Ergebnisse der Mathematik und ihrer Grenzgebiete.

3. Folge / A Series of Modern Surveys in Mathematics 76

Tuomas Hytönen
Jan van Neerven
Mark Veraar
Lutz Weis

Analysis in
Banach Spaces
Volume III: Harmonic Analysis and
Spectral Theory
Ergebnisse der Mathematik und ihrer
Grenzgebiete. 3. Folge / A Series of Modern
Surveys in Mathematics

Volume 76

Series Editors
L. Ambrosio, Pisa, Italy
V. Baladi, Paris, France
G. -M. Greuel, Kaiserslautern, Germany
M. Gromov, Bures-sur-Yvette, France
G. Huisken, Tübingen, Germany
J. Jost, Leipzig, Germany
J. Kollár, Princeton, USA
G. Laumon, Orsay, France
U. Tillmann, Oxford, UK
D. B. Zagier, Bonn, Germany
Ergebnisse der Mathematik und ihrer Grenzgebiete, now in its third sequence, aims
to provide reports, at a high level, on important topics of mathematical research. Each
book is designed as a reliable reference covering a significant area of advanced
mathematics, spelling out related open questions, and incorporating a comprehensive,
up-to-date bibliography.
Tuomas Hytönen • Jan van Neerven
Mark Veraar • Lutz Weis

Analysis in Banach Spaces

Volume III: Harmonic Analysis
and Spectral Theory
Tuomas Hytönen Jan van Neerven
Department of Mathematics and Statistics Delft Institute of Applied Mathematics
University of Helsinki Delft University of Technology
Helsinki, Finland Delft, The Netherlands

Mark Veraar Lutz Weis

Delft Institute of Applied Mathematics Department of Mathematics
Delft University of Technology Karlsruhe Institut of Technology
Delft, The Netherlands Karlsruhe, Germany

ISSN 0071-1136 ISSN 2197-5655 (electronic)

Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge / A Series of Modern Surveys in
Mathematics
ISBN 978-3-031-46597-0 ISBN 978-3-031-46598-7 (eBook)
https://doi.org/10.1007/978-3-031-46598-7

Mathematics Subject Classification (2020): 46Bxx, 35Kxx, 42Bxx

© Springer Nature Switzerland AG 2023

This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher,
whether the whole or part of the material is concerned, specifically the rights of translation, reprinting,
reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way,
and transmission or information storage and retrieval, electronic adaptation, computer software, or by
similar or dissimilar methodology now known or hereafter developed.
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, expressed or implied, with respect to the material contained herein
or for any errors or omissions that may have been made. The publisher remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Paper in this product is recyclable.

Preface

Originally conceived to be the final volume of a trilogy on Analysis in Banach

spaces, containing the applications of the infrastructure of Volumes I–II to
Harmonic and Stochastic Analysis, it was eventually evident that the body of
work that we wanted to discuss would never fit into one volume of compara-
ble size. Thus it was decided to divide the topics over two volumes, of which
the present one, subtitled Harmonic Analysis and Spectral Theory, will of-
fer a systematic treatment of Banach space-valued singular integrals, Fourier
transform, and function spaces; further develop and ramify the theory of func-
tional calculus from Volume II; and culminate in applications of these notions
and tools in the problem of maximal regularity of evolution equations. The
subsequent Volume IV will then be dedicated to the stochastic counterparts
of some of these topics.
Like the previous Volume II, the present Volume III has its time-wise
centre of gravity firmly in the present century. At the same time, we always
cover the necessary prerequisites from earlier developments, presenting a self-
contained picture rather than just a modern uppermost layer. As one might
expect, Banach spaces with the unconditional martingale differences (UMD)
property, will again manifest themselves as the most useful class of spaces
for our analysis, but many of the other Banach space properties discussed in
Volumes I–II will also feature prominently.
Our discussion of singular integrals is thoroughly influenced by the recent
notion of sparse domination, and we use this technology to prove the A2
theorem on sharp weighted norm inequalities, only obtained in 2010/2012 and
unforeseen at the time of starting our book project back in 2008. Our approach
to this theorem is still more recent, a result of a sequence of simplifications and
abstractions of the original argument achieved over the past decade. Another
main result on singular integrals is the characterisation of their boundedness
given by the T (1) theorem, which goes back to the 1980’s even in the vector-
valued setting, but was only proved in the full operator-valued generality in
2006. Once again, several subsequent extensions of the argument have taken
place, and we have tried to present a proof, while still non-trivial probably
v
vi Preface

by necessity, that combines elements from several of the existing approaches,

also highlighting intermediate results of independent interest.
The main theorem that we prove about the Fourier transform is the vector-
valued Hausdorff–Young inequality due to Bourgain, again from the 1980’s,
but we have followed a more recent framework of the argument by Hinrichs,
Pietsch and Wenzel from the late 1990’s. Equipped with this tool, we can iden-
tify the operator-valued Fourier multipliers discussed in the first two volumes
with some singular integrals treated in the present one, and thereby obtain
results like weighted norm inequalities for these operators. Many of these
corollaries are relatively recent in the literature. In the previous volumes, we
have also seen that the UMD property is characterised by the boundedness
of a few distinguished Fourier multipliers; we now extend such results to a
much broader class covered by a 2010 result of Geiss, Montgomery-Smith and
Saksman, which also allows us to complete some characterisations of UMD in
terms of the equivalence of different function spaces.
The theory of vector-valued function spaces, already hinted at in several
occasions in the previous volumes, is finally taken up here in a systematic
way. With the complete scale of the relevant function spaces at hand, we can
provide the final form with sharp end-point assumptions of several embedding
theorems that were discussed in weaker or incomplete forms in the previous
volumes. Some of these function spaces will also play an important role later
in this volume, when we treat the maximal regularity problem of evolution
equations.
The notion of functional calculus of sectorial operators was already de-
veloped at length in Volume II, but there is a lot more to add to this vast
topic, both for intrinsic and applied interest. While the necessary background
material on fractional powers dates back as far as the 1960’s, we have followed
a more modern approach of viewing these powers in the framework of the so-
called extended calculus. Of special interest is the class of operators admitting
bounded imaginary powers.
The theory of sums of operators and the operator-valued functional calcu-
lus offer significant extensions of the “vanilla” functional calculus that pave
the way for the treatment of the maximal regularity problem. We also de-
velop the perturbation theory of sectorial operators, which expands the list of
examples of concrete operators for which we can check and hence apply the
H ∞ -calculus.
The penultimate chapter presents a treatment of the maximal regularity
problem for evolution equations. The main characterisation result is from the
turn of the millennium, and so are the related counterexamples by Kalton–
Lancien, but for the latter we will follow a much simplified recent approach
by Fackler from 2014–2016. In the final chapter we present a recent theory of
parabolic nonlinear evolution equations in critical spaces based on maximal Lp
and continuous regularity. This theory was developed during the last decade
by LeCrone, Prüss, Simonett, and Wilke, and it has already turned out to lead
 Preface vii

to far-reaching improvements for several classes of parabolic partial differential

equations.
*
The main stylistic conventions of the previous volumes are adopted in the
present volume as well: Most of the time, we are quite explicit with the con-
stants appearing in our estimates, and we especially try to keep track of the
dependence on the main parameters involved. Some of these explicit quan-
titative formulations appear here for the first time. Where relevant, we also
pay more attention than many texts to the impact of the underlying scalar
field (real or complex) on the results under consideration, although a need
for this perhaps appears slightly less in this volume than in the previous two.
A notable instance is the distinction between the real and complex UMD
constants, and their relation to various multiplier norms.
*
While the previous Volumes I and II of this series were written largely in
parallel over the years 2008–2017, the major work on this Volume III only
took place after the completion of the first two books. Critical to the progress
of this endeavour was the possibility of intensive joint workshops that we held
in the rural serenity of Stiftsgut Keysermühle in Klingenmünster (June 2018,
March 2019) but also, due to other professional and personal commitments
of some of us, in the urban beat of Delft (October 2018, January 2020) and
Helsinki (January 2020). Shortly after the last two meetings, and with the
completion of this volume already on the horizon, the global pandemic broke
out. This changed our plans like so much else, and our progress on this work
was essentially halted for two years. Only at the beginning of 2023 we were
able to resume our ‘live’ writing sessions to finally bring it to completion.
Preliminary versions of parts of the material were presented in advanced
courses and lecture series at various international venues and in seminars at
our departments, and we would like to thank the students and colleagues
who attended these events for feedback that shaped and improved the fi-
nal form of the text. Special thanks go to Antonio Agresti, Sebastian Bech-
tel, Chenxi Deng, Emiel Lorist, Floris Roodenburg, Max Sauerbrey, Esmée
Theewis, Joshua Willems, and Joris van Winden, who read in detail portions
of this book. Needless to say, we take full responsibility for any remaining
errors. Lists with errata for each of the three volumes are maintained on our
personal websites.
During the writing of this book, we have benefited from external funding
by the Academy of Finland / Research Council of Finland (grants 314829 and
346314 to T.H., and the Finnish Centre of Excellence in Randomness and
Structures “FiRST”, of which T.H. is a member), the Netherlands Organisa-
tion for Scientific research (NWO) (VIDI grant 639.032.427 and VICI grant
639.212.027 to M.V.), and the Deutsche Forschungsgemeinschaft (Project-ID
258734477 – SFB 1173 to L.W.).
Delft, Helsinki, and Karlsruhe, September 15, 2023
Contents

11 Singular integral operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

11.1 Local oscillations of functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
11.1.a Sparse collections and Lerner’s formula . . . . . . . . . . . . . . . 4
11.1.b Almost orthogonality in Lp . . . . . . . . . . . . . . . . . . . . . . . . . 9
11.1.c Maximal oscillatory norms for Lp spaces . . . . . . . . . . . . . 13
11.1.d The dyadic Hardy space and BMO . . . . . . . . . . . . . . . . . . 17
11.2 Singular integrals and extrapolation of Lp0 bounds . . . . . . . . . . . 22
11.2.a Calderón–Zygmund decomposition and case p ∈ (1, p0 ) . 25
11.2.b Local oscillations of T f and case p ∈ (p0 , ∞) . . . . . . . . . . 29
11.2.c The action of singular integrals on L∞ . . . . . . . . . . . . . . . 34
11.3 Calderón–Zygmund operators and sparse bounds . . . . . . . . . . . . 41
11.3.a An abstract domination theorem . . . . . . . . . . . . . . . . . . . . 45
11.3.b Sparse operators and domination . . . . . . . . . . . . . . . . . . . . 50
11.3.c Sparse domination of Calderón–Zygmund operators . . . . 54
11.3.d Weighted norm inequalities and the A2 theorem . . . . . . . 56
11.3.e Sharpness of the A2 theorem . . . . . . . . . . . . . . . . . . . . . . . . 66
11.4 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

12 Dyadic operators and the T (1) theorem . . . . . . . . . . . . . . . . . . . 87

12.1 Dyadic singular integral operators . . . . . . . . . . . . . . . . . . . . . . . . . 88
12.1.a Haar multipliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
12.1.b Nested collections of unions of dyadic cubes . . . . . . . . . . . 102
12.1.c The elementary operators of Figiel . . . . . . . . . . . . . . . . . . . 108
12.2 Paraproducts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
12.2.a Necessary conditions for boundedness . . . . . . . . . . . . . . . . 125
12.2.b Sufficient conditions for boundedness . . . . . . . . . . . . . . . . 134
12.2.c Symmetric paraproducts . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
12.2.d Mei’s counterexample: no simple sufficient conditions . . 145
12.3 The T (1) theorem for abstract bilinear forms . . . . . . . . . . . . . . . . 147
12.3.a Weakly defined bilinear forms . . . . . . . . . . . . . . . . . . . . . . . 148
12.3.b The BCR algorithm and Figiel’s decomposition . . . . . . . 154

ix
x Contents

12.3.c Figiel’s T (1) theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160

12.3.d Improved estimates via random dyadic cubes . . . . . . . . . 167
12.4 The T (1) theorem for singular integrals . . . . . . . . . . . . . . . . . . . . 174
12.4.a Consequences of the T (1) theorem . . . . . . . . . . . . . . . . . . . 185
12.4.b The dyadic representation theorem . . . . . . . . . . . . . . . . . . 196
12.5 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208

13 The Fourier transform and multipliers . . . . . . . . . . . . . . . . . . . . . 225

13.1 Bourgain’s theorem on Fourier type . . . . . . . . . . . . . . . . . . . . . . . . 226
13.1.a Hinrichs’s inequality: breaking the trivial bound . . . . . . . 231
13.1.b The finite Fourier transform and sub-multiplicativity . . . 237
13.1.c Key lemmas for an initial uniform bound . . . . . . . . . . . . . 244
13.1.d Conclusion via duality and interpolation . . . . . . . . . . . . . 251
13.2 Fourier multipliers as singular integrals . . . . . . . . . . . . . . . . . . . . . 258
13.2.a Smooth multipliers have Calderón–Zygmund kernels . . . 260
13.2.b Mihlin multipliers have Hörmander kernels . . . . . . . . . . . 267
13.3 Necessity of UMD for multiplier theorems . . . . . . . . . . . . . . . . . . 273
13.4 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283

14 Function spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293

14.1 Summary of the main results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
14.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
14.2.a Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
14.2.b A density lemma and Young’s inequality . . . . . . . . . . . . . 303
14.2.c Inhomogeneous Littlewood–Paley sequences . . . . . . . . . . . 305
14.3 Interpolation of Lp -spaces with change of weights . . . . . . . . . . . . 310
14.3.a Complex interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
14.3.b Real interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
14.4 Besov spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
14.4.a Definitions and basic properties . . . . . . . . . . . . . . . . . . . . . 320
14.4.b Fourier multipliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326
14.4.c Embedding theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332
14.4.d Difference norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
14.4.e Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344
14.4.f Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
14.5 Besov spaces, random sums, and multipliers . . . . . . . . . . . . . . . . 352
14.5.a The Fourier transform on Besov spaces . . . . . . . . . . . . . . . 356
14.5.b Smooth functions have R-bounded ranges . . . . . . . . . . . . 360
14.6 Triebel–Lizorkin spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
14.6.a The Peetre maximal function . . . . . . . . . . . . . . . . . . . . . . . 365
14.6.b Definitions and basic properties . . . . . . . . . . . . . . . . . . . . . 370
14.6.c Fourier multipliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
14.6.d Embedding theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
14.6.e Difference norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
14.6.f Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
 Contents xi

14.6.g Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387

s s
14.6.h Pointwise multiplication by 1R+ in Bp,q and Fp,q . . . . . . 387
14.7 Bessel potential spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393
14.7.a General embedding theorems . . . . . . . . . . . . . . . . . . . . . . . 393
14.7.b Embedding theorems under geometric conditions . . . . . . 395
14.7.c Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
14.7.d Pointwise multiplication by 1R+ in H s,p . . . . . . . . . . . . . . 404
14.8 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407

15 Extended calculi and powers of operators . . . . . . . . . . . . . . . . . . 419

15.1 Extended calculi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419
15.1.a The primary calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 420
15.1.b The extended Dunford calculus . . . . . . . . . . . . . . . . . . . . . 424
15.1.c Extended calculus via compensation . . . . . . . . . . . . . . . . . 435
15.2 Fractional powers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436
15.2.a Definition and basic properties . . . . . . . . . . . . . . . . . . . . . . 437
15.2.b Representation formulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
15.3 Bounded imaginary powers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 456
15.3.a Definition and basic properties . . . . . . . . . . . . . . . . . . . . . . 457
15.3.b Identification of fractional domain spaces . . . . . . . . . . . . . 459
15.3.c Connections with sectoriality . . . . . . . . . . . . . . . . . . . . . . . 462
15.3.d Connections with almost γ-sectoriality . . . . . . . . . . . . . . . 470
15.3.e Connections with γ-sectoriality . . . . . . . . . . . . . . . . . . . . . . 472
15.3.f Connections with boundedness of the H ∞ -calculus . . . . 476
15.3.g The Hilbert space case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482
15.3.h Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 483
15.4 Strip type operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 484
15.4.a Nollau’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 485
15.4.b Monniaux’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 489
15.4.c The Dore–Venni theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 494
15.5 The bisectorial H ∞ -calculus revisited . . . . . . . . . . . . . . . . . . . . . . 497
15.5.a Spectral projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 498
15.5.b Sectoriality versus bisectoriality . . . . . . . . . . . . . . . . . . . . . 500
15.6 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 502

16 Perturbations and sums of operators . . . . . . . . . . . . . . . . . . . . . . 515

16.1 Sums of unbounded operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515
16.2 Perturbation theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 516
16.2.a Perturbations of sectorial operators . . . . . . . . . . . . . . . . . . 517
16.2.b Perturbations of the H ∞ -calculus . . . . . . . . . . . . . . . . . . . 523
16.3 Sum-of-operator theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533
16.3.a The sum of two sectorial operators . . . . . . . . . . . . . . . . . . 536
16.3.b Operator-valued H ∞ -calculus and closed sums . . . . . . . . 541
16.3.c The joint H ∞ -calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 546
16.3.d The absolute calculus and closed sums . . . . . . . . . . . . . . . 551
xii Contents

16.3.e The absolute calculus and real interpolation . . . . . . . . . . 553

16.4 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 560

17 Maximal regularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 569

17.1 The abstract Cauchy problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 570
17.2 Maximal Lp -regularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 575
17.2.a Definition and basic properties . . . . . . . . . . . . . . . . . . . . . . 575
17.2.b The initial value problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 583
17.2.c The role of semigroups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 586
17.2.d Uniformly exponentially stable semigroups . . . . . . . . . . . . 595
17.2.e Permanence properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 598
17.2.f Maximal continuous regularity . . . . . . . . . . . . . . . . . . . . . . 613
17.2.g Perturbation and time-dependent problems . . . . . . . . . . . 619
17.3 Characterisations of maximal Lp -regularity . . . . . . . . . . . . . . . . . 626
17.3.a Fourier multiplier approach . . . . . . . . . . . . . . . . . . . . . . . . . 626
17.3.b The end-point cases p = 1 and p = ∞ . . . . . . . . . . . . . . . . 633
17.3.c Sum-of-operators approach . . . . . . . . . . . . . . . . . . . . . . . . . 636
17.3.d Maximal Lp -regularity on the real line . . . . . . . . . . . . . . . 644
17.4 Examples and counterexamples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 657
17.4.a The heat semigroup and the Poisson semigroup . . . . . . . 657
17.4.b End-point maximal regularity versus containment of c0 . 660
17.4.c Analytic semigroups may fail maximal regularity . . . . . . 665
17.5 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 671

18 Nonlinear parabolic evolution equations in critical spaces . . 689

18.1 Semi-linear evolution equations with F = FTr . . . . . . . . . . . . . . . 692
18.2 Local well-posedness for quasi-linear evolution equations . . . . . . 697
18.2.a Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 697
18.2.b Main local well-posedness result . . . . . . . . . . . . . . . . . . . . . 701
18.2.c Proof of the main result . . . . . . . . . . . . . . . . . . . . . . . . . . . . 702
18.2.d Maximal solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 712
18.3 Examples and comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 717
18.4 Long-time existence for small initial data and F = Fc . . . . . . . 725
18.5 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 727

Q Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 733

K Measurable semigroups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 745

K.1 Measurable semigroups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 745
K.2 Uniform exponential stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 754
K.3 Analytic semigroups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 756
K.4 An interpolation result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 758
K.5 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 762
 Contents xiii

L The trace method for real interpolation . . . . . . . . . . . . . . . . . . . . 763

L.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 763
L.2 The trace method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 765
L.3 Reiteration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 771
L.4 Mixed derivatives and Sobolev embedding . . . . . . . . . . . . . . . . . . 774
L.4.a Results for the half-line . . . . . . . . . . . . . . . . . . . . . . . . . . . . 775
L.4.b Extension operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 779
L.4.c Results for bounded intervals . . . . . . . . . . . . . . . . . . . . . . . 780
L.5 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 782

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 783

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 820
Symbols and notations

Sets

N = {0, 1, 2, . . .} - non-negative integers

Z - integers
Q - rational numbers
R - real numbers
C - complex numbers
K - scalar field (R or C)
Z = Z ∪ {−∞, ∞} - extended integers
R+ = (0, ∞) - positive real line
BX - open unit ball
SX - unit sphere
B(x, r) - open ball centred at x with radius r
D - open unit disc
S = {z ∈ C : 0 < =z < 1} - unit strip
Σω = open sector of angle ω
Σωbi = open bisector of angle ω
T = {z ∈ C : |z| = 1} - unit circle

Vector spaces
s
Bp,q - Besov space
c0 - space of null sequences
C - space of continuous functions
C0 - space of continuous functions vanishing at infinity
C α - space of Hölder continuous functions
Cb - space of bounded continuous functions
Cc - space of continuous functions with compact support
Cc∞ - space of test functions with compact support
C p - Schatten class
γ(H, X) - space of γ-radonifying operators

xv
xvi Symbols and notations

γ(S; X) - shorthand for γ(L2 (S), X)

γ∞ (H, X) - space of almost summing operators
γ∞ (S; X) - shorthand for γ∞ (L2 (S), X)
E - space of primary functions
H - Hilbert space
H s,p - Bessel potential space
H p - Hardy space
`p - space of p-summable sequences
p
`N - space of p-summable finite sequences
H p - Hardy spaces on a sector or strip
Lp - Lebesgue space
Lp,q - Lorentz space
Lp,∞ - weak-Lp
b pw (I; X) - weighted Lp
L
L1 - space of inverse Fourier transforms
Lip - space of Lipschitz continuous functions
L (X, Y ) - space of bounded linear operators
MLp - space of Fourier multipliers
M - Mihlin class
S - space of Schwartz functions
S 0 - space of tempered distributions
s
Fp,q - Triebel–Lizorkin space
W k,p - Sobolev space
W s,p - Sobolev-Slobodetskii space
X, Y , . . . - Banach spaces
XC - complexification
XCγ,p - Gaussian complexification
X ∗ , Y ∗ , . . . - dual Banach spaces
X , Y , . . . - strongly continuous semigroup dual spaces
X ⊗ Y - tensor product
[X0 , X1 ]θ - complex interpolation space
(X0 , X1 )θ,p , (X0 , X1 )θ,p0 ,p1 - real interpolation spaces

Measure theory and probability

A - σ-algebra
dfn = fn − fn−1 - nth martingale difference
n - signs in K, i.e., scalars in K of modulus one
εn - Rademacher variables with values in K
E - expectation
F , G , . . . - σ-algebras
Ff - collection of sets in F on which f is integrable
E(·|·) - conditional expectation
γn - Gaussian variables
hI - Haar function
 Symbols and notations xvii

µ - measure
kµk - variation of a measure
(Ω, A , P) - probability space
P - probability measure
rn - real Rademacher variables
(S, A , µ) - measure space
σ(f, g, . . . ) - σ-algebra generated by the functions f, g, . . .
σ(C ) - σ-algebra generated by the collection C
τ - stopping time
wα - Walsh function

Norms and pairings

| · | - modulus, Euclidean norm

k · k = k · kX - norm in a Banach space X
k · kp = k · kLp - Lp -norm
h·, ·i - duality
(·|·) - inner product in a Hilbert space
a · b - inner product of a, b ∈ Rd

Operators

A - closed linear operator

A∗ - adjoint operator
A - part of A∗ in X
D(A) - domain of A
∇ - gradient
∆ - Laplace operator
γ(T ) - γ-bound of the operator family T
γp (T ) - γ-bound of T with respect to the Lp -norm
D - dyadic system
∂j = ∂/∂xj - partial derivative with respect to xj
∂ α - partial derivative with multi-index α
E(·|·) - conditional expectation
F - Fourier transform
F −1 - inverse Fourier transform
H - Hilbert transform
He - periodic Hilbert transform
Js - Bessel potential operator
`2 (T ) - `2 -bound of the operator family T
L (X, Y ) - space of bounded operators from X to Y
Mσ,A - sectoriality constant of A at angle σ
N(A) - null space of A
R(T ) - R-bound of the operator family T
Rp (T ) - R-bound of T with respect to the Lp -norm
R(A) - range of A
xviii Symbols and notations

Rj - jth Riesz transform

R(λ, A) = (λ − A)−1 - resolvent of A at λ
%(A) - resolvent set of A
σ(A) - spectrum of A
S, T , . . . - bounded linear operators
S(t), T (t), . . . - semigroup operators
S ∗ (t), T ∗ (t), . . . - adjoint semigroup operators on the dual space X ∗
S (t), T (t), . . . - their parts in the strongly continuous dual X
T ∗ - adjoint of the operator T
T ? - Hilbert space (hermitian) adjoint of Hilbert space operator T
Tm - Fourier multiplier operator associated with multiplier m
T ⊗ IX - tensor extension of T
ω(A) - angle of sectoriality of A
ωR (A), ωγ (A) - angles of R- and γ-sectoriality of A
ωBIP (A) - angle of bounded imaginary powers of A
ωH ∞ (A) - angle of the H ∞ -calculus of A
ω bi (A) - angle of bisectoriality of A

Constants and inequalities

αp,X - Pisier contraction property constant

±
αp,X - upper and lower Pisier contraction property constant
βp,X - UMD constant
R
βp,X - UMD constant with signs ±1
±
βp,X - upper and lower randomised UMD constant
cq,X - cotype q constant
γ
cq,X - Gaussian cotype q constant
∆p,X - triangular contraction property constant
~p,X - norm of the Hilbert transform on Lp (R; X)
Kp,X - K-convexity constant
γ
Kp,X - Gaussian K-convexity constant
κp,q - Kahane–Khintchine constant
κR
p,q - idem, for real Rademacher variables
γ
κp,q - idem, for Gaussian sums
κp,q,X - idem, for a fixed Banach space X
τp,X - type p constant
γ
τp,X - Gaussian type p constant
0
ϕp,X (Rd ) - norm of the Fourier transform F : Lp (Rd ; X) → Lp (Rd ; X).

Miscellaneous

,→ - continuous embedding
1A - indicator function
a . b - ∃C such that a 6 Cb
a .p,P b - ∃C, depending on p and P , such that a 6 Cb
 Symbols and notations xix

C - generic constant
{ - complement
d(x, y) - distance
f ? - maximal function
fe - reflected function
fbb - Fourier transform
f - inverse Fourier transform
f ∗ g - convolution
= - imaginary part
M f - Hardy–Littlewood maximal function
p0 = p/(p − 1) - conjugate exponent
p∗ = max{p, p0 }
℘ - good set-bound
< - real part
s ∧ t = min{s, t}
s ∨ t = max{s, t}
x - generic element of X
x∗ - generic element of X ∗
x ⊗ y - elementary tensor
x+ , x− , |x| - positive part, negative part, and modulus of x
w - weight
wα - power weight t 7→ tα
Standing assumptions

Throughout this book, two conventions will be in force.

1. Unless stated otherwise, the scalar field K can be real or complex. Results
which do not explicitly specify the scalar field to be real or complex are
true over both the real and complex scalars.
2. In the context of randomisation, a Rademacher variable is a uniformly
distributed random variable taking values in the set {z ∈ K : |z| = 1}.
Such variables are denoted by the letter ε. Thus, whenever we work over
R it is understood that ε is a real Rademacher variable, i.e.,
1
P(ε = 1) = P(ε = −1) = ,
2
and whenever we work over C it is understood that ε is a complex
Rademacher variable (also called a Steinhaus variable), i.e.,
1
P(a < arg(ε) < b) = (b − a).
2π
Occasionally we need to use real Rademacher variables when working over
the complex scalars. In those instances we will always denote these with
the letter r. Similar conventions are in force with respect to Gaussian
random variables: a Gaussian random variable is a standard normal real-
valued variable when working over R and a standard normal complex-
valued variable when working over C.

xxi
11
Singular integral operators

Various operators of Analysis, many of them already encountered in these

volumes, take the generic form
Z
T f (s) = K(s, t)f (t) dt. (11.1)
Rd

The mapping properties of T will of course heavily depend on the assumptions

made on the kernel K that we will discuss in more detail in this chapter. A
general feature of the different conditions is that the kernel is allowed to blow
up on the ‘diagonal’ {(x, x) : x ∈ Rd }, so that its natural domain of definition
is the set
Ṙ2d := {(s, t) ∈ Rd × Rd : s 6= t}.
This blow-up is one of the reasons for referring to (11.1) as a singular inte-
gral; in general this formula requires a careful interpretation and will only be
meaningful under restrictions on f and s.
In the prominent special case of a convolution kernel K(s, t) = K(s − t),
the operator (11.1) takes (at least formally, and under reasonable assumptions
also rigorously) a simple representation “on the Fourier transform side”:

[
Tcf (ξ) = K ∗ f (ξ) = K(ξ)
b fb(ξ) =: m(ξ)fb(ξ);

thus T = Tm can be identified with a Fourier multiplier; they have been

studied extensively in Chapter 5 and Section 8.3.
The motivations to investigate singular integral operators in the non-
transformed representation (11.1) are at least threefold. First, it allows for
a wider class of examples beyond those of the convolution form. Second, even
when the alternative Fourier multiplier representation is available in principle,
an operator may naturally arise in the form (11.1), and identifying or estimat-
ing the corresponding multiplier explicitly may not be feasible in practise, as
the Fourier transform is not isomorphic between the natural function spaces
for the kernel K and the multiplier m. Finally, and perhaps most importantly,

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 1

T. Hytönen et al., Analysis in Banach Spaces, Ergebnisse der Mathematik und ihrer
Grenzgebiete. 3. Folge / A Series of Modern Surveys in Mathematics 76,
https://doi.org/10.1007/978-3-031-46598-7_1
2 11 Singular integral operators

even for multiplier operators, the point-of-view of singular integrals gives us

access to new methods and conclusions.
An overarching theme is of this chapter is extrapolation: As soon as an
operator (11.1), with natural assumptions on the kernel K, is bounded on
a single space Lp0 (Rd ; X), it will be automatically bounded on several more
spaces, including Lp (Rd ; X) for other exponents p ∈ (1, ∞) (with certain sub-
stitute results at the end-points p ∈ {1, ∞}), and even their weighted versions
Lp (w; X), where w is an arbitrary weight in the Muckenhoupt class Ap (see
Appendix J). These results will be used to deduce analogous extrapolation re-
sults for maximal Lp -regularity of the abstract Cauchy problem in Chapter 17.
In terms of Banach spaces, this chapter deals with relatively general re-
sults, most of which are valid without restrictions of the class of admissible
spaces. Such restrictions, and notably the ubiquitous UMD condition, will
reappear in the subsequent chapters, when searching for conditions to verify
the boundedness of (11.1) on just one Lp0 (Rd ; X), to serve as an input to the
extrapolation results that we develop in the chapter at hand.

11.1 Local oscillations of functions

A characteristic feature of singular integrals, the main topic of this chapter, is
that their boundedness properties depend not only naive size estimates but on
rather delicate cancellations between different oscillatory components. Before
we dwell into a deeper study of there operators, we dedicate this section to a
general treatment of oscillations of functions per se; this will streamline the
subsequent discussion, where the results of this section will be put into action
in the context of operator norm estimates.
Given f ∈ L0 (Rd ; X) and λ > 0, we define the following measure of oscil-
lation of f on a cube Q,

oscλ (f ; Q) := inf inf k(f − c)1Q\E k∞ .

c∈X |E|6λ|Q|

Here, and in many occasions below where we will use the same notation, it is
understood that the supremum is taken over all measurable subsets E of Q
satisfying the stated requirement that |E| 6 λ|Q|. The idea is to quantify how
much f deviates from a constant, if we ignore its (possibly wild) behaviour
on an exceptional set of controlled proportion. The above way of measuring
oscillations is essentially ‘minimal’ in that it can be controlled by average Lq
oscillations for any q > 0:

Lemma 11.1.1. For any q ∈ (0, ∞), we have

k(f − c)1Q kLq,∞

oscλ (f ; Q) 6 inf .
c∈X (λ|Q|)1/q
 11.1 Local oscillations of functions 3

Proof. For a fixed c, let g := (f − c)1Q . If we choose t := kgkLq,∞ /(λ|Q|)1/q ,

then
kgkqLq,∞
|Et | := |{kgk > t}| 6 = λ|Q|
tq
But then it is clear that

inf kg1Q\E k∞ 6 kg1Q\Et k∞ 6 t,

|E|6λ|Q|

which is precisely the claimed bound. 

Given a real-valued f ∈ L0 (Rd ; R), any m ∈ R such that

|Q ∩ {f 6 m}| > 21 |Q|, |Q ∩ {f > m}| > 21 |Q|

is called a median of f on the cube (or more general set of finite positive
measure) Q ⊆ Rd . One routinely checks that a median always exists but may
fail to be unique.

Lemma 11.1.2. If λ ∈ (0, 21 ) and mf ∈ R is a median of f ∈ L0 (Q; R) on

Q, then
inf k(f − mf )1Q\E k∞ 6 2 oscλ (f ; Q).
|E|6λ|Q|

Proof. Let c ∈ R be arbitrary. Then f − mf = f − c − (mf − c) and hence

inf k(f − mf )1Q\E k∞ 6 inf k(f − c)1Q\E k∞ + |mf − c|.

|E|6λ|Q| |E|6λ|Q|

Note that mf − c is a median of g := f − c on Q. Hence it suffices to check

that the median mg always satisfies

|mg | 6 kg1Q\E k∞

whenever |E| 6 λ|Q| and λ < 21 . If mg > 0, then

|Q ∩ {|g| > |mg |} \ E| > |Q ∩ {g > mg } \ E| > 21 |Q| − |E| > ( 21 − λ)|Q| > 0

and thus kg1Q\E k∞ > |mg |. If mg < 0, the argument is the same, just replac-
ing the second step above by |Q ∩ {g 6 mg } \ E|. 
The previous lemma motivates the following:

Definition 11.1.3. Let X be a Banach space and f ∈ L0 (Q; X). A vector

m ∈ X is called a λ-pseudomedian of f on Q if

inf k(f − m)1Q\E k∞ 6 2 oscλ (f ; Q).

|E|6λ|Q|

Indeed, Lemma 11.1.2 says that the usual median is a λ-pseudomedian for
every λ ∈ (0, 21 ). Concerning existence in the general case, we have:
4 11 Singular integral operators

Lemma 11.1.4. Let X be a Banach space, f ∈ L0 (Q; X) and λ ∈ (0, 21 ).

Then f has a λ-pseudomedian on Q.

Proof. If oscλ (f ; Q) > 0, this is obvious, since we can always come within any
positive distance from the infimum. So only the case oscλ (f ; Q) = 0 needs
attention. In this case, there we can find a sequence of vectors cn ∈ X and
sets En ⊆ Q with |En | 6 λ|Q| such that k(f − cn )1Q\En k∞ → 0. Since
|En ∪ Em | 6 2λ|Q| < |Q|, any Q \ (En ∪ Em ) has positive measure, and thus

kcn − cm k = k(cn − cm )1Q\(En ∪Em ) k∞

6 k(f − cn )1Q\En k∞ + k(f − cm )1Q\Em k∞ → 0.

Thus (cn )n>1 is a Cauchy sequence and hence convergent to some c ∈ X. But
then

inf k(f − c)1Q\E k∞ 6 lim inf k(f − c)1Q\En k∞

|E|6λ|Q| n→∞
 
6 lim inf k(f − cn )1Q\En k∞ + kcn − ck = 0,
n→∞

and thus this limit c is a λ-pseudomedian. 

Lemma 11.1.5. Let X be a Banach space, let f ∈ L0 (Rd ; X) and λ ∈ (0, 21 ),

and let mf (Q) be a λ-pseudomedian of f on Q. Then

E 0 := Q ∩ {kf − mf (Q)k > 2 oscλ (f ; Q)}

satisfies |E 0 | 6 λ|Q|.

Proof. Suppose for contradiction that |E 0 | > λ|Q|. Denoting

E ε := Q ∩ {kf − mf (Q)k > 2 oscλ (f ; Q) + ε}

S∞
we have E 0 = n=1 E 1/n , so that by continuity of measure, we also have
|E ε | > λ|Q| for some ε = 1/n > 0.
Let |E| 6 λ|Q|. Then

k(f − mf (Q))1Q\E k∞ > (2 oscλ (f ; Q) + ε)k1E ε \E k∞ = 2 oscλ (f ; Q) + ε,

since |E ε \ E| > |E ε | − |E| > λ|Q| − λ|Q| = 0. Taking the infimum over all
|E| 6 λ|Q|, we contradict the definition of a λ-pseudomedian. 

11.1.a Sparse collections and Lerner’s formula

Let us recall and expand the terminology related to dyadic cubes that we
introduced in Chapter 3.

Definition 11.1.6. A dyadic system of cubes on Rd is a collection D =

j∈Z Dj , where
S
 11.1 Local oscillations of functions 5

(i) each Dj is a partition of Rd of the form

n o
Dj = sj + 2−j (m + [0, 1)d ) : m ∈ Zd ,

(ii) each Dj+1 refines the previous Dj .

When sj = 0 for all j ∈ Z, we refer to the corresponding D as the standard
dyadic system, and denote it by D 0 .

Remark 11.1.7. One might like to replace (i) in Definition 11.1.6 by the “more
intrinsic”
(iii) each Dj is a partition of Rd consisting of left-closed, right-open cubes of
side-length 2−j .
When d = 1, one can check that (i) and (iii) are equivalent. But, for d > 1,
condition (iii) is strictly more general. For instance

Dj := 2−j m + [0, 1)2 + (0, α1[0,∞) (m1 )) : m ∈ Z2 , α ∈ R,



where all cubes in the right half-plane are shifted in the y-direction by a fixed
amount α ∈ R relative to the standard dyadic cubes, would qualify for (iii) but
not for (i). The preference over one or the other definition may be a question
of taste; we choose to work with Definition 11.1.6 as stated.

We will work be working with an arbitrary dyadic system as in Definition

11.1.6. For many purposes, the reader who so wishes may think of the standard
dyadic system.
[
D 0 := Dj0 , Dj0 := {2−j ([0, 1)d + k) : k ∈ Zd }, j ∈ Z,
j∈Z

but here and there we will also make use of other systems, which makes it
convenient to deal with a generic system from the beginning. For any given
cube, we may speak of its dyadic subcubes, by which we understand all cubes
obtained by repeatedly bisecting the edges of Q. We will use the notation
D(Q) for the collection of all dyadic subcubes of a cube Q. If Q belongs to a
dyadic system D, then

D(Q) = {Q0 ∈ D : Q0 ⊆ Q}.

Definition 11.1.8. A quadrant of a dyadic system D of Rd is the union of

any strictly increasing sequence Q1 ( Q2 ( Q3 ( . . . of cubes Qi ∈ D.

Remark 11.1.9. The standard dyadic system D 0 has 2d quadrants of the form
S1 × · · · × Sd , where Si ∈ {(−∞, 0), [0, inf ty)} for each i ∈ {1, . . . , d}. It is
also easy to construct dyadic systems, where Rd is the only quadrant.
6 11 Singular integral operators

The dyadic Hardy–Littlewood maximal function is defined by

Z Z
1
MD f (x) = sup hkf kiQ , hf iQ := − f := f,
Q∈D:x∈Q Q |Q| Q

where the supremum is taken over all dyadic cubes containing x. Here, and
throughout this chapter, unless indicated otherwise, integrals are taken with
respect to Lebesgue measure and are abbreviated in theRabove way to unbur-
den notation. Thus, when g is an integrable function, Q g is shorthand for
g(x) dx. When integrating over all of Rd we will even write g for Rd g.
R R R
Q

Definition 11.1.10. A collection S of sets S ⊆ Rd of finite measure is called

γ-sparse, if each S ∈ S has a measurable subset E(S) ⊆ S of size |E(S)| >
γ|S| such that the sets E(S) are pairwise disjoint.
While the definition can be made for general measurable sets, we will be
mostly concerned with the case when S ⊆ D is a subcollection of the dyadic
cubes of Rd .
A disjoint collection is obviously 1-sparse with E(S) = S. The usefulness of
general γ-sparse collections comes from the fact that, on the one hand, they
are easier to create than genuinely disjoint collections while, on the other
hand, for the purposes of Lp estimates they are essentially as good as disjoint
ones. This is quantified by the following:
Proposition 11.1.11. Let S ⊆ D be a γ-sparse collection of dyadic cubes S
with disjoint subsets |E(S)| > γ|S|.
(1) If aS > 0, then for all p ∈ (0, ∞),
(
X X γ −1 p, p ∈ [1, ∞),
a S 1S 6 cp,γ aS 1E(S) , where cp,γ =
S∈S
p
S∈S
p γ −1/p , p ∈ (0, 1).

(2) If f > 0, then for all p ∈ (1, ∞),

X 1/p
hf ipS |S| 6 γ −1/p p0 kf kp .
S∈S

Proof of Proposition 11.1.11. If p ∈ [1, ∞), we dualise the left side of (1)
0
against φ ∈ Lp :
Z X Z
 X 1 X
a S 1S φ = aS |S|− φ 6 aS |E(S)| inf MD φ
S γ S
S∈S S∈S S∈S
Z X
1  1 X
6 aS 1E(S) MD φ 6 aS 1E(S) kMD φkp0 ,
γ γ p
S∈S S∈S

where kMD φkp0 6 pkφkp0 by Doob’s maximal inequality (Theorem 3.2.2; cf.
the explanations preceding Theorem 3.2.27).
 11.1 Local oscillations of functions 7

If p ∈ (0, 1), then the left side of (1) can be estimated by

Z X p Z X X p 1 X p
a S 1S 6 aSp 1S = aS |S| 6 aS |E(S)|
γ
S∈S S∈S S∈S S∈S
Z X Z p
1 1 X
= apS 1E(S) = aS 1E(S) ,
γ γ
S∈S S∈S

and taking the pth root completes the proof of (1).

For (2), we use hf iS 6 inf z∈S MD f (z) and |S| 6 γ −1 |E(S)| to find that
Z
X p 1 X 1 1
hf iS |S| 6 (MD f )p dx 6 kMD f kpp 6 (p0 )p kf kpp ,
γ E(S) γ γ
S∈S S∈S

again by Doob’s inequality in the last step. 

The different notions introduced above come together in the following useful
estimate, which is the main result of this section:

Theorem 11.1.12 (Lerner’s formula). Let X be a Banach space, Q0 ⊆ Rd

be a cube and f ∈ L0 (Q0 ; X). Then there is a 21 -sparse subcollection S ⊆
D(Q0 ) such that, almost everywhere,
X
1Q0 kf − mf (Q0 )k 6 4 oscλ (f ; S)1S , λ = 2−2−d ,
S∈S

where mf (Q0 ) is any λ-pseudomedian of f on Q0 .

By Lemma 11.1.2, if X = R, we can take mf (Q0 ) to be a usual median of f .

Proof. We begin with a preliminary observation. For any collection of disjoint
sets Qj ∈ D(Q0 ), we have the identity

1Q0 (f − mf (Q0 )) = 1Q0 \Sj Qj (f − mf (Q0 ))

X
+ 1Qj (mf (Qj ) − mf (Q0 ))
j (11.2)
X
+ 1Qj (f − mf (Qj )).
j

Turning to the actual proof, let

n o
E 0 := Q0 ∩ kf − mf (Q0 )k > 2 oscλ (f ; Q0 )

so that |E 0 | 6 λ|Q0 | by Lemma 11.1.5.

For α ∈ (0, 1) to be chosen, let Q1j be the maximal cubes in D(Q0 ) such
that |Q1j ∩ E 0 | > α|Q1j |. Since any two dyadic cubes are either disjoint, or
one is contained in the other, dyadic cubes that are maximal with respect
8 11 Singular integral operators

to some property are necessarily disjoint; hence our preliminary observation

applies to Qj = Q1j . Moreover, by definition of the dyadic maximal operator,
we have MD 1E 0 (x) > α, if and only if x is contained in some dyadic Q with
h1E 0 iQ > α, if and only if it is contained in a maximal dyadic cube with this
property. Hence
[
Q1j = {MD 1E 0 > α},
j

so that by Doob’s inequality

X 1 λ
|Q1j | 6 k1E 0 k1 6 |Q0 |.
j
α α

By Lebesgue’s differentiation theorem, almost every point of E is contained

in some Q1j , and hence

1Q0 \Sj Q1j kf − mf (Q0 )k 6 1Q0 \Sj Q1j 2 oscλ (f ; Q0 )

almost everywhere.
By the maximality of the Q1j , their parent cubes Q
b 1 satisfy the opposite
j
bound |Qb 1 ∩ E| 6 α|Q
b 1 |. Hence in particular
j j

|Q1j ∩ E 0 | 6 |Q
b 1 ∩ E 0 | 6 α|Q
j
b 1 | = 2d α|Q1 |.
j j

Let also n o
Ej1 = Q1j ∩ kf − mf (Q1j )k > 2 oscλ (f ; Q1j )

so that |Ej1 | 6 λ|Q1j | by Lemma 11.1.5. It follows that

|Q1j ∩ (E 0 ∪ Ej1 )| 6 (2d α + λ)|Q1j |.

If 2d α + λ < 1, then Q1j \ (E 0 ∪ Ej1 ) has positive measure, and for any x in
this set, we have both

kf (x) − mf (Q0 )k 6 2 oscλ (f ; Q0 ), kf (x) − mf (Q1j )k 6 2 oscλ (f ; Q1j ).

Since such points x exist, it follows in particular that

kmf (Q1j ) − mf (Q0 )k 6 2 oscλ (f ; Q0 ) + 2 oscλ (f ; Q1j ).

Substituting this to (11.2), we have

 11.1 Local oscillations of functions 9
1Q0 kf − mf (Q0 )k 6 1Q0 \Sj Qj1 2 oscλ (f ; Q0 )
X  
+ 1Q1j 2 oscλ (f ; Q0 ) + 2 oscλ (f ; Qj )
j
X
+ 1Qj1 kf − mf (Q1j )k
j (11.3)
X
= 1Q0 2 oscλ (f ; Q0 ) + 1Q1j 2 oscλ (f ; Q1j )
j
X
+ 1Qj1 kf − mf (Qj1 )k,
j

where each term in the last sum has exactly the same form as the left hand
side and allows to iterate the same consideration.
Assuming that we have proved
N
X −1 X X
1Q0 kf − mf (Q0 )k 6 4 1Qjn oscλ (f ; Qnj ) + 2 1QjN oscλ (f ; QjN )
n=0 j j
X
+ 1QN
j
kf − mf (QN
j )k,
j

where each Qjn is contained in some Qn−1

i and
X λ n−1
|Qnj | 6 |Q |, (11.4)
n−1
α i
j:Qn
j ⊆Qi

applying (11.3) to each QN 0

j in place of Q yields the analogue of the previous
display with N + 1 in place of N .
N
P
The support of the final error term has measure at most j |Qj | 6
(λ/α)N |Q0 |, so if λ/α < 1, this error term tends to zero pointwise almost
everywhere. Hence, in the limit, we have
∞ X
X
1Q0 kf − mf (Q0 )k 6 4 1Qnj oscλ (f ; Qnj ).
n=0 j

Choosing α = 2λ, (11.4) shows that the collection {Qnj }n,j is 21 -sparse, and
with λ = 2−2−d , we also have 2d α + λ = (2d+1 + 1)λ = 2−1 + 2−1−d < 1, as
required. This concludes the proof. 

11.1.b Almost orthogonality in Lp

In a Hilbert space H such as H = L2 (Rd ), orthogonality of elements hi implies

the fundamental Pythagorean identity
10 11 Singular integral operators
X X 1/2
hi = khi k2H .
H
i i

As we have seen in the previous Volumes, Lp analogues of this identity tend to

either take the form of a one-sided estimate only, or, insisting in a two-sided
equivalence, require the introduction of some randomised norm. In contrast to
this, it may come as a surprise that sparse collections lead to relatively simple
constructions that allow almost complete Lp analogues of the Pythagorean
identity in certain situations.
We introduce some additional notation. The following definition is mean-
ingful for any subcollection S ⊆ D of the dyadic cubes, but it will prove itself
particularly useful when S is sparse.

Definition 11.1.13. For any subcollection S ⊆ D of dyadic cubes, we have

the following notions:
(1) For each S ∈ S , let chS (S) ⊆ S (the S -children of S) denote the
collection of all maximal S 0 ∈ S such 0
S that S ( 0S.
(2) For each S ∈ S , let ES (S) := S \ S 0 ∈chS (S) S .
(3) For each Q ∈ D, let πS (Q) denote the minimal S ∈ S such that S ⊇ Q.

When S = D, we reproduce the familiar notion chD = ch of dyadic children.

