Progress in Mathematics
314
Veronique Fischer
Michael Ruzhansky
Quantization
on Nilpotent
Lie Groups
Ferran Sunyer i Balaguer
Award winning monograph
Progress in Mathematics
Volume 314
Series Editors
Hyman Bass, University of Michigan, Ann Arbor, USA
JiangHua Lu, The University of Hong Kong, Hong Kong SAR, China
Joseph Oesterl, Universit Pierre et Marie Curie, Paris, France
Yuri Tschinkel, Courant Institute of Mathematical Sciences, New York, USA
More information about this series at http://www.springer.com/series/4848
Veronique Fischer Michael Ruzhansky
Quantization on Nilpotent
Lie Groups
Veronique Fischer
Department of Mathematics
University of Bath
Bath, UK
Michael Ruzhansky
Department of Mathematics
Imperial College London
London, UK
ISSN 07431643
ISSN 2296505X (electronic)
Progress in Mathematics
ISBN 9783319295572
ISBN 9783319295589 (eBook)
DOI 10.1007/9783319295589
Library of Congress Control Number: 2016932499
Mathematics Subject Classification (2010): 22C05, 22E25, 35A17, 35H10, 35K08, 35R03, 35S05, 43A15,
43A22, 43A77, 43A80, 46E35, 46L10, 47G30, 47L80
The Editor(s) (if applicable) and The Author(s) 2016. This book is published open access.
Open Access This book is distributed under the terms of the Creative Commons Attribution 4.0 International
License (http://creativecommons.org/licenses/by/4.0/), which permits use, duplication, adaptation, distribution
and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and
the source, a link is provided to the Creative Commons license and any changes made are indicated.
The images or other third party material in this book are included in the works Creative Commons license,
unless indicated otherwise in the credit line; if such material is not included in the works Creative Commons
license and the respective action is not permitted by statutory regulation, users will need to obtain permission
from the license holder to duplicate, adapt or reproduce the material.
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.
The publisher, the authors and the editors are safe to assume that the advice and information in this book are
believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors
give a warranty, express or implied, with respect to the material contained herein or for any errors or
omissions that may have been made.
Printed on acidfree paper
This book is published under the trade name Birkhuser.
The registered company is Springer International Publishing AG Switzerland (www.birkhauserscience.com)
Ferran Sunyer i Balaguer (19121967) was a selftaught Catalan mathematician who, in spite of a
serious physical disability, was very active in research
in classical mathematical analysis, an area in which
he acquired international recognition. His heirs created the Fundaci
o Ferran Sunyer i Balaguer inside
the Institut dEstudis Catalans to honor the memory
of Ferran Sunyer i Balaguer and to promote mathematical research.
Each year, the Fundacio Ferran Sunyer i Balaguer
and the Institut dEstudis Catalans award an international research prize for a mathematical monograph of expository nature. The prizewinning monographs are published in this series. Details about the
prize and the Fundacio Ferran Sunyer i Balaguer can
be found at
http://ffsb.espais.iec.cat/EN
This book has been awarded the
Ferran Sunyer i Balaguer 2014 prize.
The members of the scientic commitee
of the 2014 prize were:
Alejandro Adem
University of British Columbia
Hyman Bass
University of Michigan
N
uria Fagella
Universitat de Barcelona
Eero Saksman
University of Helsinki
Yuri Tschinkel
Courant Institute of Mathematical Sciences,
New York University
Ferran Sunyer i Balaguer Prize winners since 2004:
2004
Guy David
Singular Sets of Minimizers for the
MumfordShah Functional, PM 233
2005
Antonio Ambrosetti and Andrea Malchiodi
Perturbation Methods and Semilinear
Elliptic Problems on Rn , PM 240
Jose Seade
On the Topology of Isolated Singularities in
Analytic Spaces, PM 241
2006
Xiaonan Ma and George Marinescu
Holomorphic Morse Inequalities and
Bergman Kernels, PM 254
2007
Rosa MiroRoig
Determinantal Ideals, PM 264
2008
Luis Barreira
Dimension and Recurrence in Hyperbolic
Dynamics, PM 272
2009
Timothy D. Browning
Quantitative Arithmetic of Projective Varieties, PM 277
2010
Carlo Mantegazza
Lecture Notes on Mean Curvature Flow,
PM 290
2011
Jayce Getz and Mark Goresky
Hilbert Modular Forms with Coecients in
Intersection Homology and Quadratic Base
Change, PM 298
2012
Angel Cano, Juan Pablo Navarrete and
Jose Seade
Complex Kleinian Groups, PM 303
2013
Xavier Tolsa
Analytic capacity, the Cauchy transform,
and nonhomogeneous Calder
onZygmund
theory, PM 307
Preface
The purpose of this monograph is to give an exposition of the global quantization
of operators on nilpotent homogeneous Lie groups. We also present the background
analysis on homogeneous and graded nilpotent Lie groups. The analysis on homogeneous nilpotent Lie groups drew a considerable attention from the 70s onwards.
Research went in several directions, most notably in harmonic analysis and in the
study of hypoellipticity and solvability of partial dierential equations. Over the
decades the subject has been developing on dierent levels with advances in the
analysis on the Heisenberg group, stratied Lie groups, graded Lie groups, and
general homogeneous Lie groups.
In the last years analysis on homogeneous Lie groups and also on other types
of Lie groups has received another boost with newly found applications and further
advances in many topics. Examples of this boost are subelliptic estimates, multiplier theorems, index formulae, nonlinear problems, potential theory, and symbolic
calculi tracing full symbols of operators. In particular, the latter has produced further applications in the study of linear and nonlinear partial dierential equations,
requiring the knowledge of lower order terms of the operators.
Because of the current advances, it seems to us that a systematic exposition of
the recently developed quantizations on Lie groups is now desirable. This requires
bringing together various parts of the theory in the right generality, and extending
notions and techniques known in particular cases, for instance on compact Lie
groups or on the Heisenberg group.
In order to do so, we start with a review of the recent developments in
the global quantization on compact Lie groups. In this, we follow mostly the
development of this subject in the monograph [RT10a] by Turunen and the second
author, as well as its further progress in subsequent papers. After a necessary
exposition of the background analysis on graded and homogeneous Lie groups, we
present the quantization on general graded Lie groups. As the nal part of the
monograph, we work out details of the general theory developed in this book in
the particular case of the Heisenberg group.
In the introduction, we will provide a link between, on one hand, the symbolic
calculus of matrix valued symbols on compact Lie groups with, on the other hand,
vii
viii
Preface
dierent approaches to the symbolic calculus on the Heisenberg group for instance.
We will also motivate further our choices of presentation from the point of view
of the development of the theory and of its applications.
We would like to thank Fulvio Ricci for discussions and for useful comments
on the historical overview of parts of the subject that we tried to present in the
introduction. We would also like to thank Gerald Folland for comments leading to
improvements of some parts of the monograph.
Finally, it is our pleasure to acknowledge the nancial support by EPSRC
(grant EP/K039407/1), Marie Curie FP7 (Project PseudodiOperatorS  301599),
and by the Leverhulme Trust (grant RPG201402) at dierent stages of preparing
this monograph.
Veronique Fischer
Michael Ruzhansky
London, 2015
Open Access. This chapter is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/),
which permits use, duplication, adaptation, distribution and reproduction in any medium
or format, as long as you give appropriate credit to the original author(s) and the source,
a link is provided to the Creative Commons license and any changes made are indicated.
The images or other third party material in this chapter are included in the works
Creative Commons license, unless indicated otherwise in the credit line; if such material
is not included in the works Creative Commons license and the respective action is not
permitted by statutory regulation, users will need to obtain permission from the license
holder to duplicate, adapt or reproduce the material.
Contents
Preface
vii
Introduction
Notation and conventions
1
13
Preliminaries on Lie groups
1.1 Lie groups, representations, and Fourier transform . . . . . .
1.2 Lie algebras and vector elds . . . . . . . . . . . . . . . . . .
1.3 Universal enveloping algebra and dierential operators . . . .
1.4 Distributions and Schwartz kernel theorem . . . . . . . . . .
1.5 Convolutions . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.6 Nilpotent Lie groups and algebras . . . . . . . . . . . . . . .
1.7 Smooth vectors and innitesimal representations . . . . . . .
1.8 Plancherel theorem . . . . . . . . . . . . . . . . . . . . . . . .
1.8.1 Orbit method . . . . . . . . . . . . . . . . . . . . . . .
1.8.2 Plancherel theorem and group von Neumann algebras
1.8.3 Fields of operators acting on smooth vectors . . . . .
2 Quantization on compact Lie groups
2.1 Fourier analysis on compact Lie groups . . . . . . . .
2.1.1 Characters and tensor products . . . . . . . .
2.1.2 PeterWeyl theorem . . . . . . . . . . . . . .
2.1.3 Spaces of functions and distributions on G . .
. . . . . . . .
2.1.4 p spaces on the unitary dual G
2.2 Pseudodierential operators on compact Lie groups
2.2.1 Symbols and quantization . . . . . . . . . . .
2.2.2 Dierence operators and symbol classes . . .
2.2.3 Symbolic calculus, ellipticity, hypoellipticity .
2.2.4 Fourier multipliers and Lp boundedness . . .
2.2.5 Sharp G
arding inequality . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
15
15
22
24
29
30
34
37
43
44
46
52
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
57
58
58
60
63
67
71
71
74
78
82
88
ix
x
3
Contents
Homogeneous Lie groups
3.1 Graded and homogeneous Lie groups . . . . . . . . . . . . . . . . .
3.1.1 Denition and examples of graded Lie groups . . . . . . . .
3.1.2 Denition and examples of homogeneous Lie groups . . . .
3.1.3 Homogeneous structure . . . . . . . . . . . . . . . . . . . .
3.1.4 Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.5 Invariant dierential operators on homogeneous Lie groups
3.1.6 Homogeneous quasinorms . . . . . . . . . . . . . . . . . . .
3.1.7 Polar coordinates . . . . . . . . . . . . . . . . . . . . . . . .
3.1.8 Mean value theorem and Taylor expansion . . . . . . . . . .
3.1.9 Schwartz space and tempered distributions . . . . . . . . .
3.1.10 Approximation of the identity . . . . . . . . . . . . . . . . .
3.2 Operators on homogeneous Lie groups . . . . . . . . . . . . . . . .
3.2.1 Leftinvariant operators on homogeneous Lie groups . . . .
3.2.2 Leftinvariant homogeneous operators . . . . . . . . . . . .
3.2.3 Singular integral operators on homogeneous Lie groups . . .
3.2.4 Principal value distribution . . . . . . . . . . . . . . . . . .
3.2.5 Operators of type = 0 . . . . . . . . . . . . . . . . . . . .
3.2.6 Properties of kernels of type , Re [0, Q) . . . . . . . . .
3.2.7 Fundamental solutions of homogeneous dierential operators
3.2.8 Liouvilles theorem on homogeneous Lie groups . . . . . . .
4 Rockland operators and Sobolev spaces
4.1 Rockland operators . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.1 Denition of Rockland operators . . . . . . . . . . . . .
4.1.2 Examples of Rockland operators . . . . . . . . . . . . .
4.1.3 Hypoellipticity and functional calculus . . . . . . . . . .
4.2 Positive Rockland operators . . . . . . . . . . . . . . . . . . . .
4.2.1 First properties . . . . . . . . . . . . . . . . . . . . . . .
4.2.2 The heat semigroup and the heat kernel . . . . . . . .
4.2.3 Proof of the heat kernel theorem and its corollaries . . .
4.3 Fractional powers of positive Rockland operators . . . . . . . .
4.3.1 Positive Rockland operators on Lp . . . . . . . . . . . .
4.3.2 Fractional powers of operators Rp . . . . . . . . . . . .
4.3.3 Imaginary powers of Rp and I + Rp . . . . . . . . . . .
4.3.4 Riesz and Bessel potentials . . . . . . . . . . . . . . . .
4.4 Sobolev spaces on graded Lie groups . . . . . . . . . . . . . . .
4.4.1 (Inhomogeneous) Sobolev spaces . . . . . . . . . . . . .
4.4.2 Interpolation between inhomogeneous Sobolev spaces . .
4.4.3 Homogeneous Sobolev spaces . . . . . . . . . . . . . . .
4.4.4 Operators acting on Sobolev spaces . . . . . . . . . . . .
4.4.5 Independence in Rockland operators and integer orders
4.4.6 Sobolev embeddings . . . . . . . . . . . . . . . . . . . .
4.4.7 List of properties for the Sobolev spaces . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
91
92
92
94
100
103
105
109
116
119
125
129
132
132
136
140
146
151
154
158
167
171
172
172
174
177
183
183
185
187
198
198
203
206
211
218
218
225
228
233
236
239
245
Contents
4.5
4.4.8 Right invariant Rockland operators
Hulanickis theorem . . . . . . . . . . . .
4.5.1 Statement . . . . . . . . . . . . . .
4.5.2 Proof of Hulanickis theorem . . .
4.5.3 Proof of Corollary 4.5.2 . . . . . .
xi
and Sobolev spaces
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
249
251
251
252
269
5 Quantization on graded Lie groups
271
5.1 Symbols and quantization . . . . . . . . . . . . . . . . . . . . . . . 272
5.1.1 Fourier transform on Sobolev spaces . . . . . . . . . . . . . 273
5.1.2 The spaces Ka,b (G), LL (L2a (G), L2b (G)), and L
a,b (G) . . . . 285
5.1.3 Symbols and associated kernels . . . . . . . . . . . . . . . . 294
5.1.4 Quantization formula . . . . . . . . . . . . . . . . . . . . . 296
m
5.2 Symbol classes S,
and operator classes m
, . . . . . . . . . . . . 300
5.2.1 Dierence operators . . . . . . . . . . . . . . . . . . . . . . 300
m
5.2.2 Symbol classes S,
. . . . . . . . . . . . . . . . . . . . . . . 306
5.2.3 Operator classes m
. . . . . . . . . . . . . . . . . . . . . 309
,
5.2.4 First examples . . . . . . . . . . . . . . . . . . . . . . . . . 313
5.2.5 First properties of symbol classes . . . . . . . . . . . . . . . 316
5.3 Spectral multipliers in positive Rockland operators . . . . . . . . . 319
5.3.1 Multipliers in one positive Rockland operator . . . . . . . . 319
5.3.2 Joint multipliers . . . . . . . . . . . . . . . . . . . . . . . . 327
5.4 Kernels of pseudodierential operators . . . . . . . . . . . . . . . . 330
5.4.1 Estimates of the kernels . . . . . . . . . . . . . . . . . . . . 330
5.4.2 Smoothing operators and symbols . . . . . . . . . . . . . . 339
5.4.3 Pseudodierential operators as limits of smoothing operators341
5.4.4 Operators in 0 as singular integral operators . . . . . . . . 345
5.5 Symbolic calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
5.5.1 Asymptotic sums of symbols . . . . . . . . . . . . . . . . . 351
5.5.2 Composition of pseudodierential operators . . . . . . . . . 353
5.5.3 Adjoint of a pseudodierential operator . . . . . . . . . . . 364
m
5.5.4 Simplication of the denition of S,
. . . . . . . . . . . . 371
5.6 Amplitudes and amplitude operators . . . . . . . . . . . . . . . . . 374
5.6.1 Denition and quantization . . . . . . . . . . . . . . . . . . 374
5.6.2 Amplitude classes . . . . . . . . . . . . . . . . . . . . . . . 379
5.6.3 Properties of amplitude classes and kernels . . . . . . . . . 381
5.6.4 Link between symbols and amplitudes . . . . . . . . . . . . 384
5.7 Calder
onVaillancourt theorem . . . . . . . . . . . . . . . . . . . . 385
5.7.1 Analogue of the decomposition into unit cubes . . . . . . . 387
0
5.7.2 Proof of the case S0,0
. . . . . . . . . . . . . . . . . . . . . 389
5.7.3 A bilinear estimate . . . . . . . . . . . . . . . . . . . . . . . 396
0
5.7.4 Proof of the case S,
. . . . . . . . . . . . . . . . . . . . . 403
5.8 Parametrices, ellipticity and hypoellipticity . . . . . . . . . . . . . 409
5.8.1 Ellipticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 410
5.8.2 Parametrix . . . . . . . . . . . . . . . . . . . . . . . . . . . 417
xii
Contents
5.8.3
Subelliptic estimates and hypoellipticity . . . . . . . . . . . 423
Pseudodierential operators on the Heisenberg group
6.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.1.1 Descriptions of the Heisenberg group . . . . . . . . . . .
6.1.2 Heisenberg Lie algebra and the stratied structure . . .
6.2 Dual of the Heisenberg group . . . . . . . . . . . . . . . . . . .
6.2.1 Schr
odinger representations . . . . . . . . . . . . . .
6.2.2 Group Fourier transform on the Heisenberg group . . .
6.2.3 Plancherel measure . . . . . . . . . . . . . . . . . . . . .
6.3 Dierence operators . . . . . . . . . . . . . . . . . . . . . . . .
6.3.1 Dierence operators xj and yj . . . . . . . . . . . . .
6.3.2 Dierence operator t . . . . . . . . . . . . . . . . . . .
6.3.3 Formulae . . . . . . . . . . . . . . . . . . . . . . . . . .
6.4 Shubin classes . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.4.1 WeylHormander calculus . . . . . . . . . . . . . . . . .
n
6.4.2 Shubin classes m
(R ) and the harmonic oscillator . . .
6.4.3 Shubin Sobolev spaces . . . . . . . . . . . . . . . . . . .
n
6.4.4 The Shubin classes m
, (R ) . . . . . . . . . . . . . .
6.4.5 Commutator characterisation of Shubin classes . . . .
m
6.5 Quantization and symbol classes S,
on the Heisenberg group
6.5.1 Quantization on the Heisenberg group . . . . . . . . . .
m
m
= S,
(Hn ) .
6.5.2 An equivalent family of seminorms on S,
m
6.5.3 Characterisation of S, (Hn ) . . . . . . . . . . . . . . .
6.6 Parametrices . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.6.1 Condition for ellipticity . . . . . . . . . . . . . . . . . .
6.6.2 Condition for hypoellipticity . . . . . . . . . . . . . . .
6.6.3 Subelliptic estimates and hypoellipticity . . . . . . . . .
A Miscellaneous
A.1 General properties of hypoelliptic operators . .
A.2 Semigroups of operators . . . . . . . . . . . . .
A.3 Fractional powers of operators . . . . . . . . . .
A.4 Singular integrals (according to CoifmanWeiss)
A.5 Almost orthogonality . . . . . . . . . . . . . . .
A.6 Interpolation of analytic families of operators .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
427
428
428
430
431
432
433
441
443
443
447
453
455
455
459
461
468
473
475
475
477
478
481
481
483
486
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
491
491
493
495
499
503
506
B Group C and von Neumann algebras
B.1 Direct integral of Hilbert spaces . . . . . . . . . . . . . . .
B.1.1 Convention: Hilbert spaces are assumed separable
B.1.2 Measurable elds of vectors . . . . . . . . . . . . .
B.1.3 Direct integral of tensor products of Hilbert spaces
B.1.4 Separability of a direct integral of Hilbert spaces .
B.1.5 Measurable elds of operators . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
509
509
509
510
511
514
515
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Contents
B.1.6 Integral of representations . . . . . . . . . . . . . . . . . . .
B.2 C  and von Neumann algebras . . . . . . . . . . . . . . . . . . . .
B.2.1 Generalities on algebras . . . . . . . . . . . . . . . . . . . .
B.2.2 C algebras . . . . . . . . . . . . . . . . . . . . . . . . . . .
B.2.3 Group C algebras . . . . . . . . . . . . . . . . . . . . . . .
B.2.4 Von Neumann algebras . . . . . . . . . . . . . . . . . . . .
B.2.5 Group von Neumann algebra . . . . . . . . . . . . . . . . .
B.2.6 Decomposition of group von Neumann algebras and abstract
Plancherel theorem . . . . . . . . . . . . . . . . . . . . . . .
xiii
516
517
517
520
521
524
525
527
Schr
odinger representations and Weyl quantization
531
Explicit symbolic calculus on the Heisenberg group
532
List of quantizations
533
Bibliography
535
Index
553
Introduction
Nilpotent Lie groups appear naturally in the analysis of manifolds and provide an
abstract setting for many notions of Euclidean analysis. As is generally the case
when studying analysis on nilpotent Lie groups, we restrict ourselves to the very
large subclass of homogeneous (nilpotent) Lie groups, that is, Lie groups equipped
with a family of dilations compatible with the group structure. They are the groups
appearing in practice in the applications (some of them are described below).
From the point of view of general harmonic analysis, working in this setting also
leads to the distillation of the results of the Euclidean harmonic analysis depending
only on the group and dilation structures.
In order to motivate the work presented in this monograph, we focus our
attention in this introduction on three aspects of the analysis on nilpotent Lie
groups: the use of nilpotent Lie groups as local models for manifolds, questions
regarding hypoellipticity of dierential operators, and the development of pseudodierential operators in this setting. We only outline the historical developments
of ideas and results related to these topics, and on a number of occasions we refer
to other sources for more complete descriptions. We end this introduction with the
main topic of this monograph: the development of a pseudodierential calculus
on homogeneous Lie groups.
Nilpotent Lie groups by themselves and as local models
It has been realised for a long time that the analysis on nilpotent Lie groups can
be eectively used to prove subelliptic estimates for operators such as sums of
squares of vector elds on manifolds. Such ideas started coming to light in the
works on the construction of parametrices for the KohnLaplacian b (the Laplacian associated to the tangential CR complex on the boundary X of a strictly
pseudoconvex domain), which was shown earlier by J. J. Kohn to be hypoelliptic
(see e.g. an exposition by Kohn [Koh73] on the analytic and smooth hypoellipticities). Thus, the corresponding parametrices and subsequent subelliptic estimates
have been obtained by Folland and Stein in [FS74] by rst establishing a version
of the results for a family of subLaplacians on the Heisenberg group, and then
for the KohnLaplacian b by replacing X locally by the Heisenberg group. These
ideas soon led to powerful generalisations. The general techniques for approximat1
Introduction
ing vector elds on a manifold by leftinvariant operators on a nilpotent Lie group
have been developed by Rothschild and Stein in [RS76]. Here the dimension of
the nilpotent Lie group is normally larger than that of the manifold, and a rst
step of such a construction is to perform the lifting of vector elds to the group.
Consequently, this approach allowed one to produce parametrices for the original
dierential operator on the manifold by using the analysis on homogeneous Lie
groups. A more geometric version of these constructions has been carried out by
Folland in [Fol77b], see also Goodman [Goo76] for the presentation of nilpotent
Lie algebras as tangent spaces (of subRiemannian manifolds). The functional analytic background for the analysis in the stratied setting was laid down by Folland
in [Fol75]. A general approach to studying geometries appearing from systems of
vector elds has been developed by Nigel, Stein and Wainger [NSW85].
Thus, one of the motivations for carrying out the analysis and the calculus of
operators on nilpotent Lie groups comes from the study of dierential operators on
CR (CauchyRiemann) or contact manifolds, modelling locally the operators there
on homogeneous invariant convolution operators on nilpotent groups. In practice
and from this motivation, only nilpotent Lie groups endowed with some compatible
structure of dilations, i.e. homogeneous Lie groups, are considered. This will be
also the setting of our present exposition.
The simplest example (apart from Rn ) of a nilpotent Lie group is the Heisenberg group, and the harmonic analysis there is a very well researched topic. We do
not intend to make an overview of the subject here, but we refer to the books of
Stein [Ste93] and Thangavelu [Tha98] for an introduction to the harmonic analysis
on the Heisenberg group and for the historic development of the area. Elements
of the harmonic analysis on dierent groups can be also found in Taylors book
[Tay86]. The Heisenberg group enters many applied areas, including various aspects of quantum mechanics, signal analysis, optics, thermodynamics; we refer to
the recent book of Binz and Pods [BP08] for an overview of this subject. We mention another recent book by Calin, Chang and Greiner [CCG07] containing many
explicit calculations related to the Heisenberg group and its subRiemannian geometry, as well as a subRiemannian treatment in Capogna, Danielli, Pauls and
Tyson [CDPT07]. As such, in this monograph we will deal with the Heisenberg
group almost exclusively in the context of pseudodierential operators, and we
refer to excellent surveys of Folland [Fol77a] and Howe [How80] on the role played
by the Heisenberg group in the theory of partial dierential equations and in harmonic analysis, as well as to Follands book [Fol89] for its relation to the theory
of pseudodierential operators on Rn through the Weyl quantization. See also a
more recent short survey by Semmes [Sem03] and a book by Krantz [Kra09].
Wellposedness questions for hyperbolic partial dierential equations on the
Heisenberg group have been considered parallel to their Euclidean counterparts.
For example, the conditions for the wellposedness of the wave equation for the
Laplacian associated to the b complex have been found by Nachman [Nac82],
the Lp estimates for the wave equation for the subLaplacian have been established by M
uller and Stein [MS99], the smoothness of the Schrodinger kernel has
Introduction
been analysed by Sikora and Zienkiewicz [SZ02], a spacetime estimate for the
Schrodinger equation has been obtained by Zienkiewicz [Zie04], etc. Nonlinear
wave and Schr
odinger equations and Strichartz estimates have been analysed on
the Heisenberg group as well, see e.g. Zuily [Zui93], Bahouri, Gerard and Xu
[BGX00] and Furioli, Melzi and Veneruso [FMV07], as well as other equations,
e.g. the GinzburgLandau equation by Birindelli and Valdinoci [BV08], quasilinear equations by Capogna [Cap99], etc.
The Hardy spaces on homogeneous Lie groups and the surrounding harmonic
analysis have been investigated by Folland and Stein in their monograph [FS82]. In
general, there are dierent machineries available depending on a degree of generality: the stratied Lie groups enjoy additional hypoellipticity techniques going back
to H
ormanders celebrated sum of the squares theorem, while on the Heisenberg
group explicit expressions from its representation theory can be used.
A typical example of such dierent degrees of generality within homogeneous
Lie groups is, for instance, a problem of characterising the Hardy space H 1 in
L1 by families of singular integrals. Thus, in [CG84], Christ and Geller presented
sucient conditions for general homogeneous Lie groups, gave explicit examples of
(generalised) Riesz transforms for such a family of integral operators on stratied
Lie groups, and derived further necessary and sucient conditions on the Heisenberg group in terms of its representation theory (see also further work by Christ
[Chr84]).
A related aspect of harmonic analysis, the CalderonZygmund theory on
homogeneous Lie groups, has a long history as well. Again, this started with the
analysis of convolution operators (with earlier works e.g. by Kor
anyi and V
agi
[KV71] in the nilpotent direction), but in this book we will adopt an utilitarian
approach, and the setting of Coifman and Weiss [CW71a] of spaces of homogeneous
type will be sucient for our purposes (see Section 3.2.3 and Section A.4).
Proceeding with this part of the introduction on general homogeneous Lie
groups, let us follow Folland and Stein [FS82] and mention another important
occurrence of homogeneous Lie groups. If G is a noncompact real connected semisimple Lie group, its Iwasawa decomposition G = KAN contains the homogeneous
Lie group N whose family of dilations comes from an appropriate oneparameter
subgroup of the abelian group A (more precisely, if g = k k is the Cartan
decomposition of the Lie algebra g, the decomposition G = KAN corresponds
to the Iwasawa decomposition of the Lie algebra, g = k + a + n, where a is the
maximal abelian subalgebra of k , and the nilpotent Lie algebra n is the sum
of the positive root spaces corresponding to eigenvalues of a acting on g). This
decomposition generalises the decomposition of a real matrix as a product of an
orthogonal, diagonal, and an upper triangular with 1 at the diagonal matrix.
Furthermore, the the symmetric space G/K has the homogeneous nilpotent Lie
group N as its boundary in the sense that N may be identied with a dense
subset of the maximal boundary of G/K. As we show in Section 6.1.1 for no = 1,
if G = SU(no + 1, 1), G/K may be identied with the unit ball in Cno +1 and the
Heisenberg group Hno acts simply transitively on the complex sphere of Cno +1
Introduction
where one point has been excluded. This provides a link between the Heisenberg
group Hno , the analysis of the complex spheres, and the group SU(no + 1, 1)
or, more generally, between general semisimple Lie groups and homogeneous Lie
groups as boundaries of their symmetric spaces. For example, harmonic functions
on the symmetric space G/K can be represented by convolution operators on N
(see e.g. the survey of Koranyi [Kor72]).
Our setting contains the realm of Carnot groups as this class of groups
consists of the stratied Lie groups equipped with a specied metrics on the
rst layer, see e.g. Gromov [Gro96] for a survey on geometric analysis of Carnot
groups. Our setting includes any class of stratied Lie groups, for instance Hgroups, HeisenbergKaplan groups, Metiviertype groups [Met80], liform groups,
as well as Kolmogorovtype groups appearing in the study of hypoelliptic ultraparabolic operators including the KolmogorovFokkerPlanck operator (see Kolmogorov [Kol34], Lanconelli and Polidoro [LP94]). We refer to the book [BLU07]
by Bonglioli, Lanconelli and Uguzonni for a detailed consideration of these groups
and of their subLaplacians as well as related operators.
Hypoellipticity and Rockland operators
On compact Lie groups, the Fourier analysis and the symbolic calculus developed
in [RT10a] are based on the Laplacian and on the growth rate of its eigenvalues.
While on compact Lie groups the Laplacians (or the Casimir element) are operators
naturally associated to the group, it is no longer the case in the nilpotent setting.
Thus, on nilpotent Lie groups it is natural to work with operators associated
with the group through its Lie algebra structure. On stratied Lie groups these
are the subLaplacians, and such operators are not elliptic but hypoelliptic. More
generally, on graded Lie groups invariant hypoelliptic dierential operators are the
socalled Rockland operators.
Indeed, in [Roc78], Rockland showed that if T is a homogeneous leftinvariant
dierential operators on the Heisenberg group, then the hypoellipticity of T and
T t is equivalent to a condition now called the Rockland condition (see Denition
4.1.1). He also asked whether this equivalence would be true for more general
homogeneous Lie groups. Soon after, Beals showed in [Bea77b] that the hypoellipticity of a homogeneous leftinvariant dierential operator on any homogeneous
Lie group implies the Rockland condition. In the same paper he also showed that
the converse holds in some steptwo cases. Eventually in [HN79], Heler and Nourrigat settled what has become known as Rocklands conjecture by proving that
the hypoellipticity is equivalent to the Rockland condition (see Section 4.1.3). At
the same time, it was shown by Miller [Mil80] that in the setting of homogeneous
Lie groups, the existence of an operator satisfying the Rockland condition (hence
of an invariant hypoelliptic dierential operator in view of Heler and Nourrigats
result), implies that the group is graded, see also Section 4.1.1. This means, altogether, that the setting of graded Lie groups is the right generality for marrying
the harmonic analysis techniques with those coming from the theory of partial
Introduction
dierential equations.
A number of wellknown functional inequalities can be extended to the graded
setting, for example, see Bahouri, FermanianKammerer and Gallagher [BFKG12b].
Also, there are many contributions to questions of solvability related to the hypoellipticity problem: for a good introduction to local and nonlocal solvability
questions on nilpotent Lie groups see Corwin and Rothschild [CR81] and, missing to mention many contributions, for a more recent discussion of the topic see
M
uller, Peloso and Ricci [MPR99].
The hypoellipticity of second order operators is a very well researched subject. Its beginning may be traced to the 19th century with the diusion problems
in probability arising in Kolmogorovs work [Kol34]. H
ormander made a major
contribution [Hor67b] to the subject which then developed rapidly after that (see
e.g. the book of Oleinik and Radkevich [OR73]) until nowadays. We will not be
concerned much with these nor with the solvability problems in this book, since
one of topics of importance to us will be Rockland operators of an arbitrary degree,
and we will be giving more relevant references as we go along.
Here we want to mention that the question of the analytic hypoellipticity
turns out to be more involved than that in the smooth setting. In general, if a
graded Lie group is not stratied, there are no homogeneous analytic hypoelliptic leftinvariant dierential operators, a result by Heler [Hel82]. For stratied
Lie groups, the situation is roughly as follows: for Htype groups the analytic hypoellipticity is equivalent to the smooth hypoellipticity, while for step 3 (and
an additional assumption that the second stratum is onedimensional) the subLaplacians are not analytic hypoelliptic, see Metivier [Met80] and Heler [Hel82],
respectively, and the discussions therein. For the KohnLaplacian b in the Neumann problem as well as for higher order operators in this setting the analytic
hypoellipticity was shown earlier by Tartako [Tar78, Tar80]. Below we will mention a few more facts concerning the analytic hypoellipticity in the framework of
the analytic calculus of pseudodierential operators.
Pseudodierential operators
Several versions of the smooth calculi of pseudodierential operators on the
Heisenberg group have been considered over the years. An earlier attempt yielding
the calculus of invariant operators with symbols on the dual g of the Lie algebra
of the group was made by Strichartz [Str72]. A calculus for (rightinvariant) operators has been also constructed by Melin [Mel81] yielding parametrices for operators
elliptic in the socalled generating directions. In particular, the symbolic calculus
for invariant operators on stratied and graded Lie groups developed by Melin
further in [Mel83] provided a simpler proof of many of Heler and Nourrigats
arguments.
The question of a general symbolic calculus for convolution operators on
nilpotent Lie groups was raised by Howe in [How84], who also tackled questions
related to the Calder
onVaillancourt theorem. A more recent development of the
Introduction
calculus for invariant operators on homogeneous Lie groups and applications to
the corresponding symbolic conditions for the L2 boundedness of convolution operators was given by Glowacki in [Glo04] and [Glo07]. All this analysis applies to
invariant operators and employs the Euclidean Fourier transform yielding a symbol on the dual g . The symbol classes of such operators on the group are dened
as coming from the usual H
ormander classes on the (Euclidean) vector space g .
They satisfy the spectral invariance properties and yield further useful generalisations of parametrix constructions, see Glowacki [Glo12]. An approach to Melins
operators on nilpotent Lie groups from the point of view of the Weyl calculus was
done by Manchon, with further applications to the Weyl spectral asymptotics for
the innitesimal representations of elliptic operators in his calculus, see [Man91].
There exists also a calculus of leftinvariant integral operators on the Heisenberg
group, using Laguerre polynomials for its Fourier analysis, see Beals, Gaveau,
Greiner and Vauthier [BGGV86], or using Lerays quadratic Fourier transform by
Gaveau, Greiner and Vauthier [GGV86].
While these are mostly the calculi of invariant operators, the geometric considerations require one to also understand operators in the noninvariant setting.
However, here the amount of knowledge is more limited and most of the symbolic
calculus is restricted to the Heisenberg group. Dynins construction of certain operators on the Heisenberg group in [Dyn76] (see also [Dyn78]), was also developed
by Folland into considering metaHeisenberg groups in [Fol94]. Beside this, a noninvariant pseudodierential calculus on any homogeneous Lie group was developed
by Christ, Geller, Glowacki and Polin in [CGGP92] but this is not symbolic since
the operator classes are dened via properties of the kernel. In the revised version of [Tay84], Taylor described several (noninvariant) operator calculi and, in
a dierent direction, he also noted a way to develop symbolic calculi: using the
representations of the group, he denes a general quantization and symbols on any
unimodular type I group (by quantization, we mean a procedure which associates
an operator with a symbol). He illustrated this on the Heisenberg group and obtained there several important applications for, e.g., the study of hypoellipticity.
He used the fact that, because of the properties of the Schr
odinger representations
of the Heisenberg group, a symbol is a family of operators in the Euclidean space,
themselves given by symbols via the Weyl quantization. Recently, the denition
of suitable classes of Shubin type for these Weylsymbols led to another version
of the calculus on the Heisenberg group by Bahouri, FermanianKammerer and
Gallagher [BFKG12a].
A calculus of pseudodierential operators on the Heisenberg group in the
analytic setting was developed by Geller [Gel90], with applications to the analytic
hypoellipticity and further extensions of the calculus to real analytic CR manifolds. In particular, this implies the analytic hypoellipticity of the KohnLaplacian
on qforms on pseudoconvex real analytic manifolds. It also implies that the Szeg
o
projection preserves analyticity, recovering earlier results on the relations between
Szeg
o projections and b operator by Greiner, Kohn and Stein [GKS75], in turn
related to the solvability of the Lewy equation. The analytic hypoellipticity of the
Introduction
complex boundary KohnLaplacian on the (p, q)forms arising in the Neumann
2
problem was proved earlier by Tartako [Tar78] using L methods and by Tr`eves
[Tr`e78] using the calculus. Here we note that a corresponding Euclidean symbolic
calculus with applications to the propagation of analytic singularities and corresponding version of Fourier integral operators has been developed by Sjostrand
[Sj
o82], and the propagation of the analytic wave front set for the subLaplacian
was studied by Grigis and Sjostrand [GS85].
The analysis of the nilpotent setting can be extended to more general manifolds. Here, a typical application of the analysis on the Heisenberg groups is to
questions on contact manifolds. Indeed, a contact structure denes a grading on
the space of vector elds assigning them the degree of one or two. Locally, a contact manifold is dieomorphic to the Heisenberg group, and the principal symbol
of a dierential operator on the contact manifold is its higher order terms, with
the calculus on the contact manifold induced by that on the Heisenberg group,
at least on the principal symbol level. The ellipticity condition for an operator on
a contact manifold is thus replaced by the Rockland condition for the homogeneous principal part of the corresponding operator on the Heisenberg group. Such
constructions can be carried out in more general settings, in particular on the
socalled Heisenberg manifolds, which are smooth manifolds with a distinguished
hyperplane bundle. The calculus of operators in this setting was carried out by
Beals and Greiner in [BG88], in particular also generalising the calculus on CR
manifolds needed for the construction of parametrices for the KohnLaplacian b .
A recent advance in this direction mostly aimed at the second order operators
on Heisenberg manifolds, with a more intrinsic notion of the principal symbol of
such operators, was made by Ponge [Pon08]. Examples of such analysis include
CR manifolds and contact manifolds, with applications to the KohnLaplacian, the
GoverGraham operators, the contact Laplacian associated to the Rumin complex,
as well as to more general Rockland operators.
Moreover, such operators are subelliptic and their index may be calculated
by the AtiyahSinger index formula, see van Erp [vE10a]. The explicit knowledge of the BargmannFock representations of the Heisenberg group allows one
to construct the necessary Heisenberg calculus adapted to subelliplic operators in
this setting, leading to the index formula also for subelliptic pseudodierential
operators on contact manifolds, see van Erp [vE10b].
The calculus of pseudodierential operators on homogeneous Lie groups in
terms of their kernels developed by Christ, Geller, Glowacki and Polin in [CGGP92]
extended the parametrix construction of Heler and Nourrigat in [HN79] and some
properties of Taylors calculus from [Tay84] in the case of the Heisenberg group.
However, this calculus is not symbolic since it is based on the properties of the
kernel. The same is true for the analysis of operators on unimodular Lie groups
considered by Meladze and Shubin [MS87] where the operator classes were dened
in terms of local properties of the kernels.
Recently in [MR15], Mantoiu and the second author developed a more general
quantization scheme on general locally compact unimodular type I groups, thus
Introduction
encompassing in particular the cases of compact Lie groups by the second author
and Turunen [RT10a] and nilpotent Lie groups including the one developed in this
monograph. Moreover, the quantizations there allow one to deal with analogues
of both KohnNirenberg and Weyl quantizations. However, due to the generality
the scope of available results at the moment is much more limited than the one
presented in [RT10a] in the compact case, or in this book. The type I assumption
is useful for having a rich machinery concerning the abstract Plancherel theorem
(see Section 1.8.2), however it can be dropped for some questions: e.g. for an Lp Lq Fourier multiplier theorem on general locally compact separable unimodular
groups (without type I assumption) see Akylzhanov and Ruzhansky [AR15]. For
nilpotent Lie groups, the relation between such quantizations and Melins quantization described above has been also established in [MR15].
Quantization on homogeneous Lie groups and the book structure
Most of the above works that concern the noninvariant symbolic calculi of operators on nilpotent Lie groups, are restricted to the Heisenberg groups or to
manifolds having the Heisenberg group as a local model (except for the calculi
which are not symbolic). One of the reasons is that they rely in an essential way
on the explicit formulae for representations of the Heisenberg group. However, in
all the motivating aspects described above graded Lie groups appear as well as the
Heisenberg group. Also, graded groups appear as local models once one is dealing
with operators which are not in the form of sum of squares even on manifolds
such as the Heisenberg manifolds.
Recently, in [RT10a, RT13], the second author and Turunen developed a
global symbolic calculus on any compact Lie group. They dened symbol classes
so that the quantization procedure, analogous to the KohnNirenberg quantization on Rn , makes sense on compact Lie groups, and the resulting operators form
an algebra with properties close enough to the one enjoyed by the Euclidean
H
ormander calculus. In particular, one can also recover the Hormander classes of
pseudodierential operators on compact Lie groups viewed as compact manifolds
through conditions imposed on global full matrixvalued symbols. This approach
works for any compact Lie group and is intrinsic to the group in the sense that it
does not depend on pseudodierential calculus of (Euclidean) H
ormander classes
on its Lie algebra. While relying on the representation theory of the group, the
quantization and the calculus do not depend on explicit formulae for its representations. It does not depend either on the available Riemannian structure. This
gives an advantage over the calculi expressed in terms of a xed connection of
a manifold, such as the one developed by Widom [Wid80], Safarov [Saf97], and
Sharafutdinov [Sha05], see also the survey of McKeag and Safarov [MS11].
The crucial and new ingredient in the denition of symbol classes in [RT10a]
was the introduction and systematic use of dierence operators in order to replace
the Euclidean derivatives in the Fourier variables by analogous operators acting
on the unitary dual of the group. These dierence operators allow one to express
Introduction
the pseudodierential behaviour directly on the group and are very natural from
the point of view of the CalderonZygmund theory. These operators and their
properties on compact Lie groups will be reviewed in Section 2.2.2.
It is not possible however to extend readily the results of the compact case
developed in [RT10a] to the nilpotent context. Indeed, the global analysis on a
noncompact setting is usually more challenging than in the case of a compact
manifold. In the specic case of Lie groups, the dual of a noncompact group
is no longer discrete and the unitary irreducible representations may be innite
dimensional, and are often so. More problematically there is no Laplacian and
one expects to replace it by a subLaplacian on stratied Lie groups or, more
generally, by a positive Rockland operator on graded Lie groups; such operators
are not central.
Thus, in this book we study the global quantization of operators on graded
Lie groups, in particular aiming at developing an intrinsic symbolic calculus of such
operators. This is done in Chapter 5. As noted earlier, the graded Lie groups is a
natural generality for such analysis since we can still make a full use of the Rockland operators as well as from the representation theory which is well understood
e.g. by Kirillovs orbit method [Kir04]. The consequent Fourier analysis is then
also well understood from earlier works on the Plancherel formula on nilpotent
and even on more general locally compact unimodular type I groups; an overview
of this topic is given in Section 1.7.
Summarising very briey the results presented in Chapter 5, we introduce a
m
of symbols and of the
global quantization on graded Lie groups and classes S,
m
m
corresponding operators in , = Op S, such that for each (, ) with
1
0 and = 1, we have an operator calculus, in the sense that the set mR m
,
forms an algebra of operators, stable under taking the adjoint, and acting on the
Sobolev spaces in such a way that the loss of derivatives is controlled by the order
of the operator. Moreover, the operators that are elliptic or hypoelliptic within
these classes allow for a parametrix construction whose symbol can be obtained
from the symbol of the original operator. Some applications of the constructed
calculus are contained in Chapter 5, see also the authors paper [FR13] for further
applications to lower bounds of operators on graded Lie groups. A preliminary
very brief outline of the constructions here was given in [FR14a].
To lay down the necessary foundation for the quantization of operators and
symbols, we also make an exposition of the construction of the Sobolev spaces
on graded Lie groups based on positive Rockland operators. Such construction
has been previously done on stratied Lie groups by Folland [Fol75], for Sobolev
spaces based on the (leftinvariant) subLaplacian. SubLaplacians in this context
are (up to a sign) a particular case of positive Rockland operators on stratied
groups and our results coincide with Follands in this case. However if we follow
Follands treatment but now in the more general context of graded Lie groups,
beside the appearance of several technical problems, we would be led to make further assumptions. One of them would be that the degree of the positive Rockland
operator must be less than the homogeneous dimension Q of the group, < Q
10
Introduction
(assuming < Q ensures the uniqueness of its homogeneous fundamental solution). In fact, Goodman makes such an assumption in his treatment of Sobolev
spaces on graded Lie group in [Goo76]. In order to avoid this assumption and also
to deal with other issues, we need to develop other arguments in the study of the
powers of a general positive Rockland operator R and of I + R and in the study of
the associated Sobolev spaces. This is done under no assumptions on the relation
between and Q in Sections 4.3 and 4.4.
The analysis of Sobolev spaces is based on the heat kernel associated with
a positive Rockland operator. We make an exposition of this topic in Section
4.2.2. Our presentation there follows essentially the arguments of Folland and
Stein [FS82]. The heat kernels are not necessarily positive functions and the heat
semigroup does not necessarily correspond to a martingale as in the stratied
case or more generally for sums of squares of vector elds with Hormanders condition. Such sums of squares have been analysed in much more general settings.
For example, we can refer to the book of Varopoulos, SaloCoste and Coulhon
[VSCC92] for a treatment of the heat kernel on unimodular Lie groups of polynomial growth, and the usually associated to it estimates, such as Harnack and
Sobolev inequalities. For another point of view, allowing dealing with heat kernels
associated to more general subelliptic second order order dierential operators we
refer to Dungey, ter Elst and Robinsons book [DtER03].
Overall, the majority of the background material can be (sometimes even
more easily) introduced in the setting of homogeneous Lie groups and we discuss
these in Chapter 3. Our treatment in this chapter is inspired by those of Folland
and Stein [FS82] and Ricci [Ric] but is slightly more general than that in the
existing literature since we allow kernels and operators to have complexvalued
homogeneity degrees. This allows us to treat complex powers of operators later
on, e.g. in Section 4.3.3.
We assume that the reader is familiar with analysis at a graduate level, e.g. as
presented in the books of Rudin [Rud87, Rud91], Reed and Simon [RS80, RS75],
or Folland [Fol99]. Nevertheless, we make a brief exposition of topics, mostly from
the representation theory of groups, to remind the reader of the necessary concepts
used in later parts of the book and to x the terminology and notation. This is
done in Chapter 1, and references to more material are given throughout.
The exposition of the (matrix) quantization on compact Lie groups from
[RT10a] and its related works is given in Chapter 2. This serves both as an introduction to the topic as well as provides motivation and examples for some of the
concepts presented later in the book.
Chapter 6 is devoted to presenting an application of the general theory developed in Chapter 5 to the concrete setting of the Heisenberg groups. Some results
from this chapter have been announced in the authors paper [FR14b] and this
chapter provides their proofs. We give the necessary preliminaries of the analysis
on the Heisenberg group, including a description of its dual using Schrodinger
representations, with further concrete expressions for the Plancherel measure and
odinger representations
Plancherel formula. For the Heisenberg group Hn , its Schr
Introduction
11
are acting on the space L2 (Rn ), thus yielding symbols acting on the Schwartz
space S(Rn ), the space of smooth vectors of . In turn, these symbols can be conveniently described using their Weyl quantization, giving a notion of scalarvalued
symbols. In this particular case of the Heisenberg group, the symbol classes of
Chapter 5 can be characterised by the property that these symbols belong to
some Shubin spaces, more precisely, a semiclassicaltype version of the usual
Shubin classes of symbols. Consequently, this is applied to giving criteria for ellipticity and hypoellipticity of operators on the Heisenberg group in terms of the
invertibility properties of their symbols. We provide a list of examples to show
the applicability of these results in several settings.
Open Access. This chapter is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/),
which permits use, duplication, adaptation, distribution and reproduction in any medium
or format, as long as you give appropriate credit to the original author(s) and the source,
a link is provided to the Creative Commons license and any changes made are indicated.
The images or other third party material in this chapter are included in the works
Creative Commons license, unless indicated otherwise in the credit line; if such material
is not included in the works Creative Commons license and the respective action is not
permitted by statutory regulation, users will need to obtain permission from the license
holder to duplicate, adapt or reproduce the material.
Notation and conventions
N = {1, 2, 3, . . .}
N0 = N {0}
R = R\{0}
R+ = (0, )
C+ = {z C, Re z > 0}
0! = 1
For = (1 , . . . , N ), ! = 1 ! N !,  = 1 + + N
[] is the homogeneous degree, dened in (3.12)
x is the smallest n Z such that n > x
x is the largest n Z such that n x
M = max{ : Nn0 with [] M }, as dened in (3.35)
A
B means there is some c > 0 such that c1 A B cA
aj := j cj aj denotes a (nite) linear combination with some (irrelevant) conj
stants cj
Unit elements: e on general groups, and 0 on nilpotent groups
x is the deltadistribution at x: x () = (x)
j,k is the Kronecker delta: j,k = 0 for j = k, and j,j = 1
L (H1 , H2 ) is the space of all linear continuous mappings from H1 to H2
L (H) := L (H, H)
U (H) is the space of unitary mappings in L (H)
G is a Lie group, g is its Lie algebra, and
U(g) is its universal enveloping algebra dened in Section 1.3
T, T and T t are dened by (1.8), (1.9), and (1.10) for an element T U(g) and
in Denition 1.3.1 for an operator T on L2 (G)
Rep G is the set of all strongly continuous unitary irreducible representations of
the group G
is the unitary dual of G, i.e. Rep G modulo the equivalence of representations
G
is the space of square integrable elds on G
with respect to the Plancherel
L2 (G)
measure, see (1.29)
LL (L2 (G)), and K(G) are the realisations of the von Neumann algebra
L (G),
of the leftof the group G as the spaces of the bounded elds of operators on G,
13
14
Notation and conventions
invariant operators in L (L2 (G)), and of the convolution kernels corresponding to
the latter, respectively, see Section 1.8.2
2
2
L
a,b (G), LL (La (G), Lb (G)), and Ka,b (G) are the Sobolev versions of the above,
see Section 5.1.2
Di k (G) is the space of leftinvariant dierential operators of order k
is the space of dierence operators on G
of order k
di k (G)
Cc (G) is the space of continuous functions on G with compact support
Co (G) is the space of continuous functions on G vanishing at innity (see Denition
3.1.57)
D(G) = Cc (G) or D(R) = Cc (R) are spaces of smooth compactly supported
functions
C (G, F ) denotes the set of smooth functions from G to a Frechet space F
Lp (G) or simply Lp for 1 p is the usual Lebesgue space on G with norm
Lp = Lp (G) = p
(, )L2 = (, )L2 (G) is the Hilbert sesquilinear form on L2 (G) corresponding to the
norm L2
M (G) the Banach space of regular complex measures on G endowed with the total
mass M (G)
, denotes the distributional duality
Q denotes the homogeneous dimension of a homogeneous Lie group. After Section
6.4.2, it denotes the harmonic oscillator
sup always means the essential supremum with respect to the corresponding measure
Open Access. This chapter is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/),
which permits use, duplication, adaptation, distribution and reproduction in any medium
or format, as long as you give appropriate credit to the original author(s) and the source,
a link is provided to the Creative Commons license and any changes made are indicated.
The images or other third party material in this chapter are included in the works
Creative Commons license, unless indicated otherwise in the credit line; if such material
is not included in the works Creative Commons license and the respective action is not
permitted by statutory regulation, users will need to obtain permission from the license
holder to duplicate, adapt or reproduce the material.
Chapter 1
Preliminaries on Lie groups
In this chapter we provide the reader with basic preliminary facts about Lie groups
that we will be using in the sequel. At the same time, it gives us a chance to x
the notation for the rest of the monograph. The topics presented here are all wellknown and we decided to give a brief account without proofs referring the reader
for more details to excellent sources where this material is treated from dierent
points of view; for example, the monographs by Chevalley [Che99], Fegan [Feg91],
Nomizu [Nom56], Pontryagin [Pon66], to mention only a few. Thus, this chapter
can also serve as a quick and informal introduction to the subject, and we refer to
monographs [RT10a] for an undergraduate level introduction to general Lie groups
and their representation theory, and to Corwin and Greenleaf [CG90] or Goodman
[Goo76] for a rather comprehensive treatment of nilpotent Lie groups. The groups
that we are dealing with in the monograph are either compact or nilpotent Lie
groups, so we can restrict our attention to unimodular Lie groups only.
The choice of material is adapted to our subsequent needs and, after giving
basic denitions, we go straight to discussing convolutions, invariant dierential
operators, and elements of the representation theory. More information on compact or homogeneous nilpotent Lie groups will be given in relevant chapters at
appropriate places. In particular Section 3.1.1 will provide examples and basic
properties of graded nilpotent Lie groups. Relevant monographs to consult on invariant dierential operators and related harmonic analysis may be Helgasons
books [Hel84b, Hel01] or Wallachs [Wal73].
1.1
Lie groups, representations, and Fourier transform
A Lie group G is a smooth manifold endowed with the smooth mappings
G G (x, y) xy G
and
G x x1 G
satisfying, for all x, y, z G, the properties
The Editor(s) (if applicable) and The Author(s) 2016
V. Fischer, M. Ruzhansky, Quantization on Nilpotent Lie Groups, Progress in Mathematics 314,
DOI 10.1007/9783319295589_1
15
16
Chapter 1. Preliminaries on Lie groups
1. x(yz) = (xy)z;
2. ex = xe = x;
3. xx1 = x1 x = e,
where e G is an element of the group called the unit element. To avoid unnecessary technicalities at a few places, we will always assume that G is connected,
although we sometimes will emphasise it explicitly. A compact Lie group is a Lie
group which is compact as a manifold.
Lie groups are naturally topological groups. Recall that a topological group
G is a topological set G endowed with the continuous mappings
G G (x, y) xy G
and
G x x1 G
satisfying, for all x, y, z G, the same properties 1., 2. and 3. as above. When
the topology of a topological group is locally compact (i.e. every point has a
compact neighbourhood), we say that the group is locally compact. Lie groups are
(Hausdor) locally compact.
Representations
A representation of a group G on a Hilbert space H = {0} is a homomorphism
of G into the group of bounded linear operators on H with bounded inverse.
This means that
for every x G, the linear mapping (x) : H H is bounded and has
bounded inverse,
for any x, y G, we have (xy) = (x)(y).
A representation of a group G is unitary when (x) is unitary for every
x G. Hence a unitary representation of a group G is a homomorphism
Hom(G, U (H )), which means that
for every x G, the linear mapping (x) : H H is unitary:
(x)1 = (x) ;
for any x, y G, we have (xy) = (x)(y).
Here and everywhere, if H is a topological vector space, L (H) denotes the
space of all continuous linear operators H H, and U (H) the space of unitary
ones, with respect to the inner product on H. For two dierent topological vector
spaces H1 and H2 , we denote by L (H1 , H2 ) the space of all linear continuous
mappings from H1 to H2 .
An invariant subspace for a representation is a vector subspace W H
such that (x)W W holds for every x G. A representation is called irreducible when it has no closed invariant subspaces.
1.1. Lie groups, representations, and Fourier transform
17
Let us give the prototype example of a representation which is not irreducible.
If j Hom(G, U (Hj )) is a family of representations, then using the direct sum
H :=
H j
with the induced inner product, we get a representation which is the direct sum
of j :
=
j Hom(G, U (H )), (x)Hj = j (x).
j
Naturally, a sum of several j s can not be irreducible as each Hj is a closed
invariant subspace of H .
If the space H is nite dimensional, the representation is said to be nite
dimensional and its dimension/degree is dened by
d := dim H .
The trivial representation, sometimes denoted by 1, is given by the group homomorphism G x 1 C, and its dimension is one. If H is innite dimensional,
then the representation is said to be innite dimensional.
Two representations 1 and 2 are said to be equivalent if there exists a
bounded linear mapping A : H1 H2 between their representation spaces with
a bounded inverse such that the relation
A1 (x) = 2 (x)A
(1.1)
holds for all x G. In this case we write
1 2
or, more precisely sometimes,
1 A 2
and denote their equivalence class by [1 ] = [2 ]. For unitary representations,
A is assumed to be unitary as well. A bounded linear mapping with bounded
inverse satisfying the relation (1.1) is sometimes called an intertwining operator or
intertwiner. The set of bounded linear mappings A with bounded inverse satisfying
the relation (1.1) is denoted by Hom(1 , 2 ).
Note that for any representation , Hom(, ) contains at least IH , C,
where IH is the identity mapping on H .
We now assume that the group G is topological. A representation of G is
continuous if the mapping
G H H
(x, v)
(x)v
18
Chapter 1. Preliminaries on Lie groups
is continuous. A representation of G is called strongly continuous if the mapping
: G L (H ) is continuous for the strong operator topology in L (H ), that
is, if the mapping
G H
x (x)v
is continuous for all v H .
A continuous representation is strongly continuous. The converse is true for
unitary representations. Indeed, if is a unitary representation of G, then we have
for any x, x0 G and v, v0 H ,
(x)v (x0 )v0 H
=
=
1
(x0 )((x1
0 x)v v0 )H = (x0 x)v v0 H
1
(x1
0 x)(v v0 ) + ((x0 x)v0 v0 )H
1
(x1
0 x)(v v0 )H + (x0 x)v0 v0 H
= v v0 H + (x1
0 x)v0 v0 H ,
having used only the unitarity of and the triangle inequality. This shows that if
a representation of G is unitary and strongly continuous then it is continuous.
Schurs lemma: Let be a strongly continuous unitary representation of a topological group G on a Hilbert space H . The representation is irreducible if and
only if the only bounded linear operators on H commuting with all (x), x G,
are the scalar operators. Equivalently,
irreducible Hom(, ) = {IH : C}.
The set of all equivalence classes of strongly continuous irreducible unitary
representations of G is called the unitary dual of G or just dual of G and is denoted
by G.
Later, we will give more details on representations of compact or nilpotent
Lie groups and their dual.
The unitary dual of G is never a group unless G is commutative. However,
has a natural structure of a
if G is a commutative locally compact group, then G
commutative locally compact group and we have
Pontryagin duality: if G is a commutative locally compact group, then G
G.
For most of the statements in the sequel, if they hold for one representation,
they will also hold for all equivalent representations. That is why we may simplify
instead of [] G
for its equivalence class.
the notation a little writing G
In this case we can think of as either any representative from its class or the
equivalence class itself. If we need to work with a particular representation from
an equivalence class (for example the one diagonalising certain operators in a
particular choice of the basis in H ) we will specify this explicitly.
1.1. Lie groups, representations, and Fourier transform
19
Haar measure
A fundamental fact, valid on general locally compact groups, is the existence of
an invariant measure, called Haar measure:
Theorem 1.1.1. Let G be a locally compact group. Then there exists a nonzero leftinvariant measure on G, and it is unique up to a positive constant. More precisely,
there exists a positive Radon measure on G satisfying
xA = A for every Borel set A G and every x G,
where A denotes the measure of the set A with respect to this Radon measure.
In the sequel, we denote this measure by dx, dy, etc., depending on the variable
of integration. Then, for every x G and every continuous compactly supported
function f on G, we have
f (xy)dy =
f (y)dy.
G
We x one of such measures. In this monograph, we will be only dealing with
either compact or nilpotent Lie groups, in which case it can be shown that the
Haar measure is also rightinvariant:
Ax = A for every Borel set A G and every x G,
and also
G
f (yx)dy =
f (y)dy;
such groups are called unimodular. Since the mapping f G f (y 1 )dy is positive, leftinvariant, and normalised, by uniqueness we must also have
f (y 1 )dy =
f (y)dy.
G
For a more general denition of a modular function we can refer to Denition
B.2.10. Here we can summarise a few properties of (unimodular) groups:
Any Lie group is a locally compact (Hausdor) group.
Any compact (Hausdor) group is a locally compact (Hausdor) group and
it is also unimodular.
Any abelian locally compact (Hausdor) group is unimodular.
Any nilpotent or semisimple Lie group is unimodular.
If 1 p , Lp (G) or simply Lp denote the usual Lebesgue space on G
with respect to the Haar measure, with the norm
Lp = Lp (G) = p ,
20
Chapter 1. Preliminaries on Lie groups
given for p [1, ) by
f p =
and for p = by
1/p
f (x) dx
f = sup f (x).
xG
Here the supremum refers to the essential supremum with respect to the Haar
measure.
The Hilbert sesquilinear form on L2 (G) is denoted by
(f1 , f2 )L2 = (f1 , f2 )L2 (G) =
f1 (x)f2 (x)dx.
G
Example 1.1.2. Let us give an important example of socalled left and right regular
representations leading to the notions of left and rightinvariant operators. We
dene the left and right regular representations of G on L2 (G), L , R : G
U (L2 (G)), respectively, by
L (x)f (y) := f (x1 y)
and
R (x)f (y) := f (yx).
Denition 1.1.3. An operator A is called left (right, resp.) invariant if it commutes
with the left (right, resp.) regular representation of G.
Fourier analysis
For f L1 (G) we dene its Fourier coecient or group Fourier transform at the
strongly continuous unitary representation as
f (x)(x) dx.
(1.2)
FG f () f() (f ) :=
G
More precisely, we can write
(f()v1 , v2 )H =
G
f (x)((x) v1 , v2 )H dx.
This gives a linear mapping f() : H H . If the representation is nite
dimensional, then after a choice of a basis in the representation space H , the
Fourier coecient f() can be also viewed as a matrix f() Cd d .
Remark 1.1.4. The choice of taking the adjoint (x) in (1.2) is natural if we think
n = { (x) = e2ix }Zn
of the unitary dual of the torus Tn = Rn /Zn being T
n
Z , and the Fourier transform on the torus dened by
f( ) f() =
e2ix f (x)dx =
f (x) (x) dx.
Tn
Tn
1.1. Lie groups, representations, and Fourier transform
21
In other contexts, the other choice, that is, integrating against (x) instead of
(x) , may be made. This is the case for instance in the study of C algebras
associated with groups.
Remark 1.1.5. We note that the Fourier coecient f() depends on the choice of
the representation from its equivalence class []. Namely, if 1 2 , so that
2 (x) = U 1 1 (x)U
for some unitary U and all x G, then
f(2 ) = U 1 f(1 )U.
This means that strictly speaking, we need to look at Fourier coecients modulo
conjugations induced by the equivalence of representations. This should, however,
cause no problems, and we refer to Remark 2.2.1 for more discussion on this.
Recalling that the Fourier transform on Rn maps translations to modulations,
f L1 (G) and x G, then
here we have an analogous property, namely, if G,
f
( x)() = (x)f() and f
(x )() = f()(x),
(1.3)
whenever the right hand side makes sense. Let us show these properties by a formal
argument, which can be made rigorous on Lie groups, see the proof of Proposition
1.7.6, (iv). We have
(x)f () =
f (y)(x)(y) dy
G
f (y)(yx1 ) dy
=
G
f (yx)(y) dy
=
G
=
as well as
f
(x )()
f
( x)(),
=
G
=
=
G
(1.4)
f (xy)(y) dy
f (y)(x1 y) dy
f (y)(y) (x)dy
f()(x).
(1.5)
We will continue with a more detailed discussion of the Fourier transform on
compact Lie groups in Section 2.1, on nilpotent Lie groups in Section 1.8.1 and,
more generally, on a separable locally compact connected, unimodular, amenable
group G of type I in Section 1.8.2.
22
Chapter 1. Preliminaries on Lie groups
1.2 Lie algebras and vector elds
A (real) Lie algebra is a real vector space V endowed with a bilinear mapping
V V (a, b) [a, b] V,
called the commutator of a and b, such that
[a, a] = 0 for every a V ;
Jacobi identity: [a, [b, c]] + [b, [c, a]] + [c, [a, b]] = 0 for all a, b, c V .
By writing [a + b, a + b] = [a, a] + [a, b] + [b, a] + [b, b] we see that the rst property
is equivalent to the condition that
a, b V
[a, b] = [b, a].
We now proceed to equip the tangent space of G (at every point) with a Lie
algebra structure. A map X(x) : C (G) R is called a tangent vector to G at
x G if
X(x) (f + g) = X(x) f + X(x) g;
X(x) (f g) = X(x) (f )g(x) + f (x)X(x) (g).
The notation X(x) is used only in this section and the reason for its choice is that
we want to reserve the notation Xx for derivatives, to be used later.
The space of all tangent vectors at x is a nite dimensional vector space of
dimension equal to the dimension of G as a manifold; the nite dimensionality can
be seen by passing to local coordinates. This vector space is denoted by Tx G. The
disjoint union,
T G :=
Tx G
xG
is a vector bundle over X, called the tangent bundle. The canonical projection
proj : T G G is given by proj X(x) := x. If Ux is a (suciently small) open
neighbourhood of x in G, we can trivialise the vector bundle T G by proj1 (Ux )
Ux E with a vector space E of dimension equal to that of G. This induces the
manifold structure on T G.
A (smooth) vector eld on G is a (smooth) section of T G, i.e. a (smooth)
mapping X : G T G such that X(x) X(x) Tx G. It acts on C (G) by
(Xf )(x) := (X(x) f )(x),
f C (G).
There is a bracket structure on the space of vector elds acting on C (G) given
by
[X, Y ](x)(f ) := X(x)Y f Y (x)Xf, x G,
leading to the corresponding (smooth) vector eld [X, Y ] : G T G given by
x [X, Y ](x). One can readily check that [X, X] = 0 for every vector eld X and
1.2. Lie algebras and vector elds
23
that the introduced bracket satises the Jacobi identity. This bracket [, ] is called
the commutator bracket for vector elds.
We now recall that G is also a group, and relate vector elds to the group
structure. First, we dene the left and right translations by an element y G:
Ly , Ry : G G,
Ly (x) := yx, Ry (x) := xy.
Consequently, their derivatives are the mappings
dLy , dRy : T G T G such that dLy L (Tx G, Tyx G), dRy L (Tx G, Txy G).
Now, a vector eld X : G T G is called leftinvariant if it commutes with the
left translations, in the sense that
X Ly = dLy X
y G.
(1.6)
A similar construction leads to the notion of rightinvariant vector elds, satisfying
X Ry = dRy X
for all y G.
It follows that once a leftinvariant vector eld is dened at any one point, by
the leftinvariance it is uniquely determined at all points. Thus, the mapping X
X(e) is a onetoone correspondence between leftinvariant vector elds on G and
the tangent space Te G at the unit element e G. Conversely, given X(e) Te G,
the vector eld X dened by (1.6) is automatically smooth and, by denition,
leftinvariant. With this identication, we can now simplify the notation for leftinvariant vector elds X, writing X also for its value X(e) at the unit element. It
can be readily checked that if X and Y are leftinvariant vector elds, so is also
their commutator [X, Y ].
Denition 1.2.1. The Lie algebra g of the Lie group G is the space Te G equipped
with the commutator [, ] induced by the commutator bracket of vector elds.
We now dene the exponential mapping expG . For X g, consider the initial
value problem for a function : [0, ) G, > 0, given by the ordinary dierential
equation determined by the leftinvariant vector eld associated with X:
(t) = X((t)) ,
(0) = e.
From the theory of ordinary dierential equations we know that this equation is
uniquely solvable on some interval [0, ) and the solution depends smoothly on
X(e) . Moreover, we notice that we can increase the interval of existence by taking
smaller vectors X(e) , in particular, in such a way that the solution exists on the
interval [0, 1]. In this case we set expG X := (1). Altogether, it follows that the
mapping expG is a smooth dieomorphism from some open neighbourhood of 0 g
to some open neighbourhood of e G.
24
Chapter 1. Preliminaries on Lie groups
Now, each vector X g can be viewed as a leftinvariant dierential operator
on C (G) dened by
Xf (x) :=
d
f (x expG (tX))t=0 .
dt
(1.7)
Indeed, it can be readily checked that XL (y) = L (y)X for all y G. Analogously, the same vector X g denes a rightinvariant dierential operator, which
we denote by
(x) := d f (expG (tX) x)t=0 .
Xf
dt
Thus, throughout this book, we will be interpreting the Lie algebra g = Te G of G
as the vector space of rst order leftinvariant partial dierential operators on G.
The space of all leftinvariant vector elds will be sometimes denoted by D(G) or
by Di 1 (G), and the space of all rightinvariant vector elds by D(G).
1.3
Universal enveloping algebra and dierential
operators
Roughly speaking, the universal enveloping algebra of a Lie algebra g is the natural
noncommutative polynomial algebra on g. If g is the Lie algebra of a Lie group G,
then, similarly to the interpretation of g as the space of leftinvariant derivatives
on G, the universal enveloping algebra U(g) of the Lie algebra of G will be also
interpreted as the vector space of leftinvariant partial dierential operators on
G of nite order. The associative algebra will be generated as a complex algebra
over g, so that we could write U(gC ) for it, where gC denotes the complexication
of g. However, we will simplify the notation writing U(g), and will later use the
PoincareBirkhoWitt theorem to identify it with the leftinvariant dierential
operators on G with complex coecients. Let us now formalise these statements.
The following construction is algebraic and works for any real Lie algebra g.
Let us denote the mfold tensor product of gC by m gC := gC gC , and let
T :=
m gC
m=0
be the tensor product algebra of g, which means that T is the linear span of the
elements of the form
00 1 +
Km
M
mk Xmk1 Xmkm ,
m=1 k=1
where 1 is the formal unit element of T , mk C, Xmkj g, and M, KM N.
This T becomes an associative algebra with the product
(X1 Xp )(Y1 Yq ) := X1 Xp Y1 Yq
1.3. Universal enveloping algebra and dierential operators
25
extended to a uniquely determined bilinear mapping T T T . We now want to
induce the commutator structure on T : let I be the twosided ideal in T spanned
by the set
O := {X Y Y X [X, Y ] : X, Y g},
i.e. I is the smallest vector subspace of T such that
O I;
for every J I and T T we have JT, T J I.
The quotient algebra
U(g) := T /I
is called the universal enveloping algebra of g; the quotient mapping
: T T T + I U(g) = T /I,
restricted to g, g : g U(g), is called the canonical mapping of g. This gives the
embedding of g into U(g):
AdoIwasawa theorem: the canonical mapping g : g U(g) is injective.
Let n = dim G and let {Xj }nj=1 be a basis of the Lie algebra g of G. Regarded
as rst order leftinvariant derivatives, they give rise to higher order leftinvariant
dierential operators
X = X11 . . . Xnn ,
= (1 , . . . , n ) Nn0 .
The converse is also true (for a stronger version of this see e.g. [Bou98, Ch 1, Sec.
2.7]):
PoincareBirkhoWitt theorem: any leftinvariant dierential operator T on G
can be written in a unique way as a nite sum
T =
c X ,
Nn
0
where all but a nite number of the coecients c C are zero. This gives an
identication between the universal enveloping algebra U(g) and the space of leftinvariant dierential operators on G.
We denote the space of all leftinvariant dierential operators of order k by
Di k (G).
If T is as above, we dene three new elements T, T , and T t of U(g) via
T :=
c (Xn )n . . . (X1 )1 ,
(1.8)
c (Xn )n . . . (X1 )1 ,
(1.9)
Nn
0
T :=
Nn
0
26
Chapter 1. Preliminaries on Lie groups
and
T t :=
c (Xn )n . . . (X1 )1 .
(1.10)
Nn
0
These T and T t are called the (formal) adjoint and transpose operators of T ,
respectively. Naturally, they coincide with the natural transpose and formal adjoint
operators of their corresponding leftinvariant vector elds. Recall that the latter
operators are dened via:
Denition 1.3.1. Let T be an operator T on L2 (G) with domain D(G) (T may be
unbounded, D(G) Dom T ). The natural transpose and formal adjoint operators
of T are the operators T t and T on L2 (G) dened via
T , = , T t
and
(T , )L2 (G) = (, T )L2 (G) ,
, D(G).
We also dene the operator T on L2 (G) via
T := T ,
for , Dom T .
Note that we also have, e.g.,
T = {T t } = {T}t
and so on. Denoting
f(x) := f (x1 ),
the left and right invariant dierential operators are related by
(x) = (X f)(x1 )
Xf
and hence
f (x) = (1) (X f)(x1 ).
X
(1.11)
Indeed, we can write
X f(x) =
d
d
)(x1 ),
f ((x expG (tX))1 )t=0 = f (expG (tX)x1 )t=0 = (Xf
dt
dt
implying (1.11).
For any X g identied with a leftinvariant vector eld, we have
y {f (xy)} = d f (xetX y)t=0 = Xx {f (xy)}.
X
dt
Recursively, we obtain
{f (xy)} = X {f (xy)}.
X
y
x
(1.12)
The rst order dierential operators are formally skewsymmetric:
1 )f2 =
2 ),
(Xf1 )f2 =
f1 (Xf2 ) and
(Xf
f1 (Xf
G
1.3. Universal enveloping algebra and dierential operators
27
so that from (1.11) we also have
(x) = (X f)(x1 ) = (X t f)(x1 ).
Xf
We now summarise several further notions and their properties that will be
of use to us in the sequel:
there is a natural representation of the Lie group G acting on its Lie algebra
g, called the adjoint representation. To introduce it, rst dene the inner
automorphism Ix (y) := xyx1 . We have Ix : G G and Ixy = Ix Iy . Its
dierential at e gives a linear mapping from Te G to Te G, and we denote it
by
Ad(x) := (dIx )e : g g.
We have Ad(e) = I and Ad(xy) = Ad(x)Ad(y), so that Ad : G L (g)
becomes a representation of G on g;
the left and right multiplications on G are related by
x expG X = exp(Ad(x)X)x,
x G, X g;
a Lie group G is called a linear Lie group if it is a closed subgroup of GL(n, C);
the adjoint representation of such G is given by
Ad(X)Y = XY X 1
as multiplication of matrices;
universality of unitary groups: any compact Lie group is isomorphic to a
subgroup of U (N ), the group of (N N )unitary matrices, for some N N;
let ad : g L(g) be the linear mapping dened by
ad(X)Y := [X, Y ];
then d(Ad)e = ad; see also Denition 1.7.4;
the Killing form of the Lie algebra g is the bilinear mapping B : g g R
dened by
B(X, Y ) := Tr(ad(X) ad(Y ));
it satises
B(X, Y ) = B(Y, X) and B(X, [Y, Z]) = B([X, Y ], Z)
and is invariant under the adjoint representation of G, namely,
B(X, Y ) = B(Ad(x)(X), Ad(x)(Y )) for all x G, X, Y g;
28
Chapter 1. Preliminaries on Lie groups
A connected Lie group G is called semisimple if B is nondegenerate; a
connected semisimple group G is compact if and only if B is negative denite.
The Adinvariance of the Killing form has its consequences. On one hand,
any bilinear form on g can be extended to a bilinear (nonnecessarily positive
denite) metric on G by left translations. It is automatically leftinvariant. On
the other hand, if the form on g is Adinvariant, then the extended metric is
also rightinvariant. Thus, we can conclude that the Killing form induces a biinvariant metric on G. By the last property above, if G is semisimple, the Killing
form is nondegenerate, and hence the corresponding metric is pseudoRiemannian.
Moreover, if G is a connected semisimple compact Lie group, the positivedenite
form B induces the biinvariant Riemannian metric on G.
For the basis {Xj }nj=1 as above, let us dene Rij := B(Xi , Xj ). If the group
G is semisimple, the matrix (Rij ) is invertible, and we denote its inverse by R1 .
This leads to another vector space basis on g given by
X i :=
(R1 )ij Xj ,
j=1
and to the socalled Casimir element of U(g) dened by
:=
Xi X i .
i=1
It has the crucial property: is independent of the choice of the basis {Xj }, and
T = T for all T U(g).
We nish this section with the formula for the group product which will be useful
for us, especially in the nilpotent case:
Theorem 1.3.2 (BakerCampbellHausdor formula). Let G be a Lie group with
Lie algebra g. There exists a neighbourhood V of 0 in g such that for any X, Y V ,
we have
n
(1)n+1
( j=1 (pj + qj ))1
expG X expG Y = expG
n
p1 !q1 ! . . . pn !qn !
n
n>0
p,qN0
pi +qi >0
(adX)p1 (adY )q1 . . . (adX)pn (adY )qn 1 Y .
The equality holds whenever the sum on the righthand side is convergent.
Writing rst few terms explicitly, we have
expG X expG Y
1
1
1
= expG X + Y + [X, Y ] + [[X, Y ], Y ] [[X, Y ], X] + . . . .
2
12
12
1.4. Distributions and Schwartz kernel theorem
1.4
29
Distributions and Schwartz kernel theorem
Here we x the notation concerning distributions. For an extensive analysis of
spaces of distributions and their properties on manifolds we refer to [Hor03].
The space of smooth functions compactly supported in a smooth manifold M
will be denoted by D(M ). Throughout the book, any smooth manifold is assumed
to be paracompact (i.e. every open cover has an open renement that is locally
nite) and this allows us to consider the space of distributions D (M ) as the dual
of D(M ). Note that any Lie group is paracompact.
If u D (M ) and D(M ), we shall denote the evaluation of u on by
u, , or even by u, M when we wish to be precise; however, we shall usually
pretend that the distributions are functions and write
u(x)(x)dx, u D (M ), D(M ).
u, =
M
The Schwartz space S(Rn ) of rapidly decreasing functions will be equipped
with a family of seminorms dened by
N
(1.13)
sup
(1 + x)
f (x) .
f S(Rn ),N :=
x
N, xRn
Its dual, the space of tempered distributions, is denoted by S (Rn ).
Theorem 1.4.1 (Schwartz kernel theorem). We have the following statements:
Let T : S(Rn ) S (Rn ) be a continuous linear operator. Then there exists a
unique distribution S (Rn Rn ) such that
(x, y)(y)dy.
T (x) =
Rn
In other words, T is an integral operator with kernel . The converse is also
true.
Let M be a smooth connected manifold and let T : D(M ) D (M ) be a
continuous linear operator. There exists a unique distribution D (M M )
such that
T (x) =
(x, y)(y)dy.
M
In other words, T is an integral operator with kernel . The converse also is
true.
In both cases, the map T is an isomorphism of topological vector space.
We refer to e.g. [Tre67] for further details. We will also give a version of this
theorem on Lie groups for leftinvariant operators in Corollary 3.2.1.
30
Chapter 1. Preliminaries on Lie groups
Let be an open set in Rn or in M . We say that u D () is supported in the
set K if u, = 0 for all D() such that = 0 on K. The smallest closed
set in which u is supported is called the support of u and is denoted by supp u.
The space of compactly supported distributions on M is denoted by E (M ), and
the duality between E (M ) and C (M ) will still be denoted by , .
We write u Dj () for the space of distributions of order j on , which
means that for any compact subset K of ,
C > 0
D(K)
u,  CC j (K) ,
but j does not depend on K. An important property of such distributions, useful
for us, is the following
Proposition 1.4.2. If a distribution u Dj (Rn ) has support supp u = {0}, then
there exist constants a C such that
u=
a 0 ,
j
where 0 () = (0) is the deltadistribution at zero.
1.5 Convolutions
Let f, g L1 (G) be integrable function on a locally compact group. The convolution f g is dened by
(f g)(x) :=
f (y)g(y 1 x)dy.
G
In this monograph we consider only unimodular groups. This means that the Haar
measure is both left and rightinvariant. Consequently we also have
(f g)(x) =
f (xy 1 )g(y)dy.
G
On a nilpotent or compact Lie group which is not abelian, the convolution is
not commutative: in general, f g = g f . However, apart from the lack of
commutativity, group convolution and the usual convolution on Rn share many
properties. For example, we have
f g, h =
(f g)(x) h(x) dx
G
f (y) g(y 1 x) h(x) dy dx
=
G
f, h g,
with
g(x) = g(x1 ).
(1.14)
1.5. Convolutions
31
We also have
f g, h
=
G
=
G G
=
G
f (y) g(y 1 x) h(x) dy dx
f (y) g(z) h(yz) dy dz
f (wz 1 ) g(z) h(w) dz dw
g, f h.
(1.15)
With the notation for the operation given by g(x) = g(x1 ), we also have
(f g)= g f.
(1.16)
One can readily check the following simple properties:
if f, g L1 (G) then f g L1 (G), and we have f gL1 f L1 gL1 ;
under the assumptions above, we have
(f g)(x) =
f (y 1 )g(yx)dy =
f (xy)g(y 1 )dy
G
for almost every x G;
if either f or g are continuous on G then f g is continuous on G;
f gL f L2 gL2 ;
the convolution is associative: f (g h) = (f g) h, for f, g, h L1 (G);
the convolution is commutative if and only if G is commutative;
(if G is a Lie group and ) if X is a leftinvariant vector eld, whenever it
makes sense, we have
X(f g) = f (Xg)
and
g) = (Xf
) g;
X(f
moreover, we also have
(Xf ) g = f (Xg);
the right convolution operator f f is leftinvariant; the left convolution
operator f f is rightinvariant.
To check the last statement, let us show that the right convolution operator given
via Af = f is leftinvariant:
f (y) (y 1 z 1 x)dy
L (z)Af (x) = (f )(z 1 x) =
G
f (z 1 y) (y 1 x)dy = (L (z)f ) (x) = AL (z)f (x).
=
G
32
Chapter 1. Preliminaries on Lie groups
Conversely, it follows from the Schwartz integral kernel theorem that if A is leftinvariant, it can be written as a right convolution Af = f , and if A is rightinvariant, it can be written as a left convolution Af = f , see Section 1.4 and
later Corollary 3.2.1.
With our choice of the denition of the convolution and the Fourier transform
in (1.2), one can readily check that for f, g L1 (G), we have
f
g() = g()f()
(1.17)
or, in the other notation,
(f g) = (g)(f ).
We say that an operator A is of weak type (p, p) if there is a constant C > 0
such that for every > 0 we have
{x G : Af (x) > } C
f pLp (G)
p
where {} denotes the Haar measure of a set in G.
Proposition 1.5.1 (Marcinkiewicz interpolation theorem). Let r < q and assume
that operator A is of weak types (r, r) and (q, q). Then A is bounded on Lp (G) for
all r < p < q.
An important fact, the Young inequality, relates convolution to Lp spaces:
Proposition 1.5.2 (Youngs inequality). Suppose
1 p, q, r
and
1
1 1
+ = + 1.
p q
r
If f1 Lp (G) and f2 Lq (G) then f1 f2 Lr (G) and
f1 f2 r f1 p f2 q .
If p, q (1, ) are such that
weakLq (G) condition:
1
p
1
q
> 1, f1 Lp (G), and f2 satises the
sup sq {x : f2 (x) > s} =: f2 qwLq (G) < ,
s>0
then f1 f2 Lr with r as above and
f1 f2 r f1 p f2 wLq (G) .
The proof is an easy adaptation of the Euclidean case which can be found
e.g., in [SW71] or, in the nilpotent case, in [FS82, Proposition 1.18] and [Fol75,
Proposition 1.10].
1.5. Convolutions
33
Convolution of distributions
We now dene the convolution of distributions on a Lie group G. For C (G),
we recall that
(x)
= (x1 )
and
L (x)(y) = (x1 y).
Consequently, we note that
1 y) = (y 1 x).
(L (x))(y)
= (x
It follows that we can write the convolution as
(f g)(x) = f, L (x)
g ,
and hence it make sense to dene
Denition 1.5.3. Let v D (G) and D(G). Then we dene their convolution
as
v, (x
1 ).
(v )(x) := v, L (x)
We also dene
v, (
x1 ),
( v)(x) := v, R (x1 )
where
1 ),
= (yx
R (x1 )(y)
and which is also consistent with the convolution of functions.
We note that this expression makes since since L (x), R (x1 ) and
are continuous mappings from D(G) to D(G).
For example, for the deltadistribution e at the unit element e G, it follows
that
e =
for every D(G),
since we can calculate
= (y 1 x)y=e = (x).
(e )(x) = e , L (x)
The following properties are easy to check using Denition 1.5.3:
if v D (G) and D(G), then v C (G);
if u, v, D(G), then u v, = u, v, in consistency with (1.14).
For v D (G), we now dene v D (G) by
v , := v, .
In particular, if v D (G) and D(G), then v C (G). This shows that
the following convolution of distributions is correctly dened:
34
Chapter 1. Preliminaries on Lie groups
Denition 1.5.4. Let u E (G) and v D (G). Then we dene their convolution
as
u v, := u, v, D(G).
This gives u v D (G) which is consistent with the convolution of functions
in view of (1.15). If we start with a compactly supported distribution v E (G) in
Denition 1.5.3, we arrive at the denition of the composition u v for u D (G)
and v E (G), given by the same formula as in Denition 1.5.4.
A word of caution has to be said about convolution of distributions, namely,
it is not in general associative for distributions, although it is associative for functions.
1.6
Nilpotent Lie groups and algebras
From now on, any Lie algebra g is assumed to be real and nite dimensional.
Proposition 1.6.1. The following are equivalent:
ad is a nilpotent endomorphism over g, i.e.
k N
X g
(adX)k = 0;
the lower central series of g, dened inductively by
g(1) := g,
g(j) := [g, g(j1) ],
(1.18)
terminates at 0 in a nite number of steps.
Denition 1.6.2. (i) If a Lie algebra g satises any of the equivalent conditions
of Proposition 1.6.1, then it is called nilpotent.
(ii) Moreover, if g(s+1) = {0} and g(s) = {0}, then g is said to be nilpotent of
step s.
(iii) A Lie group G is nilpotent (of step s) whenever its Lie algebra is nilpotent
(of step s).
Here are some examples of nilpotent Lie groups and their Lie algebras.
Example 1.6.3. The abelian group Rn equipped with the usual addition is nilpotent. Its Lie algebra is Rn equipped with the trivial Lie bracket.
Example 1.6.4. If no N, the Heisenberg group Hno is the Lie group whose
underlying manifold is R2no +1 and whose law is
1
h1 h2 = x1 + x2 , y1 + y2 , t1 + t2 + (x1 y2 y1 x2 ) ,
2
(1.19)
1.6. Nilpotent Lie groups and algebras
35
for h1 = (x1 , y1 , t1 ) and h2 = (x2 , y2 , t2 ) in Rno Rno R. Here, for vectors
x1 , y1 , x2 , y2 Rno , we denote by x1 y2 and y1 x2 their usual inner products on
Rno .
Its Lie algebra hno is R2no +1 equipped with the Lie bracket given by the
commutator relations of its canonical basis {X1 , . . . , Xno , Y1 , . . . , Yno , T }:
[Xj , Yj ] = T
for j = 1, . . . , no ,
and all the other Lie brackets (apart from those obtained by antisymmetry) are
trivial.
In the case no = 1, we will often simplify the notation and write X, Y, T for
the basis of h1 , etc...
Example 1.6.5. Let Tno be the group of no no matrices which are upper triangular
with 1 on the diagonal. The matrix group Tno is a nilpotent Lie group.
It can be proved that any (connected simply connected) nilpotent Lie group
can be realised as a subgroup of Tno .
Its Lie algebra tno is the space of no no matrices which are upper triangle
with 0 on the diagonal. A basis is {Ei,j , 1 i < j no } where Ei,j is the matrix
with all zero entries except the ith row and jth column which is 1.
Proposition 1.6.6. Let G be a connected simply connected nilpotent Lie group with
Lie algebra g. Then
(a) The exponential map expG is a dieomorphism from g onto G.
(b) If G is identied with g via expG , the group law (x, y) xy is a polynomial
map.
(c) If dg denotes a Lebesgue measure on the vector space g, then dg exp1
G is
a biinvariant Haar measure on G.
This proposition can be found in, e.g. [FS82, Proposition 1.2] or [CG90, Sec.
1.2].
After the choice of a basis {X1 , . . . , Xn } for g, Proposition 1.6.6, Part (a),
implies that the group G is identied with Rn via the exponential mapping; this
means that a point x = (x1 , . . . , xn ) Rn is identied with the point
expG (x1 X1 + . . . + xn Xn )
of the group. Part (b) implies that the law can be written as
x y = (P1 (x, y), P2 (x, y), . . . , Pn (x, y)),
(1.20)
where Pj : Rn Rn R, j = 1, . . . , n, are polynomial mappings given via the
BakerCampbellHausdor formula (see Theorem 1.3.2). Indeed in the nilpotent
case, since ad is nilpotent, the BakerCampbellHausdor formula is nite and
holds for any two elements of the Lie algebra.
36
Chapter 1. Preliminaries on Lie groups
Remark 1.6.7. More is known.
1. Certain choices of bases, namely the socalled JordanH
older or strongMalcev
bases ([Puk67, CG90]), lead to a triangular shaped law, that is,
P1 (x, y)
x 1 + y1 ,
P2 (x, y)
..
.
Pn (x, y)
x2 + y2 + Q2 (x1 , y1 ),
xn + yn + Qn (x1 , . . . , xn1 , y1 , . . . , yn1 ),
with Q1 , . . . , Qn polynomials.
In Chapter 3 we will see that in the particular case of homogeneous Lie
groups, with the choice of the basis made in Section 3.1.3, this fact together
with some additional homogeneous properties is proved in Proposition 3.1.24.
2. The second type of exponential coordinates
Rn (x1 , . . . , xn ) expG (x1 X1 ) . . . expG (xn Xn ) G,
may be used to identify a nilpotent Lie group with Rn after the choice of a
suitable basis as in Part 1.
In the particular case of homogeneous Lie groups, with the choice of the
basis made in Section 3.1.3, this fact together with some additional homogeneous properties is proved in Lemma 3.1.47.
3. The converse of (a) and (b) in Proposition 1.6.6 holds in the following sense:
if a Lie group G can be identied with Rn such that
(a) its law is a polynomial mapping (as in (1.20)),
(b) and for any s, t R, x Rn , the product of the two points sx and tx
is the point (s + t)x,
then the Lie group G is nilpotent [Puk67, Part. II chap. I].
However, we will not use these general facts.
Setting aside the abelian case (Rn , +), we use the multiplicative notation for
the group law of any other connected simply connected nilpotent Lie group G.
The identication of G with g leads to consider the origin 0 as the unit element
(even if the equality xx1 = 0 may look surprising at rst sight). Because of the
BakerCampbellHausdor formula (see Theorem 1.3.2), the inverse of an element
is in fact its opposite, that is, with the notation above,
x1 = (x1 , . . . , xn ).
The identication of G with g allows us to dene objects which usually live
on a vector space, for example the Schwartz class:
1.7. Smooth vectors and innitesimal representations
37
Denition 1.6.8. A Schwartz function f on G is a function f such that f expG is
a Schwartz function on g. We denote by S(G) the class of Schwartz functions. It is
naturally a Frechet space and its dual space is the space of tempered distribution
S (G).
Formally a distribution T D (G) is tempered when T expG is a tempered
distribution on g. The distribution duality is formally given by
f, =
f (x)(x)dx, f S (G), S(G).
G
The Schwartz space and the tempered distributions on a nilpotent homogeneous Lie group will be studied more thoroughly in Section 3.1.9.
1.7
Smooth vectors and innitesimal representations
In this section we describe the basics of the part of the representation theory of
noncompact Lie groups that is relevant to our context. For most statements of
this section we give proofs since understanding of these ideas will be important for
the developments of pseudodierential operators in Chapter 5. Thus, the setting
that we have in mind is that of nilpotent Lie groups, although we do not need to
make this assumption for the following discussion. For the general representation
theory of locally compact groups we can refer to, for example, the books of Knapp
[Kna01], Wallach [Wal92, Chapter 14] or Folland [Fol95].
Let us rst recall some basic denitions about dierentiability of a Banach
spacevalued function.
Denition 1.7.1. Let f be a function from on open subset of Rn to a Banach
space B with norm  B .
The function f is said to be dierentiable at xo if there exists a (necessarily unique) linear map f (xo ) : Rn B such that
1
f (x) f (xo ) f (xo )(x xo )B 0.
xxo
x xo Rn
We call f (xo ) the dierential of f at xo .
If f is dierentiable at each point of , then x f (x) is a function from
to the Banach space L (Rn , B) of linear mappings from Rn to B (recall that
linear mappings from Rn to B are automatically bounded.) We say that f is of
class C 1 if x f (x) is continuous, and that f is of class C 2 if x f (x) is of
class C 1 and so on. We say that f is of class C if f is of class C k for all k N.
These denitions extend to any open set of any smooth manifold.
As in the case of functions valued in a nite dimensional Euclidean space, we
have the basic properties for a function f as in Denition 1.7.1:
38
Chapter 1. Preliminaries on Lie groups
The function f is of class C k if and only if all of its partial derivatives of
order 1, 2, . . . , k exist and are continuous.
The chain rule holds for a composition f h where h is a mapping from an
open subset of a nite dimensional Euclidean space into .
We can now dene the smooth vectors of a representation.
Denition 1.7.2. Let G be a Lie group and let be a representation of G on a
Hilbert space H . A vector v H is said to be smooth or of type C if the
function
G x (x)v H
is of class C .
We denote by H the space of all smooth vectors of .
The following is a necessary preparation to introduce the notion of the innitesimal representation and of the operator d(X). This will be of fundamental
importance in the sequel.
Proposition 1.7.3. Let G be a Lie group with Lie algebra g. Let be a strongly
continuous representation of G on a Hilbert space H . Then for any X g and
v H , the limit
1
lim ((expG (tX))v v)
t0 t
exists in the norm topology of H and is denoted by d(X)v. Each d(X) leaves
H invariant, and d is a representation of g on H satisfying
X, Y g
d(X)d(Y ) d(Y )d(X) d ([X, Y ]) = 0.
Consequently, d extends to a representation of the Lie algebra U(g) on
d(0) = 0 and d(1) = 0.
(1.21)
H
with
Recalling the derivative with respect to X in (1.7), we may formally abbreviate writing
d(X)v = X((x)v)x=e
or even
d(X) = X(e).
(1.22)
Sketch of the proof of Proposition 1.7.3. Let v H . The function f : g H
dened by f (X) := (exp X)v is of class C , and for any X g we have
f (0)(X) = lim
t0
1
((expG (tX))v v) .
t
By denition f (0)(X) = d(X).
Since is continuous we have, using the identication of g with the space of
leftinvariant vector elds,
(x)d(X)v
=
=
1
lim (x) ((expG (tX))v v)
t
1
lim ((x expG (tX))v (x)v) = XF (x),
t0 t
t0
1.7. Smooth vectors and innitesimal representations
39
where F : G H is the function dened by F (x) := (x)v. By assumption F is
of type C thus x XF (x) is also of type C and the equality above says that
d(X)v is smooth. Hence d(X) leaves H stable. Consequently X d(X) can
be extended to an algebra homomorphism U(g) H as in the statement.
It remains to prove (1.21), i.e. that
X, Y g
d(X)d(Y ) d(Y )d(X) d ([X, Y ]) = 0.
We x X, Y g and dene a path c by
1
1
1
1
c(t) := expG (sgnt)t 2 X expG t 2 Y expG (sgnt)t 2 X expG t 2 Y .
Clearly c is dened on a neighbourhood of 0 in R and valued in G, and is of class
C 1 with c (0) = [X, Y ]. Let v H . By the chain rule the map t (c(t))v has
dierential F (e)([X, Y ]) at t = 0, where F is F (x) = (x)v as above and e is the
neutral element. Thus
1
1
d([X, Y ]) = lim ((c(t))v v) = lim 2 (c(t2 ))v v .
t0 t
t0 t
The strong continuity of implies then
lim
t0 t2
((expG (tX) expG (tY ))v (expG (tY ) expG (tX))v)
= lim (expG (tY ) expG (tX))
t0
= d([X, Y ])v.
1
(c(t2 ))v v
2
t
(1.23)
But we can also compute
(d(X)d(Y )v, u) = s=0 t=0 ((expG (sX) expG (tY ))v, u)
1
= lim ( 2 {(expG (tX) expG (tY )) (expG (tX)) (expG (tY )) + I} v, u).
t0 t
Interchanging X and Y and subtracting we nd
((d(X)d(Y ) d(Y )d(X))v, u)
(1.24)
1
= lim ( 2 {(expG (tX) expG (tY )) (expG (tY ) expG (tX))} v, u).
t0 t
Comparing this with (1.23), we obtain (1.21). This concludes the proof of
Proposition 1.7.3.
Denition 1.7.4. Let G be a Lie group with Lie algebra g and let be a strongly
continuous representation of G on a Hilbert space H . The representation d
dened in Proposition 1.7.3 is called the innitesimal representation associated
to . We will often denote it also by . Consequently, for T U(g) or for its
corresponding leftinvariant dierential operator, we write
(T ) := d(T ).
40
Chapter 1. Preliminaries on Lie groups
Example 1.7.5. For example, the innitesimal representation of Ad is ad, see Section 1.3.
We now collect some properties of the innitesimal representations.
Proposition 1.7.6. Let G be a Lie group with Lie algebra g and let be a strongly
continuous unitary representation of G on a Hilbert space H . Then we have the
following properties.
(i) For the innitesimal representation d of g on H each d(X) for X g is
skewhermitian: d(X) = d(X).
(ii) The space H of smooth vectors is invariant under (x) for every x G,
and
D U(g) v H
(x)d(D)(x)1 v = d(Ad(x)D)v.
(iii) If S is a vector subspace of H such that for all v S and X g, the limits
of t1 {(expG (tX))v v} as t 0 exist, then S H .
(iv) Let D(G). For any X g, viewed as a leftinvariant vector eld,
v H
()v H
and
d(X)()v = (X)v,
and viewing X as a rightinvariant vector eld X,
v H
()d(X)v = (X)v.
If G is a connected simply connected nilpotent Lie group, one can replace
D(G) by the Schwartz space S(G).
Proof. Let us prove Part (i). Let u, v H . The unitarity of implies
i
i
((expG (tX))v v) , u .
v, ((expG (tX))u u) =
t
t
By denition of d(X)u and d(X)v, the limits as t 0 of the left and right
hand sides are (v, id(X)u) and (id(X)v, u), respectively. Hence they are equal
and d(X) is skewhermitian. This proves Part (i).
For (ii), we rst observe that the map x (x)(xo )v is the composition of
x xxo and x (x)v. Hence H is an invariant subspace for (xo ).
Now let X g, x G and v H . Then we compute easily
1
((expG (tX)) I) (x)1 v
t
=
=
1
(x expG (tX)x1 ) I v
t
1
(x)1 ((expG (Ad(x)(tX)) I) v.
t
(x)1
1.7. Smooth vectors and innitesimal representations
41
Passing to the limit as t 0, we obtain
d(X)(x)1 v = (x)1 d(Ad(x)(tX))v.
Hence
(x)d(X)(x)1 = d(Ad(x)(tX))
on H . Using Proposition 1.7.3, we obtain a similar property for D U(g) instead
of X. This shows (ii).
For (iii), by assumption for v S the map Fv : G x (x)v is dierentiable at the neutral element e, the partial derivative in the X g direction
being
1
XFv (e) = lim {(expG (tX))v v} .
t0 t
More generally, since is strongly continuous, we have for any x G,
1
1
(x)XFv (e) = lim (x) {(expG (tX))v v} = lim {Fv (x expG (tX)) Fv (x)} .
t0 t
t0 t
Thus Fv is also dierentiable at x G and
XFv (x) = (x)XFv (e)
for any X g. This shows that the rst derivatives of Fv are continuous, thus Fv
must be of class C 1 . Furthermore,
Fv (x)(X) = (x)XFv (e).
If Fv is of class C k for k N, then the map x XFv (x) = (x)XFv (e) is of class
C k and Fv must be of class C k+1 . Inductively this shows that Fv is of type C .
This shows Part (iii).
For (iv), for any L1 (G) and x G, recalling (1.3), we have
(x)() = (( x)).
Hence for any D(G), v H and X g,
1
((expG (tX))()v ()v) =
t
( expG (tX))
t
v.
This last expression tends to (X)v as t 0. Applying (iii) to S = ()H , we
see that S H . We also have
d(X)()v = (X)v.
For the rightinvariant case, again by (1.3), we have
()(x) = ((x ))
42
Chapter 1. Preliminaries on Lie groups
for any L1 (G) and x G. Hence for any D(G), v H and X g,
(expG (tX) )
1
(()(expG (tX))v ()v) =
v.
t
t
This last expression tends to (X)v
as t 0 while the lefthand side tends to
()d(X)v if v H . This proves Part (iv) in the general case. The changes for
G connected simply connected nilpotent Lie group, and to replace D(G) by S(G)
are straightforward. This concludes the proof of Proposition 1.7.6.
In the following proposition, we show that the space of smooth vectors is
dense in the space of a strongly continuous representation. The argument is famously due to G
arding.
Proposition 1.7.7. Let G be a Lie group and let be a strongly continuous representation of G on a Hilbert space H . Then the subspace H of smooth vectors is
dense in H .
Proof. Let v H and > 0 be given. Since is strongly continuous, the set
:= {x G : (x) v vH < }
is open. We can nd a nonnegative function D(G) supported in satisfying
(x)dx = 1. Then
G
()v vH = (x)((x) v v)dx
(x)(x) v vH dx
(x)dx = .
By Proposition 1.7.6, we know that ()v is a smooth vector. This shows that
H is dense in H .
In the proof above, we have in fact showed that the vectors ()v for v H
and D(G) form a dense subspace of H . If G is nilpotent connected simply
connected, the same property holds with S(G). The nite linear combinations
of those vectors form a subspace called the G
arding subspace, which is included in
H by Proposition 1.7.6 (iv).
It turns out that the G
arding subspace is not only included in the subspace
H but is in fact equal to H . This is a consequent of the following theorem, due
to Dixmier and Malliavin [DM78]:
Theorem 1.7.8 (DixmierMalliavin). Let G be a Lie group and let be a strongly
continuous representation of G on a Hilbert space H .
The space H of smooth vectors is spanned by all the vectors of the form
()v for v H and D(G). This means that any smooth vector can be
written as a nite linear combination of vectors of the form ()v.
If G is a connected simply connected nilpotent Lie group, one can replace
D(G) by the Schwartz space S(G).
1.8. Plancherel theorem
1.8
43
Plancherel theorem
Here we discuss the Plancherel theorem for locally compact groups and for the
special case of nilpotent Lie groups. Our presentation will be rather informal.
One reason is that we decided not to present here in full detail the orbit method
yielding the representations of the nilpotent Lie groups but to limit ourselves only
to its consequences useful for our subsequent analysis. The reason behind this
choice is that it could take quite much space to prove the general results for the
orbit method and would lead us too much away from our main exposition also
risking overwhelming the reader with technical discussions somewhat irrelevant
for our purposes. In general, this subject is wellknown and we can refer to books
by Kirillov [Kir04] or by Corwin and Greenleaf [CG90] for excellent expositions
of this topic. The same reasoning applies to the abstract Plancherel theorem: it is
known in a much more general form, due to e.g. Dixmier [Dix77, Dix81], and we
will limit ourselves to describing its implications for nilpotent Lie groups relevant
to our subsequent work.
As we will see in Chapter 2, all the results of the abstract Plancherel theorem
in the case of compact groups can be recaptured there thanks to the PeterWeyl
theorem (see Theorem 2.1.1). However, for nilpotent Lie groups, even if the orbit
method provides a description of the dual of the group and of the Plancherel
measure, in our analysis we will need to use the properties of the von Neumann
algebra of the group provided by the general abstract Plancherel theorem. This
will replace the use of the Fourier coecients in the compact case.
Before we proceed, let us adopt two useful conventions. First, the set of all
strongly continuous unitary irreducible representations of a locally compact group
G will be denoted by Rep G, i.e.
Rep G = {all strongly continuous unitary irreducible representations of G}.
We have
The equivalence of representations in Rep G leads to the unitary dual G.
meaning that the expressions, when dealing with
already agreed to write G
Fourier transforms, may depend on as described in Remark 1.1.5. However,
in this section we will sometimes want to show that certain expressions do not
depend on the equivalence class of , and for this purpose we will be sometimes
distinguishing between the sets Rep G and G.
The second useful convention that we will widely use especially in Chapter
5 is that we may denote the Fourier transform in three ways, namely, we have
()
() FG ()().
Although this may seem as too much notation for the same object, the reason for
this is twofold. Firstly, the notation () is widely adopted in the representation
44
Chapter 1. Preliminaries on Lie groups
theory of C algebra associated with groups. Secondly, it becomes handy for longer
expressions as well as for expressing properties like
(T ) = (T )()
where (T ) is the innitesimal representation given in Denition 1.7.4. The no
tation ()
is useful as an analogy for the Euclidean case and will be extensively
used in the case of compact groups. When we want to write the Fourier transform
as a mapping between dierent spaces, the notation FG becomes useful.
1.8.1 Orbit method
In this section we briey discuss the idea of the orbit method and its implications
for our analysis. In general, we will not use the orbit method by itself in our
analysis, but only the existence of a Plancherel measure and some Fourier analysis
similar to the compact case as described in Section 2.1.
Let G be a connected, simply connected, nilpotent Lie group with Lie algebra
g. The orbit method describes a way to associate to a given linear functional on
g a collection of unitary irreducible representations of G which are all unitarily
equivalent between themselves. Consequently, to any element of the dual g of g,
one can associate an equivalence class of unitary irreducible representations. It
turns out that any such class is realised in this way. Furthermore, two elements
f1 , f2 g lead to the same class if and only if the two elements are in the same
orbit under the natural action of G on g ; this natural action is the socalled coadjoint representation: since the group G acts on g by the adjoint representation
Ad, it also acts on its dual g by
coAd : G g (g, f ) f (Ad1 g ) g .
This gives a onetoone correspondence between
of the group, that is, the collection of unitary
on the one hand, the dual G
irreducible representations modulo unitary equivalence, and
on the other hand, g /coAd(G), that is, the set of coadjoint orbits.
Example 1.8.1. In the case of the Heisenberg group Hno presented in Example
1.6.4, a family of representatives of all coadjoint orbits is
1. either of the form T if R\{0},
no
xj Xj + yj Yj with xj , yj R,
2. or of the form j=1
where {X1 , . . . , Xn o , Y1 , . . . , Yn o , T } is the dual basis to the canonical basis of bno
given inExample 1.6.4. To T is associated the Schrodinger representation ,
no
xj Xj +yj Yj is associated the 1dimensional representation (x, y, t)
and to j=1
exp i(xx + yy ) , where xx and yy denote the canonical scalar product on Rn .
See Section 6.2.
1.8. Plancherel theorem
45
As for Schrodinger representations, the representations constructed via the
may be
orbit method can be realised as acting on some L2 (Rm ) and the dual G
identied with g /coAd(G), or even with suitable representatives of this quotient.
is concretely described as
Thus, by the orbit method the unitary dual G
a subset of some Euclidean space. It is then possible to construct explicitly a
such that we have the Fourier inversion theorem (where we recall
measure on G
once more the notation and conventions described in the beginning of Section 1.8):
Theorem 1.8.2. Let G be a connected simply connected nilpotent Lie group. The
is then equipped with a measure called the Plancherel measure satisfying
dual G
the following property for any S(G).
The operator () ()
is trace class for any strongly continuous unitary
irreducible representation Rep G, and Tr(()) depends only on the class of ;
Tr (()) is integrable against and the following formula
the function G
holds:
Tr (()) d().
(1.25)
(0) =
G
For the explicit expression of the Plancherel measure , see, e.g., [CG90,
Theorem 4.3.9].
Applying formula (1.25) to () = f ( x) and using () = (x)(f ) in view
of (1.3), we obtain:
Corollary 1.8.3 (Fourier inversion formula). Let G be a connected simply connected
nilpotent Lie group and let be the Plancherel measure on G.
If f S(G), then (x)(f ) and (f )(x) are trace class for every x G,
Tr ((x)(f )) is integrable against , and we have
the function G
Tr ((x)(f )) d() =
Tr ((f )(x)) d().
(1.26)
f (x) =
G
G
The latter equality can be seen by the same argument as above, applied to
the function f (x ).
Example 1.8.4. In the case of the Heisenberg group Hno , the Plancherel measure
is given by integration over R\{0} against cn0 no d, with a suitable constant cno
(depending on normalisations):
Tr () no d.
(0) = cno
R\{0}
An orthonormal basis for H = L2 (Rno ) is given by the Hermite functions. The
formed by the 1dimensional representations is negligible with respect
subset of G
to the Plancherel measure. We refer to Section 6.2.3 for a more detailed discussion
as well as for the constant cno .
46
Chapter 1. Preliminaries on Lie groups
1 ), one obApplying the inversion formula to ( ), where (x) = (x
tains:
Theorem 1.8.5 (Plancherel formula). We keep the notation of Theorem 1.8.2. Let
S(G). Then the operator () is HilbertSchmidt, that is,
()HS = Tr (()() ) <
2
for any Rep G, and its HilbertSchmidt norm is constant on the equivalence
()2 is integrable against and
class of . The function G
HS
2
(x)2 dx =
()HS d().
(1.27)
G
Formula (1.27) can be extended unitarily to hold for any L2 (G), permitting the denition of the group Fourier transform of a square integrable function
on G.
Applying the inversion formula to ( ), or bilinearising the Plancherel
formula, we also obtain:
Corollary 1.8.6. Let , S(G). Then the operator ()() is trace class for
any Rep G, and its trace is constant on the equivalence class of . The function
Tr (()() ) is integrable against and
G
(, )L2 (G) =
(x)(x)dx =
Tr (()() ) d().
G
G
1.8.2 Plancherel theorem and group von Neumann algebras
In this section we describe the concept of the group von Neumann algebra that becomes handy in associating symbols with convolution kernels of invariant operators
on G. For the details of the constructions described below we refer to Dixmiers
books [Dix77, Dix81] and to Section B in the appendix of this monograph. For
the Plancherel theorem on locally compact groups with emphasis on the decomposition of reducible representations in continuous Hilbert sums, see also Bruhat
[Bru68]. A more extensive discussion of this subject is given in Appendix B.2,
more precisely in Section B.2.5. An abstract version of the Plancherel theorem is
also given in the appendix in Theorem B.2.32.
Our framework
The representation theory of a general locally compact group may be very wild.
However, in favourable cases most of the traditional Fourier analysis on compact
Lie groups (described in Section 2.1) remains valid under natural modications;
for instance, the sum over the discrete dual in the compact case is replaced by an
integral. By favourable cases we mean the following hypothesis:
1.8. Plancherel theorem
47
(H) The group G is separable locally compact,
unimodular, and of type I.
(See e.g. Dixmier [Dix77]). For our purpose, it suces to know that any Lie group
is a standard
which is either compact or nilpotent satises (H). Its unitary dual G
Borel space.
We will now present the abstract Plancherel theorem as obtained by Dixmier
in [Dix77, 18.8] and stated in Theorem B.2.32. Here, we will formulate it neither
in its logical order with the viewpoint of proving its statement nor in its full
generality since this would require introducing a lot of additional notation. Instead,
we present its consequences applicable to our setting, starting with the existence
of the Plancherel measure.
The Plancherel formula
We start by describing the part of the Plancherel theorem dealing with the Plan
is a bounded operator
cherel formula. First if Cc (G) and Rep G, then ()
on H (as the group Fourier transform of an integrable function) and one checks
easily that its HilbertSchmidt norm is constant on the class of Rep G in
Hence ()
G.
HS(H ) may be viewed as depending on G. The Plancherel
formula states that there exists a unique positive nite measure , called the
Plancherel measure, such that for any Cc (G) we have
2
2
(x) dx =
d().
(1.28)
()
G
G
HS(H )
In the compact or nilpotent case, the Plancherel measure can be described explicitly via the PeterWeyl Theorem (see Theorem 2.1.1) or the orbit method (see
Theorem 1.8.5), respectively.
The Plancherel formula in (1.28) may be reformulated in the following (more
precise) way. The group Fourier transform is an isometry from Cc (G) endowed
with the L2 (G)norm to the Hilbert space
:=
L2 (G)
HS(H )d().
(1.29)
G
is dened (see Section B.1 or, e.g., [Dix81, Part II ch. I])
Hence the space L2 (G)
as the space of measurable elds of HilbertSchmidt operators { HS(H ) :
which are square integrable in the sense that
G}
2
2HS d() < .
L2 (G)
:=
G
Here we use the usual identications of a strongly continuous irreducible unitary
and of a eld of operarepresentation from Rep G with its equivalence class in G,
tors on G with its equivalence class with respect to the Plancherel measure . One
48
Chapter 1. Preliminaries on Lie groups
can check that indeed, the properties above do not depend on a particular representative of and of the eld of operators. The Plancherel formula implies that FG
extends to an isometry on L2 (G). We keep the same notation FG for this map, allowing us to consider the Fourier transform of a square integrable function. The ab
stract Plancherel theorem states moreover that the isometry FG : L2 (G) L2 (G)
2
2
is surjective. In other words, FG maps L (G) onto L (G) isometrically.
Note that for any , L2 (G), the operator () () is trace class on H
for almost all Rep G with
Tr () ()  ()HS(H ) () HS(H ) = ()HS(H ) ()HS(H ) ,
and that Tr () ()  and Tr (() () ) are constant on the class of
Thus these traces can be viewed as being parametrised by G.
Rep G in G.
The bilinearisation of the Plancherel formula yields
(x)(x)dx =
Tr (() () ) d().
(1.30)
G
One also checks easily, for example by density of Cc (G) in L2 (G), that Formula (1.17), that is,
f
g() = g()f()
(1.31)
or, in the other notation,
(f g) = (g)(f ),
remains valid for f L1 (G) and g L2 (G) and also for f L2 (G) and g L1 (G).
We now present the parts of the Plancherel theorem (relevant for our subsequent analysis) regarding the description of the group von Neumann algebra.
Group von Neumann algebra
In this monograph, we realise the von Neumann algebra of a group G as the algebra
denoted by LL (L2 (G)) and dened as follows.
Denition 1.8.7. Let L (L2 (G)) denote the set of bounded linear operators L2 (G)
L2 (G), and let LL (L2 (G)) be the subset formed by the operators in L (L2 (G))
which are leftinvariant (in the sense of Denition 1.1.3).
Endowed with the operator norm and composition of operators, one checks
easily that LL (L2 (G)) is a von Neumann algebra, see Section B.2.5 for the exposition of its general ideas.
Given a measurable eld of uniformly bounded operators = { }, the
operator T LL (L2 (G)) dened via
T
() = (),
L2 (G),
(1.32)
1.8. Plancherel theorem
49
is in LL (L2 (G)). Using (1.30), this yields that the operator T : S(G) S (G)
can also be dened by
Tr ( () () ) d(), , L2 (G).
(1.33)
(T , )L2 (G) =
G
to LL (L2 (G)) where the space
This denes a map T from L (G)
is dened by
L (G)
denote the space of measurable elds on G
of uniDenition 1.8.8. Let L (G)
formly bounded operators = { L (H ), G}, that is,
sup L (H ) < .
(1.34)
G
Here we use the usual identications of a strongly continuous irreducible
and of a eld of
unitary representation from Rep G with its equivalence class in G,
with its equivalence class with respect to the Plancherel measure
operators on G
does not depend on a particular
. One can check that indeed, being in L (G)
representative of and of the eld of operators. In (1.34), the supremum is to be
understood as the essential supremum with respect to the Plancherel measure .
with the pointwise composition given by
We endow L (G)
:= { , G},
for
= { , G}
L (G),
= { , G},
and the essential supremum norm
L (G)
:= sup L (H ) .
(1.35)
G
We may sometimes abuse the notation and write L (G)
when no confusion is possible.
is a von Neumann algebra and that the map
One checks easily that L (G)
T LL (L2 (G)),
L (G)
is a morphism of von Neumann algebras. The Plancherel theorem implies that
this map is in fact a bijection and an isometry, and hence a von Neumann algebra
isomorphism. More precisely it yields that for any T LL (L2 (G)), there exists
(T )
a measurable eld of uniformly bounded operators { } such that for any
(T
)
2
L (G) the HilbertSchmidt operators T() and f() are equal almost
(T )
everywhere; the eld { } is unique up to a negligible set.
Note that by the Schwartz kernel theorem (see Corollary 3.2.1), an operator
T LL (L2 (G)) is of convolution type with kernel D (G),
T f = f ,
f D(G).
50
Chapter 1. Preliminaries on Lie groups
If D (G) is such that the corresponding convolution operator D(G) f f
extends to a bounded operator T on L2 (G) then T LL (L2 (G)) and we extend
the denition of the group Fourier transform by setting
().
(T ) := ()
(1.36)
We denote by K(G) the set of such distributions :
Denition 1.8.9. Let K(G) denote the space of distributions D (G) such that
the corresponding convolution operator
D(G) f f
extends to a bounded operator on L2 (G).
If G is a connected simply connected nilpotent Lie group, the Schwartz kernel
theorem (see Corollary 3.2.1), implies in fact that the distributions in K(G) are
tempered, i.e. K(G) S (G).
(x1 ) is also in K(G). If 1 , 2
If K(G), then dened via (x) =
2
K(G) and T1 , T2 LL (L (G)) denote the associated rightconvolution operator, then T1 T2 LL (L2 (G)) and we denote by 2 1 its convolution kernel.
One checks easily that this convolution product coincides or extends the already
dened convolution products in Section 1.5. Furthermore K(G) equipped with this
convolution product, the adjoint and the operator norm
K(G) := f f L (L2 (G))
(1.37)
is a von Neumann algebra. It is naturally isomorphic to LL (L2 (G)).
The part of the Plancherel theorem that we have already presented implies
that the space K(G) is a von Neumann algebra isomorphic to LL (L2 (G)) and to
Moreover, the group Fourier transform dened on K(G) gives the isomorL (G).
phism between K(G) and L (G).
Naturally, L1 (G) is embedded in K(G) since if L1 (G), then the operator
is in LL (L2 (G)). Note that Youngs inequality (see Proposition 1.5.2)
implies
(1.38)
L (G)
= K L1 (G) .
Furthermore, as FG ( ) =
(see e.g. (1.31)), there is no conict of notation
between the group Fourier transforms dened rst on L1 (G) via (1.2) and then
on K(G) in (1.36) as these group Fourier transforms coincide, since the eld of
operators associated to an operator in LL (L2 (G)) is unique.
More generally, the proof of Example 1.8.10 below shows that the space of
complex Borel measures M (G) (which contains L1 (G)) is contained in K(G), that
is,
L1 (G) M (G) K(G).
Moreover, their group Fourier transform may be dened directly via (1.39) below
or as of an element of K(G) via Denition 1.36.
1.8. Plancherel theorem
51
Example 1.8.10 (Complex Borel measures). Any complex Borel measure on G
is in K(G) and
K M (G) ,
where M (G) denotes the total mass of .
The group Fourier transform of a complex Borel measure is given in the
sense of Bochner by the integral
(x) d(x).
(1.39)
FG ()() () () :=
G
In particular, the group Fourier transform of the Dirac measure e at the neutral
element is the identity operator
e () (e ) = IH
on the representation space H . More generally, the group Fourier transform of
the Dirac measure xo at the element xo G is
xo () = (xo ).
Proof of Example 1.8.10. By Jensens inequality, for p = 1 and 2 (in fact for any
p [1, )), the operator T : D(G) extends to an Lp bounded
operator with norm .
If Cc (G), then L1 (G) (see Example 1.8.10) and we have in the
sense of Bochner, using the change of variable y = xz 1 ,
1
( ) =
(xz )(x) dxd(z) =
(y)(yz) dyd(z)
GG
GG
=
(y)(z) (y) dyd(z) =
(z) d(z)
(y)(y) dy
GG
() (),
conrming the formula for (). Since the eld of operators associated to an
operator in LL (L2 (G)) is unique, the group Fourier transform of as an element
dened in (1.39).
of K(G) is {(), G}
The abstract Plancherel theorem
We now summarise the consequences of Dixmiers abstract Plancherel theorem,
see Theorem B.2.32, that we will use:
Theorem 1.8.11 (Abstract Plancherel theorem). Let G be a Lie group satisfying
hypothesis (H). We denote by its Plancherel measure.
The Fourier transform FG extends to an isometry from L2 (G) onto
:=
L2 (G)
HS(H )d().
G
52
Chapter 1. Preliminaries on Lie groups
The Fourier transform of an element f of K(G), i.e. f D (G) such that
the operator D(G) f extends boundedly to L2 (G), has a meaning as a
eld of uniformly (essentially) bounded operators
L (G)
{f() (f ) : G}
satisfying
( f ) = (f )()
Conversely, any eld in L (G)
leads to an element
for any D(G) and G.
of K(G). Furthermore
f K = f L (L2 (G)) = sup f()L (H ) .
(1.40)
G
The Fourier transform is a von Neumann algebra isomorphism from K(G)
In particular, it is a bijection from K(G) onto L (G)
and satises
onto L (G).
f1 , f2 , f K(G)
FG (f1 f2 ) = FG (f2 )FG (f1 )
and
FG (f ) = FG (f ) ,
if f (x) = f(x1 ). Moreover
fL (G)
= f K(G) .
If G is a connected simply connected nilpotent Lie group, the elements of
K(G) are tempered distributions.
Naturally the various denitions of group Fourier transforms on L1 (G) or on
the space M (G) of regular complex measures on G, on L2 (G) or on K(G), coincide
on any intersection of these subspaces of D (G). This can be seen easily using the
and
abstract Plancherel theorem, especially the bijections FG : L2 (G) L2 (G)
FG : K(G) L (G), together with the properties of the convolution and of the
representations, especially (1.31).
1.8.3 Fields of operators acting on smooth vectors
Let us assume that the group G satises hypothesis (H) as in the previous section
and is also a Lie group. This means that G is a unimodular Lie group of type I,
for instance a compact or nilpotent Lie group.
In our subsequent analysis, we will need to consider elds of operators para but not necessarily bounded, for instance the elds given by the
metrised by G
(X) s.
The denition of elds of smooth vectors or of operators dened on smooth
vectors will be a consequence of the following lemma. For a more general setting
for measurable elds of operators see Section B.1.5.
1.8. Plancherel theorem
53
Lemma 1.8.12. Let 1 , 2 Rep G with 1 T 2 , that is, we assume that 1 and
2 are intertwined by the unitary operator T , i.e. T 1 = 2 T . Then T maps H1
onto H2 bijectively.
Proof. This is an easy consequence of the DixmierMalliavin theorem, see Theorem
1.7.8.
Lemma 1.8.12 allows us to dene elds of operators not necessarily bounded
but just dened on smooth vectors:
Denition 1.8.13. A Geld
of operators dened on smooth vectors is a family of
where
classes of operators { , G}
:= {1 : H1 H1 , 1 }
viewed as a subset of Rep G, satisfying for any two elements
for each G
1
and 2 in :
1 T 2 = 2 T = T 1 .
It is measurable when for one (and then any) choice of realisation 1 and
any vector x1 H1 , as
runs2 over G, the resulting eld {1 x1 , G} is
measurable whenever G x1 H d() < .
1
We will allow ourselves the shorthand notation
= { : H H , G}
to indicate that the Geld
of operators is dened on smooth vectors. Unless
otherwise stated, all the Gelds
of operators are assumed to be measurable and
with operators dened on smooth vectors. We may allow ourselves to write =
Note that we do not require the domain of each operator to be the
{ , G}.
whole representation space H1 but just the space of smooth vectors.
The next denition would allow us to compose such elds of operators.
Denition 1.8.14. A measurable Geld
of operators acting on the smooth vectors
such that for
is a measurable Geld of operators = { : H H , G}
any 1 Rep G, we have
1 (H1 ) H1 .
We will often abuse the notation and write
{ : H H , G}
to express the fact that the measurable Geld
of operators act on smooth vectors.
be a Geld.
Remark 1.8.15. Let = { : H H , G}
If 1 T 2 that
is, we assume that 1 and 2 are intertwined by the unitary operator T , then T
maps 1 (H1 ) onto 2 (H2 ) bijectively. Thus the range (H ) makes sense as
the collection of the equivariant ranges 1 (H ) for 1 Rep G.
Consequently, in Denition 1.8.14, it suces that 1 (H1 ) H1 for one
representation 1 for each G.
54
Chapter 1. Preliminaries on Lie groups
Remark 1.8.16. We will often consider measurable eld of operators ,s acting on
but also by another set S. When
smooth vectors and parametrised not only by G
n
this set S is a subset of some R , we say that this parametrisation is smooth
whenever the map appearing in Denition 1.8.14 above is not only measurable
but also smooth with respect to the Svariable. Note that this
with respect to G
hypothesis yields the existence of the elds of operators given by Ds ,s where Ds
is a (smooth) dierential operator on S.
and
It is clear that one can sum two elds = { : H H , G}
dened on smooth vectors. We may then write
= { : H H , G}
+ = { + : H H , G}
for the resulting eld. If and act on smooth vectors, then so does + .
It is also clear that one can compose two elds = { : H H , G}
and = { : H H , G} dened on smooth vectors if the rst one acts
on smooth vectors. We may then write
= { : H H , G}
for the resulting eld which is then dened on smooth vectors. Note that is not
obtained as the composition of two unbounded operators on H as in Denition
A.3.2 but as the composition of two operators acting on the same space H .
Almost by denition of smooth vectors, we have the following example of
measurable elds of operators acting on smooth vectors:
yields a measurable eld of
Example 1.8.17. If T U(g) then {(T ), G}
(see also Proposition
operators acting on smooth vectors and parametrised by G
1.7.3).
with {(T2 ), G}
If T1 , T2 U(g) then the composition of {(T1 ), G}
as eld of operators acting on smooth vectors is {(T1 T2 ), G}.
The denition of Fourier transform and Proposition 1.7.6 (iv) easily imply
the next example of measurable elds of operators acting on smooth vectors:
is a measurExample 1.8.18. If D(G), then = {() : H H , G}
able Geld of operators acting on smooth vectors.
If 1 , 2 D(G), then the composition of 1 with 2 as elds of operators
acting on smooth vectors is
2 1 .
If G is simply connected and nilpotent, the properties above also hold for
Schwartz functions.
always gives by restriction operators
A eld = { : H H , G}
that are dened on smooth vectors. If we start from a eld of operators = { :
dened on smooth vectors, we can not always extend it to
H H , G}
1.8. Plancherel theorem
55
operators dened on every H . However, since the space H of smooth vectors is
dense in H (see Proposition 1.7.7), each operator 1 : H1 H1 , 1 Rep G,
has a unique extension to a bounded operator on H1 provided that such an
extension exists. In this case, 2 would have the same property if 1 2 , and
the operator norm 1 L (H1 ) or the HilbertSchmidt norm 1 HS(H1 ) of 1
are constant (maybe innite) for 1 . Hence we may regard these norms as
Furthermore, if or 2 are nite,
being parametrised by G.
L (G)
L (G)
or L2 (G)
(resp.) is unique and
then the eld of bounded operators in L (G)
extends .
On a compact Lie group, any Geld
of operators is measurable and the
is discrete and
operators act on smooth vectors. This is because in this case G
countable, and all the strongly continuous irreducible representations are nite
dimensional and these have only smooth vectors, see the PeterWeyl theorem in
Theorem 2.1.1.
However on a noncompact Lie group, we can not restrict ourselves to the
case of Gelds
acting on smooth vectors in general since a noncompact Lie group
may have innite dimensional (strongly continuous irreducible) representations
which do not act on
with nonsmooth vectors and we then can nd elds in L2 (G)
smooth vectors. Indeed, in this case,
we can nd a measurable eld {v , G}
2
of nonsmooth vectors satisfying G v HS(H ) d() < , and then construct the
in L2 (G)
which does not act on smooth vectors.
eld of operators {v v , G}
Such eld of vectors {v } are easy to nd for instance on the Heisenberg group
Hn whose case is detailed in Chapter 6: in this case, almost all the representations
n may be realised on L2 (Rn ) and the space of smooth vectors then coincides
in H
with the Schwartz space S(Rn ), see Section 6.2.1.
We can give a sucient condition for a eld to act on smooth vectors:
Lemma 1.8.19. Let = { : H H } be a eld dened on smooth vectors. If
for each D(G), is a eld of operators acting on smooth vectors, that is,
= { () : H H },
then acts on smooth vectors.
Proof. Let us assume that is a eld of operators acting on smooth vectors
realised as a representation and each
for every D(G). Then, for each G
is smooth. By the DixmierMalliavin Theorem,
smooth vector v H , ()v
see Theorem 1.7.8. the nite linear combination of the vectors of the form ()v
form H . Therefore : H H , and the statement is proved.
As an application of Lemma 1.8.19, we see that the eld
xo given at the end
of Example 1.8.10 acts on smooth vectors:
Example 1.8.20. For any xo G, the eld
xo = {(xo ) : H H } L (G)
acts on smooth vectors.
56
Chapter 1. Preliminaries on Lie groups
xo )() and
Proof. Let xo G. If D(G), then by (1.4), (xo )() = (
( xo ) D(G). Thus for any v H , (xo )()v is smooth. We conclude using
Lemma 1.8.19.
To summarise, we will identify measurable Gelds
= { : H H ,
dened on smooth vectors with their possible extensions whenever possible. If
G}
the group is noncompact, we can not restrict ourselves to elds acting on smooth
vectors.
Open Access. This chapter is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/),
which permits use, duplication, adaptation, distribution and reproduction in any medium
or format, as long as you give appropriate credit to the original author(s) and the source,
a link is provided to the Creative Commons license and any changes made are indicated.
The images or other third party material in this chapter are included in the works
Creative Commons license, unless indicated otherwise in the credit line; if such material
is not included in the works Creative Commons license and the respective action is not
permitted by statutory regulation, users will need to obtain permission from the license
holder to duplicate, adapt or reproduce the material.
Chapter 2
Quantization on compact Lie
groups
In this chapter we briey review the global quantization of operators and symbols on compact Lie groups following [RT13] and [RT10a] as well as more recent
developments of this subject in this direction. Especially the monograph [RT10a]
can serve as a companion for the material presented here, so we limit ourselves to
explaining the main ideas only. This quantization yields full (nite dimensional)
matrixvalued symbols for operators due to the fact that the unitary irreducible
representations of compact Lie groups are all nite dimensional. Here, in order
to motivate the developments on nilpotent groups, which is the main subject of
the present monograph, we briey review key elements of this theory referring to
[RT10a] or to other sources for proofs and further details.
Technically, the machinery for such global quantization of operators on compact Lie groups appears to be simpler than that on graded Lie groups that we deal
with in subsequent chapters. Indeed, since the symbols can be viewed as matrices (more precisely, as linear transformations of nite dimensional representation
spaces), we do not have to worry about their domains of denitions, extensions,
and other functional analytical properties arising in the nilpotent counterpart of
the theory. Also, we have the Laplacian at our disposal, which is elliptic and biinvariant, simplifying the analysis compared to the analysis based on, for example,
the subLaplacian on the Heisenberg group, or more general Rockland operators
on graded Lie groups. On the other hand, the theory on graded Lie groups is
greatly assisted by the homogeneous structure, signicantly simplifying the analysis of appearing dierence operators and providing additional tools such as the
naturally dened dilations on the group.
When we will be talking about the quantization on graded Lie groups in
Chapter 5 we will be mostly concerned, at least in the rst stage, about assigning
an operator to a given symbol. In fact, it will be a small challenge by itself to make
The Editor(s) (if applicable) and The Author(s) 2016
V. Fischer, M. Ruzhansky, Quantization on Nilpotent Lie Groups, Progress in Mathematics 314,
DOI 10.1007/9783319295589_2
57
58
Chapter 2. Quantization on compact Lie groups
rigorous sense of a notion of a symbol there, but eventually we will show that the
correspondence between symbols and operators is onetoone. The situation on
compact Lie groups is considerably simpler in this respect. Moreover, in (2.19)
we will give a simple formula determining the symbol for a given operator. Thus,
here we may talk about quantization of both symbols and operators, with the
latter being often preferable from the point of view of applications, when we are
concerned in establishing certain properties of a given operator and use its symbol
as a tool for it.
Overall, this chapter is introductory, also serving as a motivation for the
subsequent analysis, so we only sketch the ideas and refer for a thorough treatise
with complete proofs to the monograph [RT10a] or to the papers that we point
out in relevant places.
We do not discuss here all applications of this analysis in the compact setting.
For example, we can refer to [DR14b] for applications of this analysis to Schatten
classes, rnuclearity, and trace formulae for operators on L2 (G) and Lp (G) for
compact Lie groups G. For the functional calculus of matrix symbols and operators
on G we refer to [RW14].
A related but dierent approach to the pseudodierential calculus of [RT10a]
has been also recently investigated in [Fis15]; there, a dierent notion of dierence
operators is dened intrinsically on each compact groups. This will not be discussed
here.
2.1
Fourier analysis on compact Lie groups
Throughout this chapter G is always a compact Lie group. As in Chapter 1, we
equip it with the uniquely determined probability Haar measure which is automatically biinvariant by the compactness of G. We denote it by dx. We start by
making a few remarks on the representation theory specic to compact Lie groups.
2.1.1 Characters and tensor products
An important rst addition to Section 1.1 is that for a compact group G, every
continuous irreducible unitary representation of G is nite dimensional. We denote
by d the dimension of a nite dimensional representation , d = dim H .
Another important property is the orthogonality of representation coecients
and let us choose some basis in the representation spaces
as follows. Let 1 , 2 G
d 1
d 2
so that we can view 1 , 2 as matrices 1 = ((1 )ij )i,j=1
and 2 = ((2 )kl )k,l=1
.
Then:
if 1 = 2 , then ((1 )ij , (2 )kl )L2 (G) = 0 for all i, j, k, l;
if 1 = 2 but (i, j) = (k, l), then ((1 )ij , (2 )kl )L2 (G) = 0;
2.1. Fourier analysis on compact Lie groups
59
if 1 = 2 and (i, j) = (k, l), then
((1 )ij , (2 )kl )L2 (G) =
1
,
d
with d = d1 = d2 .
For a nite dimensional continuous unitary representation : G U (H ) we
denote
(x) := Tr((x)),
the character of the representation . Characters have a number of fundamental
properties most of which follow from properties of the trace:
(e) = d ;
1 2 if and only if 1 = 2 ;
consequently, the character does not depend on the choice of the basis in
the representation space H ;
(yxy 1 ) = (x) for any x, y G;
1 2 = 1 + 2 ;
1 2 = 1 2 , with the tensor product 1 2 dened in (2.1);
a nite dimensional continuous unitary representation of G is irreducible
if and only if L2 (G) = 1.
( , )L2 (G) = 1 if 1 2 , and ( , )L2 (G) = 0 if
for 1 , 2 G,
1
2
1
2
1 2 ;
for any f L2 (G), there is the decomposition
d f ,
f=
G
given by the projections (2.7).
If we take 1 Hom(G, U (H1 )) and 2 Hom(G, U (H2 )) two nite dimensional representations of G on H1 and H2 , respectively, their tensor product 1 2
is the representation on H1 H2 , 1 2 Hom(G, U (H1 H2 )), dened by
(1 2 )(x)(v1 v2 ) := 1 (x)v1 2 (x)v2 .
(2.1)
Here the inner product on H1 H2 is induced from those on H1 and H2 by
(v1 v2 , w1 w2 )H1 H2 := (v1 , w1 )H1 (v2 , w2 )H2 .
In particular, it follows that
((1 2 )(x)(v1 v2 ), w1 w2 )H1 H2 = (1 (x)v1 , w1 )H1 (2 (x)v2 , w2 )H2 . (2.2)
60
Chapter 2. Quantization on compact Lie groups
the representation 1 2 does not have to be irreducible, and we
If 1 , 2 G,
can decompose it into irreducible ones:
1 2 =
m .
(2.3)
G
The constants m = m (1 , 2 ) are called the ClebschGordan coecients and
they determine the multiplicity of in 1 2 ,
m m
1 .
Also, we can observe that in view of the nite dimensionality only nitely many of
m s are nonzero. Combining this with (2.2), we see that the product of any of the
can be written as a nite linear
matrix coecients of representations 1 , 2 G
combination of matrix coecients of the representations from (2.3) with nonzero
ClebschGordan coecients. In fact, this can be also seen on the level of characters
providing more insight into the multiplicities m . First, for the tensor product of
1 and 2 we have 1 2 = 1 2 . Consequently, equality (2.3) implies
1 2 = 1 2 =
m
(2.4)
G
with
m = m (1 , 2 ) = (1 2 , )L2 (G) .
This equality can be now reduced to the maximal torus of G, for which we recall
Cartans maximal torus theorem: Let Tl G be an injective group homomorphism
with the largest possible l. Then two representations of G are equivalent if and
only if their restrictions to Tl are equivalent. In particular, the restriction Tl of
to Tl determines the equivalence class [].
Now, coming back to (2.4), we can conclude that we have
1  T l 2  T l =
m Tl .
G
For a compact connected Lie group G, the maximal torus is also called the Cartan
subgroup, and its dimension is denoted by rank G, the rank of G.
Explicit formulae for representations and the ClebschGordan coecients on
a number of compact groups have been presented by Vilenkin [Vil68] or Zhelobenko
[Zel73],
with further updates in [VK91, VK93] by Vilenkin and Klimyk.
2.1.2 PeterWeyl theorem
As discussed in Section 1.3, the Casimir element of the universal enveloping algebra
U(g) can be viewed as an elliptic linear second order biinvariant partial dierential
2.1. Fourier analysis on compact Lie groups
61
operator on G. If G is equipped with the uniquely determined (normalised) biinvariant Riemannian metric, the Casimir element can be viewed as its (negative
denite) LaplaceBeltrami operator, which we will denote by LG . Consequently,
for any D U(g) we have
DLG = LG D.
The fundamental result on compact groups is the PeterWeyl Theorem
[PW27] giving a decomposition of L2 (G) into eigenspaces of the Laplacian LG
on G, which we now sketch.
Theorem 2.1.1 (PeterWeyl). The space L2 (G) can be decomposed as the orthogonal
direct sum of biinvariant subspaces parametrised by G,
V , V = {x Tr(A(x)) : A Cd d },
L2 (G) =
G
the decomposition given by the Fourier series
d Tr f()(x) .
f (x) =
(2.5)
G
After a choice of the orthonormal basis in each representation space H , the set
d ij : = (ij )di,j=1
, G
(2.6)
B :=
becomes an orthonormal basis for L2 (G). For f L2 (G), the convergence of the
series in (2.5) holds for almost every x G, and also in L2 (G).
One possible idea for the proof of the PeterWeyl theorem is as follows. Let
us take B as in (2.6). Finite linear combinations of elements of B are called the
trigonometric polynomials on G, and we denote them by span(B). From the orthogonality of representations (see Section 2.1.1) we know that B is an orthonormal
set in L2 (G). It follows from (2.3) and the consequent discussion that span(B) is a
subalgebra of C(G), trivial representation is its identity, and it is involutive since
if G.
By invariance it is clear that B separates points of G. Conse G
quently, by the StoneWeierstrass theorem span(B) is dense in C(G). Therefore,
it is also dense in L2 (G), giving the basis and implying the PeterWeyl theorem.
For f L2 (G), the decomposition
f=
d f
G
given in Section 2.1.1 corresponds to the decomposition (2.5), the projections of
L2 (G) to V given by the convolution mappings
L2 (G) f f V .
(2.7)
62
Chapter 2. Quantization on compact Lie groups
The PeterWeyl theorem can be also viewed as the decomposition of left or right
regular representations of G on L2 (G) into irreducible components. Indeed, from
the homomorphism property of representations it follows that in the decomposition
L2 (G) =
d
span{ij : 1 i d },
(2.8)
j=1
G
the spans on the right hand side are L invariant, and the restriction of L to each
such space is equivalent to the representation itself. This gives the decomposition
of L into irreducible components as
L
d
G
The same is true for the decomposition of L2 (G) into R invariant subspaces
span{ij : 1 j d }, replacing the spans in (2.8).
It follows that the spaces V are biinvariant subspaces of L2 (G) and, therefore, they are eigenspaces of all biinvariant operators. In particular, they are
eigenspaces for the Laplacian LG and, by varying the basis in the representation
space H , we see that V corresponds to the same eigenvalue of LG , which we
denote by , i.e.
(2.9)
LG V = I, 0.
It is useful to introduce also the quantity corresponding to the rst order elliptic
operator (I LG )1/2 ,
(2.10)
:= (1 + )1/2 ,
so that we also have
(I LG )1/2 V = I.
The quantity and its powers become very useful in quantifying the growth/
decay of Fourier coecients, and eventually of symbols of pseudodierential operators.
Using the Fourier series expression (2.5) and the orthogonality of matrix
coecients of representations, one can readily show that the Plancherel identity
takes the form
d Tr f()
g () .
(f, g)L2 (G) =
G
From this, it becomes natural to dene the norm 2 (G)
,
f2 (G)
=
G
1/2
d f()2HS
(2.11)
2.1. Fourier analysis on compact Lie groups
63
with
Tr f()f() .
f()HS =
with the inner product
This norm denes the Hilbert space 2 (G)
d Tr (() () ),
, 2 (G),
(, )2 (G)
:=
(2.12)
G
and
2 (G)
= (, )2 (G)
=
1/2
1/2
d ()2HS
2 (G),
G
so that the Plancherel identity yields
f L2 (G) = f2 (G)
.
(2.13)
We conclude the preliminary part by recording some useful relations between the
there exists
dimensions d and the eigenvalues for representations G:
C > 0 such that
d C
dim G
2
and, even stronger,
d C
dim GrankG
2
(2.14)
The rst estimate follows immediately from the Weyl asymptotic formula for the
eigenvalue counting function for the rst order elliptic operator (I LG )1/2 on
the compact manifold G recalling that d2 is the multiplicity of the eigenvalue ,
and the second one follows with a little bit more work from the Weyl character
formula, with rank G denoting the rank of G. There is also a simple convergence
criterion
s
d2 < if and only if s > dim G,
(2.15)
G
which follows from property (ii) in Section 2.1.3 applied to the deltadistribution
e at the unit element e G.
2.1.3 Spaces of functions and distributions on G
Dierent spaces of functions and distributions can be characterised in terms of the
Fourier coecients. For this, it is convenient to introduce the space of matrices
taking into account the dimensions of representations. Thus, we set
:=
= (())G : () L (H )
= (())G : () Cd d ,
the second line valid after a choice of basis in H , and we are interested in the
images of function spaces on G in under the Fourier transform.
64
Chapter 2. Quantization on compact Lie groups
As it will be pointed out in Remark 2.2.1, we should rather consider the
quotient space / as the space of Fourier coecients, with the equivalence in
induced by the equivalence of representations. However, in order to simplify the
exposition, we will keep the notation as above.
The set can be considered as a special case of the direct sum of Hilbert
spaces described in (1.29), with the corresponding interpretation in terms of von
Neumann algebras. However, a lot of the general machinery can be simplied in the
present setting since the Fourier coecients allow the interpretation of matrices
with the dimension of each matrix equal to the
indexed over the discrete set G,
dimension of the corresponding representation.
Distributions
is dened
For any distribution u D (G), its matrix Fourier coecient at G
by
u
() := u, .
These are welldened since (x) are smooth (even analytic). This gives rise to
the Fourier transform of distributions on G but we will come to this after stating
a few properties of several function spaces.
The following equivalences are easy to obtain for spaces dened initially via
their localisations to coordinate charts, in terms of the quantity introduced in
(2.10):
i.e. if
(i) as we have already seen, f L2 (G) if and only if f 2 (G),
d f()2HS < .
G
s
if and only if
(ii) For any s R, we have f H s (G) if and only if f 2 (G)
2s
d f()2HS < .
G
(iii) f C (G) if and only if for every M > 0 there exits CM > 0 such that
f()HS CM
holds for all G.
(iv) u D (G) if and only if there exist M > 0 and C > 0 such that
u()HS C
holds for all G.
2.1. Fourier analysis on compact Lie groups
65
The second characterisation (ii) follows from (i) if we observe that f H s (G)
means that (I LG )s/2 f L2 (G), and then pass to the Fourier transform side.
The third characterisation (iii) follows if we observe that f() must satisfy (ii) for
all s and use estimates (2.14), and (iv) follows from (iii) by duality. The last two
S (G)
by
characterisations motivate to dene spaces S(G),
:= : M > 0 CM > 0 such that ()HS CM M
S(G)
and
:= : M > 0, C > 0 such that ()HS CM ,
S (G)
dened by family
with the seminormed topology on S(G)
k
d ()HS ,
pk () =
G
It follows that the Fourier inversion formula (2.5)
and the dual topology on S (G).
can be extended to the following: the Fourier transform FG in (1.2) and its inverse,
dened by
1
(FG
)(x) :=
d Tr (()(x)),
(2.16)
G
F 1 : S(G)
C (G), and are inverse to
are continuous as FG : C (G) S(G),
G
is a nuclear
each other on C (G) and S(G). In particular, this implies that S(G)
is given by
Montel space. The distributional duality between S (G) and S(G)
2 S(G).
d Tr (1 ()2 ()), 1 S (G),
1 , 2 =
G
G
The Fourier transform can be then extended to the space of distributions D (G).
by
S (G)
Thus, for u D (G), we dene FG u u
1
G , S(G),
FG u, G := u, FG
where ( )(x) = (x1 ) and , G is the distributional duality between D (G)
and C (G). Analogously, its inverse is given by
1
C (G),
FG , G := , FG ( )G , S (G),
and are
and these extended mappings are continuous between D (G) and S (G)
inverse to each other. It can be readily checked that they agree with their restrictions to spaces of test functions, explaining the appearance of the inversion
mapping .
66
Chapter 2. Quantization on compact Lie groups
Gevrey spaces and ultradistributions
Recently, Gevrey spaces of ultradierentiable functions as well as spaces of corresponding ultradistributions have been characterised as well. We say that a function
C (G) is a GevreyRoumieu ultradierentiable function, s (G), if in every local coordinate chart, its local representative C (Rn ) belongs to s (Rn ),
that is, satises the condition that there exist constants A > 0 and C > 0 such
that
 (x) CA (!)s
holds for all x Rn and all multiindices . For s = 1 we obtain the space of
analytic functions on G. As with other spaces before, s (G) is thus dened as
having its localisations in s (Rn ), and a question of its characterisation in terms
of its Fourier coecients arises.
Analogously, we say that is a GevreyBeurling ultradierentiable function,
(s) (G), if its local representatives satisfy the condition that for every A > 0
there exists CA > 0 such that
 (x) CA A (!)s
holds for all x Rn and all multiindices . For 1 s < , these spaces do
not depend on the choice of local coordinates on G in the denition, and can be
characterised as follows:
Proposition 2.1.2. Let 1 s < .
(1) We have s (G) if and only if there exist B > 0 and K > 0 such that
B
()
HS Ke
1/s
holds for all G.
(2) We have (s) (G) if and only if for every B > 0 there exists KB > 0 such
that
B1/s
()
HS KB e
holds for all G.
The space of continuous linear functionals on s (G) or (s)
(G) is called the
space of ultradistributions and is denoted by s (G) or (s)
(G) , respectively.
(G) , we note that its Fourier coecient at G
For any v s (G) or (s)
can be dened analogously to the case of distributions by
v() := v, v( ).
These are welldened since G is compact and hence (x) are analytic.
2.1. Fourier analysis on compact Lie groups
67
Proposition 2.1.3. Let 1 s < .
(1) We have v s (G) if and only if for every B > 0 there exists KB > 0 such
that
1/s
v ()HS KB eB
holds for all G.
(2) We have v (s)
(G) if and only if there exist B > 0 and K > 0 such that
v ()HS KeB
1/s
holds for all G.
Proposition 2.1.2 can be actually extended to hold for any 0 < s < ,
and we refer to [DR14a] for proofs and further details. This can be viewed also
from the point of view of general eigenfunction expansions of function of compact
manifolds, see [DR16] for the treatment of more general Komatsutype classes
of ultradierentiable functions and ultradistributions, building on an analogous
description for analytic functions by Seeley [See69].
For a review of the representation theory of compact Lie groups and further
constructions using the LittlewoodPaley decomposition based on the heat kernel
we refer to Steins book [Ste70b].
2.1.4 p spaces on the unitary dual G
For a general theory of noncommutative integration on locally compact unimodular groups we refer to Dixmier [Dix53] and Segal [Seg50, Seg53]. In this framework,
the HausdorYoung inequality has been established (see Kunze [Kun58]) for a ver based on the Schatten classes, namely, an
sion of p spaces on the unitary dual G
inequality of the type
1/p
d f()p p
f Lp (G) for 1 < p 2,
G
Sd
with an obvious modication for p = 1, where p1 + p1 = 1, and Sdp is the (d d )dimensional Schatten p class. While the theory of the above spaces is wellknown
(see e.g. Hewitt and Ross [HR70, Section 31] or Edwards [Edw72, Section 2.14]),
which
here we describe and develop a little further another class of p spaces on G
was considered in [RT10a, Section 10.3.3], to which we refer for details and proofs
of statement that we do not prove here.
by the condition
For 1 p < , we dene the space p (G)
p (G)
:=
G
1/p
p(
d
2
1
p2
()pHS
< .
68
Chapter 2. Quantization on compact Lie groups
by
For p = , we dene the space (G)
1/2
(G)
()HS < .
:= sup d
G
dened in (2.11), while the 1 (G)norm
For p = 2 we recover the space 2 (G)
becomes
d3/2
1 (G)
:=
()HS .
G
This space and the HausdorYoung inequality for it become useful in, for example,
proving Proposition 2.1.2. Also, it appears naturally in questions concerning the
convergence of the Fourier series:
then the (Fourier) series (2.16) converges absolutely
Remark 2.1.4. If 1 (G),
and uniformly on G.
On the other hand, one can show that if f C k (G) with an even k > 12 dim G,
and the Fourier series (2.5) converges uniformly. Indeed, we can
then f 1 (G)
estimate
f1 (G)
d3/2
G
((I LG )k/2 f )HS
1/2
d2
2k
G
1/2
d ((I LG )k/2 f )2HS
G
C(I LG )
k/2
f L2 (G) < ,
in view of the Plancherel formula and (2.15), provided that 2k > dim G. In fact,
the same argument shows the implication
f H s (G),
s>
1
dim G
2
f 1 (G),
with the uniform convergence of the Fourier series (2.5) of f .
as weighted sequence spaces with weights given
Regarding these p (G)spaces
by powers of d , a general theory of interpolation spaces [BL76, Theorem 5.5.1]
implies that they are interpolation spaces, namely, for any 1 p0 , p1 < , we
have
p1 (G)
= p (G),
p0 (G),
,p
where 0 < < 1 and
1
p
1
p0
p1 ,
see [RT10a, Proposition 10.3.40].
The HausdorYoung inequality holds for these spaces as well. Namely, if
1 p 2 and p1 + p1 = 1, we have
2.1. Fourier analysis on compact Lie groups
69
fp (G)
f Lp (G)
(2.17)
1
Lp (G) p (G)
FG
,
(2.18)
for all f L (G), and
for all p (G).
We give a brief argument for these. To prove (2.18), on one hand we already
have Plancherels identity (2.13). On the other hand, from (2.16) we have
1
)(x)
d ()HS (x)HS =
d3/2
(FG
.
()HS = 1 (G)
G
G
Now the SteinWeiss interpolation (see e.g. [BL76, Corollary 5.5.4]) implies (2.18).
= p (G),
1 p < .
From this, (2.17) follows using the duality p (G)
We remark that it is also possible to prove (2.17) directly by interpolation
as well. However, one needs to employ an version of the interpolation theory
with the change of measure, as e.g. in Lizorkin [Liz75].
Let us point out the continuous embeddings, similar to the usual ones:
Proposition 2.1.5. We have
q (G)
p (G)
q (G)
p (G)
and
1 p q .
Proof. We can assume p < q. Then, in the case 1 p < and q = , we can
estimate
"p
!
2 p
12
p
sup d ()HS
(G)
d 2 ()pHS = pp (G)
=
.
G
G
Let now 1 p < q < . Denoting a := dq
q (G)
q1
aq
G
12
()HS , we get
p1
ap =
G
p1
p( q2 12 )
()pHS
G
p (G)
,
completing the proof.
and the correspondFinally, we establish a relation between the family p (G)
p
p
ing Schatten family of spaces, which we denote by sch (G), dened by the norms
p
sch (G)
:=
G
1/p
d ()pS p
, 1 p < ,
70
Chapter 2. Quantization on compact Lie groups
where S p = Sdp is the (d d )dimensional Schatten pclass, and
sch (G)
:= sup ()L (H ) ,
G
We have the following relations:
Proposition 2.1.6. For 1 p 2, we have continuous embeddings as well as the
estimates
p (G)
p (G)
sch
p
and
sch (G)
p (G)
1 p 2.
2 p .
For 2 p , we have
p (G)
psch (G)
p (G)
p
and
sch (G)
Proof. For p = 2, the norms coincide since S 2 = HS. Let rst 1 p < 2. Since
() Cd d , denoting by sj its singular numbers, by the Holder inequality we
have
p2
2p
2
d
d
d
2
2p
p
spj
1
sj p = d 2 ()pHS ,
()pS p =
j=1
j=1
j=1
i.e.
2p
()S p d2p ()HS
(1 p 2).
Consequently, it follows that
2 p
pp (G)
d ()pS p
d 2 ()pHS = pp (G)
=
,
sch
G
G
proving the rst claim. Conversely, for 2 < p < , we can estimate
()2HS =
s2j
j=1
d
j=1
p2
p2
p
d
p2
p
2
1
sj 2 = d p ()2S p ,
j=1
implying
p2
()HS d2p ()S p
It follows that
pp (G)
=
G
2 p
2
()pHS
(2 < p < ).
G
d ()pS p = pp
sch (G)
proving the second claim for 2 < p < . Finally, for p = , the inequality
()HS d1/2
()L (H )
2.2. Pseudodierential operators on compact Lie groups
71
implies
1/2
(G)
()HS sup ()L (H ) =
= sup d
G
G
sch (G)
,
completing the proof.
2.2 Pseudodierential operators on compact Lie groups
In this section we look at linear continuous operators A : C (G) D (G) and a
global quantization of A yielding its full matrixvalued symbol. By the Schwartz
kernel theorem (Theorem 1.4.1) there exists a unique distribution KA D (GG)
such that
KA (x, y)f (y)dy,
Af (x) =
G
interpreted in the distributional sense. We can rewrite this as a rightconvolution
kernel operator
Af (x) =
RA (x, y 1 x)f (y)dy,
G
with
RA (x, y) = KA (x, xy 1 ),
so that
Af (x) = (f RA (x, ))(x).
2.2.1 Symbols and quantization
The idea for the following construction is that we dene the symbol of A as the
Fourier transform of its right convolution kernel in the second variable. However,
for the presentation purposes we now take a dierent route and, instead, we dene
by
the mapping A : G G
A (x, ) := (x) (A)(x),
(2.19)
with (A)(x) L (H ) dened by
(A(x)u, v)H := A((x)u, v)H
for all u, v H . After choosing a basis in the representation space H , we can
interpret this as a matrix A (x, ) Cd d and (A)ij = A(ij ), i.e. the operator
A acts on the matrix (x) componentwise, so that
A (x, )ij =
ki (x)Akj (x).
k=1
We note that the symbol in (2.19) is welldened since we can multiply the distribution A by a smooth (even analytic) matrix .
72
Chapter 2. Quantization on compact Lie groups
Remark 2.2.1. We also observe that strictly speaking, the denition (2.19) depends
on the choice of the representation from its equivalence class []. Namely, if
1 2 , so that
2 (x) = U 1 1 (x)U
for some unitary U and all x G, then
f(2 ) = U 1 f(1 )U
and, therefore,
A (x, 2 ) = U 1 A (x, 1 )U.
(2.20)
However, it can be readily checked that the quantization formula (2.22) below
remains unchanged due to the presence of the trace. So, denoting by Rep G the
set of all strongly continuous unitary irreducible representations of G, the symbol
is well dened as a mapping
A : G Rep G
or as
/
A : G G
where the equivalence on is given by the equivalence of representations on Rep G
inducing the equivalence on by conjugations, as in formula (2.20). We will disregard this technicality in the current presentation to simplify the exposition, referring to [RT10a] for a more rigorous treatment. We note, however, that if 1 2 ,
then
(2.21)
Tr 1 (x)A (x, 1 )f(1 ) = Tr 2 (x)A (x, 2 )f(2 ) .
Using the symbol A , it follows that the linear continuous operator A :
C (G) D (G) can be (de)quantized as
Af (x) =
d Tr (x)A (x, )f() .
(2.22)
G
If the operator A maps C (G) to itself and f C (G), the formula (2.22) can
be understood in the pointwise sense to hold for all x G, with the absolute
convergence of the series. It can be shown that formulae (2.19) and (2.22) imply
that A is the Fourier transform of RA , namely, we have
RA (x, y)(y) dy.
A (x, ) =
G
If the formula (2.22) holds, we will also write A = Op(A ).
In view of (2.21), the sum in (2.22) does not depend on the choice of a
representation from its equivalence class [].
2.2. Pseudodierential operators on compact Lie groups
73
Example 2.2.2. For the identity operator I we have its symbol
I (x, ) = (x) (x) = Id
is the identity matrix in Cd d , by the unitarity of (x), so that (2.22) recovers
the Fourier inversion formula (2.5) in this case. For the Laplacian LG on G, we
have
LG (x, ) = (x) LG (x) = Id
by the unitarity of and (2.9), where are the eigenvalues of LG corresponding
to . Consequently, we also have
(ILG )/2 (x, ) = Id .
Example 2.2.3. In the case of the torus G = Tn = Rn /Zn , and the representations
{ }Zn xed as in Remark 1.1.4, we see that all d = 1. Hence
A (x, ) A (x, ) = e2ix A(e2ix ) C,
(x, ) Tn Zn ,
with the quantization (2.22) becoming the toroidal quantization
e2ix A (x, ) f(),
Af (x) =
Zn
for a thorough analysis of which we refer to [RT10b] and [RT10a, Section 4].
Example 2.2.4. With our choices of denitions, the symbols of leftinvariant operators on G become independent of x. As shown in Section 1.5, if
Af = f
for some L1 (G), then it is leftinvariant. Consequently, the right convolution
kernel of A is RA (x, y) = (y) and, therefore, its Fourier transform is
A (x, ) =
().
On the other hand, if
Af = f
for some L (G), then it is rightinvariant. In this case its right convolution
kernel is RA (x, y) = (xyx1 ) and, therefore, its Fourier transform in y gives
1
A (x, ) = (x)
()(x).
The notion of the symbol A becomes already useful in stating a criterion
for the L2 boundedness for an operator A. We recall from Section 1.3 that X
denotes the leftinvariant partial dierential operators of order  corresponding
to a basis of leftinvariant vector elds X1 , , Xn , n = dim G, of the Lie algebra
74
Chapter 2. Quantization on compact Lie groups
g of G. As the derivatives with respect to these vector elds in general do not
commute, in principle we have to take into account their order in forming partial
dierential operators of higher degrees. However, we note that the subsequent
statements remain valid if we restrict our choice to
X = X11 Xnn .
We will sometimes write Xx to emphasise that the derivatives are taken with
respect to the variable x.
Theorem 2.2.5. Let G be a compact Lie group and let A : C (G) C (G) be
a linear continuous operator. Let k be an integer such that k > 12 dim G. Assume
that there is a constant C > 0 such that
Xx A (x, )L (H ) C
and all  k. Then A extends to a bounded operator from
for all (x, ) G G,
2
2
L (G) to L (G).
In this theorem and elsewhere, L (H ) denotes the operator norm of
A (x, ) L (H ) or, after a choice of the basis, the operator norm of the matrix
multiplication by the matrix A (x, ) Cd d . The appearance of the operator
norm is natural since for the convolution operators we have
h()L (H ) ,
f f hL (L2 (G)) = f h f L (L2 (G)) = sup
(2.23)
G
following from f
h() =
h()f() and Plancherels theorem.
2.2.2 Dierence operators and symbol classes
In order to describe the symbolic properties and to establish the symbolic calculus
of operators we have to replace the derivatives in frequency, used in the symbolic
calculus on Rn , by suitable operations acting on the space of Fourier coecients.
We call these operations dierence operators. Roughly speaking, this corresponds
to the idea that in the Calder
onZygmund theory, the integral kernel KA has singularities at the diagonal or, in other words, the rightconvolution kernel RA (x, )
has singularity at the unit element e of the group only. Therefore, if we form an
operator with a new integral kernel q()RA (x, ) with a smooth q C (G) satisfying q(e) = 0, the properties of this new operator should be better than those of
the original operator A.
In [RT10a], the corresponding notion of dierence operators has been introduced leading to the symbolic calculus of operators on G. However, we now
follow the ideas of [RTW14] with a slightly more general treatment of dierence
operators.
2.2. Pseudodierential operators on compact Lie groups
75
Denition 2.2.6. Let q C (G) vanish of order k N at the unit element e G,
i.e. (Dq)(e) = 0 for all leftinvariant dierential operators D Di k1 (G) of order
k 1. Then the dierence operator of order k is an operator acting on the space
of Fourier coecients by the formula
().
(q f)() := qf
We denote the set of all dierence operators of order k by di k (G).
We now dene families of rst order dierence operators replacing derivatives
in the frequency variable in the Euclidean setting.
Denition 2.2.7. A collection of rst order dierence operators q1 , . . . , q
is called admissible, if the corresponding functions q1 , . . . , q C (G)
di 1 (G)
satisfy
qj (e) = 0, dqj (e) = 0, j = 1, . . . , ,
and, moreover,
rank(dq1 (e), . . . , dq (e)) = dim G.
It follows, in particular, that e is an isolated common zero of the family {qj }j=1 .
We call an admissible collection strongly admissible, if it is the only common zero,
i.e. if
#
{x G : qj (x) = 0} = {e}.
j=1
We note that dierence operators all commute with each other. For a given
admissible collection of dierence operators we use the multiindex notation
1
1
q (x) ,
:= q1 q and q (x) := q1 (x)
the dimension of the multiindex N0 depending on the number of dierence
operators in the collection. Consequently, there exist corresponding dierential
operators X () Di  (G) such that the Taylor expansion formula
f (x) =
N 1
1 1
q (x ) X () f (e) + O(h(x)N ),
!
h(x) 0,
(2.24)
holds true for any smooth function f C (G) and any N , with h(x) the geodesic
distance from x to the identity element e. An explicit construction of operators
X () in terms of q (x) can be found in [RT10a, Section 10.6]. Operators X and
X () can be expressed in terms of each other.
Example 2.2.8. In the case of the torus, G = Tn = Rn /Zn , let
qj (x) = e2ixj 1, j = 1, . . . , n.
76
Chapter 2. Quantization on compact Lie groups
The collection {qj }nj=1 is strongly admissible, and the corresponding dierence
operators take the form
(qj )( ) (qj )() = ( + ej ) (),
j = 1, . . . , n,
n identied with Zn , where ej is the j th unit vector in Zn . The
with T
periodic Taylor expansion takes the following form (see [RT10a, Theorem 3.4.4]):
for any C (Tn ) we have
1
(e2ix 1) Xz() (z)z=0 +
(x) =
(x)(e2ix 1) ,
!
<N
=N
where C (Tn ) and
(e2ix 1) := (e2ix1 1)1 (e2ixn 1)n .
()
The operators Xz
Xz()
have the form
1)
Xz(
1
n)
Xz(
n
with
k)
Xz(
k
$
k 1
j=0
1
j .
2i zk
Example 2.2.9. For partial dierential operators, it can be readily observed that
the application of dierence operators reduces the order of symbols. Thus, let
D=
c (x)Xx , c C (G).
N
Then it was shown in [RT10a, Proposition 10.7.4] that
q D (x, ) =
c (x)
(1) (Xx q)(e)Xx (x, ).
N
In particular, if q has zero of order M at e G then Op(q D ) is of order N M .
Remark 2.2.10. We can estimate dierences in terms of original symbols: assume
that the symbol satises
:= supm ()L (H ) <
for some m R. Then for any dierence operator q dened in terms of a function
q C (G) we have the estimate
q ()L (H ) CqC +m (G) m
with a constant C independent of and q, where = (dim G)/2 is the smallest
integer larger than half the dimension of G and m is the smallest integer larger
than m. We refer to [RW14, Lemma 7.1] for the proof. However, if q vanishes at
the unit element e to some order, we can impose a much better behaviour.
2.2. Pseudodierential operators on compact Lie groups
77
The usual Hormander classes m (G) of pseudodierential operators on G
viewed as a manifold can be characterised in terms of the matrixvalued symbols.
Here we recall that A m (G) means that in every local coordinate chart U G,
n
the pullback of AU to Rn is a pseudodierential operator AU m
1,0 (R ), i.e. it
can be written as
2ix
e
a(x, )f ()d with f () =
e2ix f (x)dx,
(2.25)
AU f (x) =
Rn
Rn
m
(Rn ), i.e. satisfying
with symbol a = aU S1,0
x a(x, ) C (1 + )m
for all multiindices , , and all x, Rn .
The following characterisation was partly proved in [RT10a, RT13] (namely
(A)(C)) and completed in [RTW14] (namely (B)(C)(D)) .
Theorem 2.2.11. Let G be a compact Lie group of dimension n. Let A be a linear
continuous operator from C (G) to D (G). Then the following statements are
equivalent:
(A) A m (G).
(B) For every leftinvariant dierential operator D Di k (G) of order k and
of order l the symbol estimate
every dierence operator q di l (G)
q DA (x, )L (H ) CqD
ml
is valid.
we have
(C) For an admissible collection 1 , . . . , di 1 (G)
m
Xx A (x, )L (H ) C
for all multiindices N0 and Nn0 . Moreover,
sing supp RA (x, ) {e}.
we have
(D) For a strongly admissible collection 1 , . . . , di 1 (G)
Xx A (x, )L (H ) C
m
for all multiindices N0 and Nn0 .
m
Motivated by Theorem 2.2.11, (D), we may dene symbol classes S,
(G).
Fixing a strongly admissible collection of dierence operators
1 , . . . , di 1 (G),
78
Chapter 2. Quantization on compact Lie groups
m
we say that A S,
(G) if A (x, ) satises
Xx A (x, )L (H ) C
m+
(2.26)
and for all multiindices N and Nn . If > ,
for all (x, ) G G
0
0
this denition is independent of the choice of a strongly admissible collection
of dierence operators. The equivalence (A)(D) in Theorem 2.2.11 can be
rephrased as
m
(G).
A m (G) A S1,0
For any 0 < 1, the equivalence (B)(C)(D) in Theorem 2.2.11
m
(G) if we replace the symbolic conditions
remains valid for the symbol class S,
m
(G) with dierent
there by the condition (2.26). As we shall see later, the class S,
values of and becomes useful in a number of applications.
Theorem 2.2.5 has analogue for (, ) classes:
Theorem 2.2.12. Let 0 < 1 and let A be an operator with symbol in
m
(G). Then A is a bounded from H s (G) to H sm (G) for any s R.
S,
See [RW14, Theorem 5.1] for the proof.
2.2.3
Symbolic calculus, ellipticity, hypoellipticity
We now give elements of the symbolic calculus on the compact Lie group G. Here,
we x some strongly admissible collection of dierence operators, with correspond()
ing operators Xx coming from the Taylor expansion formula (2.24). We refer to
[RT10a, Section 10.7.3] for proofs and other variants of the calculus below. We
start with the composition.
Theorem 2.2.13. Let m1 , m2 R and 0 < . Let A, B : C (G) C (G) be
m1
m1
(G) and B S,
(G). Then
linear continuous operators with symbols A S,
m1 +m2
(G) and we have
AB S,
AB
1
( A )(X () B ),
!
where the asymptotic expansion means that for every N N we have
AB (x, )
1
m +m ()N
( A )(x, )Xx() B (x, ) S,1 2
(G).
!
<N
The composition formula together with Theorem 2.2.5 imply a criterion for
the boundedness in L2 Sobolev spaces.
2.2. Pseudodierential operators on compact Lie groups
79
Corollary 2.2.14. Let G be a compact Lie group and let A : C (G) C (G) be
a linear continuous operator. Let m R. Assume that the symbol A satises
Xx A (x, )L (H ) C
and all multiindices . Then A extends to a bounded
for all (x, ) G G,
operator from H s (G) to H sm (G), for all s R.
Let us now present a construction of amplitude operators in our setting. Let
is a matrixvalued amplitude in the
0 , 1. We say that a : G G G
m
class A, (G) if for a strongly admissible collection of dierence operators on G
we have the amplitude inequalities
Xx Xy a(x, y, )L (H ) C
m++
The corresponding
for all multiindices , , and for all (x, y, ) G G G.
amplitude operator Op(a) : C (G) D (G) is dened by
d Tr (x)
a(x, y, )f (y)(y) dy .
(2.27)
Op(a)f (x) :=
G
G
In the case a(x, y, ) = A (x, ) independent of y, we recover the quantization
(2.22), namely, we have Op(a) = A.
Theorem 2.2.15. Let a Am
, (G). If 0 < 1 and 0 1 then Op(a) is a
continuous linear operator from C (G) to C (G). Moreover, if 0 < 1,
then A = Op(a) is a pseudodierential operator with a matrixvalued symbol
m
(G), which has the asymptotic expansion
A S,
A (x, )
1
X () a(x, y, )y=x ,
! y
where the asymptotic expansion means that for every N N we have
1
m()N
X () a(x, y, )y=x S,
(G).
A (x, )
! y
<N
For the proof of this theorem we refer to [RT11]. Given the formula for the
amplitude operators in Theorem 2.2.15, the symbol of the adjoint operator can be
found as follows.
Theorem 2.2.16. Let m R and 0 < . Let A : C (G) C (G) be a linear
m
(G). Then the symbol A of the adjoint
continuous operator with symbol A S,
m
operator A satises A S, (G), and is given by
A (x, )
1
X () A (x, ) ,
! x
80
Chapter 2. Quantization on compact Lie groups
where A (x, ) is the adjoint matrix to A (x, ), and the asymptotic expansion
means that for every N N we have
A (x, )
1
m()N
X () A (x, ) S,
(G).
! x
<N
We recall that the operator A m (G) on G viewed as a manifold is elliptic
if all of its localisations to coordinate charts are (locally) elliptic. This can be
characterised in terms of the matrixvalued symbols. A combination of [RTW14,
Theorem 4.1] and [RT10a, Theorem 10.9.10] yields
Theorem 2.2.17. An operator A m (G) is elliptic if and only if its symbol
and for all such satises
A (x, ) is invertible for all but nitely many G,
A (x, )1 L (H ) C
for all x G. Furthermore, in this case, assume that
A
Aj ,
Aj mj (G).
j=0
Let B
k=0
Bk , where
B0 (x, ) = A0 (x, )1
for large , and the symbols Bk are dened recursively by
BN = B0
k
N
1 N
k=0 j=0 =N jk
1
( B )(Xx() Aj ).
! k
Then Op(Bk ) mk (G), B = Op(B ) m (G), and the operators AB I
and BA I are in (G).
One can also provide a criterion for the hypoellipticity in terms of matrixvalued symbols ([RTW14]), in analogy to the one on Rn given by Hormander
([Hor67b]).
m
Theorem 2.2.18. Let m m0 and 0 < 1. Let A Op(S,
(G)) be a
m
pseudodierential operator with symbol A S, (G) which is invertible for all
and for all such satises
but nitely many G,
A (x, )1 L (H ) C
m0
for all x G. Assume also that (for a strongly admissible collection of dierence
operators) we have
%
&
+
A (x, )1
Xx A (x, ) L (H ) C
2.2. Pseudodierential operators on compact Lie groups
81
Then there
for all multiindices , , all x G, and all but nitely many G.
m0
exists an operator B Op(S, (G)) such that AB I and BA I belong to
(G). Consequently, we have
sing supp Au = sing supp u
for all u D (G).
We nish this section with several results that are usually expected from the
calculus. The following asymptotic expansion formula was established in [RW14].
m
Proposition 2.2.19. Let j S,j (G), j N0 , 0 < 1, be a family of
m0
. Then there exists a symbol S,
(G) such that
symbols with mj
N
1
j=0
mN
j S,
(G)
for all N N0 .
The functional calculus of matrix valued symbols and its operator counterpart have been also developed in [RW14]. A notable corollary of such functional
calculus is the following
2m
(G) satises
Corollary 2.2.20. Let 0 < 1 and let m 0. Assume A S,
A (x, ) > 0 and
A (x, )1 L (H ) C2m
for all x and . Then the square root
B (x, ) =
A (x, )
m
in the sense of positive matrices is a symbol satisfying B S,
(G).
This is the corollary of the following more general result:
m
(G), m 0, is
Theorem 2.2.21. Let 0 1 and 0 < 1. Assume A S,
positive denite, invertible, and satises
A (x, )1 L (H ) Cm
for all x and for all but nitely many . Then for any number s C,
B (x, ) := A (x, )s = exp(s log A (x, ))
m
denes a symbol B S,
(G), with m = Re (ms).
In fact, the assumptions of Theorem 2.2.21 imply something stronger, namely,
that the symbol A (x, ) is parameterelliptic with respect to R ; we refer to
[RW14] for the denition of parameterellipticity in this setting, and for a more general exposition and statements of the functional calculus on compact Lie groups.
82
Chapter 2. Quantization on compact Lie groups
2.2.4 Fourier multipliers and Lp boundedness
Here we give an overview of the Lp estimates for the Fourier multipliers and for
noninvariant operators on compact Lie groups following [RW13, RW15]. We set
aside the case of biinvariant operators (or spectral multipliers) noting that there
exist many results in this direction (see e.g. N. Weiss [Wei72], Coifman and G.
Weiss [CW74], Stein [Ste70b], Cowling [Cow83], Alexopoulos [Ale94], to refer the
reader to only a few). Instead, we concentrate on the case of leftinvariant operators
(or Fourier multipliers). To the best of our knowledge the literature in this case is
much smaller, with a notable exception of a multiplier theorem for leftinvariant
operators on the group SU(2) treated by Coifman and Weiss [CW71b], Coifman
and de Guzm
an [CdG71], and appearing in more detail in the monograph by
Coifman and Weiss [CW71a]. The conditions there are formulated using specic
explicit expressions involving ClebschGordan coecients on SU(2), but they can
be recast in a much shorter form using the concept of dierence operators. It also
allows one to treat the case of general compact Lie groups. Finally we note that
there exist also results for the spectral multipliers in the subLaplacian, also on
SU(2), for which we refer to Cowling and Sikora [CS01].
First, we discuss leftinvariant operators A : C (G) D (G), so that the
matrixvalued symbol A (x, ) = A () is independent of x and can be given as
A () = (x) (A)(x) = (A)(e).
The multiplier theorems that we will present can be said to be of MihlinH
ormander type in the sense that they provide analogues of famous multiplier theorems
on Rn by Mihlin [Mih56, Mih57] and Hormander [Hor60].
In order to formulate the results, we need to x a particular collection of
rst order dierence operators associated to the elements of the unitary dual G.
we notice that the (d d )matrix
Thus, for a xed representation 0 G,
0
0
0 (x) Id0 vanishes at x = e. Consequently, we dene the dierence operators
0 D
0
= (0 Dij )i,j=1
associated with its elements,
0 Dij
:= (0 )ij i,j ,
where i,j is the Kronecker delta. For a family of dierence operators of this type,
D1 =1 Di1 j1 , D2 =2 Di2 j2 , . . . , Dm =m Dim jm ,
(2.28)
1 ik , jk d , 1 k m, we dene
with k G,
k
m
1
D := D
1 Dm .
(2.29)
The described dierence operators 0 D have a number of useful properties. For
example, they satisfy the nite Leibniz formula (while general dierence operators
2.2. Pseudodierential operators on compact Lie groups
83
satisfy only an asymptotic Leibniz formula, see [RT10a, Section 10.7.4]). Namely,
for any xed 0 , they satisfy
Dij ( ) = (Dij ) + (Dij ) +
d 0
(Dik )(Dkj ).
(2.30)
k=1
The collection of dierence operators
1 i, j d }
{0 Dij : 0 G,
0
is strongly admissible. Moreover, it has a nite strongly admissible subcollection.
Indeed, a homomorphic embedding of G into U (N ) for some N is itself a representation of G. Decomposing it into irreducible components gives the desired nite
family of 0 s.
We now formulate the rst result on the Lp boundedness of leftinvariant
operators.
Theorem 2.2.22. Let A : C (G) D (G) be a leftinvariant linear continuos
operator on a compact Lie group G, and let k denote the smallest even integer
such that k > 12 dim G. Assume that the symbol A of A satises
D A ()L (H ) C

