ (ξ, S)
E

= ξ, u R(u) + 1 (ξ, S) dμ
E

R + 1 + uR (u)
= u R(u) + 1 ξ, (ξ, S) dμ − (ξ, S)2 dμ.
E E (u)
Putting the various identities together yields the conclusion. Lemma 6.11.3 is estab-
lished.
6.11.3 Fast Diffusion and Curvature-Dimension Condition

CD(ρ, n)
On the basis of Lemma 6.11.3, the aim is now to compare the second derivative
of the entropy functional F to quantities arising from the CD(ρ, n) condition. The
situation is much simpler here in the fast diffusion case since then the various quan-
tities involved in the expression K are easier to handle. Indeed, the fast diffusion
rm
equation corresponds to (r) = r m , in which case (r) = m−1 m
r m−1 , H (r) = m−1
and R = m − 2. Therefore K = Km in Lemma 6.11.3 is equal to

Km = 2(1 − m) S 2 (ξ ) + ξ, (ξ ) − m S, (ξ )
(ξ, S)2
+ 2(2 − m) ξ, (ξ, S) − 2 .
S
This quantity Km may actually be described in more geometrical and tractable
terms. Set ζ = (v) (and thus −S = ξ + ζ ), and X = ∇ξ , Y = ∇ζ , M = ∇∇ξ ,
P = ∇∇ζ . Using the notation M · N for the scalar product in the space of symmetric
tensors, that is in coordinates (and with Einstein’s summation notation),
M · N = M ij N k gik gj ,
and the notation X # Y for the symmetric tensor product of two vectors (X # Y )ij =
2 (X Y + X Y ), we get that
1 i j j i
2 (ξ ) = |M|2 + Ric(L) · X # X,

S, (ξ ) = −2M · (X + Y ) # X,

ξ, (ξ ) = 2M · X # X,

ξ, (ξ, S) = −M · X # (X + Y ) − (M + P ) · X # X,
(S, ξ ) = −X · (X + Y ).
Then
1
Km = (1 − m) S 2 (ξ ) + (m − 1) M · X # (3X + 2Y )
2
(X · (X + Y ))2
+ (m − 2) P · (X # X) − .
S
We are now ready to use the curvature-dimension condition CD(ρ, n). We start
with an elementary computation which is a further application of the integration by
parts formula. Details are omitted.
Lemma 6.11.4 For every smooth function w,

1
w (Lξ ) dμ =
2
w 2 (ξ ) + w, (ξ ) + ξ, (ξ, w) dμ.
E E 2
As a consequence,
Proposition 6.11.5 Under the curvature-dimension condition CD(ρ, n), for every
smooth function w ≥ 0,

ρn 1
w 2 (ξ ) dμ ≥ w (ξ )dμ + w, (ξ ) dμ
E n−1 E 2(n − 1) E

1
+ ξ, (ξ, w) dμ.
n−1 E
In the case of interest, and towards the comparison with Km , we have

w = (1 − m)uS = mum and S = 1−mm
um−1 . In terms of X, Y, S, M, P ,
∇w = mu(X + Y ), ∇∇w = mu(P + M) + u2−m (X + Y ) # (X + Y ).
Therefore, the inequality of Proposition 6.11.5 reads

(1 − m)u S 2 (ξ )dμ
E

ρn m
≥ (1 − m)u S (ξ ) + uM · X # (3X + 2Y ) dμ
n−1 E n−1 E

m 1 2
+ uP · [X # X] dμ + u2−m X · (X + Y ) dμ.
n−1 E n−1 E
The next proposition summarizes the final relevant inequality on the derivative
of the generalized Fisher information.
Proposition 6.11.6 Under the curvature-dimension condition CD(ρ, n), and for
(r) = m−1
m
r m−1 ,

1 d ρn
− Iμ,F (u) ≥ mum (ξ )dμ
2 dt n−1 E

m
+ +m−2 u P · (X # X)dμ
n−1 E

m
+ +m−1 u M · X # (3X + 2Y ) dμ
n−1 E

1 m−1 2
+ + u2−m X · (X + Y ) dμ,
n−1 m E
where we recall that ξ = (u) − (v), ζ = (v), X = ∇ξ , Y = ∇ζ , M = ∇∇ξ

and P = ∇∇ζ .
In the particular case m = 1 − n1 , and if −P = −∇∇(v) ≥ a ≥ 0, then
−P · (X # X) ≥ a (ξ ), and everything simplifies miraculously to

1 d
− Iμ,F (u) ≥ ρ um (ξ )dμ + a I (u).
2 dt E
On the basis of the preceding proposition, we may now state the resulting func-
tional inequality under a curvature-dimension condition. For simplicity, we only
deal with the CD(0, n) hypothesis.
Theorem 6.11.7 In the preceding setting, under the CD(0, n) condition, for a

∇∇h ≥
a > 0, then, for H (r) = −n r
positive smooth function h such that 1−1/n ,

H (v) = −h, f > 0, and provided E vdμ = E f dμ, we have

1 2
H (f ) − H (v) − (f − v)H (v) dμ ≤ f ∇ H (f ) − H (v) dμ.
E 2a E
Proof Note that by the sign choice, the function H is convex, so that the left-hand
side in the inequality of the theorem is positive. We stay at a somewhat informal
level since justification of each step is rather tedious. Nevertheless, the principle is
the following. Under the CD(0, n) condition, Proposition 6.11.6 shows that solving
the equation

∂t u = −∇ ∗ u ∇ (u) − (v) , u0 = f,
leads to the differential inequality
d
Iμ,F (ut ) ≤ −2 a Iμ,F (ut ).
dt
It follows from the latter that
Iμ,F (ut ) ≤ e−2at Iμ,F (f )
for every t ≥ 0. Since Iμ,F (ut ) = − dt

d
F (ut ), it follows further that
∞ 1
F (u0 ) − F (u∞ ) ≤ Iμ,F (f ) e−2a(n−1)t dt = Iμ,F (f ).
0 2a
We are left
to show that indeed ut → v as t → ∞. A necessary condition is that
E vdμ = E f dμ. A more precise investigation of the entropy decay (and further
technical bounds) indeed ensures that this is the case, so that the preceding amounts
to the entropic inequality of the theorem.
Remark 6.11.8 It is somewhat strange that for the parameters of Theorem 6.10.4,
the Gagliardo-Nirenberg inequalities may be deduced either from the Sobolev in-
equality in Rn+m with ν = n + m2 relying on the CD(0, n + m) condition, or via
the fast diffusion equation from the CD(0, ν) condition of Rn (with the impres-
sion therefore of a loss of information in the CD(0, n) condition of Rn ). But, as
already observed throughout this chapter, this loss of information is actually recap-
tured by the action of the dilation group (which enters in a somewhat subtle way
in the equivalence between the entropic inequalities of Proposition 6.10.6 and the
Gagliardo-Nirenberg inequalities). Actually, as already mentioned, most of the func-

tional inequalities on Rn , after optimization under the action of dilations, yield in-
equalities equivalent to the Sobolev inequality. One illustrative example is the proof
of the sharp Euclidean logarithmic Sobolev inequality on Rn from the Gaussian
logarithmic Sobolev inequality, itself a consequence of the CD(0, ∞) condition in
Euclidean space.
Remark 6.11.9 Starting from a curvature-dimension condition CD(ρ, n) with

ρ > 0, there is no need to make use of the stabilizing term v, that is the function
h satisfying ∇∇h ≥ a > 0, to produce a Sobolev or a Gagliardo-Nirenberg inequal-
ity. Indeed, using the same evolution equation with v constant, and along the same
lines, one may get directly the Sobolev inequality with optimal constants (i.e. with
the constants one would get on spheres under the same hypothesis). Namely, for
m = 1 − n1 ,

1 d
− Iμ,F (u) ≥ ρ um (ξ )dμ.
2 dt E
Observe then that

d 2(n − 2)
u1−2/n dμ = um (ξ )dμ
dt E n(n − 1) E
which leads to

d ρ n(n − 1)
Iμ,F (u) + u1−2/n
dμ ≤ 0.
dt n−2 E
Therefore,

1−2/n 1−2/n n−2
ut dμ ≤ u0 dμ + Iμ,F (u0 ).
E E ρ n(n − 1)

If we accept finally that limt→∞ ut = E u0 dμ (which has to be proved using a more

precise analysis than the one provided here), in the limit (after a suitable rewriting
of Iμ,F (u0 )),
1−2/n
1−2/n 4 n−1 (n−2)/(2n)
u0 dμ ≤ u0 dμ + u0 dμ.
E E n(n − 2) ρ E
Changing u0 into f 2n/(n−2) leads to the tight Sobolev inequality with the sharp
constant of Theorem 6.8.3, thus providing an alternative approach.

Sobolev inequalities, starting with the founding paper [386] by S. Sobolev, are a
central theme in analysis, covering an incredibly large spectrum of both theoretical
and more applied topics. It is not within the scope of this short notice to present
exhaustive comments and historical background on Sobolev inequalities. One major
reference is the comprehensive monograph of V. Maz’ya [303] (a recent expansion
of the famous [302]) to which we refer for a complete account of the subject. The
contributions [1, 2, 96, 288, 413–415] are further relevant references. Related to the
geometric aspects emphasized in this work, we mention in addition the monographs
by T. Aubin [20, 21], E. Hebey [233, 234], and those of by L. Saloff-Coste [376],
A. Grigor’yan [217] and P. Li [284] as well as the references therein.
The optimal constants in the Sobolev inequalities on Rn are due independently
to T. Aubin [18, 19] and G. Talenti [406] (see also [364] for dimension 3). The
case of the sphere is examined in [17] (see also [20]), while the hyperbolic case
is proved in [235] by similar means as described here, relying on [382]. The Nash
inequality in Rn was introduced by J. Nash [323], the optimal constant having been
computed by E. Carlen and M. Loss [110]. The history of the Euclidean logarithmic
Sobolev inequality (Proposition 6.2.5) is closely related to that of the Gaussian log-
arithmic Sobolev inequality (see Chap. 5), which in particular goes back to [387].
Proposition 6.2.5 linking the Euclidean and Gaussian forms has been explicitly or
implicitly used by many authors [54, 106, 437, 438]. Forms of the Euclidean loga-
rithmic Sobolev inequality were used by G. Perelman in his solution of the Poincaré
conjecture.
The equivalences of Proposition 6.2.3 in Sect. 6.2 between the various forms
of Sobolev inequalities classically appeared as consequences of ultracontractive
bounds equivalent to the corresponding functional inequalities (Sect. 6.3). The di-
rect approach presented here via the slicing decomposition was introduced in [35]
and [149]. See e.g. [376, 377] for more and for bibliographical accounts. The Eu-
clidean logarithmic Sobolev inequality regarded as a limit of the sharp Sobolev in-
equality (Remark 6.2.6) was pointed out in [58] (see also [57]).
The main ultracontractive Theorem 6.3.1 of Sect. 6.3 is the end result of a se-
ries of steps and results by numerous authors including, among others, [26, 108,
138, 144, 181, 319, 323, 421]. The proof presented here is due to Nash, using the
corresponding Nash inequality, following [108]. The other authors dealt with either
Sobolev or logarithmic Sobolev inequalities, illustrated here in Proposition 6.2.3,
or Gagliardo-Nirenberg inequalities. See [26, 144, 217, 376, 377, 422] for further
details and historical background.
The Rellich-Kondrachov Theorem 6.4.3 may be found in most classical refer-
ences such as [96, 179, 203].
Proposition 6.6.1 in Sect. 6.6 is due to L. Saloff-Coste [375] (the proof here
being taken from [275]). The sharp version of the diameter bound under a Sobolev
inequality (namely that D ≤ π whenever the Sobolev inequality with the constants
of the model space Sn holds) was obtained in [39], leading to the classical Bonnet-
Myers Theorem in this context. Proposition 6.6.2 on the volume growth is part of
the folklore and may be found, for example, in [114] (see also [217, 377]).
The celebrated Li-Yau parabolic inequality (Corollary 6.7.5) was established
in [285] by means of the maximum principle. The heat flow monotonicity proof
presented here via dimensional logarithmic Sobolev inequalities for heat kernels is
taken from [40]. More on Harnack inequalities for elliptic operators on Riemannian
manifolds and precise heat kernel bounds (of the form (6.7.11) and (6.7.12)) may be
found in [144, 217, 284, 378]. The small time asymptotics (6.7.13) go back to [420]
(see [332] for the final word). Local hypercontractivity (Theorem 6.7.7) first ap-
peared in [31]. Versions of the Li-Yau inequality in sub-Riemannian geometries by
means of heat flow monotonicity have been obtained in [51].
The sharp Sobolev inequality on the sphere (6.1.2) of T. Aubin [17, 20] was ex-
tended to Riemannian manifolds with a uniform strictly positive lower bound on
the Ricci curvature in [253] via the Lévy-Gromov isoperimetric comparison The-
orem (see the Notes and References of Chap. 8). The non-linear proof of Theo-
rem 6.8.3 developed in Sect. 6.8 is adapted from the corresponding analysis by
O. Rothaus [371] for logarithmic Sobolev inequalities. The method actually orig-
inates in [202] in a partial differential equation context (see later [66]). The fi-
nal step of the argument is reminiscent of the Nash-Moser iteration principle. The
proof is presented in [26], and further developed in [185] where the subsequent
Remarks 6.8.4 and 6.8.5 are emphasized, the second one again following [371].
On the sphere, these remarks go back to W. Beckner [56] who used sharp Hardy-
Littlewood-Sobolev inequalities for this purpose. See also [66, 160, 161]. Moser-
Trudinger inequalities appeared in [320, 416] as the limiting case n = 2 (actually
any n in the suitable version in higher dimension). See also [35] in the context of the
slicing technique of Proposition 6.2.3. On the sphere S2 , the sharp inequality (6.8.8)
is sometimes referred to as Onofri’s inequality [337] (see also [56]).
Section 6.9 on conformal invariance of Sobolev inequalities describes some of
the main ideas involved in the Yamabé program developed in particular by T. Au-
bin [17, 20] and R. Schoen [380, 381] (see [21, 233, 234]). Its abstract formulation
in the context of this book is partly inspired by the paper [382] of R. Schoen and
S.-T. Yau. The formal equivalence between the classical Sobolev inequalities in Eu-
clidean, spherical and hyperbolic spaces may also be traced back to [382] and is em-
phasized in [235]. (Remark 6.9.5 on the (minimal) class of functions in the Sobolev
inequality in Euclidean space is part of the classical theory [288, 303, 449].) The
references [165, 234] present the (A, B) program which is devoted to the respective
sharpness of the constants C and A, in the notation used here, for Sobolev inequal-
ities in manifolds.
The Gagliardo-Nirenberg inequalities recalled in Sect. 6.10 were first put forward
in [192, 331]. See for instance [96, 303] for general recent references and histori-
cal background. Optimality and extremal functions of the sub-family of Gagliardo-
Nirenberg inequalities of Theorem 6.10.4 have been obtained by M. Del Pino and
J. Dolbeault [147] by an analytic study of the corresponding minimization prob-
lem. Theorem 6.10.4 has been generalized in [330] to a larger class of parameters
with arguments based on the mass transportation method of [137]. The reformula-
tion in [147] as an entropy decay along non-linear equations in Proposition 6.10.6
has been an important step in the understanding of the connections between porous
medium and fast diffusion equations, functional inequalities and convergence to
equilibrium. See also [111], and [424, 426] for links with mass transportation and
the references therein.
A comprehensive account of porous medium and fast diffusion equations is the
recent monograph [423] to which we refer for a complete background. On the basis
of the Dolbeault-Del Pino developments, Sect. 6.11, and Theorem 6.11.7, are es-
sentially due to J. Demange [155], who solved in particular all the regularity issues
necessary for the proof (see also [28, 156] and [161]). The last Remark 6.11.9 is
also due to him (unpublished). Fast diffusion has also been used to obtain Hardy-
Littlewood-Sobolev inequalities in [107], relying on the heat flow monotonicity ap-
proach to geometric Brascamp-Lieb inequalities (different from the approach de-
scribed in Sect. 4.9, p. 215) developed in [59, 109].
Part III
Related Functional, Isoperimetric
and Transportation Inequalities
Chapter 7
Generalized Functional Inequalities
Part II, devoted to Poincaré, logarithmic Sobolev and Sobolev inequalities, describes
how each of these families capture different features of the associated semigroup or
the invariant measure, in terms of convergence to equilibrium, estimates on the heat
kernels or tail behaviors of the invariant measure. There are many ways to describe
intermediate families of functional inequalities which are suited to a wide variety of
regimes as well as to more precise, or different, features. This section investigates
such families, restricting to three main examples, entropy-energy, generalized Nash
and weak Poincaré inequalities. Many other families have been developed in differ-
ent directions, each of them having its own interest. Some will be mentioned at the
end of the section.
In this chapter, we deal as usual with a Full Markov Triple (E, μ, ) as pre-
sented in Sect. 3.4, p. 168, with Dirichlet form E, infinitesimal generator L, Markov
semigroup P = (Pt )t≥0 and underlying function algebras A0 and A. Actually, the
natural framework for the investigation here is that of the Standard Markov Triple,
and besides the section on off-diagonal heat kernel bounds for which the extended
algebra A is required, all the results may be stated in this framework. We neverthe-
less simply use the terminology “Markov Triple” to cover the various instances.
The first section describes the family of functional entropy-energy inequalities
governed by a growth function , the example of the logarithmic function giv-
ing rise to the logarithmic entropy-energy inequality of the previous chapter as an
equivalent form of the standard Sobolev inequality. The family of entropy-energy in-
equalities is in particular well-suited to heat kernel bounds by means of the method
developed for hypercontractivity under logarithmic Sobolev inequalities in Chap. 5.
Off-diagonal heat kernel estimates may be achieved in the same way (Sect. 7.2).
Several examples of both entropy-energy inequalities and their associated heat ker-
nel bounds are presented in Sect. 7.3. The next section investigates generalized Nash
inequalities on the basis of the example of the classical Nash inequality in Euclidean
space, as well as weighted Nash inequalities which form a further family of interest.
Again, heat kernel bounds and tail inequalities of various types may be obtained.
Weak Poincaré inequalities are studied in Sect. 7.5 as the main minimal tool to
tighten families of standard functional inequalities. Weak Poincaré inequalities may

DOI 10.1007/978-3-319-00227-9_7,
348 7 Generalized Functional Inequalities
be further studied in their own right from the viewpoint of heat kernel bounds and
tail estimates. Section 7.6 briefly describes related families of functional inequali-
ties of interest and their relationships with the previous ones. Various applications
of the tools investigated in this chapter are illustrated, for comparison, on the model
family of (probability) measures cα e−|x| dx, α > 0, on Rn with respect to the stan-
α
dard carré du champ operator (f ) = |∇f |2 . Actually, to avoid technical regularity
issues of the potential W (x) = |x|α (at the origin), it will be more convenient to deal
with the family
α/2
dμα (x) = cα exp − 1 + |x|2 dx (7.0.1)
(with cα > 0 the normalization constant) which behave similarly at infinity. The
main results for this family in the one-dimensional case are summarized in Sect. 7.7
at the end of the chapter. The multi-dimensional picture is essentially the same with
constants depending on the dimension.
Finally, it is worth emphasizing that most inequalities investigated in this chapter
compare two functionals via a growth function which will always be C 1 increasing
and concave from (0, ∞) to R, R+ or (0, ∞). Although the C 1 hypothesis is not
strictly necessary (continuous should be enough), it will be convenient for a number
of technical steps. As already (briefly) mentioned in the proof of Proposition 4.10.4,
p. 224, a standard procedure in this regard, at least

for positive growth functions, is
to replace (concave increasing) by (r) = 2 r (s)ds which is still increasing
r 0
and concave and satisfies 12 (r) ≤ (r) ≤ (r)). In several parts of this chapter,
sharp constants are indeed not an issue and focus is instead placed on growth orders.
7.1 Inequalities Between Entropy and Energy

This section deals with generalized functional inequalities comparing the entropy
and energy of a given function in the Dirichlet space. These generalized functional
inequalities are then shown to efficiently reach uniform heat kernel bounds.
7.1.1 Entropy-Energy Inequalities
This family compares the entropy and the energy of a function (in the Dirichlet
domain D(E)) via a C 1 increasing concave growth function : (0, ∞) → R. A first
example of such an inequality is the logarithmic entropy-energy inequality (6.2.2)
of Proposition 6.2.3, p. 281, with growth function the logarithmic function. The
function (r), r > 0, may have a finite limit at r = 0, denoted (0), or converge to
−∞ as in the latter example. Recall that the entropy of a measure ν (not necessarily
finite) on a measurable space (E, F) has been defined at (5.1.1), p. 236, as

Entν (f ) = f log f dν − f dν log f dν
E E E

for all positive integrable functions f : E → R such that E f | log f |dν < ∞.
7.1 Inequalities Between Entropy and Energy 349
Definition 7.1.1 (Entropy-energy inequality) A Markov Triple (E, μ, ) is said

to satisfy an entropy-energy inequality EE()
with respect to a growth function
: (0, ∞) → R if for every f ∈ D(E) such that E f 2 dμ = 1 and E(f ) > 0,

Entμ f 2 ≤ E(f ) . (7.1.1)
As usual, it is enough to state and establish such an inequality for a family of

functions f which is dense in the domain D(E) of the Dirichlet form E (typically
the algebra A0 ).
As mentioned previously, the logarithmic entropy-energy inequality with (r) =
n
2 log(A + Cr), r ∈ (0, ∞), described in Proposition 6.2.3, p. 281, as an equivalent
form of a Sobolev inequality, is an instance of Definition 7.1.1. In particular, the
Euclidean logarithmic Sobolev inequality with respect to the Lebesgue measure on
Rn of Proposition 6.2.5, p. 284, holds for the optimal function (r) = n2 log( nπe
2r
),
r ∈ (0, ∞). Further examples of entropy-energy inequalities have occurred earlier
for specific choices of the growth function . For instance, for μ a probability mea-
sure, the (defective) logarithmic Sobolev inequality LS(C, D) of Definition 5.1.1,
p. 236, is an example of an EE() inequality with (r) = 2Cr + D, r ∈ (0, ∞).
Note also that if EE(1 ) and EE(2 ) hold for two concave functions 1 and 2 ,
then EE(1 ∧ 2 ) holds.
The entropy-energy inequalities from Definition 7.1.1 may be described in equiv-
alent linearized forms. Indeed, since in EE() is assumed to be C 1 , by concavity
(s) − (r) ≤ (r)(s − r) for all r, s > 0. The entropy-energy inequality EE()
of Definition 7.1.1 may then be linearized equivalently

as 2a family of defective log-
arithmic Sobolev inequalities, for all f ∈ D(E) with E f dμ = 1,

Entμ f 2 ≤ (r) E(f ) + (r), r ∈ (0, ∞), (7.1.2)
where (r) = (r) − r (r).

Whenever is strictly decreasing, and takes values in some interval (a, b) ⊂ R,
one may choose s = (r) and turn EE() into the family of inequalities

Entμ f 2 ≤ s E(f ) + β(s), s ∈ (a, b), (7.1.3)

f ∈ D(E), E f 2 dμ = 1, where β(s) = (−1 (s))−s −1 (s). Conversely, (7.1.3)

gives rise to the EE() inequality with

(r) = inf sr + β(s) , r ∈ (0, ∞).
s∈(a,b)
Observe that for smooth functions, β = −−1 , and so there is a complete equiv-
alence between the two formulations (7.1.2) and (7.1.3). Whenever is not bi-
jective, −1 has to be replaced by its generalized inverse, where the generalized
inverse of an increasing function T on an interval I ⊂ R is defined by

T −1 (s) = inf r ∈ I ; T (r) ≥ s (7.1.4)
for every s in the range of T .

Inequalities (7.1.2) or (7.1.3) will be considered as the linearized forms of the
entropy-energy inequality EE(φ).
When μ is a probability measure, the entropy-energy inequality EE() is thus
equivalent to a family of defective logarithmic Sobolev inequalities. According to
Proposition 5.1.3, p. 238, a tight logarithmic Sobolev inequality LS(C) implies a
Poincaré inequality P (C) while conversely, under a Poincaré inequality, a defective
logarithmic Sobolev inequality LS(C, D) may be tightened. This principle applied
to an entropy-energy inequality EE() through the equivalent descriptions (7.1.2)
or (7.1.3), shows that, under EE(), it is equivalent to ask for a Poincaré inequality
or for the function to be bounded from above by Cr, r > 0, for some constant
C > 0.
7.1.2 Uniform Heat Kernel Bounds
The main interest in the family of entropy-energy inequalities EE() is the flexi-
bility in the choice of the growth function . In particular, this family interpolates
between logarithmic Sobolev and Sobolev-type inequalities, and actually allows for
intermediate regimes, in particular of heat kernel bounds. The following first and
main result is an illustration of this principle. It is a direct extension of the hyper-
contractivity Theorem 5.2.3, p. 246, and is proved using the same Gross method.
If 1 ≤ p ≤ ∞, p ∗ denotes its conjugate exponent, so that p1 + p1∗ = 1. The norms
are understood as usual in Lp (μ), 1 ≤ p ≤ ∞. Recall the operator norm notation
· p,q from Lp (μ) into Lq (μ) from (6.3.1), p. 286.
Theorem 7.1.2 Let (E, μ, ) be a Markov Triple satisfying an entropy-energy

inequality EE() for some growth function . Recall (r) = (r) − r (r),
r ∈ (0, ∞), from (7.1.2). Then, for every 2 ≤ p ≤ q ≤ ∞ and every δ > 0,
Pt (δ) p,q ≤ em(δ)
where
⎧ qq ∗
⎪
⎪ dr
⎪t (δ) =
⎨ (δr) √ ,
pp ∗ 4 r(r − 4)
⎪ qq ∗
(7.1.5)
⎪
⎪ dr
⎩m(δ) = (δr) √ .
pp ∗ r r(r − 4)

qq ∗
pp∗
When 1 ≤ p ≤ q ≤ 2, the integrals in (7.1.5) have to be replaced by
pp ∗ qq ∗ while

pp∗
qq ∗
when 1 ≤ p ≤ 2 ≤ q ≤ ∞, they have to be replaced by 4 + 4 .
Remark
7.1.3 In Theorem 7.1.2, it is sometimes simpler to change r into
s = 1 − 4r and δ into 4δ so that (7.1.5) becomes
⎧ 2
⎪
⎪ 1 1− q δ ds
⎪
⎪ t (δ) = ,
⎨ 2 1− p
2 1−s 2 1 − s2
(7.1.6)
⎪
⎪ 1 1− q
2
⎪
⎪ =
δ
⎩ m(δ) ds.
2 1− p2 1 − s2
In this form, the formulas are valid whatever the values of the parameters
1 ≤ p ≤ q ≤ ∞, as long as the integrals converge.
We make a few comments before turning to the proof of Theorem 7.1.2. First,
the result is compatible with the semigroup property. Namely, since Pt ◦ Ps p,q ≤
Pt p,r Ps r,q when p ≤ r ≤ q, applying the conclusion separately on [0, t] and
on [0, s] yields the same bound as applying it directly on [0, t + s]. In particular, the
case p ≤ 2 ≤ q is a direct consequence of the separate results for the pairs (p, 2)
and (2, q). Moreover, thanks to the symmetry of the formulas under the change of
p and q into their conjugate exponents p ∗ and q ∗ , it is also compatible with the
symmetry property Pt p,q = Pt q ∗ ,p∗ . Therefore, everything reduces to the case
2 ≤ p ≤ q ≤ ∞.
The second set of observations actually illustrates the range of applications of
Theorem 7.1.2. Note first that whenever lim infr→∞ (r) > 0, as is the case for a
logarithmic Sobolev inequality and hypercontractivity (cf. Theorem 5.2.3, p. 246),
one may not reach q = ∞ in Theorem 7.1.2. There is actually a minimal value
of t for which hypercontractivity applies. On the other hand, the conclusion of
Theorem 7.1.2 may be applied to p = 2 and q = ∞, or p = 1 and q = ∞, as
soon as r −1 (r) is integrable at infinity (which implies that r −2 (r) is also in-
tegrable at infinity). It applies in particular to the logarithmic entropy-energy in-
equality (6.2.2), p. 281, for which (r) = n2 log(A + Cr), r ∈ (0, ∞), and yields
in this case ultracontractive bounds of the form Pt 1,∞ ≤ Ct −n/2 as described
in Sect. 6.3, p. 286. It also shows that (r) = A + Cr α , r ∈ (0, ∞), gives ultra-
contractive bounds as soon as α < 1 from which it appears that the standard log-
arithmic Sobolev case is a limiting case
∞ for ultracontractivity. It may furthermore
happen that limr→∞ (r) = 0 while (r) dr
r = ∞. In this case, the semigroup
is said to be immediately hypercontractive in the sense that for any t > 0 and any
1 < p < q < ∞, Pt is a bounded operator from Lp (μ) into Lq (μ).
Proof of Theorem 7.1.2 According to the preceding comments, it is enough to deal

with 2 ≤ p ≤ q < ∞ (the case q = ∞ being handled by letting q → ∞). The proof
closely follows that of Theorem 5.2.3, p. 246, concerning hypercontractivity. For
notational convenience, set q0 = p ≤ q = q1 . For some positive bounded function
f in D(E), let q = q(t), t ≥ 0, be smooth enough with q(0) = q0 (= p). Setting

(t) = E (Pt f )
q(t) dμ, t ≥ 0, and arguing as in the proof of Theorem 5.2.3,
q (t)
(t) = Entμ (Pt f )q(t) + (t) log (t) − q(t) q(t) − 1 Eq(t) (Pt f )
q(t)

where we set Eq (f ) = E f q−2 (f )dμ to ease the notation. Now, (7.1.2) applied
to (Pt f )q/2 yields that for all r > 0,

q q
≤q (r) − (q − 1) Eq (Pt f ) + (r) + log .
4 q
Next choose q → r(q) ∈ (0, ∞) and take q = q(t) which solves the differential
equation q (r(q)) = 4(q − 1). The preceding inequality then reads
q
(log ) ≤ r(q) + log ,
q
or equivalently

1 q
log ≤ 2
r(q) .
q q
Integrating this differential inequality between 0 and t yields
t
q
log (t) 1/q(t)
≤ log (0) 1/q(0)
+ 2
r(q) ds.
0 q
In other words, for any choice of the function q → r(q), and using q as a variable
instead of t, if
q1
dq
t= r(q)
q0 4(q − 1)
with q0 < q1 , then

q1 dq
Pt q0 ,q1 ≤ exp r(q) 2 .
q0 q
It remains to choose the map q → r(q) in order to optimize this bound when
q0 = q(0), q1 = q(t) and t are given. The computation shows that the optimal choice
δq 2
indeed does not depend on , and is given by q → r(q) = q−1 . The final result is
then obtained after a change of variable. Theorem 7.1.2 is established.
As mentioned previously, Theorem 7.1.2 is of most interest when r −1 (r) is

integrable at infinity in which case q = ∞ may be reached in the conclusion. This
main instance deserves a separate statement. As for the ultracontractive bounds of
Sect. 6.3, p. 286, recall that whenever Pt 1,∞ is finite, the operators Pt , t > 0,
are represented by (symmetric) density kernels pt (x, y), t > 0, (x, y) ∈ E × E, in
L2 (μ) such that Pt 1,∞ = pt (·, ·)L∞ (μ⊗μ) .
Corollary 7.1.4 In the setting of Theorem 7.1.2, if the entropy-energy inequality

EE() holds for a growth function such that r −1 (r) is integrable at infin-
ity, then, for any δ > 0, Pt (δ) 1,∞ ≤ em(δ) where t (δ) and m(δ) are represented
through the parameter δ > 0 as
⎧
⎪ 1 ∞ dr
⎪
⎨t (δ) = (δr) √ ,
2 1 r(r − 1)
∞ (7.1.7)
⎪
⎪ 1 dr
⎩m(δ) = (δr) √ .
2 1 r r(r − 1)
Conversely, if for some t ∈ (a, ∞), a ≥ 0, Pt 1,∞ ≤ K(t), then EE() holds for
defined by

(r) = inf 2tr + log K(t) , r ∈ (0, ∞).
t>a
Proof The first assertion follows from Theorem 7.1.2 with p = 1 and q = ∞,
changing r into 4r and δ in 4δ. Conversely, as consequence of the Riesz-Thorin
√
Theorem, Pt 1,∞ ≤ K implies that Pt 2,∞ ≤ K. By Proposition 5.2.6, p. 248,

2
it follows that whenever E f dμ = 1,

Entμ f 2 ≤ 2t E(f ) + log K(t).
Optimizing in t > a yields the announced EE() inequality.
The example of K(t) = B t −n/2 , t > 0, in Corollary 7.1.4 (as is the case under
a Sobolev inequality, cf. Theorem 6.3.1, p. 286), yields a growth function of the
form (r) = n2 log(A + Cr), r ∈ (0, ∞), in the entropy-energy inequality EE().
In general, the uniform bound on Pt 1,∞ reflects the behavior of at infinity as
t → 0, and the behavior of at 0 as t → ∞. In this way, the next statement recov-
ers from Corollary 7.1.4 the ultracontractive bounds under a Sobolev inequality of
Theorem 6.3.1, p. 286, through entropy-energy inequalities.
Corollary 7.1.5 In the setting of Theorem 7.1.2, if the entropy-energy inequality

EE() holds for a growth function such that (r) − 2r
n
is integrable at infinity,
then
C
Pt 1,∞ ≤ , 0 < t ≤ 1.
t n/2
In the same way, if (r) ∼ n
2 log r as r → 0, and if r −1 (r) is integrable at infinity,
then
C
Pt 1,∞ ≤ , t ≥ 1.
t n/2
Proof When (r) − 2r n

is integrable at infinity, (r) = n2 log r + O(r) as r → ∞.
It follows that when δ → ∞, using the form (7.1.7),

n 1 n
t (δ) = +o , m(δ) = log(δ) + O(δ),
2δ δ 2
from which the bound on Pt 1,∞ is easily deduced. The behavior at t → ∞ is
treated in the same way.
A somewhat striking application of the preceding developments concerns the Eu-

clidean logarithmic Sobolev inequality of Proposition 6.2.5, p. 284, which actually
yields optimal heat kernel bounds. The proof simply follows from the expression of
the optimal function (r) = n2 log( nπe
2r
), r ∈ (0, ∞), and the explicit values of t (δ)
and m(δ) of (7.1.7) in this case.
Corollary 7.1.6 Under the Euclidean logarithmic Sobolev inequality EE() with
(r) = n2 log( nπe
2r
), r ∈ (0, ∞),
1
Pt 1,∞ ≤ , t > 0.
(4πt)n/2
Therefore, in particular, pt (x, y) ≤ 1

(4πt)n/2
uniformly in t > 0 and (μ ⊗ μ-almost
everywhere) in (x, y) ∈ E × E.
The function (r) = n2 log( nπe2r

), r ∈ (0, ∞), thus appears as the best growth
function which bounds entropy from energy in the Euclidean case since from
the standard representation (2.1.1), p. 78, the heat kernel bound is obviously the
best possible. After simple (but tedious) computations, it also follows that, for every
1 ≤ p ≤ q ≤ ∞ and every t > 0,
* 1
1− p1 +n * + n2 ( p1 − q1 )
p p (1 − p1 ) 2
1 1 1
Pt p,q ≤ 1
− . (7.1.8)
1− q1 4πt p q
q q (1 − q1 )
These intermediate bounds are again optimal on the standard heat semigroup on Rn .
Indeed, since this bound is optimal when p = 1 and q = ∞, thanks to the compati-
bility relations between the (p, q) norms and the semigroup property, any interme-
diate bound is also optimal.
When the measure μ is finite (and normalized into a probability measure), the
growth function in the definition of an entropy-energy inequality EE() takes
positive values (since Entμ (f 2 ) ≥ 0). Furthermore, as mentioned earlier, the exis-
tence of a Poincaré inequality is equivalent to the fact that may be chosen so
that (r) ≤ Cr near 0. As it is known from Corollary 6.4.1, p. 290, as soon as
(Pt )t≥0 is ultracontractive and the measure is finite, the operators Pt , t > 0, are
Hilbert-Schmidt and therefore the generator L has a discrete spectrum. In particular,
a Poincaré inequality holds, and it may therefore be assumed that (r) ≤ Cr for
7.2 Off-diagonal Heat Kernel Bounds 355
some constant C, or equivalently that (0) < ∞. Note also that, from the general
Definition 7.1.1, tightness in EE() is achieved as soon as (0) = 0. As a conse-
quence of the discussion in Sect. 7.5 below (Corollary 7.5.7), when (0) = 0 it is
possible to choose (some other) growth function such that (r) ≤ Cr for some
C > 0.
When (0) < ∞ and r −1 (r) is integrable at infinity, Corollary 7.1.4 actually
yields bounds on the convergence to equilibrium. The following statement again
follows from a precise analysis of t (δ) and m(δ) as δ → 0 in (7.1.7). The details are
left to the reader.
Corollary 7.1.7 In the setting of Theorem 7.1.2, and whenever μ is a probability

measure, if the entropy-energy inequality EE() holds for a growth function
such that (0) < ∞ and r −1 (r) is integrable at infinity, then as t → ∞,

log Pt 1,∞ ≤ A e− (0)/2t 1 + o(t)
where A and C only depend on and are explicitly given by

∞ 1 ∞
dr (r) dr (r) dr
A = 2eC (r) and C = − 1 + .
0 r2 0 (0) r 1 (0) r
7.2 Off-diagonal Heat Kernel Bounds
This section addresses the issue of off-diagonal estimates on the density kernels
pt (x, y), t > 0, (x, y) ∈ E × E, of a Markov semigroup (Pt )t≥0 (with respect to the
invariant measure) under an entropy-energy inequality EE(). Actually, as soon as
uniform bounds are available, that is a control of pt (x, x) or pt (x, y) uniformly over
(x, y) ∈ E × E (almost everywhere), there are in general also off-diagonal estimates
taking into account the (intrinsic) distance d(x, y) between x and y associated with
the Markov Triple (E, μ, ) as considered in (3.3.9), p. 166. Recall that the distance
function (Sect. 3.3.7, p. 166) refers to the extended algebra A, justifying the Full
Markov Triple assumption (cf. Sect. 3.4, p. 168).
Gaussian bounds of the type
pt (x, y) ≤ C(t) e−d

2 (x,y)/ct
, t > 0, (x, y) ∈ E × E, (7.2.1)
(for constants C(t) > 0, c > 0) are of interest. Such heat kernel bounds were already
considered in Sect. 6.7, p. 296, under geometric features by means of Harnack-type
inequalities. In this section, a general method of providing such bounds is developed
under functional inequalities in entropy-energy form.
7.2.1 Diameter Bounds
In the finite measure case, whenever C(t) < 2 in (7.2.1) (with this in mind, re-
call Proposition 6.3.4, p. 289), Proposition 1.2.6, p. 15, provides a uniform lower
bound on p2t , and therefore such heat kernel bounds can only hold if the diam-
eter D = D(E, μ, ) of E with respect to the distance d (cf. (3.3.10), p. 167) is
finite. In this respect, we first state a bound on the diameter under an entropy-energy
inequality EE() for a suitable growth function . It is obtained exactly as Propo-
sition 6.6.1, p. 294, under logarithmic entropy-energy inequalities by means of the
Herbst argument (Proposition 5.4.1, p. 252).
Proposition 7.2.1 Let (E, μ, ) be a Markov Triple, with μ a probability

measure,
∞ satisfying an entropy-energy inequality EE() where is such that

0 (r) dr
r 3/2
< ∞. Then, for every integrable 1-Lipschitz function f and every
s > 0,
2
s s /4 dr
e dμ ≤ exp s
sf
f dμ + (r) 3/2 .
E E 4 0 r
In particular,
∞ ∞ ∞
1 dr dr dr
D≤ (r) 3/2 = (r) 1/2 = (r) .
2 0 r 0 r 0 r 3/2
Recall that a standard logarithmic Sobolev inequality LS(C) with (r) = 2Cr,
r ∈ (0, ∞), does not yield any finite diameter (for example the standard Gaussian
measure in Rn ). As for Poincaré and logarithmic Sobolev inequalities, the Laplace
transform
bounds of Proposition 7.2.1 may be used towards tails estimates on
μ(|f − E f dμ| ≥ r), r > 0, at rates reflected by the growth of . Further, and
stronger, estimates will be illustrated later in the context of generalized Nash in-
equalities.
7.2.2 Off-diagonal Heat Kernel Bounds
The next theorem presents the announced off-diagonal estimates under an entropy-
energy inequality EE(). According to the previous discussion and Proposi-
tion 7.2.1, in the finite measure case the statement is restricted to the case when
EE() implies a bounded diameter. The technique developed in the proof by a
suitable transformation of the semigroup under a Lipschitz function turns out to be
most efficient in many situations.
Theorem 7.2.2 Let (E, μ, ) be a Markov Triple satisfying

∞ an entropy-energy in-
equality EE() where is such that dr
(r) √ r
< ∞. Setting

∞
U (s) = s (r) √
dr
r
, s > 0, define τ = U −1 (d), d = d(x, y), (x, y) ∈ E × E, and
∞
1 (r)
T (d) = √ dr.
2 τ r(r − τ )
Furthermore, for 0 ≤ t ≤ T (d), let

√ 1 ∞ √ δ−τ dr
H (t, d) = t (δ − τ ) − δd + (r) δ− √
2 τ r + δ − τ r 3/2
where δ > 0 is defined from t and d so that

1 ∞ (r)
t= √ dr.
2 τ r(r + δ − τ )
Then, for 0 < t ≤ T (d(x, y)), (x, y) ∈ E × E, the density kernels pt (x, y), t > 0, of
the semigroup (Pt )t≥0 satisfy

log pt (x, y) ≤ H t, d(x, y)
for (μ ⊗ μ-almost) all (x, y) ∈ E × E.
Observe that in the finite measure case, the diameter D is bounded from above
by U (0) according to Proposition 7.2.1 so that T (d) is well-defined on [0, D]. Al-
though the explicit bound put forward in Theorem 7.2.2 appears rather involved,
Corollary 7.2.3 below will actually show that it leads to the expected precise off-
diagonal bounds under Sobolev-type inequalities.
Proof For a function h ∈ A0 such that (h) ≤ δ where δ > 0 will be specified later,
consider the new semigroup Pth f = e−h Pt (eh f ), t ≥ 0. This semigroup is no longer
Markov, but is still positivity preserving. It is symmetric with respect to the measure
e2h dμ, but this symmetry will actually not be used since we mostly work with the
initial measure μ. The first task is to reach, under the EE() inequality, the bound
h
P
t 1,∞ ≤ e
m(t,δ)
(7.2.2)
where m(t, δ) only depends on t and δ (and of course on the function itself).
Hence the density kernel pth (x, y) with respect to μ is uniformly bounded from
above by em(t,δ) . But, by construction, pth (x, y) = pt (x, y)eh(y)−h(x) so that, for all
t > 0 and (x, y) ∈ E × E,
pt (x, y) ≤ em(t,δ)+h(x)−h(y) .
Now, by definition of the distance d(x, y) ((3.3.9), p. 166),

√
inf h(x) − h(y) = − δ d(x, y).
(h)≤δ
Minimize then the previous bound on pt (x, y) on all functions h such that (h) ≤ δ
to get that
√
pt (x, y) ≤ em(t,δ)− δd(x,y)
. (7.2.3)
The last step will then amount to optimizing δ.

The challenge is therefore to reach the upper bound (7.2.2) on Pth 1,∞ . To this
end, we may proceed as in the proof of Theorem 7.1.2, retaining

the same nota-
tion. For some positive function f ∈ A0 , consider (t, q) = E (Pth f )q dμ, t ≥ 0,
where q ∈ [1, ∞). From standard computations and integration by parts, setting
g = g(t) = Pth f ,

4(q − 1)
∂t = − g q/2 dμ + q g q (h)dμ
q E E

− q(q − 2) g q−1 (g, h)dμ.
E
By |(g, h)| ≤ (g)1/2 (h)1/2 and the Cauchy-Schwarz inequality in L2 (μ),

q q/2
g (g, h)dμ = g g , h dμ
q−1 q/2
2 E E
1/2 1/2

≤ g q (h)dμ g q/2 dμ .
Linearizing this last inequality, for any α > 0,

4(q − 1) |q − 2| q/2
∂t ≤ − + g dμ + q + α|q − 2| g q (h)dμ
q α E E

4(q − 1) |q − 2|
≤ − + g q/2 dμ + δ q + α|q − 2| .
q α E
For T > 0, choose then an increasing function t → q(t) with q(0) = 1 and
q(T ) = ∞ to be specified later and consider on [0, T ] the quantity (t) =
(t, q(t)), t ≥ 0. The inequality EE() in the form (7.1.2) leads to the differential
inequality on ,

q |q − 2| 4(q − 1)
≤
(r) + − g q/2 dμ
q α q E

q q
+ (r) + δ q + α|q − 2| + log

q q
depending on the two parameters α > 0 and r > 0. These parameters may be cho-
sen to depend upon q, and by further choosing q = q(t) so that the coefficient of
q/2 )dμ
E (g in the previous inequality vanishes, we end up with

1 q δ

log ≤ 2 r(q) + q + α(q)|q − 2|
q q q
where
q|q − 2|
α(q) = . (7.2.4)
4(q − 1) − q (r(q))
Setting
q|q − 2|s
α=
4(q − 1)(s − 1)
where s = s(q) > 1 is a new function, (7.2.4) turns into
(r(q))
dt = s(q) dq.
4(q − 1)

As a consequence, PTh f ∞ ≤ em(T ) E f dμ where

∞ dq
T= s(q) r(q)
1 4(q − 1)
and
∞ 4(q − 1)
m(T ) = r(q)
1 q2

(q − 2)2 s(q) dq
+ δs(p) r(q) 1 + .
4(q − 1)(s(q) − 1) 4(q − 1)
The final step is to choose optimal functions r(q) and s(q) which minimize the
quantity m(T ) when T > 0 is fixed. The optimal choice once again does not depend
on , and is given, for some parameter 0 < κ < 1, by

qq ∗ − 4 qq ∗ s
s = s(q) = 1 + , r = r(q) = δ(1 − κ)
qq ∗ − 4κ 4(2 − s)
where as usual q ∗ is the conjugate of q, that is q ∗ = q

q−1 . After a change of variable,
taking κ ∈ (0, 1) as a parameter,
⎧
⎪ 1 ∞ dw
⎪
⎨t = (δw) √ ,
2 1−κ w(w + κ)
∞
⎪
⎪ 1 κ dw
⎩m = δκt + (δw) 1 − √ ,
2 1−κ w+κ w 3/2
or, with κ = 1 − τ
δ and v = δw,
⎧
⎪ 1 ∞ dv
⎪
⎨t = t (δ, τ ) = 2 (v) √ ,
v(v + δ − τ )
τ
∞
⎪ 1 √ δ−τ dv
⎪
⎩m = m(δ, τ ) = (v) δ − √ + (δ − τ )t.
2 τ v + δ − τ 3/2
v
√
Turning back to (7.2.3), the minimum in δ of m(δ, τ ) − δd when t (δ, τ ) is fixed
is obtained when
∞
1 ∞ (v) (τ ) (v)
d= 3/2
dv − √ = dv
2 τ v τ τ v 1/2
from which the theorem is a direct consequence. The proof is complete.
The following consequence of Theorem 7.2.2 applies to logarithmic entropy-

energy inequalities equivalent to classical Sobolev-type inequalities (cf. Proposi-
tion 6.2.3, p. 281).
Corollary 7.2.3 Under an entropy-energy inequality EE() with (r) =

2 log(A + Cr), r ∈ (0, ∞), there is a constant B > 0 such that for all 0 < t ≤ 1 and
n
(x, y) ∈ E × E,
pt (x, y) ≤ B t −n e−d
2 (x,y)/4t
.
The details, a direct consequence of Theorem 7.2.2, are left to the reader. It is
worth mentioning that the preceding corollary achieves the optimal 4t in the expo-
nential factor, although the polynomial term is only of order t −n and not t −n/2 as in
the uniform estimate (Corollary 7.1.5). This is not a failure of the method. Indeed,
on a Riemannian manifold, outside the diagonal, pt (x, y) behaves when t → 0 in
a different way at the cut-locus of x. The correct exponent is t −(n+p)/2 , where p
is the dimension of the manifold of geodesics which go from x to y. For example,
p = n − 1 on two opposite points on an n-dimensional sphere.
7.2.3 Optimal Bounds for the Harmonic Distance
The distance function d used to obtain the off-diagonal estimates in Theorem 7.2.2 is
built from the generator L of the Markov semigroup (Pt )t≥0 . Other kinds of pseudo-
distances may be considered in the same way, such as for example, the harmonic
distance d H (x, y) defined as

d H (x, y) = esssup h(x) − h(y) , (x, y) ∈ E × E, (7.2.5)
the esssup running over all functions h in A such that Lh = 0 and (h) ≤ 1. Of
course d H = 0 on a compact manifold with respect to the Laplace-Beltrami operator
since every harmonic function is then constant. But in Rn with the usual Laplace
operator, d H (x, y) = |x − y| since the linear functions are harmonic and attain the
distance between x and y. With this definition, optimal heat kernel bounds may be
achieved.
Proposition 7.2.4 Let (E, μ, ) be a Markov Triple satisfying an entropy-energy

inequality EE() for some growth function such that r −1 (r) is integrable at
infinity. Then, for every δ > 0 with m(δ) and t (δ) given by (7.1.7), and provided both
quantities are finite,
H (x,y)2 /4t (δ)
pt (δ) (x, y) ≤ em(δ)−d
for all (x, y) ∈ E × E. In particular, for (r) = n2 log( nπe
2r
), r ∈ (0, ∞), corre-
sponding to the Euclidean logarithmic Sobolev inequality (6.2.8), p. 284, for all
t >0
1
e−d (x,y) /4t
H 2
pt (x, y) ≤
(4πt)n/2
(which is the exact value of the Euclidean heat kernel).
Proof We only sketch the proof, staying at a somewhat informal level. Following
the proof of Theorem 7.2.2, change Pt f into Pt f = e−γ h Pt (eγ h f ) where now
γ
(h) ≤ 1 and Lh = 0, for some real parameter γ . Denoting by Lγ the Markov

γ
generator of the semigroup Pt , t ≥ 0, for every f > 0 in the domain D(L) and in
L1 (μ) ∩ L∞ (μ), and every q ∈ (1, ∞),

f q−1 Lγ f dμ = f q−1 Lf dμ + γ 2 (h)f q dμ.
E E E
Therefore, if, for every r > 0,

q2
Entμ f q ≤ − (r) f q−1 Lf dμ + (r) f q dμ,
4(q − 1) E E
then

q2
Entμ f q ≤ − (r) f q−1 Lγ f dμ + (r) + γ 2 f q dμ.
4(q − 1) E E
Hence, following the proof of Theorem 7.1.2, and with the same values of m and t,
γ
Pt 2
≤ em+γ t ,
1,∞
or equivalently for the density kernels

2 t−γ [h(x)−h(y)]
pt (x, y) ≤ em+γ
for every t > 0 and (x, y) ∈ E × E. It then remains to optimize first in h and then
in γ to get the desired result.
7.3 Examples
This section illustrates some of the results of the preceding sections, both at the
level of the entropy-energy inequalities themselves and of their consequences for
heat kernel bounds. We deal for simplicity with Markov Triples (E, μ, ) on the
Euclidean space E = Rn with (f ) = |∇f |2 and dμ = e−W dx where W : Rn → R
is a smooth potential.
The next proposition describes conditions on the growth of the potential W and
its derivatives in order for an entropy-energy inequality EE() to hold for some
function .
Proposition 7.3.1 In the preceding setting with dμ = e−W dx, let

1
c(s) = sup 2W (x) − |∇W |2 (x) + W (x) , s > 0.
x∈Rn 4πes
Then, an entropy-energy inequality EE() holds for any growth function

: (0, ∞) → R such that for any r > 0, there exists an s > 0 with

n 2r
(r) ≥ log s − 1 + + c(s).
2 nπes
In the preceding, the function c may of course be replaced by any larger function,
and the best choice for (r), r ∈ (0, ∞), is the concave envelope of the function

n r
inf (log s − 1) + + c(s)
s>0 2 πes
(which may then be turned into a C 1 -function).
Proof Start from the Euclidean logarithmic Sobolev inequality (6.2.8) of Proposi-
tion 6.2.5, p. 284, which

indicates that for any smooth compactly supported function
g on Rn such that Rn g 2 dx = 1,

n 2
g 2 log g 2 dx ≤ log |∇g|2 dx .
Rn 2 nπe Rn
As for (7.1.2), by concavity of the logarithmic function, for every s > 0,

n 2
g 2 log g 2 dx ≤ log s − 1 + |∇g|2 dx .
Rn 2 nπes Rn

Set now g 2 = f 2 e−W . Then Rn f 2 dμ = 1 and

1 1
|∇g| dx =
2
|∇f | dμ +
2
f W dμ −
2
f 2 |∇W |2 dμ.
Rn Rn 2 Rn 4 Rn
7.3 Examples 363
It follows that, for every s > 0,

n 2
f log f dμ ≤
2 2
log s − 1 + |∇f | dμ
2
Rn 2 nπes Rn

1
+ f 2
2W − |∇W | + W dμ
2
Rn 4πes

n 2
≤ log s − 1 + |∇f |2 dμ + c(s).
2 nπes Rn
The conclusion then follows from the very definition of .
The above proof is not specifically tied to the Euclidean case and shows how to go
from an entropy-energy inequality EE() on (E, μ, ) to a new one for (E, ν, ),
where dν = e−W dμ.
Proposition 7.3.1 may be applied in several instances of interest, including the
model family μα of (7.0.1) described in the introduction of this chapter. Indeed, if
W (x) = (1 + |x|2 )α/2 with α > 2, it is easily checked that c(s) ≤ Cs α/(α−2) , s ≥ 1.
Here and below, C > 0 depends on n and α, and may change from line to line. As a
consequence, μα with α > 2 satisfies an EE() inequality with

(r) = C 1 + r α/(2α−2) , r ∈ (0, ∞). (7.3.1)
(It may be checked furthermore that whenever α < 2, μα does not satisfy any
EE() inequality, while μ2 satisfies a logarithmic Sobolev inequality, that is
EE() with (r) = Cr, r ∈ (0, ∞).) According to Corollary 7.1.4, for 0 < t ≤ 1,
−α/(α−2)
Pt 1,∞ ≤ eCt .
Conversely, this bound for some t > 0 shows that EE() holds with (7.3.1), so that
the heat kernel bound has the right order of magnitude as t → 0. Finally, Propo-
sition 7.2.1 easily shows that μα (|x| ≥ r) ≤ Ce−r /C , r > 0, confirming again the
α
exponents of .
Along the same lines, if W (x) = |x|2 [log(1 + |x|2 )]α , α > 1, we may choose
c(s) = C(1 + ebs ), s > 0, and (r) = C(1 + r log−α (e + r)), r ∈ (0, ∞), from
1/α
which
−1/(α−1)
Pt 1,∞ ≤ exp eCt , 0 < t ≤ 1.
In the same spirit, it may be checked that for α = 2, the semigroup is not ultracon-
tractive but immediately hypercontractive (cf. p. 351).
The same method applies on a compact interval of the real line. For example, on
(0, 1) with W (x) = −m log x for small x > 0 and W (x) = −p log(1 − x) for x near
to 1, m, p > 2, there exists an entropy-energy inequality EE() with
n
(r) ≤ (C + log r)
2
when r → ∞, where n = max(m + 1, p + 1). In that case c(s) = Cs + κ2 log s where

κ = max(m, p). On the other hand, if W (x) = C1 x −a for small x > 0 and W (x) =
C2 (1 − x)−b for x near 1, a, b > 0, one may choose

(r) ≤ C 1 + r κ/(2κ+2) , r > 0,
with κ = max(a, b). This last inequality is easily seen to be compatible with bound-
edness of the diameter. These examples indicate that the behavior of the function
at infinity reflects the behavior of the measure at the boundaries of the interval, the
smaller the weight around the edges, the bigger the function at infinity.
7.4 Beyond Nash Inequalities

As described in Chap. 6 (Proposition 6.2.3, p. 281), Sobolev inequalities for a
Markov Triple (E, μ, ) may be described equivalently via logarithmic entropy-
energy inequalities, but also as families of Nash (and Gagliardo-Nirenberg) inequal-
ities. This section investigates generalized Nash inequalities on the model of general
entropy-energy inequalities of the preceding section through a growth function .
In a second step, these inequalities will also be extended so as to introduce weights.
This further generalization carries many interesting features, such as bounds on the
heat kernels when there is no a priori uniform bounds, or significant relationships
with spectral properties of the generator, as already discussed in Sect. 4.10, p. 220.
7.4.1 Generalized Nash Inequalities
The generalized Nash inequalities involve a C 1 increasing concave and positive

growth function : (0, ∞) → R+ . The function (r), r > 0, then has a finite limit
at r = 0, denoted (0).
Definition 7.4.1 (Generalized Nash inequality) A Markov Triple (E, μ, ) is said

to satisfy a (generalized) Nash inequality N () with respect to a growth function
: (0, ∞) → R+ if for every f ∈ D(E) with f 1 = 1 and E(f ) > 0,

f 2 dμ ≤ E(f ) .
E
According to the discussion following Proposition 6.2.3, p. 281, the Euclidean

Nash inequality in Rn corresponds to (r) = Cr n/(n+2) , r ∈ (0, ∞) (for some con-
stant C > 0). In the finite measure case, the tight Nash inequality corresponds to
(r) = (1 + Cr)n/(n+2) , r ∈ (0, ∞).
Observe that due to the normalization f 1 = 1, necessarily limr→∞ (r) = ∞
(since if not f 2 ≤ Cf 1 , which may only occur on discrete spaces). Fur-
thermore, it may well be that the Nash inequality N () only holds for func-
tions f such that f 22 ≥ cf 1 for some c > 0. Then may be replaced by
7.4 Beyond Nash Inequalities 365
max(, a0 ). However, the latter function is no longer concave on (0, ∞). Provided
that (r) = (r) − r (r) → ∞ as r → ∞ (which is usually the case), may be
compensated by a linear part near the origin in order to make it concave.
As for the entropy-energy inequality EE(), linearized versions of the Nash
inequality N () may be considered. Whenever is bijective (if it is not bijective,
use its generalized inverse) and takes values in some interval (a, b), then the N ()
inequality is equivalent to the family of inequalities

f 2 dμ ≤ s E(f ) + β(s), s ∈ (a, b), (7.4.1)
E
for every f ∈ D(E), f 1 = 1, where β(s) = (−1 (s)) − s −1 (s).

In this form, such inequalities have already been used in Sect. 4.10, p. 220, to
describe the spectrum of the associated generator L, and in particular to provide cri-
terions related to its discreteness. Generalized Nash inequalities will be used in these
linearized versions in the investigation of measure-capacity inequalities in Chap. 8.
For probability measures μ, generalized Nash inequalities are closely related to
Poincaré inequalities (Definition 4.2.1, p. 181).
Proposition 7.4.2 Let (E, μ, ) be a Markov Triple with μ a probability measure.

A Poincaré inequality P (C) with constant C > 0 is equivalent to a Nash inequality
N() with (r) = 1 + Cr, r ∈ (0, ∞).
Proof The proof is straightforward. Under P (C), for every f ∈ D(E) with
f 1 = 1,
2
f 2 dμ ≤ C E(f ) + f dμ ≤ C E(f ) + 1.
E E
Conversely, apply N () with (r) = 1 + Cr to 1 + εf where f is a bounded
function in D(E). For ε > 0 small enough so that 1 + εf ≥ 0, N () boils down to
2
1 + 2ε f dμ + ε 2 f 2 dμ ≤ ε 2 C E(f ) + 1 + 2ε f dμ + ε 2 f dμ
E E E E
from which the Poincaré inequality P (C) follows as ε → 0.
Remark 7.4.3 When μ is a probability, tightness in N () corresponds to the fact

that limr→0 (r) = 1, while the Poincaré inequality requires moreover that has
a bounded slope at the origin. However, as for entropy-energy inequalities, as soon
as limr→0 (r) = 1, Corollary 7.5.7 below shows that a Poincaré inequality also
holds, and therefore that the growth function may then be replaced by another
one satisfying (r) ≤ 1 + Cr, r > 0.
Generalized Nash inequalities may be compared to the entropy-energy inequali-

ties of Sect. 7.1, and are sometimes equivalent to the latter. However, the Nash fam-
ily covers more general examples than the entropy-energy inequalities since they
include Poincaré inequalities while the entropy-energy inequalities at least require

stronger logarithmic Sobolev inequalities (together with tightness).
Proposition 7.4.4 Let (E, μ, ) be a Markov Triple satisfying the entropy-energy

inequality EE(1 ). Then it also satisfies the Nash inequality N (2 ) where 2 (r) =
r[U (log r)]−1 with U the inverse function of 1 + log r, r ∈ (0, ∞). In particular,
if μ is a probability measure, the tight EE(1 ) inequality (1 (0) = 0) implies the
tight N(2 ) inequality (2 (0) = 1).
Proof Recall as in the proof of Proposition 6.2.3, p. 281, that the convex function φ :
Entμ (f 2 )
r → log(f 1/r ), r ∈ (0, 1], is such that φ ( 12 ) = − . Hence, by convexity,
f 22

f 22 Entμ (f 2 )
log ≤ .
f 21 f 22
Therefore, for f ∈ D(E) such that f 1 = 1, setting a = f 22 and b = E(f ),

EE(1 ) indicates that log a ≤ 1 ( ab ), or in other words,

b b
log b ≤ 1 + log .
a a
Since 1 is increasing, so is 1 (r) + log r, r ∈ (0, ∞), and therefore U (log b) ≤ ab ,

which immediately leads to the announced claim.
As an example, a (defective) logarithmic Sobolev inequality LS(C, D) (Defini-

tion 5.1.1, p. 236), thus with 1 (r) = 2Cr + D, r ∈ (0, ∞), leads to a Nash inequal-
ity N (2 ) with 2 (r) ∼ C1 logr r as r → ∞. In the same way, 1 (r) = n2 log Cr,
r ∈ (0, ∞), leads to 2 (r) ∼ C1 r n/(n+2) as r → ∞ (classical Nash inequalities),
while 1 (r) = Cr α , r ∈ (0, ∞), α > 0, leads to 2 (r) ∼ C1 r(log r)−1/α as r → ∞.
At least when μ is a probability measure, the implication in the preceding propo-
sition may easily be reversed to show that under a Nash inequality N (2 ), and
provided 2 log(2 ) is concave at infinity, an entropy-energy inequality EE(1 )
holds with 1 (r) = ae−r −1 2 (e ) + b, r ∈ (0, ∞), for constants a, b > 0. This
r
is an easy consequence of Proposition 5.1.8, p. 241, and the linearization forms

of both inequalities. Unfortunately, this conclusion is too rough for example to
reach the dimensional Sobolev parameter from the classical Nash inequality and
2 (r) = C(1 + r)n/(n+2) only yields 1 (r) ≤ an 2 log(C1 (1 + r)) + b, r > 0 (al-
though a = 1 may actually be reached with some further analysis).
7.4.2 Heat Kernel Bounds and Tail Inequalities Under Generalized

Nash Inequalities
As for the classical Nash inequalities in the proof of Theorem 6.3.1, p. 286, the
generalized Nash inequalities N () may be used to obtain heat kernel bounds and
ultracontractivity provided does not grow too fast at infinity. While easier to ob-
tain than under an entropy-energy inequality EE(), the bounds are in general less
precise (and do not reveal any information on the · p,q norms). However, as de-
scribed in the next section, the same method may provide non-uniform bounds on
the heat kernel when the latter is not necessarily bounded. The next statement is a
first illustration of this phenomenon.
Theorem 7.4.5 Let (E, μ, ) be a Markov Triple satisfying

∞ adrNash inequality
N() with growth function : (0, ∞) → R+ such that (r) r < ∞. Let U be
the (generalized) inverse of , defined on
((0), ∞), and define then K(s), s > 0,
∞
to be the (decreasing) inverse function of U (s) (r) dr
r . Then, for every t > 0,
Pt 21,2 ≤ K(2t).
In particular, the density kernel pt (x, y), t > 0, (x, y) ∈ E × E, of Pt with respect
to the invariant measure μ is uniformly bounded from above by K(t).