The other two notions above are uninteresting in this special case, as we simply
have ED (Q) = ∅ and πD (Q) = Q for all Q ∈ D.
We begin with a one-sided estimate:

Proposition 11.1.14. Let X be a Banach space and p ∈ [1, ∞). Let S ⊆ D

be a γ-sparse collection of dyadic cubes. For each S ∈ S , let fS ∈ Lp (Rd ; X)
be a function supported on S and constant on each S 0 ∈ chS (S). Then
X 0
X 1/p
fS 6 (1 + γ −1/p p) kfS kpLp (Rd ;X) .
Lp (Rd ;X)
S∈S S∈S

Proof. We assume that the right-hand side is finite, for otherwise there is
nothing to prove. We then assume without loss of generality that S is finite.
In fact, once we have proved the result for finite families,
P in the infinite case
it follows easily that the partial sums of the series S∈S fS (with arbitrary
enumeration) form a Cauchy sequence in Lp (Rd ; X), from which we deduce
the (unconditional) convergence of this series and the asserted norm bound.
0
Concentrating on the finite case, by dualising with g ∈ Lp (Rd ; X ∗ ), it is
equivalent to the estimate
Z X X 1/p
0
hfS , gi dx 6 (1 + γ −1/p p) kfS kpLp (Rd ;X) kgkLp0 (Rd ;X ∗ ) .
S S

Since fS is supported on S and constant on each S 0 ∈ chS (S), and since S is

partitioned by chS (S) ∪ {ES (S)}, we have
 11.1 Local oscillations of functions 11
Z X XZ
hfS , gi dx = hfS , gi dx
S∈S S∈S S
X X Z X
0
= hhfS iS 0 , hgiS 0 i|S | + 1ES (S) hfS , gi dx.
S∈S S 0 ∈chS (S) S∈S

We can estimate the second term by Hölder’s inequality and the pairwise
disjointness of the sets ES (S),
Z X X
1ES (S) hfS , gi dx 6 1ES (S) fS p d kgkLp0 (Rd ;X ∗ )
L (R ;X)
S∈S S∈S
X 1/p
= k1ES (S) fS kpLp (Rd ;X) kgkLp0 (Rd ;X ∗ ) .
S∈S

For the first term we argue as follows.

X X
hhfS iS 0 , hgiS 0 i|S 0 |
S∈S S 0 ∈chS (S)
X X 1/p  X X 0
1/p0
6 khfS iS 0 kpX |S 0 | khgiS 0 kpX ∗ |S 0 |
S∈S S 0 ∈chS (S) S∈S S 0 ∈chS (S)
X 1/p  X 0
1/p0
p
6 kfS kL p (Rd ;X) hkgkX ∗ ipS 0 |S 0 | ,
S∈S S 0 ∈S

where in the second factor we rearranged the double sum into a single sum,
observing that every S 0 ∈ S is counted at most once as a child of a unique
0
S ∈ S . The second factor is bounded by γ −1/p pkgkp0 thanks to Proposi-
tion 11.1.11(2). Summing up the bounds, we complete the proof of the direct
estimate. 
The following lemma describes useful projections and also provides prominent
examples of the functions fS featuring in Proposition 11.1.14.
Lemma 11.1.15. For S ∈ S ⊆ D and f ∈ Lloc
1
(Rd ; X), let
X
PS f := ES 0 f + 1ES (S) f. (11.5)
S 0 ∈chS (S)

Then hf iQ = hPS f iQ for all Q ∈ D such that πS (Q) = S.

Proof. From definition, we have
|S 0 ∩ Q|
Z
X 1
hPS f iQ = hf iS 0 + f dx.
|Q| |Q| Q∩ES (S)
S 0 ∈chS

Since πS (Q) = S, we have Q ⊆ S and it is not possible that Q ⊆ S 0 ∈ chS (S).

Hence S 0 ∩ Q ∈ {∅, S 0 } for all S 0 ∈ chS (S) and Q is exactly partitioned by
Q ∩ ES (S) and those S 0 ∈ chS (S) with S 0 ( Q. Thus
12 11 Singular integral operators

|S 0 ∩ Q|
Z
X 1 X
01 X
hf iS 0 = hf iS 0 |S | = f dx,
|Q| |Q| |Q| S0
S 0 ∈chS S 0 ∈chS ,S 0 ⊆Q S 0 ∈chS ,S 0 ⊆Q

and
Z Z Z
1 X 1 1
hPS f iQ = f dx + f dx = f dx,
|Q| S0 |Q| Q∩ES (S) |Q| Q
S 0 ∈chS ,S 0 ⊆Q

confirming the lemma. 

A typical way in which a sparse collection arises is via the following basic
construction:

Definition 11.1.16 (Principal cubes). Let Q0 ∈ D and f ∈ L1 (Q0 ; X).

The collection
S∞of principal cubes of f in D(Q0 ) with parameter A > 1 is the
family S = k=0 Sk constructed as follows:
(1) S0 := {Q0 }.
(2) If Sk is already defined for some k ∈ N, then
(a) for each S ∈ Sk we let
n o
chS (S) := S 0 ∈ D(S) maximal with hkf kX iS 0 > Ahkf kX iS ,

(b) and then [

Sk+1 := chS (S).
S∈Sk

The first instance of the interplay of a function and its principal cubes is the
following:

Lemma 11.1.17. Let f ∈ L1 (Q0 ; X) and S be the principal cubes of f with

parameter A > 1. Then S is (1 − A−1 )-sparse, and in fact
1
|ES (S)| > (1 − )|S|. (11.6)
A
If PS f is defined by (11.5), then kPS f kL∞ (Rd ;X) 6 2d Ahkf kX iS .

Note that (11.6) is slightly more than the mere (1 − A−1 )-sparseness of S :
it says that the disjoint subsets E(S) ⊆ S in the definition of sparseness may
be chosen as E(S) = ES (S), which is not always the case for an arbitrary
sparse family. For instance, S = {[0, 1), [0, 21 ), [ 12 , 1)} is 12 -sparse, and one can
take for instance E([0, 1)) = [ 41 , 34 ), E([0, 12 )) = [0, 41 ) and E([ 21 , 1)) = [ 43 , 1),
but ES ([0, 1)) = ∅ in this case.
Proof. By maximality, the cubes S 0 ∈ chS (S) are pairwise disjoint. From the
defining condition it follows that
 11.1 Local oscillations of functions 13
R R
X X kf kX dx
S0
kf kX dx |S|
|S 0 | 6 6 S =
Ahkf kX iS Ahkf kX iS A
S 0 ∈chS (S) S 0 ∈chS (S)

and hence X 1
|ES (S)| = |S| − |S 0 | > (1 − )|S|.
A
S 0 ∈chS (S)

If x ∈ ES (S), then x is not contained in any S 0 ∈ chS (S), and hence

hkf kX iQ 6 Ahkf kX iS for all Q ∈ D(S) with x ∈ Q. As `(Q) → 0, it fol-
lows from Lebesgue’s Differentiation Theorem that kPS f (x)kX = kf (x)kX 6
Ahkf kX iS for almost every x ∈ ES (S). If x ∈ S 0 ∈ chS (S), then fS (x) =
hf iS 0 . By the maximality of S 0 , its dyadic parent Sb0 satisfies the opposite
inequality hkf kX iSb0 6 Ahkf kX iS , and hence
Z
1
kPS f (x)kX 6 hkf kX iS = 0
0 kf kX dx
|S | S 0
2d
Z
6 kf kX dx 6 2d Ahkf kX iS .
|Sb0 | Sb0

These two cases confirm the upper bound kPS f kL∞ (Rd ;X) 6 2d Ahkf kX iS . 

11.1.c Maximal oscillatory norms for Lp spaces

Based on the oscillations studied above, we introduce the related John–

Strömberg maximal operator
#
M0,λ f (x) := sup oscλ (f ; Q),
Q3x

where the supremum is taken over all cubes containing x ∈ Rd ; a dyadic ver-
#,D
sion M0,λ is obtained by restricting the supremum to dyadic cubes Q ∈ D
only. Via this maximal operator we can obtain a useful oscillatory character-
isation of Lp (Rd ; X), which we will prove in the rest of this section:

Theorem 11.1.18. Let X be a Banach space, p ∈ (0, ∞), λ = 2−2−d , and

f ∈ L0 (Rd ; X). Then there is a constant c ∈ X such that f − c ∈ Lp (Rd ; X)
#
if and only if M0,λ f ∈ Lp (Rd ), and in this case

−1/p # #
cd kM0,λ f kLp (Rd ) 6 kf − ckLp (Rd ;X) 6 cp kM0,λ f kLp (Rd ) ,

where cp = 8p for p ∈ [1, ∞) and cp = 22+1/p for p ∈ (0, 1).

The result is also valid with Rd replaced by a cube Q0 ⊆ Rd or a quadrant
#
S ⊆ Rd , and with the supremum in the maximal operator M0,λ restricted to
cubes Q ⊆ Q0 or Q ⊆ S, respectively.
14 11 Singular integral operators

Remark 11.1.19. If we now a priori require that f ∈ Lp0 ,∞ (Rd ; X) for some
p0 ∈ (0, ∞) (unrelated to the exponent p), then the constant c ∈ X guaranteed
by Theorem 11.1.18 is necessarily 0, and thus in fact f ∈ Lp (Rd ; X).
Namely, if f ∈ Lp0 ,∞ (Rd ; X) and f − c ∈ Lp (Rd ; X), it follows that c =
f − (f − c) ∈ Lp0 ,∞ (Rd ; X) + Lp (Rd ; X), thus

|{kck > t}| 6 |{kf k > t/2}| + |{kf − ck > t/2}| < ∞

for all t > 0, which would lead to a contradiction for t ∈ (0, kck).
By Lemma 11.1.1 for any q ∈ (0, ∞), we have
k(f − c)1Q kLq,∞ kf 1Q kLq −1/q
Z 1/q
oscλ (f ; Q) 6 inf 1/q
6 1/q
= λ − kf kq .
c∈X (λ|Q|) (λ|Q|) Q

Taking the supremum over all cubes Q containing a given point, it follows
that
1/q
#
M0,λ f 6 λ−1/q Mq f, Mq f := M (kf kq ) , (11.7)
where M is the Hardy–Littlewood maximal operator. The Lp boundedness of
Mq is an easy combination of some estimates collected from Chapter 3:
Lemma 11.1.20. For all 0 < q < p < ∞, we have
   p 1/q
max kMq kLp →Lp , kMq kLp,∞ →Lp,∞ 6 3d/q+d/p .
p−q
Proof. The dyadic (in fact more general martingale) bounds for MqD on Lp
and Lp,∞ for p ∈ (q, ∞), with norm bound (p/(p − q))1/q in each case, have
been treated in Lemma 3.5.17. On the other hand, we recall from (3.36) that

M f 6 3d sup M α f,
α∈{0, 31 , 23 }d

thus  1/p
X
Mq f 6 3d/q sup Mqα f 6 3d/q [Mqα f ]p .
α∈{0, 31 , 23 }d
α∈{0, 13 , 23 }d

Hence
 X 1/p  p 1/q
kMq f kp 6 3d/q kMqα f kpp 6 3d/q+d/p kf kp ,
p−q
α∈{0, 13 , 23 }d

and, for every λ > 0,

X 1/p X 1/p
λ|{Mq f > λ}|1/p 6 λ |{Mqα f > 3−d/q λ}| 6 3d/q kMqα f kpLp,∞ ,
α α

after which the last step is exactly as in the strong-type case, now using the
weak-type boundedness of the dyadic Mqα instead. 
 11.1 Local oscillations of functions 15
#
Proposition 11.1.21. The operator M0,λ is bounded from Lp (Rd ; X) to
1/p
Lp (Rd ) and from Lp,∞ (Rd ; X) to Lp,∞ (Rd ), with norm at most cd,λ , where
cd,λ is a constant depending only on d and λ.

The first half of Theorem 11.1.18 is immediate from this proposition (with the
choice λ = 2−2−d so that cd,λ = cd ), combined with the trivial observation
# #
that M0,λ f = M0,λ (f − c) for any constant c ∈ X.
Proof. Let Y ∈ {Lp , Lp,∞ }. By (11.7) and Lemma 11.1.20, we have
 p 1/q
#
kM0,λ f kY 6 λ−1/q kMq f kY 6 λ−1/q 3d/q+d/p kf kY .
p−q

With, say, q = 21 p, the right hand side takes the form (λ−2 33d 22 )1/p kf kY . 
Towards the deduction of a global Lp estimate from local ones, we record:

Lemma 11.1.22. Let X be a Banach space and p ∈ (0, ∞). Suppose that
f ∈ Lploc (Rd ; X) satisfies

k1Q (f − cQ )kp 6 K

for some constants cQ ∈ X and all cubes Q ⊆ Rd . Then there is a constant

c ∈ X such that f − c ∈ Lp (Rd ; X) and

kf − ckp 6 K.

Proof.
S∞ Consider an increasing sequence of cubes Q1 ⊆ Q2 ⊆ . . . such that
d
n=1 n = R . If m 6 n, then
Q

kcQm − cQn k = |Qm |−1/p k1Qm (cQm − cQn )kp

 
6 |Qm |−1/p k1Qm (f − cQm )kp + k1Qn (f − cQn )kp
6 |Qm |−1/p 2K → 0 as m → ∞.

Hence (cQn )n>1 is a Cauchy sequence and thus convergent to some c ∈ X.

Now Fatou’s lemma shows that
Z Z Z
p p
kf − ck = lim 1Qn kf − cQn k 6 lim inf kf − cQn kp 6 K,
Rd Rd n→∞ n→∞ Qn

which completes the proof. 

We can now prove the remaining half of Theorem 11.1.18, which we restate
as:
16 11 Singular integral operators

Proposition 11.1.23. Let f ∈ L0 (Rd ; X), λ = 2−2−d , and suppose that

#
M0,λ f ∈ Lp (Rd ) for some p ∈ (0, ∞). Then there is a constant c ∈ X such
that f − c ∈ Lp (Rd ; X) and
(
# 8p, p ∈ [1, ∞),
kf − ckLp (Rd ;X) 6 cp kM0,λ f kLp (Rd ) , cp = 2+1/p
2 , p ∈ (0, 1).

The result also holds with Rd replaced by a cube Q0 ⊆ Rd or a quadrant

#
S ⊆ Rd , and with the supremum in the maximal operator M0,λ restricted to
cubes contained in Q0 or S, respectively.
Proof. Consider a fixed cube Q0 ⊆ Rd . By Lerner’s formula (Theorem
11.1.12), there is a 21 -sparse subcollection S ⊆ D(Q0 ) such that
X
1Q0 kf − mf (Q0 )k 6 4 1S oscλ (f ; S),
S∈S

whenever mf (Q0 ) is a λ-pseudomedian of f on Q0 . Taking Lp norms and

using Proposition 11.1.11 (with γ = 21 ), we get
X
k1Q0 (f − mf (Q0 ))kp 6 4 1S oscλ (f ; S)
p
S∈S
X #
6 4cp, 21 1E(S) oscλ (f ; S) 6 4cp, 12 kM0,λ f kp .
p
S∈S

This estimate is uniform with respect to the choice of Q0 ⊆ Rd ; hence we can

apply Lemma 11.1.22 with cQ = mf (Q) to complete the proof.
The variant in the case of a cube or a quadrant in place of Rd is immediate
by inspection of the argument. 
We conclude this section with an end-point analogue of Theorem 11.1.18 for
the space BMO(Rd ; X) in place of Lp (Rd ; X). Recall that we have previously
defined the space BMO(Rd ; X) of functions of bounded mean oscillation as
the class of functions f ∈ L1loc (Rd ; X) such that
Z
kf kBMO(Rd ;X) := sup inf − kf − ck < ∞.
Q c∈X Q

Proposition 11.1.24. Let X be a Banach space, λ = 2−2−d , and f ∈

#
L0 (Rd ; X). Then f ∈ BMO(Rd ; X) if and only if M0,λ f ∈ L∞ (Rd ), and
# #
λkM0,λ f kL∞ (Rd ) 6 kf kBMO(Rd ;X) 6 8kM0,λ f k∞ .

Proof. From Lemma 11.1.1 it is immediate that

Z
1
oscλ (f ; Q) 6 inf − kf − ck,
λ c∈X Q
 11.1 Local oscillations of functions 17

from which the first claimed inequality follows by taking the supremum over
all cubes Q ⊆ Rd .
In the other direction, given a cube Q ⊆ Rd , Lerner’s formula (Theorem
11.1.12) guarantees that
Z
4 X
− kf − mf (Q)k 6 |S| oscλ (f ; S)
Q |Q|
S∈S
4 X # #
6 2|E(S)|kM0,λ f k∞ 6 8kM0,λ f k∞ ,
|Q|
S∈S

and taking the supremum over all cubes Q proves the second bound. 

11.1.d The dyadic Hardy space and BMO

Often an efficient way of capturing the relevant local oscillations of a function
is in terms of the following notion:
Definition 11.1.25 (Atom). A function a : Rd → X is called a (normalised)
1
HD -atom if
(i) supp a ⊆ Q for some Q ∈ D;
∞ d
R ∈ L (R ; X) (and kak∞ 6 1/|Q|);
(ii) a
(iii) Q a = 0.
It is immediate that a normalised atom satisfies kak1 6 1. If a 6= 0 is an atom
a
supported on Q ∈ D, then is a normalised atom. Out of these atoms
|Q|kak∞
we can then construct a useful subspace of L1 (Rd ; X):
Definition 11.1.26 (Atomic Hardy space). The atomic Hardy space
1
HD,at (Rd ; X)
consists of all f ∈ L1 (Rd ; X) that admit a representation
∞
X  X∞ 
f= αk = λk a k ,
k=1 k=1
1 d 1
absolutely convergent in L (R ; X), where each αk is an HD -atom supported
in some Qk ∈ D (or each ak is a normalised HD -atom and λk ∈ K) with
1

∞
X ∞
X 
kαk k∞ |Qk | < ∞ |λk | < ∞ .
k=1 k=1

The norm in this space is defined as

∞
X  ∞
X 
kf kHD,at
1 := inf kαk k∞ |Qk | = inf |λk |
k=1 k=1

where the infimum is taken over all such representations.

18 11 Singular integral operators

It is immediate that the two versions of the definition are equivalent via the
correspondence λk = kαk k∞ |Qk | and ak = λ−1k αk .
A disadvantage of this definition is the difficulty of checking the mem-
1
bership of a given function in HD,at (Rd ; X), as doing this via the definition
would require one to construct the atomic decomposition, which might not be
an easy task. The following notion is much more amenable to this:
Definition 11.1.27 (Maximal Hardy space). The maximal Hardy space
1
HD,max (Rd ; X)

consists of all f ∈ L1 (Rd ; X) for which also the (cancellative) dyadic maximal
function
MD f (x) := sup 1Q (x)khf iQ kX
Q∈D
1 d
satisfies MD f ∈ L (R ). The norm in this space is defined as
kf kHD,max
1 := kMD f kL1 (Rd ) .
1
Theorem 11.1.28. Let X be a Banach space. The spaces HD,at (Rd ; X) and
1 d
HD,max (R ; X) are equal with equivalent norms; in fact
d
khkHD,max
1 (Rd ;X) 6 khkHD,at
1 (Rd ;X) 6 6 · 2 · khkHD,max
1 (Rd ;X) .

Proof. Suppose
R first that a ∈ L∞ (Rd ; X) satisfies supp a ⊆ Q for some dyadic
cube and a = 0. Then haiR 6= 0 only if R ( Q, and hence supp MD a ⊆ Q
as well. It follows that
kMD ak1 6 |Q|kMD ak∞ 6 |Q|kak∞ .
P∞
If h = i=1 ai is a series of such function on cubes Qi , then by sublinearity
∞
X ∞
X
khkHD,max
1 (Rd ;E) = kMD hk1 6 kMD ai k1 6 |Qi |kai k∞ ,
i=1 i=1

and taking the infimum over all such representations of h shows that
khkHD,max
1 (Rd ;X) 6 khkHD,at
1 (Rd ;X) .

1
In the other direction, suppose that h ∈ Hmax (Rd ; X). Given λ > 0, let Qλ
be the collection of maximal dyadic cubes Q such that khhiQ kX > λ. Then
X 1 1
|Q| = |{MD h > λ}| 6 kMD hkL1 (Rd ) = khkHD,max
1 (Rd ;X) .
λ λ
Q∈Qλ

Let Qbλ be the collection of maximal dyadic cubes that have a child in Qλ .
Thus these cubes do not belong to Qλ themselves. Hence khhiQ kE 6 λ for
Q ∈ Qbλ , and also
 11.1 Local oscillations of functions 19
X X X
|Q| 6 |Q|
b = 2d |Q| = 2d |{M h > λ}|.
Q∈Q
bλ Q∈Qλ Q∈Qλ

Let then
X X
gλ := 1{(S Qbλ ) h + 1Q hhiQ , bλ := 1Q (h − hhiQ ).
Q∈Q
bλ Q∈Q
bλ

By definition of MD , we have khhiQ kX 6 MD h(x) whenever x ∈ Q ∈ D.

As `(Q) → 0, this gives kf (x)kX 6 MD h(x) at a.e. x by the Lebesgue Differ-
entiation Theorem. Thus kgλ (x)k6 MD h(x) almost everywhere. On the other
hand, we have khhiQ kX 6 λ for Q ∈ Qbλ , and MD h(x) 6 λ for x ∈ {( Qbλ );
S
thus in fact kgλ kX 6 min(λ, MD h) almost everywhere, where MD h ∈ L1 (Rd ).
Moreover, gλ = h on {MD h 6 λ} → Rd as λ → ∞, and hence
(
h, λ → ∞,
gλ →
0, λ → 0,

pointwise, and by dominated convergence also in L1 (Rd ; X). Thus

X X
h= (g2k+1 − g2k ) = (b2k − b2k+1 )
k∈Z k∈Z
X X X 
= 1Q (h − hhiQ ) − 1R (h − hhiR )
k∈Z Q∈Q
bk R∈Q
b k+1
2 2
X X  X 
= 1Q\S Qb k+1 (h − hhiQ ) + 1R (hhiR − hhiQ )
2
k∈Z Q∈Q R∈Q
2k
b b k+1
2
R⊆Q
X X
=: ak,Q .
k∈Z Q∈Q
2k
b

aQ = 0 and kaQ k∞ 6 2k+1 + 2k = 3 · 2k . Hence

R
Here supp aQ ⊆ Q,
X X X X
khkHD,at
1 6 |Q|kak,Q k∞ 6 3 · 2k |Q|
k∈Z Q∈Q k∈Z Q∈Q
2k
b bk
2
X
6 3 · 2k · 2d |{MD h > 2k }|
k∈Z
X Z 2k
d
6 3·2·2 |{MD h > t}| dt
k∈Z 2k−1

= 6 · 2d kMD hkL1 (Rd ) = 6 · 2d khkHD,max

1 (Rd ;X) .


1
Corollary 11.1.29. The space HD,at (Rd ; X) = HD,max
1
(Rd ; X) is complete.
20 11 Singular integral operators

1
Proof. It in enough to prove this for HD,max (Rd ; X). Since kf (x)kX 6
d
MD f (x) at a.e. x ∈ R , we have kf kL1 (Rd ;X) 6 kf kHD,max
1 (Rd ;X) . Hence, if
∞ 1
(fn )n=1 is a Cauchy sequence in HD,max (Rd ; X), it is also a Cauchy sequence
in L (R ; X) and thus kfn − f k1 → 0 for some f ∈ L1 (Rd ; X). Since h iQ is
1 d

continuous from L1 (Rd ; X) to X, we have for all x ∈ Q ∈ D we have, for each

1
h ∈ HD,max (Rd ; X),

khf − hiQ kX = lim khfn − hiQ kX 6 lim inf MD (fn − h)(x);

n→∞ n→∞

hence MD (f −h)(x) 6 lim inf n→∞ MD (fn −h)(x), and thus by Fatou’s lemma

kMD (f − h)kL1 (Rd ) 6 lim inf kMD (fn − h)kL1 (Rd ) .

n→∞

1
With h = 0, this shows that f ∈ HD,max (Rd ; X). With h = fm , we find that

lim kMD (f − fm )kL1 (Rd ) 6 lim lim inf kMD (fn − fm )kL1 (Rd ) = 0,
m→∞ m→∞ n→∞

1
and hence fm → f in HD,max (Rd ; X). 

Theorem 11.1.30. Let X be a Banach space. The duality

Z ∞ Z
X n N o
hb, hi := lim hbN , hi = hb, ai i, bN := min 1, b
N →∞
i=1
kbkX ∗

between b ∈ BMOD (Rd ; X ∗ ) and h ∈ HD,at

1
(Rd ; X) is well defined, and realises
d ∗
BMOD (R ; X ) with the norm
Z
kbkBMOD (Rd ;X) := sup inf − kb − ckX
Q∈D c∈X Q

1
as an isometric subspace of (HD,at (Rd ; X))∗ .

Proof. Since all norms BMO norms appearing in this proof are dyadic, we
drop the subscript D for the benefit of slightly lighter notation.

Part 1: Estimating the dual norm by the BMO norm

If supp ai ⊆ Qi ∈ D and ai = 0, we have

R

Z Z Z
hb, ai i = hb − c, ai i 6 − kb − ckX ∗ |Qi |kai k∞
Qi Qi

for all c ∈ E ∗ . Taking the infimum over c ∈ E ∗ it follows that

Z
hb, ai i 6 kbkBMO |Qi |kai k∞
 11.1 Local oscillations of functions 21
P∞ R d ∗
P∞
and hence i=1 hb, ai i converges for b ∈ BMO(R ; E ) and i=1 ai ∈
1
Hat (Rd ; E).
One checks that kbN − cN kX ∗ 6 2kb − ckX ∗ , whence
Z Z Z
inf∗ − kbN − ckE ∗ 6 inf∗ − kbN − cN kE ∗ 6 2 inf∗ − kb − ckE ∗ ,
c∈E Q c∈E Q c∈E Q

so that bN ∈ (BMO ∩L∞ )(Rd ; X ∗ ) and

Z
hbN , ai i 6 kbN kBMO |Qi |kai k∞ 6 2kbkBMO |Qi |kai k∞ .

Thus
∞ Z
X ∞
X Z ∞ Z
X Z
hb, ai i = lim hbN , ai i = lim hbN , ai i = lim hbN , hi,
N →∞ N →∞ N →∞
i=1 i=1 i=1

where the first two identities use dominated convergence in L1 (Qi ) and in `1 ,
respectively, and the last one follows from the convergence of the series h =
P ∞ 1 d ∞ d ∗ 1 d ∗
i=1 ai in L (R ; E), and the fact that bN ∈ L (R ; E ) ⊆ (L (R ; E)) .
This shows in particular that the pairing of hb, hi is independent of the par-
ticular series representation of h, and hence well defined. Taking the infimum
over all representations in the estimate
∞
X
|hb, hi| 6 kbkBMO |Qi |kai k∞ ,
i=1

we find that
kbk(Hat
1 (Rd ;X))∗ 6 kbkBMO(Rd ;X ∗ ) . (11.8)

Part 2: Estimating the BMO norm by the dual norm

For the converse estimate, consider a cube Q and suppose first that s ∈
L1 (Q; X ∗ ) is a simple function, thus measurable with respect to a finite σ-
algebra F of Q. The advantage of this setting is that, for a finite σ-algebra,
0
we have the duality (Lp (F ; X))∗ = Lp (F ; X ∗ ) for an arbitrary Banach space
X and for every p ∈ [1, ∞], including in particular p = ∞. Now inf c∈E ∗ ks −
ckL1 (Q;X ∗ ) is the norm of the equivalence class [s] ∈ L1 (F ; X ∗ )/X ∗ , where
L1 (F ; X ∗ ) = (L∞ (F ; X))∗ .
We claim that the quotient space above is the dual of the subspace
L0∞ (F ; X) ⊆ L∞ (F ; X) of functions with mean zero. In fact, recall from
Proposition B.1.4 that for any subspace Y ⊆ Z, we have the identification
Y ∗ = Z ∗ /Y ⊥ , the quotient of Z ∗ with the annihilator Y ⊥ of Y in Z ∗ . Now
Z = L∞ (F ; X) for a finite σ-algebra F , in which case Z ∗ = L1 (F ; X ∗ ). To
identify Y ⊥ for Y = L0∞ (F R ; X), it is easy to check that the only functions
f ∈ L1 (F ; X ∗ ) for which hf, gi = 0 for all g ∈ L∞0 (F ; X) are the constant
functions. Thus indeed L1 (F ; X ∗ )/X ∗ = (L0∞ (F ; X))∗ , and hence
22 11 Singular integral operators
Z
inf ks − ckL1 (Q;X ∗ ) = k[s]kL1 (F ;X ∗ )/X ∗ = sup hs, gi .
c∈X ∗ g∈L∞0 (F ;X)
kgk∞ 61

Now, given b ∈ BMO(Rd ; X ∗ ) and a cube Q, we choose a simple s ∈

L1 (Q; X ∗ ) such that kb − skL1 (Q;X ∗ ) 6 ε. Then

inf kb − ckL1 (Q;X ∗ ) 6 inf ∗ ks − ckL1 (Q;X ∗ ) + ε

c∈X ∗ c∈X
Z Z
6 sup hs, gi + ε 6 sup hb, gi + 2ε.
g∈L0∞ (Q;X) g∈L0∞ (Q;X)
kgk∞ 61 kgk∞ 61

But each g ∈ L0∞ (Q; X) is an HD 1

-atom, and hence
Z
hb, gi 6 kbk(HD,at
1 (Rd ;X))∗ kgkHD,at
1 (Rd ;X) 6 kbk(HD,at
1 (Rd ;X))∗ kgk∞ |Q|.

Dividing by |Q| and letting ε → 0, we obtain

Z
inf ∗ − kb − ckX ∗ 6 kbk(HD,at
1 (Rd ;X))∗ ,
c∈X Q

and hence the estimate converse to (11.8). 

11.2 Singular integrals and extrapolation of Lp0 bounds

In this section we study a fairly broad class of kernels satisfying a relatively
general integrability condition first introduced by Hörmander. Nevertheless,
this condition turns out to be strong enough to yield a fundamental extrapo-
lation property of singular integral operators: once bounded on one Lp0 space,
they remain bounded on the full scale of Lp spaces for p ∈ (1, ∞), together
with appropriate end-point estimates for p = 1 and p = ∞.
The precise classes of kernels relevant are described in the following defi-
nition. We recall that Ṙ2d = R2d \ {(t, t) : t ∈ Rd }.

Definition 11.2.1. Let X and Y be Banach spaces, p0 ∈ [1, ∞], and consider

K : Ṙ2d → L (X, Y ), T ∈ L (Lp0 (Rd ; X), Lp0 ,∞ (Rd ; Y )).

(1) We say that T has kernel K, or that K is the kernel of T , if for ev-
ery f ∈ Lpc 0 (Rd ; X) and almost every s at a positive distance from
supp f the following holds: for every functional y ∗ ∈ Y ∗ , the function
t 7→ hK(s, t)f (t), y ∗ i is integrable, and
Z
hT f (s), y i = hK(s, t)f (t), y ∗ i dt.
∗
 11.2 Singular integrals and extrapolation of Lp0 bounds 23

(2) We say that K is a Hörmander (resp. operator-Hörmander) kernel, or

satisfies the Hörmander (resp. operator-Hörmander) condition, if the fol-
lowing estimate holds for all x ∈ X and t, t0 ∈ Rd with a fixed constant c
independent of these quantities:
Z
k[K(s, t) − K(s, t0 )]xkY ds 6 ckxkX
|s−t|>2|t−t0 |
 Z  (11.9)
0
resp. kK(s, t) − K(s, t )kL (X,Y ) dx 6 c .
|s−t|>2|t−t0 |

The smallest admissible c is denoted by kKkHör (resp. kKkHörop ).

(3) We say that K is a dual Hörmander (resp. dual operator-Hörmander)
kernel, or satisfies the dual Hörmander (resp. dual operator-Hörmander)
condition, if the following estimate holds for every y ∗ ∈ Y ∗ and s, s0 ∈ Rd
with a fixed constant c0 independent of these quantities:
Z
k[K(s, t)∗ − K(s0 , t)∗ ]y ∗ kX ∗ dt 6 c0 ky ∗ kY ∗
|t−s|>2|s−s0 |
 Z  (11.10)
resp. kK(s, t) − K(s0 , t)kL (X,Y ) dt 6 c0
|t−s|>2|s−s0 |

The smallest admissible c0 is denoted by kKkHör∗ (resp. kKkHör∗op ).

(4) If Q ⊆ Rd is a cube or a quadrant, we make analogous definitions with
each occurrence of Rd replaced by Q; in particular, with Ṙ2d by {(s, t) ∈
Q × Q : s 6= t}, and the integrals extended over Q only, while keeping the
other integrations conditions in force. In this situation, we say that K is
a (dual/operator) Hörmander kernel on Q, respectively.
Remark 11.2.2. If K is a (dual/operator) Hörmander kernel, then its restric-
tion to {(s, t) ∈ Q × Q : s 6= t} is a (dual/operator) Hörmander kernel on Q.
Example 11.2.3. A kernel K(x, y) that only depends on the difference x − y,
i.e., K(x − y) = k(x − y) for some function k, is called a convolution kernel.
For such kernels, after simple changes of variables, the Hörmander and dual
Hörmander conditions take the forms
Z
k[k(s − t) − k(s)]xkY ds 6 ckxkX ,
|s|>2|t|
Z
k[k(s − t)∗ − k(s)∗ ]y ∗ kX ∗ ds 6 c0 ky ∗ kY ∗ ,
|s|>2|t|

and similar reformulations of the operator Hörmander conditions are obvious.

The role of these conditions in the extrapolation of Lp -boundedness is sum-
marised in the next theorem. Before stating the result, we make a remark
concerning the extension of the action of operators from Lp0 (Rd ; X) to
L∞ (Rd ; X). An inherent obstacle here is that the intersection Lp0 (Rd ; X) ∩
L∞ (Rd ; X) is not dense in L∞ (Rd ; X). As a substitute we have:
24 11 Singular integral operators

Lemma 11.2.4. Let X be a Banach space. The closure of Lp (Rd ; X) ∩

L∞ (Rd ; X) in L∞ (Rd ; X) is independent of p ∈ (0, ∞), and it coincides with

∞ L∞ (Rd ;X)
L̄fin (Rd ; X) := L∞ d
fin (R ; X) , where
L∞ d
fin (R ; X)
∞ d
:= {f ∈ L (R ; X) : |{f 6= 0}| < ∞}.

Proof. It is clear that L∞ d p d ∞ d

fin (R ; X) ⊆ L (R ; X) ∩ L (R ; X), and taking the
closures of both sides proves one side of the claim.
Conversely, let p ∈ (0, ∞), a function f ∈ Lp (Rd ; X) ∩ L∞ (Rd ; X), and
ε > 0 be given. Now

Fε := {kf (·)kX > ε} 6 ε−p kf kpLp (Rd ;X) < ∞,

and hence fε := 1Fε f ∈ L∞ d

fin (R ; X). On the other hand, it is clear that

kf − fε kL∞ (Rd ;X) = k1{Fε f kL∞ (Rd ;X) 6 ε.

Since this can be done for any ε > 0, we find that f belongs to the L∞ (Rd ; X)-
closure of L∞ d p d ∞ d
fin (R ; X). Since f ∈ L (R ; X) ∩ L (R ; X), this whole intersec-
tion belongs to the said closure, and then so does the closure of this intersec-
tion. This completes the proof. 

Theorem 11.2.5 (Calderón–Zygmund). Let X and Y be Banach spaces

and p0 ∈ [1, ∞]. Let

T ∈ L (Lp0 (Rd ; X), Lp0 ,∞ (Rd ; Y ))

(where L∞,∞ := L∞ ) with norm A0 := kT kL (Lp0 (Rd ;X),Lp0 ,∞ (Rd ;Y )) .

(1) If T has a Hörmander kernel K, then
(a) T extends uniquely to T ∈ L (Lp (Rd ; X), Lp (Rd ; Y )) for all p ∈ (1, p0 ),
and
 p0 − 1 1/p
kT kL (Lp (Rd ;X),Lp (Rd ;Y )) 6 cd (A0 + kKkHör );
(p0 − p)(p − 1)

(b) T extends uniquely to T ∈ L (L1 (Rd ; X), L1,∞ (Rd ; Y )) and

kT kL (L1 (Rd ;X),L1,∞ (Rd ;Y )) 6 cd (A0 + kKkHör ).

(2) If T has a dual Hörmander kernel K, then

(a) T extends uniquely to T ∈ L (Lp (Rd ; X), Lp (Rd ; Y )) for all p ∈
(p0 , ∞), and
 p 1/p
0
kT kL (Lp (Rd ;X),Lp (Rd ;Y )) 6 cd p (A0 + kKkHör∗ );
p − p0
 11.2 Singular integrals and extrapolation of Lp0 bounds 25

(b) T extends uniquely to T ∈ L (L̄fin∞

(Rd ; X), BMO(Rd ; Y )), where the
∞ d
space L̄fin (R ; X) is as in Lemma 11.2.4, and
∞ (Rd ;X),BMO(Rd ;Y )) 6 cd (A0 + kKkHör )kf kL∞ (Rd ;X)
kT kL (L̄fin
for all f in this space.
(3) If T has a kernel K that satisfies both the Hörmander and the dual
Hörmander conditions, then for all p ∈ (1, ∞), T extends uniquely to
T ∈ L (Lp (Rd ; X), Lp (Rd ; Y )), and
kT kL (Lp (Rd ;X),Lp (Rd ;Y )) 6 cd · pp0 · (A0 + kKkHör + kKkHör∗ ).
(4) All claims remain valid when Rd is replaced either by a cube or a quad-
rant throughout. In this case, it suffices to relax the Hörmander conditions
accordingly, as in Definition 11.2.1(4).
The rest of this section is dedicated to a case-by-case proof of the different as-
sertions of Theorem 11.2.5. For the proof of (1), we introduce the fundamental
Calderón–Zygmund decomposition in Proposition 11.2.6. The proof of (2), in
turn, depends on the notion of local oscillations developed in Section 11.1. The
result of (2b) does not directly allow the extension of T to all of L∞ (Rd ; X)
since Lp0 (Rd ; X) ∩ L∞ (Rd ; X) is not dense in this space; see Theorem 11.2.9
for results in this direction. The proof of (3) is essentially a combination of
(1) and (2), but note that this case provides additional information about
p = p0 (bootstrapping the initial weak-type bound into a strong-type one)
and improves the quantitative estimates for p close to p0 , where the bounds
provided by (1) and (2) blow up as p → p0 . Finally, the claims (4) will be dealt
with by indicating the relevant modifications in the proofs of (1) through (3).
As it turns out, these modifications are fairly minor, although in the case of
(1) they might not be entirely obvious.

11.2.a Calderón–Zygmund decomposition and case p ∈ (1, p0 )

The key to extrapolating in this range is the following classical result:
Proposition 11.2.6 (Calderón–Zygmund decomposition). Let X be a
Banach space. Given f ∈ L1 (Rd ; X) and λ > 0, there exists a decomposition
f = g + b, where
kgk∞ 6 2d λ, kgk1 6 kf k1 ,
P
and b = i bi , where
Z X 1 X
supp bi ⊆ Qi , bi = 0, |Qi | 6 kf k1 , kbi k1 6 2kf k1
i
λ i

for some disjoint dyadic cubes Qi . If f is simple, then all biR are also simple.
If f ∈ L1 (Q0 ; X) for some cube Q0 ⊆ Rd and λ > 2−d −Q0 kf k, then the
cubes Qi can be chosen as dyadic subcubes of the initial Q0 , and the function
g to be supported on Q0 .
If f ∈ L1 (S; X) for some quadrant of Rd , then we have Qi ⊆ S.
26 11 Singular integral operators

Proof. Let Qi ∈ D be the maximal dyadic cubes such that

R
−
Qi
kf k > λ. Then
they are pairwise disjoint, and
X 1
|Qi | = |{MD f > λ}| 6 kf k1 .
i
λ

We define bi := 1Qi (f − hf iQi ) (which is clearly simple if f is), whence the

first two properties of bi are clear, and it remains to estimate
X X X Z
kbi k1 6 (k1Qi f k1 + |Qi |khf iQi k) 6 2 kf k 6 2kf k1
i i i Qi

by the disjointness of the cubes. To ensure that f = g + b, we must then define

X
g := 1{(Si Qi ) f + 1Qi hf iQi ,
i
S
where the terms areR disjointly supported. If x ∈ {( i Qi ), then all dyadic
cubes Q 3 x satisfy −Q |f | 6 λ, and thus
Z
kg(x)k = kf (x)k = lim − kf k 6 λ
Q3x Q
`(Q)→0

at almost every such x by the Lebesgue Differentiation Theorem 2.3.4 (or in

fact just the scalar-valued version, since we apply it to the function kf (·)k
rather than f itself). On the other hand, the maximality of Qi implies that
R
its dyadic parent Qb i satisfies the opposite inequality, −b |f | 6 λ. Thus
Qi

|Q
bi |
Z Z
1 1
kg(x)kX = khf iQi kX 6 kf kX 6 · kf kX 6 2d · λ
|Qi | Qi |Qi | |Qbi | Q
bi

for x ∈ Qi , and we see that kg(x)k 6 2d λ in both cases. Moreover,

Z X Z XZ
kgk1 = S
kf k + |Qi |khf iQi k 6 S
kf k + kf k = kf k1
{( i Qi ) i {( i Qi ) i Qi

by the disjointness of the cubes.R

If f ∈ L1 (Q0 ; X) and λ > −Q0 kf k, then the maximal dyadic subcubes
R
Qi of Q0 with −Qi kf k > λ, are necessarily strict subcubes of Q0 , and the
same proof produces a decomposition with the claimed additional properties.
If λ ∈ [2−d , 1)−Q0 kf k, then we let the family {Qi }i consist of the initial cube
R

Q0 only, so that g := P hf iQ0 1Q0 and b := (f − hf iQ0 )1Q0 . Then kgk∞ =

khf iQ0 k 6 2d λ and i |Qi | = |Q0 | 6 λ−1 kf k1 by the two assumed bounds
on λ. The last claim of the theorem is obvious. 
We can now give:
 11.2 Singular integrals and extrapolation of Lp0 bounds 27

Proof of Theorem 11.2.5(1). Our plan is to first prove the weak-type result
(1b), and then obtain the strong-type bound (1a) via the Marcinkiewicz In-
terpolation Theorem 2.2.3.
For f ∈ Lp0 (Rd ; X) ∩ L1 (Rd ; X) and λ > 0, we estimate λ|{kT f k > λ}|.
Let f = g + b the Calderón–Zygmund decomposition of f at level αλ
(instead of λ), where α is to be determined. Then
1/p0 1/p0 0 1/p0
kgkp0 6 kgk∞ 0 kgk1 6 (2d αλ)1/p0 kf k1 ,
p0 d p0 d
P in particular g ∈ L (R ; X), and thus b = f − g ∈ L (R ; X). Since b =
so
i bi and the bi are disjointly supported, it follows that each bi also belongs to
Lp0 (Rd ; X) and the identity b = i bi also holds in the sense of convergence
P
in Lp0 (Rd ; X). The assumption that T ∈ L (Lp0 (Rd ; X), Lp0 ,∞ (Rd ; Y )) then
implies that
X X
T f = T (g + b) = T g + T b, Tb = T bi = T bi .
i i

If Qi are the corresponding cubes, let Bi be the concentric ball of twice

the diameter and O∗ := i Bi . Then
S

|{kT f k > λ}| 6 |{kT gk > λ/2}| + |{kT bk > λ/2} \ O∗ | + |O∗ |, (11.11)

where the last term satisfies

X X cd
|O∗ | 6 |Bi | = cd |Qi | 6 kf k1 .
i i
αλ

For the middle term, we have

kT bk
Z Z Z
2X 2X
|{kT bk > λ/2} \ O∗ | 6 6 kT bi k 6 kT bi k.
{O ∗ λ/2 λ i {O∗ λ i {Bi

In order to estimate the ith term here, we denote by zi the common centre
of the cube Qi and the ball Bi . Now the integral representation of T bi (s) is
available at s ∈ {Bi . Explicitly, for each y ∗ ∈ Y ∗ ,
Z Z
hT bi (s), y ∗ i = hK(s, t)bi (t), y ∗ i dt = h[K(s, t) − K(s, zi )]bi (t), y ∗ i dt,
R
where the last step follows from the fact that bi (t) dt = 0. Thus
Z
kT bi (s)kY 6 k[K(s, t) − K(s, zi )]bi (t)kY dt
Qi

and hence
Z Z Z
kT bi (s)kY ds 6 k[K(s, t) − K(s, zi )]bi (t)kY ds dt
{Bi Qi {Bi
28 11 Singular integral operators
Z
6 kKkHör kbi (t)kX dt,
Qi

since |s − zi | > 2 diam(Qi ) > 2|t − zi | for s ∈ {Bi and t ∈ Qi . Substituting

back, it follows that
Z XZ
2X 2 2 4
kT bi k 6 kKkHör kbi k = kKkHör kbk1 6 kKkHör kf k1 .
λ i {Bi λ i Qi λ λ

It remains to estimate |{kT gk > λ/2}|. If p0 < ∞, we have

Ap00 2p0 p0
|{kT gk > λ/2}| 6 kgkp0
p 6 A · (2d αλ)p0 −1 kf k1 ,
(λ/2)p0 0
λp 0 0
so that altogether
 (2A · 2d α)p0 cd  kf k1
0
|{kT f k > λ}| 6 d
+ 4kKkHör + ,
2 α α λ
where we are still free to choose α > 0. Taking

α = 2−d−1 /A0 (11.12)

leads to
kf k1
|{kT f k > λ}| 6 (cd A0 + 4kKkHör ) . (11.13)
λ
If p0 = ∞, we observe that kT gk∞ 6 A0 kgk∞ 6 A0 2d αλ, so that the
same choice of α guarantees that |{kT gk > λ/2}| = 0. Thus, in this case, we
only need to estimate the last two terms in (11.11), and these have exactly
the same bounds in the case p0 < ∞ that was already handled.
We have hence confirmed (11.13) for all f ∈ Lp0 (Rd ; X) ∩ L1 (Rd ; X)
and λ > 0, and this proves the existence of a unique bounded extension
T ∈ L (L1 (Rd ; X), L1,∞ (Rd ; X)) by the density of Lp0 (Rd ; X) ∩ L1 (Rd ; X) in
L1 (Rd ; X). This completes the proof of (1b).
(1b) in case (4): Let then Rd be replaced by a cube Q0 . Note that

kT f kL1,∞ (Q0 ;Y ) := sup λ|Q0 ∩ {|T f | > λ}|.

λ>0
R
If λ 6 2A0 −Q0 kf k, then
Z
λ|Q0 ∩ {|T f | > λ}| 6 2A0 − kf k × |Q0 | = 2A0 kf k1 (11.14)
Q0
R
If λ > 2A0 −Q0 kf k and α is as in (11.12), then
Z
αλ > 2−d − kf k
Q0
 11.2 Singular integrals and extrapolation of Lp0 bounds 29

is in the admissible range to have Calderón–Zygmund decomposition at level

αλ fully localised within the cube Q0 (Proposition 11.2.6). Thus, the earlier
argument for the full space Rd localises to Q0 to produce the same conclusion
(11.13), but with the integral defining kKkHör restricted to Q0 only. A com-
bination with (11.14) shows that this estimate holds for all λ > 0, and hence
we have the desired weak-type bound on Q0 .
The case of a quadrant S is an immediate variant of the case of Rd , since
Proposition 11.2.6 guarantees that the Calderón–Zygmund decomposition is
localised to this quadrant for all values of the level parameter.
(1a): A direct application of Marcinkiewicz Interpolation Theorem 2.2.3 (with
1 in place of p0 , and p0 in place of p1 ) shows that
 c (A + kKk ) 1−θ  A θ
d 0 Hör 0
kT kL (Lp (Rd ;X),Lp (Rd ;Y ) 6 c(θ, 1, p0 ) ,
1−θ θ
where θ ∈ (0, 1) is such that 1/p = (1 − θ)/1 + θ/p0 ,
n p−1 p0 − p o p1
c(θ, 1, p0 ) = p0p0 −1
(p0 − p)(p − 1)
1
if p0 ∈ (1, ∞), and c(θ, 1, ∞) = (p − 1)− p . By the arithmetic–geometric mean
inequality, we have
 1 1−θ  1 θ 1 1
61−θ + θ = 2, (11.15)
1−θ θ 1−θ θ
1

and by elementary calculus one verifies that p0p0 −1 6 e for p0 ∈ (1, ∞). Sub-
stituting these estimates, we obtain
n p0 − p o p1
kT kL (Lp (Rd ;X),Lp (Rd ;Y ) 6 2e · cd · (A0 + kKkHör ),
(p0 − p)(p − 1)
which coincides with the claim after redefining cd . Since the Marcinkiewicz
Interpolation Theorem 2.2.3 is valid for general measure spaces, the same
argument applies equally well in the case of a cube or a quadrant as the
underlying domain. 