(2.31)
Then the operator A is of weak type
for all multiindices  k and all G.
(1,1) and is bounded on Lp (G) for all 1 < p < .
We note that by Theorem 2.2.11, imposing conditions (2.31) for all multiindices would imply that A is a leftinvariant pseudodierential operator in
Hormanders class, A 0 (G), for which the Lp boundedness would follow from
the corresponding Lp boundedness in Rn for its localisations. However, imposing
conditions (2.31) for multiindices  k still assures that the operator A is of
Calder
onZygmund type (in the sense of Coifman and Weiss, see Section A.4). The
proof of the Lp boundedness for 1 < p 2 follows by Marcinkiewicz interpolation
theorem (see Proposition 1.5.1) from the L2 boundedness (and hence also weak
(2,2) type) in Theorem 2.2.5, and from weak (1,1) type, which becomes, therefore,
the main task.
For 2 < p < , the result follows by duality. Before we give an idea behind
the proof of the weak (1,1) type, let us formulate several corollaries from Theorem
2.2.22. We recall that the Sobolev space W p,s (G) on G is the usual Sobolev space
on G as a manifold dened by requiring all the localisations to belong to the
Euclidean space W p,s (Rn ) = (I LRn )s/2 Lp (Rn ), where LRn is the Laplacian on
Rn and s R.
Corollary 2.2.23. Let A : C (G) D (G) be a leftinvariant linear continuous
operator on a compact Lie group G. Let 0 1 and let k denote the smallest
84
Chapter 2. Quantization on compact Lie groups
even integer such that k >
1
2
dim G. Assume that the symbol A of A satises
D A ()L (H ) C

Then the operator A extends to a
for all multiindices  k and all G.
bounded operator from the Sobolev space W p,r (G) to Lp (G) for any 1 < p < ,
with
1 1
r = k(1 ) .
p 2
Example 2.2.24. Let
Lsub = X 2 + Y 2
be a subLaplacian on SU(2). Then it was shown in [RTW14] that it has a
(SU(2)). Consequently,
parametrix with the matrixvalued symbol in the class S 1
1
2 ,0
for any 1 < p < , Corollary 2.2.23 implies the subelliptic estimate
f
p,s+1 1 1 
p
2 (SU(2))
Cp Lsub f W p,s (SU(2)) ,
where the estimate is extended from s = 0 to any s R by the calculus. We refer to
[RTW14] for the construction and discussion of parametrices for other operators,
including the heat and the wave operator, dAlambertian, and some higher order
operators, on SU(2) and on S3 , and to [RW13, RW15] for the corresponding Lp estimates.
Example 2.2.25. Let (, , ) be the standard Euler angles on SU(2), see e.g.
[RT10a, Chapter 11] for a detailed treatment of SU(2). Thus, we have 0 < 2,
0 , and 2 < 2, and every element
a b
u = u(, , ) =
SU(2)
b a
is parametrised in such a way that
2a
a = 1 + cos , 2ab = iei sin , 2ab = iei sin .
Conversely, we can also write
cos( 2 )ei(+)/2
u(, , ) =
i sin( 2 )ei()/2
i sin( 2 )ei()/2
cos( 2 )ei(+)/2
SU(2).
Let X be a leftinvariant vector eld on G normalised in such a way that X =
/ with respect to the Killing form. It was shown in [RTW14] that for C,
the operator X + is invertible if and only if i
1
Z,
2
2.2. Pseudodierential operators on compact Lie groups
85
and, moreover, for such , the inverse (X + )1 has its matrixvalued symbol in
0
(SU(2)). The same conclusion remains true if we replace SU(2) by
the class S0,0
3
S , with the corresponding selection of Eulers angles. Then, Corollary 2.2.23 and
the calculus imply the subelliptic estimate
f W p,s (S3 ) Cp (X + )f
p,s+2 1 1 
p
2 (S3 )
1 < p < , s R.
There is an analogue of this estimate on arbitrary compact Lie groups, see [RW15].
Let us briey indicate an idea behind the proof of Theorem 2.2.22. In order
to use the theory of singular integral operators (according to Coifman and Weiss,
see Section A.4), we rst dene a suitable quasidistance on G.
Let Ad : G U(g) be the adjoint representation of G. Then by the PeterWeyl theorem it can be decomposed as a direct sum of irreducible representations,
,
Ad = (dim Z(G))1
0
where Z(G) is the centre of G, 1 is the trivial representation, and 0 is an index set
for the representations entering in this decomposition. Then we dene a smooth
nonnegative function
(d (x)),
(2.32)
r2 (x) := dim G Tr Ad(x) =
0
which is central, nondegenerate, and vanishes of the second order at the unit
element e G. It can be then checked that the function
d(x, y) := r(x1 y)
is the quasidistance in the sense of Section A.4. Consequently, one can check that
the operator A satises Calder
onZygmund conditions of spaces of homogeneous
type, in terms of the quasidistance above. Such a verication relies heavily on the
developed symbolic calculus, Leibniz rules for dierence operators, and criteria for
the weak (1,1) type in terms of suitably dened molliers. We refer to [RW15] for
further details of this construction.
Using the function r(x), one can rene the statement of Theorem 2.2.22.
Thus, let us dene the dierence operator associated with r2 (x), namely,
1
,
!
:= r2 = FG r2 (x) FG
is the second order dierence operator.
and we have that !
di 2 (G)
Theorem 2.2.26. Let A : C (G) D (G) be a leftinvariant linear continuous
operator on a compact Lie group G, and let k denote the smallest even integer
such that k > 12 dim G. Assume that the symbol A of A satises
!
k/2 A ()L (H ) C
(2.33)
86
Chapter 2. Quantization on compact Lie groups
as well as
D A ()L (H ) C

(2.34)
Then the operator A is of weak
for all multiindices  k 1 and all G.
type (1,1) and is bounded on Lp (G) for all 1 < p < .
We note that, comparing (2.33) to the condition (2.31) in Theorem 2.2.22,
only a single dierence condition of order k is required in Theorem 2.2.26. This
has interesting consequences, already in the case of the torus, as we will show in
Example 2.2.27.
Moreover, the assumption (2.34) can be rened further: namely, to form a
strongly admissible family of rst order dierence operators giving D in (2.28)
and (2.29), it is enough to take only k 0 , the set of the irreducible components
of the adjoint representation.
In all the theorems of this section an assumption that k is an even integer is
present. This seems to be related to the technical part of the argument, namely,
to the usage of the second order dierence operator !
that is naturally related to
the quasimetric on G as well as satises the nite Leibniz formula. The latter can
be derived from (2.30) using the decomposition
!
=
Dii ,
0 i=1
which follows from the denition of r2 (x) in (2.32). Thus, it satises
!
( ) = (!
) + (!
)
( Dij )( Dji ),
0 i,j=1
and becomes instrumental in establishing the relation between assumption (2.33)
and properties of the integral kernel of A in terms of the quasimetric. However, we
note also that the condition on the even number of analogous expressions appears
already in the multiplier theorem for biinvariant operators, established by rather
dierent methods by N. Weiss [Wei72].
Example 2.2.27. Let us consider now the case of the torus, G = Tn . In this case,
the leftinvariant operators take the form
2ix
Af (x) =
e
()f () with f () =
e2ix f (x)dx.
Tn
Zn
or, in other words,
() = ()f(), Zn .
Af
We take
r2 (x) = 2n
n
j=1
(e2ixj + e2ixj ),
2.2. Pseudodierential operators on compact Lie groups
so that
!
() = 2n()
87
(( + ej ) + ( ej )),
j=1
is
where Zn and ej is the j th unit vector in Zn . The appearing operator !
rather curious since it replaces the assumptions usually imposed on all highest
order dierence conditions as, for example, in the suitably modied toroidal version of Hormanders multiplier theorem% [H
& or60] (where one would need to make
assumptions on all dierences of order n2 + 1), or in Marcienkiewicz version of
multiplier theorem of Nikolskii [Nik77, Section 1.5.3] (where one imposes dierence
conditions up to order n). To clarify the nature of the operator !
, we give the
examples for T2 and T3 . As a consequence of Theorem 2.2.26 we get the following
statements. Let 1 < p < . Assume that
() C and  ( + ej ) () C,
for all Z2 and j = 1, 2, or Z3 and j = 1, 2, 3, respectively. Furthermore,
assume that
2 ()
1
(( + ej ) + ( ej )) C
4 j=1
for T2 ,
2 ()
1
(( + ej ) + ( ej )) C
6 j=1
for T3 ,
or
respectively. Then the operator A is bounded on Lp (T2 ) or Lp (T3 ), respectively.
We now drop the assumption of leftinvariance and consider general linear continuous operators from C (G) to D (G). Then we can assure the Lp boundedness provided we complement the dierences in with derivatives with
respect to x.
Theorem 2.2.28. Let A : C (G) D (G) be a linear continuous operator on a
compact Lie group G, and let k denote the smallest even integer such that k >
1
dim G
be an integer. Assume that the symbol
2 dim G. Let 1 < p < and let l >
p
A of A satises
k
Xx !
k/2 A (x, )L (H ) C
(2.35)
as well as
Xx D A (x, )L (H ) C