∞
Observe from the hypotheses that since U ((0)) = 0, 0 (r) dr
r = ∞, so that
the function K is a decreasing bijection from (0, ∞) onto ((0), ∞).
Proof The proof is very similar to that of the corresponding result for the classical
Nash inequalities in Theorem

6.3.1, p. 286, so we shall only
sketch it. For a positive
function f in D(E) with E f dμ = 1, consider (t) = E (Pt f )2 dμ, t ≥ 0. The

N() inequality turns into the differential inequality ≤ (− 2 ). Fix now t > 0.
Since K(2t) ∈ [(0), ∞), if (t) ≤ (0) there is nothing to prove. If not, since
is decreasing, (s) > (0) for s ∈ (0, t). Therefore, by definition of the inverse
function U , (s) ≤ −2U ( (s)) for s ∈ (0, t). Introducing the decreasing function
∞ ∞
du dr
H (s) = = (r) , s ∈ (0), ∞ ,
s U (u) −1 (s) r
it follows that H ( (s)) ≥ 2 when (s) ≥ (0). Therefore, (t) ≤ (0) or

H ( (t)) ≥ 2t + H ( (0)) ≥ 2t, t > 0. By definition of K, (t) ≤ K(2t) and the
upper bounds on the heat kernels are then deduced as usual.

∞ dr
Remark 7.4.6 When the function is such that (r) r = ∞, the power of
the preceding strategy actually still yields useful, although weaker, results. Namely,
choosing any real a
> 0 and letting then K(s), s > 0, be the (decreasing) inverse
function of H (s) = U (s) (r) dr
a
r , for any f with f 1 = 1 and any t > 0,

Pt f 22 ≤ K H f 22 + 2t .
The latter is even equivalent to the generalized Nash inequality N () as t → 0.

The advantage of the previous form with respect of this last one is that under the
hypothesis of Theorem 7.4.5, the bound on Pt f 22 no longer depends on f 22 .
Generalized Nash inequalities may also be used to obtain tail estimates for Lip-
schitz functions. The following statement goes somewhat beyond Proposition 7.2.1
in case of entropy-energy inequalities.
Proposition 7.4.7 Let (E, μ, ) be a Markov Triple, with μ a probability measure,

satisfying a Nash inequality N (), and let f be a 1-Lipschitz function on E. Let
r0 > 0 be such that q0 = μ(f ≥ r0 ) ≤ 2(0)
1
. Then, for every r ≥ r0 ,

μ(f ≥ r) ≤ 2F (r − r0 ) log 2
where F is the inverse function of

q0
1
−1/2
G(v) = u3 −1 du, v ∈ (0, q0 ).
v/2 2u

1
Whenever 02(0) [u3 −1 ( 2u
1 −1/2
)] du < ∞, then f is bounded (μ-almost every-
where), and in particular the diameter of (E, μ, ) is finite.

q0
In this statement, F is defined on (0, A), where A = 0 [u3 −1 ( 2u
1 −1/2
)] du.
When A < ∞, it is extended by 0 on [A, ∞).
Proof For r, s > 0, apply N () to g = 1s [(f − r)+ ∧ s] where f ∈ D(E). Since
1{f ≥r+s} ≤ g ≤ 1{f ≥r} and (g) ≤ s −2 1{f ≥r} , using that is concave, and there-
fore that u−1 (u) is decreasing, it follows that

1
q(r + s) ≤ q 2 (r) 2
s q(r)
where q(u) = μ(f ≥ u), u > 0. Choose then

−1/2
−1 1
s = q(r)
2q(r)
so that q(r + s) ≤ q(r)

2 . Now, with r(q) = inf{u > 0 ; q(u) ≤ q}, q > 0, which plays
the role of the inverse function of q(r), the previous inequality indicates that
−1/2
q −1 1
r ≤ r(q) + q .
2 2q
Recalling that u−1 (u) is decreasing,

−1/2 −1/2 q
1 1 1 du
q −1 = q −1
2q 2q log 2 q/2 u
q −1/2
1 3 −1 1
≤ u du.
log 2 q/2 2u
Applying the same procedure on ( q4 , q2 ) and iterating yields that

−1/2
−k q
3 −1 1
r 2 q − r(q) log 2 ≤ u du
2−k q 2u
for any k ≥ 1. Using that r(q) is decreasing, it finally follows that for any q1 < q
−1/2
q 1
r(q1 ) − r(q) log 2 ≤ u3 −1 du
q1 /2 2u
which yields the announced claim.
To conclude this section, we illustrate some of the previous results again on the
family μα defined in (7.0.1), in dimension one for simplicity. As mentioned ear-
lier (p. 363), these measures μα with α ∈ [1, 2) on the real line do not satisfy any
entropy-energy inequality (with respect to the standard carré du champ operator
(f ) = f 2 ). However an entropy-energy inequality EE(1 ) holds when α ≥ 2
with 1 (r) = C(1 + r α/(2α−2) ), r ∈ (0, ∞), and in turn a generalized Nash in-
equality N (2 ) with 2 ∼ r(log r)−2(α−1)/α at infinity (Proposition 7.4.4). For this
growth function, the tail estimates produced by Proposition 7.4.7 provide the correct
behavior of μα .
Using arguments close to those developed for the Muckenhoupt criterion
(cf. Sect. 4.5.1, p. 194), it is not difficult to see that this Nash inequality actu-
ally extends to every α ∈ [1, 2) (and is optimal according to the tail estimates of
Proposition 7.4.7). We only sketch the method, which may be used in other, similar,
instances. In what follows, the constant C > 0 depends only on α and may vary
from place to place.

It is enough to deal with a smooth function f on R such that
f (0) = 0 and R f 2 dμα = 1. We may moreover work separately on [0, ∞) and

∞
(−∞, 0], say [0, ∞) (thus assuming 0 f 2 dμα = 1). If p(x) = e−(1+x ) is the
2 α/2
density of μα (up to the normalizing constant), the only ingredient is actually the
tail estimate μα ([x, ∞)) ≤ Cx 1−α p(x), x > 0. For some parameter a > 0, since
f (0) = 0, write
∞ ∞
f 2 dμα = f 2 1{|f |≤a} dμα
0 0
∞ ∞ (7.4.2)

+2 f (x)f (x) 1{|f |>a} dμα dx.
0 x
The first integral is bounded above by af 1 , while for the second, use that
∞

1{|f |>a} dμα ≤ min μα [x, ∞) , a −2 ≤ C min p(x)x 1−α , a −2 .
x
If x0 = x0 (a) is the point where x 1−α p(x) = a −2 , then the right-hand side in the
latter inequality is bounded from above by aCp(x)
2 p(x ) on [0, x0 ) and by Cx0
1−α
p(x)
0
on [x0 , ∞), hence bounded everywhere by x01−α p(x). By the Cauchy-Schwarz in-
equality, the second integral on the right-hand side of (7.4.2) is then bounded from
above by Cx01−α E(f )1/2 , and therefore
∞
1= f 2 dμα ≤ af 1 + Cx01−α E(f )1/2 .
0
It remains to replace x0 by its value in terms of a and to optimize in a to get that

∞

f 2 dμα ≤ E(f )
0
for every f ∈ D(E) with f 1 = 1 and with (r) ≤ Cr(log r)−2(α−1)/α for large r.
The announced claim follows.
7.4.3 Weighted Nash Inequalities
The main advantage of generalized Nash inequalities with respect to entropy-energy

inequalities is that they may address measures which do not satisfy any logarith-
mic Sobolev inequality. However, as far as ultracontractive bounds are concerned,
they do not offer much more up to now. But with some further mild extension,
the same kind of ideas may lead to heat kernel upper bounds which are not uni-
form. In particular, non-ultracontractive settings such as E = R, (f ) = f 2 and
dμα = cα e−(1+x ) , α ∈ (1, 2], on the line may be covered.
2 α/2
In this respect, this short section is concerned with an extension of the Nash
inequalities by the introduction of a weight function which offers more flexibility
and produces further heat kernel bounds. It deals with a Markov Triple (E, μ, )
where μ is a probability measure. Here, a weight function will be a positive (mea-
surable) function w : E → R+ such that Pt w ≤ L(t)w, t ≥ 0, for some increas-
ing and positive function L on R+ . The next definition of a generalized weighted
Nash inequality is entirely similar to the generalized Nash inequality provided the
L1 (μ)-norm is weighted by w. It involves similarly a C 1 increasing and concave
growth function : (0, ∞) → R+ .
Definition 7.4.8 (Weighted Nash inequality) A Markov Triple (E, μ, ), with μ
a probability measure, is said to satisfy a (generalized) weighted Nash inequality
N(, w) with growth

function : (0, ∞) → R+ and weight w : E → R+ if, for
every f ∈ D(E) with E w|f |dμ = 1 and E(f ) > 0,

f 2 dμ ≤ E(f ) .
E
Such inequalities were already considered earlier in Sect. 4.10 (cf. (4.10.4),
p. 224) in connection with emptyness of the essential spectrum. The extra condition
Pt w ≤ L(t)w on the weight, together with special properties of the growth function
, will actually allow for quantitative estimates on the (discrete) spectrum, as well
as pointwise estimates on the density kernels pt (x, y).
The next statement is very similar to Theorem 7.4.5 and borrows from the latter
the definition of K(t), t ≥ 0.
Proposition 7.4.9 Let (E, μ, ) be a Markov Triple satisfying a weighted Nash in-
equality N(, w) with growth function : (0, ∞) → R+ such that r −1 (r) is inte-
grable at infinity. Then (Pt )t≥0 has density kernels pt (x, y), t > 0, (x, y) ∈ E × E,
such that, for every t > 0 and (x, y) ∈ E × E,
2
t
pt (x, y) ≤ K(t)L w(x)w(y), (7.4.3)
2
Proof The proof follows that of Theorem 7.4.5 and the

method developed for the
proof of Theorem 7.2.2. For f positive in D(E) with E f wdμ = 1 and t > 0,

let (s) = L(t)−2 E (Ps f )

2 dμ, s ∈ [0, t]. By symmetry

and the hypothesis on the
weight w, for every s ≤ t, E wPs f dμ ≤ L(t) E wf dμ. Arguing then exactly as
in the proof of Theorem 7.4.5 now under N (, w) and using that r → r −1 (r) is
decreasing yields that
Pt f 22 ≤ K(2t) L(t)2
for every t > 0. Introduce then the new semigroup Qt f = w1 Pt (wf ), t ≥ 0, sym-
metric with respect to the weighted measure dν = w 2 dμ. In terms of the semigroup
(Qt )t≥0 , and with Lp -norms with respect to ν, the preceding reads
Qt f 22 ≤ K(2t)L(t)2
for every t > 0 and f such that f 1 = 1. This leads to an ultracontractive bound
on Q2t , t > 0, and therefore to the uniform bound K(2t)L(t)2 on its kernel with
respect to ν. But if the semigroup (Qt )t≥0 has a kernel bounded from above by
then (Pt )t≥0 has a density kernel, with respect to μ, which is bounded from
K,
above by K w(x)w(y). The conclusion then easily follows and Proposition 7.4.9 is
established.
Proposition 7.4.9 is most useful when w ∈ L2 (μ), in which case the condition
Pt w ≤ ect w, t ≥ 0, holds as soon as w ∈ D(L) and Lw ≤ cw. The function w is thus
similar to the Lyapunov functions of Sect. 4.6, p. 201. Furthermore, since r −1 (r),
r > 0, is decreasing (as is concave and increasing), the condition Pt w ≤ L(t)w,
t ≥ 0, may be replaced by Pt w1 ≤ L1 (t)w1 , t ≥ 0, for some new weight w1 ≥ w.
In particular, adding a constant, it may always be assumed that w is bounded from
below by some strictly positive constant.
Corollary 7.4.10 Under the hypotheses of Proposition 7.4.9, and assuming fur-
thermore that w ∈ L2 (μ), the semigroup P = (Pt )t≥0 is Hilbert-Schmidt, and if
(λk )k∈N denotes the sequence of eigenvalues of −L,
2
−λk t t
e ≤ K(t)L w 2 dμ
2 E
k∈N
for every t > 0.
The proof is immediate. As w ∈ L2 (μ), the upper bound (7.4.3) on pt (x, y) is in

L2 (μ ⊗ μ), and therefore Pt is Hilbert-Schmidt for
every t > 0. The upper
bound
on the eigenvalues follows from the trace formula E pt (x, x)dμ(x) = k∈N e−λk t
(see (1.7.4), p. 32).
Two-sided estimates may be obtained as in the case of standard Nash inequalities
as soon as a Poincaré inequality holds. Recalling from the proof of Proposition 7.4.9
the semigroup Qt f = w1 Pt (wf ), t ≥ 0, symmetric with respect to the measure dν =
w 2 dμ, since (Qt )t≥0 and (Pt )t≥0 share the same spectral properties, if a Poincaré
inequality holds for (Pt )t≥0 with constant C > 0, then Qt − Q∞ 2,2 ≤ Ke−t/C ,

t ≥ 0, with Q∞ f = w1 E wf dμ, the operator norm being understood in L2 (ν).

As soon as QT is bounded from L1 (ν) into L2 (ν) for some T > 0, then using the
decomposition
Qt+T − Q∞ = QT (Qt − Q∞ )QT ,
Qt − Q∞ 1,∞ ≤ Ke−t/C for large t ≥ T . A uniform bound for the modulus of
the density kernel of Qt − Q∞ follows for large t, which in turn may be translated
into a bound for the density kernel of Pt . The conclusion is summarized in the next
statement with a proof similar to the one given for Proposition 6.3.4, p. 289.
Corollary 7.4.11 Under the hypotheses of Proposition 7.4.9, assuming further-

more that (E, μ, ) satisfies a Poincaré inequality P (C) for some C > 0, there
exist a constant K > 0 and T > 0 such that, for t ≥ T , the density kernel pt (x, y)
of Pt , satisfies

pt (x, y) − 1 ≤ K e−C/t w(x)w(y)
for all (x, y) ∈ E × E.
Of course, since in general w is not bounded and goes to infinity at infinity, the
lower bound is only useful (for a given t > 0) on a compact subset of E × E, this
subset enlarging to the full space when t goes to infinity.
7.5 Weak Poincaré Inequalities 373
We conclude this section with some examples. Starting from a classical Nash
inequality for a given operator L, it is not hard to deduce weighted Nash inequalities
for the operator L − (W, ·), that is to pass from a Nash inequality for some measure
μ to a weighted Nash inequality for the (weighted) measure e−W dμ. The procedure
is similar to the technique developed for entropy-energy inequalities in Sect. 7.3. We
briefly illustrate it in Rn for the Lebesgue measure dx and the usual Nash inequality
f 22+n ≤ Cn ∇f n2 f 21
from (6.2.4), p. 281 (the norms being understood with respect to dx). Considering
dμ = e−W dx, choose w = eW/2 as weight function. Changing then f into f e−W/2
in the above Nash inequality, after integration by parts, it follows that for every
smooth compactly supported function f : Rn → R,
(n+2)/n 4/n
2
f dμ ≤ Cn |∇f | dμ +
2 2
f R dμ |f |w dμ
Rn Rn Rn Rn
(7.4.4)
where R = 12 W − 14 |∇W |2 . Provided then R is uniformly bounded, say by c > 0,

a weighted Nash inequality for μ with weight w = eW/2 and growth function
(r) = C(1 + r)n/(n+2) , r ∈ (0, ∞), holds, where C > 0 depends on n, Cn and
c. There is no claim here that the weight w satisfies Pt w ≤ L(t)w, which has to
be studied separately (and is in general easy to check in practice). The weight w is
never in L2 (μ), so that the latter is on the other hand of little use for the control of
the spectrum.
We do not expand more on weighted Nash inequalities. On the real line R with
(f ) = f 2 and dμα (x) = p(x)dx with p(x) = cα e−(1+x ) (defined at (7.0.1)),
2 α/2
one may obtain weighted Nash inequalities with w(x) = p(x)1/2 (1 + x 2 )−β for any
β > 0. In this case, for constants C > 0 and δ ∈ (0, 1) depending on α and β, one
may choose (r) = C(1 + r)δ , r ∈ (0, ∞). Thanks to the introduction of the weight,
the function may be chosen to be similar to the one used for the classical Nash
inequalities, even in the case α > 2 where ultracontractive bounds hold. But it also
extends to non-ultracontractive settings corresponding to 1 < α ≤ 2. When α = 1
however, it is known that the spectrum of the associated operator is not discrete (cf.
p. 189).
7.5 Weak Poincaré Inequalities
For simplicity of exposition, we deal here with a Markov Triple (E, μ, ) (Standard
would be enough) where μ is a probability measure. Similar (useful) statements may
be obtained in the sigma-finite case (in which case variances have to be replaced by
L2 (μ)-norms).
Many functional inequalities encountered throughout this book may be used to

control convergence to equilibrium. This property is related to tightness in the in-
equalities. As often used and described, tightness is a way to express how functions
f converge to a constant when E(f ) → 0. Up to now, we mainly used Poincaré
inequalities to tighten the various functional inequalities under investigation (log-
arithmic Sobolev, Sobolev, entropy-energy and so on, cf. e.g. Proposition 5.1.3,
p. 238, and Proposition 6.2.2, p. 280). The weak Poincaré inequalities developed
in this section actually appear as the minimal requirement to ensure tightness. Also,
by Proposition 4.4.2, p. 190, Poincaré inequalities require at least exponential decay
on the tail behavior of the measure μ. When dealing with sub-exponential or poly-
nomial decays, none of the inequalities investigated so far reveal any information
about convergence to equilibrium. On the other hand, ergodicity, as presented in
Definition 3.1.11, p. 135, is the minimal requirement that one may ask for in order
to obtain convergence to equilibrium. But this property is purely qualitative. Weak
Poincaré inequalities developed here are in some sense a way to quantify ergod-
icity, in the weakest possible sense. Finally, weak Poincaré inequalities essentially
aim at capturing the behavior of Poincaré inequality constants on large compact
sets (typically on large balls in Euclidean space). A Poincaré inequality holds when
those constants do not diverge at infinity, and weak Poincaré inequalities describe
the behavior of those constants when they diverge.
This section thus develops the tool of weak Poincaré inequalities, how they are
used to tighten functional inequalities, and what type of information they entail.
7.5.1 Weak Poincaré Inequalities
At a technical level, weak Poincaré inequalities control how a sequence (fk )k∈N of
(suitable) functions converge to a constant in the weakest sense (in measure) when
E(fk ) → 0. Since E(f ) = E(f − a) for any a ∈ R, the control of E(fk ) cannot say
anything about the control of this constant, justifying the investigation of a tool such
as weak Poincaré inequalities.
Before presenting the notion of weak Poincaré inequality itself, it is useful to
study, as a preliminary investigation, some aspects of convergence in measure. Re-
call that a sequence (fk )k∈N of real-valued measurable functions on (E, F, μ) con-
verges in measure (or in probability, or in L0 (μ)) to a measurable function g, if, for
ε > 0, limk→∞ μ(|fk − g| ≥ ε) = 0. Convergence in measure

any is metrizable, and
E [|f − g| ∧ 1]dμ is one example of a distance defining the L 0 (μ)-topology.
The first statement tries to characterize the oscillation of a given function under
convergence in measure. Dealing with functions which are a priori not integrable,
we work instead with medians. Given f measurable on (E, F), the set of m ∈ R
such that μ(f ≥ m) ≥ 12 and μ(f ≤ m) ≥ 12 is a non-empty bounded interval (pos-
sibly reduced to one point). For simplicity, we agree to define the median m(f ) of
f to be the middle point of this interval. On the other hand, for a given measurable
function f : E → R, set τ (f ) = (f ∧ 1) ∨ (−1), and introduce

Tμ (f ) = Varμ τ (f ) and Tμ∗ (f ) = sup Varμ τ (f − a) .
a∈R
Obviously Tμ (f ) ≤ Tμ∗ (f ) ≤ 1.
Proposition 7.5.1 If a sequence (fk )k∈N of measurable functions on (E, F) con-

verges to 0 in measure, then limk→∞ m(fk ) = 0 and limk→∞ Tμ∗ (fk ) = 0. Con-
versely, if limk→∞ m(fk ) = 0 and limk→∞ Tμ (fk ) = 0, then (fk )k∈N converges to 0
in measure.

If the sequence (fk )k∈N is uniformly integrable, m(fk ) may be replaced
by E fk dμ in the latter implication.
Recall that a sequence (fk )k∈N in L

1 (μ) (where μ is a probability measure) is
uniformly integrable if limc→∞ supk∈N {|fk |>c} |fk |dμ = 0.
Proof Assuming that fk → 0 in measure, if 0 < ε ≤ 13 , for every k large enough,

μ(|fk | ≥ ε) ≤ ε ≤ 13 and therefore |m(fk )| ≤ ε. Furthermore, since the map
r → τ (r) is 1-Lipschitz, for every a ∈ R, |τ (fk − a) − τ (−a)| ≤ ε on a set of mea-
sure larger than 1 − ε and is bounded by 2 on its complement. It easily follows that
for every k large enough, Varμ (τ (fk − a)) ≤ 5ε uniformly in a ∈ R, yielding the
first claim.
For the converse implication, under the condition
limk→∞ Varμ (τ (fk )) = 0,
by Markov’s inequality, lim
k→∞ μ(|τ (f k ) − E τ (f k )dμ| ≥ ε) = 0 for every
ε > 0. Since the sequence E τ (fk )dμ, k ∈ N, is bounded by 1, there exists a
subsequence which converges to some α ∈ R. If α = 0, then along a subse-
quence limk→∞
μ(|τ (fk ) − α| ≥ ε) = 0 contradicting limk→∞ m(fk ) = 0. There-
fore limk→∞ E τ (fk )dμ = 0, and hence τ (fk ) → 0 in measure and thus also
fk → 0. The last claim of the proposition

follows in the same way. Namely,
if along a subsequence limk→∞ E τ (fk )dμ = 2α = 0, for example α > 0, it
would follow

that limk→∞ μ(fk ≤ α) = 0, and
then by uniform integrability that
limk→∞ {fk ≤α} fk dμ = 0. Then lim infk→∞ E fk dμ ≥ α > 0 contradicting that

limk→∞ E fk dμ = 0.
Having set up the preliminaries, we introduce the notion of a weak Poincaré

inequality, denoted by W P in what follows. Recall that throughout this section
μ is a probability measure. Recall also that Tμ (f ) = Varμ (τ (f )) for a measur-
able function f : E → R. As in the preceding sections, weak Poincaré inequalities
involve an increasing and concave (not necessarily C 1 a priori) growth function
! : (0, ∞) → R+ .
Definition 7.5.2 (Weak Poincaré inequality) A Markov Triple (E, μ, ), with
μ a probability measure, is said to satisfy a weak Poincaré inequality W P (!)
with respect to a growth function ! : (0, ∞) → R+ bounded by 1 and such that
limr→0 !(r) = !(0) = 0 if, for every f ∈ D(E),

Tμ (f ) ≤ ! E(f ) .
It is important to observe that the concavity assumption on ! in this definition is

not necessary since the concave envelope of a bounded increasing function which
converges to 0 at 0 shares the same properties. The restriction ! ≤ 1 is obvious
since Tμ (f ) ≤ 1.
Weak Poincaré inequalities may be described in various equivalent forms. First
note that since E(f − a) = E(f ) for every a ∈ R, Tμ may be replaced by Tμ∗ in
Definition 7.5.2. By the definition of τ , Definition 7.5.2 is also equivalent to saying
that

Varμ (f ) ≤ ! E(f ) (7.5.1)
for every f ∈ D(E) such
that f ∞ ≤ 1. It is then also equivalent for all bounded
f ’s with either f − E f dμ∞ ≤ 1 or osc(f ) = sup f − inf f ≤ 1. By lineariza-
tion, it also amounts to the existence of a decreasing rate function s → γ (s) on R+
such that
Varμ (f ) ≤ s E(f ) + γ (s), s ∈ R+ , (7.5.2)
for these respective classes of functions. For instance, for the class of bounded func-
tions, since Varμ (f ) ≤ f 2∞ , it can be assumed that γ is furthermore bounded
by 1. The important feature is the behavior of the rate function γ as s → ∞ re-
flecting the behavior of !(r) near 0 in the weak Poincaré inequality, the fact that
limr→0 !(r) = 0 being equivalent to lims→∞ γ (s) = 0. This linearized form (7.5.2)
will turn out to be particularly useful in the next chapter, Sect. 8.4, p. 403.
Weak Poincaré inequalities quantify convergence in measure under convergence
of the Dirichlet form as illustrated by the following main claim.
Proposition 7.5.3 A Markov Triple (E, μ, ) satisfies a weak Poincaré in-

equality W P (!) for some growth function ! (bounded by 1 and such that
limr→0 !(r) = !(0) = 0) if and only if for any sequence (fk )k∈N in D(E) such
that limk→∞ E(fk ) = 0 together with limk→∞ m(fk ) = 0, fk → 0 in measure.
The same
is true if limk→∞ m(fk ) = 0 is replaced by supk∈N f ∞ ≤ 1 and
limk→∞ E fk dμ = 0.
Proof One direction is obvious thanks to Proposition 7.5.1. For the converse impli-
cation, consider

!(r) = sup Tμ (f ) ; E(f ) ≤ r , r ∈ (0, ∞).
! is clearly increasing and bounded (by 1), and the fact that limr→0 !(r) = 0 is a
direct consequence of the hypothesis. Indeed, if not, there exist ε > 0 and a sequence
(fk )k∈N in D(E) such that limk→∞ E(fk ) = 0 and Tμ (fk ) ≥ ε for every k. But now,
each fk may be translated so that m(fk ) = 0, thus contradicting the hypothesis.
It remains to change ! into

its concave envelope as mentioned above. When the
median m(fk ) is replaced by E fk dμ, the same proof holds changing ! into

!(r) = sup Varμ (f ) ; f ∞ ≤ 1, E(f ) ≤ r , r ∈ (0, ∞).
The proof is complete.
Corollary 7.5.4 A tight Nash inequality N () implies a weak Poincaré inequality.
The same holds with a tight entropy-energy inequality EE().
The Nash case is an immediate consequence of the preceding proposition. The

entropy-energy case follows from the fact that a tight EE() implies a tight N ()
(Corollary 7.4.4).
The relation with standard Poincaré inequalities is described in the following
proposition.
Proposition 7.5.5 A Markov Triple (E, μ, ) satisfies a Poincaré inequality P (C)

if and only if it satisfies a weak Poincaré inequality W P (!) with !(r) ≤ Cr, r > 0.
Proof Under a Poincaré inequality P (C), Tμ (f ) ≤ CE(f ) for every f ∈ D(E)

and therefore W P (!) holds with !(r) = min(Cr, 1), r ∈ (0, ∞). Conversely, it is
enough by homogeneity to establish the Poincaré inequality for functions bounded
by 1, so that the conclusion follows from the hypothesis !(r) ≤ Cr, r > 0.
7.5.2 Tightening with Weak Poincaré Inequalities
Weak Poincaré inequalities are aimed at tightening functional inequalities. The fol-
lowing is an illustration of the principle in the context of (linearized) Nash inequal-
ities. It may be applied similarly to most functional inequalities investigated in this
monograph.
We start with a technical proposition describing the main tightening step.
Proposition 7.5.6 Let (E, μ, ) be a Markov Triple such that, for constants
s1 , β1 > 0,
f 22 ≤ s1 E(f ) + β1 f 21 (7.5.3)
for every f ∈ D(E). Assume furthermore that for some s2 > 0 and 0 < γ2 < 1
β1 +1 ,
and any f ∈ D(E) bounded by 1,
Varμ (f ) ≤ s2 E(f ) + γ2 . (7.5.4)
Then a Poincaré inequality P (C) holds with constant

2(s1 + s2 )
C= √ .
1 − γ2 (β1 + 1)
On the basis of this result, a general tightening statement is summarized in the

next corollary.
Corollary 7.5.7 (Tightening with a weak Poincaré inequality) Assume that

(E, μ, ) satisfies a generalized Nash inequality N () or an entropy-energy in-
equality EE() for some growth function . Then, if a weak Poincaré inequality
holds, (E, μ, ) satisfies a (true) Poincaré inequality.
In particular, whenever a tight Nash inequality N (1 ) holds, or a tight entropy-
energy inequality EE(2 ), then 1 (r) may be replaced by min(1 + Cr, 1 (r)) and
2 (r) by min(Cr, 2 (r)), r ∈ (0, ∞), for some constant C > 0.
To deduce this corollary from Proposition 7.5.6, simply observe that by lin-
earization, both a generalized Nash or entropy-energy inequality imply (7.5.3),
while (7.5.4) holds under a weak Poincaré inequality since in its linearized
form (7.5.2), lims→∞ γ (s) = 0. Concerning the second part of the statement, by
Corollary 7.5.4, a weak Poincaré inequality holds under a tight Nash inequality
N(1 ) or a tight EE(2 ) inequality, and therefore also a Poincaré inequality. Then
1 may be chosen such that 1 (r) ≤ Cr + 1 (by Proposition 7.4.2) and 2 (r) ≤ Cr
(from the discussion p. 350), r > 0, for some constant C > 0.
Proof of Proposition 7.5.6 Let f ∈ D(E) be such that E f dμ = 0 and E f 2 dμ = 1.

f
Set, for R > 0, fR = R τ ( R ), which coincides with f whenever |f | ≤ R, and
dR = f − fR . Since
both fR and dR are contractions of f , E(fR ) ≤ E(f ) and
E(dR ) ≤ E(f ). Since E f 2 dμ = 1, by the Cauchy-Schwarz and Markov inequali-
ties,

|dR |dμ = |f |1{|f |≥R} dμ − Rμ(|f | ≥ R)
E E
1
≤ μ(|f | ≥ R) − Rμ(|f | ≥ R) ≤ .
4R

On the other hand,

since E f dμ = 0, we also have that

| E fR dμ| = | E dR dμ| ≤ 1 −1
4R . Applying (7.5.4) to R fR then yields

1
fR2 dμ ≤ + s2 E(f ) + γ2 R 2 ,
E 16R 2
while, applying (7.5.3) to dR produces

β1
dR2 dμ ≤ + s1 E(f ).
E 16R 2
Now fR dR = RdR so that

1= f 2 dμ = fR2 dμ + dR2 dμ + 2R dR dμ
E E E E

1
≤ fR2 dμ + dR2 dμ + .
E E 2
It follows from the preceding that
β1 + 1 1
1 ≤ (s1 + s2 )E(f ) + + γ2 R 2 +
16R 2 2
1 1
≤ (s1 + s2 )E(f ) + γ2 (β1 + 1) +
2 2
√
after optimizing in R > 0. Hence 1 −
γ2 (β1 + 1) ≤ 2(s1
+ s2 )E(f ) which amounts
to the Poincaré inequality P (C) since E f dμ = 0 and E f 2 dμ = 1. The proof is
complete.
In the following, we address the general question of how to obtain a weak

Poincaré inequality. Clearly, weak Poincaré inequalities have much to do with
Poincaré inequalities, and the results presented below rely on various tools and ideas
developed for Poincaré inequalities in Chap. 4. The first statement describes how to
obtain a weak Poincaré inequality from Poincaré inequalities on large sets.
Proposition 7.5.8 Let (E, μ, ) be a Markov Triple such that for some measurable
set K ⊂ E with μ(K) > 0, some constant C > 0 and any f ∈ D(E),
2
1
f dμ ≤ C E(f ) +
2
f dμ . (7.5.5)
K μ(K) K
Then, for any bounded function f ∈ D(E) with mean zero,

μ(K c )
f 2 dμ ≤ C E(f ) + f 2∞ .
E μ(K)
Therefore, if (7.5.5) holds for a sequence (K )∈N of measurable subsets of E such
that lim→∞ μ(Kc ) = 0, a weak Poincaré inequality holds.
Proof The proof is rather immediate. Write

f dμ =
2
f dμ +
2
f 2 dμ
E K Kc
2
1
≤ C E(f ) + f dμ + μ K c f 2∞ .
μ(K) K

On the other hand, since E f dμ = 0, | K f dμ| = | K c f dμ| ≤ f ∞ μ(K c ),

from which the claim follows. The conclusion then follows from the various char-
acterizations of weak Poincaré inequalities.
Remark 7.5.9 In many situations, a stronger inequality than (7.5.5) is considered,

namely
2
1
Varμ (f ) ≤ C (f )dμ + f dμ
K μ(K) K
for some measurable set K ⊂ E and every f ∈ D(E). In Rn or a smooth manifold,
when K is a set with compact closure and regular boundary, this corresponds to the
Poincaré inequality on K with Neumann boundary conditions. Indeed, it is not hard
to see that any smooth function on K with zero normal derivative at the boundary
may be extended to a smooth function with compact support in a neighborhood of
K, and conversely that any smooth function f defined in a neighborhood of K may
be approximated by a sequence (fk )k∈N of smooth functions
satisfying

Neumann
conditions on the boundary of K such that all the quantities f 2 dμ,

K k K (fk )dμ
and K fk dμ converge to the respective quantities for f . It is therefore often use-
ful to have such estimates on Poincaré inequalities on balls with Neumann bound-
ary conditions, for example. These may be achieved under curvature-dimension hy-
potheses on balls whenever the set K is convex.
While many probability measures do satisfy a weak Poincaré inequality W P (!),

the associated growth function ! may even be explicitly described. The idea, the
same as for Poincaré or logarithmic Sobolev inequalities, consists of a suitable adap-
tation of the Muckenhoupt criterion to control Poincaré constants on balls. Results
and classical examples will be illustrated in the next chapter in Proposition 8.4.4,
p. 409.
7.5.3 Heat Kernel Bounds and Tail Inequalities Under Weak

Poincaré Inequalities
To conclude this section, we briefly describe how weak Poincaré inequalities are
related to heat kernel bounds, convergence to equilibrium and tail estimates. The
strategy is entirely similar to the one developed under generalized or weighted Nash
inequalities in the preceding sections. The procedure applies similarly to any func-
tional inequality of the form

Q(f ) ≤ E(f )
for any f ∈ D(E) where Q is a positive functional which is closed under (Pt )t≥0 and
for which Q(Pt f ) ≤ M(t)Q(f ), t ≥ 0, provided (r) is increasing and r −1 (r)
decreasing, r > 0. According to the various descriptions of weak Poincaré in-

2
equalities, Q(f ) might be either f 2∞ , or f − fE dμ∞ or osc(f )2 for which
M(t) = 1.
The following is then obtained by slightly modifying the proof of Theorem 7.4.5.
For simplicity, the growth function ! in the weak Poincaré inequality W P (!) is
assumed here to be C 1 .
Proposition 7.5.10 Let (E, μ, ) be a Markov Triple satisfying a weak Poincaré

inequality W P (!) for some growth function !. Then, for any mean-zero bounded
function f ∈ D(E) and any t > 0,
Pt f 22 ≤ K(2t)f 2∞

∞
where K(s), s > 0, is the inverse function of U (s) ! (r) dr
r and U is the inverse
function of ! defined on [0, a] where a ∈ (0, 1]. Conversely, if for some t > 0,
Pt f 22 ≤ K(t)f 2∞ for any bounded mean-zero measurable function f : E → R,
then, for any f ∈ D(E) with f ∞ ≤ 1,
Varμ (f ) ≤ 4t E(f ) + 2K(t).
The converse part of the statement is a direct application of the

L2 (μ)-bound (4.2.3), p. 184. This proposition tells us that a weak Poincaré in-
equality W P (!) holds if and only if Varμ (Pt f ) ≤ M(t)f 2∞ for any bounded
mean-zero measurable function f : E → R and any t > 0 with some function M(t)
such that limt→∞ M(t) = 0.
Even when there is no Poincaré inequality, it is possible to control convergence
to equilibrium in L2 (μ) provided that the initial function f is bounded. In general,
this convergence will appear at a sub-exponential or polynomial rate, and this rate is
related to the tail behavior of the invariant measure μ. For example, if !(r) ≤ Cr α
for 0 < r ≤ 1 with α ∈ (0, 1), then K(t) ∼ C1 t −α/(1−α) for large t, while if
K(t) ≤ Ct −α/(1−α) for large t, a W P (!) inequality holds with !(r) ≤ C1 r α , r > 0.
Finally, weak Poincaré inequalities also carry bounds on tails of Lipschitz func-
tions following the usual procedure in this regard. The next statement is for example
established by arguments similar to the proof of Proposition 7.4.7.
Proposition 7.5.11 Let (E, μ, ) be a Markov Triple satisfying a weak Poincaré

inequality W P (!), and let f be a 1-Lipschitz function on E. Choose r0 > 0 such
that μ(f ≥ r0 ) ≤ max( 14 , supr>0 !(r)) = q0 . Then, for any r > r0 ,

μ(f ≥ r) ≤ 2F (r − r0 ) log 2
where F is the inverse function of

q0 −1/2
u
G(v) = u !−1 du, v ∈ (0, q0 ).
v/2 4
Proof Arguing as in Proposition 7.4.7, apply W P (!) to g = 1s (((f − r) ∨ 0) ∧ s),

r, s > 0. With q(u) = μ(f ≥ u), u > 0,

q(r)
q(r + s) ≤ q(r) + !
2
.
s2
Choose r large enough so that q(r) ≤ 14 and then s such that !( q(r)
s2
)≤ q(r)
4 which
may be achieved for
1/2
q(r)
s= .
!−1 ( q(r)
4 )
q(r)
Hence q(r + s) ≤ 2 and the conclusion follows as in Proposition 7.4.7.
For measures with polynomial decay, this bound is optimal up to constants.

For example, if dμ(x) = cα (1 + x 2 )−α/2 dx, α > 1 then !(r) ≤ Cr (α−1)/(α+1) ,
0 < r < 1, and for such !, we get μ(f ≥ r) ≤ Cr (1−α) as r → ∞, which is the cor-
rect order. For the family μα , α > 0, of (7.0.1) on R, !(r) ≤ Cr(− log r)2(1−α)/α ,
r > 0, from which μα (f ≥ r) ≤ Ce−cr as r → ∞ which again is of the correct
α
order of magnitude up to constants.
7.6 Further Families of Functional Inequalities
This last section briefly surveys related families of functional inequalities which
have been developed for different purposes. As for the inequalities investigated in
the previous sections, each such family has its own interest and is more or less
better suited to specific issues (convergence to equilibrium, heat kernel bounds or
tail estimates). Only a few properties and illustrations are outlined here. In par-
ticular, the guideline is the study of the family of probability measures dμα (x) =
cα e−(1+x ) dx defined in (7.0.1) on the real line for α ∈ [1, 2] which do satisfy
2 α/2
a Poincaré inequality, but satisfy a logarithmic Sobolev inequality only for α = 2

(cf. Theorem 4.5.1, p. 194, and Theorem 5.4.5, p. 256). This family has already
been examined in the context of entropy-energy, generalized Nash or weighted Nash
inequalities in Sects. 7.1 and 7.4. One feature of the various functional inequalities
presented below is that they clearly distinguish the values of α, in particular in terms
of tail behavior and convergence to equilibrium. A summary of the various proper-
ties of this family according to the parameter α will be presented in the next section.
The exposition here is rather sketchy and we refer to the appropriate bibliography
in the Notes and References for further details.
The various families of interest are presented in the usual general context of a
Markov Triple (E, μ, ) where μ is a probability measure.
7.6 Further Families of Functional Inequalities 383
7.6.1 -Entropy Inequality
The first family comprises the -entropy inequalities. Such an inequality depends
on some convex C 2 function defined on an open interval

I of (0, ∞). For every
measurable function f : E → I in L1 (μ) such that E |(f )|dμ < ∞, define the
-entropy of f as

Entμ (f ) =

(f )dμ − f dμ . (7.6.1)
E E
The usual variance Varμ is obtained with (r) = r 2 on I = R and the entropy Entμ
((5.1.1), p. 236) for (r) = r log r on I = (0, ∞).
Given then a function : I → R as above, a Markov Triple (E, μ, ) is said to
satisfy a -entropy inequality with constant C > 0 if

C
Entμ (f ) ≤

(f ) (f )dμ (7.6.2)
2 E
for every I -valued function f ∈ D(E). According to the previous examples of the
function , this definition appears as a direct extension of Poincaré and logarith-
mic Sobolev inequalities. To provide interesting consequences and applications, it
is necessary to supplement the convexity assumption of by the further hypothesis
that is C 4 and that 1 is concave on I . This is clearly the case for the examples of
(r) = r 2 on I = R and (r) = r log r on I = (0, ∞). In the following, only such
admissible functions : I → R on an interval I ⊂ R are considered.
The -entropy inequalities satisfy the standard stability properties by bounded
perturbation and by tensorization (which is less easy), analogous to the correspond-
ing ones for Poincaré and logarithmic Sobolev inequalities. They may also be char-
acterized by an exponential decay along the underlying semigroup, and hold under
curvature conditions analogous to those for Poincaré and logarithmic Sobolev in-
equalities discussed in Chaps. 4 and 5. The next statement summarizes the latter
two results.
Proposition 7.6.1 Let (E, μ, ) be a Markov Triple with Markov semigroup

(Pt )t≥0 and let be an admissible function on I ⊂ R. The -entropy inequal-
ity (7.6.2) with constant C > 0 is equivalent to the exponential decay
−t/C
μ (Pt f ) ≤ e
Ent Ent
μ (f ),

for every t ≥ 0 and every I -valued function f ∈ L1 (μ) such that E |(f )|dμ < ∞.
Furthermore, under the curvature condition CD(ρ, ∞) for some ρ > 0, the Triple
(E, μ, ) satisfies a -entropy inequality (7.6.2) with constant C = ρ1 .
The proof of the first part of the proposition is similar to that of Theorem 5.2.1,
p. 244. The second part relies on the arguments put forward in the proof Proposi-
tion 5.5.3, p. 261, via the analysis of the function (s) = Ps ((Pt−s f )), s ∈ [0, t],
t > 0, for f : E → I in Aconst
0 , and the use of the admissibility property of .
By Theorem 4.2.5, p. 183, a Poincaré inequality for (E, μ, ) is equivalent to the

exponential decay in L2 (μ) of the associated Markov semigroup (Pt )t≥0 , while a
logarithmic Sobolev inequality is equivalent, by Theorem 5.2.1, p. 244, to the expo-
nential decay of the classical entropy. The main interest of -entropy inequalities is
their role in describing an exponential decay of intermediate norms between L2 (μ)
and L1 (μ).
7.6.2 Beckner Inequality
One example of interest in the family of -entropy inequalities concerns the choice
of
rp − 1
(r) = , r ∈ (0, ∞),
p−1
which is admissible only if p ∈ (1, 2], and includes the case (r) = r log r in the
limit p = 1. For this family of functions , the -entropy inequality takes the form
(after changing f ≥ 0 into f 2/p and letting q = p2 ∈ [1, 2])
2/q
1
f dμ −
2
|f | dμ
q
≤ C E(f ) (7.6.3)
2−q E E
for every f ∈ D(E), which is known as the Beckner inequality Bq (C), q ∈ [1, 2],
with constant C > 0. According to the study of Sect. 7.4, B1 (C) corresponds to the
Poincaré inequality P (C), while (in the limit) B2 (C) amounts to the logarithmic
Sobolev inequality LS(C). The inequality Bq (C) is also part of the family (6.8.7)
of Remark 6.8.4, p. 312 (with n = ∞).
By convexity of the map r → log(f 1/r ), r ∈ (0, 1], the expression
2/q
1
f dμ −
2
|f | dμ
q
1
q − 1
2 E E
is increasing in q ∈ [1, 2]. In particular, if μ satisfies a logarithmic Sobolev in-

equality LS(C), then it satisfies a Beckner inequality Bq ( q2 C) for every q ∈ [1, 2).
Conversely, by Jensen’s inequality, a Poincaré inequality P (C) implies a Beckner
inequality Bq ( 2−qC
) for every q ∈ [1, 2). Therefore, apart from the precise values
of the constants, every member of this family of Beckner inequalities for q ∈ [1, 2)
is equivalent to a Poincaré inequality. The limiting case q = 2 corresponding to the
logarithmic Sobolev inequality plays a special role. Note however that under the cur-
vature condition CD(ρ, ∞), ρ > 0, a logarithmic Sobolev inequality LS( ρ1 ) holds
(Proposition 5.7.1, p. 268) while Proposition 7.6.1 produces a Beckner inequality
Bq ( ρ1 ) for every q ∈ [1, 2], thus better than seen as consequence of LS( ρ1 ).
7.6 Further Families of Functional Inequalities 385
7.6.3 Latała-Oleszkiewicz and Modified Logarithmic Sobolev

Inequalities
According to the preceding discussion, in terms of tail behaviour, Beckner inequal-

ities do not discriminate between members of the model family of probability mea-
sures μα of (7.0.1) on the real line for α ∈ [1, 2] (which do satisfy a Poincaré in-
equality but satisfy a logarithmic Sobolev inequality only for α = 2). Two main
options have been investigated in order to describe a family of intermediate inequal-
ities between Poincaré and logarithmic Sobolev, which are in particular well-suited
to the scale of the family μα , α ∈ [1, 2].
A first strategy consists of looking precisely at the behavior of the constants in
Beckner inequalities when the parameter q approaches 2. Say, for instance, that a
Markov Triple (E, μ, ) satisfies a Latała-Oleszkiewicz inequality LO(a, C) with
parameter a ∈ [0, 1] and constant C > 0 if for every f ∈ D(E),
2/q
1
sup f dμ −
2
|f | dμ
q
≤ C E(f ). (7.6.4)
q∈[1,2) (2 − q)
a
E E
Using as above Hölder’s inequality, it is easily checked that a Poincaré inequality

P (C) coincides with LO(0, C), that a logarithmic Sobolev inequality LS(C) im-
plies LO(1, C), while conversely LO(1, C) implies LS(2C). Moreover, the family
LO(a, C) gets stronger (with the same C) as a increases. It may be shown (see
Sect. 8.4, p. 403) that the measures μα satisfy a Latała-Oleszkiewicz inequality
LO(a, C) with parameter a ∈ [0, 1] if and only if α ≥ 2−a 2
.
A second approach amounts to keeping the (classical) entropy and modifying the
energy, yielding modified logarithmic Sobolev inequalities. For simplicity, we con-
sider probability measures μ on Rn , and say that μ satisfies a modified logarithmic
Sobolev inequality with parameter p ≥ 1 and constant C > 0 if for every smooth
function f : Rn → R,
n
∂i f
Entμ f 2 ≤ C Hp f 2 dμ (7.6.5)
Rn f
i=1
∗
where Hp (r) = r 2 if |r| ≤ 1 and Hp (r) = |r|p if |r| > 1, p1 + p1∗ = 1. Of course,
for p = 2, we are left with the usual logarithmic Sobolev inequality. The measure
μα (on R or Rn ) satisfies the modified logarithmic Sobolev inequality (7.6.5) if and
only if α ≥ p.
The main interest in both the Latała-Oleszkiewicz and modified logarithmic
Sobolev inequalities is that they are stable under products and capture the correct
tail estimates of the given measure. That is, if the Markov Triple (E, μ, ) satis-
fies either a Latała-Oleszkiewicz inequality LO(a, C) (for a ∈ [0, 1]), or a modified
logarithmic Sobolev inequality with parameter p ≥ 1 and constant C > 0, then for
every Lipschitz function f on E, and every r ≥ 0,
p
−r p /C1 f Lip
μ f≥ f dμ + r ≤ C1 e
E
where C1 > 0 only depends only on C and p (and a = 2(1 − p1 )). The modified
logarithmic Sobolev inequality implies the concentration result for all p > 1 while
the Latała-Oleszkiewicz inequalities are valid only for p ∈ [1, 2]. Again, such tail
estimates capture the correct behaviors of the model measures μα .
7.7 Summary for the Model Example μα

This brief section summarizes some of the conclusions of this and the preceding
chapters in terms of functional inequalities, heat kernel bounds and tail behaviors
for the model family (7.0.1) of probability measures
α/2
dμα (x) = cα exp − 1 + |x|2 dx, α > 0,
on the Borel sets of Rn . Recall that the density of μα is essentially e−|x| but it is
α
convenient to smooth out the potential to avoid regularity issues. As for the Markov
Triple, the carré du champ operator is the usual one (f ) = |∇f |2 on smooth func-
tions f on Rn . The associated Markov semigroup is denoted by (Pt )t≥0 .
According to (7.4.4), for all α > 0, μα satisfies a weighted Nash inequality
N() with growth function (r) = r n/(n+2) , r ∈ (0, ∞), and weight w(x) =
e−(1+|x| ) /2 , x ∈ Rn . In particular, following Proposition 7.4.9, the semigroup
2 α/2
(Pt )t≥0 has density kernels pt (x, y), t > 0, (x, y) ∈ Rn × Rn , with respect to μα
satisfying
pt (x, y) ≤ C(t)w(x)w(y)
with C(t) > 0. This inequality however does not yield any significant information
about the spectrum since the weight is not in L2 (μα ).
We next distinguish the various ranges of α > 0, and for simplicity restrict to
the one-dimensional case. Growth and rate functions , !, γ , β of the various
inequalities are only given for small or large values of the variables. Constants C > 0
may change from line to line.
• α ∈ (0, 1). Only a weak Poincaré inequality is satisfied here with growth func-
tion !(r) = Cr(− log r)2(1−α)/α for small values of r > 0. Equivalently, in
the linearized form (7.5.2), γ (s) = exp(−Cs α/(2−2α) ) for large values of s.
Poincaré, logarithmic Sobolev, Latała-Oleszkiewicz and modified logarithmic
Sobolev inequalities are not satisfied. Moreover the essential spectrum is not
empty (cf. Corollary 4.10.9, p. 229).
• α = 1. The measure μ1 behaves as the symmetric exponential measure.
A Poincaré inequality P (C) with constant C > 0 holds (cf. (4.4.3), p. 189),
as well as a Latała-Oleszkiewicz inequality with a = 0 or a modified logarith-
mic Sobolev inequalities with p = 1. A weak Poincaré inequality holds with
!(r) = Cr, r ∈ (0, ∞), or γ (s) = 0 for every s ≥ C (with C the Poincaré con-
stant). The measure μ1 does not satisfy any logarithmic Sobolev inequalities (see
for example Theorem 5.4.5, p 256). Even if there is a spectral gap, the essential
spectrum is not empty (cf. p. 229).
• α ∈ (1, 2). A Poincaré inequality P (C) holds (see Theorem 4.5.1, p. 194), and
then also a weak Poincaré inequality with !(r) = Cr, r ∈ (0, ∞). Both the
Latała-Oleszkiewicz inequality for any 0 ≤ a ≤ 2(1 − α1 ), and the modified log-
arithmic Sobolev inequality for any 1 ≤ p ≤ α, hold. On the other hand, there is
no logarithmic Sobolev inequality or entropy-energy inequality. However, a gen-
eralized Nash inequality N () holds with (r) = Cr(log r)−2(α−1)/α for large
r > 0, as well as, for any κ ∈ (0, 1), a weighted Nash inequality W N () with
(r) = C(1 + r)κ , r ∈ (0, ∞), and weight
−σ −(1+x 2 )α/2 /2
w(x) = 1 + x 2 e , x ∈ R,
where σ > 0. For σ large enough, w ∈ L2 (μα ) so that the operator Pt is Hilbert-
Schmidt for any t > 0 (while not ultracontractive). In particular, the essential
spectrum is empty (cf. the end of Sect. 7.4.3).
• α = 2. The measure μ2 behaves as the Gaussian measure and in particular satis-
fies a Poincaré and a logarithmic Sobolev inequality. All the properties of the case
α ∈ (1, 2) extend to α = 2, but due to the additional logarithmic Sobolev inequal-
ity, the semigroup (Pt )t≥0 is hypercontractive (while still not ultracontractive),
cf. Chap. 5.
• α > 2. Both the Poincaré and the logarithmic Sobolev inequalities hold. As
for α ∈ (1, 2], a weak Poincaré inequality holds with !(r) = Cr, r ∈ (0, ∞),
where C > 0 is the Poincaré constant. The Latała-Oleszkiewicz inequality for
any a ∈ [0, 1] and the modified logarithmic Sobolev inequality for any p ∈ [1, α]
hold. An entropy-energy inequality EE() holds with (r) = C(1 + r α/(2α−2) ),
r ∈ (0, ∞), (cf. Sect. 7.3) as well as a generalized Nash inequality N () with
(r) = Cr(log r)2(1−α)/α for large r > 0 (cf. p. 366). (Weighted Nash inequal-
ities may be established for suitable weights but are not useful as compared to
the generalized Nash inequality.) The semigroup (Pt )t≥0 is ultracontractive, the
density kernels are uniformly bounded for every t > 0, and hence the spectrum is
discrete.

This chapter provides an introduction to and a brief exposition of some of the de-
velopments concerning functional inequalities in the framework of this monograph
over the last 20 years. It is by no means a complete account of these fruitful activi-
ties.
Generalized functional inequalities in the form of entropy-energy or Nash in-
equalities emerged in various places in the literature. Among early sources are the
book [144] by E. Davies and the lecture notes [26]. In particular, E. Davies de-
scribes in [144] the method based on families of logarithmic Sobolev inequalities
(entropy-energy inequalities) to obtain on and off-diagonal heat kernel bounds, later
refined in [26], on the basis of Gross’ equivalence between logarithmic Sobolev in-
equality and hypercontractivity. The results of Sects. 7.1 and 7.2 along these lines
mainly follow [26]. The optimal Euclidean bounds under the sharp Euclidean log-
arithmic Sobolev inequality presented in Corollary 7.1.6 and Proposition 7.2.4 are
taken from [34]. The optimal heat kernel bounds (7.1.8) are part of a much deeper
investigation of Gaussian kernels by E. Lieb [287]. More on heat kernel bounds in
manifolds and metric spaces may be found in [217].
The examples of Sect. 7.3 are taken from [265] (where a more probabilistic view-
point is emphasized) and [34]. Later and deeper refinements along these lines in-
clude in particular [46].
The ultracontractive bounds from the generalized Nash inequality (Theo-
rem 7.4.5 in Sect. 7.4) are due to M. Tomisaki [411] and T. Coulhon [139]. The
generalized Nash inequalities have also been considered in the equivalent lin-
earized form (7.4.1) called super Poincaré inequalities, extensively developed by
F.-Y. Wang (cf. [431] for a complete account as well as further bibliographical
information in the context of probability theory). They are presented here in the
equivalent language of generalized Nash inequalities which offer a somewhat more
flexible treatment. The results of Sect. 7.4.1, in particular tail estimates, are mainly
suitable translations of Wang’s main contributions in this context. Weighted Nash
inequalities were investigated in [430] and more recently in [32]. The results of
Sect. 7.4.3 are essentially taken from these references. Other types of generalized
Nash inequalities have been considered in the literature, in particular for infinite-
dimensional models [63, 254, 445].
Weak Poincaré inequalities, and variations in the form of positivity improving
properties, as presented in Sect. 7.5 were first used by S. Aida [4] and P. Math-
ieu [299] as a tool to tighten defective logarithmic Sobolev inequalities. The argu-
ment was further refined in for instance [361], and then extensively developed in the
monograph [431] to which we refer for further information and complete references.
Section 7.5 tries to offer a synthesized treatment to tighten arbitrary functional in-
equalities. The topic of Remark 7.5.9 has been studied extensively in the literature
(cf. e.g. [431] and the references therein).
Section 7.6 collects various forms of functional inequalities which have been
studied in recent years. Each form has its own advantages and privileged applica-
tions. The -entropic inequalities were first considered in [36], and further devel-
oped in [84, 120, 238]. The Beckner inequality appeared in [55] for the Gaussian
measure (cf. also in a manifold setting [66]) and the Latała-Oleszkiewicz inequality
was introduced in [272]. Modified logarithmic Sobolev inequalities were considered
in, among other references, [79] and [441], and later extended in [48, 199–201, 261].
See also [80] for further developments in connection with isoperimetric bounds. We
refer to these references for the corresponding tensorization and tail properties, in
particular with the measures μα as model examples.
To conclude, and in addition to the families of Sect. 7.6, we mention here
the further family of F -Sobolev inequalities. It is established in [46] that for
α ∈ (1, 2), the probability measure μα on the line satisfies

f log 1 + f dμα − f dμα log 1 + f dμα ≤ C f dμα
2 β 2 2 β 2 2
R R R R
for every smooth f : R → R, where β = 2(1 − α1 ). The suitable general form for a
Markov Triple (E, μ, ) appears as

f 2 F f 2 dμ ≤ C E(f ) + D, (7.8.1)
E

for every f ∈ D(E) with E f 2 dμ = 1, where F : R+ → R is an increasing func-

tion satisfying F (1) = 0. Such inequalities entail useful regularity properties. For
example, the Markov semigroup (Pt )t≥0 associated with the model with invari-
ant measure μα is not hypercontractive for α ∈ [1, 2). Nevertheless, the latter
F -Sobolev inequality may be shown to imply intermediate hypercontractivity prop-
erties. F -Sobolev inequalities as in (7.8.1) go back to J. Rosen [365] and have
recently been investigated in particular in [46, 436] in the analysis of hypercon-
tractivity in intermediate Orlicz spaces. See also [360] for related Orlicz-Sobolev
inequalities and sub-Gaussian bounds.
Chapter 8
Capacity and Isoperimetric-Type Inequalities
The capacity of a set is a way to measure its size from the point of view of potential
theory. The theme of this chapter is inequalities comparing measure and capacity
uniformly over a given class of sets as equivalent forms of functional inequalities.
The passage from sets to functions is usually performed through the use of level sets
of functions and co-area formulas. Inequalities on sets often offer more flexibility
and a more transparent description of the hierarchy between functional inequalities.
Moreover, measure-capacity inequalities allow for the description of more general
forms of functional inequalities, dealing with Orlicz norms or with different mea-
sures on Lp -spaces and Dirichlet forms, and it is in these contexts that they are
most naturally studied. Furthermore, measure-capacity inequalities produce criteria
to satisfy a functional inequality much in the spirit of the Muckenhoupt character-
izations in dimension one (cf. Sect. 4.5.1, p. 194, and Sect. 5.4, p. 252), although
these criteria might appear less useful for exhibiting examples in higher dimension.
Capacities, and measure-capacity inequalities, may be investigated (and have
been considered in the literature) in rather large generality, but we only concen-
trate here on the so-called 2-capacities which capture the relevant Dirichlet form
information on sets. A second form of interest are the 1-capacities which are related
to boundary or surface measures. Measure-capacity inequalities in this case are then
isoperimetric-type inequalities comparing the measure of a set with the measure
of its boundary. Indeed, the first part of this chapter suitably adapts to the Markov
Triple framework some of the basic tools and results of the theory of capacities,
translating functional inequalities as measure-capacity inequalities. The second part
of the chapter is devoted to the particular case of Gaussian isoperimetric-type in-
equalities as an illustration of the power of heat flow techniques.
The context of this chapter is the traditional one of a (Standard or Full) Markov
Triple (E, μ, ) with Dirichlet form E, generator L and semigroup P = (Pt )t≥0 , as
summarized in Sect. 3.4, p. 168. Full Markov Triples are actually only used in the
study of the local heat kernel measures in the second part of this chapter, starting
with Sect. 8.5, where curvature-dimension inequalities come into play. To translate,
for comparison, the functional isoperimetric-type inequalities into more (classical)

DOI 10.1007/978-3-319-00227-9_8,
392 8 Capacity and Isoperimetric-Type Inequalities
geometric statements, we will sometimes assume in addition that the intrinsic dis-
tance d defines a Polish topology on (E, μ, ).
As announced, the first part of this chapter mainly revisits from the viewpoint
of sets some of the most important functional inequalities studied in the former
chapters. The first section introduces the basic notions associated with capacities and
the main technical tool to transfer (and back) functional inequalities into measure-
capacity inequalities. This tool may be seen as the essence of the slicing method
developed in previous chapters and is related to the famous co-area formulas in
geometric measure theory. In Sect. 8.2, Sobolev-type inequalities are described via
measure-capacity inequalities by a simple Orlicz space duality argument. Poincaré
and logarithmic Sobolev inequalities are handled in Sect. 8.3, and Nash and weak
Poincaré inequalities are examined in Sect. 8.4.
Replacing the 2-capacity by the boundary measure, measure-capacity inequali-
ties turn into isoperimetric-type inequalities. In this direction, Sect. 8.5 investigates,
with the heat kernel tools under a curvature condition, measure-capacity inequal-
ities of isoperimetric-type, leading in particular to the Gaussian isoperimetric in-
equality as well as to comparison results under the curvature condition CD(ρ, ∞).
On the basis of these developments, Sect. 8.6 revisits the infinite-dimensional Har-
nack inequalities of Chap. 5, Sect. 5.6, p. 265, by providing a new approach via
a reverse isoperimetric-type inequality along the semigroup. Combining the heat
kernel isoperimetric inequality with its reverse form yields furthermore an isoperi-
metric-type Harnack inequality. Finally, Sect. 8.7 addresses the relationships be-
tween Poincaré and logarithmic Sobolev inequalities and the resulting concentra-
tion properties raised in Chaps. 4 and 5. While the latter are not enough in general
to entail Poincaré and logarithmic Sobolev inequalities, they actually do under pos-
itive curvature bounds. More precise comparisons of the isoperimetric functions are
moreover available in this context.
8.1 Capacity Inequalities and Co-area Formulas
This section introduces the notions of capacities and of measure-capacity inequali-

ties, and the technical tool of co-area formulas for the transfer from set inequalities
to function inequalities.
8.1.1 Measure-Capacity Inequalities

For a Markov Triple (E, μ, ) with associated Dirichlet form E(f ) = E (f )dμ,
f ∈ D(E), we first introduce the notion of the capacity of a measurable subset of E.
Capacities related to Lp (μ)-norms of the gradient for any p ≥ 1 are defined in the
same way, but as announced, only the value p = 2 is considered here.
8.1 Capacity Inequalities and Co-area Formulas 393
Given a measurable subset A ⊂ E, set

Capμ (A) = inf (f )dμ = inf E(f ) (8.1.1)
E
where the infimum is taken over all functions f in D(E) such that 1A ≤ f ≤ 1 (the
latter expression is +∞ if the infimum is taken over the empty set. The quantity
Capμ (A) is usually referred to as the 2-capacity of A relative to the measure μ and
the Dirichlet form E or carré du champ operator .
For measurable subsets A ⊂ B ⊂ E, the capacity Capμ (A, B) is defined by re-
stricting E to B, and is referred to as the 2-capacity of A with respect to B relative
to μ and E or . In particular Capμ (A, E) = Capμ (A). From the very definition,
Capμ (A, B) is increasing in A and decreasing in B.
If the constant function 1 belongs to D(E), as is the case for example when the
measure μ is finite, then Capμ (A) = 0 for every measurable set A ⊂ E. In order
to overcome this difficulty, introduce a modified capacity defined on all measurable
sets A with μ(A) ≤ 12 by
" #
∗ 1
Capμ (A) = inf Capμ (A, B) ; A ⊂ B, μ(B) ≤ . (8.1.2)
2
For any A such that μ(A) ≤ 12 , Capμ (A) ≤ Cap∗μ (A).