11.2.b Local oscillations of T f and case p ∈ (p0 , ∞)

We next turn to the study of extrapolation of the boundedness to p > p0 ,

which will involve the dual Hörmander condition. A reader familiar with
the scalar-valued counterpart of the theory might expect a duality argu-
ment at this point. While this might not be strictly out of question here,
either, one should note that at least some number of technicalities would
have to be tackled by such an approach. To begin with, the adjoint of
T ∈ L (Lp (Rd ; X), Lp (Rd ; Y )) would be an operator
30 11 Singular integral operators

T ∗ ∈ L (Lp (Rd ; Y )∗ , Lp (Rd ; X)∗ ),

0
where each Lp (Rd ; Z)∗ is in general a larger space than Lp (Rd ; Z ∗ ), unless
additional assumptions are imposed on Z ∗ (see Section 1.3). Rather than
dwelling into such issues, we prefer a direct approach within the original spaces
of X and Y valued functions that we are interested in.
We still need to settle a technical issue about the validity of the integral
representation of T f (x) for certain non-compactly supported functions f :

Lemma 11.2.7. Let X and Y be Banach spaces and p0 ∈ [1, ∞]. Let T ∈
L (Lp0 (Rd ; X), Lp0 ,∞ (Rd ; Y )) be an operator with dual Hörmander kernel K.
If B ⊆ Rd is a ball and f ∈ Lp0 (Rd ; X) ∩ L∞ (Rd ; X) is supported in {B, then
for almost all s, s0 ∈ 21 B, we have
Z
hT f (s) − T f (s0 ), y ∗ i = h[K(s, t) − K(s0 , t)]f (t), y ∗ i dt ∀y ∗ ∈ Y ∗ .
{B

Proof.
S∞ Consider an increasing sequence of balls B1 ⊆ B2 ⊆ . . . such that
d
n=1 n = R , and let fn := 1Bn f . Since fn = 1{B fn is compactly supported
B
away from B, for almost every s ∈ 21 B we have
Z
∗
hT fn (s), y i = hK(s, t)fn (t), y ∗ i dt ∀y ∗ ∈ Y ∗ .
{B

Thus, for almost every s, s0 ∈ 21 B, the following holds for every y ∗ ∈ Y ∗ :

Z
0 ∗
hT fn (s) − T fn (s ), y i = hfn (t), [K(s, t)∗ − K(s0 , t)∗ ]y ∗ i dt. (11.16)
{B

Now consider the limit n → ∞. Since fn → f in Lp0 (Rd ; X) and T ∈

L (Lp0 (Rd ; X), Lp0 ,∞ (Rd ; Y )), we have T fn → T f in Lp0 ,∞ (Rd ; Y ). Hence
a subsequence, which we keep denoting simply by fn , also satisfies T fn (s) →
T f (s) at almost every s ∈ 21 B. This means that

LHS(11.16) → hT f (s) − T f (s0 ), y ∗ i.

It is also clear that fn (t) → f (t) pointwise. On the other hand, the integrand
in (11.16) is pointwise dominated by

(k[K(s, t)∗ − K(zB , t)∗ ]y ∗ kY ∗ + k[K(s0 , t)∗ − K(zB , t)∗ ]y ∗ kY ∗ )kf k∞ ,

which is integrable over t ∈ {B (thus |t−zB | > rB > 2 max{|s−zB |, |s0 −zB |})
by the dual Hörmander condition. Hence
Z
RHS(11.16) → hf (t), [K(s, t)∗ − K(s0 , t)∗ ]y ∗ i dt
{B

by dominated convergence. The equality of the limits is what we claimed. 

 11.2 Singular integrals and extrapolation of Lp0 bounds 31

Recall the John–Strömberg maximal function and the local oscillations

#
M0,λ f (x) = sup oscλ (f ; Q), oscλ (f ; Q) := inf inf k(f − c)1Q\E k∞ .
Q3x c∈X |E|6λ|Q|

The following lemma contains the technical core of the upper extrapolation:

Lemma 11.2.8. Under the assumptions of Theorem 11.2.5(2), for all f ∈

Lp0 (Rd ; X) ∩ L∞ (Rd ; X) we have
# 1/p
kM0,λ (T f )k∞ 6 (cd,λ 0 A0 + 2kKkHör∗ )kf k∞ .

If Rd is replaced by a cube Q0 ⊆ Rd or a quadrant S ⊆ Rd , the conclusion

remains valid with the following modifications:
#
(a) in the maximal operator M0,λ , the supremum is restricted to cubes Q con-
tained in the initial cube Q0 or the quadrant S;
(b) in the Hörmander norm kKkHör∗ , the variables and the integrals are again
restricted to Q0 or S.

Proof. Let f ∈ Lp0 (Rd ; X) ∩ L∞ (Rd ; X) and let Q ⊆ Rd be a cube. Let B be

a ball with the same centre and three time the diameter. We decompose

T f = T (1B f ) + [T (1{B f ) − T (1{B f )(z)] + c,

where c = T (1{B f )(z), and z ∈ Q is fixed as one of the (almost all) points of
Q where the conclusion of Lemma 11.2.7 is valid for the function 1{B f . Thus

k(T f − c)1Q\E k∞ 6 kT (1B f )1Q\E k∞ + k[T (1{B f ) − T (1{B f )(z)]1Q k∞ .

For the first term, we observe that

kT (1B f )kLp0 ,∞ 6 A0 k1B f kp0 6 A0 |B|1/p0 kf k∞ ,

and hence
 A kf k p0
0 ∞
|EΛ | := |{kT (1B f )k > Λ}| 6 cd |Q| 6 λ|Q|
Λ
if we choose Λ := (cd /λ)1/p0 A0 kf k∞ . We conclude that

kT (1B f )1Q\EΛ k∞ 6 (cd /λ)1/p0 A0 kf k∞ .

For the other term, we estimate pointwise at almost every s ∈ Q where

the conclusion of Lemma 11.2.7 is valid. Recalling that z ∈ Q was also chosen
in this way and dualising against y ∗ ∈ Y ∗ , we get
Z
∗
|hT (1{B f )(s) − T (1{B f )(z), y i| = hf (t), [K(s, t)∗ − K(z, t)∗ ]y ∗ i dt
{B
32 11 Singular integral operators
Z
6 k[K(s, t)∗ − K(z, t)∗ ]y ∗ kX ∗ dtkf k∞
{B
6 2kKkHör∗ ky ∗ kX ∗ kf k∞ .

Taking the supremum over y ∗ in the unit ball of Y ∗ and the essential supre-
mum over s ∈ Q, we arrive at

k1Q [T (1{B f ) − T (1{B f )(z)]k∞ 6 2kKkHör∗ kf k∞ .

Hence altogether

oscλ (T f ; Q) 6 k(T f − c)1Q\EΛ k∞ 6 (cd /λ)1/p0 A0 kf k∞ + 2kKkHör∗ kf k∞ ,

and taking the supremum over all Q ⊆ Rd proves the lemma.

The modifications in the case of a cube Q0 or a quadrant S in place of
Rd are immediate by inspection. We note that the balls B featuring in the
argument may extend beyond Q0 or S; one simply thinks of B ∩ Q0 or B ∩ S
in this case, while the complement {B will be replaced by Q0 \ B or S \ B,
respectively. 
Proof of Theorem 11.2.5(2a). Let us first consider the mapping properties of
#
the sub-linear operator M0,λ ◦ T , where λ = 2−2−d .
By assumption, T : L (Rd ; X) → Lp0 ,∞ (Rd ; Y ) is bounded (with norm
p0
#
A0 ), and Proposition 11.1.21 gives the boundedness of M0,λ : Lp0 ,∞ (Rd ; Y ) →
1/p
Lp0 ,∞ (Rd ) (with norm bounded by cd,λ 0 6 cd , since λ depends only on d, and
#
1/p0 6 1); thus the composition M0,λ ◦ T : Lp0 (Rd ; X) → Lp0 ,∞ (Rd ) is also
bounded (with norm at most cd A0 ).
#
On the other hand, the previous Lemma 11.2.8 says that M0,λ ◦T :
1/p
Lp0 (Rd ; X)∩L∞ (Rd ; X) → L∞ (Rd ) is bounded (with norm at most cd,λ 0 A0 +
kKkHör∗ 6 cd A0 + kKkHör∗ ), where the subspace Lp0 (Rd ; X) ∩ L∞ (Rd ; X) ⊆
L∞ (Rd ; X) is equipped with the norm of L∞ (Rd ; X).
This is essentially a setting to apply the Marcinkiewicz Interpolation The-
orem 2.2.3: by inspection, one checks that the relaxed assumption
#
M0,λ ◦ T : Lp0 (Rd ; X) ∩ L∞ (Rd ; X) → L∞ (Rd )

#
(in place of M0,λ ◦ T : L∞ (Rd ; X) → L∞ (Rd )) allows us to deduce the relaxed
conclusion
#
M0,λ ◦ T : Lp0 (Rd ; X) ∩ Lp (Rd ; X) → Lp (Rd ), p ∈ (p0 , ∞), (11.17)

where Lp0 (Rd ; X) ∩ Lp (Rd ; X) ⊆ Lp (Rd ; X) is equipped with the norm of

Lp (Rd ; X). In fact, the proof of the Marcinkiewicz Interpolation Theorem
2.2.3 is based on decomposing a function f in the domain space into the two
truncations, at varying level t,
 11.2 Singular integrals and extrapolation of Lp0 bounds 33

f

fet := f − t · 1{kf k>t} ,
kf k
f
fet := f · 1{kf k6t} + t · 1{kf k>t} ,
kf k

and it is immediate to verify that, if f ∈ Lp0 (Rd ; X), these remain in the
space Lp0 (Rd ; X), in addition to the other function space memberships used
in the proof of Theorem 2.2.3.
If θ ∈ (0, 1) is such that 1/p = (1 − θ)/p0 + θ/∞ = (1 − θ)/p0 , the
Marcinkiewicz Interpolation Theorem 2.2.3 shows that the norm of the oper-
ator in (11.17) is at most

 kM # ◦ T kLp0 (Rd ;X)→Lp0 ,∞ (Rd ) 1−θ

0,λ
c(θ, p0 , ∞) ×
1−θ
 kM # ◦ T k(Lp0 ∩L∞ )(Rd ;X)→L∞ (Rd ) θ
0,λ
×
θ
 c A 1−θ  c A + kKk ∗ θ
d 0 d 0 Hör
6 c(θ, p0 , ∞)
1−θ θ
6 c(θ, p0 , ∞) · 2 · (cd A0 + kKkHör∗ )

by the arithmetic–geometric mean inequality (11.15) in the last step. More-

over, still from Theorem 2.2.3 and the identity Γ (x + 1) = xΓ (x),
n Γ (p − p )Γ (p + 1) o1/p
0 0
c(θ, p0 , ∞) = = {p0 B(p − p0 , p0 )}1/p ,
Γ (p)

where the beta function is

Z 1
Γ (p − p0 )Γ (p0 )
B(p − p0 , p0 ) = = up−p0 −1 (1 − u)p0 −1 du
Γ (p) 0
Z 1
1
6 up−p0 −1 du = ,
0 p − p0

since p0 > 1 here. Substituting back (and redefining cd ), we find that the
norm of the operator in (11.17) is at most
 p 1/p
0
(cd A0 + 2kKkHör∗ ).
p − p0
Now Theorem 11.1.18, together with Remark 11.1.19 and the a priori
condition that T f ∈ Lp0 ,∞ (Rd ; X), show that
#
kT f kLp (Rd ;Y ) 6 8pkM0,λ (T f )kLp (Rd )
 p 1/p
0
6p (cd A0 + 16kKkHör∗ )kf kLp (Rd ;X)
p − p0
34 11 Singular integral operators

for all f ∈ Lp (Rd ; X) ∩ Lp0 (Rd ; X) and p ∈ (p0 , ∞). Since this is a dense
subspace of Lp (Rd ; X), the operator T has a unique extension to this space,
with the same norm estimate above.
The case of a cube or a quadrant in place of Rd follows by the same
argument, since all results quoted are also valid in these settings. 
It is also immediate from Lemma 11.2.8 and Proposition 11.1.24 that
#
kT f kBMO(Rd ;X) 6 8kM0,λ (T f )kL∞ (Rd ) 6 (cd A0 + 8kKkHör∗ )kf kL∞ (Rd ;Y )

for all f ∈ Lp0 (Rd ; X) ∩ L∞ (Rd ; X). Since this is not a dense subspace of
L∞ (Rd ; X), extending this estimate, and indeed the very meaning of “T f ”,
to all f ∈ L∞ (Rd ; X) requires an additional effort, to which we turn in Section
11.2.c below.
Proof of Theorem 11.2.5(3). We now assume that K satisfies both Hörmander
and dual-Hörmander conditions, and hence we have access to both cases (1)
and (2) that we already proved. By Theorem 11.2.5(1b), we have

kT kL (L1 (Rd ;X),L1,∞ (Rd ;Y ) 6 cd (A0 + kKkHör ).

We now use this estimate (rather than the original assumption) as input
to Theorem 11.2.5(2a), i.e., we apply the latter with 1 in place of p0 and
cd (A0 + kKkHör ) in place of A0 . This gives, for all p ∈ (1, ∞), the estimate
 1 1/p  
kT kL (Lp (Rd ;X),Lp (Rd ;Y ) 6 cd p cd (A0 + kKkHör ) + kKkHör∗
p−1
2 0
6 cd pp (A0 + kKkHör + kKkHör∗ ),

where we estimated
 1 1/p  p 1/p
6 = (p0 )1/p 6 p0 .
p−1 p−1
The conclusion agrees with the claim, after redefining cd .
The case of a cube or a quadrant in place of Rd is immediate, since both
(1) and (2) of the theorem, which we used above, were already proved in these
cases as well. 

11.2.c The action of singular integrals on L∞

The goal of this section is to establish the following theorem, in which indistin-
guishability of BMO(Rd ; X) functions only differing by an additive constant
manifests itself.

Theorem 11.2.9. Let X and Y be Banach spaces, p0 ∈ (1, ∞), and T ∈

L (Lp0 (Rd ; X), Lp0 (Rd ; Y )) be an operator with a dual Hörmander kernel K.
Suppose, moreover, at least one of the following:
 11.2 Singular integrals and extrapolation of Lp0 bounds 35

(1) Y does not contain a copy of c0 , or

(2) K is a dual operator-Hörmander kernel.
Then there is an operator Te ∈ L (L∞ (Rd ; X), BMOp0 (Rd ; Y )/Y ) of norm at
most (cd A0 + kKkHör∗ ) such that
(a) for all f ∈ Lp0 (Rd ; X) ∩ L∞ (Rd ; X), we have T f ≡ Tef modulo constants,
(b) for all f ∈ L∞ (Rd ; X) and g ∈ Lc,0 ∞
(Rd ; Y ∗ ) (compactly supported with
vanishing integral), we have

hTef, gi = lim hT (1En f ), gi (11.18)

n→∞

for any bounded measurable sets En ⊆ Rd such that dist({En , 0) → ∞.

Remark 11.2.10.
(1) By the John–Nirenberg inequality, the target space BMOp (Rd ; Y )/Y of Te
is independent of the value of p ∈ [1, ∞); however, the estimate for the
operator norm need not be, and we specifically state it with p = p0 .
(2) The left-hand side of (11.18) could be more pedantically written as “hh, gi,
where h ∈ [Tef ] is arbitrary”: the vanishing integral of g guarantees that
this expression is independent of the choice of h.
(3) The boundedness requirement on T in Theorem 11.2.9 may seem stronger
than in Theorem 11.2.5(2) (where it was only assumed that T maps bound-
edly into the larger space Lp0 ,∞ (Rd ; Y ) and for some p0 in the larger range
[1, ∞]), but this is only superficial, as we can always arrange ourselves to
be in the situation of Theorem 11.2.9 even under the apparently weaker
boundedness hypothesis:
First, if p0 = ∞, there is nothing to prove, as we can simply take
Te = T , which already maps into L∞ (Rd ; Y ) ⊆ BMO(Rd ; Y ). If, on
the other hand, p0 ∈ [1, ∞), Theorem 11.2.5(2a) guarantees that T ∈
L (Lp (Rd ; X), Lp (Rd ; Y )) for all p ∈ (p0 , ∞) ⊆ (1, ∞), and choosing one
such p as a new p0 , we are in the situation assumed in Theorem 11.2.9.

To deal with the equivalence classes modulo additive constants, it is convenient

to make the following preliminary observation:

Lemma 11.2.11. Let S be a set and X be a Banach spaces. There is a bijec-

tive linear correspondence between the following two classes of objects:
(1) equivalence classes [b] of functions b : S → X, where

[b] = {f : S → X; s 7→ f (s) − b(s) is constant on S},

(2) functions ∆ : S × S → X with the property

∆(s, t) + ∆(t, u) = ∆(s, u) ∀ s, t, u ∈ S. (11.19)

36 11 Singular integral operators

This correspondence is realised by

[s 7→ b(s)] ↔ (s, t) 7→ ∆(s, t) = b(s) − b(t).
Proof. To every [b], we associate ∆(s, t) := b(s) − b(t), and it is clear that this
is independent of the chosen representative of the equivalence class.
For the other direction, it is convenient to first record some additional
algebraic relations automatically satisfied by ∆. Taking s = t = u, we have
2∆(s, s) = ∆(s, s), and hence ∆(s, s) = 0 for all s ∈ S. Then taking u = s,
we have ∆(s, t) + ∆(t, s) = ∆(s, s) = 0, and hence ∆(s, t) = −∆(t, s) for all
s, t ∈ S. Now, to every ∆, we associate [∆(·, t)], where each t ∈ S defines the
same equivalence class. Indeed,
∆(s, t) − ∆(s, u) = ∆(u, s) + ∆(s, t) = ∆(u, t),
which is constant as a function of s ∈ S. It is immediate to verify that these
operations sending [b] to ∆, and ∆ to [b], are inverses of each other. 
For S ⊆ Rd (where we are mainly interested in the case that S = Rd or one
of its dyadic quadrants), we define
n
^ p (S; X) := ∆ ∈ Lloc
BMO 1
(S × S; X) with property (11.19),
Z 1/p o
k∆k∗,p := sup − k∆(s, t)kpX ds dt <∞
Q⊆S Q×Q
cube

^ p (S; X) by replacing “Q ⊆ S cube” by “Q ∈ D(S)”.

and BMO D

Lemma 11.2.12. Under the correspondence [b] ↔ ∆ of functions as in

Lemma 11.2.11, we have the correspondence of spaces:

BMOp (Rd ; X)/X ' BMO

^ p (Rd ; X),

with the equivalence of norms

kbkBMOp (Rd ;X) 6 k∆k∗,p 6 2kbkBMOp (Rd ;X) . (11.20)
The similar correspondence is valid with any of the dyadic quadrants S in
place of Rd and the dyadic BMOpD (both with and without tilde) in place of
BMOp .
Proof. For each cube Q ⊆ Rd , we have
Z 1/p Z Z p 1/p
inf − kb(s) − ckpX ds 6 − b(s) − − b(t) dt ds
c∈X Q Q Q X
Z Z 1/p
6 − − kb(s) − b(t)kpX ds dt
Q Q
Z Z p 1/p Z 1/p
p
= − − (b(s) − c) − (b(t) − c) ds dt 6 2 − kb(s) − ckX ds
Q Q X Q
 11.2 Singular integrals and extrapolation of Lp0 bounds 37

and taking the infimum over c ∈ X on the right, and then the supremum
over all cubes Q ⊆ Rd of the whole chain, we derive (11.20). The dyadic case
follows by taking the supremum over Q ∈ D(S) instead. 
In view of Lemma 11.2.12, the construction of an extension

Te ∈ L (L∞ (Rd ; X), BMO(Rd ; Y ))

of T ∈ L (Lp0 (Rd ; X), Lp0 (Rd ; Y )) is reduced to the construction of ∆T ∈

L (L∞ (Rd ; X), BMO(R
^ d ; X)) such that

∆T f (s, u) = T f (s) − T f (u) ∀ f ∈ Lp0 (Rd ; X) ∩ L∞ (Rd ; X).

It is convenient to define this as a priori mapping into Y ∗∗ -valued functions:

Lemma 11.2.13. For f ∈ L∞ (Rd ; X), y ∗ ∈ Y ∗ and s, u ∈ Rd , the expression

hy ∗ , ∆T f (s, u)i := hT (1B f )(s) − T (1B f )(u), y ∗ i

(11.21)
Z
+ h[K(s, t) − K(u, t)]f (t), y ∗ i dt,
{B

is independent of the auxiliary ball B with s, u ∈ 21 B.

Proof. With f, y ∗ , s, u fixed, let us temporarily denote the expression of inter-

est by δ(B). If B and B 0 are two such balls, we can choose a third such ball B 00
that contains both of them. So it is enough to prove the equality δ(B) = δ(B 0 )
for balls B ⊆ B 0 , hence {B 0 ⊆ {B. Note that ({B) \ ({B 0 ) = B 0 \ B. Then

δ(B 0 ) − δ(B) = hT (1B 0 \B f )(s) − T (1B 0 \B f )(u), y ∗ i

Z Z 
+ − h[K(s, t) − K(u, t)]f (t), y ∗ i dt,
{B 0 {B

where the difference of the integrals is

Z
h[K(u, t) − K(s, t)]f (t), y ∗ i dt = hT (1B 0 \B f )(u) − T (1B 0 \B f )(s), y ∗ i,
B 0 \B

which exactly cancels out the first term in the formula of δ(B 0 ) − δ(B). 
Let us then check how ∆T compares with the original T on the intersection
of their domains of definition:
Lemma 11.2.14. If f ∈ Lp0 (Rd ; X) ∩ L∞ (Rd ; X), then

∆T f (s, u) = T f (s) − T f (u).

38 11 Singular integral operators

Proof. Under the stated assumptions, Lemma 11.2.7 guarantees that

Z
h[K(s, t) − K(u, t)]f (t), y ∗ i dt = hT (1{B f )(s) − T (1{B f )(u), y ∗ i
{B

for almost all s, u ∈ 21 B and all y ∗ ∈ Y ∗ , and hence

hy ∗ , ∆T f (s, u)i = hT (1B f )(s) − T (1B f )(u), y ∗ i
+ hT (1{B f )(s) − T (1{B f )(u), y ∗ i = hT f (s) − T f (u), y ∗ i
Since this is true for all y ∗ ∈ Y ∗ , the claimed identity follows. 
To justify that the a priori Y ∗∗ -valued function ∆T f actually takes values
in Y , we invoke the following corollary of the Bessaga–Pelczyński Theorem
1.2.40. This is where the condition c0 6⊆ Y comes to use:
Proposition 11.2.15. Let Y be a Banach space that does not contain an
isomorphic copy of c0 . If yj ∈ Y satisfy
∞
X
|hyj , y ∗ i| < ∞ ∀y ∗ ∈ Y ∗ , (11.22)
j=1
P∞
then the series j=1 yj converges in norm in Y .
Proof. Let us first note that the condition (11.22) says that y ∗ 7→ (hyj , y ∗ i)∞
j=1
defines a linear operator from Y ∗ into `1 , which is easily seen to be closed,
and therefore bounded. Thus the closed graph theorem improves (11.22) to
∞
X
|hyj , y ∗ i| 6 Cky ∗ kY ∗ ∀y ∗ ∈ Y ∗ .
j=1
P∞ Pn
If j=1 yj does not converge, then the partial sums j=1 yj fail the Cauchy
criterion, and hence we can find m1 < n1 < m2 < . . . and δ > 0 such that
nk
X
kvk kY > δ > 0, vk := yj . (11.23)
j=mk

On the other hand, for any k = ±1 and any y ∗ ∈ Y ∗ , we also have

K
DX E K
X ∞
X
 k vk , y ∗ 6 |hvk , y ∗ i| 6 |hyj , y ∗ i| 6 Cky ∗ kY ∗ ;
k=1 k=1 j=1

hence
K
X
 k vk 6 C. (11.24)
Y
k=1
But the two conditions (11.23) and (11.24) are precisely those of the Bessaga–
Pelczyński Theorem 1.2.40 that guarantee the containment of an isomorphic
copy of c0 in span(vk )∞
k=1 ⊆ Y . This contradicts the assumption on Y . 
 11.2 Singular integrals and extrapolation of Lp0 bounds 39

After this interlude, we return to the main topic of this section:

Lemma 11.2.16. Under the assumptions of Theorem 11.2.9, for every f ∈

L∞ (Rd ; X), the function ∆T f in (11.21) is well defined, takes values in Y ⊆
Y ∗∗ , is strongly measurable, and satisfies

k∆T f kLp0 (Q×Q;Y ) 6 (cd A0 + kKkHör∗ )kf k∞ |Q|2/p0

for every cube Q ⊆ Rd .

Proof. Let B be the ball concentric with Q and with twice the diameter of Q;
hence Q ⊆ 21 B. From the assumption that T ∈ L (Lp0 (Rd ; X), Lp0 (Rd ; Y ))
and f ∈ L∞ (Rd ; X), it is immediate that T (1B f ) ∈ Lp0 (Rd ; Y ) and

kT (1B f )kp0 6 A0 k1B f kp0 6 A0 |B|1/p0 kf k∞ ,

so that
k(s, u) 7→ T (1B f )(s) − T (1B f )(u)kLp0 (Q×Q;Y )
6 2|Q|1/p0 kT (1B f )kLp0 (Rd ;Y ) 6 cd A0 |Q|2/p0 kf k∞ ,

The more delicate matter is the integral in (11.21). Certainly this inte-
gral exists, since the dual Hörmander condition guarantees that [K(s, t)∗ −
K(u, t)∗ ]y ∗ is jointly measurable and belongs to L1 ({B, dt; Y ∗ ) uniformly in
(s, u) ∈ Q, while f ∈ L∞ (Rd ; Y ) by assumption. An immediate estimate
with the dual Hörmander condition shows that this integral is bounded by
kKkHör∗ kf k∞ ky ∗ kY ∗ , uniformly in x ∈ Q, and hence defines a Y ∗∗ -valued
function h(s, u) with the pointwise bound

kh(s, u)kY ∗∗ 6 kKkHör∗ kf k∞ . (11.25)

What remains is to justify the Y -valuedness and the strong measurability

of this weakly defined function. To this end, we write fn = 12n B\2n−1 B f ,
p0 d
P
so that 1{B f = n>1 fn , say pointwise. Since each fn ∈ L (R ; X) ∩
L∞ (Rd ; X), we can apply Lemma 11.2.7 to see that
Z
h[K(s, t) − K(u, t)]fn (t), y ∗ i dt
{B
= hT fn (s) − T fn (u), y ∗ i =: hhn (s, u), y ∗ i

is the pairing of y ∗ with a Y -valued, strongly measurable function hn (s, u).

If we denote by h the a priori Y ∗∗ -valued function defined by
Z
hy , h(s, u)i := − h[K(s, t) − K(u, t)]f (t), y ∗ i dt,
∗
{B

then
40 11 Singular integral operators
∞
X
hy ∗ , h(s, u)i = hhn (s, u), y ∗ i ∀y ∗ ∈ Y ∗ . (11.26)
n=1

If K satisfies the dual operator-Hörmander condition, then

X∞ Z
khn (s, u)k 6 kK(s, t) − K(u, t)kkf k∞ dt 6 2kKkHörop
∗ kf k∞ ,

n=1 {B
P∞
so the series n=1 hn (s, u) converges absolutely, and hence in norm. Un-
der the mere dual Hörmander condition, but with the assumption that Y
does
P∞ not contain an isomorphic copy of c0 , the needed norm convergence of
n=1 hn (s, u) follows by Proposition 11.2.15 and the bound
∞
X Z
∗
|hhn (s, u), y i| 6 |hf (y), [K(s, t)∗ − K(u, t)∗ ]y ∗ i| dt
n=1 {B

6 2kKkHör∗ ky ∗ kY ∗ kf k∞ < ∞ ∀y ∗ ∈ Y ∗ .
P∞
In both cases, by (11.26), the limit of n=1 hn (s, u) must be h(s, u). Thus, as
a pointwise limit of Y -valued strongly measurable functions, h itself must be
both Y -valued and strongly measurable. Once these qualitative properties are
verified, the quantitative Lp0 (Q × Q; Y ) estimate is immediate by integrating
over Q × Q the already observed pointwise bound (11.25). 
Now we are prepared to complete:
p0
Proof of Theorem 11.2.9. The operator ∆T : L∞ (Rd ; X) → Lloc (Rd+d ; Y ) is
well defined by Lemma 11.2.16 and satisfies

k∆T f k∗,p0 6 (cd A0 + kKkHör∗ )kf k∞

for the norm defined in Lemma 11.2.12. By Lemma 11.2.12, we obtain a

bounded linear operator Te ∈ L (L∞ (Rd ; X), BMOp0 (Rd ; Y )/Y ), with the
same norm bound, by setting

Tef := [∆T f (·, u)] (the equivalence class modulo constants), (11.27)

where the choice of u ∈ Rd is irrelevant. By Lemma 11.2.14, we have

∆T f (s, u) = T f (s) − T f (u) for f ∈ Lp0 (Rd ; X) ∩ L∞ (Rd ; X), and hence
Tef = [T f ] in this case. This completes the proof of Claim (a) of the theorem.
∞
As for Claim (b), we note that pairing a g ∈ Lc,0 (Rd ; Y ∗ ) with an element
p0 d
of BMO (R ; Y )/Y is well defined, and independent of the representative of
the equivalence class, since the integral of g against any constant c ∈ Y will
vanish. By the assumptions on En , we can choose balls Bn := B(0, rn ) :=
B(0, dist({En , 0)) ⊆ En with rn → ∞. Let n be so large that supp g ⊆ 21 Bn .
Since Te is linear, we have

hTef, gi = hTe(1En f ), gi + hTe(1{En f ), gi =: In + IIn .

 11.3 Calderón–Zygmund operators and sparse bounds 41

By Claim (a), which we already proved, we have

In = hT (1En f ), gi.

For IIn , recalling the construction of Te from (11.27) with u = 0, and then the
definition of ∆T f (s, u) from (11.21) with B = Bn , we have

IIn = h∆T (1{En f )(·, 0), gi

Z
= h∆T (1{En f )(s, 0), g(s)i ds
d
ZR
= hT (1Bn 1{En f )(s) − T (1Bn 1{En f )(0), g(s)i ds
Rd
Z Z
+ h[K(s, t) − K(0, t)](1{En f )(t), g(s)i dt ds
Rd {Bn
=: IIIn + IVn = 0 + IVn ,

since Bn ⊆ En . Finally,
Z Z
|IVn | 6 kf kL∞ (Rd ;X) k[K(s, t) − K(0, t)]∗ g(s)kX ∗ dt ds.
Rd {Bn

For every fixed s ∈ supp g ⊆ 21 Bn , the inner integral is bounded by

kKkHör∗ kg(s)kY ∗ , and, as n → ∞, it converges to 0 by dominated conver-
gence; the same is also true for s ∈
/ supp g, since both the integral and the
upper bound vanish in this case. Thus also the double integral converges to 0
by another application of dominated convergence.
Altogether, we have seen that

hTef, gi − hT (1En f ), gi = IIn = IVn → 0 as n → ∞,

which concludes the proof of the remaining Claim (b) of Theorem 11.2.9. 

11.3 Calderón–Zygmund operators and sparse bounds

The goal of this section is to derive a powerful pointwise domination of

Calderón–Zygmund operators by simple averaging operators over sparse fam-
ilies of dyadic cubes; from this domination, norm estimates for Calderón–
Zygmund operators in various different spaces follow almost instantly.
The assumptions that we have to make on the kernel of the operator in
order to carry out this programme are somewhat stronger than those needed
for the Lp extrapolation of the previous section:

Definition 11.3.1 (Calderón–Zygmund kernel). Let Z be a Banach

space, and K : Ṙ2d → Z. We define the quantities
42 11 Singular integral operators

cK := sup{|s − t|d · kK(s, t)k : (s, t) ∈ Ṙ2d },

and, for u ∈ [0, 21 ],

n o
1
ωK (u) := sup |s − t|d kK(s, t) − K(s0 , t)k : |s − s0 | 6 u|s − t| ,
n o
2
ωK (u) := sup |s − t|d kK(s, t) − K(s, t0 )k : |t − t0 | 6 u|s − t| ,
i
ωK (u) := max ωK (u).
i=1,2

For K ∈ C 1 (Ṙ2d ; Z), let further

c1K := sup{|s − t|d+1 k∇s K(s, t)k : s 6= t},

c2K := sup{|s − t|d+1 k∇t K(s, t)k : s 6= t}.

We say that a kernel K with cK < ∞ is

(i) a standard kernel if ωK (u) 6 cδ uδ for some δ ∈ (0, 1],
(ii) a Dini kernel if ωK satisfies the Dini condition
Z 1/2
du
kωK kDini := ωK (u) < ∞,
0 u

(iii) a C 1 -Calderón–Zygmund kernel if K ∈ C 1 (Ṙ2d ; Z) and ciK < ∞, i = 1, 2,

(iv) an ω-Calderón–Zygmund kernel if ωK 6 ω,
i
(v) an (ω1 , ω2 )-Calderón–Zygmund kernel if ωK 6 ωi , i = 1, 2.
We also apply these notions to kernels K defined on {(s, t) : s, t ∈ S, s 6= t},
where S is either a cube or a quadrant of Rd ; in this case, each supremum
above is taken only over the respective domain of definition.

It is immediate that a standard kernel is a Dini kernel with kωkDini 6 δ −1 cδ .

Remark 11.3.2. For a convolution kernel K(x, y) = k(x − y), we have

cK = sup{|s|d kk(s)k : s 6= 0},

ciK = sup{|s|d+1 k∇k(s)k : s 6= 0}, i = 1, 2,
i d
ωK (u) = ωK (u) = sup{|s| kk(s) − k(s − t)k : |t| 6 u|s|}, i = 1, 2,

with no difference between i = 1 and i = 2 in the last two formulas.

Lemma 11.3.3.
∞
i 1 1
X
ωK ( ) 6 (1 + 2d )cK , i
ωK (2−k ) 6 kω i kDini .
2 log 2 K
k=2
 11.3 Calderón–Zygmund operators and sparse bounds 43

Proof. If |t − t0 | 6 21 |s − t|, then |s − t| 6 |s − t0 | + |t − t0 | 6 |s − t0 | + 12 |s − t|,

and hence |s − t| 6 2|s − t0 |. Thus

|s − t|d kK(s, t) − K(s, t0 )k 6 cK + 2d |s − t0 |d kK(s, t0 )k 6 (1 + 2d )cK ,

2 1
and hence ωK ( 2 ) 6 (1 + 2d )cK . The proof for ωK
1
is entirely similar.
i
If ω is increasing, which is obviously the case with ω = ωK , it follows that

ω(2−k−1 ) 6 ω(u), u ∈ (2−k−1 , 2−k ),

hence
Z 2−k
du
ω(2−k−1 ) log 2 6 ω(u) ,
2−k−1 u
and thus
∞ ∞ Z 1/2
X X 1 du
ω(2−k ) = ω(2−1−k ) 6 ω(u) .
log 2 0 u
k=2 k=1

Lemma 11.3.4. For K : Ṙ2d → Z = L (X, Y ), we have:

1
(1) If kωK kDini < ∞, then K is a dual operator-Hörmander kernel, and
1
kKkHör∗op 6 σd−1 kωK kDini .
2
(2) If kωK kDini < ∞, then K is an operator-Hörmander kernel, and
2
kKkHörop 6 σd−1 kωK kDini .

(3) Every standard kernel is a Dini kernel with

cK  cδ 
kωkDini 6 2d+1 1 + log+ d+1 .
δ 2 cK

(4) Every C 1 -Calderón–Zygmund kernel is a standard kernel with

i
ωK (u) 6 2d+1 ciK · u

and a Dini kernel with

i
 ci 
kωK kDini 6 2d+1 cK 1 + log+ K .
cK

Here σd−1 is the (d − 1)-dimensional measure of the unit sphere in Rd . The

same conclusions hold with Ṙ2d replaced by Ṡ 2 := {(s, t) : s, t ∈ S, s 6= t},
where S is either a cube or a quadrant of Rd , and both the Dini and the
Hörmander conditions are modified by restricting the variables to the respective
domain of definition.
44 11 Singular integral operators

Note that, in concrete situations, the constants cδ or ciK are often much larger
than cK . The point of the bounds in parts (3) and (4) is that these larger
constants contribute to the Dini bounds only logarithmically.
Proof. We will first prove (2); the proof of (1) is analogous.
Z
kK(x, y) − K(x, y 0 )k dx
|x−y|>2|y−y 0 |
Z  |y − y 0 | 1
2
6 ωK dx
|x−y|>2|y−y 0 | |x − y| |x − y|d
Z ∞  |y − y 0 |  dr Z 12
2 2 dt 2
= σd−1 ωK = σd−1 ωK (t)d = σd−1 kωK kDini
0
2|y−y | r r 0 t

and this is the required bound.

For the remaining claims, we begin with the following observation. For
|x − x0 | 6 u|x − y| and v ∈ [0, 1], we have
1
|x + v(x0 − x) − y| > |x − y| − |x0 − x| > (1 − u)|x − y| > |x − y|.
2
This implies the crude bound

cK cK cK 2d+1 cK
kK(x0 , y) − K(x, y)k 6 + 6 (2 d
+ 1) 6 .
|x0 − y|d |x − y|d |x − y|d |x − y|d

This shows that ωK i

(u) 6 2d+1 cK for all u ∈ [0, 21 ] and i = 1, and the proof
for i = 2 is similar.
(3): By the previous observation, denoting c0 := 2d+1 cK , the standard
estimate ω(u) 6 cδ uδ bootstraps to ω(u) 6 min{c0 , cδ uδ }. If c0 6 cδ , then
Z (c0 /cδ )1/δ Z 1
du δ du
kωkDini 6 cδ u + c0
0 u (c0 /cδ ) 1/δ u
cδ c0  cδ 1/δ
 c0  cδ 
= + c0 log = 1 + log .
δ cδ c0 δ c0
R1
If c0 > cδ , we simply estimate kωkDini 6 0 cδ uδ−1 du = cδ /δ 6 c0 /δ. Hence,
in each case, we have
c0  cδ 
kωkDini 6 1 + log+ .
δ c0
We will prove (4) in the case i = 1, the case of i = 2 is analogous. Hence
 11.3 Calderón–Zygmund operators and sparse bounds 45
.1
kK(x0 , y) − K(x, y)k = K(x + u(x0 − x), y)
v=0
Z 1
= (x0 − x) · ∇x K(x + v(x0 − x), y) dv
0
1
c1K
Z
0
6 |x − x| dv
0 |x + v(x0 − x) − y|d+1
1
c1K 2d+1 c1K
Z
6 u|x − y| dv = u .
0 ( 12 |x − y|)d+1 |x − y|d

This is the claimed standard estimate, and the Dini estimate follows from part
(3) with δ = 1 and cδ = 2d+1 ciK .
The version with a cube or a quadrant follows with the same argument by
simply restricting all the variables and the integrals to the relevant domain of
definition. 
In particular, Dini kernels satisfy both Hörmander and dual Hörmander con-
ditions, and hence all the results of the previous section apply to them:
Corollary 11.3.5 (Calderón–Zygmund). Let X and Y be Banach spaces
and p0 ∈ [1, ∞]. Let T ∈ L (Lp0 (Rd ; X), Lp0 ,∞ (Rd ; Y )) be an operator with
a Calderón–Zygmund kernel K. Then all conclusions of Theorem 11.2.5 hold
2 1
with kKkHör replaced by kωK kDini and kKkHör∗ by kωK kDini in the estimates.
Proof. This follows at once from Theorem 11.2.5, where the same conclusions
are deduced for Hörmander and/or dual Hörmander kernels K, and Lemma
11.3.4, where these assumptions are verified for under the Dini conditions. 

11.3.a An abstract domination theorem

We will first present an abstract form of the domination theorem, i.e., we

postulate the relevant properties of the operator needed to carry out the
proof, and only then return to the question of checking these properties in the
concrete case of Calderón–Zygmund operators.
We will formulate the theorem for positive sub-linear operators mapping
a linear space of X-valued functions into L0 (Rd ; R+ ). By this we mean that
for all functions f and g we have that T f > 0 is a non-negative function,
T (αf ) = |α|T f for constants α, and T (f + g) 6 T f + T g for all f, g in the
domain of T . Note that if T is a linear operator mapping into L0 (Rd ; Y ), then
the operator f 7→ kT f (·)kY is a positive sub-linear one, and this is the way
that such operators will be naturally covered by the theory.
Theorem 11.3.6 (Abstract sparse domination). Let X be a Banach
space, let T be a positive sub-linear operator from L1 (Rd ; X) into L0 (Rd ; R+ ),
and consider the associated maximal operator

MT# f (x) := sup sup |T (1{(5Q) f )(y) − T (1{(5Q) f )(z)|. (11.28)

Q3x y,z∈Q
46 11 Singular integral operators

Suppose that both T and MT# are bounded from L1 (Rd ; X) to L1,∞ (Rd ). Then
for every boundedly supported f ∈ L1 (Rd ) and ε ∈ (0, 1), there is a (1 − ε)-
sparse family S of dyadic cubes such that, almost everywhere,

8 · 10d · cT X
Z
Tf 6 1S − kf k,
ε 5S S∈S

where
cT := kT k1→1,∞ + kMT# k1→1,∞ . (11.29)

The heart of Theorem 11.3.6 is contained in the following lemma:

Lemma 11.3.7. Under the assumptions of Theorem 11.3.6, for any f ∈

L1loc (Rd ; X), any cube Q0 and ε ∈ (0, 1), there are disjoint subcubes Qj ∈
D(Q0 ) such that X
|Qj | 6 ε|Q0 | (11.30)
j

and, almost everywhere,

4 · 10d cT 
Z X Z  X
1Q0 T (15Q0 f ) 6 1Q0 − kf k+ 1Qj − kf k + 1Qj T (15Qj f ),
ε 5Q0 j 5Qj j

where cT was defined in (11.29).

Proof. Given a cube Q0 , consider any disjoint family of its subcubes Qj ∈

D(Q0 ). Then we have
X
1Q0 T (15Q0 f ) = 1Q0 \Sj Qj T (15Q0 f ) + 1Qj T (15Q0 f )
j
X X (11.31)
6 1Q0 \ S
j Qj T (15Q0 f ) + 1Qj T (15Q0 \5Qj f ) + 1Qj T (15Qj f ),
j j

and
1Qj T (15Q0 \5Qj f ) 6 1Qj [inf MT# (15Q0 f ) + inf T (15Q0 \5Qj f )]
Qj Qj
(11.32)
6 1Qj [inf MT# (15Q0 f ) + inf {T (15Q0 f ) + T (15Qj f )}]
Qj Qj

where we used sublinearity and the definition of MT# to get the estimates.
Note that no convergence issues arise when viewing the above lines in the
pointwise sense.
The last term in (11.31) already has the correct form, and it remains to
choose the cubes Qj in such a way that we have (11.30) as well as
Z
X cd cT
1Q0 \Sj Qj T (15Q0 f ) + 1Qj T (15Q0 \5Qj f ) 6 1Q0 − kf k.
j
ε 5Q0
 11.3 Calderón–Zygmund operators and sparse bounds 47

For a λ > 0 to be chosen and every Q ∈ D(Q0 ), we define F (Q) ⊆ Q by

F (Q) := Q ∩ [{T (15Q f ) > λhkf ki5Q } ∪ {MT# (15Q f ) > λhkf ki5Q }].

Thus, by the assumed L1 (Rd ; X) to L1,∞ (R) bounds,

|F (Q)| 6 |{T (15Q f ) > λhkf ki5Q }| + |{MT# (15Q f ) > λhkf ki5Q }|
k15Q f k1 5d (11.33)
6 (kT k1→1,∞ + kMT k1→1,∞ ) = cT · |Q|.
λhkf ki5Q λ

Let then Qj ∈ D(Q0 ) be the maximal dyadic subcubes such that

|Qj ∩ F (Q0 )|
> 2−d−1 .
|Qj |
The cubes Qj are disjoint, so that
X X |Qj ∩ F (Q0 )| 2 · 10d
|Qj | 6 −d−1
6 2d+1 |F (Q0 )| 6 cT · |Q0 | = ε|Q0 |,
j j
2 λ

which is (11.30), if we choose

2 · 10d
λ := cT .
ε
Substituting back to (11.33), this choice gives in particular that

|F (Q)| 6 2−d−1 |Q|.

Since 1F (Q0 ) 6 MD (1F (Q0 ) ) almost everywhere, we see that F (Q0 ) is

contained in j Qj = {MD (1F (Q0 ) ) > 2−d−1 }, except perhaps for a subset of
S
measure zero. In particular, we have (a.e.)

1Q0 \Sj Qj T (15Q0 f ) 6 1Q0 \Sj Qj λhkf ki5Q0 . (11.34)

On the other hand, the maximality of Qj implies that its dyadic parent Q
bj
satisfies the opposite inequality, and hence

|Qj ∩ F (Q0 )| |Q
b j ∩ F (Q0 )| 2−d−1 1
6 6 −d = .
|Qj | −d
2 | Qj |
b 2 2

But also |F (Qj )| 6 2−d−1 |Qj | 6 41 |Qj |, and hence

1 1
|Qj \ [F (Q0 ) ∪ F (Qj )]| > (1 − − )|Qj | > 0.
2 4
With any zj in the non-empty set Qj \ [F (Q0 ) ∪ F (Qj )], we can now complete
the estimation of (11.32) as follows:
48 11 Singular integral operators

1Qj T (15Q0 \5Qj f ) 6 1Qj [MT# (15Q0 f )(zj ) + T (15Q0 f )(zj ) + T (15Qj f )(zj )]
6 1Qj [λhkf ki5Q0 + λhkf ki5Q0 + λhkf ki5Qj ],

where we used the bounds for MT# (15Q0 f ) and T (15Q0 f ) on {F (Q0 ) that
follow directly from the definition of these sets, and the analogous bound for
T (15Qj f ) on {F (Qj ). Hence
X X
1Qj T (15Q0 \5Qj f ) 6 1Sj Qj 2λhkf ki5Q0 + 1Qj λhkf ki5Qj ,
j j

and together with (11.31), (11.34) and the choice of λ, this completes the
proof of the lemma. 
Iterating the previous lemma, we obtain:
Lemma 11.3.8. Under the assumptions of Theorem 11.3.6, for any cube Q0
1
and f ∈ Lloc (Rd ; X) and ε ∈ (0, 1), there is a (1 − ε)-sparse subcollection
S (Q0 ) ⊆ D(Q0 ) such that, almost everywhere,

8 · 10d cT X
Z
1Q0 T (15Q0 f ) 6 1S − kf k.
ε 5S
S∈S (Q0 )

Proof. By Lemma 11.3.7, almost everywhere we have

Z Z
cd cT  X  X
1Q0 T (15Q0 f ) 6 1Q0 − kf k + 1Qj − kf k + 1Q1j T (15Q1j f )
ε 5Q0 j 5Qj j

for disjoint subcubes Q1j ∈ D(Q0 ) such that

X
|Q1j | 6 ε|Q0 |,
j

and cd = 4 · 10d Applying the same estimate to each Q1j in place of Q0 , and
continuing by induction, almost everywhere we obtain
Z N −1 X Z
cd cT  X
1Q0 T (15Q0 f ) 6 1Q0 − kf k + 2 1Qnj − kf k
ε 5Q0 n=1 j 5Qnj
Z  (11.35)
X X
+ 1Q N
k
− kf k + 1QN k
T (15QN
k
f ),
k 5Qn
k k

where the Qnj are dyadic subcubes of some Qn−1

i in such that
X
|Qnj | 6 ε|Qn−1
i |.
n−1
j:Qn
j ⊆Qi

In particular,
 11.3 Calderón–Zygmund operators and sparse bounds 49
X X
|Qnj | 6 ε |Qn−1
i | 6 . . . 6 εn |Q0 |,
j i

so that the support of the last term in (11.35) becomes negligible in the limit
N → ∞. Thus, almost everywhere, we have
∞ Z
cd cT X X
1Q0 T (15Q0 f ) 6 2 1Qnj − kf k, (11.36)
ε n=0 j 5Qn
j

where the pairwise disjoint subsets

[
Ejn := Qnj \ Qn+1
k
k

have measure |Ejn | > (1 − ε)|Qnj |. In other words, the cubes Qnj form a (1 − ε)-
sparse subcollection S (Q0 ) ⊆ D(Q0 ), and (11.36) is precisely the estimate
asserted in the lemma. 
In order to pass from the local Lemma 11.3.8 to the global Theorem 11.3.6,
we use:

Lemma 11.3.9. Let E ⊆ Rd satisfy 0 < diam(E) < ∞. Then there is a

partition Q of Rd by dyadic cubes Q such that E ⊆ 5Q for every Q ∈ Q.