(2.36)
and for all multiindices and with  k 1 and  l. Then
for all G
the operator A is bounded on Lp (G).
We refer to [RW15] for the detailed proofs of all the results in this section.
88
Chapter 2. Quantization on compact Lie groups
2.2.5 Sharp G
arding inequality
The sharp G
arding inequality on Rn is an important lower bound for operators
with positive symbols, nding many applications in the theory of partial dierential equations of elliptic, parabolic and hyperbolic types. The original G
arding
inequality for elliptic operators has been established by G
arding in [G
ar53]. It says
m
(Rn ), 0 < 1, is a symbol satisfying
that if p S,
Re p(x, ) cm ,
c > 0, for all x Rn and large enough, then the corresponding pseudodierential
operator
e2ix p(x, )f()d
p(x, D)f (x) =
Rn
satises the following lower bound: for every s R and every compact set K Rn
there exist some constants c0 , c1 such that
Re (p(x, D)f, f )L2 (Rn ) c0 f 2H m/2 (Rn ) c1 f 2H s (Rn )
(2.37)
holds for all f D(K). Its improvement, the socalled sharp G
arding inequality was
m
obtained by Hormander in [H
or66]. It says that if p S,
(Rn ), 0 < 1,
is a nonnegative symbol, p(x, ) 0 for all x, Rn , then the corresponding
pseudodierential operator satises the lower bound
Re (p(x, D)f, f )L2 (Rn ) cf 2H (m())/2 (Rn )
(2.38)
for all f D(K). This inequality was further generalised to systems by Lax and
Nirenberg [LN66], Kumanogo [Kg81], and Vaillancourt [Vai70]. It has been also
2
extended to regain two derivatives for the class S1,0
(Rn ) by Feerman and Phong
[FP78]. For expositions concerning sharp G
arding inequalities with dierent proofs
we refer to monographs of Kumanogo [Kg81], Taylor [Tay81], Lerner [Ler10], or
Friedrichs notes [Fri70]. There is also an approach based on constructions in space
variables rather than in frequency one, developed by Nagase [Nag77].
The situation with G
arding inequalities on manifolds is more complicated.
The main problem is that the assumption that the symbol of a pseudodierential
operator is nonnegative is harder to formulate since the full symbol is not invariantly dened. For second order pseudodierential operators, under the nonnegativity assumption on the principal symbol and certain nondegeneracy assumptions on the subprincipal symbol, a lower bound now known as MelinHormander inequality has been obtained by Melin [Mel71] and Hormander [H
or77].
The nondegeneracy conditions on the subprincipal symbol can be somehow relaxed, see [MPP07].
Nevertheless, in our setting we are assisted by the fact that the algebraic
structure of a Lie group gives us the notion of the full symbol in (2.19). This
2.2. Pseudodierential operators on compact Lie groups
89
symbol, however, is not needed for the standard G
arding inequality (2.37) since the
ellipticity is determined by the principal symbol only. Thus, the standard G
arding
inequality (2.37) on compact Lie groups has been established in [BGJR89] using
Langlands results for semigroups on Lie groups [Lan60].
Let us rst look at a possible assumption for the positivity of an operator in
the invariant situation. If an operator A is given by the convolution Af = f ,
we obtain
, f)2 (G)
d Tr f()
() f() ,
(Af, f )L2 (G) = ( f, f )L2 (G) = (f
=
G
where we used the Plancherel identity (2.13). On the other hand, according to
Section 1.5, A is rightinvariant, and according to Example 2.2.4 its symbol is
()(x). Thus, we get that A is a positive operator if and only if
A (x, ) = (x)
i.e. when (
the matrix
() is positive for all G,
()v, v)H 0 for all v H .
But this means that the symbol A is positive, A (x, ) 0 for all (x, ) G G.
Analogously, for leftinvariant operators Af = f , one sees that
(Af, f )L2 (G) = (f , f )L2 (G) = (
f, f)2 (G)
d Tr f()
() f() .
=
G
() is positive.
So again, A is a positive operator if and only if its symbol A () =
This motivates a hypothesis that the positivity of the matrixvalued symbol
on G would be an analogue of the positivity of the KohnNirenberg symbol on Rn .
Indeed, we have the following criterion, which for noninvariant operators becomes
a sucient condition:
Theorem 2.2.29. Let A m (G) be such that its matrixvalued symbol A is
positive, i.e.
A (x, ) 0 for all (x, ) G G.
Then there exists a constant c such that
Re (Af, f )L2 (G) cf 2H (m1)/2 (G)
for all f C (G).
The usual proofs of the sharp G
arding inequality on Rn (that is, the proofs
not relying on the antiWick quantization) make use of a positive approximation
of a pseudodierential operator, the socalled Friedrichs symmetrisation, approximating an operator with nonnegative symbol of order m by a positive operator
modulo an error of order m 1. This construction, indeed, allows one to gain
one derivative needed for the sharp G
arding inequality. Unfortunately, such an
approximation in the frequency variable seems to be less useful on a Lie group G
is not well adapted for such purpose. However, one
because the unitary dual G
90
Chapter 2. Quantization on compact Lie groups
can carry out, instead, a symmetrisation in the space variables using the symbolic
calculus of operators for the construction. In particular, it relies heavily on dealing
m
with the symbol class S1,
1 (G) dened in Section 2.2.3.
2
arding inequality leads to sevAs in the case of operators on Rn , the sharp G
eral further conclusions concerning the L2 boundedness of operators. For example,
pseudodierential operators of the rst order are bounded on L2 (Rn ) provided
their matrixvalued symbols are bounded:
Corollary 2.2.30. Let A 1 (G) be such that its matrixvalued symbol A satises
A (x, )L (H ) C
Then the operator A is bounded from L2 (G) to L2 (G).
for all (x, ) G G.
It can be also used to determine constants as bounds for operator norms
of mappings between L2 Sobolev spaces. For the proofs of the statements in this
section, as well as for further details we refer the reader to [RT11].
m
In the above, we concentrated on symbol classes S1,0
(G) of type (1, 0). However, certain conclusions can be made also for operators with symbols of type
(, ).
Proposition 2.2.31 (G
ardings inequality on G). Let 0 < 1 and m > 0.
2m
Let A Op S,
(G) be elliptic and such that A (x, ) 0 for all x and conitely
many . Then there are constants c1 , c2 > 0 such that for any function f H m (G)
the inequality
Re (Af, f )L2 c1 f 2H m c2 f 2L2
holds true.
The statement follows by the calculus from its special case m = . We
refer to [RW14, Corollary 6.2] for the proofs.
Open Access. This chapter is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/),
which permits use, duplication, adaptation, distribution and reproduction in any medium
or format, as long as you give appropriate credit to the original author(s) and the source,
a link is provided to the Creative Commons license and any changes made are indicated.
The images or other third party material in this chapter are included in the works
Creative Commons license, unless indicated otherwise in the credit line; if such material
is not included in the works Creative Commons license and the respective action is not
permitted by statutory regulation, users will need to obtain permission from the license
holder to duplicate, adapt or reproduce the material.
Chapter 3
Homogeneous Lie groups
By denition a homogeneous Lie group is a Lie group equipped with a family of
dilations compatible with the group law. The abelian group (Rn , +) is the very
rst example of homogeneous Lie group. Homogeneous Lie groups have proved to
be a natural setting to generalise many questions of Euclidean harmonic analysis.
Indeed, having both the group and dilation structures allows one to introduce
many notions coming from the Euclidean harmonic analysis. There are several
important dierences between the Euclidean setting and the one of homogeneous
Lie groups. For instance the operators appearing in the latter setting are usually
more singular than their Euclidean counterparts. However it is possible to adapt
the technique in harmonic analysis to still treat many questions in this more
abstract setting.
As explained in the introduction (see also Chapter 4), we will in fact study
operators on a subclass of the homogeneous Lie group, more precisely on graded
Lie groups. A graded Lie group is a Lie group whose Lie algebra admits a (N)gradation. Graded Lie groups are homogeneous and in fact the relevant structure
for the analysis of graded Lie groups is their natural homogeneous structure and
this justies presenting the general setting of homogeneous Lie groups. From the
point of view of applications, the class of graded Lie groups contains many interesting examples, in fact all the ones given in the introduction. Indeed these groups
appear naturally in the geometry of certain symmetric domains and in some subelliptic partial dierential equations. Moreover, they serve as local models for contact
manifolds and CR manifolds, or for more general Heisenberg manifolds, see the
discussion in the Introduction.
The references for this chapter of the monograph are [FS82, ch. I] and
[Goo76], as well as Fulvio Riccis lecture notes [Ric]. However, our conventions
and notation do not always follow the ones of these references. The treatment
in this chapter is, overall, more general than that in the above literature since
we also consider distributions and kernels of complex homogeneous degrees and
adapt our analysis for subsequent applications to Sobolev spaces and to the op The Editor(s) (if applicable) and The Author(s) 2016
V. Fischer, M. Ruzhansky, Quantization on Nilpotent Lie Groups, Progress in Mathematics 314,
DOI 10.1007/9783319295589_3
91
92
Chapter 3. Homogeneous Lie groups
erator quantization developed in the following chapters. Especially, our study of
complex homogeneities allows us to deal with complex powers of operators (e.g.
in Section 4.3.2).
3.1
Graded and homogeneous Lie groups
In this section we present the denition and the rst properties of graded Lie
groups. Since many of their properties can be explained in the more general setting
of homogeneous Lie groups, we will also present these groups.
3.1.1 Denition and examples of graded Lie groups
We start with denitions and examples of graded and stratied Lie groups.
Denition 3.1.1. (i) A Lie algebra g is graded when it is endowed with a vector
space decomposition (where all but nitely many of the Vj s are {0}):
g=
Vj
such that
[Vi , Vj ] Vi+j .
j=1
(ii) A Lie group is graded when it is a connected simply connected Lie group
whose Lie algebra is graded.
The condition that the group is connected and simply connected is technical
but important to ensure that the exponential mapping is a global dieomorphism
between the group and its Lie algebra.
The classical examples of graded Lie groups and algebras are the following.
Example 3.1.2 (Abelian case). The abelian group (Rn , +) is graded: its Lie algebra
Rn is trivially graded, i.e. V1 = Rn .
Example 3.1.3 (Heisenberg group). The Heisenberg group Hno given in Example
1.6.4 is graded: its Lie algebra hno can be decomposed as
h no = V 1 V 2
where
o
V1 = ni=1
RXi RYi
and
V2 = RT.
(For the notation, see Example 1.6.4 in Section 1.6.)
Example 3.1.4 (Upper triangular matrices). The group Tno of no no matrices
which are upper triangular with 1 on the diagonal is graded: its Lie algebra tno of
no no upper triangular matrices with 0 on the diagonal is graded by
tno = V1 . . . Vno 1
where
no j
Vj = i=1
REi,i+j .
(For the notation, see Example 1.6.5 in Section 1.6.) The vector space Vj is formed
by the matrices with only nonzero coecients on the jth upper odiagonal.
3.1. Graded and homogeneous Lie groups
93
As we will show in Proposition 3.1.10, a graded Lie algebra (hence possessing
a natural dilation structure) must be nilpotent. The converse is not true, see
Remark 3.1.6, Part 2.
Examples 3.1.23.1.4 are stratied in the following sense:
Denition 3.1.5. (i) A Lie algebra g is stratied when g is graded, g =
j=1 Vj ,
and the rst stratum V1 generates g as an algebra. This means that every
element of g can be written as a linear combination of iterated Lie brackets
of various elements of V1 .
(ii) A Lie group is stratied when it is a connected simply connected Lie group
whose Lie algebra is stratied.
Remark 3.1.6. Let us make the following comments on existence and uniqueness
of gradations.
1. A gradation over a Lie algebra is not unique: the same Lie algebra may
admit dierent gradations. For example, any vector space decomposition of
Rn yields a graded structure on the group (Rn , +). More convincingly, we
can decompose the 3 dimensional Heisenberg Lie algebra h1 as
h1 =
3
Vj
with
V1 = RX1 , V2 = RY1 , V3 = RT.
j=1
This last example can be easily generalised to nd several gradations on the
Heisenberg groups Hno , no = 2, 3, . . . , which are not the classical ones given
in Example 3.1.3. Another example would be
h1 =
8
Vj
with
V3 = RX1 , V5 = RY1 , V8 = RT,
(3.1)
j=1
and all the other Vj = {0}.
2. A gradation may not even exist. The rst obstruction is that the existence
of a gradation implies nilpotency; in other words, a graded Lie group or a
graded Lie algebra are nilpotent, as we shall see in the sequel (see Proposition
3.1.10). Even then, a gradation of a nilpotent Lie algebra may not exist. As a
curiosity, let us mention that the (dimensionally) lowest nilpotent Lie algebra
which is not graded is the seven dimensional Lie algebra given by the following
commutator relations:
[X1 , Xj ] = Xj+1
for j = 2, . . . , 6,
[X2 , X3 ] = X6 ,
[X2 , X4 ] = [X5 , X2 ] = [X3 , X4 ] = X7 .
They dene a seven dimensional nilpotent Lie algebra of step 6 (with basis
{X1 , . . . , X7 }). It is the (dimensionally) lowest nilpotent Lie algebra which
is not graded. See, more generally, [Goo76, ch.I 3.2].
94
Chapter 3. Homogeneous Lie groups
3. To go back to the problem of uniqueness, dierent gradations may lead to
morally equivalent decompositions. For instance, if a Lie algebra g is graded
by g =
j=1 Vj then it is also graded by g = j=1 Wj where W2j +1 = {0} and
W2j = Vj . This last example motivates the presentation of homogeneous Lie
groups: indeed graded Lie groups are homogeneous and the natural homogeneous structure for the graded Lie algebra
g =
j=1 Vj = j=1 Wj
is the same for the two gradations.
Moreover, the relevant structure for the analysis of graded Lie groups
is their natural homogeneous structure.
4. There are plenty of graded Lie groups which are not stratied, simply because
the rst vector subspace of the gradation may not generate the whole Lie
algebra (it may be {0} for example). This can also be seen in terms of dilations
dened in Section 3.1.2. Moreover, a direct product of two stratied Lie
groups is graded but may be not stratied as their stratication structures
may not match. We refer to Remark 3.1.13 for further comments on this
topic.
3.1.2 Denition and examples of homogeneous Lie groups
We now deal with a more general subclass of Lie groups, namely the class of
homogeneous Lie groups.
Denition 3.1.7.
mappings
(i) A family of dilations of a Lie algebra g is a family of linear
{Dr , r > 0}
from g to itself which satises:
the mappings are of the form
Dr = Exp(A ln r) =
1
(ln(r)A) ,
!
=0
where A is a diagonalisable linear operator on g with positive eigenvalues, Exp denotes the exponential of matrices and ln(r) the natural
logarithm of r > 0,
each Dr is a morphism of the Lie algebra g, that is, a linear mapping
from g to itself which respects the Lie bracket:
X, Y g, r > 0
[Dr X, Dr Y ] = Dr [X, Y ].
(ii) A homogeneous Lie group is a connected simply connected Lie group whose
Lie algebra is equipped with dilations.
3.1. Graded and homogeneous Lie groups
95
(iii) We call the eigenvalues of A the dilations weights or weights. The set of
dilations weights, or in other worlds, the set of eigenvalues of A is denoted
by WA .
We can realise the mappings A and Dr in a basis of Aeigenvectors as the
diagonal matrices
r 1
1
2
r 2
and
D
.
r
..
..
.
.
n
r n
The dilations weights are 1 , . . . , n .
Remark 3.1.8. Note that if {Dr } is a family of dilations of the Lie algebra g, then
r := Dr := Exp(A ln r) denes a new family of dilations {D
r , r > 0} for any
D
> 0. By adjusting if necessary, we may assume that the dilations weights
satisfy certain properties in order to compare dierent families of dilations and
in order to x one of such families. For example in [FS82], it is assumed that the
minimum eigenvalue is 1.
Graded Lie algebras are naturally equipped with dilations: if the Lie algebra
g is graded by
g =
j=1 Vj ,
then we dene the dilations
Dr := Exp(A ln r)
where A is the operator dened by AX = jX for X Vj .
The converse is true:
Lemma 3.1.9. If a Lie algebra g has a family of dilations such that the weights are
all rational, then g has a natural gradation.
Proof. By adjusting the weights (see Remark 3.1.8), we may assume that all the
eigenvalues are positive integers. Then the decomposition in eigenspaces gives the
the gradation of the Lie algebra.
Before discussing the dilations in the examples given in Section 3.1.1 and
other examples of homogeneous Lie groups, let us state the following crucial property.
Proposition 3.1.10. The following holds:
(i) A Lie algebra equipped with a family of dilations is nilpotent.
(ii) A homogeneous Lie group is a nilpotent Lie group.
96
Chapter 3. Homogeneous Lie groups
Proof of Proposition 3.1.10. Let {Dr = Exp(A ln r)} be the family of dilations.
By Remark 3.1.8, we may assume that the smallest weight is 1. For WA let
/ WA then we set
W g be the corresponding eigenspace of A. If R but
W := {0}.
Thus Dr X = r X for X W . Moreover, if X W and Y W then
Dr [X, Y ] = [Dr X, Dr Y ] = r+ [X, Y ]
and hence
[W , W ] W+ .
In particular, since 1 for WA , we see that the ideals in the lower series of
g (see (1.18)) satisfy
g(j) aj Wa .
Since the set WA is nite, it follows that g(j) = {0} for j suciently large. Consequently the Lie algebra g and its corresponding Lie group G are nilpotent.
Let G be a homogeneous Lie group with Lie algebra g endowed with dilations
{Dr }r>0 . By Proposition 3.1.10, the connected simply connected Lie group G
is nilpotent. We can transport the dilations to the group using the exponential
mapping expG = exp of G (see Proposition 1.6.6 (a)) in the following way: the
maps
r > 0,
expG Dr exp1
G ,
are automorphisms of the group G; we shall denote them also by Dr and call them
dilations on G. This explains why homogeneous Lie groups are often presented as
Lie groups endowed with dilations.
We may write
rx := Dr (x)
for r > 0 and x G.
The dilations on the group or on the Lie algebra satisfy
Drs = Dr Ds ,
r, s > 0.
As explained above, Examples 3.1.2, 3.1.3 and, 3.1.4 are naturally homogeneous Lie groups:
In Example 3.1.2: The abelian group (Rn , +) is homogeneous when equipped with
the usual dilations Dr x = rx, r > 0, x Rn .
In Example 3.1.3: The Heisenberg group Hno is homogeneous when equipped with
the dilations
rh = (rx, ry, r2 t),
h = (x, y, t) Rno Rno R.
The corresponding dilations on the Heisenberg Lie algebra hno are given by
Dr (Xj ) = rXj , Dr (Yj ) = rYj , j = 1, . . . , no , and Dr (T ) = r2 T.
3.1. Graded and homogeneous Lie groups
97
In Example 3.1.4: The group Tno is homogeneous when equipped with the dilations
dened by
[Dr (M )]i,j = rji [M ]i,j
1 i < j n o , M T no .
The corresponding dilations on the Lie algebra tno are given by
Dr (Ei,j ) = rji Ei,j
1 i < j no .
As already seen for the graded Lie groups, the same homogeneous Lie group
may admit various homogeneous structures, that is, a nilpotent Lie group or algebra may admit dierent families of dilations, even after renormalisation of the
eigenvalues (see Remark 3.1.8). This can already be seen from the examples in
the graded case (see Remark 3.1.6 part 1). These examples can be generalised as
follows.
Example 3.1.11. On Rn we can dene
Dr (x1 , . . . , xn ) = (r1 x1 , . . . , rn xn ),
where 0 < 1 . . . n , and on Hno we can dene
Dr (x1 , . . . , xno , y1 , . . . , yno , t) = (r1 x1 , . . . , rno xno , r1 y1 , . . . , rno yno , r t),
where j > 0, j > 0 and j + j = for all j = 1, . . . , no .
These families of dilations give graded structures whenever the weights j
for Rn and j , j , for Hno are all rational or, more generally, all in Q+ for
a xed R+ . From this remark it is not dicult to construct a homogeneous
nongraded structure: on R3 , consider the diagonal 3 3 matrix A with entries,
e.g., 1 and and 1 + .
Example 3.1.12. Continuing the example above, choosing the j and j s rational
in a certain way, it is also possible to nd a homogeneous structure for Hno such
that the corresponding gradation of hno =
j=1 Vj does exist but is necessarily
such that V1 = {0}: we choose j , j positive integers dierent from 1 but with
1 as greatest common divisor (for instance for no = 2, take 1 = 3, 2 = 2, 1 =
5, 2 = 6 and = 8). As an illustration for Corollary 4.1.10 in the sequel, with
this example, the homogeneous dimension is Q = 3 + 2 + 5 + 6 + 8 = 24 while the
least common multiple is o = 2 3 5 = 30, so we have here Q < o .
If nothing is specied, we assume that the groups (Rn , +) and Hno are endowed with their classical structure of graded Lie groups as described in Examples
3.1.2 and 3.1.3.
Remark 3.1.13. We continue with several comments following those given in Remark 3.1.6.
98
Chapter 3. Homogeneous Lie groups
1. The converse of Proposition 3.1.10 does not hold, namely, not every nilpotent
Lie algebra or group admits a family of dilations. An example of a nine dimensional nilpotent Lie algebra which does not admit any family of dilations
is due to Dyer [Dye70].
2. A direct product of two stratied Lie groups is graded but may be not stratied as their stratication structures may not match. This can be also seen
on the level of dilations dened in Section 3.1.2. Jumping ahead and using
the notion of homogeneous operators, we see that this remark may be an
advantage for example when considering the subLaplacian L = X 2 + Y 2 on
the Heisenberg group H1 . Then the operator
L + tk
for k N odd, becomes homogeneous on the direct product H1 R when it
is equipped with the dilation structure which is not the one of a stratied
Lie group, see Lemma 4.2.11 or, more generally, Remark 4.2.12.
3. In our denition of a homogeneous structure we started with dilations dened
on the Lie algebra inducing dilations on the Lie group. If we start with a Lie
group the situation may become slightly more involved. For example, R3 with
the group law
xy = (arcsinh(sinh(x1 ) + sinh(y1 )), x2 + y2 + sinh(x1 )y3 , x3 + y3 )
is a 2step nilpotent stratied Lie group, the rst stratum given by
X = cosh(x1 )1 x1 ,
Y = sinh(x1 )x2 + x3 ,
and their commutator is
T = [X, Y ] = x2 .
It may seem like there is no obvious homogeneous structure on this group
but we can see it going to its Lie algebra which is isomorphic to the Lie
algebra h1 of the Heisenberg group H1 . Consequently, the above group itself
is isomorphic to H1 with the corresponding dilation structure.
4. In fact, the same argument as above shows that if we dened a stratied
Lie group by saying that there is a collection of vector elds on it stratied
with respect to their commutation relations, then for every such stratied
Lie group there always exists a homogeneous stratied Lie group isomorphic
to it. Indeed, since the Lie algebra is stratied and has a natural dilation
structure with integer weights, we obtain the required homogeneous Lie group
by exponentiating this Lie algebra. We refer to e.g. [BLU07, Theorem 2.2.18]
for a detailed proof of this.
Rening the proof of Proposition 3.1.10, we can obtain the following technical result which gives the existence of an adapted basis of eigenvectors for the
dilations.
3.1. Graded and homogeneous Lie groups
99
Lemma 3.1.14. Let g be a Lie algebra endowed with a family of dilations {Dr , r >
0}. Then there exists a basis {X1 , . . . , Xn } of g, positive numbers 1 , . . . , n > 0,
and an integer n with 1 n n such that
t > 0
j = 1, . . . , n
Dt (Xj ) = tj Xj ,
(3.2)
and
[g, g] RXn +1 . . . RXn .
(3.3)
Moreover, X1 , . . . , Xn generate the algebra g, that is, any element of g can be
written as a linear combination of these vectors together with all their iterated Lie
brackets.
This result and its proof are due to ter Elst and Robinson (see [tER97,
Lemma 2.2]). Condition (3.2) says that {Xj }nj=1 is a basis of eigenvectors for the
mapping A given by
Dr = Exp(A ln r).
Condition (3.3) says that this basis can be chosen so that the rst n vectors of
this basis generate the whole Lie algebra and the others span (linearly) the derived
algebra [g, g].
Proof of Lemma 3.1.14. We continue with the notation of the proof of Proposition 3.1.10. For each weight WA , we choose a basis
{Y,1 , . . . , Y,d , Y,d +1 , . . . , Y,d } of W
such that {Y,d +1 , . . . , Y,d } is a basis of the subspace
"
!
#
[W , W ] .
W
Span
+ =
Since g = WA W , we have by construction that
[g, g] Span {Y,j : WA , d + 1 j d } .
Let h be the Lie algebra generated by
{Y,j : WA , 1 j d } .
(3.4)
We now label and order the weights, that is, we write
WA = {1 , . . . , m }
with 1 1 < . . . < m . It follows by induction on N = 1, 2 . . . , m that N
j=1 Wj
is contained in h and hence h = g and the set (3.4) generate (algebraically) g.
A basis with the required property is given by
Y1 ,1 , . . . , Y1 ,d , . . . , Ym ,1 , . . . , Ym ,dm
1
for X1 , . . . , Xn ,
and
Y1 ,d1 +1 , . . . , Y1 ,d1 , . . . , Ym ,dm +1 , . . . , Ym ,dm
for Xn +1 , . . . , Xn .
100
Chapter 3. Homogeneous Lie groups
3.1.3 Homogeneous structure
In this section, we shall be working on a xed homogeneous Lie group G of dimension n with dilations
{Dr = Exp(A ln r)}.
We denote by 1 , . . . , n the weights, listed in increasing order and with each value
listed as many times as its multiplicity, and we assume without loss of generality
(see Remark 3.1.8) that 1 1. Thus,
1 1 2 . . . n .
(3.5)
If the group G is graded, then the weights are also assumed to be integers with
one as their greatest common divisor (again see Remark 3.1.8).
By Proposition 3.1.10 the Lie group G is nilpotent connected simply connected. Thus it may be identied with Rn equipped with a polynomial law, using
the exponential mapping expG of the group (see Section 1.6). With this identication its unit element is 0 Rn and it may also be denoted by 0G or simply by
0.
We x a basis {X1 , . . . , Xn } of g such that
AXj = j Xj
for each j. This yields a Lebesgue measure on g and a Haar measure on G by
Proposition 1.6.6. If x or g denotes a point in G the Haar measure is denoted by
dx or dg. The Haar measure of a measurable subset S of G is denoted by S.
We easily check that
Q
Dr (S) = r S,
f (rx)dx = r
G
f (x)dx,
(3.6)
where
Q = 1 + . . . + n = TrA.
(3.7)
The number Q is larger (or equal) than the usual dimension of the group:
n = dim G Q,
and may replace it for certain questions of analysis. For this reason the number Q
is called the homogeneous dimension of G.
Homogeneity
Any function dened on G or on G\{0} can be composed with the dilations Dr .
Using property (3.6) of the Haar measure and the dilations, we have for any
measurable functions f and on G, provided that the integrals exist,
(f Dr )(x) (x) dx = rQ
f (x) (D r1 )(x) dx.
(3.8)
G
3.1. Graded and homogeneous Lie groups
101
Therefore, we can extend the map f f Dr to distributions via
f Dr , := rQ f, D r1 ,
f D (G), D(G).
(3.9)
We can now dene the homogeneity of a function or a distribution in the
same way:
Denition 3.1.15. Let C.
(i) A function f on G\{0} or a distribution f D (G) is homogeneous of degree
C (or homogeneous) when
f Dr = r f
for any r > 0.
(ii) A linear operator T : D(G) D (G) is homogeneous of degree C (or
homogeneous) when
T ( Dr ) = r (T ) Dr
for any D(G), r > 0.
Remark 3.1.16. We will also say that a linear operator T : E F , where E is a
Frechet space containing D(G) as a dense subset, and F is a Frechet space included
in D (G), is homogeneous of degree C when its restriction as an operator from
D(G) to D (G) is. For example, it will apply to the situation when T is a linear
operator from Lp (G) to some Lq (G).
Example 3.1.17 (Coordinate function). The coordinate function xj = [x]j given
by
G x = (x1 , . . . , xn ) xj = [x]j ,
(3.10)
is homogeneous of degree j .
Example 3.1.18 (Koranyi norm). The function dened on the Heisenberg group
Hno by
1/4
2
,
Hno (x, y, t) x2 + y2 + t2
where x and y denote the canonical norms of x and y in Rno , is homogeneous
of degree 1. It is sometimes called the Koranyi norm.
Example 3.1.19 (Haar measure). Equality (3.8) shows that the Haar measure,
viewed as a tempered distribution, is a homogeneous distribution of degree Q (see
(3.7)). We can write this informally as
d(rx) = rQ dx,
see (3.6).
102
Chapter 3. Homogeneous Lie groups
Example 3.1.20 (Dirac measure at 0). The Dirac measure at 0 is the probability
measure 0 given by
f d0 = f (0).
G
It is homogeneous of degree Q since for any D(G) and r > 0, we have
1
0 Dr , = rQ 0 , D r1 = rQ ( 0) = rQ (0) = rQ 0 , .
r
Example 3.1.21 (Invariant vector elds). Let X g be viewed as a leftinvariant
(cf. Section 1.3). We assume
vector eld X or a rightinvariant vector eld X
that X is in the j eigenspace of A. Then the left and rightinvariant dierential
are homogeneous of degree j . Indeed,
operators X and X
X(f Dr ) (x)
=
=
t=0 {f Dr (x expG (tX))} = t=0 {f (rx expG (rj tX))}
rj t =0 {f (rx expG (t X))} = rj (Xf )(rx),
and similarly for X.
The following properties are very easy to check:
Lemma 3.1.22. (i) Whenever it makes sense, the product of two functions, distributions or operators of degrees 1 and 2 is homogeneous of degree 1 2 .
(ii) Let T : D(G) D (G) be a homogeneous operator. Then its formal adjoint
and transpose T and T t , given by
(T f )g =
f (T g),
(T f )g =
f (T t g), f, g D(G),
G
are also homogeneous with degree and respectively.
Consequently for any nonzero multiindex = (1 , . . . , n ) Nn0 \{0}, the
function
n
1
(3.11)
x := x
1 . . . xn ,
and the operators
:= X
1 . . . X
n ,
:=
...
, X := X11 . . . Xnn and X
n
1
x
x1
xn
are homogeneous of degree
[] := 1 1 + . . . + n n .
(3.12)
Formula (3.12) denes the homogeneous degree of the multiindex . It is usually
dierent from the length of given by
 := 1 + . . . + n .
3.1. Graded and homogeneous Lie groups
103
are dened
For = 0, the function x and the operators ( x
) , X , X
to be equal, respectively, to the constant function 1 and the identity operator I,
which are of degree [] := 0.
) , X
With this convention for each Nn0 , the dierential operators ( x
and X are of order  but of homogeneous degree [].
One easily checks for 1 , 2 Nn0 that
[1 ] + [2 ] = [1 + 2 ],
1  + 2  = 1 + 2 .
Proposition 3.1.23. Let the operator T be homogeneous of degree T and let f be
a function or a distribution homogeneous of degree f . Then, whenever T f makes
sense, the distribution T f is homogeneous of degree f T .
In particular, if f D (G) is homogeneous of degree , then
f, f
X f, X
are homogeneous of degree [].
Proof. The rst claim follows from the formal calculation
(T f ) Dr = rT T (f Dr ) = rT T (rf f ) = rT +f T f.
f and f are well
The second claim follows from the rst one since X , X
dened on distributions and are homogeneous of the same degree [] given by
(3.12).
3.1.4 Polynomials
By Propositions 3.1.10 and 1.6.6 we already know that the group law is polynomial.
This means that each [xy]j is a polynomial in the coordinates of x and of y. The
homogeneous structure implies certain additional properties of this polynomial.
Proposition 3.1.24. For any j = 1, . . . , n, we have
[xy]j = xj + yj +
cj,, x y .
,Nn
0 \{0}
[]+[]=j
In particular, this sum over [] and [] can involve only coordinates in x or y with
degrees of homogeneity strictly less than j .
For example,
for 1 :
[xy]1
x 1 + y1 ,
for 2 :
[xy]2
x 2 + y2 +
c2,, x y ,
[]=[]=1
for 3 :
[xy]3
x 3 + y3 +
[]=1 , []=2
or []=2 , []=1
c3,, x y ,
104
Chapter 3. Homogeneous Lie groups
and so on.
Proof. Let j = 1, . . . , n. From the BakerCampbellHausdor formula (see Theorem 1.3.2) applied to the two vectors X = x1 X1 + . . . + xn Xn and Y = y1 X1 +
. . . + yn Xn of g, we have with our notation that
[xy]j = xj + yj + Rj (x, y)
where Rj (x, y) is a polynomial in x1 , y1 , . . . , xn , yn . Moreover, Rj must be a nite
linear combination of monomials x y with  +  2:
cj,, x y .
Rj (x, y) =
,Nn
0
+2
We now use the dilations. Since the function xj is homogeneous of degree j ,
we easily check
Rj (rx, ry) = rj Rj (x, y)
for any r > 0 and this forces all the coecients cj,, with [] + [] = j to be
zero. The formula follows.
Recursively using Proposition 3.1.24, we obtain for any Nn0 \{0}:
n
1
c1 ,2 ()x1 y 2 ,
(3.13)
(xy) = [xy]
1 . . . [xy]n =
1 ,2 Nn
0
[1 ]+[2 ]=[]
with
c1 ,0 () =
0 if 1 =
1 if 1 =
and
c0,2 () =
0 if 2 =
.
1 if 2 =
(3.14)
Denition 3.1.25. A function P on G is a polynomial if P expG is a polynomial
on g.
For example the coordinate functions x1 , . . . , xn dened in (3.10) or, more
generally, the monomials x dened in (3.11) are (homogeneous) polynomials on
G.
It is clear that every polynomial P on G can be written as a unique nite
linear combination of the monomials x , that is,
c x ,
(3.15)
P =
Nn
0
where all but nitely many of the coecients c C vanish. The homogeneous
degree of a polynomial P written as (3.15) is
D P := max{[] : Nn0 with c = 0},
3.1. Graded and homogeneous Lie groups
105
which is often dierent from its isotropic degree:
d P := max{ : Nn0 with c = 0}.
For example on Hno , 1 + t is a polynomial of homogeneous degree 2 but
isotropic degree 1.
Denition 3.1.26. We denote by P(G) the set of all polynomials on G. For any
M 0 we denote by PM the set of polynomials P on G such that D P M
iso
the set of polynomials on G such that d P M . We also dene in
and by PM
the same way P<M , P=M , PM and so on, and similarly for P iso .
It is clear that P(G) is an algebra, for pointwise multiplication, which is
generated by the xj s.
It is not dicult to see:
iso
of P are nite dimensional with
Lemma 3.1.27. The subspaces PM and PM
bases {x : N0 , [] M } and {x : Nn0 ,  M }, respectively.
Furthermore,
iso
Pn M .
M 0
PM PM
Proof. The rst part of the lemma is clear. For the second, because of (3.5), we
have
 [] n .
(3.16)
Nn0
Therefore,
P P
d P D P n d P,
and the inclusions follow.
By Proposition 3.1.24, [xy]j is in Pj as a function of x for each y, and also
as a function of y for each x. Hence each subspace PM is invariant under left and
iso
iso
(unless PM
C or G = (Rn , +));
right translation. This is not the case for PM
consequently, it will not be of much use to us.
3.1.5 Invariant dierential operators on homogeneous Lie groups
We now investigate expressions for left and rightinvariant operators on homogeneous Lie groups.
j , for any
Proposition 3.1.28. The left and rightinvariant vector elds Xj and X
j = 1, . . . , n, can be written as
+
Pj,k
=
+
Pj,k
Xj =
xj
xk
xj
xk
j
X
1kn
j <k
1kn
j <k
1kn
j <k
1kn
j <k
+
Qj,k
=
+
Qj,k ,
xj
xk
xj
xk
106
Chapter 3. Homogeneous Lie groups
where Pj,k and Qj,k are homogeneous polynomials on G of homogeneous degree
k j > 0.
Proof. For any x G, we denote by Lx : G G the lefttranslation, i.e. Lx (y) =
xy. Let j = 1, . . . , n. Recall that Xj is the dierential operator invariant under
lefttranslation which agrees with x
at 0, that is, for any f C (G) and xo G,
j
we have
f
(Xj f ) Lxo (0) = Xj (f Lxo )(0) and Xj (f )(0) =
(0).
xj
Thus
(Xj f )(xo )
(Xj f ) Lxo (0) = Xj (f Lxo )(0) =
n
f
[xo x]k
(xo )
(0),
xk
xj
(f Lxo )(0)
xj
k=1
by the chain rule. But by Proposition 3.1.24,
[xo x]k
[xo ]k + xk +
(0)
(0) =
ck,, xo x
xj
xj
,Nn
0 \{0}
=
j,k +
[]+[]=k
ck,, x
o,
=ej , Nn
0 \{0}
[]+[]=k
where ej is the multiindex with 1 in the jth place and zeros elsewhere, and j,k
is the Kronecker delta. The assertion for Xj now follows immediately, and the
j is proved in the same way using right translations.
assertion for X
Proposition 3.1.28 gives, in particular,
for n :
Xn
for n1 :
Xn1
for n2 :
Xn2
,
xn
+ Pn1,n
,
xn1
xn
+ Pn2,n1
+ Pn2,n
,
xn2
xn1
xn
so that
xn
xn1
xn2
Xn ,
Xn1 Pn1,n Xn ,
Xn2 Pn2,n1 (Xn1 Pn1,n Xn ) Pn2,n Xn ,
3.1. Graded and homogeneous Lie groups
107
and so forth, with similar formulae for the rightinvariant vector elds. This shows
j s,
s of the same sort as for the Xj s and X
that there are formulas for the x
j
that is,
k ,
j +
= Xj +
pj,k Xk = X
qj,k X
(3.17)
xj
1kn
j <k
1kn
j <k
where pj,k and qj,k are homogeneous polynomials on G of homogeneous degree
k j > 0.
Remark 3.1.29. 1. Given the formulae above and the condition on the degree,
it is not dicult to see that the Pj,k and Qj,k in Proposition 3.1.28 and the
pj,k and qj,k in (3.17), with k > j , are polynomials in (x1 , . . . , xk1 ) and
k and respectively.
commute with Xk , X
xk
2. The rst part of Proposition 3.1.28 and its proof are valid for any nilpotent
Lie group (see Remark 1.6.7, part (1)). In our setting here, the homogeneous structure implies the additional property that the Pj,k and Qj,k are
homogeneous.
Corollary 3.1.30. For any Nn0 \{0},
X
Nn
0 , 
[][]
=
P, X
p, ,
X
Nn
0 , 
[][]
Q, X =
Nn
0 , 
[][]
X q, ,
Nn
0 , 
[][]
where P, , p, , Q, , q, are homogeneous polynomials of homogeneous degree
[] [].
Proof. By Proposition 3.1.28 we obtain recursively for any Nn0 \{0} that
X =
Nn
0 , 
[][]
P,
(3.18)
with P, homogeneous polynomial of degree [] []. Similar formulae yield X
in terms of the x s.
Recursively from (3.17), we also obtain similar formulae for x
in terms
.
of the X or X
The assertion comes form combining these formulae, with a similar argument
for p, and q, .
108
Chapter 3. Homogeneous Lie groups
Corollary 3.1.31. For any M 0, the maps
(ii)
(iii)
{X P (0)}Nn , []M ,
0
X P (0)
,
n
(i)
P (0)
Nn
0 , []M
N0 , []M
are linear isomorphisms from PM to Cdim PM . Also, the maps
(ii)
(iii)
{X P (0)}Nn , []=M ,
0
X P (0)
,
n
(i)
P (0)
Nn
0 , []=M
N0 , []=M
are linear isomorphisms from P=M to Cdim P=M .
Proof. By Lemma 3.1.27, the vector subspace PM of P is nite dimensional, with
basis {x : Nn0 , [] M }. Hence case (i) is a simple consequence of Taylors
Theorem on Rn .
Note that in the formula (3.18), P, is a constant function when [] = []
and P, (0) = 0 when [] > []. Hence
X 0 =
P,
Nn
0 , 
[]=[]
.
0
and .
We have similar result from the other formulae relating X , X
x
Cases (ii) and (iii) follow from these observations together with case (i). The
case of the homogeneous polynomials of order M is similar.
We may use the following property without referring to it.
Corollary 3.1.32. Let , Nn0 . The dierential operator X X is a linear combination of X with [] Nn0 , [] = [] + []:
X X =
c,, X .
(3.19)
Nn
0 , +
[]=[]+[]
is a linear combination of X
with [] Nn ,
X
The dierential operator X
0
  +  and [] = [] + [].
3.1. Graded and homogeneous Lie groups
109
Proof. The dierential operator X X is a leftinvariant dierential operator of
order  +  by (3.18), and it is a linear combination of X ,   +  (see
Section 1.3),
X X =
c,, X .
Nn
0 , +
By homogeneity, for any r > 0 and any function f C (G), we have on one
hand,
X X (f Dr ) = r[]+[] (X X f ) Dr ,
and on the other hand,
X X (f Dr )
c,, X (f Dr )
Nn
0 , +
c,, r[] (X f ) Dr .
Nn
0 , +
Choosing f suitably (for example f being polynomials of homogeneous degree
at most [] + [], see Corollary 3.1.31), this implies that if [] + [] = [] then
c,, = 0, showing (3.19).
The property for the rightinvariant vector elds is similar.
3.1.6 Homogeneous quasinorms
We can dene an Euclidean norm E on g by declaring the Xj s to be orthonormal.
We may also regard this norm as a function on G via the exponential mapping,
that is,
xE =  exp1
G xE .
However, this norm is of limited use for our purposes, since it does not interact in
a simple fashion with dilations. We therefore dene:
Denition 3.1.33. A homogeneous quasinorm is a continuous nonnegative function
G x x [0, ),
satisfying
(i) (symmetric) x1  = x for all x G,
(ii) (1homogeneous) rx = rx for all x G and r > 0,
(iii) (denite) x = 0 if and only if x = 0.
The  ball centred at x G with radius R > 0 is dened by
B(x, R) := {y G : x1 y < R}.
110
Chapter 3. Homogeneous Lie groups
Remark 3.1.34. With such denition, we have for any x, xo G, R > 0,
xo B(x, R) = B(xo x, R),
(3.20)
since
1 1
z xo B(x, R) x1
xo z < R z B(xo x, R).
o z B(x, R) x
In particular, we see that
B(x, r) = xB(0, r).
It is also easy to check that
B(0, r) = Dr (B(0, 1)).
Note that in our denition of quasiballs, we choose to privilege the left
translations. Indeed, the set {y G : yx1  < R} may also be dened as
a quasiball but one would have to use the right translation instead of the left
xo translation to have a similar property to (3.20).
An important example of a quasinorm is given by Example 3.1.18 on the
Heisenberg group Hno . More generally, on any homogeneous Lie group, the following functions are homogeneous quasinorms:
(x1 , . . . , xn )p =
p1
xj 
p
j
(3.21)
j=1
for 0 < p < , and for p = :
1
(x1 , . . . , xn ) = max xj  j .
1jn
(3.22)
In Denition 3.1.33 we do not require a homogeneous quasinorm to be
smooth away from the origin but some authors do. Quasinorms with added regularity always exist as well but, in fact, a distinction between dierent quasinorms
is usually irrelevant for many questions of analysis because of the following property:
Proposition 3.1.35. (i) Every homogeneous Lie group G admits a homogeneous
quasinorm that is smooth away from the unit element.
(ii) Any two homogeneous quasinorms   and   on G are mutually equivalent:
in the sense that
a, b > 0
x G
ax x bx .
3.1. Graded and homogeneous Lie groups
111
Proof. Let us consider the function
(r, x) = Dr x2E =
r2j x2j .
j=1
Let us x x = 0. The function (r, x) is continuous, strictly increasing in r
and satises
(r, x) 0 and (r, x) +.
r+
r0
Therefore, there is a unique r > 0 such that Dr xE = 1. We set xo := r1 .
Hence we have dened a map
G\{0} x x1
o (0, )
which is the implicit function for (r, x) = 1. This map is smooth since the
function (r, x) is smooth from (0, +) G\{0} to (0, ) and r (r, x) is always
dierent from zero. Setting 0G o := 0, the map  o clearly satises the properties
of Denition 3.1.33. This shows part (i).
For Part (ii), it is sucient to prove that any homogeneous quasinorm is
equivalent to  o constructed above. Before doing so, we observe that the unit
spheres in the Euclidean norm and the homogeneous quasinorm  o coincide,
that is,
S := {x G : xE = 1} = {x G : xo = 1}.
Let   be any other homogeneous norm. Since it is a denite function (see
(iii) of Denition 3.1.33) its restriction to S is never zero. By compactness of S
and continuity of  , there are constants a, b > 0 such that
x S
a x b.
For any x G\{0}, let t > 0 be given by t1 = xo . We have Dt x S, and thus
a Dt x b
and
axo = t1 a x t1 b = bxo .
The conclusion of Part (ii) follows.
Remark 3.1.36. If G is graded, the formula (3.21) for p = 21 . . . n gives another
concrete example of a homogeneous quasinorm smooth away from the origin since
x xpp is then a polynomial in the coordinate functions {xj }.
Proposition 3.1.35 and our examples of homogeneous quasinorms show that
the usual Euclidean topology coincides with the topology associated with any
homogeneous quasinorm:
112
Chapter 3. Homogeneous Lie groups
Proposition 3.1.37. If   is a homogeneous quasinorm on G Rn , the topology
induced by the  balls
B(x, R) := {y G : x1 y < R},
x G and R > 0, coincides with the Euclidean topology of Rn .
Any closed ball or sphere for any homogeneous quasinorm is compact. It
is also bounded with respect to any norm of the vector space Rn or any other
homogeneous quasinorm on G.
Proof of Proposition 3.1.37. It is a routine exercise of topology to check that the
equivalence of norm given in Proposition 3.1.35 implies that the topology induced
by the balls of two dierent homogeneous quasinorms coincide. Hence we can
choose the norm   given by (3.22) and the corresponding balls
B (x, R) := {y G : x1 y < R}.
We also consider the supremum Euclidean norm given by
(x1 , . . . , xn )E, = max xj ,
1jn
and its corresponding balls
BE, (x, R) := {y G :  x + yE, < R}.
That the topologies induced by the two families of balls
{B (x, R)}xG,R>0
and
{BE, (x, R)}xG,R>0
must coincide follows from the following two observations. Firstly it is easy to
check for any R (0, 1)
1
B (0, R 1 ) BE, (0, R) B (0, R n ).
Secondly for each x G, the mappings x : y x1 y and E,x : y x + y are
two smooth dieomorphisms of Rn . Hence these mappings are continuous with
continuous inverses (with respect to the Euclidean topology). Furthermore, by
Remark 3.1.34, we have
x (B (x, R)) = B (0, R)
and
E,x (BE, (x, R)) = BE, (0, R).
The second part of the statement follows from the rst and from the continuity of homogeneous quasinorms.
The next proposition justies the terminology of quasinorm by stating that
every homogeneous quasinorm satises the triangle inequality up to a constant,
the other properties of a norm being already satised.
3.1. Graded and homogeneous Lie groups
113
Proposition 3.1.38. If   is a homogeneous quasinorm on G, there is a constant
C > 0 such that
xy C (x + y)
x, y G.
:= {x : x 1} be its associated closed
Proof. Let  be a quasinorm on G. Let B
unit ball. By Proposition 3.1.37, B is compact. As the product law is continuous
is also compact. Therefore, there is a
(even polynomial), the set {xy : x, y B}
constant C > 0 such that
x, y B
xy C.
Let x, y G. If both of them are 0, there is nothing to prove. If not, let t > 0 be
so that
given by t1 = x + y > 0. Then Dt (x) and Dt (y) are in B,
txy = Dt (xy) = Dt (x)Dt (y) C,
and this concludes the proof.
Note that the constant C in Proposition 3.1.38 satises necessarily C 1
since 0 = 0 implies x Cx for all x G. It is natural to ask whether
a homogeneous Lie group G may admit a homogeneous quasinorm   which
is actually a norm or, equivalently, which satises the triangle inequality with
constant C = 1. For instance, on the Heisenberg group Hno , the homogeneous
quasinorm given in Example 3.1.18 turns out to be a norm (cf. [Cyg81]). In the
stratied case, the norm built from the control distance of the subLaplacian, often
called the CarnotCaratheodory distance, is also 1homogeneous (see, e.g., [Pan89]
or [BLU07, Section 5.2]). This can be generalised to all homogeneous Lie groups.
Theorem 3.1.39. Let G be a homogeneous Lie group. Then there exist a homogeneous quasinorm on G which is a norm, that is, a homogeneous quasinorm  
which satises the triangle inequality
xy x + y
x, y G.
A proof of Theorem 3.1.39 by Hebisch and Sikora uses the correspondence
between homogeneous norms and convex sets, see [HS90]. Here we sketch a dierent proof. Its idea may be viewed as an adaptation of a part of the proof that the
control distance in the stratied case is a distance. Our proof may be simpler than
the stratied case though, since we dene a distance without using horizontal
curves.
Sketch of the proof of Theorem 3.1.39. If : [0, T ] G is a smooth curve, its
tangent vector (to ) at (to ) is usually dened as the element of the tangent
space T(to ) G at (to ) such that
d
, f C (G).
(to )(f ) = f ((t))
dt
t=to
114
Chapter 3. Homogeneous Lie groups
It is more convenient for us to identify the tangent vector of at (to ) with an
element of the Lie algebra g = T0 G. We therefore dene (to ) g via
d
1
(to )(f ) := f ((to ) (t))
, f C (G).
dt
t=to
We now x a basis {Xj }nj=1 of g such that Dr Xj = rj Xj . We also dene
the map   : g [0, ) by
1
X := max xj  j ,
j=1,...,n
X=
xj Xj g.
j=1
Given a piecewise smooth curve : [0, T ] G, we dene its length adapted
to the group structure by
:=
()

(t) dt.
If x and y are in G, we denote by d(x, y) the inmum of the lengths ()
of the piecewise smooth curves joining x and y. Since two points x and y can
always be joined by a smooth compact curve, e.g. (t) = ((1 t)x) ty, the quantity
d(x, y) is always nite. Hence we have obtained a map d : G G [0, ). It is a
routine exercise to check that d is symmetric and satises the triangle inequality
in the sense that we have for all x, y, z G, that
d(x, y) = d(y, x)
and
d(x, y) d(x, z) + d(z, y).
r ()) = r()
and (z)
= (),
thus we
Moreover, one can check easily that (D
also have for all x, y, z G and r > 0, that
d(zx, zy) = d(x, y)
and
d(rx, ry) = rd(x, y).
(3.23)
Let us show that d is nondegenerate, that is, d(x, y) = 0 x = y. First let
E be the Euclidean norm on g Rn such that the basis {Xj }nj=1 is orthonormal.
We endow each tangent space Tx G with the Euclidean norm obtained by left
translation of the Euclidean norm  E . Hence we have for any smooth curve at
any point to
 (to )T(to ) G = 
(to )E .
n
Now we see that if X = j=1 xj Xj g is such that
XE, := max xj  1,
j=1,...,n
then
XE
XE, X .
3.1. Graded and homogeneous Lie groups
115
This implies that if : [0, T ] G is a smooth curve satisfying
 (t)T(t) G < 1,
t [0, T ]
(3.24)
then
() C (),
(3.25)
where is the usual length
() :=
0
 (t)T(t) G dt,
and C > 0 a positive constant independent of .
Let dG be the Riemaniann distance induced by our choice of metric on the
manifold G, that is, the inmum of the lengths () of the piecewise smooth curves
joining x and y. Very well known results in Riemaniann geometry imply that
dG induces the same topology as the Euclidean topology. Moreover, there exists
a small open set containing 0 such that any point in may be joined to 0 by
a smooth curve satisfying (3.24) at any point. Then (3.25) yields that we have
dG (0, x) Cd(0, x) for any x . This implies that d is nondegenerate since d
is invariant under lefttranslation and is 1homogeneous in the sense of (3.23),
Checking that the associated map x x = d(0, x) is a quasinorm concludes
the sketch of the proof of Theorem 3.1.39.
Even if homogeneous norms do exist, it is often preferable to use homogeneous
quasinorms. Because the triangle inequality is up to a constant in this case, we
do not necessarily have the inequality xy x Cy. However, the following
lemma may help:
Proposition 3.1.40. We x a homogeneous quasinorm   on G. For any f
C 1 (G\{0}) homogeneous of degree C, for any b (0, 1) there is a constant
C = Cb > 0 such that
f (xy) f (x) Cy xRe 1
whenever
y bx.
Indeed, applying it to a C 1 (G\{0}) homogeneous quasinorm, we obtain
b (0, 1) C = Cb > 0 x, y G y bx = xy x Cy. (3.26)
Proof of Proposition 3.1.40. Let f C 1 (G\{0}). Both sides of the desired inequality are homogeneous of degree Re so it suces to assume that x = 1
and y b. By Proposition 3.1.37 and the continuity of multiplication, the set
{xy : x = 1 and y b} is a compact which does not contain 0. So by the
(Euclidean) mean value theorem on Rn , we get
f (xy) f (x) CyE .
We conclude using the next lemma.
116
Chapter 3. Homogeneous Lie groups
The next lemma shows that locally a homogeneous quasinorm and the Euclidean norm are comparable:
Lemma 3.1.41. We x a homogeneous quasinorm   on G. Then there exist
C1 , C2 > 0 such that
1
C1 xE x C2 xEn
whenever
x 1.
Proof of Lemma 3.1.41. By Proposition 3.1.37, the unit sphere {y : y = 1} is
compact and does not contain 0. Hence the Euclidean norm assumes a positive
maximum C11 and a positive minimum C2n on it, for some C1 , C2 > 0.
Let x G. We may assume x = 0. Then we can write it as x = ry with
y = 1 and r = x. We observe that since
ry2E =
yj2 r2j ,
j=1
we have if r 1
rn yE ryE ryE .
Hence for r = x 1, we get
xE = ryE ryE xC11
xE = ryE rn yE xn C2n ,
and
implying the statement.
3.1.7 Polar coordinates
There is an analogue of polar coordinates on homogeneous Lie groups.
Proposition 3.1.42. Let G be a homogeneous Lie group equipped with a homogeneous quasinorm  . Then there is a (unique) positive Borel measure on the
unit sphere
S := {x G : x = 1},
such that for all f L1 (G), we have
f (x)dx =
f (ry)rQ1 d(y)dr.
G
(3.27)
In order to prove this claim, we start with the following averaging property:
Lemma 3.1.43. Let G be a homogeneous Lie group equipped with a homogeneous
quasinorm  . If f is a locally integrable function on G\{0}, homogeneous of
degree Q, then there exists a constant mf C (the average value of f ) such that
for all u L1 ((0, ), r1 dr), we have
f (x)u(x)dx = mf
u(r)r1 dr.
(3.28)
G
3.1. Graded and homogeneous Lie groups
117
The proof of Lemma 3.1.43 yields the formula for mf in terms of the homogeneous quasinorm  ,
f (x)dx.
(3.29)
mf =
1xe
However, in Lemma 3.1.45 we will give an invariant meaning to this value.
Proof of Lemma 3.1.43. Let f be locally integrable function on G\{0}, homogeneous of degree Q. We set for any r > 0,
1
f (x)dx
if r 1,
1xr
(r) :=
rx1 f (x)dx if r < 1.
The mapping : (0, ) C is continuous and one easily checks that
(rs) = (r) + (s)
for all r, s > 0,
by making the change of variable x sx and using the homogeneity of f . It
follows that (r) = (e) ln r and we set
mf := (e).
Then the equation (3.28) is easily satised when u is the characteristic function
of an interval. By taking the linear combinations and limits of such functions, the
equation (3.28) is also satised when u L1 ((0, ), r1 dr).
Proof of Proposition 3.1.42. For any continuous function f on the unit sphere S,
we dene the homogeneous function f on G\{0} by
f(x) := xQ f (x1 x).
Then f satises the hypotheses of Lemma 3.1.43. The map f mf is clearly a
positive functional on the space of continuous functions on S. Hence it is given
by integration against a regular positive measure (see, e.g. [Rud87, ch.VI]).
For u L1 ((0, ), r1 dr), we have
f (x1 x)u(x)dx =
f(x)xQ u(x)dx = mf
rQ1 u(r)dr
r=0
f (y)u(r)rQ1 d(y)dr.
=
0
Since linear combinations of functions of the form f (x1 x)u(x) are dense in
L1 (G), the proposition follows.
We view the formula (3.27) as a change in polar coordinates.
118
Chapter 3. Homogeneous Lie groups
Example 3.1.44. For 0 < a < b < and C, we have
1
(b a ) if = 0
Q
x
dx = C
ln ab
if = 0
a<x<b
with C = (S).
And if R and f is a measurable function on G such that f (x) = O(xQ )
then f is integrable either near if < 0, or near 0 if > 0.
The measure in the polar coordinates decomposition actually has a smooth
density. We will not need this fact and will not prove it here, but refer to [FR66]
and [Goo80].
Now, the polar change of coordinates depends on the choice of a homogeneous
quasinorm to x the unit sphere. But it turns out that the average value of the
(Q)homogeneous function considered in Lemma 3.1.43 does not. Let us prove
this fact for the sake of completeness.
Lemma 3.1.45. Let G be a homogeneous Lie group and let f be a locally integrable
function on G\{0}, homogeneous of degree Q.
Given a homogeneous quasinorm, let be the Radon measure on the unit
sphere S giving the polar change of coordinate (3.27). Then the average value of
f dened in (3.28) is given by
f d.
(3.30)
mf =
S
This average value mf is independent of the choice of the homogeneous quasinorm.
Proof of Lemma 3.1.45. For any homogeneous quasinorm, using the polar change
of coordinates (3.27), we obtain
b
f (x)dx =
f (rx)d(x)rQ1 dr
a<x<b
=
a
f (x)d(x)r1 dr =
b
a
r1 dr
S
f (x)d(x) =
ln
b
a
mf .
This shows (3.30), taking a = 1 and b = e, see (3.29) and the proof of Lemma
3.1.43.
Let   and   be two homogeneous quasinorms on G. We denote by
r := {x G : x r} and
B
r := {x G : x r},
B
the closed balls around 0 of radius r for   and   , respectively. By Proposi1 . We also
a B
tion 3.1.35, Part (ii), there exists a constant a > 0 such that B
B
2 and, with the usual sign convention for integration, we have
a B
have B
2a
=
=
+
.
2 \B
1
B
2 \B
B
a
1 \B
B
a
2 \B
B
2a
\B
B
a
2a
1 \B
B
a
3.1. Graded and homogeneous Lie groups
119
Using the homogeneities of f and of the Haar measure, we see, after the changes
of variables x = 2y and x = az, that
f (x)dx =
f (y)dy and
f (x)dx =
f (z)dz.
2 \B
B
2a
\B
B
a
2a
1 \B
B
a
Hence
\B
B
2
1
2 \B
1
B
f=
\B
B
2
1
f.
Using the rst computations of this proof, the left and right hand sides are equal to
(ln b/a)mf and (ln b/a)mf , respectively, where mf and mf are the average values
for   and   . Thus mf = mf .
3.1.8 Mean value theorem and Taylor expansion
Here we prove the mean value theorem and describe the Taylor series on homogeneous Lie groups. Naturally, the space C 1 (G) here is the space of functions f such
that Xj f are continuous on G for all j, etc. The following mean value theorem
can be partly viewed as a renement of Proposition 3.1.40.
Proposition 3.1.46. We x a homogeneous quasinorm   on G. There exist group
constants C0 > 0 and > 1 such that for all f C 1 (G) and all x, y G, we have
f (xy) f (x) C0
yj sup (Xj f )(xz).
j=1
zy
In order to prove this proposition, we rst prove the following property.
Lemma 3.1.47. The map : Rn G dened by
(t1 , . . . , tn ) = expG (t1 X1 ) expG (t2 X2 ) . . . expG (tn Xn ),
is a global dieomorphism.
Moreover, xing a homogeneous quasinorm   on G, there is a constant
C1 > 0 such that
(t1 , . . . , tn ) Rn , j = 1, . . . , n,
tj  j C1 (t1 , . . . , tn ).
The rst part of the lemma is true for any nilpotent Lie group (see Remark
1.6.7 Part (ii)). But we will not use this fact here.
Proof. Clearly the map is smooth. By the BakerCampbellHausdor formula
(see Theorem 1.3.2), the dierential d(0) : Rn T0 G is the isomorphism
d(0)(t1 , . . . , tn ) =
n
j=1
tj X j  0 ,
120
Chapter 3. Homogeneous Lie groups
so that is a local dieomorphism near 0 (this is true for any Lie group). More
precisely, there exist , C > 0 such that is a dieomorphism from U to the ball
B := {x G : x < } with
1 (B ) = U {(t1 , . . . , tn ) :
max tj  j < C }.
j=1,...,n
We now use the dilations and for any r > 0, we see that
(r1 t1 , . . . , rn tn )
hence
expG (r1 t1 X1 ) . . . expG (rn tn Xn )
(r expG (t1 X1 )) . . . (r expG (tn Xn ))
r (expG (t1 X1 ) . . . expG (tn Xn )) ,
(r1 t1 , . . . , rn tn ) = r(t1 , . . . , tn ).
(3.31)
If (t1 , . . . , tn ) = (s1 , . . . , sn ), formula (3.31) implies that for all r > 0, we
have
(r1 t1 , . . . , rn tn ) = (r1 s1 , . . . , rn sn ).
For r suciently small, this forces tj = sj for all j since is a dieomorphism on
U . So the map : Rn G is injective.
Moreover, any x G\{0} can be written as
x = ry
2
x
r :=
with
and
y := r1 x B (U ).
2
We may write y = (s1 , . . . sn ) with sj  j C and formula (3.31) then implies
1
that x = (t1 , . . . , tn ) is in (Rn ) with tj := rj sj satisfying tj  j C r. Setting
C1 = 2C /, the assertion follows.
Proof of Proposition 3.1.46. First let us assume that y = expG (tXj ). Then
f (xy) f (x)
=
=
t
=
0
s =s {f (x expG (s Xj ))} ds
s =0 {f (x expG (sXj ) expG (s Xj ))} ds
Xj f (x expG (sXj ))ds,
and hence
f (xy) f (x)
t sup Xj f (x expG (sXj ))
t sup Xj f (xz).
0st
zy
3.1. Graded and homogeneous Lie groups
121
Since  expG (sXj ) = s j  expG Xj  and hence y = t j  expG Xj , setting
C2 := max  expG Xk k ,
k=1,...,n
we obtain
f (xy) f (x) C2 yj sup Xj f (xz).
(3.32)
zy
We now prove the general case, so let y be any point of G. By Lemma 3.1.47,
it can be written uniquely as y = y1 y2 . . . yn with yj = expG (tj Xj ), and hence
1
yj  = t j  expG Xj  C1 C3 y
C3 := max  expG Xk ,
where
k=1,...,n
(3.33)
and C1 is as in Lemma 3.1.47. We write
f (xy) f (x) f (xy1 . . . yn ) f (xy1 . . . yn1 )
+f (xy1 . . . yn1 ) f (xy1 . . . yn2 ) + . . . + f (xy1 ) f (x),
and applying (3.32) to each term, we obtain
f (xy) f (x)
C2 yj j sup Xj f (xy1 . . . yj1 z).
zyj 
j=1
Let C4 1 be the constant of the triangle inequality for  (see Proposition 3.1.38).
If z yj , then z = y1 . . . yj1 z satises
z  C4 (y1 . . . yj1  + yj ) C4 C4 (y1 . . . yj2  + yj1 ) + yj 
C42 (y1 . . . yj2  + yj1  + yj ) . . . C4j1 (y1  + y2  + . . . yj )
C4j1 jC1 C3 y,
using (3.33). Therefore, setting := C4n nC1 C3 , using again (3.33), we have obtained
f (xy) f (x) C2
(C1 C3 y)
j=1
sup Xj f (xz ),
z y
completing the proof.
Remark 3.1.48. Let us make the following remarks.
1. In the same way, we can prove the following version of Proposition 3.1.46 for
rightinvariant vector elds: a homogeneous quasinorm   being xed on G,
there exists group constants C > 0 and b > 0 such that for all f C 1 (G)
and all x, y G, we have
f (yx) f (x) C
n
j=1
j f )(zx).
yj sup (X
zby
122
Chapter 3. Homogeneous Lie groups
2. If the homogeneous Lie group G is stratied, a more precise version of the
mean value theorem exists involving only the vector elds of the rst stratum,
see Folland and Stein [FS82, (1.41)], but we will not use this fact here.
3. The statement and the proof of the mean value theorem can easily be adapted
to hold for functions which are valued in a Banach space, the modulus being
replaced by the Banach norm.
Taylor expansion
In view of Corollary 3.1.31, we can dene Taylor polynomials:
Denition 3.1.49. The Taylor polynomial of a suitable function f at a point x G
of homogeneous degree M N0 is the unique P PM such that
Nn0 , [] M
X P (0) = X f (x).
More precisely, we have dened the left Taylor polynomial, and a similar
yields the right Taylor
denition using the rightinvariant dierential operators X
polynomial. However, in this monograph we will use only left Taylor polynomials.
We may use the following notation for the Taylor polynomial P of a function
f at x and for its remainder of order M :
(f )
Px,M := P
and
(f )
Rx,M (y) := f (xy) P (y).
(3.34)
(f )
For instance, Px,M (0) = f (x). We will also extend the notation for negative M
with
(f )
(f )
Px,M := 0 and Rx,M (y) := f (xy) when M < 0.
With this notation, we easily see (whenever it makes sense), the following
properties.
Lemma 3.1.50. For any M N0 , Nn0 and suitable function f , we have
(X f )
X Px,M = Px,M []
(f )
and
(X f )
X Rx,M = Rx,M [] .
(f )
Proof. It is easy to check that the polynomial Po := X Px,M is homogeneous of
degree M []. Furthermore, using (3.19), it satises for every Nn0 , such that
[] + [] M , the equality
(f )
X Po (0)
=
=
X X Px,M (0)
(f )
c,, X Px,M (0) =
(f )
+
[]=[]+[]
=
This shows the claim.
c,, X f (x)
+
[]=[]+[]
X X f (x).
3.1. Graded and homogeneous Lie groups
123
In Denition 3.1.49 the suitable functions f are distributions on a neighbourhood of x in G whose (distributional) derivatives X f are continuous in a
neighbourhood of x for [] M . We will see in the sequel that in order to control
(uniformly) a remainder of a function f of order M we would like f to be at least
(k + 1) times continuously dierentiable, i.e. f C k+1 (G), where k N0 is equal
to
(3.35)
M := max{ : Nn0 with [] M };
this is indeed a maximum over a nite set because of (3.16).
We can now state and prove Taylors inequality.
Theorem 3.1.51. We x a homogeneous quasinorm   on G and obtain a corresponding constant from the mean value theorem (see Proposition 3.1.46). For any
M N0 , there is a constant CM > 0 such that for all functions f C M +1 (G)
and all x, y G, we have
(f )
y[]
sup
(X f ) (xz) ,
Rx,M (y) CM
z M +1 y
M +1
[]>M
(f )
where Rx,M and M are dened by (3.34) and (3.35).
Theorem 3.1.51 for M = 0 boils down exactly to the mean value theorem as
stated in Proposition 3.1.46. Similar comments as in Remark 3.1.48 for the mean
value theorem are also valid for Taylors inequality.
(f )
Proof. Under the hypothesis of the theorem, a remainder Rx,M is always C 1 and
vanishes at 0. Let us apply the mean value theorem (see Proposition 3.1.46) at the
(X
(f )
f)
(X
f)
j0
1
, and so on as long as
point 0 to the remainders Rx,M , Rx,Mj0j , Rx,Mj(
j0 +j1 )
0
M (j0 + . . . + jk ) 0; using this together with Lemma 3.1.50, we obtain
n
(Xj f )
(f )
j0
0
y0 
sup Rx,M j (y1 ) ,
Rx,M (y0 ) C0
(Xj f )
0
R
(y
)
x,M j0 1
(Xj ...Xj0 f )
k
R
x,M (j0 +...+jk ) (yk )
C0
j0 =1
n
y1 y0 
y1 j1
j1 =1
..
.
C0
n
jk =1
yk 
jk
sup
y2 y1 
sup
yk+1 yk 
We combine these inequalities together, to obtain
(f )
y0 j0 +...+jk sup
Rx,M (y0 ) C0k+1 k
ji =1,...,n
i=0,...,k+1
(Xj Xj f )
1
0
R
(y
)
x,M (j0 +j1 ) 2 ,
(Xjk+1 ...Xj0 f )
R
x,M (j0 +...+jk+1 ) (yk ) .
yk+1  k+1 y0 
(X jk+1 ...X j0 f )
Rx,M (j +...+j
0
k+1
(y
)
.
k
)
124
Chapter 3. Homogeneous Lie groups
The process stops exactly for k = M by the very denition of M . For this value
of k, Corollary 3.1.32 and the change of discrete variable := j0 ej0 +. . . jk+1 ejk+1
(where ej denotes the multiindex with 1 in the jth place and zeros elsewhere)
yield the result.
Remark 3.1.52. 1. We can consider Taylor polynomials for rightinvariant vector elds. The corresponding Taylor estimates would then approximate f (yx)
with a polynomial in y. See Part 1 of Remark 3.1.48, about the mean value
theorem for the case of order 0. Note that in Theorem 3.1.51 we consider
f (xy) and its approximation by a polynomial in y.
2. If the homogeneous Lie group G is stratied, a more precise versions of Taylors inequality exists involving only the vector elds of the rst stratum, see
Folland and Stein [FS82, (1.41)], but we will not use this fact here.
3. The statement and the proof of Theorem 3.1.51 can easily be adapted to
hold for functions which are valued in a Banach space, the modulus being
replaced by the Banach norm.
4. One can derive explicit formulae for Taylors polynomials and the remainders
on homogeneous Lie groups, see [Bon09] (see also [ACC05] for the case of
Carnot groups), but we do not require these here.
As a corollary of Theorem 3.1.51 that will be useful to us later, the rightderivatives of Taylor polynomials and of the remainder will have the following
properties, slightly dierent from those for the left derivatives in Lemma 3.1.50.
Corollary 3.1.53. Let f C (G). For any M N0 and Nn0 , we have
P (f ) = P (Xx f (x ))
X
x,M
0,M []
and
R(f ) = R(Xx f (x )) .
X
x,M
0,M []
Proof. Recall from (1.12) that for any X g identied with a leftinvariant vector
eld, we have
y {f (xy)} = d f (xetX y)t=0 = Xx {f (xy)},
X
dt
and recursively, we obtain
y {f (xy)} = Xx {f (xy)}.
X
(3.36)
Therefore, we have
P (f ) (y) P (Xx f (x )) (y)
X
x,M
0,M []
y f (xy) R(f ) (y) Xx f (xy) R(Xx f (x )) (y)
=X
x,M
0,M []
R(f ) (y) + R(Xx f (x )) (y).
= X
x,M
0,M []
(3.37)
3.1. Graded and homogeneous Lie groups
125
By Corollary 3.1.30, we can write
R(f ) (y) =
X
x,M
Q, (y)X Rx,M (y)
(f )
, [][]
(X f )
, [][]
Q, (y)Rx,M [] (y),
where each Q, is a homogeneous polynomial of degree [] [].
Fixing a homogeneous quasinorm   on G, the Taylor inequality (Theorem
(X f (x ))
(X f )
and Rx,M [] implies that, for y 1,
3.1.51) applied to R0,Mx []
(X f (x ))
R0,Mx [] (y) CyM []+1
Hence
and
(X f )
Rx,M [] (y) CyM []+1 .
M []+1
R
.
X
x,M (y) Cy
(f )
Going back to (3.37), we have obtained that its left hand side can be estimated as
(X f (x ))
M []+1
x
P
.
X
x,M (y) P0,M [] (y) Cy
(f )
(X f (x ))
x
P
But X
x,M (y) P0,M [] (y) is a polynomial of homogeneous degree at most
M []. Therefore, this polynomial is identically 0. This concludes the proof of
Corollary 3.1.53.
(f )
3.1.9 Schwartz space and tempered distributions
The Schwartz space on a homogeneous Lie group G is dened as the Schwartz space
on any connected simply connected nilpotent Lie group, namely, by identifying G
with the underlying vector space of its Lie algebra (see Denition 1.6.8). The
vector space S(G) is naturally endowed with a Frechet topology dened by any of
a number of families of seminorms.
In the traditional Schwartz seminorm on Rn (see (1.13)) we can replace
(without changing anything for the Frechet topology):
and the isotropic degree  by X and the homogeneous degree [],
x
respectively, in view of Section 3.1.5,
the Euclidean norm by the norm  p given in (3.21), and then by any homogeneous norm since homogeneous quasinorms are equivalent (cf. Proposition 3.1.35).
Hence we choose the following family of seminorms for S(G), where G is a
homogeneous Lie group:
f S(G),N :=
sup
[]N, xG
(1 + x)N X f (x)
(N N0 ),
126
Chapter 3. Homogeneous Lie groups
after having xed a homogeneous quasinorm   on G.
Another equivalent family is given by a similar denition with the right replacing X .
invariant vector elds X
The following lemma proves, in particular, that translations, taking the inverse, and convolutions, are continuous operations on Schwartz functions.
Lemma 3.1.54. Let f S(G) and N N. Then we have
f y
CN (1 + y)N f S(G),N
(y G),
(3.38)
S(G),N
CN f S(G),(n +1)N where f(x) = f (x1 ), (3.39)
f
S(G),N
f y
CN (1 + y)(n +1)N f S(G),(n +1)2 N (y G). (3.40)
S(G),N
Moreover,
f y f
y0 0
S(G),N
f y f
and
S(G),N
y0 0.
(3.41)
The group convolution of two Schwartz functions f1 , f2 S(G) satises
f1 f2 S(G),N CN f1 S(G),N +Q+1 f2 S(G),N .
(3.42)
Proof. Let Co 1 be the constant of the triangle inequality, cf. Proposition 3.1.38.
We have easily that
(1 + x) Co (1 + y)(1 + yx).
x, y G
Thus,
f y
S(G),N
sup
[]N, xG
(Co (1 + y)(1 + yx)) X f (yx)
CoN (1 + y)N f S(G),N .
This shows (3.38).
For (3.39), using (1.11) and Corollary 3.1.30, we have
f )(x1 )
sup (1 + x)N (X
f
S(G),N
[]N, xG
sup
[]N, xG
CN
(1 + x)N  Q, X f (x1 )
Nn
0 , 
[][]
sup
[]n N, xG
(1 + x )N +[] X f (x )
by homogeneity of the polynomials Q, and (3.16).
Since f y = (f y 1 ), we deduce (3.40) from (3.38) and (3.39).
(3.43)
3.1. Graded and homogeneous Lie groups
127
By the mean value theorem (cf. Proposition 3.1.46),
f y f
=
sup (1 + x)N X f (yx) X f (x)
S(G),N
[]N, xG
n
j=1
n
j=1
yj
sup
(1 + x)N (Xj X f )(xz)
[]N
xG, zy
yj f S(G),N +n ,
(3.44)
and this proves (3.41) for the left invariance. The proof is similar for the right
invariance and is left to the reader.
Since using (3.43) we have
N
(1 + x) X (f1 f2 )(x)
(1 + x)N f1 (y) X f2 (y 1 x)dy
G
(1 + y)N f1 (y)(1 + y 1 x)N X f2 (y 1 x)dy
CoN
G
N
N
Co sup (1 + z) X f2 (z) (1 + y)N f1 (y)dy,
zG
we obtain (3.42) by the convergence in Example 3.1.44.
The space of tempered distributions S (G) is the (continuous) dual of S(G).
Hence a linear form f on S(G) is in S (G) if and only if
N N0 , C > 0
S(G)
f,  CS(G),N .
(3.45)
The topology of S (G) is given by the family of seminorms given by
f S (G),N := sup{f, , S(G),N 1},
f S (G), N N0 .
Now, with these denitions, we can repeat the construction in Section 1.5
and dene convolution of a distribution in S (G) with the test function in S(G).
Then we have
Lemma 3.1.55. For any f S (G) there exist N N and C > 0 such that
S(G)
x G
( f )(x) C(1 + x)N S(G),N .
(3.46)
The constant C may be chosen of the form C = C f S (G),N for some C and
N independent of f .
For any f S (G) and S(G), f C (G). Moreover, if f f
in S (G) then for any S(G),
f f
128
Chapter 3. Homogeneous Lie groups
in C (G).
Furthermore, if f S (G) is compactly supported then f S(G) for any
S(G).
Proof. Let f S (G) and S(G). By denition of the convolution in Denition
1.5.3 and continuity of f (see (3.45)) we have
( f )(x)
x1 )S(G),N
x1 ) C(
f, (
S(G),( +1)2 N
C(1 + x1 )(n +1)N
C(1 + x)
(n +1)N
S(G),(n +1)3 N
(by (3.40))
(by (3.39)).
This shows (3.46). Consequently
( f ) = (X
) f
X
is also bounded for every Nn0 and hence f is smooth. The convergence
statement then follows from the denition of the convolution for distributions.
Let us now assume that the distribution f is compactly supported. Its support
is included in the ball of radius R for R large enough. There exists N N0 such
that
1
1
sup
((xy ))
( f )(x) = f, ( x ) C
y
yR, N
CR
sup
Xy {(xy 1 )} ,
yR, []n N
using (3.16) and (3.17). By (1.11), we have
{(xy 1 )} = (1) (X )(xy 1 ),
X
y
and so for every M N0 with M [], we obtain
Xy {(xy 1 )} = X (xy 1 ) S(G),M (1 + xy 1 )M .
By (3.43), we have also
(1 + xy 1 )1 Co (1 + y)(1 + x)1 .
Therefore, for every M N with M n N we get
( f )(x)
CR sup CoM (1 + y)M (1 + x)M S(G),M
CR
(1
This shows f S(G).
yR
+ x)M S(G),M .
3.1. Graded and homogeneous Lie groups
129
We note that there are certainly dierent ways of introducing the topology
of the Schwartz spaces by dierent choices of families of seminorms.
Lemma 3.1.56. Other families of Schwartz seminorms dening the same Frechet
topology on S(G) are
max[],[]N x X p
max[],[]N X x p
max[]N (1 +  )N X p
(for the rst two we dont need a homogeneous quasinorm) where p [1, ].
Proof. The rst two families with the usual Euclidean derivatives instead of leftinvariant vector elds are known to give the Frechet topologies. Therefore, by e.g.
using Proposition 3.1.28, this is also the case for the rst two families.
The last family would certainly be equivalent to the rst one for the homogeneous quasinorm  p in (3.21), for p being a multiple of 1 , . . . , n , since xpp
is a polynomial. Therefore, the last family also yields the Frechet topology for any
choice of homogeneous quasinorm since any two homogeneous quasinorms are
equivalent by Proposition 3.1.35.
3.1.10 Approximation of the identity
The family of dilations gives an easy way to dene approximations to the identity.
If is a function on G and t > 0, we dene t by
t := tQ Dt1
If is integrable then
i.e.
t (x) = tQ (t1 x).
t is independent of t.
We denote by Co (G) the space of continuous functions on G which vanish at
innity:
Denition 3.1.57. We denote by Co (G) the space of continuous function f : G C
such that for every > 0 there exists a compact set K outside which we have
f  < .
Endowed with the supremum norm = L (G) , Co (G) is a Banach
space.
We also denote by Cc (G) the space of continuous and compactly supported
functions on G. It is easy to see that Cc (G) is dense in Lp (G) for p [1, ) and
in Co (G) (in which case we set p = ).
Lemma 3.1.58. Let L1 (G) and G = c.
130
Chapter 3. Homogeneous Lie groups
(i) For every f Lp (G) with 1 p < or every f Co (G) with p = , we
have
in Lp (G) or Co (G), i.e.
t f cf
t0
t f cf Lp (G) 0.
t0
The same holds for f t .
(ii) If S(G), then for any S(G) and f S (G), we have
t c
t0
in S(G)
and
t f cf
t0
in S (G).
The same holds for t and f t .
The proof is very similar to its Euclidean counterpart.
Proof. Let L1 (G) and c = G . If f Cc (G) then
tQ (t1 y)f (y 1 x)dy cf (x)
(t f )(x) cf (x) =
G
1
(z)f ((tz) x)dz
(z)dzf (x)
=
G
G
=
(z) f ((tz)1 x) f (x) dz.
G
Hence by the Minkowski inequality we have
(z) f ((tz)1 ) f p dz.
t f cf p
G
Since f ((tz)1 ) f p 2f p , this shows (i) for any f Cc (G) by the
Lebesgue dominated convergence theorem. Let f be in Lp (G) or Co (G) (in this
case p = ). By density of Cc (G), for any > 0, we can nd f Cc (G) such
that f f p . We have
t (f f )p t 1 f f p 1 ,
thus
t f cf p
t (f f )p + cf f p + t f cf p
(1 + c) + t f cf p .
Since t f cf p 0 as t 0, there exists > 0 such that
t (0, )
t f cf p < .
Hence if 0 < t < , we have
t f cf p (1 + c + 1).
3.1. Graded and homogeneous Lie groups
131
This shows the convergence of t f cf for any f Lp (G) or Co (G).
With the notation for the operation given by g(x) = g(x1 ), we also have
(f g)= g f.
we obtain the convergence of
Hence applying the previous result to f and ,
f t cf .
Let us prove (ii) for , S(G). We have as above
(t )(x) c(x) =
(z) ((tz)1 x) (x) dz,
G
thus
t cS(G),N
(z) ((tz)1 ) S(G),N dz
(z) C
n
j=1
(tz)1 j S(G),N +n dz
by (3.44). And this shows
t cS(G),N C
n
j=1
S(G),Q+1+j S(G),N +n tj 0.
t0
Hence we have obtained the convergence of t c. As above, applying the
we obtain the convergence of t .
previous result to and ,
Let f S (G). By (1.14) for distributions, we see for any S(G), that
f t , = f, t cf,
t0
by the convergence just shown above. This shows that f t converges to f in
we obtain the converS (G). As above, applying the previous result to f and ,
gence of f t .
In the sequel we will need (only in the proof of Theorem 4.4.9) the following collection of technical results. Recall that a simple function is a measurable
function which takes only a nite number of values.
Lemma 3.1.59. Let B denote the space of simple and compactly supported functions
on G. Then we have the following properties.
(i) The space B is dense in Lp (G) for any p [1, ).
(ii) If S(G) and f B, then f and f are in S(G).
132
Chapter 3. Homogeneous Lie groups
(iii) For every f B and p [1, ],
t f (
t0
)f
in Lp (G). The same holds for f t .
Proof. Part (i) is wellknown (see, e.g., Rudin [Rud87, ch. 1]).
As a convolution of a Schwartz function with a compactly supported tempered distribution f B, f and f are Schwartz by Lemma 3.1.55. This
proves (ii).
Part (iii) follows from Lemma 3.1.58 (i) for 1 p < . For the case p = ,
we proceed as in the rst part of the proof of Lemma 3.1.58 (i) taking f not in
Cc (G) but a simple function with compact support.
Remark 3.1.60. In Section 4.2.2 we will see that the heat semigroup associated
to a positive Rockland operator gives an approximation of the identity ht , t > 0,
which is commutative:
ht hs = hs ht = hs+t .
3.2
Operators on homogeneous Lie groups
In this section we analyse operators on a (xed) homogeneous Lie group G. We
rst study sucient conditions for a linear operator to extend boundedly from
some Lp space to an Lq space. We will be particularly interested in the case of
leftinvariant homogeneous linear operators. In the last section, we will focus our
attention on such operators which are furthermore dierential and on the possible
existence of their fundamental solutions. As an application, we will give a version
of Liouvilles Theorem which holds on homogeneous Lie groups. All these results
have wellknown Euclidean counterparts.
All the operators we consider here will be linear so we will not emphasise
their linearity in every statement.
3.2.1 Leftinvariant operators on homogeneous Lie groups
The Schwartz kernel theorem (see Theorem 1.4.1) says that, under very mild
hypothesis, an operator on a smooth manifold has an integral representation. An
easy consequence is that a leftinvariant operator on a Lie group has a convolution
kernel.
Corollary 3.2.1 (Kernel theorem on Lie groups). We have the following statements.
Let G be a connected Lie group and let T : D(G) D (G) be a continuous
linear operator which is invariant under lefttranslations, i.e.
xo G, f D(G)
T (f (xo )) = (T f )(xo ).
3.2. Operators on homogeneous Lie groups
133
Then there exists a unique distribution D (G) such that
(y 1 x)f1 (y)dy.
T f1 : x f1 (x) =
G
In other words, T is a convolution operator with (right convolution) kernel .
The converse is also true.
Let G be a connected simply connected nilpotent Lie group identied with Rn
endowed with a polynomial law (see Proposition 1.6.6). Let T : S(G) S (G)
be a continuous linear operator which is invariant under left translations, i.e.
xo G, f S(G)
T (f (xo )) = (T f )(xo ).
Then there exists a unique distribution S (Rn ) such that
(y 1 x)f1 (y)dy.
T f1 : x f1 (x) =
G
In other words, T is a convolution operator with (right convolution) kernel .
The converse is also true.
In both cases, for any test function f1 , the function T f1 is smooth. Furthermore, the map T is an isomorphism of topological vector spaces.
A similar statement holds for rightinvariant operators.
We omit the proof: it relies on approaching the kernels (x, y) by continuous
functions for which the invariance forces them to be of the form (y 1 x). The
converses are much easier and have been shown in Section 1.5.
In this monograph, we will often use the following notation:
Denition 3.2.2. Let T be an operator on a connected Lie group G which is continuous as an operator D(G) D (G) or as S(G) S (G). Its right convolution
kernel , as given in Corollary 3.2.1, is denoted by
T 0 = .
In the case of leftinvariant dierential operators, we obtain easily the following properties.
Proposition 3.2.3. If T is a leftinvariant dierential operator on a connected Lie
group G, then its kernel is by denition the distribution T 0 D (G) such that
D(G)
T = T 0 .
The distribution T 0 S (G) is supported at the origin. The equality
f T 0 = T f
134
Chapter 3. Homogeneous Lie groups
holds for any f E (G), the lefthand side being the group convolution of a distribution with a compactly supported distribution. The equality
T 0 f = Tf
for the rightinvariant dierential operator corresponding to T also holds for any
f E (G).
The kernel of T t 0 is given formally by
T t 0 (x) = T 0 (x1 ).
If T = X , for a leftinvariant vector eld X on G and N, then the
distribution (1) X 0 (x1 ) is the left convolution kernel of the rightinvariant
dierential operator T.
We can also see from (1.14) and Denition 1.5.4 that the adjoint of the
bounded on L2 (G) operator T f = f is the convolution operator T f = f
,
well dened on D(G), with the right convolution kernel given by
(x) =
(x1 ).
(3.47)
The transpose operation is dened in Denition A.1.5, and for leftinvariant
dierential operators it takes the form given by (1.10). Clearly the transpose of a
leftinvariant dierential operator on G is a leftinvariant dierential operator on
G.
Proof. A leftinvariant dierential operator is necessarily continuous as D(G)
D(G). Hence it admits the kernel T 0 . We have for D(G) with (x)
= (x1 )
that
= ( T 0 )(0) = T (0).
T 0 ,
So if 0
/ supp then T 0 , = 0. This shows that T 0 is supported at 0.
If , D(G), then
T 0 , = T , = , T t = , T t 0 .
By (1.14) this shows that T t 0 = (T 0 ). Furthermore, if f D (G), then
T f, = f, T t = f, T t 0 = f, (T 0 ) = f T 0 , .
This shows T f = f T 0 .
Now we can check easily (see (1.11)) that
= (X f)
Xf
and, more generally,
f = (1) (X f)
X
3.2. Operators on homogeneous Lie groups
135
for N. Since the equality (f g)= g f holds as long as it makes sense, this
shows that
(1) (X 0 ) f = Tf.
In fact, our primary concern will be to study operators of a dierent nature,
and their possible extensions to some Lp spaces. This (i.e. the Lp boundedness) is
certainly not the case for general dierential operators.
Assuming that an operator is continuous as S(G) S (G) or as D(G)
D (G) is in practice a very mild hypothesis. It ensures that a potential extension
into a bounded operator Lp (G) Lq (G) is necessarily unique, by density of D(G)
in Lp (G). Hence we may abuse the notation, and keep the same notation for an
operator which is continuous as S(G) S (G) or as D(G) D (G) and its
possible extension, once we have proved that it gives a bounded operator from
Lp (G) to Lq (G).
We want to study in the context of homogeneous Lie groups the condition
which implies that an operator as above extends to a bounded operator from
Lp (G) to Lq (G).
As the next proposition shows, only the case p q is interesting.
Proposition 3.2.4. Let G be a homogeneous Lie group and let T be a linear leftinvariant operator bounded from Lp (G) to Lq (G), for some (given) nite p, q
[1, ). If p > q then T = 0.
The proof is based on the following lemma:
Lemma 3.2.5. Let f Lp (G) with 1 p < . Then
1
lim f f (x )Lp (G) = 2 p f Lp (G) .
Proof of Lemma 3.2.5. First let us assume that the function f is continuous with
compact support E. For xo G, the function f (xo ) is continuous and supported
1
= {yz : y E, z E 1 }, then f and
in x1
o E. Therefore, if xo is not in EE
f (xo ) have disjoint supports, and
f f (xo )pp =
f p +
f (xo )p = 2f pp .
E
x1
o E
Now we assume that f Lp (G). For each suciently small > 0, let f be a
continuous function with compact support E {x 1 } satisfying f f p <
. We claim that for any suciently small > 0, we have
1
1
(3.48)
xo  > 21 = f f (xo )p 2 p f p (2 + 2 p ).
136
Chapter 3. Homogeneous Lie groups
Indeed, using the triangle inequality, we obtain
1
1
1
f f (xo )p 2 p f p f f (xo )p 2 p f p + 2 p f p f p .
For the last term of the righthand side we have
f p f p f f p < ,
whereas for the rst term, if xo E E1 , using the rst part of the proof and
then the triangle inequality, we get
f f (xo )p 2 p1 f p = f f (xo )p f f (xo )p
(f f (xo )) (f f (xo ))p
f f p + f (xo ) f (xo )p < 2.
This shows (3.48) and concludes the proof of Lemma 3.2.5.
Proof of Proposition 3.2.4. Let f D(G). As T is leftinvariant, we have
(T f )(xo ) T f q = T f (xo ) f q T L (Lp (G),Lq (G)) f (xo ) f p .
Taking the limits as xo tends to innity, by Lemma 3.2.5, we get
1
2 q T f q T L (Lp (G),Lq (G)) 2 p f p .
But then
T L (Lp (G),Lq (G)) 2 p q T L (Lp (G),Lq (G)) .
1
Hence p > q implies T L (Lp (G),Lq (G)) = 0 and T = 0.
As in the Euclidean case, Proposition 3.2.4 is all that can be proved in the
general framework of leftinvariant bounded operators from Lp (G) to Lq (G). However, if we add the property of homogeneity more can be said and we now focus
our attention on this case.
3.2.2 Leftinvariant homogeneous operators
The next statement says that if the operator T is leftinvariant, homogeneous
and bounded from Lp (G) to Lq (G), then the indices p and q must be related in
the same way as in the Euclidean case but with the topological dimension being
replaced by the homogeneous dimension Q.
Proposition 3.2.6. Let T be a leftinvariant linear operator on G which is bounded
from Lp (G) to Lq (G) for some (given) nite p, q [1, ). If T is homogeneous of
degree C (and T = 0), then
Re
1 1
=
.
q
p
Q
3.2. Operators on homogeneous Lie groups
137
Proof. We compute easily,
Q
f Dt p = t p f p ,
f Lp (G), t > 0.
Thus, since T is homogeneous of degree , we have
Q
tRe q T f q = t T f Dt q = T f Dt q T L (Lp (G),Lq (G)) f Dt p
Q
= T L (Lp (G),Lq (G)) t p f p ,
so
t > 0
T L (Lp (G),Lq (G)) tRe + q p T L (Lp (G),Lq (G)) .
Hence we must have
Re +
Q Q
=0
q
p
as claimed.
Combining together Propositions 3.2.4 and 3.2.6, we see that it makes sense
to restrict ones attention to
Re
(1, 0].
Q
The case Re = 0 is the most delicate and we leave it aside for the moment (see
Section 3.2.5). We shall discuss instead the case
Q < Re < 0.
Let us observe that the homogeneity of the operator is equivalent to the
homogeneity of its kernel:
Lemma 3.2.7. Let T be a continuous leftinvariant linear operator as S(G)
S (G) or as D(G) D (G), where G is a homogeneous Lie group. Then T is homogeneous if and only if its (right) convolution kernel is (Q+)homogeneous.
Proof. On one hand we have
T (f (r ))(x) =
f (ry)(y 1 x)dy,
and on the other hand,
T f (rx)
f (z)(z rx)dz =
f (ry)((ry)1 rx)rQ dy
G
rQ
f (ry)( Dr )(y 1 x)dy.
G
Now the statement follows from these and the uniqueness of the kernel.
138
Chapter 3. Homogeneous Lie groups
The following proposition gives a sucient condition on the homogeneous
kernel so that the corresponding leftinvariant homogeneous operator extends to
a bounded operator from Lp (G) to Lq (G).
Proposition 3.2.8. Let T be a linear continuous operator as S(G) S (G) or as
D(G) D (G) on a homogeneous Lie group G. We assume that the operator T
is leftinvariant and homogeneous of degree , that
Re (Q, 0),
and that the (right convolution) kernel of T is continuous away from the origin.
Then T extends to a bounded operator from Lp (G) to Lq (G) whenever p, q
(1, ) satisfy
Re
1 1
=
.
q
p
Q
The integral kernel then can also be identied with a locally integrable
function at the origin.
We observe that, by Corollary 3.2.1, is a distribution (in S (G) or D (G))
on G. The hypothesis on says that its restriction to G\{0} coincides with a
continuous function o on G\{0}.
Proof of Proposition 3.2.8. We x a homogeneous norm   on G. We denote by
R := {x : x R} and S := {x : x = 1} the ball of radius R and the unit
B
sphere around 0. By Lemma 3.2.7, o is a continuous homogeneous function of
degree (Q + ) on G\{0}. Denoting by C its maximum on the unit sphere, we
have
C
.
x G\{0}
o (x)
Q+Re
x
Hence o denes a locally integrable function on G, even around 0, and we keep
the same notation for this function. Therefore, the distribution = o on
G is, in fact, supported at the origin. It is also homogeneous of degree Q .
Due to the compact support of ,  , f  is controlled by some C k norm of f on
a xed small neighbourhood of the origin. But, because of its homogeneity, and
using (3.9), we get
t > 0
, f = tQ D 1t , f = t , f Dt .
Letting t tend to 0, the C k norms of f Dt remain bounded, so that , f = 0
since Re < 0. This shows that = 0 and so = o .
Note that the weak Lr (G)norm of is nite for r = Q/(Q + Re ). Indeed,
if s > 0,
C
o (x) > s = xQ+Re ,
s
3.2. Operators on homogeneous Lie groups
so that
{x : o (x) > s} B
139
1
(C/s) Q+Re
Q
Q+Re
c C
,
s
with c = B1 , and hence
Q
o wLr (G) c C Q+Re
with r =
Q
.
Q + Re
The proposition is now easy using the generalisation of Youngs inequalities (see
Proposition 1.5.2), so that we get that T is bounded from Lp (G) to Lq (G) for
1
Re
1 1
= 1=
,
q
p
r
Q
as claimed.
We may use the usual vocabulary for homogeneous kernels as in [Fol75] and
[FS82]:
Denition 3.2.9. Let G be a homogeneous Lie group and let C.
A distribution D (G) which is smooth away from the origin and homogeneous of degree Q is called a kernel of type on G.
A (right) convolution operator T : D(G) D (G) whose convolution kernel
is of type is called an operator of type . That is, T is given via
T () = ,
where kernel of type .
Remark 3.2.10. We will mainly be interested in the Lp Lq boundedness of
operators of type . Thus, by Propositions 3.2.4 and 3.2.6, we will restrict ourselves
to C with Re [0, Q).
If Re (0, Q), then a ( Q)homogeneous function in C (G\{0}) is
integrable on a neighbourhood of 0 and hence extends to a distribution in D (G),
see the proof of Proposition 3.2.8. Hence, in the case Re (0, Q), the restriction
to G\{0} yields a onetoone correspondence between the ( Q)homogeneous
functions in C (G\{0}) and the kernels of type .
We will see in Remark 3.2.29 that the case Re = 0 is more subtle.
In view of Lemma 3.2.7 and Proposition 3.2.8, we have the following statement for operators of type with Re (0, Q).
Corollary 3.2.11. Let G be a homogeneous Lie group and let C with
Re (0, Q).
Any operator of type is ()homogeneous and extends to a bounded operator
from Lp (G) to Lq (G) whenever p, q (1, ) satisfy
Re
1 1
=
.
p q
Q
140
Chapter 3. Homogeneous Lie groups
As we said earlier the case of a leftinvariant operator which is homogeneous
of degree 0 is more complicated and is postponed until the end of Section 3.2.4. In
the meantime, we make a useful parenthesis about the Calder
onZygmund theory
in our context.
3.2.3 Singular integral operators on homogeneous Lie groups
In the case of R, a famous example of a leftinvariant 0homogeneous operator is the
Hilbert transform. This particular example has motivated the development of the
theory of singular integrals in the Euclidean case as well as in other more general
settings. In Section A.4, the interested reader will nd a brief presentation of this
theory in the setting of spaces of homogeneous type (due to Coifman and Weiss). In
this section here, we check that homogeneous Lie groups are spaces of homogeneous
type and we obtain the corresponding theorem of singular integrals together with
some useful consequences for leftinvariant operators. We also propose a denition
of CalderonZygmund kernels on homogeneous Lie groups, thereby extending the
one on Euclidean spaces (cf. Section A.4).
First let us check that homogeneous Lie groups equipped with a quasinorm
are spaces of homogeneous type in the sense of Denition A.4.2 and that the Haar
measure is doubling (see Section A.4):
Lemma 3.2.12. Let G be a homogeneous Lie groups and let   be a quasinorm.
Then the set G endowed with the usual Euclidean topology together with the quasidistance
d : (x, y) y 1 x
is a space of homogeneous type and the Haar measure has the doubling property
given in (A.5).
Proof of Lemma 3.2.12. We keep the notation of the statement. The dening
properties of a quasinorm and the fact that it satises the triangular inequality up
to a constant (see Proposition 3.1.38) imply easily that d is indeed a quasidistance
on G in the sense of Denition A.4.1. By Proposition 3.1.37, the corresponding
quasiballs B(x, r) := {y G : d(x, y) < r}, x G, r > 0, generate the usual
topology of the underlying Euclidean space. Hence the rst property listed in
Denition A.4.2 is satised.
By Remark 3.1.34, the quasiballs satisfy B(x, r) = xB(0, r) and B(0, r) =
Dr (B(0, 1)). By (3.6), the volume of B(0, r) is B(0, r) = rQ B(0, 1). Hence we
have obtained that the volume of any open quasiball is B(x, r) = rQ B(0, 1).
This implies that the Haar measure satises the doubling condition given in (A.5).
We can now conclude the proof of the statement with Lemma A.4.3.
Lemma 3.2.12 implies that we can apply the theorem of singular integrals on
spaces of homogeneous type recalled in Theorem A.4.4 and we obtain:
3.2. Operators on homogeneous Lie groups
141
Theorem 3.2.13 (Singular integrals). Let G be a homogeneous Lie group and let T
be a bounded linear operator on L2 (G), i.e.
Co
f L2
T f 2 Co f 2 .
(3.49)
We assume that the integral kernel of T coincides with a locally integrable
function away from the diagonal, that is, on (G G)\{(x, y) G G : x = y}.
We also assume that there exist C1 , C2 > 0 satisfying
y, yo G
(x, y) (x, yo )dx C2 ,
(3.50)
yo1 x>C1 yo1 y
for a quasinorm  .
Then for all p, 1 < p 2, T extends to a bounded operator on Lp because
Ap > 0
f L2 Lp
T f p Ap f p ;
for p = 1, the operator T extends to a weaktype (1,1) operator since
A1 > 0
f L2 L1
{x : T f (x) > } A1
f 1
;
the constants Ap , 1 p 2, depend only on Co , C1 and C2 .
Remark 3.2.14.
The L2 boundedness, that is, Condition (3.49), implies that
the operator satises the Schwartz kernel theorem (see Theorem 1.4.1) and
thus yields the existence of a distributional integral kernel. We still need to
assume that this distribution is locally integrable away from the diagonal.
Since any two quasinorms on G are equivalent (see Proposition 3.1.35), if
the kernel condition in (3.50) holds for one quasinorm, it then holds for any
quasinorm (maybe with dierent constants C1 , C2 ).
As recalled in Section A.4, the notion of CalderonZygmund kernels in the
Euclidean setting appear naturally as sucient conditions (often satised in practice) for (A.7) to be satised by the kernel of the operator and the kernel of its
formal adjoint. This leads us to dene the Calder
onZygmund kernels in our setting
as follows:
Denition 3.2.15. A Calder
onZygmund kernel on a homogeneous Lie group G is
a measurable function o dened on (G G)\{(x, y) G G : x = y} satisfying
for some , 0 < 1, C1 > 0, A > 0, and a homogeneous quasinorm   the
inequalities
o (x, y)
o (x, y) o (x , y)
o (x, y) o (x, y )
Ay 1 xQ ,
x1 x 
A 1 Q+
y x
y 1 y 
A 1 Q+
y x
if C1 x1 x  y 1 x,
if C1 y 1 y  y 1 x.
142
Chapter 3. Homogeneous Lie groups
A linear continuous operator T as D(G) D (G) or as S(G) S (G) is called
a Calder
onZygmund operator if its integral kernel coincides with a Calder
onZygmund kernel on (G G)\{(x, y) G G : x = y}.
Remark 3.2.16. 1. In other words, we have modied the denition of a classical
Calder
onZygmund kernel (as in Section A.4)
by replacing the Euclidean norm by a homogeneous quasinorm
and, more importantly, the topological (Euclidean) dimension of the
underlying space n by the homogeneous dimension Q.
2. By equivalence of homogeneous quasinorms, see Proposition 3.1.35, the definition does not depend on a particular choice of a homogeneous quasinorm
as we can change the constants C1 , A.
As in the Euclidean case, we have
Proposition 3.2.17. Let G be a homogeneous Lie group and let T be a bounded
linear operator on L2 (G).
If T is a Calder
onZygmund operator on G (in the sense of Denition 3.2.15),
then T is bounded on Lp (G), p (1, ), and weaktype (1,1).
Proof of Proposition 3.2.17. Let T be a bounded operator on L2 (G) and :
(x, y) (x, y) its distributional kernel. Then its formal adjoint T is also bounded
on L2 (G) with the same operator norm. Furthermore its distributional kernel is
(y, x). We assume that coincides with a CalderonZygmund
() : (x, y)
kernel o away from the diagonal. We x a quasinorm  . The rst inequality in
()
Denition 3.2.15 shows that o and o coincide with locally integrable functions
away from the diagonal. Using the last inequality, we have for any y, yo G,
y 1 yo 
o (x, y) o (x, yo )dx A
1 Q+ dx
yo1 xC1 yo1 y
yo1 xC1 yo1 y yo x
and, using the change of variable x = yo1 x, we have
1
dx
=
x (Q+) dx
1 Q+
yo1 xC1 yo1 y yo x
x C1 yo1 y
x (Q+) dx
x C1 yo1 y
+
=c
r=C1 yo1 y
r(Q+) rQ1 dr = c1 yo1 y ,
having also used the polar coordinates (Proposition 3.1.42) with c denoting the
mass of the Borel measure on the unit sphere, and c1 a new constant (of C1 ,
and Q). Hence we have obtained
o (x, y) o (x, yo )dx c1 A.
yo1 xC1 yo1 y
3.2. Operators on homogeneous Lie groups
143
()
Similarly for o , we have
()
o (x, y)
yo1 xC1 yo1 y
()
o (x, yo )dx
o (y, x)
yo1 xC1 yo1 y
o (yo , x)dx
yo1 y
dx,
yo1 xQ+
yo1 xC1 yo1 y
having used the second inequality in Denition 3.2.15. The same computation as
()
above shows that the last lefthand side is bounded by c1 A. Hence o and o
satisfy (3.50). Proposition 3.2.17 now follows from Theorem 3.2.13.
Remark 3.2.18. As in the Euclidean case, Calder
onZygmund kernels do not necessarily satisfy the other condition of the L2 boundedness (see (3.49)) and a condition of cancellation is needed in addition to the CalderonZygmund condition to
ensure the L2 boundedness. Indeed, one can prove adapting the Euclidean case (see
onZygmund
the proof of Proposition 1 in [Ste93, ch.VII 3]) that if o is a Calder
kernel satisfying the inequality
c > 0
o (x, y) cy 1 xQ ,
x = y
then there does not exist an L2 bounded operator T having o as its kernel.
The following statement gives sucient conditions for a kernel to be Calder
onZygmund in terms of derivatives:
Lemma 3.2.19. Let G be a homogeneous Lie group. If o is a continuously dierentiable function on (G G)\{(x, y) G G : x = y} satisfying the inequalities
for any x, y G, x = y, j = 1, . . . , n,
o (x, y)
Ay 1 xQ ,
(Xj )x o (x, y)
Ay 1 x(Q+j ) ,
(Xj )y o (x, y)
Ay 1 x(Q+j ) ,
onfor some constant A > 0 and homogeneous quasinorm  , then o is a Calder
Zygmund kernel in the sense of Denition 3.2.15 with = 1.
Again, if these inequalities are satised for one quasinorm, then they are
satised for all quasinorms, maybe with dierent constants A > 0.
Proof of Lemma 3.2.19. We x a quasinorm  . We assume that it is a norm
without loss of generality because of the remark just above and the existence of a
homogeneous norm (Theorem 3.1.39); although we could give a proof without this
hypothesis, it simplies the constants below. Let o be as in the statement. Using
the Taylor expansion (Theorem 3.1.51) or the Mean Value Theorem (Proposition
3.1.46), we have
o (x , y) o (x, y) Co
n
j=1
x1 x j
sup
zx1 x 
(Xj )x1 =xz o (x1 , y).
144
Chapter 3. Homogeneous Lie groups
Using the second inequality in the statement, we have
sup
zx1 x 
(Xj )x1 =xz o (x1 , y) A
sup
zx1 x 
y 1 xz(Q+j ) .
The reverse triangle inequality yields
y 1 xz y 1 x z
1 1
y x
2
if z
1 1
y x.
2
Hence, if 2x1 x  y 1 x, then we have
sup
zx1 x 
y 1 xz(Q+j ) 2Q+j y 1 x(Q+j ) ,
and we have obtained
o (x, y) o (x , y)
Co
x1 x j 2Q+j y 1 x(Q+j )
j=1
Co
(2)(j 1) 2Q+j x1 x y 1 x(Q1) .
j=1
This shows the second inequality in Denition 3.2.15.
We proceed in a similar way to prove the third inequality in Denition 3.2.15:
the Taylor expansion yields
o (x, y) o (x, y ) Co
y 1 y j
j=1
sup
zy 1 y 
(Xj )y1 =yz o (x, y1 )
while one checks easily
sup
zy 1 y 
(Xj )y1 =yz o (x, y1 )
(yz)1 x(Q+j )
A2Q+j y 1 x(Q+j ) ,
sup
zy 1 y 
when 2y 1 y  y 1 x. We conclude in the same way as above and this shows
onZygmund kernel.
that o is a Calder
Corollary 3.2.20. Let G be a homogeneous Lie group and let be a continuously
dierentiable function on G\{0}. If satises for any x G\{0}, j = 1, . . . , n,
(x)
AxQ ,
Xj (x)
j (x)
X
Ax(Q+j ) ,
Ax(Q+j ) ,
for some constant A > 0 and homogeneous quasinorm  , then
o : (x, y) (y 1 x)
is a Calder
onZygmund kernel in the sense of Denition 3.2.15 with = 1.
3.2. Operators on homogeneous Lie groups
145
Corollary 3.2.20 will be useful when dealing with convolution kernels which
are smooth away from the origin, in particular when they are also (Q)homogeneous, see Theorem 3.2.30.
Proof of Corollary 3.2.20. Keeping the notation of the statement, using properties
(1.11) of left and right invariant vector elds, we have
(Xj )x o (x, y)
(Xj )y o (x, y)
(Xj )(y 1 x),
j )(y 1 x).
(X
The statement now follows easily from Lemma 3.2.19.
Often, the convolution kernel decays quickly enough at innity and the main
singularity to deal with is about the origin. The next statement is an illustration
of this idea:
Corollary 3.2.21. Let G be a homogeneous Lie group and let T be a linear operator
which is bounded on L2 (G) and invariant under left translations.
We assume that its distributional convolution kernel coincides on G\{0} with
a continuously dierentiable function which satises
(x)dx A,
x1/2
sup xQ (x)
A,
A,
0<x1
sup xQ+j Xj (x)
j = 1, . . . , n,
0<x1
for some constant A > 0 and a homogeneous quasinorm  . Then T is bounded
on Lp (G), p (1, ), and is weaktype (1,1).
Proof. Let D(G) be [0, 1]valued
function such that 0 on {x 1}
and 1 on {x 1/2}. As x1/2 (x)dx is nite, (1 ) is integrable
and the convolution operator with convolution kernel (1 ) is bounded on
Lp (G) for p [1, ]. Hence it suces to prove that the kernel o given via
onZygmund.
o (x, y) = ()(y 1 x) is Calder
From the estimates satised by , it is clear that the quantities
sup
xG\{0}
xQ ()(x)
and
sup
xG\{0}
x(Q+j ) Xj ()(x)
j may be expressed as a combination of Xk with homogeneous
are nite. As each X
polynomial coecients, see Section 3.1.5, we have for any (regular enough) function
f with compact support
sup
xG\{0}
j f (x) C
x(Q+j ) X
sup
xG\{0}
k=1,...,n
x(Q+k ) Xk f (x).
146
Chapter 3. Homogeneous Lie groups
j ()(x) are also bounded.
Consequently, the quantities supxG\{0} x(Q+j ) X
onApplying Lemma 3.2.19 to o dened above, one checks easily that it is a Calder
Zygmund kernel. Applying Proposition 3.2.17 concludes the proof of Corollary
3.2.21.
This closes our parenthesis about the CalderonZygmund theory in our context, and we can go back to the study of leftinvariant homogeneous operators,
this time of homogeneous degree 0.
3.2.4 Principal value distribution
As we will see in the sequel, many interesting operators for our analysis on a
homogeneous Lie group G will be given by convolution operators with (right convolution distributional) kernels homogeneous of degree with Re = Q. In most
of the interesting cases, the distribution will be given by a locally integrable
function away from the origin; denoting by o the restriction of to G\{0}, one
may wonder if there is a onetoone correspondence between and o . As in the
Euclidean case, this leads to the notion of the principal value distribution and we
adapt the ideas here to t the homogeneous context; in particular, the topological
(Euclidean) dimension is replaced by the homogeneous dimension Q.
So the question is: Considering a locally integrable function o on G\{0}
which is homogeneous of degree with Re = Q, does there exist a distribution
D (G) on G, homogeneous of the same degree on G, whose restriction to
G\{0} coincides with o ? that is,
, f =
o (x)f (x)dx,
G\{0}
whenever f D(G) and 0 supp f . In other words, can the functional
n
D(R \{0}) f
o (x)f (x)dx
G\{0}
be extended to a continuous functional on D(Rn )?
Remark 3.2.22. 1. We observe that if such an extension exists, it is not unique
in general. For = Q, the reason is that the Dirac 0 at the origin is
homogeneous of degree Q (see Example 3.1.20), so that if is a solution,
then + c0 for any constant c is another solution. (However, see Proposition 3.2.27.)
2. The second observation is that the answer is negative in general:
Example 3.2.23. Let  be some xed homogeneous quasinorm on G smooth away
from the origin. The function dened by o (x) = x with = Q + i , R, is
homogeneous of degree on G\{0} but can not be extended into a homogeneous
distribution D (G) of homogeneous degree .
3.2. Operators on homogeneous Lie groups
147
Proof of Example 3.2.23. Indeed, let us assume that such a distribution exists
for this o . Homogeneity of degree Q + i means that
, Dt = ti , ,
t > 0, D(G).
Let B := {x G : x < } be the ball around 0 of radius . Let D(G) be a
realvalued function supported on D2 (B )\B , such that
((x) (2x)) xQ dx = 0.
G
We now dene
(x) := xi (x)
and
f := 2i ( D2 ), x G\{0}.
Immediately we notice that
f (x) = xi ((x) (2x))
and, therefore, both and f are supported inside D4 (B )\B and are smooth.
We compute
((x) (2x)) xQ dx = 0
o , f =
G
by the choice of . On the other hand,
, f = , 2i , D2 = 0.
We have obtained a contradiction.
The next statement answers the question above under the assumption that
o is also continuous on G\{0}.
Proposition 3.2.24. Let G be a homogeneous Lie group and let o be a continuous
homogeneous function on G\{0} of degree with Re = Q.
Then o extends to a homogeneous distribution in D (G) if and only if its
average value, dened in Lemmata 3.1.43 and 3.1.45, is mo = 0.
Proof. Let us x a homogeneous quasinorm  . We denote by the measure on
the unit sphere S = {x : x = 1} which gives the polar change of coordinates
(see Proposition 3.1.42) and  its total mass.
By Lemma 3.1.41, there exists c > 0 such that
x 1 = xE cx.
(3.51)
First let us assume mo = 0. Therefore, for any a, b [0, ),
a<x<b
o (x)dx =
b
r=a
S
o (rx)d(x)r
Q1
dr = mo
b
r=a
r rQ1 dr = 0,
148
Chapter 3. Homogeneous Lie groups
see Section 3.1.7. We claim that, for each f D(G),
o (x)f (x)dx < .
lim
0
(3.52)
x>
Indeed, let us check the Cauchy condition for 0 < < . We see that
o (x)f (x)dx
o (x)f (x)dx =
o (x)f (x)dx
x>
<x<
x>
=
o (x) (f (x) f (0)) dx
<x<
o (x) f (x) f (0) dx.
<x<
The (Euclidean) mean value theorem and the estimate (3.51) imply
f (x) f (0) f xE f cx
if x < 1.
Since o is homogeneous with Re = Q, denoting by Co the maximum of o 
on the unit sphere {x : x = 1}, we have
x G\{0}
o (x) Co xQ .
Hence if < 1,
o (x)f (x)dx
o (x)f (x)dx
f cCo x1Q dx
x>
x>
<x<
= f cCo ( ).
This implies the Cauchy condition. Therefore, Claim (3.52) is proved and we denote the limit by
o (x)f (x)dx, f D(G).
(3.53)
, f := lim
0
x>
This clearly denes a linear functional. Moreover, this functional is continuous
R = {x : x R} for R large enough,
since if f D(G) is supported in a ball B
then, for < 1,
o (x)f (x)dx
o (x)f (x)dx +
o (x)f (x)dx
<x<1
x>
1<x
f cCo (1 ) + Co
f (x)dx
1<xR
CR (f + f ).
3.2. Operators on homogeneous Lie groups
149
For the converse, we proceed by contradiction: let us assume that exists
0. Then
and that mo =
mo
o
x

is a continuous homogeneous distribution of G\{0} of degree with mean average
m o
mo
x d(x) =
o (x)d(x)
d(x)
o (x)