Usually, capacities are defined on topological spaces, and first for open sets and
continuous functions. The preceding definition has the advantage of being consid-
ered in the context of (Standard) Markov Triples, but the price to pay is that such
capacities might be quite hard to compute at this level. This is the case even in Rn
for the Lebesgue measure and the Euclidean carré du champ operator (beyond the
case where A and B are two concentric balls). Thus, in order to get some feeling
about the definition of capacity, it might be useful to illustrate it on the example
of the real line E = R with a measure μ admitting a strictly positive and smooth
density p with respect to the Lebesgue measure and with carré du champ operator
(f ) = f 2 on smooth functions f . Then, for every measurable set A ⊂ [b, ∞),
b ∈ R, such that essinf(A) > b,
−1
essinf(A) dx
Capμ A, [b, ∞) = . (8.1.3)
b p(x)
The critical point in this identity is the essential infimum which requires us to
suitably approximate functions in the Dirichlet
domain by smooth functions. To
a dx
check (8.1.3), let a = essinf(A) > b and C = b p(x) . First, the estimate is at-

y dx
tained for the choice of f such that f (y) = 0 for y ≤ b, f (y) = C −1 b p(x) if
y ∈ [b, a] and f (y) = 1 for y ≥ a (this function is not in the dense algebra A0 of
smooth and compactly supported functions but may easily be replaced by a suitable
smooth approximation). It remains to establish the lower bound. The difficulty, as
announced, is to replace f ∈ D(E) so that 1A ≤ f ≤ 1B by some smooth (or at
least continuous) function with the same property. To this end, observe first that if
(fk )k∈N is a sequence of smooth compactly supported functions converging to f

in D(E), then (fk )k∈N converges in L2 (μ) to some function g. Then, the function

x
fˆ(x) = b g(u)du (with fˆ = 0 on {x < b}), which is the scalar product in L2 (μ) of
g and p −1 1[b,x] , is continuous and coincides with f almost everywhere on R. More-

over E(f ) = R g 2 dμ. By continuity, fˆ(a) = 1 since f = 1 almost everywhere on

x∧a
A. Now, f¯(x) = b g(t)dt (again with f¯ = 0 on {x < b}) clearly also satisfies

a
1A ≤ f¯ ≤ 1B and b g 2 dμ = E(f¯) ≤ E(f ). By the Cauchy-Schwarz inequality,
a a 1/2 a 1/2
dx
1= g(x)dx ≤ 2
g dμ .
b b b p(x)
Therefore Cap(A, [b, ∞)) ≥ C −1 which completes the claim.

Even if no topology is used to define capacities, concrete estimates for a given
set can thus only be reasonably achieved with topological arguments. In particular,
it is not clear in the Markov Triple framework that the infimum in the definition of
capacities may be restricted to functions in A0 . Fortunately, capacities will only be
used as an abstract tool below, without ever making use of a single explicit estimate.
Having defined capacities, we may introduce the notion of a measure-capacity
inequality, or more simply a capacity inequality, with respect to some increasing
growth function : R+ → R+ , in this context.
Definition 8.1.1 (Measure-capacity inequality) A Markov Triple (E, μ, ) is said

to satisfy a measure-capacity inequality with growth function : R+ → R+ if for
every measurable set A in E such that μ(A) < ∞,

μ(A) ≤ Capμ (A). (8.1.4)
In the finite measure case,

μ(A) ≤ Cap∗μ (A) (8.1.5)
for all A with μ(A) ≤ 12 . (In particular, it is enough in this case to know on [0, 12 ].)
One flexibility of measure-capacity inequalities is that one may consider entirely

similar inequalities where the capacity of the set is evaluated under another measure
ν on (E, F). If ν has infinite mass, the inequality turns into

ν(A) ≤ Capμ (A)
for every measurable set A in E (such that ν(A) < ∞). If ν is finite and normalized
into a probability measure, set
" #
∗ 1
Capμ,ν (A) = inf Capμ (A, B) ; A ⊂ B, ν(B) ≤
2
8.1 Capacity Inequalities and Co-area Formulas 395
and the second part of the definition then takes the form

ν(A) ≤ Cap∗μ,ν (A)
for ν(A) ≤ 12 . There are several benefits of such generalizations in applications.

Note that when μ is absolutely continuous with respect to ν, this extension amounts
to setting inequalities for ν with a conformal change of metric (cf. Sect. 6.9, p. 313).
Throughout this section, we deal for simplicity of exposition with the case of a single
measure μ, but repeatedly point out the possible extensions to a pair of measures.
8.1.2 Co-area Formulas
The purpose of this chapter is to relate Definition 8.1.1 to the functional inequal-
ities studied in the previous chapters. When μ has infinite mass, Sect. 8.2 studies
how (8.1.4) of Definition 8.1.1 is an alternative description of a Sobolev inequality
(for a suitable choice of the growth function ). When μ is a probability mea-
sure, (8.1.5) will be related, in Sects. 8.3 and 8.4, according again to the growth
of , to Poincaré, logarithmic Sobolev and others classical inequalities (such as
generalized Nash or weak Poincaré inequalities).
The key to this transfer is a suitable description of the energy of a function in
terms of capacities of its level sets, through versions of the so-called co-area for-
mula for 2-capacities, and achieved at the technical level by the slicing method. The
standard form of the co-area formula will be presented below in Theorem 8.5.1, but
a first formula at the level of 2-capacities is stated next.
Theorem 8.1.2 (Co-area formula I) For every function f ∈ D(E),

∞
Capμ (Nr ) rdr ≤ 6 E(f ) = 6 (f )dμ (8.1.6)
0 E
where Nr = {x ∈ E, |f (x)| > r}, r > 0.
Proof We use the classical slicing principle on the basis of Proposition 3.1.17,
p. 137. Letting f ∈ D(E), since the sets Nr are decreasing in r,
∞ 2k+1
2 Capμ (Nr ) rdr = Capμ (Nr )d r 2
0 k
k∈Z 2

≤ 22k+2 − 22k Capμ (N2k ) (8.1.7)
k∈Z

=3 22k Capμ (N2k ).
k∈Z
For every k ∈ Z, let fk = 1

2k
[(|f | − 2k )+ ∧ 2k ]. By construction,
1N2k < fk−1 ≤ 1N2k−1
and therefore, for every k,

Capμ (N2k ) ≤ (fk−1 )dμ ≤ 2−2k+2 (f )dμ. (8.1.8)
E N2k−1 \N2k
By means of Proposition 3.1.17, p. 137, combining (8.1.7) and (8.1.8) yields

∞
2 Capμ (Nr )r dr ≤ 12 (f )dμ
0 E
which is the claim.
The optimal constant in (8.1.6) is known to be 2 in the classical case of Rn (or

an open set in Rn ) with the standard gradient and a measure absolutely continuous
with respect to the Lebesgue measure, with however a different argument.
Theorem 8.1.2 can easily be generalized to functions with support B ⊂ E by
replacing Capμ (Nr ) by Capμ (Nr , B). In particular, if the support of f ∈ D(E) is of
μ-measure less than 12 , then
∞
Cap∗μ (Nr ) rdr ≤ 6 E(f ).
0
8.2 Capacity and Sobolev Inequalities

This section revisits some of the Sobolev inequalities of Chaps. 6 and 7 in terms of
measure-capacity inequalities. As announced in the introduction, one of the advan-
tages of capacities is their flexibility in describing functional inequalities directly
on subsets of the state space. In particular, it is possible to deal with more general
(Orlicz) spaces and with different measures on subsets and capacities respectively.
Dealing with Orlicz spaces, for π : R+ → R+ , assumed (for simplicity) to be
strictly increasing and continuous, such that π(0) = 0, set
r
(r) = π(u)du, r ∈ R+ .
0
Setting υ = π −1 to be the inverse of π , define

r
ϒ(r) = υ(v)dv, r ∈ R+ .
0
The two functions and ϒ are convex increasing on R+ and dual with respect to
the Legendre-Fenchel transform (∗∗ (r) = ϒ ∗ (r) = sups∈R+ [rs − (s)], r ∈ R+ ),
8.2 Capacity and Sobolev Inequalities 397
and form a pair of Orlicz functions. If μ is a measure on (E, F), the Orlicz space
L (μ) is defined as the space of measurable functions f : E → R such that
" #
f = sup |f |g dμ ; g ≥ 0, ϒ(g)dμ ≤ 1 < ∞. (8.2.1)
E E
It is easily checked, using convexity and Jensen’s inequality, that for every measur-
able subset A ⊂ E with μ(A) < ∞,

1
1A = μ(A) ϒ −1 . (8.2.2)
μ(A)
∗
As a main example, if 1 < p < ∞ and 1
p + p1∗ = 1, for ϒ(r) = r p , then (r) =
(p−1)p−1
pp r p , r ∈ R+ , L (μ) is the usual Lebesgue space Lp (μ) and · = · p .
(The results below for the case p = 1 may simply be obtained by taking the limit
p → 1.) Observe also that when ϒ(r) = er − 1, r ∈ R+ , comparison with the vari-
ational formula for entropy (5.1.3), p. 236, shows that for a probability measure μ
and f ≥ 0, Entμ (f ) ≤ f , although these quantities are not equal in general.
The following proposition is a model example of the equivalence between capac-
ity and functional inequalities.
Proposition 8.2.1 Let (E, μ, ) be a Markov Triple and (, ϒ) be a pair of dual
Orlicz functions. The following assertions are equivalent.
(i) There is a constant C1 > 0 such that for all measurable subsets A ⊂ E with
μ(A) < ∞,

−1 1
μ(A) ϒ ≤ C1 Capμ (A). (8.2.3)
μ(A)
(ii) There is a constant C2 > 0 such that for all functions f ∈ D(E),
2
f ≤ C2 E(f ). (8.2.4)

Moreover the constants C1 and C2 satisfy C1 ≤ C2 ≤ 12C1 .

The equivalence still holds whenever (r) = r, r ∈ R+ , in which case
ϒ −1 (r) = 1 and f 2 is replaced by f 22 .
As announced before, this statement, as well as most similar statements through-

out this chapter, may be extended to a pair of measures (μ, ν) replacing μ on the
left-hand side by some other measure ν on (E, F). The modifications are minor and
left to the reader.
Proposition 8.2.1 takes a particularly simple form in Lp (μ)-spaces, 1 < p < ∞,
as the equivalence between
μ(A)1/p ≤ C1 Capμ (A)

for all measurable subsets A ⊂ E and

2
f ≤ C2 E(f )
p
for all functions f ∈ D(E), with moreover C1 ≤ C2 ≤ 12C1 . The case p = 1 is

included in the limit p → 1, p > 1. Thus, in particular, the standard Sobolev in-
equality (6.1.1), p. 278, in Euclidean space

f 2p ≤ Cn |∇f |2 dx
Rn
for all functions f : Rn → R in the associated Dirichlet domain and with p = 2n

n−2 ,
n > 2, is translated equivalently as a capacity inequality,
voln (A)(n−2)/n ≤ Cn Capdx (A)
for all bounded measurable subsets A ⊂ Rn . While not necessarily of practical value
in order to reach the standard Sobolev inequality, this equivalence produces a some-
what different view of its meaning.
Proof of Proposition 8.2.1 Start with the implication from (i) to (ii). For any func-
tion f ∈ D(E), setting as above Nr = {|f | > r}, r > 0,
∞
2
f = sup f 2 g dμ = 2 sup g dμ rdr

E 0 Nr
∞
≤2 sup g dμ rdr,
0 Nr

the supremum is taken over all positive measurable functions g : E → R such
where
that E ϒ(g)dμ ≤ 1. Hence,

2 ∞
f ≤ 2 1Nr rdr

0
and therefore by (8.2.2),

2 ∞ 1
f ≤ 2 μ(Nr ) ϒ −1 rdr.
μ(Nr )
0
The conclusion then follows from (8.2.3) and Theorem 8.1.2 with C2 = 12C1 . Turn-
ing to the converse implication, given f ∈ D(E) such that 1A ≤ f ≤ 1 for A ⊂ E,
2
f ≥ 1A

from which (8.2.3) follows with C1 = C2 by the definition of the capacity

and (8.2.2). Proposition 8.2.1 is proved.
8.3 Capacity and Poincaré and Logarithmic Sobolev Inequalities 399
Remark 8.2.2 Replacing E by some measurable subset B, Proposition 8.2.1 can be

generalized for functions with a fixed support. Indeed, given a measurable B ⊂ E,
there is an equivalence between the existence of a constant C1 > 0 such that for all
measurable subsets A ⊂ B with μ(A) < ∞,

1
μ(A) ϒ −1 ≤ C1 Capμ (A, B)
μ(A)
and the existence of a constant C2 > 0 such that for all functions f ∈ D(E) with
support in B,
2
f ≤ C2 E(f ). (8.2.5)

The optimal constants C1 and C2 similarly satisfy C1 ≤ C2 ≤ 12 C1 .
In the context of smooth (weighted) Riemannian manifolds, when (r) = r and
B = O is an open set, (8.2.5) turns into

f 2 ≤ C2 E(f ),
O
for all functions f ∈ D(E) with compact support in O. The optimal constant C2 > 0
describes a lower bound of the spectrum of the restriction of the operator L on
O with Dirichlet boundary conditions in the spirit of the Faber-Krahn inequalities
(Remark 6.2.4, p. 284). More precisely, in a classical formulation,
1
Cap# (O) ≤ λ(O) ≤ Cap# (O)
12
Cap (A,O )
where Cap# (O) = infA⊂O μ(A) μ
and λ(O) = inf
Ef(f2 dμ
)
where the infimum
O
runs over all non-zero functions f ∈ D(E) with compact support in O.
8.3 Capacity and Poincaré and Logarithmic Sobolev Inequalities

In the spirit of the previous equivalence between Sobolev-type inequalities and ca-
pacity inequalities, this section addresses similarly the capacity inequalities describ-
ing Poincaré and logarithmic Sobolev inequalities. The main difference between the
preceding section on Sobolev inequalities is that both variance and entropy, which
define Poincaré and logarithmic Sobolev inequalities, are not exactly Orlicz norms
and must be compared with such norms for suitable Orlicz functions. Below, μ is a
probability measure on (E, F).
8.3.1 Capacity and Poincaré Inequalities
The first statement characterizes Poincaré inequalities in terms of capacities. Recall

from Definition 4.2.1, p. 181, that a Markov Triple (E, μ, ) satisfies a Poincaré
inequality P (C) for some constant C > 0 if

Varμ (f ) ≤ C E(f ) = (f )dμ
E
for all f ∈ D(E).
Proposition 8.3.1 Let (E, μ, ) be a Markov Triple. Assume that for some con-
stant C1 > 0,
μ(A) ≤ C1 Cap∗μ (A) (8.3.1)
for all subsets A ⊂ E such that μ(A) ≤ 12 . Then μ satisfies a Poincaré inequality
P (C2 ) with C2 = 12C1 . Conversely, under a Poincaré inequality P (C2 ), (8.3.1)
holds with C1 = 2C2 .
Proof Let f ∈ D(E) and denote by m a median of f for μ (cf. p. 374). Setting
F+ = (f − m)+ and F− = (f − m)− ,

Varμ (f ) ≤ (f − m)2 dμ = F+2 dμ + F−2 dμ.
E E E
Denoting by B+ the support of F+ , (8.3.1) applied to B+ (since μ(B+ ) ≤ 12 ) to-

gether with Proposition 8.2.1 and Remark 8.2.2 yield that

F+ dμ ≤ 12 C1 (F+ )dμ.
2
E E
With the same inequality for F− , the announced Poincaré inequality P (C2 ) holds
with C2 = 12C1 , using the fact, from (3.1.17), p. 129, that

(F+ )dμ + (F− )dμ ≤ E(f ). (8.3.2)
E E
Conversely, if μ satisfies a Poincaré inequality P (C2 ), let B ⊂ E be such that

μ(B) ≤ 12 and A ⊂ B. If f ∈ D(E) satisfies 1A ≤ f ≤ 1B , the support of f is
contained in B and thus, by the Cauchy-Schwarz inequality,
2
f dμ ≤ μ(B) f 2 dμ.
E E
Therefore

1
Varμ (f ) ≥ 1 − μ(B) f 2 dμ ≥ μ(A)
E 2
and, under the Poincaré inequality P (C2 ), 2C2 E(f ) ≥ μ(A). Taking the infimum
over all f ’s as above yields the announced claim. Proposition 8.3.1 is established.
8.3 Capacity and Poincaré and Logarithmic Sobolev Inequalities 401
As initiated in the previous section, we may deal as easily with generalized

Poincaré inequalities involving different measures for variance and energy of the
type, for ν another probability measure on (E, F),

Varν (f ) ≤ C E(f ) = C (f )dμ
E
for some C > 0 and every f in the Dirichlet domain D(E) of the Markov Triple
(E, μ, ). This inequality is then characterized via Proposition 8.3.1 by
ν(A) ≤ C1 Cap∗μ,ν (A)
for all sets A ⊂ E such that ν(A) ≤ 12 . In particular, μ need not be finite in this
formulation.
8.3.2 Capacity and Logarithmic Sobolev Inequalities
We turn to the corresponding statement for logarithmic Sobolev inequalities which

is slightly more delicate, in particular due to the fact that the Orlicz norm related
to the exponential function is not exactly the entropy. Recall from Definition 5.1.1,
p. 236, that a Markov Triple (E, μ, ) satisfies a logarithmic Sobolev inequality
LS(C) for some constant C > 0 if

2
Entμ f ≤ 2C E(f ) = (f )dμ
E
for all f ∈ D(E).
Proposition 8.3.2 Let (E, μ, ) be a Markov Triple. Assume that for some con-
stant C1 > 0,

e2
μ(A) log 1 + ≤ C1 Cap∗μ (A) (8.3.3)
μ(A)
for all subsets A ⊂ E such that μ(A) ≤ 12 . Then μ satisfies a logarithmic Sobolev
inequality LS(C2 ) with C2 = 12C1 . Conversely, under a logarithmic Sobolev in-
equality LS(C2 ), (8.3.3) holds with C1 = 8C2 .
Again, there is a version of this statement for a pair of measures (μ, ν).
Proof The scheme of proof is rather similar to that of Proposition 8.3.1. One starting
point is Lemma 5.1.4, p. 239, which indicates that for all a ∈ R,

2
Entμ f ≤ Entμ (f − a) + 2 (f − a)2 dμ.
2
E
In particular for a = m a median of f with respect to μ, the variational formula for

entropy (5.1.3), p. 236, yields that

Entμ f 2 ≤ sup (f − m)2 g dμ
E

g dμ ≤ e2 .
where the supremum runs over all measurable functions g such that Ee

Since E (f − m)2 gdμ ≤ {g≥0} (f − m)2 gdμ, setting h = g 1{g≥0} ,

Entμ f 2 ≤ sup (f − m)2 h dμ (8.3.4)
E
where the supremum

now runs over the set H of all measurable positive functions
h such that E eh dμ ≤ e2 + 1.
Recall F+ = (f − m)+ and F− = (f − m)− from the proof of Proposition 8.3.1
so that

Entμ f 2 ≤ sup F+2 h dμ + sup F−2 h dμ, (8.3.5)
h∈H+ B+ h∈H− B−
where B+ (respectively B− ) is the support of F+ (respectively F− ) and H

+ (respec-
tively H− ) is the set of all measurable positive functions h such that B+ eh dμ ≤

e2 + 1 (respectively B− eh dμ ≤ e2 + 1). A direct computation indicates that for any

measurable set A ⊂ B+

e2
sup h dμ = μ(A) log 1 +
h∈H+ A μ(A)
2
(choose h = log(1 + μ(A)
e
)1A ). Combining with (8.3.3), for all A ⊂ B+ and h ∈ H+ ,
A hdμ ≤ C1 Capμ (A, B+ ), from which, by Proposition 8.2.1 and Remark 8.2.2,

F+2 h dμ ≤ 12 C1 (F+ )dμ.
B+ B+
Optimize then over h ∈ H+ and add the same bound for F− to reach LS(12C1 ) due
to (8.3.5) and (8.3.2).
Turning to the converse, let B ⊂ E such that μ(B) ≤ 12 , A ⊂ B and f ∈ D(E)
such that 1B ≥ f ≥ 1A . Then (5.1.3), p. 236, implies that

1 − μ(B)
Entμ f 2 ≥ sup g dμ = μ(A) log 1 +
A 2μ(A)

1
≥ μ(A) log 1 +
2μ(A)
8.4 Capacity and Further Functional Inequalities 403
where

the supremum may be taken over all positive measurable functions g such
that B eg dμ ≤ 1. Now, for every 0 < μ(A) ≤ 12 ,

1 e2 1
μ(A) log 1 + ≤ μ(A) log 1 +
4 μ(A) 2μ(A)
so that the claim immediately follows. The proposition is therefore established.
One further advantage of measure-capacity inequalities is that they allow for

some immediate comparison. Indeed, it is clear from Propositions 8.3.1 and 8.3.2
that a logarithmic Sobolev inequality is stronger than a Poincaré inequality. Sharp
constants as in Proposition 5.1.3, p. 238, are however somewhat lost in this proce-
dure.
8.4 Capacity and Further Functional Inequalities

There are many other functional inequalities which can be seen as a capacity in-
equalities with statements analogous to the preceding characterizations. This is the
case, for example, for the entropy-energy inequalities of Sect. 7.1, p. 348, which
may be handled as in Proposition 8.3.2. In this section, we focus on Nash, weak
Poincaré and Latała-Oleszkiewicz inequalities as presented in Chap. 7.
Again, capacities allow us to deal as easily with pairs of measures (ν, μ) where
(E, μ, ) is a Markov Triple and ν a measure on (E, F), with μ possibly infinite
(and ν finite). For simplicity, we only deal with the case where ν = μ is a probability
measure.
8.4.1 Capacity and Nash Inequalities
We start with the generalized Nash inequalities N () of Definition 7.4.1, p. 364,
that is for a (C 1 increasing and concave) growth function on (0, ∞),

f 2 dμ ≤ E(f )
E
for every f ∈ D(E) such that f 1 = 1. It will

turn out to be more convenient to
work with the linearized form (7.4.1), p. 365, E f 2 dμ ≤ s E(f ) + β(s) which holds
for all functions f ∈ D(E) with f 1 = 1 where β is the corresponding decreasing
rate function. Actually, capacity inequalities translate best for the (generalized) in-
verse function δ = β −1 of the decreasing rate function β. Hence, we consider here
the (generalized) Nash inequality in the form

f 2 dμ ≤ δ(u)E(f ) + u (8.4.1)
E
which holds for all functions f ∈ D(E) with f 1 = 1 with rate function
δ : [1, ∞) → (0, ∞) (since μ is a probability measure, only the values u ≥ 1 are
relevant).
The next statement is the announced equivalence between Nash and suitable ca-
pacity inequalities.

and let δ : [1, ∞) → (0, ∞) be decreasing and such that u → uδ(u) is increasing.
Assume that for all measurable subsets A ⊂ E with μ(A) ≤ 12 ,
μ(A)
1
≤ Cap∗μ (A). (8.4.2)
δ( μ(A) )
Then μ satisfies a generalized Nash inequality (8.4.1) with associated rate func-
tion 12 δ. Conversely, assume furthermore that there exists a q ≥ 4 such that
qδ(qu) ≥ 4 δ(u) for all u ≥ 1. Then, if μ satisfies a generalized Nash in-
equality (8.4.1) with rate function δ, then it satisfies (8.4.2) with rate function
u → 2qδ( u2 ).
Observe that for the classical Euclidean Nash inequality (6.2.3), p. 281 (with
A = 0), for which (r) = Cr n/(n+2) (although this case is formally excluded here
since the statement is restricted for convenience to probability measures), it holds
that μ(A)(n−2)/n ≤ C Cap(A), which is of the same form as the inequality arising
from Sobolev inequalities (Proposition 8.2.1). The conclusion here therefore recov-
ers the equivalence between Sobolev and Nash inequalities studied in Sect. 6.2,
p. 279.
Proof Given f ∈ D(E), let m be a median of f for μ. For every u ≥ 1,
2 2

f 2 dμ − u |f |dμ = Varμ |f | − (u − 1) |f |dμ
E E E
2
≤ F dμ − (u − 1)
2
|F |dμ
E E
where F = f − m. Since g ≥ 0,
2 " #
1
|F |dμ = inf F g dμ ; g ≥ 0,
2
dμ ≤ 1 ,
E E E g
for every u ≥ 1,
2
F 2 dμ − (u − 1) |F |dμ
E E
" #
−1 1
= sup F h dμ ; h ≤ 1, (1 − h) dμ ≤
2
E E u−1
" #
1
≤ sup F 2 g dμ ; g ∈ [0, 1], (1 − g)−1 dμ ≤ 1 + .
E E u−1
It follows that
2
F 2 dμ − (u − 1) |F |dμ
E E
" #
1
≤ sup F+2 g dμ ; g ∈ [0, 1], (1 − g)−1 dμ ≤ 1 +
B+ E u−1
" #
1
+ sup F−2 g dμ ; g ∈ [0, 1], (1 − g)−1 dμ ≤ 1 + ,
B− E u−1
where B+ (respectively B− ) is the support of F+ = (f − m)+ (respectively

F− = (f − m)− ). For u = 1, the suprema run over all functions g ∈ [0, 1] with-
out any assumption on the integral.
As for the logarithmic Sobolev inequalities in the previous section, by a direct
computation, for any measurable subset A ⊂ E,
" #
1
sup g dμ ; g ∈ [0, 1], (1 − g)−1 dμ ≤ 1 +
A E u−1
(8.4.3)
μ(A)
=
1 + (u − 1)μ(A)
(this and similar formulas remain valid in the limit case u = 1). Now, for every u ≥ 1
and a ∈ (0, 12 ], using the fact that δ(u) is decreasing and uδ(u) is increasing, it is
easily checked according to whether u ≤ a1 or not that
a aδ(u)
≤ .
1 + (u − 1)a δ( a1 )

Therefore, for every function g ∈ [0, 1] satisfying E (1 − g)−1 dμ ≤ 1 + u−1

1
, since
μ(B+ ) ≤ 2 , the hypothesis (8.4.2) and (8.4.3) imply that, for every A ⊂ B+ and
1
u ≥ 1,

μ̃(A) = g dμ ≤ δ(u) Capμ (A, B+ ).
A
Proposition 8.2.1 and Remark 8.2.2 applied to μ and μ̃ then yield that for every
u ≥ 1,

F+ g dμ ≤ 12 δ(u)
2
(F+ )dμ.
B+ B+
Optimizing over g together with the same inequality for F− allows us to conclude
the first assertion of the proposition.
Turning to the converse statement, let A ⊂ B with μ(B) ≤ 12 and let as usual f
in D(E) be such that 1A ≤ f ≤ 1B . Define for every k ∈ N, fk = (f − 2−k )+ ∧ 2−k
and Nk = {f > 2−k }. For k ∈ N, (8.4.1) applied to fk shows that, for every u ≥ 1,
2
fk2 dμ ≤ δ(u) (fk )dμ + u |fk |dμ .
E E E

Since ( E |fk |dμ)2 ≤ μ(Nk ) E fk2 dμ, choosing u = 2μ(1Nk ) this inequality be-
comes

1 1 1
2−2k−1 μ(Nk−1 ) ≤ fk2 dμ ≤ δ E(fk ) ≤ δ E(f )
2 E 2μ(Nk ) 2μ(Nk )
for every k ∈ N. Lemma 8.4.2 below for k = 0 shows that μ(N0 ) ≤

2 qδ( 2μ(1N0 ) )E(f ). Since μ(A) ≤ μ(N0 ) and u → uδ(u) is increasing, it follows
that
μ(A)
1
≤ E(f ).
2q δ( 2μ(A) )
Optimizing over all f ’s as above then yields the converse statement of the proposi-
tion. Proposition 8.4.1 is therefore established.
For the proof to be complete, it remains to prove the following technical lemma.
Lemma 8.4.2 Assume that δ : [1, ∞) → R+ is decreasing, that u → uδ(u) is in-

creasing and that there exits a q ≥ 4 such that qδ(qu) ≥ 4 δ(u) for all u ≥ 1. If for
some increasing sequence (ak )k∈N of real numbers such that 0 < ak ≤ 12 , k ∈ N, and
some constant C > 0 and every k ∈ N,

−2k 1
2 ak−1 ≤ C δ , (8.4.4)
2ak
then, for every k ∈ N,

1
2−2k ak ≤ Cq δ . (8.4.5)
2ak
Proof Define the sequence δk = δ( 2a1k ), k ∈ N, which, by the hypotheses, is increas-

δk ak
ing and such that ak , k ∈ N, is decreasing. The aim is to prove that δk ≤ Cq22k for
every k. Since ak ≤ 12 and q ≥ 4, this is true if 2−2k ≤ 8Cδk and the latter clearly
holds if k is large enough. It remains to see that if (8.4.5) holds for some k (≥ 1),
it also holds for k − 1. To this end, using (8.4.4) and the fact that u → uδ(u) is
increasing, we have
ak−1 C 22k δk
≤ .
δk−1 δ( 12k )
2C2 δ k
By the recurrence hypothesis and since u → δ(u) is decreasing, it follows that
ak−1 C 22k δk
≤ .
δk−1 δ( 2aq k )
The final conclusion is reached using that δ(qu) ≥ q4 δ(u).
8.4.2 Capacity and Weak Poincaré Inequalities
We turn to the analogous discussion for the weak Poincaré inequalities of Defini-
tion 7.5.2, p. 375, that is (for μ a probability measure),

Varμ (f ) ≤ ! E(f )
for every bounded functions f in D(E) such that (for example) osc(f ) =
sup f − inf f = 1, where ! : (0, ∞) → R+ is a given (C 1 ) concave increasing
growth function bounded by 1 and such that limr→0 !(r) = !(0) = 0. Again, it
will be better to work with the linearized version (7.5.2), p. 376, or rather, with the
(generalized) inverse = γ −1 of the rate function γ there. Hence, we agree here
that (E, μ, ) satisfies a weak Poincaré inequality if
Varμ (f ) ≤ (u)E(f ) + u, u ∈ (0, 1), (8.4.6)
for every bounded function f in D(E) such that osc(f ) = 1 where

: (0, 1) → (0, ∞) is decreasing.
The following statement is then an analogue of Proposition 8.4.1 above. Again,
there is an analogous versions for a pair of measures.

and let : (0, 1) → (0, ∞) be decreasing. Assume that for all measurable subsets
A ⊂ E with μ(A) ≤ 12 ,
μ(A)
≤ Cap∗μ (A). (8.4.7)
(μ(A))
Then μ satisfies a weak Poincaré inequality with rate function 12 . Conversely, if
μ satisfies a weak Poincaré inequality (8.4.6) with rate function , then it satis-
fies (8.4.7) with rate function u → 4 ( u4 ).
Proof The proof follows the same pattern as the previous ones. The second part
of the proposition is rather easy. Namely, for A ⊂ B ⊂ E such that μ(B) = 12 , if
1A ≤ f ≤ 1B , then osc(f ) = 1. Therefore, as for Poincaré inequalities in Proposi-

tion 8.3.1,

1 μ(A)
sup − u ≤ Capμ (A, B).
u∈(0, 1 )
(u) 2
4
μ(A)
The choice of u = 4then yields the claim.
Turning to the direct implication, assuming the capacity inequality (8.4.7), let
f ∈ D(E) and let m be a median of f with respect to μ. As in previous similar
proofs, set F+ = (f − m)+ and F− = (f − m)− so that

Varμ (f ) ≤ (f − m) dμ =2
F+ dμ + F−2 dμ.
2
E E E

We handle the term E F+2 dμ, E F−2 dμ being treated similarly. Denote by B+ the
support of F+ thus satisfying μ(B+ ) ≤ 12 . Fixing u ∈ (0, 1), define

c = c(u) = inf r ≥ 0 ; μ(F+ > r) ≤ u .

If c = 0, then E F+2 dμ ≤ u(sup F+ )2 since μ(B+ ) ≤ u. If c > 0, define for

a ∈ (0, 1) and k ∈ N, Nk = {F+ > ca k } and uk = μ(Nk ). Observe that
1
u0 ≤ u ≤ u1 ≤ · · · ≤ uk ≤ · · · ≤ μ(B+ ) ≤ .
2
!
By the decomposition {F+ > 0} = N0 ∪ k∈N (Nk+1 \ Nk ),

F+2 dμ = F+2 dμ + F+2 dμ
E N0 k≥0 Nk+1 \Nk

≤ u0 (sup F+ )2 + c2 a 2k (uk+1 − uk ).
k∈N
Since u0 ≤ u and
1 − a 2 2k
a 2k (uk+1 − uk ) = a (uk − u0 ),
a2
k∈N k≥1
the latter inequality turns into

c2 (1 − a 2 ) 2k
F+2 dμ ≤ u(sup F+ )2 + a (uk − u0 ).
E a2
k≥1
Since : (0, 1) → (0, ∞) is positive and decreasing, for any integer k ≥ 1, the map
uk −u0 + θ u0
θ → (u+θ (uk −u)) is increasing on (0, 1). Hence
uk − u0 uk
≤ ≤ Cap∗μ (Nk ) ≤ Capμ (Nk , Nk+1 ),
(u) (uk )
and
uk − u0 ≤ (u) Capμ (Nk , Nk+1 ).
Define then, for every k ∈ N,
1
fk = F+ − c a k+1 + ∧ c a k (1 − a) .
c a k (1 − a)
Then 1Nk ≤ fk ≤ 1Nk+1 and, by definition,

Capμ (Nk , Nk+1 ) ≤ (fk )dμ
Nk+1 \Nk

1
≤ (F+ )dμ.
c2 a 2k (1 − a)2 Nk+1 \Nk
Collecting the various estimates, it finally follows that

1+a
F+ dμ ≤ 2
2
(u) (F+ )dμ + u(sup F+ )2 .
E a (1 − a) B+
After optimizing on a ∈ (0, 1), the inequality becomes

F+2 dμ ≤ 12 (u) (F+ )dμ + u (sup F+ )2 .
E B+
Together with the corresponding bound for F− , it follows by (8.3.2) that

Varμ f 2 ≤ 12 (u)E(f ) + u (sup f − m)2 + (inf f − m)2 .
The conclusion then follows since if p = sup f and q = inf f , then for any
m ∈ [p, q], (p − m)2 + (q − m)2 ≤ (p − q)2 . The proof is complete.
As mentioned in Proposition 7.5.8, p. 379, any reasonable probability measure

satisfies a weak Poincaré inequality. However, to describe the associated growth
function !, or rate function of (8.4.6), is not an easy task in general. In dimension
one, however, since capacities can be made explicit, the description of ! or is
often reduced to a Muckenhoupt criterion as for Poincaré or logarithmic Sobolev
inequalities (cf. Theorem 4.5.1, p. 194, and Theorem 5.4.5, p. 256). The following
result is essentially established in the same way and there is also a version of this
characterization for a pair of measures.
Proposition 8.4.4 Let μ be a probability measure on the Borel sets of R with

strictly positive density p with respect to the Lebesgue measure and median m. De-
note by C > 0 the optimal constant of the weak Poincaré inequality in its linearized
form

Varμ (f ) ≤ C (u) f 2 dμ + u, u ∈ (0, 1),
R
for every bounded function f ∈ D(E) such that osc(f ) = 1. Then, there exist uni-
versal constants c, c > 0 such that
c max(b− , b+ ) ≤ C ≤ c max(B− , B+ ),
where
x
1 1
b+ = sup μ [x, +∞) dt,
x≥m (μ([x, +∞))/4) m p(t)
x
1 1
B+ = sup μ [x, +∞) dt,
x≥m (μ([x, +∞))) m p(t)
and similarly,

1 m 1
b− = sup μ (−∞, x] dt,
x≤m (μ((−∞, x])/4) x p(t)
and

1 m 1
B− = sup μ (−∞, x] dt.
x≤m (μ((−∞, x])) x p(t)
What this proposition tells us is that in order to determine a rate function , it is

necessary that B+
and B− are finite, which amounts to the comparison as x → ∞
∞ 1
of μ([x, ∞)) and x p(t) dt together with a similar estimate as x → −∞. As an
illustration, the model example dμα (x) = cα e−(1+x ) dx on the line (summarized
2 α/2
in Sect. 7.7, p. 386) satisfies a weak Poincaré inequality for 0 < α < 1 with rate
function (u) = C (− log u)2(1−α)/α , u ∈ (0, 1). Alternatively, the growth function
! is given by !(r) = (C r(− log r)2(1−α)/α ) ∧ 1, r ∈ (0, ∞), for some other con-
stant C > 0. When α = 1, since μ1 satisfies a Poincaré inequality P (C), (u) = C
for all u ∈ (0, 1). As a further example, if dμ(x) = cγ (1 + x 2 )−γ /2 dx with γ > 1, a
weak Poincaré inequality holds with rate function (u) = C u2/(1−γ ) , u ∈ (0, 1), or
equivalently !(r) = (C r (γ −1)/(γ +1) ) ∧ 1, r ∈ (0, ∞).
8.4.3 Capacity and Latała-Oleszkiewicz Inequalities
In the last part of this section, we analogously briefly translate Latała-Oleszkiewicz-

type inequalities in terms of capacity inequalities. Recall that, according to (7.6.4),
p. 385, a Markov Triple (E, μ, ) with μ a probability measure is said to satisfy
a Latała-Oleszkiewicz inequality LO(a, C) with parameter a ∈ [0, 1] and constant
C > 0 if
2/q
1
sup f 2
dμ − |f |q
dμ ≤ C E(f ). (8.4.8)
q∈[1,2) (2 − q)
a
E E
8.5 Gaussian Isoperimetric-Type Inequalities Under a Curvature Condition 411
for every f ∈ D(E). The following statement is in the same spirit as those for gen-
eralized Nash or weak Poincaré inequalities. We omit the proof.
Proposition 8.4.5 Let (E, μ, ) be a Markov Triple, with μ a probability measure.

Assume that for some constant C1 > 0 and a ∈ [0, 1],
a
1
μ(A) log 1 + ≤ C1 Cap∗μ (A) (8.4.9)
μ(A)
for all measurable sets A ⊂ E such that μ(A) ≤ 12 . Then μ satisfies a Latała-
Oleszkiewicz inequality LO(a, cC1 ) where c > 0 is a numerical constant. Con-
versely, under the Latała-Oleszkiewicz inequality LO(a, C2 ) of (8.4.8) with C2 > 0,
the inequality (8.4.9) holds with constant C1 = c C2 where c > 0 is numerical.
This result recovers the fact that Latała-Oleszkiewicz inequalities are an inter-
polation between Poincaré and logarithmic Sobolev inequalities according to the
parameter a ∈ [0, 1]. Moreover, following Proposition 8.4.1, it provides a link with
generalized Nash inequalities. Namely, if μ satisfies LO(a, C), a ∈ [0, 1], C > 0,
then it satisfies a generalized Nash inequality in the linearized form (8.4.1) with rate
function
−a
δ(u) = c C log(1 + u) , u ∈ (1, ∞),
where c > 0 is a numerical constant. With the tools developed above, Latała-
Oleszkiewicz inequalities may be shown to be characterized similarly through a
Muckenhoupt-type criterion. In particular, the model family μα on the real line sat-
isfies LO(a, C) if and only if α ≥ 2−a
2
.
8.5 Gaussian Isoperimetric-Type Inequalities Under

a Curvature Condition
In the second part of this chapter, starting with this section, we consider another form
of measure-capacity inequalities: the isoperimetric-type inequalities. The p = 1
definition of capacity may be viewed as a boundary or surface measure so that a
capacity inequality compares in this case measures and surface measures, which is
the essence of isoperimetric inequalities.
To address these issues, the first sub-section collects general properties on bound-
ary measures and co-area formulas. Our main focus is then Gaussian-type isoperi-
metric inequalities for heat kernel measures under curvature condition, providing in
particular a semigroup proof of the classical Gaussian isoperimetric inequality.
The setting of this and the following sections is that of a Full Markov Triple
(E, μ, ). As in all this work, the emphasis is on functional inequalities, and the
main results here actually describe functional forms of isoperimetric properties (for
heat kernel and invariant measures).
However, it is of interest to relate these functional inequalities to the more geo-

metric and traditional statements. To this end, and only for the matter of compari-
son, it is convenient to supplement the Markov Triple framework with convenient
topological features. As we know, a Markov Triple (E, μ, ) includes the notion of
intrinsic distance d (Sect. 3.3.7, p. 166). In general, d is not always a true metric.
When it will be useful and necessary to suitably describe the geometric conclusions
from the functional ones, we will therefore assume that the intrinsic distance d de-
fines a true Polish metric on E, and that μ is a non-atomic Borel measure on (E, d).
We furthermore assume that the √ functions in A0 are continuous (in the metric d)
and that for functions f in A, (f ) may be identified with a modulus of gradient
|f (x) − f (y)|
(f )(x) = lim sup (8.5.1)
d(x,y)→0 d(x, y)
(for every x ∈ E). In this setting, and when μ is a probability, Lipschitz means that
|f (x) − f (y)| ≤ Cd(x, y) for some C > 0 on the support of μ ⊗ μ, and it will
be assumed in addition that bounded Lipschitz functions (in this sense) are in the
Dirichlet domain D(E).
This description suitably covers the examples of weighted Riemannian manifolds
with (f ) the usual Riemannian length squared of the gradient of a smooth function
f (cf. Sect. C.4, p. 509). These assumptions will be implicit as soon as geometric
statements on sets (involving isoperimetric neighborhood and surface measure) are
considered below, and they are in particular in force in the first sub-section. It should
be pointed out that these hypotheses might rule out natural infinite-dimensional ex-
amples such as Wiener spaces (cf. Sect. 2.7.2, p. 108). Finite-dimensional approxi-
mations may nevertheless be developed to cover such instances on the basis of the
dimension-free inequalities described here (for example for Gaussian measures).
8.5.1 Isoperimetric-Type Inequalities
As explained in the previous introduction, an isoperimetric-type inequality is a

measure-capacity inequality where the 2-capacity is replaced by a boundary mea-
sure identified with a 1-capacity.
In the (true) intrinsic metric d, and together with (8.5.1), whenever f ap-

√ set A ⊂ E (for example
proximates the characteristic function of a closed
fε = (1 − 1ε d(·, A))+ and ε → 0), the integral E (f )dμ approaches the so-
called Minkowski (exterior) boundary measure of the set A defined by
1
μ+ (A) = lim inf μ(Aε ) − μ(A) (8.5.2)
ε→0 ε
where Aε = {x ∈ E ; d(x, A) < ε}, ε > 0, is the ε-(open) neighborhood of A ⊂ E.
The quantity μ+ (A) describes the surface measure of A.
The measure μ is then said to satisfy an isoperimetric-type inequality if there

exists a (non-trivial) positive function J on R+ , called an isoperimetric function
for μ, such that

J μ(A) ≤ μ+ (A) (8.5.3)
for every measurable (closed) subset A of E (of finite measure). Equivalently, define
the isoperimetric profile Iμ of μ on (E, d) (or of (E, μ, )) as the pointwise maxi-
mal function J in (8.5.3), so that an isoperimetric-type inequality bounds from be-
low the isoperimetric profile Iμ . When equality is achieved in (8.5.3) for some non-
trivial sub-family of sets A, called the extremal sets, determining Iμ , the inequal-
ity (8.5.3) expresses a true isoperimetric statement, comparing the surface measure
of a set to that of an extremal set with the same measure.
For a given measure on the Borel sets of a metric space (E, d), not that many
isoperimetric profiles are known. One may mention the standard isoperimetric
inequality for the Lebesgue measure on Rn given by the isoperimetric function
1/n
nωn v (n−1)/n , v ∈ R+ , where ωn is the volume of the Euclidean unit ball, and
for which balls are the extremal sets. Similarly, (geodesic) balls are the extremal
sets of the isoperimetric problem on the sphere Sn ⊂ Rn+1 with respect to the uni-
form measure and the isoperimetric profile is again described by its value on these
extremal sets. The next sub-section will feature the further example of the isoperi-
metric profile of Gaussian measures (for which half-spaces are the extremal sets).
Another example is that of the exponential measure dμ(x) = e−x dx on R+ for
which Iμ (v) = min(v, 1 − v), v ∈ [0, 1].
The theorem below presents a co-area formula involving the surface measure μ+
instead of the 2-capacity of Theorem 8.1.2. Both the statement itself and its proof
are purely metric.
Theorem 8.5.1 (Co-area formula II) In the metric space setting (E, d) induced by
the Markov Triple (E, μ, ) as described above, for every Lipschitz function f on
E,
+∞
μ+ (Nr )dr ≤ (f ) dμ (8.5.4)
−∞ E
where Nr = {x ∈ E, f (x) > r}, r ∈ R.
In the context of a general Markov Triple, there is no reason why (8.5.4) should
be an equality. Equality is known in specific instances such as Rn with a measure
μ which is absolutely continuous with respect to the Lebesgue measure, in which
case (8.5.4) is classically referred to as a co-area formula.
Proof Assume first that f : E → R is bounded and moreover, without loss of gen-
erality, positive. For each ε > 0, set fε (x) = supd(x,y)<ε f (y), x ∈ E. Since f is
Lipschitz, fε is finite and lower semi-continuous. From (8.5.1), for every x ∈ E,
fε (x) − f (x)
lim sup ≤ (f )(x).
ε→0 ε
The dominated convergence Theorem then implies that

1
lim sup (fε − f )dμ ≤ (f ) dμ.
ε→0 ε E E
On the other hand, for all r > 0,

x ∈ E ; fε (x) > r = x ∈ E ; f (x) > r ε = (Nr )ε ,

∞
the ε-neighborhood of Nr . Since E f dμ = μ(Nr )dr and E fε dμ =
∞ 0
0 μ((Nr )ε )dr,
∞
1 1
(fε − f )dμ = μ (Nr )ε − μ(Nr ) dr.
ε E ε 0
By Fatou’s Lemma,
∞
1 ∞ μ((Nr )ε ) − μ(Nr )
lim inf μ (Nr )ε − μ(Nr ) dr ≥ lim inf dr
ε→0 ε 0 0 ε→0 ε
∞
= μ+ (Nr )dr.
0
The desired inequality follows. Whenever f is unbounded, apply the preceding to

the sequence of Lipschitz functions fk = (f ∧ ak ) ∨ (−ak ), ak ∈ N (where the se-
quence (ak )k∈N is chosen such that limk→∞ ak = ∞ and μ(f = ak ) = 0) and let
k → ∞. Theorem 8.5.1 is established.
Theorem 8.5.1 is typically used to provide evidence that a lower bound on the
isoperimetric profile Iμ may imply standard functional inequalities for μ. A clas-
sical example is that of Lebesgue measure in Rn , n > 1. Namely, if f : Rn → R is
smooth and compactly supported, then by the co-area formula, denoting respectively
by voln and vol+
n the volume and surface measures,
∞

vol+
n |f | > r dr ≤ |∇f |dx.
0 Rn
By the isoperimetric inequality, for every r > 0,
(n−1)/n
≤ vol+
1/n
n
ωn voln |f | > r n |f | > r
from which (after integration in r and a little bit of work)

1/n
ωn f p
n ≤ |∇f |dx
Rn
with p = n−1 n
. Changing f into f 2 then yields the standard Sobolev inequal-
ity (6.1.1), p. 278, in Rn (however not with its best constant).
Some more care has to be taken for finite measures. One classical example in
this case is the following statement expressing that whenever Iμ may be compared
to the isoperimetric profile of the exponential measure, then it satisfies a Poincaré
inequality. Similar statements may be obtained for different behaviors of the isoperi-
metric function and different functional inequalities. For example, as will be used
below in Theorem 8.7.2, comparison with the Gaussian isoperimetric profile yields
logarithmic Sobolev inequalities.
Proposition 8.5.2 In the preceding context with μ a probability measure, assume

that there exists an isoperimetric function J for μ such that, for some c > 0
J (v) ≥ c min(v, 1 − v), v ∈ [0, 1].
Then μ satisfies a Poincaré inequality P ( c42 ).
Proof By Theorem 8.5.1 and the hypothesis, for any positive and Lipschitz g,
∞

c min μ(g > r), 1 − μ(g > r) dr ≤ (g) dμ. (8.5.5)
0 E
Let now f : E → R be Lipschitz and denote by m a median of f for μ. Set as usual

F+ = (f − m)+ and F− = (f − m)− so that f − m = F+ − F− . By the definition
of the median, for every r > 0,
1 1
μ (F+ )2 > r ≤ and μ (F− )2 > r ≤ .
2 2
Hence, (8.5.5) applied to g = (F+ )2 and g = (F− )2 yields

c |f − m| dμ = c (F+ ) dμ + c (F− )2 dμ
2 2
E E E
∞ ∞

=c μ (F+ )2 ≥ r dr + c μ (F− )2 ≥ r dr (8.5.6)
0 0

≤ (F+ ) dμ +
2 (F− )2 dμ.
E E
By the Cauchy-Schwarz inequality,

(F+ ) dμ = 2 F+ (F+ ) dμ
2
E E
1/2 1/2
≤2 F+2 dμ (F+ )dμ
E E
and similarly for F− . By (8.3.2), the right-hand side of (8.5.6) is less than or equal to
1/2 1/2
2 |f − m| dμ
2
(f )dμ .
E E
Therefore, for every median m of f ,

c2
|f − m|2 dμ ≤ (f )dμ.
4 E E

Since the variance of f is the infimum of E |f − a|2 dμ over a ∈ R, it follows that

μ satisfies a Poincaré inequality P ( c42 ). Since in the considered setting any function
f in A0 is Lipschitz, the claim follows.
8.5.2 Gaussian Isoperimetric-Type Inequalities Under CD(ρ, ∞)
While it is in general a difficult issue to determine extremal sets for an isoperimetric

problem such as (8.5.3), however one might expect to find comparison theorems
involving a given isoperimetric function which might not be the optimal one. The
theme of this section is comparison with the Gaussian isoperimetric profile under
curvature conditions. Accordingly, we will speak of a Gaussian isoperimetric-type
inequality for a Markov Triple (E, μ, ) if (8.5.3) holds for J the isoperimetric
profile of the (standard) Gaussian measure (see below).
The standard Gaussian measure on the Borel sets of Rn satisfies, with respect to
the standard carré du champ operator (f ) = |∇f |2 on smooth functions, numer-
ous functional inequalities of interest, in particular the logarithmic Sobolev inequal-
ity (Proposition 5.5.1, p. 258). The powerful curvature tools allowed in Sects. 5.5
and 5.7, p. 257 and p. 268, for the extension of these inequalities to Markov Triples
(E, μ, ) under the curvature condition CD(ρ, ∞) with ρ > 0. This covers in par-
ticular the case of probability measures dμ = e−W dx on Rn where the smooth po-
tential W is such that uniformly ∇∇W ≥ ρ Id > 0, ρ > 0 (as symmetric matrices).
The main conclusion of the investigation here is a stronger property in the form of
the isoperimetric inequality for the standard Gaussian measure and its extension, as
a comparison property, to positively curved diffusion operators. The key will again
be provided by the -calculus on heat kernel measures via a suitable functional de-
scription of isoperimetry. The heat flow proof which will be developed towards this
goal therefore also entails a proof of the isoperimetric inequality for the Gaussian
measure itself.
To state the main result, we introduce the isoperimetric profile of Gaussian mea-

r
sures. Let (r) = −∞ e−x /2 √dx , r ∈ R, be the distribution function of the stan-
2
2π
dard Gaussian measure on the real line and ϕ = its density with respect to the
Lebesgue measure. Then, as will be justified below,

−1 1
I = ϕ ◦ : [0, 1] → 0, √ (8.5.7)
2π
defines the Gaussian isoperimetric profile (in any dimension). Note that the function
I is concave continuous, symmetric with respect to the vertical line going through
2 , and such that I(0) = I(1) = 0, and satisfies the fundamental differential equation
1
I I = −1.
√
The main result of this section is that ρ I is an isoperimetric function for the
invariant probability measure μ of a Markov Triple (E, μ, ) satisfying the cur-
vature condition CD(ρ, ∞) for some ρ > 0. To reach this goal, we use an ad-hoc
functional description of isoperimetry, known as a Bobkov inequality, and, as for
Poincaré and logarithmic Sobolev inequalities, we establish such Bobkov inequali-
ties for heat kernel measures under curvature conditions.
Theorem 8.5.3 (Local Bobkov inequalities) Let (E, μ, ) be a Markov Triple with
semigroup P = (Pt )t≥0 . The following assertions are equivalent.
(ii) For every function f in A0 with values in [0, 1], every (or some) α ≥ 0 and
every t ≥ 0,

I 2 (Pt f ) + α (Pt f ) ≤ Pt I 2 (f ) + cα (t)(f ) (8.5.8)
where
1 − e−2ρt
cα (t) = + α e−2ρt , t ≥0
ρ
(with cα (t) = 2t + α whenever ρ = 0).
Before turning to the proof of this result, let us illustrate its isoperimetric content
in the form of the following corollary, which follows from the choice of α = ρ1 with
ρ > 0 (so that cα (t) = ρ1 for every t ≥ 0) and letting t → ∞ by ergodicity. In this
context, μ is a probability measure.
Corollary 8.5.4 (Bobkov inequality under CD(ρ, ∞)) Let (E, μ, ) be a Markov
Triple, with μ a probability measure, satisfying the curvature condition CD(ρ, ∞)
for some ρ > 0. Then, for every f in D(E) with values in [0, 1],

√
ρI f dμ ≤ ρ I 2 (f ) + (f ) dμ. (8.5.9)
E E
In a metric context, applying this corollary to functions approximating the

characteristic function of a measurable (closed) subset A in E yields that (using
I(0) = I(1) = 0),
√
ρ I μ(A) ≤ μ+ (A) (8.5.10)
where we recall the surface measure μ+ (A) as defined by the Minkowski content
formula (8.5.2). In an alternative integrated form, for any measurable (closed) subset
A in E and any r > 0,

√
−1 ◦ μ(Ar ) ≥ −1 ◦ μ(A) + ρr (8.5.11)
+
(since drd
−1 ◦ μ(Ar )|r=0 = Iμ(μ(A))
(A)
).
Note that, for the Gaussian measure in Euclidean space itself, integrating (8.5.10)
along the level sets of a function yields conversely, due to the dimension-free charac-
ter of the Gaussian isoperimetric profile I, the functional form (8.5.9) which there-
fore fully describes Gaussian isoperimetry.
The preceding conclusion may be summarized as follows.
Corollary 8.5.5 (Lévy-Gromov isoperimetric comparison Theorem) In the pre-

ceding metric context, under the curvature condition CD(ρ, ∞) for some ρ > 0,
√
the isoperimetric profile Iμ of (E, μ, ) is bounded from below by ρ I where I
is the Gaussian isoperimetric profile.
The corollary applies in particular to measures dμ = e−W dx on Rn such that

∇∇W ≥ ρ Id, ρ > 0. The result actually covers the optimal isoperimetric inequality
for the standard Gaussian measure dμ(x) = (2π)−n/2 e−|x| /2 dx on Rn for which
2
ρ = 1 and the half-spaces H = {x ∈ Rn ; x · u ≤ r}, r ∈ R, u unit vector in Rn ,

achieve equality in (8.5.10) (since μ(H ) = (r) and μ+ (H ) = ϕ(r)), providing
therefore a purely functional proof of the Gaussian isoperimetric inequality itself.
Note that since the measures of half-spaces are one-dimensional, the isoperimetric
statement is dimension-free. The previous isoperimetric application may be devel-
oped similarly at the level of the heat kernel measures in Theorem 8.5.3. In partic-
ular, for the Euclidean semigroup in Rn , we recover in the same way the Gaussian
isoperimetric inequality.
Furthermore, it should be pointed out that the Bobkov inequality (8.5.9) in Corol-
lary 8.5.4 may actually be studied in its own right, as for Poincaré or logarithmic
Sobolev inequalities. In particular, it may be shown to be stable under products,
justifying its relevance towards dimension-free isoperimetric properties. It also im-
proves upon the logarithmic Sobolev inequality, preserving constants, whenapplied
to εf 2 with a Taylor expansion as ε → 0 using the asymptotics I(v) ∼ v 2 log v1
as v → 0. In this way, Corollary 8.5.4 implies the logarithmic Sobolev inequality
of Proposition 5.7.1, p. 268, in the same context (and similarly Theorem 8.5.3 im-
proves upon Theorem 5.5.2, p. 259).
We next turn to the semigroup properties of Theorem 8.5.3 and to its proof. First
observe that, as for the Poincaré and logarithmic Sobolev inequalities for heat kernel
measures, since ccα (s) (t) = cα (s + t), (8.5.8) is stable under (Pt )t≥0 . With the same
Taylor expansion as the one used above under the invariant measure, (ii) improves
upon the Poincaré and logarithmic Sobolev inequalities of Theorem 4.7.2, p. 209,
and Theorem 5.5.2, p. 259, and therefore (8.5.8), which holds for every f in A0 with
values in [0, 1], every t > 0 and some fixed α ≥ 0, ensures conversely the curvature
condition CD(ρ, ∞). This may also be checked directly by a suitable (although
somewhat tedious) Taylor expansion at t = 0 together with the further asymptotics
f = 12 + εg as ε → 0. In addition, whenever α → ∞, we recover the fundamental

strong gradient bound of Theorem 3.3.18, p. 163,

(Pt f ) ≤ e−ρt Pt (f )
for any function f ∈ A0 and any t ≥ 0. In particular, these various arguments settle
the implication from (ii) to (i).
We now address the converse implication in the proof of Theorem 8.5.3. The ar-
gument is very similar to the semigroup monotonicity proofs of the local Poincaré
and logarithmic Sobolev inequalities. Basic use will be made of the differential
equation I I = −1. It will be convenient to record the following technical lemma.
Lemma 8.5.6 Let be smooth on R3 (or some open set in R3 ), f ∈ A0 and t > 0
be fixed. Then
d
Ps s, Pt−s f, (Pt−s f ) = Ps (K)
ds
with

K = ∂1 + 2 ∂3 2 (g) + ∂22 (g) + 2 ∂2 ∂3 g, (g) + ∂32 (g)
where we wrote on the right-hand side g = Pt−s f and for (s, g, (g)).
Proof We have
d
Ps s, Pt−s f, (Pt−s f )
ds

d
= Ps L s, Pt−s f, (Pt−s f ) + s, Pt−s f, (Pt−s f ) .
ds
By the diffusion property and using the notation of the statement,

d
L s, Pt−s f, (Pt−s f ) + s, Pt−s f, (Pt−s f )
ds

= ∂2 Lg + ∂3 L(g) + ∂22 (g) + 2 ∂2 ∂3 g, (g)

+ ∂32 (g) + ∂1 − ∂2 Lg − 2 ∂3 (g, Lg).
The lemma then follows by the very definition of the 2 operator.
Proof of Theorem 8.5.3 Assume the curvature CD(ρ, ∞) condition for some
ρ ∈ R. Let f ∈ Aconst+
0 with values in [0, 1] and t > 0 be fixed. For every s ∈ [0, t],
set

(s) = Ps I 2 (Pt−s f ) + cα (s)(Pt−s f ) .
(Strictly speaking, it would be necessary to replace (Pt−s f ) by (Pt−s f ) + ε for

some ε > 0.) Since cα (0) = α, it will be enough to show that is increasing. Apply
Lemma 8.5.6 with

(s, u, v) = I 2 (u) + cα (s)v, (s, u, v) ∈ (0, t) × (0, 1) × (0, ∞).
It is immediate that, with I = I(u), I = I (u),
cα cα
∂1 = v, ∂2 = I I , ∂3 = ,
2 2
and (using I I = −1)
cα
3 ∂22 = − I 2 I 2 + 2 I 2 − 1 , 3 ∂2 ∂3 = − I I ,
2
cα2
3 ∂32 = − .
4
Therefore, K of Lemma 8.5.6 satisfies
cα (s)
3K = 2 (g) + 2 cα (s)2 (g)
2

− I 2 (g)I 2 (g)(g) + 2 I 2 (g) − 1 (g)
cα (s)2
− cα (s)I(g)I (g) g, (g) − (g) ,
4
where = (s, g, (g)) (and g = Pt−s f ). Since

= s, g, (g) = I 2 (g) + cα (s)(g),
it follows that
c (s)
3 K = I 2 (g) + cα (s)(g) α (g)
2

+ I 2 (g) + cα (s)(g) cα (s)2 (g)

− I 2 (g)I 2 (g)(g) + I 2 (g) + cα (s)(g) I 2 (g) − 1 (g)
cα (s)2
− cα (s)I(g)I (g) g, (g) − (g) .
4
8.6 Harnack Inequalities Revisited 421
Therefore, after some algebra,

cα (s)
K = cα (s)(g) cα (s)2 (g) − 1 −
3
(g)
2

cα (s)2 cα (s)
− (g) + I (g) cα (s)2 (g) − 1 −
2
(g)
4 2

− cα (s) I (g)I(g) g, (g) + cα (s) I 2 (g)(g)2 .
c (s)
By the very definition of cα (s), 1 − α2 = ρ cα (s) so that

1
3 K = cα (s)2 (g) 2 (g) − ρ (g) − (g)
4
2
+ cα (s) I (g) 2 (g) − ρ (g) − I (g)I(g) g, (g) + I 2 (g)(g)2 .
Now, as a consequence of the reinforced curvature condition (3.3.6), p. 159,

4 (f ) 2 (f ) − ρ (f ) ≥ (f ) , f ∈ A,
applied twice,

((g))
2
K ≥ cα (s) I (g)
3 2
− I (g)I(g) g, (g) + I (g)(g) .
2
4(g)
The right-hand-side of this inequality is a quadratic form in I(g) and I (g) which
is positive since, as (f ) ≥ 0 for every f ,
2
f, (f ) ≤ (f ) (f ) .
Hence K ≥ 0 and thus (s) = Ps (K) ≥ 0 for every s ∈ [0, t]. This completes the
proof of Theorem 8.5.3.
The natural challenge after Theorem 8.5.3 would be to investigate similar con-
clusions under the curvature-dimension condition CD(ρ, n) (for some finite n). The
aim would be to reach in this way the isoperimetric inequality on the n-sphere,
and the corresponding comparison statement for manifolds with dimension n and
with a strictly positive lower bound on the Ricci curvature. However, this program
is mainly open since in particular a functional description of isoperimetry on the
sphere suitable for the -calculus is still missing. The analysis should probably also
involve the fast diffusion equation rather than the heat flow as in Sect. 6.11, p. 330.
8.6 Harnack Inequalities Revisited

Section 5.6, p. 265, investigated Harnack inequalities for infinite-dimensional
Markov generators with curvature CD(ρ, ∞). In particular, it was observed there
(see (5.6.4), p. 267) that the reverse forms of logarithmic Sobolev inequalities for
heat kernel measures may be used to get quite close to the optimal statement. To-
gether with the refined Gaussian isoperimetric function I, we show in this section
how to obtain these Harnack inequalities by means of a suitable reverse form of The-
orem 8.5.3. In addition, the conjunction of Theorem 8.5.3 together with its reverse
form will lead to an isoperimetric-type Harnack inequality.
The following statement is therefore a kind of reverse Bobkov inequality along
−2ρt
the semigroup. Recall c0 (t) = 1−eρ (= 2t if ρ = 0) from Theorem 8.5.3.
Proposition 8.6.1 Let (E, μ, ) be a Markov Triple satisfying the curvature con-
dition CD(ρ, ∞) for some ρ ∈ R. Then, for every f in A0 and every t ≥ 0,
2 2
I(Pt f ) − Pt I(f ) ≥ e2ρt c0 (t) (Pt f ). (8.6.1)
By the usual extension procedure, the inequality (for t > 0) applies to any measur-
able function f with values in [0, 1]. As a consequence, for every bounded measur-
able function f and every t > 0,
e−2ρt
(Pt f ) ≤ f 2∞ . (8.6.2)
2πc0 (t)
Proof The proof relies on the standard heat flow interpolation argument. For
f ∈ Aconst+
0 with values in [0, 1] and t > 0, write
t
2 2 d 2
I(Pt f ) − Pt I(f ) = − Ps I(Pt−s f ) ds.
0 ds
By the change of variables formula for the diffusion operator L,

d 2
− Ps I(Pt−s f )
ds

= −2Ps I(Pt−s f ) Ps L I(Pt−s f ) − I (Pt−s f ) LPt−s f

= −2Ps I(Pt−s f ) Ps I (Pt−s f ) (Pt−s f )

(Pt−s f )
= 2Ps I(Pt−s f ) Ps
I(Pt−s f )
where we used that I I = −1 in the last step. Since Ps is given by a kernel, it
satisfies a Cauchy-Schwarz inequality, and hence
2
A 2
Ps (B)Ps ≥ Ps (A) , A, B ≥ 0.
B