Proof. Consider all dyadic cubes Q ∈ D with the property that E 6⊆ 2Q.
Clearly all cubes with diam(Q) < 21 diam(E) will satisfy this condition. On the
other hand, every cube Q ∈ D is contained in some Q e ∈ D such that E ⊆ 2Q:e
if we fix some x ∈ Q and then r > 0 large enough so that E ⊆ B(x, r), then
e ⊇ Q with `(Q)
it suffices to take Q e > 2r, since then 2Q e ⊇ B(x, 1 `(Q))
e ⊇ E.
2
Let Q be the collection of maximal dyadic cubes with the property that
E 6⊆ 2Q. Maximality implies disjointness, and from what we just checked, it
follows that every x ∈ Rd is contained in some Q ∈ Q, so these cubes form a
partition of Rd .
Since Q is maximal, its dyadic parent Q b satisfies E ⊆ 2Q.b It remains to
observe that 2Q b ⊆ 5Q to complete the proof. 
We now return to:
Proof of Theorem 11.3.6. If f ≡ 0, there is nothing to prove, so fix a non-
zero, compactly supported f ∈ L1c (Rd ; X). Thus E = supp f satisfies 0 <
diam(E) < ∞ as required to apply Lemma 11.3.9. This lemma produces a
partition Q ⊆ D of Rd such that supp f ⊆ 5Q, and thus 15Q f = f , for every
Q ∈ Q. This means that
X X
Tf = 1Q T f = 1Q T (15Q f ).
Q∈Q Q∈Q
50 11 Singular integral operators

Now Lemma 11.3.8 applies to each term on the right, producing (1 − ε)-sparse
subcollections S (Q) ⊆ D(Q) for each Q ∈ Q, and
X cd cT X Z Z
X cd cT X
1Q T (15Q f ) 6 1S − kf k = 1S − kf k,
ε 5S ε 5S
Q∈Q Q∈Q S∈S (Q) S∈S

where S := Q∈Q S (Q) and cd = 8 · 10d . It is immediate that this union of

S
disjointly supported sparse collections remains sparse, as the same pairwise
disjoint subsets E(S) ⊆ S remain pairwise disjoint also among all S ∈ S . 

11.3.b Sparse operators and domination

With Theorem 11.3.6 at our disposal, the following notion should not appear
too alien to the reader:

Definition 11.3.10 (Sparse operator). Given a sparse collection of sets

S ⊆ D, the associated sparse operator is
X Z
AS f := 1S − f.
S∈S S

More generally, with a dilation factor % > 1, we define

X Z
A%S f := 1S − f.
S∈S %S

In contrast to most other operators that we encounter, the boundedness prop-

erties of the sparse operators tend to be extremely easy. As a first illustration,
0
we check the Lp boundedness of AS by dualising against g ∈ Lp :

|E(S)|
Z XZ Z X
AS f · g = − f · − g · |S| 6 inf MD f · inf MD g ·
S S γ
S∈S S S S∈S
Z
1 1 1
6 MD f · MD g 6 kMD f kp · kMD gkp0 6 p0 kf kp · pkgkp0 .
γ γ γ

This shows that kAS kp→p 6 γ −1 pp0 , where γ is the sparseness parameter;
since AS is manifestly positive, it suffices to consider positive functions above,
and the same bound persists for vector-valued functions.
Looking back at the statement of Theorem 11.3.6, it almost says that
T f 6 c · AS kf k under the assumptions of the theorem, but the presence of
the expanded cubes 5S prevents this from being strictly true in the stated
form. While the variant of a sparse operator implicitly appearing in Theorem
11.3.6 would be almost as good as AS for many purposes, the use of the more
symmetric (indeed, self-dual) operators AS as in Definition 11.3.10 is often
preferred.
 11.3 Calderón–Zygmund operators and sparse bounds 51

A trivial way to achieve this in Theorem 11.3.6 is to dominate 1S 6

15S , after which the same cube 5S will appear in both the indicator and the
integral. These cubes will still be sparse, if only with a smaller parameter γ =
5−d (1 − ε), since the disjoint major subsets E(S) ⊆ S ⊆ 5S satisfy |E(S)| >
(1−ε)|S| = (1−ε)5−d |5S| and hence also qualify for the disjoint major subsets
of the expanded cubes 5S. An apparent loss in this construction is the fact
that these 5S are no longer dyadic cubes. Even this problem, however, can be
fixed, by a variant of the shifted dyadic cubes that we introduced in Definition
3.2.25. Recall that the standard dyadic system is
[
D 0 := Dj0 , Dj0 := {2−j ([0, 1)d + m) : m ∈ Zd }.
j∈Z

We will need the case N = 5 of the following statement, but we record the gen-
eral formulation for convenience of reference, as the case N = 3 also features
in various applications.
Proposition 11.3.11 (Dilated dyadic cubes). Let N ∈ Z+ be odd. Then
the collection of N -fold concentric dilations {N Q : Q ∈ D(Rd )} can be par-
titioned into N d subcollections D n;N , n ∈ ZNd
, each of which has the same
covering and nestedness properties as D, namely,
[ n;N
D n;N = Dj ,
j∈Z

where for each j ∈ Z:

(1) Djn;N is a partition of Rd consisting cubes of side-length N · 2−j , and
n;N
(2) Dj+1 is a refinement of Djn;N .

Proof. Since Dj (Rd ) = {I1 × · · · × Id : Ii ∈ Dj (R)} and N (I1 × · · · × Id ) =

N I1 × · · · × N Id , it suffices to verify the case d = 1. In the calculation that
follows, we will need to dilate an interval I = [c − r, c + r) both by the
algebraic multiplication a · I = {a · t : t ∈ I} = [ac − ar, ac + ar) and by
the concentric dilation, for which we temporarily adopt the heavier notation
a I = [c − ar, c + ar) for the sake of distinction.
With these notations fixed, we have

{N I : I ∈ Dj } = {N 2−j ([0, 1) + m) : m ∈ Z}
= {2−j ([−N 0 , N 0 + 1) + m) : m ∈ Z} N := 2N 0 + 1

= {2−j ([0, N ) + m − N 0 ) : m ∈ Z}
= {2−j ([0, N ) + m) : m ∈ Z}
n  m o
= N 2−j [0, 1) + :m∈Z .
N
The sought-after partition of this collection is now achieved as follows: For
each n ∈ ZN = {0, 1, . . . , N − 1} and j ∈ Z, we define
52 11 Singular integral operators
n  α(n, j)  o
Djn;N := N 2−j [0, 1) + k + :k∈Z (11.37)
N
for appropriate α(n, j) ∈ ZN to be shortly determined. It is clear that each
Djn;N satisfies (1) from the statement of the Proposition, no matter how we
choose α(n, j). To ensure (2), it suffices to check that the left (or equivalently
right) half of any I ∈ Djn;N belongs to Dj+1
n;N
. For a generic I as written above,
the left half will be
 α(n, j)   2α(n, j) 
N 2−j [0, 21 ) + k + = N 2−j−1 [0, 1) + 2k + .
N N
n;N
For this to be in Dj+1 , it is necessary and sufficient that

2α(n, j) ≡ α(n, j + 1) mod N (11.38)

If we specify α(n, 0) := n, all other α(n, j), j ∈ Z \ {0} will be uniquely

determined by (11.38), since 2 has a multiplicative inverse in ZN for odd N .
Indeed, the solution is given by

α(n, j) ≡ 2j n mod N, (11.39)

where the negative powers are interpreted in the sense of the multiplicative
inverse mod N .
For each j ∈ Z, the map n 7→ 2j n mod N is a bijection on ZN , and thus
N
[ n  a o
Djn;N = N 2−j [0, 1) + k + : k ∈ Z, a ∈ ZN
n=0
N
n  m  o
= N 2−j [0, 1) + : m ∈ Z = {N I : I ∈ Dj },
N
so indeed {N I : I ∈ D} is a disjoint union of the collections D n;N , n ∈ ZN ,
and we already checked that each D n;N has the properties (1) and (2). 
Remark 11.3.12 (Shifted dyadic cubes). The cube families D n;N constructed
above are close relatives of the shifted dyadic cubes of Definition 3.2.25, and
they satisfy a variant of the Covering Lemma 3.2.26:
Given an odd N ∈ Z+ , for every cube Q ⊆ Rd , there exist a vector
n ∈ ZdN and a cube D ∈ D n;N such that
N 2N
`(Q) < `(D) 6 `(Q) and Q ⊆ D. (11.40)
N −1 N −1
In fact, let R ∈ D be a cube of side-length `(R) ∈ (`(Q)/(2N 0 ), `(Q)/N 0 ] that
contains the centre zQ of Q, where N = 2N 0 + 1 as before. Then D = N R ∈
D n;N for some n ∈ ZdN , and D contains the cube of side-length 2N 0 `(R) >
`(Q) centred at zQ ; thus D ⊇ Q, and `(D) = N `(R) lies exactly in the range
asserted in (11.40).
 11.3 Calderón–Zygmund operators and sparse bounds 53

Also note that both the partition and refinement properties (1) and (2) of
Proposition 11.3.11 of each D n;N , as well as the covering property of every
cube Q ⊆ Rd by a cube in some D n;N , remain invariant if we drop the algebraic
dilation factor N in (11.37), so as to be back to cubes of side-length 2−j .
When N = 3, this reproduces precisely the shifted dyadic cubes of Definition
3.2.25; since 2 ≡ −1 mod 3, (11.39) reduces in this case to the simpler form
α(n, j) = (−1)j n, where reference to modular arithmetic can be avoided.
%
It is now easy to show that the sparse operators with a dilation, AS , may
always be dominated by a finite number of the simple sparse operators AS n .
It is technically convenient to take an odd integer N for the dilation factor.
This causes little loss of generality since, choosing N > %, we can always
dominate Z  N dZ
− f6 − f
%Q % NQ

and hence A%S f 6 (N/%)d AN

S f for f > 0.

Lemma 11.3.13. Let S ⊆ D be ε-sparse for some ε ∈ (0, 1), and N ∈ Z+

be odd. Then there are N −d ε-sparse collections S n ⊆ D n;N for each n ∈ ZdN
such that, for every non-negative f ∈ L1loc (Rd ),
X
ANSf 6 AS n f
n∈Zd
N

Proof. We note that the collection {5Q : Q ∈ S } is N −d ε-sparse, with

the same disjoint subsets E(Q) ⊆ Q ⊆ N Q that satisfy |E(Q)| > ε|Q| =
εN −d |N
S Q|. By Proposition 11.3.11, we have a decomposition {N Q : Q ∈
D} = n∈Zd D n;N into dyadic systems D n;N . We then define S n := {N Q :
N
Q ∈ S } ∩ D n;5 . Thus
X Z X Z X X Z
AN S f = 1Q − f 6 1 N Q − f = 1 N Q − f
NQ NQ NQ
Q∈S Q∈S n∈Zd
N
Q∈S
N Q∈D n;N
X X Z X
= 1Q0 − f = AS n f.
Q0 ∈S n Q0
n∈Zd
N n∈Zd
N


We can now reformulate Theorem 11.3.6 in terms of sparse operators:

Theorem 11.3.14 (Abstract sparse domination II). Let X be a Ba-

nach space, and let T be a positive sub-linear operator from L1 (Rd ; X) into
L0 (Rd ; R+ ), and consider the associated Lerner’s maximal operator

MT# f (x) := sup sup |T (1{(5Q) f )(y) − T (1{(5Q) f )(z)|.

Q3x y,z∈Q
54 11 Singular integral operators

Suppose that both T and MT are bounded from L1 (Rd ; X) to L1,∞ (Rd ). Then
for every boundedly supported f ∈ L1 (Rd ), there is a 5−1 -sparse collection
S ⊆ D and, for every n ∈ Zd5 , a 5−d−1 -sparse collection S n ⊆ D n;5 of the
dyadic systems as in Proposition 11.3.11, such that almost everywhere
X
5
T f 6 10d+1 cT AS kf k 6 10d+1 cT AS n kf k,
n∈Zd
5

where cT := kT k1→1,∞ + kMT# k1→1,∞ .

Proof. Choosing ε = 4/5 in Theorem 11.3.6, we find a 51 -sparse collection

S ⊆ D such that

8 · 10d · cT X
Z
Tf 6 1S − kf k = 10d+1 cT A5S kf k.
4/5 5S
S∈S

This is the first claim, and the second one follows from Lemma 11.3.13. 

11.3.c Sparse domination of Calderón–Zygmund operators

The goal of this section is to specialise the abstract Theorem 11.3.14 to the
case of Calderón–Zygmund operators in the following form:

Theorem 11.3.15 (Sparse domination of singular integrals). Let X

and Y be Banach spaces, p0 ∈ [1, ∞], and let

T ∈ L (Lp0 (Rd ; X), Lp0 ,∞ (Rd ; Y ))

be an operator with a Dini kernel K. Then for every boundedly supported

f ∈ L1 (Rd ), there is a 5−1 -sparse collection S ⊆ D and, for every n ∈ Zd5 ,
a 5−d−1 -sparse collection S n ⊆ D n;5 of the dyadic systems as in Proposition
11.3.11, such that almost everywhere
X
kT f kY 6 cd,T A5S kf kX 6 cd,T AS n kf k,
n∈Zd
5

where

cd,T 6 cd kT kLp0 (Rd ;X)→Lp0 ,∞ (Rd ;Y ) + cK + kωkDini
with cK and ω as in Definition 11.3.1.
The result remains true if Rd is systematically replaced by a cube or a quad-
rant of Rd , both in the function spaces where the boundedness is considered,
and in the definition of the kernel bounds cK and kωK kDini .

Proof. By Theorem 11.3.14, applied to the positive sub-linear operator U :

f 7→ kT f (·)kY , the result follows if we can estimate kU kL1 →L1,∞ and
kMU kL1 →L1,∞ by the bound for cd,T given above. For the former, this is
 11.3 Calderón–Zygmund operators and sparse bounds 55

immediate from the Calderón–Zygmund Theorem 11.2.5 and Lemma 11.3.4,

which show that

kU kL1 (Rd ;X)→L1,∞ (Rd ) = kT kL1 (Rd ;X)→L1,∞ (Rd ;Y )

6 cd kT kLp0 (Rd ;X)→Lp0 ,∞ (Rd ;Y ) + kKkHör


6 cd kT kLp0 (Rd ;X)→Lp0 ,∞ (Rd ;Y ) + kKkDini .

For MU , we first observe that, by the triangle inequality,

|U (1{5Q f )(y) − U (1{5Q f )(z)| = |kT (1{5Q f )(y)kY − kT (1{5Q f )(z)kY |

6 kT (1{5Q f )(y) − T (1{5Q f )(z)kY .

Hence, taking the supremum over y, z ∈ Q and then over cubes Q 3 x, it

follows that

MU f (x) 6 MT# f (x) := sup sup kT (1{(5Q) f )(y) − T (1{(5Q) f )(z)kY .

Q3x y,z∈Q

The norm estimate of the latter is the content of the following lemma. 

Lemma 11.3.16. Let X and Y be Banach spaces, p0 ∈ [1, ∞], and let T be an
operator with a Dini kernel K : Ṙ2d → L (X, Y ). Then the maximal operator

MT# f (x) := sup sup kT (1{(5Q) f )(y) − T (1{(5Q) f )(z)kY

Q3x y,z∈Q

satisfies
MT# f (x) 6 cd (cK + kωK kDini )M f (x)
and
kMT# kL1 (Rd ;X)→L1,∞ (Rd ) 6 cd (cK + kωK kDini ).
The result remains true if Rd is systematically replaced by a cube Q0 ⊆ Rd
or a quadrant S ⊆ Rd , both in the function spaces where the boundedness is
considered, and in the definition of the kernel bounds cK and kωK kDini .

Proof. For x, x0 , x1 ∈ Q, we have

T (1{(5Q) f )(x0 ) − T (1{(5Q) f )(x1 )

1
X
= (−1)j [T (1{(5Q) f )(xj ) − T (1{(5Q) f )(x)]
j=0
1
X Z
j
= (−1) [K(xj , y) − K(x, y)]f (y) dy,
j=0 {(5Q)

where
56 11 Singular integral operators
Z
[K(xj , y) − K(x, y)]f (y) dy
{(5Q)
Z
6 √ k[K(xj , y) − K(x, y)]f (y)k dy
{B(x,4 d`(Q))
Z
+ √ k[K(xj , y) − K(x, y)]f (y)k dy =: I + II
B(x,4 d`(Q))\(5Q)
√
where, observing that |xj − x| < d`(Q) 6 14 |x − y| for x, xj ∈ Q and
√
y ∈ {B(x, 4 d`(Q)),
Z  |x − x|  1
1 j
I6 √ ω K kf (y)k dy
{B(x,4 d`(Q)) |x − y| |x − y|d
X∞ Z  √d`(Q)  kf (y)k dy
1
6 √ √ ωK √ √
k=2 2
k d`(Q)6|y−x|<2k+1 d`(Q) 2k d`(Q) (2k d`(Q))d
X∞ Z
6 ωK1
(2−k )cd − √ kf (y)k dy
k=2 B(x,2k+1 d`(Q))
∞
X
6 cd M f (x) 1
ωK (2−k ) 6 cd M f (x)kωK
1
kDini ,
k=2

by Lemma 11.3.3 in the last step. On the other hand, since |xj − y|, |x − y| >
2`(Q) for x, xj ∈ Q and y ∈ / 5Q, we obtain
Z
2
II 6 √ cK d
kf (y)k dy
B(x,2 d`(Q))\(5Q) (2`(Q))
Z
6 cK cd − √ kf (y)k dy 6 cK cd M f (x).
B(x,2 d`(Q))

These bounds give the pointwise estimate for MT# f (x), and the norm estimate
is then immediate from the corresponding bound for the Hardy–Littlewood
maximal operator M .
The case of a cube or a quadrant in place of Rd follows by inspection of
the same argument: if all variables under consideration are restricted like this,
it is evident that only the corresponding restrictions of the kernel conditions
will be needed to make the estimates. 

11.3.d Weighted norm inequalities and the A2 theorem

We are now ready to provide the main application of the sparse domination
of Calderón–Zygmund operators: their weighted norm inequalities with an
optimal dependence of the weight. A function w ∈ L1loc (Rd ) is called a weight
if w(x) ∈ (0, ∞) almost everywhere. We recall from Appendix J the following
definition, which we now extend to the local situation as well:
 11.3 Calderón–Zygmund operators and sparse bounds 57

Definition 11.3.17. For p ∈ (1, ∞) the Muckenhoupt Ap characteristic of a

weight w is defined by
Z Z 0
p−1
[w]Ap := sup − w(x) dx − w1−p (x) dx ,
Q Q Q

where the supremum is over all (axes-parallel) cubes Q ⊆ Rd . We say that w

is an Ap weight if [w]Ap < ∞.
For a cube or quadrant Q0 ⊆ Rd , we define the local weight characteristic
[w]Ap (Q0 ) and the weight class Ap (Q0 ) in a similar way, but restricting the
supremum to cubes Q ⊆ Q0 only.

For the treatment of weighted norm inequalities, it is useful to introduce the

following simple but far-reaching idea:

Remark 11.3.18 (Dual weight trick). Given an operator T , a weight w and an

exponent p ∈ (1, ∞), consider an inequality of the form

kT (h)kLp (w) 6 CkhkLp (w) ∀h ∈ Lp (w). (11.41)

If σ is another weight, we observe that h = f σ is in Lp (w) if and only if

f ∈ Lp (σ p w). With this substitution, the previous estimate becomes

kT (f σ)kLp (w) 6 Ckf σkLp (w) = Ckf kLp (σp w) ∀f ∈ Lp (σ p w).

Equating the weights inside the operator and on the right hand side, we want
to arrange that σ = σ p w, i.e., that σ = w−1/(p−1) ; this is called the (Lp -)dual
weight of w. With this choice, the previous display reduces to

kT (f σ)kLp (w) 6 Ckf kLp (σ) ∀f ∈ Lp (σ), σ := w−1/(p−1) . (11.42)

Applying duality in Lp (w), yet another equivalent condition is given by the

conveniently symmetric formulation
Z
0
T (f σ) · gw 6 Ckf kLp (σ) kgkLp0 (w) ∀f ∈ Lp (σ), g ∈ Lp (w). (11.43)

Thus all three formulations (11.41), (11.42) and (11.43) are equivalent.

We now give the A2 theorem for the sparse operators AS . The simplicity
of this argument is a manifestation of the usefulness of dominating other
operators by the sparse ones.

Theorem 11.3.19 (Cruz-Uribe–Martell–Pérez). Let ε ∈ (0, 1) and S ⊆

D be ε-sparse. Let N ∈ Z+ be odd. If w ∈ A2 , then the sparse operator AS
N
is
2
bounded on L (w), and

N 4 2d
kAS kL (L2 (w)) 6 N [w]A2 .
ε
58 11 Singular integral operators

Proof. By the dual weight trick (Remark 11.3.18), with σ := w−1 we need to
prove that
Z
4 2d
ANS (f σ) · gw 6 N [w]A2 kf kL2 (σ) kgkL2 (w) ∀f ∈ L2 (σ), g ∈ L2 (w).
ε
Since AS is a positive operator, both g and h may be taken to be positive,
and there are no subtle convergence issues in the computation that follows.
We first observe that
Z Z
1 σ(Q) 1
hf σiQ = fσ = f σ = hσiQ hf iσQ ,
|Q| Q |Q| σ(Q) Q

where σ(Q) = Q σ and hf iσQ is the average of f with respect to the measure
R

induced by the weight σ. We denote the corresponding dyadic maximal op-

σ
erator by MD f := supQ∈D 1Q hf iσQ ; this operator is bounded on L2 (σ) with
norm 2 according to Doob’s maximal inequality (Theorem 3.2.2, cf. explana-
tions preceding Theorem 3.2.27) with p = p0 = 2.
We can then estimate, using that [w]A2 = supQ hwiQ hσiQ by definition,
Z X Z
AN
S (f σ) · gw = hf σi NQ 1Q · gw
Q∈S Rd
X
= hf σiN Q hgwiQ |Q|
Q∈S
X
= hσiN Q hwiQ hf iσN Q hgiw
Q |Q|,
Q∈S

where
hσiN Q hwiQ 6 hσiN Q hwiN Q N d 6 [w]A2 N d .
Hence
|E(Q)|
Z X
AN d
S (f σ) · gw 6 N [w]A2 hf iσN Q hgiw
Q ,
ε
Q∈S

where
hgiw w
Q 6 inf MD g(z)
z∈Q

by definition of the dyadic maximal operator. As for hf iσN Q , we observe by

Proposition 11.3.11 that the dilated cube N Q belongs to one of the N d dyadic
system D n;N , where n ∈ ZdN , and the average over N Q is then something that
appears in the corresponding maximal operator MD n;N . Hence
X |E(Q)|
hf iσN Q hgiw
Q (11.44)
ε
Q∈S
X X |E(Q)|
= hf iσN Q hgiw
Q
ε
n∈Zd
N
Q∈S
N Q∈D n;N
 11.3 Calderón–Zygmund operators and sparse bounds 59
X X
σ w |E(Q)|
6 inf MD n;N f · inf MD g ·
d
Q Q ε
n∈ZN Q∈S
N Q∈D n;N
Z
1 X X
σ w
6 MD n;N f · MD g
ε d E(Q)
n∈ZN Q∈S
N Q∈D n;N
Z
1 X σ w 1/2 1/2
6 MD n;N f · MD g · σ w
ε d R d
n∈ZN
1 X σ w
6 kMD n;N f kL2 (σ) kMD gkL2 (w)
ε d
n∈ZN
1 X
6 2kf kL2 (σ) · 2kgkL2 (w)
ε dn∈ZN
4 d
= N kf kL2 (σ) kgkL2 (w) .
ε
Substituting back, this gives the claimed bound for kAS kL (L2 (w)) . 

Corollary 11.3.20. Let ε ∈ (0, 1) and S ⊆ D be ε-sparse, and let Q0 ∈ D.

If N ∈ Z+ is odd, p ∈ (1, ∞), and w ∈ Ap , then the sparse operator AN
S is
bounded on Lp (w), and

N 2d max(1, p−1
1
)
kAN
S kL (Lp (w)) 6 cd,p [w]Ap .
ε
Proof. This is an immediate consequence of Theorem 11.3.19 and Rubio de
Francia’s Extrapolation Theorem J.2.1. (In the latter, φpr and cpr should be
replaced by φdpr and cdpr ; the omission of dependence on d is a systematic
typo in Theorem J.2.1 and its proof. This explains a need a constant cd,p
rather than just cp in the statement of the corollary.) 
It is also useful to record the following localised version:

Proposition 11.3.21. Let ε ∈ (0, 1) and S ⊆ D be ε-sparse, and let Q0 ∈ D.

If N ∈ Z+ is odd and w ∈ A2 (Q0 ), then the sparse operator AN
S is bounded
on L2 (Q0 , w), and
4 
kAN
S kL (L2 (Q0 ,w)) 6 N 2d + 1 [w]A2 (Q0 ) .
ε
The same result is true if the cube Q0 is replaced by a quadrant of Rd .

We start with a simple:

Lemma 11.3.22. For every Q ∈ D(Q0 ), there exists a cube Q

e such that
N Q ∩ Q0 ⊆ Q ⊆ Q0 and `(Q) = min{N `(Q), `(Q0 )}.
e e
60 11 Singular integral operators

Proof. If N Q ⊆ Q0 , we take Q e := N Q, and if N `(Q) > `(Q0 ), we define

Q := Q0 .
e
Let us finally consider Q ∈ D(Q0 ) such that N Q 6⊆ Q0 but N `(Q) < `(Q0 ).
Let first d = 1, so that both Q0 = [a, b) and Q are intervals. If N Q extends to
the left of a, then Qe := [a, a + N `(Q)) satisfies the desired properties. If N Q
extends to the right of b, then Q e := [b − N `(Q), b) works. For general d > 1
with Q = I1 ×· · ·× Id and Q0 = J1 ×· · ·× Jd , we take Q e := Ie1 ×· · ·× Ied , where
each Ii is built relative to the respective interval Ji as in the one-dimensional
e
construction just given. This completes the proof. 
Proof of Proposition 11.3.21. The norm on the left is the L2 (w)-norm of the
operator f 7→ 1Q0 AN S (1Q0 f ), i.e., both the domain and the range of the
operator is restricted to functions supported on Q0 . Since Q0 ∈ D, each
Q ∈ D ⊆ S that contributes to 1Q0 AS (1Q0 f ) satisfies either Q ⊆ Q0 or
Q ) Q0 . Letting S 0 := {Q ∈ S : Q ⊆ Q0 }, we hence have
Z
AN
S (1Q0 f σ) · gw
Q0
Z Z X
N
6 AS 0 (1Q0 f σ) · gw + h1Q0 f iN Q · gw =: I + II.
Q0 Q0 Q)Q
0

By the dual weight trick with σ = w−1 , estimating the left-hand side uniformly
over f ∈ L2 (Q0 , σ) and g ∈ L2 (Q0 , w) of unit norm is equivalent to bounding
N
kAS kL (L2 (Q0 ,w)) .
Term II is dominated by
X Z X |Q0 |Z ∞
X Z
− (1Q0 f ) = − (1Q0 f ) = 2−kd − (1Q0 f ),
|Q| N Q0
Q)Q0 N Q Q)Q0 N Q0 k=1

P∞ P∞
where k=1 2−kd 6 k=1 2−k = 1. Thus

II 6 1Q0 h1Q0 f iN Q0 kgkL2 (w) ,

L2 (w)

where
w(Q0 )1/2
Z
1Q0 h1Q0 f iN Q0 = f w1/2 σ 1/2
L2 (w) |N Q0 | Q0

w(Q0 )1/2
6 kf kL2 (Q0 ,w) σ(Q0 )1/2
|Q0 |
1/2
6 [w]A2 (Q0 ) kf kL2 (Q0 ,w) 6 [w]A2 (Q0 ) kf kL2 (Q0 ,w) .

We then turn to the main part I involving S 0 := {Q ∈ S : Q ⊆ Q0 }. We

can largely follow the proof of Theorem 11.3.19, but some care is needed to
 11.3 Calderón–Zygmund operators and sparse bounds 61

ensure that we only apply the A2 condition to cubes contained in Q0 , which

need not be the case with the dilated cubes N Q. We start with
X Z
I= h1Q0 f σiN Q 1Q · gw
Q∈S 0 Q0
Z Z
X 1
= fσ · gw
|N Q| N Q∩Q0 Q
Q∈S 0
X σ(N Q ∩ Q0 ) w(Q)
= hf iσN Q∩Q0 · hgiw
Q |Q|
0
|N Q| |Q|
Q∈S

By Lemma 11.3.22, for every Q ∈ S 0 ⊆ D(Q0 ), there is a cube Q

e such that
Q ⊆ N Q ∩ Q0 ⊆ Q ⊆ Q0 and `(Q) 6 N `(Q). Thus
e e

σ(N Q ∩ Q0 ) 6 σ(Q),
e w(Q) 6 w(Q),
e e 6 |N Q| = N d |Q|.
|Q|

Hence
σ(N Q ∩ Q0 ) w(Q) σ(Q)
e w(Q)
e
6 N d 6 [w]A2 (Q0 ) N d ,
|N Q| |Q| |Q|
e |Q|e

since Q
e is a cube contained in Q0 . Substituting back, and using sparseness, it
follows that
|E(Q)|
Z X
AN d
S 0 (1Q0 f σ) · gw 6 N [w]A2 (Q0 ) hf iσN Q∩Q0 · hgiw
Q .
Q0 0
ε
Q∈S

As in the proof of Theorem 11.3.19, we have hgiw w

Q 6 inf z∈Q MD(Q0 ) g. Also, us-
ing Proposition 11.3.11, each N Q belongs to one of the dilated dyadic systems
D n;N , where n ∈ ZdN . A key observation is that then also

C n;N := {N Q ∩ Q0 : Q ∈ D, N Q ∈ D n;N }

is a nested family with set-theoretic (if not geometric) properties matching

those of D(Q0 ): Each of the subfamilies

Ckn;N := {N Q ∩ Q0 ∈ C n;N : `(Q) = 2−k `(Q0 )}

n;N
is a partition of Q0 , and each Ck+1 refines the previous Ckn;N . Thus, the
corresponding maximal operators

MCσ n;N f := sup 1R hf iσR

R∈C n;N

are still instances of the Doob maximal operator with respect on abstract
filtered spaces. Repeating the computation (11.44) mutatis mutandis, we then
obtain
62 11 Singular integral operators
X |E(Q)|
hf iσN Q∩Q0 hgiw
Q
ε
Q∈S 0
1 X
6 kMCσ n;N f kL2 (σ) kMD
w
gkL2 (w)
ε d
n∈ZN
1 X 4
6 2kf kL2 (σ) · 2kgkL2 (w) = N d kf kL2 (σ) kgkL2 (w) .
ε d
ε
n∈ZN

Hence
4
I 6 N d [w]A2 (Q0 ) · N d kf kL2 (σ) kgkL2 (w) .
ε
In combination with the bound
II 6 [w]A2 (Q0 ) kf kL2 (σ) kgkL2 (w) .
Recalling that
Z Z Z X
AN
S (1Q 0 f σ) · gw 6 AN
S 0 (1Q0 f σ) · gw + h1Q0 f iN Q · gw
Q0 Q0 Q0 Q)Q
0

and the dual weight trick, we conclude the proof in the case of a cube.
If Q0 is replaced by a quadrant S, we note by density that it suffices to
consider the integrals above compactly supported f and g. But then, if Q0 is a
sufficiently large cube contained in the quadrant and having one corner at the
corner of the quadrant, then such f and g will be supported in Q0 . Thus the
previous considerations apply and give a bound in terms of [w]A2 (Q0 ) , which
is clearly dominated by [w]A2 (S) . 
An extension of Proposition 11.3.21 to p 6= 2 follows, in principle, by Rubio de
Francia’s Extrapolation Theorem J.2.1 just like Corollary 11.3.20 from The-
orem 11.3.19. Since Theorem J.2.1 was formulated for global Ap (Rd ) weights
only, we include some remarks about its local version. As a rule, all dyadic
considerations carry over without any change. However, one needs to play a
little attention to the interplay of dyadic and non-dyadic cubes in the local
setting. The following is a local variant of the Covering Lemma 3.2.26:
Lemma 11.3.23. For cubes Q ⊆ Q0 ⊆ Rd , there exist a vector α ∈ {0, 31 , 23 }d
and a dyadic cube
`(P )
log2
D ∈ D α (Q0 ) := {P + α(−1) `(Q0 )
`(P ) : P ∈ D(Q0 )} (11.45)
(the shifted dyadic cubes from Definition 3.2.25) such that
`(D) 6 3`(Q) and Q ⊆ D ⊆ Q0 .
`(P )
log
In (11.45), the point of the factor (−1) 2 `(Q0 ) is simply to alternate between
±1 with each consecutive generation of the dyadic cubes. We refer the reader
to the discussion preceding Lemma 3.2.26 for why such a factor is needed.
 11.3 Calderón–Zygmund operators and sparse bounds 63

Proof. If 3`(Q) > `(Q0 ), then clearly D := Q0 ∈ D(Q0 ) = D 0 (Q0 ) satisfies

the required properties.
Let then 3`(Q) < `(Q0 ). By Lemma 3.2.26 (a global version of the lemma
that we are proving), there exists a cube D as asserted, expect that we do
not know whether D ⊆ Q0 or not. If yes, then we are done, so suppose that
D 6⊆ Q0 . We will check that an appropriate shift of D will be a cube that we
are looking for.
Let first d = 1 so that Q0 = [a, b) as well as Q and D are just intervals. If
D extends to the left of a, then we can take D0 := [a, a + `(D)) ∈ D(Q0 ), and
if D extends to the right of b, then we can take D0 := [b − `(D), b) ∈ D(Q0 ).
Let then d > 1 be arbitrary, Q = I1 ×· · ·×Id ⊆ D = J1 ×· · ·×Jd ∈ D α (Rd ),
and Q0 := K1 × · · · × Kd . For each i ∈ {1, . . . , d}, we run the previous
construction: If Ji ⊆ Ki , we let Ji0 := Ji ∈ D αi (R). If Ji 6⊆ Ki , we let Ji0 be
the interval of lengths `(Ji ) that meets the same end-point of Ki as Ji . Then
0
Ji0 ∈ D(Ki ). Defining D0 := J10 × · · · × Jd0 , we have D0 ∈ D α , where αi0 = αi
if Ji ⊆ Ki and αi0 = 0 otherwise. This D0 in place of D satisfies the claimed
properties, and the proof of the lemma is complete. 
As in (3.36), we can now easily dominate the local maximal operator
Z
MQ0 f (x) := sup 1Q (x)− kf (y)k dy
Q⊆Q0 Q
cube

by the local dyadic maximal operators

Z
α
MQ 0
f (x) := sup 1 P (x)− kf (y)k dy, α ∈ {0, 31 , 23 }d
P ∈D α (Q0 ) P
P ⊆Q0

with
X
MQ0 f 6 3d max α
MQ 0
f 6 3d α
MQ 0
f. (11.46)
1 2
α∈{{0, 3 , 3 }d 1 2
α∈{{0, 3 , 3 }d

Proposition 11.3.24. Let p, r ∈ (1, ∞) and cube Q0 ⊆ Rd be a cube. Then

1/(p−1)
(1) kMQ0 f kLp (Q0 ,w) 6 cd p0 [w]Ap kf kLp (Q0 ,w) ;
(2) if a pair of functions (f, h) satisfies

khkLr (Q0 ,w) 6 φr ([w]Ar (Q0 ) )kf kLr (Q0 ,w)

for all w ∈ Ar (Q0 ), where φr is a non-negative increasing function, then

khkLp (Q0 ,w) 6 φdpr ([w]Ap )kf kLp (Q0 ,w)

for all w ∈ Ap (Q0 ), where each φdpr is a non-negative increasing function.

r−1
In particular, if φr (t) = cr tτ , then φdpr (t) 6 cdpr tτ max{ p−1 ,1} .
64 11 Singular integral operators

Proof. (1) follows by repeating the proof of Theorem J.1.1: the dyadic con-
siderations are unchanged, and in the last step of the proof, one replaces an
application of (3.36) by its localised version (11.46).
The proof of (2) is the same as the proof of Theorem J.2.1, except that the
all references to the maximal operator M are replaced by the local version
MQ0 and, accordingly, all applications of Theorem J.1.1 by case (1) of the
proposition that we already proved. (We note that the φpr and cpr should
be replaced by φdpr and cdpr already in Theorem J.2.1; the omission of the
dependence on d is a systematic typo in Theorem J.2.1 and its proof.) 

Corollary 11.3.25. Let ε ∈ (0, 1) and S ⊆ D be ε-sparse, and let Q0 ∈ D.

If N ∈ Z+ is odd, p ∈ (1, ∞), and w ∈ Ap (Q0 ), then the sparse operator AN
S
is bounded on Lp (Q0 , w), and
4  max(1, 1 )
kAN
S kL (Lp (Q0 ,w)) 6 cd,p N 2d + 1 [w]Ap (Q0 )p−1 .
ε
The same result is true if Q0 is replaced by a quadrant of Rd .

Proof. The case of a cube is immediate from case p = 2 established in Propo-

sition 11.3.21 and extrapolation established in Proposition 11.3.24(2). The
case of a quadrant follows from this by the same considerations as in the last
paragraph of the proof of Proposition 11.3.21. 
Thanks to sparse domination, we also obtain the corresponding results for
Calderón–Zygmund operators:

Theorem 11.3.26 (A2 theorem). Let X and Y be Banach spaces, p0 ∈

[1, ∞], and let
T ∈ L (Lp0 (Rd ; X), Lp0 ,∞ (Rd ; Y ))
with norm N0 be an operator with a Dini kernel K. Then for every p ∈ (1, ∞)
and every w ∈ Ap , the operator T extends uniquely to

T ∈ L (Lp (w; X), Lp (w; Y ))

with norm estimate

  max(1, 1 )
p−1
kT kL (Lp (w;X),Lp (w;Y )) 6 cd,p N0 + cK + kωK kDini [w]Ap

where cK , ωK are as in Definition 11.3.1.

The result remain true if Rd is systematically replaced by a cube Q0 ⊆ Rd
or a quadrant S ⊆ Rd , as the domain of the function spaces, in the definition of
the Calderón–Zygmund constants cK and kωK kDini , as well as in the definition
of the weight class Ap .
 11.3 Calderón–Zygmund operators and sparse bounds 65

Proof. Let us first consider the global case. Let f ∈ Lpc (w; X) be supported
on a compact set F . Denoting by σ = w−1/(p−1) the dual weight, we have
Z Z
0 0
kf k = kf kw1/p σ 1/p 6 kf kLp (w) σ(K)1/p < ∞,
K

so that f ∈ L1c (Rd ; X) as well, and T f is well defined by the Calderón–

Zygmund theorem 11.2.5. Then Theorem 11.3.15 guarantees the existence of
a 51 -sparse collection S ⊆ D such that, pointwise almost everywhere,

kT f (x)kY 6 cd cT (A5S kf kX )(x), cT = N0 + cK + kωkDini .

Thus, by Corollary 11.3.20, we have

kT f (x)kLp (w;Y ) 6 cd cT kA5S (kf kX )kLp (w)

1
max(1, p−1 )
6 cd cT cd,p [w]Ap kf kX (11.47)
Lp (w)
1
max(1, p−1 )
= cd,p cT [w]Ap kf kLp (w;X) .

Recalling the definition of cT , this is the required norm estimate for T re-
stricted to Lpc (w; X); since this subspace is dense in Lp (w; X), it allows to
uniquely extend T to the whole space with the same norm.
The proof in the case of a cube or a quadrant in place of Rd remains the
same, just using the local Corollary 11.3.25 in place of Corollary 11.3.20 to
replace (11.47) by

kT f (x)kLp (Q0 ,w;Y ) 6 cd cT kA5S (kf kX )kLp (Q0 ,w)

1
max(1, )
6 cd cT cd,p [w]Ap (Q0 )p−1 kf kX
Lp (Q0 ,w)
max(1, 1 )
= cd,p cT [w]Ap (Q0 )p−1 kf kLp (Q0 ,w;X) .

Corollary 11.3.27 (A2 theorem for the Hilbert transform). Let X be

a UMD space, p ∈ (1, ∞) and w ∈ Ap (R). Then the Hilbert transform
Z
1 f (t)
Hf (s) = lim dt
ε↓0 π |s−t|>ε s − t

extends uniquely to H ∈ L (Lp (w; X)) with

1
max(1, p−1 )
kHkL (Lp (w;X)) 6 cp [w]Ap ~2,X , ~2,X := kHkL (L2 (R;X)) .

Proof. Recall that the Hilbert transform is bounded on L2 (R; X) when X is

a UMD space (Theorem 5.1.13). In particular, taking T = H and p0 = 2
66 11 Singular integral operators

in Theorem 11.3.26, we have N0 6 ~2,X < ∞, using the notation from the
statement of that theorem. The kernel of the Hilbert transform is K(s, t) =
1
s−t , so that cK = 1 qualifies for the constant in Definition 11.3.1. Moreover,

1 1 s0 − s |s0 − t| 1
− 0 = 0
62 ∀|s − s0 | 6 |s − t|,
s−t s −t (s − t)(s − t) |s − t|2 2

so that we can take the modulus of continuity ω1 (u) = 2u in Definition 11.3.1.

Checking that ω2 (u) = 2u also works in entirely similar. Thus kωkDini =
R1
0
2u du
u = 2. Finally, it is easy to check that the norm ~2,X = kHkL (L (R;X))
2

is at least 1, say by Proposition 5.2.2, which says that H acts as multiplication

by −i on functions with Fourier transform supported on R+ . Thus N0 + cK +
kωkDini 6 ~2,X + 1 + 2 6 4~2,X . Substituting this into the result of Theorem
11.3.26 gives the claimed bound for kHkL (Lp (w;X)) . 

11.3.e Sharpness of the A2 theorem

Already in the scalar-valued case X = K, Corollary 11.3.27, and hence The-

orem 11.3.26, is sharp in its dependence on the weight characteristic [w]Ap .
In order to see this, we need to know about the behaviour of [w]Ap for some
concrete examples of weights, for which we can also estimate the weighted
norm of the Hilbert transform. The following important power weights will
serve this purpose:

Example 11.3.28 (Power weights). Let α ∈ R, p ∈ (1, ∞), w(x) = |x|α for
x ∈ Rd , and σ(x) = w(x)−1/(p−1) = |x|−α/(p−1) . Then

w ∈ Ap (Rd ) ⇔ w, σ ∈ L1loc (Rd ) ⇔ −d < α < d(p − 1),

and if these equivalent conditions holds, then

1  1 p−1
cd,p [w]Ap 6 6 Cd,p [w]Ap .
1+α p−1−α
To verify the claims of this example, we make use of the following:

Lemma 11.3.29. If Q ⊆ Rd is any cube, and Q̃ is a cube of the same size

centred at the origin, then

`(Q)−γ
Z Z
− |x| dx 6 − |x|−γ dx hd
−γ
, γ ∈ [0, d),
Q Q̃ d−γ
Z Z
− |x|γ dx > − |x|γ dx hd,Γ `(Q)γ , γ ∈ [0, Γ ], Γ > 0.
Q Q̃

Qd Qd
Proof. Let Q = i=1 Ii and Q̃ = i=1 I˜d . Then Q is the disjoint union of the
sets QI := i∈I (Ii ∩ I˜i ) × i∈{I (Ij \ I˜j ), where I ranges over all subsets
Q Q
 11.3 Calderón–Zygmund operators and sparse bounds 67

of {1, . . . , d}, and {I := {1, . . . , d} \ I . Of course Q̃ is a similar union over

Q̃I , defined by interchanging the roles of Ij and I˜j in QI .
Since `(Ij ) = `(I˜j ) is the common side-length of Q and Q̃, it follows that
also |Ij \ I˜j | = |I˜j \ Ij |. Since I˜j is centred at the origin, if xj ∈ Ij \ I˜j and
x̃j ∈ I˜j \ Ij , then |x̃j | 6 |xj |.
d
Now all x = (xi )i=1 ∈ QI are in measure-preserving correspondence with
x̃ = (x̃i )di=1 ∈ Q̃I , such that |xi | = |x̃i | for all i ∈ I , and |xj | > |x̃j | for all
j ∈ {I ; hence altogether |x| > |x̃|.
This implies inequalities like the first ones on each line of the lemma, for
QI and Q̃I in place of Q and Q̃, and thus also these inequalities as claimed,
by summing over all I ⊆ {1, . . . , d}.
To estimate the integrals over Q̃, we note that B(0, 21 `(Q)) ⊆ Q̃ ⊆
√
B(0, 12 d`(Q)), where, for α > −d,
Z cd `(Q)
(cd `(Q))d+α
Z
|x|α dx = rα rd−1 σd−1 dr = σd−1 ,
B(0,cd `(Q)) 0 d+α
thus
`(Q)α
Z √ `(Q)α
2−d−α σd−1 6 − |x|α dx 6 (2−1 d)d+α σd−1 .
d+α Q̃ d+α
For α = −γ ∈ (−d, 0], the quantities multiplying `(Q)α /(d+α) = `(Q)−γ /(d−
γ) are clearly uniformly bounded from above and away from zero, with bounds
depending on d only. Similarly, for α = γ ∈ [0, Γ ], the quantities multiplying
`(Q)α = `(Q)γ have this property, with bounds depending on d and Γ only.

Proof of Example 11.3.28. The second ⇔ in the claim is immediate.
Note that at least one of w and σ is |x| to a non-negative exponent, and
therefore locally integrable with a strictly positive integral over every cube Q.
Thus, in order that [w]Ap is finite, it is necessary that the other of the two
functions is locally integrable as well, showing the first ⇒ in the claim.
It remains to check that −d < α < d(p − 1) implies that w ∈ Ap (Rd ),
together with the claimed estimate for [w]Ap .
Let first √α > 0, and denote δQ := dist(Q,√0)/`(Q). For x ∈ Q, we have
|x| 6 (δQ + d)`(Q), and thus −Q w 6 (δQ + d)α `(Q)α . If δQ > 0, we also
R
−1 −α
have |x|−1 6 δQ `(Q)−1 , and hence (−Q σ)p−1 6 δQ `(Q)−1 . Thus
R

√ 
Z Z p−1 √ −α
 d α
sup − w − σ 6 sup (δQ + d)α δQ = 1+
Q:δQ >δ Q Q Q:δQ >δ δ
On the other hand, for any cube Q, it follows from Lemma 11.3.29 that
Z Z p−1 √  2d d α
− p−1
p−1
sup w − σ 6 sup (δQ + d)α `(Q)α α `(Q)
Q:δQ 6δ Q Q Q:δQ 6δ d − p−1
√  2d d  p−1
= (δ + d)α α
d − p−1
68 11 Singular integral operators

Fixing some δ = δd,p , it is then immediate that

cd,p c0d,p
[w]Ap 6 α p−1 =
(d − p−1 ) [d(p − 1) − α]p−1

For a matching lower bound, it is enough to consider just the unit cube Q, in
which case the estimates of Lemma 11.3.29 apply with Γ = d(p − 1) to give
that
Z Z p−1 p−1 p−1
 1  1
[w]Ap > − w − σ hd,p 1 · α h d,p .
Q Q d − p−1 d(p − 1) − α
1
This completes the proof for α ∈ [0, d(p − 1)), noting that 1+α hd,p 1 in this
case.
For α = −γ < 0, we note that
γ
n 1 p0 −1 op−1
[|x|−γ ]Ap = [|x| p−1 ]p−1
Ap0 hd,p 0 γ
d(p − 1) − p−1
p−1 1
= hd,p
d−γ d+α
γ
by applying the previous case to p−1 > 0 and p0 in place of α and p, and
noting that (p − 1)(p0 − 1) = 1. 
We are now fully equipped to confirm the sharpness of Corollary 11.3.27.

Proposition 11.3.30 (Buckley). Fix p ∈ (1, ∞), and suppose that φ :

[1, ∞) → [1, ∞) is an increasing function such that

kHkL (Lp (w)) 6 φ([w]Ap ) ∀w ∈ Ap ,

or even just for all power weights in Ap . Then

1
φ(t) > cp · tmax(1, p−1 ) ∀t > 1.