 S
S
S
= mo mo = 0.
Hence it admits an extension into a homogeneous distribution by the rst part
of the proof. But this would imply that x has such an extension and this is
impossible by Example 3.2.23.
Remark 3.2.25. (i) In view of the proof above, the hypothesis of continuity in
Proposition 3.2.24 (and also in Proposition 3.2.27) can be relaxed into the
following condition: o is locally integrable and locally bounded on G\{0}.
This ensures that all the computations make sense and, since the unit
sphere of a given homogeneous quasinorm is compact, o  is bounded there.
We will not use this fact.
(ii) By Lemma 3.1.45 the condition mo = 0 is independent of the homogeneous
quasinorm. However, the distribution dened in (3.53) depends on the choice
of a particular homogeneous quasinorm. For instance, one can show that the
function on R2 given in polar coordinates by
o (rei ) =
cos 4
,
r2
admits two dierent extensions via the procedure (3.53) when considering
x + y.
the Euclidean norm (x, y) (x2 + y 2 )1/2 and the 1 norm (x, y)
Denition 3.2.26. The distribution given in (3.53) is called a principal value distribution denoted by
p.v. o (x).
The notation is ambiguous unless a homogeneous norm is specied.
The next proposition states that, modulo a Dirac distribution at the origin,
the only possible extension is the principal value distribution:
Proposition 3.2.27. Let be a homogeneous distribution of degree with Re =
Q on a homogeneous Lie group G. We assume that the restriction of to G\{0}
coincides with a continuous function o .
Then o is homogeneous of degree on G\{0} and mo = 0. Moreover, after
the choice of a homogeneous norm,
(x) = p.v. o (x) + co ,
for some constant c C, with c = 0 if = Q.
150
Chapter 3. Homogeneous Lie groups
Proof. By Proposition 3.2.24, mo = 0. Then
:= p.v. o
is also homogeneous of degree and supported at the origin.
Let f D(G) with f (0) = 0. Due to the compact support of ,  , f  is
controlled by some C k norm of f on a xed small neighbourhood of the origin.
But, because of its homogeneity of degree with Re = Q,
t > 0
 , f  =  , f Dt .
Letting t tend to 0, the note that the C k norms of f Dt remain bounded. Let us
show that as t 0, we actually have , f Dt 0. We claim that f Dt 0
in C k (U ) for a neighbourhood U of 0. Indeed,
X (f Dt ) = t[] (X f ) Dt 0
as t 0,
provided that = 0. On the other hand, also (f Dt )(x) = f (tx) f (0) = 0
as t 0, and same for the L norm over the set U . Thus, we have proved that
, f = 0 for any f D(G) vanishing at 0.
We now x a function D(G) with (0) = 1. For any f D(G),
, f = , f f (0) + f (0) , = f (0) , ,
since f f (0) D(G) vanishes at 0. This shows = c0 where c = , . But
0 is homogeneous of degree Q, see Example 3.1.20, whereas is homogeneous
of degree . So c = 0 if = Q.
have
Alternatively, we can also argue as follows. By Proposition 1.4.2 we must
a 0
= p.v. o =
j
for some j and some constants a . Now, we know by Example 3.1.20 that 0 is
homogeneous of degree Q, and by Proposition 3.1.23 that 0 is homogeneous
of degree Q[]. Since is homogeneous of degree Q, it follows that all a = 0
for Q [] = . The statement now follows since, if = Q, we must have all
a = 0, and if = Q, we take c = a0 .
Using the vocabulary of kernels of type , see Denition 3.2.9, Proposition
3.2.24 implies easily:
Corollary 3.2.28. Let G be a homogeneous Lie group and let o be a smooth homogeneous function on G\{0} of degree with Re = Q. Then o extends to
a homogeneous distribution in D (G) if and only if its average value, dened in
Lemmata 3.1.43 and 3.1.45, is mo = 0. In this case, the extension is a kernel of
type .
3.2. Operators on homogeneous Lie groups
151
Remark 3.2.29. Remark 3.2.10 explained the correspondence between the kernels
of type and their restriction to G\{0} in the case Re (0, Q).
With Corollary 3.2.28, we obtain the case Re = 0: the restriction to G\{0}
yields a correspondence between
the ( Q)homogeneous functions in C (G\{0}) with vanishing mean value
and the kernels of type .
It is onetoone if = 0 but if = 0, we have to consider the kernels of type
modulo C0 .
3.2.5 Operators of type = 0
We can now go back to our original motivation, that is, a condition on a leftinvariant homogeneous operator of degree 0 to obtain continuity on every Lp (G).
Our condition here is that the operator is of type 0, or more generally of type ,
Re = 0.
Theorem 3.2.30. Let G be a homogeneous Lie group and let C with
Re = 0.
Any operator of type on G is ()homogeneous and extends to a bounded operator on Lp (G), p (1, ).
The proof consists in showing that the operator is Calder
onZygmund (in
the sense of Denition 3.2.15) and bounded on L2 (G). Note that the cancellation
condition (see Remark 3.2.18), is provided by mo = 0, see Proposition 3.2.27.
Proof. Let D (G) be a kernel of type , Re = 0. We denote by o its
smooth restriction to G\{0}. One checks easily that o satises the hypotheses of
onZygmund kernel in the sense of
Corollary 3.2.20. Consequently, o is a Calder
Denition 3.2.15. By the Singular Integral Theorem, more precisely its form given
in Proposition 3.2.17, to prove the Lp boundedness for every p (1, ), it suces
to prove the case p = 2.
Fixing a homogeneous norm   smooth away from the origin, by Proposition 3.2.27, we may assume that is the principal value distribution of o (see
Denition 3.2.26). We want to show that
f f p.v. o
is bounded on L2 (G). For this, we will apply the CotlarStein lemma (see Theorem A.5.2) to the operators
Tj : f f Kj ,
j Z,
where
Kj (x) = o (x)12j x2j+1 (x).
152
Chapter 3. Homogeneous Lie groups
We claim that
max Tj Tk L (L2 (G)) , Tj Tk L (L2 (G)) C2jk .
(3.54)
Assuming this claim, by the CotlarStein lemma, j Tj denes a bounded oper
ator on L2 (G) and its (right convolution) kernel is j Kj which coincides, as a
distribution, with p.v. o = . This would conclude the proof.
Let us start to prove Claim (3.54). It is not dicult to see (see (3.47)) that
the adjoint of the operator Tj on L2 (G) is the convolution operator with right
convolution kernel given by
j (x1 ),
Kj (x) = K
which is compactly supported. Therefore, the operators Tj Tk and Tj Tk are convolution operators with kernels Kk Kj and Kk Kj , respectively. We observe
that, by homogeneity of o , for any j N0 ,
Kj (x) = 2jQ K0 (2j x)
and so
Kj L1 (G) = K0 L1 (G) .
By the Young convolution inequality (see Proposition 1.5.2), the operators Tj ,
Tj Tk and Tj Tk are bounded on L2 (G) with operator norms
Tj L (L2 (G))
Tj Tk L (L2 (G))
Tj Tk L (L2 (G))
Kj 1 = K0 1 ,
Kk Kj 1 Kk 1 Kj 1 = K0 21 ,
Kk Kj 1 Kk 1 Kj 1 = K0 21 .
In order to prove Claim (3.54) we need to obtain a better decay for Kk Kj 1
and Kk Kj 1 when j and k are far apart. Since Kk Kj 1 = Kj Kk 1 and
Kk Kj 1 = Kj Kk 1 , we may assume k > j. Quantitatively we assume that
C1 2jk+1 < 1/2 where C1 1 is the constant appearing in (3.26) for b = 1/2.
We observe that the cancellation condition mo = 0 implies
Kk (x)dx =
o (x)dx = mo ln 2 = 0,
G
2k x2k+1
and so
Kk Kj (x)
Kk (y)Kj (y 1 x)dy = Kk (y) Kj (y 1 x) Kj (x) dy
G
G
Kk (y) Kj (y 1 x) Kj (x) dy
G
Co yQ Kj (y 1 x) Kj (x) dy,
2k y2k+1
3.2. Operators on homogeneous Lie groups
153
where Co is the maximum of o  on the unit sphere {x = 1}. Thus after the
change of variable z = 2k y,
Kk Kj (x)
Co zQ Kj ((2k z)1 x) Kj (x) dz.
1z2
We want to estimate the L1 norm with respect to x of the last expression. Hence
we now look at
k 1
Kj ((2 z) x) Kj (x) dx =
Kj x1 2k z Kj (x1 ) dx1 ,
G
j
after the change of variable x = x1
1 . Using Kj = 2 K0 Dj and the change of
j
variable x2 = 2 x1 , we obtain
Kj x1 2k z Kj (x1 ) dx1 =
K0 x2 2k+j z K0 (x2 ) dx2 .
G
Let A0 = {1 x 2} be the annulus with radii 1 and 2 around 0 and write
momentarily y 1 = 2k+j z with z A0 . We can write the last integral as
K0 (xy 1 ) K0 (x) dx =
+
+
.
A0 (A0 y)
A0 \(A0 y)
(A0 y)\A0
For the last two integrals, we see with a change of variable x = x y 1 that
K0 (x y 1 ) dx =
=
K0 (x) dx =
,
A0 \(A0 y)
A0 \(A0 y)
and
K0 
A0 \(A0 y)
Thus
G
(A0 y)\A0
xy 1 >2
1x2
Co xQ dx +
K0 (xy 1 ) K0 (x) dx =
+2Co
xy 1 >2
1x2
(A0 y)\A0
xy 1 <1
1x2
K0 (xy
A0 (A0 y)
xQ dx +
Co xQ dx.
) K0 (x)dx
xy 1 <1
1x2
xQ dx .
Since y 1 is relatively small, by (3.26) we get for the two integrals above
+
+
1
1
xy >2
1x2
xy <1
1x2
= ln
2C1 y<x2
1x<1+C1 y
2
+ ln(1 + C1 y) Cy,
2 C1 y
(3.55)
154
Chapter 3. Homogeneous Lie groups
(see Example 3.1.44), whereas by Proposition 3.1.40 we have for any x A0 ,
K0 (xy 1 ) K0 (x) Cy xQ1 ,
and so
A0 (A0 y)
K0 (xy 1 ) K0 (x) dx Cy
xQ1 dx Cy.
1x2
We have obtained that the expression (3.55) is up to a constant less than 2k+j
when C1 2jk+1 < 1/2 (and y 1 = 2k+j z, z A0 ). This estimate gives
Q
K0 (x 2k+j z) K0 (x) dx dz
Kk Kj 1 Co
z
zA0
G
Co
zQ+1 C2k+j dz C2k+j .
zA0
With a very minor modication, we can show in the same way that Kk Kj 1
C2k+j .
This shows Claim (3.54) and concludes the proof of Theorem 3.2.30.
Remark 3.2.31. In view of the proof, we can relax the smoothness condition in
the hypotheses of Theorem 3.2.30: it suces to assume that o C 1 (G\{0}).
This ensures that we can apply Propositions 3.2.27 and 3.1.40 during the
proof.
3.2.6 Properties of kernels of type , Re [0, Q)
The kernels and operators of type have been dened in Denition 3.2.9. Summarising results of the previous section, namely Corollary 3.2.11 for Re (0, Q),
and Theorem 3.2.30 for Re = 0, we can unite them as
Corollary 3.2.32. Let G be a homogeneous Lie group and let C with
Re [0, Q).
Any operator of type on G is ()homogeneous and extends to a bounded operator from Lp (G) to Lq (G) provided that
1 1
Re
=
,
p q
Q
1 < p q < .
When considering kernels of type , we have regularly used the following
property: if is a kernel of type then, xing a homogeneous quasinorm   on
G, admits a maximum C on the unit sphere {x = 1}, and by homogeneity we
have
(3.56)
x G\{0}
(x) C xRe Q .
3.2. Operators on homogeneous Lie groups
155
In particular, it is locally integrable if Re > 0 and denes a distribution on the
whole group G in this case. In the case when Re = 0, by Proposition 3.2.27,
also denes a distribution on G of the form
= p.v. 1 + c0 ,
where 1 is of type with vanishing average value and c C is a constant.
We can also deduce the type of a kernel from the following lemma:
Lemma 3.2.33. Let be a kernel of type with Re (0, Q). Let T be a
homogeneous dierential operator of homogeneous degree T . If Re T [0, Q)
then T denes a kernel of type T .
Proof. Clearly T is a (Q + T )homogeneous distribution which coincides
with a smooth function away from 0.
Remark 3.2.34. We have obtained certain properties of convolution operators with
kernels of type in Corollary 3.2.11 for Re (0, Q), and in Theorem 3.2.30 for
Re = 0. When composing two such types of operators, we have to deal with
the convolution of two kernels and this is a problematic question in general. Indeed, the problems about convolving distributions on a noncompact Lie group
are essentially the same as in the case of the abelian convolution on Rn . The convolution 1 2 of two distributions 1 , 2 D (G) is well dened as a distribution
provided that at most one of them has compact support, see Section 1.5. However, additional assumptions must be imposed in order to dene convolutions of
distributions with noncompact supports. Furthermore, the associative law
(1 2 ) 3 = 1 (2 3 ),
(3.57)
holds when at most one of the j s has noncompact support, but not necessarily
when only one of the j s has compact support even if each convolution in (3.57)
could have a meaning.
The following proposition establishes that there is no such pathology appearing when considering convolution with kernel of type with Re [0, Q). This
will be useful in the sequel.
Proposition 3.2.35. Let G be a homogeneous Lie group.
(i) Suppose C with 0 Re < Q, p 1, q > 1, and r 1 given by
1 1 Re
1
= +
1.
r
p q
Q
If is a kernel of type , f Lp (G), and g Lq (G), then f (g ) and
(f g) are well dened as elements of Lr (G), and they are equal.
156
Chapter 3. Homogeneous Lie groups
(ii) Suppose 1 is a kernel of type 1 C with Re 1 > 0 and 2 is a kernel of
type 2 C with Re 2 0. We assume Re (1 + 2 ) < Q. Then 1 2 is
well dened as a kernel of type 1 + 2 . Moreover, if f Lp (G) where
1 < p < Q/(Re (1 + 2 ))
then (f 1 ) 2 and f (1 2 ) belong to Lq (G),
1
1 Re (1 + 2 )
=
,
q
p
Q
and they are equal.
Proof. Let us prove Part (i). By Corollary 3.2.11, Theorem 3.2.30 and Youngs
inequality (see Proposition 1.5.2), the mappings (f, g) f (g ) and (f, g)
(f g) are continuous from Lp (G) Lq (G) to Lr (G). They coincide when they
have compact support, and hence in general.
Let us prove Part (ii). We x a homogeneous quasinorm   smooth away
from the origin. We will use the general properties of kernels of type explained
at the beginning of this section, especially estimate (3.56).
Let x = 0 be given. We can nd > 0 such that the balls
B(0, ) := {y : y < }
and
B(x, ) := {y : xy 1  < },
do not intersect. We note that these balls are dierent from those in Denition
3.1.33 (that are used throughout this book) but in this proof only, it will be more
convenient for us to work with the balls dened as above.
If Re 1 , Re 2 > 0, then both 1 and 2 are locally integrable and
for y B(0, ),
yRe 2 Q
1
1 (xy )2 (y) Cx,
for y B(x, ),
xy 1 Re 1 Q
/ B(0, ) B(x, ).
O(yRe (1 +2 )2Q ) y
Thus we can integrate 1 (xy 1 )2 (y) against dy on B(0, ), B(x, ) and outside of
B(0, ) B(x, ) to obtain the sum of three integrals absolutely convergent:
+
+ y> 1 (xy 1 )2 (y)dy := (x).
yB(0,)
yB(x,)
xy 1 >
This denes (x) which is independent of small enough.
If Re 2 = 0, by Proposition 3.2.27, we may assume that 2 is the principal
value of a homogeneous distribution with mean average 0 (see also Denition 3.2.26
and (3.53)). In this case, by smoothness of 1 away from 0 and Proposition 3.1.40,
1 (xy 1 ) 1 (x) 2 (y) Cx, y1Q for y B(0, ),
3.2. Operators on homogeneous Lie groups
157
and we obtain again the sum of three integrals absolutely convergent:
1 (xy 1 ) 1 (x) 2 (y)dy +
yB(0,)
+
yB(x,)
y>
xy 1 >
1 (xy 1 )2 (y)dy =: (x).
This denes (x) which is independent of small enough.
In both cases, we have dened a function on G\{0}. A simple change of
variables shows that is homogeneous of degree 1 + 2 Q (this is left to the
reader interested in checking this fact).
Let us x 1 D(G) with 1 1 on B(0, /2) and 1 0 on the complement
of B(0, ). We x again x = 0 and we set 2 (y) = 1 (xy 1 ). Then 1 and 2 have
disjoint supports and for Re 2 > 0 it is easy to check that for z B(x, /2) we
have (z) = I1 + I2 + I3 , where
1 (y)1 (zy 1 )2 (y)dy,
I1 =
G
I2 =
2 (y)1 (zy 1 )2 (y)dy =
2 (y 1 z)1 (y)2 (y 1 z)dy,
G
G
1
(1 1 (y) 2 (y))1 (zy )2 (y)dy,
I3 =
G
with a similar formula for Re 2 = 0. The integrands of I1 , I2 , and I3 depend
smoothly on z. Furthermore, one checks easily that their derivatives in z remains
integrable. This shows that is smooth near each point x = 0. Since Re (1 +2 ) >
0, is locally integrable on the whole group G. Hence the distribution D (G)
is a kernel of type 1 + 2 .
We can check easily for D(G),
2 = 2 ,
1 .
, = 1 ,
So having (1.14) and (1.15) we dene 1 2 := .
Let f Lp (G) where p > 1 and
1 Re (1 + 2 )
1
=
> 0.
q
p
Q
We observe that (f 1 ) 2 and f are in Lq (G) by Corollary 3.2.11, Theorem 3.2.30, and Youngs inequality (see Proposition 1.5.2). To complete the proof,
it suces to show that the distributions (f 1 ) 2 and f (1 2 ) are equal.
For this purpose, we write 1 = 01 +
1 with
01 := 1 1x1
and
1 := 1 1x>1 .
158
Chapter 3. Homogeneous Lie groups
r+
If r = Q/(Q Re 1 ) then 01 Lr (G) and
(G) for any > 0. We
1 L
take so small that r > 1 and
p1 + (r + )1 Re 2 /Q 1 > 0.
By Part (i), (f 01 ) 2 and f (01 2 ) coincide as elements of Ls (G) where
s1 = p1 + (r )1 Re 2 /Q 1.
t
And (f
1 ) 2 and f (1 2 ) coincide as elements of L (G) where
t1 = p1 + (r + )1 Re 2 /Q 1.
Thus (f 1 )2 and f coincide as elements of Ls (G) and Lt (G). This concludes
the proof of Part (ii) and of Proposition 3.2.35.
3.2.7 Fundamental solutions of homogeneous dierential operators
On open sets or manifolds, general results about the existence of fundamental
kernels of operators hold, see e.g. [Tre67, Theorems 52.1 and 52.2]. On a Lie group,
we can study the case when the fundamental kernels are of the form (xy) in the
abelian case and (y 1 x) on a general Lie group, where is a distribution, often
called a fundamental solution. It is sometimes possible and desirable to obtain
the existence of such fundamental solutions for left or right invariant dierential
operators.
In this section, we rst give a denition and two general statements valid
on any connected Lie group, and then analyse in more detail the situation on
homogeneous Lie groups.
Denition 3.2.36. Let L be a leftinvariant dierential operator on a connected
Lie group G. A distribution in D (G) is called a (global) fundamental solution
of L if
L = 0 .
A distribution
on a neighbourhood of 0 is called a local fundamental solution
of L (at 0) if L
= 0 on .
On (Rn , +), global fundamental solutions are often called Green functions.
Example 3.2.37. Fundamental solutions for the Laplacian = j j2 on Rn are
wellknown
cn
if n 3
xn2 + p(x)
G(x) =
if n = 2
c2 ln x + p(x)
x1[0,) (x) + p(x) if n = 1
where cn is a (known) constant of n, p is any polynomial of degree 1, and  
the Euclidean norm on Rn .
3.2. Operators on homogeneous Lie groups
159
Example 3.2.37 shows that fundamental solutions are not unique, unless some
hypotheses, e.g. homogeneity (besides existence), are added.
Although, in practice, computing fundamental solutions is usually dicult,
they are useful and important objects.
Lemma 3.2.38. Let L be a leftinvariant dierential operator with smooth coecients on a connected Lie group G.
1. If L admits a fundamental solution , then for every distribution u D (G)
with compact support, the convolution f = u D (G) satises
Lf = u
on G.
2. An operator L admits a local fundamental solution if and only if it is locally
solvable at every point.
For the denition of locally solvability, see Denition A.1.4.
Proof. For the rst statement,
L u = u L = u 0 = u.
For the second statement, if L is locally solvable, then at least at the origin,
one can solve L
= 0 and this shows that L admits a local fundamental solution.
Conversely, let us assume that L admits a local fundamental solution
on
the open neighbourhood of 0. We can always nd a function D() such that
= 1 on an open neighbourhood 1 of 0; we dene 1 D () by 1 :=
and view 1 also as a distribution with compact support. Then it is easy to check
that L1 = 0 on 1 but that
L1 = 0 + ,
where is a distribution whose support does not intersect 1 .
Let 0 be an open neighbourhood of 0 such that
1
y : x, y 0 } 1 .
1
0 0 = {x
We can always nd a function 1 D(0 ) which is equal to 1 on a neighbourhood
0 0 of 0.
If now u D (G), then the convolution f = (1 u) 1 is well dened and
Lf = 1 u + 1 u ,
showing that Lf = 1 u on 0 and hence Lf = u on 0 . Hence L is locally solvable
at 0. By leftinvariance, it is locally solvable at any point.
160
Chapter 3. Homogeneous Lie groups
Because of the duality between hypoellipticity and solvability, local fundamental solutions exist under the following condition:
Proposition 3.2.39. Let L be a leftinvariant hypoelliptic operator on a connected
Lie group G. Then Lt is also leftinvariant and it has a local fundamental solution.
Proof. The rst statement follows easily from the denition of Lt , and the second
from the duality between solvability and hypoellipticity (cf. Theorem A.1.3) and
Lemma 3.2.38.
The next theorem describes some property of existence and uniqueness of
global fundamental solutions in the context of homogeneous Lie groups.
Theorem 3.2.40. Let L be a homogeneous leftinvariant dierential operator on
a homogeneous Lie group G. We assume that the operators L and Lt are hypoelliptic on a neighbourhood of 0. Then L admits a fundamental solution S (G)
satisfying:
(a) if < Q, the distribution is homogeneous of degree Q and unique,
(b) if Q, = o + p(x) ln x where
(i) o S (G) is a homogeneous distribution of degree Q, which is
smooth away from 0,
(ii) p is a polynomial of degree Q and,
(iii)   is any homogeneous quasinorm, smooth away from the origin.
Necessarily is smooth on G\{0}.
Remark 3.2.41. In case (a), the unique homogeneous fundamental solution is a
kernel of type , with the uniqueness understood in the class of homogeneous
distributions of degree Q. For case (b), Example 3.2.37 shows that one can not
hope to always have a homogeneous fundamental solution.
The rest of this section is devoted to the proof of Theorem 3.2.40.
The proofs of Parts (a) and (b) as presented here mainly follow the original
proofs of these results due to Folland in [Fol75] and Geller in [Gel83], respectively.
Proof of Theorem 3.2.40 Part (a). Let L be as in the statement and let < Q.
By Proposition 3.2.39, L admits a local fundamental solution at 0: there exist a
neighbourhood of 0 and a distribution
D () such that L
= 0 on . Note
that by the hypoellipticity of L,
as well as any fundamental solution coincide
with a smooth function away form 0. By shrinking if necessary, we may assume
that after having xed a homogeneous quasinorm, is a ball around 0. So if
x and r (0, 1] then rx .
Folland observed that if exists then the distribution h :=
annihilates
L on , so it must be smooth on , while
(rx) rQ h(rx)
(x) = rQ
3.2. Operators on homogeneous Lie groups
yields
161
(rx)
(x) = lim rQ
r0
and
h(x) =
(x) lim rQ
(rx).
r0
Going back to our proof, Follands idea was to dene hr D () by
rQ
Dr
hr :=
on \{0}, r (0, 1],
which makes sense in view of the smoothness of
on \{0}. Since for any test
function D(),
(r )), = rQ (L
)(r )), = L
, (r1 ) = (r1 0) = (0),
L(rQ
we have Lhr = 0 0 = 0. So hr is in NL () C () where the D () and
C () topologies agree, see Theorem A.1.6. Let us show that
lim hr h C ();
(3.58)
r0
for this it suces to show that {hr } is a Cauchy family in D ().
We observe that if s r, we have
hs (x) hr (x)
=
=
=
rQ
(rx) sQ
(sx)
s Q s
rx
rx
rQ
r
r
rQ h rs (rx).
(3.59)
In particular, setting s = r2 in (3.59) we obtain
hr2 = rQ hr Dr + hr .
This formula yields, rst by substituting r by r2 ,
hr 4
r2(Q) hr2 Dr2 + hr2
r2(Q) rQ hr Dr Dr2 + hr Dr2 + rQ hr Dr + hr
r3(Q) hr Dr3 + r2(Q) hr Dr2 + rQ hr Dr + hr .
Continuing inductively, we obtain
hr 2 =
2
1
rk(Q) hr Drk .
k=0
This implies
n N0
sup
x(1)
hr2 (x) (1 rQ )1
sup
x(1)
hr (x),
162
Chapter 3. Homogeneous Lie groups
and, since any s
r [ 14 , 12 ],
s
1
2
1
2
can be expressed as s = r2 for some N0 and some
sup
x(1)
hs (x) (1 2Q )1
sup
x(1)
1
1
4 r 2
hr (x).
Now the SchwartzTreves lemma (see Theorem A.1.6) implies that the topologies of D () and C () on
NL () = {f D () : T f = 0} C ()
coincide. Since r hr is clearly continuous from (0, 1] to D () NL (), {hr , r
[ 14 , 12 ]} and {hr , r [ 12 , 1]} are compact in D(). Therefore, we have
sup
x(1)
0<s1
hs (x)
sup
x(1)
0<s 12
(1 2Q )1
hs (x) +
sup
x(1)
1
1
4 r 2
sup
x(1)
1
2 s1
hr (x) +
hs (x)
sup
x(1)
1
2 s1
hs (x) = C < ,
that is, the hr s are uniformly bounded on (1 ). But if s < r, (3.59) implies
sup hs (x) hr (x) rQ sup h rs (rx) C rQ 0.
x(1)
x(1)
r0
This shows that {hr }r(0,1] is a Cauchy family of C(K) for any compact subset
K of . Therefore, {hr }r(0,1] is a Cauchy family of D () and Claim (3.58) is
proved. Let h C () be the limit of {hr }. Necessarily Lh = 0. We set
:=
h D ().
Now, on one hand
L = L
Lh = 0
and is smooth on \{0}, and on the other,
(rx),
(x) = lim rQ
r0
so if s (0, 1], then
(sx) = lim r
r0
Q
r
(srx) = lim
(r x) = sQ (x).
r =rs0
s
By requiring that the formula (sx) = sQ (x) holds for all s > 0 and x = 0, we
can extend into a distribution dened on the whole space. The homogeneity of
L guarantees that the equation L = 0 holds globally.
Finally, if 1 were another fundamental solution of L satisfying (a), then
1 would be ( Q)homogeneous with Q < 0; 1 would also be
smooth even at 0 since it annihilates L on G. Thus 1 = 0.
3.2. Operators on homogeneous Lie groups
163
Proof of Theorem 3.2.40 Part (b). Let L be as in the statement and let Q.
Let also
, and hr be dened as in the proof of part (a).
Geller noticed that Follands idea could be adapted by taking higher order
derivatives. Indeed from (3.59), we have
X hs (x) X hr (x) = rQ+[] X h rs (rx);
if Nn0 is large so that Q + [] > 0, we can proceed as for hr in the proof
of Part (a) and obtain that {X hr }r(0,1] is a Cauchy family of C ().
If [] Q, the C ()family {X hr }r(0,1] may not be Cauchy but by
Taylors theorem at the origin for homogeneous Lie groups, cf. Theorem 3.1.51,
(h )
x[]
sup
(X hr ) (z) ,
hr (x) P0,Mr (x) CM
M +1
[]>M
z M +1 x
for any x such that x and M +1 x are in the ball . Choosing M = Q and
(h )
setting the polynomial pr (x) := P0,Mr (x) and the ball := (M +1) , this
shows that the C ( )family {hr pr }r(0,1] is Cauchy. We set
C ( ) h := lim (hr pr ),
r0
o :=
h D( ).
Note that Lpr = 0, since the polynomial pr is of degree Q and the dierential
operator L is homogeneous. Therefore, Lo = 0 in and o C ( \{0}).
Furthermore, if [] > Q and x \{0} then
o (x) = lim rQ+[]
(rx),
r0
x
x
so if s (0, 1],
Q+[]
o (sx) = lim r
(rsx)
r0
x
x
Q+[]
Q[]
= lim
(r x) = s
(x).
r =rs0
s
x
x
One could describe this property as x
o being homogeneous on \{0}. We
conclude the proof by applying Lemma 3.2.42 below.
In order to state Lemma 3.2.42, we rst dene the set W of all the possible
homogeneous degrees [], Nn0 ,
W := {1 1 + . . . + n n : 1 , . . . , n N0 }.
(3.60)
In other words, W is the additive semigroup of R generated by 0 and WA .
For instance, in the abelian case (Rn , +) or on the Heisenberg group Hno ,
with our conventions, W = N0 . This is also the case for a stratied Lie group or
for a graded Lie group with g1 nontrivial.
164
Chapter 3. Homogeneous Lie groups
Lemma 3.2.42. Let B be an open ball around the origin of a homogeneous Lie
group G equipped with a smooth homogeneous quasinorm  . We consider the
sets of functions K dened by
if R\W
K := {f C (B\{0}) : f is homogeneous} ,
K := {f C (B\{0}) : f = f1 + p(x) ln x ,
if W
where f1 is homogeneous and p is a homogeneous polynomial} ,
where W was dened in (3.60), and we say that a function f on B or B\{0} is
homogeneous when f Ds = s f on B for all s (0, 1).
f K[] with [] > , then
For any R and f C (B\{0}), if x
there exists p P< such that f p K .
Recall (see Denition 3.1.26) that P<M denotes the set of polynomials P on
G such that D P < M . It is empty if M < 0.
Proof of Lemma 3.2.42. By induction it suces to prove that for any R and
f C (B\{0}),
(f pj )
Kj with pj P<j for all j = 1, . . . , n
xj
= f p K for some p P< .
(3.61)
To prove (3.61), we start by showing that for any f C (B\{0}),
f
Kj for all j = 1, . . . , n = f c K for some c C.
xj
(3.62)
By convention (see Denition 3.1.26), a homogeneous polynomial of homogeneous
degree which is not in W is 0. With this in mind we continue the proof of (3.62) in
a unied way. We consider f C (B\{0}) satisfying the hypothesis of (3.62): for
f
Kj and there exists pj P=j such that f pj ln  
each j = 1, . . . , n, x
j
is a homogeneous function on \{0}. We dene
A(r, x) := f (rx) r f (x),
x B, r (0, 1].
We see that
A(r, x)
xj
=
=
f
f
(rx) r
(x)
xj
xj
rj pj (rx) ln rx r pj (x) ln x = r pj (x) ln r.
r j
Note that for any j, k we have
pj
pk
=
xk
xj
since
A(r, x) =
A(r, x).
xk xj
xj xk
3.2. Operators on homogeneous Lie groups
165
Because of this observation we can adapt the proof of the Poincare Lemma to
construct the polynomial
q(x) := c
k xk pk (x),
(3.63)
k=1
which is homogeneous and satises
q
xj
=
=
ct=1 (pj (tx)) + cj pj (x) = c( j )pj (x) + cj pj (x)
pj (x),
k=1
k x k
pk (x)
pj (x)
+ cj pj (x) = c
k x k
+ cj pj (x)
xj
xk
k=1
by choosing c = 1 if = 0; if = 0, the polynomials pj and q are zero. So we
have
(A(r, x) q(x)r ln r) = 0 for all j = 1, . . . , n.
xj
Therefore,
A(r, x) = q(x)r ln r + a(r)
for some a C ((0, 1]).
Replacing f by f (r ln r) q we may assume that q = 0 in all the cases, so that
r (0, 1], x B
f (rx) r f (x) = a(r).
(3.64)
Now if 0 < r, s < 1, then using the formula just above twice, we get
a(rs)
=
=
f (rsx) (rs) f (x) = a(r) + r f (sx) (rs) f (x)
a(r) + r (a(s) + s f (x)) (rs) f (x)
a(r) + r a(s).
Solving this functional equation and setting
fo (x) := f (x) a(x)
(x G\{0}),
for a particular solution a, we check easily that fo is homogeneous:
If = 0, then a satises the functional equation
a(rs) = a(r) + a(s)
and must, therefore, be of the form a(r) = C ln(r) for some constant C C.
Using (3.64) we obtain
fo (rx) = f (rx) a(rx) = f (x) + a(r) a(rx) = f (x) C ln x = fo (x).
166
Chapter 3. Homogeneous Lie groups
If = 0, then a satises the functional equation
a(r) + r a(s) = a(s) + s a(r)
and must therefore be of the form a(r) = C(1 r ) for some constant C C.
Using (3.64) we obtain
fo (rx)
f (rx) C(1 rx ) = r f (x) + C(1 r ) C(1 rx )
r (f (x) C(1 x )) = r fo (x).
Hence (3.62) is proved and we can now go back to showing the main claim, that
is, the one given in (3.61). Let f and pj be as in the hypotheses of (3.61).
First we see that if < 0, then all the polynomials pj are zero and, inspired
by the construction of q above, we check easily that
!
"
n
f
f
k x k
,
1
=
xj
xk
xj
k=1
thus f and
n
1
k=1
f
k xk x
must coincide so (3.61) is proved in this case.
k
Let us assume 0. We claim that
j, k = 1, . . . , n
pj
pk
=
.
xj
xk
(3.65)
This is certainly true if j k < 0 since both are zero in this case. If instead
j k 0 then the polynomial
pk
pj
f
f
=
pk
pj
,
xj
xk
xj
xk
xk
xj
is in Kj k and thus must be zero. Indeed if a polynomial p is in some Ka then
either a W and then p = 0, or a W and p(rx) is a polynomial in r of degree
a with ra p(rx) unbounded unless p = 0; in both cases, p = 0.
Therefore, we can construct q as above by (3.63) so that
q
xj
= pj . Then
(f q)
f
=
pj Kj for all j = 1, . . . , n,
xj
xj
so f q K by (3.62).
This concludes the proof of Claim (3.61) and of Lemma 3.2.42.
Remark 3.2.43. The class of functions K dened in Lemma 3.2.42 is also used in
the denition of the calculus by Christ et al. [CGGP92].
As an application of Theorem 3.2.40, let us extend Liouvilles Theorem to
homogeneous Lie groups.
3.2. Operators on homogeneous Lie groups
167
3.2.8 Liouvilles theorem on homogeneous Lie groups
Let us consider the following statement and proof of Liouvilles Theorem in Rn :
Theorem 3.2.44 (Liouville). Every harmonic tempered distribution is a polynomial.
This means that if f S (Rn ) and f = 0 in the sense of distributions where
is the canonical Laplacian, then f is a polynomial on Rn .
Proof. Let f S (Rn ) with f = 0. Then 2 f = 0 where f is the Euclidean
Fourier transform of f S (Rn ) on Rn . Hence the distribution f is supported at
the origin and must be a linear combination of derivatives of the Dirac distribution
at 0, see Proposition 1.4.2. Consequently f is a polynomial.
Liouvilles Theorem and its proof given above are also valid for any homogeneous elliptic constantcoecient dierential operator on Rn . We now show the
following generalisation for homogeneous Lie groups:
Theorem 3.2.45 (Liouville theorem on homogeneous Lie groups). Let L be a homogeneous leftinvariant dierential operator on a homogeneous Lie group G. We
assume that L and Lt are hypoelliptic on G. If the distribution f S (G) satises
Lf = 0 then f is a polynomial.
The rest of this section is devoted to the proof of Theorem 3.2.45. We follow
the proof given by Geller in [Gel83].
Let denote the Euclidean Fourier transform on Rn (cf. (2.25)). In view of
the proof of Theorem 3.2.44, we want to show that the distribution f is supported
at 0. For this purpose, it suces to show that any test function S(G) whose
Euclidean Fourier transform is supported away from 0, that is, supp 0, can be
written as Lt for some S(G). Indeed, denoting momentarily (x) = x for
x G identied with Rn , and by the inverse Fourier transform on Rn , we have
and
= , so that supp = supp ,
= f, = f, Lt = Lf, = 0.
f,
The set of functions with 0 supp is contained in
n
n
n
(0) = 0, N0 .
So (R ) := S(R ) :
We observe that the space So (Rn ) can be also described in terms of the group
structure using the identication of G with Rn , as
n
So (R ) = So (G) = { S(G) :
x (x)dx = 0, Nn0 }.
G
) (0) with c a known nonzero constant. Here dx
Indeed Rn x (x)dx = c (
denotes the Lebesgue measure on Rn and the Haar measure on G since these two
measures coincide via the identication of G with Rn .
168
Chapter 3. Homogeneous Lie groups
By Theorem 3.2.40, the operator Lt has a fundamental solution S (G)
satisfying Part (a) or (b) of the statement. Thus we need only showing that for
any So (G), the function := is not only smooth (cf. Lemma 3.1.55) but
also Schwartz. This is done in the following lemma:
Lemma 3.2.46. If So (G) is a Schwartz function and S (G) is a homogeneous distribution smooth away from the origin or a distribution of the form
= p(x) ln x where p is a polynomial and   a homogeneous quasinorm smooth
away from the origin, then S(G).
The end of this section is devoted to the proof of Lemma 3.2.46; this relies
on consequences of the following versions of Hadamards Lemma for S(Rn ) and
So (Rn ):
Lemma 3.2.47 (Hadamard). Let f S(Rn ) with f = 0. Then f can be written
as
n
fj
with fj S(Rn )
f=
x
j
j=1
In addition, if f So (Rn ), each function fj can be also taken in So (Rn ).
Proof of Lemma 3.2.47. We x o D(Rn ) such that o () = 1 if  1 and
o () = 0 if  > 2. Since f = 0 we have f(0) = 0 and
f() = o f + (1 o )f = (o f) (o f)(0) + (1 o )f.
We can write
(o f)() (o f)(0) =
n
j
o f (t) dt =
and
(1 o )f() =
j=1
j2
j=1
We set
hj () :=
0
1 o ()
f ()
2
1
0
(here 2 =
(o f)
(t)dt,
j
n
j2 ).
j=1
(o f)
1 o ()
(t)dt + j
f ().
j
2
The rst term is compactly supported (in the ball of radius 2), whereas the second
one is well dened and is identically 0 on the
unit ball. Since both terms are
smooth, hj S(Rn ). We have obtained f = j j hj . We dene fj S(Rn ) such
fj
that fj = cj hj where the constant cj is such that j = cj j . Hence f =
.
Moreover, since
j xj
hj (0) =
we see that if f So (Rn ) then fj So (Rn ).
f (0),
j
3.2. Operators on homogeneous Lie groups
169
We will use the following consequence of Lemma 3.2.47 (in fact only the
second point):
Corollary 3.2.48.
If f So (Rn ), then for any M N0 ,
f=
f
with f So (Rn ).
x
=M
If f So (G) where G is a homogeneous Lie groups, then for any M 1, we
can write f as a nite sum
X f
f=
[]>M
with f So (G).
Proof of Corollary 3.2.48. Both points are obtained recursively, the rst one from
Lemma 3.2.47 and the second from
n the following observation: if f So (G), there
exists gj So (G) such that f = j=1 Xj gj . Indeed writing f as in Lemma 3.2.47
and using (3.17) with Remark 3.1.29 (1), we set
(pj,k fj )
gj := fj +
1kn
j <k
and we see that gj So (G).
We can now prove Lemma 3.2.46.
Proof of Lemma 3.2.46. Let be a distribution as in the statement. We can always
decompose as the sum of 0 + , where 0 has compact support and is
smooth. Indeed, let D(G) be identically 1 on a neighbourhood of the origin
and dene 0 by
0 , := , .
Then
:= 0
coincides with (1 )o , where o is a smooth function on G\{0} either homogeneous or of the form p(x) ln x; we denote by the homogeneous degree of the
function o or of the polynomial p.
Let So (G). Since the distribution 0 is compactly supported, we get, by
Lemma 3.1.55, that 0 S(G). Since, by Corollary 3.2.48, we can write as
a (nite) linear combination of X f with f So (G) and [] as large as we want.
We observe that
(X f ) = f X
170
Chapter 3. Homogeneous Lie groups
and that for [] larger that  + N + 1 for N N0 xed, we have
(x) CN (1 + x)N .
X
Thus
(X f ) (x)
f X (x) = f (y)X (y x)dy
G
f (y)CN (1 + y 1 x)N dy
G
CN CoN (1 + x)N
f (y)(1 + y)N dy,
G
by (3.43). This shows that S(G).
Hence Lemma 3.2.46 and Theorem 3.2.45 are proved.
Open Access. This chapter is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/),
which permits use, duplication, adaptation, distribution and reproduction in any medium
or format, as long as you give appropriate credit to the original author(s) and the source,
a link is provided to the Creative Commons license and any changes made are indicated.
The images or other third party material in this chapter are included in the works
Creative Commons license, unless indicated otherwise in the credit line; if such material
is not included in the works Creative Commons license and the respective action is not
permitted by statutory regulation, users will need to obtain permission from the license
holder to duplicate, adapt or reproduce the material.
Chapter 4
Rockland operators and Sobolev
spaces
In this chapter, we study a special type of operators: the (homogeneous) Rockland
operators. These operators can be viewed as a generalisation of subLaplacians to
the nonstratied but still homogeneous (graded) setting. The terminology comes
from a property conjectured by Rockland and eventually proved by Heler and
Nourrigat in [HN79], see Section 4.1.3.
First, we discuss these operators in general. Subsequently, we concentrate on
positive Rockland operators and study the heat semigroup, the Bessel and Riesz
potentials and the Sobolev spaces naturally associated with a positive Rockland
operator. Most results concerning the heat semigroup are known [FS82, ch.3.B].
To the authors knowledge, however, this chapter is the rst systematic presentation of the fractional powers and the homogeneous and inhomogeneous Sobolev
spaces associated with a positive Rockland operator on a graded Lie group.
In fact, this appears to be the greatest generality for such constructions, since
the existence of a Rockland (dierential) operator on a homogeneous Lie group
implies that the group must admit a graded structure, see Proposition 4.1.3. In
the case of stratied Lie groups, Sobolev spaces have been developed by Folland
[Fol75] for 1 < p < , for the Rockland operator being a subLaplacian (see also
[Sak79]). Since subLaplacians are not always available on graded Lie groups, our
constructions are based on general positive Rockland operators. In particular, this
allows one to still cover the case of stratied Lie groups, but permitting taking
Rockland operators other than a canonical subLaplacian.
Although we dene Sobolev spaces using a xed Rockland operator, Theorem 4.4.20 shows that these spaces are actually independent of the choice of a
homogeneous positive Rockland operator.
The Editor(s) (if applicable) and The Author(s) 2016
V. Fischer, M. Ruzhansky, Quantization on Nilpotent Lie Groups, Progress in Mathematics 314,
DOI 10.1007/9783319295589_4
171
172
Chapter 4. Rockland operators and Sobolev spaces
4.1 Rockland operators
We start with the discussion of general Rockland operators, giving denitions,
examples, and then relating them to the hypoellipticity questions.
4.1.1 Denition of Rockland operators
The rst denition of a Rockland operator uses the representations of the group.
We use the notation which has become quite conventional nowadays in this part
of the theory of group representations and which is explained in Section 1.7. In
denotes the unitary dual of G and H the smooth vectors of a
particular, G
see Denition 1.7.2. For a leftinvariant dierential operator
representation G,
T we will denote (T ) := d(T ), see Denition 1.7.4.
Denition 4.1.1. Let T be a leftinvariant dierential operator on a Lie group G.
Then T satises the Rockland condition when
except for the trivial representation,
(R) for each representation G,
the operator (T ) is injective on H , that is,
v H
(T )v = 0 = v = 0.
There is a similar denition of the Rockland condition for rightinvariant
dierential operators, and also for left or rightinvariant L2 (G)bounded operators
(for the latter, see Glowacki [Glo89, Glo91]). See also Section 4.4.8.
Denition 4.1.2. Let G be a homogeneous Lie group. A Rockland operator R on
G is a leftinvariant dierential operator which is homogeneous of positive degree
and satises the Rockland condition.
Some other authors may
operators as op dene nonhomogeneous Rockland
erators of the form R = [] c X with the main term []= c X satisfying the Rockland property given in (R). Here we have chosen to assume that a
Rockland operator is homogeneous to study directly the main term.
We will give examples of Rockland operators in Section 4.1.2. Before this,
we show that their existence on a homogeneous Lie group implies that the group
is graded and that the weights could be chosen in N. This property inuences
the examples we can produce, and the subsequent development of the theory of
pseudodierential operators.
Proposition 4.1.3. Let G be a homogeneous Lie group. If there exists a Rockland
operator on G then the group G is graded.
Furthermore, the dilations weights 1 , . . . , n satisfy
a 1 1 = . . . = a n n
for some integers a1 , . . . , an .
4.1. Rockland operators
173
This property was shown by Miller in [Mil80], with a small gap in the proof
later corrected by ter Elst and Robinson (see [tER97]).
Proof of Proposition 4.1.3. Let G be a homogeneous Lie group. Its Lie algebra g is
endowed with the dilations Dr = Exp(ln rA). Let the number n and {X1 , . . . , Xn }
be the basis described in Lemma 3.1.14. We assume that there exists a homogeneous Rockland operator R which we can write as
c X .
R=
[]=
We x an integer j n . Let : g R be the linear functional such that
j. Since Xj
/ [g, g],
(Xk ) = j,k , that is, (Xj ) = 1 while (Xk ) = 0 for any k =
is identically zero on [g, g]. We set for any X g:
(expG X) := exp (i(X)) .
This denes a onedimensional representation of G. Indeed, if x, y G, we can
write x = expG X and y = expG Y and we have
xy = expG X expG Y = expG (X + Y + Z)
with Z [g, g] by the BakerCampbellHausdor formula (see Theorem 1.3.2).
Thus, (Z) = 0 and we obtain
(xy)
exp (i(X + Y + Z)) = exp (i(X) + i(Y ))
exp (i(X)) exp (i(Y )) = (x)(y).
So is a onedimensional representation of G and we see that
(Xk ) = t=0 (etXk ) = t=0 exp (i(tXk )) = t=0 exp (it(Xk )) = ij,k .
As is a nontrivial onedimensional representation of G and R satises the
Rockland condition,
(R) =
c (X )
[]=
must be nonzero. We see that (X ) is always zero unless is of the form aej for
a N where ej is the multiindex with 1 in the jth place and zeros elsewhere; in
this case [] = j a. So must be of the form = j a for some integer a = aj N
which may depend on j. And this is true for any j = 1, . . . , n .
Since X1 , . . . , Xn generate the Lie algebra g, the other weights are linear
combinations with coecients in N0 of the j s, j n . This shows that the operators Dr = Exp( lnr A) are dilations over g with rational weights. By Lemma 3.1.9,
the group G is graded.
174
Chapter 4. Rockland operators and Sobolev spaces
Remark 4.1.4. Proposition 4.1.3 and Remark 3.1.8 imply that the natural context
for the study of Rockland operators is a graded Lie group endowed with a family
of dilations with integer weights.
One may further assume that the weights have no common divisor other than
1 but we do not assume so unless we specify it.
From the proof of Proposition 4.1.3, we see:
Corollary 4.1.5. Let G be a graded Lie group and let {X1 , . . . , Xn } be the basis
described in Lemma 3.1.14. We keep the notation of the lemma.
The homogeneous degree of any Rockland operator is a multiple of 1 , . . . , n .
If R is a Rockland operator satisfying Rt = R then its homogeneous degree
is even.
4.1.2 Examples of Rockland operators
On (Rn , +), it is easy to see that Rockland dierential operators are exactly the operators P (i1 , . . . , in ) where P is a polynomial which is homogeneous (for the
standard dilations) and does not vanish except at zero. For instance homogeneous
elliptic operators on Rn with constant coecients are Rockland operators. More
generally, let us prove that subLaplacians on a stratied Lie group are Rockland
operators. First let us recall their denition.
Denition 4.1.6. If G is a stratied Lie group with a given basis Z1 , . . . , Zp for the
rst stratum of its Lie algebra, then the leftinvariant dierential operator on G
given by
Z12 + . . . + Zp2
is called a subLaplacian.
For example, the canonical subLaplacian of the Heisenberg group Hno is
X12 + Y12 + . . . + Xn2o + Yn2o ,
see Examples 1.6.4, 3.1.2 and 3.1.3 for our notation regarding the Heisenberg
group.
Lemma 4.1.7. Any subLaplacian on a stratied Lie group is a Rockland operator
of homogeneous degree 2.
This could be seen as a consequence of famous powerful theorems, namely
from combining H
ormanders sums of squares and HelerNourrigat (see Theorems A.1.2 and 4.1.12 in the sequel) but we prefer to give a direct and easy proof.
Proof. Let
R = Z12 + . . . + Zp2
be a subLaplacian on the stratied Lie group G, where Z1 , . . . , Zp is a given basis
for the rst stratum V1 of the Lie algebra of G.
4.1. Rockland operators
175
Clearly R is a homogeneous leftinvariant dierential operator of degree 2.
Let G\{1}
and v H be such that (R)v = 0. Then
0
=
=
((R)v, v)H = ((Z1 )2 v, v)H + . . . + ((Zp )2 v, v)H
((Z1 )v, (Z1 )v)H . . . ((Zp )v, (Zp )v)H
(Z1 )v2H . . . (Zp )v2H ,
and hence
(Z1 )v = . . . = (Zp )v = 0.
Since {Z1 , . . . , Zp } generates linearly the rst stratum V1 of g and V1 generates
g as a Lie algebra, we see that (X)v = 0 for any vector X g. But since is
nontrivial and irreducible, this forces v to be zero.
Looking at the proof of Lemma 4.1.7, it is not dicult to construct the
classical Rockland dierential operators on graded Lie groups G:
Lemma 4.1.8. Let G be a graded Lie group of dimension n, i.e. G Rn . We denote
by {Dr }r>0 the natural family of dilations on its Lie algebra g, and by 1 , . . . , n
its weights. We x a basis {X1 , . . . , Xn } of g satisfying
Dr Xj = rj Xj , j = 1, . . . , n, r > 0.
If o is any common multiple of 1 , . . . , n , the operator
n
j=1
o
2
(1) j cj Xj
with
cj > 0,
(4.1)
is a Rockland operator of homogeneous degree 2o .
Proof. The operator R given in (4.1) is clearly a homogeneous leftinvariant dif
and v H be such
ferential operator of homogeneous degree 2o . Let G\{1}
that (R)v = 0. Then
0
((R)v, v)H =
(1) j cj ((Xj )
o
2
j
v, v)H
j=1
cj (Xj ) j vH ,
j=1
o
and hence (Xj ) j v = 0 for j = 1, . . . , n.
Let us observe the following simple fact regarding any positive integer p and
any Z U(g): the hypothesis (Z)p v = 0 implies that
if p is odd then (Z)p+1 v = (Z)(Z)p v = 0,
176
Chapter 4. Rockland operators and Sobolev spaces
whereas if p is even then
p
0 = ((Z)p v, v)H = (1)p/2 ((Z) 2 v, (Z) 2 v)H = (1)p/2 (Z) 2 v2H ,
p
and hence (Z) 2 v = 0.
Applying this argument inductively on Z = Xj and p = o /j , o /2j , . . . ,
we obtain that (Xj )v = 0 for each j. Hence v = 0.
Remark 4.1.9. By Proposition 4.1.3 and its proof, if a homogeneous Lie group G
admits a Rockland operator, then, up to rescaling the dilations (cf. Remark 3.1.8),
we may assume that the group G is graded and endowed with its natural family
of dilations {Dr }r>0 . Lemma 4.1.8 gives the converse: on such a group, we can
always nd a Rockland operator.
The proof of Lemma 4.1.8 can easily be modied using an adapted basis
constructed in Lemma 3.1.14 to obtain
Corollary 4.1.10. Let G be a graded Lie group endowed with a family of dilations
{Dr }r>0 . Let {X1 , . . . , Xn } be a basis of g as in Lemma 3.1.14. In particular, the
vectors X1 , . . . , Xn generate the Lie algebra g.
If o is any common multiple of 1 , . . . , n , the operator
n
j=1
o
2
(1) j Xj
(4.2)
is a Rockland operator of homogeneous degree 2o .
If the group G is stratied, the vectors X1 , . . . , Xn span linearly the rst
stratum and we obtain the subLaplacian if we choose o = 1 .
From one Rockland operator, we can construct many since powers of a Rockland operator or its complex conjugate operator are Rockland:
Lemma 4.1.11. Let R be a Rockland operator on a graded Lie group G endowed
with a family of dilations with integer weights. Then the operators Rk for any
are also Rockland operators.
k N and R
as an element of U(g) was dened in (1.8).
The operator R
and Rk are leftinvariant homogeneous dierential operProof. It is clear that R
ators on G.
Let G\{1}.
We have
= (R).
(R)
This holds in fact for any leftinvariant dierential operator viewed as an element
is Rockland. For the case of Rk , let v H be such
of U(g). Therefore, R
k
that (R )v = 0. Applying recursively the simple fact explained in the proof of
Lemma 4.1.8, we obtain (R)v = 0 and this implies v = 0 because R is Rockland.
Therefore, Rk is also Rockland.
4.1. Rockland operators
177
4.1.3 Hypoellipticity and functional calculus
The analysis of leftinvariant homogeneous operators on a nilpotent graded Lie
group has played a very important role in the understanding of hypoellipticity.
We refer the interested reader on this subject to the lecture notes by Heler and
Nier [HN05]. For the denition of hypoellipticity, see Section A.1.
In [Roc78], Rockland showed that if T is a homogeneous leftinvariant differential operators on the Heisenberg group Hno , then the hypoellipticity of T
and T t is equivalent to the Rockland condition (see Denition 4.1.1). He also
asked whether this equivalence would be true for more general homogeneous Lie
groups. Just afterwards, Beals showed [Bea77b] that the hypoellipticity of a homogeneous leftinvariant dierential operator on any homogeneous Lie group implies
the Rockland condition. At the same time he also showed that the converse holds
in some steptwo cases. Eventually in [HN79], Heler and Nourrigat settled what
has become Rocklands conjecture by proving the following equivalence:
Theorem 4.1.12. Let R be a leftinvariant and homogeneous dierential operator on a homogeneous Lie group G. The hypoellipticity of R is equivalent to R
satisfying the Rockland condition.
In this case, any operator of the form
c X ,
R+
[]<
where is the degree of homogeneity of R and c any complex number, is also
hypoelliptic.
via Kirillovs orbit
The proof of Theorem 4.1.12 relies on the description of G
method.
Remark 4.1.13. 1. The hypotheses of Theorem 4.1.12 with the existence of a
Rockland operator imply that the family of dilations of the group may be
rescaled to have integer weights and consequently that the group may be
viewed as graded, see Proposition 4.1.3. When describing properties of a
Rockland operator R on a homogeneous Lie group G, unless stated otherwise,
we will always assume that the group G is graded in such a way that the
operator R is homogeneous for the natural family of dilations (with integer
weights).
2. Combining the theorems of HellferNourrigat and of Hormander (see Theorems 4.1.12 and A.1.2) gives another proof that the subLaplacians are Rockland operators, see Lemma 4.1.7.
3. If R is a Rockland operator formally selfadjoint, i.e. R = R as elements of
must also be Rockland by Lemma 4.1.11. Hence Theorem
U(g), then Rt = R
4.1.12 implies that any formally selfadjoint Rockland operator satises the
hypothesis of Theorem 3.2.40 and thus admits fundamental solutions. It also
satises the hypothesis of the Liouville theorem as in Theorem 3.2.45.
178
Chapter 4. Rockland operators and Sobolev spaces
4. Let us also mention an alternative reformulation of the HellferNourrigat
theorem given by Rothschild [Rot83]: a leftinvariant homogeneous operator
R on a graded Lie group G is hypoelliptic if and only if there is no nonconstant bounded function f on G such that Rf = 0 on G. The proof of
this relies on the Liouville theorem from Section 3.2.8. Essentially, in one
direction this is Beals result as above, while in the other it will follow from
Corollary 4.3.4.
Along the proof of Theorem 4.1.12 (see [HN79, Estimate (6.1)]), Heler and
Nourrigat also showed the following property which will be used in the sequel.
Corollary 4.1.14. Let G be a graded Lie group endowed with a family of dilations
with integer weights. Let R be a Rockland operator G of homogeneous degree .
Then there exists C > 0 such that
X 2L2 (G) C R2L2 (G) + 2L2 (G) .
S(G)
[]=
After developing the Sobolev spaces on G, we will be actually able to prove
its Lp version, see Lemma 4.4.19.
The following property of Rockland dierential operators is technically important and relies on hypoellipticity.
Proposition 4.1.15. Let R be a Rockland operator on a graded Lie group G. We
assume that R is formally selfadjoint. Let be a strongly continuous unitary
representation of G.
Then the operators R and (R) densely dened on D(G) L2 (G) and H
H , respectively, are essentially selfadjoint.
That R is formally selfadjoint means that R = R as elements of the universal enveloping algebra U(g), see (1.9).
Before we prove it, let us point out its consequences:
Corollary 4.1.16 (Functional calculus of Rockland operators and their Fourier
transform). Let R be a Rockland operator on a graded Lie group G. We assume that R is formally selfadjoint as an element of U(g). Then R is essentially
selfadjoint on L2 (G) and we denote by R2 its selfadjoint extension on L2 (G).
Moreover, for each strongly continuous unitary representation of G, (R) is
essentially selfadjoint on H and we keep the same notation for its selfadjoint
extension. Let E, E be the spectral measures of R2 and (R):
dE() and (R) =
dE ().
R2 =
R
For any Borel subset B R, the orthogonal projection E(B) is leftinvariant
hence E(B) LL (L2 (G)). The group Fourier transform of its convolution kernel
E(B)0 K(G) is
FG (E(B)0 )() = E (B).
4.1. Rockland operators
179
If is a measurable function on R, the spectral multiplier operator (R2 ) is
dened by
()dE(),
(R2 ) :=
R
and its domain
Dom((R2 )) is the space of function f L2 (G) such that the
integral R ()2 d(E()f, f ) is nite. It satises for all f Dom((R2 )) and
r > 0:
f (r ) Dom((r R2 ))
and
(R2 )f = (r R2 ) (f (r )) (r1 ).
(4.3)
If 1 is another strongly continuous representation such that 1 T , that
is, T is a unitary operator satisfying T 1 = T , then T E1 = E T and we have
for any measurable function the equality
T (1 (R)) = ((R))T.
(4.4)
Let L (R) be any measurable bounded function. Then the spectral multiplier operator (R2 ) is in LL (L2 (G)), that is, it is bounded on L2 (G) and leftinvariant. Its convolution kernel denoted by (R2 )o is the unique tempered distribution (R2 )o S (G) such that
f S(G)
(R2 )f = f (R2 )o .
In fact (R2 )o K(G) and its group Fourier transform is
()dE ().
F{(R2 )o }() = ((R)) =
(4.5)
Consequently, for any f L2 (G),
F{(R2 )f }() = ((R))f().
(4.6)
We have for any r > 0 and x G:
(r R2 )o (x) = rQ (R2 )o (r1 x).
(4.7)
For any L (R),
{(R2 )0 } = (R)
0,
where
{(R2 )0 } (x) = (R2 )0 (x).
(4.8)
If is also realvalued, then (R2 ) is a selfadjoint operator and its kernel satises
(R2 )o = ((R2 )o ) , that is, in the sense of distributions,
(R2 )o (x) = (R2 )o (x1 ).
If is realvalued and furthermore if Rt = R, then (R2 )o is realvalued
(as a distribution).
180
Chapter 4. Rockland operators and Sobolev spaces
Remark 4.1.17. For any measurable function : R C such that for every
of
1 Rep G, the domain of (1 (R)) contains H1 , the corresponding Geld
operators {((R)) : H H } is well dened in the sense of Denition 1.8.13
because of (4.4). This is the case for instance if is bounded since in this case
(1 (R)) is a bounded and therefore dened on the whole space H1 .
The rest of this section is devoted to the proof of Proposition 4.1.15 and
Corollary 4.1.16; it may be skipped at rst reading. Proposition 4.1.15 follows
from a Theorem by Nelson and Stinespring [NS59, Theorem 2.2] regarding elliptic
operators on Lie groups as well as the adaptation of its proof due to Folland and
Stein [FS82, ch.3.B] to our case. Let us sketch briey the ideas for the sake of
completeness. Nelson and Stinesprings Theorem can be reformulated here as the
following:
Proposition 4.1.18. Let R be a Rockland operator on a graded Lie group G. We
assume that R is formally selfadjoint as an element of U(g).
If is a strongly continuous unitary representation of G, then the closure of
(R ) is the adjoint of (R).
Proof of Proposition 4.1.18. Let v H be orthogonal to the range of (R) + I.
Then for all D(G),
(R + I)(x) ((x) v, v)H dx.
0 = (((R) + I)()v, v)H =
G
In other words, the continuous function f dened by
f (x) := ((x) v, v)H = (v, (x)v)H ,
x G,
is a solution in the sense of distributions of the partial dierential equation (R +
I)f = 0. By Theorem 4.1.12, the operator R+I is hypoelliptic. Hence f is smooth
on G and the equation (R + I)f = 0 holds in the ordinary pointwise sense. We
observe that for any X U(g) identied with a leftinvariant vector eld we have
Xf (x) = t=0 v, (xetX )v H = (v, (x)(X)v)H .
Thus,
(R + I)f (x) = (v, (x)(R)v)H + (v, (x)v)H .
Therefore, (R + I)f (0) = 0 implies
(v, (R)v)H = (v, v)H = v2H .
If R can be written as S S for some nonconstant S U(g), then the lefthand
side is equal to (S)v2 so v = 0. In the general case, we apply the argument
above to R R = R2 which is also a Rockland operator by Lemma 4.1.11, and we
obtain the desired conclusion thanks to the following lemma applied to T = (R),
T = (R ) and D = H .
4.1. Rockland operators
181
Lemma 4.1.19. Let D be a dense vector subspace of a Hilbert space H. Let T and
T be two linear operators on H, whose domains are D and whose ranges are
contained in D such that T is contained in the adjoint of T . If T T is essentially
selfadjoint then the closure of T is the adjoint of T .
Proof of Lemma 4.1.19. We denote by T the adjoint of T . Let (u, v) be an element
of the graph of T which is orthogonal to the graph of T . This means
v = T u
and
w D
(u, w)H + (v, T w)H = 0.
In particular, for w = T x with x D, we obtain
0 = (u, T x)H + (v, T T x)H = (v, x)H + (v, T T x)H ,
x D.
But it is not dicult to see that I + T T has a dense range. Consequently v = 0.
So (u, w)H = 0 for all w D and therefore u = 0. This shows that the graph of
T contains no nonzero element orthogonal to the graph of T ; hence the closure
of T is T .
Proof of Proposition 4.1.15. We apply Proposition 4.1.18 to the left regular action
on L2 (G) and the strongly continuous unitary representation of G.
Proof of Corollary 4.1.16. Applying the spectral theorem to the selfadjoint operators R2 and (R) (see, e.g., Rudin [Rud91, Part III]) we obtain the spectral
measures E and E together with the denition of the spectral multipliers.
For each xo G and r > 0 we set for any Borel set B R and any function
f L2 (G),
E (xo ) (B)f
E (r) (B)f
:=
:=
(E(B)) (f (xo )) (x1
o ),
E(r B) (f (r )) (r1 ),
where the dilation of a subset of R is dened in the usual sense. It is not dicult
to check that this denes new spectral measures E (xo ) and E (r) and, that for any
function f S(G),
1
dE (xo ) ()f =
d (E()) (f (xo )) (x1
o ) = R2 (f (xo )) (xo )
R
R (f (xo )) (x1
o ) = Rf = R2 f,
dE (r) ()f =
(r )d (E()) (f (r )) (r1 ) = r R2 (f (r )) (r1 )
=
r R (f (r )) (r1 ) = Rf = R2 f,
since R is leftinvariant and homogeneous. By density of S(G) in L2 (G), we have
obtained for any f L2 (G) that
dE (xo ) ()f = R2 f and
dE (r) ()f = R2 f.
R
182
Chapter 4. Rockland operators and Sobolev spaces
By uniqueness of the spectral measure of R2 , the spectral measures E (xo ) , E (r)
and E coincide. For E (r) this implies (4.3).
For E (xo ) this means that for each Borel subset B R, the projection E(B)
is a leftinvariant operator on L2 (G). By the Plancherel theorem (see Section 1.8.2)
the group Fourier transform of its convolution kernel E(B)0 K(G) satises
f L2 (G)
(E(B)f ) = (E(B)0 )(f ).
(4.9)
It is not dicult, using the uniqueness of the group Fourier transform, to check
that
F : B (E(B)0 ) =: F (B),
is a spectral measure on H . Equality (4.9) can be rewritten for any f L2 (G) as
FG
()dE()f () =
()dF () f(),
(4.10)
R
with = 1B , that is, the characteristic function of a Borel subset B R. Hence
Equality (4.10) also holds for a nite linear combination of characteristic functions,
and then, passing through the limit carefully, for any L (R) with f L2 (G)
and () = for f S(G). The latter yields
dF () f () = FG
dE()f ()
R
= FG (R2 f )() = (R)f().
Since the space H of smooth vectors is linearly spanned by elements of the form
f()v, f S(G), v H (see Theorem 1.7.8), we have on H
dF () = (R).
R
The uniqueness of the spectral measure E shows that
E (B) = F (B) = (E(B)0 ).
Equality (4.5) follows from (4.10) for L (R).
(T )
If 1 T , then we set E := T E1 T 1 , where E1 denotes the spectral
(T )
measure of 1 (R). We check easily that E is a spectral measure on H and that
(T )
dE = T
dE1 T 1 = T 1 (R)T 1 = T 1 T 1 (R) = (R).
R
The property of the spectral measure E , that is, its uniqueness and the functional
(T )
calculus, shows that E = E and that (4.4) holds.
The rest of the statement follows from the Schwartz kernel theorem (see
Corollary 3.2.1) and basic properties of the convolution.
4.2. Positive Rockland operators
4.2
183
Positive Rockland operators
In this section we concentrate on positive Rockland operators, i.e. Rockland operators which are positive in the operator sense. Positive Rockland operators always
exist on a graded Lie group, see Remark 4.2.4 below. Among Rockland operators,
positive ones enjoy a number of additional useful properties. In particular, in this
section, we analyse the heat semigroup associated to a positive Rockland operator
and the corresponding heat kernel.
4.2.1
First properties
We shall be interested in Rockland dierential operators which are positive in the
sense of operators:
Denition 4.2.1. An operator T on a Hilbert space H is positive when for any
vectors v, v1 , v2 H in the domain of T , we have
(T v1 , v2 )H = (v1 , T v2 )H
and
(T v, v)H 0.
In the case of leftinvariant dierential operator, this is easily equivalent to
Proposition 4.2.2. Let T be a leftinvariant dierential operator on a Lie group G.
Then T is positive on L2 (G) when T is formally selfadjoint, that is, T = T in
U(g), and satises
f D(G)
T f (x)f (x) dx 0.
G
For the denition of T , see (1.9).
The following properties of positive operators are easy to prove:
Lemma 4.2.3. 1. A linear combination with nonnegative coecients of positive
operators is a positive operator.
p
2. If X is a leftinvariant vector eld and p 2N0 , then the operator (1) 2 X p
is positive on G.
3. If T is a positive dierential operator on G then for any k N the dierential
operator T k is also positive.
Proof. The rst property is clear.
The second is true since each invariant vector eld is essentially skewsymmetric, see Section 1.3.
Let us prove the third property. Let T be a positive dierential operator and
k N. Clearly T k is also formally selfadjoint and we obtain recursively if k = 2:
T f (x)2 dx,
T k f (x)f (x)dx =
T f (x)T f (x)dx =
G
184
Chapter 4. Rockland operators and Sobolev spaces
which is necessarily nonnegative, whereas if k = 2 + 1,
T f (x)f (x)dx =
T (T f (x)) T f (x)dx,
which is nonnegative since T is positive.
We observe that the signs of the coecients of a positive dierential operator
can not be guessed, as the example (1 2 )2 on R2 shows.
Remark 4.2.4. By Lemma 4.2.3, Parts 1 and 2, we see that the examples in Section
4.1.2 yield positive Rockland operators. For instance, on stratied Lie groups,
the subLaplacians give operators R with R positive and Rockland. Also, the
operators in (4.1) and (4.2) give positive Rockland operators. In particular, this
shows that any graded Lie group admits a positive Rockland operator.
We may obtain other positive Rockland operators as powers of those since a
direct consequence of Lemma 4.1.11 and Lemma 4.2.3, Part 3, is the following
Lemma 4.2.5. Let R be a positive Rockland operator on a graded Lie group G.
= Rt are also positive Rockland operators.
Then Rk for every k N and R
We x a positive Rockland operator R. By Proposition 4.2.2, R is essentially
selfadjoint and we may adopt the same notation as in Corollary 4.1.16. Since R
is positive, the spectrum of R2 is included in [0, ) and we have
R2 =
dE().
Proposition 4.2.6. Let R be a positive Rockland operator on a graded Lie group G.
then the operator (R) is positive. Furthermore, if is nontrivial and
If G,
((R)v, v)H = 0
then v = 0.
Proof. By Proposition 4.1.15, (E(B)) = E (B). Since E is supported in [0, )
then so is E and the operator (R) is positive:
v
((R)v, v)H =
d(E ()v, v)H 0.
If ((R)v, v)H = 0 then the (real nonnegative) measure (E ()v, v)H is concentrated on { = 0} and this means that v = E (0)v is in the nullspace of (R).
Thus v = 0 since R satises the Rockland condition and is nontrivial.
4.2. Positive Rockland operators
185
4.2.2 The heat semigroup and the heat kernel
In this section, we x a positive Rockland operator R which is homogeneous of
degree N.
By the functional calculus (see Corollary 4.1.16), we dene the multipliers
etR2 :=
et dE(), t > 0.
0
We then have
etR2 L (L2 (G)) sup et  = 1
and
etR2 esR2 = e(t+s)R2 ,
since es et = e(t+s) . Thus {etR2 }t>0 is a contraction semigroup of operators on L2 (G) (see Section A.2). This semigroup is often called the heat semigroup. The corresponding convolution kernels ht S (G), t > 0, are called heat
kernels. We summarise its main properties in the following theorem:
Theorem 4.2.7. Let R be a positive Rockland operator on a graded Lie group G.
Then the heat kernels ht associated with R satisfy the following properties.
Each function ht is Schwartz and we have
s, t > 0
x G, t, r > 0
ht h s
hr t (rx)
x G
G
ht+s ,
(4.11)
(4.12)
ht (x)
ht
ht (x)dx
1.
ht (x),
(x1 ),
(4.13)
(4.14)
The function h : G R C dened by
ht (x) if t > 0 and x G,
h(x, t) :=
0
if t 0 and x G,
is smooth on (G R)\{(0, 0)} and satises
(R + t )h = 0,0 ,
where 0,0 is the deltadistribution at (0, 0) G R.
Having xed a homogeneous norm   on G, we have for any N N0 , Nn0
and N0 , that
C = C,N, > 0
t (0, 1]
sup t X ht (x) C,N tN .
(4.15)
x=1
The proof of Theorem 4.2.7 is given in the next section. We nish this section
with some comments and some corollaries of this theorem.
186
Chapter 4. Rockland operators and Sobolev spaces
Remark 4.2.8. 1. If the group is stratied and R = L where L is a subLaplacian, then R is of order two and the proof relies on Hunts theorem
[Hun56], cf. [FS82, ch1.G]. In this case, the heat kernel is realvalued and
moreover nonnegative. The heat semigroup is then a semigroup of contraction which preserves positivity.
2. The behaviour of the heat kernel in the general case is quite well understood.
For instance, it can be extended to the complex righthalf plane. Then the
heat kernel hz with z C, Re z > 0 decays exponentially. See [Dzi93, DHZ94,
AtER94].
3. Since R2 is a positive operator, only the values of L (R) on [0, ) are
taken into account for the multipliers (R2 ). But in fact, the value at 0 can
be neglected too, as a consequence of the property of the heat kernel. Indeed,
from ht S(G) and (4.12), it is not dicult to show
f ht L2 (G) 0,
t
rst for f D(G) and then by density for any f L2 . This shows
etR2 f L2 (G) 0,
t
and therefore we have
0
dE()L2 (G) 0.
0
4. Another consequence of the heat kernel being Schwartz, proved in [HJL85],
is that the spectrum of (R) is discrete and lies in (0, ) for any G\{1}.
Indeed, it is easy to see that (R) is the innitesimal generator of the semigroup {(etR )}t>0 in H and that (etR ) = (ht ) is a compact operator
since ht S(G) (for this last property, see [CG90, Theorem 4.2.1]).
Moreover, strong properties of the eigenvalue distributions of (R) are
known, see [tER97].
Theorem 4.2.7 shows that the functions ht provide a commutative approximation of the identity, see Remark 3.1.60. We already know that {etR2 }t>0 is
a strongly continuous contraction semigroup. Moreover, we have the following
properties for any p:
Corollary 4.2.9. The operators
f f ht , t > 0,
form a strongly continuous semigroup on Lp (G) for any p [1, ) and on Co (G).
Furthermore, for any f D(G) and any p [1, ] (nite or innite), we have
the convergence
1
(f ht f ) Rf t0 0.
(4.16)
t
p
4.2. Positive Rockland operators
187
Finally, we formulate a simple but useful corollary of Theorem 4.2.7.
Corollary 4.2.10. Setting r = t in (4.12), we get
1
ht (x) = t h1 (t x)
x G, t > 0
(4.17)
and
for x G\{0} xed, Xx h(x, t) =
Q+[]
O(t ) as t ,
O(tN ) for all N N0 as t 0.
(4.18)
Inequalities (4.18) are also valid for any x in a xed compact subset of G\{0}.
4.2.3 Proof of the heat kernel theorem and its corollaries
This section is entirely devoted to the proofs of Theorem 4.2.7 and Corollaries 4.2.9
and 4.2.10. This may be skipped at rst reading. The proofs essentially follow the
arguments of Folland and Stein [FS82, Ch. 4. B].
Since ht is the convolution kernel of the R2 multiplier operator, Corollary
4.1.16 yield that ht S (G) is a distribution which satises Properties (4.12) and
(4.13) for each t > 0 xed. Note that (4.12) easily yields (4.17).
By the Schwartz kernel theorem (see Corollary 3.2.1), since (0, ) t
etR2 L (L2 (G)) is a strongly continuous mapping, the function (0, ) t
ht S (G) is continuous. Consequently the mapping (t, x) ht (x) is a distribution on (0, ) G.
By the properties of semigroups (cf. Proposition A.2.3 (4)), we have
D(G), t > 0,
t (etR2 ) = R2 (etR2 ) = R(etR2 ).
Taking this equation at 0G shows that (t, x) ht (x) is a solution in the sense of
distributions of the equation (t + R)f = 0 on (0, ) G.
The next lemma is independent of the rest of the proof and shows that t +R
can be turned into a Rockland operator:
Lemma 4.2.11. Let R be a positive Rockland operator on a graded Lie group G.
We equip the group H := G R (which is the direct product of the groups G and
(R, +)) with the dilations
Dr (x, t) := (rx, r t),
x G, t R.
The group H has become a homogeneous Lie group and the operators R + t
and R t are Rockland operators on H.
R:
Proof of Lemma 4.2.11. The dual of H is easily seen to be isomorphic to G
and R, we can construct the representation = , of H on
if G
H = H by (x, t) := eit (x);
188
Chapter 4. Rockland operators and Sobolev spaces
can be realised into a representation of
conversely, any representation H
the form , .
We observe that H = H , (R) = (R), and (t ) = i.
Let = , H.
If v H is such that (R + t )v = 0 then
0 = ((R t )v, v)H = ((R)v, v)H i(v, v)H = ((R)v, v)H iv2H .
Since, by Proposition 4.2.6, ((R)v, v)H 0, the real part of the previous equal
ities is ((R)v, v)H = 0. Again by Proposition 4.2.6, necessarily v = 0.
Remark 4.2.12. A similar proof implies that R tk for k N odd is a Rockland
operator on the group G R endowed with the dilations Dr (x, t) = (rx, r/k t).
Corollary 4.2.13. The distribution (t, x) ht (x) is smooth on (0, ) G and
satises the equation
(t + R)f = 0.
Furthermore, for any t > 0, ht L2 (G) and
2
Q
h1 (x)2 dx < .
ht (x) dx = t
G
(4.19)
Proof. The operator t + R is Rockland on G R by Lemma 4.2.11, therefore
hypoelliptic by the HellferNourrigat theorem (see Theorem 4.1.12). Since the
distribution (t, x) ht (x) is a solution of the equation (t +R)f = 0 on (0, )G,
it is in fact smooth.
Since R is a positive Rockland operator, Rt is also a positive Rockland
operator (see Lemma 4.2.5) and we can apply Lemma 4.2.11 to both. Therefore,
R + t and its transpose are Rockland and thus hypoelliptic on G R. By the
SchwartzTreves theorem (see Theorem A.1.6), the distribution topology on G
(0, ) and the C topology agree on the the nullspace of R + t
N = {f D (G (0, )) : (R + t )f = 0}.
Since (0, ) t ht S (G) is continuous and (t, x) ht (x) is smooth on
(0, ) G, the mapping T dened via
tR2
)(x) =
ht (x)(x)dx, L2 (G), x G, t > 0,
T (x, t) = (e
G
is continuous from L (G) to D (G (0, )). Furthermore, the semigroup properties imply that the range of T lies in N . Therefore, the mapping
(x)h1 (x)dx,
L2 (G) T (0, 1) =
2
is a continuous functional. Hence h1 must be square integrable.
By homogeneity (see (4.17)), for any t > 0, we see that ht L2 (G) as a
consequence of Corollary 4.2.10 and (4.19) must hold.
4.2. Positive Rockland operators
189
We now dene the function h : G R C as in the statement of Theorem
4.2.7 by
ht (x) if t > 0 and x G,
h(x, t) :=
0
if t 0 and x G.
By Corollary 4.2.13, the function h is smooth on G (R\{0}) and satises the
equation (R + t )h = 0 on G (R\{0}). However, it is not obvious that it is a
distribution on G R. Our next goal is to prove that it is indeed a distribution
and that it satises the equation (R + t )h = 0 on G R.
It is easy to prove that h is a distribution under the assumption > Q/2
since it is then locally integrable:
Lemma 4.2.14. If > Q/2, then h is locally integrable on G R.
Proof of Lemma 4.2.14. We assume > Q/2. We see that for any > 0 and
R > 0, using the homogeneity property given in (4.19),
0
x<R
h(x, t)dxdt
" 12 !
!
x<R
1
2
B(0, 1) R
= CR
ht (x) dx
" 12
Q/2
Q
Q/2 1 2
h1 (x) dx
2
dt
1dx
x<R
12
t 2 dt
since we assumed > Q/2. This shows that h is locally integrable on G R and
hence denes a distribution.
If we know that h is a distribution, being a solution of (R + t )h = 0,0 is
almost granted:
Lemma 4.2.15. Let us assume that h D (G R) is a distribution and that
either h1 L2 (G) and > Q/2,
or h1 L1 (G) (without restriction on > Q/2).
Then h satises the equation
(R + t )h = 0,0
as a distribution.
The proof of Lemma 4.2.15 will require the following technical property which
is independent of the rest of the proof:
Lemma 4.2.16. Let R be a positive Rockland operator on a graded Lie group G
Rn with homogeneous degree . If m n2 , the functions in the domain of Rm
are continuous on , i.e.
Dom(Rm ) C(),
190
Chapter 4. Rockland operators and Sobolev spaces
where C() denotes the space of continuous functions on . Furthermore, for any
compact subset of G, there exists a constant C = C,R,G,m such that
Dom(Rm )
sup (x) C (L2 + Rm L2 ) .
This is a (very) weak form of Sobolev embeddings. We will later on obtain
stronger results in Theorem 4.4.25. The proof below uses Corollary 4.1.14 showed
by Heler and Nourrigat during their proof of Theorem 4.1.12.
Proof of Lemma 4.2.16. By the classical Sobolev embedding theorem on Rn , see
e.g. [Ste70a, p.124], if L2 (Rn ) together with x L2 (Rn ) for any multiindex
satisfying  n2 , then may be modied on a set of zero measure so that
the resulting function, still denoted by , is continuous.
Furthermore, for any compact subset of G, we may choose a closed ball
B(0, R) strictly containing , and there exists a constant C = C,R independent
of such that
sup  C
x L2 (B(0,R)) .
 n
2
As the abelian derivatives may be expressed as linear combination of leftinvariant ones, see Section 3.1.5, there exists another constant C = CR such that
x L2 (B(0,R)) C
X L2 (B(0,R))
 n
2
 n
2
for any such that the righthand side makes sense. By the corollary of the HelerNourrigat theorem applied to Rm (see Corollary 4.1.14, see also Lemma 4.2.5),
there exists C = CR,m > 0 such that
S(G)
X L2 (G) C Rm L2 (G) + L2 (G) .
[]m
The last two properties yield easily
x L2 (B(0,R)) C Rm L2 (G) + L2 (G) ,
 n
2
for any function L2 (G) for which the righthand side makes sense, for some
constant C = CR,R,m independent of , as long as m n2 . Together with
the embedding property recalled at the beginning of the proof, this shows Lemma
4.2.16.
We can now go back to the proof of the heat kernel theorem, and more
precisely, the proof of Lemma 4.2.15.
4.2. Positive Rockland operators
191
Proof of Lemma 4.2.15. If we set for each > 0 and (x, t) G R,
h(x, t) if t > ,
h() (x, t) :=
0
if t ,
it is clear that this denes a distribution h() D (GR) and that {h() } converges
to h in D (G R) as tends to 0. To prove that
(R + t )h = 0,0 ,
it suces to show that (R + t )h() converges to 0,0 in D (G R) as tends to
0; this means:
D(G R)
D
h() , (Rt t ) = (R + t )h() , (0).
0
Using the translation of the group H = G R which is the direct product of the
groups G and (R, +), this is equivalent to the pointwise convergence in H:
D(H), (x, t) H
(R + t )( h() )(x, t) (x, t),
0
(4.20)
since
(R + t )( h() )(x, t) = ((R + t )h() )(x, t) = (R + t )h() , ((x, t) 1 ).
The above convolution is in H, given by
( h() ) (x, t) =
(y, u) h() ((y, u)1 (x, t))dydu
G R
t
(y, u) h(y 1 x, u + t)dydu.
=
u=
We see that
(R + t )( h() )(x, t)
t
u=
=
+
G
(y, u) (Rx + t )h(y 1 x, u + t)dydu
(y, t ) h(y 1 x, )dy,
and the rst term of the right hand side is zero since (R + t )h = 0 on G (0, )
and R + t is leftinvariant on H. Hence
(R + t )( h() )(x, t) = (, t ) h (x),
(4.21)
using the convolution in H and G for the left and right hand sides respectively.
We now x t and set (y) := (y, t ). Then
(, t ) h = h ,
192
Chapter 4. Rockland operators and Sobolev spaces
and we can write
h 0 = ( 0 ) h (0 h 0 ).
(4.22)
For the rst term in the righthand side of (4.22), we need to separate the case
h1 L2 (G) with > Q/2 from the case h1 L1 (G). Indeed if h1 L2 (G) with
> Q/2, then by (4.19),
Q
h 2 = 2 h1 2
and the CauchySchwartz inequality yields
( 0 ) h 0 2 h 2 .
We easily obtain 0 2 C as D(G R). Thus
Q
( 0 ) h C 1 2 0 0,
since we assumed > Q/2. If h1 L1 (G), then by (4.19), h 1 = h1 1 and the
Holder inequality yields
( 0 ) h 0 h 1 = h1 1 0 .
Again 0 2 C as D(G R) thus
( 0 ) h C 0 0.
For the second term in the righthand side of (4.22), the functional calculus
of R2 yields the convergence in L2 (G)
0 h = eR2 0 0 0 .
As R2 commutes with the R2 multiplier eR2 and since 0 D(G), R2 0 = R0 ,
we know that 0 h = eR2 0 Dom(R2 ) and moreover
L2 (G)
(R0 ) h = (R2 0 ) h = eR2 R2 0 = R2 eR2 0 0 R2 0 .
More generally, for any m N, 0 h = eR2 0 Dom(Rm
2 ) and
L2 (G)
R2
0 0 Rm
Rm
2 e
2 0 .
By Lemma 4.2.16, this implies that 0 h 0 is continuous on G. Furthermore,
for any compact subset of G Rn and any m N with m >
n2 , we have
sup 0 h 0  C (0 h 0 2 + Rm (0 h 0 )2 ) 0 0.
Hence we have obtained that both terms on the righthand side of (4.22)
go to zero for the supremum norm on any compact subset of G. Therefore, the
expression in (4.21) tends to
(R + t )( h() )(x, t) 0 (, t ) h (x),
for t xed, locally in x. This is even stronger than the pointwise convergence in H
we wanted in (4.20) and concludes the proof of Lemma 4.2.15.
4.2. Positive Rockland operators
193
Corollary 4.2.17. Under the hypothesis of Lemma 4.2.15, h is smooth on (G
R)\{(0, 0)} and satises (4.15) and (4.18). Moreover, each function ht is Schwartz
on G and
ht (x)dx = 1.
G
Proof of Corollary 4.2.17. By Lemma 4.2.15, the distribution h annihilates the
hypoelliptic operator R+t on (GR)\{0}, and thus h is smooth on (GR)\{0}.
Since h(x, t) = 0 for t 0, this implies that h(x, t) vanish to innite order as t 0:
x G\{0}, N N0
> 0, C > 0 t (0, )
h(x, t) CtN .
We can choose = 1 since h is smooth on G (0, ). In fact this estimate remains
) h(x, t). It is also uniform in x when x runs over a
true for any xderivatives ( x
xed compact set which does not contain 0. Choosing this compact set to be the
unit sphere of a given quasinorm  , we have
h(x, t) CtN .
N N0 C > 0 t (0, 1] sup
x
x=1
) by the leftinvariant ones, see Section
We may replace the abelian derivatives ( x
3.1.5. This implies (4.15).
Using the homogeneity of h (see Property (4.12) which was already proven
and Proposition 3.1.23), we have
X h(x, t) = rQ[] X hr t (rx),
x G, r > 0
and so, in particular, if x 1 then we obtain, because of (4.15), that
X h1 (x) = xQ+[] X hx (x1 x) C,N xQ+[]N .
Since h1 is smooth on G, this shows that h1 is Schwartz. This is also the case for
ht by homogeneity, see (4.17). Note that the same homogeneity property together
with (4.15) implies (4.18).
Since each function ht satises the homogeneity property given in (4.17) and
is integrable, the functions ht form a commutative approximation of the identity,
see Remark 3.1.60. In particular,
with c =
ht t0 c
G
in L2 (G),
h1 (x)dx. Since we know
ht = etR2 0
in L2 (G),
this constant c must be equal to 1. By homogeneity,
ht (x)dx =
h1 (x)dx = c = 1.
t > 0
G
194
Chapter 4. Rockland operators and Sobolev spaces
Lemmata 4.2.14 and 4.2.15 imply Theorem 4.2.7 and Corollary 4.2.10 under
the assumption > Q/2. We now need to remove this assumption. For this, we will
use the following formula which is a consequence of the principle of subordination:
Lemma 4.2.18. For any > 0, we have
s
2
e
e 4s ds.
e =
s
0
(4.23)
Sketch of the proof of Lemma 4.2.18. We follow [Ste70a, p.61]. We start from the
well known identity
ix
e
e =
dx,
(4.24)
2
1 + x
which is an application of the Residue theorem to the function
z
eiz
.
z2 + 1
In (4.24) we replace 1 + x2 using
1
=
1 + x2
)u
du,
and we obtain the double integral
eix
e =
e(1+x
e(1+x
)u
du dx.
One can show that it is possible to invert the order of integration:
2
u
e
eix ex u dx du.
e =
0
It is well known that the inner integral in dx is equal to
2
e 4u
.
u
And this shows (4.23).
We can now nish the proofs of Theorem 4.2.7 and Corollary 4.2.10.
End of the proofs of Theorem 4.2.7 and Corollary 4.2.10. Since the case > Q/2
is already proven, we may assume Q/2.
m
For any m N0 , R2 is a positive Rockland operator (see Lemma 4.2.5),
with homogeneous degree 2m . We denote by Km the function on G R giving
2m
its heat kernel in the sense that if t > 0, Km (, t) S (G) is the kernel of etR2
4.2. Positive Rockland operators
195
and if t 0 then Km (x, t) = 0 for any x G. This is possible since, by Corollary
4.2.13, Km is smooth on G (0, ). By homogeneity, it will always satisfy
Q
Km (x, t) = t 2m Km (t 2m x, 1).
x G, t > 0
In (4.23), replacing by t2
et
2m1
(4.25)
m1
, one nds that
s
m
t 2 2
e
e 4s ds.
=
s
0
Using the functional calculus on R, that is, integrating against the spectral measure dE() of R2 , we obtain formally that for any nonnegative integer m N0
and t > 0,
s
2m1
2m
t2
e
e 4s R2 ds,
etR2
=
(4.26)
s
0
and for the kernels of these operators,
s
e
t2
Km (x, )ds.
(4.27)
Km1 (x, t) =
4s
s
0
It is not dicult to see that Formulae (4.26) and (4.27) hold as operators and continuous integrable functions respectively when, for instance, Km (, t) is integrable
on G for each t > 0 and
s
e
t2
Km (, )L1 (G) ds < .
4s
s
0
Indeed under this hypothesis, Km1 (, t) is integrable on G for any xed t > 0 and
s
e
t2
Km (, )L1 (G) ds < .
(4.28)
Km1 (, t)L1 (G)
4s
s
0
It is then a standard procedure to make sense of (4.26) by rst integrating over
[0, N ] and then letting N tend to innity.
We rst assume that 2m > Q/2, so that the conclusion of Theorem 4.2.7
holds for Km . In particular, Km (, 1) S(G) and by homogeneity, the L1 norm of
Km (, t) is
G
Km (x, t)dx =
Km (x, 1)dx,
is nite and independent of t. Therefore
s
s
e
e
t2
ds,
Km (x, )dxds =
Km (x, 1)dx
4s
s
s
G
G
0
0
is nite. Consequently Formula (4.27) holds and by (4.28),
s
e
ds < .
Km1 (t, )L1 (G)
Km (x, 1)dx
s
G
0
196
Chapter 4. Rockland operators and Sobolev spaces
By homogeneity, G Km1 (x, t)dx must also be independent of t > 0, while it is
identically zero if t 0. This implies that Km1 is locally integrable on G R
and that Km1 (, 1) L1 (G). By Lemmata 4.2.14 and 4.2.15, Km1 satisfy the
properties of the heat kernel described in Theorem 4.2.7 and Corollary 4.2.10.
Now we can repeat the same reasoning with m replaced successively by m
1, m 2, . . . , 2, 1. Since K0 = h, this concludes the proofs of Theorem 4.2.7 and
Corollary 4.2.10.
We still have to show Corollary 4.2.9.
Proof of Corollary 4.2.9. Since the heat kernels ht , t > 0, form a commutative
approximation of the identity (see Theorem 4.2.7 and Remark 3.1.60 in Section
3.1.10), the operators f f ht , t > 0, form a strongly continuous semigroup
on Lp (G) for any p [1, ) and on Co (G), see Lemma 3.1.58. It is naturally
equibounded by h1 since
f ht p f p ht 1
and
ht 1 = h1 .
Let us prove the convergence in (4.16) for p = . Let f D(G). By Lemma
4.2.16, for any compact subset G,
1
sup (f ht f ) Rf
t
1
1 m
m+1
C (f ht f ) Rf + R (f ht f ) R
f ,
t
t
2
2
where m is an integer such that m n2 . Since D(G) Dom(R) and
etR2 f = f ht ,
we have for any integer m N0 that
1 m
1
+1
R (f ht f ) Rm +1 f = Rm
etR2 f f Rm
f
2
2
t
t
1 tR2 m
1
m +1
=
R2 f Rm
f=
e
(Rm f ) ht Rm f Rm +1 f
2 f R2
t
t
t0 0 in L2 (G).
Therefore,
1
sup (f ht f ) Rf t0 0.
t
We x a quasinorm  . By Part 2 of Remark 3.2.16 and the existence of a
homogeneous norm (Theorem 3.1.39), without loss of generality, we may assume
  to be also a norm, that is, the triangular inequality is satised with constant 1;
although we could give a proof without this hypothesis, it simplies the constants
4.2. Positive Rockland operators
197
R be a closed ball about 0 of radius R which contains the support of
below. Let B
2R the closed ball about 0 and with radius 2R. If x , then
f . We choose = B
R ,
since f is supported in B
1
1
1
(f ht f ) Rf (x) = f ht (x) =
f (y)ht (y 1 x)dy,
t
t
t yR
hence
1
f
1
f ht (x) f
h
(y
x)dy
=
h1 (z)dz,
t
t
1
t
t
yR
xt z 1 R
as ht satises (4.17). Note that {z : xt z 1  R} {z : t z > R/2} since
1
t z R/2 = xt z 1  x t z 1 
1
Therefore
3
R.
2
1
xt z 1 R
h1 (z)dz
z>t R/2
h1 (z)dz.
Since h1 is Schwartz, we must have
C
z G\{0}
h1 (z) Cza ,
for a = Q + 2 for instance. This together with the polar change of variable (cf.
Proposition 3.1.42) yield
h
(z)dz
C
raQ1 dr = C t2 .
1
1
1
z>t R/2
r=t R/2
Consequently, denoting by c the complement of in G, we have
1
sup (f ht f ) Rf C t t0 0.
c
t
This shows the convergence in (4.16) for p = .
We proceed in a similar way to prove the convergence in (4.16) for p nite.
R . We decompose
As above we x f D(G) supported in B
1
(f ht f ) Rf p
t
1
1
c ).
(f ht f ) Rf Lp (B2R ) + (f ht f ) Rf Lp (B2R
t
t
For the rst term,
1
2R  p1 1 (f ht f ) Rf 0,
(f ht f ) Rf Lp (B2R ) B
t0
t
t
198
Chapter 4. Rockland operators and Sobolev spaces
as we have already proved the convergence in (4.16) for p = . For the second
term, we obtain for the reasons explained in the case p = that
1
1
c ) =
c )
f ht Lp (B2R
(f ht f ) Rf Lp (B2R
t
t
p " p1
!
1
1
=
f (y) ht (y x)dy dx
t
x>2R y<R
!
"p " p1
!
Q
1
C
f (y) t t (y 1 x)a dy dx
t
x>2R
y<R
" p1
!
Q
Ct1+p( + ) f L1
a
x>2R
(x R)ap dx
where we have used that the reverse triangle inequality
y 1 x x y x R.
Consequently we obtain the convergence in (4.16) for p nite if we choose a large
enough.
4.3
Fractional powers of positive Rockland operators
In this section we aim at dening fractional powers of positive Rockland operators.
We will carry out the construction on the scale of Lp spaces for 1 p ,
with L (G) substituted by the space Co (G) of continuous functions vanishing at
innity. The extension of a positive Rockland operator R to Lp (G) will be denoted
by Rp , and rst we discuss the essential properties of such an extension. Then we
dene its complex powers. Before studying the corresponding Riesz and Bessel
potentials, we will show that imaginary powers are continuous operators on Lp ,
p (1, ).
4.3.1 Positive Rockland operators on Lp
We start by dening the analogue Rp of the operator R on Lp (G).
Denition 4.3.1. Let R be a positive Rockland operator on a graded Lie group G.
For p [1, ), we denote by Rp the operator such that Rp is the innitesimal generator of the semigroup of operators f f ht , t > 0, on Lp (G).
We also denote by Ro the operator such that Ro is the innitesimal
generator of the semigroup of operators f f ht , t > 0, on Co (G).
For the moment it seems that R2 denotes the selfadjoint extension of R on
L2 (G) and minus the generator of f f ht , t > 0, on L2 (G). In the sequel, in
4.3. Fractional powers of positive Rockland operators
199
fact in Theorem 4.3.3 below, we show that the two operators coincide and there
is no conict of notation.
The case p = is somewhat irrelevant and will be often replaced by p = o ,
especially when using duality. The next lemma aims at clarifying this point.
Lemma 4.3.2.
If p (1, ), any bounded linear functional on Lp (G) can be
realised by integration against a function in Lp (G), where p is the conjugate
exponent of p, that is, p1 + p1 = 1. Consequently, the dual Lp (G) of Lp (G)
may be identied with Lp (G) and the corresponding norms coincide.
If p = 1, any bounded linear functional on L1 (G) can be realised by integration against a bounded function on G. Consequently, the dual L1 (G) of
L1 (G) may be identied with L (G) and the corresponding norms coincide.
In particular, L1 (G) contains Co (G).
If p = o , any bounded linear functional on Co (G) can be realised by integration against a regular complex measure. Consequently, the dual Co (G)
of Co (G) may be identied with the Banach space M (G) of regular complex
measures endowed with the total mass M (G) as its norm, and the corresponding norms coincide. With this identication, Co (G) contains L1 (G)
and the corresponding norms coincide.
Proof. See, e.g., Rudin [Rud87, ch.6].
We can now describe the properties of Rp .
Theorem 4.3.3. Let R be a positive Rockland operator on a graded Lie group G.
In this statement, p [1, ) {o }.
(i) The semigroup {f f ht }t>0 is strongly continuous and equicontinuous
on Lp (G) if p [1, ) or on Co (G) if p = o :
t > 0, f Lp (G) or Co (G)
f ht p h1 1 f p .
Consequently, the operator Rp is closed. The domain of Rp contains D(G),
and for f D(G) we have Rp f = Rf .
p is the innitesimal generator of the strongly continuous
(ii) The operator R
t }t>0 on Lp (G).
semigroup {f f h
(iii) We use the identications of Lemma 4.3.2. If p (1, ) then the dual of Rp
p . The dual of R restricted to L1 (G) is R
1 . The dual of R1 restricted
is R
o
to Co (G) L (G) is Ro .
(iv) If p [1, ), the operator Rp is the maximal restriction of R to Lp (G), that
is, the domain of Rp consists of all the functions f Lp (G) such that the
distributional derivative Rf is in Lp (G) and Rp f = Rf .
The operator Ro is the maximal restriction of R to Co (G), that is,
the domain of Ro consists of all the functions f Co (G) such that the
distributional derivative Rf is in Co (G) and Rp f = Rf .
200
Chapter 4. Rockland operators and Sobolev spaces
(v) If p [1, ), the operator Rp is the smallest closed extension of RD(G) on
Lp (G). For p = 2, R2 is the selfadjoint extension of R on L2 (G).
Proof. Part (i) is a consequence of Corollary 4.2.9, see also Section A.2.
Part (i) implies, intertwining with the complex conjugate, that {f f
t }t>0 is also a strongly continuous semigroup on Lp (G). On D(G), its innitesh
= Rt which is a positive Rockland operator (see
imal operator coincide with R
Lemma 4.2.5) and it is easy to see that
t.
etR2 = etR2 = ht = h
D(G), t > 0
This shows Part (ii).
For Part (iii), we observe that using (1.14) and (4.13), we have
f1 , f2 D(G)
t .
f1 ht , f2 = f1 , f2 h
(4.29)
Thus we have for any f, g D(G) and p [1, ) {o }
1
1
1
t g = 1 f, etR p g g.
(etRp f f ), g = f ht f, g = f, g h
t
t
t
t
Here the brackets refer to the duality in the sense of distributions or, equivalently,
to the duality explained in Lemma 4.3.2. Taking the limit as t 0 of the rst and
last expressions proves Part (iii).
We now prove Part (iv) for any p [1, ) {o }. Let f Dom(Rp ) and
and by Part
D(G). Since R is formally selfadjoint, we know that Rt = R,
(i), we have Rq = R for any q [1, ) {o }. Thus by Part (iii) we have
p = f, Rt = Rf, ,
Rp f, = f, R
and Rp f = Rf in the sense of distributions. Thus
Dom(Rp ) {f Lp (G) : Rf Lp (G)}.
We now prove the reverse inclusion. Let f Lp (G) such that Rf Lp (G).
Let also D(G). The following computations are justied by the properties of
R and ht (see Theorem 4.2.7), Fubinis Theorem, and (4.29):
t
s )ds
s ( h
f ht f, = f, ht = f,
0
t
t
s )ds
R( hs )ds = f, R
( h
= f,
0
0
t
t
s ds =
s ds
h
Rf, h
= Rf,
0
0
t
t
(Rf ) hs , ds = (Rf ) hs ds, .
=
0
4.3. Fractional powers of positive Rockland operators
Therefore,
f ht f =
201
(Rf ) hs ds.
Let us recall the following general property:
t if t xt is a continuous mapping
from [0, ) to a Banach space X , then 1t 0 xs ds converges to x0 in the strong
topology of X as t 0. We apply this property to X = Lp (G) and t (Rf ) ht ;
the hypotheses are indeed satised because of the properties of the heat kernel,
see Theorem 4.2.7. Hence we have the following convergence in Lp (G):
1
1
(f ht f ) =
t
t
(Rf ) hs ds Rf.
t0
This shows f Dom(Rp ) and concludes the proof of (iv).
Part (v) follows from (iv). This also shows that the selfadjoint extension
of R coincides with R2 as dened in Denition 4.3.1 and concludes the proof of
Theorem 4.3.3.
Theorem 4.3.3 has the following couple of corollaries which will enable us to
dene the fractional powers of Rp .
Corollary 4.3.4. We keep the same setting and notation as in Theorem 4.3.3.
(i) The operator Rp is injective on Lp (G) for p [1, ) and Ro is injective
on Co (G), namely,
for p [1, ) {o } :
f Dom(Rp )
Rp f = 0 = f = 0.
(ii) If p (1, ) then the operator Rp has dense range in Lp (G). The operator
Ro has dense range in Co (G). The closure of the range of R1 is the closed
subspace { L1 (G) : G = 0} of L1 (G).
Proof. Let f Dom(Rp ) be such that Rp f = 0 for p [1, ) {o }. By
Theorem 4.3.3 (iv), f S (G) and Rf = 0. In Remark 4.1.13 (3), we noticed that
any positive Rockland operator satises the hypotheses of Liouvilles Theorem for
homogeneous Lie groups, that is, Theorem 3.2.45. Consequently f is a polynomial.
Since f is also in Lp (G) for p [1, ) or in Co (G) for p = o , f must be identically
zero. This proves (i).
For (ii), let be a bounded linear functional on Lp (G) if p [1, ) or on
Co (G) if p = o such that vanishes identically on Range(Rp ). Then can be
realised as the integration against a function f Lp (G) if p [1, ) or a measure
also denoted by f M (G) if p = o , see Lemma 4.3.2. Using the distributional
notation, we have
() = f,
Lp (G)
or
Co (G).
202
Chapter 4. Rockland operators and Sobolev spaces
Then for any D(G), we know that Dom(Rp ) and Rp = R by Theorem 4.3.3 (i) thus
,
0 = (Rp ()) = f, R() = Rf,
This shows that Rf
= 0. Applying again Liouvilles Theorem, this
since Rt = R.
(see Lemma 4.2.5), this shows that f is
time to the positive Rockland operator R
a polynomial. For p (1, ), f being also a function in Lp (G), this implies that
f 0. For p = o , f M (G), this shows that f is an integrable polynomial on G
hence f 0. For p = 1, f being a measurable bounded function and a polynomial,
f must be constant, i.e. f c for some c C. This shows that if p (1, ){o }
p
then = 0 and Range(R
p ) is dense in L (G) or Co (G), whereas if p = 1 then
1
: L (G) c G . This shows (ii) for p (1, ) {o }.
Let us study more precisely the case p = 1. It is easy to see that
X(x)dx =
(x) (X1)(x)dx = 0
G
for any
holds for any L1 (G) such that X L1 (G). Consequently,
Dom(R1 ), we know that and R are in L1 (G) thus G R1 = 0. So the range
of R1 is included in
0
1
S := L (G) :
= 0 Range(R1 ).
G
Moreover, if 1 a bounded linear functional on S such that 1 is identically 0
on Range(R1 ), by the HahnBanach Theorem (see, e.g. [Rud87, Theorem 5.16]),
it can be extended into a bounded linear function on L1 (G). As vanishes
identically on Range(R1 ) S, we have already proven that must be of the form
: L1 (G) c
G
for some constant c C and its restriction to S is 1 0. This concludes the
proof of Part (ii).
Eventually, let us prove that the operator Rp is Komatsunonnegative, see
hypothesis (iii) in Section A.3:
Corollary 4.3.5. For p [1, ) {o }, and any > 0, the operator I + Rp is
invertible on Lp (G), p [1, ), and Co (G) for p = o , and the operator norm of
(I + Rp )1 is
(I + Rp )1 h1 1 .
Proof. Integrating the formula
( + )1 =
et(+) dt,
4.3. Fractional powers of positive Rockland operators
against the spectral measure dE() of R2 , we have formally
1
et(I+R2 ) dt,
(I + R2 ) =
203
(4.30)
and the convolution kernel of the operator on the righthand side is (still formally)
given by
(x) :=
et ht (x)dt.
From the properties of the heat kernel ht (see Theorem 4.2.7 and Corollary
4.2.10), we see that the function dened just above is continuous on G and
that
h1
t
< .
e ht 1 dt = h1
et dt =
1
0
0
As L1 (G), it is a routine exercise to show that the operator
et(I+R2 ) dt
is bounded on L2 (G) with convolution kernel (it suces to consider integration
over [0, N ] with N ). Moreover, Formula (4.30) holds in L (L2 (G)).
For any D(G) and p [1, ) {o }, Theorem 4.3.3 (iv) implies
(I + Rp ) = (I + R) = (I + R2 ) D(G),
thus
((I + Rp )) = ((I + R2 )) = .
This yields that the operator (I + Rp )1 : is bounded on Lp (G) if
p [1, ) and on Co (G) if p = o . Furthermore, its operator norm is
(I + Rp )1 1 h1 1 ,
completing the proof.
4.3.2 Fractional powers of operators Rp
We now apply the general theory of fractional powers outlined in Section A.3 to
the operators Rp and I + Rp .
Theorem 4.3.6. Let R be a positive Rockland operator on a graded Lie group G.
We consider the operators Rp dened in Denition 4.3.1. Let p [1, ) {o }.
1. Let A denote either R or I + R.
204
Chapter 4. Rockland operators and Sobolev spaces
(a) For every a C, the operator Aap is closed and injective with (Aap )1 =
0
N
Aa
p . We have Ap = I, and for any N N, Ap coincides with the usual
powers of dierential operators on S(G) and Dom(AN ) Range(AN )
is dense in Range(Ap ).
(b) For any a, b C, in the sense of operator graph, we have Aap Abp Aa+b
p .
a b
a+b
If Range(Ap ) is dense then the closure of Ap Ap is Ap .
(c) Let ao C+ .
If Range(Aapo ) then Dom(Aap ) for all a C with 0 <
Re a < Re ao and the function a Aap is holomorphic in {a
C : Re ao < Re a < 0}.
If Dom(Aapo ) then Dom(Aap ) for all a C with 0 < Re a <
Re ao and the function a Aap is holomorphic in {a C : 0 <
Re a < Re ao }.
(d) For every a C, the operator Aap is invariant under left translations.
(e) If p (1, ) then the dual of Ap is Ap . The dual of Ao restricted to
L1 (G) is A1 . The dual of A1 restricted to Co (G) L (G) is Ao .
(f ) If a, b C+ with Re b > Re a, then
C = Ca,b > 0
Dom(Abp )
Aap C1 Re b Abp Re b .
Re a
Re a
(g) For any a C+ , Dom(Aap ) contains S(G).
(h) If f Dom(Aap )Lq (G) for some q [1, ){o }, then f Dom(Aaq )
if and only if Aap f Lq (G), in which case Aap f = Aaq f .
2. For each a C+ , the operators (I + Rp )a and Rap are unbounded and their
domains satisfy for all > 0,
Dom [(I + Rp )a ] = Dom(Rap ) = Dom [(Rp + I)a ] .
3. If 0 < Re a < 1 and Range(Rp ) then
1
a
ta1 etRp dt,
Rp =
(a) 0
N
in the sense that limN 0 converges in the norm of Lp (G) or Co (G).
4. If a C+ , then the operator (I + Rp )a is bounded and for any X with
X = Lp (G) or Co (G), we have
1
(I + Rp )a =
ta1 et(I+Rp ) dt,
(a) 0
in the sense of absolute convergence:
ta1 et(I+Rp ) X dt < .
0
4.3. Fractional powers of positive Rockland operators
205
5. For any a, b C, the two (possibly unbounded) operators Rap and (I + Rp )b
commute.
6. For any a C, the operator Rap is homogeneous of degree a.
Recall (see Denition A.3.2) that the two (possibly unbounded) operators A
and B commute when
x Dom(AB) Dom(BA) = ABx = BAx,
and that the domain of the product AB of two (possibly unbounded) operators A
and B on the same Banach space X is formed by the elements x X such that
x Dom(B) and Bx Dom(A).
Proof. The operator Rp is closed and densely dened by Theorem 4.3.3 (i), it is
injective by Corollary 4.3.4 and Komatsunonnegative in the sense of Section A.3
(iii) by Corollary 4.3.5. Therefore, Rp satises the hypotheses of Theorem A.3.4.
Moreover, I + Rp also satises these hypotheses by Remark A.3.3, and (I + Rp )
generates an exponentially stable semigroup:
et(I+Rp ) et etRp h1 1 et .
Most of the statements then follow from the general properties of fractional
powers constructed via the Balakrishnan formulae recalled in Section A.3. More
precisely, from the Balakrishnan formula, for any N N, AN
p coincides with the
usual powers of dierential operators on S(G) and Part (1a) follows from Theorem
A.3.4 (1) and (2) and Remark A.3.1.
The duality properties explained in Part (1e) for p (1, ) hold for the
Balakrishnan operators hence they hold for their maximal closure. The cases of
p = 1, o are similar and this proves Part (1e). The properties in Parts (1d), (5)
and (6) hold for the Balakrishnan operators hence they hold for their maximal
closure and these parts are proved.
Part (1b) follows from Theorem A.3.4 (4).
Part (1c) follows from Theorem A.3.4 (5).
Part (1f) follows from Theorem A.3.4 (6).
Part (1g) follows from Parts (1a) and (1c).
Part (1h) is certainly true for any f S(G) and Re a > 0 via the Balakrishnan formulae. By analyticity (see Part (1c)) it is true for any a C. The density
of D(G) in Lp (G) (or Co (G) if p = o ) together with the maximality of Aap and
the uniqueness of distributional convergence imply the result.
Part (2) follow from Theorem A.3.4 (8).
Parts (3) and (4) follows from Theorem A.3.4 (10).
This concludes the proof of Theorem 4.3.6.
206
Chapter 4. Rockland operators and Sobolev spaces
4.3.3 Imaginary powers of Rp and I + Rp
In this section, we show that imaginary powers of a positive Rockland operator
R as well as I + R are bounded operators on Lp (G), p (1, ). We prove this
as a consequence of the theorem of singular integrals on homogeneous groups, see
Section 3.2.3.
We start by showing that if R is a positive Rockland operator, then the
imaginary powers of I + Rp are bounded on Lp (G):
Proposition 4.3.7. Let R be a positive Rockland operator on a graded Lie group G.
For any R and p (1, ), the operator (I + Rp )i is bounded on Lp (G). For
any p (1, ), there exists C = Cp,R > 0 and > 0 such that
(I + Rp )i L (Lp (G)) Ce  .
For any p (1, ) and a C, Dom((I + Rp )a ) = Dom((I + Rp )Re a ).
The following technical result will be useful in the proof of Proposition 4.3.7
and in other proofs (see Sections 4.3.4 and 4.4.4).
Lemma 4.3.8. Let R be a positive Rockland operator on a graded Lie group G. Let
ht be its heat kernel as in Section 4.2.2.
1. For any homogeneous quasinorm  , any multiindex Nn0 , and any real
, there exists a constant C > 0 such that
number a with 0 < a < Q+[]
ta1 X ht (x)dt CxQ[]+a .
For any homogeneous quasinorm  , any multiindex Nn0 , there
exists a constant C > 0 such that
X ht (x)et dt CxQ[] .
0
2. For any homogeneous quasinorm , any multiindex Nn0 , and any t > 0,
we have
x1/2
X ht (x)dx t
[]
X h1 L1 .
3. For any homogeneous quasinorm  , any multiindex Nn0 , any N N
and any t (0, 1), there exists a constant C > 0 such that
X ht (x)dx CtN .
x1/2
4.3. Fractional powers of positive Rockland operators
207
Proof of Lemma 4.3.8 . Let us prove Part 1. We write
x
ta1 X ht (x)dt =
+
0
x
For the second integral, we use the property of homogeneity of ht (see (4.12) or
(4.17))
Q+[]
1
=
ta1 X h1 (t x)dt
x
x
Q+[]
Q + []
a)1 X h1 x(a ) .
As h1 S(G), X h1 is nite. For the rst integral, we use again (4.12) to
obtain
x
x
x
dt
=
ta1 x(Q+[]) X hx t
x
0
0
C1 a1 x(a
Q+[]
)
where C1 := supy=1,0t1 1 X ht1 (y) is nite by (4.15). Combining the two
estimates above shows the estimates for the rst integral in Part 1. We proceed in
the same way for the second one:
x
t
X ht (x)e dt =
+
.
0
We have (with C1 as above)
x
Q+[]
(a )
C1 x
0
x
et dt = C1 x(a
Q+[]
)
(1 ex )
whereas
x
x
C1 x
(a
Q+[]
)
X h1 (x )
X h1 x
Q+[]
(Q+[])
x
et dt = X h1 x(Q+[]) ex
We conclude in the same way as above and Part 1 is proved.
Let us prove Part 2. The property of homogeneity of ht (see (4.17)) together
with h1 S(G) imply
[]+Q
1
X ht (x)dx =
X h1 (t x)t dx
x1/2
x1/2
[]
[]
= t
X h1 (x )dx t
X h1 ,
1
t x 1/2
208
Chapter 4. Rockland operators and Sobolev spaces
having used the change of variable x = t x. This shows Part 2.
1
Let us prove Part 3. The properties of the heat kernel, especially (4.12) and
(4.15), imply
X ht (x) = x[]Q X hx t (x1 x) Cx[]Q (x t)N ,
N
if x 1/2 and t (0, 1) where C = supx =1,0<t <1 t X ht (x ) is nite.
Hence
N
X ht (x)dx Ct
x[]QN dx.
x1/2
x1/2
This shows Part 3 and concludes the proof of Lemma 4.3.8.
Proof of Proposition 4.3.7. By Theorem 4.3.6 (1), to show that (I + Rp )i is
bounded on Lp (G) for some p (1, ) and R, it suces to show that (I+R2 )i
can be extended to an Lp bounded operator. To do this, we will show that Corollary 3.2.21 can be applied to (I + R2 )i .
By functional calculus, (I + R2 )i is bounded on L2 (G). Part 1 of Lemma
4.3.8 together with the formula
> 0
i =
ti et dt,
(1 i ) 0
and the functional calculus of R2 imply that the right convolution kernel of (I +
R2 )i is the tempered distribution which coincides with the smooth function
away from 0 given via
1
(x) =
ti (I + R)ht (x)et dt,
x = 0.
(4.31)
(1 i ) 0
Using this formula, we have
1
(x)dx (1 i )
x1/2
t=0
x1/2
(ht (x) + Rht (x))et dxdt.
By Part 2 of Lemma 4.3.8, (and h1 being Schwartz), the integrals
ht (x)et dxdt and
Rht (x)et dxdt,
t=0
t=1
x1/2
x1/2
are nite. By Part 3 of Lemma 4.3.8, the integral
1
t=0
x1/2
is nite. This shows that
Rht (x)e dxdt C
x1/2
(x)dx is nite.
1
t=0
t0 dt = C,
4.3. Fractional powers of positive Rockland operators
209
Using (4.31), we also obtain easily that
sup xQ+[] X (x) (1 i )1 sup xQ+[]
0<x<1
0<x<1
X ht (x) + X Rht (x)dt,
and the righthand side is nite by Lemma 4.3.8. Note that if we denote by =
,R the kernel of (I + R2 )i , then we have
,R (x1 ) = ,R (x),
using the formula in (4.31) and
t (x)
((I + R)ht ) (x1 ) = ((I t )ht ) (x1 ) = (I t )h
= ((I t )ht ) (x) = ((I + R)ht ) (x),
where we have used (4.13). Hence we also have that each quantity
(x) =
sup xQ+[] X
0<x<1
sup xQ+[] X ,R (x)
0<x<1
is nite.
The estimates above show that satises the hypotheses of Corollary 3.2.21
and therefore the operator (I + R2 )i is bounded on Lp (G), p (1, ). The properties of the semigroup (see Theorem A.3.4 (3)) imply the rest of the statement
in Proposition 4.3.7.
Let us now prove the homogeneous case, that is, that the imaginary powers
of a positive Rockland operator are bounded on Lp (G):
Proposition 4.3.9. Let R be a positive Rockland operator on a graded Lie group
p
G. For any R and p (1, ), the operator Ri
p is bounded on L (G). For any
p (1, ), there exists C = Cp,R > 0 and > 0 such that
R
 