Therefore, with A = (Pt−s f ) and B = I(Pt−s f ),
t
2 2 2
I(Pt f ) − Pt I(f ) ≥ 2 Ps (Pt−s f ) ds.
0
8.6 Harnack Inequalities Revisited 423
√
By √the strong gradient bounds of Theorem 3.3.18, p. 163, Ps ( (g) ) ≥
eρs (Ps g). With g = Pt−s f , it follows that
t
2 2
I(Pt f ) − Pt I(f ) ≥ 2 e2ρs ds (Pt f )
0
which is the result. The gradient bound (8.6.2) is sharp on the example of the
Ornstein-Uhlenbeck semigroup (2.7.3), p. 104, for f the characteristic function of
a half-space.
It may be shown that as t → 0 (8.6.1) implies in return the curvature condition

CD(ρ, ∞). Also, by applying it to εf and letting ε → 0, one recovers the reverse
logarithmic Sobolev inequality (5.5.6) of Theorem 5.5.2, p. 259.
On the basis of Proposition 8.6.1, the aim is now to fully recover the Harnack
inequalities of Theorem 5.6.1, p. 265. Indeed, in terms of gradient bounds, since
(−1 ) = I1 , for every f ∈ A0 with values in [0, 1] and t > 0,
(Pt f ) 1
−1 ◦ Pt f = ≤
I(Pt f )2 e2ρt c0 (t)
as a consequence of (8.6.1). In particular, for x, y ∈ E and d(x, y) the intrinsic
distance between x and y,
d(x, y)
−1 ◦ Pt f (x) ≤ −1 ◦ Pt f (y) + √ (8.6.3)
eρt c0 (t)
which extends to any measurable function f with values in [0, 1]. This Lipschitz
property actually entails the announced Harnack inequalities. Set
δ = e−ρt c0 (t)−1/2 d(x, y), so that (8.6.3) reads as

Pt f (x) ≤ −1 ◦ Pt f (y) + δ .
Now apply this inequality with f replaced by 1{f ≥r} , r ≥ 0, for a positive measur-
able function f on E. Denoting by ν the distribution of f under Pt at the point y
(that is ν(B) = Pt (1{f ∈B} )(y) for every Borel set B in R),

Pt (1{f ≥r} )(x) ≤ −1 ν [r, ∞) + δ .
Integrating in r ≥ 0 and using Fubini’s Theorem,

∞ −1 (ν([r,∞)))+δ du
e−u
2 /2
Pt f (x) ≤ √ dr
0 −∞ 2π
∞ ∞
du
1{u≤−1 (ν([r,∞)))+δ} dr e−u /2 √ .
2
=
−∞ 0 2π
Change u into u + δ to get
∞ ∞
−δ 2 /2 −δu du
1{(u)≤ν([r,∞))} dr e−u /2 √ .
2
Pt f (x) ≤ e e
−∞ 0 2π
Changing u into −u and denoting by F the distribution function of ν, it follows that

∞ ∞
−δ 2 /2 du
1{F (r)≤(u)} dr e−u /2 √ .
2
Pt f (x) ≤ e e δu
−∞ 0 2π
After the further change of variable v = (u),
1 ∞
−δ 2 /2 δ −1 (v)
Pt f (x) ≤ e e 1{F (r)≤v} dr dv.
0 0
The next statement summarizes the conclusions reached so far.
Theorem 8.6.2 Let (E, μ, ) be a Markov Triple satisfying the curvature condition
CD(ρ, ∞) for some ρ ∈ R. For every positive measurable function f on E, every
t > 0 and every x, y ∈ E,
∞
−δ 2 /2 −1
Pt f (x) ≤ e eδ ◦F (s) s dF (s)
0
where δ = e−ρt c0 (t)−1/2 d(x, y) and F is the distribution function of f under Pt at

the point y.
Although not expressed in a very tractable form, Theorem 8.6.2 may be viewed
as the root of the various Harnack inequalities in this context. For example, by the
Cauchy-Schwarz inequality,
∞ ∞ 1/2 ∞ 1/2
δ −1 ◦F (s) 2δ −1 ◦F (s)
e s dF (s) ≤ e dF (s) 2
s dF (s)
0 0 0
1/2
δ2
≤ e Pt f 2 (y)
since
∞ 1 ∞
2δ−1 ◦F (s) 2δ−1 (v) 2 /2 du 2
e dF (s) = e dv = e2δu−u √ = e2δ .
0 0 −∞ 2π
The preceding therefore exactly recovers Wang’s Harnack inequality of Theo-
rem 5.6.1, p. 265, for α = 2,

2 d(x, y)2
(Pt f ) (x) ≤ Pt f (y) exp 2ρt
2
.
e c0 (t)
By Hölder’s inequality rather than the Cauchy-Schwarz inequality, one obtains the
whole family of inequalities with α > 1. Using the entropic inequality (5.1.2),
p. 236, yields similarly the log-Harnack inequality (5.6.2) of Remark 5.6.2, p. 266.
With respect to the original argument of Theorem 5.6.1, the proof here avoids inter-
polation along geodesics and holds true in the general context of Markov Triples.
8.7 From Concentration to Isoperimetry 425
As announced, Theorem 8.6.2 may be coupled with Theorem 8.5.3 to reach a

kind of isoperimetric Harnack inequality. We only sketch the argument which as-
sumes the necessary metric structure.
Theorem 8.6.3 (Isoperimetric Harnack inequality) Let (E, μ, ) be a Markov

Triple satisfying the curvature condition CD(ρ, ∞) for some ρ ∈ R. For every mea-
surable (closed) set A in E, every t ≥ 0 and every x, y ∈ E with d(x, y) > 0,
Pt (1A )(x) ≤ Pt (1Adt )(y)
where dt = e−ρt d(x, y). In particular, when ρ = 0,
Pt (1A )(x) ≤ Pt (1Ad(x,y) )(y).
Proof In its integrated form (cf. (8.5.11)), Theorem 8.5.3 implies that for any y ∈ E,
any measurable set A ⊂ E, any t > 0 and any r > 0,
r
−1 ◦ Pt (1Ar )(y) ≥ −1 ◦ Pt (1A )(y) + √
c0 (t)
where Ar is the r-neighborhood of A with respect to the distance d. On the other
hand, the Lipschitz property (8.6.3) applied to f = 1A ensures that
−1 ◦ Pt (1A )(x) ≤ −1 ◦ Pt (1A )(y) + δ
where δ = e−ρt c0 (t)−1/2 d(x, y). The combination of these two inequalities imme-
diately yields the result.
8.7 From Concentration to Isoperimetry

In this last section, we address the questions raised in Remarks 4.4.5, p. 193,
and 5.4.4, p. 255, concerning possible converses from concentration properties to-
wards Poincaré or logarithmic Sobolev inequalities under curvature hypotheses. Ac-
tually, the investigation will provide a complete picture of the links between measure
concentration and isoperimetric-type inequalities for a Markov Triple satisfying the
curvature condition CD(0, ∞). Since we will be dealing at some point with isoperi-
metric enlargements, it is convenient to further examine the metric setting empha-
sized in the introduction of Sect. 8.5. A typical illustrative setting is the case of a
weighted Riemannian manifold with dμ = e−W dμg and ∇∇W ≥ 0 (in particular
log-concave probability measures on Rn ).
We saw in Chap. 4, Proposition 4.4.2, p. 190, and (4.4.6), p. 192, that under a
Poincaré inequality P (C), the probability μ entails exponential concentration in the
sense that any Lipschitz function f : E → R is integrable and satisfies
√
μ f≥ f dμ + r ≤ 3 e−r/ C f Lip , r ≥ 0.
E
This section investigates the converse implication. While it may not be expected
to hold in general, as is clear for example from the characterization of Poincaré
inequalities on the line (cf. Theorem 4.5.1, p. 194), surprisingly a converse does
hold under suitable curvature assumptions.
In order to properly state the result, we say that (E, μ, ) has exponential con-
centration if there are constants C, c > 0 such that for every integrable 1-Lipschitz
function f on E,

μ f≥ f dμ + r ≤ C e−cr , r ≥ 0. (8.7.1)
E
It should be mentioned that the relevant constant in this property is c which controls
the exponential decay. The first constant C is usually easily handled.
The following is then the announced converse implication. Recall the isoperimet-
ric profile Iμ of a measure μ.
Theorem 8.7.1 (Milman’s Theorem I) Let (E, μ, ) be a Markov Triple, with μ a

probability measure, satisfying the curvature condition CD(0, ∞). If (E, μ, ) has
exponential concentration with constants C, c > 0, then
Iμ (v) ≥ c min(v, 1 − v), v ∈ [0, 1], (8.7.2)
where c > 0 only depends on C, c. (In other words, the isoperimetric profile of
(E, μ, ) is bounded from below, up to the constant c , by the isoperimetric profile
of the exponential measure). In particular, (E, μ, ) satisfies a Poincaré inequality
P (C ) where C > 0 only depends on C, c.
By Theorem 4.6.3, p. 203, a log-concave probability measure on Rn satisfies

a Poincaré inequality with a constant depending on the dimension. Theorem 8.7.1
indicates that under exponential concentration, the Poincaré constant only depends
on the concentration parameters.
There is a corresponding statement for Gaussian concentration. Say that (E, μ, )
has Gaussian concentration if there are C, c > 0 such that for every integrable
1-Lipschitz function f : E → R,

f dμ + r ≤ C e−cr , r ≥ 0.
2
μ f≥ (8.7.3)
E
As shown in Proposition 5.4.1, p. 252, under a logarithmic Sobolev inequality

LS(C ), (E, μ, ) has Gaussian concentration with C = 2 and c = 2C
1
. Again, there
is a converse under the curvature condition CD(0, ∞).
Theorem 8.7.2 (Milman’s Theorem II) Let (E, μ, ) be a Markov Triple, with μ
a probability measure, satisfying the curvature condition CD(0, ∞). If (E, μ, )
has Gaussian concentration with constants C, c > 0, then
Iμ ≥ c I,
8.7 From Concentration to Isoperimetry 427
where c > 0 only depends on C, c and I is the Gaussian isoperimetric function

defined at (8.5.7). (In other words, (E, μ, ) satisfies a Gaussian isoperimetric-
type inequality.) In particular, (E, μ, ) satisfies a logarithmic Sobolev inequality
LS(C ) where C > 0 only depends on C, c.
Both Theorems 8.7.1 and 8.7.2 bound the isoperimetric functions Iμ for the
Markov Triple (E, μ, ) from below by the isoperimetric functions of exponential
and (standard) Gaussian measure. It is not too difficult to verify, either directly or via
the subsequent Poincaré or logarithmic Sobolev inequalities, that conversely such
lower bounds ensure the exponential, respectively Gaussian, concentration proper-
ties (8.7.1) and (8.7.3). As such, Theorems 8.7.1 and 8.7.2 provide a complete con-
nection between concentration properties and isoperimetric-type inequalities under
positive curvature bounds.
The schemes of the proofs of Theorems 8.7.1 and 8.7.2 are rather similar. For
simplicity, we only concentrate on the first one on exponential concentration.
Proof of Theorem 8.7.1 By Proposition 8.5.2, the Poincaré inequality is a conse-

quence of the lower bound (8.7.2) on the isoperimetric function. It is therefore
enough to prove the latter.
Let f in Aconst+
0 . For every t ≥ 0, by the usual heat flow interpolation and de
Bruijn’s identity (5.2.3), p. 245,
t
d
f log f dμ − Pt f log Pt f dμ = − Ps f log Ps f dμ ds
E E 0 ds E
t
(Ps f )
= dμ ds.
0 E Ps f
By the reverse local logarithmic Sobolev inequality (5.5.5) of Theorem 5.5.2, p. 259,
whenever 0 < ε ≤ f ≤ 1 where 0 < ε < 1, for every s > 0,
,
1 1
(Ps f ) ≤ log Ps f.
s ε
Hence
,
(Ps f ) 1 1
dμ ≤ log (Ps f ) dμ
E Ps f s ε E
,
1 1
≤ log (f ) dμ
s ε E
where the last inequality

√ √from the fact that, under CD(0, ∞), the strong
follows
gradient bound (Ps f ) ≤ Ps ( (f )) holds (cf. Theorem 3.3.18, p. 163). It fol-
lows from the preceding that for every t > 0 and 0 < ε < 1, and every function f in
Aconst+
0 such that ε ≤ f ≤ 1,
,
1
f log f dμ − Pt f log Pt f dμ ≤ 2 t log (f ) dμ. (8.7.4)
E E ε E

Recall next (5.6.4), p. 267, which tells us that −ψ = − log P2t f is
1
√
2 t
-Lipschitz. (Note that Proposition 8.6.1 may also be used for this task.) By
the exponential concentration hypothesis (8.7.1), for every r ≥ 0,
√
μ(ψ ≤ m − r) ≤ C e−2cr t

where m = E ψdμ. By convexity of the map u → log u2 on (0, 1],

m ≥ log
f2 dμ so that, for every r > 0 and t > 0,
E

2 √
μ ψ ≤ log
− r ≤ C e−2cr t .
E f dμ

This implies that, for every 0 ≤ r ≤ log
2
f dμ
and t > 0,
E
√
μ Pt f ≥ 2 f dμ er ≤ C e−2cr t .
2
E

2
Therefore, whenever 0 < δ ≤ 1 is such that δ ≥ 2 E f dμ er ,

1 1
Pt f log dμ ≥ log Pt f dμ
E Pt f δ {Pt f ≤δ}

1
= f dμ − Pt f dμ log (8.7.5)
E {Pt f >δ} δ
√

1
≥ f dμ − C e−2cr t log .
E δ
2
Let δ = [2(ε + (1 − ε)μ(A))]1/2 er , which we assume to be less than 1. The pre-
vious inequalities (8.7.4) and (8.7.5) extend to D(E). Now, for a closed set A ⊂ E,
apply these to f = max(1A , ε), 0 < ε < 1, actually first to some suitable Lipschitz
approximations fη = max(ε, (1 − η1 d(·, A))+ ) with η → 0, η > 0 (which belong to
D(E) by hypothesis). For this choice of f , (8.7.4) and (8.7.5) (together with (8.5.2))
imply that
,
1 √
−2cr t 1 1
−ε log + (1 − ε)μ(A) − C e log ≤ 2 t log μ+ (A).
ε δ ε
It remains to suitably optimize the various parameters. The simple choices

ε = μ(A)2 and r 2 = 18 log 1ε = 14 log μ(A)
1
satisfy δ ≤ 2μ(A)1/4 ≤ 1, for μ(A) ≤ 16
1
,

and r ≤ log
f dμ . Moreover, for every t > 0 and μ(A) ≤ 16
2 1
E

1 1 μ(A) √ 1 √ √
−2μ(A)2 log + − C e−2cr t log ≤ 4 2 r t μ+ (A).
μ(A) 4 2 16μ(A)
1
As a consequence, for t > 0 well chosen of the order of log μ(A) (for instance satis-
√
fying r t = 2c log μ(A) ), there exists a c > 0 depending only on C, c > 0 such that
1 4C
for every set A with 0 < μ(A) ≤ c ,
c μ(A) ≤ μ+ (A). (8.7.6)
Similar, and actually easier, arguments show that (8.7.6) may be extended to ev-
ery set A with 0 < μ(A) ≤ 12 . Taking the complement yields (8.7.2) and therefore
Theorem 8.7.1.
The notion of capacity has been developed in the second part of the 20th century
in various parts of mathematics, including functional analysis, potential theory, har-
monic analysis and geometric measure theory. A comprehensive account on the
notion of measure-capacity inequalities and the various developments in analysis,
together with bibliographical references, is the monograph [303] by V. G. Maz’ya.
Early developments in the context of geometric measure theory are considered by
H. Federer in [184]. See also [97, 98].
The first co-area formula of Theorem 8.1.2 may be found in [303, Chap. 2].
The topic of Sect. 8.2 linking classical Sobolev inequalities and capacity inequal-
ities is also discussed in [216, 303], where in particular the general Proposition 8.2.1
on capacity inequalities and pairs of Orlicz functions is emphasized. The classical
Faber-Krahn inequality compares the first eigenvalue of the Laplacian on a bounded
open domain in Rn with Dirichlet boundary conditions to that of the ball with the
same volume (cf. e.g. [121, 303]). As alluded to in Remark 8.2.2, such inequalities
may be studied similarly in the context of measure-capacity inequalities to formu-
late equivalently functional inequalities [35, 114, 216, 303].
The corresponding statements for Poincaré and logarithmic Sobolev inequalities
in Sect. 8.3 have been developed more recently in the works of F. Barthe, P. Cattiaux
and C. Roberto [46, 49]. Their results may be understood as a generalization of the
one-dimensional characterizations of B. Muckenhoupt [321] and S. Bobkov and
F. Götze [77] discussed respectively in Chap. 4, Sect. 4.5, p. 193, and Chap. 5,
Sect. 5.4, p. 252.
Generalized Nash inequalities from the point of view of capacity inequalities are
studied in [47, 207, 450]. For the corresponding results for weak Poincaré inequal-
ities, see [45]. Characterizations in dimension one of Latała-Oleszkiewicz inequali-
ties are considered in [49].
Extensions of the method have been studied for further families of functional
inequalities including F -Sobolev inequalities (cf. [46]), weak logarithmic Sobolev
inequalities (cf. [116]) or Lq -logarithmic Sobolev inequalities (cf. [162]).
The co-area formula of Theorem 8.5.1 in Sect. 8.5 is a standard statement
which may be found in numerous references including [97, 122, 184, 303]. See
also [78, 295, 313]. We refer to these references for an account of the historical de-
velopments. Proposition 8.5.2 linking (exponential) isoperimetric-type bounds with
Poincaré inequalities is an observation going back to J. Cheeger [124] in Rieman-
nian geometry (see also [99, 442]). It has been the source of many extensions on
the basis of the same principle (cf. for instance [78, 312]). The monograph [78]
by S. Bobkov and C. Houdré studies in particular a variety of statements along
these lines connecting isoperimetric-type inequalities and functional inequalities,
including those for Gaussian measures of Sect. 8.5.2 (see below). Moreover, the
metric framework of Sect. 8.5 (surface measure, co-area formulas etc.) is care-
fully described there and we refer the reader to it for all the necessary techni-
cal details. General introductions to isoperimetric and geometric inequalities in-
clude [98, 122, 195, 338].
The isoperimetric inequality for Gaussian measures (Sect. 8.5.2) is due to
C. Borell [87] and V. Sudakov and B. Cirel’son [401]. The functional form of
the Gaussian isoperimetric inequality illustrated in Corollary 8.5.4 was introduced
by S. Bobkov [73] who established it first on the two-point space and then in the
limit for the Gaussian measure by the central limit Theorem (following the orig-
inal approach of L. Gross [224] in his proof of the logarithmic Sobolev inequal-
ity, cf. the Notes and References in Chap. 5). On the basis of the Bobkov in-
equality, the local inequalities of Theorem 8.5.3 were established in [38]. Corol-
lary 8.5.5 may be considered as the infinite-dimensional extension of a famous re-
sult of P. Lévy and M. Gromov [220, 281, 316] comparing the isoperimetric profile
of a Riemannian manifold with a strictly positive lower bound on the Ricci curva-
ture to that of the sphere with the same (constant) curvature and dimension (see
also [60, 193, 221, 315]). A purely Markov operator proof of this statement under
the curvature-dimension condition CD(ρ, n), ρ > 0, n < ∞, is yet to be found. Al-
ternative proofs of Corollary 8.5.5 for a probability measure dμ = e−W dx on Rn
with ∇∇W ≥ ρ Id, ρ > 0, have been provided in [103] by mass transportation meth-
ods (more precisely using Theorem 9.3.4, p. 447, in the next chapter), in [75] via
the localization method of [293] and in [318] with a geometric derivation. See [315]
for a recent complete geometric picture of isoperimetric comparison theorems with
families of one-dimensional models under curvature-dimension conditions and di-
ameter bounds.
Section 8.6 is taken from [37]. Proposition 8.6.1 already appeared in [38] (in an
alternative proof of Gaussian isoperimetry).
Theorems 8.7.1 and 8.7.2 of Sect. 8.7 connecting isoperimetric and concentra-
tion properties in spaces with positive curvature are due to E. Milman [312, 314] in
a (weighted) Riemannian manifold setting. His results go far beyond the statements
presented here and cover a large spectrum of isoperimetric regimes and functional
inequalities. Preliminary contributions in this context go back to [38, 99, 273] (see
also [279]). The semigroup proof of Theorems 8.7.1 and 8.7.2 presented here is
taken from [280]. It should be pointed out that, in a (weighted) Riemannian manifold
framework with (extended) positive Ricci curvature, the isoperimetric profile of the
(weighted) Riemannian measure is always concave as established by V. Bayle [52]
(see [312, 314]). Therefore, in this case, the weakest concentration rate actually
implies its comparison with that of the exponential measure (and thus exponential
concentration). Earlier steps in the relationships between (Gaussian) measure con-
centration and logarithmic Sobolev inequalities are due to F.-Y. Wang [428] on the
basis of his Theorem 5.6.1, p. 265 (see also [431], [14, Chap. 7] and [80]).
Chapter 9
Optimal Transportation and Functional
Inequalities
This chapter is a brief investigation of the links between optimal transportation

methods and functional inequalities of Poincaré, logarithmic Sobolev or Sobolev-
type. Optimal transportation tools and ideas have arisen from the geometric analy-
sis of partial differential equations, and the study of gradient flows in Wasserstein
spaces and of interpolation along the geodesics of optimal transport have been par-
ticularly useful in establishing some of the functional and geometric inequalities
recorded in this monograph.
Along these lines, optimal transportation will be used here in particular to present
an alternative approach to sharp Sobolev or Gagliardo-Nirenberg inequalities in
Euclidean space. Transportation cost inequalities comparing relative entropy and
Wasserstein distances between probability measures are further investigated and we
study their relationships to logarithmic Sobolev inequalities, hypercontractivity and
measure concentration.
This chapter is not a complete exposition of optimal transportation (in particular
most proofs of general results are omitted) or of its applications to functional in-
equalities. It only aims at giving a flavor of some of the results and methods in this
context, and mainly concentrates on the links with functional inequalities.
In contrast to the rest of this work, this chapter will be presented in somewhat
specific frameworks adapted to the various results. More precisely, the general the-
ory of optimal transportation, as well as several transportation cost inequalities, may
(and will) be presented in the classical context of topological Polish spaces. On the
other hand, smooth (complete connected, weighted) Riemannian manifolds, often
even Euclidean spaces, form a convenient framework in which we can easily deal
with transportation maps and basic Hamilton-Jacobi equations and their (viscos-
ity) solutions by Hopf-Lax formulas. Recent developments (see Sects. 9.8 and 9.9)
have addressed extensions to general metric measure spaces involving refined non-
smooth analysis. Nevertheless, to better highlight the main results, we restrict here
to the smooth Riemannian setting as soon as these analytic tools are exploited.
Finally, some statements may be given in the (Full) Markov Triple (E, μ, )
framework of this monograph. While the general measurable Markov Triple struc-
ture might be pushed further to cover most of the results in this chapter (see again

DOI 10.1007/978-3-319-00227-9_9,
434 9 Optimal Transportation and Functional Inequalities
the Notes and References for recent progress in this direction), for the sake of clar-
ity, the exposition is restricted to the specific smooth frameworks, indicating when
such extensions are possible.
The first section introduces the basic definitions and general results (without
proofs) of optimal transportation as well as the basic tool of the Brenier map. The
main topic of transportation cost inequalities and first examples, in particular for
Gaussian measures, are discussed in Sect. 9.2. Section 9.3 develops the tool of op-
timal transportation to establish logarithmic Sobolev and Sobolev inequalities (in
Euclidean space) with sharp constants. Non-linear Hamilton-Jacobi equations are
briefly presented in Sect. 9.4, while the subsequent section emphasizes a hypercon-
tractivity property of solutions of Hamilton-Jacobi equations analogous to the one
for linear heat equations. The preceding results are then applied in Sect. 9.6 to in-
vestigate the relationships between (quadratic) transportation cost inequalities and
logarithmic Sobolev inequalities. Section 9.7 investigates contraction properties in
Wasserstein space along the heat semigroup under a curvature condition by means
of commutation between the heat and Hopf-Lax semigroups. Section 9.8 is a very
brief overview of recent developments towards a notion of Ricci curvature lower
bounds based on optimal transportation.
9.1 Optimal Transportation

This section presents some of the basic definitions and fundamental theorems (with-
out proofs) concerning optimal transportation costs and couplings.
9.1.1 Optimal Transportation
Although we will mostly deal with Euclidean or Riemannian spaces in the various
illustrations, in order to develop some of the preliminaries of optimal transportation,
a natural topological framework is that of a Polish space (complete separable met-
ric space) (E, d) equipped with its Borel σ -field F . The product space E × E is
equipped with the product Borel σ -field. P(E) denotes the set of probability mea-
sures on (E, F).
Definition 9.1.1 (Optimal transportation cost) Let c : E × E → [0, ∞] be a lower

semi-continuous function and let μ, ν ∈ P(E). The optimal transportation cost for
transporting μ onto ν with cost c is defined by

Tc (μ, ν) = inf c(x, y)dπ(x, y) ∈ [0, ∞] (9.1.1)
E×E
where the infimum runs over the set of probability measures π ∈ P(E × E) with
respective marginals μ and ν. That is, for all bounded measurable functions u and
9.1 Optimal Transportation 435
v on E,

u(x) + v(y) dπ(x, y) = u dμ + v dν.
E×E E E
The function c is called a cost function and a measure π ∈ P(E × E) with marginals
μ and ν is called a coupling of (μ, ν).
Note that the set over which we take the infimum in (9.1.1) is not empty, the
product measure μ ⊗ ν being a coupling of μ and ν. The first result concerns the
existence of such an optimal coupling.
Proposition 9.1.2 (Optimal coupling) Given μ, ν ∈ P(E), there exists an optimal

coupling π of (μ, ν) such that

Tc (μ, ν) = c dπ.
E×E
The optimal transportation cost admits a powerful dual representation.
Theorem 9.1.3 (Kantorovich’s Theorem) Let c : E × E → [0, ∞] be a cost func-

tion and let μ, ν ∈ P(E). Then

Tc (μ, ν) = sup u dμ − v dν ∈ [0, ∞] (9.1.2)
E E
where the supremum runs over all bounded continuous functions u and v (or in
L1 (μ) and L1 (ν) respectively) satisfying, for all (x, y) ∈ E × E,
u(x) − v(y) ≤ c(x, y). (9.1.3)
If the cost function is given by a distance function d̄(x, y) = c(x, y) on E × E

then Kantorovich’s Theorem takes an another formulation.
Theorem 9.1.4 (Kantorovich-Rubinstein Theorem) Let d̄ be a distance on E,

lower semi-continuous with respect to d (not necessarily the initial distance of
(E, d)). Then

Td̄ (μ, ν) = sup u dμ − u dν
E E
where the supremum runs over all 1-Lipschitz (with respect to d̄) functions u. (The
functions u may be assumed furthermore to be bounded.)
As already indicated by the preceding theorem, in this short exposition of optimal

transportation we will not be interested in the most general cost functions c, and
mainly consider c = d p where d is the metric on E and p ≥ 1 (actually, we even
restrict ourselves below to p = 1 or 2). Denote by Pp (E) the space of probability

measures μ on E with finite p-th moment (i.e. such that E d(x, x0 )p dμ(x) < ∞
for some x0 ∈ E). Therefore, for the choice of c = d p , the optimal transportation
cost gives rise to the so-called Wasserstein distance
1/p
Wp (μ, ν) = inf d(x, y)p dπ(x, y) = Td p (μ, ν)1/p , (9.1.4)
E×E
where as above the infimum runs over all couplings π of (μ, ν) such that
μ, ν ∈ Pp (E). Note, as is easily seen, that if μ ∈ Pp (E) and Wp (μ, ν) < ∞, then
necessarily ν ∈ Pp (E). From a more probabilistic point of view,
1/p
Wp (μ, ν) = inf E d(X, Y )p
where the infimum is over all random variables X and Y with respective laws μ
and ν. On Pp (E), the Wasserstein distance Wp metrizes the weak convergence
topology together with convergence of the respective p-th moments, defining the
corresponding Wasserstein space.
The Kantorovich-Rubinstein Theorem 9.1.4 is of particular interest in two special
cases. First, choose d̄ = d where d is the distance on the Polish space E. This
choice gives rise to the Wasserstein distance W1 . Another choice is the trivial metric
d̄(x, y) = 1x=y , (x, y) ∈ E × E, which yields the total variation distance between
μ and ν,

1 dν
Td̄ (μ, ν) = 1− dμ = μ − νTV = sup μ(A) − ν(A). (9.1.5)
2 E dμ A∈F
9.1.2 The Brenier Map
One main question in optimal transportation is to find an explicit coupling π of

(μ, ν) optimizing (9.1.1) or equivalently, in the dual formulation, to find optimal
functions u and v such that (9.1.2) is an equality. An even more stringent require-
ment, known as Monge’s problem, is to find a (measurable) map T : E → E pushing
forward μ onto ν and such that π = (Id, T )#μ is an optimal coupling. By definition
of the push-forward #,

h(x, y)dπ(x, y) = h x, T (x) dμ(x)
E×E E
for all, say, bounded measurable functions h : E × E → R. The following theorem,

emphasizing the Brenier map, answers this question in Euclidean (or Riemannian)
spaces for the quadratic cost c(x, y) = |x − y|2 , (x, y) ∈ Rn × Rn .
Theorem 9.1.5 (Brenier’s map) Let μ and ν be two probability measures on the
Borel sets of Rn with μ absolutely continuous with respect to the Lebesgue measure
9.1 Optimal Transportation 437
and such that W2 (μ, ν) < ∞. Then there exists a convex function φ : Rn → R such
that T = ∇φ maps μ onto ν (denoted T #μ = ν), where, here, ∇φ is considered as
a map Rn → Rn . In other words, for every bounded measurable function h on Rn ,

h dν = h(∇φ)dμ. (9.1.6)
Rn Rn
Moreover, T = ∇φ is uniquely determined μ-almost everywhere and if μ and ν have

a finite second moment, T = ∇φ is optimal in the sense that

W2 (μ, ν)2 = x − T (x)2 dμ(x). (9.1.7)
Rn
Brenier’s Theorem may be suitably extended in a Riemannian manifold for the

cost given by the square of the Riemannian distance. The Brenier map T = ∇φ then
has to be replaced by the geodesic exponential map generated by the gradient of a
function satisfying a suitable convexity property (which would be |x|2 − φ in the
2
Euclidean case).
If μ and ν have densities f and g with respect to the Lebesgue measure on Rn ,
according to (9.1.6), for every bounded measurable map h : Rn → R,

h(y)g(y)dy = h ∇φ(x) f (x)dx.
Rn Rn
Whenever the change of variable y = ∇φ(x) is licit, the preceding leads to the so-
called Monge-Ampère equation
f = g(∇φ) det(∇∇φ) (9.1.8)
(where ∇∇φ is the matrix of the second derivatives of φ). However, the map
T = ∇φ exists only almost everywhere, and may not be differentiable in any usual
sense. The Monge-Ampère equation may then be understood in a generalized sense.
Actually it can be proved that φ is locally Lipschitz on the interior of its domain and
that the Monge-Ampère equation is valid f dx-almost everywhere with ∇∇φ being
understood as the Hessian of φ in the sense of Aleksandrov (the absolutely contin-
uous part of the distributional Hessian of φ). Alternatively, suitable assumptions on
f and g ensure its validity as in the next statement.
In the following, we say that a function f defined on an open set O in Rn belongs
to C k,α (O) for some k ∈ N if f ∈ C k (O) and all its derivatives up to order k are
locally Hölder continuous with exponent α ∈ (0, 1).
Theorem 9.1.6 (Caffarelli’s regularity properties) Let μ and ν be probability mea-

sures on the Borel set of Rn with respective densities f and g defined on bounded
open sets Oμ and Oν in Rn such that ε ≤ f, g ≤ 1ε for some ε > 0. Then, if
f ∈ C 1,α (Oμ ) and g ∈ C 1,α (Oν ) with α ∈ (0, 1), and if Oμ is a convex set, the
Brenier map φ is C 2,α (Oμ ) and the Monge-Ampère equation (9.1.8) holds on Oμ .
Moreover, if f and g are C ∞ , then φ is also C ∞ .
9.2 Transportation Cost Inequalities
This section addresses transportation cost inequalities which are built on the cost
functions discussed in the preceding section. More precisely, transportation cost in-
equalities compare a transportation cost distance to a fluctuation distance expressed
by (relative) entropy.
The first example of a transportation cost inequality is the Pinsker-Csizsár-
Kullback inequality (5.2.2), p. 244,
1
μ − ν2TV ≤ H(ν | μ),
2
where we recall that, for probability measures μ, ν on a metric space (E, d),

dν
H(ν | μ) = log dν
E dμ
if ν μ and H(ν | μ) = ∞ if not, is the relative entropy of ν with respect to

μ (cf. (5.2.1), p. 244). Now, as discussed in (9.1.5), the total variation distance
may be interpreted as a Wasserstein distance W1 with respect to the trivial metric
d̄(x, y) = 1x=y on E × E.
On the basis of this example, one may more generally consider transportation
cost inequalities T1 (C) for a given μ ∈ P1 (E) as
W1 (μ, ν)2 ≤ 2C H(ν | μ) (9.2.1)
for some C > 0 and all ν ∈ P(E). (Note that under this inequality, H(ν | μ) = ∞
whenever ν ∈ / P1 (E).) This family of transportation cost inequalities will in partic-
ular be studied below in connection with concentration inequalities.
However, the connections between transportation cost inequalities, partial differ-
ential equations and functional inequalities as discussed in this work are actually
of most interest for a quadratic cost and the associated W2 Wasserstein distance
(with respect to the Euclidean or Riemannian structures). The first basic example in
this setting is the following quadratic transportation cost inequality for the Gaussian
measure in Rn . Recall that P2 (Rn ) denotes the set of probability measures on Rn
with a second moment.
Theorem 9.2.1 (Talagrand’s inequality) Let dμ(x) = (2π)−n/2 e−|x| /2 dx be the

2
standard Gaussian measure on the Borel sets of Rn . Then, for any ν ∈ P(Rn ),
W2 (μ, ν)2 ≤ 2 H(ν | μ).
It is instructive to first present a one-dimensional proof this theorem.

Proof Let n = 1. Let f ≥ 0 be such that R f dμ = 1 and W2 (μ, ν) < ∞ where

dν = f dμ. In this one-dimensional setting, the idea is to describe the explicit map
9.2 Transportation Cost Inequalities 439
pushing forward μ onto ν. To this end, by standard approximation arguments, as-

sume that f > 0 everywhere. We may therefore uniquely define T : R → R by

ν −∞, T (x) = μ (−∞, x] , x ∈ R,
so that ν is the image measure of μ under the strictly increasing differentiable

map T . Since then, for every bounded measurable h : R → R,

−y 2 /2

dy = h T (x) e−x /2 dx,
2
h(y)f (y) e
R R
by the change of variables formula y = T (x),

f T (x) e−T (x) /2 T (x) = e−x /2 ,
2 2
x ∈ R.
This identity is the Monge-Ampère equation (9.1.8) in this case. Hence, taking log-
arithms, for every x,
1 x2
log f T (x) + log T (x) − T (x)2 = − .
2 2
Integrating with respect to μ, and using that T #μ = ν,

1
log f dν = T (x)2 − x 2 dμ − log T dμ.
R 2 R R
Integrating by parts with

respect to the
Gaussian measure μ classically yields
(cf. (2.7.7), p. 107) that R x(T − x)dμ = R (T − 1)dμ so that

1 2
H(ν | μ) = log f dν = x − T (x) dμ + T − 1 − log T dμ
R 2 R R

1 2
≥ x − T (x) dμ
2 R
where we used that r − 1 − log r ≥ 0, r > 0. Now, since ν is the image of μ under
T , the image measure π of μ under the map x → (x, T (x)) has marginals μ and ν
respectively so that

x − T (x)2 dμ = |x − y|2 dπ ≥ W2 (μ, ν)2 .
R R×R
The theorem is therefore established (in the one-dimensional case).
The extension to arbitrary dimension n raises interesting issues, and different

options. The first option is to tensorize the Talagrand inequality of Theorem 9.2.1 as
was done in the earlier chapters on Poincaré and logarithmic Sobolev inequalities.
This may indeed be performed, either directly or on a dual formulation as a con-
sequence of the Kantorovich duality. We follow this route below, before studying
another option in the next section. To suitably present the arguments, let us intro-
duce the general definition of a quadratic transportation cost inequality (covering

Theorem 9.2.1 for Gaussian measures). As before, (E, d) denotes a Polish space.
Recall the space P2 (E) of probability measures on the Borel sets of E with a finite
second moment.
Definition 9.2.2 (Quadratic transportation cost inequality) We say that μ ∈ P2 (E)

satisfies a quadratic transportation cost inequality T2 (C) with constant C > 0 if for
every ν ∈ P(E),
W2 (μ, ν)2 ≤ 2C H(ν | μ).
Talagrand’s inequality of Theorem 9.2.1 tells us that the standard Gaussian mea-
sure on Rn satisfies T2 (1) (justifying again the choice in the normalization of the
constant C).
The following equivalent dual formulation of Definition 9.2.2 will turn out to
be important and useful. With this task in mind, we introduce, for any continuous
function f : E → R, the infimum-convolutions
" #
1
Qt f (x) = Qt (f )(x) = inf f (y) + d(x, y)2 , t > 0, x ∈ E. (9.2.2)
y∈E 2t
Note the homogeneity Qt f = 1t Q1 (tf ), t > 0. The operators Qt , t > 0, will be

identified in Sect. 9.4 as the fundamental solutions of Hamilton-Jacobi equations
(in smooth settings). Their relevance at this point is expressed by the Kantorovich
duality Theorem 9.1.3 which indicates that

1
W2 (μ, ν) = sup
2
Q1 f dν − f dμ (9.2.3)
2 E E
where the supremum runs over all bounded continuous functions f : E → R.
Proposition 9.2.3 (Dual formulation of the quadratic transportation cost inequality)

A probability measure μ in P2 (E) satisfies the quadratic transportation cost in-
equality T2 (C) for some C > 0 if and only if for every bounded continuous function
f : E → R,

eQC f dμ ≤ e E f dμ . (9.2.4)
E
dν
Proof We first prove (9.2.4) from T2 (C). According to (9.2.3), if dμ = g, the
quadratic transportation cost inequality T2 (C) expresses that

Q1 f g dμ − f dμ ≤ C H(ν | μ) = C g log g dμ.
E E E
Choosing
e(Q1 f )/C
g=
(Q1 f )/C dμ
Ee
9.2 Transportation Cost Inequalities 441
in the latter shows that

1
log e(Q1 f )/C dμ ≤ f dμ.
E C E
f
The conclusion follows since by homogeneity C1 Q1 f = QC ( C ). Conversely, (9.2.4)

h
implies that E e dμ ≤ 1 with h = QC f − E f dμ. If g = dμ where ν ∈ P(E), the

dν
variational formula (5.1.3), p. 236, yields that

QC f g dμ − f dμ = gh dμ ≤ g log g dμ = H(ν | μ).
E E E E
f
Again by means of (9.2.3), the T2 (C) inequality follows after changing f into C.
Before turning to the tensorization issues, note that a quadratic transportation cost
inequality T2 (C) (for μ ∈ P2 (E)) implies an inequality T1 (C) of the form (9.2.1),
with the same constant C > 0. Moreover, following the scheme of proof of the above
Proposition 9.2.3, such an inequality T1 (C) may be translated to the equivalent

2
esf dμ ≤ es E f dμ+Cs /2 (9.2.5)
E
for every s ∈ R and every (bounded) 1-Lipschiz function f : E → R. Such T1 in-

equalities appear therefore as consequences of logarithmic Sobolev inequalities by
the Herbst argument of Proposition 5.4.1, p. 252. In particular, they equivalently de-
scribe Gaussian concentration properties in the sense of (8.7.3), p. 426. Indeed, as
already illustrated in Sect. 5.4, p. 252, (9.2.5) implies by Markov’s inequality that

f dμ + r ≤ e−r /2C , r ≥ 0,
2
μ f≥ (9.2.6)
E
for every integrable 1-Lipschitz function f : E → R. After integration, these con-

centration properties are equivalent to transportation cost inequalities T1 in the
form (9.2.5), up to the numerical constants.
On the basis of the characterization of Proposition 9.2.3, the stability of quadratic
transportation cost inequalities under products is easily obtained. It may be worth-
while to observe that T1 inequalities, being equivalent to (9.2.5), do not tensorize
independently of the dimension of the product.
Proposition 9.2.4 (Stability under product) Let μ1 and μ2 be probability measures

on metric spaces (E1 , d1 ) and (E2 , d2 ) satisfying respectively T2 (C1 ) and T2 (C2 )
for constants C1 > 0 and C2 > 0. Then μ1 ⊗ μ2 on (E1 × E2 , d1 ⊕ d2 ) where
d1 ⊕ d2 = (d12 + d22 )1/2 satisfies T2 (C) with C = max(C1 , C2 ).
Proof We work accordingly with the dual formulation of Proposition 9.2.3. Let
f : E1 × E2 → R be bounded and continuous. For C = max(C1 , C2 ) and
(x1 , x2 ) ∈ E1 × E2 ,
QC f (x1 , x2 )
" #
1 1
≤ inf f (y1 , y2 ) + d1 (x1 , y1 ) +
2 2
d2 (x2 , y2 ) .
(y1 ,y2 )∈E1 ×E2 2C1 2C2
For every x2 ∈ E2 , define z1 ∈ E1 → Q C2 f (z1 , x2 ) as

" #
1
QC2 f (z1 , x2 ) = QC2 f (z1 , ·)(x2 ) = inf f (z1 , y2 ) + d2 (x2 , y2 )2
y2 ∈E2 2C2
so that

QC f (x1 , x2 ) ≤ QC1 QC2 f (·, x2 ) (x1 ).
By the dual transportation cost inequality (9.2.4) along the x1 variable,

e QC f (x1 ,x2 )
dμ1 (x1 ) ≤ eQC1 (QC2 (f (·,x2 ))(x1 ) dμ1 (x1 )
E1 E1

≤ exp
QC2 f (x1 , x2 )dμ1 (x1 ) .
E1
C2 is an infimum,
Now, since Q

QC2 f (x1 , x2 )dμ1 (x1 ) ≤ QC2 f (x1 , ·)dμ1 (x1 ) (x2 ),
E1 E1
so that integrating in dμ2 (x2 ) and applying Proposition 9.2.3 again, this time along
the variable x2 , yields the conclusion.
Together with Proposition 9.2.4, we may then conclude the proof of Talagrand’s
inequality of Theorem 9.2.1 in any dimension. Now, another way to obtain this
theorem is to try to perform a transportation proof directly in dimension n with
the help of the Brenier transportation map. This is the approach followed in the
next section, which will prove useful not only for transportation cost inequalities
but also for other functional inequalities such as logarithmic Sobolev and Sobolev
inequalities.
9.3 Transportation Proofs of Functional Inequalities

As announced, this section develops the mass transportation method in order to
reach the preceding transportation cost inequality for Gaussian measures as well
as, in an alternative way, some of the functional inequalities considered so far such
as logarithmic Sobolev and Sobolev inequalities (in Euclidean or Riemannian set-
tings). The key idea is the use of the Brenier map T = ∇φ and of the associated
9.3 Transportation Proofs of Functional Inequalities 443
Monge-Ampère equation (9.1.8) together with the convexity properties of φ. In par-

ticular, log-concavity of the map Q → det(Q) or concavity of Q → det(Q)1/n on
the set of symmetric positive n × n matrices Q will provide the technical step to-
wards the various conclusions. The precise arguments usually require some refined
regularity analysis, which we do not necessarily enter into here. To some extent,
these may be handled either by Theorem 9.1.6 or by using the Hessian in the sense
of Aleksandrov.
9.3.1 HWI Inequality
This section introduces a new inequality which encompasses several inequalities

of interest such as logarithmic Sobolev and transportation cost inequalities. The
following statement is a general formulation from which the various cases will be
easily drawn.
For simplicity we deal here with measures on Rn , but the methods and results
may be extended to a (weighted) Riemannian setting. Let dμ = e−W dx be a proba-
bility measure on the Borel sets of Rn where W : Rn → R is a smooth (of class C 2 )
potential such that ∇∇W ≥ ρ Id as symmetric matrices for some ρ ∈ R (equiva-
2
lently, W (x) − ρ|x| −W dx
2 is convex). In other words, R with the measure dμ = e
n
and the standard carré du champ operator (f ) = |∇f | of the diffusion operator
2
L = − ∇W · ∇ invariant with respect to μ is a Markov Triple (as a weighted Rie-

mannian manifold) of curvature CD(ρ, ∞). The standard Gaussian measure corre-
sponds as usual to ρ = 1. The following theorem is established in the same way as
Theorem 9.2.1 using the Brenier map in place of the explicit one-dimensional trans-
port. If f : Rn → R+ is measurable, for simplicity we denote by f μ the measure
with density f with respect to μ.
Theorem 9.3.1 Let dμ = e−W dx be a probability measure on the Borel sets of Rn

where W : Rn → R is a smooth potential such that ∇∇W ≥ ρ Id for some ρ ∈ R.
Furthermore, let f, g : Rn → R+ be probability densities with respect to μ of class
C 1 and compactly supported. If T (x) = x + ∇θ (x) is the Brenier map pushing f μ
onto gμ, then

ρ
H(gμ | μ) ≥ H(f μ | μ) + ∇f · ∇θ dμ + W2 (f μ, gμ)2 .
Rn 2
Proof We only sketch the argument, without justifying several of the regularity is-
sues involved in the proof. Since T pushes f μ onto gμ, the corresponding Monge-
Ampère equation (9.1.8) reads

f (x)e−W (x) = g T (x) e−W (x+∇θ(x)) det Id +∇∇θ (x) , x ∈ Rn
(under the suitable regularity assumptions or in a generalized sense). As in the proof

of Theorem 9.2.1, taking logarithms,

log g T (x) = log f (x) + W x + ∇θ (x) − W (x) − log det Id +∇∇θ (x) .
Note that the matrix Id +∇∇θ is positive at any point since the Brenier map φ is
convex. Hence

log det Id +∇∇θ (x) ≤ θ (x).
Furthermore, by the convexity assumption on W , for every x ∈ Rn ,
ρ 2
W x + ∇θ (x) − W (x) ≥ ∇W (x) · ∇θ (x) + ∇θ (x) .
2
Therefore,
ρ 2
log g T (x) ≥ log f (x) + ∇W (x) · ∇θ (x) − θ (x) + ∇θ (x) .
2
After integration with respect to f μ,

g log g dμ ≥ f log f dμ
Rn Rn

ρ
− [θ − ∇W · ∇θ ]f dμ + W2 (f μ, gμ)2
Rn 2
where
we used that T is the Brenier map sending f μ onto gμ so that, by (9.1.7),

R n |∇θ |2 f dμ = W2 (f μ, gμ)2 . Finally, by integration by parts for the operator
L = − ∇W · ∇ with invariant measure μ,

[θ − ∇W · ∇θ ]f dμ = − ∇θ · ∇f dμ
Rn Rn
which yields the claim. The proof of Theorem 9.3.1 is complete.
Theorem 9.3.1 covers a variety of interesting cases according to the respective

choices of f and g. First choosing f so that it approaches the constant function
1 smoothly, we obtain a generalization of Talagrand’s inequality (Theorem 9.2.1)
under the curvature condition CD(ρ, ∞) for some ρ > 0.

Then μ satisfies the quadratic transportation cost inequality T2 ( ρ1 ).
Another instance of interest is the case where g approaches 1. We then end up

with

ρ
H(f μ | μ) ≤ − ∇f · ∇θ dμ − W2 (f μ, μ)2 .
Rn 2

By the Cauchy-Schwarz inequality, − Rn ∇f · ∇θ dμ is bounded from above by

1/2 1/2
|∇f |2
|∇θ |2 f dμ dμ = W2 (f μ, μ) Iμ (f )1/2
Rn Rn f
where

|∇f |2
Iμ (f ) = dμ
Rn f
is the Fisher information (5.1.6), p. 237, of f (or f μ) with respect to μ. (Recall that
according to Remark 5.1.2, p. 238, this Fisher information Iμ (f ) has to be under-
stood as a suitable limit of strictly positive functions decreasing to f .) The resulting
inequality is known as the HWI inequality (involving entropy H, Wasserstein dis-
tance W and Fisher information Iμ ) and is exhibited in the next corollary.
Corollary 9.3.3 (HWI inequality) Let dμ = e−W dx be a probability measure on

the Borel sets of Rn with a finite second moment where W : Rn → R is a smooth
potential such that ∇∇W ≥ ρ Id for some ρ ∈ R. For every probability density
f : Rn → R+ with respect to μ such that W2 (f μ, μ) < ∞,
ρ
H(f μ | μ) ≤ W2 (f μ, μ) Iμ (f )1/2 − W2 (f μ, μ)2 . (9.3.1)
2
The HWI inequality actually admits an alternative proof which relies on the
semigroup tools developed in this work, allowing in particular with the material
presented in Sect. 8.6, p. 421, for the extension to the setting of a (Full) Markov
Triple (E, μ, ) under the curvature condition CD(ρ, ∞). We briefly outline the
argument keeping for simplicity the Euclidean notation, all the arguments holding
true in the same way in a Markov Triple.
Alternative proof of Corollary 9.3.3 Denote by P = (Pt )t≥0 the Markov semigroup
with generator L = − ∇W · ∇. Let T
> 0 and let f be a smooth bounded and
strictly positive function on Rn such that Rn f dμ = 1. Following de Bruijn’s iden-
tity (5.7.3), p. 269,
T
Entμ (f ) = Iμ (Pt f )dt + Entμ (PT f ).
0
Since ∇∇W ≥ ρ Id, and thus the CD(ρ, ∞) criterion holds, it follows from the
exponential decay (5.7.4), p. 269, of Fisher information that
Entμ (f ) ≤ α(T ) Iμ (f ) + Entμ (PT f ) (9.3.2)

−2ρT
where α(T ) = 1−e2ρ (= T if ρ = 0). The idea is now to control Entμ (PT f ) by
a mass transportation factor with the help of the log-Harnack inequality (5.6.2),
p. 266. The latter applied to PT f yields, for x, y ∈ Rn ,
|x − y|2
PT (log PT f )(x) ≤ log P2T f (y) +
2β(T )
where β(T ) = e ρ −1 (= 2T if ρ = 0). In other words, taking the infimum over

2ρT
y ∈ Rn , using the infimum-convolution notation (9.2.2),
PT (log PT f ) ≤ Qβ(T ) (log P2T f ).
Therefore, by time reversibility,

Entμ (PT f ) = f PT (log PT f )dμ
Rn

≤ f Qβ(T ) (log P2T f )dμ
Rn

1
≤ f Q1 β(T ) log P2T f dμ.
β(T ) Rn

Since furthermore, by Jensen’s inequality, Rn log P2T f dμ ≤ log( Rn P2T f dμ) = 0,

the Kantorovich duality in the form (9.2.3) ensures that
1
Entμ (PT f ) ≤ W2 (f μ, μ)2 .
2β(T )
) = 2α(T ) − ρ, combining this estimate with (9.3.2) and optimizing in

1 1
Since β(T
T > 0 yields the conclusion.
The HWI inequality is a bridge between quadratic transportation cost inequali-

ties and logarithmic Sobolev inequalities. When ρ > 0, using the fact that for real
2
s2
numbers rs ≤ ρr2 + 2ρ , the HWI inequality implies that
1
H(f μ | μ) ≤ Iμ (f ),
2ρ
which is the logarithmic Sobolev inequality for the measure dμ = e−W dx (as de-
duced in Corollary 5.7.2, p. 268, from the curvature condition CD(ρ, ∞) for some
ρ > 0). When only ρ ≥ 0, the quadratic transportation cost inequality T2 (C) still im-
plies a logarithmic Sobolev inequality, but with a weaker constant, namely LS(4C)
(the implication actually still holds as soon as ρ > − C1 ). With a less precise depen-
dence on the constants, this implication actually holds if μ only satisfies the T1 (C)
transportation cost inequality (9.2.1) as a consequence of Theorem 8.7.2, p. 426. In
general, however, a quadratic transportation cost inequality is not enough to ensure
the validity of a logarithmic Sobolev inequality (see Sect. 9.6 below and the Notes
and References).
To conclude this section, and from a somewhat different perspective, it is worth

mentioning (without proof) the following result which easily allows us to deduce
quadratic transportation cost, logarithmic Sobolev and even isoperimetric-type in-
equalities from their counterparts for Gaussian measures.
Theorem 9.3.4 (Caffarelli’s contraction Theorem) Let dμ = e−W dx be a prob-

ability measure on the Borel sets of Rn where W : Rn → R is a smooth poten-
tial such that ∇∇W ≥ ρ Id for some ρ > 0. Then the Brenier transportation map
T = ∇φ : Rn → Rn pushing the standard Gaussian measure on Rn onto μ is Lips-
chitz, with Lipschitz coefficient √1ρ .
9.3.2 Sobolev Inequality in Rn
The transportation method developed in the preceding section to recover in partic-

ular logarithmic Sobolev inequalities may be used similarly to directly obtain the
standard Sobolev inequality on Rn with optimal constants, thus providing another
approach. One further interesting feature of the method is that it allows us to simi-
larly deal with arbitrary norms on Rn (with optimal constants), something which is
not allowed by the semigroup approach restricted to Euclidean metrics. We there-
fore present here the result and the transportation proof of Sobolev inequalities in
this framework. Although not discussed in detail here, the transportation proof leads
to a unique description of the extremal functions.
Consider therefore an arbitrary norm · on Rn . Denote by · ∗ its dual norm
defined by
x∗ = sup x · y, x ∈ Rn ,
y∈Rn ,y≤1
where x · y is as usual the Euclidean scalar product in Rn . The proof below will
make use of Young’s inequality
p∗
xp y∗
x·y ≤ + (9.3.3)
p p∗
for x, y ∈ Rn , 1 ≤ p ≤ ∞, 1
p + 1
p∗ = 1, and of Hölder’s inequality for (suitable)
functions F, G : Rn → Rn ,
1/p 1/p∗
p∗
F · G dx ≤ F dx
p
G∗ dx . (9.3.4)
Rn Rn Rn
The next statement is the announced version of the Sobolev inequality (6.1.1),
p. 278, on Rn (n > 2) with respect to such an arbitrary norm. We refer to Re-
mark 6.9.5, p. 319, for the class of functions satisfying the Sobolev inequality in
Euclidean space.
Theorem 9.3.5 (Sobolev inequality) Let n > 2 and p = n−2 2n

. For every smooth
function f : Rn → R with f ∈ Lp (dx) and ∇f ∗ ∈ L2 (dx),

f p ≤ Cn ∇f ∗ 2
where the norms are understood to be with respect to the Lebesgue measure.
The sharp constant Cn > 0 is achieved on the extremal functions fσ,b,x0 (x) =
(σ 2 + b x − x0 2 )(2−n)/2 , x ∈ Rn , σ > 0, b > 0 and x0 ∈ Rn .
Proof Before starting the proof, note that the regularity properties of Theorem 9.1.6
will not be enough since the various densities should be smooth and compactly
supported and not bounded from below by a strictly positive constant. A careful
proof requires the use of distributional gradients, Laplacians and Hessians in the
sense of Aleksandrov. As in the previous sub-section, this will be mostly implicit
below and the proof therefore only emphasizes the main ideas and arguments.
It is enough to establish the theorem for positive functions. Let u, v be smooth
compactly supported probability densities with respect to the Lebesgue measure.
According to Theorem 9.1.5, there is a convex function φ : Rn → R such that
T = ∇φ, the Brenier map, is the optimal map from udx to vdx. The associated
Monge-Ampère equation (9.1.8) reads
u = v(∇φ) det(∇∇φ).
Hence,
v −1/n (∇φ) = u−1/n det(∇∇φ)1/n .
The arithmetic-geometric inequality (for the determinant of symmetric positive ma-
trices) indicates that
1
det(∇∇φ)1/n ≤ φ
n
since φ is the trace of ∇∇φ (which is at the heart of the CD(0, n) condition).
Therefore, after integration with respect to the measure udx,

−1/n 1
v (∇φ)u dx ≤ u1−(1/n) φ dx. (9.3.5)
Rn n Rn
Since T = ∇φ maps udx onto vdx,

−1/n
v (∇φ)u dx = v 1−(1/n) dx.
Rn Rn
Integrating the right-hand side of (9.3.5) by parts therefore yields that

1
v 1−(1/n) dx ≤ − ∇ u1−(1/n) · ∇φ dx. (9.3.6)
Rn n Rn
Note that since u and v are compactly supported, ∇φ is bounded.
Let now f and g be two positive smooth compactly supported functions on Rn

such that f p = u and g p = v (p = n−22n
) are probability densities. Then (9.3.6) reads
as

2(n − 1)
g 2(n−1)/(n−2)
dx ≤ − f n/(n−2) ∇f · ∇φ dx.
Rn n(n − 2) Rn
The Cauchy-Schwarz inequality (cf. (9.3.4)) implies that
1/2 1/2
2(n − 1)
g 2(n−1)/(n−2)
dx ≤ ∇f 2∗ dx f ∇φ dx
p 2
Rn n(n − 2) Rn Rn
and thus, by construction of T = ∇φ,

1/2 1/2
2(n − 1)
g 2(n−1)/(n−2) dx ≤ ∇f 2∗ dx g p x2 dx .
Rn n(n − 2) Rn Rn
Hence, removing the normalization condition on f , it follows after a density argu-

ment that for every positive smooth function f and every positive function g such
that g p is a probability density (both compactly supported),
1/2
2(n − 1)
f p ≤ ∇f 2∗ dx
n(n − 2) Rn
1/2 −1 (9.3.7)
× g x dx
p 2
g 2(n−1)/(n−2)
dx .
Rn Rn
Now, it may be observed that whenever f = g = fσ,b,x0 for suitable choices of

σ and b in order that g p is a probability density, ∇φ = x. Then, even though these
densities are not compactly supported, the preceding computations are actually still
valid for these specific functions and (9.3.5) as well as the subsequent Cauchy-
Schwarz inequality are equalities in this case. Therefore, the choice of g = fσ,b,x0
in the preceding yields the Sobolev inequality with its optimal constant. Moreover,
following this reasoning, it may be shown that the functions fσ,b,x0 are the only
extremal functions (see the Notes and References).
9.3.3 Sobolev Trace Inequality in Rn
The same strategy may be used to derive an (optimal) Sobolev trace inequality on the
half-space, illustrating the power of the transportation argument. Set below Rn+ =
Rn−1 × [0, ∞) and ∂Rn+ = Rn−1 × {0} and consider as before · a norm on Rn .
Denote a generic point x ∈ Rn+ by x = (y, s) so that y ∈ Rn−1 is identified with
(y, 0) ∈ ∂R+n.
Theorem 9.3.6 (Sobolev trace inequality) Let n > 2 and p = n−2 2n

, p̃ = 2(n−1)
n−2 .
For every smooth function f : R+ → R with f ∈ L (Rn+ , dx) and
n p
∇f ∗ ∈ L2 (Rn+ , dx),

f Lp̃ (∂Rn+ ) ≤ Cn ∇f ∗ L2 (Rn ) . (9.3.8)
+
The sharp constant Cn > 0 is attained for the extremal functions h(x) = x − e2−n ,
x ∈ Rn+ , with e = (0, . . . , 0, −1) ∈ Rn .
Proof The starting point is similar to the proof of Theorem 9.3.5 for the Sobolev
inequality. We again make use of the Monge-Ampère equation (9.1.8) in the sense
of Aleksandrov without any further notice.
Let u and v be two smooth probability densities with respect to the Lebesgue
measure on Rn+ with compact support, and denote by T = ∇φ the Brenier map
from udx onto vdx. As in (9.3.5), integrating with respect to the measure udx,

1
v 1−(1/n)
dx ≤ u1−(1/n) φ dx. (9.3.9)
Rn+ n Rn+
To obtain an inequality at the boundary, let ψ = φ − e · x. Since ψ = φ, an

integration by parts on the right-hand side yields

1
v 1−(1/n) dx ≤ − ∇ u1−(1/n) · ∇(φ − e · x)dx
Rn+ n Rn+
(9.3.10)
1
+ u1−(1/n) (∇φ − e) · e dx.
n ∂Rn+
Again the integration by parts is justified since u and v are compactly supported,
and the map φ may be extended to the full space as a convex function so that ∇φ
has a bounded variation at the boundary. Since T = ∇φ is the optimal map between
udx and vdx, ∇φ ∈ Rn+ and then ∇φ · e ≤ 0 on ∂Rn+ , so that the previous inequality
yields

1
v 1−(1/n)
dx ≤ − ∇ u1−(1/n) · ∇(φ − e · x)dx
R+n n R+ n
(9.3.11)
1
− u 1−(1/n)
dx.
n ∂Rn+
Let now f and g be two positive smooth compactly supported functions such
that u = f p and v = g p (p = n−22n
) are probability densities. The Cauchy-Schwarz
inequality and the properties of the transportation map T = ∇φ then imply, as in the
proof of Theorem 9.3.5, that
1/2

f p̃ dx ≤ p̃ g p x − e2 dx ∇f ∗ −n g p̃ dx.
L2 (Rn+ )
∂Rn+ Rn+ Rn+
9.4 Hamilton-Jacobi Equations 451
Therefore, removing the

normalization on f , for any pair of smooth positive func-
tions f and g such that Rn g p dx = 1, and with κ = ∇f ∗ L2 (Rn+ ) /f Lp (Rn+ ) ,
+
p̃
f Lp̃ (∂Rn )

+
≤ κ −p̃ κA(g) − B(g)
∇f ∗ p̃ 2 n
L (R ) +
where
1/2
A(g) = p̃ g x − e dx
p 2
and B(g) = n g p̃ dx.
Rn+ Rn+
B(g)
The worse case is when κ = 2(n−1)
n A(g) , for which it follows that for all positive g
p
such that g is a probability density,

f p̃ n ≤ Cn (g) ∇f ∗ 2 n
L (∂R+ ) L (R+ )
where Cn (g) > 0. The optimal Sobolev trace inequality follows with the choice of
g = ch for a suitable normalization constant c > 0 since when f = g = ch, all the
previous inequalities turn into equalities. As in the proof of Theorem 9.3.5, even if h
is not compactly supported, all the arguments are justified since in this case ∇φ = x.
The proof of Theorem 9.3.6 is then easily completed along these lines.
Remark 9.3.7 The preceding transportation method may be developed similarly to

reach the general family of optimal Sobolev inequalities in Rn with Lq -norms of
the gradient, 1 ≤ q < n, and arbitrary norms on Rn ,

f p ≤ Cn,q ∇f ∗ q
qn
where now p = n−q . Similarly, the Sobolev trace inequality

f Lp̃ (∂Rn+ ) ≤ Cn,q ∇f ∗ Lq (Rn )
+
now with p̃ = q(n−1)

n−q is established in the same way. The strategy also includes the
family of Gagliardo-Nirenberg inequalities of Theorem 6.10.4, p. 326, again with
Lq -norms of gradients and arbitrary norms on Rn .
9.4 Hamilton-Jacobi Equations