Proof. Let σ = w−1/(p−1) denote the dual weight. Using the dualised formu-
lation (11.43) of the Lp (w)-boundedness of T = H, and choosing f and g
with positively separated compact supports, so that the kernel representation
is available, we have
ZZ
1 f (y)σ(y)g(x)w(x)
dx dy 6 φ([w]Ap )kf kLp (σ) kgkLp0 (w) (11.48)
π x−y
for all such f and g. If these functions are non-negative with supp f ⊆ R−
and supp g ⊆ R+ , then the integrand is non-negative, and by monotone con-
vergence (11.48) persists even if the supports of f and g meet at the origin.
The crucial point in bounding the Hilbert transform form below is the
following observation: if h(y) = |y|−α 1(−1,0) (y), then for x ∈ (0, 1),
 11.3 Calderón–Zygmund operators and sparse bounds 69
1 x
y −α y −α 1 x−α
Z Z
1 1
Hh(x) = dy > dy = , (11.49)
π 0 x+y π 0 2x 2π 1 − α
1
which is essentially h again, but with a factor 1−α that blows up as α → 1−.
We now “test” (11.48) with two choices of (f, g, σ, w), so that (f σ, gw) is
either (|y|−α 1(−1,0) , 1(0,1) ) or (1(−1,0) , |y|−α 1(0,1) ), with α ∈ [0, 1). In either
case (11.49) shows that
1
x−α
Z
1 1 1
LHS(11.48) > dx = ,
2π 0 1−α 2π (1 − α)2

where we have accumulated a quadratic blow-up.

To estimate the right hand side of (11.48), we need to specify the in-
dividual functions, not just the products f σ and gw. In the first case, let
f = 1(−1,0) and σ(y) = w(y)−1/(p−1) = |y|−α ; thus w(y) = |y|α(p−1) and
g(y) = 1(0,1) (y)w(y)−1 = 1(0,1) (y)|y|−α(p−1) . Then
Z 1 1/p  Z 1 1/p0
0
kf kLp (σ) kgkLp0 (w) = x−α dx xα(p−1)(1−p ) dx
0 0 (11.50)
= 1/(1 − α).

noting that (p − 1)(p0 − 1) = 1, and Example 11.3.28 shows that [w]Ap 6

cp /(1 − α)p−1 . Thus, altogether, we have

1 1  cp  1
6 (11.48) 6 φ . (11.51)
2π (1 − α)2 (1 − α)p−1 1 − α

Denoting t = cp /(1 − α)p−1 , this reduces to

φ(t) > c̃p t1/(p−1) ∀t > cp . (11.52)

Since H 2 = −I, it is clear that kHkL (Lp (w)) > 1, and hence φ(t) > 1 >
c0p t1/(p−1) for t ∈ [1, cp ) as well.
In the second case, we take g = 1(0,1) and w(x) = σ(x)1−p = |x|−α ; thus
0 0
σ(x) = |x|α/(p−1) = |x|α(p −1) and f (x) = 1(−1,0) (x)|x|−α(p −1) . A computa-
tion like (11.50) gives exactly the same final result, only with a slightly differ-
ent intermediate step, and Example 11.3.28 shows that [w]Ap 6 cp /(1 − α).
With this quantity inside φ in (11.51), the substitution t = cp /(1 − α) then
gives
φ(t) > c̃p t ∀t > cp , (11.53)
and the same bound for t ∈ [1, cp ) follows from H 2 = −I as before. The two
lower bounds (11.52) and (11.53) together prove the proposition. 
70 11 Singular integral operators

11.4 Notes
Given the emphasis of these volumes in analysis of functions having their
range in a Banach space, we have chosen to keep the consideration related to
the domain of the functions relatively simple, concentrating on the canonical
case of the Euclidean space Rd and, with specific applications in the later
chapters in mind, its rather special subdomains—cubes and quadrants—only.
However, much of this theory could be developed on far more general do-
mains, notably on spaces of homogeneous type (espaces de nature homogène)
introduced by Coifman and Weiss [1971] and extensively studied ever since.
Since our treatment is heavily based on the dyadic cubes on Rd , we recall
that analogous constructions are also available in the mentioned generality.
The construction of a fixed family of sets, sharing the essential properties of
the standard dyadic cubes of Rd , is due to Christ [1990]. We also make use
of “adjacent” and “random” families of dyadic cubes; a reasonably compre-
hensive account of their analogues in spaces of homogeneous type is provided
by Hytönen and Kairema [2012] with several variants and elaborations due to
Auscher and Hytönen [2013], Hytönen and Martikainen [2012], Hytönen and
Tapiola [2014], and Nazarov, Reznikov, and Volberg [2013].

Section 11.1

This section deals with relatively classical topics but with some modern
flavour. In particular, the local oscillation decomposition of Theorem 11.1.12
dates essentially back to Lerner [2010] in the scalar-valued case. The vector-
valued generalisation, introducing the notion of λ-pseudomedian, was first
found by Hänninen and Hytönen [2014]. Our present proof streamlines the
original one.
Proposition 11.1.14 was proved by Katz and Pereyra [1999] in the scalar-
valued case via a multilinear estimate, and by Hänninen and Hytönen [2016]
as stated.
Theorem 11.1.30 on the vector-valued H 1 –BMO duality is essentially from
Bourgain [1986], although the present proof is different. In this circle of ideas,
we have only covered the relatively elementary part of the theory that does
not require any assumptions on the underlying Banach space. Note that The-
orem 11.1.30 says that BMOD (Rd ; X ∗ ) can be identified with an isometric
1
subspace of (HD,at (Rd ; X))∗ . The same proof works in the non-dyadic case,
where arbitrary cubes are allowed both in the definition of BMO and of the
Hardy space atoms. To describe the full dual (Hat 1
(Rd ; X))∗ , Blasco [1988] de-
fines a class of Banach space Y -valued measures BM O(Rd ; Y ). Among other
things, he shows that (Hat1
(Rd ; X))∗ = BM O(Rd ; X ∗ ) for every Banach space
X, whereas BM O(R ; Y ) = BMO(Rd ; Y ), if and only if Y has the Radon–
d

Nikodým property. A recent account with more information on the Banach

space valued H 1 and BMO can be found in Chapter 7 of Pisier [2016].
 11.4 Notes 71

Section 11.2

The material of this section is predominantly classical, and most of the results
would have been available in essentially the present form by the 1980’s, if not
earlier, even in the Banach space valued setting. The scalar-valued origins, of
course, date much further back.
The essence of Theorem 11.2.5 comes from Calderón and Zygmund [1952],
who consider the scalar-valued case (X = Y = L (X, Y ) = C)
and Dini kernels
x−y
of the special form K(x, y) = K(x − y) = |x − y|−d Ω |x−y| , where moreover
Ω dσ = 0. In contrast to Theorem 11.2.5, which extrapolates other Lp -
R
S n−1
bounds from an assumed a priori Lp0 -bound, Calderón and Zygmund [1952]
obtained their Lp -boundedness conclusions unconditionally, i.e., they also de-
duce the initial Lp0 -bound for p0 = 2 from their special assumptions on the
kernel. Once this is achieved, the extrapolation to other Lp -bounds is carried
out in much the same way as in the present treatment, particularly in the case
p < p0 . The fact that the extrapolation part of Calderón and Zygmund [1952]
argument remains valid under more general assumptions on the kernel was
observed by Hörmander [1960], who introduced the conditions, now bearing
his name, in Definition 11.2.1 in the case of scalar-valued convolution kernels
K(x, y) = K(x − y). What we have called the (operator-)Hörmander class
Hör was designated as K 1 by Hörmander [1960], who also defines a family of
related conditions K a with a parameter a ∈ [1, ∞]. Just like Hör = K 1 is rel-
evant for the extrapolation of Lp -boundedness, the condition K a permits the
extrapolation of Lp -to-Lq boundedness from one pair (p, q) with p1 − 1q = 1− a1
to other such pairs.
The first Banach space-valued generalisations, which used the operator-
Hörmander conditions, were found by Schwartz [1961] and, apparently inde-
pendently, by Benedek, Calderón, and Panzone [1962]. According to Garcı́a-
Cuerva and Rubio de Francia [1985], the fact that the mere Hörmander con-
dition (involving integrals of kK(s, t)x − K(s0 , t)xkY rather than kK(s, t) −
K(s0 , t)kL (X,Y ) ) is sufficient for results like Theorem 11.2.5 “should have been
observed by anyone trying to adapt the proof of [the Calderón–Zygmund the-
orem] to the vector valued case”, yet they “do not emphasize very much the
interest of this weaker condition since, in most of the applications of vec-
tor valued singular integrals, [the operator Hörmander condition] does hold.”
Rubio de Francia, Ruiz, and Torrea [1986] provided, in their own words, an
“updated review” of Benedek et al. [1962], incorporating several new devel-
opments in singular integrals into the vector-valued theory, and in particular
explicitly dealing with two-variable kernels K(s, t), as we have done here. Our
considerations related to c0 in Theorem 11.2.9 were inspired by Girardi and
Weis [2004].
A version of Theorem 11.2.5 for convolution kernels K(s, t) = K(s − t) is
also presented by Grafakos [2008], where (in contrast to our approach) the
upper extrapolation is achieved by a duality argument, and the interested
reader is referred to this work for details of that approach. Grafakos [2008] is
72 11 Singular integral operators

also explicit about the norm estimate in Theorem 11.2.5(3); this is certainly
well known, but often not spelled out in many references.

Section 11.3
The main body of this section consists of results from the 2010’s. Since the dis-
covery of the original forms of many of these results, there has been significant
activity in generalising and streamlining their proofs, as well as developing en-
tirely new approaches. As a result, our order of presentation deviates from the
historical timeline in favour of a smoother mathematical story. A main result
of this section is certainly the A2 Theorem 11.3.26, but the various Sparse
Domination Theorems 11.3.6, 11.3.14, and 11.3.15, originally developed as
tools for proving the A2 Theorem 11.3.26, have by now established them-
selves as results of intrinsic value and models for desirable type of domination
to search for in other situations.
Prehistory of the A2 theorem
In its scalar-valued and qualitative form (i.e., saying that T is bounded on
Lp (w), but without tracking the estimate for the operator norm), the result
goes back to Hunt, Muckenhoupt, and Wheeden [1973] in the special case
that T is the Hilbert transform (as in Corollary 11.3.27) and to Coifman and
Fefferman [1974] for all standard Calderón–Zygmund operators of convolu-
tion type. The question of sharp dependence of the weighted operator norms
kT kL (Lp (w)) on the weights constant [w]Ap was raised by Buckley [1993],
who settled the case of the Hardy–Littlewood maximal operator (Theorem
J.1.1) and obtained non-matching upper and lower bounds for Calderón–
Zygmund operators. In particular, Proposition 11.3.30 saying that an esti-
1
max(1, p−1 )
mate for kT kL (Lp (w)) can be no better than [w]Ap , is essentially from
Buckley [1993]. In many papers, results of this type a stated in a slightly
weaker form along the lines that “the power of [w]Ap can be no better than
1
max(1, p−1 )”. However, in some related questions, the sharp estimate is known
to exhibit behaviour different from a pure power law.
The question of Buckley [1993] gained new interest through the work of
Astala, Iwaniec, and Saksman [2001], who considered the following problem:
Let O ⊆ C be a domain and k ∈ (0, 1). What is the minimal q such that all
1,q ¯ | 6 k|∂f | (referred to as weakly quasiregular)
functions f ∈ Wloc (O) with |∂f
1,2
must in fact belong to f ∈ Wloc (O) (and then be called simply quasiregular)?
By results of Astala [1994], q > 1 + k suffices; by examples due to Iwaniec
and Martin [1993], q < 1 + k does not, leaving q = 1 + k as the critical case.
Astala, Iwaniec, and Saksman [2001] proved that q = 1+k is still sufficient for
the said self-improvement, under their conjecture that the Beurling–Ahlfors
transform
Z
1 f (y) dA(y)
Bf (z) := − lim , D(z, ε) := {y ∈ C : |y − z| < ε}
π ε→0 C\D(z,ε) (z − y)2
 11.4 Notes 73

satisfies the upper bound

kBkL (Lp (w)) 6 cp [w]Ap , p ∈ [2, ∞). (11.54)

Special cases of the A2 theorem

Shortly after being posed, the conjecture of Astala et al. [2001] was verified by
Petermichl and Volberg [2002], and another proof was found by Dragičević and
Volberg [2003]. Already Petermichl and Volberg [2002] observed that (11.54)
as stated may be derived from its special case p = 2 by keeping track of
the constants in the proof of Rubio de Francia’s extrapolation theorem as
presented, e.g., by Garcı́a-Cuerva and Rubio de Francia [1985]. This idea was
systematised by Dragičević, Grafakos, Pereyra, and Petermichl [2005], whose
results were treated in Appendix J and applied in the section under discussion.
The positive results for the Beurling–Ahlfors transform inspired the ques-
tion of sharp weighted bounds for other operators, and the special role of the
exponent p = 2 as the critical case for extrapolation gave rise to the name
“A2 conjecture”, several further cases of which were settled over the next few
years. In particular, the Hilbert transform (the scalar-valued case of Corollary
11.3.27) and the Riesz transforms were handled by Petermichl [2007, 2008], a
general class of sufficiently smooth odd kernels on R by Vagharshakyan [2010],
and powers of the Beurling–Ahlfors operator by Dragičević [2011]. All these
results relied on
(A) ad hoc representation formulas of special singular integrals in terms of
simple “dyadic shifts” as in the representation of Petermichl [2000] for
the Hilbert transform (see Theorem 5.1.13 and (5.20)), and
(B) Bellman function techniques for sharp weighted bounds of these shifts.
The component (B) behind these results was first challenged by Lacey, Peter-
michl, and Reguera [2010], who replaced it with
(C) “corona decompositions” to verify the “testing conditions” in a
(D) dyadic two-weight T (1) theorem of Nazarov, Treil, and Volberg [2008].
Shortly after, a much simpler alternative to either (B) or (C)–(D) was found
by Cruz-Uribe, Martell, and Pérez [2010], who in turn replaced it by methods
largely similar to the ones that we have used here:
(E) domination of dyadic shifts from (A) (not yet of singular integrals di-
rectly) by the sparse operators AS , and
(F) estimating kAS kL (L2 (w)) as in Theorem 11.3.19, whose proof follows
closely the original one from Cruz-Uribe et al. [2010],
However, component (A) of the original proofs remained unchallenged and,
being somewhat ad hoc for the specific singular integrals considered thus far,
restricted their extension to wider classes of operators.
74 11 Singular integral operators

The general A2 theorem

These limitations of (A) were overcome by Hytönen [2012], who found

(G) a general dyadic representation formula (a variant of which will be pre-
sented in Theorem 12.4.27) of all standard Calderón–Zygmund operators
in terms of a series of dyadic shifts of increasing complexity.
Moreover, (C) and (D) had to be replaced by
(C0 ) refinements of (C) to control the general shifts produced by (G), and
(D0 ) a difficult two-weight T (1) theorem of Pérez, Treil, and Volberg [2010]
about the singular integral itself, rather than the dyadic shifts as in (D).
A combination of (G), (C0 ), and (D0 ) gave the first proof of the A2 Theorem
11.3.26 for all standard Calderón–Zygmund operators in the scalar case.
In a matter of months since the announcement of Hytönen [2012] in 7/2010,
several variants and extensions were found. Streamlined versions and certain
improvements of the original approach were obtained in Hytönen, Pérez, Treil,
and Volberg [2014], Hytönen and Pérez [2013], and Hytönen [2017], which
appeared in arXiv in 10/2010, 3/2011, and 8/2011, respectively. At the same
time, alternatives to (C0 ) and (D0 ) by
(B0 ) elaborations of (B) with good control on the shift complexity
were obtained by Nazarov and Volberg [2013] (arXiv 4/2011) and Treil [2013]
(arXiv 5/2011), and these were used by Nazarov, Reznikov, and Volberg [2013]
(arXiv 6/2011) to give an extension of the A2 theorem to doubling metric
space domains in place of Rd . (Thus, the versions with a cube or a quadrant
that we have stated in Theorem 11.3.26 are but very particular instances of
the general domains in which the result may be formulated.)
Still over the same hectic months, Hytönen, Lacey, Martikainen, Orponen,
Reguera, Sawyer, and Uriarte-Tuero [2012] (arXiv 3/2011) combined the ap-
proach of Hytönen [2012] with input from the time–frequency techniques of
Lacey and Thiele [2000] to extend the A2 theorem to maximally truncated
Calderón–Zygmund operators
Z
T# f (x) = sup kTε f (x)k, Tε f (x) = K(x, y)f (y) dy. (11.55)
ε>0 |x−y|>ε

However, these results were shortly superseded by Hytönen and Lacey [2012]
(arXiv 6/2011) by a new approach combining (G) with elaborations of (E)
and (F) from the approach of Cruz-Uribe et al. [2010]:
(E0 ) domination of the general dyadic shifts from (G) by operators (essentially
like) AN
S , where arbitrarily large N appear, and
0 N
(F ) estimating kAS kL (L2 (w)) with bounds polynomial in log N (which re-
quires much more delicate analysis than Theorem 11.3.19).
 11.4 Notes 75

As a curiosity, the term “sparse” in its present usage seems to have been
introduced by Hytönen and Lacey [2012] (line below (∗) on page 2042). This
was pointed out by Andrei Lerner in his survey talk at the “AIM Workshop on
sparse domination of singular integrals” in San José, California, in 10/2017.

Simpler proofs

The difficulties with arbitrarily high shift complexity N , which seemed un-
avoidable in the general A2 theorem until this point, were finally eliminated
by Lerner [2013a,b] (arXiv 2/2012). These papers provide two different proofs
of the same main result, stating that

kT# f kF 6 cd,T sup kAS f kF , (11.56)

D,S

where T# is the maximal truncation (11.55) of a standard Calderón–Zygmund

operator, F is any Banach function space of Rd , and the supremum is taken
over all dyadic systems D and their sparse subcollections S . With T in place
of T# , this is slightly weaker than the pointwise estimate of Theorem 11.3.15
but, taking F = Lp (w), quite sufficient for bounding T (or T# ) on Lp (w).
The first proof of (11.56) by Lerner [2013a] still started with (G) and (E0 ),
but then proceeded with the key new idea of
N ∗
(H) domination of the adjoints (AS ) by the simple operators AS = A∗S .
(The fact that the argument passes through the adjoint is where the Banach
function space F is needed, while everything else can be estimated pointwise.)
The A2 estimate can then be completed by the simple step (F).
At the same time, Hytönen, Lacey, and Pérez [2013] found a way of re-
placing the initial steps (G) and (E0 ) by
(I) direct domination of the singular integral by an infinite series of operators
(essentially like) AN
S with arbitrarily large N .

Thus, a self-contained proof of the A2 theorem is obtained by concatenating

the steps (I), (H), and (F), and these constitute the simple proof of the A2
conjecture presented by Lerner [2013b]. As soon as things started falling into
the right place, the progress was very fast, and the preprints of the just dis-
cussed papers appeared in the arXiv essentially over a weekend in February
2012: Lerner [2013a] on Thursday 9th, Hytönen et al. [2013] on Friday 10th,
and Lerner [2013b] on Monday 13th.
The simple proof of Lerner [2013b] also admitted the first extension of
the A2 theorem to the weighted Bochner space Lp (w; X) by Hänninen and
Hytönen [2014]. At the time, the main difficulty with this Banach space valued
extension was the dependence of the sparse domination (I), via its use of
Lerner’s local oscillation formula (Theorem 11.1.12), on the notion of median.
Thus, a workable vector-valued version of this concept had to be developed;
it is reproduced in Section 11.1.
76 11 Singular integral operators

Pointwise sparse domination

Although not a necessity for proving the A2 theorem, the possibility of replac-
ing (11.56) by pointwise domination presented itself as a natural question,
which attracted some interest. This was independently achieved by Conde-
Alonso and Rey [2016] (arXiv 9/2014) and Lerner and Nazarov [2019] (also
announced and circulated around the same time in 2014, although in arXiv
only in 8/2015). These results still slightly deviated from Theorem 11.3.15 by
requiring a stronger form of the Dini condition,
Z 1/2
1
dt
ω(t) log2 < ∞.
0 t t
All Dini kernels were first covered by the “elementary” (but not so easy)
proof of Lacey [2017] (arXiv 1/2015), which was further quantified (in terms
of dependence on kωkDini ) by Hytönen, Roncal, and Tapiola [2017] (arXiv
10/2015) and remarkably simplified again by Lerner [2016] (arXiv 12/2015).
In proving Theorem 11.3.15, we have followed the further simplification due
to Lerner and Ombrosi [2020]. One advantage of their approach is a reduc-
tion of the prerequisites from classical Calderón–Zygmund theory necessary
to run their argument. On the technical level, this is achieved by replacing
the maximal operator

MT f (x) = sup sup T (1{5Q f )(y)

Q3x y∈Q

of Lerner [2016] by its “sharp” version MT# defined in (11.28). While MT#
can be estimated relatively directly, bounding the larger MT f originally re-
quired non-trivial classical results about the maximal truncations (11.55).
However, it was later observed by Almeida, Betancor, Fariña, and Rodrı́guez-
Mesa [2022] that the bounds for the two operators are actually equivalent
under general assumptions only involving the bounds for T that are used in
the theory anyway. Although not explicitly discussed by Lerner and Ombrosi
[2020], the present vector-valued extensions of their results, leading to Theo-
rems 11.3.15 and 11.3.26, involved little additional effort; this is in contrast to
the first vector-valued A2 theorem by Hänninen and Hytönen [2014]. Further
abstractions are due to Lorist [2021] and Lerner, Lorist, and Ombrosi [2022];
the latter work also explicitly addresses the vector-valued case.

Routes to sharpness in weighted estimates

There are some alternative routes to see the sharpness result of Proposition
11.3.30, which goes back to Buckley [1993] well before the matching upper
bounds were known. Luque, Pérez, and Rela [2015] made the curious observa-
tion that this can also be achieved without exhibiting any explicit examples
in the weighted situation, but studying instead the asymptotics of the un-
weighted norms kT kLp →Lp as p → 1 and p → ∞. This depends on a variant
 11.4 Notes 77

of Rubio de Francia’s Extrapolation Theorem J.2.1, where one keeps track

of the p-dependence in the estimates for kT kLp →Lp given by extrapolating a
bound of the type
kT kLp0 (w)→Lp0 (w) 6 φ([w]Aq0 ),
where q0 can also be different from p0 . Via contraposition, a lower bound
for kT kLp →Lp imposes a lower bound for φ. This quantitative weighted-to-
unweighted extrapolation was already used earlier by Fefferman and Pipher
[1997] in the “positive” direction to obtain sharp unweighted Lp -norm asymp-
totics for some operators by studying their weighted behaviour. They also ob-
tained a certain predecessor of the A2 Theorem 11.3.26 with kT kL (L2 (w)) 6
cd,T [w]A1 , where
Z  
[w]A1 := kM w/wk∞ = sup − w ess sup w−1
Q Q Q
Z Z p−1
> sup − w − w−1/(p−1) = [w]Ap ∀p ∈ (1, ∞).
Q Q Q

Further results

For a while, it might have seemed that the new sharp weighted technology
was essentially restricted to the class of Calderón–Zygmund operators. A cer-
tain discouragement against further extensions came from an observation of
Orponen [2013] that if an operator T has a dyadic representation (G) in the
sense of Hytönen [2012], then T must necessarily be a Calderón–Zygmund op-
erator. However, as soon as the role of (G) in the A2 theorem was challenged
by other methods, the door was also open for extensions beyond the standard
Calderón–Zygmund realm. Nevertheless, few could probably have expected
how far this theory could indeed be extended.
As an application of the sharp weighted estimates for Dini kernels discussed
above, Hytönen, Roncal, and Tapiola [2017] (arXiv 10/2015) showed that
rough homogeneous singular integrals
Z
Ω(t/|t|)
TΩ f (s) := p. v. f (s − t) dt, Ω ∈ L∞
0 (S
d−1
).
Rd |t|d

satisfy the weighted norm inequality

kTΩ kL (L2 (w)) 6 cd kΩk∞ φ([w]A2 )

with φ(u) 6 u2 . Although dealing with a class of operators outside the direct
scope of the sparse domination technology of the time, this result may never-
theless be seen as stretching those methods, rather than introducing genuinely
new ones, in that the operator TΩ was decomposed into a series of pieces in
the scope of the previously available tools by following a classical approach to
qualitative versions of similar results by Duoandikoetxea and Rubio de Francia
78 11 Singular integral operators

[1986], and Watson [1990]. A more intrinsic approach has been subsequently
developed by Conde-Alonso, Culiuc, Di Plinio, and Ou [2017], but φ(u) 6 u2
seems to remain the best available bound at the time of writing. In the other
direction, Honzı́k [2023] constructed examples of symbols Ω and weights w
to show that φ(u) > u3/2 ; hence the quantitative behaviour of TΩ is defi-
nitely different from the linear A2 theorem for standard Calderón–Zygmund
operators, but their precise bounds remain open.
Already a few weeks before Hytönen, Roncal, and Tapiola [2017] (late
10/2015 in arXiv), a far-reaching approach to sparse domination of a wide class
of operators had been revealed by Bernicot, Frey, and Petermichl [2016] (early
10/2015 in arXiv). They observed that several operators that act boundedly
in Lp only in some range (p0 , q0 ) ( (1, ∞) (and thus are definitely outside the
Calderón–Zygmund class by Theorem 11.2.5) can be proved to possess sparse
form domination of the type
X Z 1/p0 Z 0
1/q00
|hT f, gi| 6 C |Q| − |f |p0 − |g|q0 .
Q∈S 5Q 5Q

This in turn implies weighted norm inequalities of the form

α
kT f kLp (w) 6 C [w]Ap/p0 [w]RH(q0 /p)0 kf kLp (q) , p ∈ (p0 , q0 ),

where [w]RHt is the best constant in the reverse Hölder inequality

Z 1/t Z
t
− w 6 C− w,
Q Q

and α = α(p0 , q0 , p) is a certain explicit exponent depending on the indicated

quantities only.
Typical examples in the scope of the theory of Bernicot, Frey, and Peter-
michl [2016] are various “singular non-integral operators” arising in harmonic
analysis adapted to operators other than the classical Laplacian, e.g., gener-
alised Riesz transforms ∇L−1/2 , where L could be a second-order divergence-
form operator L = − div(A∇) with bounded coefficient matrix A, or a
Schrödinger operator L = −∆ + V with some potential V .
After the key observation that it is possible to go beyond Calderón–
Zygmund theory at all, sparse domination results and weighted norm inequal-
ities, as a corollary, for several different types of operators have been obtained:
• rough singular integrals (Conde-Alonso, Culiuc, Di Plinio, and Ou [2017],
Di Plinio, Hytönen, and Li [2020a]);
• Bochner–Riesz multipliers (Benea, Bernicot, and Luque [2017], Conde-
Alonso et al. [2017], Lacey, Mena, and Reguera [2019]);
• oscillatory integrals (Lacey and Spencer [2017], Krause, Lacey, and Wierdl
[2019]);
• bilinear Hilbert transforms and related phase-space objects (Culiuc, Di Plinio,
and Ou [2018a], Di Plinio, Do, and Uraltsev [2018]);
 11.4 Notes 79

• singular integrals along curves, Radon transforms (Cladek and Ou [2018],

Culiuc, Kesler, and Lacey [2019], Oberlin [2019], Anderson, Hu, and Roos
[2021]);
• spherical maximal operators both on Rd (Lacey [2019], Beltran, Ober-
lin, Roncal, Seeger, and Stovall [2022a], Borges, Foster, Ou, Pipher, and
Zhou [2023]) and on the Heisenberg group (Bagchi, Hait, Roncal, and
Thangavelu [2021], Ganguly and Thangavelu [2021]);
• pseudo-differential operators (Beltran and Cladek [2020]).
A relatively general theory has been developed by Beltran, Roos, and Seeger
[2022b], who also explicitly discuss Banach space valued operators.

Product space theory

A related direction, in which a weighted theory of singular integrals is well

developed since the works of Fefferman and Stein [1982] and Fefferman [1987,
1988], yet the sparse domination technology has met obstacles, consists of
the theory of product space or multi-parameter singular integrals modelled
after the product Hilbert transform H1 ⊗ H2 (where Hi denotes the Hilbert
transform in the ith variable of R2 ). Natural maximal operators in this theory
are the strong maximal operator
Z
M∗ f (s) := sup 1R (s)− kf (t)k dt.
R rectangle R

and its dyadic version, where the rectangles are restricted to be dyadic (i.e.,
products of dyadic intervals). Barron, Conde-Alonso, Ou, and Rey [2019] have
shown that it is impossible to dominate the strong dyadic maximal operator
by sparse forms based on rectangles with sides parallel to the axes, which
presents an obstacle to sparse techniques in this setting. While the most obvi-
ous extension of sparse domination is thus excluded, it was shown by Barron
and Pipher [2017] that one canR still obtain a workable substitute
R by replac-
ing the dominating averages −R |f | of f with the averages −R Sf of its dyadic
square function Sf on the right-hand side.
On the other hand, the original dyadic representation (G), while largely su-
perseded by sparse technology in applications to standard Calderón–Zygmund
operators, remains available, after natural modifications, in the product space
theory, as first proved by Martikainen [2012b] in the two-parameter case and
extended to arbitrarily many parameters by Ou [2017]. A vector-valued ap-
proach to this theory has been developed by Hytönen, Martikainen, and Vuori-
nen [2019a].

Sparse domination versus causality

While the current mainstream in sparse domination, evidenced by the previ-

ous list, consists of proving and applying domination for ever wider classes
of operators, one may also pose a somewhat opposite question: Suppose that
80 11 Singular integral operators

a given (say, standard Calderón–Zygmund) operator T possesses some addi-

tional properties. Can this be reflected in the dominating sparse operator as
well? A concrete instance of such an additional property is causality. Suppose
for simplicity that d = 1, and that K(s, t) is non-zero only if s > r; thus T f (s)
depends only on the “past” values f (t) with t < s. If T is a Calderón–Zygmund
operator, then it satisfies the sparse domination T f (s) 6 cT A5S f (s) by the
5
general theory. However, the dominating sparse operator AS is no longer
causal. Is it possible to exploit the causality of T to obtain a sharper form
of sparse domination, where this causality is preserved also in the right-hand
side? Some partial (but far from complete) results in this direction have been
obtained by Hytönen and Rosén [2023].
Aimar, Forzani, and Martı́n-Reyes [1997] have shown that causal Calderón–
Zygmund operators remain bounded on the weighted space Lp (w) for the
larger class of one-sided Ap weights, defined by the finiteness of
Z c  Z b p−1
1  1
[w]Ap− := sup p
w w− p−1 ,
−∞<a<b<c<∞ (c − a) b a

but the optimal bound for the operator norm kT kL (Lp (w)) in terms of [w]A−
p
remains open. In analogy with the A2 Theorem 11.3.26, it is natural to make:

Conjecture 11.4.1 (One-sided A2 conjecture of Chen, Han, and Lacey [2020]).

For all causal Calderón–Zygmund operators,
1
kT kL (Lp (w)) 6 cT ([w]A−
p
)max(1, p−1 ) .

Partial results for Haar multipliers (see Section 12.1.a) in place of Calderón–
Zygmund operators are obtained by Chen et al. [2020], but beyond that the
conjecture remains open.
Causal operators appear very naturally; e.g., the operator-valued kernel

K(s, t) = 1R+ (s − t)Ae−(s−t)A ,

of relevance to the maximal regularity problem studied in Chapter 17, has

this form. A theory of one-sided singular integrals applicable to this operator-
valued situation has been developed by Chill and Król [2018].

Matrix weighted spaces and convex body domination

Let W : Rd → RN ×N be a matrix weight, i.e., measurable and positive definite

almost everywhere, and f : Rd → RN be measurable. The norm
Z
kf k2L2 (W ) := hW (t)f (t), f (t)i dt
d
ZR
1 1
= |W (t) 2 f (t)|2 dt = kW 2 f k2L2 (Rd ;RN ) ,
Rd
 11.4 Notes 81

appears naturally from the prediction theory for multivariate stationary

stochastic processes n ∈ Z 7→ ξn ∈ L2 (Ω; RN ) developed by Wiener and
Masani [1958], where stationarity means that Γn−k := Eξn ξkT ∈ RN ×N de-
pends only on the difference of the discrete times n, k ∈ Z. If W is the density
of the spectral measure of the process, i.e., Γk = Wc (k) are the Fourier coef-
1 N ×N
ficients of W ∈ L (T; R ), the boundedness of the Hilbert transform on
L2 (W ) is equivalent to a positive angle between the past and the future of the
process. Even for N = 1, this problem was only solved 15 years later by Hunt,
Muckenhoupt, and Wheeden [1973], who characterised this boundedness in
terms of the A2 condition. For N > 1, it took over 20 more years before the
solution was obtained by Treil and Volberg [1997], who identified the correct
analogue of the A2 condition in the matrix-valued case:
1/2 1/2
[W ]A2 := sup |hW iQ hW −1 iQ |2 ,
Q

where | | is (say) the operator norm on L (RN ) (but the choice of the norm
on RN ×N is irrelevant, as they are all equivalent).
With the natural definition
1
kf kLp (W ) := kW p f kLp (Rd ;RN ) ,

one is led to inquire about the boundedness of the Hilbert transform on

Lp (W ). The characterising matrix-Ap condition, identified via different ap-
proaches by Nazarov and Treil [1996] and Volberg [1997], is less intuitive for
p 6= 2. It is perhaps most easily formulated with the help of the classical the-
orem of John [1948], which guarantees that every norm on RN is equivalent
(with constants depending only on N ) to a Euclidean norm, whose unit ball
is a linear transformation of the standard unit ball. If W is a matrix weight
1
and V := W p , it is easy to see that
Z 1/p
e ∈ Rn 7→ − |V (t)e|p dt
Q

is a norm, and hence, by the theorem of John [1948], there is a positive definite
reducing operator [V ]Q,p ∈ RN ×N , such that
Z 1/p √
|[V ]Q,p e| 6 − |V (t)e|p dt 6 N · |[V ]Q,p e|.
Q

The matrix-Ap condition may then be defined by the finiteness of the constant
1
[W ]Ap := sup |[V ]Q,p [V −1 ]Q,p0 |p , V := W p .
Q

1
The reader is invited to check that [V ]Q,p = hV p iQ
p
if N = 1 or p = 2 (but not
in general otherwise), so that the different definitions of Ap are consistent. It
82 11 Singular integral operators

is possible to give an equivalent definition of the matrix Ap condition with-

out reference to reducing operators, but one would still need them to prove
anything interesting, which is why we prefer to state the definition as above.
While the qualitative boundedness of the Hilbert transform, and in fact
of more general Calderón–Zygmund operators, on Lp (W ) was settled in the
mentioned papers, the proof of the scalar-valued A2 theorem raised the natural
question of its extension to the matrix-weighted case. This remains open, but
several related results have been achieved.
While sparse domination is perfectly applicable to vector-valued (even Ba-
nach space valued) functions, as we have seen in this chapter, it loses essential
directional information, which makes it ill-suited for matrix-weighted consid-
erations. To address this drawback, Nazarov, Petermichl, Treil, and Volberg
[2017] invented a refined notion of convex body domination, where the averages
hkf kiQ are replaced by the related convex bodies
n o
hφf iQ : kφkL∞ (Q) 6 1 ⊆ RN , f ∈ L1 (Q; RN ).

Convex body domination of T is most easily stated in its bilinear form, as an

elaboration of the sparse form domination
X
|hT f, gi| 6 cd,T |Q|h|f |i5Q h|g|i5Q
Q∈S
X 1 ZZ (11.57)
= c0d,T |f (s)||g(t)| ds dt.
|Q| 5Q×5Q
Q∈S

Convex body domination of T can now be stated in the form

X cd,N,T ZZ
|hT f, gi| 6 sup φ(s)(s) · ψ(t)g(t) ds dt , (11.58)
|Q| kφk∞ 61 5Q×5Q
Q∈S
kψk∞ 61

with the important difference that we take the dot product of f (s), g(t) ∈ Rn
first, and only then the absolute value of the result; this allows for critical
directional cancellation compared to (11.57).
The proof of Nazarov, Petermichl, Treil, and Volberg [2017] (arXiv 1/2017),
that standard Calderón–Zygmund operators satisfy (11.58), follows the same
lines as the proof of Theorem 11.3.15 but with important elaborations at a
few selected points, making again use of the ellipsoid theorem of John [1948].
On the other hand, with (11.58) available, Nazarov et al. [2017] can prove the
bound
3/2
kT kL (L2 (W )) 6 cd,T [W ]A2 ,
which remains the best available matrix-weighted estimate for Calderón–
Zygmund operators (or even just for the Hilbert transform) at the time of
writing. A variant of the same results was also obtained by Culiuc, Di Plinio,
and Ou [2018b], seemingly earlier (arXiv 10/2016) but not independently; ac-
cording to their acknowledgment, the concept of domination by convex body
 11.4 Notes 83

averages was introduced to these authors by Sergei Treil during his seminar
talk at Brown University in the Spring of 2016.
Since then, further applications and extensions of convex body dom-
ination have been explored by Cruz-Uribe, Isralowitz, and Moen [2018],
Di Plinio, Hytönen, and Li [2020a], Isralowitz, Pott, and Rivera-Rı́os [2021],
Isralowitz, Pott, and Treil [2022], and Muller and Rivera-Rı́os [2022]. Impor-
tantly, Bownik and Cruz-Uribe [2022] extended the Rubio de Francia algo-
rithm (Proposition J.2.2), and its key application to weighted extrapolation
(Theorem J.2.1), to matrix-valued weights, by further development of the
convex body philosophy.
An abstract framework for convex body domination has been proposed by
Hytönen [2023], allowing also Banach space valued functions in the theory.
While genuinely operator-valued weights in infinite dimensions seem to be
out of reach, this framework allows the treatment of RN ×N -valued weights
on spaces of X N -valued functions. In particular, the following simultaneous
extensions of the boundedness of the Hilbert transform on the Banach space
valued L2 (R; X) by Burkholder [1983], and on the matrix-weighted L2 (W ) by
Treil and Volberg [1997], is obtained there.

Theorem 11.4.2. Let X be a UMD space, and W : Rd → RN ×N be a matrix

A2 weight. Then the Hilbert transform H extends boundedly to
n 1
o
L2 (W ; X N ) := f : R → X N : kf kL2 (W ;X N ) := kW 2 f kL2 (R;X N ) < ∞

3/2 2 3/2
and satisfies kHkL (L2 (W ;X N )) 6 cN ~2,X [W ]A2 6 cN β2,X [W ]A2 , where
~2,X = kHkL2 (R;X) and β2,X is the UMD constant.

The stated quantitative formulation in terms of ~2,X is not explicit in Hytönen

[2023], but can be tracked in the proof, in a similar way as in Corollary 11.3.27
in the text.

A summary of sharp weighted bounds for classical operators

Our discussion above has been focused on norms of Calderón–Zygmund singu-

lar integrals and their various extensions, viewed as operators on a weighted
Lp (w) (or matrix-weighted Lp (W )) space; these are referred to as strong-type
bounds. We will briefly summarise results in two closely related directions.
First, one may inquire about the corresponding weak-type bounds, i.e., op-
erator norms in L (Lp (w), Lp,∞ (w)). These are obviously dominated by the
strong-type norms, but the point is that the optimal weak-type norms may
be significantly smaller in some cases, which gives these questions an indepen-
dent interest. Second, one may pose the same questions for various square-
functions, which could be viewed as part of the extended family of (vector-
valued, when suitably interpreted) Calderón–Zygmund operators; however, it
turns out that these operators are actually slightly “better” in terms of the
84 11 Singular integral operators

dependence of their norms on the weight constant. A basic example is the

dyadic square function
X 1/2
Sf (x) := |DQ f (x)|2 ,
Q∈D

(where the operators DQ are defined in (12.1) and discussed extensively in

Chapter 12), but several other classical square functions satisfy the same
weighted bounds; we refer the reader to the papers quoted below for details.
A summary of the sharp bounds known for these operators is as follows:

Singular integrals:

For p ∈ (1, ∞) and w ∈ Ap , the sharp estimates in Lp (w) are:

1
max(1,
p−1 )
(1) the strong-type bound is [w]Ap (Hytönen [2012]);
(2) the weak-type bound is [w]Ap (Hytönen, Lacey, Martikainen, Orponen,
Reguera, Sawyer, and Uriarte-Tuero [2012]);
(3) the weak-type L1 (w) bound is [w]A1 (1 + log[w]A1 ) (the upper bound was
proved by Lerner, Ombrosi, and Pérez [2009], its sharpness is due to
Lerner, Nazarov, and Ombrosi [2020]).
A speculative linear-in-[w]A1 bound in (3) was known as the A1 conjecture,
or the weak Muckenhoupt–Wheeden conjecture. The original conjecture, dis-
proved by Reguera [2011] and Reguera and Thiele [2012], was about the
boundedness of T : L1 (M w) → L1,∞ (w) for any weight w. This holds for
M in place of T (Theorem 3.2.27), which motivated the conjecture.

Square functions:

For the range of p as specified and w ∈ Ap , the sharp estimates in Lp (w) are:
1
max( 12 , p−1 )
(4) the strong-type bound is [w]Ap for p ∈ (1, ∞) (Lerner [2011]);
max( 1 , 1 )
(5) the weak-type bound is [w]Ap 2 p for p ∈ [1, ∞) \ {2} (p = 1: Chanillo
and Wheeden [1987], Wilson [2007, 2008]; p ∈ (1, 2): Lacey and Scurry
[2012]; p > 2: Hytönen and Li [2018]);
1 1
(6) the weak-type L2 (w) bound is at most [w]A2 2 (1 + log[w]A1 ) 2 (Domingo-
Salazar, Lacey, and Rey [2016]), but its sharpness seems to remain open
(see Ivanisvili and Volberg [2018] for partial related results).

In contrast to singular integrals, the bounds at p = 1 above are consequences of

the stronger statement that S : L1 (M w) → L1,∞ (w) is bounded for any weight
w, i.e., the Muckenhoupt–Wheeden conjecture holds for square functions. This
also explains the (implicit) appearance of sharp weighted bounds in Chanillo
and Wheeden [1987], long before this became a fashionable topic.
 11.4 Notes 85

For matrix-weights, the only known sharp estimates among these exam-
ples, at the time of writing, seem to be the square function bounds (4) for
p ∈ (1, 2]; this was proved by Hytönen, Petermichl, and Volberg [2019b] for
p = 2 and extended by Isralowitz [2020] to p ∈ (1, 2).
12
Dyadic operators and the T (1) theorem

In Chapter 11, we have mainly dealt with a situation, where a bounded linear
operator on some Lp0 (Rd ; X) space is given, and we have then explored its
bounded extensions to other spaces including Lp (Rd ; X) for p 6= p0 . We now
turn to a somewhat different (and often more difficult) question of recognising
such bounded operators to begin with.
Before addressing this question for the Calderón–Zygmund type operators
of the kind studied in Chapter 11, we investigate a number of related objects
in a simpler dyadic model. Besides serving as an introduction to some of the
key techniques, it turns out that these dyadic operators can be, and will be,
also used as building blocks of the proper singular integral operators towards
the end of the chapter.
The dyadic operators will be of two essentially different types. The first
class, which we vaguely refer to as “dyadic singular integrals” in Section 12.1,
consist of a somewhat diverse family of relatives of the prototype dyadic shifts
encountered in Chapter 5, where they we used to represent the prototype sin-
gular integral given by the Hilbert transform. It is thus only natural that a
family of dyadic operators generalising this basic dyadic shift will serve as
building block of the Calderón–Zygmund family of singular integrals gener-
alising the Hilbert transform. Martingale techniques vaguely reminiscent of
those in Section 5.1, but of somewhat higher complexity probably by neces-
sity, will feature in the argument to put the UMD property of the underlying
Banach space into action.
The second class of dyadic operators consists of so-called paraproducts,
which we discuss in Section 12.2. These are new creatures of the non-
convolution realm that we have entered and they will vanish (as we will see) as
soon as we occasionally specialise our considerations to singular integral of the
convolution form. However, for the representation the full class of Calderón–
Zygmund operators they will turn out be quite essential.
The chapter will culminate in a lengthy treatment of the so-called T (1)
theorem, a general criterion for boundedness of singular integral operators.
We will first discuss a version for abstract bilinear form in Section 12.3, and
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 87
T. Hytönen et al., Analysis in Banach Spaces, Ergebnisse der Mathematik und ihrer
Grenzgebiete. 3. Folge / A Series of Modern Surveys in Mathematics 76,
https://doi.org/10.1007/978-3-031-46598-7_2
88 12 Dyadic operators and the T (1) theorem

only then, in the final Section 12.4, turn to the task of checking the assump-
tions of the abstract result for singular integral operators with a Calderón–
Zygmund kernel, of the kind that we met Chapter 11. However, in order to
establish boundedness on Lp (Rd ; X) from scratch, rather than extrapolating
it from another Lp0 (Rd ; X) space where it was already known (as in Chapter
11), somewhat stronger versions of the Calderón–Zygmund conditions will be
needed, and the notion of R-boundedness from Chapter 8 will, once again,
play a prominent role. While the results of this chapter will generically be
established in arbitrary UMD spaces, it turns out that additional information
about type and cotype, as studied in Chapter 7 can be traded against the pre-
cise kernel conditions, so that slightly larger classes of kernels are admissible
under conditions of type and cotype of the underlying space.

12.1 Dyadic singular integral operators

In this section, we introduce and study a family of dyadic models of singular
integrals, starting from the simplest case of Haar multipliers and proceeding
to their more complicated relatives. All these operators will eventual come
together as parts of a decomposition of general singular integral operators
towards the end of the chapter.
Since our aim is not to assume, but to prove, the Lp -boundedness of the
relevant operators, we will first define their action on appropriate spaces of
test functions only.

Definition 12.1.1 (Classes of simple functions). For a collection C of

bounded Borel subsets of Rd , let
n o
S(C ; X) := span 1C ⊗ x : C ∈ C , x ∈ X ,
n Z o
S0 (C ; X) := f ∈ S(C ; X) : f (t) dt = 0 ,
Rd
Sloc (C ; X) := {f ∈ 1
Lloc (Rd ; X) : 1C f ∈ S(C ; X) for all C ∈ C },
∞
S∞ (C ; X) := Sloc (C ; X) ∩ L (Rd ; X).

It is easy to see that S(C ; X) ⊆ Lp (Rd ; X) for all p ∈ [1, ∞], and that

S0 (C ; X) ⊆ S(C ; X) ⊆ S∞ (C ; X) ⊆ Sloc (C ; X).

Our primary case of interest will be when C = D is a collection of dyadic

cubes of Rd in the sense of Definition 11.1.6. In this case, S(C ; X) is dense
in Lp (Rd ; X) for all p ∈ [1, ∞). In (12.2) below, we will add yet another
space S00 (D; X) ⊆ S0 (D; X) to this list, but its introduction requires some
preliminaries.
 12.1 Dyadic singular integral operators 89

12.1.a Haar multipliers

We begin with what is arguably the simplest class of operators deserving the
name of “dyadic singular integrals”. In essence, we have encountered these
operators already, at least implicitly on the one-dimensional domain space
R1 , where we dealt with operators of the form
X
f 7→ I hf, hI ihI
I∈D

and showed their uniform boundedness on Lp (R; X) for arbitrary unimodular

coefficients I , assuming that p ∈ (1, ∞) and X is a UMD space (see Theorem
4.2.13). We now wish to extend these consideration to the general Euclidean
domain Rd . This hardly presents any new challenges, and mainly serves as a
warm-up for the subsequent considerations.
We first recall and extend the notation related to conditional expectations
and martingale differences over the dyadic filtration of Rd . For any cube

Q = aQ + `(Q)[0, 1)d ,

with sidelength `(Q) > 0 and “lower left” corner aQ ∈ Rd , we denote by

n 1 o
ch(Q) := aQ + `(Q)([0, 1)d + α) : α ∈ {0, 1}d
2
the collection of its 2d “children” obtained by bisecting each of the intervals
in the Cartesian product defining Q. In particular, for

Q ∈ Dk := {2−k ([0, 1)d + n) : n ∈ Zd },

we have
ch(Q) = {Q0 ∈ Dk+1 : Q0 ⊆ Q}.
For every cube Q, we define the conditional expectation and martingale
difference projections (acting on f ∈ L1loc (Rd ; X))
Z X
EQ f := 1Q − f dx, DQ f := EQ0 f − EQ f. (12.1)
Q Q0 ∈ch(Q)

Then for every k ∈ Z, we let

X
Ek f := E(f |σ(Dk )) = EQ f,
Q∈Dk
X
Dk f := Ek+1 f − Ek f = DQ f.
Q∈Dk

We still want to express the martingale difference projections DQ in terms

of vector-valued extensions of rank-one operators on scalar-valued functions.
90 12 Dyadic operators and the T (1) theorem

In dimension d = 1, the operators already have this form, as we recall from

Lemma 4.2.11 and the preceding discussion:
DI f = hf, hI ihI , hI = |I|−1/2 (1I− − 1I+ ),
where hI is called the Haar function associated with the interval I.
In higher dimensions, there are various ways of constructing analogues of
the Haar functions. For the present purposes, a standard tensor construction
suffices. In d = 1, we denote
hI1 := hI , h0I := |I|−1/2 1I .
Lemma 12.1.2. In general dimension d > 1, the (tensor-)Haar functions
d
(α ,...,α )
Y
α
hQ 1
(x) = hI1 ×···×Id
d
(x1 , . . . , xd ) := hIαii (xi ), α = (α1 , . . . , αd ) ∈ {0, 1}d .
i=1

1
satisfy the following identity for all f ∈ Lloc (Rd ; X):
X X
α
DQ f = hf, hQ ihα
Q =:
α
DQ f.
α∈{0,1}d \{0} α∈{0,1}d \{0}

Proof. From the (obvious) orthogonality of one-dimensional Haar functions,

it follows that
d
Y d
Y
β βi
hhα
Q , hQ i = hhα
Ii , h Ii i =
i
δαi ,βi = δα,β .
i=1 i=1

Let HQ be the space of scalar-valued functions supported on Q, constant on

each dyadic child of Q, and of mean zero. Clearly dim HQ = (2d − 1) and
α
hQ ∈ HQ for each α ∈ {0, 1}d \ {0}. Since these hQ α
are orthonormal and their
number is equal to dim HQ , they must form an orthonormal basis of HQ . On
the other hand, one easily verifies that DQ is the orthogonal projection of
L2 (Rd ) onto HQ , so in particular DQ f = f for all f ∈ HQ . Since the hQ α

form an orthonormal basis, the claimed identity is true for all f ∈ HQ . If

1
f ∈ Lloc (Rd ; X) and x∗ ∈ X ∗ , then hDQ f, x∗ i ∈ HQ and thus
X D X E
hDQ f, x∗ i = hhDQ f, x∗ i, hQ
α α
ihQ = α
hDQ f, hQ ihα
Q, x
∗
,
α∈{0,1}d \{0} α∈{0,1}d \{0}

where
X
hDQ f, hα
Qi =
α
hEQ0 f, hQ i − hEQ f, hQ i
Q0 ∈ch(Q)
X
= hf, EQ0 hα α
Q i = hf, hQ i.
Q0 ∈ch(Q)

The claimed identity follows, since the functionals x∗ ∈ X ∗ separate the points
x ∈ X by the Hahn–Banach theorem. 
 12.1 Dyadic singular integral operators 91

α
The functions hQ , with α ∈ {0, 1}d \ {0}, are referred to as cancellative Haar
functions, as they all have vanishing mean. In contrast, hQ 0
= |Q|−1/2 1Q is the
non-cancellative Haar function on Q. In the wavelet literature, the cancellative
Haar functions are special cases of mother wavelets, while the non-cancellative
Haar function is the father wavelet.