Ri
.
p L (Lp (G)) Ce
a
For any p (1, ) and a C, Dom(Rap ) = Dom(RRe
).
p
Proof of Proposition 4.3.9. Let p (1, ) and R. Let us denote by Rp,i the
(possibly unbounded) operator given as the strong limit in Lp (G) of ( + Rp )i
as 0, for Dom(( + Rp )i ) for any (0, 0 ) for some small 0 > 0 and
such that this strong limit exists. The domain of Rp,i is naturally the space of
all those functions . Note that the homogeneity of R implies
( + Rp )i = i (I + 1 Rp )i = i (I + Rp )i {(1/ )}(1/ ),
for any > 0 and any Lp (G) such that
(1/ ) Dom((I + Rp )i ).
210
Chapter 4. Rockland operators and Sobolev spaces
By Proposition 4.3.7, Dom((I + Rp )i ) = Lp (G) and the operator (I + Rp )i is
bounded. Therefore for all Lp (G) and > 0, is in Dom(( + Rp )i ) and we
have
( + Rp )i Lp (G)
(I + Rp )i {(1/ )}(1/ )Lp (G)
p (I + Rp )i {(1/ )}Lp (G)
p (I + Rp )i L (Lp (G)) (1/ )Lp (G)
= (I + Rp )i L (Lp (G)) Lp (G) .
Consequently, Rp,i extends to a bounded operator on Lp (G). By Theorem
p
A.3.4 (9), this implies that Ri
p is also a bounded operator on L (G) as Rp has
dense range and domain by Corollary 4.3.4. As in the inhomogeneous case, the
properties of the semigroup (see Theorem A.3.4 (3)) imply the rest of the statement in Proposition 4.3.9.
Given the proof of Proposition 4.3.7, one would be tempted to study the
p
convolution kernel of the operator Ri
2 in order to show the L boundedness in
the proof of Proposition 4.3.9. Indeed, following the same arguments as in the
proof of Proposition 4.3.7, one shows that the kernel of Ri
2 coincides away from
the origin with the smooth function
1
ti Rht (x)dt.
G\{0} x
(1 i ) 0
However, this function can not be in general a kernel of type i : already for the
usual Laplacian on (Rn , +) it is not the case. Indeed, in the Euclidean case, this
function is radial and nonzero and its average on the sphere can therefore not
vanish.
In the stratied case, Folland proved the Lp boundedness of imaginary powers
of the subLaplacian L and I + (L) using general properties of semigroups preserving positivity together with the Laplace transform see [Fol75, Proposition 3.14
and Lemma 3.13]. More precisely, the boundedness follows from the LittlewoodPaley theory and the study of square functions associated with the semigroup.
Note that in the case of a subLaplacian, the proof in [Fol75] yields a bound of
the operator norm by (1 i )1 up to a constant of p.
In our case, we applied a consequence of the theorem of Singular Integrals via
Corollary 3.2.20 to obtain the Lp boundedness of the imaginary powers of I + R
and we have shown
i
Ri
p L (Lp (G)) (I + Rp ) L (Lp (G)) ,
p (1, ),
in the proof of Proposition 4.3.9. We can follow the constants in the proof of the
theorem of Singular Integrals (see Remark A.4.5 (2)) as well as in our application
to show that (I + Rp )i L (Lp (G)) is bounded up to a constant of p, by
(1 + (1 i )1 )2 p 2  .
1
4.3. Fractional powers of positive Rockland operators
211
However, we do not need these precise bounds as the bounds obtained from
the general theory of semigroups as stated in Propositions 4.3.7 and 4.3.9 will
be sucient for our purpose in the proofs of interpolation properties for Sobolev
spaces in Theorem 4.4.9 and Proposition 4.4.15.
4.3.4 Riesz and Bessel potentials
We mimic the usual terminology in the Euclidean setting, to dene the Riesz and
Bessel potentials associated with a positive Rockland operator.
Denition 4.3.10. Let R be a positive Rockland operator of homogeneous degree
. We call the operators Ra/ for {a C, 0 < Re a < Q} and (I + R)a/ for
a C+ , the Riesz potential and the Bessel potential, respectively.
In the sequel we will denote their kernels by Ia and Ba , respectively, as
dened in the following:
Corollary 4.3.11. We keep the setting and notation of Theorem 4.3.3.
(i) Let a C with 0 < Re a < Q. The integral
a
1
t 1 ht (x)dt
Ia (x) :=
a
( ) 0
converges absolutely for every x = 0. This denes a distribution Ia which is
smooth away from the origin and (a Q)homogeneous.
For any p (1, ), if S(G) or, more generally, if Lq (G)Lp (G)
where q [1, ) is given by 1q p1 = ReQa , then
a
Dom(Rp )
Rp = Ia Lp (G).
and
Consequently,
S(G)
Rp Lp (G)
and
= (Rp ) Ia .
(ii) Let a C+ . The integral
Ba (x) :=
1
( a )
t 1 et ht (x)dt
a
converges absolutely for every x = 0 and denes an integrable function Ba on
G. The function Ba is always smooth away from 0.
If Re a > Q, Ba is also smooth at 0.
If Re a > Q/2, then Ba is square integrable: Ba L2 (G).
212
Chapter 4. Rockland operators and Sobolev spaces
All the operators (I + Rp )a/ , p [1, ) {o }, are bounded convolution operators with the same (right convolution) kernel Ba .
If a, b C+ , then as integrable functions, we have
Ba Bb = Ba+b .
Remark 4.3.12. In other words for Part (i), Ia is a kernel of type a and
Ra/
0 = I a .
p
This shows that if < Q, I1 is a fundamental solution of R, in fact, the unique
homogeneous fundamental solution (cf. Theorem 3.2.40).
Note that we will show in Lemma 4.5.9 that more generally X Ba L2 (G)
whenever Re a > [] + Q/2, as well as other L1 estimates.
Proof of Corollary 4.3.11. The absolute convergence and the smoothness of Ia
and Ba follow from Lemma 4.3.8.
For the homogeneity of Ia , we use (4.12) and the change of variable s = r t,
to get
a
1
t 1 ht (rx)dt
Ia (rx) =
(a/) 0
a
1
(r s) 1 rQ hs (x)r ds = raQ Ia (x).
=
(a/) 0
Hence Ia is a kernel of type a with 0 < Re a < Q (see Denition 3.2.9).
By Lemma 3.2.7, the operator S(G) Ia is homogeneous of degree
a, and by Proposition 3.2.8, it admits a bounded extension Lq (G) Lp (G)
when p1 1q = ReQ(a) .
a
Let RQ (S(G)). By Theorem 4.3.6, the function a Rp is analytic
on the strip {z C, 0 < Re z < Q} and coincides there with
a
1
t 1 ht dt.
a
( a ) 0
But since the integral dening Ia (x) is absolutely convergent for all x G\{0},
we have
a
1
a C, Re a (0, Q),
t 1 ht dt = Ia ,
( a ) 0
and a Ia is analytic on the strip {0 < Re a < Q}.
Hence we have obtained that
a
Rp = Ia
4.3. Fractional powers of positive Rockland operators
213
holds for Re a (0, Q) and for any RQ (S(G)). Note that RQ (S(G)) is dense
in any Lr (G), r (1, ) as it suces to apply Corollary 4.3.4 (ii) to the positive
Rockland operator RQ . Then Corollary 3.2.32 concludes the proof of Part (i).
By Theorem 4.2.7,
G
for all t > 0, so
Ba (x)dx
G
1
( a )
ht  =
Re a
1
et
h1  <
ht (x)dx dt =
( Re a )
h1 L1 ,
( a )
(4.32)
and Ba is integrable.
By Theorem 4.3.6 Part (4), the integrable function Ba is the convolution
kernel of (I + Rp )a/ .
Let us show the square integrability of Ba . We compute for any R > 0:
(a/)2
Ba (x)2 dx =
(a/)Ba (x)(a/)Ba (x)dx
x<R
=
x<R 0
=
0
a
1
a
1
x<R
et ht (x)dt
e(t+s)
s (x)ds dx
s 1 es h
x<R
s (x)dx dtds.
ht (x)h
From the properties of the heat kernel (see (4.13) and (4.11)) we see that
s (x)dx =
ht (x)h
ht (x)hs (x1 )dx ht hs (0),
x<R
x<R
and ht hs (0) = ht+s (0) = (t + s) h1 (0).
Therefore,
Ba (x)2 dx =
G
h1 (0)
=
(a/)2
h1 (0)
(a/)2
1
s =0
a
1
0
(1 s )
s 1 t 1 e(t+s) (t + s) dtds
ds
u=0
eu u2(
Q
Re a
1) +1
du,
(4.33)
after the change of variables u = s + t and s = s/u. The integrals over s and u
converge when Re a > Q/2. Thus Ba is square integrable under this condition.
The rest of the proof of Corollary 4.3.11 follows easily from the properties of
the fractional powers of I + R.
The proof of Corollary 4.3.11 implies:
Corollary 4.3.13. We keep the notation of Corollary 4.3.11 and h1 denotes the
heat kernel at time t = 1 of R.
214
Chapter 4. Rockland operators and Sobolev spaces
1. For any a C+ , the operator norm of (I + Rp ) on Lp (G) if p [1, ) or
on Co (G) if p = o is bounded by Ba 1 and we have
a
Ba L1 (G)
2. If Re a > Q/2,
!
Ba L2 (G) =
( Re a )
h1 L1 (G) .
 a 
( 2ReaQ )
h1 (0)
a
( 2Re
)
"1/2
.
3. If p (1, 2) and a > Q(1 p1 ) then Ba Lp (G).
Proof. The rst statement follows from (4.32).
For the second part, Estimate (4.33) yields
Ba 22 = h1 (0)Ca ,
where
Ca
=
=
(a/)2
s
(1 s ) 1 ds
a
s =0
a 2 ( a )( a ) 2Re a

( a + a )
Q
u=0
eu u2
Q
Re a
1
du
thanks to the properties of the Gamma function (see equality (A.4)). We notice
that
a a
a 2
a a
=
= 
 .
Thus the constant Ca simplies into
Ca =
( 2ReaQ )
.
( a + a )
This shows the second part.
The third part is obtained by complex interpolation between Parts 1 and 2.
More precisely, we x a > 0 and b > Q/2 and we consider the linear functional
dened on simple functions in L1 (G) via
Tz =
Baz+b(1z) (x)(x)
G
for any z C, Re z [0, 1]. We have
Tz  Baz+b(1z) 1 .
4.3. Fractional powers of positive Rockland operators
215
Before applying Part 1 to Baz+b(1z) 1 , let us mention that the Stirling
formula (A.3) implies that for any w C+ ,
(Re w)
w ( Ree w )Re w
(w)
Re w ( we )w 
Re w 12
Re w
wwRe w 
w
Re w 12
Re w
exp Im w ln w .
w
This together with Part 1 then yield
ln Tz  ln(Baz+b(1z) 1 ) (1 + Im z) ln(1 + Im z),
thus {Tz } is an admissible family of operator (in the sense of Section A.6). The
same arguments also show that
T1+iy  (1 + y) + 2 exp (cy ln(1 + y)) ,
a
where c is a constant of a, b, .
The CauchySchwartz estimate and Part 2 yield
Tiy  Baiy+b(1iy) 2 2 ,
and Part 2 implies that the quantity
!
Baiy+b(1iy) 2 =
h1 (0)
( 2bQ
)
( 2b
)
"1/2
,
is independent of y. Hence we can apply Theorem A.6.1 to {Tz }: Tt extends to an
qt
Lqt bounded operator where t (0, 1) and q1t = 1t
2 . Therefore Bat+b(1t) L
where qt is the dual exponent to qt , i.e. q1t + q1 = 1. This shows Part 3 and
t
concludes the proof of Corollary 4.3.13.
We nish this section with some technical properties which will be useful in
the sequel. The rst one is easy to check.
Lemma 4.3.14. If R is a positive Rockland operator with Ba being the kernel of the
is also a positive Rockland
Bessel potential as given in Corollary 4.3.11, then R
operator and Ba is the kernel of the Bessel potential associated to R.
Lemma 4.3.15. We keep the notation of Corollary 4.3.11. If a C+ , then the
function
x xN Ba (x)
216
Chapter 4. Rockland operators and Sobolev spaces
is integrable on G, where   denotes any homogeneous quasinorm on G and N
is any positive integer. Consequently, for any S(G), the function Ba is
Schwartz and
Ba
acts continuously from S(G) to itself.
Note that we will show in Lemma 4.5.9 that, more generally,
xb X Ba L1 (G)
and that
X Ba L2 (G)
for Re a + b > [],
for Re a > [] + Q/2.
Proof of Lemma 4.3.15. Let   be a homogeneous quasinorm on G and N N.
We see that
Re a
1
N
1 t
x Ba (x)dx
t
e
xN ht (x)dx dt,
( a ) 0
G
G
and using the homogeneity of the heat kernel (see (4.17)) and the change of vari1
ables y = t x, we get
1
N
xN ht (x)dx =
t yN h1 (y)dy = cN t ,
G
where cN = yN h1 (y)L1 (dy) is a nite constant since h1 S(G). Thus,
G
xN Ba (x)dx
cN
( a )
Re a
N
1+
et dt < ,
and x xN Ba (x) is integrable.
Let Co 1 denote the constant in the triangle inequality for   (see Proposition 3.1.38 and also Inequality (3.43)). Let also S(G). We have for any N N
and Nn0 :
(1 + x)N X
[ Ba ] (x) = (1 + x)N X
Ba (x)
(1 + x)N X
Ba  (x)
(1 +  )N Ba (x) (x)
CoN (1 +  )N X
CoN (1 +  )N X
(1 +  )N Ba L1 (G) .
This shows that that Ba S(G) and that Ba is continuous as a map
of S(G) to itself (for a description of the Schwartz class, see Section 3.1.9).
4.3. Fractional powers of positive Rockland operators
217
Corollary 4.3.16. We keep the notation of Corollary 4.3.11.
For any a C and p [1, ) {o }, Dom(I + Rp )a S(G) and, moreover,
(I + Rp )a (S(G)) = S(G).
(4.34)
Furthermore on S(G), (I + Rp )a does not depend on p [1, ) {o } and acts
continuously on S(G).
If a C+ , we have
(I + Rp )a ( Ba ) = ((I + Rp )a ) Ba =
(p [1, ) {o }).
(4.35)
Proof. Formula (4.35) holds for each p [1, ) {o } by Theorem 4.3.6 and
Corollary 4.3.11.
Let us show (4.34) in the case of a = N N. By Theorem 4.3.6 (1a), we have
the equality (I + Rp )N = (I + R)N for any S(G) and p (1, ). Hence
(I + Rp )N (S(G)) = (I + R)N (S(G)). The inclusion (I + R)N (S(G)) S(G) is
immediate. The converse follows easily from Lemma 4.3.15 together with (4.35).
This proves (4.34) for a = N N. This implies that for any N N, S(G) is
included in
&
%
&
%
Dom (I + Rp )N Range (I + Rp )N
and we can apply the analyticity results (Part (1c)) of Theorem 4.3.6: xing
S(G), the function a (I + Rp )a is holomorphic in {a C : N < Re a < N }.
We observe that by Corollary 4.3.11 (ii), if N < Re a < 0, all the functions
(I + Rp )a coincide with Ba for any p [1, ) {o }. This shows that for
each a C xed, (I + Rp )a is independent of p. Furthermore, it is Schwartz.
Indeed if Re a < 0 this follow from Lemma 4.3.15. If Re a 0, we write a = ao + a
with ao N and Re a < 0 and we have in the sense of operators
(I + R)a (I + R)ao (I + R)a .
The operator (I + R)ao is a dierential operator, hence maps S(G) to itself, and
the operator (I + R)a maps S(G) to itself by Lemma 4.3.15. Thus in any case
(I + Rp )a S(G) and is independent of p.
We have obtained that (I+Rp )a (S(G)) S(G) for any p (1, ), a C. As
{(I+Rp )a }1 = (I+Rp )a by Theorem 4.3.6 (1a), this proves the equality in (4.34)
for any a C. Lemma 4.3.15 says that this action is continuous if Re a < 0. This
is also the case for Re a 0 since we can proceed as above and write a = ao + a
with ao N and Re a < 0, the action of (I + R)ao being continuous on S(G). This
concludes the proof of Corollary 4.3.16.
Corollary 4.3.16 implies that the following denition makes sense.
Denition 4.3.17. Let R be a positive Rockland operator of homogeneous degree
and let s R. For any tempered distribution f S (G), we denote by (I + R)s/ f
the tempered distribution dened by
s/ ,
(I + R)s/ f, = f, (I + R)
S(G).
218
Chapter 4. Rockland operators and Sobolev spaces
4.4 Sobolev spaces on graded Lie groups
In this section we dene the (homogeneous and inhomogeneous) Sobolev spaces
associated to a positive Rockland operator R and show that they satisfy similar
properties to the Euclidean Sobolev spaces and to the Sobolev spaces dened
and studied by Folland [Fol75] on stratied Lie groups. In Section 4.4.5, we show
that the constructed spaces are actually independent of the choice of a positive
Rockland operator R on a graded Lie group with which we start our construction.
In Section 4.4.7, we list the main properties of our Sobolev spaces.
4.4.1 (Inhomogeneous) Sobolev spaces
We rst need the following lemma:
Lemma 4.4.1. We keep the notation of Theorem 4.3.6. For any s R and p
s
[1, ) {o }, the domain of the operator (I + Rp ) contains S(G), and the map
s
f (I + Rp ) f Lp (G)
denes a norm on S(G). We denote it by
s
f Lps (G) := (I + Rp ) f Lp (G) .
Moreover, any sequence in S(G) which is Cauchy for Lps (G) is convergent in
S (G).
We have allowed ourselves to write L (G) = Lo (G) for the supremum
norm. We may also write or o .
s
Proof. By Corollary 4.3.16, the domain of (I + Rp ) contains S(G). Since the
s
s
operator (I + Rp ) is linear, it is easy to check that the map f (I + Rp ) f p
is nonnegative and satises the triangle inequality. Since (I + Rp )s/ is injective
by Theorem 4.3.6, Part (1), we have that f Lps (G) = 0 implies f = 0.
Clearly Lp0 (G) = p , so in the case of s = 0 a Cauchy sequence of
Schwartz functions converges in Lp norm, thus also in S (G).
s
Let us assume s > 0. By Corollary 4.3.11 (ii), the operator (I + Rp ) is
bounded on Lp (G). Hence we have
Lp (G) C Lps (G)
on S(G). Consequently a Lps (G) Cauchy sequence of Schwartz functions converge
in Lp norm thus in S (G).
Now let us assume s < 0. Let {f }N be a sequence of Schwartz functions
which is Cauchy for the norm Lps (G) . By (4.35) we have
s
f = (I + Rp ) f Bs .
4.4. Sobolev spaces on graded Lie groups
219
Furthermore, if S(G) then using (1.14) and (4.13), we have
s
(I + Rp ) f (x) ( Bs ) (x) dx.
f (x)(x)dx =
G
(4.36)
By assumption the sequence {(I+Rp ) f }N is Lp (G) Cauchy thus convergent
in Lp (G). By Lemma 4.3.15, Bs S(G). Therefore, the righthand side of
(4.36) is convergent as . Hence the scalar sequence f , converges for any
S(G). This shows that the sequence {f } converges in S (G).
Lemma 4.4.1 allows us to dene the (inhomogeneous) Sobolev spaces:
Denition 4.4.2. Let R be a positive Rockland operator on a graded Lie group
G. We consider its Lp analogue Rp and the powers of (I + Rp )a as dened in
Theorems 4.3.3 and 4.3.6. Let s R.
If p [1, ), the Sobolev space Lps,R (G) is the subspace of S (G) obtained
by completion of S(G) with respect to the Sobolev norm
s
f Lps,R (G) := (I + Rp ) f Lp (G) ,
f S(G).
o
If p = o , the Sobolev space L
s,R (G) is the subspace of S (G) obtained by
completion of S(G) with respect to the Sobolev norm
s
o
:= (I + Ro ) f L (G) ,
f L
s,R (G)
f S(G).
When the Rockland operator R is xed, we may allow ourselves to drop the
index R in Lps,R (G) = Lps (G) to simplify the notation.
We will see later that the Sobolev spaces actually do not depend on the
Rockland operator R, see Theorem 4.4.20.
By construction the Sobolev space Lps (G) endowed with the Sobolev norm is
a Banach space which contains S(G) as a dense subspace and is included in S (G).
The Sobolev spaces share many properties with their Euclidean counterparts.
Theorem 4.4.3. Let R be a positive Rockland operator of homogeneous degree
on a graded Lie group G. We consider the associated Sobolev spaces Lps (G) for
p [1, ) {o } and s R.
1. If s = 0, then Lp0 (G) = Lp (G) for p [1, ) with Lp0 (G) = Lp (G) , and
o
L
0 (G) = Co (G) with L0 o (G) = L (G) .
2. If s > 0, then for any a C with Re a = s, we have
a
%
a&
Lps (G) = Dom (I + Rp ) = Dom(Rp ) Lp (G),
and the following norms are equivalent to Lps (G) :
s
f f Lp (G) + (I + Rp ) f Lp (G) , f f Lp (G) + Rp f Lp (G) .
220
Chapter 4. Rockland operators and Sobolev spaces
3. Let s R and f S (G).
Given p (1, ), we have f Lps (G) if and only if the tempered distribution (I + Rp )s/ f dened in Denition 4.3.17 is in Lp (G), in the
sense that the linear mapping
p )s/
S(G) (I + R)s/ f, = f, (I + R
extends to a bounded functional on Lp (G) where p is the conjugate
exponent of p.
f L1s (G) if and only if (I + R1 )s/ f L1 (G) in the sense that the
linear mapping
)s/
S(G) (I + R)s/ f, = f, (I + R
o
extends to a bounded functional on Co (G) and is realised as a measure
given by an integrable function.
s/
o
f Co (G) in the sense that the
f L
s (G) if and only if (I + Ro )
linear mapping
1 )s/
S(G) (I + R)s/ f, = f, (I + R
extends to a bounded functional on L1 (G) and is realised as integration
against functions in Co (G).
4. If a, b R with a < b and p [1, ) {o }, then the following continuous
strict inclusions hold
S(G) Lpb (G) Lpa (G) S (G),
and an equivalent norm for Lpb (G) is
ba
Lpb (G) f f Lpa (G) + Rp f Lpa (G) .
5. For p [1, ) {o } and any a, b, c R with a < c < b, there exists a
positive constant C = Ca,b,c such that for any f Lpb , we have f Lpc Lpa
and
f Lp ,
f Lpc Cf 1
Lp
a
b
where := (c a)/(b a).
In Theorem 4.4.20, we will see that the denition of the Sobolev spaces and
their properties given in Theorem 4.4.3 hold independently of the chosen Rockland
operator R.
From now on, we will often use the notation Lp0 (G) since this allows us not to
distinguish between the cases Lp0 (G) = Lp (G) when p [1, ) and Lp0 (G) = Co (G)
when p = o .
In the proof of Part (2) of Theorem 4.4.3, we will need the following exercise
in functional analysis:
4.4. Sobolev spaces on graded Lie groups
221
Lemma 4.4.4. Let T1 and T2 be two linear operators between two Banach spaces
X Y. We assume that T1 and T2 are densely dened and share the same domain.
We also assume that they are both closed injective operators and that T2 is bijective
with a bounded inverse. Then the graph norms of T1 and T2 are equivalent, that
is,
C > 0 x Dom(T1 ) = Dom(T2 )
(x + T2 x) x + T1 x C(x + T2 x).
Sketch of the proof of Lemma 4.4.4. One can check easily that T := T1 T21 denes a closed linear operator T : Y Y dened on the whole space Y. By the
closed graph theorem (see, e.g., [Rud91, Theorem 2.15] or [RS80, Thm III. 12]),
T is bounded. Furthermore, T is injective as the composition of two injective
operators. It may not have a closed range in Y but one checks easily that the
operator
Y X Y
,
(T21 , T ) :
y (T21 y, T y)
has a closed range in X Y. Hence the restriction of (T21 , T ) onto its image is
bounded with a bounded inverse (see e.g. [RS80, Thm III. 11]). Consequently,
T21 y + T y
y
for any element y Y, in particular of the form y = T2 x, x Dom(T2 ).
We can now prove Theorem 4.4.3.
0
Proof of Theorem 4.4.3. Part (1) is true since (I+Rp ) = I. Let us prove Part (2).
So let s > 0. Clearly Lps (G) coincides with the domain of the unbounded operator
s
(I + Rp ) (see Theorem 4.3.6 (2)) hence it is a proper subspace of Lp (G). As the
s
operator (I + Rp ) is bounded on Lp (G), we have Lp (G) C Lps (G) on
p
Ls (G). So Lp (G) + Lps (G) is a norm on Lps (G) which is equivalent to the
s
Sobolev norm. Theorem 4.3.6 implies that Rp and (I+Rp ) satisfy the hypotheses
of Lemma 4.4.4. This shows part (2).
Part (3) follows from Part (2) and the duality properties of the spaces Lp (G)
and Co (G) in the case s 0. We now consider the case s < 0. By Lemma 4.3.15
and Corollary 4.3.11 (and also Lemma 4.3.14), the mapping
p )s/ = f, Bs
Ts,p ,f : S(G) f, (I + R
is well dened for any f S (G). If Ts,p ,f admits a bounded extension to a
functional on Lp0 (G), then we denote this extension Ts,p ,f and we have
Ts,p ,f L (Lp ,C) = f Lps (G) .
0
(4.37)
222
Chapter 4. Rockland operators and Sobolev spaces
This is certainly so if f S(G). Furthermore a sequence {f }N of Schwartz functions is convergent for the Sobolev norm Lps (G) if and only if {Ts,p ,f } is con
vergent in Lp0 (G) (see Lemma 4.3.2). In the case of convergence, by Lemma 4.4.1,
{f }N converges in the sense of distributions. Denoting this limit by f S (G),
we have
7
6
= Ts,p ,f .
lim Ts,p ,f
S(G)
It is easy to see, by linearity of f1 Ts,p ,f1 and (4.37), that Ts,p ,f extends to a
continuous functional on Lp0 (G).
Conversely, let us consider a distribution f S (G) such that Ts,p ,f extends
to a bounded functional Ts,p ,f on Lp0 (G). If {f }N is a sequence of Schwartz
functions converging to f in S (G), then
lim Ts,p ,f () = Ts,p ,f ()
for every S(G), and using the density of S(G) in Lp0 (G) and the BanachSteinhaus Theorem, this shows that {Ts,p ,f } converges to Ts,p ,f in the norm of
the dual of Lp0 (G). This shows the case s < 0 and concludes the proof of Part (3).
Let us show Part (4). Let a b and p [1, ) {o }. By Theorem 4.3.6
(1), we have in the sense of operators
a
(I + Rp ) (I + Rp )
Since the operator (I + Rp )
f Lpa (G)
ab
(I + Rp ) .
is bounded, we have for any f S(G)
(I + Rp ) f p = (I + Rp )
(I + Rp )
ab
ab
ab
(I + Rp ) f p
L (Lp0 ) (I + Rp ) f p = (I + Rp )
ab
L (Lp0 ) f Lpb .
By density of S(G), this implies the continuous inclusion Lpb Lpa . Note that we
also have if a < b
f Lpb (G)
ba
(I + Rp )
(I + Rp ) f Lp (G) + Rp (I + Rp ) f Lp (G) ,
(I + Rp ) f p = (I + Rp ) f Lpba (G)
ba
by Part (2) above for any f S(G). By Theorem 4.3.6 (5), we can commute
ba
the operators Rp and (I + Rp ) in this last expression. Consequently, we have
obtained for any f S(G),
ba
f Lpb (G)
f Lpa (G) + Rp f Lpa (G) .
4.4. Sobolev spaces on graded Lie groups
223
ba
By density of S(G), this holds for any f Lpb (G). Since the operator Rp is
unbounded, this also implies the strict inclusions given in Part (4).
Part (5) follows from Theorem 4.3.6 (1f) for the case of a = 0. For f Lpb ,
a
we then apply this to b a, c a instead of b and c and := (I + Rp ) f Lpba
instead of f .
This concludes the proof of this part and of the whole theorem.
Theorem 4.4.3 has the two following corollaries. The rst one is an easy
consequence of Part (3).
Corollary 4.4.5. We keep the setting and notation of Theorem 4.4.3. Let s < 0
and p [1, ) {o }. Let f S (G).
The tempered distribution f is in Lps (G) if and only if the mapping
S(G) f, Bs
extends to a bounded linear functional on Lp0 (G) with the additional property that
for p = 1, this functional on Co (G) is realised as a measure given by an
integrable function,
if p = o , this functional on L1 (G) is realised by integration against a function in Co (G).
Corollary 4.4.6. We keep the setting and notation of Theorem 4.4.3. Let s R
and p [1, ) {o }. Then D(G) is dense in Lps (G).
Proof of Corollary 4.4.6. This is certainly true for s 0 (see the proof of Parts
(1) and (2) of Theorem 4.4.3). For s < 0, it suces to proceed as in the last part
of the proof of Part (3) with a sequence of functions f D(G).
Theorem 4.4.3, especially Part (3), implies the following property regarding
duality of Sobolev spaces. This will be improved in Proposition 4.4.22 once we
show in Theorem 4.4.20 that the Sobolev spaces are indeed independent of the
considered Rockland operator.
Lemma 4.4.7. Let R be a positive Rockland operator on a graded Lie group G. We
consider the associated Sobolev spaces Lps,R (G). If s R and p (1, ), the dual
space of Lps,R (G) is isomorphic to Lps,R (G) via the distributional duality, where
p is the conjugate exponent of p, p1 + p1 = 1.
Proof of Lemma 4.4.7. Clearly if f Lps,R (G) then for any S(G),
p ) (I + R
p ) = (I + Rp ) f, (I + R
p )
f, = f, (I + R
s
by Theorem 4.3.6. Hence by Theorem 4.4.3 Part (3),
s
p ) s p
f,  (I + Rp ) f p (I + R
224
Chapter 4. Rockland operators and Sobolev spaces
and the linear function S(G) f, extends to a bounded linear functional
on Lps,R (G). Conversely, let be a bounded linear functional on Lps,R (G). Then
since
p )s/ S(G) = S(G) Lp (G),
(I + R
s,R
p )s/ is
see Corollary 4.3.16 and Denition 4.4.2, the linear functional (I + R
well dened on S(G) and satises for any S(G),
p )s/ 
p )s/ () =  (I + R
 (I + R
p )s/ p
C(I + R
L
s,R
= CLp .
0
p )s/ extends into a bounded linear functional on Lp (G).
Therefore, (I + R
0
In the next statement, we show how to produce functions and converging
sequences of Sobolev spaces using the convolution:
Proposition 4.4.8. We keep the setting and notation of Theorem 4.4.3. Here a R
and p [1, ) {o }.
(i) If f Lp0 (G) and S(G), then f Lpa for any a and p.
(ii) If f Lpa (G) and S(G), then
a
a
(I + Rp ) ( f ) = (I + Rp ) f ,
(4.38)
and f Lpa (G) with
Furthermore, if
f Lpa (G) L1 (G) f Lpa (G) .
(4.39)
= 1, writing
(x) := Q (1 x)
for each > 0, then { f } converges to f in Lpa (G) as 0.
Proof of Proposition 4.4.8. Let us prove Part (i). Here f Lp0 (G). By density of
S(G) in Lp0 (G), we can nd a sequence of Schwartz functions {f } converging to
f in Lp0 norm. Then f S(G) and for any N N,
RN (f ) = f RN f RN
in Lp0 (G),
N
p
thus RN
p (f ) = f R L (G) and
p
f Lp0 (G) + RN
p (f )L0 (G) < .
By Theorem 4.4.3 (4), this shows that f is in LpN for any N N, hence in
any pSobolev spaces. This proves (i).
4.4. Sobolev spaces on graded Lie groups
225
Let us prove Part (ii). We observe that both sides of Formula (4.38) always
make sense as convolutions of a Schwartz function with a tempered distribution.
Let us rst assume that f S(G). Formula (4.38) is true if a < 0 by Corollary
a
4.3.11 (ii) since then the (I + Rp ) is a convolution operator with an integrable
a
convolution kernel. Formula (4.38) is also true if a N0 as in this case (I+Rp ) is
a leftinvariant dierential operator by Theorem 4.3.6 (1a). Hence Formula (4.38)
holds for any a > 0 by writing a = a0 + a , a0 N0 , a < 0, and
a
(I + Rp ) f = (I + Rp )
a0
a
(I + Rp ) f.
Together with Corollary 4.3.16, this shows that Formulae (4.38) and consequently
(4.39) hold for any a R and f S(G).
By density of S(G) in Lps (G) and (4.39), this shows that Formulae (4.38) and
(4.39) hold for any f Lps (G).
Hence
f Lpa (G) with Lpa norm 1 f Lpa (G) .
If G = 1, by Lemma 3.1.58 (i),
a
f f Lpa (G) = (I + Rp ) ( f f )p
a
a
= (I + Rp ) f (I + Rp ) f p 0 0,
that is, { f } converges to f in Lpa (G) as 0. This proves (ii).
4.4.2 Interpolation between inhomogeneous Sobolev spaces
In this section, we prove that interpolation between Sobolev spaces Lpa (G) works
in the same way as its Euclidean counterpart.
Theorem 4.4.9. Let R and Q be two positive Rockland operators on two graded Lie
groups G and F . We consider their associated Sobolev spaces Lpa (G) and Lqb (F ).
Let p0 , p1 , q0 , q1 (1, ) and let a0 , a1 , b0 , b1 be real numbers.
We also consider a linear mapping T from Lpa00 (G) + Lpa11 (G) to locally integrable functions on F . We assume that T maps Lpa00 (G) and Lpa11 (G) boundedly
into Lqb00 (F ) and Lqb11 (F ), respectively.
Then T extends uniquely to a bounded mapping from Lpat (G) to Lqbt (F ) for
t [0, 1] where at , bt , pt , qt are dened by
1 1
1 1
1 1
at , bt , ,
= (1 t) a0 , b0 , ,
+ t a1 , b1 , ,
.
pt q t
p0 q 0
p1 q 1
The idea of the proof is similar to the one of the Euclidean or stratied cases,
see [Fol75, Theorem 4.7]. Some arguments will be modied since our estimates
for (I + R)i L (Lp ) are dierent from the ones obtained by Folland in [Fol75].
For this, compare Corollary 4.3.13 and Proposition 4.3.7 in this monograph with
[Fol75, Proposition 4.3].
226
Chapter 4. Rockland operators and Sobolev spaces
Proof of Theorem 4.4.9. By duality (see Lemma 4.4.7) and up to a change of
notation, it suces to prove the case
a1 a0
and
b1 b 0 .
(4.40)
This fact is left to the reader to check. The idea is to interpolate between the
operators formally given by
bz
Tz = (I + Q) Q T (I + R)
az
(4.41)
where R and Q denote the degrees of homogeneity of R and Q, respectively, and
the complex numbers az and bz are dened by
(az , bz ) := z (a1 , b1 ) + (1 z) (a0 , b0 ) ,
for z in the strip
S := {z C : Re z [0, 1]}.
In (4.41), we have abused the notation regarding the fractional powers of I + Rp
and I + Qq and removed p and q. This is possible by Corollary 4.3.16 and density
of the Schwartz space in each Sobolev space. Hence (4.41) makes sense. We will
use complex interpolation given by Theorem A.6.1, which requires to start with
the space B of compactly supported simple functions on G (see Remark A.6.2).
To solve this technical problem we proceed as in the proof of [Fol75, Theorem 4.7]:
we will use the convolution of a function in B with a bump function depending
on at the end of the proof.
The hypotheses on T give that the operator norms
bj
T L (Lpaj ,Lqj ) = (I + Q) Q T (I + R)
j
bj
L (Lpj ,Lqj ) ,
j = 0, 1,
are nite.
By Corollary 4.3.16, for any S(G) and S(F ), we have
Tz , = T (I + R)
N az
bz
M + Q (I + Q)
M
(I + R)N , (I + Q)
for any M, N Z. In particular, for M and N large enough, Theorem 4.3.6 implies
that
S z Tz ,
is analytic. With M = N N large enough, for instance the smallest integer with
N > a1 , a0 , b1 , b0 , we get
Tz ,  A(z) B(z) T L (Lpa1 ,Lq1 ) LpN1 LqN1 ,
1
b1
where A(z) and B(z) denote the operator norms
A(z) := (I + R)
N +
az +a1
R
L (Lp1 )
and
M +
B(z) := (I + Q)
bz b1
Q
L (Lq1 ) .
4.4. Sobolev spaces on graded Lie groups
227
We can write
A(z) = (I + R)(+z) L (Lp1 )
with = N
a1 a0
a1 a0
> 0, =
0.
R
R
Thus
A(z)
(I + R)(+Re z) L (Lp1 ) (I + R)Im z L (Lp1 )
h1 L1 eIm z ,
by Corollary 4.3.13 and Proposition 4.3.7 using the notation of their statements.
We have a similar property for B(z). This implies easily that there exists a constant
C depending on , , a1 , a0 , b1 , b0 and F, G, R, Q such that we have
ln Tz ,  C(1 + Im z).
z S
We now estimate operator norms of Tz for z on the boundary of the strip,
that is, z = j + iy, j = 0, 1, y R:
Tz L (Lpj ,Lqj )
bz
= (I + Q) Q T (I + R)
= (I + Q)
bz bj
Q
(I + Qqj )
az
(I + Q)
bz bj
Q
bj
Q
L (Lpj ,Lqj )
aj
T (I + R) R (I + R)
aj az
R
L (Lqj ) T L (Lpaj ,Lqj ) (I + Rpj )
bj
= (I + Qqj )
iy
b1 b0
Q
L (Lpj ,Lqj )
aj az
R
L (Lpj )
L (Lqj ) T L (Lpaj ,Lqj ) (I + Rpj )
j
iy
a0 a1
R
bj
L (Lpj ) .
Proposition 4.3.7 then implies
Tj+iy L (Lpj ,Lqj ) CT L (Lpaj ,Lqj ) e
j
a1 a0
R
y Q
bj
b0 b1
R
y
where C, R and Q are positive constants obtained from the applications of
Proposition 4.3.7 to R and Q.
The
end of the proof is now classical. We x a nonnegative function S(G)
with G = 1 and write
(x) := Q (1 x)
for > 0. If f B, then f S(G) (see Lemma 3.1.59) and we can set for any
> 0, z S,
Tz, f := Tz (f ) .
Clearly Tz, satisfy the hypotheses of Theorem A.6.1 (see also Remark A.6.2).
Thus for any t [0, 1], there exists a constant Mt > 0 independent of such that
f B
Tt, f qt Mt f pt .
228
Chapter 4. Rockland operators and Sobolev spaces
For p (1, ), we consider the space Vp of functions of the form = f ,
with f B and > 0, satisfying f p 2f p . By Lemma 3.1.59, the space
Vp contains S(G) and is dense in Lp (G) for p (1, ). Going back to the proof
of Theorem 4.4.9, we have obtained for any t [0, 1] and = f Vpt , that
Tt qt = Tt, f qt Mt f pt 2Mt pt .
This shows that Tt extends to a bounded operator from Lpt (G) to Lqt (G).
As a consequence of the interpolation properties, we have
Corollary 4.4.10. Let S (G) and let T be its associated convolution operator
T : S(G) .
Let also a R, p (1, ) and let { , Z} be a sequence of real numbers which
tends to as . Assume that for any Z, the operator T extends
continuously to a bounded operator Lp (G) Lpa+ (G). Then the operator T
extends continuously to a bounded operator Lp (G) Lpa+ (G) for any R.
Furthermore, for any c 0, we have
sup T L (Lp ,Lpa+ ) Cc max T L (Lp ,Lpa+ ) , T L (Lp ,Lpa+ )
c
where N0 is the smallest integer such that c and c.
4.4.3 Homogeneous Sobolev spaces
Here we dene the homogeneous version of our Sobolev spaces and obtain their
rst properties. Many proofs are obtained by adapting the corresponding inhomogeneous cases and we may therefore allow ourselves to present them more
succinctly. For technical reasons explained below, the denition of homogeneous
Sobolev spaces is restricted to the case p (1, ).
As in the inhomogeneous case, we rst need the following lemma:
Lemma 4.4.11. We keep the notation of Theorem 4.3.6.
s
1. For any s R and p [1, ) {o }, the map f Rp f Lp (G) denes a
s
norm on S(G) Dom(Rp ). We denote it by
s
f L ps (G) := Rp f Lp (G) .
s
2. For any s 0 and p [1, ){o }, S(G)Dom(Rp ) contains Rs (S(G))
which is dense in Range(Rp ) for Lp (G) , and any sequence in S(G)
s
Dom(Rp ) which is Cauchy for L ps (G) is convergent in S (G).
4.4. Sobolev spaces on graded Lie groups
229
s
3. If s > 0 and p (1, ), then S(G) Dom(Rp ) and any sequence in S(G)
which is Cauchy for L ps (G) is convergent in S (G).
s
Proof of Lemma 4.4.11. The fact that the map f Rp f Lp (G) denes a norm
on S(G) follows easily from Theorem 4.3.6 Part (1).
In the case s = 0, L p (G) = Lp (G) and Part 2 is proved in this case.
0
Let s < 0 and p [1, ) {o }. By Theorem 4.3.6 (especially Parts
s
(1a) and (1c)), for any N N with N > s/, Dom(R ) contains RN (S(G))
s
and RN (S(G)) is dense in Range(Rp ). Consequently S(G) Dom(Rp ) contains
RN (S(G)) and is dense in Range(Rp ). Let p be the dual exponent of p, i.e.
1
1
p + p = 1 with the usual extension. Theorem 4.3.6 (1), and the duality properties
p
imply
of L as well as Rt = R
s
p ,
f,  Rp f Lp (G) R
L (G)
p
s
for any f S(G) Dom(Rp ) and S(G). Furthermore, as S(G)
s
Dom(Rp ), Theorem 4.3.6 (1) also yields for any S(G)
s
p
R
L (G)
p
s
s
p , R
p
max R
L (G)
L (G)
p
p
max
[]=
s
s
,
X Lp (G)
for some constant C = CN,R . We have obtained that
s
f,  CRp f Lp (G)
max
[]=N,N +1
X Lp (G)
for any f S(G) Dom(Rp ) and S(G). This together with the properties of
the Schwartz space (see Section 3.1.9) easily implies Part 2.
s
Let s > 0. By Theorem 4.3.6 (1g), S(G) Dom(Rp ).
Let p (1, ). By Corollary 4.3.11 Part (i), if s (0, Q
p ), then there exists
C > 0 such that
f S(G)
f Lq (G) CRp f Lp (G) = Cf L ps (G) ,
where q (1, ) is such that
s
1 1
= .
p q
Q
Note that q is indeed in (1, ) as s < Q
p . Hence if {f } S(G) is Cauchy for
L ps (G) , then {f } S(G) is Cauchy for Lq (G) thus in S (G). This shows
Part 3 for any s > 0, p (1, ) satisfying ps < Q.
230
and
Chapter 4. Rockland operators and Sobolev spaces
Q
Q
If s [N Q
p , (N + 1) p ) for some N N0 , we write s = s1 + s with s (0, p )
s1 [(N 1)
Q
Q
,N )
p
p
and by Corollary 4.3.11 Part (i) with Theorem 4.3.6 (1), we have
f S(G)
C = Cs ,p
where q (1, ) is such that
s1
Rq f Lq CRp f Lp (G) ,
s
1 1
= .
q
p
Q
Hence if {f } S(G) is Cauchy for L ps (G) , then {f } S(G) is Cauchy for
L qs (G) . Note that
1
NQ
NQ
<
.
s1
p
q
Recursively, this shows Part 3.
The use of Corollary 4.3.11 in the proof above requires p (1, ). Moreover,
by Corollary 4.3.4 (ii), the range of Rp is dense in Lp (G) for p (1, o ]. As we
want to have a unied presentation for all the homogeneous spaces of any exponent
s R, we restrict the parameter p to be in (1, ) only.
Denition 4.4.12. Let R be a Rockland operator of homogeneous degree on
a graded Lie group G, and let p (1, ). We denote by L ps,R (G) the space of
s
tempered distribution obtained by the completion of S(G) Dom(Rp ) for the
norm
s
f L ps (G) := Rp f p , f S(G) Dom(Rs/
p ).
As in the inhomogeneous case, we will write L ps (G) or L ps,R but often omit
the reference to the Rockland operator R. We will see in Theorem 4.4.20 that
the homogeneous Sobolev spaces do not depend on a specic R. Adapting the
inhomogeneous case, one obtains easily:
Proposition 4.4.13. Let G be a graded Lie group of homogeneous dimension Q. Let
R be a positive Rockland operator of homogeneous degree on G. Let p (1, )
and s R.
1. We have
)
L ps (G) S (G).
S(G) Dom(Rs/
p
Equipped with the homogeneous Sobolev norm L ps (G) , the space L ps (G)
s/
is a Banach space which contains S(G) Dom(Rp ) as dense subspace.
s/
2. If s > Q/p then S(G) Dom(Rp ) L ps (G). If s < 0 then S(G)
s
Dom(Rp ) contains Rs (S(G)) which is dense in Lp (G).
4.4. Sobolev spaces on graded Lie groups
231
3. If s = 0, then L p0 (G) = Lp (G) for p (1, ) with L p (G) = Lp (G) .
0
s/
4. Let s R, p (1, ) and f S (G). If f L ps (G) then Rp f Lp (G) in
the sense that the linear mapping
s/
s/
S(G) Dom(R
)
f, R
p
p
is densely dened on Lp (G) and extends to a bounded functional on Lp (G)
where p is the conjugate exponent of p. The converse is also true.
5. If 1 < p < q < and a, b R with
1 1
b a = Q( ),
p q
then we have the continuous inclusion
L pb L qa
that is, for every f L pb , we have f L qa and there exists a constant C =
Ca,b,p,q,G > 0 independent of f such that
f L qa Cf L p .
b
6. For p (1, ) and any a, b, c R with a < c < b, there exists a positive
constant C = Ca,b,c such that we have for any f L pb
1
f L pc Cf L
p
p f L
a
where := (c a)/(b a).
Proof of Proposition 4.4.13. Parts (1), (2), and (3) follow from Lemma 4.4.11 and
its proof. Part (4) follows easily by duality and Lemma 4.4.11. Parts (5) and (6)
are an easy consequence of the property of the fractional powers of R on the Lp s/
spaces (cf. Theorem 4.3.6) and the operator Rp , s (0, Q), being of type s and
independent of p (cf. Corollary 4.3.11 (i)).
Note that Part (2) of Proposition 4.4.13 can not be improved in general as
s
the inclusions S(G) Dom(Rp ) or S(G) L sp (G) can not hold in general for
any group G as they do not hold in the Euclidean case i.e. G = (Rn , +) with
the usual dilations. Indeed in the case of Rn , p = 2, one can construct Schwartz
functions which can not be in L 2s with s < n/2. It suces to consider a function
S(G) satisfying ()
1 on a neighbourhood of 0 since then s ()
is not
square integrable about 0 for s < n/2.
As in the homogeneous case (see Lemma 4.4.7), Part (4) of Proposition 4.4.13
above implies the following property regarding duality of Sobolev spaces. This
will be improved in Proposition 4.4.22 once we know (see Theorem 4.4.20) that
homogeneous Sobolev spaces are indeed independent of the considered Rockland
operator.
232
Chapter 4. Rockland operators and Sobolev spaces
Lemma 4.4.14. Let R be a positive Rockland operator on a graded Lie group G.
We consider the associated homogeneous Sobolev spaces L ps,R (G). If s R and p
(1, ), the dual space of L p (G) is isomorphic to L p (G) via the distributional
s,R
duality, where p is the conjugate exponent of p, i.e.
s,R
1
1
p + p
= 1.
The following interpolation property can be proved after a careful modication of the inhomogeneous proof:
Proposition 4.4.15. Let R and Q be two positive Rockland operators on two graded
Lie groups G and F respectively. We consider their associated homogeneous Sobolev
spaces L pa (G) and L qb (F ). Let p0 , p1 , q0 , q1 (1, ) and a0 , a1 , b0 , b1 R.
We also consider a linear mapping T from L pa00 (G) + L pa11 (G) to locally integrable functions on F . We assume that T maps L pa00 (G) and L pa11 (G) boundedly
into L qb00 (F ) and L qb11 (F ), respectively.
Then T extends uniquely to a bounded mapping from L pat (G) to L qbt (F ) for
t [0, 1], where at , bt , pt , qt are dened by
1 1
1 1
1 1
= (1 t) a0 , b0 , ,
+ t a1 , b1 , ,
.
at , bt , ,
pt q t
p0 q 0
p1 q 1
Sketch of the proof of Proposition 4.4.15. By duality (see Lemma 4.4.14) and up
to a change of notation, it suces to prove the case a1 a0 and b1 b0 . The idea
is to interpolate between the operators formally given by
Tz = Q
b1 b0
Q
b0
Q Q T R
a0 a1
R
z S,
(4.42)
with the same notation for R , Q , az , bz and S as in the proof of Theorem 4.4.9.
In (4.42), we have abused the notation regarding the fractional powers of Rp and
Qq and removed p and q thanks to by Theorem 4.3.6 (1). Moreover, Theorem 4.3.6
implies that on S(G), each operator Tz , z S, coincides with
Tz = Q
(1z)
b0 b1
Q
b1
Q Q T R
(1z)
a1 a0
R
and that for any S(G) and S(F ), z Tz , is analytic on S. We also
have
bz b1
az +a1
Q Lq1 .
Tz ,  T L (L pa1 ,L q1 ) R R Lp1 Q
1
b1
Note that Re az + a1 0 thus we have
R
az +a1
R
Lp1
R
Re az +a1
R
Lp1 R
1
Lp1 R Lp1 e
Im az
R
Im a 
z
R
Lp1
by Theorem 4.3.6 (1f) with N the smallest integer strictly greater than Re az +a1
and = (Re az + a1 )/N , and by Proposition 4.3.9 using the notation of its
4.4. Sobolev spaces on graded Lie groups
233
bz b1
Q q and all these estimates imply
statement. We have similar bounds for Q
1
easily that there exists a constant depending on , , a1 , a0 , b1 , b0 such that
z S
ln Tz ,  C(1 + Im z).
For the estimate on the boundary of the strip, that is, z = j + iy, j = 0, 1, y R,
we see as in the proof of Theorem 4.4.9:
iy
Tz L (Lpj ,Lqj ) Qqj
b1 b0
Q
iy
L (Lqj ) T L (L paj ,L qj ) Rpj
j
a0 a1
R
bj
L (Lpj ) .
Proposition 4.3.9 then implies
Tj+iy L (Lpj ,Lqj ) CT L (L paj ,L qj ) e
j
bj
a1 a0
R
y Q
b0 b1
R
y
where C, R and Q are positive constants obtained from the applications of
Proposition 4.3.9 to R and Q. We conclude the proof in the same way as for
Theorem 4.4.9.
4.4.4 Operators acting on Sobolev spaces
In this section we show that leftinvariant dierential operators act continuously
on homogeneous and inhomogeneous Sobolev spaces. We will also show a similar
property for operators of type , Re = 0.
In the statements and in the proofs of this section, we keep the same notation for an operator dened on a dense subset of some Lp space and its possible
bounded extensions to some Sobolev spaces in order to ease the notation.
Theorem 4.4.16. Let G be a graded Lie group.
1. Let T be a leftinvariant dierential operator of homogeneous degree T . Then
for every p (1, ) and s R, T maps continuously Lps+T (G) to Lps (G).
Fixing a positive Rockland operator R in order to dene the Sobolev norms,
it means that
C = Cs,p,T > 0
S(G)
T Lps (G) CLps+
(G) .
2. Let T be a T homogeneous leftinvariant dierential operator. Then for every
p (1, ) and s R, T maps continuously L ps+T (G) to L ps (G). Fixing a
positive Rockland operator R in order to dene the Sobolev norms, it means
that
C = Cs,p,T > 0
L ps+T (G)
T L ps (G) CL p
s+T
(G) .
We start the proof of Theorem 4.4.16 with studying the case of T = Xj . This
uses the denition and properties of kernel of type 0, see Section 3.2.5.
234
Chapter 4. Rockland operators and Sobolev spaces
Lemma 4.4.17. Let R be a positive Rockland operator on a graded Lie group G
and Ia the kernel of its Riesz operator as in Corollary 4.3.11.
1. For any j = 1, . . . , n, Xj Ij is a kernel of type 0.
2. If is a kernel of type 0, then, for any j = 1, . . . , n, Xj Ij is a kernel
of type 0 and, more generally, for any multiindex Nn0 , the kernel
()1
()n
X I[1 ] . . . I[n ]
is of type 0.
is an operator
3. If T is an operator of type 0, then, for any N N, RN T RN
2
of type 0 hence it is bounded on Lp (G), p (1, ).
4. For any j = 1, . . . , n and for any N N0 , RN Xj R2
type 0.
is an operator of
In Part 2, we have used the notation
m
f () = f . . . f
8 9: ;
m
times
Proof of Lemma 4.4.17. We adopt the notation of the statement. By Corollary
4.3.11 (i), Ij is a kernel of type j (0, Q) hence, by Lemma 3.2.33, Xj Ij is a
kernel of type 0. This shows Part 1.
More generally, if is a kernel of type 0, then Ij is a kernel of type j
by Proposition 3.2.35 (ii) hence by Lemma 3.2.33, Xj ( Ij ) is a kernel of type
0. Iterating this procedure shows Part 2.
Let T be an operator of type 0. We denote by its kernel. Let N N.
The operator RN can be written as a linear combination of X , Nn0 with
[] = N . Using the spectral calculus of R to dene and decompose RN
2 , this
can
be
written
as
a
linear
combination
over
shows that the operator RN T RN
2
[] = N of the operators X T R2 1 . . . R2 n whose kernel can be written
()1
()n
is of
as X I[1 ] . . . I[n ] . Part 2 implies that the operator RN T RN
2
type 0. By Theorem 3.2.30, it is a bounded operator on Lp (G), p (1, ). This
shows Part 3.
Part 4 follows from combining Parts 1 and 3.
We can now nish the proof of Theorem 4.4.16.
Proof of Theorem 4.4.16. By Lemma 4.4.17, Part 4, RN Xj R2
is an operator
of type 0, hence bounded on Lp (G), p (1, ). The transpose of this operator is
(RN Xj R2
N t
) = R
2
N ,
Xj R
4.4. Sobolev spaces on graded Lie groups
235
By duality, this operator is Lp bounded where
since Xjt = Xj and Rt = R.
1
1
p + p = 1. As R is also a positive Rockland operator, see Lemma 4.1.11, we
Hence we have obtained that the operators
can exchange the r
ole of R and R.
R N Xj R2
and R2
Xj RN are bounded on Lp (G) for any p (1, )
and N N. This shows that Xj maps L pj +N to L pN and L pN to L pj N
continuously. The properties of interpolation, cf. Proposition 4.4.15, imply that
Xj maps L pj +s to L ps continuously for any s R, p (1, ) and j = 1, . . . , n.
Interpreting any X as a composition of operators Xj shows Part (2) for any
T = X , Nn0 , with T = []. As any T homogeneous leftinvariant dierential
operator is a linear combination of X , Nn0 , with T = [], this shows Part
(2).
Let us show Part (1). Let Nn0 . If s > 0, then by Theorem 4.4.3 (4) and
Part (2), we have for any S(G)
X Lps
X Lp + X L ps
L p + L p
[]
s+[]
Lp[] + Lps+[]
Lps+[] .
This shows that X maps Lps+[] to Lps continuously for any s > 0, p (1, ) and
any Nn0 . The transpose (X )t of X is a linear combination of X , [] = [],
and will also have the same properties. By duality, this shows that X maps Lps
to Lp(s+[]) continuously for any s > 0, p (1, ) and any Nn0 . Together
with the properties of interpolation (cf. Theorem 4.4.9), this shows that X maps
Lps+[] to Lps continuously for any s R, p (1, ) and any Nn0 .
As any left invariant dierential operator can be written as a linear combination of monomials X , this implies Part (1) and concludes the proof of Theorem
4.4.16.
The ideas of the proofs above can be adapted to the proof of the following
properties for the operators of type 0:
Theorem 4.4.18. Let T be an operator of type C on a graded Lie group G
with Re = 0. Then for every p (1, ) and s R, T maps continuously Lps (G)
to Lps (G) and L ps (G) to L ps (G). Fixing a positive Rockland operator R in order to
dene the Sobolev norms, it means that there exists C = Cs,p,T > 0 satisfying
S(G)
and
L ps
T Lps (G) CLps (G)
T L ps (G) CL ps (G) .
Proof. Let T be a operator of type T C with Re T = 0. Proceeding as in the
proof of Lemma 4.4.17 Part 3 yields that for any N N, the operator RN T RN
2
236
Chapter 4. Rockland operators and Sobolev spaces
is of type T . We can apply this to the transpose T t of T as well as the operator
and RN T t RN
T t is also of type . By Theorem 3.2.30, the operators RN T RN
2
2
p
p
p
are bounded on L (G). This shows that T maps Ls to Ls continuously for s = N
and s = N , N N0 . By interpolation, this holds for any s R and this shows
the statement for the homogeneous Sobolev spaces. If s > 0, then by Theorem
4.4.3 (4), using the continuity on homogeneous Sobolev spaces which has just
been proven, we have for any S(G)
T Lps T Lp + T L ps Lp + L ps Lps .
This shows that T maps Lps to Lps continuously for any s > 0, p (1, ). Applying
this to T t , by duality, we also obtain this property for s < 0. The case s = 0 follows
from Theorem 3.2.30. This concludes the proof of Theorem 4.4.18.
Theorem 4.4.18 extends the result of Theorem 3.2.30, that is, the boundedness on Lp (G) of an operator of type T , Re T = 0, from Lp spaces to Sobolev
spaces. Let us comment on similar results in related contexts:
In the case of Rn (and similarly for compact Lie groups), the continuity on
Sobolev spaces would be easy since T would commute with the Laplace
operator but the homogeneous setting requires a more substantial argument.
Theorem 4.4.18 was shown by Folland in [Fol75, Theorem 4.9] on any stratied Lie group and for = 0. However, the proof in that context uses the
existence of a positive Rockland operator with a unique homogeneous fundamental solution, namely the (any) sublaplacian. If we wanted to follow
closely the same line of arguments, we would have to assume that the group is
equipped with a Rockland operator with homogeneous degree with < Q,
see Remark 4.3.12. This is not always the case for a graded Lie group as
the example of the three dimensional Heisenberg group with gradation (3.1)
shows.
The proof above is valid under no restriction in the graded case. Somehow
the use of the homogeneous fundamental solution in the stratied case is
replaced by the kernel of the Riesz potentials together with the properties of
the Sobolev spaces proved so far.
4.4.5 Independence in Rockland operators and integer orders
In this Section, we show that the homogeneous and inhomogeneous Sobolev spaces
do not depend on a particular choice of a Rockland operator. Consequently Theorems 4.4.3, 4.4.9, 4.4.16, and 4.4.18, Corollaries 4.4.6 and 4.4.10, Propositions
4.4.8 and 4.4.13 and 4.4.15, hold independently of any chosen Rockland operator
R.
We will need the following property:
4.4. Sobolev spaces on graded Lie groups
237
Lemma 4.4.19. Let R be a Rockland operator on G of homogeneous degree and let
N0 , p (1, ). Then the space Lp (G) is the collection of functions f Lp (G)
such that X f Lp (G) for any Nn0 with [] = . Moreover, the map
X p
[]=
is a norm on L p (G) which is equivalent to the homogeneous Sobolev norm and
the map
X p
p +
[]=
is a norm on Lp (G) which is equivalent to the Sobolev norm.
Proof of Lemma 4.4.19. Writing
R =
c, X
[]=
we have on one hand,
S(G)
R p max c, 
X p .
(4.43)
[]=
On the other hand, by Theorem 4.4.16 (2), for any Nn0 , the operator X maps
continuously L p[] (G) to L p (G), hence
C > 0
S(G)
X p CL p .
[]
[]=
This shows the property of Lemma 4.4.19 for homogeneous Sobolev spaces.
Adding Lp on both sides of (4.43) implies by Theorem 4.4.3, Part (2):
C > 0 S(G)
Lp C Lp +
X p .
[]=
On the other hand, by Theorem 4.4.16 (1), for any Nn0 , the operator X maps
continuously Lp[] (G) to Lp (G), hence
C > 0
S(G)
X p CLp[] .
[]=
This shows the property of Lemma 4.4.19 for inhomogeneous Sobolev spaces and
concludes the proof of Lemma 4.4.19.
238
Chapter 4. Rockland operators and Sobolev spaces
One may wonder whether Lemma 4.4.19 would be true not only for integer
exponents of the form s = but for any integer s. In fact other inhomogeneous
Sobolev spaces on a graded Lie group were dened by Goodman in [Goo76, Section
III. 5.4] following this idea. More precisely the Lp GoodmanSobolev space of order
s N0 is given via the norm
X p
(4.44)
[]s
Goodmans denition does not use Rockland operators but makes sense only for
integer exponents.
The Lp GoodmanSobolev space of integer order s certainly contains Lps (G).
Indeed, proceeding almost as in the proof of Lemma 4.4.19, using Theorem 4.4.16
and Theorem 4.4.3, we have
s N0 C = Cs > 0 S(G)
X p CLps .
[]s
In fact, adapting the rest of the proof of Lemma 4.4.19, one could show easily
that the Lp GoodmanSobolev space of order s N0 with s proportional to the
homogeneous degree of a positive Rockland operator coincides with our Sobolev
spaces Lps (G). Moreover, on any stratied Lie group, for any nonnegative integer
s without further restriction, they would coincide as well, see [Fol75, Theorem
4.10].
However, this equality between GoodmanSobolev spaces and our Sobolev
spaces is not true on any general graded Lie group. For instance this does not hold
on a graded Lie groups whose weights are all strictly greater than 1. Indeed the
Lp GoodmanSobolev space of order s = 1 is Lp (G) which contains Lp1 (G) strictly
(see Theorem 4.4.3 (4)). An example of such a graded Lie group was given by the
gradation of the three dimensional Heisenberg group via (3.1).
We can now show the main result of this section, that is, that the Sobolev
spaces on graded Lie groups are independent of the chosen positive Rockland
operators.
Theorem 4.4.20. Let G be a graded Lie group and p (1, ). The homogeneous
Lp Sobolev spaces on G associated with any positive Rockland operators coincide.
The inhomogeneous Lp Sobolev spaces on G associated with any positive Rockland
operators coincide. Moreover, in the homogeneous and inhomogeneous cases, the
Sobolev norms associated to two positive Rockland operators are equivalent.
Proof of Theorem 4.4.20. Positive Rockland operators always exist, see Remark
4.2.4 Let R1 and R2 be two positive Rockland operators on G of homogeneous
degrees 1 and 2 , respectively. By Lemma 4.2.5, R12 and R21 are two positive
Rockland operators with the same homogeneous degree = 1 2 . Their associated
homogeneous (respectively inhomogeneous) Sobolev spaces of exponent = 1 2
4.4. Sobolev spaces on graded Lie groups
239
for any N0 coincide and have equivalent norms by Lemma 4.4.19. By interpolation (see Proposition 4.4.15, respectively Theorem 4.4.9), this is true for any
Sobolev spaces of exponent s 0, and by duality (see Lemma 4.4.14, respectively
Lemma 4.4.7) for any exponent s R.
Corollary 4.4.21. Let R(1) and R(2) be two positive Rockland operators on a graded
Then sfor any s s
Lie group G with degrees of homogeneity 1s and 2 , respectively.
s
C and p (1, ), the operators (I + R(1) ) 1 (I + R(2) ) 2 and (R(1) ) 1 (R(2) ) 2
extend boundedly on Lp (G).
Proof of Corollary 4.4.21. Let us prove the inhomogeneous case rst. For any a
a
(2)
R, we view the operator (I+Rp ) 2 as a bounded operator from Lp (G) to Lpa (G)
a
(1)
and use the norm f (I + Rp ) 1 f p on Lpa (G). This shows that the operator
s
s
(I + R(1) ) 1 (I + R(2) ) 2 is bounded on Lp (G), p (1, ) for s = a R. The
case of s C follows from Proposition 4.3.7.
Let us prove the homogeneous case. For any a R, we view the operaa
(2)
tor (Rp ) 2 as a bounded operator from Lp (G) to L pa (G) and use the norm
s
s
(1) a
f (Rp ) 1 f p on L pa (G). This shows that the operator (R(1) ) 1 (R(2) ) 2 is
bounded on Lp (G), p (1, ) for s = a R. The case of s C follows from
Proposition 4.3.9.
Thanks to Theorem 4.4.20, we can now improve our duality result given in
Lemmata 4.4.7 and 4.4.14:
Proposition 4.4.22. Let Lps (G) and L ps (G), p (1, ) and s R, be the inhomogeneous and homogeneous Sobolev spaces on a graded Lie group G, respectively.
For any s R and p (1, ), the dual space of Lps (G) is isomorphic to
p
Ls (G) via the distributional duality, and the dual space of L ps (G) is isomorphic
to L ps (G) via the distributional duality. Here p is the conjugate exponent of p if
p (1, ), i.e. p1 + p1 = 1. Consequently the Banach spaces Lps (G) and L ps (G) are
reexive.
4.4.6 Sobolev embeddings
In this section, we show local embeddings between the (inhomogeneous) Sobolev
spaces and their Euclidean counterparts, and global embeddings in the form of an
analogue of the classical fractional integration theorems of HardyLittlewood and
Sobolev.
Local results
Recalling that G has a local topological structure of Rn , one can wonder what
is the relation between our Sobolev spaces Lps (G) and their Euclidean counterparts Lps (Rn ). The latter can also be seen as Sobolev spaces associated by the
240
Chapter 4. Rockland operators and Sobolev spaces
described construction to the abelian group (Rn , +), with Rockland operator being the Laplacian on Rn .
By Proposition 3.1.28 the coecients of vector elds Xj with respect to
the abelian derivatives xk are polynomials in the coordinate functions x , and
conversely the coecients of xj s with respect to derivatives Xk are polynomials
in the coordinate functions x s. Hence, we can not expect any global embeddings
between Lps (G) and Lps (Rn ).
It is convenient to dene the local Sobolev spaces for s R and p (1, )
as
Lps,loc (G) := {f D (G) : f Lps (G) for all D(G)}.
(4.45)
The following proposition shows that Lps,loc (G) contains Lps (G).
Proposition 4.4.23. For any D(G), p (1, ) and s R, the operator f f
dened for f S(G) extends continuously into a bounded map from Lps (G) to itself.
Consequently, we have
Lps (G) Lps,loc (G).
Proof. The Leibniz rule for the Xj s and the continuous inclusions in Theorem
4.4.3 (4) imply easily that for any xed Nn0 there exist a constant C = C, > 0
> 0 such that
and a constant C = C,
f D(G) X (f )p C
X f p C f Lp[] (G) .
[][]
Lemma 4.4.19 yields the existence of a constant C = C,
> 0 such that
f D(G)
f Lp (G) C f Lp (G)
for any integer N0 and any degree of homogeneity of a Rockland operator.
This shows the statement for the case s = . The case s > 0 follows by
interpolation (see Theorem 4.4.9), and the case s < 0 by duality (see Proposition
4.4.22).
We can now compare locally the Sobolev spaces on graded Lie groups and
on their abelian counterpart:
Theorem 4.4.24 (Local Sobolev embeddings). For any p (1, ) and s R,
Lps/1 ,loc (Rn ) Lps,loc (G) Lps/n ,loc (Rn ).
Above, Lps,loc (Rn ) denotes the usual local Sobolev spaces, or equivalently the
spaces dened by (4.45) in the case of the abelian (graded) Lie group (Rn , +).
Recall that 1 and n are respectively the smallest and the largest weights of the
dilations. In particular, in the stratied case, 1 = 1 and n coincides with the
number of steps in the stratication, and with the step of the nilpotent Lie group
G. Hence in the stratied case we recover Theorem 4.16 in [Fol75].
4.4. Sobolev spaces on graded Lie groups
241
Proof of Theorem 4.4.24. It suces to show that the mapping f f dened
on D(G) extends boundedly from Lps/1 (Rn ) to Lps (G) and from Lps (G) to
Lps/n ,loc (Rn ). By duality and interpolation (see Theorem 4.4.9 and Proposition
4.4.22), it suces to show this for a sequence of increasing positive integers s.
For the Lps/1 (Rn ) Lps (G) case, we assume that s is divisible by the homogeneous degree of a positive Rockland operator. Then we use Lemma 4.4.19,
the fact that the X may be written as a combination of the x with polynomial
coecients in the x s and that max[]s  = s/1 .
For the case of Lps (G) Lps/n ,loc (Rn ), we use the fact that the abelian
derivative x ,  s, may be written as a combination over the X ,  s, with
polynomial coecients in the x s, that X maps Lp Lp[] boundedly together
with maxs [] = sn .
Proceeding as in [Fol75, p.192], one can convince oneself that Theorem 4.4.24
can not be improved.
Global results
In this section, we show the analogue of the classical fractional integration theorems of HardyLittlewood and Sobolev. The stratied case was proved by Folland
in [Fol75] (mainly Theorem 4.17 therein).
Theorem 4.4.25 (Sobolev embeddings). Let G be a graded Lie group with homogeneous dimension Q.
(i) If 1 < p < q < and a, b R with
1 1
b a = Q( )
p q
then we have the continuous inclusion
Lpb Lqa ,
that is, for every f Lpb , we have f Lqa and there exists a constant C =
Ca,b,p,q,G > 0 independent of f such that
f Lqa Cf Lpb .
(ii) If p (1, ) and
s > Q/p
then we have the inclusion
Lps (C(G) L (G)) ,
242
Chapter 4. Rockland operators and Sobolev spaces
in the sense that any function f Lps (G) admits a bounded continuous representative on G (still denoted by f ). Furthermore, there exists a constant
C = Cs,p,G > 0 independent of f such that
f Cf Lps (G) .
Proof. Let us rst prove Part (i). We x a positive Rockland operator R of homogeneous degree and we assume that b > a and p, q (1, ) satisfy ba = Q( p1 1q ).
By Proposition 4.4.13 (5),
a
Rq Lq CRp Lp .
We can apply this to (a, b) and to (0, b a). Adding the two corresponding estimates, we obtain
ba
a
b
Lq + Rq Lq C Rp Lp + Rp Lp .
Since b, a, and b a are positive, by Theorem 4.4.3 (4), the lefthand side is equivalent to Lqa and both terms in the righthand side are CLpb . Therefore,
we have obtained that
C = Ca,b,p,q,R
S(G)
Lqa CLpb .
By density of S(G) in the Sobolev spaces, this shows Part (i).
Let us prove Part (ii). Let p (1, ) and s > Q/p. By Corollary 4.3.13, we
know that
Bs L1 (G) Lp (G),
where p is the conjugate exponent of p. For any f Lps (G), we have
fs := (I + Rp ) f Lp
s
and
f = (I + Rp ) fs = fs Bs .
s
Therefore, by Holders inequality,
f fs p Bs p = Bs p f Lps .
Moreover, for almost every x, we have
fs (y)Bs (y 1 x)dy =
fs (xz 1 )Bs (z)dz.
f (x) =
G
Thus for almost every x, x , we have
1
1
fs (xz ) fs (x z ) Bs (z)dz
f (x) f (x ) =
G
Bs p fs (x ) fs (x )p .
4.4. Sobolev spaces on graded Lie groups
243
As the left regular representation is continuous (see Example 1.1.2) we have
fs (x ) fs (x )Lp (G) x x 0,
thus almost surely
f (x) f (x ) x x 0.
Hence we can modify f so that it becomes a continuous function. This concludes
the proof.
From the Sobolev embeddings (Theorem 4.4.25 (ii)) and the description of
Sobolev spaces with integer exponent (Lemma 4.4.19) the following property follows easily:
Corollary 4.4.26. Let G be a graded Lie group, p (1, ) and s N. We assume
that s is proportional to the homogeneous degree of a positive Rockland operator,
that is, s N, and that s > Q/p.
Then if f is a distribution on G such that f Lp (G) and X f Lp (G) when
Nn0 satises [] = s, then f admits a bounded continuous representative (still
denoted by f ). Furthermore, there exists a constant C = Cs,p,G > 0 independent
of f such that
X f p .
f C f p +
[]=s
The Sobolev embeddings, especially Corollary 4.4.26, enables us to dene
Schwartz seminorms not only in terms of the supremum norm, but also in terms
of any Lp norms:
Proposition 4.4.27. Let   be a homogeneous norm on a graded Lie group G. For
any p [1, ], a > 0 and k N0 , the mapping
S(G) S,a,k,p :=
(1 +  )a X p
[]k
is a continuous seminorm on the Frechet space S(G).
Moreover, let us x p [1, ] and two sequences {kj }jN , {aj }jN , of nonnegative integers and positive numbers, respectively, which go to innity. Then the
family of seminorms S,aj ,kj ,p , j N, yields the usual topology on S(G).
Proof of Proposition 4.4.27. One can check easily that the property
1 p, q , a > 0, k N0 , a > 0, k N0 , C > 0,
S,a,k,p S,a ,k ,q ,
(4.46)
is a consequence of the following observations (applied to X instead of ):
244
Chapter 4. Rockland operators and Sobolev spaces
1. If p and q are nite, by H
olders inequality, we have
(1 +  )a p C(1 +  )a q
where C is a nite constant of the group G, p and q. In fact C is explicitly
given by
C = (1 +  )
Q+1
r
r =
B(0, 1)
(1 + )
(Q+1) Q1
r1
d
,
with r (1, ) such that
1
p
1
q
+ 1r .
2. If p is nite and q = , we also have
(1 +  )a p C(1 +  )a+Q+1
where C = (1 +  )Q1 p is a nite constant.
3. In the case q is nite and p = , let us prove that
(1 +  )a Cs,p
(1 +  )a X p .
(4.47)
[]s
Indeed rst we notice that, by equivalence of the homogeneous quasinorms
(see Proposition 3.1.35), we may assume that the quasinorm is smooth away
from 0. We x a function D(G) such that
1 if x 1,
(x) =
0 if x 2.
We have easily
(1 +  )a C ( + (1 ) a ) .
(4.48)
By Corollary 4.4.26, there exist an integer s N such that
Cs,p
X ()p .
[]s
By the Leibniz rule (which is valid for any vector eld) and H
olders inequality, we have
X ()p C
X 1 X 2 p
[1 ]+[2 ][]
C,p
[1 ]+[2 ][]
X 1 p X 2 .
4.4. Sobolev spaces on graded Lie groups
245
Hence
Cs,p,
X p .
(4.49)
[]s
Following the same line of arguments, we have
(1 ) a Cs,p
X ((1 ) a )p
Cs,p
[]s
X
X 2 {(1 ) a }p
[1 ]+[2 ]s
Cs,p
(1 +  )a X 1 p (1 +  )a X 2 {(1 ) a } .
[1 ]+[2 ]s
All the norms above are nite since X 2 {(1 ) a }(x) = 0 if x 1
and for x 1,
X 2 {(1 ) a }(x) C2
X 3 (1 )(x) X 4  a (x)
[3 ]+[4 ]=[2 ]
C 2
X 3 (1 ) xa[4 ] ,
[3 ]+[4 ]=[2 ]
since X 4  a is a homogeneous function of degree a [4 ]. Hence we have
obtained
(1 +  )a X p .
(1 ) a Cs,p,
[]s
Together with (4.48) and (4.49), this shows (4.47).
4. If p = q is nite or innite, (4.46) is trivial.
Hence Property (4.46) holds. We also have directly for p = q [1, ] and
any 0 < a a , k k ,
S,a,k,p S,a ,k ,p .
Consequently we can assume a to be an integer in (4.46). This clearly implies that
any family of seminorms S,aj ,kj ,p , j N, yields the same topology as the family
of seminorms S,N,N, , N N. The latter is easily equivalent to the topology
given by the family of seminorms S(G),N dened in Section 3.1.9. This is the
usual topology on S(G).
4.4.7 List of properties for the Sobolev spaces
In this section, we list the important properties of Sobolev spaces we have already
obtained and also give some easy consequences regarding the special case of p = 2.
246
Chapter 4. Rockland operators and Sobolev spaces
Theorem 4.4.28. Let G be a graded Lie group with homogeneous dimension Q.
1. Let p (1, ) and s R. The inhomogeneous Sobolev space Lps (G) is a
Banach space satisfying
S(G) Lps (G) S (G).
The homogeneous Sobolev space L ps (G) is a Banach space satisfying
p
(S(G) Dom(Rs/
p )) Ls (G) S (G).
Norms on the Banach spaces Lps (G) and L ps (G) are given respectively by
s
(I + Rp ) Lp (G)
and
Rp Lp (G) ,
for any positive Rockland operator R (whose homogeneous degree is denoted
by ). All these homogeneous norms are equivalent, all these inhomogeneous
norms are equivalent.
The continuous inclusions Lpa (G) Lpb (G) holds for any a b and
p (1, ).
2. If s = 0 and p (1, ), then L p0 (G) = Lp0 (G) = Lp (G) with L p (G) =
0
Lp0 (G) = Lp (G) .
3. If s > 0 and p (1, ), then we have
Lps (G) = L ps (G) Lp (G),
and the inhomogeneous Sobolev norm (associated with a positive Rockland
operator) is equivalent to
Lps (G)
Lp (G) + L ps (G) .
4. If T is a leftinvariant dierential operator of homogeneous degree T , then
T maps continuously Lps+T (G) to Lps (G) for every s R, p (1, ).
If T is a T homogeneous leftinvariant dierential operator, then T
maps continuously L ps+T (G) to L ps (G) for every s R, p (1, ).
5. If 1 < p < q < and a, b R with b a = Q( p1 1q ), then we have the
continuous inclusions
L pb L qa
and
Lpb Lqa .
If p (1, ) and s > Q/p then we have the following inclusion:
Lps (C(G) L (G)) ,
4.4. Sobolev spaces on graded Lie groups
247
in the sense that any function f Lps (G) admits a bounded continuous representative on G (still denoted by f ). Furthermore, there exists a constant
C = Cs,p,G > 0 independent of f such that
f Cf Lps (G) .
6. For p (1, ) and any a, b, c R with a < c < b, there exists a positive
constant C = Ca,b,c such that we have for any f L pb
1
f L pc Cf L
p
p f L
a
and for any f
Lpb
f Lpc Cf 1
f Lp
Lp
a
b
where := (c a)/(b a).
7. (GagliardoNirenberg inequality) If q, r (1, ) and 0 < < s then there
exists C > 0 such that we have
f Lq (G) L rs (G)
where := 1
1
f L p Cf Lq f L
r ,
s
and p (1, ) is given via
1
p
1
r .
8. Let s be an integer which is proportional to the homogeneous degree of a
positive Rockland operator. Let p (1, ). Let f S (G).
The membership of f in Lps (G) is equivalent to f Lp (G) and X f
Lp (G), Nn0 , [] = s. Furthermore
X p
p +
[]=
is a norm on the Banach space Lps (G).
The membership of f in L ps (G) is equivalent to X f Lp (G), Nn0 ,
[] = s. Furthermore
X p
[]=
is a norm on the Banach space L ps (G).
9. (Interpolation) The inhomogeneous and homogeneous Sobolev spaces satisfy
the properties of interpolation in the sense of Theorem 4.4.9 and Proposition
4.4.15 respectively.
10. (Duality) Let s R. Let p (1, ) and p its conjugate exponent. The dual
space of L ps (G) is isomorphic to L ps (G) via the distributional duality, and the
dual space of Lps (G) is isomorphic to L ps (G) via the distributional duality,
Consequently, the Banach spaces Lps (G) and L ps (G) are reexive.
248
Chapter 4. Rockland operators and Sobolev spaces
Proof. Parts (1), (2), (3), and (6) follow from Theorem 4.4.3, Proposition 4.4.13
and Theorem 4.4.20.
Part (4) follows from Theorem 4.4.16 and Proposition 4.4.13.
Part (5) follows from Theorem 4.4.25 and Proposition 4.4.13 (5).
Part (7) follows from Parts (5) and (6).
Part (8) follows from Theorem 4.4.20.
For Part (9), see Theorem 4.4.9 and Proposition 4.4.15.
Part (10) follows from Lemmata 4.4.7 and 4.4.14 together with Theorem
4.4.20.
Properties of L2s (G)
Here we discuss some special feature of the case Lp (G), p = 2. Indeed L2 (G) is
a Hilbert space where one can use the spectral analysis of a positive Rockland
operator.
Many of the proofs in Chapter 4 could be simplied if we had restricted
the study to the case Lp with p = 2. For instance, let us consider a positive
Rockland operator R and its selfadjoint extension R2 on L2 (G). One can dene
the fractional powers of R2 and I + R2 by functional analysis. Then one can
obtain the properties of the kernels of the Riesz and Bessel potentials with similar
methods as in Corollary 4.3.11.
In this case, one would not need to use the general theory of fractional powers
of an operator recalled in Section A.3. Even if it is not useful, let us mention that
the proof that R2 satises the hypotheses of Theorem A.3.4 is easy in this case:
it follows directly from the LumerPhillips Theorem (see Theorem A.2.5) together
with the heat semigroup {etR2 }t>0 being an L2 (G)contraction semigroup by
functional analysis.
The proof of the properties of the associated Sobolev spaces L2s (G) would
be the same in this particular case, maybe slightly helped occasionally by the
Holder inequality being replaced by the CauchySchwartz inequality. A noticeable
exception is that Lemma 4.4.19 can be obtained directly in the case Lp , p = 2,
from the estimates due to Heler and Nourrigat (see Corollary 4.1.14).
The main dierence between L2 and Lp Sobolev spaces is the structure of
Hilbert spaces of L2s (G) whereas the other Sobolev spaces Lps (G) are only Banach
spaces:
Proposition 4.4.29 (Hilbert space L2s ). Let G be a graded Lie group.
For any s R, L2s (G) is a Hilbert space with the inner product given by
s
s
(I + R2 ) f (x) (I + R2 ) g(x)dx,
(f, g)L2s (G) :=
G
and
L 2s (G)
is a Hilbert space with the inner product given by
s
s
R2 f (x) R2 g(x)dx,
(f, g)L 2 (G) :=
s
4.4. Sobolev spaces on graded Lie groups
249
where R is a positive Rockland operator of homogeneous degree .
If s > 0, an equivalent inner product on L2s (G) is
s
s
(f, g)L2s (G) :=
f (x) g(x)dx +
R2 f (x) R2 g(x)dx.
G
If s = with N0 , an equivalent inner product on L2s (G) is
(f, g) = (f, g)L2 (G) +
(X f, X g)L2 (G) ,
[]=
and an equivalent inner product on L 2s (G) is
(f, g) =
(X f, X g)L2 (G) .
[]=
Proposition 4.4.29 is easily checked, using the structure of Hilbert space of
L2 (G) and, for the last property, simplifying the proof of Lemma 4.4.19.
4.4.8 Right invariant Rockland operators and Sobolev spaces
We could have started with rightinvariant (homogeneous) Rockland operators R
instead of R. We discuss here some links between the two operators and their
Sobolev spaces.
Since both left and right invariant Rockland operators are dierential op .
erators, we can relate them by Formulae (1.11) for the derivatives X and X
Then, given our analysis of R, we can give some immediate properties of the
rightinvariant operator R:
Proposition 4.4.30. Let R be a positive Rockland operator. For any S(G),
1
R(x)
= (Rt {(1 )})(x1 ) = (R{(
)})(x1 ),
Therefore, the spectral measure E
of R
is given by
because Rt = R.
1
E()(x)
= (E{(
)})(x1 ),
L2 (G), x G.
and R are linked by
Consequently, the multipliers of R
1
)})(x1 ).
m(R)()(x)
= (m(R){(
(4.50)
commute strongly, that is, their spectral measures E
The operators R and R
and E commute. Moreover, for functions f, g S (G) and a C, we have
Ra (f g)
a (f g)
R
(Ra f ) g
f Ra g,
a f ) g,
(R
a g.
f R
250
Chapter 4. Rockland operators and Sobolev spaces
We can give a rightinvariant version of Denition 4.3.17:
Denition 4.4.31. Let R be a positive Rockland operator of homogeneous degree
s/ f
and let s R. For any tempered distribution f S (G), we denote by (I + R)
the tempered distribution dened by
s/ ,
s/ f, := f, (I + R)
(I + R)
S(G).
The Sobolev spaces that we have introduced are based on the Sobolev spaces
corresponding to leftinvariant vector elds and leftinvariant positive Rockland
ps (G) dened via
operators. We could have considered the right Sobolev spaces L
the Sobolev norms
s/ f Lp .
f (I + R)
The relations between left and right vector elds in (1.11) easily implies that
if f Lp (G) is such that X f Lp (G) then f : x f (x1 ) is in Lp (G) and
f Lp (G). By Lemma 4.4.19, we see that the map f f must map
satises X
ps for any p (1, ) and s a multiple of the homogeneous
continuously Lps L
degrees of positive Rockland operators.
More generally, the spectral calculus, see (4.50), implies
2 )s/ f (x) = (I + R2 )s/ f(x1 ),
(I + R
f S(G),
where, again, f(x) = f (x1 ), and thus for any p (1, o ),
p )s/ f Lp (G) = (I + Rp )s/ fLp (G) ,
(I + R
f S(G).
p for any p (1, )
This easily implies that f f maps continuously Lps L
s
and any real exponent s R. This is also an involution: f = f . Hence the map
Lps (G)
f
ps (G)
L
f
is an isomorphism of vector spaces.
Even if the left and right Sobolev spaces are isomorphic, they are not equal in
general. Note that in the commutative case of G = Rn , both left and right Sobolev
spaces coincide. It is also the case on compact Lie groups, where the Sobolev spaces
are associated with the LaplaceBeltrami operator (which is central) and coincide
with localisation of the Euclidean Sobolev spaces [RT10a]. This is no longer the
case in the nilpotent setting. Indeed, below we give an example of functions f
ps (G).
(necessarily not symmetric, that is, f = f ), in some Lps (G) but not in L
Example 4.4.32. Let us consider the three dimensional Heisenberg group H1 and
the canonical basis X, Y, T of its Lie algebra (see Example 1.6.4). Then X =
= x + y t thus X
X = yt .
x y2 t whereas X
2
The Sobolev spaces are then associated with the natural subLaplacian X 2 +
Y 2 , see Example 6.1.1. Hence it is covered by the work of Folland [Fol75] on Sobolev
4.5. Hulanickis theorem
251
spaces associated with subLaplacian on stratied Lie groups and consequently,
L21 (G) is the space of functions f L2 (H1 ) such that Xf and Y f are both in
L2 (H1 ) [Fol75, Corollary 4.13].
One can nd a smooth function C (R) such that , L2 (R) but
2
z  (z)2 dz = . For instance, we consider = 1 where is a suitable
R
smoothing function (i.e. D(G) is valued in [0, 1] with a small support around
0), and the graph of the function 1 is given by isosceles triangles parametrised
by N, with vertex at points (, ), and base on the horizontal axis and with
length 2/ . We then choose , R with 2 (3, 1) and 2 > 2 + 1. We
also x a smooth function : R [0, 1] supported on [1/2, 2] with (1) = 1. We
dene f C (R3 ) via
yx
+ t (x)(t).
f (x, y, t) =
2
One checks easily that f , Xf and Y f are square integrable hence f L21 (H1 ).
X = yt f , this shows that (X +
However yt f is not square integrable. As X
2
2
2 (H1 ).
X)f
/ L (H1 ) and Xf can not be in L thus f is not in L
1
4.5
Hulanickis theorem
We now turn our attention to Hulanickis theorem which will be useful in the next
chapter when we deal with pseudodierential operators on graded Lie groups.
An important consequence of Hulanickis theorem is the fact that a Schwartz
multiplier in (the L2 selfadjoint extension of) a positive Rockland operator has
a Schwartz kernel. This section is devoted to the statement and the proof of
Hulanickis theorem and its consequence regarding Schwartz multiplier.
From now on, we will allow ourselves to keep the same notation R for a
positive Rockland operator and its selfadjoint extension R2 on L2 (G) when no
confusion is possible. In particular, when we dene functions of R2 (see Corollary
4.1.16), that is, a multiplier m(R2 ) dened using the spectral measure of R2 where
m L (R+ ) is a function, we may often write
m(R2 ) = m(R),
in order to ease the notation. Furthermore, we denote the corresponding rightconvolution kernel of this operator by
m(R)o .
4.5.1 Statement
Hulanicki proved in [Hul84] that if multipliers m satisfy Marcinkiewicz properties,
then the kernels of m(R) satisfy certain estimates:
252
Chapter 4. Rockland operators and Sobolev spaces
Theorem 4.5.1 (Hulanicki). Let R be a positive Rockland operator on a graded Lie
group G. Let  be a xed homogeneous quasinorm on G. For any M1 N, M2 0
N0 such that
there exist C = CM1 ,M2 > 0 and k = kM1 ,M2 N0 , k = kM
1 ,M2
k
for any m C [0, ), we have
X m(R)o (x) (1 + x)M2 dx C sup (1 + )  m(),
[]M1
>0
=0,...,k
=0,...,k
in the sense that if the righthand side is nite then the lefthand side is also nite
and the inequality holds.
The main consequence of Theorem 4.5.1 is the following:
Corollary 4.5.2. Let R be a positive Rockland operator on a graded Lie group G.
If S(R) then the kernel (R)o of (R) is Schwartz. Furthermore, the map
associating a multiplier function with its kernel
S(R) (R)o S(G),
(4.51)
is continuous between the Schwartz spaces.
The continuity of (4.51) means that for any continuous seminorm on
S(G) there exist C > 0 and N N such that for any m S(R) we have
m(R)o C
sup
xR,N
(1 + x)N m(x).
Examples of such Schwartz seminorms are S(G),N , N N, dened in Section
3.1.9, and S,a,k,p , a > 0, k N0 , p [1, ], dened in Proposition 4.4.27.
For completeness sake, we include the proofs of Theorem 4.5.1 and Corollary
4.5.2 below. Before this, let us notice that Corollary 4.5.2 implies that the heat
kernel of any Rockland operator is Schwartz. However, we will see that the proofs
of Theorem 4.5.1 and Corollary 4.5.2 rely on the properties of the Bessel potentials
which have been shown, in turn, using the properties of the heat kernel. Beside
the properties of the Bessel potentials, the proof uses the functional calculus of R
and the structure of G.
4.5.2 Proof of Hulanickis theorem
This section is devoted to the proof of Theorem 4.5.1 and can be skipped at rst
reading.
We follow the essence of [Hul84], but we modify the original proof to take
into account our presentation of the properties of Rockland operators as well as to
bring some (small) simplications. We also do not present some results obtained
in [Hul84] on groups of polynomial growth. One of these simplications is the fact
that we x a quasinorm   which we assume to be a norm. Indeed, it is clear
4.5. Hulanickis theorem
253
from the equivalence of quasinorms (see Proposition 3.1.35) that it suces to
prove Hulanickis theorem for one quasinorm for it to hold for any quasinorm.
As a homogeneous norm exists by Theorem 3.1.39, we may assume that   is
a norm without loss of generality. We could do without this but it simplies the
constants in the next pages.
First step
The rst step in the proof can be summarised with the following lemma:
Lemma 4.5.3. Let m : [0, +) C be a function and let o N. We dene the
function F : (, 1) C by
1
1
if 0 < < 1,
m o 1
F () :=
0
if 0,
and we have
[0, )
m() = F (1 + )o .
Furthermore, the following holds.
1. The function F extends to a continuous function on R if and only if m is
continuous on [0, ) and lim+ m() = 0.
2. The function F extends to a C 1 function on R if and only if m is C 1 on
[0, ) with lim+ m() = 0 and lim+ (1 + )1+o m () = 0.
Let k N. If m C k [0, +) and
lim (1 + )1+j+ko m(j) () = 0 for j = 1, . . . , k ,
then the function F extends to a function in C k (R)
3. Let k N and m C k [0, ). We assume that the suprema
sup(1 + )2+j+ko m(j) (),
0
j = 0, . . . , k.
are nite. Then we can construct an extension to R, still denoted by F , such
that the function F C k (R) is supported in [0, 2] and satises F (0) = 0 and
for every Z,
F () C(1 + )k sup (1 + )1+j+ko m(j) (),
0
j=0,...,k
where C = Ck,o is a positive constant independent of m. Here F (), Z,
denotes the Fourier coecients of F in the sense of
d
F () :=
F ()ei .
2
254
Chapter 4. Rockland operators and Sobolev spaces
Proof. Part (1) is easy to prove. Part (2) in the case of k = 1 follows easily from
the following observations.
If = (1 + )0 , > 0 then
F () F (0)
= (1 + )o m().
We can compute formally for (0, 1):
F () =
1 o
1
o +1
1
m o 1 ,
and in particular if = (1 + )0 , > 0, then
F () =
1
(1 + )1+o m ().
o
The general case of Part (2) follows from the following observation: F (k ) () is a
linear combination over j = 1, . . . , k of
1
1
1+j
1
1
o (k j)j( o +1) m(j) o 1 = o k m(j) o 1 .
The details are left to the reader.
Let us prove Part (3). Let m C k [0, ). Let Pk be the Taylor expansion
of m at 0, that is, Pk is the polynomial of degree k such that we have for > 0
small,
m() = Pk () + o(k ).
We x an arbitrary smooth function supported in [0, 2] and satisfying 1 on
[0, 1]. We construct an extension of F , still denoted F , by setting
0
if 0,
1o
m
1
if 0 < < 1,
F () :=
1
Pk o 1 () if 1.
We assume that the suprema given in the statement of Part 3 are nite. Clearly
F C k (R) is supported in [0, 2]. The proof of Part 2 implies easily
F
(k )
C
sup(1 + )1+j+o k m(j) (),
j=1 0
where the constant C = Ck ,o , > 0 is independent on m.
(4.52)
4.5. Hulanickis theorem
255
The Fourier coecient of F at 0 is
d
F (0) =
F ()
2
2
1
1
1
d
d
+
m( o 1)
Pk o 1 ()
=
2
2
0
1
2
d
o
d
1 o
m()
+
P
1
() .
=
k
+1
o
2
(1
+
)
2
0
1
We can always assume that the function was chosen so that
2
1
d
d
o
=
Pk o 1 ()
m()
.
2
2 (1 + )o +1
1
0
Indeed, it suces to replace by + c1 where 1 D(R) is supported in (1, 2)
and c a well chosen constant.
It is a simple exercise using integration by parts to show that the Fourier
coecients may be estimated by
k = 0, . . . , k
C = Ck > 0 Z
F () C(1 + )k F (k ) .
This together with (4.52) concludes the proof of Part (3).
Second step
The second step consists in noticing that, with the notation of Lemma 4.5.3,
studying the multiplier m(R) and using the Fourier series of F leads to consider the
o
and, more precisely, the properties of its convolution kernel.
operator ei(I+R)
Lemma 4.5.4. Let R be a positive Rockland operator on a graded Lie group G. Let
o N and Fo () := ei 1, R. Then, for any Z, the convolution kernel
of Fo ((I + R)o ) is an integrable function:
Fo ((I + R)o )o L1 (G).
Proof of Lemma 4.5.4. Since Fo () =
:=
=
j=1
(i)j
j! ,
we have at least formally
Fo ((I + R)o ) o
(i)j
(i)j
(I + R)jo 0 =
Bjo ,
j!
j!
j=1
j=1
where Ba is the convolution kernel of the Bessel potentials, see Section 4.3.4, and
is the degree of homogeneity of R. In fact, by Corollary 4.3.11, we know that
a C+
Ba L1 (G)
and
j
Bjo = Bo . . . Bo := B
.
o
256
Chapter 4. Rockland operators and Sobolev spaces
Thus in the Banach algebra L1 (G) endowed with the convolution product, the
series
(i)j j
(i)j
Bo =
Bjo ,
j!
j!
j=1
j=1
is convergent in the L1 norm. It is then a routine exercise to justify that is in
L1 (G) and is also the convolution kernel of the multiplier Fo ((I + R)o ).
Unfortunately the brute force estimate
Fo ((I + R)o )o L1 (G)
j
j=1
j!
Bo jL1 (G) = exp(Bo L1 (G) ) 1,
is exponential in and would be of no use for us. However, we notice that we can
already modify the proof above to show:
Lemma 4.5.5. We keep the notation of Lemma 4.5.4 and its proof. If   is a
homogeneous quasinorm on G, then for each Z and ao > 0, the function
Fo ((I + R)o )o is integrable against a weight of the form (1 +  )ao . Moreover,
if o > Q/2, where Q is the homogeneous dimension of G, then the function
Fo ((I + R)o )o is in L2 (G).
Proof of Lemma 4.5.4. One checks easily that if : G [1, ) is a continuous
function satisfying
C = C
x, y G
(xy) C(x)(y),
the subspace L1 (w) of L1 (G) of functions f which are integrable against w, is a
Banach algebra for the norm
f (x)(x)dx.
f L1 () =
G
Examples of such s are precisely weights of the form (1 +  )a with   being a
quasinorm on G. By Lemma 4.3.15, for any a C+ and ao > 0,
Ba L1 ((1 +  )ao ).
We keep the notation of the proof of Lemma 4.5.4 and proceed in the similar way
but using L1 ((1 +  )ao ) instead of L1 (G):
L1 ((1+)ao )
j
j=1
which is nite.
j!
Bo jL1 ((1+)ao ) = exp(Bo L1 ((1+)ao ) ) 1,
4.5. Hulanickis theorem
257
Let us now show that L2 (G). We have
2
j
j=1
and
(j1)
j
2 Bo 1 Bo
B
o
j!
j
B
2
o
2 . . . Bo j1
1 Bo 2 .
We also know by Corollary 4.3.11 that Ba L2 (G) whenever Re a > Q/2. Thus
in this case,
j
Bo j1
2
1 Bo 2 < ,
j!
j=1
nishing the proof.
Using only the properties of Banach algebras, we obtain again that
Fo ((I + R)o )o (1 +  )ao and
Fo ((I + R)o )o 2 ,
G
explode exponentially with which is not good enough for our subsequent analysis.
However, we are going to show that G Fo ((I+R)o )o (1+)ao actually grows
polynomially in (see Lemma 4.5.6) and this is the crucial technical point in the
proof of Theorem 4.5.1.
Main technical lemma
Lemma 4.5.6. We keep the notation of Lemmata 4.5.4 and 4.5.5. We x a homogeneous quasinorm   on G and assume that o > Q/2. Then for ao 0,
Q
(1 + x)ao Fo ((I + R)o )o (x)dx C3(ao + 2 +1) ,
(4.53)
G
where C = Cao ,o ,R,G, is a positive constant independent of Z.
In the proof of Lemma 4.5.6 we will need the following easy lemma:
Lemma 4.5.7. If 1 , . . . , 2m are 2m measures in M (G), then for any continuous
function ,
d1 . . . 2m
G
1
=
d 2 (y11 ) . . . 2m (ym
) d1 (y1 ) . . . d2m1 (ym ),
Gm
whenever the right or the left hand side is nite.
258
Chapter 4. Rockland operators and Sobolev spaces
Here we have denoted by (y ) the ylefttranslated measure of a positive or
complex Borel measure , that is, the measure given by
(x)d1 (yx) =
(y 1 x)d1 (x),
Cc (G).
xG
Proof of Lemma 4.5.7. First let us observe that if 1 , . . . , m are m measures in
M (G), then, recursively, one can show readily that
d(1 . . . m ) =
(x1 . . . xm ) d1 (x1 ) . . . dm (xm ).
(4.54)
G
Gm
If 1 , . . . , 2m are 2m measures in M (G), then applying (4.54) for these 2m measures yields
d(1 . . . 2m )
G
(x1 x2 . . . x2m1 x2m )
=
G2m
d1 (x1 ) d2 (x2 ) . . . d2m1 (x2m1 ) d2m (x2m )
=
G
Gm
(x2 x4 . . . x2m )
1
1
d2 (x1
1 x2 )d4 (x3 x4 ) . . . d2m (x2m1 x2m )
d1 (x1 ) . . . d2m1 (x2m1 ),
after the change of variables x2 = x1 x2 , . . . , x2m = x2m1 x2m . Using (4.54), we
recognise an iterated convolution in the quantity in parenthesis.
We now turn our attention to showing Lemma 4.5.6.
Proof of Lemma 4.5.6. By Theorem 3.1.39, we may assume that the homogeneous
quasinorm   is also a norm, that is, we assume that the triangle inequality holds
with a constant = 1. Moreover, we notice that it suces to prove (4.53) with
N. Indeed, as
:= Fo ((I + R)o )0 = ei(I+R)
we have
= ei(I+R)
o
o
0 0 ,
o )0 ,
0 0 = Fo ((I + R)
and 0 = 0. Since
Fo () = ei 1 = (ei ) 1 = (Fo () + 1) 1,
we have in the unital Banach algebra C0 L1 (G),
= (1 + o ) 0 .
4.5. Hulanickis theorem
259
Let us x N. We can decompose
1 = f 0 + f ,
where f0 and f are the integrable functions dened via
f0 (x) := 1 (x)1x2
and
f (x) := 1 (x)1x>2 .
(4.55)
We can also write
1 + 0 = 0 + f
where
0 := f0 + 0 .
We now develop the noncommutative convolution product in C0 L1 (G),
(1 + o )
=
=
(0 + f ) = (0 + f ) . . . (0 + f )
1
1
f
. . .
f
,
0
0
,
where the summation is over all sequences = (1 , . . . , ), = (1 , . . . , ), of 0
and 1 such that 1 + . . . + + 1 + . . . + = .
Let us x two such sequences and . By Lemma 4.5.7 applied to
2j1 = 0 j ,
and
j
2j = f
,
and = ao where the positive function C(G) is dened by
(x) := 1 + x,
x G,
we have for any ao > 0,
1
1
ao d
f
. . .
f
0
0
G
1
1
=
ao d
(y
)
.
.
.
(y
)
0
1
0
G
1
df
(y1 ) . . . df
(y ).
Let us also x (y1 , . . . , y ) G . We notice that for any M (G) we have
(0 (y )) = (y ),
since for any Cc (G),
d {(0 (y )) } =
(xz) d0 (yx) d(z) =
(y 1 x) d(x).
G
260
Chapter 4. Rockland operators and Sobolev spaces
Therefore, if 1 = 0,
1
2 2
1
(y
)
(y
)
0
1
0
2
=
=
=
2
2
0 (y11 )
(y
)
0
2
1
2
) (y11 )
0 (y2
1 1
2
(y
y
)
,
0
2
1
whereas if 1 = 1, we have
1
2
1
2
)
)
0 (y1
0 (y2
2
1
2 2
2
= 0 (y11 )
(y
)
+
f
(y
)
(y
)
0
0
2
1
0
2
1
2 1 1
2 2
= 0 (y2 y1 ) + f0 (y1 ) 0 (y2 ) .
Recursively, we nd that
1
1
) . . .
)
0 (y1
0 (y
=
(y
)
f0 (y11 . . . yj j ) 0 j+1 (yj+1j+1 ) . . .
0
j:j =1
when = 0. If = 0, we compute directly:
1
1
) . . .
)
=
0 (y1
0 (y
=
so that
o (y11 ) . . . o (y )
o y . . . y11 ,
1
1
ao d
(y
)
.
.
.
(y
)
0
1
0
G
a o
= ao (y11 . . . y ) = 1 + y11 . . . y 
a o
a o
j
1
1 + y1  + . . . + y 
1 + max yj 
,
j =1,...,
since the quasinorm   is assumed to be a norm.
If j = 0, we notice that the measure given by
f0 (y11 . . . yj j ) 0 j+1 (yj+1j+1 ) . . .
(y
)
,
0
(4.56)
2 where we denote
is compactly supported. Indeed recall that f0 is supported in B
R := {x G : x R} a closed ball of radius R about 0 for the chosen norm.
by B
Therefore
2 .
supp 0 {0} supp f0 B
4.5. Hulanickis theorem
261
The general properties
1 , 2 M (G)
supp(1 2 )
(supp1 )(supp2 )
= {x1 x2 : x1 supp1 , x2 supp2 },
and
supp ( y 1 ) = (supp)y = {xy : x supp},
M (G), y G
imply that the measure in (4.56) is supported in
2 y 1 . . . B
2 y . . . y j .
2 y j1 B
B
1
1
j1
j1
j
Using the properties of the norm  , it is easy to check that the above subset of
G is included in the closed ball about 0 of radius
(
j )2 + (
j =1
since
j =1
j yj ) 3 + max yj ,
j =1,...,
j and j =1 j are .
R then
Note that if f is a measurable function supported in B

ao (x)f (x)dx
(1 + x)ao f (x)dx
G
xR
(1 + R)ao
f (x)dx
j =1
xR
ao
R 1/2 f 2 .
(1 + R) B
We apply this to the function f in (4.56) and R = 3 + maxj =1,..., yj . This
leads us to look at the L2 norm of this function which can be written as
f = T (y ) . . . T j+1 j+1
f0 (y11 . . . yj j ) .
0
(yj+1
Here we have used our usual notation for the convolution operator T :
with rightconvolution (distributional) kernel . Since such convolution operators
are leftinvariant, we have
T
j
j
(yj
L (L2 (G)) = Tj L (L2 (G)) = T0 Lj(L2 (G)) .
Again by leftinvariance,
f0 (y1 1 . . . yj j )
L2 (G)
= f0 L2 (G) 1 L2 (G) ,
which is nite by Lemma 4.5.5 since we assume o > Q/2. Therefore,
j<j j
.
f 2 1 L2 (G) T0 L (L
2 (G))
262
Chapter 4. Rockland operators and Sobolev spaces
We have obtained that for any (y1 , . . . , y ) G , , with = 0,
ao d f0 (y 1 . . . y j ) j+1 (y j+1 ) . . . (y )
1
0
0
j
j+1
G
ao
3 + max
B
1/2
1 + 3 + max yj 
j =1,..., yj 
j =1,...,
j:j =1
j<j j
T0 L (L
1 L2 (G)
2 (G))
ao +Q/2
j<j j
C 1 + 3 + max yj 
T0 L (L
,
2 (G))
j =1,...,
1
1 1/2 . We now integrate against df
(y1 ) . . . df
(y ) to
with C = 1 L2 (G) B
obtain
1
ao d1 f
. . . 0 f
0
G
1 1
ao
d 0 (y1 ) . . . 0 (y )
G
1
df
(y1 ) . . . df
(y )
j<j j
CT0 L (L
2 (G))
ao +Q/2
1
1 + 3 + max yj 
df
(y1 ) . . . df
(y ).
j =1,...,
G
Let us estimate this last integral:
1 + + max yj 
ao +Q/2
G
j =1,...,
1 + 3 + (q + 1)
ao +Q/2
1
df
(y1 ) . . . df
(y )
1
df
(y1 ) . . . df
(y ).
qmaxj yj <q+1
q=0
For each of these integrals, we see that either
0 if = 0, q 1 or = 0, q < 2 ,
1
df
(y1 ) . . . df
(y ) =
1 if = 0, q = 0,
qmax y <q+1
j
or if = 0 and q 2 ,
1
1
df (y1 ) . . . df (y ) f 1
qmaxj yj <q+1
1
f 1
(1 + q)a
G
1  a ,
qyq+1
1 (y)dy
4.5. Hulanickis theorem
263
for any a > 0 that will be suitably chosen. Indeed, by Lemma 4.5.5, this last
integral is nite. Hence if = 0, then
ao +Q/2
1
1 + 3 + max yj 
df
(y1 ) . . . df
(y )
j =1,...,
G
1
f 1
ao +Q/2
1 
1 + 3 + (q + 1)
(1 + q)a .
a
q=2
We choose a := 2(ao + Q/2 + 2) so that the sum
1 + 3 + (q + 1)
ao +Q/2
q=2
(1 + 2)ao +Q/2
(1 + q)a (1 + 2)ao +Q/2
(q + 1)ao +Q/2a
q=2
xao +Q/2a dx (1 + 2)ao +Q/2
2(ao +Q/2a+1)
,
ao + Q/2 a + 1
is nite and bounded independently of . We set
a
+Q/2
o
1 + 3 + (q + 1)
Ca := max 1,
1  a max
(1 + q)a .
N
q=2
We have obtained in the case and both nonzero:
1
ao d1 f
. . . 0 f
0
G
1
j<j j
CCa
T0 L (L
f 1
2 (G))
j:j =1
1
CCa max(1, T0 L (L2 (G)) ) f 1
whereas in the case = 0 and = 0,
j<j j
1
ao d1 f
.
.
.
f
T
(1 + 3 )ao +Q/2
0 L (L2 (G))
0
0
G
j:j =1
C(1 + 3 )ao +Q/2 max(1, T0 L (L2 (G)) ) ,
and in the case =
0 and = 0,
1
ao d1 f
.
.
.
f
0
0
G
a o
1
df
(y1 ) . . . df
(y )
1 + max yj j 
G
j =1,...,
(1 + q)
ao
q=2
1
CCa f 1
1
f 1
G
1 
(1 + q)a
264
Chapter 4. Rockland operators and Sobolev spaces
We can now sum over and to obtain
1
  ao
CCa f 1
G
=0
C(1 + 3 )ao +Q/2 max(1, T0 L (L2 (G)) )
=0
1
CCa max(1, T0 L (L2 (G)) ) f 1
=0
=0
We need to estimate the operator norm of
T 0 = I + T 1 T f .
By functional calculus, the operator
I + T1 = I + Fo ((I + R)o ) = exp(i(I + R)o ),
has norm
I + T1 L (L2 (G))
exp(i(I + R)o )L (L2 (G))
sup  exp(i(1 + )o ) 1.
0
For Tf , we have
Tf L (L2 (G)) f L1 (G) .
Hence
To L (L2 (G)) I + T1 L (L2 (G)) + Tf L (L2 (G)) 1 + f L1 (G) ,
and
f 1 + max(1, T0 L (L2 (G)) ) 1 + 2f 1 .
Let us also estimate
f L1 (G)
1
dx
1 (x)(x)
1 + x
x2
1
1 .
1 + 2 G
By Lemma 4.5.5,
c :=
G
1 
(4.57)
4.5. Hulanickis theorem
265
is nite. Thus we have obtained so far that
1
  ao CCa
(1 + 2f 1 ) f 1
G
=0,
Q
+CCa (1 + 3 )ao + 2 +1
CCa
(1 + 2f 1 )
=0
1+
=0,
2c
1 + 2