This section presents the basic material on Hamilton-Jacobi equations and their re-
lationships to optimal transport at the level of the dual formulation expressed by
Theorem 9.1.3. The main aspects may be presented in the context of a (smooth
complete connected) Riemannian manifold (M, g), Rn for example, with d the Rie-
mannian or Euclidean distance.
9.4.1 Hamilton-Jacobi Equations and Hopf-Lax Formulas
In its general form, a Hamilton-Jacobi equation is a non-linear partial differential

equation of the form
∂t u + H (t, x, u, ∇u) = 0, (9.4.1)
u = u(t, x), t > 0, x ∈ M, where the function H is called a Hamiltonian. We will
only be interested here in the very particular case when H : R+ → R+ only depends
on the length |∇u| of the gradient of u. That is, given the initial condition in the form
of a continuous and bounded function f on M, the equation takes the form
-
∂t u + H (|∇u|) = 0 in (0, ∞) × M,
(9.4.2)
u=f on {t = 0} × M.
Existence and properties of the solutions are summarized in the next proposition.
Proposition 9.4.1 (Hopf-Lax formula) Assume that H : R+ → R+ is a C 2 increas-

ing and convex function satisfying H (0) = 0 and
H (r)
lim = ∞.
r→∞ r
Let f : M → R be bounded and continuous and, for t > 0 and x ∈ M, set
-
u(t, x) = infy∈M {f (y) + tH ∗ ( d(x,y)
t )} for (t, x) ∈ (0, ∞) × M,
(9.4.3)
u(0, x) = f (x) for x ∈ M,
where H ∗ is the Legendre-Fenchel transform of H , H ∗ (r) = sups∈R+ [rs − H (s)],

r ∈ R+ . Then for all t > 0, x → u(t, x) is Lipschitz and, almost everywhere in the
variable x ∈ M, u satisfies the Hamilton-Jacobi equation (9.4.2) starting from f .
If the initial condition f is a C 2 Lipschitz map such that f Lip ≤ (H ∗ ) (∞),

then it may be shown moreover that u satisfies the Hamilton-Jacobi equation down
to t = 0.
For any (bounded and continuous) initial condition f , the solution given by u
presented in Proposition 9.4.1 is called the Hopf-Lax infimum-convolution repre-
sentation of the solutions of the Hamilton-Jacobi equation (9.4.2). The Hopf-Lax
formula actually defines a non-linear semigroup starting from f , and we denote, by
similarity with Markov semigroups,
" #
d(x, y)
Qt f (x) = inf f (y) + tH ∗
(H )
, t > 0, x ∈ M,
y∈M t
(H )
together with Q0 f = f . The semigroup property is then expressed by the fact
(H ) (H ) (H )
that Qt ◦ Qs = Qt+s , t, s ≥ 0. We will not be interested in the most general
9.4 Hamilton-Jacobi Equations 453
Hamiltonians, and actually only concentrate below (as for Wasserstein distances in
2
the preceding section) on the quadratic example H (r) = r2 for which H ∗ = H . In
(H )
this case, Qt will be denoted more simply by Qt , so that
" #
1
Qt f (x) = inf f (y) + d(x, y)2 , t > 0, x ∈ M. (9.4.4)
y∈M 2t
As already emphasized in (9.2.3), the link between the Hopf-Lax formula and
optimal transportation is made clear on the dual Kantorovich problem. Indeed, for
c(x, y) = H (d(x, y)), (x, y) ∈ M × M, Theorem 9.1.3 implies that for all probabil-
ity measures μ, ν ∈ P(M),

(H )
Tc (μ, ν) = sup Q1 f dμ − f dν (9.4.5)
M M
where the supremum is taken over all bounded continuous functions f : M → R.
9.4.2 Vanishing Viscosity
In order to get a flavor of the connections between Hamilton-Jacobi and heat equa-
tions, and therefore between transportation inequalities and functional inequalities
related to heat kernels as developed earlier in this monograph, consider the example
of a diffusion semigroup (Pt )t≥0 with generator L = g − ∇W · ∇ on a Rieman-
nian manifold (M, g) with carré du champ operator (f ) = |∇f |2 . The Hamilton-
2
Jacobi equation, with H (r) = r2 , r ∈ R+ , with a bounded continuous initial condi-
tion f : M → R takes here the form
1
∂t u + |∇u|2 = 0, u(0, ·) = f (·). (9.4.6)
2
Consider now, for every ε > 0,

uε = −2ε log Pεt e−f/2ε , (9.4.7)
for which, by the heat equation and the change of variables formula,
1 2
∂t uε = ε Luε − ∇uε , uε (0, ·) = f.
2
As ε → 0, it is expected that uε will approach the solution u of (9.4.6) in a rea-
sonable sense. This approximation is classically known as the vanishing viscosity
method. Note that the limiting solution u given by the infimum-convolution Qt f is
independent of the potential W and thus of the invariant measure dμ = e−W dμg of
the diffusion operator L = g − ∇W · ∇. In particular, this asymptotics is explicit
on the basic Brownian and Ornstein-Uhlenbeck examples.
In a more probabilistic interpretation, there is a clear picture of the preceding

asymptotics ε → 0 via the Laplace-Varadhan asymptotics with rate precisely de-
scribed by the infimum convolution of f with the quadratic large deviation rate
function for the heat semigroup according to (6.7.13), p. 303,
lim 4t log pt (x, y) = −d 2 (x, y),

t→0
where pt (x, y), t > 0, (x, y) ∈ M × M, are the density kernels with respect to the
reference measure μ. Under this asymptotics, uε in (9.4.7) takes the form

d 2 (x, ·)
u (x) = − log
ε
exp −f − + o(1)
, x∈M
2t 1/2ε
(where the norm is understood to be with respect to μ). Using the fact that
Lp -norms converge to L∞ -norms as p → ∞, it turns out in the end that uε con-
verges when ε → 0 to the infimum-convolution Qt f of (9.4.4) which recovers the
Hopf-Lax formulation. Of course, this should only be considered as an heuristic, the
precise approximation having to be set up more carefully (see Sect. 9.5 below).
9.5 Hypercontractivity of Solutions of Hamilton-Jacobi

Equations
This section addresses smoothing properties of the solutions of Hamilton-Jacobi
equations (9.4.2) in the form of the Hopf-Lax semigroup (Qt )t>0 similar to the hy-
percontractivity property of linear semigroups. As presented in Sect. 5.2.2, p. 246,
a logarithmic Sobolev inequality for the invariant measure of a Markov semigroup
(Pt )t≥0 may be translated equivalently into hypercontractive bounds on the latter.
We discuss here the analogous description for (Qt )t>0 . In continuation of the pre-
vious section, we work below on a (smooth complete connected) Riemannian man-
ifold (M, g) with Riemannian measure μg and distance d.
For simplicity, we only deal with the Hamilton-Jacobi equation for the quadratic
cost, leading by Proposition 9.4.1 to the Hopf-Lax representation
" #
1
Qt f (x) = inf f (y) + d(x, y)2 , t > 0, x ∈ M,
y∈M 2t
of (9.4.4). Following the naive approximation of the Hamilton-Jacobi equation de-

scribed in Sect. 9.4, the results described below should not look too surprising.
Recall from Definition 5.1.1, p. 236, that, after a change of functions, a proba-
bility measure μ on the Borel sets of M satisfies a logarithmic Sobolev inequality
LS(C) with constant C > 0 if for every smooth function f on M,

f C
Entμ e ≤ |∇f |2 ef dμ.
2 M
9.5 Hypercontractivity of Solutions of Hamilton-Jacobi Equations 455
The following statement is then the analogue of the (linear) hypercontractivity The-
orem 5.2.3, p. 246.
Theorem 9.5.1 (Hypercontractivity of Solutions of Hamilton-Jacobi Equations)

Let μ be a probability measure on (M, g) (absolutely continuous with respect to the
Riemannian measure μg ). If μ satisfies the logarithmic Sobolev inequality LS(C),
then for all t > 0, all a > 0, and all bounded continuous functions f on M,
Q f
e t t ≤ ef . (9.5.1)
a+ C a
Conversely, if (9.5.1) holds for some a > 0 and all t > 0 and all f bounded and
continuous, then μ satisfies LS(C).
Proof The proof is similar to the proof of Theorem 5.2.3, p. 246. The derivative
along Pt is replaced by the derivative along Qt , which together with the Hamilton-
Jacobi equation (9.4.2) yields that

q
∂t eq Qt f dμ = − |∇Qt f |2 eq Qt f dμ.
M 2 M
Then, setting

= (t, q) = eq Qt f dμ, t ≥ 0, q > 0,
M
the LS(C) inequality applied to q Qt f expresses that
q ∂q − log ≤ −Cq ∂t .
For the choice of q = q(t) = a + Ct , it is then immediately checked from this in-
equality that H (t) = q(t)
1
log( (t, q(t))) is decreasing in t > 0. Inequality (9.5.1) is
then simply H (t) ≤ H (0).
Conversely, assuming that, for some a > 0, (9.5.1) holds for any t > 0 and any
initial condition f , the time derivative at t = 0 yields

Ca 2
Entμ eaf ≤ |∇f |2 eaf dμ,
2 M
which amounts (as a > 0) to the logarithmic Sobolev inequality LS(C). Theo-
rem 9.5.1 is therefore established. (Note that the absolute continuity of μ is im-
plicitly used in the Hamilton-Jacobi equation, which only holds almost everywhere.
More refined analyses developed in works mentioned in the Notes and References
actually allow us to suitably extend the conclusion without this assumption.)
While Theorem 9.5.1 describes, under a logarithmic Sobolev inequality for μ,

a hypercontractivity property for the semigroup (Qt )t>0 of the classical Hamilton-
Jacobi equation similar to hypercontractivity of the semigroup (Pt )t≥0 with gener-
ator L and invariant measure μ, it is actually possible to derive hypercontractivity
of (Qt )t>0 directly from that for (Pt )t≥0 using the vanishing viscosity method out-
lined in Sect. 9.4. Indeed, under the logarithmic Sobolev inequality LS(C), apply
the reverse hypercontractivity property described in Remark 5.2.4, p. 247, stating
that
Pt hq ≥ hp
for any strictly positive h : M → R and any −∞ < q 0, let 0 > p = −2εa > q = −2εb and t > 0 so that
1 + 2εb
e2εt/C = .
1 + 2εa
It follows that, with uε = −2ε log Pεt (e−f/2ε ) from (9.4.7),

uε
e ≤ ef .
b a
As ε → 0, uε → Qt f where t > 0 is such that b = a + Ct , and we recover Theo-

rem 9.5.1 in the limit ε → 0.
The preceding hypercontractive property of the Hopf-Lax solutions (Qt )t>0 of
the classical Hamilton-Jacobi equation may be developed similarly at the level of
the Euclidean logarithmic Sobolev inequality (6.2.8), p. 284, providing analogues
for Qt of the optimal heat kernel bounds in Theorem 6.3.1, p. 286. We work below
in Rn with the Lebesgue measure dx. Recall thus the Euclidean logarithmic Sobolev
inequality from Proposition 6.2.5, p. 284,

2 n 2
Entdx f ≤ log |∇f | dx
2
(9.5.2)
2 nπe Rn

which holds for f on Rn in the associated Dirichlet domain with Rn f 2 dx = 1.

Arguing as in the proof of Theorem 9.5.1, one similarly concludes the following
statement.
Theorem 9.5.2 For every bounded continuous function f : Rn → R, any 0 < α ≤ β

and any t > 0,
n β−α n α+β
Q f β − α 2 βα α 2 αβ
e t ≤ ef (9.5.3)
β 2πt β α
(where the norms are understood here with respect to the Lebesgue measure). Fur-
thermore, as β → ∞ (with α = 1 for example), for every x ∈ Rn ,

n 1
Qt f (x) ≤ log ef dx + log .
Rn 2 2πt
Proof By scaling, it is enough to assume that t = 1. By concavity of the logarithmic

function, the Euclidean logarithmic Sobolev inequality (9.5.2) is as usual linearized
9.5 Hypercontractivity of Solutions of Hamilton-Jacobi Equations 457
as

n n r
Entdx ef ≤ |∇f |2 ef dx + log ef dx, r > 0.
2r Rn 2 2πe2 n Rn
αβ
Setting (t) = eQt f q(t) with q(t) = (α−β)t+β , 0 < α ≤ β, this family of logarith-

mic Sobolev inequalities with r = nq (t) implies as in the proof of Theorem 9.5.1
that

nq (t) q (t)
(t) ≤ (t) 2 log .
2q (t) 2πe2
The conclusion follows after integration over [0, 1]. Note that the limiting case
β = ∞ may also be established directly since by definition of Qt f (x),
2 /2t
eQt f (x)−|x−y| ≤ ef (y)
for every x, y ∈ Rn and t > 0, so that integration in y yields the claim.
It should be noted that (9.5.3) of Theorem 9.5.2, which holds for every t ≥ 0 and
β ≥ α for a fixed α > 0, is formally equivalent to the Euclidean logarithmic Sobolev
inequality.
To conclude this section we briefly describe a local version of the main hyper-
contractivity Theorem 9.5.1 for solutions of the Hamilton-Jacobi equations in the
spirit of local hypercontractivity for the linear semigroup (Pt )t≥0 of Theorem 5.5.5,
p. 262. The result in particular produces a characterization of the curvature condi-
tion CD(ρ, ∞). We still work in the context of a (weighted) Riemannian manifold
(M, g).
Theorem 9.5.3 (Local hypercontractivity) Consider the Markov Triple consist-

ing of a smooth complete connected Riemannian manifold (M, g) and the diffusion
operator L = g − ∇W · ∇, where W is a smooth potential, with invariant mea-
sure dμ = e−W dμg and Markov semigroup (Pt )t≥0 . The following assertions are
equivalent.
(ii) For all q > p > 0 and u, t > 0 satisfying q = p + 1−eρt−2ρu ,
1/q 1/p
Pu eq Qt f ≤ Pu epf , (9.5.4)
for all bounded continuous functions f on M. When ρ = 0 the condition has to

be replaced by q = p + 2u
t
.
Proof We only sketch the argument following the proofs of Theorem 5.5.5,
p. 262, and Theorem 9.5.1. Theorem 5.5.2, p. 259, indicates that (i) is equiva-
lent to the local logarithmic Sobolev inequality (5.5.5), p. 260. Fixing u > 0, let
(s) = Pu (eq(s)Qs f )1/q(s) , s ∈ [0, t], where q is affine and satisfies q (s) = 1−eρ−2ρu .
Arguing as in the proof of Theorem 9.5.1, the local logarithmic Sobolev inequal-
ity (5.5.5), p. 260, at time u readily implies that (s) ≤ 0 on [0, t], proving (ii).
Conversely, a simple asymptotics at t = 0 of (9.5.4) implies (5.5.5), p. 260, and
thus (i).
9.6 Transportation Cost and Logarithmic Sobolev Inequalities

On the basis of the preceding developments, this section examines the connec-
tions between logarithmic Sobolev inequalities and (quadratic) transportation cost
inequalities, as well as stability properties of transportation cost inequalities. We
again deal with the setting of a (smooth complete connected) Riemannian manifold
(M, g).
9.6.1 From Logarithmic Sobolev to Transportation Cost

Inequalities
The first main result is that a logarithmic Sobolev inequality is stronger than a
quadratic transportation cost inequality.
Theorem 9.6.1 (Otto-Villani Theorem) Let μ be a probability measure on M. If

μ satisfies a logarithmic Sobolev inequality LS(C) for some constant C > 0, then it
satisfies a quadratic transportation cost inequality T2 (C) with the same constant C.
The proof is actually an immediate consequence of the hypercontractivity Theo-

rem 9.5.1 (assuming in addition that μ is absolutely continuous with respect to μg ).
Indeed, if μ satisfies LS(C), we may let a → 0 and t = C therein to obtain that for
every bounded continuous function f on M,

eQC f dμ ≤ e M f dμ .
M
But this is exactly the dual formulation of the T2 (C) inequality as expressed in
Proposition 9.2.3.
We will present below an alternative proof of the Otto-Villani Theorem. How-
ever, before we begin, it should be pointed out that the converse implication in The-
orem 9.5.1 only works when a > 0, that is a transportation cost inequality T2 (C)
does not in general imply in return a logarithmic Sobolev inequality (and explicit
counterexamples may be given on the real line). Nevertheless, a T2 (C) inequality
implies a Poincaré inequality P (C).
Proposition 9.6.2 Let μ be a probability measure on M. If μ satisfies a quadratic

transportation cost inequality T2 (C) for some constant C > 0, then it satisfies a
Poincaré inequality P (C) with the same constant C.
9.6 Transportation Cost and Logarithmic Sobolev Inequalities 459
The proof again relies on the dual description of T2 (C) which yields that for
every bounded continuous function f on M, and every t > 0,

et QCt f dμ ≤ et M f dμ .
M
A second order Taylor expansion at t = 0 together with the Hamilton-Jacobi equa-

tion immediately yields the claim.
While the converse of Theorem 9.5.1 does not hold in general, a partial result in
this direction is of use in the study of perturbation properties of transportation cost
inequalities, which we briefly describe without proof. Namely, whenever a = 0, that
is starting from T2 (C), for f bounded and continuous and t > 0,

t/C
etQt f/C dμ ≤ e(t/C) M f dμ
≤ ef dμ
M M
where we used Jensen’s inequality in the last inequality. By the entropic inequal-
ity (5.1.2), p. 236,

t
ef Qt f dμ ≤ Entμ ef + ef dμ log etQt f/C dμ .
C M M M
By the preceding, after some algebra, it follows that

t2 1
Entμ ef ≤ [f − Qt f ]ef dμ (9.6.1)
t −C M t
for every t > C. While we cannot let t → 0 in (9.6.1) to reach a logarithmic Sobolev
inequality, it turns out that, after some work, (9.6.1) implies in return a transportation
cost inequality T2 (C1 ) for some C1 > 0 only depending on C. In other words, (9.6.1)
is a kind of ersatz of the logarithmic Sobolev inequality replacing the generator L by
t (Id −Qt ) which is equivalent to T2 (C). The argument relies on a refined analysis of
1
the Herbst argument of Sect. 5.4, p. 252, but we skip the details here (see Sect. 9.9).
Now, due to Lemma 5.1.7, p. 240, and since f − Qt f ≥ 0, it is clear that (9.6.1)
is stable under bounded perturbations of μ. As a consequence, similarly to Poincaré
and logarithmic Sobolev inequalities, quadratic transportation cost inequalities T2
are stable under bounded perturbations.
Proposition 9.6.3 (Bounded perturbation) Assume that a probability measure μ

on M satisfies a quadratic transportation cost inequality T2 (C) for some C > 0. Let
μ1 be a probability measure with density h with respect to μ such that b1 ≤ h ≤ b
for some constant b > 0. Then μ1 satisfies T2 (C1 ) where C1 > 0 only depends on C
and b.
As for Poincaré and logarithmic Sobolev inequalities, if h = ek , the dependence

on b may be replaced by dependence on osc(k) = sup k − inf k.
9.6.2 Dimension-Free Transportation Inequalities
Towards an alternative approach to the Otto-Villani Theorem 9.6.1, we next inves-

tigate the relationships between the T1 and T2 transportation cost inequalities of
independent interest. The main statement is the following result which clarifies the
difference between T1 and T2 inequalities under tensorization. As mentioned be-
fore Proposition 9.2.4 establishing the stability under products of T2 inequalities,
T1 inequalities do not tensorize independently of the dimension. If (E, d) is a met-
ric (Polish) space, the k-fold product space E k is assumed to be equipped with the
2 -metric (d 2 ⊕ · · · ⊕ d 2 )1/2 .
Theorem 9.6.4 (Gozlan’s Theorem) Let μ be a probability measure on the Borel

sets of (E, d) such that μ⊗k satisfies the transportation cost inequality T1 (C)
of (9.2.1) on E k uniformly in k ≥ 1. Then μ satisfies the quadratic transportation
cost inequality T2 (C).
The proof shows that the conclusion actually holds under a (uniform) Gaussian
measure concentration property (8.7.3), p. 426 (essentially equivalent to T1 (C)).
Proof Let f : E → R be positive (bounded measurable) and let ε > 0. Set

" #
A = x ∈ E ; f (x) ≤ (1 + ε) f dμ .
E
Then 1 − μ(A) ≤ 1
1+ε and, for every t > 0 and every x ∈ E,
" #
1 1
Qt f (x) ≤ inf f (y) + d(x, y) ≤ (1 + ε) f dμ + d(x, A)2
2
y∈A 2t E 2t
where d(x, A) is the distance from the point x to the set A. For every r > 0, the map
h(x) = min(d(x, A), r), x ∈ E, is 1-Lipschitz and

r
h dμ ≤ r 1 − μ(A) ≤ .
E 1+ε
Hence, if s = εr
1+ε ,

r
μ d(·, A) ≥ r = μ(h ≥ r) = μ h ≥ +s ≤μ h≥ hdμ + s ,
1+ε E
so that by the dual formulation (9.2.5) of the T1 inequality and its application (9.2.6)
via Markov’s inequality,

μ d(·, A) ≥ r ≤ e−s /2C = e−ε r /2C(1+ε)
2 2 2 2
9.6 Transportation Cost and Logarithmic Sobolev Inequalities 461

for every r > 0. On the basis of this tail estimate, we may actually control E hdμ
in a finer way as
r

h dμ = μ d(·, A) ≥ u du
E 0
, √
r
−ε 2 u2 /2C(1+ε)2 2 C(1 + ε)
≤ e du ≤ = r0 (ε)
0 π ε
to get similarly that, for every r ≥ r0 (ε),

μ d(·, A) ≥ r = μ(h ≥ r) ≤ μ h ≥ hdμ + r − r0 (ε)
E
≤ e−(r−r0 (ε))
2 /2C
.
In particular, for every t > C,

2 /2t
ed(·,A) dμ ≤ K
E
where K = K(C, t, ε). Therefore, in conclusion, for t > C,

eQt f dμ ≤ K e(1+ε) E f dμ .
E
By the hypothesis, apply the latter on (E k , μ⊗k ), with respect to the 2 -metric
(thus with K uniform in k ≥ 1), to
f (x1 , . . . , xk ) = g(x1 ) + · · · + g(xk ), (x1 , . . . , xk ) ∈ E k ,
for a given positive (bounded measurable) function g : E → R. Since
Qt f (x1 , . . . , xk ) = Qt g(x1 ) + · · · + Qt g(xk ),
it follows that

eQt g dμ ≤ K 1/k e(1+ε) E gdμ

.
E
Since K = K(C, t, ε) is independent of k, as k → ∞, and then ε → 0, t → C, the
inequality T2 (C) in its dual formulation (Proposition 9.2.3) follows. The proof is
complete.
As announced, the preceding Theorem 9.6.4 may be used to provide an alterna-

tive proof of Theorem 9.6.1 linking logarithmic Sobolev inequalities and quadratic
transportation inequalities, in particular going beyond smooth settings. Indeed, un-
der a logarithmic Sobolev inequality for μ, by the tensorization Proposition 5.2.7,
p. 249, and the Herbst argument, Proposition 5.4.1, p. 252, μ⊗k satisfies a T1 in-
equality independently of k, and thus a T2 inequality by Theorem 9.6.4 (with the
same constant). In this form, Theorem 9.6.1 may be extended to the Markov Triple
structure (E, μ, ) (provided the intrinsic distance d defines a Polish topology
on E).
9.7 Heat Flow Contraction in Wasserstein Space

This section is concerned with contraction properties in Wasserstein distance along
the heat flow (Pt )t≥0 of a diffusion operator with some curvature. We work again in
the context on a (smooth complete connected) Riemannian manifold (M, g) with a
diffusion operator L = g − ∇W · ∇ where W is a smooth potential on M, which is
the infinitesimal generator of a Markov semigroup (Pt )t≥0 . As will be pointed out,
the results here may be developed in the usual setup of a (Full) Markov Triple but
by homogeneity with the rest of this chapter, we stick to the preceding framework.
The first step towards contractions is a basic commutation property between the
heat and Hamilton-Jacobi semigroups under curvature conditions.
Proposition 9.7.1 (Commutation property) In the preceding weighted Riemannian

setting, the following assertions are equivalent.
(ii) For every bounded continuous function f : M → R and every t, s ≥ 0,
Pt (Qs f ) ≤ Qe2ρt s (Pt f ). (9.7.1)
Proof For simplicity we only prove the result for ρ = 0, the general result being
obtained in the same way. The most important part is the proof of (ii) under the
curvature condition. Assume for simplicity that f ≥ 0. To this end, use the log-
Harnack inequality (5.6.2), p. 266, which applied to eu Q1 f yields
1
Pt (Q1 f ) ≤ Q2t log Pt eu Q1 f
u
for every t, u > 0. Now, the local hypercontractivity inequality (9.5.4) as p → 0
implies that
1 1
log Pt e 2t Q1 f ≤ Pt f
2t
for every t > 0. Combining the inequalities with u = 2t1 yields (i) for s = 1. Arbi-
trary values of s are obtained by homogeneity. Conversely, under (ii), the s deriva-
tive in (9.7.1) at s = 0 implies in return, by the Hamilton-Jacobi equation, the gra-
dient bound |∇Pt f |2 ≤ Pt (|∇f |2 ) which is equivalent to the curvature dimension
CD(0, ∞). The proof is complete.
Under the curvature condition CD(ρ, ∞), the commutation (9.7.1) actually ex-
tends to the (Full) Markov Triple setting (E, μ, ) on the basis of the developments
9.7 Heat Flow Contraction in Wasserstein Space 463
in Sect. 8.6, p. 421. Indeed (assuming in addition that the intrinsic distance de-
fines a Polish topology on E), apply the isoperimetric Harnack inequality of The-
orem 8.6.3, p. 425, to A = {Q1 f ≥ u}, u ≥ 0, for f : E → R bounded continuous
and positive. If z ∈ Aε (where ε > 0), there exists an a ∈ A such that d(z, a) < ε so
that, by definition of the infimum-convolution Q1 ,
ε2 d(z, a)2
f (z) + ≥ f (z) + ≥ Q1 f (a) ≥ u.
2 2
Hence
" #
d(x, y)2
{Q1 f ≥ u} d(x,y) ≤ f ≥ u +
2
for all (x, y) ∈ E × E with d(x, y) > 0. Again taking ρ = 0 for simplicity, Theo-
rem 8.6.3 after integration with respect to u ≥ 0 then yields
d(x, y)2
Pt (Q1 f )(x) ≤ Pt f (y) +
2
for every x and y in E. In other words, Pt (Q1 f ) ≤ Q1 (Pt f ) which amounts to (ii).
On the basis of the commutation property (9.7.1) between the heat and Hopf-
Lax semigroups of the last proposition, we next obtain contraction in Wasserstein
distance, which is again shown to be equivalent to the curvature condition. Recall
that f μ denotes for simplicity the measure with density f with respect to μ.
Theorem 9.7.2 (Heat flow contraction in Wasserstein space) Consider the Markov
Triple consisting of a smooth complete connected Riemannian manifold (M, g) and
of a diffusion operator L = g − ∇W · ∇, where W is a smooth potential, with
invariant measure dμ = e−W dμg and Markov semigroup (Pt )t≥0 . The following
assertions are equivalent.
(ii) For every t ≥ 0 and all probability densities f and g with respect to μ with a
finite second moment,
W2 (Pt f μ, Pt gμ) ≤ e−ρt W2 (f μ, gμ). (9.7.2)
Proof We start with (i). For any bounded continuous h : M → R, by time reversibil-
ity and (9.7.1)

Q1 h Pt f dμ − h Pt g dμ = Pt (Q1 h)f dμ − Pt h g dμ
M M M M

≤ Qe2ρt (Pt h)f dμ − Pt h g dμ.
M M
Since by homogeneity Qe2ρt (Pt h) = e−2ρt Q1 (e2ρt Pt h), the Kantorovich dual de-
scription of the Wasserstein distance (Theorem 9.1.3) yields the contraction prop-
erty (9.7.2).
Assume conversely that (ii) holds, that is, again by Kantorovich duality and sym-
metry,

Pt (Q1 h)f dμ − Pt h g dμ ≤ e−ρt W2 (f μ, gμ)
M M
for any bounded continuous h : M → R. Given x, y ∈ M, by a suitable approxima-
tion, let f μ and gμ approach respectively Dirac masses at x and y so that the latter
inequality turns into
d(x, y)2
Pt (Q1 h)(x) − Pt h(y) ≤ e−2ρt
2
(alternatively take ν1 = δx and ν2 = δy in (9.7.4) below). Taking the infimum over y
then exactly expresses the commutation property (9.7.1), which is equivalent to the
CD(ρ, ∞) condition. The theorem is established.
With similar arguments, it may be shown that the curvature-dimension condition

CD(0, n) is equivalent to the contraction bound
√ √ 2
W22 (Pt f μ, Ps gμ) ≤ W22 (f μ, gμ) + n t− s , (9.7.3)
for any t, s ≥ 0 and any probability densities f and g with respect to μ.
Remark 9.7.3 Theorem 9.7.2 expresses equivalently that, under the curvature con-
dition CD(ρ, ∞),

W2 Pt∗ ν1 , Pt∗ ν2 ≤ e−ρt W2 (ν1 , ν2 ) (9.7.4)
for all probability measures ν1 and ν2 in P2 (M), where according to (1.1.2), p. 9,
Pt∗ ν1 , respectively Pt∗ ν2 , is the law at time t ≥ 0 of the underlying Markov process
{Xtx ; t ≥ 0, x ∈ M} with initial data ν1 , respectively ν2 . It should also be observed
that the preceding inequalities extend, via the commutation property, to the whole
family of Wasserstein distances Wp , 1 ≤ p < ∞, and even to more general costs.
9.8 Curvature of Metric Measure Spaces

This last section briefly outlines recent developments on notions of Ricci curvature
lower bounds based on optimal transport. These notions, providing an alternative
approach to Ricci curvature in Riemannian spaces, may actually be extended to the
context of metric measure spaces, justifying their interest.
To start with, let M = Rn and dμ = e−W dx be a probability measure on the
Borel sets of Rn for a smooth potential W : Rn → R. If μ0 and μ1 are absolutely
continuous probability measures on Rn , recall T = ∇φ, the Brenier map pushing
μ0 onto μ1 . For any θ ∈ [0, 1], interpolate between μ0 and μ1 by letting

μθ = (1 − θ ) Id +θ T #μ0 .
9.8 Curvature of Metric Measure Spaces 465
One then speaks of the geodesics of optimal transportation joining μ0 to μ1 .

Namely, since the Brenier map achieves optimality in the Wasserstein distance W2 ,
W2 (μ0 , μθ ) + W2 (μθ , μ1 ) = W2 (μ0 , μ1 ),
justifying the terminology (see below).

As we have seen, under suitable regularity assumptions, μθ has a density fθ
(with respect to the Lebesgue measure) satisfying

f0 (x) = fθ (1 − θ )x + θ ∇φ(x) ) det (1 − θ ) Id +θ ∇∇φ(x)
f0 (x)dx-almost everywhere. Arguing then as in the proof of Theorem 9.3.1, it may

be shown that whenever ∇∇W ≥ 0, the relative entropy H(· | μ) with respect to μ
is convex along the geodesics of optimal transportation in the sense that
H(μθ | μ) ≤ (1 − θ )H(μ0 | μ) + θ H(μ1 | μ) (9.8.1)
for every θ ∈ [0, 1].

The preceding convexity property (9.8.1) has taken the name of displacement
convexity (convexity along the geodesics of optimal transportation). Together with
the suitable Riemannian extension of the Brenier map, (9.8.1) extends to (weighted)
Riemannian manifolds (M, g) under the convexity property Ric +∇∇W ≥ 0. Actu-
ally, under Ric +∇∇W ≥ ρ g, ρ ∈ R, the displacement convexity of order ρ has to
be understood with respect to the Wasserstein distance W2 as
H(μθ | μ) ≤ (1 − θ )H(μ0 | μ) + θ H(μ1 | μ)

ρ θ (1 − θ ) (9.8.2)
− W2 (μ0 , μ1 )2 , θ ∈ [0, 1].
2
Now, these displacement convexity inequalities may be formulated in purely met-
ric terms, opening the extension to abstract metric measure spaces. A key observa-
tion in this regard is that the displacement inequality (9.8.2) in Riemannian man-
ifolds characterizes in return the Ricci curvature lower bound Ric +∇∇W ≥ ρ g.
Therefore, (9.8.2) provides an equivalent (synthetic, as opposed to the more local
definition of Riemannian curvature) definition of Ricci curvature lower bounds. This
interpolation along the geodesics of optimal transport moreover shares some simi-
larities with the semigroup interpolation along the heat flow. The optimal transporta-
tion formalism actually induces a formal Riemannian structure on the Wasserstein
space P2 (E) for which convexity of entropy provides the corresponding curvature
lower bounds.
To formulate the extension to metric measure spaces, it is necessary to clarify
a few definitions from metric geometry. In a metric (Polish) space (E, d), a curve
joining two points x, y ∈ E is a continuous map γ : [0, 1] → E such that γ (0) = x
and γ (1) = y. The length of γ is defined by L(γ ) = sup ki=1 d(γ (si−1 ), γ (si ))
where the supremum runs over all subdivisions 0 = s0 < s1 < · · · < sk = 1. Such
a curve γ is called a geodesic if L(γ ) = d(γ (0), γ (1)). The metric space (E, d)
is called a length space if, for all (x, y) ∈ E × E, d(x, y) = inf L(γ ) where the
infimum is taken over all curves γ joining x to y. The space (E, d) is called
geodesic if all points x, y ∈ E are connected by a geodesic. In a geodesic space,
there exist middle points, or θ -middle points (θ ∈ [0, 1]), z between x, y ∈ E, i.e.
d(x, z) = θ d(x, y), d(z, y) = (1 − θ )d(x, y), and such a point z is necessarily on
a geodesic joining x to y. The Wasserstein space P2 (E) of probability measures
on the Borel sets of E with a finite second moment equipped with the Wasserstein
distance W2 is a length space if (E, d) is a length space.
In the following, μ is a fixed reference probability measure on (E, d), so that
(E, d, μ) is a metric measure space. According to the preceding developments in
a Riemannian context, we say that (E, d, μ) is of Ricci curvature bounded from
below by ρ, ρ ∈ R, if for every pair (μ0 , μ1 ) in P2 (E) (with supports contained in
the support of μ), there exists a geodesic (μθ )θ∈[0,1] in P2 (E) joining μ0 and μ1
such that
ρ θ (1 − θ )
H(μθ | μ) ≤ (1 − θ )H(μ0 | μ) + θ H(μ1 | μ) − W2 (μ0 , μ1 )2 (9.8.3)
2
for every θ ∈ [0, 1].
As already mentioned above, this definition is thus equivalent to the Ricci cur-
vature lower bound Ric +∇∇W ≥ ρ g in Riemannian manifolds. Among the most
noticeable aspects of this synthetic definition, not discussed here, it may be shown
to be stable under the so-called Gromov-Hausdorff convergence (a weak notion of
convergence of metric measure spaces for which a limit of Riemannian manifolds is
not necessarily a Riemannian manifold anymore). The preceding definition of Ricci
curvature lower bound also allows for Poincaré or logarithmic Sobolev inequalities
as investigated in this monograph. Further definitions may also be used to include
curvature and dimension of a metric measure space as the CD(ρ, n) condition (see
the Notes and References).
This chapter and the notes and references here are only a rough introduction to
mass transportation and transportation cost inequalities. General references on mass
transportation are the comprehensive books by S. T. Rachev [351], S. T. Rachev
and L. Rüschendorf [352, 353], and C. Villani [424, 426]. The first references are
more oriented towards probabilistic issues. The monographs by C. Villani present
a complete and deep investigation of modern mass transportation theory and its
rich interplay with partial differential equations, probability theory and geometry.
A recent account of transportation cost inequalities with bibliographical references
is the survey [210] by N. Gozlan and C. Léonard.
The reader may find in the preceding references the basic Kantorovich Theo-
rem 9.1.3 and Kantorovich-Rubinstein Theorem 9.1.4 of Sect. 9.1, presented in
greater generality. The existence of a transport map as the gradient of a convex
function (Theorem 9.1.5) goes back to M. Knott and C. Smith [267], L. Rüschen-
dorf and S. T. Rachev [373] and Y. Brenier [94], the latter having emphasized the
geometric relevance to partial differential equations. The manifold case is due to
R. McCann [304]. The associated Monge-Ampère equation (9.1.8) with the Hessian
of φ in the sense of Aleksandrov is proved in [305], on the basis of results in [180]
and the regularity properties of the Brenier map as expressed in Theorem 9.1.6 due
to L. Caffarelli [101, 102] (cf. [424, 426] for details and numerous further contribu-
tions).
The quadratic transportation cost inequality for the Gaussian measure (Theo-
rem 9.2.1) is due to M. Talagrand [404] with the one-dimensional proof together
with tensorization. A multi-dimensional transportation proof, further extended to
strictly log-concave measures, appeared in [71, 133] (see also below). The dual for-
mulation of the quadratic transportation cost inequality T2 (Proposition 9.2.3) was
introduced by S. Bobkov and F. Götze in [77]. The corresponding statement for T1
may also be found there. The link between the latter and measure concentration goes
back to K. Marton [298] (cf. [278, 426]). See [210, 213] for further developments
on transportation cost inequalities.
The new HWI inequality of Corollary 9.3.3 was introduced in the seminal pa-
per [340] by F. Otto and C. Villani. In this work, deep links between mass transporta-
tion and functional inequalities were identified on the basis of the Otto differential
calculus in the Wasserstein space of probability measures [259, 339]. The proof of
Theorem 9.3.1 given here, and its consequence for logarithmic Sobolev inequalities
and quadratic transportation cost inequalities, is due to D. Cordero-Erausquin [133].
The alternative heat flow proof of the HWI inequality comes from [76]. Caffarelli’s
contraction Theorem 9.3.4 is proved in [103] (see also [424]). On the basis of the
transportation proof of the HWI and logarithmic Sobolev inequalities, the optimal
transportation method to obtain classical Sobolev inequalities with their extremal
functions was developed by D. Cordero-Erausquin, B. Nazaret and C. Villani [137]
(where in particular more details on the integration by parts formula (9.3.6) may
be found). The optimal Sobolev trace inequality of Theorem 9.3.6 is due inde-
pendently to J. F. Escobar and W. Beckner [56, 177]. The proof presented in the
monograph and its generalization in Lp (as presented in Remark 9.3.7) is due
to B. Nazaret [324]. The transportation mass method has been further developed
to investigate a wide variety of inequalities between entropy functionals and en-
ergy in a partial differential equation context (see e.g. [3, 112, 113, 134, 296], and
cf. [424, 426]). It should be noted that, as far as functional inequalities such as
Sobolev-type inequalities are concerned, alternative maps may be considered in the
transportation proofs. In particular, the earlier triangular Knothe map [266] can be
used to this end, as M. Gromov did in a proof of the standard isoperimetric in-
equality in Euclidean space [316] (see also [137, 426]). The optimal Brenier map
becomes relevant as soon as more (Riemannian) geometric features enter into the
picture (see [426] for much more in this regard).
Classical properties of Hamilton-Jacobi equations as briefly outlined in Sect. 9.4
such as the Hopf-Lax infimum-convolution formula and the vanishing viscosity
method are presented, for example, in [44, 104, 179] (see also [424, 426] or [291]
in the context of length spaces). For recent developments in metric structures, see
[11, 12, 43, 212]).
Hypercontractivity of solutions of Hamilton-Jacobi equations in Sect. 9.5 have
been emphasized in [76] towards a new proof of the Otto-Villani Theorem. Theo-
rem 9.5.1 appears there whereas ultracontractivity of the Hamilton-Jacobi solutions
is examined in [197] and in a more general context in [188, 198]. In the latter pa-
pers, generalizations of the Euclidean logarithmic Sobolev inequality in Lp (dx),
originally introduced in [148], are also considered. Local hypercontractivity for
Hamilton-Jacobi solutions, Theorem 9.5.3, is investigated in [31].
A first contraction result in Wasserstein distance (along the Boltzmann equation)
goes back to H. Tanaka [407, 408]. The Wasserstein contraction property put for-
ward in Theorem 9.7.2 may be traced back to the investigation [339] of the heat
flow as a gradient flow in the Wasserstein space, further developed in this spirit
in [112, 113, 341] and [9, 175] in connection with the Evolutionary Variational
Inequality. For coupling arguments, see [128, 129, 427, 431]. The crucial equiva-
lence with curvature bounds as expressed by Theorem 9.7.2 is due to M. von Re-
nesse and K.-T. Sturm [427]. See [424, 426] for more on Wasserstein contraction
properties. The proof given here, based of the commutation property between the
heat and Hopf-Lax semigroup (Proposition 9.7.1) is due to K. Kuwada [270] (see
also [12, 37]). Contraction under any curvature-dimension is studied, among other
recent developments, in [85, 176, 271].
The major link between quadratic transportation inequality and logarithmic
Sobolev inequality, at the source of many developments at the interface between
partial differential equations, probability theory and geometry, is due to F. Otto
and C. Villani [340]. The proof presented here is taken from [76]. That logarithmic
Sobolev inequalities are actually strictly stronger that quadratic transportation cost
inequalities is due to P. Cattiaux and A. Guillin [117] (see [208] for a direct charac-
terization). The perturbation Proposition 9.6.3 is due to N. Gozlan, C. Roberto and
P.-M. Samson [211], and was part of the early motivation of [340]. Theorem 9.6.4
is due to N. Gozlan [206], following the preliminary investigation [209] (cf. [210]).
Section 9.8 is a very short overview of recent deep investigations involving a no-
tion of Ricci curvature lower bounds in metric measure spaces due to J. Lott and
C. Villani [292] and K.-T. Sturm [399, 400]. Preliminary steps consisted of the dis-
placement convexity (9.8.1) by R. McCann [304], the heat flow as gradient flow of
entropy with respect to the Wasserstein metric by R. Jordan, D. Kinderlehrer and
F. Otto [259] and the Otto calculus [339], the links with functional inequalities by
F. Otto and C. Villani [340], the Borell-Brascamp-Lieb inequalities by D. Cordero-
Erausquin, R. McCann and M. Schmuckenschläger [135, 136] and the equiva-
lence with the Riemannian Ricci curvature lower bounds by M. von Renesse and
K.-T. Sturm [427]. This has been the starting point of many fruitful developments in
recent years. The monumental monograph [426] by C. Villani is a complete expo-
sition to which the reader is warmly referred for both mathematical developments
and historical background. See also [9].
In connection with the Markov operator curvature-dimension viewpoint empha-
sized in this book, it should be mentioned that recent developments by L. Ambrosio,
N. Gigli and G. Savaré [10, 11] actually conclude that in a suitable sub-class of met-
ric measure spaces for which the gradient energy is Hilbertian, the curvature as con-
vexity of relative entropy (9.8.3) is equivalent to the Markov operator curvature con-
dition CD(ρ, ∞), thus providing an even deeper link between Markov semigroups
and optimal transportation. The approach relies on the identification of the heat
flow as gradient flow of either Dirichlet energy or entropy in curved spaces. An im-
proved version of contraction in Wasserstein distance in the form of the Evolution-
ary Variational Inequality combining the heat flow interpolation along the geodesics
of optimal transportation provides one key step in this direction. This investigation
partially relies on the extension to a non-smooth framework of some of the ba-
sic semigroup tools described in this monograph by means of refined non-smooth
analysis in metric measure spaces. The case of curvature-dimension CD(ρ, n) is
investigated by M. Erbar, K. Kuwada and K.-T. Sturm [176, 271]. We refer to these
and the previously mentioned references for a complete account of these deep de-
velopments.
Appendices
The following appendices are devoted to the three basic aspects of this monograph,
the analytic aspect of semigroups of operators, the probabilistic aspect of diffusion
processes and their generators, and the geometric aspect of Laplace-type operators
and their curvature properties. These appendices only briefly feature a few central
objects and properties, in perspective with the main developments in the core of
the book. In particular, the analysis of Markov semigroups and their infinitesimal
generators may be found there. In this interplay between these different viewpoints,
the aim here is to help the reader to access these topics in a balanced way. It is of
course strongly advised to supplement these short surveys with more complete and
detailed references. A few such references are indicated at the end of each appendix.
The first appendix is thus devoted to some basics of the analysis of semigroups of
bounded operators on a Banach space, the second is a brief introduction to stochas-
tic calculus while the third presents some basic notions in differential and Rieman-
nian geometry. Note that the last two sections of the third appendix describe in this
geometric context some features of the 2 operator, in particular the reinforced cur-
vature condition, which are critically used in the core of the book.
Appendix A
Semigroups of Bounded Operators on a Banach
Space
This appendix presents some of the basics of the analysis of semigroups of operators
on a Banach space. It provides in particular the necessary material for the investi-
gation of Markov semigroups and their infinitesimal generators as developed in the
core of the book. The appendix surveys the Hille-Yosida theory, symmetric and
self-adjoint operators and their Friedrichs extensions, spectral decompositions and
Hilbert-Schmidt operators. Some general references on these topics are provided at
the end of the appendix.
A.1 The Hille-Yosida Theory

The Hille-Yosida theory is devoted to the study of semigroups of bounded opera-
tors on Banach spaces. The application to Markov semigroups mainly deals with
Lp -spaces, and moreover with contraction semigroups (as is the case for Markov
semigroups when the reference measure is invariant). For convenience, we restrict
to separable Banach spaces in what follows, although this assumption may be re-
moved in many places.
Thus let B be a (real separable) Banach space with norm · .
Definition A.1.1 (Semigroup) A family (Pt )t≥0 of linear operators Pt : B → B,

t ≥ 0, is called a contraction semigroup provided the following holds:
(i) For all t ≥ 0 and all x ∈ B, Pt x ≤ x.
(ii) For all t, s ≥ 0, Pt ◦ Ps = Pt+s .
(iii) For all x ∈ B, limt→0 Pt x = x (= P0 x).
For simplicity here (and throughout the book) Pt x = Pt (x). Note that linearity
together with the contraction property imply that for any x ∈ B, t → Pt x is contin-
uous in B since for every s ≥ 0,

Pt+s x − Pt x = Pt (Ps x − x) ≤ Ps x − x,

DOI 10.1007/978-3-319-00227-9,
474 A Semigroups of Bounded Operators on a Banach Space
and similarly Pt−s x − Pt x ≤ x − Ps x for 0 ≤ s ≤ t.

Given such a contraction semigroup (Pt )t≥0 , the Hille-Yosida theory ensures the
existence of a dense linear subspace of B on which the function t → Pt x has a
bounded derivative at t = 0 (and therefore at any point t > 0 by the semigroup
property). The derivative at t = 0 is a linear operator L, often unbounded. It is called
the infinitesimal generator of the semigroup P = (Pt )t≥0 and the dense domain on
which this generator is defined is called the domain of the generator L. Formally, Pt
may be described as et L .
To briefly understand these results, it is convenient to introduce new operators
acting on B. Given a bounded measure ν on the Borel sets of R+ , define the operator
Pν as
∞
Pν = Pt dν(t).
0

∞
Hence Pν x = 0 Pt x dν(t), x ∈ B. The latter quantity has to be understood as a
limit of Riemann sums

Pti x ν [ai , ai+1 )
i
along a sequence of partitions (ai ) which approximate the measure ν. When ν is

a finite measure, this operator is bounded on B with norm bounded from above by
|ν|([0, ∞)). In this extended framework, the semigroup property (ii) may then be
translated into
∞ ∞
Pν Pν = Pν Pν = Pt+s dν(t)dν (s) (A.1.1)
0 0
for bounded measures ν, ν on R+ . A special case is the so-called resolvent opera-

tor, defined for λ > 0 as
∞
Rλ = Pt e−λt dt (A.1.2)
0
(thus corresponding to the choice of dν(t) = e−λt dt). Anticipating the next results,
note that formally Rλ = (λ Id −L)−1 . Indeed, for any t ≥ 0 and x ∈ B,
∞
Rλ Pt x = eλt Ps x e−λs ds.
t
The time derivative at t = 0 then yields L(Rλ x) = λRλ x − x.

Now (A.1.1) and (A.1.2) yield the resolvent equation, that is, for λ, λ > 0,

Rλ − Rλ = λ − λ Rλ Rλ = λ − λ Rλ Rλ . (A.1.3)
In other words, Rλ = Rλ (Id +(λ − λ)Rλ ). This identity indicates that the image of
Rλ in B is included in the image of Rλ , and exchanging the roles of λ and λ , that
this image is independent of λ. Call this image D.
A.2 Symmetric Operators 475

∞
Since Pt x → x as t → 0 by (i) and λRλ x = 0 Ps/λ x e−s ds, it follows that for
every x ∈ B,
lim λRλ x = x.
λ→∞
Therefore the image D of Rλ is dense in B. Fix λ > 0, and define L on D by setting,

for x = Rλ y,
L x = λx − y.
Then L x may indeed be seen as the derivative of Pt x at t = 0. Indeed,
∞
Rλ Pt x = eλt Ps x e−λs ds,
t
so that the time derivative at t = 0 gives ∂t Rλ Pt x |t=0 = λRλ x − x. Therefore for

any t > 0,
∂t Pt x = LPt x = Pt L x.
A bit more is available. D is, in fact, precisely the set of points x in B for which
Pt x admits a derivative at t = 0. Furthermore, Rλ is a (bi-) continuous bijection
between B and the domain D when the latter is given the domain topology defined
from the norm by
xD = x + Lx, x ∈ D. (A.1.4)
The space D is then described as the domain of the operator L and denoted by D(L).
Moreover, the knowledge of D(L) and of the action of L on it completely determines
the semigroup P = (Pt )t≥0 as the unique solution of the (parabolic) heat equation
(with respect to L)
∂t Pt x = LPt x (A.1.5)
for x ∈ D(L), and subsequently for x ∈ B. The operator L is called accordingly the
infinitesimal generator of the semigroup P with domain D(L).
A.2 Symmetric Operators
In the special case when B is a (real) Hilbert space H with scalar product ·, ·, one
may consider semigroups of bounded symmetric operators. Their generators are
then (unbounded) self-adjoint operators on H. In this book, H will mainly appear
as an L2 (μ)-space for some (positive) measure μ on a measurable space (E, F),
but what follows has nothing to do with this particular case. (Unbounded) self-
adjoint operators have special features, such as, for example, spectral decomposi-
tions (cf. Sect. A.4).
The first point we need to pay attention to is the difference between symmetry
and self-adjointness of unbounded operators. A linear operator A defined on a dense
subspace D(A) of H is said to be symmetric if for all x, y in D(A),
Ax, y = x, Ay. (A.2.1)
For a symmetric operator A, its adjoint operator A∗ is defined on the domain D(A∗ )
consisting of the points x ∈ H for which there exists a finite constant C(x) such that
for any y ∈ D(A), |x, Ay| ≤ C(x)y. Since A is symmetric, D(A) ⊂ D(A∗ ). On
D(A∗ ), A∗ is then defined by duality
∗
A x, y = x, Ay.
More precisely, the map y → x, Ay, which is defined on the dense subspace D(A),
may be uniquely extended into a continuous linear form on H, and thus may be
represented as the scalar product with some element A∗ x ∈ H. By the symmetry
assumption, A∗ x = Ax for every x ∈ D(A) and A∗ is therefore an extension of A.
Moreover, A∗ is a closed operator on its domain, meaning that whenever (xk )k∈N
is a sequence in D(A∗ ) converging to x such that (A∗ xk )k∈N converges to y, then
x ∈ D(A∗ ) and A∗ x = y. Thus, in particular, densely defined symmetric operators
are closable (they admit a closed extension).
The following definition makes precise the main difference between symmetry
and self-adjointness.
Definition A.2.1 (Self-adjoint operator) In the preceding setting, a symmetric op-

erator A is self-adjoint if D(A∗ ) = D(A).
A self-adjoint operator therefore has no symmetric extension other than itself.

Except for bounded operators (bounded symmetric operators are self-adjoint), there
is a huge difference between symmetric and self-adjoint operators. Every symmet-
ric operator considered in this book admits at least one self-adjoint extension (in
the sense that its domain may be extended until this extension is self-adjoint), but
this extension may not be unique. Therefore, the description of an operator L on
a domain which is too small (even if dense in H) may not ensure the complete
description of the semigroup P = (Pt )t≥0 with infinitesimal generator L.
A semigroup (Pt )t≥0 is said to be symmetric if the operators Pt , t ≥ 0, are sym-
metric. In any case, the generator L of a symmetric semigroup (Pt )t≥0 of contrac-
tions on H, defined on its domain D(L), is always self-adjoint. This is a conse-
quence of the fact that the operators Pt , t ≥ 0, themselves are self-adjoint, and of
the very definition of the domain. The true problem is that in general the full do-
main of L is unknown. What is usually given is the description of L on a dense
subspace D0 of H, and the question is to determine whether D0 is dense in D(L)
for the topology of the domain. Such a dense subspace D0 is called a core. A core
completely determines a unique self-adjoint extension, and therefore a unique semi-
group P = (Pt )t≥0 (consisting of symmetric bounded operators) with infinitesimal
generator L. When a symmetric linear operator L defined on a dense subset D0 of H
has a unique self-adjoint extension to some larger domain D(L), then L is said to be
essentially self-adjoint (this notion refers both to L and D0 ). Determining if a given
A.3 Friedrichs Extension of Positive Operators 477
operator L defined on some domain D0 is essentially self-adjoint is investigated in

Sect. A.5.
A.3 Friedrichs Extension of Positive Operators

The generators L of symmetric Markov semigroups considered in this book share
some further positivity property. To stick with the standard notations, set here
A = −L. The operators A are then positive in the following sense.
Definition A.3.1 (Positive operator) Let H be a (real) Hilbert space, and D0 be a

dense linear subspace of H on which is given a symmetric linear operator A. The
operator A is said to be positive if, for any x ∈ D0 , x, Ax ≥ 0.
Symmetric positive operators A may be extended to self-adjoint operators in a

minimal way (which does not mean that the domain of the extension is minimal,
since in general the domains of the various self-adjoint extensions of A are not com-
parable). Indeed, define D0 as the set of points x ∈ H which are limits of sequences
(xk )k∈N in D0 for which the sequence (Axk )k∈N converges. It is easily checked,
using the symmetry of A, that this limit does not depend on the sequence (xk )k∈N .
This procedure thus defines an extension Ā of the operator A on D0 , which is closed
with respect to the topology of the domain given by the norm (A.1.4). The operator
Ā is a closed extension of A, however it is not self-adjoint in general so that A has
to be further extended.
With this aim in mind, on the linear space D0 , consider the norm
x, xA = x2A = x, x + x, Āx
and then take the completion of D0 under this norm. Since the norm · A
is larger than the norm in H, this completion naturally imbeds into H. The
resulting completion thus defines a new Hilbert space H1 which is such that
D0 ⊂ H1 ⊂ H. Define then the domain D(A) as the set of points x ∈ H1 for
which the map (x) : y → x, yA is bounded for the H-topology (that is such that
|(x)(y)| ≤ C(x)y for some C(x) < ∞). For these points x, (x)(y) may be rep-
resented as T (x), y for some linear operator T defined on D(A) which is easily
seen to be self-adjoint by construction. The self-adjoint extension of A is then given
by T − Id, and the domain D(A) of A satisfies D0 ⊂ D(A) ⊂ H1 . This self-adjoint
extension is called the Friedrichs extension of A.
The Friedrichs extension is not in general the unique self-adjoint extension of A,
and two different self-adjoint extensions may have domains which cannot be com-
pared. However, the Friedrichs extension is minimal in some other sense, namely
in terms of Dirichlet domains. The construction of the Hilbert space H1 above may
actually be performed for any symmetric (and therefore any self-adjoint) extension
of A and is called the domain of the Dirichlet form. It is this domain of the Dirichlet
form which is minimal for the Friedrichs extension. When dealing with operators
on a bounded domain, with D0 the set of smooth compactly supported functions

on it, this corresponds to the Dirichlet boundary condition, the Neumann boundary
condition corresponding to the maximal corresponding H1 space. The space D0 is
a core as soon as these minimal and maximal extensions coincide.
A.4 Spectral Decompositions

When A is a linear operator defined on a dense domain D(A) of a Banach space
B, the spectrum of A is defined according to the following definition. It is neces-
sary here to deal with complex Banach spaces, for which the developments of the
preceding sections are similar.
Definition A.4.1 (Spectrum) Given a linear operator A on B, the resolvent set ρ(A)
of A is the set of all complex values λ ∈ C such that the range of λ Id −A is dense
in B and such that λ Id −A has a bounded inverse Rλ = (λ Id −A)−1 . The spectrum
σ (A) of A is C \ ρ(A).
When the operator (A, D(A)) is closed (for the domain topology), if λ ∈ ρ(A),
then Rλ (λ Id −A) = B. Therefore the inverse (λ Id −A)−1 is everywhere defined.
Also, the resolvent set ρ(A) is always an open set in C.
A complex number λ may be in the spectrum σ (A) of A for three different rea-
sons. If there is a non-zero solution x ∈ B to Ax = λx, then λ is called an eigenvalue
(of A) and such an x ∈ B (x = 0) is an eigenvector (associated to the eigenvalue
λ). The set of eigenvalues forms the so-called point spectrum. In case the range of
λ Id −A is dense in B but the inverse is not bounded, the set of those λ’s forms the
continuous spectrum. Finally, it may happen that the range of λ Id −A is not dense,
but λ is not an eigenvalue. The set of those λ’s forms the residual spectrum.
On real Banach spaces, as is the case throughout this book, one has to look at the
complexification of the space. Fortunately, this is irrelevant for self-adjoint operators
on separable Hilbert spaces for which the spectrum is always real, and the residual
spectrum is empty.
Self-adjoint operators are quite similar to symmetric matrices in finite dimension
which may be diagonalized in orthonormal bases of eigenvectors associated to real
eigenvalues. Things are of course more intricate in infinite dimension where there
may not exist any eigenvectors. Think for example of the operator Af = −f on the
line R, which is self-adjoint on L2 (dx) for the Lebesgue measure dx (the space of
smooth compactly supported functions being a core), and for which eigenvectors are
not in L2 (dx). Diagonalization then has to be replaced by what is called a spectral
decomposition.
We restrict ourselves here to positive self-adjoint operators A on a (real, separa-
ble) Hilbert space H as described in the previous section.
A spectral decomposition is then an increasing family (Hλ )λ≥0 of closed linear
subspaces of H, right-continuous
! in the sense that λ >λ Hλ = Hλ . It is further-
more required that λ≥0 Hλ is dense in H. Consider then, for every λ ≥ 0, the
A.4 Spectral Decompositions 479
orthogonal projection Eλ onto Hλ . For any x ∈ H, the map λ → Eλ x is then right-

continuous and converges to x when λ converges to ∞. Furthermore, the map
λ → Eλ x2 = Eλ x, x
is bounded and increasing (and positive). Therefore, for any pair of points
(x, y) ∈ H × H, the map λ → Eλ x, y is right-continuous and of bounded vari-
ations as a difference of bounded increasing functions
1
Eλ x, y = Eλ (x + y), (x + y) − Eλ x, x − Eλ y, y .
2
Then given any bounded measurable function ψ : R+ → R, and any x, y ∈ H, one
may define the Stieltjes integral
∞
ψ(λ)dEλ x, y.
0
By convention, we assume that Eλ = {0} for λ < 0, so

∞that the previous integral
has to be understood, extending ψ by 0 on (−∞, 0), as −∞ ψ(λ)dEλ x, y, taking
into account the possible gap at λ = 0.
This construction defines by duality a bounded symmetric linear operator which
may be written symbolically as
∞
= ψ(λ) dEλ ,
0

∞
therefore satisfying x, y = 0 ψ(λ)dEλ x, y for every x, y ∈ H. When ψ is

∞
unbounded, the operator = 0 ψ(λ) dEλ is unbounded on the domain
" ∞ #
x ∈ H; ψ(λ) dEλ x, x < ∞ .
2
(A.4.1)
0
The main existence result is the following statement.
Theorem A.4.2 (Spectral decomposition) Given a positive self-adjoint operator A

on H, there exists a spectral decomposition (Eλ )λ≥0 such that
∞
A= λ dEλ .
0
For any measurable function ψ on R+ , one may then define the operator ψ(A) as
∞
ψ(A) = ψ(λ) dEλ
0

∞
on the domain (A.4.1) on which it satisfies ψ(A)x2 = 0 ψ(λ)2 dEλ x, x.
When A is a positive and self-adjoint operator, the semigroup Pt = e−tA , t ≥ 0,

with infinitesimal generator −A is defined via the spectral resolution (Eλ )λ≥0 by
∞
Pt = e−λt dEλ , t ≥ 0.
0
In the context of this monograph, it is convenient to feature a few norms as-

sociated with a Markov semigroup P = (Pt )t≥0 on (E, F, μ) with invariant and
symmetric measure μ, and its generator A = −L on L2 (μ) in terms of the preced-
ing spectral decomposition. (See Chaps. 1 and 3 for the full description of Markov
semigroups and their infinitesimal generators.) Some of the following identities and
inequalities are presented and used repeatedly throughout the various chapters. For
example, on suitable classes of functions f : E → R,
∞
(−Lf )2 dμ = λ2 dEλ f, f ,
E 0
and also
∞
2
E(f, f ) = f (−Lf )dμ = λdEλ f, f = (−L)1/2 f dμ.
E 0 E
On the semigroup Pt = et L , t ≥ 0, itself, for every t ≥ 0,

(Pt − Id)f 2 ≤ t 2 Lf 2 ,
and using that (1 − e−λt )2 ≤ 2λt,

(Pt − Id)f 2 ≤ 2t E(f, f ).
The relationship of the spectral decomposition Theorem A.4.2 with the spectrum
σ (A) of A as described in Definition A.4.1 is the following. For a positive real
number λ, λ ∈ σ (A) if and only if, for any ε > 0, the dimension of Hλ+ε is strictly
larger than the dimension of Hλ−ε . Furthermore,
! λ belongs to the point spectrum if
and only if the closed space Hλ− spanned by λ <λ Hλ is such that Hλ− = Hλ . In
that case, any point x ∈ Hλ which is orthogonal to Hλ− is an eigenvector of A, with
eigenvalue λ.
The spectrum σ (A) of a positive self-adjoint operator A may also be decomposed
in another way, much more useful in practice.
Definition A.4.3 (Essential spectrum) Let A be a positive self-adjoint operator

on H. A number λ ∈ σ (A) belongs to the essential spectrum σess (A) if, for any
ε > 0, the dimension of the orthogonal of Hλ−ε in Hλ+ε is infinite. The comple-
ment σd (A) = σ (A) \ σess (A) is called the discrete spectrum.
According to this definition, a point lies in the discrete spectrum if it is isolated

in the point spectrum and if the corresponding eigenspace is finite-dimensional. If
A.5 Essentially Self-adjoint Operators 481
λ ∈ σd (A), then the set of solutions x of the equation Ax = λx is non-zero and

finite-dimensional. The following useful criterion characterizes the elements of the
spectrum and of the essential spectrum.
Theorem A.4.4 (Weyl’s criterion) Let A be a positive self-adjoint operator with

domain D(A) ⊂ H. Then λ ∈ σ (A) if and only if for any ε > 0, there exists an
x ∈ D(A) with x = 1 such that Ax − λx < ε. An element λ ∈ σess (A) if and
only if, for any ε > 0, there exists an orthonormal sequence (xk )k∈N in D(A) ⊂ H
such that Axk − λxk < ε for every k ∈ N.
As an illustration, let H = L2 (dx), and let A = − where is the Laplace

operator (more precisely the closure of this operator when defined on smooth func-
tions with compact support). The operator A = − is self-adjoint and its spectral
decomposition is given by the family (Hλ )λ>0 where, for each λ > 0, Hλ√ is the
√set
of functions in L2 (dx) with Fourier transform compactly supported on [− λ, λ].
The point spectrum is empty and σess = [0, ∞).
As a consequence of Theorem A.4.4, when σess (A) is empty, the operator A is
diagonalizable. That is, there exist a sequence (λk )k∈N in σ (A) and an orthonormal
basis (ek )k∈N of D(A) ⊂ H such that Aek = λk ek for every k. Since A is positive,
λk ≥ 0, and it may be assumed for simplicity that the λk ’s are ordered increasingly
0 ≤ λ0 ≤ · · · ≤ λk ≤ · · · . It is actually more convenient to distinguish the different
eigenvalues, that is 0 ≤ λ0 < · · · < λk < · · · and to consider for any k ∈ N, the vec-
tor space Fk spanned by the eigenvectors associated to the eigenvalues λ0 , . . . , λk .
In this case, the spectral decomposition of A may be described in the following
simplified way. Namely, Ek is the orthogonal projection onto Fk and Eλ = Ek on
[λk , λk+1 ) and is constant between two jumps, which occur at the values λk . If
ψ : R+ → R, ψ(A) is the operator which maps ek into ψ(λk )ek , k ∈ N, which is the
natural way of thinking of ψ(A), for example when ψ is a polynomial function.
A.5 Essentially Self-adjoint Operators
As mentioned earlier, given a symmetric operator L on some dense linear space D0

of a Hilbert space H, a central question is whether the knowledge of L on D0 is
enough to determine a unique self-adjoint extension of L (or equivalently a unique
symmetric semigroup P = (Pt )t≥0 with generator L)? In other words, is D0 a core?
The discussion below concentrates on positive symmetric operators A (A = −L in
the case of Markov generators in the framework of this book).
In this setting, one may always consider the closure of D0 with respect to the
domain topology given by the norm (A.1.4) (already used when dealing with the
Friedrichs extension). This closure represents the minimal extension of A one may
think of. The adjoint operator A∗ of A and its domain may therefore be defined as
before and A∗ is always an extension of A. This leads to the following definition.
Definition A.5.1 (Essentially self-adjoint operator) In the preceding setting, a pos-

itive symmetric operator A is said to be essentially self-adjoint on D0 if the minimal
extension is self-adjoint.
Now let A be an unbounded positive symmetric operator defined on some dense

domain D(A) in a complex Hilbert space H. If A is symmetric, dim(Ker(λ Id −A∗ )),
as a function C → R, is constant on %(λ) > 0 and %(λ) < 0. Furthermore,
Ker(λ Id −A∗ ) is the orthogonal of Im(λ̄ Id −A). In general, choose two particu-
lar complex numbers, for example λ = ±i, and consider the kernels K+ and K− of
±i Id −A∗ . One fundamental (although quite easy) result is the following.
Theorem A.5.2 (Von Neumann’s Theorem) For a closed symmetric operator A

defined on D(A), the domain D(A∗ ) is D(A) + K+ + K− . In particular, A is self-
adjoint if and only if the two linear spaces K+ and K− are 0.
A useful criterion for an operator A defined on D0 to be essentially self-adjoint

is then given by the following property.
Proposition A.5.3 Let A be a positive symmetric operator defined on some dense

linear subspace D0 . Then A is essentially self-adjoint as soon as there exists a real
number λ such that λ Id −A∗ is injective.
Indeed, since the resolvent set is an open subset in C, λ Id −A∗ is still injective
for some imaginary complex number close to λ with positive or negative imaginary
part. Therefore, both K+ and K− are reduced to 0 and the conclusion follows from
Theorem A.5.2. In particular, in order to ensure self-adjointness, it is enough to
exhibit some real number λ for which the equation λx = A∗ x admits no non-zero
solutions (in other words, if x is such that λy, x = Ay, x and |y, x| ≤ Cy
for all y ∈ D0 , then x must be 0).
We conclude this section with a classical illustration. A lot of attention has been
paid over the last century to the class of so-called Schrödinger operators on Rn , that
is, operators of the form Af = −f + Vf , where is the standard Laplacian and
V is some L1loc -function on Rn . The main results on this family of operators are
summarized in the next statement.
Theorem A.5.4 Assume that V ≥ 0 and that V ∈ L2loc . Then A = − + V is sym-

metric in L2 with respect to the Lebesgue measure and essentially self-adjoint on
the class of smooth functions with compact support.
The statement may be further completed with a description of the bottom of the
p
essential spectrum. That is, as soon as V ∈ Lloc with p > n2 and
lim essinf|x|>r V (x) > −∞,

r→∞
A.6 Compact and Hilbert-Schmidt Operators 483
then

n f Af dx
inf σess (A) = sup inf
R
∞ 2
K⊂Rn f ∈Cc (K ),f =0 Rn f dx
c
where K ranges among all compact subsets of Rn and where Cc∞ (K c ) denotes the
class of smooth functions compactly supported in K c = Rn \ K. Operators A on
Rn sharing this property are called Persson operators in the literature and aspects
of their analysis are discussed in Sect. 4.10.3, p. 227. In particular, for such V , the
spectrum is discrete as soon as V goes to ∞ at infinity.
A.6 Compact and Hilbert-Schmidt Operators
When B1 and B2 are Banach spaces, a linear operator A : B1 → B2 is said to be

compact if the image of the unit ball (of B1 ) is relatively compact (in B2 ). If an
operator A may be written as QK or KQ, with Q bounded and K compact, then
it is compact. In addition, a strong limit of compact operators is compact. In other
words, if (Ak )k∈N is a sequence of compact operators and if limk→∞ Ak − A = 0
(for the operator norm), then A is compact. Furthermore, if the image of an operator
is finite-dimensional (a finite range operator), this operator is compact. Actually, on
a Hilbert space a compact operator is always a strong limit of finite range operators.
In the context of this book, the following statement will be relevant.
Theorem A.6.1 For a compact symmetric operator on a Hilbert space, the spec-
trum consists of a sequence of eigenvalues (λk )k∈N , the only limit point possibly not
in the discrete spectrum being 0. The non-zero eigenvalues form a (possibly finite)
sequence converging to 0 at infinity, have finite-dimensional eigenspaces, and there
is an orthonormal basis of eigenvectors. Conversely, an operator with a sequence of
eigenvalues converging to 0 is compact.
Hilbert-Schmidt operators are significant examples of compact symmetric oper-

ators.
Definition A.6.2 (Hilbert-Schmidt operator) A symmetric (bounded) operator K

defined on a separable Hilbert space H is Hilbert-Schmidt if, for some orthonormal
basis (ek )k∈N of H,

Kek 2 < ∞.
k∈N
This property is independent of the chosen basis.
Proposition A.6.3 Hilbert-Schmidt operators are compact.