Lemma 12.1.3. Let X be a Banach space and p ∈ (1, ∞). Then the space of
finite linear combinations of cancellative Haar functions with X-coefficients,
n o
S00 (D; X) := span hQ α
⊗ x : Q ∈ D, α ∈ {0, 1}d \ {0}, x ∈ X , (12.2)

is dense in Lp (Rd ; X).

Proof. The filtration generated by the dyadic cubes, (Fk )k∈Z := (σ(D Sk ))k∈Z is

σ-finite with respect to the Lebesgue measure on Rd , and F∞ := σ k∈Z Fk
is the Borel σ-algebra of Rd . Hence Ek f → f in Lp (Rd ; X) as k → ∞ for all
f ∈ Lp (Rd ; X) by the forward convergence of generated martingales (Theorem
3.3.2). On the other hand, F−∞ := k∈Z Fk contains only sets of Lebesgue
T
measure 0 (the empty set) or ∞ (the quadrants, and their unions), which
means (by definition) that the Lebesgue measure is purely infinite on F−∞ .
Thus Ek f → 0 in Lp (Rd ; X) as k → −∞ for all f ∈ Lp (Rd ; X) by the
backward convergence of martingales (Theorem 3.3.5).
Combining these observation P about the limits at ±∞, it follows that func-
M −1
tions of the form EM f − Em f = k=m Dk f are dense in Lp (Rd ; X). Next, we
make the following observations about each Dk appearing in this expansion.
First, for any P ∈ Dm , multiplication with 1P commutes with Dk ; second,
∞
1P Dk f is a finite linear combination of some DQ f , and finally, if (Pi )i=1 is
PN
an enumeration of Dm , then i=1 1Pi f → f in L (R ; X) as N → ∞. Thus
p d

finite linear combinations of DQ f are dense in Lp (Rd ; X). Finally, Lemma

12.1.2 shows that DQ f ∈ S00 (D; X), and completes the proof. 

Remark 12.1.4. One can check that

n Z o
S00 (D; X) = f ∈ S(D; X) : f = 0 for each quadrant D of Rd .
D

In particular, if D is a connected tree of dyadic cubes (i.e., every two cubes

are contained in a common bigger dyadic cube), then S00 (D; X) = S0 (D; X).
Making this connectedness assumption would slightly simplify some consid-
erations, but have the disadvantage of excluding the standard dyadic system
(cf. Remark 11.1.9).

After these preparatory considerations, we are in a position to prove the first

non-trivial estimates for operators of dyadic singular integral type. As one
expects, the UMD property is used, but in this first estimate still in a relatively
straightforward manner.
92 12 Dyadic operators and the T (1) theorem

Proposition 12.1.5. Let X be a UMD space, p ∈ (1, ∞), and f ∈ S00 (D; X).
For any any α ∈ {0, 1}d \ {0} and coefficients λQ ∈ K, we have the estimates
X
α
λQ hhQ , f ihα
Q 6 βp,X sup |λQ |kf kLp (Rd ;X) ,
Lp (Rd ;X) Q∈D
Q∈D
X
+
εQ hhα α
Q , f ihQ 6 βp,X kf kLp (Rd ;X) .
Lp (Ω×Rd ;X)
Q∈D

Proof. Let us denote

X γ
Dα α α
Q f := hhQ , f ihQ ,
−α
DQ f := DQ f − Dα
Qf = hhQ , f ihγQ .
γ∈{0,1}d \{0,α}

−α γ
Then (Dα Q f, DQ f ) is a martingale difference sequence on Q, as each hQ with
α
γ∈/ {0, α} has average zero on the sets where hQ is constant. Appropriately
−α
enumerated, (Dα Q f, DQ f )Q∈D also forms a martingale difference sequence.
Estimating its martingale transform by a multiplying sequence of 0’s and 1’s,
we obtain
X X 
α
λ Q DQ f p d = λQ · DQα
f + 0 · D−α
Q f p d
L (R ;X) L (R ;X)
Q∈D Q∈D
X 
α −α
6 βp,X DQ f + DQ f ,
Lp (Rd ;X)
Q∈D

For the other claim, we argue by the contraction principle and the ran-
domised UMD inequality to see that
X X 
0 −α
εQ Dα
Qf p d
6 ε α
Q DQ f + εQ DQ f
p d
L (Ω×R ;X) L (Ω×R ;X)
Q∈F Q∈D
X 
+ −α
6 βp,X Dα
Q f + DQ f ,
Lp (Rd ;X)
Q∈D

and in both cases we conclude by observing that

X  X
α −α
DQ f + DQ f = DQ f = f.
Q∈D Q∈D


For operator-valued coefficients λQ ∈ L (X, Y ), the following variants of R-
boundedness turn out to be relevant:

Definition 12.1.6. For p ∈ (1, ∞) and an operator family λ = (λQ )Q∈C ⊆

L (X, Y ) indexed by a collection C of bounded Borel subsets of Rd , we de-
note by DR p (λ) and E R p (λ) the smallest admissible constants such that the
 12.1 Dyadic singular integral operators 93

following estimates hold for all finitely non-zero families (xQ )Q∈C ⊆ X and
∗
(yQ )Q∈C ⊆ Y ∗ :
X
∗
|Q||hλQ xQ , yQ i|
Q∈C
X X
6 DR p (λ) ε Q x Q 1Q ∗
εQ yQ 1Q ,
Lp (Ω×Rd ;X) Lp0 (Ω×Rd ;Y ∗ )
Q∈C Q∈C

and
X X
ε Q λ Q x Q 1Q 6 E R p (λ) εQ xQ 1Q .
Lp (Ω×Rd ;Y ) Lp (Ω×Rd ;X)
Q∈C Q∈C

We refer to DR p (λ) as the DR p -bound of λ, and say that λ is DR p -bounded

if DR p (λ) < ∞. The same convention applies to E R p in place of DR p .
Remark 12.1.7. The primary case of interest will be when C = D is a system
of dyadic cubes. In this case, it is useful to observe at once that the defining
inequality of E R p (λ) immediately extends to Haar functions hQ α
in place of
the indicators 1Q :
X X
εQ λQ xQ hγQ p d
6 E R p (λ) ε Q x Q hα
Q p d
.
L (Ω×R ;Y ) L (Ω×R ;X)
Q∈D Q∈D

−1/2
Proof. We have hα α
Q = sgn(hQ )|Q| 1Q and hence, by the contraction prin-
ciple,
X X
α
ε Q zQ h Q = εQ |Q|−1/2 zQ 1Q
Lp (Ω×Rd ;Z) Lp (Ω×Rd ;Z)
Q∈D Q∈D

for both (zQ , Z) = {(xQ , X), (λQ xQ , Y )}. Using this twice, with both α and
γ, and in between the defining inequality of E R p (λ) for |Q|−1/2 xQ in place
xQ , yields the claim. 
These notions are weaker than R-boundedness; we will shortly see that the
converse fails in general.
Lemma 12.1.8. For all Banach spaces X and Y , all operator families λ =
(λQ )Q∈C ⊆ L (X, Y ) and their adjoints λ∗ := (λQ
∗
)Q∈D ⊆ L (Y ∗ , X ∗ ), and
all p ∈ (1, ∞), we have

sup kλQ k 6 DR p (λ) 6 min{E R p (λ), E R p0 (λ∗ )},

Q∈C

E R p (λ) 6 kx 7→ Rp ({λQ : Q 3 x})kL∞ (Rd ) 6 Rp (λ).

Proof. The last two estimates are immediate. The first estimate follows by
testing the defining condition of DR p with only one non-zero pair (xQ , yQ ∗
) at
a time. To see that DR p (λ) 6 E R p (λ), for suitable scalars |ηQ | = 1, we have
94 12 Dyadic operators and the T (1) theorem
X XZ
∗ ∗
|Q||hλQ xQ , yQ i| = ηQ hλQ xQ 1Q , yQ 1Q i
Q∈C Q∈C
Z DX X E
∗
=E εQ ηQ λQ xQ 1Q , ε R yR 1R
Q∈C R∈C
X X
∗
6 ε Q η Q λ Q x Q 1Q ε R yR 1R
Lp (Ω×Rd ;X) Lp0 (Ω×Rd ;Y ∗ )
Q∈C R∈C
X X
6 E R p (λ) εQ xQ 1Q ∗
εR yR 1R ,
Lp (Ω×Rd ;X) Lp0 (Ω×Rd ;Y ∗ )
Q∈C R∈C

where we used Kahane’s contraction principle and the definition of E R p (λ)

to pull out the scalar ηQ and the operators λQ in the last step. Since
∗
hλQ xQ , yQ ∗
i = hxQ , λ∗Q yQ i, and DR p0 (λ) is defined by testing the expressions
on the right over a more general choice of x∗∗ Q ∈ X
∗∗
in place of xQ ∈ X, it
follows that
DR p (λ) 6 DR p0 (λ∗ ) 6 E R p0 (λ∗ )
by using what we already proved, but with λ∗ in place of λ. 

Corollary 12.1.9. If λ = (λQ )Q∈C ⊆ L (X) consists of scalar multiples of

the identity, then

sup |λQ | = DR p (λ) = E R p (λ) = Rp (λ).

Q∈C

Proof. Lemma 12.1.9 shows that we have this chain with “6” in place of “=”
throughout. On the other hand, Kahane’s contraction principle guarantees
that Rp (λ) = supQ∈C |λQ |. Thus we have equality throughout. 
The following example of DR p -bounded families will play a role in our inves-
tigation of criteria for boundedness of singular integral operators; the uniform
boundedness of the quantities |Q|−1 hT 1Q , 1Q i is classically known as the weak
boundedness property of the operator T .

Example 12.1.10. Suppose that T ∈ L (Lp (Rd ; X), Lp (Rd ; Y )), and define
hT (1Q ), 1Q i ∈ L (X, Y ) by
Z
hT 1Q , 1Q i : x 7→ hT (1Q x), 1Q i = T (1Q x) ∈ Y.
Q

For any collection C of bounded Borel subsets of Rd , it follows that

n hT 1 , 1 i o 
Q Q
DR p 6 kT kL (Lp (Rd ;X),Lp (Rd ;Y )) .
|Q| Q∈C

Proof. With suitable scalars |ηQ | = 1, we have

 12.1 Dyadic singular integral operators 95
D hT 1 , 1 i E
Q Q
X X
∗ ∗
|Q| xQ , y Q = ηQ hT (1Q xQ ), 1Q yQ i
|Q|
Q∈C Q∈C
D X X E
∗
=E T ε Q η Q x Q 1Q , εR yR 1R
Q∈C R∈C
X X
∗
6 T ε Q η Q x Q 1Q ε R yR 1R ,
Lp (Ω×Rd ;Y ) Lp0 (Ω×Rd ;Y )
Q∈C R∈C

where
X
T εQ ηQ xQ 1Q
Lp (Ω×Rd ;Y )
Q∈C
X
6 kT kL (Lp (Rd ;X),Lp (Rd ;Y )) ε Q x Q 1Q
Lp (Ω×Rd ;X)
Q∈C

by the assumed boundedness of T and Kahane’s contraction principle with

the coefficients ηQ . 

While Example 12.1.10 will only play a role later, the weakening of R-
boundedness has the following immediate application:

Theorem 12.1.11 (Haar multipliers). Let X and Y be UMD spaces and

p ∈ (1, ∞). For α, γ ∈ {0, 1}d \ {0} and λ = (λQ )Q∈D ⊆ L (X, Y ) , consider
the operator X
Hαγ
λ : f 7→ λQ hf, hα γ
Q ihQ , (12.3)
Q∈D

initially mapping S00 (D; X) into S00 (D; Y ). Then Hαγ

λ extends to a bounded
operator on Lp (Rd ; X) if and only if DR p (λ) < ∞, and in this case

DR p (λ)
− 6 kHαγ
λ kL (Lp (Rd ;X),Lp (Rd ;Y )) 6 βp,X βp0 ,Y ∗ DR p (λ).
+ +
(12.4)
βp,X βp−0 ,Y ∗

Proof. Dualising Hαγ p d ∗

λ f ∈ S00 (D; X) ⊆ L (R ; X) with g ∈ S00 (D; Y ) ⊆
p0 d ∗
L (R ; Y ), we arrive at
XD E
|hHαγ
λ f, gi| = λQ hf, hα γ
Q i, hhQ , gi
Q∈D
D hf, hα i hhγ , gi E
Q Q
X
= |Q| λQ ,
|Q|1/2 |Q|1/2
Q∈D
(12.5)
X 1Q
6 DR p (λ) εQ hf, hα
Qi ×
|Q|1/2 Lp (Ω×Rd ;X)
Q∈D
X 1Q
× εQ hg, hα
Qi .
|Q|1/2 Lp0 (Ω×Rd ;Y ∗ )
Q∈D
96 12 Dyadic operators and the T (1) theorem

α
For a fixed s ∈ Rd , the sequences (εQ 1Q (s)/|Q|1/2 )Q∈D and (εQ hQ )Q∈D have
equal distribution; thus
X 1Q X
εQ hf, hα
Qi = εQ hf, hα α
Q ihQ
|Q|1/2 Lp (Ω×Rd ;X) Lp (Ω×Rd ;X)
Q∈D Q∈D
+
6 βp,X kf kLp (Rd ;X)

by Proposition 12.1.5 in the last step. Similarly, the last term in (12.5) is
dominated by βp+0 ,Y ∗ kgkLp0 (Rd ;Y ∗ ) . Hence

|hHαγ
λ f, gi| 6 DR p (λ)βp,X βp0 ,Y ∗ kf kLp (Rd ;X) kgkLp (Rd ;Y ∗ ) ,
+ +
0

which proves the second estimate in (12.4).

∗
Conversely, for finitely non-zero families (xQ )Q∈D ⊆ X and (yQ )Q∈D ⊆
∗ ∗ ∗
Y , we choose scalar |ηQ | = 1 such that |hλQ xQ , yQ i| = ηQ hλQ xQ , yQ i and
consider the functions
X X
∗ γ
f := |Q|1/2 ηQ xQ hαQ ∈ S00 (D; X), g := |Q|1/2 yQ hQ ∈ S00 (D; Y ∗ ).
Q∈D Q∈D

Then
X
Hαγ
λ f = |Q|1/2 ηQ λQ xQ hγQ ,
Q∈D
X
hHαγ
λ f, gi = ∗
|Q|ηQ hλQ xQ , yQ i,
Q∈D

and hence
X
∗
|Q||hλQ xQ , yQ i| 6 kHαγ
λ kL (Lp (Rd ;X),Lp (Rd ;Y ) kf kLp (Rd ;X) kgkLp (Rd ;Y ∗ ) ,
0

Q∈D

where
X
−
kf kLp (Rd ;X) 6 βp,X εQ |Q|1/2 ηQ xQ hα
Q
Lp (Ω×Rd ;X)
Q∈D
X
−
= βp,X εQ xQ 1Q
Lp (Ω×Rd ;X)
Q∈D

by a similar equidistribution property as before. Similarly, we have

X
kgkLp0 (Rd ;Y ∗ ) 6 βp−0 ,Y ∗ ∗
ε Q yQ 1Q 0
,
Lp (Ω×Rd ;Y ∗ )
Q∈D

and combining the bounds, we have proved the first estimate in (12.4). 
Remark 12.1.12. Under stronger assumptions on the coefficients λ, one can
improve the dependence on the UMD constants:
 12.1 Dyadic singular integral operators 97

(1) If X = Y , α = γ, and λ ⊆ K · IX is bounded, then Hαα

λ extends to a
bounded operator on Lp (Rd ; X) of norm at most

kHαα
λ kL (Lp (Rd ;X)) 6 βp,X kλk∞ .

(2) If λ ⊆ L (X, Y ) is R-bounded, then Hαγ λ extends to a bounded operator

from Lp (Rd ; X) to Lp (Rd ; Y ) of norm at most

kHλαγ kL (Lp (Rd ;X),Lp (Rd ;Y )) 6 βp,Y

− +
βp,X E R p (λ),
−
where a partial advantage over Theorem 12.1.11 comes from βp,Y 6 βp+0 ,Y ∗ .

Proof. (1): This is a restatement of the first estimate in Proposition 12.1.5.

(2): Since (hγQ )Q∈D is a martingale difference sequence, using the defin-
ing properties of various constants and the definition of E R p (λ) via Remark
12.1.7, we have
X γ
α
λQ hf, hQ ihQ p dL (R ;Y )
Q∈D
X γ
−
6 βp,Y εQ λQ hf, hα
Q ihQ
Lp (Ω×Rd ;X)
Q∈D
X
−
6 βp,Y E R p (λ) εQ hf, hα α
Q ihQ
Lp (Ω×Rd ;Y )
Q∈D
−
6 βp,Y E R p (λ)βp,X
+
kf kLp (Rd ;X) ,

where, in the last step, we used the second estimate in Proposition 12.1.5. 
Here is a nice class of examples of coefficients satisfying the dyadic R-
boundedness condition:

Proposition 12.1.13. Let Y be a UMD space and p ∈ (1, ∞). Let b ∈

L∞ (Rd ; L (X, Y )), let a = (aQ )Q∈D ∈ `∞ (D; L∞ (Rd )), and

λ := (λQ )Q∈D := (haQ biQ )Q∈D .

Then
E R p ((haQ biQ )Q∈D ) 6 βp,Y
+
kak`∞ (L∞ ) kbkL∞ (Rd ;L (X,Y ))
Thus, for α, γ ∈ {0, 1}d \{0}, the Haar multiplier Hαγ
λ extends to a bounded
operator from Lp (Rd ; X) to Lp (Rd ; Y ) of norm at most

kHλαγ kL (Lp (Rd ;X),Lp (Rd ;Y )) 6 βp,Y

− +
βp,Y +
βp,X kak`∞ (L∞ ) kbkL∞ (Rd ;L (X,Y )) .

Proof. The second claim is immediate from the first one in combination with
Remark 12.1.12(2), so we concentrate on the first one. We may assume by
scaling that kaQ kL∞ (Rd ) 6 1. Then
98 12 Dyadic operators and the T (1) theorem
X X
εQ haQ biQ xQ 1Q = εQ EQ (aQ bxQ 1Q )
Lp (Rd ;Y ) Lp (Rd ;Y )
Q∈D Q∈D
X
+
6 βp,Y b ε Q x Q 1Q
Lp (Rd ;Y )
Q∈D
X
+
6 βp,Y kbkL∞ (Rd ;L (X,Y ) εQ xQ 1Q ,
Lp (Rd ;Y )
Q∈D

where, in the first estimate, we applied Stein’s inequality (Theorem 4.2.23)

followed by Kahane’s contraction principle with the scalar coefficients aQ . 
The following result shows that result of Proposition 12.1.13 cannot be im-
proved to usual R-boundedness; thus the notions DR p and E R p represent
genuine relaxations:

Proposition 12.1.14. For non-zero Banach spaces X and Y , the following

are equivalent:
(1) X has type 2 and Y has cotype 2;
(2) for every b ∈ L∞ (0, 1; L (X, Y )), the set {hbiQ : Q ∈ D([0, 1))} is R-
bounded;
(3) for every b ∈ L∞ (0, 1; L (X, Y )), the function
 
x 7→ R hbiQ : x ∈ Q ∈ D([0, 1))

is essentially bounded.

Proof. (1)⇒(2): For b ∈ L∞ (0, 1; L (X, Y )), it is clear that the {hbiQ : Q ∈
D([0, 1))} is uniformly bounded. Under the assumption (1), this implies R-
boundedness by Proposition 8.6.1.
(2)⇒(3): This is clear.
(3)⇒(1): From the definition of R-boundedness, it is immediate that
R(T ) = sup({R(F ) : F ⊆ T finite}). So if some collection T is not R-
bounded, it hasSfinite subcollections Fn with R(Fn ) → ∞. Then the count-
∞
able collection n=1 Fn ⊆ T also fails to be R-bounded.
If (1) is not satisfied, then Proposition 8.6.1 says that the unit ball of
B̄L (X,Y ) of L (X, Y ) is not R-bounded. By what we just observed, this means
∞
that we can find a sequence {uk }k=0 ⊆ B̄L (X,Y ) that fails to be R-bounded.
4 1
Let vk := 3 uk − 3 uk+1 and
∞
X
b := vj 1[4−j−1 ,4−j ) .
j=0

Then b ∈ L∞ (Rd ; L (X, Y )) and kbk∞ = supk kvk k 6 5

3 supk kuk k = 5
3.
Moreover,
 12.1 Dyadic singular integral operators 99
∞ ∞ ∞
X 3 3  X −j 4 X 1 
hbi[0,4−k ) = 4k 4−j vj = 4k 4 uj − 4−j uj+1 = uk .
4 4 3 3
j=k j=k j=k

Then for each n, we have

kx 7→ R({hbiI : x ∈ I ∈ D([0, 1))})kL∞ (0,1)
> R({hbiI : [0, 4−n ) ∈ I ∈ D([0, 1))}) > R({hbi[0,4−k ) }nk=0 ) = R({uk }nk=0 ),
and hence
∞ = R({uk }∞
k=0 ) = sup R({uk }k=0 )
n
n∈N
6 kx 7→ R({hbiI : x ∈ I ∈ D([0, 1))})kL∞ (0,1) .
Thus (3) fails, and by contraposition this proves the claimed implication. 

Comparison of DR p and E R p

In the rest of this section, we make a further comparison of the two relaxed
notions of R-boundedness from Definition 12.1.6.. When Y is a UMD space—
an assumption that we make a good part of the time—, these notions turn
out to be equivalent. The universal bound DR p (λ) 6 E R p (λ) was already
observed in Lemma 12.1.8. The reverse estimate could be achieved essentially
by concatenating a couple of results that we have treated earlier in these
volumes, but it turns out that a slightly sharper quantitative bound can be
achieved by also revisiting their proofs to establish the following proposition:
Proposition 12.1.15. Let Y be a UMD space and p ∈ (1, ∞). Let E0 :=
{∅, Ω} be the trivial σ-algebra of a probability space (Ω, A , P) supporting a
Rademacher sequence (εn )N N
n=1 , and (Fn )n=1 be a σ-finite filtration of some
measure space (S, F , µ). Then, for all f ∈ Lp (Ω × S; Y ), we have
N
X
εn E(εn f |E0 × Fn ) +
6 βp,Y kf kLp (Ω×S;Y ) .
Lp (Ω×S;Y )
n=1

Proof. Let En := σ(ε1 , . . . , εn ) for n = 1, . . . , N . Then

E(εn f |E0 × Fn ) = E(E(εn f |En × Fn )|E0 × Fn )
= E(εn E(f |En × Fn )|E0 × FN ),
where in the last step we note that for both k ∈ {n, N }, the conditional expec-
tation of the function inside, given E0 × Fk , is obtained by simply integrating
out the dependence on ω ∈ Ω. On the other hand, we have
E(εn E(f |En−1 ⊗ Fn )|E0 × FN )
= E(E(εn E(f |En−1 × Fn )|En−1 × FN )|E0 × FN )
= E(E(εn |En−1 × FN )E(f |En−1 × Fn )|E0 × FN )
= E(0 · E(f |En−1 × Fn )|E0 × FN ) = 0.
100 12 Dyadic operators and the T (1) theorem

Thus
E(εn f |E0 × Fn ) = E(εn [E(f |En × Fn ) − E(f |En−1 × Fn )]|E0 × FN )
= E(εn [E(f |G2n ) − E(f |G2n−1 )]|E0 × FN )
= E(εn d2n |E0 × FN ),

where (
E(f |Gk ) − E(f |Gk−1 ), k = 2, . . . , 2N,
dk :=
E(f |G1 ), k = 1,
are martingale differences relative to a filtration (Gk )2N
k=1 on Ω × S defined by

G2n := En × Fn , G2n−1 := En−1 × Fn .

Then, noting that E( |E0 × FN ) is constant in ω ∈ Ω, and denoting by

k=1 another Rademacher sequence on some (Ω , A , P ), we have
(ε0k )2N 0 0 0

N
X
εn E(εn f |E0 × Fn )
Lp (Ω×S;Y )
n=1
N
X
= ε02n E(εn d2n |E0 × FN )
Lp (Ω 0 ×Ω×S;Y )
n=1
N
X
= E( ε02n εn d2n |E0 × FN )
Lp (Ω 0 ×Ω×S;Y )
n=1
N
X N
X
6 ε02n εn d2n = 0
ε2n d2n
Lp (Ω 0 ×Ω×S;Y ) Lp (Ω 0 ×Ω×S;Y )
n=1 n=1
2N
X 2N
X
6 εk0 dk +
6 βp,Y dk
Lp (Ω 0 ×Ω×S;Y ) Lp (Ω×S;Y )
k=1 k=1
+ +
= βp,Y kE(f |G2N )kLp (Ω×S;Y ) 6 βp,Y kf kLp (Ω×S;Y ) ,

where, in the four estimates, we applied the contractivity of conditional ex-

pectation on Lp , Kahane’s contraction principle with coefficients {0, 1}, the
+
definition of the UMD constant βp,Y , and again the contractivity of condi-
tional expectation on Lp . 

Remark 12.1.16. Proposition 12.1.15 is a simultaneous generalisation of Stein’s

inequality (Theorem 4.2.23),
N
X N
X
+
εn E(fn |Fn ) 6 βp,Y εn f n , (12.6)
Lp (Ω×S;Y ) Lp (Ω×S;Y )
n=1 n=1

for all fn ∈ Lp (S; Y ), and the K-convexity inequality for UMD spaces (Propo-
sition 4.3.10),
 12.1 Dyadic singular integral operators 101
N
X
+
εn E(εn f ) 6 Kp,Y kf kLp (Ω;Y ) , Kp,Y 6 βp,Y , (12.7)
Lp (Ω;Y )
n=1

for all f ∈ Lp (Ω; Y ).

Namely, (12.6) is obtained from Proposition 12.1.15 by taking f =
PN
k=1 εk ⊗ fk , in which case E(εn f |E0 × Fn ) = E(fn |Fn ), while (12.7) is
the special case where S = {s} is a singleton, or in other words f is indepen-
dent of s ∈ S. Moreover, Proposition 12.1.15 shows that (12.7) holds equally
well with real or complex Rademacher variable εn , provided only that we use
−
the UMD constant βp,Y defined in terms of the same variables; in contrast,
the proof of Proposition 4.3.10 was written for the real Rademacher variables
rn and made some explicit (although not essential) use of this choice.
Qualitatively, Proposition 12.1.15 could also be derived from the said two
results, but with a quantitatively weaker conclusion; namely,
N
X
εn E(εn f |E0 × Fn )
Lp (Ω×S;Y )
n=1
N
X
= εn0 E(E(εn f |E0 × F )|E0 × Fn )
Lp (Ω 0 ×Ω×S;Y )
n=1
N
X
+
6 βp,Y ε0n E(εn f |E0 × F )
Lp (Ω 0 ×Ω×S;Y )
n=1
N
X
+ +
= βp,Y εn E(εn f |E0 ) 6 βp,Y Kp,Y kf kLp (S;Lp (Ω;Y )) ,
Lp (S;Lp (Ω;Y ))
n=1

using the K-convexity inequality in Lp (Ω; Y ), pointwise at each s ∈ S, in the

last step.

Corollary 12.1.17. If Y is a UMD space and λ = (λQ )Q∈D ⊆ L (X, Y ),

then
DR p (λ) 6 E R p (λ) 6 βp+0 ,Y ∗ DR p (λ).

Proof. We already proved the first inequality in Lemma 12.1.8. For the second
inequality, we first note that, by Fubini’s theorem,
X X
εQ zQ 1Q p d
= εn(Q) zQ 1Q p d
, (12.8)
L (Ω×R ;Z) L (Ω×R ;Z)
Q∈D Q∈D

where n(Q) ∈ Z is such that Q ∈ Dn : This is because, pointwise at each

s ∈ Rd , there is exactly one dyadic Q 3 s of each generation n ∈ Z, and we
can replace the sequence (εQ )Q3s by the equidistributed sequence (εn )n∈Z =
(εn(Q) )Q3s . For zQ = λQ xQ and Z = Y , we then dualise the right-hand side
0
of (12.8) with G ∈ Lp (Ω × Rd ; Y ∗ ):
102 12 Dyadic operators and the T (1) theorem
DX E X D E
εn(Q) λQ xQ 1Q , G = λQ xQ , hE(εn(Q) G)iQ |Q|
Q∈D Q∈Dn
X
6 DR p (λ) εn(Q) xQ 1Q × (12.9)
Lp (Ω×Rd ;X)
Q∈D
X
× εn(Q) hE(εn(Q) G)iQ 1Q .
Lp0 (Ω×Rd ;Y ∗ )
Q∈D

0
In the Lp (Ω × Rd ; Y ∗ ) norm on the right, we write
X X X
εn(Q) hE(εn(Q) G)iQ 1Q = εn E(E(εn G)|σ(Dn ))1Q
Q∈D n∈Z Q∈Dn
X X
= εn E(E(εn G)|σ(Dn )) = εn E(εn G|{∅, Ω} × σ(Dn )).
n∈Z n∈Z

Thus, by a direct application of Proposition 12.1.15 in the UMD space Y ∗ in

place of Y , it follows that
X
εn(Q) hE(εn(Q) G)iQ 1Q 0
6 βp+0 ,Y ∗ kGkLp0 (Ω×Rd ;Y ∗ ) .
Lp (Ω×Rd ;Y ∗ )
Q∈D

Substituting back to (12.9), it follows by duality that

X
εn(Q) λQ xQ 1Q
Lp (Ω×Rd ;Y )
Q∈D
X
6 βp+0 ,Y ∗ DR p (λ) εn(Q) xQ 1Q ,
Lp (Ω×Rd ;X)
Q∈D

and we can replace n(Q) by Q on both sides according to (12.8) to obtain the
claimed result. 

12.1.b Nested collections of unions of dyadic cubes

Before proceeding to more complicated dyadic singular integrals, we de-

vote this intermediate section to elementary, although not entirely trivial,
geometric–combinatorial considerations related to the dyadic cubes. Collect-
ing the relevant auxiliary results here for easy reference will allow our subse-
quent analysis to flow with a nice tempo without annoying interruptions.

Definition 12.1.18 (Nestedness). We say that two set E, F are nested if

E ∩ F ∈ {∅, E, F }. A collection E of sets is called nested if any E, F ∈ E
have this property.

The fact that the collection D of dyadic cubes enjoys this property underlies
many considerations that we have encountered in these volumes.
 12.1 Dyadic singular integral operators 103

In the dyadic analysis of a singular integral operators that we undertake

in this section, we will also need to deal with unions Q1 ∪ Q2 of two dyadic
cubes of the same size. A moment’s thought confirms that two such sets will
not be nested in general, yet quite frequently they still enjoy this property.
Accordingly, a key to the related considerations will be the decomposition
of collections of pairs of dyadic cubes into controllably many subcollections,
where the nestedness of the unions Q1 ∪ Q2 is valid.

Definition 12.1.19 (Strong nestedness). Let Q1 ∪ Q2 and R1 ∪ R2 be two

unions of some Qi , Ri ∈ D with `(Q1 ) = `(Q2 ) and `(R1 ) = `(R2 ). We say
that E and F are strongly nested if they are equal, or disjoint, or one of them,
say Q1 ∪Q2 , is contained not just in R1 ∪R2 but in a dyadic child of R1 or R2 .
A collection of such unions is called strongly nested if any two of its members
have this property.

Note that the dyadic cubes themselves, contained in this definition as a degen-
erate case with Q2 = Q1 , clearly satisfy this strong nestedness. This notion
is relevant for considerations dealing with Haar functions which, as we recall,
are constant on the dyadic children of their supporting dyadic cubes; thus,
if Q1 ∪ Q2 and R1 ∪ R2 are strongly nested, unequal but intersecting, then
the smaller union is entirely contained in a set of constant value for any Haar
function related to the larger union.
Our first (relatively simple) decomposition into strongly nested subcollec-
tions is the following:

Lemma 12.1.20. Suppose that, for some n ∈ N:

(a) F ⊆ D is a finite subcollection;
(b) φ : F → D is an injection with `(φ(Q)) = `(Q) for all Q ∈ F ;
(c) if Q, R ∈ F and `(Q) < `(R), then `(Q) < 2−n `(R);
and
φ(Q) ⊆ Q(n) ∀ Q ∈ F. (12.10)
Then F can be partitioned into 3 subcollections Fi such that each collection
{Q ∪ φ(Q) : Q ∈ Fi } is strongly nested.

Proof. Step 1 – Let all assumptions of the lemma be in force until further
notice. For each Q ∈ F ∪ φ(F ), we define a label r(Q) ∈ {0, 1, 2} such that
r(Q) 6= r(φ(Q)) for every Q ∈ F unless φ(Q) = Q. This ensures that Q∪φ(Q)
and R ∪ φ(R) are disjoint whenever Q, R ∈ F are two different cubes with
r(Q) = r(R) and `(Q) = `(R).
Indeed, Q 6= R implies φ(Q) 6= φ(R). Since different dyadic cubes of equal
size are disjoint, this implies that Q ∩ R = ∅ = φ(Q) ∩ φ(R). If φ(Q) = Q
or φ(R) = R, this already shows that Q ∪ φ(Q) and R ∪ φ(R) are disjoint. If
φ(Q) 6= Q and φ(R) 6= R, then r(φ(Q)) 6= r(Q) = r(R) implies φ(Q) 6= R and
similarly φ(R) 6= Q. By equal size again, this implies that φ(Q) ∩ R = ∅ =
Q ∩ φ(R), giving the (strong) nestedness when `(Q) = `(R).
104 12 Dyadic operators and the T (1) theorem

To define such r(R), let us denote φ◦0 (Q) = Q, φ◦k (Q) = φ(φ◦(k−1) (Q))
for k > 1. An orbit of φ is a set {φ◦k (Q) : k = 0, . . . , K}, where Q ∈ F and
either φ◦(K+1) (Q) = Q (in this case the orbit is cyclic), or Q ∈ / φ(F ) and
φ◦K (Q) ∈/ F . For all Q ∈ F ∪ φ(F ), we define r(Q) ∈ {0, 1, 2} by alternating
the values 0 and 1 on both non-cyclic orbits and cyclic orbits of even length,
and in addition using the value 2 once on cyclic orbits of odd length. In this
way, we ensure that r(Q) 6= r(φ(Q)) for any Q ∈ F unless Q = φ(Q).
Step 2 – It remains to check the strong nestedness in the case of Q, R ∈ F
with `(Q) < `(R), hence `(Q) < 2−n `(R). If Q ∪ φ(Q) intersect R ∪ φ(R), then
one of P ∈ {Q, φ(Q)} intersects one of S ∈ {R, φ(R)}. Since `(P ) < 2−n `(S)
and the cubes are dyadic, this implies that P (n) ( S. Since φ(Q) ⊆ Q(n) , we
have φ(Q)(n) = Q(n) , and hence Q ∪ φ(Q) ⊆ Q(n) ( S, confirming strong
nestedness in the case of `(Q) < `(R). 
In the lack of (12.10), the situation is somewhat more complicated. Suitable
substitute conditions are provided in the following:

Lemma 12.1.21. Assume conditions (a) through (c) as well as:

(d) φ(Q) ⊆ 3Q(n) for all Q ∈ F ;
(e) 3Q ⊆ Q(n) for all Q ∈ F ∪ φ(F ).
Then F can be partitioned into nine subcollections Fi such that each collection
{Q ∪ φ(Q) : Q ∈ Fi } is strongly nested.

Proof. Step 1 – We define the label r(Q) ∈ {0, 1, 2} exactly as in the proof of
Lemma 12.1.20 to ensure that r(Q) 6= r(φ(Q)) unless Q = φ(Q). This gives
the nestedness of the sets Q ∪ φ(Q) for cubes of a fixed sidelength, as before.
Step 2 – We claim that, for each Q ∈ F ∪ φ(F ), there can be at most one
R ∈ F ∪ φ(F ) such that

Q ( R, 3Q(n) 6⊆ RQ , (12.11)

where RQ is the unique dyadic child of R that contains Q ( R.

In fact, let R be as above, and Q ( R ( S ∈ F ∪ φ(F ), thus Q(n) ( R,
(n)
R ( S by (c). By (e) applied to the cube R, we then have 3Q(n) ⊆ 3R ⊆
(n)
R ⊆ SR , so indeed S will not satisfy the condition (12.11) that R does, and
this proves the uniqueness of R.
Step 3 – For each P ∈ F , we define a second label s(P ) ∈ {0, 1, 2} in such
a way that if (r(P ), s(P )) = (r(S), s(S)), then (12.11) does not hold for either
R = S or R = φ(S). This will ensure strong nestedness for the subcollection
with constant pairs of labels (r(P ), s(P )).
Indeed, suppose that P, S ∈ F have (r(P ), s(P )) = (r(S), s(S)) where
`(P ) < `(S) and P ∪ φ(P ) intersects S ∪ φ(S). Hence (at least) one of Q ∈
{P, φ(P )} intersects (at least) one of R ∈ {S, φ(S)} and thus Q ( R. By (d)
and the failure of (12.11), we have P ∪ φ(P ) ⊆ (1 + 2n+1 Q) ⊆ RQ .
 12.1 Dyadic singular integral operators 105

The required second label s(P ) is defined for each P ∈ F as follows. For
all P ∈ F of maximal size, let s(P ) := 0. Recursively, we proceed to the unla-
belled cubes P ∈ F of maximal size. For these cubes, we first check whether
(12.11) occurs with either Q = P or Q = φ(P ), and some R ∈ F ∪ φ(F ).
It could happen that R ∈ F , or R = φ(S) with S ∈ F , or both. We then
require that s(P ) is chosen so that (r(P ), s(P )) ∈
/ {(r(R), s(R)), (r(S), s(S))}.
If S = R, this is clearly one restriction on S(P ). But if S 6= R = φ(S), then
r(R) 6= r(S) by the alternating choice of r along the orbits, and we still get
at most one restriction of the possible value of s(P ). Since different R and S
may arise from the case Q = P and Q = φ(P ) we get altogether at most two
restrictions on s(P ), and we can declare that s(P ) is the smallest remaining
number in {0, 1, 2}. 
The next result relaxes the assumptions even further, at the cost of compli-
cating the conclusions:

Lemma 12.1.22. Assume conditions (a) through (d). Then F can be par-
titioned into 144 subcollections Fi , and on each of them we have injections
φi,j : Fi → D, j = 0, 1, 2, 3, where φi,0 (Q) = Q and φi,3 (Q) = φ(Q) such that
each collection
{φi,j (Q) ∪ φi,j+1 (Q) : Q ∈ Fi } (12.12)
is strongly nested.

Proof. The idea is to combine the special cases treated in the two previous
Lemmas 12.1.20 and 12.1.21, which had the additional assumptions (12.10)
and (e), respectively; neither is assumed now.
For every R ∈ D, consider the 2nd cubes Q ∈ D with Q(n) = R. Among
them, there are (2n − 2)d off-boundary cubes Q with 3Q ⊆ R, while the
number of boundary cubes is then
1 nd
2nd − (2n − 2)d = 2nd [1 − (1 − 21−n )d ] 6 2nd · 21−n d 6 2
2
if n > log2 (4d). When this is the case, we can define a permutation ψ : D → D
with `(ψ(Q)) = `(Q), ψ(Q) ⊆ Q(n) (as in (12.10)) such that ψ(Q) is an off-
boundary cube in Q(n) whenever Q is a boundary cube in Q(n) .
Let us first divide F into four subcollection Fu,v , where u, v ∈ {0, 1}, so
that Q ∈ Fu,v is a boundary cube in Q(n) if and only if u = 1, whereas φ(Q)
is a boundary cube in ψ(Q)(n) if and only if v = 1.
Case F0,0 : By Lemma 12.1.21, we can divide F0,0 into nine subcollections Fi
such that {Q ∪ φ(Q) : Q ∈ Fi } is strongly nested. Letting φi,1 = φi,2 = φi,3 =
φ in this case, we trivially have the strong nestedness of {φi,j (Q) ∪ φi,j+1 (Q) :
Q ∈ Fi } for j = 1, 2 (since the collection is simply φ(Fi ) ⊆ D in this case.
106 12 Dyadic operators and the T (1) theorem

Case F0,1 : On the collection F0,1 , we consider the map ψ ◦ φ and observe
that it also satisfies (d); indeed, φ(Q) ⊆ 3Q(n) lies inside one of the dyadic
neighbours of Q(n) , and ψ keeps it inside this same nth generation ancestor.
Since φ(Q) is a boundary cube in φ(Q)(n) for Q ∈ F0,1 by definition of this
collection, ψ(φ(Q)) is off-boundary in φ(Q)(n) = ψ(φ(Q))(n) by definition of
ψ, and hence (F0,1 , ψ ◦ φ) also satisfies (e) in place of (F , φ). Then Lemma
12.1.21 shows that F0,1 can be divided into nine subcollections Fa0 such that
each {Q ∪ ψ(φ(Q)) : Q ∈ Fa0 } is strongly nested. On the other hand, we can
write

{ψ(φ(Q)) ∪ φ(Q) : Q ∈ F0,1 } = {R ∪ ψ(R) : R ∈ φ(F0,1 )}.

Here (F0,1 , ψ) satisfies the assumptions of Lemma 12.1.20, and hence φ(F0,1 )
can be divided into three subcollections Gb such that {R ∪ ψ(R) : R ∈ Gb )} is
strongly nested. This since φ is injective, this induces a decomposition of F0,1
into three subcollections where Fb00 such that {ψ(φ(Q)) ∪ φ(Q) : Q ∈ Fb00 }
is strongly nested. Then, defining Fi = Fa0 ∩ Fb00 for i = (a, b), we find that
both
{Q ∪ ψ(φ(Q)) : Q ∈ Fi }, {ψ(φ(Q)) ∪ φ(Q) : Q ∈ Fi }
are strongly nested, and there is in total 9 · 3 such collections Fi decomposing
F0,1 . So taking φi,1 = ψ ◦ φ and φi,2 = φi,3 , we have the strong nestedness of
the collections in (12.12), the case j = 2 for trivial reasons as in case F0,0 .
Case F1,0 : Similarly, on the collection F1,0 , Lemma 12.1.20 applies to the
mapping ψ to provide three subcollection Fa0 such that {Q ∪ ψ(Q) : Q ∈ Fa0 }
is strongly nested. And Lemma 12.1.21 applies to (ψ(F1,0 ), φ◦ψ −1 ) to provide
nine subcollections Fb00 such that {ψ(Q) ∪ φ(Q) : Q ∈ Fb00 } is strongly nested.
So altogether we have 3 · 9 subcollection Fi = Fa0 ∩ Fb00 such that

{Q ∪ ψ(Q) : Q ∈ Fi }, {ψ(Q) ∪ φ(Q) : Q ∈ Fi }

are strongly nested. We can hence define φi,1 = ψ, φi,2 = φi,3 = φ to get the
claimed conclusions.
Case F1,1 : Finally, on the collection F1,1 , Lemma 12.1.20 applies to both
(F1,1 : ψ) and to (ψ ◦ φ(F1,1 : ψ −1 ) to provide three subcollections Fa0 and
three other Fb00 such that {Q ∪ ψ(Q) : Q ∈ Fa0 } and {ψ(φ(Q)) ∪ φ(Q) :
Q ∈ Fb00 } are strongly nested. And we check that Lemma 12.1.21 applies to
(ψ(F1,1 ), ψ◦φ◦ψ −1 ) to provide nine subcollections Fc000 such {ψ(Q)∪ψ(φ(Q)) :
Q ∈ Fc000 } is strongly nested. Then with Fi = Fa0 ∩ Fb00 ∩ Fc000 we obtain 32 · 9
subcollections such that {Q∪ψ(Q) : Q ∈ Fi }, {ψ(Q)∪ψ(φ(Q)) : Q ∈ Fi }, and
{ψ(φ(Q)) ∪ φ(Q) : Q ∈ Fi } are strongly nested, and we can define φi,1 = ψ,
φi,2 = ψ ◦ φ, φi,3 = φ in this case.
In total we have divided F into 9 + 2 · 9 · 3 + 9 · 32 = 144 subcollections
Fi with required properties. 
Another variant of the conclusion with the same assumptions is as follows:
 12.1 Dyadic singular integral operators 107

Lemma 12.1.23. Assume conditions (a) through (d). Then F can be parti-
tioned into 33d+1 subcollections Fi such that each collection

{Q[m(i)] ∪ φ(Q)[m(i)] : Q ∈ Fi } ⊆ D m(i);3

is strongly nested, where

(1) D m(i);3 is one of the dilated dyadic systems from Proposition 11.3.11;
(2) for each P ∈ D, we denote by P [m] the unique

P [m] ∈ D m;3 with P [m] ⊇ P and `(P [m] ) = 3`(P ). (12.13)

Proof. We have Q ∪ φ(Q) ⊆ (1 + 2n+1 )Q ⊆ 3Q(n) , where Q(n) is the nth

generation dyadic ancestor of Q. Recall that the cubes 3R, R ∈ D, can be
split into 3d new dyadic-like systems D m;3 by Proposition 11.3.11. For each
Q ∈ F , let mQ be the index such that 3Q(n) ∈ D mQ ;3 , and let Q0 = Q[mQ ] ,
Q00 = φ(Q)[mQ ] be as in (12.13). (Thus Q0 is the three-fold expansions of one
of the neighbours of Q; any of these contains Q, and exactly one of them
belongs to the correct D mQ ;3 ; the same remark applies to Q00 and φ(Q) in
place of Q0 and Q.) Note that the same Q0 can arise from 3d different cubes
Q, and the same Q00 from 3d different φ(Q); however, by dividing F into 9d
subcollections F a , we ensure that Q is uniquely determined by Q0 , and φ(Q)
by Q00 , within each F a .
Let us then consider the collections F a,m = {Q0 : Q ∈ F a , mQ = m} ⊆
D m;3
for the 3d different values of m. We can define Φ : F a,m → D m;3 by
Φ(Q ) = Q00 ; this is well-defined since Q0 uniquely determines Q, which de-
0

termines φ(Q) and then Q00 . The map Φ is also injective, since Q00 uniquely
determines φ(Q), which (since φ is injective) determines Q and then Q0 . More-
over, we have

`(Φ(Q0 )) = `(Q00 ) = 3`(φ(Q)) = 3`(Q) = `(Q0 ).