Since
c
1 + 2
1

2c
1 + 2
=0


c
2c
3 ao + Q
+1
2
CCa (1 + )
1+
2
1+
1 + 2
,
3c
3 ao + Q
+1
2
= CCa (1 + )
.
1+
1 + 2
Q
+CCa (1 + 3 )ao + 2 +1
1+
max 1 +
3c
< ,
N
1 + 2
we have proved what we wanted, that is,
Q
  ao C (1 + 3 )ao + 2 +1 ,
G
completing the proof of Lemma 4.5.6.
The proof of Lemma 4.5.6 can be modied to obtain a similar property for
X o Fo ((I + R)o )o :
Lemma 4.5.8. We keep the notation of Lemmata 4.5.4, 4.5.5 and 4.5.6. Then for
any o Nn0 with o > [o ] + Q/2 and ao 0, we have
Q
(4.58)
(1 + x)ao X o Fo ((I + R)o )o (x)dx C3(ao + 2 +1) ,
G
where C = Co ,ao ,o ,R,G, is a positive constant independent of Z.
The proof of Lemma 4.5.8 follows the one of Lemma 4.5.6 but with the
derivatives X o now applied to the last term of the convolution products. These
convolution products have to be understood as convolutions between compactly
supported distributions and integrable functions since we replace the denition of
f0 and f given in (4.55) by
f0 := 1 (1   )
where
and
f := 1 (1 )(1   ),
266
Chapter 4. Rockland operators and Sobolev spaces
D(R) is a nonnegative smooth function satisfying (x) = 1 if x [1, 1]
and (x) = 0 if x > 2,
and   is a homogeneous norm which is smooth away from 0 (see Proposition 3.1.35 (i) or Remark 3.1.36).
We also need to replace = 1 +  ao with
1
if x 1,
(x) =
x if x 1,
in order to make sense of the distribution X o 0 with support at 0 being applied
to the function which is smooth around 0. The conditions on the parameters
given
in Lemma 4.5.8 and the following lemma ensure that the new quantities
X o 2 and G X o  a appearing in the proof are nite.
G
Lemma 4.5.9. Let R be a Rockland operator of homogeneous degree on a graded
Lie group G and let Ba , a C+ , be the kernels of its Bessel potentials.
1. If   is a homogeneous quasinorm on G, b 0 and Nn0 with Re a + b >
[], then
xb X Ba (x)dx < .
G
2. If Nn0 with Re a > [] + Q/2, then
X Ba (x)2 dx < .
G
Proof of Lemma 4.5.9. For the rst part, we generalise the rst part of Lemma
4.3.15, that is,
Re a
1
1 et
xb X Ba (x)dx
t
xb X ht (x)dx dt,
a
(
)
0
G
G
and using the homogeneity of the heat kernel (see (4.17)) and the change of vari1
ables y = t x, we get
[]
b[]
1
b
x X ht (x)dx =
t yb t X h1 (y)dy = cb, t ,
G
where cb, = yb X h1 (y)L1 (dy) is a nite constant since h1 S(G). Thus,
cb,
x X Ba (x)dx
( a )
G
b
is nite whenever Re a + b [] > 0.
b[]
Re a
1+
et dt
4.5. Hulanickis theorem
267
For the second part of Lemma 4.5.9, we can not adapt the proof of the square
with (X h)(x1 ). We proceed
integrability of Ba since we can not relate X h(x)
dierently. Let D(G). From the properties of the heat kernel and the denition
of Ba we have
Ba (x1 ) = Ba (x) = Ba (x)
and so
(X Ba )
Ba X
G
(X ) Ba (0)
(X ) Ba
s
C I + R2 (X ) Ba ,
2
by the Sobolev embeddings, see Theorem 4.4.25 for s > Q/2. But, since Ba is the
2 ) a
, we have
rightconvolution kernel of the Bessel potential (I + R
a
s
sRe
2 (X ) Ba
2
I
+
R
(X
)
I+R
=
2
2
sRea+[]
2
C I+R
,
2
by Theorem 4.4.16. Hence we have obtained
sRea+[]
2
X Ba CC I + R
G
C1 2 ,
2
if s Re a + [] 0, from the Sobolev inclusions, see Theorem 4.4.3 (4). Hence,
under this condition, X Ba L2 (G).
Proceeding as in the proof of Lemma 4.5.5, Lemma 4.5.9 yields
X o L1 ((1 +  )ao ) if o + ao > [o ],
and X o L2 (G) if o > [o ] + Q/2.
The details of the proof of Lemma 4.5.8 are left to the interested reader.
Last step
We can now conclude the proof of Theorem 4.5.1.
End of the proof of Theorem 4.5.1. We x ao N and o Nn0 . We consider
k N and o to be chosen in terms of ao and o . Let m C k [0, ) satisfying the
268
Chapter 4. Rockland operators and Sobolev spaces
hypotheses of Lemma 4.5.3 Part 3 for o and k, and let F be the corresponding
function in C k (R). Since supp F [0, 2], we may develop F in the Fourier series:
F ()ei =
F ()ei , (, ),
F () =
Z
Z\{0}
as F (0) = 0. Since
0 = F (0) =
F (),
Z\{0}
we also have for any [0, 2] that
F () =
F ()(ei 1).
Z\{0}
This yields
m() =
o
F ()(ei(1+)
1),
0,
Z\{0}
and at least formally
m(R)o =
F () ,
(4.59)
Z\{0}
o
where := ei(I+R) 0 0 as before.
By Lemma 4.5.8, if o > [o ] + Q/2, then
F () (1 + x)ao X o (x)dx
Z\{0}
F ()3(ao + 2 +1)
Q
Z\{0}
CCk sup (1 + )1+j+ko m(j) ()
R
j=0,...,k
(1 + )k 3(ao + 2 +1) ,
Z\{0}
see Lemma 4.5.3 for the estimates of F (). This last sum converges provided that
we have chosen k > 3(ao + Q
2 + 1) + 2. We assume that we have chosen such o
and k. One can now show easily that m(R)o L1 (G) and that (4.59) holds in
L1 (G). Furthermore, X o m(R)o L1 (G) and
ao
o
(1 + x) X m(R)0 (x)dx
F () (1 + x)ao X o (x)
G
Z\{0}
Ck sup (1 + )1+j+ko m(j) ().
This concludes the proof of Theorem 4.5.1.
0
j=0,...,k
4.5. Hulanickis theorem
4.5.3
269
Proof of Corollary 4.5.2
Let us show Corollary 4.5.2.
Proof of Corollary 4.5.2. Applying Theorem 4.5.1 to the restriction of m S(R)
to [0, ), we have for any and a > 0 that
m(R)0 S,a,[],1 C sup (1 + )1+j+ko m(j) (),
0
j=0,...,k
with k := 3(a + Q
2 + 1) + 3 and o the smallest integer such that o > [] + Q/2.
Clearly the righthand side of this inequality is less than a S(R)seminorm of m, up
to a constant depending on this seminorm. By Proposition 4.4.27, this concludes
the proof of Corollary 4.5.2.
We may simplify the proof of Corollary 4.5.2 by modifying the proof of Theorem 4.5.1 and choosing F independently of k for m S(R). Indeed, it suces to
set
1
1
m o 1 1 () if > 0,
F () :=
0
if 0,
where 1 D(R) is supported in [1, 2] and satises 1 1 on [0, 1] together
with F(0) = 0.
Remark 4.5.10. Behind this technical point lays the fact that the spectrum of R2
is contained in [0, ) thus we can modify any function m as we see t on (, 0)
without changing the operator m(R).
We had already used this idea in the proof of Theorem 4.5.1 indirectly, since
dierent extensions of the function F will lead to the same formula in (4.59).
Note that we may also modify the function m at 0, see Remark 4.2.8 (3), but
we did not use this point in the proofs of Theorem 4.5.1 or Corollary 4.5.2.
Open Access. This chapter is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/),
which permits use, duplication, adaptation, distribution and reproduction in any medium
or format, as long as you give appropriate credit to the original author(s) and the source,
a link is provided to the Creative Commons license and any changes made are indicated.
The images or other third party material in this chapter are included in the works
Creative Commons license, unless indicated otherwise in the credit line; if such material
is not included in the works Creative Commons license and the respective action is not
permitted by statutory regulation, users will need to obtain permission from the license
holder to duplicate, adapt or reproduce the material.
Chapter 5
Quantization on graded Lie
groups
In this chapter we develop the theory of pseudodierential operators on graded Lie
groups. Our approach relies on using positive Rockland operators, their fractional
powers and their associated Sobolev spaces studied in Chapter 4. As we have
pointed out in the introduction, the graded Lie groups then become the natural
setting for such analysis in the context of general nilpotent Lie groups.
m
and the corresponding operator classes
The introduced symbol classes S,
m
m
, = Op S, ,
for (, )with 1 0 and = 1, have an operator calculus, in the sense that
the set mR m
, forms an algebra of operators, stable under taking the adjoint,
and acting on the Sobolev spaces in such a way that the loss of derivatives is
controlled by the order of the operator. Moreover, the operators that are elliptic
or hypoelliptic within these classes allow for a parametrix construction whose
symbol can be obtained from the symbol of the original operator.
During the construction of the pseudodierential calculus mR m
, on
graded Lie groups in this chapter, there are several diculties one has to overcome and which do not appear in the case of compact Lie groups as described in
Chapter 2. The immediate one is the need to nd a natural framework for discussing the symbols to which we will be associating the operators (quantization)
and we will do so in Section 5.1. In Section 5.2 we dene symbol classes leading
to algebras of symbols and operators and discuss their properties. The symbol
classes that we introduce are based on a positive Rockland operator on the group
and contain all the leftinvariant dierential operators. As with Sobolev spaces,
the symbol classes can be shown to be actually independent of the choice of a
positive Rockland operator used in their denition. In Section 5.3 we show that
the multipliers of Rockland operators are in the introduced symbol classes. We
The Editor(s) (if applicable) and The Author(s) 2016
V. Fischer, M. Ruzhansky, Quantization on Nilpotent Lie Groups, Progress in Mathematics 314,
DOI 10.1007/9783319295589_5
271
272
Chapter 5. Quantization on graded Lie groups
investigate the behaviour of the kernels of operators corresponding to these symbols in Section 5.4, both at 0 and at innity and show, in particular, that they
are Calder
onZygmund (in the sense of Coifman and Weiss, see Sections 3.2.3 and
A.4). The symbolic calculus is established in Section 5.5. In Section 5.7 we show
that the operators satisfy an analogue of the CalderonVaillancourt theorem. The
construction of parametrices for elliptic and hypoelliptic operators in the calculus
is carried out in Section 5.8.
Conventions
Throughout Chapter 5, G is always a graded Lie group, endowed with a family
of dilations with integer weights. Its homogeneous dimension is denoted by Q.
Also throughout, R will be a homogeneous positive Rockland operator of homogeneous degree . If G is a stratied Lie group, we can choose R = L with L
a subLaplacian, or another homogeneous positive Rockland operator. Since it is
a leftinvariant dierential operator, we denote by (R) the operator described
in Denition 1.7.4. Both R and (R) and their properties have been extensively
discussed in Chapter 4, especially Section 4.1.
Finally, when we write
sup
G
we always understand it as the essential supremum with respect to the Plancherel
measure on G.
5.1
Symbols and quantization
The global quantization naturally occurs on any unimodular Lie (or locally compact) group of type 1 thanks to the Plancherel formula, see Subsection 1.8.2 for
the Plancherel formula. The quantization was rst noticed by Michael Taylor in
[Tay86, Section I.3]. The case of locally compact type 1 groups was studied recently in [MR15]. The case of the compact Lie groups was described in Section
2.2.1. Here we describe the particular case of graded nilpotent Lie groups, with an
emphasis on the technical meaning of the objects involved. A very brief outline of
the constructions of this chapter appeared in [FR14a].
Formally, for a family of operators (x, ) on H parametrised by x G and
G, we associate the operator T = Op() given by
Tr (x)(x, )()
d().
(5.1)
T (x) :=
G
Again formally, the Fourier inversion formula implies that if (x, ) does not depend on x and is the group Fourier transform of some function , i.e. if (x, ) =
(), then Op() is the convolution operator with rightconvolution kernel , i.e.
5.1. Symbols and quantization
273
Op() = . We would like this to be true not only for (say) integrable functions
but also for quite a large class of distributions, in order
to quantize X = Op() by (x, ) = (X) ,
with (X) as in Denition 1.7.4.
The rst problem is to make sense of the objects above. The dependence of
on x is not problematic for the interpretation in the formula (5.1), but we have
identied a unitary irreducible representation with its equivalence class and the
families of operators may be measurable in Rep G but not dened for all
More worryingly, we would like to consider collections of operators which
G.
For these reasons, it may be
are unbounded, for instance such as (X) , G.
dicult to give a meaning to the formula (5.1) in general.
Thus, our rst task is to dene a large class of collections of operators (x, ),
for which we can make sense of the quantization procedure. We
x G, G,
will use the realisations
and LL (L2 (G))
K(G), L (G),
of the von Neumann algebra of the group G described in Section 1.8.2. We will
also use their generalisations
2
2
Ka,b (G), L
a,b (G), and LL (La (G), Lb (G))
which we dene in Section 5.1.2. In order to do so we use a special feature of our
setting, namely the existence of positive Rockland operators and the corresponding
L2 Sobolev spaces.
5.1.1 Fourier transform on Sobolev spaces
In Section 4.3, we have discussed in detail the fractional powers of a positive
Rockland operator R and of the operator I + R. In the sequel, we will also need to
understand powers of the operators 1 (I + R), 1 Rep G. We now address this,
and use it to extend the group Fourier transform to the Sobolev spaces L2a (G).
From now on we will keep the same notation for the operators R and 1 (R)
(where 1 Rep (G)) and their respective selfadjoint extensions, see Proposition
4.1.15. We note that by Proposition 4.2.6 the operator 1 (R) is also positive. We
can consider the powers of I + R and 1 (I + R) = I + 1 (R) as dened by the
functional calculus
a
a
a
a
(1 + ) dE(), 1 (I + R) =
(1 + ) dE1 (),
(I + R) =
0
where E and E1 are the spectral measures of R and 1 (R), respectively, and
is the homogeneous degree of R, see Corollary 4.1.16.
274
Chapter 5. Quantization on graded Lie groups
Remark 5.1.1. If a/ is a positive integer, there is no conict of notation between
the powers of 1 (I+R) as the innitesimal representation of 1 (see Denition
1.7.4) at I + R U(g)
a
and the operator 1 (I + R) dened by functional calculus.
a
Indeed, if a = , the two coincide. If a = , N, then the operator 1 (I + R)
dened by functional calculus coincides with the th power of 1 (I + R). The case
a = 0 is trivial.
a
We can describe more concretely the operators 1 (I + R) , 1 Rep G.
Lemma 5.1.2. Let R be a positive Rockland operator of homogeneous degree . As
in Corollary 4.3.11, we denote by Ba the rightconvolution kernels of its Bessel
a
potentials (I + R) , Re a > 0.
If a C with Re a < 0, then Ba is an integrable function and
1 Rep G
a
1 (I + R) = Ba (1 ).
For any a C and any 1 Rep G, the operator 1 (I + R) maps H1
a
a
onto H1 bijectively. Furthermore, the inverse of 1 (I + R) is 1 (I + R) as
operators acting on H1 .
a
Proof. Let a C, Re a < 0. Then the Bessel potential (I + R) coincides with
the bounded operator with rightconvolution kernel Ba L1 (G), see Corollary
a
4.3.11. Therefore, (I + R) LL (L2 (G)) and
a
FG {(I + R) f } = FG {f Ba } = Ba f,
f L2 (G).
Now we apply Corollary 4.1.16 with the bounded multiplier given by () =
a
(1 + ) , 0. By Equality (4.5) in Corollary 4.1.16, we obtain
a
a
FG {(I + R) f } = (I + R) f,
f L2 (G).
The injectivity of the group Fourier transform on K(G) yields that Ba () =
a
and the rst part of the statement is proved.
(I + R) for any G,
Let a C. We apply Corollary 4.1.16 with the multiplier given by () =
a
(1+) , 0. Although this multiplier is unbounded, simple modications of the
proof show that Equality (4.5) in Corollary 4.1.16 still holds for f in the domain
a
of the operator. Recall that the domain of (I + R) contains S(G) by Corollary
a
4.3.16 and moreover (I + R) S(G) = S(G). Consequently, if 1 Rep G, we have
a
1 {(I + R) f }v = 1 (I + R) 1 (f )v,
f S(G), v H1 ,
with 1 (I + R) dened spectrally. Recall that 1 (f )v H1 when f S(G) by
a
Proposition 1.7.6 (iv), hence here 1 {(I+R) f }v H as well. By Lemma 1.8.19,
a
5.1. Symbols and quantization
275
1 (I+R) maps H1 to H1 . The spectral calculus implies that as operators acting
on H1 , we have
a
1 (I + R) 1 (I + R) = IH
1
a
and
1 (I + R) 1 (I + R) = IH
.
1
a
Consequently, the inverse of 1 (I + R) is 1 (I + R) as operators dened on
a
H1 and 1 (I + R) H1 = H1 .
a
Lemma 5.1.2 and Remark 4.1.17 now imply easily
Corollary 5.1.3. Let R be a positive Rockland operator of homogeneous degree .
a
is a measurable Geld
of
For any a C, {(I + R) : H H , G}
operators acting on smooth vectors (in the sense of Denition 1.8.14).
Lemma 5.1.2 together with the Plancherel formula (see Section 1.8.2) and
Corollary 4.3.11 also imply
Corollary 5.1.4. Let R be a positive Rockland operator of homogeneous degree .
For any a R, we have
a > Q/2
a
L2 (G),
{(I + R) , G}
and also, for a > Q/2,
(I + R) L2 (G)
= Ba ()L2 (G)
= Ba L2 (G) < .
a
Note that an analogue of Corollary 5.1.4 for compact Lie groups may be
obtained by noticing that (2.15) yields
m
2m
d (I LG ) 2 2HS =
d2
< .
m > n/2
=
G
G
The following statement describes an important property of the eld {(I +
in relation with the right Sobolev spaces (see Section 4.4.8 for right
R) , G},
Sobolev spaces):
a
Proposition 5.1.5. Let R be a positive Rockland operator on G of homogeneous
degree . Let also a R.
a f L2 (G) and there exists a eld of operators
2a (G), then (I + R)
If f L
such that
{ : H H , G}
L2 (G),
{ (I + R) : H H , G}
a
(5.2)
and for almost all G,
a f }() = (I + R) a .
FG {(I + R)
(5.3)
satises (5.2) then there exists a
Conversely, if { : H H , G}
2
unique function f La (G) satisfying (5.3).
276
Chapter 5. Quantization on graded Lie groups
a
In Proposition 5.1.5, (I + R) is not obtained as the composition of
a
(possibly) unbounded operators as in Denition A.3.2. Instead, for (I + R) ,
it is viewed as the composition of a eld of operators dened on smooth vectors
with a eld of operators acting on smooth vectors, see Section 1.8.3.
In Proposition 5.1.5, we use the right Sobolev spaces associated with the
positive Rockland operator R. These spaces are in fact independent of the choice
of a positive Rockland operator used in their denition, see Sections 4.4.5 and 4.4.8.
Consequently, if (5.2) holds for one positive Rockland operator then (5.2) and (5.3)
hold for any positive Rockland operator and the Sobolev norm of f L2 (G), using
of (5.2).
one particular positive Rockland operator R, is equal to the L2 (G)norm
2a (G), then by Theorem 4.4.3 (3) (see also
Proof of Proposition 5.1.5. If f L
a f is in L2 (G) and its Fourier transform
Section 4.4.8), we have that fa := (I + R)
is a eld of bounded operators (in fact in the HilbertSchmidt class). By Lemma
a
5.1.2, (I + R) maps H onto itself. Hence we can dene
:= (fa )(I + R) ,
a
as an operator dened on H . One readily checks that the operators , G,
satisfy (5.2) and (5.3).
satises (5.2) then we dene
For the converse, if { : H H : G}
the function
a
1
L2 (G) fa := FG
{ (I + R) },
which is square integrable by the Plancherel theorem (see Theorem 1.8.11), and
the function
a fa ,
f := (I + R)
2a (G) by Theorem 4.4.3 (3). One readily checks that the function
which will be in L
f satises the properties described in the statement.
We now aim at stating and proving a property similar to Proposition 5.1.5 for
a
the left Sobolev spaces. It will use the composition of a eld with (I + R) on the
left and this is problematic when we consider any general eld = { : H
H } without utilising the composition of unbounded operators as in Denition
A.3.2. To overcome this problem, we introduce the following notion:
Denition 5.1.6. Let 1 Rep G and a R. We denote by Ha1 the Hilbert space
obtained by completion of H1 for the norm
a
Ha1 : v 1 (I + R) vH1 := vHa1 ,
where R is a positive Rockland operator on G of homogeneous degree .
5.1. Symbols and quantization
277
We may call them the H1 Sobolev spaces. Note that in the case of the
Schrodinger representation for the Heisenberg group, they coincide with Shubin
Sobolev spaces, see Section 6.4.3. More generally, if we realise an element G
2
m
as a representation 1 acting on some L (R ) via the orbit methods, see Section
1.8.1, then we view the corresponding Sobolev spaces as tempered distributions:
Ha1 S (Rm ).
The following lemma is a routine exercise.
Lemma 5.1.7. Let 1 Rep G and a R.
1. If a = 0, then Ha1 = H1 . If a > 0, we realise Ha1 as a subspace of H1 and
a
it is the domain of the operator 1 (I + R) . If a < 0, we realise Ha1 as a
a
Hilbert space containing H1 and the operator 1 (I + R) extends uniquely
to a bounded operator Ha1 H1 .
2. For any a R, realising Ha1 as in Part 1, this space is independent of
the positive Rockland operator R and two positive Rockland operators yield
equivalent norms.
3. We have the continuous inclusions
a < b = Hb 1 Ha1 .
injectively and
For any a, b R, the operator 1 (I + R) maps Hb 1 to Hba
1
are
in
duality
via
continuously. In this way, Ha1 and Ha
1
a
u, vHa
v)H .
:= (1 (I + R) u, 1 (I + R)
1
a
a
1 H1
This duality extends the H1 duality in the sense that
H 1
u Ha1 H1 , v Ha
1
u, vHa
a
1 H1
= (u, v)H1 .
4. If 2 is another strongly continuous representation such that 1 T 2 , that
is, T is a unitary operator satisfying T 1 = T2 , then T maps H1 to H2
bijectively by Lemma 1.8.12 and extends uniquely to an isometric operator
Ha1 Ha2 .
Lemma 5.1.7, especially Part 4, shows that Gelds
with domain or range on
these Sobolev spaces make sense:
Denition 5.1.8. Let a R. A Geld
of operators = { : H H , G}
dened on smooth vectors is dened on the Sobolev spaces Ha when for each
1 Rep G, the operator 1 is bounded on Ha1 in the sense that
C
v H1
1 vH1 CvHa1 .
278
Chapter 5. Quantization on graded Lie groups
Thus, by density of H1 in Ha1 , 1 extends uniquely to a bounded operator
dened on Ha1 for which we keep the same notation 1 : Ha1 H1 .
Example 5.1.9. For any positive Rockland operator of degree , the eld {(I +
a
is dened on the Sobolev spaces Ha . This is an easy consequence of
R) , G},
Lemma 5.1.7, especially Part 3.
We will allow ourselves the shorthand notation
= { : Ha H , G},
to indicate that the Geld
of operators is dened on the Sobolev spaces Ha .
Instead of Denition 5.1.8, we could also have dened Gelds
of operators
a
dened on H Sobolev spaces in a way similar to Denition 1.8.13 (where Gelds of operators dened on smooth vectors were dened). Naturally, these two
viewpoints are equivalent since H1 is dense in Ha1 .
However, in order to dene Gelds
of operators with range in the Ha Sobolev
spaces, we have to adopt the latter viewpoint in the sense that we modify Denitions 1.8.13 and 1.8.14 (in this way, we make no further assumptions on the elds
or on the Sobolev spaces):
Denition 5.1.10. Let a R.
A Geld
of operators dened on smooth vectors with range in the Sobolev
where
spaces Ha is a family of classes of operators { , G}
:= {1 : H1 Ha1 , 1 }
viewed as a subset of Rep G, satisfying for any two elements
for each G
1 and 2 in :
1 T 2 = 2 T = T 1 on H .
(Here we have kept the same notation for the intertwining operator T and
its unique extension between Sobolev spaces Ha1 Ha2 , see Lemma 5.1.7
Part 4.)
It is measurable when for one (and then any) choice of realisation 1 and
the resulting eld { v , G}
is
any vector v1 Ha1 , as runs over G,
2
measurable whenever G v Ha d() < . (Here we assume that all the
Ha norms are realised via a xed positive Rockland operator.)
Unless otherwise stated, a Geld
of operators dened on smooth vectors
with range in the Sobolev spaces Ha is always assumed measurable. We will allow
ourselves the shorthand notation
= { : H Ha , G}
5.1. Symbols and quantization
279
to indicate that the Geld
of operators has range in the Sobolev space Ha .
Naturally, if a Geld of operators is dened on smooth vectors = { :
with the usual range H = H0 , then it has range in the Sobolev
H H , G}
a
spaces H when for each 1 Rep G and any v H1 , we have 1 v Ha1 .
Moreover, the following property of composition is easy to check: if 1 has
range in Ha and 2 is dened on Ha ,
i.e.
1 = {1, : H Ha , G}
and
2 = {2, : Ha H , G},
then the following eld
2 1 := {2, 1, : H H , G}
makes sense as a Geld
of operators dened on smooth vectors. This coincides
or extends the denition of composition of elds (the rst one acting on smooth
vectors) given in Section 1.8.3.
We can apply this property of composition to = { : H Ha , G}
a
and {(I + R) , G}, see Example 5.1.9 for the latter, to obtain the Geld
dened on smooth vectors by
(I + R) = {(I + R) : H H , G}.
a
(5.4)
We can now state the proposition which will enable us to dene the group
Fourier transform of a function in a left or right Sobolev space.
Proposition 5.1.11. Let a R.
a
(L) If f L2a (G), then (I + R) f L2 (G) and there exists a eld of operators
such that
{ : H Ha , G}
a
L2 (G),
{(I + R) : H H , G}
a
a
FG {(I + R) f }() = (I + R) , for almost all G,
(5.5)
(5.6)
where R is a positive Rockland operator on G of homogeneous degree .
satises (5.5) for one positive
Conversely, if { : H Ha , G}
Rockland operator R, then there exists a unique function f L2a (G) satisfying
(5.6).
2 (G), then the (unique) eld obtained in Proposition 5.1.5 can be
(R) If f L
a
dened on Ha .
extended uniquely into a eld { : Ha H , G}
Properties (L) and (R) are independent of the choice of R.
a
In Proposition 5.1.11, (I + R) is not obtained as the composition of
(possibly) unbounded operators as in Denition A.3.2 but is understood via (5.4).
280
Chapter 5. Quantization on graded Lie groups
In Proposition 5.1.11, we use the left and right Sobolev spaces associated
with the positive Rockland operator R. These spaces are in fact independent of
the choice of a positive Rockland operator used in their denition, see Sections
4.4.5 and 4.4.8. Consequently, if (5.5) hold for one positive Rockland operator then
(5.5) and (5.6) hold for any positive Rockland operator and the Sobolev norm of
f L2 (G), using one particular positive Rockland operator R, is equal to the
of (5.5).
L2 (G)norm
Proof of Proposition 5.1.11. Property (L). If f L2a (G), then by Theorem 4.4.3
a
(3), we have that fa := (I + R) f is in L2 (G) and its Fourier transform is a eld
of bounded operators (in fact in the HilbertSchmidt class). By (5.4) we can dene
a
= { : H Ha } via := (I + R) (fa ). One readily checks that the
eld satises (5.2) and (5.3).
satises (5.2) then we dene
For the converse, if { : H Ha : G}
the function
a
1
{(I + R) },
L2 (G) fa := FG
which is square integrable by the Plancherel theorem (see Theorem 1.8.11), and
the function
a
f := (I + R) fa ,
which will be in L2a (G) by Theorem 4.4.3 (3). One readily checks that the function
f satises the properties described in the statement. This shows the property (L).
Property (R) follows easily from (5.2).
2 (G) is also
From the proof above, one can check easily that if f L2a (G) or L
a
in any of the spaces where the group Fourier transform has already been dened,
will coincide with
namely, L2 (G) or K(G), then = { : H H , G}
the group Fourier transform of f . Hence we can extend the denition of the group
Fourier transform to Sobolev spaces:
Denition 5.1.12. Let a R. The group Fourier transform of f L2a (G) or f
2 (G) is the eld of operators dened on smooth vectors given in Proposition
L
a
5.1.11.
This leads us to dene the following spaces of elds of operators:
denote the space of elds of operators with
Denition 5.1.13. (L) Let L2a (G)
range in Ha and satisfying (5.5), that is,
= { : H Ha , G},
a
L2 (G),
{(I + R) : H H , G}
for one (and then any) positive Rockland operator of homogeneous degree .
We also set
a
(5.7)
L2 (G)
:= (I + R) L2 (G)
.
a
5.1. Symbols and quantization
281
2a (G)
denote the space of elds of operators dened on Ha and satis(R) Let L
fying (5.2), that is,
= { : Ha H , G},
a
L2 (G),
{ (I + R) : H H , G}
for one (and then any) positive Rockland operator of homogeneous degree .
We also set
a
L 2 (G)
:= (I + R) L2 (G)
.
a
It is a routine exercise, using Proposition 5.1.11 and the properties of the
Sobolev spaces (see Section 4.4), to show that
Proposition 5.1.14. Let a R. If R is a positive Rockland operator of homogeneous
2
degree , the map L2 (G)
given by (5.7) is a norm on the vector space La (G).
a
is a Banach space which is independent of R.
Endowed with this norm, L2a (G)
Two norms corresponding to any two choices of Rockland operators via (5.7) are
equivalent.
The Fourier transform FG is an isomorphism between Banach spaces acting
It coincides with the usual Fourier transform on L2 (G)
from L2a (G) onto L2a (G).
for a = 0.
be in L2 (G).
Then
Let = { , G}
a
{(X) , G}
for any Nn , and
is in L2a[] (G)
0
{(I + R)s/ , G}
for any s R. Furthermore, if f = F 1 L2a (G) then
is in L2as (G)
G
FG (X f )() = (X) f()
and
FG ((I + R)s/ f )() = (I + R)s/ f().
We have similar results for the right Sobolev spaces. Furthermore the adjoint
L
2a (G)
and L
2a (G)
L2a (G)
isomorphically as Banach
map maps L2a (G)
spaces.
Recall that the tempered distributions X f and (I + R)s/ f used in the
statement just above are respectively dened via
X f, = f, {X }t ,
and
S(G),
s/ ,
(I + R)s/ f, = f, (I + R)
S(G).
(5.8)
(5.9)
282
Chapter 5. Quantization on graded Lie groups
For (5.9), see Denition 4.3.17. For (5.8), this is the composition of the formula
obtained for one vector eld (with polynomial coecients) by integration by parts.
See also (1.10) for the denition of {X }t .
In Corollary 1.8.3, we stated the inversion formula valid for any Schwartz
function on any connected simply connected Lie group. Here we weaken the hypothesis using the Sobolev spaces in the context of a graded Lie group G:
Proposition 5.1.15 (Fourier inversion formula). Let f be in the left Sobolev space
2s (G) with s > Q/2. Then for almost every
L2s (G) or in the right Sobolev space L
Rep G, the operator f () is trace class with
Trf()d() < .
(5.10)
G
Furthermore, f is continuous on G, and for every x G we have
Tr (x)f () d() =
Tr f()(x) d().
f (x) =
G
G
(5.11)
it is
In the statement above, as s > Q/2 > 0, the eld f is in L2 (G),
then a eld of bounded operators (even in HilbertSchmidt classes) and so can be
composed on the left and the right with (x). The (possibly innite) traces
Tr 1 (x)f(1 ) , Tr f(1 )1 (x) and Tr f(1 )
are equal for 1 Rep G as 1 is unitary. They are constant on the class of
and are, therefore, treated as depending on G.
They are
1 Rep G in G
nite for almost all G in view of (5.10).
Note that (5.10) implies not only that the two expressions
Tr (x)f() d() and
Tr f()(x) d()
G
G
make sense but that they are also equal by the properties of the trace since (x)
is bounded.
Proof of Proposition 5.1.15. Let R be a positive Rockland operator of homogeneous degree . Let f L2s (G) with s > Q/2. We set
s
fs := (I + R) f L2 (G).
The properties of the trace imply
s
s
Trf() = Tr (I + R) fs () (I + R) HS fs ()HS .
5.1. Symbols and quantization
283
Integrating against the Plancherel measure, we obtain by the CauchySchwartz
inequality
s
Trf()d() (I + R) L2 (G)
fs L2 (G)
.
G
By Corollary 5.1.4, Cs := (I + R) L2 (G)
is a positive nite constant. Since
fs ()L2 (G)
is equal to f L2s (G) which is nite, we have obtained (5.10).
s
Let S(G). By the Plancherel formula, especially (1.30), we have
(f, )L2 (G)
=
=
=
=
(fs , (I + R) )L2 (G)
s
Tr FG {fs }() FG {(I + R) }()
d()
G
s
(I + R) s d()
Tr (I + R) f() ()
G
d().
Tr f() ()
s
G
Note that the two functions fs and (I + R) are both square integrable so all
the traces above are nite.
We now x a nonnegative
function D(G) with compact support con
taining 0 and satisfying G = 1. We apply what precedes to := given
by
(y) := Q (1 y), > 0, y G,
and obtain
(f, )L2 (G) =
G
Tr f()
() d().
Let us show that the right handside of (5.12) converges to
Tr f()
() d() 0
Tr f() d().
G
G
Note that the right handside of (5.13) is nite by (5.10).
The integrand on the lefthand side is bounded by
()L (H ) Trf(),
()
Tr f()
and
()L (H ) L1 (G) = L1 (G) .
Hence
() L1 (G) Trf(),
Tr f()
and the righthand side is integrable by (5.10).
(5.12)
(5.13)
284
Chapter 5. Quantization on graded Lie groups
Let us show the convergence for every G
Tr f()
() 0 Tr f() .
(5.14)
In order to do this, we want to estimate the dierence
() I)
() Tr f() = Tr f() (
Tr f()
() IL (G)
Tr f () .
Since
() =
and as
G
G
(y)(y)dy =
(
y)(y)dy =
(z)(z)dz,
= 1, we have
() IL (H )
(z) ((z) I) dzL (H )
(z) (z) IL (H ) dz
(z)dz.
sup (z) IL (H )
zsupp
As is strongly continuous and supp compact, we know that
sup (z) IL (H ) 0 0.
zsupp
This implies the convergence in (5.14) for each G.
We can now apply Lebesgues dominated convergence theorem to obtain the
convergence in (5.13).
By the Sobolev embeddings (see Theorem 4.4.25), f is continuous on G and
it is a simple exercise to show that the left handside of (5.12) converges to
(f, )L2 (G) 0 f (0).
Hence we have obtained the inversion formula given in (5.11) at x = 0. Replacing
f by its left translation f (x ) which is still in L2s (G) with the same Sobolev norm,
it is then easy to obtain (5.11) for every x G.
2s (G) with s > Q/2, we set fs := (I + R)
s f L2 (G)
For the case of f L
and we obtain similar properties as above, ending by using right translations to
obtain (5.11).
5.1. Symbols and quantization
285
5.1.2 The spaces Ka,b (G), LL (L2a (G), L2b (G)), and L
a,b (G)
In this section we describe the spaces Ka,b (G), LL (L2a (G), L2b (G)) and L
a,b (G),
extending the notion of the group von Neumann algebras discussed in Section
1.8.2, to the setting of Sobolev spaces.
Denition 5.1.16 (Spaces LL (L2a (G), L2b (G)) and Ka,b (G)). Let a, b R. We denote by
LL (L2a (G), L2b (G))
the subspace of operators T L (L2a (G), L2b (G)) which are leftinvariant.
We denote by
Ka,b (G)
the subspace of tempered distributions f S (G) such that the operator S(G)
f extends to a bounded operator from L2a (G) to L2b (G).
If a positive Rockland operator R of homogeneous degree is xed, then the
Ka,b (G)norm is dened for any f Ka,b (G), as the operator norm of f
viewed as an operator from L2a (G) to L2b (G), i.e.
f Ka,b := f L (L2a (G),L2b (G)) .
(5.15)
c
Here we have considered the Sobolev norms (I + R) 2 for c = a, b for
L2a (G) and L2b (G), respectively.
The vector space LL (L2a , L2b ) is a Banach subspace of L (L2a , L2b ). Since the
Sobolev spaces L2a (G) are independent of the choice of a positive Rockland operator
R (see Section 4.4.5), so are LL (L2a (G), L2b (G)) and also Ka,b (G). However, the
norms on these spaces do depend on a choice of a positive Rockland operator R.
We may often write Ka,b instead of Ka,b (G) to ease the notation when no
confusion is possible.
We have the immediate properties:
Proposition 5.1.17.
1. If a = b = 0 then
K0,0 = K
and
LL (L2a , L2b ) = LL (L2 ).
The norms K0,0 and K (dened in (5.15) and in (1.37) respectively)
coincide. For any f K we have
f K = f K
where
f (x) = f(x1 ),
and
r > 0
f Dr K = rQ f K .
286
Chapter 5. Quantization on graded Lie groups
2. Fixing a positive Rockland operator R, the mapping f f Ka,b denes
a norm on the vector space Ka,b which becomes a Banach space. Any two
positive Rockland operators produce equivalent norms on Ka,b .
3. Let a, b R. We have the continuous inclusion
Ka,b (G) S (G).
Moreover if Tf denotes the convolution operator f for f S (G),
then the following are equivalent:
f Ka,b
Tf LL (L2a (G), L2b (G))
(I + R) Tf (I + R) LL (L2 (G))
b
a f K(G),
(I + R) (I + R)
where R is any positive Rockland operator of homogeneous degree .
4. For any c1 , c2 0 we have the inclusions
LL (L2a , L2b ) LL (L2a+c1 , L2bc2 )
and
Ka,b Ka+c1 ,bc2 .
5. If f Ka,b then X f Ka,b[] for any Nn0 and (I + R)s/ f Ka,bs
for any s R. Furthermore, X and (I + R)s/ are bounded on Ka,b :
X f Ka,b[] Ca,b,[] f Ka,b
and
f Ka,b
(I + R)s/ f Ka,bs Ca,b,s
for some positive nite constants Ca,b,[] and Ca,b,s
independent of f .
If a and b are in N0 , a norm equivalent to the Ka,b norm is
X f K ,
X
f
[]a, []b
and if a [a, 0] and b [0, b] then
f Ka ,b Ca,b,a ,b,R
X f K .
X
[]a, []b
The denitions of the tempered distributions X f and (I + R)s/ f were
recalled in (5.8) and (5.9) respectively. For the proper denition of the operators
b
a , see Denitions 4.3.17 and 4.4.31.
(I + R) , (I + R)
5.1. Symbols and quantization
287
Proof of Proposition 5.1.17. Part (1) follows from the properties of the von Neumannalgebras K(G) and LL (L2 (G)) as well as from the following two easy observations:
Q
L2 (G)
Dr 2 = r 2 2 ,
and for any f K, S(G) and r > 0,
(f Dr ) (x) = rQ
D r1 f (rx).
Part (2) is easy to check. Part (3) follows from the Schwartz kernel theorem, see
Corollary 3.2.1. Parts (4) and (5), follow easily from the properties of the Sobolev
spaces and Part (3).
We now show that we can make sense of convolution of distributions in some
Ka,b (G)spaces. The following lemma is almost immediate to check.
Lemma 5.1.18. Let f Ka,b (G) and g Kb,c (G) for a, b, c R, and let Tf :
f and Tg : g be the associated operators. Then the operator Tg Tf is
continuous from L2a (G) to L2c (G) and its rightconvolution kernel (as a continuous
linear operator from S(G) to S (G)) is denoted by h Ka,c (G).
If (fn ) and (gn ) are sequences of Schwartz functions converging to f in
Ka,b (G) and g in Kb,c (G), respectively, then h is the limit of fn gn in Ka,c (G).
Consequently, with the notation of the lemma above, h coincides with the
convolution of f with g whenever the convolution of f with g makes any technical sense, for instance, if the tempered distributions f and g (which are already
assumed to be in Ka,b (G) and Kb,c (G) respectively) satisfy
f and g are locally integrable functions with f  g L1 (G),
or at least one of the distributions f or g has compact support,
or at least one of the distributions f or g is Schwartz.
Hence we may extend the notation and dene:
Denition 5.1.19. If f Ka,b (G) and g Kb,c (G) for a, b, c R, and Tf : f ,
Tg : g are the associated operators, we denote by f g the distribution in
Ka,c (G) which is the right convolution kernel of Tg Tf .
We obtain easily the following properties:
Corollary 5.1.20. Let f Ka,b (G) and g Kb,c (G) for a, b, c R. Then we have
the following property of associativity for any S(G)
(f g) = ( f ) g,
and more generally for any h Kc,d (G) (where d R)
f (g h) = (f g) h,
288
Chapter 5. Quantization on graded Lie groups
as convolutions of an element of Ka,b (G) with an element of Kb,d (G) for the lefthand side, and of an element of Ka,c (G) with an element of Kc,d (G) for the righthand side.
The rest of this section is devoted to the denition of the group Fourier
transform of a distribution in Ka,b (G). We start by dening what will turn out to
be the image of the group Fourier transform on Ka,b (G). We recall that L (G)
is the space of measurable elds of operators on G which are uniformly bounded,
see Denition 1.8.8.
Denition 5.1.21. Let a, b R. We denote by L
a,b (G) the space of elds of opera
b
tors = { : H H , G} satisfying
C > 0 S(G)
2 CL2 (G) .
L (G)
a
(5.16)
Here we assume that a positive Rockland operator has been xed to dene the
and L2a (G).
norms on L2b (G)
For such a eld , L (G)
denotes the inmum of the constant C > 0
a,b
satisfying (5.16).
We may sometimes abuse the notation and write L (G)
when no cona,b
fusion is possible.
Note that as S(G), its group Fourier transform acts on smooth vectors,
see Example 1.8.18. Hence the composition above makes sense, see Section
1.8.3.
Naturally, the space L
a,b (G) introduced in Denition 5.1.21 is independent of
and L2a (G):
the choice of a Rockland operator used to dene the norms on L2b (G)
satises the condition in Denition
Lemma 5.1.22. If { : H Hb , G}
5.1.21 for one positive Rockland operator, then it satises the same property for
any positive Rockland operator. Moreover, if R1 and R2 are two positive Rockland
operators, and if L
and L
denote the corresponding inma,
(G)
(G)
a,b,R1
a,b,R2
then there exists C > 0 independent of such that
C 1 L
a,b,R2 (G)
L
a,b,R1 (G)
CL
a,b,R2 (G)
Proof. This follows easily from the independence of the Sobolev spaces on G and
of the positive Rockland operators, see Section 4.4.5 and Proposition 5.1.14.
G
If the eld acts on smooth vectors, we can simplify Denition 5.1.21:
be a eld acting on smooth
Lemma 5.1.23. Let = { : H H , G}
vectors. Then La,b (G) if and only if
b
a
L (G),
{(I + R) (I + R) : H H , G}
(5.17)
5.1. Symbols and quantization
289
where R is a positive Rockland operator of degree , and in this case,
L
= (I + R) (I + R) L (G)
.
b
a,b (G)
Proof. This follows easily from the density of S(G) in L2b (G).
Note that the composition in (5.17) makes sense as all the elds involved act
on smooth vectors. In Corollary 5.1.30, we will see a sucient condition (which
will be useful later) for a eld to be acting on smooth vectors.
We can now characterise the elements of Ka,b (G) in terms of L
a,b (G):
Proposition 5.1.24. Let a, b R.
(i) If L
a,b (G), then the operator T : S(G) S (G) dened via
T
() := (),
S(G), G,
(5.18)
extends uniquely to an operator in L (L2a , L2b ). Moreover,
T L (L2a ,L2b ) = L
a,b (G)
(5.19)
where the Sobolev norms are dened using a chosen positive Rockland operator R with homogeneous degree . The right convolution kernel f S (G)
of T is in Ka,b (G).
(ii) Conversely, if f Ka,b (G) then there exists a unique L
a,b (G) such that
f () = (),
S(G), G.
(5.20)
Furthermore, if f is also in any of the spaces where the group Fourier transform has already been dened, namely any Sobolev space L2a (G) or K(G),
will coincide with the group Fourier transform of f .
then = { , G}
Proof. The properties of T in Part (i) follow from the Plancherel theorem (Theorem 1.8.11) and the density of S(G) in L2 (G). The right convolution kernel
f S (G) of T is in Ka,b (G) by Proposition 5.1.17.
Conversely, let f Ka,b (G). By assumption the operator Tf : S(G)
f admits a bounded extension from L2a (G) to L2b (G). Thus the operator (I +
b
a
R) Tf (I + R) is bounded on L2 (G) and we denote by fa,b K(G) its right
a
convolution kernel. For any S(G), we have a := (I + R) S(G) by
2
Corollary 4.3.16 thus a fa,b L (G) and we have
Tf L2b (G)
with
Tf = (I + R) (a fa,b ).
b
and
Consequently FG (Tf ) L2b (G)
b
a
b
FG (Tf ) = (I + R) fa,b a = (I + R) fa,b (I + R) .
290
Chapter 5. Quantization on graded Lie groups
dened via
One checks easily that { : H Hb , G}
:= (I + R) fa,b () (I + R)
b
is in L
a,b (G) and satises (5.20). The rest of the proof of Part (ii) follows easily
from the computations above and the uniqueness of the group Fourier transforms
already dened.
Thanks to Proposition 5.1.24, we can extend the denition of the group
Fourier transform to Ka,b (G):
Denition 5.1.25 (The group Fourier transform on Ka,b (G)). The group Fourier
in
transform of f Ka,b (G) is the eld of operators { : H Hb , G}
La,b (G) associated to f by Proposition 5.1.24, and we write
f() := (f ) := ,
G.
As the next example implies, any leftinvariant vector eld is in some Ka,b (G)
and their Fourier transform can be dened via Denition 5.1.25. As is shown in
the proof below, this coincides with the innitesimal representation of the corresponding element of U(g) dened in Section 1.7.
Example 5.1.26. Let Nn0 . The operator X is in L (L2[] (G), L2 (G)) and more
generally in L (L2[]+s (G), L2s (G)) for any s R. Its right convolution kernel is the
distribution X 0 dened via (see (5.8))
X 0 , = 0 , {X }t = {X }t (0),
which is in K[],0 , and more generally in Ks+[],s for any s R. Its group Fourier
transform is
FG (X 0 )() = (X ) = (X)
and coincides with the innitesimal representation on U(g). It is in L
s+[],s (G) for
any s R.
Proof. By Theorem 4.4.16, X maps L2[] (G) continuously to L2 (G) and, more
generally, L2s+[] (G) continuously to L2s (G).
By Proposition 1.7.6, we have for any S(G)
FG (X )() = (X )()
= (X) ().
This shows that FG (X 0 ) coincides with {(X ), G}.
As our next example shows, when multipliers in a positive Rockland operator
are in LL (L2s (G), L2sb (G)), the group Fourier transform of their right convolution
kernels can also be given via the functional calculus of the Rockland operators:
5.1. Symbols and quantization
291
Example 5.1.27. Let R be a positive Rockland operator of homogeneous degree
. Let m be a measurable function on [0, ) satisfying
b
C > 0 0 m() C(1 + ) .
Then the operator m(R) dened by the functional calculus of R extends uniquely
to an operator in LL (L2s+b (G), L2s (G)) for any s R. Its right convolution kernel
m(R)0 is in Ks+b,s for any s R. Its group Fourier transform is
FG (m(R)0 )() = m((R))
dened by the functional calculus of (R). It is in L
s+b,s (G) for any s R. For a
xed s R, we have
m(R)LL (L2s+b (G),L2s (G)) = m(R)0 Ks+b,s = m((R))L
s+b,s (G)
sup(1 + )
>0
m(),
if we realise the Sobolev norms with R.
We refer to Section 4.1.3 and Corollary 4.1.16 for the properties of the functional calculus of R2 and (R).
Proof. The function m1 given by
m1 () := m()(1 + ) ,
b
0,
is measurable and bounded on [0, ). The operator m1 (R) dened by the functional calculus of R is therefore bounded on L2 (G) with
m1 (R)L (L2 (G)) sup m1 ().
0
Again from the properties of the functional calculus of R, we also have
b
m(R) m1 (R)(I + R) ,
in the sense of operators. Since Dom(I + R)b/ S(G) (see Corollary 4.3.16), this
shows that the domain of m(R) contains S(G) and that
m1 (R) = m(R)(I + R)
b
on S(G).
The properties of the functional calculus of R yield for any s R,
m1 (R)L (L2 (G))
m1 (R)L (L2s (G))
m(R)(I + R) L (L2s (G))
m(R)L (L2s+b (G),L2s (G)) .
292
Chapter 5. Quantization on graded Lie groups
By Corollary 4.1.16, the kernel of m1 (R) is the tempered distribution m1 (R)0
Adapting the proof of Corollary 4.1.16,
with Fourier transform {m1 ((R)), G}.
we see that
m1 ((R)) = m((R))(I + (R))
b
on H ,
G.
It is now straightforward to check that the kernel of the operator m(R) is in Ks+b,s
and its Fourier transform is {m((R)), G}.
Naturally, any Schwartz function is in any Ka,b and one can readily estimate
the associated norm:
Example 5.1.28. If S(G), then for any a, b R, the operator T : is
in L (L2a (G), L2b (G)), Ka,b and L
a,b . If we x a positive Rockland operator
R of homogeneous degree , then we have
L (I + R) b (I + R)
a L1 (G) < ,
T L (L2a (G),L2b (G)) = Ka,b =
a,b
where the norms on L (L2a (G), L2b (G)), Ka,b and L
a,b are dened with R.
With Denition 5.1.25, we can reformulate Proposition 5.1.24 and parts of
Proposition 5.1.17 and Corollary 5.1.20 as the following proposition.
Proposition 5.1.29. 1. Let a, b R. The Fourier transform FG maps Ka,b (G)
onto L
a,b (G). Furthermore, FG : Ka,b (G) La,b (G) is an isomorphism
between Banach spaces. In particular, for f Ka,b (G),
f Ka,b = fL
a,b (G)
It coincides with the Fourier transform on K(G) for a = b = 0.
2. If 1 L
a1 ,b1 (G) and 2 La2 ,b2 (G) with b2 = a1 , then their product 1 2
given by the Fourier transform of
makes sense as the element of La2 ,b1 (G)
1
1
(FG 2 ) (FG 1 ).
and f2 Ka ,b (G)
with b2 = a1 , then
In other words, if f1 Ka1 ,b1 (G)
2 2
the Fourier transform of f2 f1 Ka2 ,b1 (G) is
FG (f2 f1 ) = FG (f1 )FG (f2 ).
L (G).
Then we have for any Nn ,
3. Let = { : H H , G}
0
a,b
L
{(X) : H H , G}
a,b[] (G),
(5.21)
and for any s R,
L2a,bs (G).
{(I + R)s/ : H H , G}
(5.22)
5.1. Symbols and quantization
293
1
Furthermore, if f = FG
Ka,b (G) then
FG (X f )() = (X) f()
and
FG ((I + R)s/ f )() = (I + R)s/ f().
The elds of operators in (5.21) and (5.22) are understood as compositions
of elds of operators in L
a2 ,b2 and La1 ,b1 with b2 = a1 , see Part 2 and Examples
5.1.26 and 5.1.27.
With the help of Proposition 5.1.29, we can now give a usefull sucient
condition for a eld to act on smooth vectors and reformulate Corollary 4.4.10
into
Corollary 5.1.30. Let a, b R and let { , Z} be a sequence of real numbers
which tends to as . Let L
a+ ,b+ (G) for every Z. Then
is a eld of operators acting on smooth vectors:
= { : H H , G}.
Furthermore L
a+,b+ (G) for every R and for any c 0, we have
sup L
c
a+,b+ (G)
Cc max L
a+ ,b+ (G)
, L
a+ ,b+ (G)
where N0 is the smallest integer such that c and c.
n
Proof. By Proposition 5.1.29, (X) L
a+ ,b+ [] for every N0 and
for every
every Z. Thus choosing [] b, we have (X) L2 (G)
S(G). Realising G as a representation of G and xing v H , this
is smooth. Hence ()v
is smooth
implies that the mapping x (x) ()v
and acts on smooth vectors. As this holds for every D(G), so does by
Lemma 1.8.19. We conclude with Corollary 4.4.10.
We end this section with one more technical property:
Lemma 5.1.31. Let L
a,b (G) where a, b R. Let S(G). Then we have
2
Ls (G) for any s R and
(5.23)
Tr ()
d() < .
G
1
Setting f := FG
Ka,b , the function f is smooth and we have for any x G
the equality
f (x) =
Tr (x) ()
d().
G
294
Chapter 5. Quantization on graded Lie groups
Remark 5.1.32. The composition makes sense since is dened on smooth
vectors and acts on smooth vectors. The composition (x) () makes sense
since (x) is bounded and is bounded (even in Hilbert Schmidt classes) since
2 (G)
for any s, hence in particular in L2 (G).
it is stated rst that L
s
1
. Then
Proof. Let T be the operator with right convolution kernel f := FG
2
2
2
T L (La (G), Lb (G)) and T T extends to an operator in L (La (G)). For any
S(G), the denition of the adjoint and the duality between Sobolev spaces
yield
T 2L2 (G)
T T ,
L2a (G)L2a (G)
T T L (L2a (G)) L2a (G) L2a (G) .
This last expression is nite since T T L (L2a (G)) and S(G) L2s (G) for any
For any
s R. Thus T L2 (G) and its Fourier transform is L2 (G).
yields
s R, we may replace with s = (I + R)s/ S(G) and s L2 (G)
2
Ls (G).
for some s > Q/2, we obtain
2 (G)
Applying Proposition 5.1.15 to L
s
1
(5.23). Note that f := FG is a tempered distribution so f is smooth (see
Lemma 3.1.55). The group Fourier transform of f is by Proposition 5.1.29
Part 2 and Example 5.1.28. We now conclude with the inversion formula given in
Proposition 5.1.15.
5.1.3
Symbols and associated kernels
In this section we aim at establishing a onetoone correspondence between a
and a function ; this function
collection of operators parametrised by G G
will turn out to be the kernel of the operator naturally associated to . For the
abstract setting behind measurable elds of operators and some of their properties
we refer to Section B.1.6, especially to Proposition B.1.17, as well as Section 1.8.3.
Denition 5.1.33 (Symbols). A symbol is a eld of operators {(x, ) : H
depending on x G, satisfying for each x G
H , G}
a, b R
L
(x, ) := {(x, ) : H H , G}
a,b (G).
Here we use the usual identications of a strongly continuous irreducible
and of a eld
unitary representation from Rep G with its equivalence class in G,
of operators acting on the smooth vectors parametrised by G with its equivalence
class with respect to the Plancherel measure .
We will usually assume that the symbols are uniformly regular in x:
5.1. Symbols and quantization
295
Denition 5.1.34 (Continuous and smooth symbols).
is said to be continuous in x G
A symbol {(x, ) : H H , G}
whenever there exists a, b R such that
x G
L
(x, ) := {(x, ) : H H , G}
a,b (G),
and the map x (x, ) is continuous from G Rn to the Banach space
L
a,b (G).
is said to be smooth in x G
A symbol = {(x, ) : H H , G}
whenever it is a eld of operators depending smoothly in x G (see Remark
is
1.8.16) and, for every Nn0 , the eld {x (x, ) : H H , G}
continuous.
Important note: In the sequel, whenever we talk about symbols (on graded Lie
groups), we always mean the symbols which are smooth in x G in the sense of
Denition 5.1.34 unless stated otherwise.
For a symbol as in Denition 5.1.34, we will usually write
= {(x, ), (x, ) G G},
but we may sometimes abuse the notation and refer to the symbol simply as
(x, ).
is a symbol, then
Lemma 5.1.35. If = {(x, ), (x, ) G G}
1
{(x, )}
x := FG
is a tempered distribution and the map
G x x S (G)
is smooth.
In other words,
C (G, S (G)).
Here C (G, S (G)) denotes the set of smooth functions from G to S (G).
Proof. As is a smooth symbol, for every Nn0 , there exists a , b R such that
G x x (x, ) L
a ,b (G) is continuous. By Proposition 5.1.29, composing
1
this with FG implies that G x x x Ka ,b is continuous. Since the
inclusion Ka ,b S (G) is continuous, this implies that each map G x
x x S (G) is continuous. Hence G x x S (G) is smooth.
296
Chapter 5. Quantization on graded Lie groups
Denition 5.1.36 (Associated kernels). If is a symbol, then the tempered distribution
1
{(x, )} S (G)
x := FG
is called its associated kernel, sometimes its right convolution kernel, or just a
kernel. We may also call the smooth map G x x S (G) or the map
(x, y) x (y) = (x, y) the kernel associated with .
The smoothness of the map x (x, ) implies easily:
Lemma 5.1.37. If = {(x, )} is a symbol with kernel x then for any Nn0 ,
X := {Xx (x, )},
:= {X
x (x, )}, and x := {x (x, )},
X
are symbols with respective kernels
Xx x ,
x x ,
X
and
x x .
m
Examples of symbols are the symbols in the classes S,
(G) dened later on.
Here are more specic examples of symbols which do not depend on x G.
is a symbol
Example 5.1.38. If f Ka,b (G), then f = {f() : H H , G}
with kernel f .
The following are particular instances of this case:
is a symbol and its kernel is the Dirac
0 = I = {I : H H , G}
measure 0 .
is a symbol with kernel
For any Nn0 , {(X) : H H , G}
X 0 , see Example 5.1.26. It acts on smooth vectors, see Example 1.8.17, or
alternatively Example 5.1.26 together with Corollary 5.1.30.
If R is a positive Rockland operator of homogeneous degree and if m is a
measurable function on [0, ) satisfying
C > 0
0 m() C(1 + )b/ ,
is a symbol with kernel m(R)0 , see
then {m((R)) : H H , G}
Example 5.1.27. By Corollary 5.1.30, this symbol also acts on smooth vectors
{m((R)) : H H , G}.
5.1.4
Quantization formula
With the notion of symbol explained in Section 5.1.3, our quantization makes
sense:
5.1. Symbols and quantization
297
Theorem 5.1.39 (Quantization). The quantization dened by formula (5.1) makes
sense for any symbol = {(x, )}. More precisely, for any S(G) and x G,
we have
Tr (x)(x, )()
d() = x (x),
(5.24)
Op()(x) =
G
in (5.24) is welldened
where x denotes the kernel of . The integral over G
and absolutely convergent. We also have Op() C (G). Furthermore, the
quantization mapping Op() is onetoone and linear.
Proof. Lemma 5.1.31 (see also Remark 5.1.32) implies that the integral in (5.24)
is well dened, absolutely convergent and is equal to x (x).
By Lemma 3.1.55, for each x G, the function x is smooth. By Lemma
5.1.35, x x S (G) is smooth. Hence by composition, x x (x) is smooth.
The quantization is clearly linear. Since the kernel is in onetoone linear
correspondence with the operator, and by Lemma 5.1.35 also with the symbol,
the quantization Op() is onetoone.
Denition 5.1.40 (Notation). If an operator T is given by the formula (5.24) with
symbol (x, ), so that T = Op(), we will also write
= T
or
(x, ) = T (x, )
or even
= Op1 (T ).
This notation is justied since the quantization given by (5.24) is onetoone by
Theorem 5.1.39.
The operators associated with the symbols given in Example 5.1.38 are the
ones alluded to in the introduction of this Section:
Continued Example 5.1.38: If f Ka,b (G), then Op(f) is the convolution operator
f with the right convolution kernel f .
The following are particular instances of this case:
Op(I) = I and, more generally, for any Nn0 , Op((X) ) = X .
These relations can also be expressed as
I (x, ) = IH and X (x, ) = (X) .
If R is a positive Rockland operator of homogeneous degree and if m is a
measurable function on [0, ) satisfying
C > 0
0 m() C(1 + )b/ ,
then Op(m((R))) = m(R).
In these examples, the symbols are independent of x. However it is easy to
produce xdependent symbols out of them using the following two observations.
298
Chapter 5. Quantization on graded Lie groups
is a symbol and c : G C is a smooth
If = {(x, ), (x, ) G G}
is a symbol.
function, then c := {c(x)(x, ), (x, ) G G}
and = { (x, ), (x, ) G G}
are two
If = {(x, ), (x, ) G G}
symbols, then so is their sum + = {(x, ) + (x, ), (x, ) G G}.
Remark 5.1.41. 1. The observations just above together with Example 5.1.38
and its continuation above imply that any dierential operator of the form
c (x)X with smooth coecients c
(5.25)
[]M
may be quantized, in the sense that
bol and we have
[]M
[]M
c (x)X = Op
c (x)(X) is a (smooth) sym
c (x)(X) .
[]M
The dierential calculus is, by denition, the space of dierential operators of the form
b (x)x with smooth coecients b ,
d
or, equivalently, of the form (5.25), see (3.1.5). Hence, we have obtained
that the dierential calculus may be quantized. This could be viewed as the
minimum requirement for a notion of symbol and quantization on a manifold.
2. In order to achieve this, we had to consider and use elds of operators dened on smooth vectors in our denition of symbol. Indeed, for instance, the
symbol associated to a leftinvariant vector eld X is {(X)} while (X) are
dened on H but is not bounded on H .
This technicality has also the following advantage when we apply our
theory in the setting of the Heisenberg group Hno in Chapter 6. Realising
n via Schr
odinger representations, the spaces
(almost all of) its dual group H
o
of smooth vectors will coincide with the Schwartz space S(Rno ). In this context, the symbols will be operators acting on S(Rno ) (which are smoothly
parametrised by points in Hno ).
3. With our notion of symbols and quantization, we also obtain part of the
functional calculus of any Rockland operators. More precisely, if R is a positive Rockland operator, we obtain all the operators of the form m(R) with
m : [0, ) C a measurable function of (at most) polynomial growth at
innity.
5.1. Symbols and quantization
299
4. The symbol classes that we have introduced are based on the quantization
relying on writing the operators as operators with rightconvolution kernels. There is an obvious parallel theory of quantization and of the corresponding symbols and their classes suited for problems based on the rightinvariant operators. With natural modications we could have considered
at the same time rightinvariant vector elds in Part (1) above and a quantization involving leftconvolution kernels of operators, i.e. writing the same
operators but now in the form x . As an outcome, with natural modications we would obtain a parallel theory with the same parallel collection
of results to those presented here.
Op() as a limit of nice operators
The operators we have obtained as Op() for symbols are limits of nice operators in the following sense:
Lemma 5.1.42. If = {(x, )} is a symbol, we can construct explicitly a family
of symbols = { (x, )}, > 0, in such a way that
1. the kernel (x, y) of is smooth in both x and y, and compactly supported
in x,
2. if S(G) then Op( ) D(G), and
3. Op( ) Op() uniformly on any compact subset of G.
0
Proof of Lemma 5.1.42. We x a number p such that p/2 is a positive integer
divisible by all the weights 1 , . . . , n . Therefore, if  p is the quasinorm given
by (3.21), then the mapping x xpp is a phomogeneous polynomial. We also x
o Cc (R) with o 0, o = 1 on [1/2, 2] and o = 0 outside of [1/4, 4]. For
any > 0, we write
(x) := o (xpp ).
Clearly D(G).
we denote by  the distance between the coadjoint orbits correIf G,
sponding to and 1.
a basis
Applying the orbit method, one can construct explicitly for each G
(v, )=1 formed by smooth vectors and such that the eld of vectors G v,
is measurable. We denote by proj, the orthogonal projection on the subspaces
spanned by v1, , . . . , v, where is the smallest integer such that > 1 .
We consider for any (0, 1) the mapping
(x, ) := (x)11 (x, ) proj, .
By Denition 5.1.36, the symbol and the kernel are related by
FG (,x )() = (x, ).
300
Chapter 5. Quantization on graded Lie groups
By the Fourier inversion formula (1.26), the corresponding kernel is
Tr (x, ) proj, (y) d(),
,x (y) = (x, y) = (x)
1
which is smooth in x and y and compactly supported in x.
The corresponding operator is Op( ), given for any S(G) and x G by
Op( )(x) =
Tr (x) (x, )()
d()
G
Tr (x)(x, ) proj, ()
= (x)
d().
1
It is also given by
Op( )(x) = ,x (x).
Clearly Op( ) is smooth and compactly supported.
Since
Tr (x, )()
G
is integrable against , using the dominated convergence theorem, we obtain easily
the uniform convergence of Op( ) to Op() on any compact set.
5.2
m
Symbol classes S,
and operator classes m
,
m
m
= S,
(G). By
In Section 5.2, we will dene and study classes of symbols S,
applying the quantization procedure described in Section 5.1, we will then obtain
the corresponding classes of operators
m
m
, = Op(S, ).
In Section 5.5, we will show that this collection of operators mR m
, forms an
algebra and satises the usual properties expected from a symbolic calculus.
Before dening symbol classes, we need to dene dierence operators.
5.2.1 Dierence operators
On compact Lie groups the dierence operators were dened as acting on Fourier
coecients, see Denition 2.2.6. Its adaptation to our setting leads us to (densely)
dened dierence operators on Ka,b (G) viewed as elds.
m
5.2. Symbol classes S,
and operator classes m
,
301
Denition 5.2.1. For any q C (G), we set
() (qf ),
q f() := qf
for any distribution f D (G) such that f Ka,b and qf Ka ,b for some
a, b, a , b R.
Recall that if f D (G) and q C (G), then the distribution qf D (G)
is dened via
qf, := f, q, D(G),
(5.26)
which makes sense since q D(G). In Denition 5.2.1, we assume that the two
distributions f and qf are in a ,b R Ka ,b . Note that, as all the denitions of
group Fourier transform coincide, dierent values for the parameters a, b, a , b in
and
Denition 5.2.1 yield the same elds of operators {f() : H H , G}
() : H H , G}.
This justies our use of the notation q without
{qf
reference to the parameters a, b, a , b .
Remark 5.2.2. In general, it is not possible to dene an operator q on a single
, and it has to be viewed as acting on the whole elds parametrised by G.
n
For example, already on the commutative group (R , +), the dierence operators
corresponding to coordinate functions will satisfy
1
(),
Rn ,
() =
i
with appropriately chosen functions q, thus involving derivatives in the dual variable, see Example 5.2.6. Furthermore if q is not a coordinate function but for
instance a (nonzero) smooth function with compact support, the corresponding
dierence operator is not local.
Also, on the Heisenberg group Hno (see Example 1.6.4), taking q = t the
central variable, and the Schrodinger representations (see Section 6.3.2), then
t is expressed using derivatives in , see Lemma 6.3.6 and Remark 6.3.7.
Let us x a basis of g. For the notation of the following proposition we refer
to Section 3.1.3 where the spaces of polynomials on homogeneous Lie groups have
been discussed, with the set W dened in (3.60). We will dene the dierence
operators associated with the polynomials appearing in the Taylor expansions:
Proposition 5.2.3. 1. For each Nn0 , there exists a unique homogeneous polynomial q of degree [] satisfying
1 if = ,
X q (0) = , =
Nn0
0 otherwise.
2. The polynomials q , Nn0 , form a basis of P. Furthermore, for each M
W, the polynomials q , [] = M , form a basis of P[]=M .
302
Chapter 5. Quantization on graded Lie groups
3. The Taylor polynomial of a suitable function f at a point x G of homogeneous degree M W is
(f )
q (y)X f (x).
(5.27)
Px,M (y) =
[]M
4. For any Nn0 , we have for any x, y G,
c1 ,2 q1 (x)q2 (y)
q (xy) =
[1 ]+[2 ]=[]
for some coecients c1 ,2 R independent of x and y. Moreover, we have
1 if 1 =
1 if 2 =
c1 ,0 =
, c0,2 =
.
0 otherwise
0 otherwise
Proof. For each M W, by Corollary 3.1.31, there exists a unique polynomial
q P=M satisfying X q (0) = , for every Nn0 with [] = M , therefore for
every Nn0 . This shows parts (1) and (2). Part (3) follows from the denition
of a Taylor polynomial.
It remains to prove Part (4). For this it suces to consider q (xy) as a
polynomial in x and in y, using the bases (q1 (x)) and (q2 (y)). Therefore, q (xy)
can be written as a nite linear combination of q1 (x)q2 (y). Since
q ((rx)(ry)) = r[] q (xy),
this forces this linear combination to be over 1 , 2 Nn0 satisfying [1 ] + [2 ] =
[]. The conclusions about the coecients follow by setting y = 0 and then x = 0,
see also (3.14).
In the case of (Rn , +) the polynomials q are the usual normalised monomials
(1 ! . . . n !)1 x . But it is not usually the case on other groups:
Example 5.2.4. On the three dimensional Heisenberg group H1 where a point is
described as (x, y, t) R3 (see Example 1.6.4), we compute directly that for degree
1 we have
q(1,0,0) = x, q(0,1,0) = y,
and for degree 2,
1
q(2,0,0) = x2 , q(0,2,0) = y 2 , q(1,1,0) = xy, q(0,0,1) = t xy.
2
Denition 5.2.5. For each Nn0 , the dierence operators are
:= q ,
where
Nn0 ,
q (x) := q (x1 )
and q P=[] is dened in Proposition 5.2.3.
m
5.2. Symbol classes S,
and operator classes m
,
303
The dierence operators generalise the Euclidean derivatives with respect to
the Fourier variable on (Rn , +) in the following sense:
n
Example 5.2.6. Let us consider the abelian group G = (Rn , +). We identify R
n
n
with R . If the Fourier transform of a function S(R ) is given by
n
eix (x)dx, Rn ,
FG () = (2) 2
Rn
then
FG () =
Rn
Thus, coincide with the operators D
Fourier analysis on Rn .
1
FG ().
i
= 1i
usually appearing in the
eix (x) (x)dx =
Example 5.2.7. 0 is the identity operator on each Ka,b (G).
and any Nn \{0}, we
Example 5.2.8. For I = o = {I : H H , G}
0
have I = 0.
Proof. We know that I = 0 (see Example 5.1.38). The distribution q 0 is dened
by
q 0 , = 0 , q , D(G),
see (5.26). Since
0 , q = (
q )(0) = q (0) (0) = 0
0 = 0.
we must have q0 = 0. Therefore, I = q
More generally, we have
(see
Lemma 5.2.9. Let , Nn0 . Then the symbol {(X) : H H , G}
Example 5.1.38) satises
(X) = 0
if [] > [].
If [] [], then (X) is a linear combination depending only on , , of the
terms (X)2 with [2 ] = [] [], that is,
(X) =
(X)2 .
[]+[2 ]=[]
Proof of Lemma 5.2.9. We see that (X) is the group Fourier transform of
the distribution q X 0 dened via
q }(0)
q X 0 , = X 0 , q = {X }t {
for any D(G), see Example 5.1.38. This is so as long as we prove that q X 0
is in some Ka,b . Let us nd another expression for this distribution. As {X }t is
304
Chapter 5. Quantization on graded Lie groups
a []homogeneous leftinvariant dierential operators, by the Leibniz formula for
vector elds, we have
q } =
{X }t {
X 1 q X 2 .
[1 ]+[2 ]=[]
We easily see that X 1 q P=[][1 ] and, therefore, by Part (2) of Proposition
5.2.3 we have
q .
X 1 q =
[ ]=[][1 ]
Hence we have obtained
{X }t {
q } =
q X 2 ,
[1 ]+[2 ]=[]
[ ]=[][1 ]
and
q X 0 , =
(
q X 2 )(0) =
[1 ]+[2 ]=[]
[ ]=[][1 ]
X 2 (0),
[1 ]+[2 ]=[]
0=[][1 ]
with the convention that the sum is zero if there are no such 1 , 2 . Thus
q X 0 =
X 2 0 .
[1 ]+[2 ]=[]
[]=[1 ]
Since X 2 0 K[2 ],0 (see Example 5.1.26), we see that q X 0 K[],0 . Furthermore, taking the group Fourier transform we obtain
(X) =
(X)2 .
[1 ]+[2 ]=[]
[]=[1 ]
This sum is zero if there are no such 1 , 2 , for instance if [] < [].
Let us collect some properties of the dierence operators.
Proposition 5.2.10. (i) For any Nn0 , the operator is linear, its domain
of denition contains FG (S(G)) and FG (S(G)) FG (S(G)).
(ii) For any 1 , 2 Nn0 , there exist constants c1 ,2 , R, with Nn0 such
that [] = [1 ] + [2 ], so that for any S(G), we have
1 2 = 2 1 =
c1 ,2 , ,
[]=[1 ]+[2 ]
where the sum is taken over all Nn0 satisfying [] = [1 ] + [2 ].
m
5.2. Symbol classes S,
and operator classes m
,
305
(iii) For any Nn0 , there exist constants c,1 ,2 R, 1 , 2 Nn0 , with [1 ] +
[2 ] = [], such that for any 1 , 2 S(G), we have
2 =
1 2
2 ,
1
c,1 ,2 1
(5.28)
[1 ]+[2 ]=[]
where the sum is taken over all 1 , 2 Nn0 satisfying [1 ] + [2 ] = [].
Moreover,
1 if 1 =
1 if 2 =
c,1 ,0 =
, c,0,2 =
.
0 otherwise
0 otherwise
The coecients c1 ,2 , in (ii) and c,1 ,2 in (iii) are dierent in general.
We interpret Formula (5.28) as the Leibniz formula.
Proof. Since the Schwartz space is stable under multiplication by polynomials,
q is Schwartz for any S(G), and ()
= (
q ). This shows (i).
For Part (ii), we see that the polynomial q1 q2 is homogeneous of degree
[1 ] + [2 ]. Since {q , [] = M } is a basis of P=M by Proposition 5.2.3, there exist
constants c1 ,2 , R, 1 , 2 Nn0 with [1 ] + [2 ] = [], satisfying
c1 ,2 , q .
q 1 q 2 =
[1 ]+[2 ]=[]
Therefore
= (
q1 q2 )
1 2 ()
c1 ,2 , (
q )
[1 ]+[2 ]=[]
c1 ,2 , ().
[1 ]+[2 ]=[]
This and the equality q1 q2 = q2 q1 show (ii).
Let us prove (iii). By Proposition 5.2.3 (4),
q (x) (2 1 )(x) =
q (x1 y y 1 ) 2 (y) 1 (y 1 x) dy
G
c1 ,2
q2 (y 1 )2 (y) q1 (x1 y)1 (y 1 x) dy
=
[1 ]+[2 ]=[]
c1 ,2 (
q2 2 ) (
q1 1 ),
[1 ]+[2 ]=[]
with constants depending on , 1 , 2 . Taking the Fourier transform implies the
formula (5.28), with conclusions on coecients following from Proposition 5.2.3.
We will see that the dierence operators dened in Denition 5.2.5 appear
in the general asymptotic formulae for adjoint and product of pseudodierential
operators in our context, see Sections 5.5.3 and 5.5.2.
306
Chapter 5. Quantization on graded Lie groups
m
5.2.2 Symbol classes S,
m
m
In this section we dene the symbol classes S,
= S,
(G) of symbols on a graded
Lie group G and discuss their properties. We use the notation for the symbol
classes similar to the familiar ones on the Euclidean space and also on compact
Lie groups.
Let us give the formal denition of our symbol classes.
Denition 5.2.11. Let m, , R with 0 1. Let R be a positive
Rockland operator of homogeneous degree . A symbol
= {(x, ) : H H , (x, ) G G}
is called a symbol of order m and of type (, ) whenever, for each , Nn0 and
R, we have
sup Xx (x, )L
xG
,[]m[]+
(G)
< .
(5.29)
m
m
= S,
(G) is the set of symbols of order m and of type (, ).
The symbol class S,
By Corollary 5.1.30, the symbols Xx are elds acting on smooth vectors.
By Lemma 5.1.23, we can reformulate (5.29) as
sup
xG,G
(I + R)
[]m[]+
Xx (x, )(I + R) L (H ) < .
(5.30)
Recall that, as usual, the supremum in in (5.30) has to be understood as the
essential supremum with respect to the Plancherel measure.
Clearly, the converse holds: if is a symbol such that Xx are elds
m
.
acting on smooth vectors for which (5.30) holds, then is in S,
We note that condition (5.30) requires one to x a positive Rockland operator
m
R in order to x the norms of L
a ,b (G). However, the resulting class S, does not
depend on the choice of R, see Lemma 5.1.22.
If a positive Rockland operator R of homogeneous degree is xed, then we
m
and a, b, c N0 ,
set for S,
m ,a,b,c := sup sup X (x, )
S,
x
L
c xG
[]a, []b
,[]m[]+
.
(G)
This quantity is also equal to
m ,a,b,c =
S,
sup
xG, G
m ,a,b,c ,
(x, )S,
(xed)
where we dene for any symbol , a, b, c N0 , and (x, ) G G
m ,a,b,c := sup (I + R)
(x, )S,
c
[]a, []b
[]m[]+
Xx (x, )(I + R) L (H ) .
m
5.2. Symbol classes S,
and operator classes m
,
307
Here, as always, the supremum has to be understood as the essential supremum
with respect to the Plancherel measure.
Before making some comments, let us say that the classes of symbols we have
just dened have the usual structures of symbol classes.
m
is a vector space independent of any
Proposition 5.2.12. The symbol class S,
Rockland operator R used in (5.29) to consider the L
,[]m[]+ (G)norms.
We have the continuous inclusions
m 1 m2 ,
1 2 ,
1 2
Sm11,1 Sm22,2 .
(5.31)
We x a positive Rockland operator R. For any m R, , 0, the resulting
m
m ,a,b,c , a, b, c N0 , are seminorms over the vector space S
maps S,
, which
m
endow S, with the structure of a Frechet space.
m ,a,b,c , a, b, c N0 , by
We may replace the family of seminorms S,
sup sup Xx (x, )L
[]a, xG
[]b
,[]m[]+
,
(G)
a, b N0 , Z,
where the sequence { , Z} of real numbers satises .
Two dierent positive Rockland operators give equivalent families of semim
is independent of the choice of the Rockland operator
norms. The topology on S,
R.
Proof. Using Corollary 5.1.30 and Lemma 5.1.22, this is a routine exercise.
Remark 5.2.13. Let us make some comments about Denition 5.2.11:
1. In the abelian case, that is, Rn endowed with the addition law, and R = L
m
boils down to the usual Hormander
with L being the Laplace operator, S,
class, in view of the dierence operators corresponding to the derivatives, see
Example 5.2.6.
2. In the case of compact Lie groups with R being the (positive) Laplacian,
a similar denition leads to the one considered in (2.26) since the operator
(I+R) is scalar. However, here, in the case of nonabelian graded Lie groups,
the operator R can not have a scalar Fourier transform.
3. The presence of the parameter is included to facilitate proving that the
m
, with suitable restrictions on , , forms an algebra
space of symbols mR S,
of operators later on. It already has enabled us to see that the symbols are
elds of operators acting on smooth vectors and therefore can be composed
without using the composition of unbounded operators (in Denition A.3.2).
We will see in Theorem 5.5.20 that in fact we can remove this . By
m
this we mean that a symbol is in S,
if and only if the condition in (5.29)
308
Chapter 5. Quantization on graded Lie groups
m ,a,b,0 ,
holds for any , Nn0 and = 0. Furthermore, the seminorms S,
m
a, b N0 , yield the topology of S, .
4. We could have used other families of dierence operators instead of the s
m
. For instance, we could have used any family
to dene the symbol classes S,
of dierence operators associated with a family {p }Nn0 of homogeneous
polynomials on G which satisfy
for each Nn0 , p is of homogeneous degree [],
and {p }Nn0 is a basis of P(G).
Indeed, in this case, the following properties hold.
 Any q is a linear combination of p , [] = [].
 Conversely, any p is a linear combination of q , [] = [].
Thus,
 any is a linear combination of p , [] = [].
 Conversely, any p is a linear combination of , [] = [].
m
if and only if for each
It is then easy to see that a symbol is in S,
, Nn0 and R,
sup Xx p (x, )L
xG
,[]m[]+
(G)
< .
m
does not depend on
Note that this implies that the symbol class S,
a particular choice of realisation of G through a basis of g (of eigenvectors
for the dilations) but only on the graded Lie group G and its homogeneous
structure.
For such a family p , the same proof as for Proposition 5.2.10 shows
a Leibniz formula in the sense of (5.28).
Although we could use easier dierence operators to dene our symbol
classes, for instance x , Nn0 , we choose to present our analysis with the
dierence operators given in Denition 5.2.5. Note that the asymptotic
formulae for composition and adjoint in (5.57) and (5.60) will be expressed
in terms of the dierence operators and derivatives Xx .
Note that the change of dierence operators explained just above is linear, whereas in the compact case, one can use many more dierence operators
m
, see Section 2.2.2.
to dene the symbol classes S,
The type (1, 0) can be thought of as the basic class of symbols and the
types (, ) as its generalisations. There are certain limitations on the parameters (, ) coming from reasons similar to the ones in the Euclidean settings. For
type (1, 0), we set
m
,
S m := S1,0
m
5.2. Symbol classes S,
and operator classes m
,
309
and
m ,a,b,c = (x, )a,b,c , S m ,a,b,c = a,b,c , etc. . . .
(x, )S1,0
1,0
We also dene the class of smoothing symbols
Denition 5.2.14. We set
S :=
Sm.
mR
One checks easily that
S =
m
S,
,
mR
independently of and as long as 0 1 and = 0. Moreover, S
m
is equipped with the topology of projective limit induced by mR S,
, again
independently of and .
We will see in Corollary 5.4.10 that the symbols in S really deserve to be
called smoothing.
5.2.3 Operator classes m
,
The pseudodierential operators of order m R {} and type (, ) are
obtained by the quantization
Op()(x) =
G
Tr (x)(x, )()
d(),
justied in Theorem 5.1.39, from the symbols of the same order and type, that is,
m
m
, := Op(S, ).
They inherit a structure of topological vector spaces from the classes of symbols,
m ,a,b,c .
:= S,
Op()m
, ,a,b,c
For type (1, 0), we set as for the corresponding symbol classes:
m := m
1,0 .
310
Chapter 5. Quantization on graded Lie groups
Continuity on S(G)
By Theorem 5.1.39, any operator in the operator classes dened above maps
Schwartz functions to smooth functions. Let us show that in fact it acts continuously on the Schwartz space:
Theorem 5.2.15. Let T m
, where m R, 1 0. Then for any
S(G), T S(G). Moreover the operator T act continuously on S(G): for
any seminorm S(G),N there exist a constant C > 0 and a seminorm S(G),N
such that for every S(G),
T S(G),N CS(G),N .
The constant C can be chosen as C1 T m
where C1 is a constant of and the
, ,a,b,c
seminorm m
depend
on
G,
m,
,
,
and
on the seminorm S(G),N .
, ,a,b,c
In other words, the mapping T T from m
, to the space L (S(G)) of
continuous operators on S(G) is continuous (it is clearly linear).
Our proof of Theorem 5.2.15 will require the following preliminary result on
the right convolution kernels:
m
Proposition 5.2.16. Let = {(x, )} be in S,
with 1 0. Let x
denote its associated kernel. If m < Q/2 then for any x G, the distribution x
is square integrable and
m
x L2 (G)
C sup (I + R)
x L2 (G)
C sup (x, )(I + R)
G
(x, )L (H ) ,
m
G
L (H ) ,
with C = Cm > 0 a nite constant independent of and x.
The proof below will show that we can choose Cm = Bm L2 (G) the L2 norm
of the rightconvolution kernel of the Bessel potential of the positive Rockland
operator R.
Proof of Proposition 5.2.16. We write
(x, )HS
(I + R) (I + R)
m
(x, )HS
(I + R) HS (I + R)
(x, )L (H ) ,
which shows
(x, )HS sup 1 (I + R)
1 G
(x, 1 )L (H1 ) (I + R) HS .
Squaring and integrating against the Plancherel measure, we obtain
m
m
(x, )2HS d() sup 1 (I+R) (x, 1 )2L (H )
(I+R) 2HS d().
G
1 G
G
m
5.2. Symbol classes S,
and operator classes m
,
311
By the Plancherel formula and Corollary 5.1.4, if m < Q/2, we have
m
2
:=
(I + R) 2HS d() = Bm 2L2 (G) < .
Cm
G
This gives the rst estimate in the statement. For the second estimate, we write
(x, ) = (x, )(I + R)
(I + R) ,
and adapt the ideas above.
We can now prove Theorem 5.2.15.
Proof of Theorem 5.2.15. Let T m
, where m R, 1 0. Then for
any S(G), T is smooth by Theorem 5.1.39.
Let : (x, y) x (y) be the kernel associated with T . Let R be a positive
Rockland operator of homogeneous degree . The properties of R (see Sections
4.3 and 4.4.8) yield for any S(G) and x G that
(y)x (y 1 x)dy
T (x) =
G
%
&
(I + R)N {(I + R)N }(y) x (y 1 x)dy
=
G
N x }(y 1 x)dy,
{(I + R)N }(y) {(I + R)
=
G
thus, by the CauchySchwartz inequality,
N x L2 (G) .
T (x) (I + R)N L2 (G) (I + R)
N x }() = (x, )(I + R)N yields a symbol in S mN , by
Since FG {(I + R)
,
Proposition 5.2.16, we have
N x L2 (G) C sup (x, )(I + R)N L (H ) ,
(I + R)
G
whenever m N < Q/2. Note that in this case,
N +
m ,0,0,m (I + R)
sup (x, )(I + R)N L (H ) S,
L (H ) ,
m
G
and by functional calculus
(I + R)N + L (H ) sup(1 + )N + 1.
m
Thus if we choose N N0 such that N > (m +
Q
2 )/,
then
N
m ,0,0,m (I + R)
L2 (G) .
T (x) CS,
312
Chapter 5. Quantization on graded Lie groups
This shows that T is bounded.
Let Nn0 . Using the Leibniz property of vector elds, we easily obtain
X T (x) =
c1 ,2 ,
(y)Xx11=x Xx22=y1 x x1 (x2 )dy.
G
[1 ]+[2 ]=[]
As above, we can insert powers of I + R. Noticing that the symbol
1
2
2 1
N
N
FG {(I + R)
x1 Xx1 =x X x1 } = (X) Xx (x, )(I + R)
m+[1 ]+[2 ]N
is in S,
X T (x) C1
, we proceed as above to obtain
(I + R)N L2 (G) (X)2 Xx1 (x, )(I+R)N L2 (G)
[1 ]+[2 ]=[]
N
m ,0,[],m+[] (I + R)
C2 S,
L2 (G) .
as long as N > (m + [] + Q
2 )/.
Let Nn0 . Proceeding as in the proof of Proposition 5.2.3 (4), we can write
(xy) =
c,1 ,2 q1 (x) q2 (y).
[1 ]+[2 ]=[]
Using this, we easily obtain
x T (x) =
(y y 1 x) (y)x (y 1 x)dy
G
c,1 ,2
q1 (y)(y)q2 (y 1 x)x (y 1 x)dy.
=
[1 ]+[2 ]=[]
Noticing that
N {q x } = {2 (x, )} (I + R)N S mN [2 ] ,
FG {(I + R)
2
,
we can now proceed as in the rst paragraph above to obtain
N {q x }2
(I + R)N
x T (x) C1
y {q1 }2 (I + R)
2
[1 ]+[2 ]=[]
m ,[],0,m+[]
C2 (x, )S,
(I + R)N
y {q1 }2
[1 ][]
as long as N > (m + Q/2)/.
We can combine the two paragraphs above to show that for any , Nn0 ,
we have
x X T (x) C(x, )S m ,[],[],m+[]+[]
(I + R)N
y {q1 }2 ,
,
[1 ][]
m
5.2. Symbol classes S,
and operator classes m
,
313
as long as N > (m + [] + Q/2)/. By Lemma 3.1.56, we have
(I + R)N
y {q1 }2 C S(G),N
[1 ][]
for some N N depending on N and , and T is a Schwartz function. Furthermore, these estimates also imply the rest of Theorem 5.2.15.
Theorem 5.2.15 shows that composing two operators in (possibly dierent)
m
makes
sense as the composition of operators acting on the Schwartz space.
,
m2
1
We will see that in fact, the composition of T1 m
, with T2 , is T1 T2 in
1 +m2
, see Theorem 5.5.3.
m
,
We will see that our classes of pseudodierential operators are stable under
taking the formal L2 adjoint, see Theorem 5.5.12. This together with Theorem
5.2.15 will imply the continuity of our operators on the space S (G) of tempered
distributions, see Corollary 5.5.13.
Returning to our exposition, before proving that the introduced classes of
m
and of the corresponding operators mR m
symbols mR S,
, are stable under
composition and taking the adjoint, let us give some examples.
5.2.4
First examples
As it should be, mR m contains the leftinvariant
dierential operators. More
precisely, the following lemma implies that []m c X m . The coecients
c here are constant and it is easy to relax this condition with each function c
being smooth and bounded together with all of its left derivatives.
Lemma 5.2.17. For any o Nn0 , the operator X o = Op((X)0 ) is in [o ] .
Proof. By Lemma 5.2.9, we have
0
(X)o =
(X)
if [] > [o ],
if [] [o ].
[]+[2 ]=[o ]
L
Recall that, by Example 5.1.26, {(X) , G}
+[], (G) for any
is zero if [] > [o ] whereas it is in
R, Nn0 . So { (X)o , G}
for any R if [] [o ].
L+[o ][], (G)
Remark 5.2.18. Lemma 5.2.17 implies that mR m contains the leftinvariant
dierential calculus, that is, the space of leftinvariant dierential operators.
One could wonder whether it also contains the rightinvariant dierential
calculus, since we can quantize any dierential operator, see Remark 5.1.41 (1).
This is false in general, see Example 5.2.19 below. Thus, if one is interested in
dealing with problems based on the setting of rightinvariant operators one can
314
Chapter 5. Quantization on graded Lie groups
use the corresponding version of the theory based on the rightinvariant Rockland
operator, see Remark 5.1.41 (4).
Example 5.2.19. Let us consider the three dimensional Heisenberg group H1 and
the canonical basis X, Y, T of its Lie algebra (see Example 1.6.4). Then the right
can not be in mR m .
invariant vector eld X
Proof of the statement in Example 5.2.19 . We have already seen that any operator A m acts continuously on the Schwartz space, cf. Theorem 5.2.15. We will
see later (see Corollary 5.7.2) that it also acts on Sobolev spaces with a loss of
derivative controlled by its order m. By this, we mean that, if an operator A in
m is homogeneous of degree A , then we must have
s R
C > 0 f S(G)
Af L2sm Cf L2s ,
and when s + m and s are nonnegative, we realise the Sobolev norm as f L2s =
s
f L2 + R f L2 for some positive Rockland operator of degree , cf. Theorem
4.4.3 Part (2). Applying the inequality to dilated functions f Dr and letting
r yield that m A .
shows that if X
were in some m then m 1
Applying this to the case of X
2
2
2
and X would map L1 to L1m hence to L continuously. We have already shown
in the proof of Example 4.4.32 that this is not possible.
An example of a smoothing operator is given via convolution with a Schwartz
function:
Lemma 5.2.20. Let S(G). We denote by T : the corresponding
convolution operator. Its symbol T is independent of x and is given by
T () =
().
Furthermore, the mapping
S(G) T
is continuous.
Proof. For the rst part, see Example 5.1.38 and its continuation.
For any S(G), we have q S(G) for any Nn0 , and
b S(G)
(I + R)a (I + R)
for any a, b N (see also (4.34) and Proposition 4.4.30). For any m R, R
and Nn0 , we have by (1.38)
L
,[]m+
(G)
(I + R)
(I + R)
[]m+
[]m+
()(I + R) L (G)
{
(I + R)
q }L1 (G) .
m
5.2. Symbol classes S,
and operator classes m
,
315
As S(G), this L1 norm is nite and this shows that T . More precisely,
this L1 norm is less or equal to
1
B 1 (I + R)a {
q }1
if and []m+
> 0 and a = []m+
,
[]m+
B []m+ 1 (I + R) {
q }1 if and
< 0 and b = ,
where x denotes the smallest integer > x and B is the rightconvolution kernel
of the Bessel potential of R, see Corollary 4.3.11. By Proposition 4.4.27, these
quantities can be estimated by Schwartz seminorms.
More generally, the operators and symbols with kernels depending on x but
satisfying the following property are smoothing:
Lemma 5.2.21. Let : (x, y) x (y) be a smooth function on G G such that,
for each multiindex Nn0 and each Schwartz seminorm S(G),N , the following
quantity
sup Xx x S(G),N < ,
xG
is nite.
Then the symbol given via (x, ) =
x () is smoothing. Furthermore for
any seminorm S m ,a,b,c , there exists C > 0 and Nn0 , N N0 such that
S m ,a,b,c C sup Xx x S(G),N .
xG
Proof of Lemma 5.2.21. By (1.38), we have
sup (x, )L (H ) = sup
x ()L (H ) x L1 (G) .
G
G
More generally, for any 1 , 2 R, denoting by N1 , N2 N0 integers such that
1 N1 2 N2 , we have
sup (I + R)1 Xx (x, ) (I + R)2 L (H )
G
sup (I + R)N1 Xx (x, ) (I + R)N2 L (H )
G
N2 X q x }()L (H )
= sup FG {(I + R)N1 (I + R)
x
G
N2 q Xx x L1 (G) .
(I + R)N1 (I + R)
This last L1 norm is, up to a constant, less or equal than a Schwartz seminorm of
Xx x , see Section 3.1.9. This implies the statement.
In Theorem 5.4.9, we will see that the converse holds, that is, that any
smoothing operator has an associated kernel as in Lemma 5.2.21.
316
Chapter 5. Quantization on graded Lie groups
5.2.5
First properties of symbol classes
We summarise in the next theorem some properties of the symbol classes which
follow from their denition.
Theorem 5.2.22. Let 1 0.
m
(i) Let S,
have kernel x and order m R.
1. For every x G and R,
q X x K,[]m[]+ .
is in S m+[o ]
2. If o Nn0 then the symbol {Xxo (x, ), (x, ) G G}
,
with kernel Xxo x , and
m ,a,b+[ ],c .
Xxo (x, )S m+[o ] ,a,b,c Cb,o (x, )S,
o
,
is in S
3. If o Nn0 then the symbol {o (x, ), (x, ) G G}
,
with kernel qo x , and
m[o ]
m ,a+[ ],b,c .
o (x, )S m[o ] ,a,b,c Ca,o (x, )S,
o
,
4. The symbol
:= {(x, ) , (x, ) G G}
m
with kernel x given by
is in S,
x (y) =
x (y 1 ),
and
m ,a,b,c =
(x, ) S,
sup (I + R) Xx (x, )(I + R)
c
[]a, []b
[]m[]+
L (H ) .
m1
m2
(ii) Let 1 S,
and 2 S,
have kernels 1x and 2x , respectively. Then
(x, ) := 1 (x, )2 (x, )
m
denes the symbol in S,
, m = m1 + m2 , with kernel 2x 1x with the
convolution in the sense of Denition 5.1.19. Furthermore,
m
m ,a,b,c C1 (x, ) m1
(x, )S,
S
,a,b,c+a+m2 +b 2 (x, )S 2 ,a,b,c ,
,
where the constant C = Ca,b,c,m1 ,m2 > 0 does not depend on 1 , 2 .
m
5.2. Symbol classes S,
and operator classes m
,
317
Note that, in Part (ii), the composition (x, ) := 1 (x, )2 (x, ) may be
understood as the composition of two elds of operators acting on smooth vectors
as well as the composition of 1 (x, ) L
1 ,1 m1 (G) with 2 (x, ) L2 ,2 m2 (G)
for any choice of 1 , 2 R such that 1 m1 = 2 .
Proof. Properties (1), (2), (3), and (4) of (i) are straightforward to check.
Let us prove Part (ii). By Property (1) of (i), or by the denition of symbol
classes,
jx Kj ,mj +j for any j R, j = 1, 2,
thus choosing = 2 and 1 = m2 + 2 , we have by Corollary 5.1.20
2x 1x K,m+
for any
R.
Its group Fourier transform is
(1x )(2x ) = 1 (x, )2 (x, ) = (x, ).
Therefore, is a symbol with kernel 2x 1x .
Let , Nn0 and R. From the Leibniz rules for (see Proposition
5.2.10) and X , the operator
(I + R)
[]m[]+
Xx (x, )(I + R) ,
is a linear combination over 1 , 2 , 1 , 2 Nn satisfying [1 ] + [2 ] = [], [1 ] +
[2 ] = [], of terms
(I + R)
[]m[]+
Xx1 1 1 (x, )Xx2 2 2 (x, )(I + R) ,
whose operator norm is bounded by
(I + R)
[]m[]+
(I + R)
Xx1 1 1 (x, )(I + R)
[2 ]m2 [2 ]+
[2 ]m2 [2 ]+
Xx2 2 2 (x, )(I + R)
L (H )
L (H ) .
This shows that the inequality between the seminorms of , 1 and 2 given in
(ii) holds. Consequently is a symbol of order m = m1 + m2 and of type (, ),
and (ii) is proved.
A direct consequence of Part (ii) of Theorem 5.2.22 is that the symbols in
the introduced symbol classes form an algebra:
m
forms
Corollary 5.2.23. Let 1 0. The collection of symbols mR S,
an algebra.
m
is of order m R, then 0 and
Furthermore, if 0 S and S,
0 are also in S
.
318
Chapter 5. Quantization on graded Lie groups
m
The fact that the symbol classes mR S,
form
an algebra does not imply
directly the same property for the operator classes mR m
, since our quanti) is not equal in general to
zation is not an algebra morphism, that is, Op(1 2
Op(1 )Op(2 ). However, we will show that indeed mR m
, is an algebra of
operators, cf. Theorem 5.5.3, and we will often use the following property:
Lemma 5.2.24. Let 1 and 2 be as in Theorem 5.2.22, (ii). We assume that 2
Then
does not depend on x: 2 = {2 () : G}.
(x, ) := 1 (x, )2 ()
m
denes the symbol in S,
, m = m1 + m2 and
Op() = Op(1 )Op(2 )
Proof. We keep the notation of the statement. Let 1x and 2 be the convolution
kernels of 1 and 2 respectively. Hence 2 is a function on G independent of x.
By Theorem 5.2.22(ii), 2 1x is the convolution kernel of , thus
S(G)
Op()()(x) = (2 1x ).
As 2 = Op(2 ), this implies easily that Op() is the composition of Op(1 )
with Op(2 ).
The following will also be useful, for instance in the estimates for the kernels
in Section 5.4.1.
m
have kernel x . If 1 and 2 are
Corollary 5.2.25. Let 1 0. Let S,
n
in N0 , then
S m+[1 ]+[2 ]
{(X)1 (x, ) (X)2 , (x, ) G G}
,
y2 x (y). Furthermore, for any a, b, c there exists C = Ca,b,c, ,
with kernel Xy1 X
1 2
independent of such that
m ,a,b,c C m
(X)1 (x, )(X)2 S,
S, ,a,b,c+a+[1 ]+[2 ]+b .
If 2 = 0, for any a, b, c there exists C = Ca,b,c,1 independent of such that
m ,a,b,c CS m ,a,b,c .
(X)1 S,
,
Proof. The rst part follows directly from Theorem 5.2.22 Part (ii) together with
Lemma 5.2.17.
We need to show a better estimate for 2 = 0. Let , o Nn0 . By the Leibniz
formula (see (5.28)), we have
Xxo {(X)1 (x, )}
c,1 ,2 {1 (X)1 } {Xxo 2 (x, )}.
=
[1 ]+[2 ]=[]
5.3. Spectral multipliers in positive Rockland operators
319
Hence, denoting mo := m + [o ], we have
[]mo [1 ]+
(I + R)
Xxo {(X)1 (x, )}(I + R) L (H )
[]mo [1 ]+
[2 ]mo +
C
(I + R)
1 (X)1 (I + R)
L (H )
[1 ]+[2 ]=[]
(I + R)
[2 ]mo +
Xxo 2 (x, )(I + R) L (H ) .
As {(X)1 } S1,01 by Lemma 5.2.17, each quantity
[ ]
sup
c,G
(I + R)
[]mo [1 ]+
1 (X)1 (I + R)
[2 ]mo +
L (H ) <
is nite for any c > 0 and 1 , 2 Nn0 such that [1 ] + [2 ] = []. This implies
sup
c,G
C
(I + R)
[]mo [1 ]+
sup (I + R)
[2 ][]
Xxo {(X)1 (x, )}(I + R) L (H )
[2 ]mo +
c
G
Xxo 2 (x, )(I + R) L (H ) .
Taking the supremum over [] a and [] b yields the stated estimate.
5.3
Spectral multipliers in positive Rockland operators
In this section we show that multipliers in positive Rockland operators belong to
the introduced symbol classes m .
The main result is stated in Proposition 5.3.4. This will allow us to use the
LittlewoodPaley decompositions associated with a positive Rockland operator,
and therefore will enter most of the subsequent proofs.
5.3.1 Multipliers in one positive Rockland operator
The precise class of multiplier functions that we consider is the following:
Denition 5.3.1. Let Mm be the space of functions f C (R+ ) such that the
following quantities for all N0 are nite:
f Mm, :=
sup
>0, =0,...,
(1 + )m+  f ().
In other words, the class of functions f that appears in the denition above
are the functions which are smooth on R+ = (0, ) and have the symbolic bem
(R) on the real line. However, we
haviour at innity of the Hormander class S1,0
rather prefer the notation Mm in order not to create any confusion between these
m
(G) dened on the group G.
classes and the classes S,
320
Chapter 5. Quantization on graded Lie groups
Example 5.3.2. For any m R, the function (1 + )m is in Mm .
It is a routine exercise to check that Mm endowed with the family of maps
Mm, , N0 , is a Frechet space. Furthermore, it satises the following property.
Lemma 5.3.3. If f1 Mm1 and f2 Mm2 then f1 f2 Mm1 +m2 with
f1 f2 Mm1 +m2 , C f1 Mm1 , f2 Mm2 , .
Proof. This follows from the Leibniz formula for  (f1 f2 ) and from the following
inequality which holds for > 0 and 1 , 2 :
(1 + )m1 m2 +1 +2 1 f1 () 2 f2 () f1 Mm1 , f2 Mm2 , ,
which implies the claim.
The main property of this section is
Proposition 5.3.4. Let m R and let R be a positive Rockland operator of homoge
neous degree . If f M m
, then f (R) is in m and its symbol {f ((R)), G}
satises
a, b, c N0
N, C > 0 :
f ((R))a,b,c Cf M m , ,
with and C independent of f .
Proof. First let us show that it suces to show Proposition 5.3.4 for m < . If
with m , then we dene
f Mm
m2 such that
f1 () := (1 + )
m2
m2
is the smallest integer strictly larger than
f () and f2 () := (1 + )
m2
m
,
By Example 5.3.2 and Lemma 5.3.3, we see that f1 M m1 with m1 = m m2 .
By Lemma 5.2.17, we see that f2 ((R)) S m2 . If Proposition 5.3.4 holds for
m1 < , then we can apply it to f1 and hence f1 ((R)) S m1 . Thus the
product
f ((R)) = f1 ((R))f2 ((R))
is in S m1 +m2 = S m .
Therefore, as claimed above, it suces to show Proposition 5.3.4 for m < .
Now we show that we may assume that f is supported away from 0. Indeed,
if f M m
, we extend it smoothly to R and we write
f = f o + f (1 o ),
where o D(R) is identically 1 on [1, 1]. Since f o D(R), by Hulanickis
theorem (cf. Corollary 4.5.2), the kernel of (f o )(R) is Schwartz and by Lemma
5.2.20, we have (f o )(R) with suitable inequalities for the seminorms.
Thus we just have to prove the result for f (1 o ) which is supported in [1, )
where
1 + . The statement then follows from the following lemma.
5.3. Spectral multipliers in positive Rockland operators
321
Showing Proposition 5.3.4 boils then down to
Lemma 5.3.5. Let m < . If f C (R) is supported in [1, ) and satises
N0
 f () C  ,
m
then f (R) m , and for any a, b, c N0 we have
f (R)m ,a,b,c C
sup
 +  f (),
m
1, =0,...,
with = m,a,b,c N and C = Cm,a,b,c > 0 independent of f .
The proof of Lemma 5.3.5 relies on the following consequence of Hulanickis
theorem (see Theorem 4.5.1).
Lemma 5.3.6. Let R be a positive Rockland operator on a graded Lie group G.
Let m D(R) and o Nn0 . We denote by m(R)0 the kernel of the multiplier
m(R) and we set
(x) := xo m(R)0 (x).
The function is Schwartz.
For any p (1, ), N N and a R with 0 a N , there exist C > 0
and k N such that for any S(G),
a
RN ( )p C sup (1 + )k  m() Rp Lp (G) .
>0
=0,...,k
Proof of Lemma 5.3.6. By Hulanickis Theorem 4.5.1 or Corollary 4.5.2, S(G).
It suces to prove the result with X , [] = N , instead of RN . By Corollary
p, with p, a homogeneous
3.1.30, we can write X as a nite sum of X
polynomial of homogeneous degree [] [] 0. We then have
p, ) =
(X
(X ) (p, ).
X ( ) = X =
Therefore, by Proposition 4.4.30,
[]+a
[]a
p, )p
(R X ) (R
X ( )p
[]+a
[]a
p, 1 .
R X p R
By Theorem 4.4.16, Part 2,
R
And we have
R
[]+a
[]a
X p CR p .
p, 1 = R
[]a
p,
1 ,
322
Chapter 5. Quantization on graded Lie groups
see Section 4.4.8. By Theorem 4.3.6, since [] [] = N a, we obtain
R
[]a
p,
1 C
p,
1
[]a
N
[]a
RN p,
1 N .
Note that because of (4.8), we have
(x) := (1)o  xo m(R)
0 (x).
b
By Hulanickis theorem (see Theorem 4.5.1),
p,
1 and R p,
1 are
sup (1 + )k  m(),
>0
=0,...,k
[]a
p, 1 .
for a suitable k, therefore this is also the case for R
Combining all these inequalities shows the desired result.
Proof of Lemma 5.3.5. Let f be as in the statement. We need to show for any
Nn0 that the convolution operator with right convolution kernel q f (R)0
maps L2 (G) boundedly to L2[]m+ (G) for any R. It is sucient to prove this
for in a sequence going to + and (see Proposition 5.2.12) and, in fact,
only for a sequence of positive since
(
q f (R)0 ) = (1) q f(R)0 .
At the end of the proof, we will see that, because of the equivalence between the
Sobolev norms, it actually suces to prove that for a xed in this sequence, the
operators given by
[]m+
q f (R)0 ) , (5.32)
R (
(
q f (R)0 ) and R
are bounded on L2 (G). So, we rst prove this by decomposing f and applying the
CotlarStein lemma.
We x a dyadic decomposition: there exists a nonnegative function D(R)
supported in [1/2, 2] and satisfying
1
1=
j ()
where
j () := (2j ).
jN0
We set for j N0 and 1,
fj ()
:=
f ()j (),
f (j) ()
:=
fj (2j ),
gj ()
:=
f (j) ().
5.3. Spectral multipliers in positive Rockland operators
323
One obtains easily that for any j N0 and N0 , we have
fj ()
1 ( 2 f )() 2j3 ( 3 )(2j ),
m
1 +2 +3 =
 fj ()
where
C sup +  f ()
m
1
1 2 2j3 ( 3 )(2j ),
1 +2 +3 =
stands for a linear combination of its terms with some constants. As
is supported in [1/2, 2] and since
2j , we have
1 2 2j3
2j1 +2 +3 ,
so that
1 2 2j3 ( 3 )(2j ) C, 2j .
1 +2 +3 =
Therefore, we have obtained
 fj () C sup +  f () 2j .
m
1
Hence, for each j N0 , f (j) is smooth and supported in [1/2, 2], and satises for
any N0 the estimate
 f (j) () = 2j fj () C sup +  f ().
m
1
Consequently, each gj is smooth and supported in [1/2, 2], and satises
N0
sup  gj () C sup +  f ().
m
(5.33)
1
[ 12 ,2]
=0,...,
Clearly f () is the sum of the terms
2j gj (2j ) = f ()j ()
m
over j N0 and this sum is uniformly locally nite with respect to . Furthermore,
since the functions f and gj are continuous and bounded, the operators f (R) and
gj (2j R) dened by the functional calculus are bounded on L2 (G) by Corollary
4.1.16. Therefore, we have in the strong operator topology of L (L2 (G)) that
f (R) =
j=0
2j gj (2j R),
m
324
Chapter 5. Quantization on graded Lie groups
and in K(G) or S (G) that
f (R)o =
2j gj (2j R)o .
m
j=0
We x Nn0 . For each j N0 , by Hulanickis theorem (see Corollary 4.5.2),
gj (2 R)o is Schwartz, thus so is
j
Kj := 2j q gj (2j R)o
m
and also (see (4.8))
j (x1 ) = (1) 2j q gj (2j R)o (x1 ).
Kj =: Kj (x) = K
m
We claim that for any a, b R satisfying
either b N0 and a [0, b)
or b 0 and a <
b/
there exist N and C > 0 such that for all j N0 , we have
j
R Kj K C(2 )m[]a+b sup +  f (),
R
a
(5.34)
1
K .
and the same is true for R R
j
a
Let us prove this claim. By homogeneity (see (4.3)), we see that
j
gj (2j R)o (x) = (2 )Q gj (R)o (2 x),
thus
Kj (x)
2j (2 )[] q (2 x) (2 )Q gj (R)o (2 x)
q gj (R)o ) (2 x).
(2 )m[]+Q (
m
j
(see
More generally, by Part (7) of Theorem 4.3.6 for R and consequently for R
(4.50)) we have
a R b {
a R b Kj = (2 j )m[]+Qa+b R
q gj (R)o } D j ,
R
2
whenever it makes sense (that is, Kj is in the L2 domain of R such that R Kj
a ). Consequently, by Proposition 5.1.17 (1), with norms
is in the L2 domain of R
possibly innite, we have
a R b {
a R b Kj K = (2 j )m[]a+b
q gj (R)o } .
R
R
K
5.3. Spectral multipliers in positive Rockland operators
325
a R b Kj ) = R
b R a K for any a, b whenever it makes sense, or by the
Since (R
j
same argument as above, we also have
a R b K K = (2 j )m[]a+b
a R b {
q gj (R)o } .
R
R
j
K
Therefore, if b N0 and a [0, b), by Lemma 5.3.6, there exist = a,b N such
that
a b
q gj (R)o }
Ca,b sup (1 + )  gj ()
R R {
K
>0
=0,...,
Ca,b
sup  gj (),
>0
=0,...,
since each gj is supported in [1/2, 2]. As gj satises (5.33), we have shown Claim
(5.34) in the case b N0 and a [0, b).
If a <
b/ then we can apply the result we have just obtained to (
b/)
and b/ . Using Theorem 4.3.6 we then have for any S(G), with :=
b b 1 , that
b
R 2
CR 1
R 2
2
1+
a
C sup  gj () R 2
>0
=0,...,
for some . This shows Claim (5.34) in the case a <
b/.
We set Tj : S(G) Kj . We want to apply the CotlarStein lemma
(Theorem A.5.2) to two families of L2 (G)bounded operators: rst to Tj , j N0 ,
and then to
Kj , j N0 .
Tj,, : R R
where N is such that := [] m + > 0.
Let us check the hypothesis of the CotlarStein lemma for Tj . By Claim (5.34)
for a = b = 0, there exists N0 such that for any j, k N0 ,
max Tj Tk L (L2 (G)) , Tj Tk L (L2 (G))
C max Tj L (L2 (G)) Tk L (L2 (G)) , Tj L (L2 (G)) Tk L (L2 (G))
C2
C2
j+k
(m[])
(sup +  f ())2
m
1
jk
(m[])
(sup +  f ())2 ,
m
1
326
Chapter 5. Quantization on graded Lie groups
since m [] < 0.
Let us check the hypothesis of the CotlarStein lemma for Tj,, . By Propo
sition 4.4.30 the right convolution kernel of the operator Tj,,
Tk,, is given by
Kk ) (R
R Kj ) = (R
R Kk ) (R
2
R Kj ).
(R R
Therefore, its operator norm is
R Kk K R
2
Tj,,
Tk,, L (L2 (G)) R
R Kj K .
j
k
m
2 (m[]+) 2 (m[]+2) sup +  f () ,
1
for some , thanks to Claim (5.34) with a = b = N and with b = 2 =
2[] 2m + and a = . So we have obtained
Tk,, L (L2 (G)) 2
Tj,,
kj
(m[])
m +
 f () .
sup
1
Since the adjoint of Tj,,
Tk,, is Tk,,
Tj,, , we may replace k j above by
k j.
whose right
We proceed in a similar way for the operator norm of Tj,, Tk,,
convolution kernel is
Kk ) (R
Kk ) (R
R Kj ) = (R 2
R
K ).
R
(R R
j
Therefore, we obtain
Tk,, L (L2 (G)) , Tj,, Tk,,
L (L2 (G))
max Tj,,
C2
kj
(m[])
(sup +  f ())2 .
m
1
By the CotlarStein lemma (see Theorem A.5.2),
Tj and j Tj,, converge in the strong operator topology of L (L2 (G)) and the resulting operators
have operator norms, up to a constant, less or equal than
sup +  f ().
m
1,k
Clearly
Tj and j Tj,, coincide on S(G) with the operators in (5.32), respectively. Using the equivalence between the two Sobolev norms (Theorem 4.4.3, Part
5.3. Spectral multipliers in positive Rockland operators
327
4), this implies
(
q f (R)0 )L2 (G)
C (
q f (R)0 )2 + R ( (
q f (R)0 ) 2
m
C sup +  f () 2 + R 2
C sup +  f ()L2 (G) .
1
1
We have obtained that the convolution operator with the right convolution
kernel q f (R)0 maps L2 (G) boundedly to L2m[]+ (G) for any N such that
m [] + > 0, with operator norm less or equal than
sup
1,
+  f (),
m
up to a constant, with depending on . This concludes the proof of Lemma
5.3.5.
Hence the proof of Proposition 5.3.4 is now complete.
Looking back at the proof of Proposition 5.3.4, we see that we can assume
that f depends on x G in the following way:
Corollary 5.3.7. Let R be a positive Rockland operator of homogeneous degree .
Let m R and 0 1. Let
f : G R+ (x, ) fx () C
be a smooth function. We assume that for every Nn0 , Xx fx M m+[] . Then
m
(x, ) = fx ((R)) denes a symbol in S1,
which satises
a, b, c N0
N, C > 0 :
m ,a,b,c C sup X fx M
S1,
,
x
m+[] ,
[]b
with and C independent of f .
5.3.2 Joint multipliers
To a certain extent, we can tensorise the property in Proposition 5.3.4. But we need
to dene the tensorisation of the space Mm and the multipliers of two Rockland
operators.
First, we dene the space Mm1 Mm2 of functions f C (R+ R+ ) such
that
f Mm1 Mm2 , :=
sup
1 ,2 >0
1 ,2 =0,...,
(1 + 1 )m1 +1 (1 + 2 )m2 +2 11 22 f (1 , 2 ),
328
Chapter 5. Quantization on graded Lie groups
is nite for every N0 . It is a routine exercise to check that Mm1 Mm2 is a
Frechet space.
Secondly, we observe that if L and R are two Rockland operators on G which
commute strongly, meaning that their spectral measures EL and ER commute,
then we can dene their common spectral measure EL,R via
EL,R (B1 B2 ) := EL (B1 )ER (B2 ),
for B1 , B2 Borel subsets of R,
and we can also dene the multipliers in L and R by
f (L, R) :=
f (1 , 2 )dEL,R (1 , 2 ),
R+ R+
for any f L (R+ R+ ).
Corollary 5.3.8. Let L and R be two positive Rockland operators on G of respective
degrees L and R . We assume that L and R commute strongly, that is, their
spectral measures EL and ER commute. If f M m
1 M m2 then f (L, R) is in
m1 +m2 . Furthermore, we have for any a, b, c N0 ,
f (L, R)m1 +m2 ,a,b,c Cf M m1 M m2 , ,
L
where and C > 0 are independent of f .
Proof. By uniqueness, the spectral measure EL,R is invariant under left transla its group Fourier transform, we see that the
tions. Denoting by (EL,R ) for G
group Fourier transform of a multiplier f (L, R) for f L (R+ R+ ) is
(f (L, R)) =
f (1 , 2 )d(EL,R )(1 , 2 ),
R+ R+
since it is true for a function f of the form f (1 , 2 ) = f1 (1 )f2 (2 ) with f1 , f2
L (R+ ), by Corollary 5.3.7.
We x C (R) supported in [ 12 , 12 ] such that
R
( + j ) = 1.
j Z
We also x another function C (R) supported in [1, 1] such that = 1 on
[ 12 , 12 ]. For any j , k Z, we dene j ,k C (R) by
j ,k ( ) := eik ( j ) ( j ).
It is easy to show that for any N0 there exists C = C > 0 such that
j , k Z
j ,k Mm , C(1 + k ) (1 + j )m+ .
5.3. Spectral multipliers in positive Rockland operators
329
Since the symbols form an algebra (see Section 5.2.5), and by Proposition 5.3.4,
writing m = m1 + m2 , we have for any j1 , j2 , k1 , k2 Z:
j1 ,k1 ((L))j2 ,k2 ((R))S m ,a,b,c
Cj1 ,k1 ((L))S m1 ,a1 ,b1 ,c1 j2 ,k2 ((R))S m2 ,a2 ,b2 ,c2
C(1 + k1 )1 (1 + j1 )
1 +1
L
(1 + k2 )2 (1 + j2 )
m2
+2
(5.35)
for some 1 , 2 N0 .
Let f be as in the statement. We extend f to a smooth function supported
in (1, )2 and decompose it as a locally nite sum:
f=
fj
where fj () = f ()(1 j1 )(2 j2 ), = (1 , 2 ).
jZ2
For each j Z, we view fj (+j) as a smooth function supported in [1, 1][1, 1]
and we expand it in the Fourier series
fj ( + j) =
cj,k eik .
kZ2
The hypothesis on f implies that for any 1 , 2 N0 , we have
cj,k 
C1 ,2 f M m1 M m2 ,1 +2 (1 + k1 )1 (1 + k2 )2
(5.36)
m1
(1 + j1 ) L
m2
(1 + j2 ) R
We have obtained that (taking dierent s)
cj,k j1 ,k1 M m1 ,1 j2 ,k2 M m2 ,2 < .
L
j,kZ2
We have therefore obtained the following decomposition of f in the Frechet
space M m
1 M m2 ,
f (1 , 2 ) =
cj,k j1 ,k1 (1 )j2 ,k2 (2 ).
j,kZ2
And so for any a, b, c with 1 , 2 as in (5.35),
f ((L), (R))S m ,a,b,c
cj,k j1 ,k1 ((L))j2 ,k2 ((R))S m ,a,b,c
j,kZ2
cj,k C(1 + k1 )1 (1 + j1 )
1 +1
L
(1 + k2 )2 (1 + j2 )
m2
R
+2
j,kZ2
Cf M m1 M m2 ,1 +2 +4 ,
L
by (5.37) with 1 + 2 and 2 + 2. This shows that f ((L), (R)) S m and the
desired inequalities for the seminorms.
330
Chapter 5. Quantization on graded Lie groups
Corollary 5.3.8 could be generalised by considering a nite family of positive
Rockland operators which commute strongly between themselves (i.e. with commuting spectral measures), with symbols possibly depending on x in a similar way
to Corollary 5.3.7.
5.4 Kernels of pseudodierential operators
In this section we obtain estimates for the kernels of operators in the classes m
,
(cf. Section 5.4.1) and some consequences for smoothing operators (cf. Section
5.4.2) and for operators of CalderonZygmund type in the calculus (cf. Section
5.4.4). We will also show the Lp boundedness of 0 in Section 5.4.4.
For technical reasons which will become apparent in Section 5.5.2, we will
also consider the seminorms:
S m,R ,a,b :=
,
sup
(x,)GG
[]a,[]b
Xx (x, )(I + R)
m[]+[]
L (H ) ,
(5.37)
where R is a positive Rockland operator of homogeneous degree . The superscript
m
seminorms,
R indicates that the powers of I + R are on the right. As for the S,
this is a seminorm which is equivalent to a similar seminorm for another positive
Rockland operator.
5.4.1 Estimates of the kernels
This section is devoted to describing the behaviour of the kernel of an operator
m
with symbol in the class S,
. As usual in this chapter, G is a graded Lie group of
homogeneous dimension Q. Our results in this section may be summarised in the
following theorem.
m
with 1 0, = 0. Then
Theorem 5.4.1. Let = {(x, )} be in S,
its associated kernel : (x, y) x (y) is smooth on G (G\{0}). We also x a
homogeneous quasinorm   on G.
(i) Away from 0, x has a Schwartz decay:
M N C > 0, a, b, c N : (x, y) G G
M
m ,a,b,c y
y > 1 = x (y) C supG (x, )S,
.
(ii) Near 0, we have
if Q+m > 0, x behaves like y
such that
(x, y) G (G\{0})
Q+m
: there exists C > 0 and a, b, c N
m ,a,b,c y
x (y) C sup (x, )S,
G
Q+m
5.4. Kernels of pseudodierential operators
331
if Q + m = 0, x behaves like ln y: there exists C > 0 and a, b, c N
such that
(x, y) G (G\{0})
m ,a,b,c ln y;
x (y) C sup (x, )S,
G
if Q + m < 0, x is continuous on G and bounded:
m ,0,0,0 .
sup x (z) C sup (x, )S,
zG
G
m ,a,b,c in (i) and (ii)
Moreover, it is possible to replace the seminorm S,
with a seminorm S m,R ,a,b given in (5.37).
,
Remark 5.4.2. Using Theorem 5.2.22 (i) Parts (3) and (2), and Corollary 5.2.25,
2 (X o q (y)x (y)).
we obtain similar properties for Xy1 X
y
x
We start the proof of Theorem 5.4.1 with consequences of Proposition 5.2.16
as preliminary results on the right convolution kernels and then proceed to analysing the behaviour of these kernels both at zero and at innity.
Proposition 5.2.16 has the following consequences:
m
with 1 0. Let x denote
Corollary 5.4.3. Let = {(x, )} be in S,
its associated kernel.
1. If , 1 , 2 , o Nn0 are such that
m [] + [1 ] + [2 ] + [o ] < Q/2,
z2 (Xxo q (z)x (z)) is square integrable and for
then the distribution Xz1 X
every x G we have
2
1 2 o
m ,a,b,c
Xz Xz (Xx q (z)x (z)) dz C sup (x, )2S,
G
where a = [], b = [o ], c = []+[1 ]+[2 ]+[o ] and C = Cm,,1 ,2 ,o > 0
is a constant independent of and x. If 1 = 0 then we may replace the
m ,a,b,c with a seminorm m,R
seminorm S,
S
,a,b given in (5.37).
,
2. For any , 1 , 2 , o
Nn0
satisfying
m [] + [1 ] + [2 ] + [o ] < Q,
z2 Xxo q (z)x (z) is continuous on G for every
the distribution z Xz1 X
x G and we have
z2 Xxo q (z)x (z) C sup (x, )S m ,[],[ ],[ ] ,
sup Xz1 X
o
2
,
zG
G
332
Chapter 5. Quantization on graded Lie groups
where C = Cm,,1 ,2 ,o > 0 is a constant independent of and x. If 1 =
m ,[],[ ],[ ] with the seminorm
0 then we may replace the seminorm S,
o
2
S m,R ,[],[o ] , see (5.37).
,
Consequently, if > 0 then the map : (x, y) x (y) is smooth on
G (G \{0}).
Proof. Part (1) follows from Proposition 5.2.16 together with Theorem 5.2.22 (i)
Parts (3) and (2), and Corollary 5.2.25 . Now by the Sobolev inequality in Theorem
4.4.25 (ii), if the righthand side of the following inequality is nite:
s
z2 Xxo q (z)x (z) C
z2 Xxo q (z)x (z)
sup Xz1 X
(I+Rz ) Xz1 X
zG
L2 (dz)
for s > Q/2, then the distribution
z2 Xxo q (z)x (z)
z Xz1 X
is continuous and the inequality of Part (2) holds. By Theorem 4.4.16,
s
z2 Xxo q (z)x (z)
2
(I + Rz ) Xz1 X
L (dz)
s+[1 ]
[2 ]
Xxo q (z)x (z)
C (I + R) (I + R)
2
L (dz)
s+[1 ]
[2 ]
C (I + R) Xxo (x, )(I + R) 2
,
L (G)
by the Plancherel formula (1.28). By Proposition 5.2.16 (together with Theorem
5.2.22 (ii)) as long as
m + s + [1 ] [] + [o ] + [2 ] < Q/2,
since
(I + R)
s+[1 ]
(I + R)
[2 ]
Xxo q (z)x (z)
is the kernel of the symbol
(I + R)
s+[1 ]
Xxo (x, )(I + R)
we have
s+[1 ]
[2 ]
(I + R) Xxo (x, )(I + R)
L2 (G)
[2 ]
m ,[],[ ],[ ] ,
C(x, )S,
o
2
if s + [1 ] [] m [o ] [2 ]. This shows Part (2).
5.4. Kernels of pseudodierential operators
333
Estimates at innity
We will now prove better estimates for the kernel than the ones stated in Corollary
5.4.3. First let us show that the kernel has a Schwartz decay away from the origin.
m
Proposition 5.4.4. Let = {(x, )} be in S,
with 1 0. Let x denote
its associated kernel.
We assume that > 0 and we x a homogeneous quasinorm   on G. Then
for any M R and any , 1 , 2 , o Nn0 there exist C > 0 and a, b, c N
independent of such that for all x G and z G satisfying z 1, we have
1 2 o
M
m ,a,b,c z
.
Xz Xz (Xx q (z)x (z)) C sup (x, )S,
G
m ,a,b,c with a
Furthermore, if 1 = 0 then we may replace the seminorm S,
seminorm S m,R ,a,b given in (5.37).
,
Proof. We start by proving the stated result for = 1 = 2 = o = 0 and for
the homogeneous quasinorm  p given by (3.21). Here p > 0 is a positive number
to be chosen suitably. We also x a number bo > 0 and a function o C (R)
valued in [0, 1] with o 0 on (, 12 ] and o 1 on [1, ). We set
p
(x) := o (bp
o xp ).
Therefore, is a smooth function on G such that (z) = 1 if zp bo . Consequently,
M
sup zM
p x (z) sup zp x (z)(z)
zp bo
zG
(z)(z)
Xz zM
x
p
[ ]Q/2
L2 (G,dz)
(5.38)
by the Sobolev inequality in Theorem 4.4.25.
We study each term separately. We assume that p/2 is a positive integer
divisible by all the weights 1 , . . . , n and we introduce the polynomial
zpp =
zj  j
j=1
and its inverse, so that
Xz
M p p
zM
zp zp zp x (z) (z)
p x (z)(z) = Xz
2 p
p
zp x (z) ,
Xz 1 zM
=
p zp (z) Xz
[1 ]+[2 ]=[ ]
334
Chapter 5. Quantization on graded Lie groups
means taking a linear combination, that is, a sum involving some conwhere
stants. We observe that, using a polar change of coordinates,
p
Xz 1 zM
p zp (z) L2 (G,dz) <
as long as 2(M p [1 ]) + Q 1 < 1. We assume that p has been chosen so that
2(M p) + Q < 0. Therefore, all these L2 norms can be viewed as constants. By
the CauchySchwartz inequality and the properties of Sobolev spaces, we obtain
C
Xz 2 zpp x (z) L2 (G,dz)
Xz zM
p x (z)(z) L2 (G,dz)
[2 ][ ]
Xz 2 {
q x } 2 ,
[2 ][ ] []p
p
n
since zpp = j=1 zj j is a polynomial of homogeneous degree p. Therefore, by
Corollary 5.4.3 Part (1), we get
m ,p,0,p+[ ]
Xz zM
p x (z)(z) L2 (G,dz) C sup (x, )S,
G
if p m > Q/2 + [ ]. We choose p accordingly. Combining this with (5.38) yields
m ,p,0,p+Q/2 .
sup zM
p x (z) C sup (x, )S,
zp bo
G
Therefore, we have obtained the result for the homogeneous norm  p and =
1 = 2 = o = 0.
The full result follows for any homogeneous norm and indices , 1 , 2 , o
from the equivalence of any two homogeneous norms and by Theorem 5.2.22 (i)
Parts (3) and (2), and Corollary 5.2.25.
Remark 5.4.5. 1. During the proof of Proposition 5.4.4, we have obtained the
following statement which is quantitatively more precise. We keep the setting
of Proposition 5.4.4. Then for any M R and bo > 0, there exists C =
CM,bo ,m > 0 such that
m ,p,0,p+Q/2 ,
sup zM
p x (z) C sup (x, )S,
zp bo
G
where p N is the smallest positive integer such that p/2 is divisible by all
the weights 1 , . . . , n and p > max(Q/2 + M, 1 (m + Q + 1)).
2. Combining Part (1) above, Theorem 5.2.22 (i) Parts (3) and (2), and Corollary 5.2.25, it is possible (but not necessarily useful) to obtain a concrete
expression for the numbers a, b, c appearing in Proposition 5.4.4, in terms of
m, , , , 1 , 2 , o and of Q.
Furthermore, the same statement is true for z bo for an arbitrary
lower bound bo > 0. However, the constant C may depend on bo .
5.4. Kernels of pseudodierential operators
335
Estimates at the origin
We now prove a singular estimate for the kernel near the origin which is (therefore)
not covered by Corollary 5.4.3 (2).
m
with 1 0. Let x denote
Proposition 5.4.6. Let = {(x, )} be in S,
its associated kernel.
We assume that > 0 and we x a homogeneous quasinorm   on G. Then
for any , 1 , 2 , o Nn0 with Q + m + [o ] [] + [1 ] + [2 ] 0 there exist
a constant C > 0 and computable integers a, b, c N0 independent of such that
for all x G and z G\{0}, we have that if
Q + m + [o ] [] + [1 ] + [2 ] > 0,
then
Q+m+[o ][]+[1 ]+[2 ]
1 2 o
m ,a,b,c z
,
Xz Xz (Xx q (z)x (z)) C sup (x, )S,
G
and if
Q + m + [o ] [] + [1 ] + [2 ] = 0,
then
1 2 o
m ,a,b,c ln z.
Xz Xz (Xx q (z)x (z)) C sup (x, )S,
G
m ,a,b,c with
In both estimates, if 1 = 0 then we may replace the seminorm S,
a seminorm S m,R ,a,b given in (5.37).
,
During the proof of Proposition 5.4.6, we will need the following technical
lemma which is of interest on its own.
m
with 1 0. Let D(R) and
Lemma 5.4.7. Let = {(x, )} be in S,
co > 0. We also x a positive Rockland operator R of homogeneous degree with
m
.
corresponding seminorms for the symbol classes S,
Then for any N0 , the symbols given by
L, (x, ) := (2co (R))(x, )
and
R, (x, ) := (x, )(2co (R)),
are in S . Moreover, for any m1 R and a, b, c N0 , there exists a constant
C = Cm,m1 ,,,a,b,c,,co > 0 such that for any N0 we have
co
m ,a,b,c 2
L, (x, )S m1 ,a,b,c C sup (x, )S,
,
(mm1 )
G
The same holds for R, (x, ), but with a possibly dierent seminorm on the right
hand side.
Only for R, (x, ), we also have for the seminorm S m,R ,a,b given in (5.37),
,
the estimate
co
R, (x, )S m1 ,R ,a,b C sup (x, )S m,R ,a,b 2
,
G
(mm1 )
336
Chapter 5. Quantization on graded Lie groups
Proof of Lemma 5.4.7. For each N0 , the symbol (2co (R)) is in S by
Proposition 5.3.4. Therefore, by Theorem 5.2.22 (ii) and the inclusions (5.31), L,
and R, are in S .
Let us x o , o Nn0 and R. By the Leibniz formula (see (5.28)),
[o ]m1 [o ]+
Xxo o L, (I + R)
[o ]m1 [o ]+
= (I + R)
Xxo o (2co (R))(x, ) (I + R)
[o ]m1 [o ]+
=
c1 ,2 (I + R)
1 (2co (R))
(I + R)
[1 ]+[2 ]=[o ]
Xxo 2 (x, )(I + R) .
Therefore, taking the operator norm, we obtain
[o ]m1 [o ]+
(I + R)
Xxo o L, (I + R)
[o ]m1 [o ]+
[2 ]m[o ]+
C
(I + R)
1 (2co (R))(I + R)
[1 ]+[2 ]=[o ]
(I + R)
[2 ]m[o ]+
Xxo 2 (x, )(I + R)
m ,[ ],[ ],
C(x, )S,
o
o
[o ]m1 [o ]+
[2 ]m[o ]+
(I + R)
1 (2co (R))(I + R)
.
[1 ]+[2 ]=[o ]
By Proposition 5.3.4,
(I + R)
[o ]m1 [o ]+
1 (2co (R))(I + R)
[2 ]m[o ]+
co
C(2
)M m2 ,k ,
for some k, where m2 is such that
[1 ] m2 = [o ] m1 [o ] + ([2 ] m [o ] + ),
that is,
m2 = m1 m + [1 ](1 ).
Now, we can estimate
(2co )M m2 ,k
m2
k ((2co ))
m2
2co k ( k )(2co )
sup
(1 + )k
sup
(1 + )k
>0, k =0,...,k
>0, k =0,...,k
C2co
m2
5.4. Kernels of pseudodierential operators
Therefore,
(I + R)
[]m1 +
337
1 (2co (L))(I + R)
[1 ]+[2 ]=[o ]
2co
m1 m+[1 ](1)
[2 ]m[o ]+
C2co
m1 m
[1 ]+[2 ]=[o ]
and we have shown that
(I + R)
[o ]m1 [o ]+
Xxo o L, (I + R)
co
m ,[ ],[ ], 2
Co (x, )S,
o
o
m1 m
The desired property for L, follows easily. The property for R, may be obtained
by similar methods and its proof is left to the reader.
Proof of Proposition 5.4.6. By Theorem 5.2.22 (i) Parts (3) and (2), and Corollary
5.2.25, it suces to show the statement for = 1 = 2 = o = 0. By equivalence
of homogeneous quasinorms (Proposition 3.1.35), we may assume that the homogeneous quasinorm is  p given by (3.21) where p > 0 is such that p/2 is the
smallest positive integer divisible by all the weights 1 , . . . , n . Since x decays
faster than any polynomial away from the origin (more precisely see Proposition
5.4.4), it suces to prove the result for zp < 1.
m
So let S,
with Q + m 0. By Lemma 5.4.11 (to be shown in Section
5.4.2) we may assume that the kernel : (x, y) x (y) of is smooth on G G
and compactly supported in x. By Proposition 5.4.4 it is also Schwartz in y.
We x a positive Rockland operator R of homogeneous degree and a dyadic
decomposition of its spectrum: we choose two functions 0 , 1 D(R) supported
in [1, 1] and [1/2, 2], respectively, both valued in [0, 1] and satisfying
> 0
() = 1,
=0
where for N we set
() := 1 (2(1) ).
For each N0 , the symbol ((R)) is in S by Proposition 5.3.4 and its kernel
(R)0 is Schwartz by Corollary 4.5.2. Furthermore, by the functional calculus,
N
operator topology of L (L2 (G)) to the identity
=0 (R) converges in the strong
N
operator I as N , and thus =0 (R)0 converges in K(G) and in S (G) to
the Dirac measure 0 at the origin as N .
By Theorem 5.2.22 (ii), the symbol given by
(x, ) := (x, ) ((R)),
(x, ) G G,
338
Chapter 5. Quantization on graded Lie groups
is in S . The kernel associated with is given by
(x, y) = ,x (y) = ( (R)0 ) x (y).
N
For each x, we have ,x S(G). The sum =0 ,x converges in S (G) to x as
N since
N
N
Op( (x, )) = Op((x, ))
(R)
=0
=0
converges to Op((x, )) in the strong operator topology of L (L2 (G), L2m (G)).
This convergence is in fact stronger. Indeed, by Lemma 5.4.7,
(mm1 )
m ,a ,b ,c 2
,
S m1 ,a,b,c C sup S,
,
thus
G
S m1 ,a,b,c <
,
m1
if m1 > m.
Consequently, the sum is convergent in S, and, xing x G,
the sum supzS ,x (z) is convergent where S is any compact subset of G\{0}
by Proposition
5.4.4 or more precisely therst part in Remark 5.4.5. Necessarily,
the limit of is and the limit of ,x for the uniform convergence on
any compact subset of G\{0} is x with
x (z)
,x (z),
z G\{0}.
=0
By Corollary 5.4.3 (2), for any m1 < Q and r N0 , we have
sup zpr
sup (x, )S m1 ,0,0,0
p ,x (z) C
zG
[]=pr G
Cc,r 2(mm1 pr)
(5.39)
m ,pr,0,0 .
by Lemma 5.4.7 and its proof, with c,r := supG (x, )S,
We write zp 2o in the sense that o N0 is the only integer satisfying
zp (2(o +1) , 2o ].
Let us assume that Q + m > 0. We use (5.39) with r = 0 and m1 such that
m m1 = (Q + m)/. In particular,
Q
1
< Q.
m1 = m(1 )
The sum over = 0, . . . , o 1, can be estimated as
o 1
,x (z)
=0
o 1
Cc,0 2(mm1 ) c,0 2o (mm1 )
=0
Q+m
Cc,0 zp
5.4. Kernels of pseudodierential operators
339
We now choose r N and m1 < Q such that
m m1 pr < 0
pr(1 ) + m m1 =
and
Q+m
.
More precisely, we set r := (m + Q)/(p) , that is, r is the largest integer strictly
greater than (m + Q)/(p), while m1 is dened by the equality just above; in
particular,
m m1 >
Q+m
Q+m
(1 )
thus
m1 < Q.
We may use (5.39) and sum over = o , o + 1 . . . , to get
zpr
p ,x (z) Cc,r
=o
2(mm1 pr) Cc,r 2o (mm1 pr) .
=o
Therefore, we obtain
,x (z)
Cc,r 2o (mm1 pr) zpr
p
1 pr)
Cc,r zpr(mm
= Cc,r zp
p
=o
Q+m
This yields the desired estimate for x when Q + m < 0.
Let us assume that Q + m = 0. Using (5.39) with r = 0 and m1 = m, we
obtain
o 1
,x (z)
=0
o 1
Cc,0 2(mm1 ) c,0 o
=0
Cc,0 ln zp .
Proceeding as above for the sum over o , we obtain that =o ,x (z) is
bounded. This yields the desired estimate for x in the case Q + m = 0.
Remark 5.4.8. It is possible to obtain a concrete expression for the numbers a, b, c
appearing in Proposition 5.4.6, in terms of m, , , , 1 , 2 , o and of Q.
5.4.2 Smoothing operators and symbols
The kernel estimates obtained in Section 5.4.1 allow us to characterise smoothing
operators in terms of their kernels. Moreover they also imply that the operators
in map the tempered distribution to smooth functions and enable the construction of sequences of smoothing operators converging in m
,
340
Chapter 5. Quantization on graded Lie groups
Theorem 5.4.9. 1. If T , then its associated kernel : (x, y) x (y) is a
smooth function on G G such that for each x G, y x (y) is Schwartz.
Moreover, for each multiindex Nn0 and each Schwartz seminorm
S(G),N , there exist a constant C > 0 and a seminorm S m ,a,b,c (both
independent of T ) such that
sup Xx x S(G),N CS m ,a,b,c .
xG
The converse is true, see Lemma 5.2.21.
2. If T , then T extends to a continuous mapping from S (G) to C (G)
via
T f (x) = f x (x)
where f S (G), x G, and x is the kernel associated with T .
Furthermore, for any compact subset K G and any multiindex
Nn0 , there exists a constant C > 0 and a seminorm S (G),N such that
sup  T f (x) Cf S (G),N .
xK
Moreover C can be chosen as C1 S m ,a,b,c , and C1 > 0 and N can be chosen
independently of f and T .
Part 1 may be rephrased as stating that the map between the smoothing
operators and their associated kernels is a Frechet isomorphism between
and the space Cb (G, S(G)) of functions C (G G) satisfying
sup Xx x S(G),N < .
xG
Here Cb (G, S(G)) is endowed with the Frechet structure given via the seminorms
max sup Xx x S(G),N < ,
[]N xG
N N0 .
Part 2 may be rephrased as stating that the mapping T T from to
the space L (S (G), C (G)) of linear continuous mappings from S (G) to C (G)
is continuous (it is clearly linear).
Proof. Part 1 follows easily from Theorem 5.4.1 and Remark 5.4.2. By Lemma
3.1.55, for any tempered distribution f S (G), the function f x is smooth on
G and the function x f x (x) is smooth on G. Hence T extends to S (G) and
T f C if f S (G).
Note that Lemma 3.1.55 also implies the existence of a positive constant C
and N N0 such that
f x (z) C(1 + z)N f S (G),N x S(G),N .
5.4. Kernels of pseudodierential operators
341
Using the Leibniz property for vector elds, one checks easily that for any
multiindex Nn0 , we have
c,1 ,2 Xx11=x (f Xx22=x x2 )(x1 ).
X (T f )(x) =
[1 ]+[2 ]=[]
Thus, proceeding as above, passing from left derivatives to the right, and using
Lemma 3.1.55, we get
1 (f (X 2 x ))(x1 )
(1 + x)[1 ] (X
X (T f )(x) C
x1 =x
x2 =x
2
[1 ]+[2 ]=[]
1 f ) (X 2 x )(x1 )
(1 + x)[1 ] (X
x1 =x
x2 =x
2
[1 ]+[2 ]=[]
1 f S (G),N X 2 x S(G),N
(1 + x)[1 ]+N X
x2 =x
2
[1 ]+[2 ]=[]
C(1 + x)N2 f S (G),N1 Xx22=x x2 S(G),N
with a new constant C > 0 and integers N2 , N1 , N N0 . This shows that f T f
is continuous from S (G) to C (G).
Using Part 1, the inequality above also shows the continuity of T T from
to the space of continuous mappings from S (G) to C (G). This concludes
the proof of Theorem 5.4.9.
Using the stability of taking the adjoint, reasoning by duality from Part 2
of Theorem 5.4.9, will yield the fact that smoothing operators map distributions
with compact support to Schwartz functions, see Corollary 5.5.13.
Note that the proof of Part 2 of Theorem 5.4.9 yields the more precise result:
Corollary 5.4.10. If T and f S (G), then T f is smooth and all its
leftderivatives X T f , Nn0 , have polynomial growth. More precisely, for any
multiindex Nn0 , there exist a constant C > 0, and integer M N0 and a
seminorm S (G),N such that
X T f (x) C(1 + x)M f S (G),N .
Moreover C can be chosen as C1 S m ,a,b,c , and C1 > 0 and N, M can be chosen
independently of f and T .
5.4.3 Pseudodierential operators as limits of smoothing operators
In the proof of Lemma 5.1.42, for a given symbol , we constructed a sequence of
symbols such that Op( ) is a sequence of nice operators converging towards
m
, then we can construct
Op() in a certain sense. If we assume that S,
342
Chapter 5. Quantization on graded Lie groups
a sequence of smoothing operators with a convergence in m
, described in the
next lemma and its corollary. These operators are therefore nice since they have
Schwartz associated kernels in the sense of Theorem 5.4.9.
m
, then we can conLemma 5.4.11. Let 1 0. If = {(x, )} is in S,
struct a family = { (x, )}, > 0, in S
, satisfying the following properties:
1. For each > 0, the xsupport of each is compact, or in other words,
the function x supG (x, )L (H ) is zero outside a compact set in
G. Hence the kernel : (x, y) ,x (y) associated with each symbol is
Schwartz on G G and compactly supported in x.
2. For any seminorm S m1 ,a,b,c , there exist a constant C = Ca,b,c,m,m1 , > 0
,
such that
m ,a,b,c
S m1 ,a,b,c CS,
(0, 1)
m1 m
and when m m1 ,
(0, 1)
m ,a,b,c+a
S m1 ,a,b,c CS,
m1 m
Here is the degree of homogeneity of the positive Rockland operator used to
dene the seminorms.
Consequently, when m < m1 , the convergence as 0 holds
m1
.
in S,
3. If S(G) then Op( ) D(G) and the convergence
Op( ) Op()
0
holds uniformly on any compact subset of G and also in S(G).
Remark 5.4.12. As the construction will show, the symbols are constructed
independently of the order m R.
Proof of Lemma 5.4.11. We consider the function on G constructed in Lemma
5.1.42. Let D(R) be such that 1 on [0, 1]. Let R be a positive Rockland
m
. We set
operator. Let S,
(x, ) = (x)(x, )( (R)).
Arguing as in Lemma 5.4.7 and its proof yields that
{(x, )( (R)), (x, ) G G}
is in S . Moreover, for any m1 R and a, b, c N0 , there exists a constant
C = Cm,m1 ,,,a,b,c, > 0 such that for any N0 we have
m ,a,b,c
(x, )( (R))S m1 ,a,b,c C sup (x, )S,
,
G
m1 m
5.4. Kernels of pseudodierential operators
343
From this, it is clear that Property (1) and the rst estimate in Property (2) hold.
Let us prove the second estimate in Property (2). We notice that
(I + R)
m1
((x, )( (R)) (x, )) L (H )
= (I + R)
(I + R)
m1
m
1
(x, ) (( (R)) I) L (H )
(x, )(I + R)
(I + R)
mm1
m1 m
L (H )
(( (R)) I) L (H ) ,
and the spectral calculus properties (cf. Corollary 4.1.16) imply
sup (I + R)
mm1
G
= (I + R)
mm1
(( (R)) I) L (H )
(( R) I) L (L2 (G)) sup(1 + )
mm1
>0
() 1.
One checks easily that
sup(1 + )
mm1
>0
() 1
1 sup (1 + )
mm1
>1
t(1 +
mm1
C
m1 m
provided that m m1 0. Hence
sup
(x,)GG
(I + R)
m1
((x, )( (R)) (x, )) L (H )
m ,0,0,m m
CS,
1
m1 m
More generally, we can introduce derivatives in x and dierence operators and use
the Leibniz properties (cf. Proposition 5.2.10):
Xx ((x, )( (R)) (x, ))
c,1 ,2 Xx 1 (x, ) 2 (( (R)) I),
=
[1 ]+[2 ]=[]
so that the quantity
(I + R)
m1 +[][]
Xx ((x, )( (R)) (x, )) (I + R) L (H )
is, up to a constant, less or equal to the sum over [1 ] + [2 ] = [] of
(I + R)
m1 +[][]
(I + R)
Xx 1 (x, )(I + R)
m m[2 ]+
1
m1 m[2 ]+
L (H )
2 (( (R)) I)(I + R) L (H ) .
344
Chapter 5. Quantization on graded Lie groups
Applying Proposition 5.3.4, we obtain
(I + R)
m1 m[2 ]+
2 (( (R)) I)(I + R) L (H ) C
mm1
Collecting the estimates and taking the supremum over [] a, [] b, 
c yield the second estimate in Property (2).
Property (3) follows from Property (2) and the continuity of Op()
m1
to L (S(G)), see Theorem 5.2.15.
from S,
Keeping the notation of Lemma 5.4.11, we can also show that the kernels
converge in some sense towards the kernel of . In order to make this more precise,
let us dene the space Cb (G, S (G)) as the space of functions x x S (G)
such that for each x G, y x (y) is a tempered distribution and, for any
Nn0 , the map x Xx x is continuous and bounded on G. This denition is
motivated by the following property:
m
Lemma 5.4.13. If S,
then its associated kernel = () is in Cb (G, S (G))
dened above. Furthermore, the map
()
m
from S,
to Cb (G, S (G)) is continuous.
Naturally, we have endowed Cb (G, S (G)) with the structure of Frechet
space given by the seminorms
max sup Xx x S (G),N ,
[]N xG
N N0 .
Proof of Lemma 5.4.13. By Lemma 5.1.35, if is a symbol then its kernel is in
C (G, S (G)). Adapting slightly its proof yields
sup Xx x S (G) C sup Xx (x, )L
xG
xG
0,m[]
.
(G)
As the inverse Fourier transform is onetoone and continuous from L
0,m[] (G)
()
m
to S (G), this shows the continuity of the map from S, to Cb (G, S (G)).
We can now express the convergence in distribution of the sequence of kernels
constructed in the proof of Lemma 5.4.11:
Corollary 5.4.14. We keep the notation of Lemma 5.4.11. The sequence of kernels
converges towards the kernel associated with in Cb (G, S (G)). If > 0,
the convergence is also uniform on any compact subset of G (G\{0}).
m1
Proof. The statement follows from the convergence of to in S,
for m1 < m
by Part 2 of Lemma 5.4.11, together with Lemma 5.4.13 for the rst part and
Corollary 5.4.3 for the second part.
5.4. Kernels of pseudodierential operators
345
5.4.4 Operators in 0 as singular integral operators
From the kernel estimates obtained in Section 5.4.1, one can show easily that the
operators in 0 are Calder
onZygmund, and generalise this to some classes m
, ,
2
see Theorem 5.4.16. We are then led to study the L boundedness.
First let us notice that thanks to the kernel estimates, our operators admit
a representation as singular integrals in the following sense:
Lemma 5.4.15. Let x be the kernel associated with T m
, with m R and
1 0 with = 0. For any f S (G) and any x0 G such that f 0 on
a neighbourhood of x0 , the integral
f (y)x0 (y 1 x0 )dy
G
makes distributional sense and denes a smooth function at x0 .
This coincides with T f if f S(G).
Proof. Let T and x be as in the statement. Let f S (G) and x0 G. We
assume that there exists a bounded open set 2 containing x0 and where f 0.
1 , and
1 2 .
Let 1 2 be open subsets of 2 such that x0 ,
We can nd 1 , D(G) such that 1 1 on 1 but 1 0 outside 2 , 1
on but 0 outside 1 . At least formally, we have
(x)
f (y)x (y 1 x)dy =
f (y) (x)(1 1 )(y)x (y 1 x)dy,
G
since f 0 on {1 = 1}. Clearly the function (x, y) (x)(1 1 )(y) is smooth
on G G and supported away from the diagonal {(x, y) G G : x = y}. By
Theorem 5.4.1, the function
y (x)(1 1 )(y)x (y 1 x),
is Schwartz and this yields a smooth mapping G S(G) (which is also compactly
supported). The rest of the statement follows easily.
In Corollary 5.5.13, we will see that an operator in m
, extends naturally
to S (G). Lemma 5.4.15 and its proof above will then imply that the operator
admits a singular representation for any tempered distribution in the sense that
the following formula makes sense and holds
T f (x) =
f (y)x (y 1 x)dy,
for any f S (G) and any x G such that f 0 on a neighbourhood of x. We
will not use this.
onWe can now give sucient condition for operator in some m
, to be Calder
Zygmund.
346
Chapter 5. Quantization on graded Lie groups
Theorem 5.4.16. 1. If T 0 then the operator T is Calder
onZygmund in the
sense of Denition 3.2.15.
2. If T m
, with
m ( 1)Q,
1 0 and = 0, then the operator T is Calder
onZygmund in the
sense of Denition 3.2.15.
In Parts 1 and 2, the constants appearing in the Denition 3.2.15 are = 1
and, up to constants of the group, given by seminorms of T m
, .
Proof. We x a homogeneous quasinorm   on G.
Let T 0 . We denote by its associated kernel. Then its integral kernel o
is formally given via o (x, y) = x (y 1 x). By Theorem 5.4.1, for any two distinct
points y, x G, we have
o (x, y) = x (y 1 x) Cy 1 xQ .
Using Remark 5.4.2 as well and the Leibniz property for vector elds, we obtain
(Xj )x o (x, y) (Xj )x1 =x x1 (y 1 x) + (Xj )x2 =x x (y 1 x2 ) Cy 1 x(Q+j ) ,
and
j )z=y1 x x (z) Cy 1 x(Q+j ) .
(Xj )y o (x, y) (X
Hence o satises the hypotheses of Lemma 3.2.19. This shows Part 1.
Let us now assume that T m
, . Again, let be its associated kernel. Let
C (G) be supported in the unit ball {x G : x 1} and such that 1
on {x G : x 1/2}. By Theorem 5.4.1 and Remark 5.4.2 together with Lemma
5.2.21, the operator given by {(1 )} is smoothing (as = 0) hence it is
a Calder
onZygmund operator by Part 1. Thus we just have to study the operator
{}. Its integral kernel is o given via
o (x, y) = (y 1 x)x (y 1 x).
Proceeding as above, in particular by Theorem 5.4.1, we have
o (x, y)
(Xj )y o (x, y)
=
=
(x )(y 1 x) y 1 x
j )z=y1 x x (z) y
(X
Q+m
x
,
Q+m+j
and o is supported on {(x, y) G : y 1 x 1} where we have
(Xj )x o (x, y)
(Xj )x1 =x x1 (y 1 x) + (Xj )x2 =x x (y 1 x2 )
y 1 x
Q+m+j
y 1 x
Q+m+j
+ y 1 x
Q+m+j
y 1 x
Q+m
since y 1 x 1. Hence if (Q + m)/ Q, we can apply Lemma 3.2.19.
5.4. Kernels of pseudodierential operators
347
In order to apply the singular integrals theorem (Theorem A.4.4), we still
need to show that the operators are L2 bounded. In the case (, ) = (1, 0), it is
not very dicult to adapt the Euclidean case to show that the operators in 0
are L2 bounded.
Theorem 5.4.17. If T 0 then T extends to a bounded operator on L2 (G).
Furthermore, there exist constants C > 0 and a, b, c N0 of the group such that
T f L2 (G) CT m ,a,b,c f L2 (G) .
f S(G)
During the proof of Theorem 5.4.17, we will need the following observation:
Lemma 5.4.18. The collection of operators 0 is invariant under left translations
in the sense that
T 0 = xo G
xo T x1
0 ,
o
where
xo : f f (xo ).
Furthermore, if x is the kernel of T and = Op1 (T ) is its symbol, then the
has xo x as kernel and (xo x, ) as symbol, and
operator xo T x1
o
0 ,a,b,c .
T 0 ,a,b,c = xo T x1
o
Proof of Lemma 5.4.18. Let T 0 and let x be its kernel. Then
f (x)
xo T x1
o
=
=
=
T (x1
f )(xo x) = (x1
f ) xo x (xo x)
o
o
1
f (x1
xo x)dy
o y)xo x (y
G
f (z)xo x (z 1 x)dz
G
after the change of variable z = x1
o y. Therefore
f (x) = f xo x (x).
xo T x1
o
Since FG (xo x )() = (xo x, ) if denotes the symbol of T , we see that xo x is the
and the corresponding
kernel associated to the symbol {(xo x, ), (x, ) G G}
1
operator is xo T xo . The rest of the statement follows easily.
Proof of Theorem 5.4.17. The proof follows the Euclidean case as given in [Ste93,
ch. VI 2]. Let T 0 and let = Op1 (T ) be its symbol. We claim that it
suces to show Theorem 5.4.17 under the additional assumption that the kernel
associated with is smooth in x and Schwartz in y, and such that G x
x S(G) is smooth. Indeed, this would imply that Theorem 5.4.17 is proved
for each operator T = Op( ) where is as in Lemma 5.4.11. The properties
(2) and (3) in Lemma 5.4.11 allow to pass through the limit as 0 and imply
then the theorem. This shows our earlier claim and hence we may assume that
G x x S(G) is smooth.
348
Chapter 5. Quantization on graded Lie groups
We x   to be the homogeneous quasinorm  p given by (3.21), where
p > 0 is such that p/2 is the smallest positive integer divisible by all the weights
1 , . . . , n . The balls are dened by B(xo , r) := {x G : x1 xo  < r}. We denote
by Co 1 a constant such that for all x, y G, we have
x
= xy x Co y,
2
see the triangle inequality in Proposition 3.1.38 and its converse (3.26).
xy Co (x + y)
y
and
Let f S(G) and let us write it as
f = f1 + f2 ,
where f1 and f2 are two smooth functions supported in B(0, 4Co ) and outside of
B(0, 2Co ), respectively, and satisfying f1 , f2  f .
First, we claim that there exists a constant C > 0 of the group such that
T f1 (x)2 dx C2S 0 ,0,Q/2,0 f1 2L2 (G) .
(5.40)
B(0,1)
Let us prove this. We x a function D(G) which is identically 1 on B(0, 1).
Then
2
T f1 (x) dx
(x) f1 x (x)2 dx
B(0,1)
B(0,1)
sup (z) f1 z (x)2 dx.
B(0,1) zG
We now use the Sobolev inequality in Theorem 4.4.25 to get
2
2
Xz {(z) f1 z (x)} dz.
sup (z) f1 z (x) C
zG
Since
[]Q/2
Xz {(z) f1 z (x)} = f1 Xz {(z)z }(x),
we have obtained
T f1 (x)2 dx
B(0,1)
B(0,1)
=C
[]Q/2
[]Q/2
f1 Xz {(z)z }(x) dzdx
f1 Xz {(z)z }(x) dxdz,
2
B(0,1)
by Fubinis property. But the integral over B(0, 1) can be estimated using Plancherels Theorem (see Theorem 1.8.11) by
2
2
f1 Xz {(z)z }(x) dx f1 Xz {(z)z }2
B(0,1)
2
(Xz {(z)z })2L (G)
f1 2 .
5.4. Kernels of pseudodierential operators
349
Now the Leibniz formula for Xz gives
(Xz {(z)z })L (L2 (G))
c1 ,2 (X 1 (z)Xz2 z }) L (L2 (G))
[1 ]+[2 ]=[]
C max Xz z } L (L2 (G))
X 1 (z).
[][]
[1 ][]
Since Xz z = Xz (z, ), we have obtained
f1 Xz {(z)z }(x) dx
2
B(0,1)
2
C max Xz (z, )2L (G)
f1 2
[][]
[1 ][]
Therefore,
B(0,1)
T f1 (x)2 dx
C
C
max
f1 Xz {(z)z }(x) dxdz
2
[]Q/2
X 1 (z)2 .
B(0,1)
2
sup Xz (z, )2L (G)
f1 2 .
[]Q/2 zG
This concludes the proof of Claim (5.40).
Secondly, we claim that for any r N, there exists a constant C = Cr > 0
such that
T f2 (x)2 dx C2S 0 ,pr,0,pr (1 +  )pr f2 2L2 (G) .
(5.41)
B(0,1)
Let us prove this. We write
T f2 (x) =
y B(0,2C
/
o)
f2 (y)y 1 xpr ( pr x )(y 1 x)dy.
If x B(0, 1) and y
/ B(0, 2Co ), then
y 1  y 1 x Co x Co
thus
y 1 x y Co
1
1
y (1 + y),
2
4
and
T f2 (x)
y B(0,2C
/
o)
f2 (y)
1
(1 + y)
4
pr
( pr x )(y 1 x) dy
4pr (1 +  )pr f2 L2 (G) ( pr x )L2 (G) ,
350
Chapter 5. Quantization on graded Lie groups
after having used the CauchySchwartz inequality. Integrating the square of the
lefthand side over x B(0, 1), and taking the supremum over x B(0, 1) of the
righthand side, we obtain
T f2 (x)2 dx 42pr sup  pr x 2L2 (G) (1 +  )pr f2 2L2 (G) . (5.42)
xB(0,1)
B(0,1)
Now writing zpr
p =
(z),
[]=pr c q
we have
 pr x 2L2 (G) Cr
q x 2L2 (G)
[]=pr
and since by Corollary 5.4.3 (1), if [] > Q/2,
q x 2L2 (G) C sup (x, )2S m
G
, ,[],0,[]
we have obtained that if pr > Q/2, then
sup
xB(0,1)
 pr x 2L2 (G) Cr 2S 0 ,pr,0,pr .
This and (5.42) show Claim (5.41).
Now, combining together Claims (5.40) and (5.41), we obtain
T f (x)2 dx Cr T 20 ,pr,Q/2,pr (1 +  )pr f 2L2 (G) ,
B(0,1)
and this is so for any f S(G). Therefore, by Lemma 5.4.18 (and its notation),
we have for any xo G, that
T f (x)2 dx =
T f (x)2 dx =
T f (xo x )2 dx
1
B(xo ,1)
x x<1
B(0,1)
o
2
=
xo (T f )(x ) dx =
(xo T x1
)(xo f )(x )2 dx
o
B(0,1)
B(0,1)
Cr xo T x1
20 ,pr,Q/2,pr (1 +  )pr xo f 2L2 (G)
o
Cr T 20 ,pr,Q/2,pr (1 +  )pr xo f 2L2 (G) .
Integrating over xo G, we obtain for the left hand side,
T f (x)2 dxdxo =
1x1
T f (x)2 dxdxo
o x<1
G B(xo ,1)
G G
=
1y<1 T f (x)2 dxdy = B(0, 1)T f 22 ,
G
5.5. Symbolic calculus
351
and for the last term in the right hand side,
pr
2
(1 + x)pr f (xo x)2 dxdxo
(1 +  ) xo f L2 (G) dxo =
G
G G
= f 22 (1 + x)2pr dx.
G
Assuming 2pr + Q < 0, this last integral is nite.
We have obtained that if r > Q/2p (for instance r = Q/2p ) then pr > Q/2
and
B(0, 1)T f 22 CT 20 ,pr,Q/2,pr f 22 .
This concludes the proof of Theorem 5.4.17.
Remark 5.4.19. More precisely we have obtained that if T , then
0
T f 2 CT 0 ,pr,Q/2,pr f 2 ,
Q
where r := 2p
, and p R is such that p/2 is the smallest positive integer divisible
by all the weights 1 , . . . , n .
Theorem 5.4.16 and Theorem 5.4.17 show that any operator of order 0 and of
type (1,0) satises the hypotheses of the singular integrals theorem, see Sections
3.2.3 and A.4. Therefore, we have the following corollary:
Corollary 5.4.20. If T 0 then T extends to a bounded operator on Lp (G) for
any p (1, ). Furthermore, there exist constants a, b, c N0 such that
p (1, )
5.5
C > 0
f S(G)
T f Lp (G) CT 0 ,a,b,c f Lp (G) .
Symbolic calculus
In this section we present elements of the symbolic calculus of operators with
m
. In particular, we will discuss asymptotic sums of
symbols in the classes S,
symbols, adjoints, and compositions.
5.5.1 Asymptotic sums of symbols
We now establish a nilpotent analogue of the asymptotic sum of symbols of decreasing orders going to .
Theorem 5.5.1. We assume 1 0. Let {j }jN0 be a sequence of symbols
m
m0
,
such that j S,j with mj strictly decreasing to . Then there exists S,
, such that
unique modulo S
M N
M
j=0
j S,M +1 .
(5.43)
352
Chapter 5. Quantization on graded Lie groups
Denition 5.5.2. Under the hypotheses and conclusions of Theorem 5.5.1, we write
j .
Proof. We keep the notation of the statement. We also x a positive Rockland
operator R of homogeneous degree on G. Let C (R) with (,1/2) = 0
and [1,) = 1. We x t (0, 1).
Let us check that for any seminorm S m0 ,a,b,c , there exists a constant
,
C = Ca,b,c > 0 such that for any t (0, 1) and any j N, we have
j (x, )(t(R))S m0 ,a,b,c Cj (x, )S m0 ,a,b,c+a+m0 mj t
,
m0 mj
(5.44)
Indeed, from the Leibniz formula (see Formula (5.28)), we obtain easily
[o ]m0 [o ]+
(I + R)
Xxo o (j (x, )(t(R))) (I + R) L (H )
(I + R)
[o ]m0 [o ]+
Xxo 1 j (x, )
[1 ]+[2 ]=[o ]
2 (t(R)) (I + R) L (H )
j (x, )S m0 ,[1 ],[o ],([o ][1 ])+m0 mj +
,
[1 ]+[2 ]=[o ]
(I + R)
[2 ]m0 +mj +
2 (t(R))(I + R) L (H ) .
By the functional calculus, we have
(I + R)
[2 ]m0 +mj +
(I + R)
2 (t(R))(I + R) L (H )
[2 ]m0 +mj +
sup (1 + )
k k
>0
m0 +mj
+k
2 (t(R))(I + R) L (H )
k {(t)} t
m0 mj
by Proposition 5.3.4 for some k N0 . This shows (5.44).
Let us choose strictly increasing sequences {a }, {b } and {c } of positive
integers. For each there exists C > 0 such that for any j N and t (0, 1), we
have
j (x, )(t(R))S m0 ,a ,b ,c C j (x, )S m0 ,a ,b ,c +a +m0 mj t
,
m0 mj
We may assume that the constants C are increasing with .
We now choose a decreasing sequence of numbers {tj } such that for any
j N,
tj (0, 2j )
and
m0 mj
Cj sup j (x, )S m0 ,aj ,bj ,cj +aj +m0 mj tj
xG
G
2j .
5.5. Symbolic calculus
353
For any j N, we dene the symbols
j (x, ) := j (x, )(tj (R)).
For any N, the sum
j=0
j S m0 ,a ,b ,c
,
j=0
j S m0 ,a ,b ,c +
,
2j ,
j=+1
m0
is a Frechet space, we obtain that
is nite. Since S,
:=
j ,
j=0
m0
is a symbol in S,
.
Starting the sequence at mM +1 , the same proof gives
j=M +1
j S,M +1 .
By Proposition 5.3.4, each symbol given by (1 )(tj (R)) is in S . Thus by
Theorem 5.2.22 (ii) and the inclusions (5.31), each symbol given by j (x, )(1
)(tj (R)) is in S . Therefore, the symbol given by
(x, )
M
j=0
j (x, ) =
j (x, )(1 )(tj (R)) +
j=0
j (x, ),
j=M +1
is in S,M +1 . This shows (5.43) for .
If is another symbol as in the statement of the theorem, then for any
M N,
M
M
=
j
j
j=0
j=0
is in S mM +1 . Thus S .
We note that the proof above does not produce a symbol depending continuously on {j }, the same as in the abelian case.
5.5.2 Composition of pseudodierential operators
In this section, we show that the class of operators mR m
, is an algebra:
354
Chapter 5. Quantization on graded Lie groups
1
Theorem 5.5.3. Let 1 0 with = 1 and m1 , m2 R. If T1 m
, and
m2
T2 , are two pseudodierential operators of type (, ), then their composi1 +m2
. Moreover, the mapping
tion T1 T2 is in m
,
(T1 , T2 ) T1 T2
m2
m1 +m2
1
is continuous from m
.
, , to ,
Since any operator in m
, maps S(G) to itself continuously (see Theorem
m2
1
5.2.15), the composition of any two operators in m
, and , denes an operator
in L (S(G)).
Let us start the proof of Theorem 5.5.3 with observing that the symbol
of T1 T2 is necessarily known and unique at least formally or under favourable
conditions such as between smoothing operators:
Lemma 5.5.4. Let 1 and 2 be two symbols in S and let 1 and 2 be their
associated kernels. We set
2,xz1 (yz 1 )1,x (z)dz,
x, y G.
x (y) :=
G
Then (x, ) = (x ) denes a smooth symbol in the sense of Denition 5.1.34.
Furthermore, it satises
Op(1 )Op(2 ) = Op().
and
(x, ) =
1,x (z)(z) 2 (xz 1 , ) dz,
(5.45)
In particular, if 2 (x, ) is independent of x then 1 2 = 1 2 .
We will often write
:= 1 2 .
Proof of Lemma 5.5.4. We keep the notation of the statement. Clearly : (x, y)
x (y) is smooth on GG, compactly supported in x. Furthermore, x is integrable
in y since
x (y)dy
2 (xz 1 , yz 1 )1 (x, z)dzdy
G
G G
2,xz1 (w)dw 1 (x, z)dz
G G
max