Hilbert-Schmidt
operators are thus operators for which the sequence (λk )k∈N of
eigenvalues satisfy k∈N λ2k < ∞. When the Hilbert space is L2 (μ) over a mea-
sure space (E, F, μ), a Hilbert-Schmidt operator K may be represented by a kernel
k(x, y) ∈ L2 (μ ⊗ μ) such that

Kf (x) = k(x, y)f (y)dμ(y), x ∈ E,
E
for every f ∈ L2 (μ). Indeed, if (e )∈N is a sequence

of eigenvectors associated
with the eigenvalues (λ )∈N , then k(x, y) = ∈N λ e (x)e (y), and the fact
that K is Hilbert-Schmidt is equivalent to the fact that this series converges in
L2 (μ ⊗ μ). Conversely, such operators K represented by L2 (μ ⊗ μ)-kernels are
compact. Indeed, given a sequence (fm )m∈N bounded in L2 (μ), there exists a sub-
sequence (fmp )p∈N which converges weakly in L2 (μ). Then Kfmp (x) converges as

p → ∞ for every x ∈ E such that E k 2 (x, y) dμ(y) < ∞ and

1/2

Kfm (x) ≤ C 2
k (x, y) dμ(y)
p
E
where C = supm∈N fm 2 < ∞. Since k(x, y) ∈ L2 (μ ⊗ μ), it follows that Kfmp
does indeed converge in L2 (μ). By Theorem A.6.1, there exists an orthonormal
basis of eigenvectors of L2 (μ) with corresponding sequence of eigenvalues (λ )∈N .
Using standard Hilbertian tools, it is easily seen that

λ2 = k 2 (x, y)dμ(x)dμ(y)
∈N E E
from which the claim follows.

These compact operators should be distinguished from those defined by some
L2 (μ)-kernel as described for example
in Definition 1.2.4, p. 14, where it is only
required that for μ-almost every x ∈ E, E k 2 (x, y)dμ(y) < ∞.
In the context of this book, the generators L of Markov semigroups P = (Pt )t≥0
are in general not Hilbert-Schmidt (they are not even bounded) whereas, for t > 0,
the semigroup Pt may be Hilbert-Schmidt and therefore compact. The following
statement describes conditions for Pt to be compact.
Theorem A.6.4 For a symmetric semigroup P = (Pt )t≥0 with (unbounded) in-
finitesimal generator L such that −L is positive with domain D(L), the following
are equivalent:
(i) σess (−L) = ∅.
(ii) For all, or some, t > 0, Pt is compact.
(iii) For all, or some, λ > 0, the resolvent (λ Id −L)−1 is compact.
At the level of the spectral decomposition, the equivalent conditions in Theo-

rem A.6.4 amount to the fact that the eigenvalues of −L go to infinity. In par-
ticular, when Pt is compact for some t > 0, the spectrum of −L is discrete and
A.7 Notes and References 485
forms a sequence of eigenvalues (λk )k∈N going to infinity, associated to eigenspaces

k with finite
E dimension nk , k ∈ N. Then, Pt is Hilbert-Schmidt if and only if
n e −2λk t < ∞.
k∈N k
A.7 Notes and References

As mentioned earlier, the material briefly presented here forms a classical part of
functional analysis. General references on these topics, used during the preparation
of this appendix, are [95, 96, 100, 143, 154, 168, 169, 189, 190, 205, 241, 264, 344,
355, 356, 359, 444]. More recent references include, among many others, [142, 174,
244, 419].
Somewhat more specifically, a standard reference on the Hille-Yosida theory is
the monograph [444] by K. Yosida. Most of the material on symmetric and self-
adjoint operators may be found in the volumes [355] by M. Reed and B. Simon
or [169] by N. Dunford and J.T. Schwartz. The definition of essential spectrum
seems to go back to H. Weyl.
Appendix B
Elements of Stochastic Calculus
Stochastic calculus is a wide topic. For the reader not familiar with advanced prob-
ability theory, we briefly present in this appendix some basic notions adapted to the
calculus related to Brownian motion. This short exposition surveys stochastic inte-
grals with respect to Brownian motion, Itô’s formula, stochastic differential equa-
tions and diffusion processes. In particular, these elements are aimed at a better
understanding of the links with diffusion operators, semigroups and heat kernels as
informally described in Chap. 1. The Notes and References include pointers to the
literature and more complete accounts.
B.1 Brownian Motion and Stochastic Integrals
A real-valued Brownian motion defined on some probability space ( , , P)

is a family (Bt )t≥0 of real-valued random variables (i.e. measurable maps)
Bt : → R, t ≥ 0, with the properties that B0 = 0, for almost all ω ∈ the map
t → Bt (ω) is continuous, and for any 0 < t1 < · · · < tk , the random variables
Bt1 , Bt2 − Bt1 , . . . , Btk − Btk−1 are independent centered Gaussian random variables
with respective variances t1 , t2 − t1 , . . . , tk − tk−1 . That is, the distribution of the
random vector (Bt1 , Bt2 − Bt1 , . . . Btk − Btk−1 ) has density

1 1 x12 xk2
exp − + ··· +
(2π)k/2 [t1 (t2 − t1 ) · · · (tk − tk−1 )]1/2 2 t1 tk − tk−1
with respect to the Lebesgue measure on Rk .

Brownian motion (Bt )t≥0 with values in Rn is constructed similarly, the random
vectors Bt1 , Bt2 − Bt1 , . . . Btk − Btk−1 now being independent centered Gaussian vec-
tors with respective covariance matrices
t1 Id, (t2 − t1 ) Id, . . . , (tk − tk−1 ) Id

DOI 10.1007/978-3-319-00227-9,
488 B Elements of Stochastic Calculus
where Id is the identity matrix of dimension n. Equivalently, such a Brownian mo-

tion (Bt )t≥0 in Rn may be described as Bt = (Bt1 , . . . , Btn ), t ≥ 0, where (Bti )t≥0 ,
1 ≤ i ≤ n, is a collection of n independent one-dimensional Brownian motions.
It is not obvious that such a process (Bt )t≥0 exists, in particular with respect to
continuity of the trajectories. From a functional analysis point of view, we may as
well consider that such a stochastic process (Bt )t≥0 induces a probability measure
on the space C([0, ∞)) of continuous functions on [0, ∞) with values in R or Rn
(that is, we may always choose = C([0, ∞)) and set Bt (ω) to be ω(t)). This
Gaussian measure on the Borel sets of C([0, ∞)) is known as the Wiener measure.
Although continuous, the functions t → Bt (ω) are very irregular. With probabil-
ity 1, they do not have a bounded variation on any interval. In particular, expressions
like dB t
dt do not make sense. Nevertheless, there is a way to define stochastic inte-
grals, that is expressions like
t
Hs (ω)dBs (ω)
0
for some stochastic processes Hs (ω) provided they “depend only on the past”.
To understand this notion, it is necessary to introduce the so-called filtration
Ft , t ≥ 0, of Brownian motion consisting of the sub-σ -fields of defined as
Ft = σ (Bs ; 0 ≤ s ≤ t) with, by convention, F∞ = σ (Bs ; s ≥ 0). In other words,
Ft is the smallest σ -field for which all the random variables Bs for s ≤ t are mea-
surable, and thus contains all information up to time t (as far as observations on
the process (Bs )s≥0 are concerned). The collection (Ft )t≥0 defines an increasing
family of σ -fields and, for technical reasons, it is usually enlarged so to contain all
the elements of F∞ which have 0 probability. Sometimes, the σ -field Ft for every
t is also enlarged in order to include non-constant variables in F0 , which are in gen-
eral independent of F∞ , and this is systematically the case when solving stochastic
differential equations with non-constant initial data.
Definition B.1.1 (Martingale) A martingale (relative to the filtration (Ft )t≥0 ) is a

real-valued process (Mt )t≥0 defined on ( , , P), such that for any t ≥ 0, Mt is
P-integrable, Mt is Ft -measurable (the process (Mt )t≥0 is then said to be adapted
to the filtration (Ft )t≥0 ) and for all s < t,
Ms = E(Mt | Fs ).
Here E(Mt | Fs ) denotes the conditional expectation of the integrable random

variable Mt with respect to the σ -field Fs (defined almost surely). In particular (and
this is an important feature of the martingale property), for a martingale (Mt )t≥0 ,
E(Mt ) is constant and E(Mt ) = E(M0 ) for every t > 0.
The Brownian motion itself is a martingale. Martingales may be constructed in
many ways. For example, if X is an integrable random variable on ( , , P), then
Mt = E(X | Ft ), t ≥ 0, is a martingale (with respect to (Ft )t≥0 ). Many martingales
may be represented in this way. In fact, if for some p > 1, supt E(|Mt |p ) < ∞,
then there exists a random variable M∞ ∈ Lp (P) such that Mt = E(M∞ | Ft ) for
B.1 Brownian Motion and Stochastic Integrals 489
every t ≥ 0, and moreover Mt converges to M∞ as t → ∞ (almost surely and in

Lp (P)). A specific property of the filtration (Ft )t≥0 of Brownian motion is that
any martingale relative to it is continuous, which makes the analysis of processes
constructed with Brownian motion much more pleasant than the general theory.
One further way to construct martingales is via the stochastic integrals (B.1.1).
Such integrals are defined for any processes Hs (ω), s ≥ 0, ω ∈ , assumed to be
continuous (more general conditions are available), adapted to the filtration (Ft )t≥0
in the sense that ω → Hs is Fs -measurable for any s, and square integrable
t
E Hs2 ds < ∞
0
(for every t ≥ 0). Given such a process H = (Hs )s≥0 , the stochastic integrals
t
Xt = Hs dBs , t ≥ 0, (B.1.1)
0
define a new stochastic process, again with continuous paths (even if H itself is not
continuous). Although the integral Xt does not make sense for any particular choice
of ω since the map t → Bt (ω) has no bounded variation on any interval, the usual
way to define Xt is to consider it as a suitable limit of Riemann sums. Indeed, start
with Hs (ω) piecewise constant in the sense that

p−1
Hs (ω) = Ki (ω)1(ti ,ti+1 ] (s)
i=0
for some finite sequence 0 = t0 ≤ t1 < · · · < tp and random variables K1 , . . . , Kp−1 .
In order for this process to be adapted, assume that Ki is Fti -measurable for every
i, and of course in L2 (P). In this case, one directly defines
∞
p−1
Hs dBs = Ki (Bti+1 − Bti )
0 i=0
and
∞
Xt = Hs 1[0,t] (s)dBs , t ≥ 0.
0
It is easily checked that the process t → Xt is a martingale and that

t
E Xt2 = E Hs2 ds
0
for every t ≥ 0. This identity thus defines an isometry between the piecewise con-
stant processes H and L2 (P) random variables, which may then be extended by
L2 (P)-continuity to general (continuous) processes H .
Via this construction, the resulting process (Xt )t≥0 is then a continuous square
integrable martingale (with respect to the filtration (Ft )t≥0 ). In general X0 = 0. It is
however convenient to add some F0 -measurable random variable X0 and to consider
the process
t
Xt = X0 + Hs dBs , t ≥ 0,
0
with initial value X0 . These stochastic integrals are in general as irregular (in the
time variable) as the original Brownian motion (Bt )t≥0 itself.
A simple special case is obtained with the choice of non-random

∞ functions
Hs = hs , s ≥ 0. In this case, the random variable X(h) = 0 hs dBs , known as

∞
a Wiener integral, is a Gaussian variable with mean 0 and variance 0 h2s ds (pro-
vided it is finite). The covariance of two such variables X(h) and X(k) is given
similarly by
∞

E X(h)X(k) = hs ks ds.
0
This identity thus yields an embedding of L2 ([0, ∞), dt) into the L2 (P)-space of
random variables defined on ( , , P) giving rise to the so-called reproducing ker-
nel Hilbert space or Cameron-Martin Hilbert space associated to Brownian motion.
More on the Cameron-Martin Hilbert space in the context of this book may be found
in Sect. 2.7.2, p. 108.
In general, it is not enough to construct stochastic integrals only for square inte-
grable processes H and it is required to “localize” this procedure. This localization
procedure is performed with the notion of stopping time.
Definition B.1.2 (Stopping time) A stopping time, with respect to the filtration
(Ft )t≥0 , is a random variable T : → [0, ∞] such that, for any t ≥ 0, the set
{T ≤ t} = {ω ∈ ; T (ω) ≤ t} is Ft -measurable.
It is important to allow stopping times to take the value +∞. A typical stopping
time is

T (ω) = inf t ≥ 0 ; Xt (ω) ∈ A , ω ∈ ,
where (Xt )t≥0 is a continuous process (with values in R or Rn , for example) adapted
to the filtration (Ft )t≥0 and A any Borel set in R or Rn (with the usual convention
that inf{∅} = +∞).
Associated with a stopping time T , one introduces the σ -field FT of events which
take place before T , defined as

FT = A ∈ F∞ ; A ∩ {T ≤ t} ∈ Ft for every t ≥ 0 .
Clearly T is FT -measurable.
When (Xt )t≥0 is an adapted stochastic process (with respect to (Ft )t≥0 ), this
process may be stopped at any stopping time T (for (Ft )t≥0 ) by setting
Xt∧T (ω) = X(ω)min(t,T (ω)) , t ≥ 0, ω ∈ .
B.2 The Itô Formula 491
Set XT = X(ω)T (ω) , being careful with this expression when T (ω) = ∞ (in gen-
eral assume that some X∞ random variable is given). This procedure thus pro-
duces a new adapted stochastic process (Xt∧T )t≥0 , and it is a basic result, known
as Doob’s stopping time Theorem, that if (Xt )t≥0 is a continuous martingale, so
is the new process (Xt∧T )t≥0 (relative to the underlying filtration (Ft )t≥0 ). More-
over, when Xt = E(X∞ | Ft ), t ≥ 0, for some integrable random variable X∞ , then
XT = E(X∞ | FT ).
Local martingales are then defined as adapted processes (Xt )t≥0 for which there
exists an increasing sequence (Tk )k∈N of stopping times (relative to (Ft )t≥0 ) which
converges to ∞ and for which the processes (Xt∧Tk )t≥0 , k ∈ N, are martingales.
Stochastic integrals of the type (B.1.1)
t
Hs dXs , t ≥ 0,
0
with respect to such a local martingale (Xt )t≥0 are then defined for (continuous)
adapted processes (Hs )s≥0 provided that
Tk
E Hs2 ds <∞
0
for any k. The resulting stochastic integral is then again a local martingale. This
construction yields much more freedom to define stochastic integrals (and later so-
lutions of stochastic differential equations) although one can no longer be sure that
a local martingale (Xt )t≥0 has a constant expectation. To reach this conclusion usu-
ally requires further integrability properties, such as supt≥0 E(|Xt |p ) < ∞ for some
p > 1.
B.2 The Itô Formula

The fundamental property of the stochastic calculus is the famous Itô formula which
expresses a change of variables adapted to stochastic integration.
Theorem B.2.1 (Itô’s formula) Given stochastic integrals

t
Xt = X0 + Hs dBs , t ≥ 0,
0
as constructed before, if f : R → R is a C 2 function, then
f (Xt ) = f (X0 ) + Mt + At , t ≥ 0,
where
t t
1
Mt = f (Xs )Hs dBs and At = f (Xs )Hs2 ds.
0 2 0
In particular, (Mt )t≥0 is a local martingale and (At )t≥0 has locally bounded varia-
tions.
In other terms, using the differential notation dXt = Ht dBt , and introducing the
“bracket”
dX, Xt = Ht2 dt,
Itô’s formula indicates that for a C 2 function f : R → R,
1
df (Xt ) = f (Xt )dXt + f (Xt )dX, Xt .
2
The quadratic notation dX, X is natural under the interpretation
dX, Xt = dX, dXt = H dB, H dBt = Ht2 dt
(in particular dB, dBt = dt).
t There is a more intrinsic definition of the increasing process X, Xt =

0 dX, Xs , t ≥ 0, as the limit under subdivisions 0 = t0 < t1 < · · · < tk = t of
the interval [0, t] of

k−1
(Xti+1 − Xti )2
i=0
when the mesh of the subdivision goes to 0. It is easily seen that such a quantity is
0 as soon as (Xt )t≥0 is a continuous bounded variation process (and therefore this
bracket extracts the purely martingale part of the process (Xt )t≥0 in the same way
as the carré du champ operator of Definition 1.4.2, p. 20, extracts the second order
part of a differential operator L).
The change of variables formula expressed by Itô’s formula is valid for the larger
class of so-called semi-martingales. A process (Xt )t≥0 is called a semi-martingale
(with respect to a given filtration (Ft )t≥0 ) if it may be decomposed into a sum
Xt = X0 + Mt + At , t ≥ 0,
where (Mt )t≥0 is a local martingale and (At )t≥0 has locally bounded variations. For
example, as illustrated by Itô’s formula, a process (Xt )t≥0 given by
t t
Xt = X0 + Hs dBs + Ks ds, t ≥ 0, (B.2.1)
0 0

t
for suitable stochastic processes (Hs )s≥0 and (Ks )s≥0 such that 0 Hs dBs , t ≥ 0, is

t
a local martingale and 0 Ks ds, t ≥ 0, has bounded variations, is a semi-martingale.
Replacing in Itô’s formula dXt = Ht dBt by dXt = dMt + dAt shows that the class
of semi-martingales is stable under the action of C 2 functions.
B.3 Stochastic Differential Equations 493
One may also consider stochastic integrals driven by semi-martingales, for ex-
ample in the framework of example (B.2.1), set, for an adapted process (Rs )s≥0 ,
t t t
(R · X)t = Rs dXt = Rs Hs dBs + Rs Ks ds, t ≥ 0.
0 0 0
The differential rules then take the form
d(R · X)t = Rt dXt , dR · X, R · Xt = Rt2 dX, Xt = Rt2 Ht2 dt.
The Itô chain rule formula is thus quite different than the formula for differentiable
processes for which the second order term vanishes. The full strength of stochastic
calculus relies on this specific feature, which deeply connects stochastic integration
to second order differential operators.
When dealing with multivariate processes, introduce (by polarization) for a pair
(Xt )t≥0 , (Yt )t≥0 of semi-martingales, the bounded variation process
1
dX, Y t = dX + Y, X + Y t − dX, Xt − dY, Y t , t ≥ 0.
2
The multi-variable form of Itô’s chain rule formula for a vector Xt = (Xt1 , . . . , Xtn ),
t ≥ 0, of semi-martingales and f : Rk → R a C 2 function is then

k
1 2
k

df (Xt ) = ∂i f (Xt )dYti + ∂ij f (Xt )d X i , X j t .
2
i=1 i,j =1
This formula is of course closely related to the change of variables formula for
diffusion processes (1.11.4), p. 43, in Chap. 1.
B.3 Stochastic Differential Equations
Stochastic integrals lead to the solution of stochastic differential equations. Stochas-

tic differential equations provide a natural framework for the probabilistic anal-
ysis of diffusions processes and their Markov generators as further developed in
Sect. B.4 (see also Sect. 1.10, p. 38).
In dimension one, the simplest stochastic differential equation appears in the
form
dXt = σ (Xt )dBt + b(Xt )dt, X0 = x,
where σ and b are smooth functions on the real line (random coefficients σ may
also be considered although they are not addressed here). In other words,
t t
Xt = X0 + σ (Xs )dBs + b(Xs )ds, t ≥ 0.
0 0
Stochastic calculus and stochastic differential equations related to the study of

diffusion semigroups deal mostly with vector-valued Brownian motion
p
Bt = (Bt1 , . . . , Bt ), t ≥ 0, thus consisting of p independent copies of the standard
Brownian motion. Stochastic integrals with respect to those p independent Brown-
ian motions may be considered simultaneously and the corresponding bracket rule
is dB i , B j t = δ ij dt, 1 ≤ i, j ≤ p. The prototype stochastic differential equation
driven by such a multi-dimensional Brownian motion (Bt )t≥0 is then described as a
vector-valued process Xt = (Xt1 , . . . , Xtn ), t ≥ 0, which is a solution of
j

p
j j
dXt = σi (Xt )dBti + bj (Xt )dt, X0 = x j , 1 ≤ j ≤ n.
i=1
j
Here σ = σ (x) = (σi (x))1≤i,j ≤n is an n × p matrix depending on x ∈ Rn and
b = b(x) = (bj (x))1≤j ≤n is a vector depending on x ∈ Rn satisfying suitable growth
and smoothness assumptions (see Theorem B.3.1 below). The preceding equation is
then summarized more simply in the form
dXt = σ (Xt )dBt + b(Xt )dt, X0 = x, (B.3.1)
and the solution is the vector-valued process denoted by {Xtx ; t ≥ 0, x ∈ Rn } to de-
scribe dependence on the initial condition x (there should be no confusion with the
j -th coordinate of Xt ).
There are many ways in which existence and uniqueness of such stochastic differ-
ential equations may be discussed. The existence may be thought of as the existence
of a Brownian motion (Bt )t≥0 on some probability space ( , , P) such that this
holds. The uniqueness may also be regarded in terms of the uniqueness of the laws of
the processes {Xtx ; t ≥ 0, x ∈ Rn }. The discussion here is restricted to the simplest
notions. Say that {Xtx ; t ≥ 0, x ∈ Rn } is a solution of (B.3.1) if it satisfies the equa-
tion path-wise (that is for almost every ω ∈ in the probability space on which the
Brownian motion is defined). Moreover, say that the solution is (path-wise) unique
if for any initial value x, two solutions X and X of (B.3.1) are such that, outside a
set of probability 0, the maps t → Xt and t → Xt coincide. In this context, the main
conclusions may be described by the following statement. There, · denotes any
norm on points, matrices and vectors which are compatible with the usual topology.
Theorem B.3.1 (Existence and uniqueness of solutions of stochastic differential

equations) Assume that the maps x → σ (x) and x → b(x) on Rn have a linear
growth

σ (x) + b(x) ≤ D 1 + x (B.3.2)
for some finite constant D and all x ∈ Rn and are locally Lipschitz in the sense that
for any compact set K ⊂ Rn , there exists a constant CK > 0 such that

σ (x) − σ (y) + b(x) − b(y) ≤ CK x − y (B.3.3)
for all x, y ∈ K. Then, there exists a unique solution {Xtx ; t ≥ 0, x ∈ Rn } of (B.3.1)
defined on the time interval [0, ∞).
B.4 Diffusion Processes 495
Solutions {Xtx ; t ≥ 0, x ∈ Rn } of stochastic differential equations (B.3.1) satisfy

many properties, in particular the strong Markov property. This property indicates
that for any finite stopping time T of the given filtration (Ft )t≥0 , the law of the
process (Yt = XT +t )t≥0 conditioned on FT only depends on XT itself, and this
law conditioned to {XT = x} is the law of (Xtx )t≥0 . The process (Yt )t≥0 is thus a
solution of the same stochastic differential equation (B.3.1) but with initial value
Y0 = XT (with of course some other Brownian motion driving the equation, namely
(BT +t − BT )t≥0 ). In other words, the process starts afresh at any stopping time. This
strong Markov property also allows for the definition of solutions of such stochas-
tic differential equations up to a given stopping time. As will be developed below,
this is an important tool in the investigation of stochastic differential equations on
manifolds where there is no global chart.
With the latter task in mind, Theorem B.3.1 is not always enough. For example,
it is often necessary to work on some open subset O in Rn and to define the solution
up to the stopping time T which is the time at which the process leaves the set O.
In this case, only the Lipschitz property (B.3.3) is necessary, and uniqueness and
existence then only hold up to this stopping time which is called the explosion time
of the process. When O is Rn itself, this explosion time is the time at which the
process reaches infinity, which could be finite whenever (B.3.2) is not satisfied.
The preceding processes solutions of stochastic differential equations may also
be constructed on manifolds, not only on open sets in Rn , replacing σ and b by a
system of vector fields. This is usually achieved in local charts via localization as
briefly outlined below in Sect. B.4.
B.4 Diffusion Processes
The link between stochastic differential equations and diffusion processes is one of
the main features of this monograph. Solutions of stochastic differential equations
with smooth coefficients define stochastic processes driven by second order differ-
ential operators. Solutions of the associated heat equation give rise to semigroups of
operators and heat kernels.
Let {Xtx ; t ≥ 0, x ∈ Rn } be a solution of the stochastic differential equa-
tion (B.3.1) with coefficients σ and b and driven by a p-dimensional Brownian
motion (Bt )t≥0 . For every smooth function f : Rn → R, by Itô’s formula,

p
n
j
df (Xt ) = σi (Xt )∂j f (Xt )dBti + Lf (Xt )dt, X0 = x,
i=1 j =1
where L is the second order operator
1 i j 2
n p n
Lf = σk σk ∂ij f + bi ∂i f.
2
i,j =1 k=1 i=1
Now, the first term in the expression of df (Xt ) gives rise to a local martingale
f
(Mt )t≥0 so that

f
t
f Xtx = Mt + Lf Xsx ds, t ≥ 0, x ∈ Rn ,
0
and these processes are therefore always semi-martingales. The process

{Xtx ; t ≥ 0, x ∈ Rn } is then said to solve the martingale problem associated with
f
the operator L. As soon as we know that the local martingale (Mt )t≥0 is indeed a
true martingale, for example as soon as f and Lf are bounded functions, it follows
by taking expectations that, for every t ≥ 0,

x 1 t x
E f Xt = f (x) + E Lf Xs ds.
2 0
This formula tells us that the process {Xtx ; t ≥ 0, x ∈ Rn } is a so-called diffusion
process with generator 12 L. In other words, setting

Pt f (x) = E f Xtx = E f (Xt ) | X0 = x , t ≥ 0, x ∈ Rn ,
it holds that
t
1
Pt f (x) = f (x) + Ps Lf (x)ds, t ≥ 0, x ∈ Rn ,
2 0
and therefore Pt f the solution of the (parabolic) heat equation ∂t = 12 L starting

from f . This construction thus produces a process whose law represents the heat
kernels associated with the generator 12 L. It should be mentioned that throughout
this monograph, in spite of this probabilistic representation, we work with the gen-
erator L rather than 12 L.
The Stratonovich integral is a way to get simpler and easier expressions in the
preceding when dealing with operators given as a sum of squares of vector fields (in
Hörmander form, cf. (1.10.2), p. 41). The associated Stratonovich notation ◦ is used
when both the integrator (Yt )t≥0 and the integrand (Ht )t≥0 are semi-martingales in
which case the Stratonovich integral reads
t t
1 t
Xt = X0 + Hs ◦ dYs = X0 + Hs dYs + dH, Y s , t ≥ 0.
0 0 2 0
Theorems on the existence and uniqueness of solutions of stochastic differential
equations under the Stratonovich integral are exactly the same as those under the Itô
form. Concerning the change of variables formula in the Stratonovich formalism, if
for example the real-valued diffusion process (Xt )t≥0 solves dXt = σ (Xt ) ◦ dBt ,
then for any smooth f ,
1
df (Xt ) = f (Xt )σ (Xt )dBt + f (Xt )σ (Xt )σ (Xt ) + f (Xt )σ 2 (Xt ) dt.
2
B.4 Diffusion Processes 497
As mentioned above, the Stratonovich formulation is a much more convenient

way to represent the operator L in the Hörmander form
1 2
d
L= Zj + Z0
2
j =1
where, for j = 0, 1, . . . , d, Zj is a vector field (in other words, the vector Zj (x) =
(Z 1 (x), . . . , Zjn (x)) is identified to the first order differential operator Zj f =
nj i
i=1 Zj ∂i f ). Indeed, in this case, as soon as

d
j
dXt = Zj (Xt ) ◦ dBt + Z0 (Xt )dt
j =1
j
where (Bt )t≥0 , j = 1, . . . , d, are independent Brownian motions, the Itô formula
immediately yields that
t
f
f (Xt ) = Mt + Lf (Xs )ds
0
f
where (Mt )t≥0 is the local martingale
f

d
j
dMt = ∂j f (Xt )Zj (Xt )dBt .
j =1
The process (Xt )t≥0 thus similarly solves the martingale problem with respect to
the operator L.
When dealing with stochastic processes and stochastic differential equations on
a smooth manifold, one usually uses the localization procedure. In some open set
O on which there exists a local chart, one solves the stochastic differential equation
up to the stopping time T = inf{t ≥ 0 ; Xt ∈ / O} and then changes the local chart.
The change of variable laws according to Itô’s formula assert that the result is in-
dependent of the choice of coordinate system. Then, the strong Markov property
at time T is used to start the process again in the new chart with initial value XT .
In such a way, the diffusion process is well defined as long as it does not reach
the boundary of the manifold (or infinity), and then specific arguments are required
to analyze whether or not the process reaches this boundary in finite time or not.
(However, the global growth condition of Theorem B.3.1 provides a useful criterion
towards non-explosion.) When dealing with Laplace operators on compact or com-
plete Riemannian manifolds, there exist some more global procedures (embedding
the manifold in some Euclidean space with larger dimensions, lifting the process to
the frame fiber bundle etc.), each of them having its specific advantages.
As already mentioned, solutions of stochastic differential equations are very ir-
regular in the time variable t (as irregular as the Brownian paths). However, they
depend smoothly on the initial condition x. In fact, solving for example the stochas-
tic differential equation

d
j
dXt = Zj (Xt ) ◦ dBt + Z0 (Xt )dt, X0 = x,
j =1
for vector fields Z0 , Z1 , . . . , Zd , the solution {Xtx ; t ≥ 0, x ∈ Rn } is as smooth in the

variable x as the vector fields Zj are. For example, if every Zj is C k for some k ≥ 2,
so is the process (Xtx )t≥0 , as long as the solution exists on some time interval [0, T ]
(which may be random when T is a stopping time). This explains why the laws
of the random variables Xtx , described by Markov kernels pt (x, dy) (cf. Sect. 1.2,
p. 9), are in general as smooth as the vector fields Zj themselves. On the other hand,
to study densities pt (x, y) with respect to some fixed reference measure dm(y),
the regularity in the y variable requires more hypotheses on the vector fields Zj ,
for example ellipticity (or hypo-ellipticity, cf. Sect. 1.12, p. 49). The probabilistic
analysis of those densities in terms of the variable y (especially the mere existence of
those densities) requires more sophisticated tools, for example Malliavin calculus,
or stochastic calculus of variation, which go far beyond this elementary introduction.
B.5 Notes and References

The material presented in this short appendix is rather standard and may be found
in any classic reference on stochastic calculus, including [132, 153, 252, 255, 263,
336, 350, 358, 362, 363] etc. For accounts of the stochastic calculus of variations
(Malliavin calculus), see [142, 297, 334].
Appendix C
Basic Notions in Differential and Riemannian
Geometry
This appendix presents some basic notions of differential geometry which are used
in many parts of this book. Even for the reader only interested in diffusions in Rn
or open sets in Rn , it may sometimes be useful to consider the case of manifolds.
For example, it is much easier to obtain the optimal Sobolev inequality on spheres
than on Euclidean spaces, and then to carry those optimal inequalities from spheres
to Rn by conformal transformations as described in Sect. 6.9, p. 313. It is also true
that the analysis of Markov semigroups on compact Riemannian manifolds without
boundaries (like spheres, toruses and so on) is in many respects much easier than the
analysis of the same objects on Rn or open sets of Rn . (This is somewhat similar to
the difference in the analysis of Markov chains on a finite set or on an infinite set.)
Unfortunately, on manifolds there is no way to locate a point through a unique chart,
thus explaining why it is necessary to develop an (apparently) somewhat heavy ma-
chinery. Finally, there are some computation rules which have a natural interpreta-
tion in terms of differential geometry, even if only the case of Rn is considered (such
as connections, curvature tensors, 2 operators etc.).
This appendix thus briefly presents the language of differentiable and Rieman-
nian manifolds, and some of the classical rules. It also introduces the notion of
Ricci curvature and its connection with the curvature-dimension conditions exten-
sively used throughout this work. It should be mentioned in particular that the last
two sections outline arguments and results on the 2 operator and the reinforced
curvature-dimension condition, which are central to this book. As for the previous
appendices, this appendix is far from a complete and formal exposition of differen-
tiable and Riemannian manifolds and only emphasizes some basic elements neces-
sary for the further developments. The reader is referred to standard textbooks on the
subject for complete details, some of which are listed in the Notes and References.
In several places in this appendix, as well as in the corresponding parts of this
book, index conventions are used according to the covariant or contravariant rules of
differential geometry. (In particular, coordinates of vectors are indicated with upper
indices. For simplicity, in the core of the book, lower indices are usually used.) The
j
Kronecker symbol δi is equal to 1 if i = j , and 0 otherwise. The Einstein notation
summarizes a summation over an index when it appears twice, one time up, one time

DOI 10.1007/978-3-319-00227-9,
500 C Basic Notions in Differential and Riemannian Geometry
down (usually from 1 to the dimension of the underlying structure). For example,
Z i ei is a shorthand for ni=1 Z i ei . In the same spirit

n
g ij ei ej = g ij ei ej .
i,j =1
C.1 Differentiable Manifolds
An n-dimensional differentiable manifold is a topological space in which a point is

located through coordinates x ∈ Rn , called a chart. It is not always possible to spec-
ify such a localization globally so sometimes points might be located in two differ-
ent charts. For example, the unit sphere Sn ⊂ Rn+1 is an n-differentiable manifold.
A point x = (x 1 , . . . , x n+1 ) ∈ Sn such that x n+1 > 0 and x 1 > 0 may be located
through its projection on the hyperplane {x n+1 = 0}, but also through its projection
on {x 1 = 0}. Those two projections are in fact points in the n-dimensional open unit
ball (in Rn ). It is then necessary to establish a correspondence between these two
points to assert that they represent the same point on Sn . In this example, a point
(y 1 , . . . , y n ) in the unit ball corresponds to

2 2
y1, . . . , yn, 1 − y1 − · · · − yn ∈ Sn
in the first projection, and to

2 2
1 − y 1 − · · · − y n , y 1 , . . . , y n ∈ Sn
in the second one. Therefore, this correspondence defines a map ψ from the in-
tersection of the unit ball with {y 1 > 0} into the intersection of the unit ball with
{y n > 0} given by

1 2 2 2
ψ y , . . . , y = y , . . . , y , 1 − y1 − · · · − yn
n n
.
The key point in differential geometry is to be able to perform computations

which are independent of the chosen chart. Therefore, objects from one chart to an-
other have to be identified, and some basic rules when changing charts have to be
prescribed. The conclusions should then be independent of such changes of coordi-
nates. For example, when solving stochastic differential equations, one has to make
sure that the solution (or at least its distribution) does not depend on the chosen
chart (cf. Sect. B.4). According to these requirements, one definition of differen-
tiable manifolds is the following.
Definition C.1.1 (Differentiable manifold) A C k (k ≥ 0) n-dimensional differen-

tiable manifold M is a topological space endowed with a countable family of open
C.1 Differentiable Manifolds 501
sets (Oi ), such that each Oi is homeomorphic to some open set in Rn via a home-
omorphism ψi . Every point x ∈ M belongs to at least one Oi , and the changes of
coordinates
ψj ◦ ψi−1 : ψi (Oi ∩ Oj ) → ψj (Oi ∩ Oj )
are C k -diffeomorphisms in Rn .
A function from one manifold to another (in particular curves from R into a
manifold) is C k if it is C k in the given charts. In the following and throughout the
monograph, mostly C ∞ connected manifolds are considered (C k manifolds for ev-
ery k) without further mention. “Smooth” is usually understood as C ∞ below (and
throughout the book).
The first objects of importance in this context are tangent vectors. A C 1 curve on a
differentiable manifold M is a C 1 map from R, or an open interval I ⊂ R containing
0, into M. It is given in a local chart by coordinates γ (t) = (x i (t))1≤i≤n , t ∈ I .
The derivative at t = 0 is the tangent vector (Z i )1≤i≤n of the curve at γ (0) = x0 .
Through a change of coordinates given by y = y(x) in a neighborhood of x0 , this
tangent vector will have coordinates (V j )1≤j ≤n at the point y0 = y(x0 ). Then (recall
the Einstein convention),
j
V j = Ji Z i , 1 ≤ j ≤ n,
j j
where (Ji ) = ( ∂y )
∂x i x=x0
is the Jacobian matrix at the point x0 .
The set of all tangent vectors to all curves passing through a given point x0 is
an n-dimensional vector space which is called the tangent space of M at x0 and
is denoted Tx0 (M). Its dual vector space (the linear forms on it) is spanned by the
differentials of functions f : M → R. Actually, given such a smooth function f and
a curve γ (t) onto M with derivative Z = (Z i )1≤i≤n at t = 0, the derivative at t = 0
of the function f (γ (t)) is nothing else than
∂f
Zi = Z i ∂i f.
∂x i
The form (more precisely the 1-form) w = df with coordinates wi = ∂i f ,
1 ≤ i ≤ n, in the local coordinate system x follows a different rule in the change
of variables from x to y = y(x). Namely, in the coordinate system y, df has com-
ponents ηj , 1 ≤ j ≤ n, with
j
ηj J i = wi , 1 ≤ i ≤ n,
which is precisely the inverse of the change of coordinates rule for vectors.
A vector field Z is for any x ∈ M a tangent vector Z(x) ∈ Tx (M) depending
in general smoothly on the point x, meaning that in a local system of coordi-
nates, its components are smooth functions. To any vector field Z is associated a
first order differential operator f → Zf given in a local system of coordinates by
Zf = Z i ∂i f . Thanks to the change of variables rules for vectors and forms, this
expression is independent of the system of coordinates in which it is considered.
In this monograph, more complicated objects than vectors and 1-forms will be
considered, namely tensors with many indices, some up (like vectors), some down
(like 1-forms). Tensors are required in particular to introduce Riemannian metrics,
which are simply Euclidean structures on the tangent space Tx (M) moving smoothly
with the point x ∈ M. The next section describes these objects first in the classical
Euclidean framework.
C.2 Some Elementary Euclidean Geometry
This section is devoted to some basic notions concerning tensor calculus in Eu-
clidean spaces. On Rn , or any n-dimensional vector space E, a Euclidean structure
is a strictly positive (positive-definite in the standard terminology) symmetric bilin-
ear form G(Z, Y ) which to any pair (Z, Y ) of vectors in E associates a real number
G(Z, Y ) such that G(Z, Z) ≥ 0 for any Z ∈ E and G(Z, Z) = 0 only for Z = 0.
Given any basis e = (ei )1≤i≤n in E, such a strictly positive symmetric bilinear
form G is represented by the positive-definite symmetric matrix gij = G(ei , ej ),
1 ≤ i, j ≤ n, so that if a vector Z is Z = Z i ei (that is (Z i )1≤i≤n are the coordinates
of Z in the basis e), then
G(Z, Z) = Z i Z j gij .
Such a non-degenerate bilinear form G provides an isomorphism G between E
and its dual space E ∗ through G (Z)(Y ) = G(Z, Y ), Z, Y ∈ E. The linear map
G is represented by the matrix G in the dual basis e∗ . The Euclidean structure G
on E may then be transferred to a Euclidean structure G∗ on E ∗ by
−1 −1
G∗ (V , W ) = G G (V ), G (W ) , (V , W ) ∈ E ∗ × E ∗ .
It turns out that if G has matrix (gij )1≤i,j ≤n in a basis e, the matrix of G∗ in the dual
basis e∗ is the inverse matrix of G, denoted (g ij )1≤i,j ≤n = ((gij )1≤i,j ≤n )−1 . When
V and W have respective coordinates (Vi )1≤i≤n and (Wi )1≤i≤n in the dual basis
e∗ , then G∗ (V , W ) = g ij Vi Wj and the correspondence G (Z) = V is such that
Vi = gij Z j , Z i = g ij Vj , 1 ≤ i ≤ n. This operation of identification between E and
E ∗ is called the lifting or lowering of indices. Observe also that since g ij gj k = δki ,
the lifting and lowering operators are inverse to each other.
Multi-linear forms acting both on E and E ∗ may be investigated similarly.
For example, a trilinear map T (Z, Y, V ) on E × E × E ∗ would have coordinates
Tij k = T (ei , ej , e∗k ) such that
T (Z, Y, V ) = Tij k Z i Y j Vk .
Quite often, it is not important to precisely locate the position of the upper and lower
indices. When it is implicit or not necessary to specify, this aspect will simply be
omitted, as is usually the case. Such an object is called a tensor (here a 3-tensor). It
may have indices up or down, depending on whether it acts for some component on
C.2 Some Elementary Euclidean Geometry 503
E or E ∗ . The spaces of multi-linear tensors may be regarded as tensor products of

copies of E and E ∗ (in the previous example E ∗ ⊗ E ∗ ⊗ E, with basis e∗ ⊗ e∗ ⊗ e,
and the coefficients Tij k are just the coordinates on T in this basis). Via the identi-
fication between E and E ∗ one may always raise or lower any component of such a
tensor. In the previous example, the components of S(Z, Y, X) = T (Z, Y, G (X))
would be Sij k = Tij gk . With this operation in mind, we may always assume that a
given tensor has all its indices up or down.
Tensors may be multiplied (in fact through tensor products). For example, a
3-tensor T (Z, Y, X) may be multiplied by a 2-tensor S(Z , Y ) to give rise to a
5-tensor R = T ⊗ S via R(Z, Y, X, Z , Y ) = T (Z, Y, X)S(Z , Y ). The coordinates
of the new tensor are just the term by term products of the coordinates of the two
initial tensors. Tensors may also be contracted. The simplest example of contraction
is the action of a linear form V on a vector Z, which may be written in coordinates
as V (Z) = Vi Z i . A tensor T = Tijk may be contracted as Tkii , or Tiki (they are not
j
the same in general). For tensors of the form Mi which correspond to linear trans-
formations from E to E, or E ∗ to E ∗ , the contraction operation is just the trace.
The quadratic forms G and G∗ defined on E and E ∗ may be used to transfer the
Euclidean metric onto some new Euclidean metric on the vector space of the tensors
of any form, which means for example that if e is an orthonormal basis of E, then
e ⊗ · · · ⊗ e is a basis of E ⊗ · · · ⊗ E, which is again orthonormal (of course, there is a
more intrinsic definition for this Euclidean metric). To compute the norm of a tensor
T in any basis, one has to apply the metric G as many times as there are indices.
For example, if T = Tij k is defined in E ⊗ E ⊗ E ∗ , then its norm is given by
|T |2 = Tij k Tp q g i g jp gkq .
(The notation | · | is used throughout the monograph to denote the Euclidean norm
of vectors and tensors.) It is easier to use the operation of lifting or lowering indices
to obtain simpler expressions such as |T |2 = Tij k T ij k , or Tij k T ij k (the multiplica-
tion by the matrix g ij or the matrix gij being hidden in the operation of lifting or
lowering indices).
Recall finally that the Lebesgue measure is in general defined on a vector space
up to some scaling constant (as the unique Radon measure which is invariant
under translation). It is well-defined on any Euclidean space, so that for exam-
ple, that the measure of the cube C constructed on an orthonormal basis, that is
C = {t i ei ; 0 ≤ ti ≤ 1}, has measure 1. This property is independent of the cho-
sen orthonormal basis (due to the fact that the Lebesgue measure in Rn is in-
variant under orthogonal transformations). In a non-orthonormal basis, with the
metric given by the matrix G = (gij )1≤i,j ≤n , the Lebesgue measure is given by
det(G)1/2 dx 1 · · · dx n .
In Euclidean spaces, it is thus easier to develop the associated calculus in or-
thonormal bases. But in Riemannian geometry, it is impossible in general to choose
a system of variables for which the tangent spaces, spanned by the vectors ∂i , are
orthonormal everywhere in the neighborhood of any point. This is why the calculus
has to be developed in non-orthonormal bases.
C.3 Basic Notions in Riemannian Geometry
In Riemannian geometry, the preceding notions from Euclidean geometry evolve as

the metric G on the tangent space Tx (M) now depends (smoothly) on the point x
in the manifold M. More precisely, a Riemannian metric on a n-dimensional dif-
ferentiable manifold M is a Euclidean metric G(x) given on any tangent space
Tx (M), x ∈ M. In a local system of coordinates, it is given by a positive-definite
symmetric matrix G(x) = (gij (x))1≤i,j ≤n , such that if Z ∈ Tx (M) has coordinates
(Z i (x))1≤i≤n , then
|Z|2 = G(Z, Z) = gij Z i Z j
(for simplicity, dependence upon x is omitted here). The entries of the matrix G(x)
are supposed to be as smooth in the parameter x ∈ M as the manifold is, C ∞ for sim-
plicity in this appendix (and in most parts of the book). By ordinary computation,
this is then also the case for the inverse matrix G∗ (x) = (g ij (x))1≤i,j ≤n denoted
below by g = G∗ and called the co-metric. Outside the Euclidean case, it is always
possible to choose a coordinate system such that, at a given point x, G(x) is the iden-
tity matrix, but it will in general not be possible to make it the identity everywhere in
a neighborhood of any given point. Nevertheless, when computing quantities such
as traces or norms of tensors at a given point x ∈ M, it is often convenient to assume
that g(x) = Id.
Definition C.3.1 (Riemannian manifold) A pair (M, G) comprising a smooth

(connected) n-dimensional differentiable manifold M and a (smooth) Riemannian
metric G = G(x), x ∈ M is called a Riemannian manifold (with dimension n).
In the framework of this monograph, working with functions rather than points,
the relevant Riemannian object is the co-metric g (thus with upper indices) rather
than the metric G. It is indeed the co-metric which naturally enters into the descrip-
tion of the Markov generators and their carré du champ operators (cf. Sects. 1.10
and 1.11, p. 38 and p. 42.) We thus prefer to write (M, g) for the basic Riemannian
structure.
Due to the dependence on x ∈ M, it is important to be able to follow various
quantities in the tangent spaces, such as metrics and tensors, in different coordi-
nate systems. In a given coordinate system, a tensor T = T (x) is represented by
···i
its coordinates Tji11···j m
(x), with as many functions of x ∈ M as there are possi-
ble values for the multi-indices i1 · · · i , j1 · · · jm , each of them varying between
1 and the dimension n (this range for the various indices is not always indicated
in similar expressions below). For example, a vector field Z(x) = (Z i (x))1≤i≤n
has one index, and represents a first order differential operator. A matrix field like
G(x) = (gij (x))1≤i,j ≤n has two indices. Recall that tensors with more indices may
be considered similarly and appear naturally under the operation of tensor products
such as (Z i gk ) (corresponding to the tensor product Z ⊗ G of the vector field Z
and the metric G). Here and below, dependence on x ∈ M is often omitted.
C.3 Basic Notions in Riemannian Geometry 505
Under a change of coordinates y(x), the coordinates of a tensor transform using

∂y ∂y i
the Jacobian matrix J = ∂x = (Jji ) = ( ∂x j ), 1 ≤ i, j ≤ n, with the rule of multipli-
cation by J for any index down and by its inverse for any index up. For example, de-
noting by J¯ the inverse matrix of J (which is ∂x ∂y ), a vector field Z with coordinates
i
(Z )1≤i≤n in the coordinates x, would have coordinates Z j = Z i J j , 1 ≤ j ≤ n, in
i
the new system y such that
∂f j
i ∂y ∂f j ∂f .
Zi = Z =Z
∂x i ∂x i ∂y j ∂y j
Similarly, a 2-tensor T = (Tij )1≤i,j ≤n would be changed into Tk J¯ik J¯j , 1 ≤ i, j ≤ n.
A simple reminder is the change of variables formula
∂f ∂f ∂x j
=
∂y i ∂x j ∂y i
which amounts to the change of variables formula for tensors with one index down.
Upper indices are called covariant and lower indices are called contravariant.
Since the Riemannian metric G(x) = (gij (x))1≤i,j ≤n is non-degenerate, the in-
verse matrix g(x) = (g ij (x))1≤i,j ≤n corresponds to the metric on the dual of the
space of 1-forms. Recall that, as vector fields correspond to first order differen-
tial operators, a vector field Z in a given coordinate system may be written as
Zf = Z i ∂i f (in short Z = Z i ∂i ) and the operators ∂i , 1 ≤ i ≤ n, form a linear
basis of this space of vector fields (or tangent vectors) in this system of coordinates.
In the dual space, a basis is given by the forms dx i , 1 ≤ i ≤ n, and
df = ∂i f dx i , Z, df = Z i ∂i f.
Therefore, the duality action between a vector field Z and a 1-form df is what was
denoted by Zf . Note that not every 1-form is of the form df . For example, when g
and f are smooth functions, gdf is a 1-form, but may not in general be written as
dh.
Recall that the indices for vectors are up and for forms are down. As in the
Euclidean case, to compute the length of a 1-form w = (wi )1≤i≤n , one uses the
inverse matrix g = (g ij (x))1≤i,j ≤n
|w|2 = g ij wi wj .
In particular, this formula defines the length |df |2 . As mentioned above, in the con-
text of this monograph, while working with second order differential operators, we
will be dealing precisely with the co-metric g via the formula
(f, f ) = |df |2 = g ij ∂i f ∂j f
for the carré du champ operator . This corresponds to a Riemannian metric only
when the underlying operator is elliptic. Some of the formulas defined in Rieman-
nian geometry however still make sense for degenerate operators, although the com-
putations are much easier in the language of Riemannian geometry (cf. Sect. C.5
below).
As already discussed, the metric and its inverse may be used to lift or lower
indices. For example, one may lift in this way the form df to yield the gradient
vector denoted by ∇f with coordinates
∇ i f = g ij ∂j f, 1 ≤ i ≤ n.
In particular ∇ i f ∂i g = (f, g). Since, as in Euclidean geometry, lifting or lowering

indices do not change the norms of tensors, it holds that |∇f |2 = |df |2 . It is however
of importance to clearly distinguish df and ∇f .
The partial derivative of a function is a tensor. But in general, the partial deriva-
tive of a tensor is not a tensor as can be seen for example in the change of vari-
able formula for ∂ij2 f which involves the second derivative of y(x). To compensate
this effect, it is necessary to introduce the notion of a connection. On vector fields
Z = (Z j )1≤j ≤n for example, ∂i Z j has to be replaced by the more complicated ob-
ject ∇i Z j defined by
j
∇i Z j = ∂i Z j + γik Z k , 1 ≤ i, j ≤ n,
where the coefficients γijk will satisfy a specific change of variables formula, dif-
ferent from the one for tensors, in order for the resulting operation ∇i Z i to behave
as a tensor. Those coefficients γijk = γi j k , 1 ≤ i, j, k ≤ n, are called the Christoffel
symbols of the connection. With tensors w = (wj )1≤j ≤n with one index down, signs
have to be changed. For example,
∇i wj = ∂i wj − γijk wk , 1 ≤ i, j ≤ n.
This yields in particular the following rule, for Z a vector field and w a 1-form,
∇(Z · w) = (∇Z) · w + Z · ∇w,
meaning in coordinates

∂i Z j wj = ∇i Z j · wj + Z j (∇i wj ), 1 ≤ i ≤ n.
For tensors T with many indices, the rule may be applied on any index. For example,
j
∇i T j k = ∂i T j k + γi T k + γik T j .
The rule (changing sign for indices down) is applied as many times as there are
indices in the tensor. In this way,
∇(T ⊗ U ) = (∇T ) ⊗ U + T ⊗ (∇U ),
being especially careful with the position of the indices in the resulting tensor.
The first fundamental theorem in Riemannian geometry is the existence of a par-
ticular connection.
C.3 Basic Notions in Riemannian Geometry 507
Theorem C.3.2 (Levi-Civita connection) On a Riemannian manifold (M, g), there

exists a unique connection (the Levi-Civita connection) such that the derivative of
the metric is zero, that is ∇(g ij ) = 0 (or equivalently ∇(gij ) = 0), and for any
smooth function f : M → R, ∇i ∇j f = ∇j ∇i f = ∇∇ij f (the connection is said to
be torsion-free). In this case, the resulting operation (∇∇ ij f = ∇ i ∇ j f )1≤i,j ≤n is
also symmetric and is called (in this book) the Hessian ∇∇f of f (sometimes also
denoted by Hess(f )). The symbols of this connection are given by
1 kp
γijk = g (∂i gjp + ∂j gip − ∂p gij ), 1 ≤ i, j, k ≤ n.
2
The derivative of any quantity according to this connection rule is called the
covariant derivative.
Of course, when the metric is constant (that is, on Euclidean spaces), the con-
nection is just the usual derivative, as long as the system of coordinates is such that
(g ij ) is constant. But even in this case, in some other coordinate system (for exam-
ple polar coordinates in Rn ), the usual derivatives differ from the covariant ones.
The connection rules may be used to compute what corresponds to the usual deriva-
tive under a change of coordinates without coming back to the standard system of
coordinates. The only necessary information is the matrix g(x) = (g ij (x)).
There is often no need to explicitly compute the Christoffel symbols. This con-
nection is nevertheless a powerful tool when working with and differentiating ten-
sors. Thanks to the rules of the Riemannian connection, the operation ∇ commutes
with lifting and lowering of indices, and also with the contraction of indices.
The Levi-Civita connection is chosen to be torsion-free, meaning that second
derivatives of functions (0-tensors) are symmetric 2-tensors. This is no longer the
case with derivatives of higher order tensors. For example, for a vector field Z,
∇i ∇j Z k may differ from ∇j ∇i Z k . Similarly, for a smooth function f , ∇i ∇j ∇k f is
in general not symmetric with respect to (i, j, k) as is the case in Euclidean spaces.
Actually, the symmetry defect (∇i ∇j − ∇j ∇i )Z k of a vector field Z gives rise to a
4-tensor R = (Rij k )1≤i,j,k,≤n , independent of the vector field Z, such that
(∇i ∇j − ∇j ∇i )Z k = Rij k Z .
This tensor R is the Riemann curvature tensor. It may be expressed (in a com-
plicated way) in terms of the metric (gij ) and its first and second derivatives in a
coordinate system.
A fundamental theorem of geometry asserts that the Riemann tensor R entirely
characterizes the metric. In particular, when it vanishes, the metric is Euclidean,
which means that there exists (locally) a change of coordinates in which the metric is
constant. It is also true that when it is constant (strictly) positive or constant (strictly)
negative, the metric is spherical or hyperbolic (provided the notion of constant tensor
is properly described). The Riemann curvature tensor R has a lot of symmetries. In
particular, the tensor (Rij k ) obtained by lowering the index k is anti-symmetric in
(i, j ), anti-symmetric in (k, ) and symmetric under the exchange of the pairs (i, j )
and (k, ). There are many other identities involving the covariant derivatives of R,
for example the so-called Bianchi identities, not described here.
Most of the tools and results in this book do not use the full Riemann tensor R,
but a simpler one, known as the Ricci 2-tensor, which is the trace of R defined as
Ricij = Rki k j , 1 ≤ i, j ≤ n. (C.3.1)
The Ricci tensor is a symmetric tensor. While defined here with lower indices, de-
pending on the context it may be convenient to use Ricij for which we retain the
same notation Ric without risk of confusion. In low dimension n ≤ 3, due to the
many symmetries of the Riemann tensor R, it also characterizes the metric. To say
that the Ricci tensor is constant (equal to some constant ρ) amounts to saying that
Ricij = ρ gij . Therefore, in dimension less than or equal to 3, a constant Ricci ten-
sor indicates that the metric is a metric of a sphere when ρ > 0 and a metric of
hyperbolic space when ρ < 0 (ρ = 0 corresponding to the flat Euclidean space).
The Ricci tensor of (M, g) is usually denoted by Ricg to emphasize the underlying
(co-) metric, or more simply Ric (to avoid confusion with indices).
To say that the Ricci tensor is bounded below by ρ ∈ R means that the tensor
Ric −ρ g is positive. In other words, the lowest eigenvalue of the 2-tensor Ric in a
system of coordinates where at a point x, g(x) is the identity, is bounded from below
by ρ. This is usually denoted as Ric ≥ ρ g. In this definition, ρ may be a function,
although in most parts of this book only constant ρ is considered.
A further trace operation on this Ricci tensor produces the scalar curvature, that
is
scg = g ij Ricij = Ricii . (C.3.2)
The function scg (x), x ∈ M, plays an important role in conformal invariance of
Sobolev inequalities (Chap. 6). In dimension 2, it is also enough to characterize the
metric, as for the Riemann tensor in general and the Ricci tensor in dimension 3.
(In dimension one, there is only one metric after a change of variables, and it is the
usual, Euclidean, flat metric.)
This Ricci tensor appears crucially in the analysis of Markov semigroups through
the Bochner-Lichnerowicz formula. This formula connects the Laplace operator
with the Ricci curvature. In the same way as in Rn where it is given on a smooth
function f as the trace of the symmetric matrix (Hessian) (∂ij2 f ) , the
1≤i,j ≤n
Laplace-Beltrami operator (or Laplacian) g on a Riemannian manifold (M, g)
is defined on smooth functions f : M → R as
g f = g ij ∇i ∇j f = ∇ i ∇i f = ∇i ∇ i f. (C.3.3)
This definition is coordinate-free and one would obtain the same operator in any
coordinate system, using the change of variables rules for tensors.
The Laplace-Beltrami operator g is presented in Sect. 1.11, p. 42, in a different
way as the differential operator whose carré du champ operator is (f, f ) = |df |2
and which is invariant with respect to the Riemannian measure. In a local system of
C.4 Riemannian Distance 509
coordinates (x i )1≤i≤n , the Riemannian measure has density det(g)−1/2 with respect
to the Lebesgue measure dx 1 · · · dx n where det(g) is the determinant of the ma-
trix g = (g ij ). This is coherent with the representation of the Lebesgue measure in
Euclidean space when the coordinate system is not orthogonal. Then, the Rieman-
nian measure, often denoted by μg in this monograph, corresponds to the Lebesgue
measure on the tangent space Tx (M) equipped with the Euclidean metric associated
with g. These two definitions of the Laplacian g are equivalent.
The following theorem introduces the Bochner-Lichnerowicz formula, which in-
spires many of the developments in this book. Indeed, it is the basis for the geomet-
ric and functional analysis of curvature developed in the context of the and 2
operators in Sect. C.6 below.
Theorem C.3.3 (Bochner-Lichnerowicz formula) For any smooth function f on

the Riemannian manifold (M, g),
1
g |∇f |2 = ∇f · ∇(g f ) + |∇∇f |2 + Ricg (∇f, ∇f ).
2
C.4 Riemannian Distance
Riemmanian manifolds admit a natural structure as metric spaces. Let (M, g) be a

Riemannian manifold. Given the metric G = (gij )1≤i,j ≤n , one may compute at any
point the length of a tangent vector. Given a smooth curve t → c(t) with values in
M, if ċ(t) = dc(t)
dt denotes the tangent vector at time t, its length computed with
respect to the metric at the point c(t) is
1/2
ċ(t) = gij ċi ċj .
c(t)
By comparison with the Euclidean case, the length of the curve c between 0 and t
may be defined as
t

ċ(s)ds.
0
Curves of minimal length are called geodesics. In a given system of coordinates, the
point c(t) = (ci (t))1≤i≤n along a geodesic satisfies the differential equation (Euler-
Lagrange equation)
d 2 ci (t) j
i dc dc
k
+ γ jk = 0, 1 ≤ i ≤ n,
dt 2 dt dt
where γjik are the Christoffel symbols. Geodesics are also curves of minimal energy,

t
minimizing 0 |ċ(s)|2 ds.
It is not always the case that there is a geodesic going from one point to an-
other. For example, in the Euclidean plane, if one removes the point 0, there is no
geodesic connecting x to −x since such a geodesic would be a straight line through