Thus F a,m ⊆ D m;3 and Φ satisfy properties (a) and (b) in place of F ⊆ D
and φ, and the scale-separation property (c) is clearly inherited by Φ from φ.
Moreover, the nth D m;3 -ancestor of both Φ(Q0 ) = Q00 and Q0 is clearly 3Q(n)
by construction, and hence Φ satisfies condition (12.10) of Lemma 12.1.20. The
said lemma guarantees that F a,m can be split into 3 subcollections Fja,m , so
that each
{Q0 ∪ Φ(Q0 ) : Q0 ∈ Fja,m } ⊆ D m;3
is strongly nested. Writing i = (a, m, j), and defining

Fi := {Q ∈ F : mQ = m, Q[m] ∈ Fja,m },

these are precisely the collections that we wanted to construct. Since a takes 9d
values, m takes 3d values, and j takes 3 values, the number of these collections
is 9d · 3d · 3 = 33d+1 , as claimed. 
108 12 Dyadic operators and the T (1) theorem

Remark 12.1.24. In each of the Lemmas 12.1.20 through 12.1.23, we can drop
assumption (c) at the cost of multiplying the required number of decomposing
subcollections Fi by n + 1.
Proof. For any F ⊆ D, consider the n + 1 subcollection F k := {Q ∈ F :
log2 `(Q) ≡ k mod (n + 1)}. Each of these clearly satisfies (c). Moreover, any
of the other properties (a) through (e) as well as (12.10), if valid for F , is
clearly inherited by each F k . Thus, if F satisfies the assumptions of any of
these lemmas with the possible exception of (c), then each F k satisfies all of
the relevant assumptions, and the lemma in question provides a decomposition
of F k into some Fik with appropriate nestedness conditions. The Sn required
decomposition of the original F is then obtained simply as F = k=0 i Fik ,
S
and clearly the number of collections inS this decomposition is n + 1 times as
many as in the decompositions F k = i Fik given by the lemmas. 

12.1.c The elementary operators of Figiel

We will now study another family of dyadic singular integral operators with
more complicated interactions between Haar functions at different locations.
The first class of these operators combines the action of a Haar multiplier
with a translation of the Haar functions. One might be tempted to refer to
such operators as dyadic or Haar “shifts”, but this name has been adopted
for a somewhat different class of operators in the literature.
While the parameter n attached with these operators may appear like a
technical detail at this point, it is essential for subsequent applications that
one obtains a good dependence on n.
Theorem 12.1.25 (Figiel). Let φ : D → D be an injection with `(φ(Q)) =
`(Q) and φ(Q) ⊆ 3Q(n) for some n ∈ N. Let X and Y be a UMD spaces, and
let p ∈ (1, ∞). Let λ = (λQ )Q∈D ⊆ L (X, Y ). Consider the mapping
αγ
X γ
Tφλ f= λQ hf, hα
Q ihφ(Q) , (12.14)
Q∈D

initially from S00 (D; X) to S00 (D; Y ). Let Ad := 6 · (81)d .

(0) If λ is R-bounded, or more generally if
(
λφ−1 (Q) , Q ∈ φ(D),
min{E R p (λ), E R p (λφ−1 )} < ∞, (λφ−1 )Q :=
0, else,
αγ
then Tφλ extends boundedly from Lp (Rd ; X) to Lp (Rd ; Y ), with norm
αγ αγ − +
kTφλ k := kTφλ kL (Lp (Rd ;X),Lp (Rd ;Y )) 6 Ad (n + 1)βp,Y βp,X C(X, Y, p; λ),
where
+
C(X, Y, p; λ) := min{βp,X · E R p (λφ−1 ), +
βp,Y · E R p (λ)}
+ +
6 min{βp,X , βp,Y }Rp (λ);
 12.1 Dyadic singular integral operators 109

(1) If, in addition, Y has type t ∈ [1, p] and X has cotype q ∈ [p, ∞], or one
of them has both, then we also have the estimate
αγ − +
kTφλ k 6 Ad (n + 1)1/t−1/q βp,Y βp,X C(X, Y, p, q, t; λ)
where
n
+
C(X, Y, p, q, t; λ) := min τt,X;p · βp,X · cq,X;p · E R p (λφ−1 ),
+
τt,Y ;p · βp,X · cq,X;p · E R p (λφ−1 ),
+
τt,Y ;p · βp,Y · cq,X;p · E R p (λ),
o
+
τt,Y ;p · βp,Y · cq,Y ;p · E R p (λ)
6 C(X, Y, p, q, t) · Rp (λ),
and
n
+ +
C(X, Y, p, q, t) := min τt,X;p βp,X cq,X;p , τt,Y ;p βp,X cq,X;p ,
o (12.15)
+ +
τt,Y ;p βp,Y cq,X;p , τt,Y ;p βp,Y cq,Y ;p .

(2) If, in addition, λQ 6= 0 only when φ(Q) ⊆ Q(n) , then we have the alterna-
tive norm estimate
αγ 0
kTφλ +
k 6 3 · βp,Y βp,X min{cq,X;p , cq,Y ;p }(n + 1)1/q E R p (λ).
0
(3) For all f ∈ Lp (Rd ; X) and g ∈ Lp (Rd ; Y ∗ ), the extended operator has the
absolutely convergent representation
αγ
XD γ
E
hTφλ f, gi = λQ hf, hα Q i, hg, hφ(Q) i .
Q∈D

When kf kLp (Rd ;X) 6 and kgkLp0 (Rd ;Y ∗ ) 6 1, the corresponding absolute
value series is dominated by the same upper bounds as those given for
αγ
kTφλ kL (Lp (Rd ;X),Lp (Rd ;Y )) above.
Remark 12.1.26. (1) In the prominent special case that X = Y , we have
C(X, X, p, q, t; λ) = C(X, X, p, q, t) · min{E R p (λφ−1 ), E R p (λ)},
+
C(X, X, p, q, t) = τt,X;p · βp,X · cq,X;p .

(2) Case (0) of Theorem 12.1.25 is a special case of (1) using the trivial type
and cotype exponents t = 1, q = ∞ with corresponding constants equal
to one. The role of non-trivial type and cotype is to relax the dependence
on the parameter n. The estimate obtained in case (2) is not strictly
comparable to the other two bounds; its main advantage over the other
two is achieving a quadratic bound in terms of the UMD constants, in
contrast to the cubic bound in the other cases.
110 12 Dyadic operators and the T (1) theorem

(3) Recalling the Haar multipliers Hαγ

λ from Theorem 12.1.11, one can check
that, for any θ ∈ {0, 1}d \ {0},
αγ θγ
Tφλ = Tφ1 ◦ Hλαθ = Hλθγ −1 ◦ Tφ1
αθ
φ

where 1 is the constant sequence of all ones. Hence, for the qualitative
conclusion of Theorem 12.1.25, it would suffice to consider just X = Y
and λ = 1, and then combine this special case with Theorem 12.1.11;
however, the reader will quickly realise that this approach would produce
a higher power of the UMD constants in the quantitative conclusion.
Before going into the proof, let us still formulate a corollary in the important
special case when φ : D → D is a bijection:
Corollary 12.1.27. Let φ : D → D be a bijection with `(φ(Q)) = `(Q) and
φ(Q) ⊆ 3Q(n) for some n ∈ N. Let X and Y be a UMD spaces, and let
p ∈ (1, ∞). Suppose that Y has type t ∈ [1, p] and X has cotype q ∈ [p, ∞], or
one of them has both. Let λ = (λQ )Q∈D ⊆ L (X, Y ) be R-bounded, consider
αγ
the mapping Tφλ as in (12.14), and let
αγ αγ
kTφλ k := kTφλ kL (Lp (Rd ;X),Lp (Rd ;Y )) .

(1) We have the norm estimate

αγ
kTφλ k 6 6 · 34d βp,X βp,Y (n + 1)1/t−1/q min{CRp (λ), C ∗ Rp∗0 (λ)}

where

C = C(12.15) (X, Y, p, q, t), C ∗ := C(12.15) (Y ∗ , X ∗ , p0 , t0 , q 0 ).

(2) If, in addition, λQ 6= 0 only when φ(Q) ⊆ Q(n) , then we have the alterna-
tive norm estimate
n 0
o
αγ
kTφλ k 6 3 · βp,X βp,Y min C(n + 1)1/q Rp (λ), C ∗ (n + 1)1/t Rp∗0 (λ) .

where

C = min{cq,X;p , cq,Y ;p }, C ∗ = min{ct0 ,Y ∗ ;p0 , ct0 ,X ∗ ;p0 }

Proof. The first versions of both bounds (i.e, using the first item of the respec-
tive minimums) above are simply those of Theorem 12.1.25, cases (1) and (2),
±
where we estimated all UMD constants by βp,Z 6 βp,Z . The second versions
of both bounds then follow by duality: When φ : D → D is a bijection, one
directly verifies that
αγ ∗
(Tφλ ) = Tφγα
−1 ,λ∗
φ−1

is an operator of the same form, acting from Q00 (Rd ; Y ∗ ) to Q00 (Rd ; X ∗ ) and
0 0
eventually from Lp (Rd ; Y ∗ ) to Lp (Rd ; X ∗ ). If Z ∈ {X, Y } has type t, then Z ∗
 12.1 Dyadic singular integral operators 111

has cotype t0 with ct0 ,Z ∗ ;p0 6 τt,Z;p . (See Proposition 7.1.13; it is formulated
for p = t, but the same short argument is easily modified to give the general
statement.) If a UMD space Z has cotype q, then it has martingale type q
(Proposition 4.3.13), hence Z ∗ has martingale cotype q 0 (Proposition 3.5.29),
and thus cotype q 0 (as observed right before Proposition 4.3.13). Thus we can
apply the case already handled, with (Y ∗ , X ∗ , p0 , t0 , q 0 ) in place of (X, Y, p, q, t),
to get
αγ
kTφλ kL (Lp (Rd ;X),Lp (Rd ;Y )) = kTφγα
−1 ,λ kL (Lp0 (Rd ;Y ∗ ),Lp0 (Rd ;X ∗ ))
φ−1
0 0
6 6 · 34d βp0 ,Y ∗ βp0 ,X ∗ (n + 1)1/q −1/t C(Y ∗ , X ∗ , p0 , t0 , q 0 )Rp0 (λ∗ ).

The claim then follows from βp0 ,Z ∗ = βp,Z and 1/q 0 − 1/t0 = 1/t − 1/q.
The second version of the second bound is obtained from the first version
in the entirely similar way by duality. 
Proof of Theorem 12.1.25. Claim (0) is the special case t = 1, q = ∞ of (1),
so we only need to prove the latter of the two. Let F be a finite collection of
dyadic cubes. Then F and φ satisfy the assumptions of Lemma 12.1.23, except
possibly the scale separation (c). By Remark 12.1.24, the lemma still applies
to produce 33d+1 (n + 1) subcollections Fi ⊆ F with the properties given in
γ
Lemma 12.1.23. Let us write xQ = hf, hα Q i. Since the functions (hQ )Q∈F form
a martingale difference sequence, we have
X γ
X
λQ xQ hφ(Q) p d
6 β −
p,Y εQ λQ xQ hγφ(Q) p d
.
L (R ;Y ) L (Ω×R ;Y )
Q∈F Q∈F

From this point on, we have some flexibility as to when we want to “pull out”
the coefficients λQ . For this reason, let us write zQ ∈ Z for a generic choice
of either zQ = λQ xQ ∈ Y or zQ = xQ ∈ X. We then continue with
X X X
εQ zQ hγφ(Q) = εi0 εQ zQ h0φ(Q)
Lp (Ω×Rd ;Z) Lp (Ω 0 ×Ω×Rd ;Z)
Q∈F i Q∈Fi
X X t 1/t
6 τt,Z;p εQ xQ h0φ(Q) ,
Lp (Ω×Rd ;Z)
i Q∈Fi

where, in the two steps above, we used the facts that

1. when multiplied by the random sign εQ , both the independent random
sign εi0 and the possible difference of the signs of hα 0
φ(Q) (t) and hφ(Q) (t) are
invisible to the norm; and
2. whenever Z has type t ∈ [1, p], then so has Lp (S; Z) (here: S = Ω × Rd ),
and τt,Lp (S;Z);p 6 τt,Z;p by Proposition 7.1.4.
For Q ∈ Fi , let us denote by E(Q) = Q[m(i)] ∪ φ(Q)[m(i)] the sets provided
by Lemma 12.1.23 that form a strongly nested family, as guaranteed by the
said lemma. In particular E(Q) ⊇ Q ∪ φ(Q) and |E(Q)| 6 2 · 3d |Q|. (The
112 12 Dyadic operators and the T (1) theorem

inequality is due to the fact that the cubes Q[m(i)] and φ(Q)[m(i)] are not
necessarily different.) Hence

2 · 3d
Z
1φ(Q) 6 1φ(Q) |Q| = 1φ(Q) 2 · 3d − 1Q 6 2 · 3d EE(Q) 1Q ,
|E(Q)| E(Q)

where the EE(Q) are conditional expectations associated with a nested family,
and hence with a filtration. This allows us to use Stein’s inequality (Theorem
4.2.23) to the effect that
X
εQ zQ h0φ(Q)
Lp (Ω×Rd ;Z)
Q∈Fi
X
6 2 · 3d εQ zQ EE(Q) h0Q (12.16)
Lp (Ω×Rd ;Z)
Q∈Fi
X
+
6 2 · 3d · βp,Z εQ zQ h0Q
Lp (Ω×Rd ;Z)
Q∈Fi

Then
X X t 1/t
εQ zQ h0Q
Lp (Ω×Rd ;Z)
i Q∈Fi
X X q 1/q
6 (33d+1 (n + 1))1/t−1/q εQ zQ h0Q
Lp (Ω×Rd ;Z)
i Q∈Fi
X X
6 (33d+1 (n + 1))1/t−1/q cq,Z;p ε0i εQ zQ h0Q ,
Lp (Ω 0 ×Ω×Rd ;Z)
i Q∈Fi

where, in the two steps above, we used

1. Hölder’s inequality and counting of terms in the other factor; and
2. an application of the cotype q property of Z, recalling that this implies
cotype q for Lp (S; Z) (here: S = Ω × Rd ) with cq,Lp (S;Z);p 6 cq,Z;p when
q ∈ [p, ∞] by Proposition 7.1.4.
By the invisibility of signs multiplying a random εQ , the last norm here is
X X X
ε0i εQ zQ h0Q = ε Q zQ h α
Q .
Lp (Ω 0 ×Ω×Rd ;Z) Lp (Ω×Rd ;Z)
i Q∈Fi Q∈F

If we did not already pull out the coefficients λQ , we do it at this point, after
which we are left with
X
+
ε Q x Q hα
Q 6 βp,X kf kLp (Rd ;X) ,
Lp (Ω×Rd ;X)
Q∈F

where the last step was a direct application of Proposition 12.1.5.

 12.1 Dyadic singular integral operators 113

It remains to collect the various coefficients that we accumulated. In any

− +
case, the first estimate gave βp,Y and the last one βp,X , but depending on
where we pull out the coefficients λQ , we may use the constant of the space
X or Y in place of the generic Z.
If we pull out the λQ before the application of Stein’s inequality in (12.16),
then λQ is the coefficient of hγφ(Q) , hence the coefficient of hR
γ
is λφ−1 (R) , and
thus an application of Remark 12.1.7 produces the factor E R p (λφ−1 ). On the
other hand, pulling out the λQ only after (12.16) leads to a “direct” application
of Remark 12.1.7 and the factor E R p (λ).
Aside from the numerical factors 2 · 3d and (33d+1 (n + 1))1/t−1/q , we get
one of the following:

E R p (λφ−1 ) × τt,X;p × βp,X

+
× cq,X;p ,
τt,Y ;p × E R p (λφ−1 ) × βp,X
+
× cq,X;p ,
+
τt,Y ;p × βp,Y × E R p (λ) × cq,X;p ,
+
τt,Y ;p × βp,Y × cq,Y ;p × E R p (λ),

where the order of the constants reflects the order of applying the related
estimates: Before pulling out the coefficients λQ , we apply estimates on the
Y side, and after that on the X side. Taking the minimum of the four terms,
we arrive at the assertion of the theorem.
The alternative estimate (2): In order to make efficient use of the additional
assumption φ(Q) ⊆ Q(n) when λQ 6= 0, we will need to modify the preceding
considerations at various points.
Let F be a finite collection of dyadic cubes, and F λ := {Q ∈ F : λQ 6==
0}. Then F λ and φ satisfy the assumptions of Lemma 12.1.20, except possibly
the scale separation (c). By Remark 12.1.24, the lemma still applies to pro-
duce 3(n + 1) subcollections Fiλ ⊆ F λ with the properties given in Lemma
α
12.1.20. Let us write xQ = hf, hQ i. In the first step, we simply use the triangle
inequality:
X γ
X X γ
λQ xQ hφ(Q) p d
6 λQ xQ hφ(Q) p d
.
L (R ;Y ) L (R ;Y )
Q∈F i Q∈Fiλ

The more interesting deviations from the previous case begin now.
Note that hQ α
= |Q|−1/2 (1Q+
α
±
− 1Qα− ) for suitable subsets Qα ⊆ Q with
± 1
|Qα | = 2 |Q|. If Q 6= φ(Q), we see that
1 α 1 −1/2
d+
Q := (h + hγφ(Q) ) = |Q| (1Q+ + − 1 −
Qα ∪φ(Q)− ),
2 Q 2 α ∪φ(Q)γ γ

1 α γ 1 −1/2
d−
Q := (hQ − hφ(Q) ) = |Q| (1Q+ − − 1 −
α ∪φ(Q)γ Qα ∪φ(Q)+ )
2 2 γ

form a martingale difference sequence (in either order) on Q ∪ φ(Q), since

either function has average zero on the sets where the other one is constant.
114 12 Dyadic operators and the T (1) theorem

If Q = φ(Q) but α 6= γ, then each of the sets Q± ± 1

α ∩ Qγ has measure 4 |Q|,
and once again
1 α
d+
Q := (h + hγQ ) = |Q|−1/2 (1Q+ + − 1 −
Qα ∩Q− ),
2 Q α ∩Qγ γ

1 α γ
d−
Q := (hQ − hQ ) = |Q|
−1/2
(1Q+ − − 1 −
α ∩Qγ Qα ∩Q+ )
2 γ

form a martingale difference sequence (in either order) on Q∪φ(Q) = Q, since

either function has average zero on the sets where the other one is constant.
Finally, if Q = φ(Q) and α = γ, then the same definition gives d+ α
Q = hQ ,
−
dQ = 0, which is also a (rather trivial) martingale difference sequence.
The conclusion of Lemma 12.1.20, that each {Q ∪ φ(Q) : Q ∈ Fiλ } is
−
strongly nested, guarantees that the whole collection {d+ Q , dQ }Q∈Fiλ can be
organised into a martingale difference sequence. Hence
X
λQ xQ hγφ(Q) p d
L (R ;Y )
Q∈Fiλ
X
−
= λQ xQ (d+
Q − dQ )
Lp (Rd ;Y )
Q∈Fiλ
X (12.17)
−
6 βp,Y εQ λQ xQ (d+
Q + dQ )
Lp (Ω×Rd ;Y )
Q∈Fiλ
X
= βp,Y ε Q λ Q x Q hα
Q ,
Lp (Ω×Rd ;Y )
Q∈Fiλ

where we used the definition of UMD with signs ±εQ multiplying the mar-
tingale differences d± Q , followed by taking an average over the εQ . (It might
appear at first glance that we could have used just the one-sided UMD− prop-
−
erty to arrive at the same conclusion with the smaller constant βp,Y , but this
−
is not the case: an application of the one-sided UMD property would give us
independent random signs, say ε± ±
Q , in front of each dQ , and this is not what
we want.)
For zQ ∈ {xQ , λQ xQ } and Z ∈ {X, Y } we then have
X X
ε Q zQ h α
Q p d
L (Ω×R ;Z)
i Q∈Fiλ

0
X X q 1/q
6 (3(n + 1))1/q ε Q zQ h α
Q
Lp (Ω×Rd ;Z)
i Q∈Fiλ
(12.18)
0 X X
6 (3(n + 1))1/q cq,Z;p ε0i ε Q zQ h α
Q
Lp (Ω×Rd ;Z)
i Q∈Fiλ
0 X
= (3(n + 1))1/q cq,Z;p ε Q zQ h α
Q ,
Lp (Ω×Rd ;Z)
Q∈F λ
 12.1 Dyadic singular integral operators 115

using in the last step the fact that the F λ = i Fiλ is a disjoint partition, so
S
the independent random signs εQ with Q ∈ F λ do not “see” the multiplying
signs ε0i . Hence, pulling out the λQ either at the beginning or at the end of
(12.18) (but in any case only after having replaced the translated hγφ(Q) by
hαQ in (12.17), which in contrast to what happened in the previous case of the
proof), we obtain
X γ
X X
λQ xQ hφ(Q) 6 βp,Y ε Q λ Q x Q hα
Q
Lp (Rd ;Y ) Lp (Rd ;Y )
Q∈F i Q∈Fiλ
0 X
6 βp,Y E R p (λ)(3(n + 1))1/q min{cq,X;p , cq,Y ;p } α
ε Q x Q hQ .
Lp (Rd ;X)
Q∈F λ

Finally, recalling that xQ = hf, hα

Q i and using the contraction principle to
replace F λ ⊆ F by the finite set F = {Q ∈ D : hf, hα Q i 6= 0}, we obtain from
Proposition 12.1.5 that
X X
α +
ε Q x Q hQ p d
6 ε Q x Q hα
Q p d
6 βp,X kf kLp (Rd ;X) ,
L (R ;X) L (R ;X)
Q∈F λ Q∈D

which concludes the estimate.

The representation (3): Let first F ⊆ D be finite. For suitable ηQ ∈ K with
|ηQ | = 1, we have
X D γ
E D X
γ
E
λQ hf, hα
Q i, hg, h φ(Q) i = η Q λ Q hf, h α
ih
Q φ(Q) , g
Q∈F Q∈F
αγ
= hTηλ,φ PF f, gi,

where (ηλ)(Q) := ηQ λQ , and

X
PF f := hf, hθQ ihQ
θ
∈ S00 (D; X)
Q∈F
θ∈{0,1}d \{0}

αγ
is a Haar projection of f ; the action of Tηλ,φ is thus well-defined via the initial
definition on this space. From the previous part of the theorem that we already
proved, we have
X D γ
E
λQ hf, hα
Q i, hg, h φ(Q) i
Q∈F
αγ
6 kTηλ,φ kL (Lp (Rd ;X),Lp (Rd ;Y )) kPF f kLp (Rd ;X) kgkLp (Rd ;Y ∗ ) .

We now apply this estimate with the increasing sequence of finite sets

FN := {Q ∈ D : 2−N < `(Q) 6 2N , dist∞ (Q, 0) 6 2N },

116 12 Dyadic operators and the T (1) theorem
∞
[
whose union is FN = D. The corresponding projection can be expressed
N =1
as [
PFN f = 1FN (EN − E−N )f, FN := Q,
Q∈D−N
dist∞ (Q,0)62N

and this is seen to satisfy kPFN f kLp (Rd ;X) 6 2kf kLp (Rd ;X) and PFN → f in
Lp (Rd ; X) as N → ∞. Thus
X D γ
E
λQ hf, hα
Q i, hg, h φ(Q) i
Q∈D
X D E
γ
= lim λQ hf, hα
Q i, hg, hφ(Q) i
N →∞
Q∈FN
αγ
6 kTηλ,φ kL (Lp (Rd ;X),Lp (Rd ;Y )) lim kPF f kLp (Rd ;X) kgkLp (Rd ;Y ∗ )
N →∞
αγ
= kTηλ,φ kL (Lp (Rd ;X),Lp (Rd ;Y )) kf kLp (Rd ;X) kgkLp (Rd ;Y ∗ ) ,
αγ αγ
where Tηλ,φ has the same norm estimate as Tλ,φ , since

E R p (ηλ) = E R p (λ), E R p ((ηλ)φ−1 ) = E R p (λφ−1 )

by the contraction principle.

Thus we have shown the claimed absolute convergence, and hence the
bilinear form X
αγ
tλφ (f, g) := α
hλQ hf, hQ i, hg, hγQ ii
Q∈D

is well-defined and bounded from Lp (Rd ; X)×Lp (Rd ; Y ) to K. So is the bilinear

αγ αγ
form hTλφ f, gi, where Tλφ denotes the bounded extension of the operator
initially defined on S00 (D; X). Moreover, these bilinear forms clearly coincide
when f ∈ S00 (D; X) and g ∈ S00 (D; Y ∗ ). By density, they must coincide for
all f and g, and the proof is complete. 
The second class of operators that we deal with in this section have the addi-
tional twist of “tearing apart” the supports of Haar functions. The relevance
of this feature will be justified in the appearance of this type of operators in
the proof of the T (1) theorem further below.

Theorem 12.1.28 (Figiel). Let φ : D → D be an injection with `(φ(Q)) =

`(Q) and φ(Q) ⊆ 3Q(n) for some n ∈ N. Let X and Y be a UMD spaces and
p ∈ (1, ∞). Let λ = (λQ )Q∈D ⊆ L (X, Y ), and consider the mapping
γ
X
Uφλ : f 7→ λQ hf, hγQ i(h0φ(Q) − hQ
0
), (12.19)
Q∈D

initially from S00 (D; X) to S0 (D; Y ). Let Bd := 5200 · (81)d .

 12.1 Dyadic singular integral operators 117

(0) If λ ⊆ L (X, Y ) is R-bounded, or more generally if E R p (λ) < ∞, then

γ
Uφλ extends boundedly from Lp (Rd ; X) to Lp (Rd ; Y ) with norm

kUφγ k := kUφγ kL (Lp (Rd ;X),Lp (Rd ;Y ))

−
6 Bd · (n + 1) · βp,Y +
· βp,X +
· min{βp,X Rp (λ), βp,Y
+
E R p (λ)}.

(1) If, in addition, X or Y has cotype q ∈ [p, ∞], then we also have
(
γ 1−1/q − + C(X, Y, p, q) · Rp (λ),
kUφ k 6 Bd (n + 1) βp,Y βp,X +
βp,Y · min{cq,X;p , cq,Y ;p } · E R p (λ),

where
n o
+ + +
C(X, Y, p, q) := min βp,X cq,X;p , βp,Y cq,X;p , βp,Y cq,Y ;p
(12.20)
= C(12.15) (X, Y, p, q, 1).

(2) If, in addition, we have λQ 6= 0 only when φ(Q) ⊆ Q(n) , then we have the
alternative norm estimate

kUφγ k 6 6 · (n + 1)1−1/q · βp,Y · βp,X

+
· min{cq,X;p , cq,Y ;p } · E R p (λ).
0
(3) For all f ∈ Lp (Rd ; X) and g ∈ Lp (Rd ; Y ∗ ), the extended operator has the
absolutely convergent representation
γ
XD E
hUφλ f, gi = λQ hf, hγQ i, hg, h0φ(Q) − h0Q i .
Q∈D

When kf kLp (Rd ;X) 6 1 and kgkLp0 (Rd ;Y ∗ ) 6 1, the corresponding absolute
value series is dominated by the same upper bounds as those given for
γ
kUφλ kL (Lp (Rd ;X),Lp (Rd ;Y )) above.

Remark 12.1.29. We have observations analogous to Remark 12.1.26:

+
(1) When X = Y , we have C(X, X, p, q) = βp,X cq,X;p .
(2) Case (0) of Theorem 12.1.28 is a special case of (1) using the cotype
exponent q = ∞ with corresponding constant equal to one. The role of
finite cotype is to relax the dependence on the parameter n. As in Theorem
12.1.25(2), the main point of the alternative bound (2) to improve the
cubic dependence on the UMD constants to a quadratic one; in contrast to
the situation in Theorem 12.1.25(2), when X = Y , the present alternative
bound (2) is a strict improvement of (1), in view of the fact that βp,X 6
− +
βp,X βp,X (Proposition 4.2.3).
(3) Recalling the Haar multipliers Hλαγ from Theorem 12.1.11, one can check
that, for any θ ∈ {0, 1}d \ {0},
γ
Uφλ θ
= Uφ1 ◦ Hλγθ ,
118 12 Dyadic operators and the T (1) theorem

where 1 is the constant sequence of all ones. Hence, for the qualitative
conclusion of Theorem 12.1.28, it would suffice to consider just X = Y
and λ = 1, and then combine this special case with Theorem 12.1.11;
however, the reader will quickly realise that this approach would produce
a higher power of the UMD constants in the quantitative conclusion.
(4) In contrast to Theorem 12.1.25, our proof of Theorem 12.1.28 does not
allow replacing the assumptions on λ by E R p (λφ−1 ) < ∞. The related
issue of when in the argument, and under what assumptions, we may pull
out the coefficients λQ , is shortly discussed inside the proof.

Proof of Theorem 12.1.28. Claim (0) is the special case q = ∞ of (1), so it

suffices to consider the latter of these two claims. Let F ⊆ D be finite. An
additional challenge compared to the proof of Theorem 12.1.25 is that, unlike
α
the Haar functions hφ(Q) , the functions h0φ(Q) − h0Q do not necessarily form a
martingale difference sequence, preventing a straightforward introduction of
the random signs in the initial step. Instead, a decomposition of F is necessary
from the beginning.
Let us denote by F k = {Q ∈ F : log2 `(Q) ≡ k mod (n + 1)} the
scale-separated subcollections of F as in Remark 12.1.24. Then F k and φ
satisfy the assumptions of both Lemmas 12.1.22 and 12.1.23. Let us denote
the decomposing subcollections of Fk provided by Lemma 12.1.22 by Aak and
those provided by Lemma 12.1.23 by Bbk , let Fik = Aak ∩ Bbk for i = (a, b),
and let Fi consists of an enumeration of all these Fik . The total number of
these Fi is then 144 · 33d+1 · (n + 1), and they satisfy the conclusions of both
Lemmas 12.1.22 and 12.1.23.
We first make use of Lemma 12.1.22. For Q ∈ Fi , we have
2
X
0
hφ(Q) − h0Q = h0φi,3 (Q) − h0φi,0 (Q) = (hφ0 i,j+1 (Q) − h0φi,j (Q) ),
j=0

where each collection {φi,j (Q) ∪ φi,j+1 (Q) : Q ∈ Fi } is strongly nested. But
this implies that each

(h0φi,j+1 (Q) − h0φi,j (Q) )Q∈Fi

is (or can be enumerated as) a martingale difference sequence. Note that here
it is important that a smaller union φi,j+1 (Q)∪φi,j (Q) is not just contained in
a larger φi,j+1 (R)∪φi,j (R), but entirely in (a dyadic child of) one of φi,j+1 (R)
or φi,j (R), where the function h0φi,j+1 (R) − hφ0 i,j (R) is constant.
Using this martingale difference property, we can then proceed as in the
proof of Theorem 12.1.25. Let us abbreviate xQ := hf, hγQ i ∈ X and yQ :=
λQ x Q ∈ Y .
 12.1 Dyadic singular integral operators 119
X
yQ (h0φ(Q) − hQ
0
)
Lp (Rd ;Y )
Q∈F
2 X
X X
6 yQ (h0φi,j+1 (Q) − h0φi,j (Q) )
Lp (Rd ;Y )
j=0 i Q∈Fi
2 X
X X
−
6 βp,Y εQ yQ (h0φi,j+1 (Q) − hφ0 i,j (Q) )
Lp (Ω×Rd ;Y )
j=0 i Q∈Fi
3
(
−
X X X α0 = α3 = 1,
6 βp,Y αj εQ yQ h0φi,j (Q) ,
j=0 i Q∈Fi
Lp (Ω×Rd ;Y ) α1 = α2 = 2,

where the first and the last steps were simply triangle inequalities.
As in the proof of Theorem 12.1.25, we have some flexibility on when to
pull out the coefficients λQ , and we again proceed with a generic choice of
zQ ∈ Z for either yQ ∈ Y or xQ ∈ X. The norm to be estimated has exactly
the same form as what we estimated (12.16) in the proof of Theorem 12.1.25
(using Lemma 12.1.23 in this step), and we can there read the bound
X
εQ zQ hφ0 i,j (Q)
Lp (Ω×Rd ;Z)
Q∈Fi
X (12.21)
+
6 2 · 3d · βp,Z 0
ε Q zQ h Q .
Lp (Ω×Rd ;Z)
Q∈Fi

By Hölder’s inequality, we have

X X
εQ zQ h0Q p
L (Ω×Rd ;Z)
i Q∈Fi

1/q0  X X q 1/q
6 144 · 33d+1 · (n + 1) 0
ε Q zQ h Q .
Lp (Ω×Rd ;Z)
i Q∈Fi

Invoking cotype q of Z, and recalling that this implies cotype q of Lp (S; Z)

(here: S = Ω × Rd ) with constant cq,Lp (S;Z);p 6 cq,Z;p when q ∈ [p, ∞] by
Proposition 7.1.4, we continue with
X X q 1/q
εQ zQ h0Q
Lp (Ω×Rd ;Z)
i Q∈Fi
X X
6 cq,Lp (Z);p εi0 εQ zQ h0Q
Lp (Ω 0 ×Ω×Rd ;Z)
i Q∈Fi
X
= cq,Lp (Z);p εQ zQ hγQ .
Lp (Ω×Rd ;Z)
Q∈F

It is no later than here that we should to pull out the coefficients λQ , after
which we are left with the final step, based on Proposition 12.1.5, that
120 12 Dyadic operators and the T (1) theorem
X γ +
ε Q x Q hQ 6 βp,X kf kLp (Rd ;X) .
Lp (Ω×Rd ;X)
Q∈F

Under the assumption of R-boundedness of λ, depending on the moment

of pulling out the coefficients λQ , the constants that we accumulate in the
various steps with the option of estimating in Z ∈ {X, Y } produce, aside

1/q0
from the numerical factors 2 · 3d and 144 · 33d+1 · (n + 1) , one of the
products

Rp (λ) · βp,X
+
· cq,X;p ,
+
βp,Y · Rp (λ) · cq,X;p ,
+
βp,Y · cq,Y ;p · Rp (λ).

In the latter two versions, i.e., pulling out the λQ only after making the step
(12.21) with Z = Y , we might as well replace Rp (λ) by E R p (λ), thus leading
to the possible upper bounds
+
βp,Y · E R p (λ) · cq,X;p ,
+
βp,Y · cq,Y ;p · E R p (λ).

(On the other hand, if we wanted to pull out the λQ before step (12.21), and
thus apply (12.21) with Z = X, the coefficient λQ would be multiplying a
Haar function h0φi,j (Q) ; this would lead to a constant of the type E R p (λφ−1 ),
i,j
where φi,j need not be the original φ from the assumptions of the theorem, but
one of the auxiliary mappings produced by Lemma 12.1.22. This would lead
to an unreasonably technical formulation of probably little practical value,
which is why we have not included the resulting alternative upper bound in
the statement of the theorem.)
Altogether, choosing the best of the possible alternative estimates, we
arrive at
X
xQ (h0φ(Q) − h0Q ) p d kf k−1 Lp (Rd ;X)
L (R ;X)
Q∈F
3
−
X
1/q0 +
6 βp,X αj (2 · 3d )(144 · 33d+1 · (n + 1) βp,X ×
j=0
(
C(X, Y, p, q)Rp (λ),
× +
βp,Y min{cq,X;p , cq,Y ;p }E R p (λ),
P3
where C(X, Y, p, q) is as in the statement of the Theorem, and j=0 αj =
1 + 2 + 2 + 1 = 6.
The alternative estimate (2): As in the previous proof of Theorem 12.1.25(2),
we construct some auxiliary martingale differences. The initial considerations
are identical:
 12.1 Dyadic singular integral operators 121

Let again F be a finite collection of dyadic cubes, and F λ := {Q ∈ F :

λQ 6= 0}. Then F λ and φ satisfy the assumptions of Lemma 12.1.20, except
possibly the scale separation (c). By Remark 12.1.24, the lemma still applies
to produce 3(n + 1) subcollections Fiλ ⊆ F λ with the properties given in
Lemma 12.1.20. Let us write xQ = hf, hα Q i. In the first step, we simply use
the triangle inequality:
X X X
0 0 0
λQ xQ (hφ(Q) − hQ ) 6 λQ xQ (hφ(Q) − h0Q ) .
Lp (Rd ;Y ) Lp (Rd ;Y )
Q∈F i Q∈Fiλ

The slight symmetry break between hα 0 0

Q and hφ(Q) − hQ is also reflected
in the construction of the auxiliary martingale differences. As in the proof of
Theorem 12.1.25(2), we denote Q± α
α := Q ∩ {sgn(hQ ) = ±1}. If φ(Q) 6= Q, we
choose
1 1 −1/2
dQ := |Q| (1φ(Q)∪Q+ − 3 · 1Q− ),
3 α α

1
d2Q := |Q|−1/2 (−1φ(Q) + 2 · 1Q+ ),
3 α

where d2Q has average zero on the sets where d1Q is constant; note that, unlike
in the proof of Theorem 12.1.25(2), the order matters now. Moreover, we can
recover the original functions by
1 −1/2
d1Q + d2Q = = hα


|Q| (1 − 1)1φ(Q) + (1 + 2)1Q+ − 3 · 1Q − Q,
3 α α

1
d1Q − 2d2Q = |Q|−1/2 (1 + 2)1φ(Q) ) + (1 − 4)1Q+ = h0φ(Q) − h0Q .


− 3 · 1Q −
3 α α

If φ(Q) = Q, then h0φ(Q) − h0Q = 0, and we can simply set d1Q := hα

Q and
2
dQ = 0, and the original functions are recovered by

hα 1 1 2
Q = dQ = dQ + dQ , h0φ(Q) − h0Q = 0 = 0 · d1Q − 2d2Q .

The conclusion of Lemma 12.1.20, that {Q ∪ φ(Q) : Q ∈ Fiλ } is strongly

nested, ensures that the full collection {d1Q , d2Q }Q∈Fiλ , appropriately enumer-
ated, is a martingale difference sequence. Hence
X
λQ xQ (h0φ(Q) − h0Q ) p d
L (R ;Y )
Q∈Fiλ
X
−
λQ xQ (1 − δQ,φ(Q) )d+


= Q − 2 · dQ
Lp (Rd ;Y )
Q∈Fiλ
X (12.22)
−
6 2βp,Y εQ λQ xQ (d+
Q + dQ )
Lp (Ω×Rd ;Y )
Q∈Fiλ
X
= 2βp,Y ε Q λ Q x Q hα
Q
Lp (Ω×Rd ;Y )
Q∈Fiλ
122 12 Dyadic operators and the T (1) theorem

as an application of the definition of UMD via martingale transforms with a

multiplying sequences of numbers {0, 1, −2} × εQ , and averaging over inde-
pendent random εQ .
Except for the factor 2, the right side of (12.22) coincides with the right
side of (12.17) from the proof of Theorem 12.1.25(2). Hence the rest of the
estimate can be concluded by repeating the said proof verbatim.
The representation (3): This is proved in the same way as the corresponding
part of Theorem 12.1.25. 

12.2 Paraproducts
The notion of paraproducts arises from a number of considerations. Here we
choose a point of departure that also motivates their name: they are objects
that arise from a decomposition of the ordinary pointwise product of functions.
While paraproducts certainly look more complicated than the regular product,
it turns out that in certain respects they actually behave better. Another
motivation is the key role that these objects play in the T (1) theorem in
Section 12.3. Some further connections will be discussed in the Notes.

Proposition 12.2.1. Let b ∈ L1loc (Rd ; L (X, Y )), where X and Y are Banach
spaces, and let f ∈ S00 (D; X). Then
X
bf = Hbαγ f + Πb f + Πb∗ f, (12.23)
α,γ∈{0,1}d \{0}

where Hαγ
b are Haar multipliers of the form
X
Hbαγ f := hsgn(hα γ α γ
Q hQ )biQ hf, hQ ihQ ,
Q∈D

and the remaining terms are the paraproducts

X
Πb f := hb, hα α
Q ihf iQ hQ ,
Q∈D
α∈{0,1}d \{0}

X 1Q
Πb∗ f := hb, hα α
Q ihf, hQ i ,
|Q|
Q∈D
α∈{0,1}d \{0}

where the series of Πb∗ f is finitely non-zero, and the non-zero terms in Πb f
are attached to cubes contained in finitely many maximal ones, and the series
converges (at least) conditionally along any decreasing order of the dyadic
cubes contained in these maximal ones.
 12.2 Paraproducts 123

The notation Πb∗ is motivated by the easily verified duality relation

hΠb∗ f, gi = hf, Πb∗ gi, f ∈ S00 (D; X), g ∈ S00 (D; Y ),

where b∗ ∈ L∞ (Rd ; L (Y ∗ , X ∗ )) is the pointwise adjoint of b.

Remark 12.2.2. The diagonal α = γ of the sum in (12.23) is

X X
Hbαα f = α
hbiQ hf, hQ ihαQ
α∈{0,1}d \{0} Q∈D
α∈{0,1}d \{0}

This has formally the same structure as Πb f , but with the roles of b and f
reversed, and hence (12.23) could be also written in the form
X
bf = Hαγ ∗
b f + Πf b + Πb f + Πb f,
α,γ∈{0,1}d \{0}
α6=γ

where the summation is empty in dimension d = 1 (since there is only one

possible value of α ∈ {0, 1}\{0}). It is also evident that Πb∗ f is symmetric in b
and f , and hence a more symmetric notation could also be preferred. However,
we shall not pursue this point of view any further, since the roles played by
the two functions b and f will be quite different in our main applications, so
that such symmetries would be only misleading.

Proof of Proposition 12.2.1. It suffices to prove this for f = x ⊗ hθR . Then

X
bf = (b − hbiR )f + hbiR f = hb, hα α θ θ
Q ix ⊗ hQ hR + hbiR x ⊗ hR ,
Q⊆R
α∈{0,1}d \{0}

where the series converges (at least) conditionally along any decreasing order
of the dyadic cubes Q ⊆ R, by the Martingale Converge Theorem 3.3.2, since
this is a martingale difference expansion of the function 1R (b − hbiR )x ∈
L1 (Rd ; Y ).
We observe that
α θ
hQ hR = hα θ
Q hhR iQ ∀Q ( R,

whereas
θ θ 1R hα+θ
hR hR = , hα θ
R hR
R
, ∀α 6= θ,
|R| |R|1/2
where we use modulo 2 addition in {0, 1}d . Hence
X X
hb, hα α θ
Q ix ⊗ hQ hR = hb, hα α
Q ihf iQ ⊗ hQ = Πb f,
Q(R Q∈D
α∈{0,1}d \{0} α∈{0,1}d \{0}
124 12 Dyadic operators and the T (1) theorem

θ
observing that hf iQ = hhR iQ x = 0 unless Q ( R. Moreover,
X
hb, hα α θ θ
R ix ⊗ hR hR + hbiR x ⊗ hR
α∈{0,1}d \{0}
 1R X hα+θ  hb, h0 i
= hb, hθR ix ⊗ + hb, hα
R ix ⊗
R
+ R
x ⊗ hθR
|R| |R|1/2 |R|1/2
α∈{0,1}d \{0,θ}

1R X hα+θ
= hb, hθR ix ⊗ + hb, hα
R ix ⊗
R
.
|R| |R|1/2
α∈{0,1}d \{θ}

Using the orthogonality of the Haar functions, we see that

1R X 1Q
hb, hθR ix ⊗ = hb, hα α
Q ihf, hQ i ⊗ = Πb∗ f.
|R| |Q|
Q∈D
α∈{0,1}d \{0}

Finally, with the change of variable γ := α + θ

X hα+θ X hγR
hb, hα
R ix ⊗
R
= hb, hγ+θ
R ix ⊗
|R|1/2 |R|1/2
α∈{0,1}d \{θ} γ∈{0,1}d \{0}
X X
= hb sgn(hγR hθR )iR x ⊗ hγR = haθγ γ
R biR x ⊗ hR
γ∈{0,1}d \{0} γ∈{0,1}d \{0}
X X X
= haαγ α γ
Q biQ hf, hQ i ⊗ hQ = Hαγ
b f,
α,γ∈{0,1}d \{0} Q∈D α,γ∈{0,1}d \{0}

again by the orthogonality of the Haar functions in the penultimate step. 

Proposition 12.2.3. Let X and Y be UMD spaces and p ∈ (1, ∞). Let
b ∈ L∞ (Rd ; L (X, Y )). Then Λb := Πb + Πb∗ , initially defined on S00 (D; X),
extends to a bounded operator from Lp (Rd ; X) to Lp (Rd ; Y ) of norm
 
− + +
kΛb kL (Lp (Rd ;X),Lp (Rd ;Y )) 6 1 + (2d − 1)2 βp,Y βp,Y βp,X kbkL∞ (Rd ;L (X,Y )) ,

and we have the identity

X
bf = Hαγ
b f + Λb f ∀ f ∈ Lp (Rd ; X).
α,γ∈{0,1}d \{0}

We will obtain a far better estimate in Theorem 12.2.25, but it seems worth-
while recording this relatively simple bound as an illustration of the techniques
that we have developed thus far.
Proof of Proposition 12.2.3. It is clear that pointwise multiplication by b ∈
L∞ (Rd ; L (X, Y )) defines a bounded operator from Lp (Rd ; X) to Lp (Rd ; Y ),
 12.2 Paraproducts 125

for any Banach spaces X, Y and all p ∈ [1, ∞]. Moreover, the Haar multi-
plier Hαγ
b featuring in Proposition 12.2.3 have exactly the form considered in
Proposition 12.1.13, and hence

kHαγ − + +
b f kLp (Rd ;Y ) 6 βp,Y βp,Y βp,X kbkL∞ (Rd ;L (X,Y )) kf kLp (Rd ;X) .

By triangle inequality, it then follows from (12.23) that

X
kΛb f kLp (Rd ;Y ) 6 kbf kLp (Rd ;Y ) + kHbαγ f kLp (Rd ;Y )
α,γ∈{0,1}d \{0}
 
− + +
6 kbkL∞ (Rd ;L (X,Y )) 1 + (2d − 1)2 βp,Y βp,Y βp,X kf kLp (Rd ;X)

for all f ∈ S00 (D; X), and hence Λb extends to a bounded operator from
Lp (Rd ; X) to Lp (Rd ; Y )) with the asserted norm estimate. Since the claimed
identity holds (by Proposition 12.2.1) for all f ∈ S00 (D; X), and each term is
continuous with respect to the Lp (Rd ; X) norm of f (as we just showed), it is
immediate that this identity extends to all f ∈ Lp (Rd ; X). 
As we shall see later, the operator Λb is not only as good as, but actually better
than the pointwise product f 7→ bf , in the sense that it remains a bounded
operator for a broader class of functions b than just the bounded ones. As
the reader will have guessed from the introduced notation, we will also be
interested in the mapping properties of the individual paraproducts Πb and
Πb∗ .
While the paraproduct Πb arouse from our analysis of the pointwise prod-
uct with a multiplier b, in other considerations we will encounter similar series
X
α
Πf = πQ hf iQ ⊗ hα
Q
Q∈D
α∈{0,1}d \{0}

α α
with some coefficient πQ replacing the Haar coefficients hb, hQ i of a function
α α
b above. Formally, we always have πQ = hb, hQ i by choosing
X
α
“ b := Π(1) = πQ ⊗ hα
Q ”,
Q∈D
α∈{0,1}d \{0}

but giving a precise meaning for this series requires non-trivial considerations
in general, and it is hence useful not to insist in the a priori existence of
function b generating the coefficients in this way.

12.2.a Necessary conditions for boundedness

As we already saw in the analysis of the pointwise product bf , and we will

see again in the T (1) theorem below, paraproducts frequently appear in pairs
126 12 Dyadic operators and the T (1) theorem

of the form Π1 + Π2∗ , where Π1 is a paraproduct as in the previous section,

and Π2∗ is the formal adjoint of another paraproduct. In other words, we are
concerned with the operator

α 1Q
X  
1,α 2,α
Λf := πQ hf iQ hα
Q + π Q hf, h Q i . (12.24)
|Q|
Q∈D
α∈{0,1}d \{0}

Of course this covers both Π1 and Π2∗ as special cases, by simply setting some
i,α
of the coefficients πQ equal to zero.
Compared to the operator Λb featuring in Proposition 12.2.3, we now allow
1,α 2,α
possibly different coefficients πQ and πQ in the first and second term above,
as this will be relevant in the T (1) theorem. Via the duality relations

hΛf, gi = hf, Λ∗ gi = L(f, g), f ∈ S00 (D; X), g ∈ S00 (D; Y ∗ ),

we define the formal adjoint

X 
1,α ∗ 1Q 2,α ∗

Λ∗ g := (πQ ) hg, hα
Qi + (πQ ) hgiQ hα
Q (12.25)
|Q|
Q∈D
α∈{0,1}d \{0}

which has exactly the same form as Λ, only with different coefficients, and the
associated bilinear form
X D E D E
1,α 2,α
L(f, g) := πQ hf iQ , hhα α
Q , gi + πQ hf, hQ i, hgiQ . (12.26)
Q∈D
α∈{0,1}d \{0}

Lemma 12.2.4. The series (12.26) is finitely non-zero whenever

(f, g) ∈ S00 (D; X) × S(D; Y ∗ ) ∪ S(D; X) × S00 (D; Y ∗ ) .