(w)dw
1,x (z)dz.
2,x
x G
Therefore, (x, ) = (x ) denes a symbol in the sense of Denition 5.1.33.
5.5. Symbolic calculus
355
Using the Leibniz formula iteratively, one obtains easily that for any o Nn0 ,
o
Xx x (y) is a linear combination of
2 1 2,x (yz 1 )X
x1=x 1,x (z)dz,
X
[1 ] + [2 ] = [o ].
2
1
1
x2 =xz
G
Hence proceeding as above
o x (y)dy
X
x
G
max
[1 ]+[2 ]=[o ]
x2 G
2 2,x (w)dw
X
x2
2
G
1 1,x (z)dz.
X
x
This together with the link between abelian and rightinvariant derivatives (see
Section 3.1.5, especially 3.17) implies easily that is a smooth symbol in the sense
of Denition 5.1.34.
The properties of 1 and 2 (see Theorem 5.4.9) justify the equalities
T2 (y)1,x (y 1 x)dy
Op(1 )Op(2 )(x) =
G
=
(z)2,y (z 1 y)1,x (y 1 x)dzdy
G G
(z)2,xw1 (z 1 xw1 )1,x (w)dzdw
=
G G
(z)x (z 1 x)dz = x (x),
=
G
with the change of variables y 1 x = w. This yields T1 T2 = Op(). We have then
nally
x (y)(y) dy
(x, ) =
x () =
G
2,xz1 (yz 1 )1,x (z)(z) (yz 1 ) dydz
=
G G
1,x (z)(z) 2 (xz 1 , ) dz,
=
G
after an easy change of variable.
From Lemma 5.5.4 and its proof, we see that if T = Op(1 )Op(2 ) then
the symbol of T is not 1 2 in general, unless the symbol {2 (x, )} does not
depend on x G for instance. However, we can link formally with 1 and 2
in the following way: using the vectorvalued Taylor expansion (see (5.27)) for
2 (x, ) in the variable x, we have
2 (xz 1 , )
q (z 1 )Xx 2 (x, ),
356
Chapter 5. Quantization on graded Lie groups
Thus, implementing this in the expression (5.45), we obtain informally
(x, )
1,x (z)(z)
q (z 1 )Xx 2 (x, ) dz
G
q (z 1 )1,x (z)(z) dz Xx 2 (x, )
1 (x, ) Xx 2 (x, ).
We will show that in fact these formal manipulations eectively give the asymm1
potitcs, see Corollary 5.5.8. From Theorem 5.2.22, we know that if 1 S,
,
m2
2 S, then
m +m2 ()[]
1 Xx 2 S,1
(5.46)
The main problem with the informal approach above is that one needs to estimate
the remainder
1 Xx 2 .
1 2
[]M
We will rst show how to estimate this remainder in the case of > using the
following property.
Lemma 5.5.5. We x a positive Rockland operator of homogeneous degree . Let
m1 , m2 R, 1 0 with = 0 and = 1, 0 Nn0 , and M, M1 N0 . We
assume that
1
m2 +(c0 +n )
M1 < M Q m1 [0 ] + (Q + 1 ),
1
(5.47)
m2 + (c0 + n + M ) M1 < Q m1 [0 ] + (Q + M ),
where
c0 :=
max
[ ].
[02 ][0 ]
[ ][02 ],  02 
If M M1 , only the second condition may be assumed.
Then there exist a constant C > 0, and two pseudonorms S m1 ,R ,a1 ,b1 ,
,
we have
S m2 ,0,b2 ,0 , such that for any 1 , 2 S and any (x, ) G G
,
Xx0 1 2 (x, )
1 (x, ) Xx 2 (x, ) L (H )
[]M
C1 S m1 ,R ,a1 ,b1 2 S m2 ,0,b2 ,0 .
,
In the proof of Lemma 5.5.5, we will use the following easy consequence of
the estimates of the kernels given in Theorem 5.2.22.
5.5. Symbolic calculus
357
m
Lemma 5.5.6. Let S,
with 1 0 with = 0. We denote by x
its associated kernel. For any R, if + Q > max( m+Q
, 0) then there exist a
m
constant C > 0 and a seminorm S, ,a,b,c such that
G
m ,a,b,c .
z x (z)dz CS,
m ,a,b,c with m,R
We may replace S,
S
,a,b .
,
Proof of Lemma 5.5.6. We keep the notation and the statement and write
z x (z)dz =
+
.
G
z1
z<1
The
estimate for large z given in Theorem 5.4.1 easily implies that the integral
is bounded up to a constant of , m, , , by a seminorm of . The estimate
z1
for small z yield
m+Q
dz if m + Q > 0,
z1 z
z x (z)dz
z  ln zdz if m + Q = 0,
z1
z1
z dz
if m + Q < 0.
z1
Using the polar change of coordinates yields the result.
Proof of Lemma 5.5.5, case 0 = 0. By Lemma 5.5.4 and the observations that
follow, we have
1 (x, ) Xx 2 (x, )
(x, )
[]M
=
1,x (z)(z) 2 (xz 1 , )
q (z 1 )Xx 2 (x, ) dz
[]M
(,)
2
1,x (z)(z) Rx,M
(z 1 )dz,
(,)
2
where Rx,M
denotes the remainder of the (vectorvalued) Taylor expansion of
v 2 (xv, ) of order M at 0. We now introduce powers of (I + R) near (z)
(z) = (z) (I + R)M1 (I + R)M1 =
(z) (X) (I + R)M1
[]M1
and we notice that
(z) .
(z) (X) = (1) (X) (z) = (1) X
z
(5.48)
358
Chapter 5. Quantization on graded Lie groups
We integrate by parts and obtain
1 (x, ) Xx 2 (x, )
(x, )
[]M
[1 ]+[2 ]M1
[1 ]+[2 ]M1
z1=z 1,x (z1 )(z) X
z2=z R(I+R)
X
x,M
1
2
M1
1 1,x (z1 )(z) R(I+R)
X
z1 =z
x,M [2 ]
M1
2 (,)
X 2 2 (,)
(z21 )dz
(z 1 )dz
by Lemma 3.1.50. Taking the operator norm, we have
1 (x, ) Xx 2 (x, )L (H )
(x, )
[]M
(I+R)M1 X 2 2 (,)
[1 ]+[2 ]M1
1 1,x (z1 ) R
X
z1 =z
x,M [2 ]
(z 1 )L (H ) dz.
The adapted statement of Taylors estimates remains valid for vectorvalued function, see Theorem 3.1.51 and Remark 3.1.52 (3), so we have
(I+R)M1 X 2 2 (,)
Rx,M [2 ]
(z 1 )L (H )
z[] sup (I + R)M1 Xx1 Xx12 2 (x1 , )L (H ) .
x1 G
(M [2 ])+ +1
[]>(M [2 ])+
We have obtained that
(x, )
1 (x, ) Xx 2 (x, )L (H )
[]M
[]>(M [2 ])+
(M [2 ])+ +1
z1=z 1,x (z1 )dz
z[] X
1
sup (I + R)M1 Xx1 Xx12 2 (x1 , )L (H ) .
x1 G
If M [2 ] 0, the integrals above are nite by Lemma 5.5.6 and the suprema
m2
seminorm in 2 when
are bounded by a S,
and it suces
m1 + [1 ] + Q < (Q + 1 )
M1 + m2 + (n + [2 ]) 0
m1 + M1 M + Q < (Q + 1 )
M1 + m2 + (n + M1 ) 0
5.5. Symbolic calculus
359
If M [2 ] > 0, the integrals above are nite by Lemma 5.5.6 and the suprema
m2
seminorm in 2 when
are bounded by a S,
m1 + [1 ] + Q < (Q + [])
,
M1 + m2 + ([] + [2 ]) 0
and it suces
m1 + M1 + Q < (Q + M )
M1 + m2 + (n + M ) 0
Our conditions on M and M1 ensure that the sucient conditions above are
satised. Collecting the various estimates yields the statement in the case = 0
and 0 = 0.
Proof of Lemma 5.5.5, general case. Using Formula (5.45), the Leibniz property
for left invariant vector elds easily implies that
Xx0 1 2 (x, ) =
Xx01 1,x (z)(z) Xx202=x 2 (x2 z 1 , ) dz.
[01 ]+[02 ]=[0 ]
Proceeding as in the case 0 = 0, we have
1 (x, ) Xx 2 (x, )
Xx0 1 2 (x, )
[]M
X 02 2 (x2 ,)
[01 ]+[02 ]=[0 ]
x2 =x
Xx01 1,x (z)(z) R0,M
(z 1 ) dz.
Introducing the powers of (I + R), each integral on the righthand side above is
equal to
z1=z Xx01 1,x (z1 )(z)
X
1
[1 ]+[2 ]M1
(I+R)M1 Xx202=x X 2 2 (x2 ,)
R0,M [2 ]
(z 1 ) dz,
(5.49)
by Corollary 3.1.53. We use a more precise version for the Taylor remainder than
in the proof of the case 0 = 0:
(I+R)M1 Xx202=x X 2 2 (x2 ,)
R0,M [2 ]
CM
(z 1 )L (H )
z[] S(z, M1 , , 02 , 2 ),
[]>(M [2 ])+
(M [2 ])+ +1
where S(z, M1 , , 02 , 2 ) denotes the supremum
S(z, M1 , , 02 , 2 ) :=
sup (I + R)M1 Xy Xx202=x Xy2 2 (x2 y, )L (H ) .
y M +1 z
360
Chapter 5. Quantization on graded Lie groups
For any reasonable function f : G C, the denitions of left and rightinvariant
vector elds imply
y f (xy)
(5.50)
Xx f (xy) = X
and the properties of left or rightinvariant vector elds (see Section 3.1.5) then
yield
f (xy) =
Xx f (xy) = X
Q, (y)Xy f (xy),
(5.51)
y
 