0 which has been removed. Also, if there exist geodesics between two points, it is
not always the case that they are unique. For example on a sphere, geodesics are
meridian lines (big circles centered at the center of the sphere, or intersections of
the sphere with hyperplanes containing the center of the sphere), and there are in-
finitely many geodesics joining two antipodal points (x and −x). On the other hand,
when two points x and y are sufficiently close to each other, there always exists a
unique geodesic connecting them.
The (geodesic) distance between two points x and y in a Riemannian manifold
(M, g) is defined as the infimum of the lengths of all curves joining x to y. This pro-
vides the manifold M with a distance (a metric structure) d(x, y), (x, y) ∈ M × M,
called the Riemannian distance. The manifold is complete when it is complete for
this distance. When the manifold is complete, there always exists a geodesic (but
not necessarily a unique one, as in the example of the sphere) joining two points.
When two different geodesics starting from x meet at the same point y, the latter
point y is said to be in the cut-locus of x (and this is a symmetric relation between x
and y). The cut-locus of any point has measure 0 for the Riemannian measure, and
may therefore often be forgotten (however unfortunately not always). Given a point
x ∈ M, the map dx : M → R defined by dx (y) = d(x, y), y ∈ M, is not smooth in
general. It is not smooth even in the Euclidean space Rn at y = x. But, in Rn , dx2
is smooth, in contrast to what happens in general in manifolds. On a manifold, this
latter map is smooth, except at the cut-locus of x, where it is not even C 2 . The map
dx is nevertheless continuous and Lipschitz, and satisfies |∇dx | ≤ 1, with equality
everywhere except possibly at the cut-locus.
In the context of this book, as long as diffusion operators are concerned, there is
a dual description of the Riemannian distance as

d(x, y) = sup f (x) − f (y) , (x, y) ∈ M × M,
where the supremum is taken over all smooth functions f : M → R such that
|∇f | ≤ 1. This is reminiscent of the Rademacher Theorem describing Lipschitz
functions (in Euclidean space) as functions with (almost everywhere) bounded
gradients. Functions which approximate this infimum are in fact smooth approx-
imations of the distance d(x, y). This dual description is much more convenient
when considering many aspects of functional inequalities. Furthermore, the lat-
ter definition makes sense for any diffusion operator, even for non-elliptic ones
(hypo-ellipticity is however necessary to effectively define a finite distance between
points), and this formulation is the one used throughout this monograph to introduce
a natural pseudo-distance associated with any generator of a diffusion semigroup
(see (3.3.9), p. 166).
Completeness of a Riemannian manifold (M, g) with respect to the Riemannian
distance d may be described similarly according to the following proposition. This
characterization will prove most useful in the analysis of self-adjointness or gradient
bounds as developed in Sect. 3.2.2, p. 141.
C.5 The Riemannian and 2 Operators 511
Proposition C.4.1 A smooth Riemannian manifold (M, g) is complete if and only

if there exists an increasing sequence (ζk )k∈N of smooth compactly supported func-
tions on M such that |∇ζk | ≤ k1 for every k (≥ 1) and limk→∞ ζk = 1 (pointwise).
C.5 The Riemannian and 2 Operators

This section and the next one go beyond the purpose of this appendix and deal with
the central objects of this monograph, namely diffusion operators and their associ-
ated carré du champ and iterated carré du champ operators and 2 . In particular,
we refer to the core of the book for the proper definitions and discussions of dif-
fusion operators and their and 2 operators (cf. Chaps. 1 and 3). This section
however aims to describe these objects in the present Riemannian setting, connect-
ing to and justifying their geometric origin and interpretation.
Most of this book deals with diffusion operators, more precisely second order
differential operators with no constant terms. These are operators on open sets O in
Rn , or on a manifold, which may be represented in the form (in Einstein’s notation)
Lf = g ij ∂ij2 f + bi ∂i f
where (g ij (x))1≤i,j ≤n is a symmetric positive matrix and (bi (x))1≤i≤n is a vec-

tor, both depending smoothly on x ∈ O. If the matrix (g ij (x)) is everywhere non-
degenerate so that the operator L is elliptic in the sense of (1.12.2), p. 50, the matrix
g = g(x) = (g ij (x))1≤i,j ≤n gives rise to a Riemannian structure (co-metric) g on
this open set. Then L may be written as
L = g + Z
where g is the Laplace-Beltrami operator associated with g. The difference be-

tween L and g is thus a first order differential operator (vector field) Z with no
0-order term. Note that this canonical decomposition into first and second order
parts is more
difficult to set up when dealing with the Hörmander form (see e.g.
Sect. B.4) dj =1 Zj2 + Z0 .
For operators as above which are moreover symmetric with respect to some mea-
sure with smooth strictly positive density w with respect to the Riemannian measure
μg , this decomposition reads
L = g + ∇(log w).
Actually, when the operator L is of the form g + Z, and provided the coefficients
of Z are smooth, whenever it is symmetric with respect to a measure μ, this measure
has a smooth density w with respect to μg . Setting w = e−W , the vector field Z is
the gradient field −∇W and in particular ∇Z is a symmetric tensor. This necessary
condition on Z in order for L be symmetric in L2 (μ) is also sufficient on sim-
ply connected domains. A Riemannian manifold (M, g) equipped with a measure
dμ = e−W dμg is referred to as a weighted Riemannian manifold.
The preceding decomposition of elliptic operators thus amounts to the investiga-

tion of operators of the form L = g + Z on a manifold (M, g). For such an operator
L, its carré du champ operator is then defined on smooth functions f : M → R
as
(f, f ) = |df |2 = |∇f |2 . (C.5.1)
The iterated carré du champ operator 2 is introduced in (1.16.1), p. 71, as
1
2 (f, f ) = L (f, f ) − 2 (f, Lf ) (C.5.2)
2
on smooth functions f : M → R. To ease the notation, we often use (f, f ) = (f )
and 2 (f, f ) = 2 (f ). Both and 2 are extended by polarization to define bilin-
ear forms (f, g) and 2 (f, g) on pairs (f, g) of smooth functions. In the local
system of coordinates g,
(f, g) = g ij ∂i f ∂j g.
The Bochner-Lichnerowicz formula (Theorem C.3.3) shows (and actually motivates

the definition of 2 ) that
2 (f ) = |∇∇f |2 + (Ricg −∇S Z)(∇f, ∇f ) (C.5.3)
where Ricg denotes the Ricci tensor and
1
(∇S Z)ij = (∇i Zj + ∇j Zi ), 1 ≤ i, j ≤ n,
2
is the symmetric part of the (non-symmetric) tensor ∇Z. The symmetric tensor
Ricg −∇S Z on the right-hand side of (C.5.3) is denoted by
Ric(L) = Ricg −∇S Z (C.5.4)
and is often called the generalized Ricci tensor (of L = g + Z ).
Remark C.5.1 (Hessian) When working with diffusion operators L and their carré
du champ operators , there is no need to compute the Christoffel symbols for
formulas on functions. For instance, to compute the Hessian of a smooth function f
acting on the gradient of smooth functions g and h on M,
1
∇∇ ij f ∂i g ∂j h = g, (f, h) + h, (f, g) − f, (g, h) . (C.5.5)
2
This formula for Hessians is widely used in the construction of 2 operators on the
domain of L (in particular in Sect. 3.3, p. 151). The right-hand side makes sense
even in non-elliptic settings, when the (co-) metric is degenerate.
C.6 Curvature-Dimension Conditions 513
C.6 Curvature-Dimension Conditions

Operators defined on an open set in Rn or on a manifold often satisfy uniform ellip-
ticity conditions in the sense that, in a given system of coordinates, the metric (g ij )
is bounded from below by c Id for some c > 0. This condition is not invariant under
a coordinate change (and is thus not intrinsic). The intrinsic condition which is the
closest to uniform ellipticity could in many cases be replaced by a uniform (possibly
negative) lower bound on the Ricci curvature. The curvature conditions presented
below cover these instances. This section is more precisely devoted to the Rieman-
nian description of the curvature-dimension condition of a diffusion operator, which
is extensively investigated throughout this work.
Start with the Laplace-Beltrami operator L = g on an n-dimensional Rieman-
nian manifold (M, g) with Ricci tensor Ric = Ricg . In this case, thanks to the
Bochner-Lichnerowicz formula (Theorem C.3.3) and (C.5.3), the 2 operator takes
the form, on smooth functions f : M → R,
2 (f ) = |∇∇f |2 + Ric(∇f, ∇f ).
In this form, the second and first order terms are well separated, and it is easily seen
that
2 (f ) ≥ ρ (f ) = ρ |∇f |2 (C.6.1)
for all f ’s if and only if
Ric(∇f, ∇f ) ≥ ρ |∇f |2
for all f ’s. In other words, (C.6.1) holds if and only if the Ricci tensor at every point
is bounded from below by ρ.
Here ρ could be a function, satisfying, for any smooth function f , Ric(∇f, ∇f )
≥ ρ(x)(f, f ), which we denote by Ric ≥ ρ g. The best ρ(x) may be seen as the
lowest eigenvalue of the matrix Ricij . The inequality (C.6.1), equivalent to the lower
bound Ric ≥ ρ g on the Ricci curvature, is called the curvature condition CD(ρ, ∞)
throughout this book, and ρ will usually be a (real) constant.
However, there is a second parameter buried in the 2 operator. Namely, recall
that, on a given function f ,
g f = ∇ i ∇i f = Tr (∇∇f )
(where Tr is the trace). In dimension n, and in an orthonormal basis,

the Hilbert-
Schmidt norm |N |2 of the symmetric matrix N = ∇∇f is the sum nij =1 Nij2 , while

Tr(N ) = ni=1 Nii . Therefore, by the Cauchy-Schwarz inequality,

n
Tr(N ) ≤ n Nii2 ≤ n |N |2 . (C.6.2)
i=1
Note that this inequality may actually be seen in a more intrinsic way. Indeed, in
the Euclidean space of symmetric matrices N endowed with the Hilbert-Schmidt
norm |N|2 , Tr(N) is the scalar product with the matrix Id (or equivalently with the
metric g). This identity has norm | Id |2 = n, and the inequality (Tr(N ))2 ≤ n|N |2 is
nothing else than the Cauchy-Schwarz inequality in this Euclidean space.
On the basis of (C.6.2), the 2 operator of the Laplace-Beltrami operator g is
bounded from below, on every smooth function f : M → R, by
1
2 (f ) ≥ ρ (f ) + (g f )2 ,
n
where n is the dimension of the manifold and ρ the lowest eigenvalue of the
Ricci tensor. This inequality is called a curvature-dimension condition or condition
CD(ρ, n) throughout this book. Note that in this example, the best possible choice
for (ρ, n) in the CD(ρ, n) condition is the lower bound ρ on the Ricci curvature
and the dimension n of the manifold.
The preceding curvature-dimension conditions may be addressed similarly for
general operators of the form L = g + Z where Z is a vector field on an
n-dimensional manifold (M, g). From (C.5.3),
1
2 (f ) ≥ ρ (f ) + (Lf )2
m
for every smooth f : M → R if and only if m ≥ n and, as symmetric tensors,
1
Ric(L) = Ricg −∇S Z ≥ ρ g + Z ⊗ Z. (C.6.3)
m−n
Here, there is no longer a best optimal choice both for m and ρ, except in particular
cases. In particular, the parameter m may be equal to the dimension of the manifold
only for Laplace-Beltrami operators. In this sense, among elliptic operators on a
manifold, the family of Laplace operators plays a particular role.
The condition (C.6.3) above takes a simpler form for symmetric operators
L = g − ∇W · ∇ for the invariant (reversible) measure with density e−W with re-
spect to the Riemannian measure μg . Namely, setting e−W = w1m−n , (C.6.3) then
turns into
1
Ricm (L) = Ricg −(m − n) ∇∇w1 ≥ ρ g.
w1
It is worth mentioning that this tensor Ricm (L) has a simple geometric interpretation
when the dimension m is an integer. Indeed, set p = m − n, and start from the
n-dimensional manifold (M, g) together with a smooth function w1 : M → (0, ∞).
Consider then an auxiliary p-dimensional manifold (M1 , g1 ). Equip the product
M = M × M1 with the Riemannian metric
y) = G(x) + w12 (x)G1 (y)

G(x,
(such metrics are often called wrapped products). In terms of the carré du champ
operator, for a function f (x, y) : M × M1 → R in a local system of coordinates
(x i , y j ),
1

(f, f ) = g ij (x) ∂x i f ∂x j f + g1k (y) ∂y k f ∂y f.
w12 (x)
p
The Riemannian measure on M × M1 is given by w1 (x)dμg (x)dμg1 (y). In the
same way, the new Laplace operator is
1
f = g f + p (log w1 , f ) +
g1 f.
w12
The Ricci operator for this new structure on M × M1 splits into two parts, that is
. = Ric0 + Ric1
Ric
where the action of Ric0 only depends on (∂x i f ) and that of Ric1 only on (∂y j f ).
Using the techniques described in Sect. 6.9, p. 313, it may be shown that
p
Ric0 = Ricg − ∇∇w1 ,
w1
1 1
Ric1 = 4 Ricg1 − g (log w1 ) + p g (log w1 ) 2 g1 .
w1 w1
In particular, when acting on functions f depending only on the first variable x ∈ M,

the action of the Laplacian on the product space M × M1 is a function depending
only on x ∈ M and is the weighted Laplace operator g f + p (log w1 , f ). Fur-
thermore,

. m−n
Ric(∇f, ∇f ) = Ric (∇f, ∇f ) = Ricg −
0
∇∇w1 (∇f, ∇f ).
w1
Observe then that if W = −p log w1 ,
m−n 1
Ricg − ∇∇w1 = Ricg +∇∇W − ∇W ⊗ ∇W = Ricm (L).
w1 m−n
In some places throughout this book, the curvature-dimension conditions

CD(ρ, ∞) or CD(ρ, n) are used in a reinforced version. For example, the
CD(ρ, ∞) condition actually implies that, for all smooth functions f : M → R,

4 (f ) 2 (f ) − ρ (f ) ≥ (f ) . (C.6.4)
This reinforcement is rather easy to understand in this differential geometry con-
text. Namely, the CD(ρ, ∞) condition amounts to Ric(L) ≥ ρ g. The reinforced
inequality expresses that
2
4 |∇f |2 |∇∇f |2 + Ric(L)(∇f, ∇f ) − ρ|∇f |2 ≥ ∇ |∇f |2 .
Now, ∇(|∇f |2 ) is a vector whose scalar product with any vector V is

2∇∇f (∇f, V ), or in coordinates 2∇ i ∇ j f ∇j f . Its norm is bounded above by
2|∇∇f ||∇f | from which the reinforced inequality immediately follows.
Much of the spirit of this book is actually to address a reinforced inequality such
as (C.6.4) not by differential geometry tools, but rather by differential calculus. This
is in particular necessary (for the CD(ρ, ∞) condition) in infinite dimension where
the usual Riemannian geometry is not available. The proof then relies on the change
of variables formula for the 2 operator. Indeed, in the geometric description (C.5.3)
of the 2 operator, the standard differentiation rules
∇∇ψ(f ) = ψ (f )∇∇f + ψ (f )∇f ⊗ ∇f
and

T ψ(f ), φ(g) = ψ (f )φ (g)T (f, g),
for smooth functions ψ, φ : R → R, smooth functions f, g : M → R and 2-tensor T ,
lead to, for smooth functions : Rk → R and vectors f = (f1 , . . . , fk ) of smooth
functions on M,
k
k
2 (f ) = Xi Xj 2 (fi , fj ) + 2 Xi Yj H (fi )(fi , f )
i,j =1 i,j,=1
(C.6.5)

k
+ Yij Ym (fi , f ) (fj , fm )
i,j,,m=1
where
Xi = ∂i (f1 , . . . , fk ), Yij = ∂ij2 (f1 , . . . , fk )
and
1
H (f )(g, h) = g, (f, h) + h, (f, g) − (f, (g, h) . (C.6.6)
2
This last expression is directly related to the representation of the Hessian (C.5.5).
A similar formula applies to 2 − ρ instead of 2 .
Note that (C.6.5) of course takes a simpler form when dealing with a single func-
tion ψ : M → R,

2 ψ(f ) = ψ (f )2 2 (f ) + ψ (f )ψ (f ) f, (f ) + ψ (f )2 (f )2 (C.6.7)
(recall the notation (f ) = (f, f ) and 2 (f ) = 2 (f, f )), a formula which will
be used extensively in Chaps. 5 and 6.
While of a standard differential calculus nature, it should be pointed out that a
more intrinsic derivation of these change of variables formulas may be developed
on the basis of the diffusion property of the generator L as presented in Sect. 1.11,
p. 42. This point of view will be systematically emphasized throughout this work in
the form of the -calculus.
Let us now describe how the reinforced inequality (C.6.4) may be obtained from
the standard curvature condition CD(ρ, ∞) by differential calculus and the change
of variables formula (C.6.5). Indeed, given functions (f1 , . . . , fk ) and at any point
x, the function may be chosen in such a way that the coefficients Xi and Yij take
any particular value, provided the symmetries Yij = Yj i are respected (for example
just letting vary among second degree polynomials). The curvature condition
CD(ρ, ∞) therefore yields a positive quadratic form in the variables (Xi , Yij ). It is
for example a quadratic form in the nine variables Xi , Yi,j , i, j = 1, . . . , 3, i < j , in
case of three functions f1 , f2 , f3 . Illustrating the argument in this sample example,
restrict this quadratic form to the set where all the variables are 0 except X1 and Y23 .
Its determinant

2 (f1 ) − ρ (f1 ) (f2 , f3 )2 + (f2 )(f3 ) − 2H (f1 )(f2 , f3 )2
is positive, which in turn yields

H (f1 )(f2 , f3 )2 ≤ 2 (f1 ) − ρ (f1 ) (f2 )(f3 ). (C.6.8)
Now,
1
H (f1 )(f1 , f2 ) = f2 , (f1 ) ,
2
so that choosing f2 = (f1 ), it follows that

(f2 )2 ≤ 4 2 (f1 ) − ρ (f1 ) (f1 )(f2 ).
The announced reinforced inequality (C.6.4) is thus obtained.

The preceding strategy may look strange. However, it allows for an efficient ge-
ometric treatment of operators which are not Laplacians, somewhat far away from
the Riemannian geometry context. Even in a Riemannian setting, it often provides
a useful form of the curvature conditions. The technique (with variations) moreover
often provides optimal inequalities in numerous functional inequalities of interest.
This -calculus, and its numerous applications, lies at the heart of this monograph.
Finally, and to conclude this section, we observe that in several places in this
work we consider operators L given in Hörmander form

d
L= Zj2 + Z0
j =1
where Z0 , Z1 , . . . , Zd are vector fields, that is first order differential operators with
no zero-order terms. Here, the carré du champ operator is given, on smooth func-
tions, by (f ) = dj =1 (Zj f )2 . In order to have a tractable form for the 2 operator,
it is useful to introduce the symmetric second order derivatives associated to this de-
composition, that is Dij f = 12 (Zi Zj f + Zj Zi f ), 0 ≤ i, j ≤ d, and to denote the
usual commutators [Zi , Zj ] by Zij . Introduce moreover Dk,ij = 12 (Zk Zij + Zij Zk ).
Then
d
2
d
d
1
2 (f ) = Dij f + Zij f +2 Zi f Dj,j i f + Zj Z0j f.
2
i,j =1 i,j =1 j =1
To get a useful formula, as is done in the elliptic case with the use of the language
of differential geometry, it is necessary to separate the first and second order terms.
This is possible for example whenever, for 0 ≤ i, j ≤ d,

d
Zij = αijk Zk
k=1
for some (smooth) functions αijk . In this case,
d

d 2
2 (f ) = Dij f + αji k Zk f + R(f, f ),
i,j =1 k=1
where R(f, f ) (which plays the role of the Ricci tensor) is given by
d
d 2
d 2
1
R(f, f ) = αijk Zk f − αji k Zk f
4
i,j =1 k=1 k=1

d
j j
+ Zi f Zj f α0i + Zk αki .
k=1
C.7 Notes and References
Standard references on differentiable manifolds include [68, 159, 166, 167]. Ba-
sics on Riemmanian geometry and Ricci curvature may be found, for example,
in [62, 123, 126, 194, 260, 346] where in particular the construction of the Laplace-
Beltrami operator and the Bochner-Lichnerowicz formula (Theorem C.3.3) are
emphasized. Comparison methods based on Ricci curvature bounds are surveyed
in [123, 126, 448]. See also [61] for an overview of modern Riemannian geometry.
The characterization of completeness of Proposition C.4.1 is due to R. Stri-
chartz [390] (see also [23] in the context of this work).
The and 2 -calculus of Sects. C.5 and C.6 was introduced in [36] in the context
of logarithmic Sobolev inequalities and in [22, 24, 25] in the study of Riesz trans-
forms. The definition of the curvature-dimension condition CD(ρ, n) (also called
the 2 criterion), which for Laplace-Beltrami operators boils down to the Bochner
C.7 Notes and References 519
formula in the context of harmonic maps between manifolds, may be found there.
For more on Bochner-Weitzenböck formulas, see [92]. These ideas have been fur-
ther developed in the early lecture notes [26] (see also [27, 28, 277]). Some geomet-
ric properties of the tensor Ric(L) are examined in [290].
Afterword
Chicken “Gaston Gérard”

After all, the content of this book is nothing but a sequence of recipes.
Let us thus conclude this monograph with the famous recipe of the chicken à
la façon Gaston Gérard. This is a creation of the Mayor of Dijon’s wife, Madame
Reine Geneviève Gérard. After reading this book, with a white Burgundy wine, such
as Chablis premier cru Montée de Tonnerre or Chassagne Montrachet, this recipe is
perfect for a nice dinner with friends.
Preparation: 60 minutes.
Cooking time: 45 minutes.
– 1 Bresse chicken
– 50 g of butter
– 2 glasses of white wine (Jura wine, preferably Savagnin, is perfect)
– 150 g of Comté cheese, plus 50 g for gratin (cheese topping)
– 400 g of single cream
– 1 teaspoon of paprika
– 1 tablespoon of Dijon mustard
– 1 tablespoon of breadcrumbs
– salt and pepper
(i) Put the pieces of chicken in a cooking pot with the butter, the paprika, salt
and pepper. Turn them over periodically and let the preparation cook for
30 minutes.
(ii) In a separate pan, mix slowly while heating the wine, cheese, mustard and
cream. Do not let the mixture boil.
(iii) Put the chicken in a gratin dish with the above mixture. Cover the preparation
with the spare grated cheese and breadcrumbs.
(iv) Place the dish in a hot oven (240 °C) for 15 minutes.
What about the heat equation in the oven: is this some practical application of
the theory developed here, or just an exercise?

DOI 10.1007/978-3-319-00227-9,
Notation and List of Symbols
·, · Scalar product in Euclidean and Hilbert xv

space
|·| Euclidean norm of vectors and tensors xv, 503
1A Characteristic function of the set A xv
Id Identity operator or Identity matrix 9
O Open set 50
E State space 120, 168
F σ -field 120
μ Positive measure xv
dx, voln Lebesgue measure on Rn xvi
Lp (E, μ), Lp (μ) Space of pth-power integrable 10
functions on (E, F, μ)
L20 (μ) Functions in L2 (μ) with mean zero 70
· p , · Lp (μ) Lp (μ)-norm 10
· p,q Operator norm from Lp (μ) into Lq (μ) 286
· TV Total variation distance 244
(E, μ, ) Diffusion, Standard, Full Markov Triple 132, 136, 167, 168
P = (Pt )t≥0 Markov semigroup 9, 168
pt (x, dy) Markov probability kernel 8, 168
pt (x, y) Markov density kernel 14
Pt∗ Dual Markov semigroup 9
(PtD )t≥0 Semigroup with Dirichlet boundary 93
conditions
(PtN )t≥0 Semigroup with Neumann boundary 93
conditions
E Dirichlet form 30, 168
D(E) Domain of the Dirichlet form E 125, 168
· E Dirichlet norm 125, 168

DOI 10.1007/978-3-319-00227-9,
524 Notation and List of Symbols
L Infinitesimal generator/Markov 18, 168

generator
D(L) Domain of the generator L 18, 168
· D(L) Domain norm on D(L) 133, 168
L∗ Adjoint operator 24, 168
Carré du champ operator 20, 168
2 Iterated carré du champ operator 71, 168
A0 Standard algebra of functions 120, 168
Aconst
0 Adding a constant to functions in A0 154, 168
Aconst+
0 Adding a strictly positive constant to 154, 168
positive functions in A0
A Extended algebra of functions 151, 168
ESA Essentially self-adjointness property 134, 168
f Lip Lipschitz coefficient 166, 168
σess Essential spectrum 480
σd Discrete spectrum 480
(Eλ )λ∈R Spectral decomposition 479
( , , P) Probability space 7
E(·), E(· | ·) Expectation, conditional expectation 8
(Ft )t≥0 Filtration 8
(Bt )t≥0 Standard Brownian motion 487
(Xtx )t≥0 Stochastic process starting at x 8
C k (O) Space of k-times differentiable 42
functions on O
C∞ Space of infinitely differentiable 138
functions
Cc∞ Space of infinitely differentiable 95
compactly supported functions
Sn ⊂ Rn+1 n-dimensional sphere 81
σn Uniform probability measure on Sn 81
Laplacian on Rn 4
(M, g) Riemannian manifold 504
g Co-metric 504
∇ Gradient 5
∇S Symmetric derivative 73
∇∇f Hessian of f 507
H (f )(g, h) Hessian of f evaluated at (∇g, ∇h) 158
g Laplace-Beltrami operator in (M, g) 504
μg Riemannian measure on (M, g) 509
Ricg Ricci tensor of the (co)-metric g 508
Ric(L) Ricci tensor of L 514
scg Scalar curvature of (M, g) 508
d Intrinsic or Riemannian distance 166, 509, 168
dH Harmonic distance 360
Notation and List of Symbols 525
Varμ Variance 181

Entμ Entropy 236
Entμ -entropy 383
H(· | ·) Relative entropy 244
Iμ Fisher information 237
Iμ,F Generalized Fisher information 333
P (C) Poincaré inequality 181
LS(C, D), LS(C) Logarithmic Sobolev inequality 236
S(p ; A, C), Sn (A, C) Sobolev inequality 279
GNn (s, q ; A, C) Gagliardo-Nirenberg inequality 324
EE() Entropy-energy inequality 349
N() Generalized Nash inequality 364
N(, w) Weighted Nash inequality 370
W P (C) Weak Poincaré inequality 375
Bq (C) Beckner inequality 384
LO(C) Latała-Oleszkiewicz inequality 385
Capμ (A) Capacity of the set A 393
Cap∗μ (A) Modified capacity of the set A 394
Iμ Isoperimetric profile of μ 413
I Gaussian isoperimetric profile 416
P(E) Space of probability measures on 434
(E, F)
Pp (E) Space of probability measures on 436
(E, F) with a finite p-th moment
T #μ Image measure of μ by T 437
Tc Optimal transportation cost 434
Wp Wasserstein distance 436
Bibliography
1. R.A. Adams, Sobolev Spaces. Pure and Applied Mathematics, vol. 65 (Academic Press, New
York, 1975)
2. R.A. Adams, J.J.F. Fournier, Sobolev Spaces, 2nd edn. Pure and Applied Mathematics,
vol. 140 (Elsevier/Academic Press, Amsterdam, 2003)
3. M. Agueh, N. Ghoussoub, X. Kang, Geometric inequalities via a general comparison princi-
ple for interacting gases. Geom. Funct. Anal. 14(1), 215–244 (2004)
4. S. Aida, Uniform positivity improving property, Sobolev inequalities, and spectral gaps.
J. Funct. Anal. 158(1), 152–185 (1998)
5. S. Aida, K.D. Elworthy, Differential calculus on path and loop spaces. I. Logarithmic Sobolev
inequalities on path spaces. C. R. Math. Acad. Sci. Paris, Sér. I 321(1), 97–102 (1995)
6. S. Aida, T. Masuda, I. Shigekawa, Logarithmic Sobolev inequalities and exponential integra-
bility. J. Funct. Anal. 126(1), 83–101 (1994)
7. S. Aida, D.W. Stroock, Moment estimates derived from Poincaré and logarithmic Sobolev
inequalities. Math. Res. Lett. 1(1), 75–86 (1994)
8. G. Allaire, A la recherche de l’inégalité perdue. Matapli 98, 52–64 (2012)
9. L. Ambrosio, N. Gigli, G. Savaré, Gradient Flows in Metric Spaces and in the Space of
Probability Measures, 2nd edn. Lectures in Mathematics ETH Zürich (Birkhäuser, Basel,
2008)
10. L. Ambrosio, N. Gigli, G. Savaré, Bakry-Emery curvature-dimension condition and Rieman-
nian Ricci curvature bounds. Preprint, 2012
11. L. Ambrosio, N. Gigli, G. Savaré, Metric measure spaces with Riemannian Ricci curvature
bounded from below. Preprint, 2012
12. L. Ambrosio, N. Gigli, G. Savaré, Calculus and heat flow in metric measure spaces and
applications to spaces with Ricci bounds from below. Invent. Math. (2013). doi:10.1007/
s00222-013-0456-1
13. L. Ambrosio, P. Tilli, Topics on Analysis in Metric Spaces. Oxford Lecture Series in Mathe-
matics and Its Applications, vol. 25 (Oxford University Press, Oxford, 2004)
14. C. Ané, S. Blachère, D. Chafaï, P. Fougères, I. Gentil, F. Malrieu, C. Roberto, G. Schef-
fer, Sur les Inégalités de Sobolev Logarithmiques. Panoramas et Synthèses, vol. 10 (Société
Mathématique de France, Paris, 2000)
15. D. Applebaum, Lévy Processes and Stochastic Calculus, 2nd edn. Cambridge Studies in
Advanced Mathematics, vol. 116 (Cambridge University Press, Cambridge, 2009)
16. A. Arnold, P. Markowich, G. Toscani, A. Unterreiter, On convex Sobolev inequalities and
the rate of convergence to equilibrium for Fokker-Planck type equations. Commun. Partial
Differ. Equ. 26(1–2), 43–100 (2001)
17. T. Aubin, Équations différentielles non linéaires et problème de Yamabe concernant la cour-
bure scalaire. J. Math. Pures Appl. 55(3), 269–296 (1976)

DOI 10.1007/978-3-319-00227-9,
528 Bibliography
18. T. Aubin, Espaces de Sobolev sur les variétés Riemanniennes. Bull. Sci. Math. 100(2),
149–173 (1976)
19. T. Aubin, Problèmes isopérimétriques et espaces de Sobolev. J. Differ. Geom. 11(4), 573–598
(1976)
20. T. Aubin, Nonlinear Analysis on Manifolds. Monge-Ampère Equations. Grundlehren
der mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences],
vol. 252 (Springer, New York, 1982)
21. T. Aubin, Some Nonlinear Problems in Riemannian Geometry. Springer Monographs in
Mathematics (Springer, Berlin, 1998)
22. D. Bakry, Transformations de Riesz pour les semi-groupes symétriques. II. Étude sous la con-
dition 2 ≥ 0, in Séminaire de Probabilités, XIX, 1983/84. Lecture Notes in Math., vol. 1123
(Springer, Berlin, 1985), pp. 145–174
23. D. Bakry, Un critère de non-explosion pour certaines diffusions sur une variété Riemanni-
enne complète. C. R. Math. Acad. Sci. Paris, Sér. I 303(1), 23–26 (1986)
24. D. Bakry, Étude des transformations de Riesz dans les variétés Riemanniennes à courbure
de Ricci minorée, in Séminaire de Probabilités, XXI. Lecture Notes in Math., vol. 1247
25. D. Bakry, The Riesz transforms associated with second order differential operators, in Sem-
inar on Stochastic Processes, Gainesville, FL, 1988. Progr. Probab., vol. 17 (Birkhäuser,
Boston, 1989), pp. 1–43
26. D. Bakry, L’hypercontractivité et son utilisation en théorie des semigroupes, in Lectures on
Probability Theory, Saint-Flour, 1992. Lecture Notes in Math., vol. 1581 (Springer, Berlin,
1994), pp. 1–114
27. D. Bakry, On Sobolev and logarithmic Sobolev inequalities for Markov semigroups, in
New Trends in Stochastic Analysis, Charingworth, 1994 (World Sci., River Edge, 1997),
pp. 43–75
28. D. Bakry, Functional inequalities for Markov semigroups, in Probability Measures on
Groups: Recent Directions and Trends (Tata Inst. Fund. Res, Mumbai, 2006), pp. 91–147
29. D. Bakry, F. Barthe, P. Cattiaux, A. Guillin, A simple proof of the Poincaré inequality for
a large class of probability measures including the log-concave case. Electron. Commun.
Probab. 13, 60–66 (2008)
30. D. Bakry, F. Baudoin, M. Bonnefont, D. Chafaï, On gradient bounds for the heat kernel on
the Heisenberg group. J. Funct. Anal. 255(8), 1905–1938 (2008)
31. D. Bakry, F. Bolley, I. Gentil, Dimension dependent hypercontractivity for Gaussian kernels.
Probab. Theory Relat. Fields 154(3), 845–874 (2012)
32. D. Bakry, F. Bolley, I. Gentil, P. Maheux, Weighed Nash inequalities. Rev. Mat. Iberoam.
28(3), 879–906 (2012)
33. D. Bakry, P. Cattiaux, A. Guillin, Rate of convergence for ergodic continuous Markov pro-
cesses: Lyapunov versus Poincaré. J. Funct. Anal. 254(3), 727–759 (2008)
34. D. Bakry, D. Concordet, M. Ledoux, Optimal heat kernel bounds under logarithmic Sobolev
inequalities. ESAIM Probab. Stat. 1, 391–407 (1995/97) (electronic)
35. D. Bakry, T. Coulhon, M. Ledoux, L. Saloff-Coste, Sobolev inequalities in disguise. Indiana
Univ. Math. J. 44(4), 1033–1074 (1995)
36. D. Bakry, M. Émery, Diffusions hypercontractives, in Séminaire de Probabilités, XIX,
1983/1984. Lecture Notes in Math., vol. 1123 (Springer, Berlin, 1985), pp. 177–206
37. D. Bakry, I. Gentil, M. Ledoux, On Harnack inequalities and optimal transportation. Ann.
Sc. Norm. Sup. Pisa (2012). doi:10.2422/2036-2145.201210_007
38. D. Bakry, M. Ledoux, Lévy-Gromov’s isoperimetric inequality for an infinite-dimensional
diffusion generator. Invent. Math. 123(2), 259–281 (1996)
39. D. Bakry, M. Ledoux, Sobolev inequalities and Myers’s diameter theorem for an abstract
Markov generator. Duke Math. J. 85(1), 253–270 (1996)
40. D. Bakry, M. Ledoux, A logarithmic Sobolev form of the Li-Yau parabolic inequality. Rev.
Mat. Iberoam. 22(2), 683–702 (2006)
Bibliography 529
41. D. Bakry, S.Y. Orevkov, M. Zani, Orthogonal polynomials and diffusion operators. Preprint,
2013
42. D. Bakry, Z.M. Qian, Some new results on eigenvectors via dimension, diameter, and Ricci
curvature. Adv. Math. 155(1), 98–153 (2000)
43. Z.M. Balogh, A. Engulatov, L. Hunziker, O.E. Maasalo, Functional inequalities and
Hamilton–Jacobi equations in geodesic spaces. Potential Anal. 36(2), 317–337 (2012)
44. G. Barles, Solutions de Viscosité des Équations de Hamilton-Jacobi. Mathématiques & Ap-
plications (Berlin) [Mathematics & Applications], vol. 17 (Springer, Paris, 1994)
45. F. Barthe, P. Cattiaux, C. Roberto, Concentration for independent random variables with
heavy tails. Appl. Math. Res. Express 2, 39–60 (2005)
46. F. Barthe, P. Cattiaux, C. Roberto, Interpolated inequalities between exponential and Gaus-
sian, Orlicz hypercontractivity and isoperimetry. Rev. Mat. Iberoam. 22(3), 993–1067 (2006)
47. F. Barthe, P. Cattiaux, C. Roberto, Isoperimetry between exponential and Gaussian. Electron.
J. Probab. 12(44), 1212–1237 (2007) (electronic)
48. F. Barthe, A.V. Kolesnikov, Mass transport and variants of the logarithmic Sobolev inequal-
ity. J. Geom. Anal. 18(4), 921–979 (2008)
49. F. Barthe, C. Roberto, Sobolev inequalities for probability measures on the real line. Stud.
Math. 159(3), 481–497 (2003). Dedicated to Professor Aleksander Pełczyński on the occa-
sion of his 70th birthday
50. R.F. Bass, Diffusions and Elliptic Operators. Probability and Its Applications (New York)
(Springer, New York, 1998)
51. F. Baudoin, N. Garofalo, Curvature-dimension inequalities and Ricci lower bounds for sub-
Riemannian manifolds with transverse symmetries. Preprint, 2012
52. V. Bayle, A differential inequality for the isoperimetric profile. Int. Math. Res. Not. 7,
311–342 (2004)
53. M. Bebendorf, A note on the Poincaré inequality for convex domains. Z. Anal. Anwend.
22(4), 751–756 (2003)
54. W. Beckner, Inequalities in Fourier analysis. Ann. Math. (2) 102(1), 159–182 (1975)
55. W. Beckner, A generalized Poincaré inequality for Gaussian measures. Proc. Am. Math. Soc.
105(2), 397–400 (1989)
56. W. Beckner, Sharp Sobolev inequalities on the sphere and the Moser-Trudinger inequality.
Ann. Math. (2) 138(1), 213–242 (1993)
57. W. Beckner, Geometric asymptotics and the logarithmic Sobolev inequality. Forum Math.
11(1), 105–137 (1999)
58. W. Beckner, M. Pearson, On sharp Sobolev embedding and the logarithmic Sobolev inequal-
ity. Bull. Lond. Math. Soc. 30(1), 80–84 (1998)
59. J. Bennett, A. Carbery, M. Christ, T. Tao, The Brascamp-Lieb inequalities: finiteness, struc-
ture and extremals. Geom. Funct. Anal. 17(5), 1343–1415 (2008)
60. P. Bérard, G. Besson, S. Gallot, Sur une inégalité isopérimétrique qui généralise celle de Paul
Lévy-Gromov. Invent. Math. 80(2), 295–308 (1985)
61. M. Berger, A Panoramic View of Riemannian Geometry (Springer, Berlin, 2003)
62. M. Berger, P. Gauduchon, E. Mazet, Le Spectre d’Une Variété Riemannienne. Lecture Notes
in Mathematics, vol. 194 (Springer, Berlin, 1971)
63. L. Bertini, B. Zegarliński, Coercive inequalities for Gibbs measures. J. Funct. Anal. 162(2),
257–286 (1999)
64. J. Bertoin, Lévy Processes. Cambridge Tracts in Mathematics, vol. 121 (Cambridge Univer-
sity Press, Cambridge, 1996)
65. J. Bertoin, Subordinators: examples and applications, in Lectures on Probability Theory and
Statistics, Saint-Flour, 1997. Lecture Notes in Math., vol. 1717 (Springer, Berlin, 1999),
pp. 1–91
66. M.-F. Bidaut-Véron, L. Véron, Nonlinear elliptic equations on compact Riemannian mani-
folds and asymptotics of Emden equations. Invent. Math. 106(3), 489–539 (1991)
530 Bibliography
67. P. Billingsley, Probability and Measure. Wiley Series in Probability and Statistics (Wiley,
Hoboken, 2012). Anniversary edition [of MR1324786], with a foreword by Steve Lalley and
a brief biography of Billingsley by Steve Koppes
68. R.L. Bishop, R.J. Crittenden, Geometry of Manifolds. Pure and Applied Mathematics,
vol. XV (Academic Press, New York, 1964)
69. J.-M. Bismut, Large Deviations and the Malliavin Calculus. Progress in Mathematics, vol. 45
(Birkhäuser, Boston, 1984)
70. W. Blaschke, Kreis und Kugel (Chelsea, New York, 1949)
71. G. Blower, The Gaussian isoperimetric inequality and transportation. Positivity 7(3),
203–224 (2003)
72. R.M. Blumenthal, R.K. Getoor, Markov Processes and Potential Theory. Pure and Applied
Mathematics, vol. 29 (Academic Press, New York, 1968)
73. S.G. Bobkov, An isoperimetric inequality on the discrete cube, and an elementary proof of
the isoperimetric inequality in Gauss space. Ann. Probab. 25(1), 206–214 (1997)
74. S.G. Bobkov, Isoperimetric and analytic inequalities for log-concave probability measures.
Ann. Probab. 27(4), 1903–1921 (1999)
75. S.G. Bobkov, A localized proof of the isoperimetric Bakry-Ledoux inequality and some ap-
plications. Teor. Veroâtn. Ee Primen. 47(2), 340–346 (2002)
76. S.G. Bobkov, I. Gentil, M. Ledoux, Hypercontractivity of Hamilton-Jacobi equations.
J. Math. Pures Appl. 80(7), 669–696 (2001)
77. S.G. Bobkov, F. Götze, Exponential integrability and transportation cost related to logarith-
mic Sobolev inequalities. J. Funct. Anal. 163(1), 1–28 (1999)
78. S.G. Bobkov, C. Houdré, Some connections between isoperimetric and Sobolev-type in-
equalities. Mem. Am. Math. Soc. 129, 616 (1997), pp. viii+111
79. S.G. Bobkov, M. Ledoux, Poincaré’s inequalities and Talagrand’s concentration phenomenon
for the exponential distribution. Probab. Theory Relat. Fields 107(3), 383–400 (1997)
80. S.G. Bobkov, B. Zegarliński, Entropy bounds and isoperimetry. Mem. Am. Math. Soc. 176,
829 (2005)
81. T. Bodineau, B. Zegarliński, Hypercontractivity via spectral theory. Infin. Dimens. Anal.
Quantum Probab. Relat. Top. 3(1), 15–31 (2000)
82. V.I. Bogachev, Gaussian Measures. Mathematical Surveys and Monographs, vol. 62 (Amer-
ican Mathematical Society, Providence, 1998)
83. V.I. Bogachev, Measure Theory, vol. I, II (Springer, Berlin, 2007)
84. F. Bolley, I. Gentil, Phi-entropy inequalities for diffusion semigroups. J. Math. Pures Appl.
93(5), 449–473 (2010)
85. F. Bolley, I. Gentil, A. Guillin, Dimensional contraction via Markov transportation distance.
Preprint, 2013
86. L. Boltzmann, Lectures on Gas Theory (University of California Press, Berkeley, 1964).
Translated by Stephen G. Brush
87. C. Borell, The Brunn-Minkowski inequality in Gauss space. Invent. Math. 30(2), 207–216
(1975)
88. C. Borell, Positivity improving operators and hypercontractivity. Math. Z. 180(2), 225–234
(1982)
89. A.A. Borovkov, S.A. Utev, An inequality and a characterization of the normal distribution
connected with it. Teor. Veroâtn. Ee Primen. 28(2), 209–218 (1983)
90. S. Boucheron, G. Lugosi, P. Massart, Concentration Inequalities: A Nonasymptotic Theory
of Independence (Oxford University Press, Oxford, 2013)
91. N. Bouleau, F. Hirsch, Dirichlet Forms and Analysis on Wiener Space. de Gruyter Studies in
Mathematics, vol. 14 (Walter de Gruyter, Berlin, 1991)
92. J.-P. Bourguignon, The “magic” of Weitzenböck formulas, in Variational Methods, Paris,
1988. Progr. Nonlinear Differential Equations Appl., vol. 4 (Birkhäuser, Boston, 1990),
pp. 251–271
Bibliography 531
93. H.J. Brascamp, E.H. Lieb, On extensions of the Brunn-Minkowski and Prékopa-Leindler
theorems, including inequalities for log concave functions, and with an application to the
diffusion equation. J. Funct. Anal. 22(4), 366–389 (1976)
94. Y. Brenier, Polar factorization and monotone rearrangement of vector-valued functions.
Commun. Pure Appl. Math. 44(4), 375–417 (1991)
95. H. Brézis, Opérateurs Maximaux Monotones et Semi-Groupes de Contractions dans les Es-
paces de Hilbert. North-Holland Mathematics Studies, No. 5. Notas de Matemática (50)
(North-Holland, Amsterdam, 1973)
96. H. Brézis, Functional Analysis, Sobolev Spaces and Partial Differential Equations (Springer,
New York, 2011)
97. D. Burago, Y.D. Burago, S. Ivanov, A Course in Metric Geometry. Graduate Studies in Math-
ematics, vol. 33 (American Mathematical Society, Providence, 2001)
98. Y.D. Burago, V.A. Zalgaller, Geometric Inequalities. Grundlehren der mathematischen Wis-
senschaften [Fundamental Principles of Mathematical Sciences], vol. 285 (Springer, Berlin,
1988). Translated from the Russian by A.B. Sosinskiı̆, Springer Series in Soviet Mathematics
99. P. Buser, Notes: a geometric approach to invariant subspaces of orthogonal matrices. Am.
Math. Mon. 89(10), 751 (1982)
100. P.L. Butzer, H. Berens, Semi-Groups of Operators and Approximation. Die Grundlehren der
mathematischen Wissenschaften, vol. 145 (Springer, New York, 1967)
101. L.A. Caffarelli, Boundary regularity of maps with convex potentials. II. Ann. Math. (2)
144(3), 453–496 (1996)
102. L.A. Caffarelli, A priori estimates and the geometry of the Monge Ampère equation, in Non-
linear Partial Differential Equations in Differential Geometry, Park City, UT, 1992. IAS/Park
City Math. Ser., vol. 2. (Amer. Math. Soc., Providence, 1996), pp. 5–63
103. L.A. Caffarelli, Monotonicity properties of optimal transportation and the FKG and related
inequalities. Commun. Math. Phys. 214(3), 547–563 (2000)
104. P. Cannarsa, C. Sinestrari, Semiconcave Functions, Hamilton-Jacobi Equations, and Opti-
mal Control. Progress in Nonlinear Differential Equations and Their Applications, vol. 58
(Birkhäuser, Boston, 2004)
105. M. Capitaine, E.P. Hsu, M. Ledoux, Martingale representation and a simple proof of log-
arithmic Sobolev inequalities on path spaces. Electron. Commun. Probab. 2, 71–81 (1997)
(electronic)
106. E.A. Carlen, Superadditivity of Fisher’s information and logarithmic Sobolev inequalities.
J. Funct. Anal. 101(1), 194–211 (1991)
107. E.A. Carlen, J.A. Carrillo, M. Loss, Hardy-Littlewood-Sobolev inequalities via fast diffusion
flows. Proc. Natl. Acad. Sci. USA 107(46), 19696–19701 (2010)
108. E.A. Carlen, S. Kusuoka, D.W. Stroock, Upper bounds for symmetric Markov transition
functions. Ann. Inst. Henri Poincaré Probab. Stat. 23(2), 245–287 (1987)
109. E.A. Carlen, E.H. Lieb, M. Loss, A sharp analog of Young’s inequality on S N and related
entropy inequalities. J. Geom. Anal. 14(3), 487–520 (2004)
110. E.A. Carlen, M. Loss, Sharp constant in Nash’s inequality. Int. Math. Res. Not. 7, 213–215
(1993)
111. J.A. Carrillo, A. Jüngel, P.A. Markowich, G. Toscani, A. Unterreiter, Entropy dissipation
methods for degenerate parabolic problems and generalized Sobolev inequalities. Monat-
shefte Math. 133(1), 1–82 (2001)
112. J.A. Carrillo, R.J. McCann, C. Villani, Kinetic equilibration rates for granular media and
related equations: entropy dissipation and mass transportation estimates. Rev. Mat. Iberoam.
19(3), 971–1018 (2003)
113. J.A. Carrillo, R.J. McCann, C. Villani, Contractions in the 2-Wasserstein length space and
thermalization of granular media. Arch. Ration. Mech. Anal. 179, 217–263 (2006)
114. G. Carron, Inégalités isopérimétriques de Faber-Krahn et conséquences, in Actes de la Table
Ronde de Géométrie Différentielle, Luminy, 1992. Sémin. Congr., vol. 1 (Soc. Math. France,
Paris, 1996), pp. 205–232
532 Bibliography
115. P. Cattiaux, A pathwise approach of some classical inequalities. Potential Anal. 20(4),
361–394 (2004)
116. P. Cattiaux, I. Gentil, A. Guillin, Weak logarithmic Sobolev inequalities and entropic conver-
gence. Probab. Theory Relat. Fields 139(3–4), 563–603 (2007)
117. P. Cattiaux, A. Guillin, On quadratic transportation cost inequalities. J. Math. Pures Appl.
86(4), 341–361 (2006)
118. P. Cattiaux, A. Guillin, Long Time Behavior of Markov process and Functional Inequalities.
Forthcoming Monograph, 2014
119. P. Cattiaux, A. Guillin, F.-Y. Wang, L. Wu, Lyapunov conditions for super Poincaré inequal-
ities. J. Funct. Anal. 256(6), 1821–1841 (2009)
120. D. Chafaï, Entropies, convexity, and functional inequalities: on -entropies and -Sobolev
inequalities. J. Math. Kyoto Univ. 44(2), 325–363 (2004)
121. I. Chavel, Eigenvalues in Riemannian Geometry. Pure and Applied Mathematics, vol. 115
(Academic Press, Orlando, 1984). Including a chapter by Burton Randol, with an appendix
by Jozef Dodziuk
122. I. Chavel, Isoperimetric Inequalities. Cambridge Tracts in Mathematics, vol. 145 (Cambridge
University Press, Cambridge, 2001). Differential geometric and analytic perspectives
123. I. Chavel, A modern introduction, in Riemannian Geometry. Cambridge Studies in Advanced
Mathematics, vol. 98, 2nd edn. (Cambridge University Press, Cambridge, 2006)
124. J. Cheeger, A lower bound for the smallest eigenvalue of the Laplacian, in Problems in Anal-
ysis (Papers Dedicated to Salomon Bochner, 1969) (Princeton Univ. Press, Princeton, 1970),
pp. 195–199
125. J. Cheeger, Differentiability of Lipschitz functions on metric measure spaces. Geom. Funct.
Anal. 9(3), 428–517 (1999)
126. J. Cheeger, D.G. Ebin, Comparison Theorems in Riemannian Geometry (AMS Chelsea,
Providence, 2008). Revised reprint of the 1975 original
127. L.H.Y. Chen, An inequality for the multivariate normal distribution. J. Multivar. Anal. 12(2),
306–315 (1982)
128. M.-F. Chen, Trilogy of couplings and general formulas for lower bound of spectral gap, in
Probability Towards 2000, New York, 1995. Lecture Notes in Statist., vol. 128 (Springer,
New York, 1998), pp. 123–136
129. M.-F. Chen, F.-Y. Wang, Application of coupling method to the first eigenvalue on manifold.
Prog. Nat. Sci. 5(2), 227–229 (1995)
130. H. Chernoff, A note on an inequality involving the normal distribution. Ann. Probab. 9(3),
533–535 (1981)
131. T.S. Chihara, An Introduction to Orthogonal Polynomials. Mathematics and Its Applications,
vol. 13 (Gordon and Breach, New York, 1978)
132. K.L. Chung, Lectures from Markov Processes to Brownian Motion. Grundlehren der math-
ematischen Wissenschaften [Fundamental Principles of Mathematical Science], vol. 249
133. D. Cordero-Erausquin, Some applications of mass transport to Gaussian-type inequalities.
Arch. Ration. Mech. Anal. 161(3), 257–269 (2002)
134. D. Cordero-Erausquin, W. Gangbo, C. Houdré, Inequalities for generalized entropy and op-
timal transportation, in Recent Advances in the Theory and Applications of Mass Transport.
Contemp. Math., vol. 353 (Am. Math. Soc., Providence, 2004), pp. 73–94
135. D. Cordero-Erausquin, R.J. McCann, M. Schmuckenschläger, A Riemannian interpolation
inequality à la Borell, Brascamp and Lieb. Invent. Math. 146(2), 219–257 (2001)
136. D. Cordero-Erausquin, R.J. McCann, M. Schmuckenschläger, Prékopa-Leindler type in-
equalities on Riemannian manifolds, Jacobi fields, and optimal transport. Ann. Fac. Sci.
Toulouse 15(4), 613–635 (2006)
137. D. Cordero-Erausquin, B. Nazaret, C. Villani, A mass-transportation approach to sharp
Sobolev and Gagliardo-Nirenberg inequalities. Adv. Math. 182(2), 307–332 (2004)
138. T. Coulhon, Inégalités de Gagliardo-Nirenberg pour les semi-groupes d’opérateurs et appli-
cations. Potential Anal. 1(4), 343–353 (1992)
Bibliography 533
139. T. Coulhon, Ultracontractivity and Nash type inequalities. J. Funct. Anal. 141(2), 510–539
(1996)
140. R. Courant, D. Hilbert, Methoden der mathematischen Physik, vol. 2 (1937) (German)
141. T.M. Cover, J.A. Thomas, Elements of Information Theory, 2nd edn. (Wiley-Interscience,
Hoboken, 2006)
142. G. Da Prato, An Introduction to Infinite-Dimensional Analysis. Universitext (Springer, Berlin,
2006). Revised and extended from the 2001 original by Da Prato
143. E.B. Davies, One-Parameter Semigroups. London Mathematical Society Monographs,
vol. 15 (Academic Press, London, 1980)
144. E.B. Davies, Heat Kernels and Spectral Theory. Cambridge Tracts in Mathematics, vol. 92
(Cambridge University Press, Cambridge, 1989)
145. E.B. Davies, L. Gross, B. Simon, Hypercontractivity: a bibliographic review, in Ideas and
Methods in Quantum and Statistical Physics, Oslo, 1988 (Cambridge Univ. Press, Cam-
bridge, 1992), pp. 370–389
146. E.B. Davies, B. Simon, Ultracontractivity and the heat kernel for Schrödinger operators and
Dirichlet Laplacians. J. Funct. Anal. 59(2), 335–395 (1984)
147. M. Del Pino, J. Dolbeault, Best constants for Gagliardo-Nirenberg inequalities and applica-
tions to nonlinear diffusions. J. Math. Pures Appl. 81(9), 847–875 (2002)
148. M. Del Pino, J. Dolbeault, The optimal Euclidean Lp -Sobolev logarithmic inequality.
J. Funct. Anal. 197(1), 151–161 (2003)
149. H. Delin, A proof of the equivalence between Nash and Sobolev inequalities. Bull. Sci. Math.
120(4), 405–411 (1996)
150. C. Dellacherie, Capacités et Processus Stochastiques. Ergebnisse der Mathematik und ihrer
Grenzgebiete, vol. 67 (Springer, Berlin, 1972)
151. C. Dellacherie, P.-A. Meyer, Probabilities and Potential. North-Holland Mathematics Stud-
ies, vol. 29 (North-Holland, Amsterdam, 1978)
152. C. Dellacherie, P.-A. Meyer, Theory of martingales, in Probabilities and Potential. B. North-
Holland Mathematics Studies, vol. 72 (North-Holland, Amsterdam, 1982). Translated from
the French by J.P. Wilson
153. C. Dellacherie, P.-A. Meyer, Théorie du potentiel associée à une résolvante. Théorie des pro-
cessus de Markov, in Probabilités et Potentiel. 2nd edn. Publications de l’Institut de Mathé-
matiques de l’Université de Strasbourg (Hermann, Paris, 1987). Chapitres XII–XVI
154. C. Dellacherie, P.-A. Meyer, Potential theory for discrete and continuous semigroups,
in Probabilities and Potential. C. North-Holland Mathematics Studies, vol. 151 (North-
Holland, Amsterdam, 1988). Translated from the French by J.R. Norris
155. J. Demange, Porous media equation and Sobolev inequalities under negative curvature. Bull.
Sci. Math. 129(10), 804–830 (2005)
156. J. Demange, Improved Gagliardo-Nirenberg-Sobolev inequalities on manifolds with positive
curvature. J. Funct. Anal. 254(3), 593–611 (2008)
157. A. Dembo, T.M. Cover, J.A. Thomas, Information-theoretic inequalities. IEEE Trans. Inf.
Theory 37(6), 1501–1518 (1991)
158. J.-D. Deuschel, D.W. Stroock, Large Deviations. Pure and Applied Mathematics, vol. 137
(Academic Press, Boston, 1989)
159. M.P. do Carmo, Riemannian Geometry. Mathematics: Theory & Applications (Birkhäuser,
Boston, 1992). Translated from the second Portuguese edition by Francis Flaherty
160. J. Dolbeault, M.J. Esteban, M. Kowalczyk, M. Loss, Sharp interpolation inequalities on the
sphere: new methods and consequences. Chin. Ann. Math., Ser. B 34(1), 99–112 (2013)
161. J. Dolbeault, M. Esteban, M. Loss, Nonlinear flows and rigidity results on compact mani-
folds. Preprint, 2013
162. J. Dolbeault, I. Gentil, A. Guillin, F.-Y. Wang, Lq -functional inequalities and weighted
porous media equations. Potential Anal. 28(1), 35–59 (2008)
163. H. Donnelly, Exhaustion functions and the spectrum of Riemannian manifolds. Indiana Univ.
Math. J. 46(2), 505–527 (1997)
534 Bibliography
164. J.L. Doob, Classical Potential Theory and Its Probabilistic Counterpart. Grundlehren
vol. 262 (Springer, New York, 1984)
165. O. Druet, E. Hebey, The AB program in geometric analysis: sharp Sobolev inequalities and
related problems. Mem. Am. Math. Soc. 160, 761 (2002), pp. viii+98
166. B.A. Dubrovin, A.T. Fomenko, S.P. Novikov, The geometry and topology of manifolds,
in Modern Geometry—Methods and Applications. Part II. Graduate Texts in Mathematics,
vol. 104 (Springer, New York, 1985). Translated from the Russian by Robert G. Burns
167. B.A. Dubrovin, A.T. Fomenko, S.P. Novikov, The geometry of surfaces, transformation
groups, and fields, in Modern Geometry—Methods and Applications. Part I, 2nd edn. Grad-
uate Texts in Mathematics, vol. 93 (Springer, New York, 1992). Translated from the Russian
by Robert G. Burns
168. N. Dunford, J.T. Schwartz, Linear Operators. I. General Theory. Pure and Applied Math-
ematics, vol. 7 (Interscience, New York, 1958). With the assistance of W.G. Bade and
R.G. Bartle
169. N. Dunford, J.T. Schwartz, Linear Operators. Part II: Spectral Theory. Self Adjoint Op-
erators in Hilbert Space (Interscience, New York, 1963). With the assistance of William
G. Bade, and Robert G. Bartle
170. E.B. Dynkin, Markov Processes. Vols. I, II. Die Grundlehren der mathematischen Wis-
senschaften, vols. 121, 122 (Academic Press, New York, 1965). Translated with the autho-
rization and assistance of the author by J. Fabius, V. Greenberg, A. Maitra, G. Majone
171. K.D. Elworthy, Stochastic Differential Equations on Manifolds. London Mathematical Soci-
ety Lecture Note Series, vol. 70 (Cambridge University Press, Cambridge, 1982)
172. K.D. Elworthy, Geometric aspects of diffusions on manifolds, in École d’Été de Probabil-
ités de Saint-Flour XV–XVII, 1988–87. Lecture Notes in Math., vol. 1362 (Springer, Berlin,
1985), pp. 277–425
173. M. Émery, J.E. Yukich, A simple proof of the logarithmic Sobolev inequality on the circle, in
Séminaire de Probabilités, XXI. Lecture Notes in Math., vol. 1247 (Springer, Berlin, 1987),
pp. 173–175
174. K.-J. Engel, R. Nagel, One-Parameter Semigroups for Linear Evolution Equations. Graduate
Texts in Mathematics, vol. 194 (Springer, New York, 2000). With contributions by S. Bren-
dle, M. Campiti, T. Hahn, G. Metafune, G. Nickel, D. Pallara, C. Perazzoli, A. Rhandi, S. Ro-
manelli and R. Schnaubelt
175. M. Erbar, The heat equation on manifolds as a gradient flow in the Wasserstein space. Ann.
Inst. Henri Poincaré Probab. Stat. 46(1), 1–23 (2010)
176. M. Erbar, K. Kuwada, K.-T. Sturm, On the equivalence of the entropy curvature-dimension
condition and Bochner’s inequality on metric measure spaces. Preprint, 2013
177. J.F. Escobar, Sharp constant in a Sobolev trace inequality. Indiana Univ. Math. J. 37(3),
687–698 (1988)
178. S.N. Ethier, T.G. Kurtz, Characterization and convergence, in Markov Processes. Wiley Se-
ries in Probability and Mathematical Statistics: Probability and Mathematical Statistics (Wi-
ley, New York, 1986)
179. L.C. Evans, Partial Differential Equations, 2nd edn. Graduate Studies in Mathematics,
vol. 19 (American Mathematical Society, Providence, 2010)
180. L.C. Evans, R.F. Gariepy, Measure Theory and Fine Properties of Functions. Studies in
Advanced Mathematics (CRC Press, Boca Raton, 1992)
181. E.B. Fabes, D.W. Stroock, A new proof of Moser’s parabolic Harnack inequality using the
old ideas of Nash. Arch. Ration. Mech. Anal. 96(4), 327–338 (1986)
182. S. Fang, Inégalité du type de Poincaré sur l’espace des chemins Riemanniens. C. R. Math.
Acad. Sci. Paris, Sér. I 318(3), 257–260 (1994)
183. P. Federbush, A partially alternate derivation of a result of Nelson. J. Math. Phys. 10(1),
50–52 (1969)
184. H. Federer, Geometric Measure Theory. Die Grundlehren der mathematischen Wis-
senschaften, vol. 153 (Springer, New York, 1969)
Bibliography 535
185. É. Fontenas, Sur les constantes de Sobolev des variétés riemanniennes compactes et les fonc-
tions extrémales des sphères. Bull. Sci. Math. 121(2), 71–96 (1997)
186. P. Fougères, Spectral gap for log-concave probability measures on the real line, in Sémi-
naire de Probabilités XXXVIII. Lecture Notes in Math., vol. 1857 (Springer, Berlin, 2005),
pp. 95–123
187. J. Franchi, Y. Le Jan, Hyperbolic Dynamics and Brownian Motions: An Introduction (Oxford
University Press, Oxford, 2012)
188. Y. Fujita, An optimal logarithmic Sobolev inequality with Lipschitz constants. J. Funct. Anal.
261(5), 1133–1144 (2011)
189. M. Fukushima, Dirichlet Forms and Markov Processes. North-Holland Mathematical Li-
brary, vol. 23 (North-Holland, Amsterdam, 1980)
190. M. Fukushima, Y. Oshima, M. Takeda, Dirichlet Forms and Symmetric Markov Processes.
de Gruyter Studies in Mathematics, vol. 19 (Walter de Gruyter, Berlin, 2011), extended edn.
191. M.P. Gaffney, The conservation property of the heat equation on Riemannian manifolds.
Commun. Pure Appl. Math. 12, 1–11 (1959)
192. E. Gagliardo, Proprietà di alcune classi di funzioni in più variabili. Ric. Mat. 7, 102–137
(1958)
193. S. Gallot, Inégalités isopérimétriques et analytiques sur les variétés Riemanniennes.
Astérisque (1988), no. 163–164, pp. 5–6, 31–91, 281 (1989). On the geometry of differ-
entiable manifolds (Rome, 1986)
194. S. Gallot, D. Hulin, J. Lafontaine, Riemannian Geometry, 3nd edn. Universitext (Springer,
Berlin, 2004)
195. R.J. Gardner, The Brunn-Minkowski inequality. Bull. Am. Math. Soc. (N.S.) 39(3), 355–405
(2002)
196. A.M. Garsia, Martingale Inequalities: Seminar Notes on Recent Progress. Mathematics Lec-
ture Notes Series (Benjamin, Reading, 1973)
197. I. Gentil, Ultracontractive bounds on Hamilton-Jacobi solutions. Bull. Sci. Math. 126(6),
507–524 (2002)
198. I. Gentil, The general optimal Lp -Euclidean logarithmic Sobolev inequality by Hamilton-
Jacobi equations. J. Funct. Anal. 202(2), 591–599 (2003)
199. I. Gentil, From the Prékopa-Leindler inequality to modified logarithmic Sobolev inequality.
Ann. Fac. Sci. Toulouse 17(2), 291–308 (2008)
200. I. Gentil, A. Guillin, L. Miclo, Modified logarithmic Sobolev inequalities and transportation
inequalities. Probab. Theory Relat. Fields 133(3), 409–436 (2005)
201. I. Gentil, A. Guillin, L. Miclo, Modified logarithmic Sobolev inequalities in null curvature.
Rev. Mat. Iberoam. 23(1), 235–258 (2007)
202. B. Gidas, J. Spruck, Global and local behavior of positive solutions of nonlinear elliptic
equations. Commun. Pure Appl. Math. 34(4), 525–598 (1981)
203. D. Gilbarg, N.S. Trudinger, Elliptic Partial Differential Equations of Second Order. Classics
in Mathematics (Springer, Berlin, 2001). Reprint of the 1998 edition
204. J. Glimm, Boson fields with nonlinear selfinteraction in two dimensions. Commun. Math.
Phys. 8, 12–25 (1968) (English)
205. J.A. Goldstein, Semigroups of Linear Operators and Applications. Oxford Mathematical
Monographs (The Clarendon Press, New York, 1985)
206. N. Gozlan, A characterization of dimension free concentration in terms of transportation
inequalities. Ann. Probab. 37(6), 2480–2498 (2009)
207. N. Gozlan, Poincaré inequalities and dimension free concentration of measure. Ann. Inst.
Henri Poincaré Probab. Stat. 46(3), 708–739 (2010)
208. N. Gozlan, Transport-entropy inequalities on the line. Electron. J. Probab. 17(49), 1–18
(2012)
209. N. Gozlan, C. Léonard, A large deviation approach to some transportation cost inequalities.
Probab. Theory Relat. Fields 139(1–2), 235–283 (2007)
210. N. Gozlan, C. Léonard, Transport inequalities. A survey. Markov Process. Relat. Fields 16,
635–736 (2010)
536 Bibliography
211. N. Gozlan, C. Roberto, P.-M. Samson, From concentration to logarithmic Sobolev and
Poincaré inequalities. J. Funct. Anal. 260(5), 1491–1522 (2011)
212. N. Gozlan, C. Roberto, P.-M. Samson, Hamilton Jacobi equations on metric spaces and trans-
port entropy inequalities. Rev. Mat. Iberoam. (2013, to appear)
213. N. Gozlan, C. Roberto, P.-M. Samson, Characterization of Talagrand’s transport-entropy in-
equality in metric spaces. Ann. Probab. 41(5), 3112–3139 (2013)
214. A. Grigor’yan, The heat equation on noncompact Riemannian manifolds. Mat. Sb. 182(1),
55–87 (1991)
215. A. Grigor’yan, Analytic and geometric background of recurrence and non-explosion of the
Brownian motion on Riemannian manifolds. Bull. Am. Math. Soc. (N.S.) 36(2), 135–249
(1999)
216. A. Grigor’yan, Isoperimetric inequalities and capacities on Riemannian manifolds, in The
Maz’ya Anniversary Collection, Vol. 1, Rostock, 1998. Oper. Theory Adv. Appl., vol. 109
(Birkhäuser, Basel, 1999), pp. 139–153
217. A. Grigor’yan, Heat Kernel and Analysis on Manifolds. AMS/IP Studies in Advanced Math-
ematics, vol. 47 (American Mathematical Society, Providence, 2009)
218. A. Grigor’yan, Yau’s work on heat kernels, in Geometry and Analysis. No. 1. Adv. Lect.
Math. (ALM), vol. 17 (Int. Press, Somerville, 2011), pp. 113–117
219. G. Grillo, On Persson’s theorem in local Dirichlet spaces. Z. Anal. Anwend. 17(2), 329–338
(1998)
220. M. Gromov, Paul Lévy’s isoperimetric inequality. Preprint, 1980
221. M. Gromov, Metric Structures for Riemannian and Non-Riemannian Spaces. Progress in
Mathematics, vol. 152 (Birkhäuser, Boston, 1999). Based on the 1981 French original. With
appendices by M. Katz, P. Pansu and S. Semmes, Translated from the French by Sean
Michael Bates
222. M. Gromov, V.D. Milman, A topological application of the isoperimetric inequality. Am. J.
Math. 105(4), 843–854 (1983)
223. L. Gross, Existence and uniqueness of physical ground states. J. Funct. Anal. 10, 52–109
(1972)
224. L. Gross, Logarithmic Sobolev inequalities. Am. J. Math. 97(4), 1061–1083 (1975)
225. L. Gross, Logarithmic Sobolev inequalities and contractivity properties of semigroups, in
Dirichlet Forms, Varenna, 1992. Lecture Notes in Math., vol. 1563 (Springer, Berlin, 1993),
pp. 54–88
226. L. Gross, Hypercontractivity, logarithmic Sobolev inequalities, and applications: a survey
of surveys, in Diffusion, Quantum Theory, and Radically Elementary Mathematics. Math.
Notes, vol. 47 (Princeton Univ. Press, Princeton, 2006), pp. 45–73
227. L. Gross, O.S. Rothaus, Herbst inequalities for supercontractive semigroups. J. Math. Kyoto
Univ. 38(2), 295–318 (1998)
228. O. Guédon, Concentration phenomena in high dimensional geometry. ESAIM Proc. (2013,
to appear)
229. A. Guionnet, B. Zegarliński, Lectures on logarithmic Sobolev inequalities, in Séminaire de
Probabilités, XXXVI. Lecture Notes in Math., vol. 1801 (Springer, Berlin, 2003), pp. 1–134
230. P. Hajłasz, P. Koskela, Sobolev met Poincaré. Mem. Am. Math. Soc. 145, 688 (2000),
p. x+101
231. G.H. Hardy, Note on a theorem of Hilbert. Math. Z. 6(3–4), 314–317 (1920)
232. G.H. Hardy, J.E. Littlewood, G. Pólya, Inequalities, 2nd edn. (Cambridge University Press,
Cambridge, 1952)
233. E. Hebey, Sobolev Spaces on Riemannian Manifolds. Lecture Notes in Mathematics,
vol. 1635 (Springer, Berlin, 1996)
234. E. Hebey, Nonlinear Analysis on Manifolds: Sobolev Spaces and Inequalities. Courant Lec-
ture Notes in Mathematics, vol. 5 (New York University Courant Institute of Mathematical
Sciences, New York, 1999)
235. E. Hebey, M. Vaugon, The best constant problem in the Sobolev embedding theorem for
complete Riemannian manifolds. Duke Math. J. 79(1), 235–279 (1995)
Bibliography 537
236. J. Heinonen, Lectures on Analysis on Metric Spaces. Universitext (Springer, New York,
2001)
237. B. Helffer, Remarks on decay of correlations and Witten Laplacians, Brascamp-Lieb inequal-
ities and semiclassical limit. J. Funct. Anal. 155(2), 571–586 (1998)
238. B. Helffer, Semiclassical Analysis, Witten Laplacians, and Statistical Mechanics. Series in
Partial Differential Equations and Applications, vol. 1 (World Scientific, River Edge, 2002)
239. B. Helffer, J. Sjöstrand, On the correlation for Kac-like models in the convex case. J. Stat.
Phys. 74(1–2), 349–409 (1994)
240. S. Helgason, Integral geometry, invariant differential operators, and spherical functions, in
Groups and Geometric Analysis. Mathematical Surveys and Monographs, vol. 83 (American
Mathematical Society, Providence, 2000). Corrected reprint of the 1984 original
241. E. Hille, R.S. Phillips, Functional Analysis and Semi-Groups. American Mathematical So-
ciety Colloquium Publications, vol. 31 (American Mathematical Society, Providence, 1957),
rev. ed.
242. M. Hino, On short time asymptotic behavior of some symmetric diffusions on general state
spaces. Potential Anal. 16(3), 249–264 (2002)
243. F. Hirsch, Intrinsic metrics and Lipschitz functions. J. Evol. Equ. 3(1), 11–25 (2003). Dedi-
cated to Philippe Bénilan
244. P.D. Hislop, I.M. Sigal, Introduction to Spectral Theory. Applied Mathematical Sciences,
vol. 113 (Springer, New York, 1996). With applications to Schrödinger operators
245. R. Holley, D.W. Stroock, Logarithmic Sobolev inequalities and stochastic Ising models.
J. Stat. Phys. 46, 1159–1194 (1987)
246. L. Hörmander, L2 estimates and existence theorems for the ∂¯ operator. Acta Math. 113,
89–152 (1965)
247. L. Hörmander, Differential operators with constant coefficients, in The Analysis of Linear
Partial Differential Operators. II. Grundlehren der mathematischen Wissenschaften [Funda-
mental Principles of Mathematical Sciences], vol. 257 (Springer, Berlin, 1983)
248. L. Hörmander, Distribution theory and Fourier analysis, in The Analysis of Linear Partial
Differential Operators. I, 2nd edn. (Springer, Berlin, 1990). Springer Study Edition
249. L. Hörmander, An Introduction to Complex Analysis in Several Variables, 3nd edn. North-
Holland Mathematical Library, vol. 7 (North-Holland, Amsterdam, 1990)
250. E.P. Hsu, Logarithmic Sobolev inequalities on path spaces over Riemannian manifolds. Com-
mun. Math. Phys. 189(1), 9–16 (1997)
251. E.P. Hsu, Stochastic Analysis on Manifolds. Graduate Studies in Mathematics, vol. 38 (Amer-
ican Mathematical Society, Providence, 2002)
252. N. Ikeda, S. Watanabe, Stochastic Differential Equations and Diffusion Processes, 2nd edn.
North-Holland Mathematical Library, vol. 24 (North-Holland, Amsterdam, 1989)
253. S. Ilias, Constantes explicites pour les inégalités de Sobolev sur les variétés Riemanniennes
compactes. Ann. Inst. Fourier (Grenoble) 33(2), 151–165 (1983)
254. J. Inglis, M. Neklyudov, B. Zegarliński, Ergodicity for infinite particle systems with locally
conserved quantities. Infin. Dimens. Anal. Quantum Probab. Relat. Top. 15(1), 1250005
(2012), 28
255. K. Itô, H.P. McKean, Diffusion Processes and Their Sample Paths. Die Grundlehren der
mathematischen Wissenschaften, vol. 125 (Springer, Berlin, 1974). Second printing, cor-
rected
256. J. Jacod, Calcul Stochastique et Problèmes de Martingales. Lecture Notes in Mathematics,
257. S. Janson, Gaussian Hilbert Spaces. Cambridge Tracts in Mathematics, vol. 129 (Cambridge
University Press, Cambridge, 1997)
258. F. John, Partial Differential Equations, 4th edn. Applied Mathematical Sciences, vol. 1
259. R. Jordan, D. Kinderlehrer, F. Otto, The variational formulation of the Fokker-Planck equa-
tion. SIAM J. Math. Anal. 29(1), 1–17 (1998)
538 Bibliography
260. J. Jost, Riemannian Geometry and Geometric Analysis, 5th edn. Universitext (Springer,
Berlin, 2008)
261. A. Joulin, N. Privault, Functional inequalities for discrete gradients and application to the
geometric distribution. ESAIM Probab. Stat. 8, 87–101 (2004)
262. R. Kannan, L. Lovász, M. Simonovits, Isoperimetric problems for convex bodies and a lo-
calization lemma. Discrete Comput. Geom. 13(3–4), 541–559 (1995)
263. I. Karatzas, S.E. Shreve, Brownian Motion and Stochastic Calculus, 2nd edn. Graduate Texts
in Mathematics, vol. 113 (Springer, New York, 1991)
264. T. Kato, Perturbation Theory for Linear Operators, 2nd edn. Grundlehren der mathematis-
chen Wissenschaften, vol. 132 (Springer, Berlin, 1976)
265. O. Kavian, G. Kerkyacharian, B. Roynette, Quelques remarques sur l’ultracontractivité.
J. Funct. Anal. 111(1), 155–196 (1993)
266. H. Knothe, Contributions to the theory of convex bodies. Mich. Math. J. 4, 39–52 (1957)
267. M. Knott, C.S. Smith, On the optimal mapping of distributions. J. Optim. Theory Appl. 43(1),
39–49 (1984)
268. R. Koekoek, P.A. Lesky, R.F. Swarttouw, Hypergeometric Orthogonal Polynomials and Their
q-Analogues. Springer Monographs in Mathematics (Springer, Berlin, 2010). With a fore-
word by Tom H. Koornwinder
269. A. Kufner, L.-E. Persson, Weighted Inequalities of Hardy Type (World Scientific, River Edge,
2003)
270. K. Kuwada, Duality on gradient estimates and Wasserstein controls. J. Funct. Anal. 258(11),
3758–3774 (2010)
271. K. Kuwada, Space-time Wasserstein controls and Bakry-Ledoux type gradient estimates.
Preprint, 2013
272. R. Latała, K. Oleszkiewicz, Between Sobolev and Poincaré, in Geometric Aspects of Func-
tional Analysis. Lecture Notes in Math., vol. 1745 (Springer, Berlin, 2000), pp. 147–168
273. M. Ledoux, A simple analytic proof of an inequality by P. Buser. Proc. Am. Math. Soc.
121(3), 951–959 (1994)
274. M. Ledoux, L’algèbre de Lie des gradients itérés d’un générateur markovien—
développements de moyennes et entropies. Ann. Sci. Éc. Norm. Super. 28(4), 435–460
(1995)
275. M. Ledoux, Remarks on logarithmic Sobolev constants, exponential integrability and bounds
on the diameter. J. Math. Kyoto Univ. 35(2), 211–220 (1995)
276. M. Ledoux, Concentration of measure and logarithmic Sobolev inequalities, in Sémi-
naire de Probabilités, XXXIII. Lecture Notes in Math., vol. 1709 (Springer, Berlin, 1999),
pp. 120–216
277. M. Ledoux, The geometry of Markov diffusion generators. Ann. Fac. Sci. Toulouse 9(2),
305–366 (2000). Probability theory
278. M. Ledoux, The Concentration of Measure Phenomenon. Mathematical Surveys and Mono-
graphs, vol. 89 (American Mathematical Society, Providence, 2001)
279. M. Ledoux, Spectral gap, logarithmic Sobolev constant, and geometric bounds, in Surveys in
Differential Geometry, vol. IX (Int. Press, Somerville, 2004), pp. 219–240
280. M. Ledoux, From concentration to isoperimetry: semigroup proofs, in Concentration, Func-
tional Inequalities and Isoperimetry. Contemp. Math., vol. 545 (Am. Math. Soc., Providence,
2011), pp. 155–166
281. P. Lévy, Problèmes Concrets d’Analyse Fonctionnelle, 2nd edn. Des Leçons d’Analyse Fonc-
tionnelle (Gauthier-Villars, Paris, 1951) (French). Avec un complément par F. Pellegrino
282. H.-Q. Li, Estimation optimale du gradient du semi-groupe de la chaleur sur le groupe de
Heisenberg. J. Funct. Anal. 236(2), 369–394 (2006)
283. H.-Q. Li, Estimations asymptotiques du noyau de la chaleur sur les groupes de Heisenberg.
C. R. Math. Acad. Sci. Paris 344(8), 497–502 (2007)
284. P. Li, Geometric Analysis (Cambridge University Press, Cambridge, 2012)
285. P. Li, S.-T. Yau, On the parabolic kernel of the Schrödinger operator. Acta Math. 156(3–4),
153–201 (1986)
Bibliography 539
286. A. Lichnerowicz, Géométrie des Groupes de Transformations. Travaux et Recherches Math-

ématiques, vol. III (Dunod, Paris, 1958)
287. E.H. Lieb, Gaussian kernels have only Gaussian maximizers. Invent. Math. 102(1), 179–208
(1990)
288. E.H. Lieb, M. Loss, Analysis, 2nd edn. Graduate Studies in Mathematics, vol. 14 (American
Mathematical Society, Providence, 2001)
289. R.S. Liptser, A.N. Shiryayev, Theory of Martingales. Mathematics and Its Applications (So-
viet Series), vol. 49 (Kluwer Academic, Dordrecht, 1989). Translated from the Russian by
K. Dzjaparidze [Kacha Dzhaparidze]
290. J. Lott, Some geometric properties of the Bakry-Émery-Ricci tensor. Comment. Math. Helv.
78(4), 865–883 (2003)
291. J. Lott, C. Villani, Hamilton-Jacobi semigroup on length spaces and applications. J. Math.
Pures Appl. 88(3), 219–229 (2007)
292. J. Lott, C. Villani, Ricci curvature for metric-measure spaces via optimal transport. Ann.
Math. (2) 169(3), 903–991 (2009)
293. L. Lovász, M. Simonovits, Random walks in a convex body and an improved volume algo-
rithm. Random Struct. Algorithms 4(4), 359–412 (1993)
294. Z.M. Ma, M. Röckner, Introduction to the Theory of (Nonsymmetric) Dirichlet Forms. Uni-
versitext (Springer, Berlin, 1992)
295. F. Maggi, Sets of Finite Perimeter and Geometric Variational Problems (Cambridge Univer-
sity Press, Cambridge, 2012)
296. F. Maggi, C. Villani, Balls have the worst best Sobolev inequalities. II. Variants and exten-
sions. Calc. Var. Partial Differ. Equ. 31(1), 47–74 (2008)
297. P. Malliavin, Stochastic Analysis. Grundlehren der mathematischen Wissenschaften [Funda-
mental Principles of Mathematical Sciences], vol. 313 (Springer, Berlin, 1997)
298. K. Marton, Bounding d-distance by informational divergence: a method to prove measure
concentration. Ann. Probab. 24(2), 857–866 (1996)
299. P. Mathieu, Quand l’inégalité log-Sobolev implique l’inégalité de trou spectral, in Sémi-
naire de Probabilités, XXXII. Lecture Notes in Math., vol. 1686 (Springer, Berlin, 1998),
pp. 30–35
300. J. Mawhin, Henri Poincaré and partial differential equations. Nieuw Arch. Wiskd. 13(3),
159–169 (2012)
301. O. Mazet, Classification des semi-groupes de diffusion sur R associés à une famille de
polynômes orthogonaux, in Séminaire de Probabilités, XXXI. Lecture Notes in Math.,
vol. 1655 (Springer, Berlin, 1997), pp. 40–53
302. V.G. Maz’ja, Sobolev Spaces. Springer Series in Soviet Mathematics (Springer, Berlin,
1985). Translated from the Russian by T.O. Shaposhnikova
303. V.G. Maz’ya, Sobolev Spaces with Applications to Elliptic Partial Differential Equations.
Grundlehren der mathematischen Wissenschaften [Fundamental Principles of Mathematical
Sciences], vol. 342 (Springer, Heidelberg, 2011). augmented ed.
304. R.J. McCann, Existence and uniqueness of monotone measure-preserving maps. Duke Math.
J. 80(2), 309–323 (1995)
305. R.J. McCann, A convexity principle for interacting gases. Adv. Math. 128(1), 153–179
(1997)
306. H.P. McKean, An upper bound to the spectrum of on a manifold of negative curvature.
J. Differ. Geom. 4, 359–366 (1970)
307. P.-A. Meyer, Probability and Potentials (Blaisdell, Waltham, 1966)
308. P.-A. Meyer, Note sur les processus d’Ornstein-Uhlenbeck, in Seminar on Probability, XVI.
Lecture Notes in Math., vol. 920 (Springer, Berlin, 1982), pp. 95–133
309. S. Meyn, R.L. Tweedie, Markov Chains and Stochastic Stability, 2nd edn. (Cambridge Uni-
versity Press, Cambridge, 2009). With a prologue by Peter W. Glynn
310. L. Miclo, Quand est-ce que des bornes de Hardy permettent de calculer une constante de
Poincaré exacte sur la droite? Ann. Fac. Sci. Toulouse 17(1), 121–192 (2008)
311. L. Miclo, On hyperboundedness and spectrum of Markov operators. Preprint, 2013
540 Bibliography
312. E. Milman, On the role of convexity in isoperimetry, spectral gap and concentration. Invent.
Math. 177(1), 1–43 (2009)
313. E. Milman, A converse to the Maz’ya inequality for capacities under curvature lower bound,
in Around the Research of Vladimir Maz’ya. I. Int. Math. Ser. (N. Y.), vol. 11 (Springer, New
York, 2010), pp. 321–348
314. E. Milman, Isoperimetric and concentration inequalities: equivalence under curvature lower
bound. Duke Math. J. 154(2), 207–239 (2010)
315. E. Milman, Sharp isoperimetric inequalities and model spaces for curvature-dimension-
diameter condition. Preprint, 2012
316. V.D. Milman, G. Schechtman, Asymptotic Theory of Finite-Dimensional Normed Spaces.
Lecture Notes in Mathematics, vol. 1200 (Springer, Berlin, 1986). With an appendix by
M. Gromov
317. D.S. Mitrinović, J.E. Pečarić, A.M. Fink, Inequalities Involving Functions and Their Inte-
grals and Derivatives. Mathematics and Its Applications (East European Series), vol. 53
(Kluwer Academic, Dordrecht, 1991)
318. F. Morgan, Manifolds with density. Not. Am. Math. Soc. 52(8), 853–858 (2005)
319. J. Moser, A Harnack inequality for parabolic differential equations. Commun. Pure Appl.
Math. 17, 101–134 (1964)
320. J. Moser, A sharp form of an inequality by N. Trudinger. Indiana Univ. Math. J. 20,
1077–1092 (1970/71)
321. B. Muckenhoupt, Hardy’s inequality with weights. Stud. Math. 44, 31–38 (1972). Collection
of articles honoring the completion by Antoni Zygmund of 50 years of scientific activity, I
322. C.E. Mueller, F.B. Weissler, Hypercontractivity for the heat semigroup for ultraspherical
polynomials and on the n-sphere. J. Funct. Anal. 48(2), 252–283 (1982)
323. J. Nash, Continuity of solutions of parabolic and elliptic equations. Am. J. Math. 80, 931–954
(1958)
324. B. Nazaret, Best constant in Sobolev trace inequalities on the half-space. Nonlinear Anal.
65(10), 1977–1985 (2006)
325. E. Nelson, A quartic interaction in two dimensions, in Mathematical Theory of Elementary
Particles, Dedham, MA, 1965. Proc. Conf. (MIT Press, Cambridge, 1966), pp. 69–73
326. E. Nelson, Dynamical Theories of Brownian Motion (Princeton University Press, Princeton,
1967)
327. E. Nelson, The free Markoff field. J. Funct. Anal. 12, 211–227 (1973)
328. E. Nelson, Quantum fields and Markoff fields, in Partial Differential Equations, Berkeley,
CA, 1971. Proc. Sympos. Pure Math., vol. XXIII (Am. Math. Soc., Providence, 1973),
pp. 413–420
329. J. Neveu, Sur l’espérance conditionnelle par rapport à un mouvement brownien. Ann. Inst.
Henri Poincaré, Sect. B (N. S.) 12(2), 105–109 (1976)
330. V.H. Nguyen, Sharp weighted Sobolev and Gagliardo-Nirenberg inequalities on half-spaces
via mass transport and consequences. Preprint, 2013
331. L. Nirenberg, On elliptic partial differential equations. Ann. Sc. Norm. Super. Pisa, Cl. Sci.
(3) 13, 115–162 (1959)
332. J.R. Norris, Heat kernel asymptotics and the distance function in Lipschitz Riemannian man-
ifolds. Acta Math. 179(1), 79–103 (1997)
333. J.R. Norris, Markov Chains. Cambridge Series in Statistical and Probabilistic Mathematics,
vol. 2 (Cambridge University Press, Cambridge, 1998). Reprint of 1997 original
334. D. Nualart, The Malliavin Calculus and Related Topics, 2nd edn. Probability and Its Appli-
cations (New York) (Springer, Berlin, 2006)
335. M. Obata, Certain conditions for a Riemannian manifold to be isometric with a sphere.
J. Math. Soc. Jpn. 14, 333–340 (1962)
336. B. Øksendal, Stochastic Differential Equations, 6th edn. Universitext (Springer, Berlin,
2003). An introduction with applications
337. E. Onofri, On the positivity of the effective action in a theory of random surfaces. Commun.
Math. Phys. 86(3), 321–326 (1982)
Bibliography 541
338. R. Osserman, The isoperimetric inequality. Bull. Am. Math. Soc. 84(6), 1182–1238 (1978)
339. F. Otto, The geometry of dissipative evolution equations: the porous medium equation. Com-
mun. Partial Differ. Equ. 26(1–2), 101–174 (2001)
340. F. Otto, C. Villani, Generalization of an inequality by Talagrand and links with the logarith-
mic Sobolev inequality. J. Funct. Anal. 173(2), 361–400 (2000)
341. F. Otto, M. Westdickenberg, Eulerian calculus for the contraction in the Wasserstein distance.
SIAM J. Math. Anal. 37(4), 1227–1255 (2005) (electronic)
342. E.M. Ouhabaz, Analysis of Heat Equations on Domains. London Mathematical Society
Monographs Series, vol. 31 (Princeton University Press, Princeton, 2005)
343. L.E. Payne, H.F. Weinberger, An optimal Poincaré inequality for convex domains. Arch.
Ration. Mech. Anal. 5, 286–292 (1960)
344. A. Pazy, Semigroups of Linear Operators and Applications to Partial Differential Equations.
Applied Mathematical Sciences, vol. 44 (Springer, New York, 1983)
345. A. Persson, Bounds for the discrete part of the spectrum of a semi-bounded Schrödinger
operator. Math. Scand. 8, 143–153 (1960)
346. P. Petersen, Riemannian Geometry. Graduate Texts in Mathematics, vol. 171 (Springer, New
York, 1998)
347. R.G. Pinsky, Positive Harmonic Functions and Diffusion. Cambridge Studies in Advanced
Mathematics, vol. 45 (Cambridge University Press, Cambridge, 1995)
348. H. Poincaré, Sur la théorie analytique de la chaleur. C. R. Séances Acad. Sci. 104, 1753–1759
(1887)
349. H. Poincaré, Sur les equations aux derivees partielles de la physique mathematique. Am. J.
Math. 12(3), 211–294 (1890)
350. P.E. Protter, Stochastic Integration and Differential Equations, 2nd edn. Stochastic Mod-
elling and Applied Probability, vol. 21 (Springer, Berlin, 2005). Version 2.1, Corrected third
printing
351. S.T. Rachev, Probability Metrics and the Stability of Stochastic Models. Wiley Series in
Probability and Mathematical Statistics: Applied Probability and Statistics (Wiley, Chich-
ester, 1991)
352. S.T. Rachev, L. Rüschendorf, Mass Transportation Problems. Vol. I. Probability and Its Ap-
plications (New York) (Springer, New York, 1998). Theory
353. S.T. Rachev, L. Rüschendorf, Mass Transportation Problems. Vol. II. Probability and Its
Applications (New York) (Springer, New York, 1998). Applications
354. J.G. Ratcliffe, Foundations of Hyperbolic Manifolds, 2nd edn. Graduate Texts in Mathemat-
ics, vol. 149 (Springer, New York, 2006)
355. M. Reed, B. Simon, Methods of Modern Mathematical Physics. II. Fourier Analysis, Self-
Adjointness (Academic Press, New York, 1975)
356. M. Reed, B. Simon, Methods of Modern Mathematical Physics. I, 2nd edn. (Academic Press,
New York, 1980). Functional analysis
357. D. Revuz, Markov Chains, 2nd edn. North-Holland Mathematical Library, vol. 11 (North-
Holland, Amsterdam, 1984)
358. D. Revuz, M. Yor, Continuous Martingales and Brownian Motion, 3rd edn. Grundlehren
359. F. Riesz, B.Sz. Nagy, Functional Analysis (Frederick Ungar, New York, 1955). Translated
by Leo F. Boron
360. C. Roberto, B. Zegarliński, Orlicz-Sobolev inequalities for sub-Gaussian measures and er-
godicity of Markov semi-groups. J. Funct. Anal. 243(1), 28–66 (2007)
361. M. Röckner, F.-Y. Wang, Weak Poincaré inequalities and L2 -convergence rates of Markov
semigroups. J. Funct. Anal. 185(2), 564–603 (2001)
362. L.C.G. Rogers, D. Williams, Itô calculus, in Diffusions, Markov Processes, and Martingales.
Vol. 2. Wiley Series in Probability and Mathematical Statistics: Probability and Mathematical
Statistics (Wiley, New York, 1987)
542 Bibliography
363. L.C.G. Rogers, D. Williams, Foundations, in Diffusions, Markov Processes, and Martingales.
Vol. 1. Wiley Series in Probability and Mathematical Statistics: Probability and Mathematical
Statistics, 2nd edn. (Wiley, Chichester, 1994)
364. G. Rosen, Minimum value for c in the Sobolev inequality φ 3 ≤ c∇φ3 . SIAM J. Appl.
Math. 21, 30–32 (1971)
365. J. Rosen, Sobolev inequalities for weight spaces and supercontractivity. Trans. Am. Math.
Soc. 222, 367–376 (1976)
366. J.-P. Roth, Opérateurs dissipatifs et semi-groupes dans les espaces de fonctions continues.
Ann. Inst. Fourier (Grenoble) 26, 1–97 (1976)
367. O.S. Rothaus, Logarithmic Sobolev inequalities and the spectrum of Sturm-Liouville opera-
tors. J. Funct. Anal. 39(1), 42–56 (1980)
368. O.S. Rothaus, Diffusion on compact Riemannian manifolds and logarithmic Sobolev inequal-
ities. J. Funct. Anal. 42(1), 102–109 (1981)
369. O.S. Rothaus, Logarithmic Sobolev inequalities and the spectrum of Schrödinger operators.
J. Funct. Anal. 42(1), 110–120 (1981)
370. O.S. Rothaus, Analytic inequalities, isoperimetric inequalities and logarithmic Sobolev in-
equalities. J. Funct. Anal. 64(2), 296–313 (1985)
371. O.S. Rothaus, Hypercontractivity and the Bakry-Emery criterion for compact Lie groups.
J. Funct. Anal. 65(3), 358–367 (1986)
372. G. Royer, An Initiation to Logarithmic Sobolev Inequalities. SMF/AMS Texts and Mono-
graphs, vol. 14 (American Mathematical Society, Providence, 2007). Translated from the
1999 French original by Donald Babbitt
373. L. Rüschendorf, S.T. Rachev, A characterization of random variables with minimum
L2 -distance. J. Multivar. Anal. 32(1), 48–54 (1990)
374. L. Saloff-Coste, A note on Poincaré, Sobolev, and Harnack inequalities. Int. Math. Res. Not.
2, 27–38 (1992)
375. L. Saloff-Coste, Convergence to equilibrium and logarithmic Sobolev constant on manifolds
with Ricci curvature bounded below. Colloq. Math. 67(1), 109–121 (1994)
376. L. Saloff-Coste, Aspects of Sobolev-Type Inequalities. London Mathematical Society Lecture
Note Series, vol. 289 (Cambridge University Press, Cambridge, 2002)
377. L. Saloff-Coste, Sobolev inequalities in familiar and unfamiliar settings, in Sobolev Spaces
in Mathematics. I. Int. Math. Ser. (N. Y.), vol. 8 (Springer, New York, 2009), pp. 299–343
378. L. Saloff-Coste, The heat kernel and its estimates, in Probabilistic Approach to Geometry.
Adv. Stud. Pure Math., vol. 57 (Math. Soc. Japan, Tokyo, 2010), pp. 405–436
379. K.-i. Sato, Lévy Processes and Infinitely Divisible Distributions. Cambridge Studies in Ad-
vanced Mathematics, vol. 68 (Cambridge University Press, Cambridge, 1999). Translated
from the 1990 Japanese original, Revised by the author
380. R. Schoen, Conformal deformation of a Riemannian metric to constant scalar curvature.
J. Differ. Geom. 20(2), 479–495 (1984)
381. R. Schoen, Variational theory for the total scalar curvature functional for Riemannian metrics
and related topics, in Topics in Calculus of Variations, Montecatini Terme, 1987. Lecture
Notes in Math., vol. 1365 (Springer, Berlin, 1989), pp. 120–154
382. R. Schoen, S.-T. Yau, Conformally flat manifolds, Kleinian groups and scalar curvature.
Invent. Math. 92(1), 47–71 (1988)
383. C.E. Shannon, A mathematical theory of communication. Bell Syst. Tech. J. 27, 379–423,
623–656 (1948)
384. B. Simon, R. Høegh-Krohn, Hypercontractive semigroups and two dimensional self-coupled
Bose fields. J. Funct. Anal. 9, 121–180 (1972)
385. J. Sjöstrand, Correlation asymptotics and Witten Laplacians. Algebra Anal. 8(1), 160–191
(1996)
386. S. Sobolev, Sur un théorème d’analyse fonctionnelle. Mat. Sb. (N.S.) 4, 471–497 (1938).
(Russian)
387. A.J. Stam, Some inequalities satisfied by the quantities of information of Fisher and Shannon.
Inf. Control 2, 101–112 (1959)
Bibliography 543
388. E.M. Stein, Singular Integrals and Differentiability Properties of Functions. Princeton Math-
ematical Series, vol. 30 (Princeton University Press, Princeton, 1970)
389. E.M. Stein, G. Weiss, Introduction to Fourier Analysis on Euclidean Spaces. Princeton Math-
ematical Series, vol. 32 (Princeton University Press, Princeton, 1971)
390. R.S. Strichartz, Analysis of the Laplacian on the complete Riemannian manifold. J. Funct.
Anal. 52(1), 48–79 (1983)
391. D.W. Stroock, An Introduction to the Analysis of Paths on a Riemannian Manifold. Math-
ematical Surveys and Monographs, vol. 74 (American Mathematical Society, Providence,
2000)
392. D.W. Stroock, Partial Differential Equations for Probabilists. Cambridge Studies in Ad-
vanced Mathematics, vol. 112 (Cambridge University Press, Cambridge, 2008)
393. D.W. Stroock, S.R.S. Varadhan, Multidimensional Diffusion Processes. Classics in Mathe-
matics (Springer, Berlin, 2006). Reprint of the 1997 edition
394. D.W. Stroock, O. Zeitouni, Variations on a theme by Bismut. Astérisque 236, 291–301
(1996). Hommage à P. A. Meyer et J. Neveu
395. K.-T. Sturm, Analysis on local Dirichlet spaces. I. Recurrence, conservativeness and
Lp -Liouville properties. J. Reine Angew. Math. 456, 173–196 (1994)
396. K.-T. Sturm, Analysis on local Dirichlet spaces. II. Upper Gaussian estimates for the funda-
mental solutions of parabolic equations. Osaka J. Math. 32(2), 275–312 (1995)
397. K.-T. Sturm, Diffusion processes and heat kernels on metric spaces. Ann. Probab. 26(1),
1–55 (1998)
398. K.-T. Sturm, The geometric aspect of Dirichlet forms, in New Directions in Dirichlet Forms.
AMS/IP Stud. Adv. Math., vol. 8 (Amer. Math. Soc., Providence, 1998), pp. 233–277
399. K.-T. Sturm, On the geometry of metric measure spaces. I. Acta Math. 196(1), 65–131
(2006)
400. K.-T. Sturm, On the geometry of metric measure spaces. II. Acta Math. 196(1), 133–177
(2006)
401. V.N. Sudakov, B.S. Cirel’son, Extremal properties of half-spaces for spherically invariant
measures. Zap. Nauč. Semin. LOMI 41, 14–24, 165 (1974). Problems in the theory of prob-
ability distributions, II
402. G. Szegő, Orthogonal Polynomials, 4th edn. American Mathematical Society Colloquium
Publications, vol. XXIII (American Mathematical Society, Providence, 1975)
403. M. Talagrand, A new isoperimetric inequality and the concentration of measure phenomenon,
in Geometric Aspects of Functional Analysis (1989–90). Lecture Notes in Math., vol. 1469
404. M. Talagrand, Transportation cost for Gaussian and other product measures. Geom. Funct.
Anal. 6(3), 587–600 (1996)
405. G. Talenti, Osservazioni sopra una classe di disuguaglianze. Rend. Semin. Mat. Fis. Milano
39, 171–185 (1969)
406. G. Talenti, Best constant in Sobolev inequality. Ann. Mat. Pura Appl. (4) 110, 353–372
(1976)
407. H. Tanaka, An inequality for a functional of probability distributions and its application to
Kac’s one-dimensional model of a Maxwellian gas. Z. Wahrscheinlichkeitstheor. Verw. Geb.
27, 47–52 (1973)
408. H. Tanaka, Probabilistic treatment of the Boltzmann equation of Maxwellian molecules.
Z. Wahrscheinlichkeitstheor. Verw. Geb. 46(1), 67–105 (1978/79)
409. A. Terras, Harmonic Analysis on Symmetric Spaces—Euclidean Space, the Sphere, and the
Poincaré Upper Half-Plane (Springer, Berlin, 2013)
410. G. Tomaselli, A class of inequalities. Boll. Unione Mat. Ital. 2, 622–631 (1969)
411. M. Tomisaki, Comparison theorems on Dirichlet norms and their applications. Forum Math.
2(3), 277–295 (1990)
412. G. Toscani, Entropy production and the rate of convergence to equilibrium for the Fokker-
Planck equation. Q. Appl. Math. 57(3), 521–541 (1999)
544 Bibliography
413. H. Triebel, Theory of Function Spaces. Monographs in Mathematics, vol. 78 (Birkhäuser,

Basel, 1983)
414. H. Triebel, Theory of Function Spaces. II. Monographs in Mathematics, vol. 84 (Birkhäuser,
Basel, 1992)
415. H. Triebel, Interpolation Theory, Function Spaces, Differential Operators, 2nd edn. (Johann
Ambrosius Barth, Heidelberg, 1995)
416. N.S. Trudinger, On imbeddings into Orlicz spaces and some applications. J. Math. Mech. 17,
473–483 (1967)
417. A.S. Üstünel, An Introduction to Analysis on Wiener Space. Lecture Notes in Mathematics,
418. J.A. van Casteren, Markov Processes, Feller Semigroups and Evolution Equations. Series on
Concrete and Applicable Mathematics, vol. 12 (World Scientific, Hackensack, 2011)
419. J. van Neerven, The Asymptotic Behaviour of Semigroups of Linear Operators. Operator
Theory: Advances and Applications, vol. 88 (Birkhäuser, Basel, 1996)
420. S.R.S. Varadhan, On the behavior of the fundamental solution of the heat equation with
variable coefficients. Commun. Pure Appl. Math. 20, 431–455 (1967)
421. N.T. Varopoulos, Hardy-Littlewood theory for semigroups. J. Funct. Anal. 63(2), 240–260
(1985)
422. N.T. Varopoulos, L. Saloff-Coste, T. Coulhon, Analysis and Geometry on Groups. Cambridge
Tracts in Mathematics, vol. 100 (Cambridge University Press, Cambridge, 1992)
423. J.L. Vázquez, The Porous Medium Equation. Oxford Mathematical Monographs (The
Clarendon Press, Oxford, 2007). Mathematical theory
424. C. Villani, Topics in Optimal Transportation. Graduate Studies in Mathematics, vol. 58
(American Mathematical Society, Providence, 2003)
425. C. Villani, Hypocoercivity. Mem. Am. Math. Soc. 202, 950 (2009), pp. iv+141
426. C. Villani, Optimal Transport, old and new. Grundlehren der mathematischen Wis-
senschaften [Fundamental Principles of Mathematical Sciences], vol. 338 (Springer, Berlin,
2009)
427. M.-K. von Renesse, K.-T. Sturm, Transport inequalities, gradient estimates, entropy, and
Ricci curvature. Commun. Pure Appl. Math. 58(7), 923–940 (2005)
428. F.-Y. Wang, Logarithmic Sobolev inequalities on noncompact Riemannian manifolds.
Probab. Theory Relat. Fields 109(3), 417–424 (1997)
429. F.-Y. Wang, Functional inequalities for empty essential spectrum. J. Funct. Anal. 170(1),
219–245 (2000)
430. F.-Y. Wang, Functional inequalities and spectrum estimates: the infinite measure case.
J. Funct. Anal. 194(2), 288–310 (2002)
431. F.-Y. Wang, Functional Inequalities, Markov Processes, and Spectral Theory (Science Press,
Beijing, 2004)
432. F.-Y. Wang, Spectral gap for hyperbounded operators. Proc. Am. Math. Soc. 132(9),
2629–2638 (2004)
433. F.-Y. Wang, Dimension-free Harnack inequality and its applications. Front. Math. China 1(1),
53–72 (2006)
434. F.-Y. Wang, Harnack inequalities on manifolds with boundary and applications. J. Math.
Pures Appl. 94(3), 304–321 (2010)
435. F.-Y. Wang, Equivalent semigroup properties for the curvature-dimension condition. Bull.
Sci. Math. 135(6–7), 803–815 (2011)
436. F. Wang, Y. Zhang, F -Sobolev inequality for general symmetric forms. Northeast. Math. J.
19(2), 133–138 (2003)
437. F.B. Weissler, Logarithmic Sobolev inequalities for the heat-diffusion semigroup. Trans. Am.
Math. Soc. 237, 255–269 (1978)
438. F.B. Weissler, Two-point inequalities, the Hermite semigroup, and the Gauss-Weierstrass
semigroup. J. Funct. Anal. 32(1), 102–121 (1979)
439. F.B. Weissler, Logarithmic Sobolev inequalities and hypercontractive estimates on the circle.
J. Funct. Anal. 37(2), 218–234 (1980)
Bibliography 545
440. D. Williams, Probability with Martingales, Cambridge Mathematical Textbooks (Cambridge

University Press, Cambridge, 1991)
441. L. Wu, A new modified logarithmic Sobolev inequality for Poisson point processes and sev-
eral applications. Probab. Theory Relat. Fields 118(3), 427–438 (2000)
442. S.-T. Yau, Isoperimetric constants and the first eigenvalue of a compact Riemannian mani-
fold. Ann. Sci. Éc. Norm. Super. 8(4), 487–507 (1975)
443. S.-T. Yau, On the heat kernel of a complete Riemannian manifold. J. Math. Pures Appl. 57(2),
191–201 (1978)
444. K. Yosida, Functional Analysis, 6th edn. Grundlehren der mathematischen Wissenschaften
[Fundamental Principles of Mathematical Sciences], vol. 123 (Springer, Berlin, 1980)
445. B. Zegarliński, Entropy bounds for Gibbs measures with non-Gaussian tails. J. Funct. Anal.
187(2), 368–395 (2001)
446. B. Zegarliński, Analysis on extended Heisenberg group. Ann. Fac. Sci. Toulouse 20(2),
379–405 (2011)
447. A. Zettl, Sturm-Liouville Theory. Mathematical Surveys and Monographs, vol. 121 (Ameri-
can Mathematical Society, Providence, 2005)
448. S. Zhu, The comparison geometry of Ricci curvature, in Comparison Geometry, Berkeley,
CA, 1993–94. Math. Sci. Res. Inst. Publ., vol. 30 (Cambridge Univ. Press, Cambridge, 1997),
pp. 221–262
449. W.P. Ziemer, Sobolev spaces and functions of bounded variation, in Weakly Differentiable
Functions. Graduate Texts in Mathematics, vol. 120 (Springer, New York, 1989)
450. P.-A. Zitt, Super Poincaré inequalities, Orlicz norms and essential spectrum. Potential Anal.
35(1), 51–66 (2011)
Index
A C
Adjoint operator, 45, 48, 133, 141, 170 Caffarelli’s
contraction Theorem, 447
B regularity properties, 437, 467
Cameron-Martin Hilbert space, 110, 490
Beckner inequality, 312, 384, 388
Capacity, 392, 429
Bessel
Capacity inequality, 394, 397, 400, 401, 404,
operator, 95
407, 411, 429
process, 94 Carré du champ operator , 20, 34, 37, 42, 55,
semigroup, 61, 94 74, 78, 82, 84, 89, 91, 98, 103, 111,
Beta distribution, 114 114, 120, 140, 169, 512
Bobkov inequality, 417, 430 Cattiaux’s inequality, 248, 353
local, 417 Cauchy kernel, 80
reverse, 422 Change of coordinates, 56
Bochner-Lichnerowicz formula, 509, 519 Change of variables formula, 43, 124, 169,
Boundary conditions 171, 493, 516
Dirichlet, 93, 97, 99, 200, 221 Chapman-Kolmogorov equation, 16, 28, 264
Neumann, 93, 97, 100, 199, 200, 273 Chart, 500
periodic, 200, 273 Christoffel symbols, 506
Bounded perturbation Closable operator, 476
-entropy inequality, 383 Co-area formula, 395, 413, 429, 430
logarithmic Sobolev inequality, 240, 274 Co-metric, 47, 82, 504, 505
Poincaré inequality, 185, 231 Commutation property, 144–146, 173, 259
transportation cost inequality, 459, 468 Hamilton-Jacobi, 462, 468
Brascamp-Lieb inequality, 215, 232 Compact operator, 483
Completeness, 141, 157, 172, 510
Brenier’s map, 436, 447, 467
Concentration
Brownian motion, 487
exponential, 426, 431, 467
Euclidean, 78
Gaussian, 426, 431, 441, 460, 467
hyperbolic, 89 Concentration under
killed, 94 logarithmic Sobolev inequality, 255, 274
reflected, 94 Nash inequality, 368
spherical, 85 Poincaré inequality, 192, 231
Brownian semigroup weak Poincaré inequality, 381
Euclidean, 79 Conformal
hyperbolic, 89 invariance, 314, 342
spherical, 81, 85 invariant, 315, 321

DOI 10.1007/978-3-319-00227-9,
548 Index
map, 84, 92 E
Connection, 506 Eigenvalue, 478
Levi-Civita, 507 Eigenvector, 478
Connexity, 33, 140, 156, 171 Ellipticity, 40, 49–51, 511
Continuity property, 11, 54 hypo-, 51
Contraction in Wasserstein space, 463, 468 semi-, 40, 43, 49, 50
Contraction principle uniform, 50
weak hypo-, 141, 157, 172
logarithmic Sobolev inequality, 254
Entropic inequality, 236
Poincaré inequality, 192
Entropy, 236, 273, 348
Contraction semigroup, 10, 18, 473
relative, 237, 244, 438, 445
Convergence to equilibrium, 33 Entropy-energy inequality, 281, 349, 387
Core, 19, 52, 101, 134, 142, 477 Ergodic semigroup, 33
Cost function, 434 Ergodicity, 33, 70, 135, 170
Coupling, 434 ESA property, 134, 142, 157, 170
Covariant derivative, 507 Essentially self-adjoint, 95, 101, 134, 170,
Curvature condition CD(ρ, ∞), 72, 104, 111, 477, 482
144, 209, 259, 416, 443, 457, 462, Explosion time, 495
463, 513 Exponential decay
reinforced, 146, 172, 421, 515 -entropy, 383
Curvature-dimension condition CD(ρ, n), 72, entropy, 244, 274
75, 79, 87, 92, 104, 115, 159, 172, variance, 183, 231
211, 232, 268, 275, 298, 305, 337, Exponential integrability
514 eigenvector, 250, 274
Cut-locus, 510 logarithmic Sobolev inequality, 252, 274
Poincaré inequality, 190, 231
Exponential measure, 188, 231
D Extended algebra, 152
De Bruijn’s identity, 245, 274, 333 Extremal function, 259, 285, 320, 448, 467
Del Pino-Dolbeault Theorem, 326, 342
Density kernel, 14, 25, 286, 356, 371 F
Diameter, 143, 166, 173, 293, 341 F -Sobolev inequality, 388, 430
Differentiable manifold, 500 Faber-Krahn inequality, 183, 284, 399, 429
Diffusion Fast diffusion equation, 331, 342
operator, 43 Feynman-Kac formula, 64
process, 38, 496 Filtration, 488
property, 56, 122 Fisher information, 237, 333, 445
Dirichlet Fokker-Planck equation, 24, 28, 332
domain, 30, 55, 125, 155, 169 Friedrichs extension, 128, 477
form, 30, 55, 125, 169, 477
norm, 125, 169 G
2 operator, 75, 79, 104, 144, 157, 172, 173,
Displacement convexity, 465, 468
275, 512, 519
Distance
Gagliardo-Nirenberg inequality, 324, 342, 451
geodesic, 510 Gamma distribution, 111
harmonic, 360 Gaussian measure, 103, 109, 296
intrinsic, 166, 173 Generalized inverse, 58, 350
Riemannian, 166, 510 Geodesic, 509
total variation, 244, 436, 438 optimal transportation, 465
Wasserstein, 436, 445 Girsanov transformation, 62
Doob’s stopping time Theorem, 491 Good measurable space, 7, 53, 169
Dual semigroup, 9, 464 Gozlan’s Theorem, 460, 468
Duhamel’s formula, 130, 144, 145, 207, 214, Gradient, 506
232 vector field, 47, 65
Index 549
Gradient bound (strong), 105, 144–146, 163, Inversion, 91

173, 209, 259, 298, 416 Isoperimetric
Gross’ hypercontractivity Theorem, 246, 274, function, 413
387 profile, 413
Ground state transformation, 64, 107 Gaussian, 417
Growth function, 225, 348, 375 Isoperimetric inequality, 413, 430
Euclidean, 413
H exponential, 413
h-transform, 66, 95 Gaussian, 411, 417
Hamilton-Jacobi equation, 452, 454, 467 spherical, 413
Harmonic Isoperimetric-type inequality, 412, 426
distance, 360 Gaussian, 416, 426
extension, 81, 86, 116 local, 418
function, 66, 92 Itô’s formula, 491
oscillator, 107
Harnack inequality, 232, 302, 341, 421, 431 J
isoperimetric, 425, 431 Jacobi
Wang, 265, 275, 424 operator, 113
Heat equation, 4, 19, 24, 28, 55, 127, 496, 521 polynomials, 114
Heat flow, 19, 55, 127 semigroup, 115
monotonicity, 130, 144, 207, 214, 257, 261, spectrum, 114
269, 275, 299, 343, 419, 422
Heat kernel bound, 264, 302, 341, 350, 388 K
Heat semigroup, 19, 55, 127 Kannan-Lovász-Simonovits-Bobkov Theorem,
Euclidean, 79 203, 232
hyperbolic, 89 Kantorovich-Rubinstein Theorem, 435, 466
spherical, 81, 85 Kantorovich’s Theorem, 435, 466
Herbst’s argument, 252, 274, 294, 356 Kernel
Hermite polynomials, 105, 179 density, 78, 90, 97, 101, 104, 116, 118
Hessian, 146, 158, 172, 507, 512 representation, 13, 104
Hilbert-Schmidt operator, 29, 290, 355, 483
Hille-Yosida theory, 18, 473, 485 L
Hopf-Lax formula, 452, 467 Laguerre
Hörmander form, 41, 43, 51, 497, 517 operator, 111
HWI inequality, 443, 467 polynomials, 112
Hyperbolic space, 89 semigroup, 113
Hypercontractivity, 246, 273, 389 spectrum, 112
Hamilton-Jacobi, 454, 468 Laplace-Beltrami operator, 46, 81, 146, 508,
immediate, 351 514
local, 262, 275, 342, 457 Laplace-Varadhan asymptotics, 303, 453
reverse, 274 Laplacian, 46, 508
Euclidean, 78
I hyperbolic, 89, 92
Indices spherical, 81, 82, 84, 85, 115
contravariant, 505 Latała-Oleszkiewicz inequality, 385, 388, 410,
covariant, 505 430
Infimum-convolution, 467 Legendre-Fenchel transform, 397, 452
Infinitesimal generator, 18, 55, 169, 475, 496 Lévy-Gromov Theorem, 342, 418, 430
domain, 18, 52, 55, 99, 126, 155, 170, 475 Li-Yau inequality, 301, 341
Integration by parts, 27, 49, 55, 132, 169, 171 Lichnerowicz’ Theorem, 213, 215, 232
Intermediate value Theorem, 136, 173 Linearized form
Invariant measure, 10, 20, 25, 33, 35, 37, 45, entropy-energy, 349
54, 78, 82, 89, 103, 109, 111, 114, Nash, 365, 403
123 weak Poincaré, 376, 407
550 Index
Lipschitz function, 166, 173 Measure decomposition theorem, 7

Local Measure-capacity inequality, 394, 397, 400,
Bobkov inequality, 417 401, 404, 407, 411, 429
hypercontractivity, 262, 275, 304, 342, 457 Median
isoperimetric-type inequality, 418 function, 375
logarithmic Sobolev inequality, 259, 275, measure, 194, 256
299 Metric measure space, 466
Poincaré inequality, 209, 232 Milman’s Theorems, 426, 431
Local inequalities, 206 Minimal extension, 134, 221, 481
Log-concave measure, 203, 425 Monge-Ampère equation, 437, 439, 448
Log-Harnack inequality, 266 Muckenhoupt’s criterion, 231, 274, 429
Logarithmic entropy-energy inequality, 281, Latała-Oleszkiewicz inequality, 411
306 logarithmic Sobolev inequality, 256
Logarithmic Sobolev constant, 237, 259 Poincaré inequality, 194
Logarithmic Sobolev inequality, 236, 273, 401, weak Poincaré inequality, 409
429, 446, 458
defective, 236, 282, 349 N
dimensional, 296 Nash inequality, 281, 403
Euclidean, 284, 328, 341, 354, 456 generalized, 364, 387, 430
Gaussian, 258, 275 weighted, 224, 370, 388
local, 259, 275 Non-explosion, 135, 165
modified, 385, 388 O
reverse, 259 Obata’s Theorem, 214, 232, 313
segment, 273, 275 Onofri’s inequality, 342
under CD(ρ, ∞), 268 Opposite generator, 31
under CD(ρ, n), 270, 275 Optimal transportation, 433, 466
Lyapunov Orlicz
criterion, 202, 231 function, 396
function, 202, 372 space, 389, 396
Ornstein-Uhlenbeck
M operator, 103
Manifold process, 104
differentiable, 500 semigroup, 103, 178
Riemannian, 137, 504 in infinite dimension, 110
Markov spectrum, 107, 111
chain, 33 Orthogonal polynomials, 102
generator, 18 Oscillation, 185, 240, 376, 459
kernel, 12, 54 Otto-Villani Theorem, 458, 468
operator, 10, 15
process, 8, 17, 37, 54 P
semigroup, 12, 54, 128, 170 -entropy inequality, 383, 388
symmetric, 25, 55 Perron-Frobenius eigenvector, 36, 67, 222
Markov property (strong), 8, 495 Persson operator, 227, 233
Markov Triple, 28, 120, 168, 177, 235, 277, Pinsker-Csizsár-Kullback inequality, 244, 274,
347, 392 438
Compact, 167, 171 Poincaré constant, 181, 190, 200
Diffusion, 132, 169 Poincaré inequality, 181, 365, 377, 400, 429,
Full, 167, 171, 177, 235, 277, 347, 392 458
Standard, 136, 170, 177, 235, 277, 347, 392 ball, 204
Martingale, 488 convex set, 204
local, 491 exponential measure, 189
problem, 22, 39, 496 Gaussian, 180, 231
Mass conservation, 10, 54, 135, 150, 165, 170, hyperbolic space, 222, 233
173 local, 209, 232
Index 551
log-concave measure, 203 Semigroup interpolation, 130, 144, 207, 214,

segment, 199, 231 232, 257, 261, 269, 275, 299, 419,
Sturm-Liouville operator, 198 422
super, 388 Semigroup property, 10, 54
under CD(ρ, ∞), 212 Slicing, 137, 241, 283, 341, 395
under CD(ρ, n), 215, 232 Smooth, 501
weak, 375, 388, 430 function, 138, 167
Poisson formula, 86, 116 Sobolev constant, 280, 318
Polish space, 7, 53, 412, 434 Sobolev inequality, 279, 340, 397, 429
Porous medium equation, 331, 342 Euclidean, 278, 318, 341, 448, 467
Positive operator, 477 hyperbolic, 279, 319, 341
Positivity preserving, 10, 54, 127 spherical, 278, 318, 341
Potential, 63 trace, 449, 467
under CD(ρ, n), 308, 342
R Spectral
Rellich-Kondrachov Theorem, 229, 291, 341 decomposition, 31, 33, 131, 478, 479
Representation gap, 182
open ball, 91 Spectrum, 32, 221, 478
orthogonal projection, 81 continuous, 478
stereographic projection, 83, 327 discrete, 224, 228, 290, 355, 480
upper half-space, 89 essential, 223, 228, 480, 485
Reproducing kernel Hilbert space, 110, 490 point, 478
Resolvent residual, 478
operator, 68, 289, 312, 474 Stability under product
set, 478 -entropy inequality, 383
Reversible measure, 25, 35, 37, 46, 55, 78, 82, logarithmic Sobolev inequality, 249, 274
89, 103, 111, 114, 123 Poincaré inequality, 185, 231
Ricci quadratic transportation cost inequality,
curvature, 72, 73, 82, 144, 146, 314, 317, 441
464, 508, 513 Sobolev inequality, 292
tensor, 71, 82, 508, 512 State space, 7, 53
Riemann curvature tensor, 507 Stationary measure, 10, 11, 123
Riemannian Stochastic
distance, 166, 510 differential equation, 38, 493
measure, 46, 509 integral, 488
metric, 504 Stochastic completeness, 135, 165
Riemannian manifold, 137, 504 Stopping time, 490
weighted, 47, 73, 137, 213, 269, 412, 433, Stratonovich integral, 41, 496
511 Sturm-Liouville operator, 97, 198, 199, 273
Riesz potential, 69 Sub-Markov semigroup, 12, 63
Riesz-Thorin Theorem, 287, 289, 353 Subordination, subordinator, 67, 86
Rota’s Lemma, 26, 75 Super Poincaré inequality, 388
Rothaus’ Lemma, 239, 274 Surface measure, 412
Symmetric operator, 476
S
Scalar curvature, 314, 508 T
Second order differential operator, 40, 42, 139 Talagrand’s inequality, 438, 440, 444, 467
Self-adjoint operator, 476 Tangent vector, 501
Semi-martingale, 492 Tensor, 502
Semigroup, 18, 473 Tight
Markov, 12, 54, 128, 170 inequality, 182
product, 59 logarithmic Sobolev inequality, 236
quotient, 60 Sobolev inequality, 280
552 Index
Tightening, 238, 274, 280, 388 W

Time change, 58 Wang’s Harnack inequality, 265, 275, 424
Total variation distance, 244, 436, 438 Wasserstein
Trace formula, 32 contraction, 463, 468
Transportation cost inequality, 466 distance, 436, 445, 465
quadratic T2 , 438, 440, 458, 467 space, 436, 463, 465, 466
T1 , 438 Weak hypo-ellipticity, 141, 157, 172
Weak Poincaré inequality, 375, 388, 407, 430
U
Weight function, 370
Ultracontractivity, 286, 341, 353, 367, 456
Weyl’s criterion, 223, 481
Uniqueness extension, 134
Wiener
V integral, 110, 490
Vanishing viscosity, 453, 455 measure, 109, 488
Variance, 181 Wirtinger’s inequality, 200, 231, 232
Variational formula of entropy, 236
Vector field, 41, 62, 501 Z
gradient, 47, 65 Zonal function, 115

Untitled

Uploaded by

Copyright:

Available Formats

Untitled

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Untitled

Uploaded by

Copyright:

Available Formats

Grundlehren der

mathematischen Wissenschaften 348

M. Berger P. de la Harpe N.J. Hitchin

ISSN 0072-7830 ISSN 2196-9701 (electronic)

Library of Congress Control Number: 2013952461

Mathematics Subject Classification: 39B62, 39B72, 47D07, 53C21

© Springer International Publishing Switzerland 2014

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

From the probabilistic viewpoint, the family of kernels pt (x), t > 0, x ∈ Rn ,

for all bounded measurable functions f : Rn → R.

Diophantine equation x 2 + y 2 + z2 = 3xyz!

where F = log Pt−s f , G = log Pt−s g and H = θ F + (1 − θ )G. Now by standard

or sometimes as E f (x)μ(dx). The Lebesgue measure on the Borel sets of Rn is

Part I Markov Semigroups, Basics and Examples

3 Symmetric Markov Diffusion Operators . . . . . . . . . . . . . . . 119

Part III Related Functional, Isoperimetric and Transportation

Appendix B Elements of Stochastic Calculus . . . . . . . . . . . . . . 487

This chapter is introductory and descriptive. It aims to introduce, somewhat infor-

D. Bakry et al., Analysis and Geometry of Markov Diffusion Operators, 3

tion being given by the conditional expectation representation

for t ≥ 0, x ∈ E and f : E → R a suitable measurable function.

where is the standard Laplacian on Rn .

where pt (x, y) = pt (x − y), (x, y) ∈ Rn × Rn . This definition applies for example

In other words, u(x, t) = Pt f (x), t ≥ 0, x ∈ Rn , solves the partial differential (heat)

Bt1 , Bt2 − Bt1 , . . . , Btk − Btk−1 , 0 ≤ t1 < · · · < t k ,

are independent centered Gaussian vectors in Rn with respective covariance matri-

The process (x + Bt )t≥0 is (by definition) a Brownian motion starting from x ∈ Rn .

gradient operator ∇ and the semigroup Pt (on smooth functions f : Rn → R). By

In fact, as will be developed extensively, this commutation inequality corresponds

1.1 Markov Processes and Associated Semigroups

In other words, Pt+s = Pt ◦ Ps , t, s ≥ 0, which is the fundamental property of

1.2 Markov Semigroups, Invariant Measures and Kernels

1.2.1 Markov Semigroups, Invariant Measures

The fundamental object of investigation consists of a family P = (Pt )t≥0 of op-

(ii) P0 = Id, the identity operator (initial condition).

In particular, for the functions φ(r) = |r|p , r ∈ R, 1 ≤ p ≤ ∞, Pt (|f |p ) ≥ |Pt f |p .

Definition 1.2.1 (Invariant measure) Given a family P = (Pt )t≥0 on a measurable

Note that, by (1.2.1), if a measure μ is invariant for P, then, for every t ≥ 0, Pt is a

on Rn for which the invariant measure is given (up to multiplication by a constant)

This property justifies the terminology of stationary measure for μ in a probabilistic

limt→0 Pt f (x) = f (x) and limt→0 Pt f = f in L2 (μ) (a property extended then to

Definition 1.2.2 (Markov semigroup) A family of operators P = (Pt )t≥0 de-

Accordingly, the Markov semigroup P = (Pt )t≥0 defines a semigroup of contrac-

Markov operators Pt , t ≥ 0, as given in Definition 1.2.2, may be represented by

Proposition 1.2.3 (Kernel representation) Given a good measure space (E, F, μ)

where 1A (respectively 1B ) is the characteristic function of the set A (respectively

for every A, B ∈ F . Since P (1) = 1, the first marginal of ν1 is μ and ν1 may be

ν1 (dx, dy) = p(x, dy)μ(dx).

Definition 1.2.4 (Density kernel) A Markov semigroup P = (Pt )t≥0 on (E, F) is

In order for Pt f to make sense for any f ∈ L2 (m) in this definition, it

in general required that for all t > 0 and (m-almost) every x ∈ E,

Proposition 1.2.5 (Existence of density kernel) Let (E, F, m) be a good mea-

for some bounded (by M) measurable functions qj , j = 1, . . . , j . Restricted to

As a further general fact, it is worth mentioning that as soon as a Markov operator

Proposition 1.2.6 Let P be a Markov operator (thus satisfying properties (iii)

Proof Assume by homogeneity that E f dμ = 1. Setting g = Pf , g ≤ 1 + ε and

E gdμ = 1 since μ is invariant for P . For any 0 < a < 1 + ε, P f = P (g) ≥

P (1{g≥a} ) = 1 − P (1{g<a} ) ≥ 1 − (1 + ε)μ(g < a)

and by Markov’s inequality,

for some bounded (by M) measurable functions qj , j = 1, . . . , j . Restricted to

(f, g)2 ≤ (f ) (g), (f, g) ∈ A × A. (1.4.3)

while if f, g belong to the class A on which is defined,

As in the finite case, for any x ∈