[ ][]
where Q, are ([ ] [])homogeneous polynomials. Therefore
1 , [] + [02
S(z, M1 , , 02 , 2 )
z[02 ][02 ] S(M
] + [2 ]),
[02
][02 ]
02
02 
1 , [0 ]) denotes the supremum
where S(M
1 , [0 ]) :=
S(M
sup
[ ]=[
0]
sup (I + R)M1 Xx1 2 (x1 , )L (H ) .
x1 G
We then obtain that (5.49) is bounded up to a constant by
z1=z Xx01 1,x (z1 )
X
z[]
1
[1 ]+[2 ]M1
[]>(M [2 ])+
(M [2 ])+ +1
1 , [] + [02
z[02 ][02 ] S(M
] + [2 ]) dz.
[02
][02 ]
02
02 
We conclude in the same way as in the case 0 = 0.
To take into account the dierence operator, we will use the following observation.
Lemma 5.5.7. Let 1 , 2 S . For any Nn0 , (1 2 ) is a linear combination independent of 1 , 2 of (1 1 ) (2 2 ), over 1 , 2 Nn0 satisfying
[1 ] + [2 ] = []. It is the same linear combination as in the Leibniz rule (5.28).
Proof of Lemma 5.5.7. We keep the notation of Lemma 5.5.4 and adapt the proof
of the Leibniz rule for given in Proposition 5.2.10. By Proposition 5.2.3 (4),
we have
q (y)x (y) =
q (yz 1 z)2,xz1 (yz 1 )1,x (z)dz
G
q2 (yz 1 )2,xz1 (yz 1 ) q1 (z)1,x (z)dz,
=
[1 ]+[2 ]=[]
where
denotes a linear combination. Lemma 5.5.4 implies easily the statement.
5.5. Symbolic calculus
361
Proof of Theorem 5.5.3 with > . We assume > . We x a positive Rockland
operator R of homogeneous degree . Let us show that for any 0 , 0 Nn0 , and
M0 N, there exists M M0 , a constant C > 0 and seminorms S m1 ,R ,a1 ,b1 ,
,
S m2 ,a2 ,b2 ,c2 such that for any 1 , 2 S we have
,
m()M0 [0 ]+[0 ]
Xx 0 0 M (x, ) (I + R)
L (H )
C1 S m1 ,R ,a1 ,b1 2 S m2 ,a2 ,b2 ,c2 ,
(5.52)
where we have denoted m = m1 + m2 and
1 Xx 2 .
M := 1 2
[]M
By Lemma 5.5.7, it suces to show (5.52) only for 0 = 0.
Let 0 N0 and M0 N. We x m2 := m1 + ( )M0 [0 ]. As > ,
we can nd M max(M0 , 1 ) such that
(Q m1 [0 ] + (Q + M )) (m2 + (c0 + n + M )) .
This shows that we can nd M1 satisfying the second condition in (5.47) for m1 , m2
and therefore also the rst. Hence we can apply Lemma 5.5.5 to M, M1 and the
m()M0 +[0 ]
, with orders m1 and m2 . The lefthand
symbols 1 and 2 (I + R)
side of (5.52) is then bounded up to a constant by
1 S m1 ,R ,a1 ,b1 2 (I + R)
m()M0 +[0 ]
m
S,2 ,0,b2 ,0
1 S m1 ,R ,a1 ,b1 2 S m2 ,0,b2 ,c2 .
,
Hence (5.52) is proved.
Using (5.46), classical considerations imply that (5.52) yield that for any
M0 N0 , and any seminorm S mM0 (),R ,a,b , there exist a constant C > 0 and
,
two seminorms S m1 ,R ,a1 ,b1 , S m2 ,a2 ,b2 ,c2 such that for any 1 , 2 S we
,
have
M0 S mM0 (),R ,a,b C1 S m1 ,R ,a1 ,b1 2 S m2 ,a2 ,b2 ,c2 .
,
In Section 5.5.4, we will see that for any seminorm S m
a,b,
c
, ,
constant C > 0 and a seminorm S m,R ,a,b such that
(5.53)
there exist a
S m
a,b,
c
, ,
CS m,R
,a,b .
(5.54)
Inequalities (5.54) together with (5.53) and Lemma 5.4.11 (to pass from S to
m1
m2
, S,
) conclude the proof of Theorem 5.5.3 in the case > .
S,
362
Chapter 5. Quantization on graded Lie groups
Note that the proof of the case > above also shows:
m1
m2
and 2 S,
, then there
Corollary 5.5.8. We assume 1 > 0. If 1 S,
m
exists a unique symbol in S, , m = m1 + m2 , such that
Op() = Op(1 )Op(2 ).
Moreover, for any M N0 , we have
m()M
{
1 Xx 2 } S,
.
(5.55)
(5.56)
[]M
Furthermore, the mapping
1
m()M
m
S,
S,
,
{ []M 1 Xx 2 }
is continuous.
Consequently, we can also write
1 Xx 2 ,
j=0
(5.57)
[]=j
in the sense of an asymptotic expansion as in Denition 5.5.2.
The case = is more delicate to prove but relies on the same kind of
arguments as above. If = , the asymptotic formula (5.56) does not bring any
improvement and, in this sense, is not interesting.
We will need the following variation of the properties given in Lemma 5.5.6
obtained using Corollary 5.4.3 instead of Theorem 5.4.1.
m
with 1 0. We denote by x its associated
Lemma 5.5.9. Let S,
kernel. Let 0 and m < Q. Then there exist a constant C > 0 and a seminorm
m ,a,b,c such that
S,
m ,a,b,c .
z x (z)dz CS,
G
We may replace
m ,a,b,c
S,
with S m,R ,a,b
,
Proof of Lemma 5.5.9. By Part 2 of Corollary 5.4.3, z x (z) is a continuous
bounded function if m < Q hence the integral z<1 z x (z)dz is nite.
By the CauchySchwartz inequality, we have
<
<
1
1
Q
2
z
z2+Q+ 2 x (z)2 dz
z x (z)dz
z>1
z>1
[]=M
q x L2 (G) ,
z>1
5.5. Symbolic calculus
363
where M/2 N is the smallest integer divisible by 1 , . . . , n satisfying M
2+Q+ 12 , having chosen (3.21) with p = M for quasinorm. By Part 1 of Corollary
5.4.3, the sum above is nite when m M < Q/2, which holds true.
Using Lemma 5.5.9 instead of Lemma 5.5.6 in the proof of Lemma 5.5.10
produces the following result.
Lemma 5.5.10. We x a positive Rockland operator of homogeneous degree . Let
m1 R, 1 0 with = 1, 0 Nn0 , and M, M1 N0 . We assume that
m1 + M1 < Q
,
M1 + m2 + (c0 + n + max(M1 , M )) 0
where
c0 :=
[ ].
max
[02 ][0 ]
[ ][02 ],  02 
Then there exist a constant C > 0, and two seminorms S m1 ,R ,a1 ,b1 ,
,
we have
S m2 ,0,b2 ,0 , such that for any 1 , 2 S and any (x, ) G G
,
Xx0 1 2 (x, )
1 (x, ) Xx 2 (x, ) L (H )
[]M
C1 S m1 ,R ,a1 ,b1 2 S m2 ,0,b2 ,0 .
,
The details of the proof of Lemma 5.5.10 are left to the reader. The rst
inequality in the statement just above shows that we will require the ability to
choose m1 as negative as one wants. We can do this thanks to the following remark:
Lemma 5.5.11. Let 1 , 2 S . For any X g and any 1 , 2 S , we have
(1 (X)) 2 = 1 (Xx 2 ) + 1 ((X)2 ).
More generally, for any Nn0 , we have
{1 (X) } 2 =
where
1 {(X)1 Xx2 2 },
[1 ]+[2 ]=[]
denotes a linear combination independent of 1 , 2 .
Note that in the expression above, (X)1 and Xx2 commute.
Proof of Lemma 5.5.7. We keep the notation of Lemma 5.5.4. Using integration
by parts and the Leibniz formula, we obtain
z =z 1,x (z1 )(z) 2 (xz 1 , ) dz
X
(1 (X)) 2 (x, ) =
1
G
z =z (z1 ) 2 (xz 1 , ) + (z) X
z =z 2 (xz 1 , ) dz
1,x (z) X
=
G
=
G
1,x (z) (z) (X)2 (xz 1 , ) + (z) Xx2 =xz1 2 (x2 , ) dz.
364
Chapter 5. Quantization on graded Lie groups
This shows the rst formula. The next formula is obtained recursively.
We can now sketch the proof of Theorem 5.5.3 in the case = .
Sketch of the proof of Theorem 5.5.3 with = . We assume = [0, 1). Writing 1 = 1 (I + R)N (I + R)N and using Lemma 5.5.11, it suces to prove
(5.52) for m1 as negative as one wants. We proceed as in the proof of the case >
replacing Lemma 5.5.5 with Lemma 5.5.10. The details are left to the reader.
5.5.3 Adjoint of a pseudodierential operator
Here we prove that the classes m
, are stable under taking the formal adjoints of
operators.
Theorem 5.5.12. We assume 1 0 with = 1 and m R. If T m
,
then its formal adjoint T is also in m
T is
, . Moreover, the mapping T
continuous on m
, .
Recall that the formal adjoint of an operator T : S(G) S (G) is the
operator T : S(G) S (G) dened by
, S(G)
T (x) (x) dx =
(x) T (x) dx.
G
We observe that the operator T = Op() m
, maps S(G) to itself continuously (see Theorem 5.2.15) and therefore has a formal adjoint T .
Before beginning the proof of Theorem 5.5.12, let us point out some of its
consequences.
Corollary 5.5.13.
1. We assume 1 0 with = 1, and m R.
Any T m
, extends uniquely to a continuous operator on S (G). Furm
thermore the mapping T T from , to the space L (S (G)) of continuous
operators on S (G) is linear and continuous.
2. Any smoothing operator T maps continuously the space E (G) of
compactly supported distributions to the Schwartz space S(G). Furthermore
the mapping T T from to the space L (E (G), S(G)) of continuous
mappings from E (G) to S(G) is linear and continuous.
Proof of Corollary 5.5.13. We admit Theorem 5.5.12 (whose proof is given below).
The statement then follows by classical arguments of duality and Theorem 5.2.15
for Part 1, and Part 2 of Theorem 5.4.9 for Part 2.
Let us start the proof of Theorem 5.5.12 by observing that the symbol ()
of the adjoint T of T = Op() is necessarily known and unique at least formally
or under favourable conditions such as in the case of a smoothing operator:
5.5. Symbolic calculus
365
Lemma 5.5.14. Let S and let : (x, y) x (y) be its associated kernel.
We set
xy1 (y 1 ),
()
x (y) :=
x, y G.
Then () : (x, y) x (y) is smooth on GG and for every Nn0 , x X x
is continuous from G to S(G).
The symbol () dened via
()
()
() (x, ) := FG (()
x )(),
(x, ) G G,
is a smooth symbol in the sense of Denition 5.1.34 and satises
(Op()) = Op( () ).
In particular, if does not depend on x, then () = .
Note that this operation is an involution since
xy1 (y 1 ).
x (y) =
()
then we have dened the adjoint
Recall that if = {(x, ), (x, ) G G}
symbol
= {(x, ) , (x, ) G G},
(see Theorem 5.2.22). Hence we may write
(x, ) := (x, ) .
Proof of Lemma 5.5.14. By Corollary 3.1.30, we have
o
o 1 {
Xxo {()
xy1 (y 1 )} = (1)o  X
xy1 (y 1 )}
x (y)} = Xx {
y1 =y
= (1)o 
Qo , (y 1 )Xy1 =y1 {
xy1 (y 1 )}
o , [][o ]
= (1)o 
Qo , (y 1 )Xx1 =xy1 {
x1 (y 1 )},
o , [][o ]
where the Qo , s are ([o ] [])homogeneous polynomials. The regularity of
()
described in Theorem 5.4.9 implies that () : (x, y) x (y) is smooth in x and
y (but maybe not compactly supported in x), and it is also Schwartz in y in such
()
a way that all the mappings G x Xx x S(G) are continuous. Clearly
()
() (x, ) = (x ) denes a smooth symbol () .
366
Chapter 5. Quantization on graded Lie groups
Let , S(G) and let x G. The regularity of described in Theorem
5.4.9 justies easily the following computations:
(Op())(x)(x)dx =
x (x)(x)dx =
(z)x (z 1 x)(x)dzdx
G
G
G G
()
(z)
x(z1 x)1 ((z 1 x)1 )(x)dzdx
=
G G
()
(z)z (x1 z)(x)dzdx
=
G G
()
(z) z (z)dz.
=
G
This shows that Op() (z) = z (z).
()
In general, () is not the adjoint of the symbol , unless for instance
it does not depend on x G. However, we can perform formal considerations to
link () with in the following way: using the Taylor expansion for x in x (see
equality (5.27)), we obtain
()
x (y) = xy 1 (y)
q (y 1 )Xx x (y) =
q (y)Xx x (y).
we get
Thus, taking the group Fourier transform at G,
() (x, ) = (()
x )
(
q (y)Xx x (y)) =
Xx (x, ) .
m
From Theorem 5.2.22 we know that if S,
then
m()[]
Xx (x, ) S,
(5.58)
From these formal computations we see that the main problem is to estimate the
remainder coming from the use of the Taylor expansion. This is the purpose of the
following technical lemma.
Lemma 5.5.15. We x a positive Rockland operator of homogeneous degree . Let
m R, 1 0 with = 0 and = 1, 0 Nn0 , and M, M1 N0 . We
assume that M M1 and ( )M + Q > m + [0 ] + M1 + Q. Then there
m ,a,b,0 , such that for any S
exist a constant C > 0, and a seminorm S,
we have
and any (x, ) G G
m ,a,b,0 .
Xx0 () (x, )
Xx (x, ) (I + R)M1 L (H ) CS,
[]M
5.5. Symbolic calculus
367
Proof of Lemma 5.5.15, case 0 = 0. By Lemma 5.5.14 and the observations that
follow, we have
() (x, )
Xx (x, )
[]M
=
xz1 (z)
q (z 1 )Xx x (z) (z) dz
[]M
x (z)
=
G
Rx,M (z 1 )(z) dz,
x (z)
where Rx,M
denotes the
v xv (z) of order M at
remainder of the (vectorvalued) Taylor expansion of
0. Using (5.48), we can integrate by parts to obtain
()
(x, )
Xx (x, ) (I + R)M1
[]M
[1 ]+[2 ]M1
2
X
[1 ]+[2 ]M1
z1=z RXz2 =z x (z2 ) (z 1 )(z) dz
X
1
x,M
1
X 1 (z2 )
2 =z x1 x1
Rx1z=x,M
[1 ]
(z 1 )(z) dz.
Taking the operator norm, we have
() (x, )
Xx (x, ) (I + R)M1 L (H )
[]M
2
X
[1 ]+[2 ]M1
X 1 (z2 )
2 =z x1 x1
Rx1z=x,M
[1 ]
(z 1 )dz.
For z < 1, we will use Taylors theorem, see Theorem 3.1.51:
2 X 1 (z2 )
X
1
2 =z x1 x1
z2=z Xx1 x (z2 ),
Rx1z=x,M
(z
)
z[] sup Xz X
[1 ]
2
1
1
(M [1 ])+ +1
[]>(M [1 ])+
x1 G
together with the estimate for z near the origin given in Theorem 5.4.1. The link
between left and right derivatives, see (1.11), implies
z2=z Xx1 x (z2 ) = sup Xz Xz2=z Xx1 x (z2 ).
sup Xz X
x1 G
x1 G
Proceeding as in the proof of Lemma 5.5.6, we obtain that the integral
2 X 1 (z2 )
X
z2 =z x
x
Rx,M
(z 1 )dz
[1 ]
z<1
z[] sup Xx1 Xz22=z Xx11 x1 (z2 )dz
(M [1 ])+ +1
[]>(M [1 ])+
z<1
x1 G
368
Chapter 5. Quantization on graded Lie groups
is nite whenever [] + Q > (m + [2 ] + ([] + [1 ]) + Q)/ with the indices
as above. These conditions are implied by the hypotheses of the statement. The
estimates for z large given in Theorem 5.4.1 show directly that the integral
2 X 1 (z2 )
X
2 =z x1 x1
Rx1z=x,M
(z 1 )dz,
[1 ]
z>1
is nite. Collecting the various estimates yields the statement in the case =
0
and 0 = 0.
Proof of Lemma 5.5.15, general case. We proceed as above and introduce the derivatives with respect to x. We obtain
Xx0
x (z)
Xx0 () (x, )
Xx (x, ) =
R0,M
(z 1 )(z) dz.
G
[]M
And adding (I + R)M1 , we have
Xx (x, ) (I + R)M1
Xx0 () (x, )
[]M
[1 ]+[2 ]M1
X 1 X 0
xx1 (z2 )
2
X
x
2 =z x1
Rx1z=0,M
[1 ]
(z 1 )(z) dz.
Taking the operator norm, we have
Xx (x, ) (I + R)M1 L (H )
Xx0 () (x, )
[]M
[1 ]+[2 ]M1
2
X
X 1 X 0
xx (z2 )
x
2 =z x1
Rx1z=0,M
[1 ]
(z 1 )dz.
For z < 1, we use the more precise version of Taylors theorem than in the case
0 = 0:
2
X
X 1 Xx0
xx (z2 )
z2 =z x1
Rx,M
[1 ]
(z 1 )
z[]
sup
y
(M [1 ])+ +1
[]>(M [1 ])+
(M [1 ])+ +1
z
2 X 1 X 0 (z2 ).
Xy X
z2 =z y
x
xy
We proceed as in the proof of Lemma 5.5.5, that is, we use (5.51) to obtain
sup
y
(M [1 ])+ +1
[0 ][0 ]
0 0 
z
2 X 1 X 0 (z2 )
Xy X
z2 =z y
x
xy
z[0 ][0 ]
sup
x1 G
[0 ]=[]+[0 ]
We conclude by adapting the case 0 = 0.
z2=z (z2 ).
Xx10 X
x1
2
5.5. Symbolic calculus
369
To take into account the dierence operator, we will use the following observation.
Lemma 5.5.16. For any Nn0 and S , () can be written as a linear
combination (independent of ) of { }() over Nn0 , [ ] = []. This is
the same linear combination as when writing as a linear combination of
{ } .
Proof of Lemma 5.5.16. For S , let be the kernel associated with the
symbol and similarly for any other symbol.
Let us prove Part 1. We have
,xy1 (y 1 ).
{
q () ,x }(y) = q (y)
As q is a []homogeneous polynomial, by Proposition 5.2.3, q is a linear combination of q over multiindices Nn0 satisfying [ ] = []. Hence
{
q () ,x }(y) =
q ,xy1 (y 1 ) =
[ ]=[]
where
obtain
{
q }() (y),
[ ]=[]
means taking a linear combination. Taking the Fourier transform, we
FG {
q () ,x }() = () (x, ) =
{ }() .
[ ]=[]
We can now prove Theorem 5.5.12 in the case > .
Proof of Theorem 5.5.12 with > . We assume > . We x a positive Rockland
operator of homogeneous degree . Let us show that for any 0 , 0 Nn0 , and
m ,a ,b ,0 ,
M0 N, there exists M M0 , a constant C > 0 and a seminorm S,
1 1
such that for any S we have
m()M0 [0 ]+[0 ]
Xx 0 0 M (x, ) (I + R)
L (H )
m ,a ,b ,0 ,
CS,
1 1
(5.59)
where we have denoted M := () []M Xx . By Lemma 5.5.16, it
suces to show (5.59) only for 0 = 0.
Let 0 N0 and M0 N. Let M1 N0 be the smallest nonnegative integer
such that
m ( )M0 + [0 ]
M1 .
We choose M max(M0 , M1 ) such that ( )M + Q > m + [0 ] + M1 + Q.
This is possible as > . Then (5.59) follows from the application of Lemma 5.5.15
to M, M1 and the symbol .
370
Chapter 5. Quantization on graded Lie groups
Using (5.58), classical considerations imply that (5.59) yields that for any
M0 N0 , and any seminorm S mM0 (),R ,a,b , there exist a constant C > 0 and
,
m ,a ,b ,0 , such that for any 1 , 2 S
a seminorm S,
we have
1 1
m ,a ,b ,0 .
M0 S mM0 (),R ,a,b CS,
1 1
,
We can then conclude as in the proof of Theorem 5.5.3 in the case > .
In fact, we have obtained a much more precise result:
m
Corollary 5.5.17. We assume 1 > 0. If S,
, then there exists a unique
()
m
symbol in S, such that
(Op()) = Op( () ).
Furthermore, for any M N0 ,
m()M
{ () (x, )
Xx (x, )} S,
.
[]M
Moreover, the mapping
1
m()M
m
S,
S,
,
{ () (x, ) []M Xx (x, )}
is continuous.
Consequently, we can also write
()
Xx ,
j=0
(5.60)
[]=j
where the asymptotic was dened in Denition 5.5.2.
As for composition, in the case = , the asymptotic formula does not bring
any improvement and, in this sense, is not interesting. The proof of this case is
more delicate to prove but relies on the same kind of arguments as above. Using
Lemma 5.5.9 instead of Lemma 5.5.6 in the proof of Lemma 5.5.15 produces the
following result:
Lemma 5.5.18. We x a positive Rockland operator of homogeneous degree . Let
m R, 1 0 with = 1, 0 Nn0 , and M, M1 N0 . We assume that
M M1
and
m + (M + c0 ) + M1 < Q,
where
c0 :=
max
[0 ][0 ]
[ ][0 ],  0 
[ ].
5.5. Symbolic calculus
371
Then there exist a constant C > 0, and a seminorm S m,R ,a,b , such that for any
,
we have
S and any (x, ) G G
Xx0 () (x, )
Xx (x, ) (I + R)M1 L (H ) CS m,R ,a,b .
,
[]M
The details of the proof of Lemma 5.5.18 are left to the reader. The conditions
in the statement just above show that we will require the ability to choose m as
negative as one wants. We can do this thanks to the following remark.
Lemma 5.5.19. For any S and any X g, we have
{(X)}() = () (x, ) (X) {Xx }() (x, ).
More generally, for any Nn0 , we have
{(X) }() =
{Xx1 }() (X)2 ,
[1 ]+[2 ]=[]
where
denotes a linear combination independent of 1 , 2 .
Proof of Lemma 5.5.19. We keep the notation of Lemma 5.5.14. The kernel of
() (X) is given via
y () (y) = X
y {
X
xy1 (y 1 )} = Xx1 =xy1
x1 (y 1 ) Xy2 =y1
xy1 (y2 ),
x
having used (5.50) and the Leibniz property for vector elds. Hence we recognise:
y () (y) = (Xx x )() (y) (Xx )() (y),
X
x
and
() (X) = (Xx )() ((X))() .
This shows the rst formula. The second formula is obtained recursively.
We can now show sketch the proof of Theorem 5.5.3 in the case = .
Sketch of the proof of Theorem 5.5.3 with = . We assume = [0, 1). Writing = (I + R)N (I + R)N and using Lemma 5.5.19, it suces to prove (5.59)
for m as negative as one wants. We proceed as in the proof of the case >
replacing Lemma 5.5.15 with Lemma 5.5.18. The details are left to the reader.
m
5.5.4 Simplication of the denition of S,
In this section, we show that it is possible to choose = 0 in the denition
of symbols as it was pointed out in Remark 5.2.13 Part (3). This simplies the
m
given in Denition 5.2.11. We will also show
denition of the symbol classes S,
a pivotal argument in the proof of Theorems 5.5.3 and 5.5.12, namely Inequalities
(5.54).
372
Chapter 5. Quantization on graded Lie groups
Theorem 5.5.20. Let m, , R with 1 0 and = 1.
is in S m if and only if for each
(L) A symbol = {(x, ), (x, ) G G}
,
n
, N0 , the eld of operators
Xx = {Xx (x, ) : H H , (x, ) G G}
is in L
0,[]m[] (G) uniformly in x G, that is,
sup Xx (x, )L
xG
0,[]m[]
(G)
< .
(5.61)
Furthermore, the family of seminorms
m ,a,b,0 = sup sup X (x, )
S,
x
L
[]a xG
[]b
0,[]m[]
,
(G)
a, b N0 ,
m
.
yields the topology of S,
is in S m if and only if for each
(R) A symbol = {(x, ), (x, ) G G}
,
n
, N0 , the eld of operators
Xx = {Xx (x, ) : H H , (x, ) G G}
is in L
m+[][],0 (G) uniformly in x G, that is,
sup Xx (x, )L
xG
m+[][],0
(G)
< .
(5.62)
Furthermore, the family of seminorms
S m,R ,a,b = sup sup Xx (x, )L
,
[]a xG
[]b
m+[][],0
,
(G)
a, b N0 ,
m
.
yields the topology of S,
In other words,
is in S m if and only if for each
(R) a symbol = {(x, ), (x, ) G G}
,
n
, N0 , the eld of operators
Xx = {Xx (x, ) : H H , (x, ) G G}
is dened on smooth vectors and satisfy
sup
xG,G
Xx (x, )(I + R)
[]m[]
L (H ) <
for one (and then any) positive Rockland operator R of homogeneous degree
(as the symbol is given by a eld of operators dened on smooth vectors,
s
and since (I + R) acts on smooth vectors, this condition makes sense);
5.5. Symbolic calculus
373
is in S m if and only if for each
(L) a symbol = {(x, ), (x, ) G G}
,
n
, N0 , the eld of operators
Xx = {Xx (x, ) : H H[]m[] , (x, ) G G}
[]m[]
is dened on smooth vectors and has range in H
sup
xG,G
(I + R)
[]m[]
, and satises
Xx (x, )L (H ) <
for one (and then any) positive Rockland operator R of homogeneous degree
. The notion of a eld having range in a Sobolev space Hs is described in
s
Denition 5.1.10 and allows us to compose on the left with (I + R) with
s = [] m [] here, see (5.4).
Naturally, the condition does not depend on the choice of the positive Rockland
operator R.
Theorem 5.5.20 makes it considerably easier to check whether a symbol is
in one of our symbol classes. However using the denition with any has the
advantages
1. that we see easily that the symbols are elds of operators acting on smooth
vectors,
m
, m R, form an algebra (cf.
2. that we see easily that the symbols in S,
Theorem 5.2.22),
3. and that the properties for the multipliers in R in Proposition 5.3.4 are for
the denition with any .
While showing Theorem 5.5.20, we will also nish the proofs of Theorems
5.5.3 and 5.5.12. Indeed, an important argument used in the proof of Theorems
5.5.3 and 5.5.12 (i.e. the properties of stability under composition and taking the
adjoint) is Inequality (5.54) which can easily be seen as equivalent to Part 2 of
Theorem 5.5.20.
Before showing Theorem 5.5.20, let us summarise what has been shown in
the proofs of Theorems 5.5.3 and 5.5.12 up to before the use of Inequality (5.54):
1 2 S m1 +m2 ,R ,a,b
1 S m1 ,R ,a1 ,b1 2 S m2 ,a2 ,b2 ,c2 ,
(5.63)
() S m,R ,a,b
m ,a ,b ,0 ;
S,
(5.64)
these estimates are valid for any , 1 , 2 S in the sense that for any seminorm
on the left hand side, one can nd seminorms on the right.
Proof of Theorem 5.5.20. Using Estimate (5.64) together with the properties of
taking the adjoint and of the dierence operators together, one checks easily that
374
Chapter 5. Quantization on graded Lie groups
m ,a,b,0 , a, b
the two families of seminorms { S m,R ,a,b , a, b N} and { S,
,
N} yield the same topology on S and that taking the adjoint of a symbol is
continuous for this topology. Consequently, for any R, any symbol S
and any seminorm S m,R ,a,b , we have
,
(I + R) S m+,R ,a,b (I + R) S m+,R ,a1 ,b1 S m,R ,a2 ,b2 ,
,
having used (5.63) and the fact that (I + R) S . As taking the adjoint is a
continuous operator for the S m,R topology, we have obtained
(I + R) S m+,R ,a,b S m,R ,a3 ,b3 .
,
One checks easily that
a, b, c N0
m ,a,b,c max (I + R) m+,R
S,
S
,a,b ,
c
whereas
a, b N0
m ,a,b,m+a+b .
S m,R ,a,b S,
,
This easily implies that the topologies on S coming from the two families of
m ,a,b,c , a, b, c N0 } and { m,R
seminorms { S,
S
,a,b , a, b N0 } coincide. This
,
m
) concludes the proof of
together with Lemma 5.4.11 (to pass from S to S,
Theorem 5.5.20.
5.6
Amplitudes and amplitude operators
In this section, we discuss the notion of an amplitude extending that of the symbol,
to functions/operators depending on both space variables x and y. This allows
for another way of writing pseudodierential operators as amplitude operators,
analogous to Formula (2.27) in the case of compact groups. However, as in the
classical theory, or as in Theorem 2.2.15 in the case of compact groups, we can
show that amplitude operators with symbols in suitable amplitude classes reduce
to pseudodierential operator with symbols in corresponding symbol classes, with
asymptotic formulae relating amplitudes to symbols.
5.6.1 Denition and quantization
Following the Euclidean and compact cases, it is natural to dene amplitudes in
the following way, extending the notion of symbols from Denitions 5.1.33 and
5.1.34:
Denition 5.6.1. An amplitude is a eld of operators
{A(x, y, ) : H H , G}
5.6. Amplitudes and amplitude operators
375
depending on x, y G, satisfying for each x, y G
a, b R
L
A(x, y, ) := {A(x, y, ) : H H , G}
a,b (G).
is said to be continuous in
An amplitude {A(x, y, ) : H H , G}
x, y G whenever there exists a, b R such that
x, y G
L (G),
A(x, y, ) := {A(x, y, ) : H H , G}
a,b
and the map (x, y) A(x, y, ) is continuous from G G Rn Rn to the
Banach space L
a,b (G).
is said to be smooth
An amplitude A = {A(x, y, ) : H H , G}
in x, y G whenever it is a eld of operators depending smoothly on
(x, y) G G (see Remark 1.8.16) and, for every 1 , 2 Nn0 , the eld
is continuous.
{x1 y2 A(x, y, ) : H H , G}
Clearly if an amplitude A = {A(x, y, )} does not depend on y, that is,
A(x, y, ) = (x, ), then it denes a symbol = {(x, )}. More generally any
amplitude A = {A(x, y, )} denes a symbol given by (x, ) = A(x, x, ). In
Section 5.6.2, we will dene amplitude classes and give other examples of amplitudes.
Similarly to the symbol case, one can associate a kernel with an amplitude:
Denition 5.6.2. Let A be an amplitude. For each (x, y) G G, let x,y S (G)
be the unique distribution such that
FG (x,y )() = A(x, y, ).
The map G G (x, y) x,y S (G) is called its kernel.
As in the symbol case, the map G G (x, y) x,y S (G) is smooth,
see Lemma 5.1.35 for the proof of this as well as for the existence and uniqueness
of x,y in the case of symbols.
Before dening the amplitude quantization, we need to open a (quick) parenthesis to describe the following property from distribution theory:
Lemma 5.6.3. Let G G (x, y) x,y S (G) be a continuous mapping. For
each x, we consider the distribution
x dened by
x (y)(y)dy = lim
x,w (y 1 x)(y) (wy 1 )dydw,
G
0
GG
where D(G), 1 D(G), G 1 = 1 and (z) = Q (1 z), > 0.
Indeed this limit exists and is independent of the choice of 1 .
This denes a continuous map G x
x D (G).
376
Chapter 5. Quantization on graded Lie groups
Proof of Lemma 5.6.3. Since x,y S (G), there exists a seminorm S(G),N
such that
x,y ,  Cx,y,N S(G),N .
S(G)
Furthermore, since the map G G (x, y) x,y S (G) is smooth, we obtain
that the constant Cx,y,N = x,y S (G),N can be chosen locally uniform with
respect to x and y. Furthermore, xing two compacts K1 and K2 of G, there
exists a seminorm S(G),N (depending on K1 and K2 ) such that the map
((x, y), (x , y )) (K1 K2 ) (K1 K2 ) x,y x ,y S (G),N ,
is uniformly continuous. This is easily proved using a cover of the compacts K1 K2
by balls of suciently small radius, and the continuity at each centre of these balls.
For any 1 D(G), > 0 and x G, we dene the distribution T1 ,,x by
T1 ,,x () :=
GG
x,w (y 1 x)(y) (wy 1 )dydw,
where D(G) is supported in a xed compact K G. Using the change of
variable from w to z with z = 1 (wy 1 ), so that w = (z)y, we obtain
T1 ,,x () =
GG
x,(z)y (y 1 x)(y)1 (z)dydz.
Therefore, for any 1 , 2 (0, 1), we get
(T1 ,1 ,x T1 ,2 ,x )()
1
1
x,(1 z)y (y x) x,(2 z)y (y x) (y)1 (z)dydz
=
GG
sup
zsupp1
ysupp
x,(1 z)y x,(2 z)y S (G),N S(G),N 1 L1 (G) ,
where S(G),N is chosen with respect to the compact sets
{x}
and
{(z)y, [0, 1], z supp1 , y K2 }.
This shows that the scalar sequence (T1 ,,x ()) converges as 0 and that the
linear map
1 D(G) lim T1 ,,x (),
0
(5.65)
extends continuously to L1 (Ko ) C for any compact Ko G. Thus the map
given in (5.65) is given by integration against a locally bounded function on G.
5.6. Amplitudes and amplitude operators
377
Let us show that the map given in (5.65) is invariant under left or right
translation. Indeed, modifying the argument above we obtain
T1 ,,x () T1 (yo1 ),,x ()
1
x,(z)y x,((zyo ))y (y x)(y)1 (z)dydz
=
GG
sup
zsupp1
ysupp
x,(z)y x,((zyo ))y S (G),N S(G),N 1 L1 (G)
for a suitable seminorm S(G),N , (depending locally on yo ). Since the two sequences ((z)y)>0 and (((zyo ))y)>0 converge to y in G, we see that
lim T1 ,,x () = lim T1 (yo1 ),,x (),
0
0
and the same is true for right translation. Therefore, the locally bounded function
given by the mapping (5.65) is a constant which we denote by T0,x ():
1 .
lim T1 ,,x () = T0,x ()
0
One checks easily that T0,x (), D(G), supp K, denes a distribution
x D (G) which is therefore independent of 1 . Rening the argument given
above shows that
x D (G) depends continuously on x G.
If G G (x, y) x,y S (G) is a continuous mapping, we will allow
ourselves to denote the distribution dened in Lemma 5.6.3 by
x (y) := x,y (y 1 x).
This closes our parenthesis about distribution theory.
We can now dene the operator
T = AOp(A)
associated with an amplitude A = {A(x, y, )} with amplitude kernel x,y , by
(y)x,y (y 1 x)dy, D(G), x G.
(5.66)
T (x) :=
G
The quantization dened by formula (5.66) makes sense for any amplitude A =
{A(x, y, )}. Clearly the quantization mapping A AOp(A) is linear. However,
as in the Euclidean or compact cases, it is injective but not necessarily 11 since
dierent amplitudes may lead to the same operator, in contrast to the situation
for symbols, cf. Theorem 5.1.39.
378
Chapter 5. Quantization on graded Lie groups
Remark 5.6.4. If an amplitude A = {A(x, y, )} does not depend on y, that is,
A(x, y, ) = (x, ), then the corresponding symbol = {(x, )} yield the same
operator:
AOp(A) = Op()
since in this case the amplitude x,y is a function/distribution x independent of
y which coincides with the kernel of the symbol .
As in the symbol case in Lemma 5.1.42, we may see AOp(A) as a limit of
nice operators in the following sense:
Lemma 5.6.5. If A = {A(x, y, )} is an amplitude, we can construct explicitly a
family of amplitudes A = {A (x, y, )}, > 0, in such a way that
1. the kernel ,x,y (z) of A is smooth in both x, y and z, and compactly supported in x and y,
2. the associated kernel
,x (y) = ,x,y (y 1 x) is smooth and compactly supported in both x, y,
3. if S(G) then AOp(A ) D(G), and
4. AOp(A ) AOp(A) uniformly on any compact subset of G.
0
Proof of Lemma 5.6.5. We use the same notation D(G),  and proj, as
in the proof of Lemma 5.1.42. We consider for any (0, 1) the amplitude given
by
A (x, y, ) := (x) (y)11 A(x, y, ) proj, .
By Denition 5.6.2 and the Fourier inversion formula (1.26), the corresponding
kernel is
,x,y (z) = (x) (y)
Tr A(x, y, ) proj, (z) d(),
1
which is smooth in x, y and z and compactly supported in x and y. The rest follows
easily.
There is a simple relation between the amplitudes of an operator and its
adjoint, much simpler than in the symbol case:
Proposition 5.6.6. Let A be an amplitude. Then B given by
B(x, y, ) := A(y, x, )
is also an amplitude. Furthermore, the formal adjoint of the operator T = AOp(A)
is T = AOp(B). If {x,y (z)} is the kernel of A, then the kernel of B is given via
(x, y, z)
y,x (z 1 ).
5.6. Amplitudes and amplitude operators
379
Proof. On one hand, from the amplitude quantization in (5.66), we compute for
, D(G), that
(T , ) =
(y)x,y (y 1 x)(x)dy
dx = (, T ),
G
therefore
T (y) =
or, equivalently,
T (x) =
G
G
x,y (y 1 x)(x)dx
y,x (x1 y)(y)dy.
One the other hand, the amplitude kernel for B is x,y satisfying
(x,y ) = B(x, y, ) = A(y, x, ) = (y,x ) = (y,x ),
with y,x (z) =
y,x (z 1 ), and therefore,
x,y (z) = y,x (z) =
y,x (z 1 ).
By (5.66), this implies that T = AOp(B).
5.6.2 Amplitude classes
m
Again similarly to the symbol case, we may dene the amplitude classes AS,
. This
is done in analogy to Denition 5.2.11 for symbols and its equivalent reformulation
in (5.29).
Denition 5.6.7. Let m, , R with 1 1. An amplitude A is called
an amplitude of order m and of type (, ) whenever, for each , Nn0 and
R, the eld {Xx1 Xy2 A(x, y, )} is in L
,[]m([1 ]+[2 ])+ (G) uniformly
in (x, y) G, i.e. if
sup Xx1 Xy2 A(x, y, )L
x,yG
,[]m([1 ]+[2 ])+
(G)
< .
(5.67)
m
In this case, proceeding in a similar way to S,
in Section 5.2.2, we see
1 2
that the elds of operators Xx Xy A(x, y, ) act on smooth vectors and (5.67)
implies
sup (I + R)
[]m([1 ]+[2 ])+
x,yG
G
Xx1 Xy2 A(x, y, )(I + R) L (H ) < .
(5.68)
The converse also holds.
380
Chapter 5. Quantization on graded Lie groups
m
m
The amplitude class AS,
= AS,
(G) is the set of amplitudes of order m
and of type (, ). We also dene
AS :=
m
AS,
,
mR
the class of smoothing amplitudes. As in the case of symbols, the class AS is
independent of and and can be denoted just by AS .
m
is a vector
It is a routine exercise to check that each amplitude class AS,
space and that we have the inclusions
m1 m2 ,
1 2 ,
1 2
ASm11,1 ASm22,2 .
(5.69)
We assume that a positive Rockland operator R of degree is xed. If A is
an amplitude and a, b, c [0, ), we set
m ,a,b,c
A(x, y, )AS,
:= sup (I + R)
[]m([1 ]+[2 ])+
c
[]a, [1 ],[2 ]b
Xx1 Xy2 A(x, y, )(I + R) L (H ) ,
and
m ,a,b,c :=
AAS,
sup
(x,y)GG, G
m ,a,b,c .
A(x, y, )AS,
m ,a,b,c , a, b, c [0, ), are
Again, one checks easily that the resulting maps S,
m
. Furthermore, taking a, b, c as nonnegative
seminorms over the vector space AS,
m
with the structure of a Frechet space. The class of
integers, they endow AS,
smoothing amplitudes AS
is then equipped with the topology of projective
limit. Similarly to the case of symbols in Proposition 5.2.12, two dierent positive
Rockland operators give equivalent families of seminorms.
The inclusions given in (5.69) are continuous for these topologies.
m
m
are examples of amplitudes in AS,
which do not depend
Symbols in S,
m
does not depend on y,
on y. Conversely, if an amplitude A = {A(x, y, )} in AS,
m
. More
that is, A(x, y, ) = (x, ), then it denes a symbol = {(x, )} in S,
generally we check easily:
m
, then the symbol given by
Lemma 5.6.8. If A = {A(x, y, )} is in AS,
(x, ) := A(x, x, )
m
is in S,
.
A wider class of examples is given by the following property which can be
shown by an easy adaption of Proposition 5.3.4 and Corollary 5.3.7:
5.6. Amplitudes and amplitude operators
381
Corollary 5.6.9. Let R be a positive Rockland operator of degree . Let m R and
0 < 1. Let f : G G R+ (x, y, ) fx,y () C be a smooth function.
We assume that for every 1 , 2 Nn0 , we have
Xx1 Xy2 fx,y M m+([1 ]+[2 ]) ,
where M is as in Denition 5.3.1. Then
A(x, y, ) = fx,y ((R))
m
denes an amplitude A in AS1,
which satises
a, b, c N0
A
m ,a,b,c
AS1,
N, C > 0
C sup[1 ],[2 ]b Xx1 Xy2 fx,y M m+[1 +2 ] , ,
with and C independent of f .
This can also be generalised easily to multipliers in a nite family of strongly
commuting positive Rockland operators.
5.6.3
Properties of amplitude classes and kernels
One can readily prove properties for the amplitudes similar to the ones already
established for symbols. Here we note that although the subsequent properties
would follow also from Theorem 5.6.14 in the sequel and from the corresponding properties of symbols in Section 5.2.5, we now indicate what can be shown
concerning amplitudes and their classes by a simple adaptation of proofs of the
corresponding properties for symbols.
Proceeding as in Section 5.2.5, we also have the following properties for the
amplitude classes:
Proposition 5.6.10. Let 1 0 and = 1.
m
have kernel x,y . Then we have the following properties.
(i) Let A AS,
1. For every x, y G and R, q Xx1 Xy2 x,y K,[]m[1 +2 ]+ ,
where we recall the notation q (x) = q (x1 ).
2. If 1 , 2 Nn0 then the amplitude {Xx1 Xy2 A(x, y, ), (x, y, ) G
is in AS m+[1 +2 ] with kernel Xx1 Xy2 x,y , and
G G}
,
m ,a,b+[ + ],c ,
Xx1 Xy2 A(x, y, )AS m+[1 +2 ] ,a,b,c CA(x, y, )AS,
1
2
,
with C = Cb,1 ,2 .
382
Chapter 5. Quantization on graded Lie groups
is
3. If o Nn0 then the amplitude {o A(x, y, ), (x, y, ) G G G}
m[o ]
in AS,
with kernel qo x,y , and
m ,a+[ ],b,c .
o A(x, )S m[o ] ,a,b,c Ca,o A(x, )S,
o
,
is in AS m with kernel
4. The symbol {A(x, y, ) , (x, ) G G G}
,
1
y,x (z ), and
x,y given by x,y (z) =
m ,a,b,c =
A(x, y, ) AS,
sup
(I+R) Xx1 Xy2 A(x, y, )(I+R)
[]m([1 ]+[2 ])+
c
[]a, [1 ],[2 ]b
L (H ) .
m1
m2
(ii) Let A1 AS,
and A2 AS,
have kernels 1,x,y and 2,x,y , respectively.
Then
A(x, y, ) := A1 (x, y, )A2 (x, y, )
m
denes the amplitude A in S,
, m = m1 + m2 , with kernel 2,x,y 1,x,y with
the convolution in the sense of Denition 5.1.19. Furthermore,
m
m ,a,b,c CA1 (x, y, ) m1
A(x, y, )S,
S
,a,b,c+a+m2 +b A2 (x, y, )S 2 ,a,b,c ,
,
where the constant C = Ca,b,c > 0 does not depend on A1 , A2 .
A direct consequence of Part (ii) of Proposition 5.6.10 is that the amplitudes
in the introduced amplitude classes form an algebra:
Corollary
5.6.11. Let 1 0 and = 1. The collection of symbols
m
AS
, forms an algebra.
mR
m
Furthermore, if A0 AS is smoothing and A AS,
is of order m R,
then A0 A and AA0 are also in AS
.
Another consequence of Part (ii) together with Lemma 5.2.17 gives the following property:
m
Corollary 5.6.12. Let 1 0 and = 1. Let A AS,
have kernel x,y . If
n
and are in N0 , then
AS
{(X) A (X) , (x, ) G G}
,
m+[]+[]
z x,y (z). Furthermore, for any a, b, c there exists C = Ca,b,c
with kernel Xz X
independent of A such that
m ,a,b,c CA
(X) A(X) AS,
AS m
, ,a,b,c+a+[]+[]+b
5.6. Amplitudes and amplitude operators
383
Proceeding as in Section 5.4.1, taking into account the dependence in x and
y, we obtain
m
with 1 0. Let x,y
Proposition 5.6.13. Let A = {A(x, y, )} be in AS,
denote its associated kernel.
1. If , 1 , 2 , o , o Nn0 are such that
m [] + [1 ] + [2 ] + ([o ] + [o ]) < Q/2,
z2 (Xxo Xyo q (z)x,y (z)) is square integrable and
then the distribution Xz1 X
for every x G we have
2
1 2 o o
m ,a,b,c
Xz Xz (Xx Xy q (z)x,y (z)) dz C sup A(x, )2AS,
G
where a = [], b = [o ] + [o ], c = [] + [1 ] + [2 ] + ([o ] + [o ]) and
C = Cm,,1 ,2 ,o ,o > 0 is a constant independent of A and x, y.
2. For any , 1 , 2 , o , o Nn0 satisfying
m [] + [1 ] + [2 ] + ([o ] + [o ]) < Q,
z2 Xxo Xy o q (z)x,y (z) is continuous on G for
the distribution z Xz1 X
every (x, y) G G and we have
2 X o Xyo q (z)x,y (z) C sup A(x, )AS m ,[],[ ]+[ ],[ ] ,
sup Xz1 X
z
x
o
2
o
,
zG
G
where C = Cm,,1 ,2 ,o ,o > 0 is a constant independent of A and x, y.
We now assume > 0. Then the map : (x, y, z) x,y (z) is smooth
on G G (G \{0}). Fixing a homogeneous quasinorm   on G, we have the
following more precise estimates:
at innity: For any M R and any , 1 , 2 , o , o Nn0 there exist C > 0 and
a, b, c N independent of A such that for all x G and z G satisfying
z 1, we have
1 2 o o
X X (X Xy q (z)x,y (z)) C sup A(x, y, )AS m ,a,b,c zM .
z
G
at the origin: For any , 1 , 2 , o , o Nn0 with Q+m+([o ]+[o ])[]+[1 ]+
[2 ] 0 there exist a constant C > 0 and computable integers a, b, c N0
independent of A such that for all x G and z G\{0}, we have, if
Q + m + ([o ] + [o ]) [] + [1 ] + [2 ] > 0,
384
Chapter 5. Quantization on graded Lie groups
then
1 2 o o
Xz Xz (Xx Xy q (z)x,y (z))
m ,a,b,c z
C sup A(x, )AS,
])[]+[ ]+[ ]
Q+m+([o ]+[o
1
2
G
and if
Q + m + ([o ] + [o ]) [] + [1 ] + [2 ] = 0,
then
1 2 o o
m ,a,b,c ln z.
Xz Xz (Xx Xy q (z)x,y (z)) C sup A(x, y, )AS,
G
5.6.4
Link between symbols and amplitudes
Symbols can be viewed as amplitudes which do not depend on the second variable
m
m
AS,
and, by Remark 5.6.4, we have the inclusion
of the group. Then S,
m
m
m
, = Op(S, ) AOp(AS, ).
The next theorem shows the converse, namely, that the class of operators
m
) is included in m
AOp(AS,
, . Therefore this will show that the amplitude quanm
m
coincides with the symbol quantization of S,
.
tization of AS,
m
Theorem 5.6.14. Let A AS,
with 1 0, = 1. Then AOp(A) is in
m
m
such that
, , that is, there exists a (unique) symbol S,
AOp(A) = Op().
Furthermore, for any M N0 , the map
1
m()(M +1)
m
S,
AS,
,
A (x, ) []M Xy A(x, y, )y=x
is continuous. If > , we have the asymptotic expansion
Xy A(x, y, )y=x .
(x, )
The proof of Theorem 5.6.14 is in essence close to the proofs of product and
adjoint of operators in mR m
, , see Theorems 5.5.12 and 5.5.3. As for these
theorems, it is helpful to understand formally the steps of the rigorous proof.
From the amplitude quantization in (5.66), we see that if AOp(A) can be
written as Op(), then, denoting by ,x the symbol kernel and by A,x,y the
amplitude kernel, we have
(y)A,x,y (y 1 x)dy =
(xz 1 )A,x,xz1 (z)dz
AOp(A)()(x) =
G
5.7. Calder
onVaillancourt theorem
whereas
Op()()(x) =
G
(y),x (y
385
x)dy =
(xz 1 ),x (z)dz.
Therefore, formally we must have
or equivalently A,x,y (y 1 x) = ,x (y 1 x) .
A,x,xz1 (z) = ,x (z)
Using the Taylor expansion in y = xz 1 for A,x,y at x, we have (again formally)
,x (z) = A,x,xz1 (z)
q (z)Xy A,x,y (z)y=x .
(5.70)
Note that the group Fourier transform in z of each term in the sum above is
FzG {
q (z)Xy=x
A,x,y (z)}()
=
=
Xy=x
FzG {A,x,y (z)}()
Xy=x
A(x, y, ).
Taking the group Fourier transform in z on both sides of (5.70), we obtain still
formally that
(x, )
Xy A(x, y, )y=x .
As in the proofs of Theorems 5.5.12 and 5.5.3, the crucial point is to control the
remainder while using Taylors expansion. The method is similar as in the proof
of Theorem 5.5.12 and the adaptation is easy and left to the reader.
Note that Theorem 5.6.14 together with Proposition 5.6.6 give another proof
of Theorem 5.5.12. This is not surprising given the similarity between the proof
of Theorems 5.6.14 and 5.5.12.
5.7
Calder
onVaillancourt theorem
In this section, we prove the analogue of the CalderonVaillancourt theorem, now
in the setting of graded Lie groups. This extends the L2 boundedness of operators
in the class 01,0 given in Theorem 5.4.17 to the classes 0, .
Theorem 5.7.1. Let T 0, with 1 0 and = 1. Then T extends to a
bounded operator on L2 (G).
Moreover, there exist a constant C > 0 and a seminorm 0, ,a,b,c with
computable integers a, b, c N0 independent of T such that
S(G)
T L2 (G) CT 0, ,a,b,c L2 (G) .
Before showing Theorem 5.7.1, let us mention that together with the pseudodierential calculus, it implies the following boundedness on Sobolev spaces L2s .
386
Chapter 5. Quantization on graded Lie groups
Corollary 5.7.2. Let T m
, with 1 0 and = 1. Then for any s R,
the operator T extends to a continuous operator from L2s (G) to L2sm (G):
S(G)
T L2sm (G) Cs,m,, T m
L2s (G) ,
, ,a,b,c
with some (computable) integers a, b, c depending on s, m, , .
Proof of Corollary 5.7.2. Let R be a positive Rockland operator. By the composition theorem (e.g. Theorem 5.5.3), we have
(I + R)
m+s
T (I + R) 0, .
s
Therefore, by Theorem 5.7.1, we have
(I + R)
m+s
T (I + R) L (L2 (G)
s
m+s
T (I + R) 0, ,a1 ,b1 ,c1
s
(I + R)
T m
,
, ,a2 ,b2 ,c2
by Theorem 5.5.3.
Remark 5.7.3. Combining the results obtained so far, for each (, ) with 1
0 and
= 1, we have therefore obtained an operator calculus, in the sense that
the set mR m
, forms an algebra of operators, stable under taking the adjoint,
and acting on the Sobolev spaces in such a way that the loss of derivatives in L2
is controlled by the order of the operator.
Note that the L2 boundedness in the case (, ) = (1, 0) was already proved
by dierent methods, see Theorem 5.4.17 and its proof. With the same proof as in
the corollary above, one obtains easily boundedness for Lp Sobolev spaces in this
case:
Corollary 5.7.4. Let T m
1,0 . Then for any s R and p (1, ) the operator T
extends to a continuous operator from Lps (G) to Lpsm (G):
S(G)
T Lpsm (G) Cs,m,, T m
Lps (G) ,
, ,a,b,c
with some (computable) integers a, b, c depending on s, m, , .
Proof of Corollary 5.7.4. As above, (I + R)
Corollary 5.4.20 we have
(I + R)
m+s
T (I + R) L (Lp (G))
by Theorem 5.5.3.
m+s
T (I + R) 0 therefore, by
s
m+s
T (I + R) 0 ,a1 ,b1 ,c1
(I + R)
T 0 ,a2 ,b2 ,c2 ,
The rest of this section is devoted to the proof of the CalderonVaillancourt
Theorem, that is, Theorem 5.7.1. In Section 5.7.2, we prove the result for = = 0.
5.7. Calder
onVaillancourt theorem
387
The proof will rely on an analogue on G of the familiar decomposition of Rn into
unit cubes presented in Section 5.7.1. The case = (0, 1) will be proved in
Section 5.7.4 and its proof relies on the case = = 0 and on a bilinear estimate
proved in Section 5.7.3. The case of = [0, 1) will then be proved and this will
imply Theorem 5.7.1 thanks to the continuous inclusions between symbol classes
(see (5.31)).
5.7.1 Analogue of the decomposition into unit cubes
In this section, we present an analogue of the dyadic cubes, more precisely we construct a useful covering of the general homogeneous Lie group G by unit balls and
the corresponding partition of unity with a number of advantageous properties.
The proof is an adaptation of [FS82, Lemma 7.14].
Lemma 5.7.5. Let   be a xed homogeneous quasinorm on the homogeneous Lie
group G. We denote by Co 1 a constant for the triangle inequality
x, y G
xy Co (x + y).
(5.71)
Denoting by B(x, R) the  ball centred at point x with radius R,
B(x, R) := {y G : x1 y < R},
there exists a maximal family {B(xi , 2C1 o )}
i=1 of disjoint balls of radius
we choose one such family. Then the following properties hold:
1
2Co ,
and
1. The balls {B(xi , 1)}
i=1 cover G.
2. For any C 1, no point of G belongs to more than (4Co2 C)Q of the balls
{B(xi , C)}
i=1 .
3. There exists a sequence of functions i D(G), i N, such that
each i is
supported in B(xi , 2) and satises 0 i 1 while we have i=1 i = 1.
Moreover, for any Nn0 , X i is uniformly bounded in i N.
4. For any p1 > Q + 1, we have
Cp1 > 0
io N
i=1
p1
(1 + x1
Cp1 < .
io xi )
Remark 5.7.6. The conclusion of Part (4) is rough but will be sucient for our
purposes. We note, however, that if the quasinorm in Lemma 5.7.5 is actually a
norm, i.e. if the constant Co in (5.71) is equal to one, Co = 1, then the conclusion
of Part (4) of Lemma 5.7.5 holds true for all p1 > Q. This will be proved together
with the lemma.
388
Chapter 5. Quantization on graded Lie groups
Proof of Lemma 5.7.5 and of Remark 5.7.6. If x G then by maximality there
exists i such that the distance from x to B(xi , 2C1 o ) is < 1/(2Co ). Denoting by y
i , 1 ) which realises the distance, we have
a point in B(x
2Co
x1
i x
Co (x1
i y
+ y
x) < Co
1
1
+
2Co
2Co
= 1.
This proves Part (1).
If x is in all the balls B(xi , C), = 1, . . . , o , then
o
y =1
B(xi , C)
[1, o ]
x1 y Co (x1 xi  + x1
i y) Co 2C.
o
This shows that B(x, 2Co C) contains =1
B(xi , C) and, therefore, it must contain
o
1
the disjoint balls =1 B(xi , 2Co ). Taking the Haar measure and denoting c1 :=
B(0, 1), we have
o
=1
1
B(xi ,
) = o c1
2Co
1
2Co
B(x, 2Co C) = (2Co C) c1 .
This proves Part (2).
Let us x D(G)
satisfying 0 1 with = 1 on B(0, 1) and = 0
on B(0, 2). The sum i =1 (x1
i ) is locally nite by Part (2); it is a smooth
function with values between 1 and (4Co2 2)Q . We dene
(x1 x)
i (x) := i 1 .
i =1 (xi x)
This gives Part (3).
To prove Part (4), we x a point xio and observe that if x G is in one of
1
the balls B(xi , 2C1 o ) with x1
io xi  [, + 1) for some N, let us say B(xi1 , 2Co ),
then
1
1
1
x1
+ + 1).
io x Co (xi1 x + xio xi1 ) Co (
2Co
This yields the inclusion
'x1 xi [,+1) B(xi ,
io
1
1
) B(xio , Co (
+ + 1)).
2Co
2Co
The measure of the left hand side is c1 (2Co )Q card{i : x1
io xi  [, + 1)} and
the measure of the right hand side is c1 (Co ( 2C1 o + + 1))Q . Therefore,
Q
card{i : x1
io xi  [, + 1)} c .
5.7. Calder
onVaillancourt theorem
389
Now we decompose
i=1
p1
(1 + x1
=
io xi )
p1
(1 + x1
+
io xi )
x1
i xio <1
=1
x1
i xio [,+1)
p1
(1 + x1
.
io xi )
2 Q
By Part (2) the rst sum on the right hand side
o ) whereas from the
is (4Cp
observation just above, the second sum is =0 (1 + ) 1 c (1 + )Q . This last
sum being convergent whenever p1 + Q < 1, Part (4) is proved.
Let us nally prove Remark 5.7.6, that is, Part (4) of the lemma for p1 > Q
provided that Co = 1. This will follow by the same argument as above if we can
show a rened estimate
Q1
.
card{i : x1
io xi  [, + 1)} c
We claim that this estimate holds true. Since Co = 1, we can estimate
1
1
x1
io x xio xi1  xi1 x >
1
1
= .
2Co
2
We also have Co ( 2C1 o + + 1) = + 32 . Consequently, we have the inclusion
'x1 xi [,+1) B(xi ,
io
1
3
1
) B(xio , + )\B(xio , ),
2Co
2
2
with the measure on the right hand side being c1 ( + 32 )Q c1 ( 12 )Q . Therefore,
Q1
,
card{i : x1
io xi  [, + 1)} c
so that the required claim is proved.
5.7.2
0
Proof of the case S0,0
This section is devoted to the proof of the following result which is a particular case
of Theorem 5.7.1. We also give an explicit estimate on the number of derivatives
and dierences of the symbol needed for the L2 boundedness.
Proposition 5.7.7. Let T 00,0 . Then T extends to a bounded operator on L2 (G).
Furthermore, if we x a positive Rockland operator R (in order to dene the seminorms on m
, ) then
S(G)
T L2 (G) CT 00,0 ,a,b,c L2 (G) ,
where C > 0 and a, b, c N0 are independent of T . In particular, this estimate
holds with a = rpo , b = r + Q
2 , c = r, where is the degree of R, po /2 is the
smallest positive integer divisible by 1 , . . . , n , and r N0 is the smallest integer
such that rpo > Q + 1.
390
Chapter 5. Quantization on graded Lie groups
Throughout Section 5.7.2, we x the homogeneous norm   =  po given
by (3.21), where po /2 is the smallest positive integer divisible by 1 , . . . , n . We
x a maximal family {B(xi , 2C1 o )}
i=1 of disjoint balls and a sequence of functions
(i )
i=1 so that the properties of Lemma 5.7.5 hold. We also x 0 , 1 D(R)
supported in [1, 1] and [1/2, 2], respectively, such that 0 0 , 1 1 and
j () = 1 with j () := 1 (2(j1) ), j N.
j=0
0
.
Let us start the proof of Proposition 5.7.7. Let S0,0
For each I = (i, j) N N0 , we dene
I (x, ) := i (x)(x, )j ((R)).
We denote by TI and I the corresponding operator and kernel.
Roughly speaking, the parameters i and j correspond to localising in space
and frequency, respectively. The localisation in space corresponds to the covering
of G by the balls centred at the xi s, while the localisation in frequency is determined by the spectral projection of R to the L2 (G)eigenspaces corresponding to
eigenvalues close to each 2j .
It is not dicult to see that each TI is bounded on L2 (G):
Lemma 5.7.8. Each operator TI is bounded on L2 (G).
Since I is localised both in space and in frequency, we may use one of the
two localisations.
Proof of Lemma 5.7.8 using frequency localisation. Let , Nn0 . By the Leibniz
formulae for dierence operators (see Proposition 5.2.10) and for vector elds, we
have
Xx1 i (x) Xx2 1 (x, ) 2 j ((R)).
Xx I (x, ) =
[1 ]+[2 ]=[]
[1 ]+[2 ]=[]
Therefore,
[]+
(I + R) Xx I (x, )(I + R) L (H )
[]+
C
(I + R) Xx2 1 (x, ) 2 j ((R))(I + R) L (H )
[2 ][]
[1 ]+[2 ]=[]
(I + R)
[]+
Xx2 1 (x, )(I + R)
[2 ]+
L (H )
[2 ][]
[1 ]+[2 ]=[]
(I + R)
[2 ]+
2 j ((R))(I + R) L (H ) .
5.7. Calder
onVaillancourt theorem
391
Therefore, by Lemma 5.4.7, we obtain
ja/
0 ,a,b,c S 0 ,a,b,c+a 2
.
I S1,0
0,0
This shows that the operator TI is in 0 and is therefore bounded on L2 (G) by
Theorem 5.4.17.
Proof of Lemma 5.7.8 using space localisation. Another proof is to apply the fol
lowing lemma since the symbol I (x, ) has compact support in x.
Lemma 5.7.9. Let (x, ) be a symbol (in the sense of Denition 5.1.33) supported
in x S, and assume that S is compact. Then the operator norm of the associated
operator on L2 (G) is
Op()L (L2 (G)) CS1/2
sup Xx (x, )L (G)
.
xG
[] Q
2
Proof of Lemma 5.7.9. Let T = Op() and let x be the associated kernel. We
have by the Sobolev inequality in Theorem 4.4.25,
T (x)2
 x (x) sup  xo (x)
xo G
Xx x (x)2 2
.
o
o
L (dxo )
[] Q
2
Hence
T 2L2 (G)
[] Q
2
[] Q
2
CS
 Xxo xo (x)2 dxo dx
Xxo xo 2L2 (dx) dxo
sup
xo G,[] Q
2
Xx x (x)2 2
.
o
o
L (dx)
Now by Plancherels Theorem,
Xx x (x) 2
L2 (dx) Xxo (xo , )L (G)
.
o
o
L (dx)
This implies that the L2 operator norm of T is
CS1/2
sup
xo G,[] Q
2
and concludes the proof of Lemma 5.7.9.
Xxo (xo , )L (G)
,
392
Chapter 5. Quantization on graded Lie groups
Let us go back to the proof of Proposition 5.7.7. The approach is to apply
the following version of Cotlars lemma:
Lemma 5.7.10 (Cotlars lemma here). Suppose that r N0 is such that rpo > Q+1
and that there exists Ar > 0 satisfying for all (I, I ) N N0 :
rpo
.
max TI TI L (L2 (G)) , TI TI L (L2 (G)) Ar 2jj r (1 + x1
i xi )
Then T = Op() is L2 bounded with operator norm C Ar .
Lemma 5.7.10 can be easily shown, adapting for instance the proof given in
[Ste93, ch. VII 2] using Part (4) of Lemma 5.7.5. Indeed, the numbering of the
sequence of operators to which the CotlarStein lemma (see Theorem A.5.2) is
applied is not important, and the condition rpo > Q + 1 is motivated by Lemma
5.7.5, Part (4). This is left to the reader.
Lemma 5.7.11 which follows gives the operator norm for TI TI and TI TI .
Combining Lemmata 5.7.10 and 5.7.11 gives the proof of Proposition 5.7.7.
Lemma 5.7.11.
1. For any r N0 , the operator norm of TI TI on L2 (G) is
rpo
TI TI L (L2 (G)) Cr 1jj 1 (1 + x1
2S 0
i xi )
Q
0,0 ,rpo , 2 ,0
2. For any r N0 , the operator norm of TI TI on L2 (G) is
TI TI L (L2 (G)) Cr 1x1
2jj r 2S 0
xi 4Co
i
Q
0,0 ,0,r+ 2 ,r
In the proof of Lemma 5.7.11, we will also use the symbols i , i N, given
by
i (x, ) := i (x) (x, ),
and the corresponding operators Ti = Op(i ) and kernels i . We observe that i is
compactly supported in x, therefore by Lemma 5.7.9, the operator Ti is bounded
on L2 (G).
Proof of Lemma 5.7.11 Part (1). We have (see the end of Lemma 5.5.4)
TI = Op(I ) = Ti j (R),
thus
TI TI = Ti j (R)j (R)Ti .
Since j (R)j (R) = (j j )(R), this is 0 if j j  > 1. Let us assume j j  1.
We set
Ti j j := Ti (j j )(R) = Op (i (j j ) ((R))) ,
5.7. Calder
onVaillancourt theorem
393
see again the end of Lemma 5.5.4. Therefore TI TI = Ti Ti j j , and we have by the
Sobolev inequality in Theorem 4.4.25,
TI TI (x) = Ti j j (z) ix (z 1 x)dz
G
sup Ti j j (z) ixo (z 1 x)dz 1xB(xi ,2)
xo
G
1
Xx
C
T
(z)
(z
x)dz
1xB(xi ,2) .
ixo
i j j
o
2
G
[] Q
2
L (dxo )
Hence,
TI TI L2
Ti j j (z)Xx ix (z 1 x)dz 1xB(x ,2)
o
i
o
[] Q
2
L2 (dxo dx)
The idea of the proof is to use a quantity which will help the space localisation; so we introduce this quantity 1 + z 1 xrpo and its inverse, where the integer
r N is to be chosen suitably. Notice that for the inverse we have
rpo
,
(1 + z 1 xrpo )1 Cr (1 + z 1 x)rpo Cr (1 + x1
i xi )
for any z suppi and x B(xi , 2). Therefore, we obtain
Ti j j (z)Xx ix (z 1 x)dz 1xB(x ,2)
o
i
o
2
G
L (dxo dx)
1 rpo
x
1
+
z
1
= Ti j j (z)
Xxo ixo (z x)dz 1xB(xi ,2)
2
1
rp
o
1 + z x
G
L (dxo ,dx)
rpo
T
C(1 + x1
x
)
(z
)
i
1 L2 (dz )
ij j
i
1
(1 + z 1 xrpo )Xx ix (z 1 x) 1xB(x ,2) 2
o
2
2
i
o
L (dz ,dx ,dx)
2
by the observation just above and the CauchySchwartz inequality. The last term
can be estimated as
(1 + z 1 xrpo )Xx ixo (z 1 x) 1xB(x ,2) 2
2
2
i
o
L (dz2 ,dxo ,dx)
rpo
B(xi , 2) sup (1 + z  )Xxo ixo (z )L2 (dz )
xo G
rpo
Xx i (xo , )
C sup
o
L (G)
xo G
[]=0
by the Plancherel theorem and Theorem 5.2.22, since z rpo can be written as a
linear combination of q (z), [] = rpo . Combining the estimates above, we have
394
Chapter 5. Quantization on graded Lie groups
obtained
rpo
T i j j L 2
TI TI L2 C(1 + x1
i xi )
sup
Xxo (xo , )
xo G
[ ] Q
2 ,[]rpo
L (G)
The supremum is equal to S 0 ,rpo , Q ,0 . So we now want to study the operator
0,0
2
norm of Ti j j , which is equal to the operator norm of Ti j j . Since the symbol of
Ti j j is localised in space we may apply Lemma 5.7.9 and obtain
Ti j j L (L2 (G)) = Ti j j L (L2 (G)) = Op (i (j j ) ((R))) L (L2 (G))
CB(xi , 2)1/2
C
C
sup
xG
[]Q/2
sup
xG, G
[1 ]+[2 ]=[]
[]Q/2
sup
xG, G
[2 ]Q/2
Xx {i (x)(x, ) (j j ) ((R))} L (G)
X 1 i (x) Xx2 (x, )L (H ) (j j ) ((R)) L (H )
0 ,0,Q/2,0 ,
Xx2 (x, )L (H ) = CS0,0
since the X 2 i s are uniformly bounded on G and over i.
Thus, we have obtained
rpo
0 ,0,Q/2,0 2 0
TI TI L2 C(1 + x1
S0,0
i xi )
L
S
Q
0,0 ,rpo , 2 ,0
,
and this concludes the proof of the rst part of Lemma 5.7.11.
Proof of Lemma 5.7.11 Part (2). Recall that each Ix (y) is supported, with respect to x, in the ball B(xi , 2). We compute easily that the kernel of TI TI is
II (x, w) =
I xz1 (wz 1 )Ixz1 (z)dz.
G
Therefore, II is identically 0 if there is no z such that xz 1 B(xi , 2)B(xi , 2).
So if x1
i xi  > 4Co (which implies B(xi , 2) B(xi , 2) = ) then TI TI = 0. So we
x

4C
.
may assume x1
i
o
i
The idea of the proof is to use a quantity which will help the frequency
localisation; so we introduce this quantity (I + R)r and its inverse, where the
integer r N is to be chosen suitably. We can write
TI TI = TI Ti j (R) = TI Ti (I + R)r (I + R)r j (R).
By the functional calculus (see Corollary 4.1.16),
(I + R)r j (R)L (L2 (G)) = sup(1 + )r j () Cr 2j r .
0
5.7. Calder
onVaillancourt theorem
395
Thus we need to study TI Ti (I + R)r . We see that its kernel is
r i xz1 (wz 1 ) 1 (z)dz
(I + R)
x (w) =
Ixz
G
r i xw1 z (z) 1 (z 1 w)dz.
=
(I + R)
Ixw z
G
We introduce (I + R) (I + R)r on the rst term of the integrand acting on the
variable of i xw1 z , and then integrate by parts to obtain
r i xw1 z (z)
Xz11=z (I + R)r (I + R)
x (w) =
1
G
[1 ]+[2 ]+[3 ]=r
Xz22=z Xz33=z Ixw1 z2 (z31 w)dz
[1 ]+[2 ]+[3 ]=r
r i z (z)
Xz11=xw1 z (I + R)r (I + R)
1
Xz22=xw1 z (X 3 Iz2 ) (z 1 w)dz.
Reinterpreting this in terms of operators, we obtain
Op (X 3 )Xx2 I (x, )
TI Ti (I + R)r =
[1 ]+[2 ]+[3 ]=r
Op (I + R)r Xx1 i (x, )(I + R)r .
By Lemma 5.7.9,
Op (I + R)r Xx1 i (x, )(I + R)r L (L2 (G))
C
sup (I + R)r Xx Xx1 i (x, )(I + R)r L (G)
xG
[] Q
2
S 0
Q
0,0 ,0,[1 ]+ 2 ,r
and
Op (X 3 )Xx2 I (x, ) L (L2 (G))
sup (X 3 )Xx Xx2 i (x, )j ((R))L (G)
[] Q
2
sup (X 3 )(I + R)
[3 ]
[] Q
2
L (G)
(I + R)
[3 ]
Xx+2 i (x, )(I + R)
(I + R)
[3 ]
j ((R))L (G)
C 2j
[3 ]
S 0
Q
0,0 ,0,[2 ]+ 2 ,[3 ]
[3 ]
L (G)
396
Chapter 5. Quantization on graded Lie groups
by Lemma 5.4.7. Hence we have obtained
TI TI L (L2 (G))
Cr 2j
2S 0
Q
0,0 ,0,r+ 2 ,r
[1 ]+[2 ]+[3 ]=r
(jj )r
Cr 2
2S 0
Q
0,0 ,0,r+ 2 ,r
2j
[3 ]
This shows Part 2 of Lemma 5.7.11 up to the fact that we should have j j 
instead of (j j ) but this can be deduced easily by reversing the r
ole of I and I ,
and using T L (L2 (G)) = T L (L2 (G)) .
This concludes the proof of Lemma 5.7.11. Therefore, by Lemma 5.7.10,
Proposition 5.7.7 is also proved.
5.7.3 A bilinear estimate
In this section, we prove a bilinear estimate which will be the major ingredient in
the proof of the L2 boundedness for operators of orders 0 in the case = (0, 1)
in Section 5.7.4.
Note that if f, g S(G) and if N0 then the Leibniz properties together with the properties of the Sobolev spaces (cf. Theorem 4.4.28, especially
the Sobolev embeddings in Part (5)) imply
(I + R) (f g)L2 (G)
X 1 f X 2 gL2 (G)
[1 ]+[2 ]
X 1 f L (G) X 2 gL2 (G)
[1 ]+[2 ]
X 1 f H s (G) X 2 gL2 (G)
[1 ]+[2 ]
f H s+ (G) gH (G) ,
where s > Q/2. As usual, R is a positive Rockland operator of homogeneous degree
; we denote by E its spectral decomposition, see Corollary 4.1.16. Consequently,
if f, g are localised in the spectrum of R in the sense that f = E(Ii )f , g = E(Ij )g,
where Ii , Ij are the dyadic intervals given via
Ij := (2j2 , 2j ),
j N,
and
I0 := [0, 1),
(5.72)
we obtain easily
(I + R) (f g)L2 (G) f L2 (G) gL2 (G) 2(+ ) max(i,j) .
s
Our aim in this section is to prove a similar result but for ) 0:
(5.73)
5.7. Calder
onVaillancourt theorem
397
Proposition 5.7.12. Let R be a positive Rockland operator of homogeneous degree
. As usual, we denote by E its spectral decomposition. There exists a constant
C > 0 such that for any R with + Q/(2) < 0, for any i, j N0 with
i j > 3, we have
f = E(Ii )f
f, g L2 (G)
and
g = E(Ij )g
Q
= (I + R) (f g)L2 (G) Cf L2 gL2 2(+ 2 ) max(i,j) .
The intervals Ii , Ij were dened via (5.72). The proof of Proposition 5.7.12
relies on the following lemma:
Lemma 5.7.13. Let R be a positive Rockland operator. As in Corollary 4.1.16, for
any strongly continuous unitary representation 1 on G, E1 denotes the spectral
decomposition of 1 (R). There exists a gap constant a N such that for any
i, j, k N0 with k < j a and i j 4, we have
, G
E (Ii ) E (Ij ) E (Ik ) = 0.
and
, G
E (Ij ) E (Ik ) E (Ii ) = 0.
Proof of Lemma 5.7.13. We keep the notation of the statement. We also set
H1 ,j := E1 (Ij ),
j N0 ,
for any strongly continuous unitary representation 1 on G. We can write R as a
linear combination
c X ,
R=
[]=
for some complex coecients c . For any strongly continuous unitary representation 1 , we have
c 1 (X) .
1 (R) =
[]=
We consider the strongly continuous unitary representation
Let , G.
1 = . For any X g, its innitesimal representation is given via 1 (X) =
Xx=0 {1 (x)}, see Section 1.7. Consequently, we have for any u H , v H ,
1 (X)(u, v)
Xx=0 1 (x)(u, v)
Xx=0 (x)u (x)v
(X)u v + u (X)v.
In other words,
( )(X) = (X) IH + IH (X).
398
Chapter 5. Quantization on graded Lie groups
We obtain iteratively
( )(X) = (X) IH + IH (X) +
(X)1 (X)2 ,
[1 ]+[2 ]=[]
0<[1 ],[2 ]<[]
denotes a linear combination which depends only on Nn0 and on the
where
This easily implies
structure of G but not on , G.
c ( )(X)
( )(R) =
[]=
= (R) IH + IH (R) +
(X)1 (X)2 ,
[1 ]+[2 ]=
0<[1 ],[2 ]<
where
denotes a linear combination which depends only on R and on the
structure of G but not on , . Hence there exists a constant C > 0 independent
of , such that for any u H , v H , we have
( )(R)(u v)H
(R)uH vH uH (R)vH
C
(X)1 uH (X)2 vH .
[1 ]+[2 ]=
0<[1 ],[2 ]<
If u H,j then from the properties of the functional calculus of (R), we have
(R)uH uH Ij .
Furthermore, the properties of the functional calculus of R and (R) yield
(X)1 uH (X)1 E (Ij )L (H ) uH ,
and, as X 1 R
[1 ]
is bounded on L2 (G) by Theorem 4.4.16, we have
(X)1 E (Ij )L (H )
X 1 E(Ij )L (L2 (G))
X 1 R
2j
[1 ]
[1 ]
L (L2 (G)) R
[1 ]
E(Ij )L (L2 (G))
We have similar inequalities for v H,k . For any unit vectors u H,j and
v H,k with j, k N, we then have
( )(R)(u v)H 2j2 2k C1
[1 ]+[2 ]=
0<[1 ],[2 ]<
j[1 ]+k[2 ]
5.7. Calder
onVaillancourt theorem
399
where the constant C1 depends only on R and on the structure of G. We notice
that
j[1 ]+k[2 ]
[2 ]
2
= 2j
2 (kj) 2j C 2a1 ,
[1 ]+[2 ]=
0<[1 ],[2 ]<
[1 ]+[2 ]=
0<[1 ],[2 ]<
if k j a. Here C is a constant which depends on the structure of G and on
. We choose a N the smallest integer such that
CC 2a1 +2 < 1/2
and
2a+3 < 1/2.
Note that a depends only on the structure of G and on R. When k j a, we
have obtained
( )(R)(u v)H
2j2 2k C2j C 2a1
= 2j2 (1 CC 2a1 +2 ) 2k
> 2j3 2ja > 2j4 .
This implies that u v can not be in H R, for i N0 such that 2i 2j4 . This
shows the rst equality of the statement when i, j, k N. The case of k = 0 or
i = 0 requires to modify slightly some constants above and is left to the reader.
This shows the rst equality of the statement and the second follows by taking
the adjoint. This concludes the proof of Lemma 5.7.13.
Proof of Proposition 5.7.12. We keep the notation of Proposition 5.7.12 and Lemma 5.7.13. We notice that it suces to prove the statement for large enough
max(i, j) and that the roles of i and j are symmetric. Hence we may assume that
i j 4 and that j a where a is the gap constant of Lemma 5.7.13
Let f, g L2 (G) such that f = E(Ii )f and g = E(Ij )g. The inverse formula
for g yields
(I + R) (f g)(x) =
Tr (g)(I + R)x {f (x)(x)} d().
G
We also have (g) = E (Ij )(g). By the CauchySchwartz inequality and the
Plancherel formula, we obtain
(I + R) (f g)(x)2 g2L2 (G)
E (Ij )(I + R)x {f (x)(x)}2HS d().
G
Integrating on both side over x G, we have
(I + R) (f g)2L2 (G) g2L2
E (Ij )(I + R)x {f (x)(x)}2HS dxd().
G
400
Chapter 5. Quantization on graded Lie groups
we x an orthonormal basis of H , so that we can write the
For each G,
HilbertSchmidt norm as the square of the coecients of a (possibly innite dimensional) matrix. The Plancherel formula then yields
E (Ij )(I + R)x {f (x)(x)}2HS dx
G
=
[E (Ij )(I + R)x {f (x)(x)}]kl 2 dx
=
kl
kl
G
F [E (Ij )(I + R) f ]kl ( )2HS(H ) d( ),
where
F [E (Ij )(I + R) f ]kl ( )
=
(I + R)x {f (x)[E (Ij )(x)]kl } (x) dx
G
f (x)[E (Ij )(x)]kl (x) dx
= (I + R)
G
6
7
= E (Ij ) (I + R)
f (x)( )(x)dx
G
kl,
Here the notation []kl, means considering the (kl)coecients in H in the tensor
product over H H . We recognise
f (x)( )(x)dx = ( )(f )
G
thus
kl
F[E (Ij )(I + R) f ]kl ( )2HS(H )
= (E (Ij ) (I + R) ) (( )(f )) 2HS(H H ) .
So far, we have obtained
E (Ij )(I + R)x {f (x)(x)}2HS dxd()
G
G
(E (Ij ) (I + R) ) (( )(f )) 2HS(H H ) d( )d()
=
G
G
= (E (Ij ) (I + R) ) (( )(f )) HS(H H ) 2L2 (d( ),d()) .
We x a dyadic decomposition, that is, we x 0 , 1 D(R) supported in
(1, 1) and (1/2, 2), respectively, valued in [0, 1] and such that
0
k=0
k () = 1 with k () = 1 (2(k1) ) if k N.
5.7. Calder
onVaillancourt theorem
401
The series k k ( (R)) converges to IH in the strong operator topology and we
can apply the following general property:
(B C)AHS(H H )
E (Ik )CL (H ) (B k ( (R)))AHS(H H ) ,
k=0
to B = E (Ij ), C = (I + R) , and
A = ( )(f ).
We keep momentarily this notation for A and C. As E (Ik )CL (H ) 2k , we
have obtained
(E (Ij ) (I + R) ) AHS(H H ) L2 (d( ),d())
2k (E (Ij ) k ( (R))) AHS(H H ) L2 (d( ),d()) .
k=0
Now
A = (( )(f )) = E (Ii ) (( )(f )) ,
thus we can apply Lemma 5.7.13 and the sum over k above is in fact from k j a.
We claim that
Q
(E (Ij ) k ( (R))) AHS(H H ) L2 (d( ),d()) f L2 (G) 2k 2 .
(5.74)
Collecting the equalities and estimates above, (5.74) would then imply
(I + R) (f g)2L2 (G) g2L2 f 2L2 (G)
2k(+ 2 ) ,
k=ja
and would conclude the proof of Proposition 5.7.12.
Hence it just remains to prove (5.74). Natural properties of tensor product
and functional calculus yield
(E (Ij ) k ( (R))) AHS(H H )
E (Ij )L (H ) (IH k ( (R))) AHS(H H )
(IH k ( (R))) AHS(H H ) .
We notice that
(IH k ( (R))) A =
f (x)( k ( (R)) )(x)dx,
402
Chapter 5. Quantization on graded Lie groups
and introducing an orthonormal basis on H ,
[(IH k ( (R))) A],l k =
f (x) [k ( (R))]l k (x)dx
G
= F [f k ( (R))]l k ( ) = F{[f k ( (R))]l k ( 1 )}().
Therefore we have
(IH k ( (R))) AHS(H H ) 2L2 (d( ),d())
=
F [f k ( (R))]l k ( )2HS(H ) d()d( )
G
k l
G
[f k ( (R))] ( 1 )2 2
d( ),
=
lk
L (G)
G
k l
having applied the Plancherel formula in . Simple manipulations yield
2
[f k ( (R))] ( 1 )2 2
=
[f k ( (R))]l k L2 (G)
lk
L (G)
k l
k l
k l
=
f (x) [k ( (R))]l k 2 dx
f (x)2 dx
k l
 [k ( (R))]l k 2
= f 2L2 (G) k ( (R))2HS(H ) .
we can apply the Plancherel formula and obtain
Integrating over G,
[f k ( (R))] ( 1 )2 2
d( ) = f 2L2 (G) k (R)0 2L2 (G) .
lk
L (G)
G
k l
Using the properties of dilations, we have for any k N:
Q k1
k (R)0 L2 (G) = 2 2
1 (R)0 L2 (G) .
Collecting the equalities and inequalities above yields that the lefthand side of
(5.74) is
(E (Ij ) k ( (R))) AHS(H H ) L2 (d( ),d())
Q k1
f L2 (G) 2 2
1 (R)0 L2 (G) .
By Hulanickis theorem, see Corollary 4.5.2, 1 (R)0 L2 (G) is a nite constant.
This shows (5.74) and concludes the proof of Proposition 5.7.12.
5.7. Calder
onVaillancourt theorem
5.7.4
403
0
Proof of the case S,
In this section, we prove the L2 boundedness of operators in 0, with (0, 1):
0
Proposition 5.7.14. Let S,
with (0, 1). Then Op() is bounded on L2 (G)
0
;
and the operator norm is, up to a constant, less than a seminorm of S,
the parameters of the seminorm depend on but not on and could be computed
explicitly.
The rest of this section is devoted to the proof of Proposition 5.7.14. The
strategy is broadly similar to the one in [Ste93, ch VII 2.5] for the Euclidean case.
Technically, this means using analogous rescaling arguments but also replacing
certain integrations by parts on the (Euclidean) Fourier side with the bilinear
estimate obtained in Proposition 5.7.12.
Strategy of the proof
We x a dyadic decomposition, that is, we x 0 , 1 D(R) supported in (1, 1)
and (1/2, 2), respectively, valued in [0, 1] and such that
0
j () = 1 with j () = 1 (2(j1) ) if j N.
j=0
0
. We dene
Let S,
j (x, ) := (x, )j ((R))
and
Tj := Op(j ) = T j (R),
where T = Op().
It is clear that Tj Ti = T (j i )(R)T is zero if j i > 1 and the strategy
of the proof is to apply the crude version of the CotlarStein Lemma, see Proposition A.5.3. We will rst prove that the operator norms of the Tj s are uniformly
0
seminorm, see Lemma 5.7.15. Then we will show that there
bounded in j by a S,
0
seminorm such that
exist a constant C > 0 and a S,
ij>3
Tj Ti L (L2 (G)) C2S,
0 ,a,b,c .
(5.75)
These
two claims together with Proposition A.5.3 and Remark A.5.4 imply that the
(L2 (G))
series j Tj L (L2 (G)) converges in the strong operator topology of L
0 ,a,b,c . As Op() =
and that the operator norm of the sum is S,
j Tj in
the strong operator topology, this will conclude the proof of Proposition 5.7.14.
Step 1
Let us show that the operator norms of the Tj s are uniformly bounded with
respect to j:
404
Chapter 5. Quantization on graded Lie groups
Lemma 5.7.15. The operator Tj = Op(j ) is bounded on L2 (G) with operator norm
0 ,a,b,c with a, b, c as in Proposition 5.7.7.
CS,
The proof of Lemma 5.7.15 uses the following result which is of interest on
its own. In particular, it describes the action of the dilations on G.
Lemma 5.7.16. Let be a symbol with kernel x and operator T = Op(). Let
r > 0. We dene the operator
Tr : S(G) (T (r )) (r1 ).
Then (with operator norm possibly innite)
T L (L2 (G)) = Tr L (L2 (G)) .
Furthermore, the symbol of Tr is
r := Op1 (Tr )
given by
r (x, ) := r1 x, (r) ,
where the representation (r) is dened by
(r) (y) := (ry).
The kernel of r is rQ r1 x (r1 ). Moreover, we have
FG ()( (r) ) = FG rQ (r1 ) (),
FG ()( (r) )
= r[] { FG ()} ( (r) ),
f ( (r) (R))
f (r (R)),
for any Nn0 , any positive Rockland operator R of homogeneous degree , and
any reasonable functions f and (for instance f measurable bounded and in
some Ka,b ).
Proof of Lemma 5.7.16. We keep the notation of the statement. The property
T L (L2 (G)) = Tr L (L2 (G)) follows easily from (r)2 = rQ/2 2 . We compute
(ry) r1 x (y 1 r1 x)dy
(T (r )) (r1 x) =
G
(z) r1 x (r1 z 1 r1 x)rQ dz
=
G
= rQ r1 x (r1 ) (x).
Therefore, the kernel of the operator Tr is rQ r1 x (r1 ). The computation of
its symbol follows from
rQ (r1 x)(x) dx
FG rQ (r1 ) () =
G
(y)(ry) dx = FG ()( (r) ).
=
G
5.7. Calder
onVaillancourt theorem
405
The dierence operator applied to the above expression is
= FG rQ (r1 ) ()
FG ()( (r) )
= FG q () rQ (r1 ) ()
= r[] FG rQ (
q )(r1 ) ()
=
r[] { FG ()} ( (r) ).
The kernels of the operators f (R) and f (r R) are respectively f (R)o and
r f (R)o (r1 ) (see (4.3) in Corollary 4.1.16, and Example 3.1.20 for the homogeneity of o ). Since the group Fourier transform of the former is f ((R)), the
group Fourier transform of the latter is f (r (R)) = f ( (r) (R)).
Q
We can now show Lemma 5.7.15 using the rescaling arguments (together with
the lemma above) and the case = = 0.
Proof of Lemma 5.7.15. Using the Leibniz formula in Proposition 5.2.10, we rst
estimate
(I + R) Xxo o j (x, )(I + R) L (H )
([1 ][o ])
C o
(I + R) Xxo 1 (x, )(I + R)
L (H )
[1 ]+[2 ]=[o ]
qquad
(I + R)
([1 ][o ])
2 j ((R))(I + R) L (H )
[2 ]+([1 ][o ])
2j
0 ,[ ],[ ],
Co S,
o
o
[1 ]+[2 ]=[o ]
j([o ][o ])
0 ,[ ],[ ], 2
Co S,
,
o
o
(5.76)
by Lemma 5.4.7.
For each j N0 , we dene the symbol j given by setting
j
j (x, ) := j 2j x, (2 ) .
By Lemma 5.7.16, the corresponding operator Tj := Op(j ) satises
(Tj )(x) = Tj (2j ) (2j x).
Lemma 5.7.16 and Proposition 5.7.7 imply that
0 ,a,b,c ,
Tj L (L2 (G)) = Tj L (L2 (G)) Cj S0,0
(5.77)
0 ,a,b,c . By
with a, b, c as in Proposition 5.7.7. So we are led to compute j S0,0
Lemma 5.7.16, we have
Xxo o j (x, ) = 2j[o ] 2j[o ] Xxoo=2j x o
o = (2
j )
j (xo , o )
= 2j([o ][o ]) (I + 2j R)
o (I + R) Xxoo=2j x o j (xo , o )o (I + R)
o = (2
j )
(I + 2j R) ,
406
Chapter 5. Quantization on graded Lie groups
so that
(I + R) Xxo o j (x, )(I + R) L (H )
2j([o ][o ]) (I + R) (I + 2j R) L (H )
o (I + R) Xxoo=2j x o j (xo , o )o (I + R)
o = (2
j )
L (H )
(I + 2j R) (I + R) L (H ) .
By the functional calculus (Corollary 4.1.16),
(I + R) (I + 2j R) L (H )
(I + 2 R) (I + R)
L (H )
sup
0
sup
0
1+
1 + 2j
1 + 2j
1+
C2j ,
C2j ,
for any j N0 . Thus, we have obtained
(I + R) Xxo o j (x, )(I + R) L (H )
C2j([o ][o ])
sup
xo G, o G
o (I + R) Xxoo o j (xo , o )o (I + R) L (H )
0 ,[ ],[ ], ,
CS,
o
o
x G, [o ] a, [o ] b and
because of (5.76). Taking the supremum over G,
 c yields
0 ,a,b,c CS 0 ,a,b,c .
j S0,0
,
0 ,a,b,c .
With (5.77), we conclude that Tj L (L2 (G)) CS,
Step 2
Now let us prove Claim (5.75). This relies on the bilinear estimate obtained in
Proposition 5.7.12.
Proof of Claim (5.75). For each i N0 , we denote by i,x the kernel associated
with i . Then one computes easily the integral kernel Kji (x, y) of the operator
Tj Ti , that is,
(Tj Ti )f (x) =
with
Kji (x, y)f (y)dy,
Kji (x, y) =
f S(G),
j,z (x1 z)i,z (y 1 z)dz.
5.7. Calder
onVaillancourt theorem
407
By Schurs lemma [Ste93, 2.4.1], we have
Kji (x, y)dy, sup
Kji (x, y)dx ,
Tj Ti L (L2 (G)) max sup
xG G
Tj Ti 2, ,a,b,c
yG
+ max Kji (x, y),
y 1 x1
since the estimates at innity for the kernels of a pseudodierential operator
obtained in Theorem 5.4.1 for = 0 yield
Kji (x, y) Tj Ti 2, ,a1 ,b1 ,c1 y 1 xN
for any N N0 . (We have assumed that a quasinorm   has been xed on
G.) The properties of composition and of taking the adjoint of pseudodierential
operators (see Theorems 5.5.3 and 5.5.12) together with Lemma 5.4.7 yield
2
1 ,a ,b ,c i S 1 ,a ,b ,c 0
Tj Ti 2, ,a1 ,b1 ,c1 j S,
S, ,a4 ,b4 ,c4 2
2 2 2
, 3 3 3
i+j
We now analyse maxy1 x1 Kji (x, y). So let x, y G with y 1 x 1. We
x a function D(G) which is a smooth version of the indicatrix function of the
ball B(0, 10) = {z G : x1 z < 10} about 0 with radius 10, that is, we assume
that 1 on B(0, 10) and 0 on B(0, 11). Let us assume that the quasinorm is in fact a norm, that is, it satises the triangle inequality with constant
1 (although we could give a proof without this restriction, it simplies the choice
of constants and therefore avoids dwelling on unimportant technical points). We
can always decompose
j,z (x1 z)i,z (y 1 z) (x1 z) + (1 (x1 z) dz
Kji (x, y) =
zG
I1 + I2 .
We rst estimate the second integral via
1 ,a ,b ,c i S 1 ,a ,b ,c
I2  j S,
5 5 5
, 6 6 6
x1 z>10
x1 zN1 y 1 zN1 dz.
having used the estimates at innity for the kernels of a pseudodierential operator obtained in Theorem 5.4.1 for = 0. As y 1 x 1, the last integral is
just a nite constant if we choose N1 = Q + 1 for instance. We estimate the
1
seminorms with Lemma 5.4.7 and we obtain then
S,
I2  2S,
0 ,a ,b ,c 2
7 7 7
i+j
We now estimate the integral I1 :
j,z (x1 z)i,z (y 1 z)(x1 z)dz.
I1 =
G
408
Chapter 5. Quantization on graded Lie groups
It is of the form G f (z, z)dz for a given function f on G G. Simple formal
manipulations yield for any N N0
N
f (z, z)dz =
(I + R)N
z2 =z (I + R)z2 f (z, z2 )dz
G
G
N
N
(I + R)
=
z1 =z (I + R)z2 =z f (z1 , z2 )dz,
G
since R is essentially
having used integration by parts or equivalently Rt = R,
selfadjoint. Hence, we obtain formally in our case
N
N
I1 =
j,z1 (x1 z2 )i,z1 (y 1 z2 )(x1 z1 ) dz,
(I + R)
z1 =z (I + R)z2 =z
G
where N N0 is to be xed later. Note that the expression in z1 is supported in
B(x1 , 11), hence so is the integrand in z. This produces the following estimate
I1 
S(z2 )dz2
x1 z2 11
where S(z2 ) is the supremum
N (I + R)N
S(z2 ) = sup (I + R)
j,z1 (x1 z2 )i,z1 (y 1 z2 )(x1 z1 )
z1
z2
z1 G
s
+ 0
N
(I + R)
(I + R)N
j,z1 (x1 z2 )i,z1 (y 1 z2 )(x1 z1 ) 2
z1
z2
L (dz1 )
N
01
1
02
1
(I + R)z2 {Xz1
j,z1 (x z2 ) Xz1 i,z1 (y z2 )} L2 (B(x,11),dz1 ) ,
[01 ]+[02 ]
N +s0
by the properties of the Sobolev spaces, see Theorem 4.4.28, especially the Sobolev
embedding in Part (5). Here s0 N denotes the smallest integer multiple of
such that s0 > Q/2. By the CauchySchwartz inequality, as B(x, 11) has nite
volume independent of x, we obtain
01
(I + R)N
j,z (x1 z2 ) Xz02 i,z (y 1 z2 )} 2
I1 
2
z {Xz
2
[01 ]+[02 ]
N +s0
sup
L (B(x,11) ,dz1 dz2 )
01
(I + R)N
j,z1 (x1 z2 ) Xz102 i,z1 (y 1 z2 )}L2 (dz ) .
z2 {Xz1
z1 B(x,11)
[01 ]+[02 ]N +s0
Q
Choosing N > 2
, we can apply Proposition 5.7.12 to the L2 norm above, so that
01
(I + R)N
j,z1 (x1 z2 ) Xz102 i,z1 (y 1 z2 )}L2 (dz2 )
z2 {Xz1
Q
Xz101
j,z1 (z2 )L2 (dz2 ) Xz102 i,z1 (z2 )L2 (dz2 ) 2(N + 2 ) max(i,j) .
5.8. Parametrices, ellipticity and hypoellipticity
By Corollary 5.4.3, we have
Xz 01
j,z1 (z2 )L2 (dz
1
2)
409
Xx01 j S,
m ,a ,b ,c ,
7 7 7
where m is a number such that m < Q/2, for instance m := 1 Q/2. By
Lemma 5.4.7, we have (with = )
m [01 ]
j
Xx01 j S,
m ,a ,b ,c S 0 ,a8 ,b8 ,c8 2
,
7 7 7
We have similar estimates for Xz102 i,z1 (z2 )L2 (dz ) , thus
max
[01 ]+[02 ]
N +s0
Xz 01
j,z1 (z2 )L2 (dz ) Xz102 i,z1 (z2 )L2 (dz
1
2S,
0 ,a ,b ,c
9 9 9
max
2j
m [01 ]
2i
2)
m [02 ]
[01 ]+[02 ]
N +s0
max(i,j)(2m +(N +s0 ))
2S,
.
0 ,a ,b ,c 2
9 9 9
The estimates above show that the rst formal manipulations on I1 are justied
and we obtain
max(i,j)((1)N 2m +s0 + 2 )
I1  2S,
.
0 ,a ,b ,c 2
9 9 9
Consequently, we have
y
thus
i+j
Q
max(i,j)((1)N 2m +s0 + 2
)
max Kji (x, y) 2S,
+
2
2
,
0 ,a,b,c
1
x1
i+j
Q
2 + 2max(i,j)((1)N 2m +s0 + 2 ) .
Tj Ti L (L2 (G)) 2S,
0 ,a,b,c
Q
< 1.
As = (0, 1), we can choose N such that (1 )N 2m + s0 + 2
Summing over i > j + 3 and using the symmetry of the role played by i and j
yield (5.75).
Hence we have shown Proposition 5.7.14 and this concludes the proof of
Theorem 5.7.1.
5.8
Parametrices, ellipticity and hypoellipticity
In this section, we obtain statements regarding ellipticity and hypoellipticity which
are similar to the compact case presented in Section 2.2.3 where the Laplacian has
the role of the positive Rockland operator. However, on nilpotent Lie groups, since
is not discrete and the representations are often not (and can be almost never)
G
nite dimensional, the precise hypotheses become more technical to present.
410
5.8.1
Chapter 5. Quantization on graded Lie groups
Ellipticity
Roughly speaking, we dene the ellipticity by requiring that the symbol is invertible for high frequencies. These high frequencies are determined with respect to
the spectral projection E of a positive Rockland operator R, and its group Fourier
transform E , see Corollary 4.1.16.
We will use the following shorthand notation:
H,
:= E (, +)H .
(5.78)
Since E (, ) = FG (1(,) (R)0 ) yields a symbol acting on smooth vectors (see
Examples 5.1.27 and 5.1.38), H,
is a subspace of H .
We can now dene our notion of ellipticity:
Denition 5.8.1. Let R be a positive Rockland operator of homogeneous degree
. Let be a symbol given by elds of operators acting on smooth vectors, i.e.
is in some L (G)
for each x G.
(x, ) = {(x, ) : H H , G}
a,b
The symbol is said to be elliptic with respect to R of elliptic order mo
and any
if there is R such that for any R, x G, almost all G,
u H, we have
R
(I + R) (x, )uH C (I + R) (I + R)
mo
uH .
(5.79)
and u H .
with C = C,R,mo ,, independent of (x, ) G G
,
We will say that the symbol or the corresponding operator Op() is
(R, , mo )elliptic, or elliptic of elliptic order mo , or just elliptic.
The notation H,
was dened in (5.78). As H,
is a subspace of H and
since (I + R) and (x, ) are elds of operators acting on smooth vectors, the
expression in the norm of the lefthand side of (5.79) makes sense.
In our elliptic condition in Denition 5.8.1, is a symbol in the sense of
Denition 5.1.33 which is given by elds of operators acting on smooth vectors. It
m
to construct parametrices,
will be natural to consider symbols in the classes S,
see Proposition 5.8.5 and Theorem 5.8.7.
Our denition of ellipticity requires a property of xuniform partial injectivmo
+mo
ity. Of course, we note that (I + R) (I + R) = (I + R) .
Naturally, we will see shortly in Corollary 5.8.4 that it suces to check (5.79)
for a sequence of real numbers { , Z} which tends to as .
Our rst examples of elliptic operators are provided by positive Rockland
operators:
Proposition 5.8.2. Let R be a positive Rockland operator of homogeneous degree
. Then we have the following properties.
1. The operator (I + R)
elliptic order mo .
mo
, for any mo R, is elliptic with respect to R of
5.8. Parametrices, ellipticity and hypoellipticity
411
2. If f1 and f2 are complexvalued (smooth) functions on G such that
inf
xG,
f1 (x) + f2 (x)
>0
1+
for some 0,
then the dierential operator f1 (x) + f2 (x)R is (R, , )elliptic.
3. The operator E(, )R, for any > 0, is (R, , )elliptic.
More generally, if f is a complexvalued function on G such that inf G f 
> 0, then f (x)E(, )R is (R, , )elliptic.
4. Let C (R) be such that
(,1 ] = 0
and
[2 ,) = 1,
for some real numbers 1 , 2 satisfying 0 < 1 < 2 , Then the operator
(R)R is (R, 2 , )elliptic.
More generally, if f is a complexvalued function on G such that inf G f 
> 0, then f (x)(R)R is (R, 2 , )elliptic.
Proof. The symbols involved in the statement are multipliers in R. By Example
5.1.27 and Corollary 5.1.30, the corresponding symbols are symbols in the sense of
Denition 5.1.33 which are given by elds of operators acting on smooth vectors.
Hence it remains just to check the condition in (5.79).
Part (1) is easy to check using the functional calculus of (R).
Let us prove Part (2). Let , f1 , f2 , and m be as in the statement. The
properties of the functional calculus for (R) yield that, for each x G xed and
we have
u H,
(I + R) (I + R)u = x ((R))(I + R) (f1 (x) + f2 (x)(R))u,
where x L [0, ) is given by
x () =
1+
1 .
f1 (x) + f2 (x)
Our assumption implies that x is bounded on [0, ) with
C := sup x =
xG
f1 (x) + f2 (x)
xG,
1+
inf
The property of the functional calculus for (R) yields
x G
x ((R))L (H ) C.
1
< .
412
Chapter 5. Quantization on graded Lie groups
Thus we have
(I + R) (I + R)uH = x ((R))(I + R) (f1 (x) + f2 (x)(R))uH
C(I + R) (f1 (x) + f2 (x)(R))uH .
This proves Part (2).
Let us prove Part (3). The properties of the functional calculus for (R) yield
(I + R)u = ((R))E (, )(R)u,
where L [0, ) is given by
() =
1+
1(,) ().
Moreover,
(I + R)1+ uH
((R))(I + R) E (, )(R)uH
(I + R) E (, )(R)uH .
Since C = 1
is a nite positive constant, we have obtained
C(I + R)1+ uH (I + R) E (, )(R)uH .
This shows that E(, )R, is elliptic.
If f is as in the statement, we proceed as above, replacing by
x () =
1+
1(,) (),
f (x)
and C such that C 1 is equal to the righthand side of the estimate
x
1+
1
sup
:= C 1 .
inf G f 
This shows Part (3).
2 .
For Part (4), we proceed as in Part (3) replacing 1(,) by () and by
The next lemma is technical. It states that we can construct a partial inverse
of an elliptic symbol. The analogue for scalarvalued symbols would be obvious: if
a(x, ) does not vanish for  > then we can consider 1> 1/a(x, ). However,
in the context of operatorvalued symbols, we need to proceed with caution.
5.8. Parametrices, ellipticity and hypoellipticity
413
Lemma 5.8.3. Let be a symbol (R, , mo )elliptic as in Denition 5.8.1.
such that (x, )u = v then
For any v H , if there is a vector u H,
and we can
this u is necessarily unique. In this sense (x, ) is invertible on H,
set
u if v = (x, )u, u H,
,
E (, )(x, )1 (v) :=
(5.80)
.
0 if H v (x, )H,
This yields the symbol (in the sense of Denition 5.1.33) given by elds of operators
acting on smooth vectors
{E (, )(x, )1 : H H , (x, ) G G}.
(5.81)
Furthermore, for every ,
E (, )(x, )1 L
,+mo (G)
C1 ,
(5.82)
where C is the constant appearing in (5.79) of Denition 5.8.1.
If is continuous in the sense of Denition 5.1.34, then the symbol in (5.81)
is continuous in the sense of Denition 5.1.34. If is smooth, then the symbol
in (5.81) is continuous and depends smoothly on x G in the sense of Remark
1.8.16.
Proof. Recall that E (, ) = FG (1(,) (R)0 ) yields a symbol acting on smooth
vectors, see Examples 5.1.27 and 5.1.38.
, then, using (5.79), we have
If v = (x, )u where u H,
(I + R)
mo +
uH C1 (I + R) (x, )uH = C1 (I + R) vH .
is a symbol in the
It is now easy to check {E (, )(x, )1 , (x, ) G G}
sense of Denition 5.1.33 and that the estimates in (5.82) hold.
If is continuous, then one checks easily that the map
G x E (, )(x, )1 L
,+mo (G)
is continuous.
is continuous. Consequently {E (, )(x, )1 , (x, ) G G}
1
If is smooth, then {E (, )(x, ) , (x, ) G G} depends smoothly
in x G, see Remark 1.8.16.
Corollary 5.8.4. Let R be a positive Rockland operator of homogeneous degree .
The symbol satises (5.79) for each R if and only if satises (5.79) for a
sequence of real numbers { , Z} which tends to as .
We may choose the constants C such that maxc C in (5.79) is nite for
any c 0.
414
Chapter 5. Quantization on graded Lie groups
Proof. From the proof of Lemma 5.8.3, we see that satises (5.79) for if and
only if
sup E (, )(x, )1 L
<
(G)
xG
,+mo
is nite. The conclusion follows from Corollary 4.4.10.
m
The next statement says that if a symbol in some S,
is elliptic and if the
elliptic order is equal to the order m of the symbol, then we can dene a symbol
m
using the operator E (, )(x, )1 dened via (5.80). This will be the
in S,
main ingredient in the construction of a parametrix, see the proof of Theorem
5.8.7.
m
be a symbol which is
Proposition 5.8.5. Assume 1 0. Let S,
(R, , m)elliptic with respect to a positive Rockland operator R. If C (R) is
such that
(,1 ] = 0 and [2 ,) = 1,
for some real numbers 1 , 2 satisfying < 1 < 2 , then the symbol
{((R)) 1 (x, ) , (x, ) G G},
given by
((R)) 1 (x, ) := ((R))E (1 , )(x, )1 ,
m
is in S,
. Moreover, for any ao , bo N0 , we have
((R)) 1 (x, )S m ,ao ,bo ,0
,
a1 +b1 +1
a +b
C
max
C,,
(x, )S2m ,a2o ,bo ,m ,
1
a1 ,a2 ao
b1 ,b2 bo
ao +bo
where C > 0 is a positive constant depending on ao , bo , , and where the constant
C,,1 was given in (5.79).
The following lemma is helpful in the proof of Proposition 5.8.5. Indeed, in
the case of Rn , if a cuto function () on the Fourier side is constant for  >
( large enough), then its derivatives are () = 0 if  > . In our case, we can
not say anything in general. If we use ((R)) as a cuto in frequency with as
in Proposition 5.8.5 for example, it is not true in general that its ( )derivatives
will vanish on E (, ) or will be of the form 1 ((R)). However, we can show
that these derivatives are smoothing:
Lemma 5.8.6. Let C (R) satisfy [,+) = 1 for some R. Then for any
Nn0 \{0}, the symbol given by ((R)) is smoothing, i.e. is in S .
5.8. Parametrices, ellipticity and hypoellipticity
415
Proof of Lemma 5.8.6. Let Nn0 \{0}. Then I = 0 by Example 5.2.8. Therefore
((R)) = (1 )((R)).
As 1 is a smooth function such that supp(1)[0, ) is compact, the symbol
(1 )((R)) is smoothing. Hence so is (1 )((R)) and ((R)).
Proof of Proposition 5.8.5. Recall that by the Leibniz formula (Proposition 5.2.10),
we have
c1 ,2 1 1 2 2 ,
o (1 2 ) =
[1 ]+[2 ]=[o ]
with
c1 ,0 =
1 if 1 = o
,
0 otherwise
c0,2 =
1 if 2 = o
.
0 otherwise
It is also easy to see that
X o (f1 f2 ) =
c1 ,2 X 1 f1 X 2 f2 ,
[1 ]+[2 ]=[o ]
with
c1 ,0 =
1 if 1 = o
,
0 otherwise
c0,2 =
1 if 2 = o
.
0 otherwise
m
and C (R) as in the statement. By Lemma 5.8.3,
Let = (x, ) S,
the continuous symbol
{E (, )(x, )1 : H H , (x, ) G G},
depends smoothly on x G. Hence so does the continuous symbol o dened via
o (x, ) := ((R)) 1 (x, ).
Since ((R)) commutes with powers of (I + R) and
((R))L (H ) ,
we have
m
(I + R) o (x, )L (H )
m
(I + R) E (, )(x, )1 L (H )
= C01 ,
where by Lemma 5.8.3, C0 is the nite constant intervening in the ellipticity
condition for = 0 in (5.79). More generally, in this proof, C denotes the constant
depending on in (5.79), see also Corollary 5.8.4.
416
Chapter 5. Quantization on graded Lie groups
By Proposition 5.3.4, ((R)) S 0 . We also see that
((R)) = o (x, )(x, ).
(5.83)
Hence for any leftinvariant vector eld X we have
0
Xx ((R))
Xx o (x, ) (x, ) + o (x, ) Xx (x, ).
Thus
Xx o (x, )(x, ) = o (x, ) Xx (x, ),
and since (x, ) is invertible on E (1 , )H ,
Xx o (x, ) = o (x, ) {Xx (x, )} E(1 , ) 1 (x, ).
Assuming that X is homogeneous of degree d, we can take the operator norm and
estimate
(I + R)
md
Xx o (x, )L (H )
(I + R)
md
o (x, )(I + R) L (H )
(I + R) Xx (x, )(I + R) L (H )
m
(I + R) E (1 , )(x, )1 L (H )
d
1
m ,0,d,m .
Cd
C01 (x, )S,
Recursively on d = [o ], we can show similar properties for Xxo ((R))(x, )1 ,
and obtain
((R))(x, )1 S m ,0,bo ,0
,
(b +1)
b
Cbo ,
max C 1 (x, )S2m
b1 ,b2 bo
bo
, ,0,bo ,m
We can proceed in a parallel way for dierence operators. Indeed, for any
o Nn0 with o  = 1, we apply o to both sides of (5.83) and obtain
o {((R))} = o o (x, ) (x, ) + o (x, ) o {(x, )},
thus
o o (x, )
o {((R))}E(1 , ) 1 (x, )
o (x, ) {o (x, )} E(1 , ) 1 (x, ).
Then
(I + R)
[o ]+m
o o (x, )L (H ) N1 + N2 ,
5.8. Parametrices, ellipticity and hypoellipticity
417
with
N1
N2
=
=
(I + R)
[o ]+m
o {((R))}E(1 , ) 1 (x, )L (H ) ,
(I + R)
[o ]+m
o (x, ) {o (x, )} E(1 , ) 1 (x, )L (H ) .
For the rst norm, we see that
N1
(I + R)
[o ]+m
o {((R))}(I + R) L (H )
m
(I + R) E(1 , ) 1 (x, )L (H )
m
C C01 ,
since o {((R))} S by Lemma 5.8.6. For the second norm, we see that
N2
(I + R)
[o ]+m
o (x, )(I + R)
(I + R)
[o ]
(I + R)
[o ]
(x, )(I + R)
L (H )
L (H )
E(1 , ) 1 (x, )L (H )
1
m ,[ ],0,m .
C[
C01 S,
o
o]
Recursively on [o ], we can show similar properties for o ((R))(x, )1 ,
and obtain
o (x, )S m ,ao ,0,0
,
Cao ,
a1 ,a2 ao
(a1 +1)
max C
ao
a
(x, )S2m
, ,ao ,0,m
More generally, we have
Xxo o {((R))} =
c1 ,2 c1 ,2 Xx1 1 o (x, )
[1 ]+[2 ]=[o ]
[1 ]+[2 ]=[o ]
Xx2 2 (x, ).
Because of the very rst remark of this proof, we obtain X o o o in terms of
X o with [ ] < [o ] and [ ] < [o ] and of some derivatives of ((R)) and
. If we assume that we can control all the seminorms o S m ,a,b,c with a < [o ],
,
b < [o ] and any c R, then we can proceed as above introducing powers of I + R
to obtain the estimate for the seminorms of ((R))(x, )1 . Recursively this
shows Proposition 5.8.5.
5.8.2 Parametrix
In the next theorem, we show that our notion of ellipticity implies the construction
of a parametrix.
418
Chapter 5. Quantization on graded Lie groups
m
Theorem 5.8.7. Let S,
be elliptic of elliptic order m with 1 > 0.
We can construct a left parametrix B m
, for the operator A = Op(), that is,
m
there exists B , such that
BA I .
Comparing with twosided parametrices in the case of compact Lie groups
(Theorem 2.2.17), this parametrix is onesided. It was also the case in [CGGP92].
Proof. We can adapt the proof in [Tay81, 0.4] to our setting. Let C (R) be
such that (,1 ] = 0 and [2 ,) = 1 for some 1 , 2 R with < 1 < 2 .
By Proposition 5.8.5,
m
.
((R)) 1 (x, ) S,
Since ((R)) = ((R)) 1 (x, )(x, ), by Corollary 5.5.8,
()
Op ((R)) 1 (x, ) A = (R) mod,
;
now (R) = I (1 )(R) and (1 ) D([0, )) so (1 )(R) . This
shows
()
.
Op ((R)) 1 (x, ) A = I mod,
So we have
Op ((R)) 1 (x, ) A
=IU
with
()
U ,
By Theorem 5.5.1, there exists T 0, such that
T I + U + U2 + . . . + Uj + . . .
By Theorem 5.5.3,
B := T Op ((R)) 1 m
, .
Therefore, we obtain
BA = T (I U ) = I mod ,
completing the proof.
It is not dicult to construct the following examples of elliptic operators
satisfying Theorem 5.8.7 out of any Rockland operator. Indeed, combining Proposition 5.3.4 or Corollary 5.3.8 together with Proposition 5.8.2 yield
Example 5.8.8. Let R be a positive Rockland operator of homogeneous degree .
1. For any m R, the operator (I + R) m is elliptic with respect to R of
elliptic order m.
m
5.8. Parametrices, ellipticity and hypoellipticity
419
2. If f1 and f2 are complexvalued smooth functions on G such that
inf
xG,
f1 (x) + f2 (x)
> 0 for some 0,
1+
and such that X 1 f1 , X 2 f2 are bounded for each 1 , 2 Nn0 , then the
dierential operator
f1 (x) + f2 (x)R
is (R, , )elliptic.
3. Let C (R) be such that
(,1 ] = 0
and
[2 ,) = 1,
for some real numbers 1 , 2 satisfying 0 < 1 < 2 , Then the operator
(R)R is (R, 2 , )elliptic.
More generally, if f is a smooth complexvalued function on G such that
inf G f  > 0 and that X f is bounded on G for every Nn0 , then
f (x)(R)R
is elliptic with respect to R of elliptic order .
Hence all the operators in Example 5.8.8 admit a left parametrix.
We will see other concrete examples of elliptic dierential operators on the
Heisenberg group in Section 6.6.1, see Example 6.6.2.
In fact we can prove the existence of left parametrices for symbols which are
elliptic with an elliptic order lower than their order. Indeed, we can modify the
hypothesis of the ellipticity in Section 5.8.1 to obtain the analogue of H
ormanders
theorem about hypoellipticity involving lower order terms, similar to Theorem
2.2.18 in the compact case.
m
with 1 > 0. We assume that is elliptic with
Theorem 5.8.9. Let S,
respect to a positive Rockland operator R in the sense of Denition 5.8.1, and that
its elliptic order is mo m.
We also assume that the following hypothesis on the lower order terms holds:
and any
there is R such that for any R, x G, almost all G,
u H, , we have
(I + R)
[][]+
X (x, ) (I + R) uH
C,,
(x, )uH ,
(5.84)
and u H .
independent of (x, ) G G
with C,,
= C,,,,R,m
,
o ,,
o
for the operator A =
Then we can construct a left parametrix B m
,
mo
Op(), that is, there exists B , such that
BA I .
420
Chapter 5. Quantization on graded Lie groups
Proceeding as in Corollary 5.8.4, we can show easily that it suces to assume
(5.79) and (5.84) for a countable sequence which goes to + and .
Proof. Let C (R) be such that (,1 ] = 0 and [2 ,) = 1 for some
1 , 2 R with < 1 < 2 . Proceeding as in the proof of Proposition 5.8.5, we
see that
mo
o (x, ) := ((R)) 1 (x, ) S,
,
with similar estimates for the seminorms of o and .
With similar ideas, using (5.84), we claim that, for any multiindex o Nn0 ,
we have
[ ]
X o (x, ) o (x, ) S, o .
Indeed, from the proof of Proposition 5.8.5, we know that
Xo = o X E(, ) 1 ,
hence
X X o (x, ) o (x, ) = XX o (x, ) o (x, ) + X o (x, ) Xo (x, )
= XX o (x, ) o (x, ) X o (x, ) o X E(, ) 1 ,
m
seminorms
and we can use the hypothesis (5.84) on each term to control the S,
of the expression on the righthand side. For the dierence operators, from the
proof of Proposition 5.8.5, we know with o  = 1, that
o o = o ((R)) E(, ) 1 o o E(, ) 1 .
Hence
o X o (x, ) o (x, )
= X o o (x, ) o (x, ) + X o (x, ) o o (x, )
= X o o (x, ) o (x, ) X o (x, ) o o E(, ) 1
+X o (x, ) o ((R)) o ((R)) 1 ,
where o C (R) is a xed smooth function such that o [1 ,) = 1 and
o (,1 /2) = 0. While we can use the hypothesis (5.84) on the rst two terms,
we use Lemma 5.8.6 for the last term which is then smoothing. Proceeding recursively as in the proof of Proposition 5.8.5, we obtain the estimates for the sum on
the righthand side.
We now dene recursively
n (x, ) :=
0<[]n
n[] X o ,
n = 1, 2, . . .
5.8. Parametrices, ellipticity and hypoellipticity
421
mo n()
It is easy to check that each symbol n (x, ) is in S,
compact case,
and that as in the
0 n
.
Op(o )Op() I Op(1 )Op() . . . Op(n )Op() mm
,
o
whose symbol is given by the asymptotic sum
Therefore, the operator B m
,
o j=1 j is a left parametrix for A = Op().
We will see a concrete example of hypoelliptic dierential operators on the
Heisenberg group in Section 6.6.2, see Example 6.6.4.
We now note the following generalisation of Proposition 5.8.5 that we have
already used in the proof of Theorem 5.8.9.
m
Proposition 5.8.10. Assume 1 0. Let S,
be a symbol which is
(R, , mo )elliptic with respect to a positive Rockland operator R. If C (R)
is such that
(,1 ] = 0 and [2 ,) = 1,
for some real numbers 1 , 2 satisfying < 1 < 2 , then the symbol
{((R)) 1 (x, ) , (x, ) G G},
given by
((R))(x, )1 := ((R))E (1 , ) 1 (x, ),
mo
is in S,
. Moreover, for any ao , bo N0 , we have
((R)) 1 (x, )S mo ,ao ,bo ,0
,
a1 +b1 +1
a +b
C
max
C,,
(x, )S2m ,a2o ,bo ,m ,
1
a1 ,a2 ao
b1 ,b2 bo
ao +bo
where C > 0 is a positive constant depending on ao , bo , , and where the constant
C,,1 was given in (5.79).
Here the elliptic order mo and the symbol order m are dierent but the same
mo
. The proof is
results holds: one can construct a symbol ((R)) 1 (x, ) S,
easily obtained by generalising the proof of Proposition 5.8.5.
We now show that Theorem 5.8.7 has a partial inverse.
Proposition 5.8.11. Suppose that the operator A = Op() m
, , with 1 >
,
i.e.
BA
.
Then is elliptic
0, admits a left parametrix B m
,
of order m, that is, there exist a positive Rockland operator R of homogeneous
and any
degree , and R such that for any R, x G, almost all G,
we have
u H,
(I + R) (x, )uH C (I + R) (I + R) uH .
422
Chapter 5. Quantizati