In particular, we have

L(x ⊗ 1R , y ∗ ⊗ hβR ) = hπR

1,β
x, y ∗ i, L(x ⊗ hβR , y ∗ ⊗ 1R ) = hπR
2,β
x, y ∗ i

for all x ∈ X, y ∗ ∈ Y ∗ , R ∈ D and β ∈ {0, 1}d \ {0}.

Proof. By symmetry, it is enough to consider (f, g) ∈ S00 (D; X) × S(D; Y ∗ ).

We may further assume that

f = x ⊗ hβP , g = y ∗ ⊗ 1R

for some x ∈ X, y ∗ ∈ Y ∗ , P, R ∈ D and β ∈ {0, 1}d \ {0}, since general f and

g are finite linear combinations of such functions.
For such f and g, the (Q, α) term in (12.26), is given by

hπ 1,α x, y ∗ ihhβP iQ hhα

Q , 1R i + hπ
2,α
x, y ∗ ihhβP , hα
Q ih1R iQ ,
 12.2 Paraproducts 127
β
where hhP iQ 6= 0 only if Q ( P , while hhα
Q , 1R i 6= 0 only if R ( Q; finally,
β α
hhP , hQ i 6= δP,Q δα,β . Thus
X
L(x ⊗ 1P , y ∗ ⊗ hβR ) := 1,α
hπQ β
x, y ∗ ihhP α
iQ hhQ , 1R i + hπP2,β x, y ∗ ih1R iP ,
Q∈D
R(Q(P
α∈{0,1}d \{0}

which is clearly a finite sum. When P = R, the sum above is void, and we get
β 2,β
L(x ⊗ 1R , y ∗ ⊗ hR ) = hπR x, y ∗ i.

The other case follows by symmetry. 

Although our main concern is Lp boundedness, we formulate the following
lemma slightly more generally, since the additional generality comes essentially
for free.

Lemma 12.2.5. Let p, q ∈ [1, ∞). A necessary condition for L to satisfy

|L(f, g)| 6 Ckf kLp (Rd ;X) kgkLq0 (Rd ;Y ∗ ) ,

uniformly for all (f, g) of the form (x ⊗ 1Q , y ∗ ⊗ hQ

α
) and (x ⊗ hα ∗
Q , y ⊗ 1Q ),
is that
i,α 1 1 1 3
kπQ kL (X,Y ) 6 C|Q|γ , γ := − + < . (12.27)
p q 2 2

On the other hand, assuming the coefficient bound (12.27), the defining series
(12.26) of L(f, g) converges absolutely for all

(f, g) ∈ S(D; X) × S(D; Y ∗ )

Proof. We have
1,α
|hπQ x, y ∗ i| = |L(x ⊗ 1Q , y ∗ ⊗ hα
Q )|

6 Ckx ⊗ 1Q kLp (Rd ;X) ky ∗ ⊗ hα

Q kLq0 (Rd ;Y ∗ )
0
= CkxkX |Q|1/p ky ∗ kY ∗ |Q|1/q −1/2
= CkxkX ky ∗ kY ∗ |Q|1/p−1/q+1/2

and taking the supremum over ky ∗ kY ∗ 6 1 and kxkX 6 1 proves the estimate
for i = 1. The case i = 2 is entirely symmetric. Finally, note that 1/p, 1/q ∈
(0, 1] so that 1/p − 1/q < 1.
To prove the convergence, it is enough to consider f = x⊗1P , g = y ∗ ⊗1R ,
1,α
and moreover, by symmetry, just the first half of L (f, g) with coefficients πQ .
Now
1,α 1,α ∗
|hπQ hf iQ , hg, hα α
Q ii| = |hπQ x, y i|h1P iQ |h1R , hQ i|,
128 12 Dyadic operators and the T (1) theorem

where
1,α |P | |R|
|hπQ x, y ∗ i| 6 C|Q|γ kxkX ky ∗ kY ∗ , h1P iQ 6 , |h1R , hα
Q i| 6 ,
|Q| |Q|1/2
and moreover the last pairing is non-zero only if Q ) R. Hence the absolute
convergence of the series follows from the convergence of
X ∞
X
|Q|γ−3/2 = |R|γ−3/2 2kd(γ−3/2) < ∞,
Q∈D k=1
Q)R

since this is as a geometric series with γ − 3/2 < 0. 

Lemma 12.2.6. Suppose that the series defining Λf converges (even just con-
ditionally) in Lp (Rd ; Y ) for some f = 1R ⊗ x, where R ∈ D and x ∈ X. Then
X 1,α
(1R − ER )Λ(1R ⊗ x) = πQ x ⊗ hα
Q
Q⊆R
α∈{0,1}d \{0}

Proof. We have
(
1,α

1,α 2,α 1Q  πQ x ⊗ hα
Q + 0, Q ⊆ R,
1R πQ h1R ⊗ xiQ hα
Q + πQ h1R ⊗ x, hα
Qi = α
|Q| yQ,R ⊗ 1R , Q 6⊆ R,
α
for some yQ,R ∈ Y , which is not difficult to find explicitly, but it is irrelevant
for the present purposes. The assumed convergence in Lp (Rd ; Y ), and the
boundedness of the conditional expectation ER and the pointwise multiplier
1R on Lp (Rd ; Y ) guarantee that we can move (1R − ER ) inside the defining
α α
series. Since ER (yQ,R ⊗ 1R ) = yQ,R ⊗ 1R , we have
X 1,α
(1R − ER )Λ(1R ⊗ x) = πQ x ⊗ hα
Q,
Q⊆R
α∈{0,1}d \{0}

as claimed. 
α
Lemma 12.2.7. Let Y be a Banach space, and p ∈ [1, ∞). Let yQ ∈ Y for
all Q ∈ D, α ∈ {0, 1} \ {0}. For each R ∈ D and n ∈ N, consider the sum
d

X
n α
BR := yQ ⊗ hα
Q
Q⊆R
`(Q)>2−n `(R)
α∈{0,1}d \{0}

Suppose that, for every R ∈ D, we have one of the following

n
(1) BR := limn→∞ BR exists in Lp (Rd ; Y ), or
 12.2 Paraproducts 129

n
(2) Y has the Radon–Nikodým property, and supn∈N kBR kLp (Rd ;Y ) < ∞.
p
Then there exists a function b ∈ Lloc (Rd ; Y ) such that

1R (b − hbiR ) = BR , α
hb, hR α
i = yR , ∀R ∈ D, α ∈ {0, 1}d \ {0}.

If, moreover, the supremum below is finite, then b ∈ BMOD (Rd ; Y ) and
α
kyQ kY kBR − ckLp (Rd ;Y )
sup 6 kbkBMOpD (Rd ;Y ) = sup inf (12.28)
Q∈D |Q|1/2 R∈D c∈Y |R|1/p
α∈{0,1}d \{0}

n ∞
Proof. It is immediate to verify that (BR )n=0 is a martingale in Lp (Rd ; Y ).
By the Martingale Convergence Theorem 3.3.16, it follows that (2) implies
(1). Hence it suffices to prove the lemma under assumption (1).
We construct the function b via the correspondence established in Lemma
11.2.11. It is enough to construct b|S separately for each quadrant S ⊆ Rd .
So we fix a quadrant S ⊆ Rd , and let
X
∆(s, t) := (hα α α
Q (s) − hQ (t))yQ ,
Q∈D(S)
α∈{0,1}d \{0}

where we need to justify the convergence of this series in some sense. We will
prove that it converges in Lploc (S × S; Y ). To this end, note that any bounded
subset of S × S is contained in R × R for some R ∈ D(S). For s, t ∈ R, only
Q ∈ D(S) with Q ∩ R 6= ∅ can contribute to the series; moreover, if Q ) R,
then hα α α
Q is constant on R, and hence hQ (s) − hQ (t) = 0 for s, t ∈ R. Thus
X
(1R×R ∆)(s, t) = 1R×R (s, t) (hα α α
Q (s) − hQ (t))yQ
Q∈D(R)
α∈{0,1}d \{0}
(12.29)
= 1R×R (s, t)(BR (s) − BR (t)),

where the (conditional) convergence in Lp (R ×R, ds dt; Y ) follows by Fubini’s

theorem from the assumed (conditional) convergence of each BR in Lp (R; Y ).
Now that the convergence of the defining series of ∆(s, t) has been justified,
it is immediate from the defining formula that ∆(s, t) + ∆(t, u) = ∆(s, u) for
s, t, u ∈ S. By Lemma 11.2.11, we have the existence of b : S → Y such that
∆(s, t) = b(s) − b(t). substituting this into (12.29), we obtain

b(s) − b(t) = BR (s) − BR (t), for s, t ∈ R,

and hence b(·) = BR (·) + (b(t) − BR (t)) ∈ Lp (R; Y ) ⊆ L1 (R; Y ). Taking the
average over t ∈ R, it follows that

b(s) − hbiR = BR (s) − hBR iR = BR (s), for s ∈ R,

130 12 Dyadic operators and the T (1) theorem

observing that hhα Q iR = 0 for all Q ⊆ R that appear in the series of BR . Then
it also follows that

hb, hα α α α
R i = h1R (b − hbiR ), hR i = hBR , hR i = yR .

This also implies, for any c ∈ Y , that

α
kyQ kY hα
Z Z 1/p
D
Q
E 1 p
= B Q − c, 6 kB Q − ckY 6 − kB Q − ck Y ,
|Q|1/2 |Q|1/2 Y Q |Q| Q

and (12.28) follows from the identity BQ = 1Q (b − hbiQ ), which implies that

kbkBMOpD (Rd ;Y ) : = sup inf k1Q (b − c)kLp (Q;Y )

Q∈D c∈Y

= sup inf
0
k1Q (BQ − c0 )kLp (Q;Y )
Q∈D c ∈Y

by a simple change of variable. 

Proposition 12.2.8. Let X and Y be Banach spaces and p ∈ (1, ∞). Let
i,α
πQ ∈ L (X, Y ), and let Λ be defined by the formal series in (12.24).
(1) If, for some x ∈ Xm the series (12.24) defining Λf converges (even just
conditionally) in Lp (Rd ; Y ) whenever f = 1R ⊗ x for R ∈ D, and these
satisfy the testing condition

k1R Λ(1R ⊗ x)kLp (Rd ;Y ) 6 TΛx |R|1/p ,

1,α
then kπQ xkY 6 TΛx |Q|1/2 and there is bx1 ∈ BMOD (Rd ; Y ) of norm
1,α
kb1 kBMOpD (Rd ;Y ) 6 TΛx such that πQ
x
x = hbx1 , hα
Q i.
1,α
(2) If, in addition to (1), we have X = Y and πQ ∈ K, then bx1 = b1 ⊗ x for
some b1 ∈ BMOD (Rd ) that is independent of x.
(3) If, for some y ∗ ∈ Y ∗ , the series (12.25) defining Λ∗ g converges (even just
0
conditionally) in Lp (Rd ; Y ∗ ) whenever g = 1R ⊗ y ∗ for R ∈ D, and these
satisfy the testing condition
∗ 0
k1R Λ∗ (1R ⊗ y ∗ )kLp0 (Rd ;X ∗ ) 6 TΛy∗ |R|1/p ,
∗ ∗
2,α ∗ ∗
then k(πQ ) y kX ∗ 6 TΛy∗ |Q|1/2 and there is by2 ∈ BMOD (Rd ; X ∗ ) with
∗ ∗ ∗
kby2 kBMOp0 (Rd ;X ∗ ) 6 TΛy∗ and (πQ
2,α ∗ ∗
) y = hby2 , hα
Q i.
D
∗
2,α
(4) If, in addition to (3), we have X = Y and πQ ∈ K, then by2 = b2 ⊗ y ∗
for some b2 ∈ BMOD (Rd ).
Proof. (1): Let us fix an x ∈ X. By assumption and Lemma 12.2.6, the series
X 1,α
x
BR := πQ x ⊗ hα
Q = (1R − ER )Λ(1R ⊗ x)
Q⊆R
α∈{0,1}d \{0}
 12.2 Paraproducts 131

converges (conditionally) in Lp (Rd ; Y ). Since ER Λ(1R ⊗ x) is constant on R,

we have the uniform estimate
x
inf kBR − ckLp (Rd ;Y ) 6 k1R Λ(1R ⊗ x)kLp (Rd ;Y ) 6 TΛx |R|1/p .
c∈Y

By Lemma 12.2.7, there is a function bx1 ∈ BMOD (Rd ; Y ) such that

kbx1 kBMOD (Rd ;Y ) 6 TΛ ,

x
p hbx1 , hα α
Q i = πQ x

for all Q ∈ D and α ∈ {0, 1}d \ {0},and kπQα

xkY 6 TΛ |Q|1/2 kxkX , from which
α
the claimed bound for kπQ kL (X,Y ) is immediate.
(2): Under the assumptions of this case, an inspection of the previous
argument shows that all auxiliary functions in the construction of bx1 have the
form φ ⊗ x for different scalar functions φ, and hence this form also remains
in the final result.
(3)–(4) follow by repeating the proof of (1)–(2) on the dual side. 

Remark 12.2.9. In the setting of Proposition 12.2.8, if we know a priori that

1,α
πQ α
x = hb1 (·)x, hQ 1
i for some b1 ∈ Lloc,so (Rd ; L (X, Y )), then our conclusion
on b1 implies that b1 ∈ BMOD,so (R ; L (X, Y )) and
x d

kb1 kBMOD,so
p
(Rd ;Y ) 6 TΛ .

According to Proposition 12.2.8, the following is a natural necessary con-

dition for the Lp boundedness of paraproducts, even when restricted to very
special functions only.

Definition 12.2.10. We say that a paraproduct Λ as in (12.24) satisfies the

weak coefficient bound if there is a finite constant C such that
i,α
kπQ kL (X,Y ) 6 C|Q|1/2 (12.30)

for all i = 1, 2, α ∈ {0, 1}d \ {0} and Q ∈ D.

While rather far from being a sufficient condition for any interesting bounded-
ness results, this weak coefficient bound nevertheless allows us to make sense
of the defining series of the paraproduct on a sufficiently rich class of functions
for our subsequent purposes.
We have the following useful convergence result for truncated paraproducts:

Lemma 12.2.11. Suppose that πQ α

∈ L (X, Y ) satisfy (12.30). Let p ∈ (1, ∞)
p d
and f ∈ L (R ; X) be boundedly supported, and consider the truncated para-
product X
α
m Πf := πQ hf iQ hα
Q.
Q∈D
`(Q)>2−m
α∈{0,1}d \{0}
132 12 Dyadic operators and the T (1) theorem

(1) For any m ∈ Z, the series defining m Πf converges absolutely in Lp (Rd ; Y ).

(2) For 2−m > diam(supp f ), we have

km Πf kLp (Rd ;Y ) 6 cd,p CkEm f kLp (Rd ;X) , (12.31)

0
and if g ∈ Lp (Rd ; Y ∗ ), then

|hm Πf, gi| 6 cd,p CkEm f kLp (Rd ;X) kEm gkLp0 (Rd ;Y ∗ ) −→ 0, (12.32)
m→−∞

2d − 1
where C is the constant in (12.30) and cd,p = .
1 − 2−d/p0
Proof. Let us first consider (2). When 2−m > diam(supp f ), the support
supp f is contained in at most 2d cubes Ri ∈ D. Then in m Πf , we only
need to consider Q ∈ D with Q ) Ri for some (not necessarily unique)
i = 1, . . . , 2d . Then
d
2
X X
α
km Πf kLp (Rd ;Y ) = kπQ hf iQ hα
Q kLp (Rd ;Y ) ,
i=1 Q∈D
Q)Ri
α∈{0,1}d \{0}

where
α
kπQ hf iQ hα α α
Q kLp (Rd ;Y ) 6 kπQ kL (X,Y ) khf iQ kX khQ kLp (Rd )

|Q|1/p
Z Z
α 1 C
= kπQ kL (X,Y ) f 6 f ,
|Q| Ri X |Q|1/2 |Q|1/p0 Ri X

and hence
d
2 Z
X X 1
km Πf kLp (Rd ;Y ) 6 (2d − 1)C f
i=1 Ri X
Q)Ri
|Q|1/p0
d
2 ∞
(2d − 1)C
Z
X X 0
= 1/p0
f 2−kd/p
i=1
|Ri | Ri X
k=1
d
2
2d − 1 X Z
= C |Ri |1/p − f ,
2d/p0−1 i=1 Ri X

where
d
2 Z 2d
X Z p 1/p
d/p0 0
X
1/p
|Ri | − f 62 |Ri | − f = 2d/p kEm f kLp (Rd ;X) .
Ri X Ri X
i=1 i=1

This proves both the convergence of the series and the claimed bound (12.31).
 12.2 Paraproducts 133

Each term in the series defining m Πf is constant on cubes R ∈ Dm . Hence

m Πf = Em (m Πf ), and thus

|hm Πf, gi| = |hm Πf, Em gi| 6 km Πf kLp (Rd ;Y ) kEm gkLp0 (Rd ;Y ∗ ) ,

so that (12.32) follows from (12.31).

Concerning (1), it only remains to consider the part of the series with
2−m < `(Q) 6 diam(supp f ). But there are only finitely many cubes Q of
fixed side-length that intersect supp f , and hence only finitely many cubes
altogether that contribute to this remaining sub-series. Thus the absolute
convergence is trivial. 
α
Corollary 12.2.12. Suppose that πQ ∈ L (X, Y ) satisfy (12.30). Then the
series X D E
α
πQ hf iQ , hhα
Q , gi
Q∈D
α∈{0,1}d \{0}

defining hΠf, gi converges absolutely for all f ∈ S(D; X) and g ∈ S(D; Y ∗ ),

and
hΠf, Em gi = hm Πf, gi.
Proof. Let v := g − Em g. Then v ∈ S00 (D; Y ∗ ), and hence only finitely many
of the terms hhαQ , vi are non-zero. Hence it is enough to prove the convergence
−m
with Em g in place of g. Since hhα Q , Em gi = 0 when `(Q) 6 2 , this coincides
p
with the series of hm Πf, Em gi. Since f ∈ S(D; X) ⊆ L (Rd ; X) is bound-
edly supported, the series defining m Πf converges absolutely in Lp (Rd ; Y ) by
0
Lemma 12.2.11. Since Em g ∈ S(D; Y ∗ ) ⊆ Lp (Rd ; Y ∗ ) ⊆ (Lp (Rd ; Y ))∗ , the
series of the bilinear form converges absolutely in K.
The last identity follows by observing that
(
α hhQα
, gi, `(Q) > 2−m ,
hhQ , Em gi =
0, `(Q) 6 2−m ,

and the proof is complete. 

i,α
Corollary 12.2.13. Suppose that πQ∈ L (X, Y ) satisfy (12.30). Then the
two series
X D E X D E
1,α 2,α α
πQ hf iQ , hhα
Q , gi + πQ hhQ , f i, hgiQ
Q∈D Q∈D
α∈{0,1}d \{0} α∈{0,1}d \{0}

defining hΛf, gi converge absolutely for all f ∈ S(D; X) and g ∈ S(D; Y ∗ ).

Proof. The convergence of the first series is the content of Corollary 12.2.12.
The convergence of the second series follows by permuting the roles of
2,α
f ∈ S(D; X) and g ∈ S(D; Y ∗ ), and transposing πQ to the dual side, since
2,α ∗
(πQ ) ∈ L (Y , X ) satisfies the same estimate.
∗ ∗

134 12 Dyadic operators and the T (1) theorem

12.2.b Sufficient conditions for boundedness

We will then turn to exploring conditions that ensure the boundedness of the
full paraproduct Π. The obtained necessary conditions serve as a model for
the type of sufficient conditions that we are looking for.
It is convenient to begin with a reduction to finite series. When Y is
reflexive, we have
0
Lp (Rd ; Y ) = Lp (Rd ; Y ∗∗ ) ' (Lp (Rd ; Y ∗ ))∗ .
0
Since S00 (D; Y ∗ ) is dense in Lp (Rd ; Y ∗ ), it is enough to show that the action
0
of Πf is bounded on S00 (D; Y ∗ ) with respect to the norm of Lp (Rd ; Y ∗ ),
uniformly for f in the unit ball of Lp (Rd ; X). Since any fixed g ∈ S00 (D; Y ∗ )
PΠfα, it is enough
only “sees” a finite part of to prove a uniform Lp (Rd ; Y ) esti-
α
mate for the finite sums πQ hf iQ hQ . A key initial estimate in this direction
is the following:
Lemma 12.2.14. Let X be a Banach space, Y be a UMD space, and p ∈
(1, ∞). Let F be a finite collection of dyadic cubes. For all f ∈ Lp (Rd ; X)
and πQα
∈ L (X, Y ), we then have
X  X 
α − +
πQ hf iQ hα
Q p d
6 β p,Y β p,Y ε Q π α 0
Q h Q f p d .
L (R ;Y ) L (R ;Y )
Q∈F Q∈F

α
Proof. Since (πQ hf iQ hα p d
Q )Q∈F is a martingale difference sequence in L (R ; Y ),
we have
X X
α α − α
πQ hf iQ hQ 6 βp,Y ε Q πQ hf iQ hα
Q .
Lp (Rd ;Y ) Lp (Rd ×Ω;Y )
Q∈F Q∈F

Rewriting the Lp norm on the product Rd × Ω with the help of Fubini’s the-
orem, we observe that at each fixed t ∈ Rd , the sequence of random variables
α
ε Q πQ hf iQ hα
Q (t)

has the same joint distribution as

α 0 α 0
ε Q πQ hf iQ hQ (t) = εQ EQ (πQ f hQ )(t),

since the possibly different sign of hα 0

Q (t) and hQ (t) is invisible after multipli-
cation by εQ . Using this and Stein’s inequality (Theorem 4.2.23), we conclude
that
X X
α
ε Q πQ hf iQ hα
Q p d
= α
εQ EQ (πQ f h0Q ) p d
L (R ×Ω;Y ) L (R ×Ω;Y )
Q∈F Q∈F
X  X 
+ α 0 +
6 βp,Y ε Q πQ f hQ = βp,Y α
εQ h0Q πQ f .
Lp (Rd ×Ω;Y ) Lp (Rd ×Ω;Y )
Q∈F Q∈F


 12.2 Paraproducts 135

The previous lemma motivates the following. A background for the nomen-
clature will be discussed in the Notes.

Definition 12.2.15. Let p ∈ (1, ∞). For an indexed family (πQ )Q∈D in a
Banach space Z, we define the Carleson norm
1 X
k(πQ )kCarp (Rd ;Z) := sup sup εQ h0Q πQ .
Q0 ∈D F ⊆D |Q0 |1/p Q⊆Q0
Lp (Q0 ×Ω;Z)
finite
Q∈F

With the help of Theorem 3.2.17 (the John–Nirenberg inequality for adapted
sequences), one can check that any these Carleson norms are actually equiv-
alent for different values of p. We will not need this observation, since the
following proof directly shows that we can use any of these norms in our upper
bound, as we like. Our first sufficient condition for paraproduct boundedness
is stated in terms of this notion as follows:

Proposition 12.2.16 (Paraproducts vs. Carleson norms). Let X be a

Banach space, Y be a UMD space, and p, q ∈ (1, ∞). Let Π be the paraproduct
α
defined by an indexed family (πQ )Q∈D,α∈{0,1}d \{0} . In order that Π is bounded
p d p d
from L (R ; X) to L (R ; Y ), it is sufficient that (πQ α
)Q∈D satisfies the Carp
condition for every α ∈ {0, 1}d \ {0}. Moreover, we have the bound
X
−
kΠkL (Lp (Rd ;X),Lp (Rd ;Y )) 6 32 · 4d pp0 βq,Y +
βq,Y α
k(πQ )kCarq (Rd ;L (X,Y )) .
α∈{0,1}d \{0}

Proof. We are going to estimate

XD E
hΠ α f, gi = α
πQ hf iQ , hhα
Q , gi (12.33)
Q∈D

for f ∈ S00 (Rd ; X) and g ∈ S00 (Rd ; Y ∗ ); the latter guarantees that the sum
α α
is finitely nonzero. We may thus replace πQ by 1F (Q)πQ for some finite set
F ⊆ D, but we do not indicate this explicitly in the notation.

Part I: Construction of principal cubes

Let P0 be the maximal cubes appearing in this sum. We then construct cube
families Pn inductively as follows. For each P ∈ Pn , let chP (P ) be the
maximal dyadic subcubes P 0 of F such that either
Z Z Z Z
− kf kX > 4− kf kX or − kgkY ∗ > 4− kgkY ∗ .
P0 P P0 P

For each such P 0 , we have

1 n R kf k R kgk ∗ o
0 X 0 Y
|P 0 | 6 max RP , RP .
4 − kf kX − kgkY ∗
P P
136 12 Dyadic operators and the T (1) theorem

Since these P 0 are pairwise disjoint, we have

X Z Z Z
kf kX 6 kf kX = |P |− kf kX
P 0 ∈chP (P ) P0 P P

and similarly with g, and hence

X 1 1
6 (|P | + |P |) = |P |.
4 2
P 0 ∈chP (P )

Thus [ 1
EP (P ) := P \ satisfies |EP (P )| > |P |.
2
P 0 ∈chP (P )

Then we let
[ ∞
[
Pn+1 := chP (P ), P := Pn ,
P ∈Pn n=0

and the sets EP (P ), P ∈ P, are seen to be pairwise disjoint.

Now every Q with a nonzero contribution to (12.33) will be contained in
some P ∈ P0 ⊆ P. Let parP P ∈ P be the minimal such P . By construction,
it follows that
Z Z Z Z
− kf kX 6 4− kf kX , − kgkY ∗ 6 4− kgkY ∗ , if parP Q = P.
Q P Q P

For P ∈ P, let
X
PP h := 1P 0 hhiP 0 + 1EP (P ) h.
P 0 ∈chP (P )

Let h ∈ {f, g}. If u ∈ EP be a Lebesgue point of h, then all Q with u ∈ Q ⊆ P

fail the stopping criterion, and hence

kPP h(u)k = kh(u)k = lim khhiQ k 6 4hkh(·)kiP .

Q→u

On the other hand, if u ∈ P 0 ∈ chP (P ), then its dyadic parent P

c0 fails the
stopping criterion, and hence

kPP h(u)k = khhiP 0 k 6 hkh(·)kiP 0 6 2d hkh(·)kiPc0 6 2d · 4hkh(·)kiP .

Hence we conclude that

kPP h(u)k 6 4 · 2d · 1P (u)hkhkiP , h ∈ {f, g}.

If parP Q = P and Q0 ∈ chD Q, then each P 0 ∈ chP P is either disjoint

from Q (thus a fortiori from Q0 ), or strictly contained in Q, hence contained
in Q0 . Thus
 12.2 Paraproducts 137
Z X Z Z
PP h = |P 0 |hhiP 0 + h= h, parP Q = P.
Q0 0 Q0 ∩EP (P ) Q0
P ∈chP (P )
P 0 (Q0

0
Since both 1Q and hα
Q are linear combination of Q ∈ chD Q, this implies in
particular that

hf iQ = hPP f iQ , hhα α
Q , gi = hhQ , PP gi, parP Q = P.

Part II: Estimates under the principal cubes

With the principal cubes P ∈ P just constructed, we can now rearrange the
sum (12.33) as
X D X E X
hΠ α f, gi = πQα
hPP f iQ hα
Q , PP g =: IP .
P ∈P Q∈D P ∈P
parP Q=P

By Lemma 12.2.14 at the key step introducing the UMD constants, and ap-
plications of Hölder’s inequality and the properties of the principal cubes
elsewhere,
X
α α
IP 6 πQ hPP f iQ hQ kPP gkLq0 (Rd ;Y ∗ )
Lq (Rd ;Y )
Q∈D
parP Q=P
 X 
− + α 0
6 βq,Y βq,Y ε Q πQ h Q PP f kPP gkLq0 (Rd ;Y ∗ )
Lq (Rd ×Ω;Y )
Q∈D
parP Q=P
X
− + α 0
6 βq,Y βq,Y ε Q πQ hQ kPP f kL∞ (Rd ;X) ×
Lq (Rd ×Ω;L (X,Y ))
Q∈D
parP Q=P
0
× kPP gkL∞ (Rd ;Y ∗ ) |P |1/q
− + α
6 βq,Y βq,Y k(πQ )kCarq (Rd ;L (X,Y )) |P |1/q × 4 · 2d hkf kX iP ×
0
× 4 · 2d hkgkY ∗ iP |P |1/q
− +
= 16 · 4d · βq,Y βq,Y α
k(πQ )kCar2 (Rd ;L (X,Y )) hkf kX iP hkgkY ∗ iP |P |
− +
=: 16 · 4d · βq,Y βq,Y α
k(πQ )kCar2 (Rd ;L (X,Y )) × IIP .

(Note that, in the step that lead to the appearance of the Carleson norm,
α α
we made use of our implicit replacement of πQ by 1F (Q)πQ , for some finite
F ⊆ D, in the beginning of the proof.)
Finally,
138 12 Dyadic operators and the T (1) theorem
X X
IIP 6 2 hkf kX iP hkgkY ∗ iP |EP (P )|
P ∈P P ∈P
X Z
62 MD f · MD g
P ∈P EP (P )
Z
62 M D f · MD g
Rd
6 2kMD f kLp (Rd ) kMD gkLp0 (Rd )
6 2 · p0 kf kLp (Rd ;X) · pkgkLp0 (Rd ;Y ∗ ) ,

by Doob’s maximal inequality in the last step.

A combination of the estimates proves the proposition. 
To compare the necessary and sufficient conditions for paraproduct bound-
edness, we have the following relation between bounded mean oscillation and
Carleson norms.
Proposition 12.2.17 (Carleson norms vs. BMO). Let Z be a UMD
space, and p ∈ (1, ∞). If b ∈ BMOD (Rd ; Z), then (πQα
)Q∈D := (hb, hα
Q i)Q∈D
p d
satisfies the Car condition for each α ∈ {0, 1} \ {0}, and
α +
max k(πQ )kCarp (Rd ;Z) 6 βp,Z kbkBMOpD (Rd ;Z) .
α∈{0,1}d \{0}

This estimate also has a converse, but since it has no immediate use in the
present discussion, we leave the details to an interested reader.
Proof. This is a direct computation
X
0 α
ε Q hQ hb, hQ i p
L (Q0 ×Ω;Z)
Q⊆Q0
Q∈F
X
α α
6 inf ε Q hQ h1Q0 (b − c), hQ i
c∈Z Lp (Q0 ×Ω;Z)
Q⊆Q0
Q∈F
+
6 inf βp,Z k1Q0 (b − c)kLp (Rd ;Z) by Proposition 12.1.5
c∈Z
+
6 βp,Z |Q0 |1/p kbkBMOpD (Rd ;Z) .

Taking the supremum over finite F ⊆ D and Q0 ∈ D, the claimed bound

follows from the definition of Carp . 
We can now formulate conditions for the boundedness of a paraproduct Πb
in terms of function space properties of b:
Theorem 12.2.18. Let X be a Banach space, Y be a UMD space, and p ∈
(1, ∞). Let b ∈ L1loc,so (Rd ; L (X, Y )), and let Πb be the paraproduct defined
α
by the operators πQ : x 7→ hb(·)x, hα Q i. In order that Πb is bounded from
p d p d
L (R ; X) to L (R ; Y ), it is
 12.2 Paraproducts 139

(1) necessary that b ∈ BMOD,so (Rd ; L (X, Y )), and

(2) sufficient that b ∈ BMOD (Rd ; Z) for some subspace Z ,→ L (X, Y ) with
the UMD property.
Moreover, we have the quantitative bounds
kbkBMOD,so (Rd ;L (X,Y )) 6 kΠkL (Lp (Rd ;X),Lp (Rd ;Y ))
−
6 32 · 8d pp0 βq,Y +
βq,Y +
kjkL (Z,L (X,Y )) βq,Z kbkBMOD
q
(Rd ;Z) ,

where j : Z → L (X, Y ) is the embedding map and q ∈ (1, ∞) is arbitrary.

Proof. The necessary condition and the lower bound for kΠk are just restate-
ments of Proposition 12.2.8 and Remark 12.2.9.
For the sufficient condition, from Proposition 12.2.16 we obtain
X
−
kΠkL (Lp (Rd ;X),Lp (Rd ;Y )) 6 32 · 4d pp0 βq,Y +
βq,Y α
k(πQ )kCarq (Rd ;L (X,Y )) ,
α∈{0,1}d \{0}

and the assumed embedding followed by Proposition 12.2.17 give us

α α
k(πQ )kCarq (Rd ;L (X,Y )) 6 kjkL (Z,L (X,Y )) k(πQ )kCarq (Rd ;Z)
+
6 kjkL (Z,L (X,Y )) βq,Z kbkBMOqD (Rd ;Z) .

The estimate is concluded by noting that #({0, 1}d \ {0}) = 2d − 1 < 2d . 

For paraproducts defined by scalar-valued coefficients, we now obtain a com-
plete characterisation of their boundedness on UMD spaces. For p = q, the
equivalence (1) ⇔ (4) provides a partial solution of the Lp extension prob-
lem, discussed in Section 2.1, in the particular case of the paraproducts. Note,
however, it does not exclude the possibility of Lp (Rd )-bounded paraproducts
extending boundedly to other classes of spaces besides UMD.
Corollary 12.2.19. Let X be a UMD space, and p, q ∈ (1, ∞). Let Π1 , Π2∗
1,α 2,α
and Λ := Π1 + Π2∗ be paraproducts with scalar coefficients πQ , πQ ∈ K.
Then the following are equivalent:
(1) Λ ∈ L (Lp (Rd ; X));
(2) both Π1 , Π2∗ ∈ L (Lp (Rd ; X));
(3) for some bi ∈ BMO(Rd ), we have
1,α 2,α
πQ = hb1 , hα
Q i, πQ α
= hb2 , hQ i, ∀Q ∈ D, α ∈ {0, 1}d \ {0};
(4) Λ ∈ L (Lq (Rd )).
Under these equivalent conditions, we have the estimates
max kbi kBMOpi (Rd ) 6 kΛkL (Lp (Rd ;X)) ,
i=1,2 D

kΠ̃i kL (Lp (Rd ;X)) 6 32 · 8d · pp0 · βq,X

2
· βq,K · kbi kBMOqi (Rd ) ,
D

kΛkL (Lp (Rd ;X)) 6 64 · 8d · pp0 · βq,X

2
· βq,K · kΛkL (Lq (Rd )) .

where Π̃1 := Π1 , Π̃2 := Π2∗ , p1 := p, p2 := p0 , q1 := q, q2 := q 0 .

140 12 Dyadic operators and the T (1) theorem

Proof. (1) ⇒ (3): The assumed boundedness (1) and duality clearly implies
the testing conditions
kΛ(1Q ⊗ x)kLp (Rd ,X) 6 kΛkL (Lp (Rd ;X)) k1Q ⊗ xkLp (Rd ;X) ,
kΛ (1Q ⊗ x∗ )kLp0 (Rd ,X ∗ ) 6 kΛkL (Lp (Rd ;X)) k1Q ⊗ x∗ kLp0 (Rd ;X ∗ ) .
∗

Condition (3) then follows from Proposition 12.2.8, which also provides the
bounds
 
max kb1 kBMOD p
(Rd ) , kb2 kBMOp0 (Rd ) 6 kΛkL (Lp (Rd ;X)) .
D

(3) ⇒ (2): We use Theorem 12.2.18 with Y = X and Z = K · IX , which

clearly embeds into L (X) with constant one. With this choice, the theorem
shows that
−
kΠ1 kL (Lp (Rd ;X)) 6 32 · 8d · pp0 βq,X +
βq,X +
βq,K kb1 kBMOqD (Rd ) .
6 32 · 8d · pp0 βq,X
2
βq,K kb1 kBMOD
p
(Rd ) ,

±
where we also used βp,X 6 βp,X . Similarly, we have

kΠ2∗ kL (Lp (Rd ;X)) = kΠ2 kL (Lp0 (Rd ;X ∗ )) 6 32 · 8d · pp0 βq,X

2
βq,K kb2 kBMOq0 (Rd ) ,
D

using the same bound on the dual side and recalling that βq0 ,X ∗ = βq,X .
(2) ⇒ (1): This is trivial by the triangle inequality.
(3) ⇔ (4): This is the already established equivalence (3) ⇔ (1) specialised
to X = K. The final quantitative bound follows by combining the bounds
already established:
2
X
kΛkL (Lp (Rd ;X)) 6 kΠ̃i kL (Lp (Rd ;X))
i=1
2
X
6 32 · 8d · pp0 βq,X
2
βq,K kbi kBMOqi (Rd ) ,
D
i=1
2
X
6 32 · 8d · pp0 βq,X
2
βq,K kΛkL (Lq (Rd ))
i=1
P2
and i=1 32 = 64. 

12.2.c Symmetric paraproducts

In this section, we will take a closer look at the special case of the symmetric
i,α α
paraproduct Λb with equal coefficients πQ = hb, hQ i for both i = 1, 2. Our
goal is to obtain a qualitative improvement of the earlier Proposition 12.2.3.
This will require developing modest prerequisites about the projective tensor
product of Banach spaces, and we first turn to this task.
 12.2 Paraproducts 141

Definition 12.2.20. For two Banach spaces X and Z, and a bilinear form
λ : X × Z → K, we define
n o
kλkB(X,Z) := sup |λ(x, z)| : kxkX 6 1, kzkZ 6 1 ,
n o
B(X, Z) := λ : X × Z → K bilinear kλkB(X,Z) < ∞ .

Lemma 12.2.21. B(X, Z) ' L (X, Z ∗ ) ' L (Z, X ∗ ).

Proof. For u ∈ L (X, Z ∗ ), we see that
Form(u) : X × Z → K, (x, z) 7→ hux, zi
defines Form(u) ∈ B(X, Z) of norm at most kukL (X,Z ∗ ) . For λ ∈ B(X, Z),
we see that Op(λ) : X → Z ∗ : x 7→ λ(x, ·) defines Op(λ) ∈ L (X, Z ∗ ) of norm
at most kλkB(X,Z) . Both Form : L (X, Z ∗ ) → B(X, Z) and Op : B(X, Z) →
L (X, Z ∗ ) are clearly linear and we just saw that they are contractive. Since
both Form ◦ Op and Op ◦ Form are identities of the respective spaces, they
must in fact be isometries. This proves the first identification, and B(X, Z) '
L (Z, X ∗ ) follows by symmetry, since clearly B(X, Z) ' B(Z, X). 
Definition 12.2.22. For two Banach spaces X and Z, and elements x ∈ X
and z ∈ Z, we define x ⊗ x ∈ B(X, Z)∗ by
x ⊗ z : B(X, Z) → K : λ 7→ λ(x, z).
Let further
X ⊗ Z := span{x ⊗ z : x ∈ X, z ∈ Z} ⊆ B(X, Z)∗ ,
and, for all v ∈ X ⊗ Z,
n
nX n
X o
kvkX ⊗Z
b := inf kxi kX kzi kZ : v = x i ⊗ zi ,
i=1 i=1

where the infimum is over all possible representations of v of any length n.

b be the completion of X ⊗ Z with respect to this norm.
Finally, let X ⊗Z
Proposition 12.2.23. For all Banach spaces X and Z, we have
b ∗ = B(X, Z),
(X ⊗Z)
in the following sense: For all v ∈ X ⊗ Z and λ ∈ B(X, Z), the pairing
n
X n
X
hv, λi := λ(xi , zi ), if v= x i ⊗ zi ,
i=1 i=1

is well defined and extends by continuity to all v ∈ X ⊗Z.

b Conversely, every
element of (X ⊗ Z)∗ has this form, and
kλk(X⊗Z)∗ = kλkB(X,Z) .
142 12 Dyadic operators and the T (1) theorem

Proof. To check thatP hv, λi is well-defined, we need to verify that two different
na
representations v = i=1 xai ⊗ zia , a = 1, 2, result in the same right-hand side.
0 p
To see this, pick a basis (xj )j=1 for span{xia : 1 6 i 6 na , a = 1, 2} and a basis
(zk0 )qk=1 for span{zia : 1 6 i 6 na , a = 1, 2} and expand all xia and zia in the
respective basis. With the help of the Hahn–Banach theorem, pick x∗m ∈ X ∗
and zn∗ ∈ Z ∗ such that hx0j , x∗m i = δj,m and hzk0 , zn∗ , =iδk,n , and consider the
forms λm,n ( ·1 , ·2 ) = h ·1 , x∗m ih ·2 , zn∗ i ∈ B(X, Z) to see that x0j ⊗zk0 are linearly
independent in B(X, Z)∗ . Hence their coefficients must be equal in the two
expansions of v. Make the same expansions on the right-hand side, using the
bilinearity of λ, to find that both expansions lead to linear combinations with
equal coefficients of the values λ(xj0 , zk0 ).
Having verified that the action of λ on X ⊗ Z is well defined, its linearity
is clear. Moreover,
n
X n
X
|λ(xi , zi )| 6 kλkB(X,Z) kxi kX kzi kZ ,
i=1 i=1

and taking the infimum over all representations of v shows that

|hv, λi| 6 kλkB(X,Z) kvkX ⊗Z
b

for all v ∈ X ⊗ Z. From this estimate, we can uniquely extend the action of
λ to all v ∈ X ⊗Z
b by density, with the estimate

kλk(X⊗Z)∗ 6 kλkB(X,Z) .
On the other hand, we also have
|λ(x, z)| = |hx ⊗ z, λi| 6 kx ⊗ zkX ⊗Z
b kλk(X⊗Z)∗ 6 kxkX kzkZ kλk(X⊗Z)∗ ;

thus kλkB(X,Z) 6 kλk(X⊗Z)∗ , and hence in fact there is equality.

Conversely, if ξ ∈ (X ⊗ Z)∗ , we can define λ ∈ B(X, Z) by λ(x, z) :=
hx ⊗ z, ξi. From the previous construction, it is then clear that hv, λi = hv, ξi
for all v ∈ (X ⊗ Z), and hence every ξ ∈ (X ⊗ Z)∗ arises from the previous
construction. 
Corollary 12.2.24. kx ⊗ zkX ⊗Z
b = kxkX kzkZ .

Proof. We compute the norm by duality:

n o
b = sup |hx ⊗ z, ξi| : kξk(X ⊗Z)
kx ⊗ zkX ⊗Z b ∗ 61
n o
= sup |λ(x, z)| : kλkB(X,Z) 6 1 .

It is clear from the definition that |λ(x, z)| 6 kxkX kzkZ for any λ as in the
last supremum. On the other hand, the Hahn–Banach theorem guarantees the
existence of x∗ ∈ X ∗ and z ∗ ∈ Z ∗ of norm one such that hx, x∗ i = kxkX and
hz, z ∗ i = kzkZ . Then clearly λ( ·1 , ·2 ) = h ·1 , x∗ ih ·2 , z ∗ i has kλkB(X,Z) 6 1
and gives λ(x, z) = kxkX kzkZ . 
 12.2 Paraproducts 143

We are now ready to prove the following improvement of Proposition 12.2.3:

Theorem 12.2.25. Let X and Y be UMD spaces and p ∈ (1, ∞). For every
function b ∈ BMOD (Rd ; L (X, Y )), the symmetric paraproduct Λb defines a
bounded operator from Lp (Rd ; X) to Lp (Rd ; Y ) of norm

kΛb kL (Lp (Rd ;X),Lp (Rd ;Y )) 6 6 · 2d · (pp0 + βp,X

+
βp+0 ,Y ∗ )kbkBMOD (Rd ;L (X,Y ))
6 30 · 2d · βp,X βp,Y kbkBMOD (Rd ;L (X,Y ))

Proof. By density, it suffices to consider the action of Λb on f ∈ S00 (D; X),

paired with g ∈ S00 (D; Y ∗ ). We will rewrite this pairing with the help
of the projective tensor product duality between X ⊗Y b ∗ and B(X, Y ∗ ) '
L (X, Y ) = L (X, Y ), recalling that UMD spaces are reflexive (Theorem
∗∗

4.3.3). In the following computation, the summation is always over Q ∈ D

and α ∈ {0, 1}d \ {0}.
X nD E D E o
α α
hΛb f, gi = hb, hQ ihf iQ , hhQ , gi + hb, hα
Q ihf, h α
Q i, hgi Q
X,Y ∗ X,Y ∗
XD E
= hb, hα α α
Q i, hf iQ ⊗ hhQ , gi + hf, hQ i ⊗ hgiQ
L (X,Y ),X ⊗bπY ∗
D X h iE
= b, hα α α
Q hf iQ ⊗ hhQ , gi + hf, hQ i ⊗ hgiQ =: hb, hi.

On the last line, we are using the H 1 –BMO-duality from Theorem 11.1.30;
for f ∈ S00 (D; X) and g ∈ S00 (D; Y ∗ ), the summation is finite, and thus h ∈
L∞c (R ; X ⊗π Y ). Since b ∈ BMOD (R ; L (X, Y )) ⊆ Lloc (R ; L (X, Y )), the
d b ∗ d 1 d

pointwise duality product hb(u), h(u)i is integrable, and one find by dominated
convergence in the defining formula of Theorem 11.1.30 that the duality can be
computed simply as the integral of hb(u), h(u)i over Rd . Thus, an application
of Theorem 11.1.30 followed by Theorem 11.1.28 shows that

|hΛb f, gi| 6 kbkBMO(Rd ;L (X,Y )) khkHat

1 (Rd ;X⊗Y ∗ )

6 kbkBMO(Rd ;L (X,Y )) · 6 · 2d · khkHmax

1 (Rd ;X⊗Y ∗ ) ,

and it remains to estimate the H 1 norm here. Recall that

khkHmax
1 (Rd ;X⊗Y ∗ ) = kMD hkL1 (Rd ) = sup 1R khhiR kX⊗Y ∗ .
R∈D L1 (Rd )

By the properties of Haar functions, we find that

X X D h iE
α
hhiR = hQ hf iQ ⊗ hhα
Q , gi + hf, h α
Q i ⊗ hgi Q
R
Q)R α∈{0,1}d \{0}
Xh i
= hf iQ ⊗ (hgiQR − hgiQ ) + (hf iQR − hf iQ ) ⊗ hgiQ ,
Q)R

where QR is the unique dyadic child of Q that contains R.

144 12 Dyadic operators and the T (1) theorem

Next, we make the following algebraic observation:

hf iQR ⊗ hgiQR − hf iQ ⊗ hgiQ

= (hf iQR − hf iQ + hf iQ ) ⊗ (hgiQR − hgiQ + hgiQ ) − hf iQ ⊗ hgiQ
= hf iQ ⊗ (hgiQR − hgiQ ) + (hf iQR − hf iQ ) ⊗ hgiQ
+ (hf iQR − hf iQ ) ⊗ (hgiQR − hgiQ ).

Thus
Xh i
hf iR = hf iQR ⊗ hgiQR − hf iQ ⊗ hgiQ
Q)R
X
+ (hf iQR − hf iQ ) ⊗ (hgiQR − hgiQ ) =: IR + IIR .
Q)R

The sum IR is telescopic and, since f ∈ S00 (D; X) (we don’t even need the
similar property of g at this point), its terms vanish for all large enough Q.
Thus in fact

IR = hf iR ⊗ hgiR , kIR kX ⊗
b π Y ∗ = khf iR kX khgiR kY ∗

and

sup 1R kIR kX ⊗
bπY ∗ 6 kMD f · MD gkL1 (Rd )
R∈D L1 (Rd )

6 kMD f kLp (Rd ) kMD gkLp0 (Rd )

6 p0 kf kLp (Rd ;X) · pkgkLp0 (Rd ;Y ∗ )

by Doob’s maximal inequality in the last step.

Turning to IIR , we note that hf iQR −hf iQ is the constant value of DQ f (u)
for any u ∈ R, and similarly for g. As before, the summation in IIR is finitely
non-zero, and we can disentangle it with the help of a Rademacher sequence
(εQ )Q∈D as
X  X 
IIR = E εP DP f (u) ⊗ εQ DQ g(u) .
P )R Q)R

Thus
X X
kIIR kX ⊗Y
b ∗ 6E εP DP f (u) εQ DQ g(u)
X Y∗
P )R Q)R
X X
6 εP DP f (u) εQ DQ g(u)
Lp (Ω;X) Lp0 (Ω;Y ∗ )
P )R Q)R
X X
6 εP DP f (u) εQ DQ g(u) ,
Lp (Ω;X) Lp0 (Ω;Y ∗ )
P ∈D Q∈D

where the last step was an application of the contraction principle. Thus
 12.2 Paraproducts 145
X X
sup kIIR kX ⊗Y
b ∗ 6 εP DP f (u) εQ DQ g(u)
R3u Lp (Ω;X) Lp0 (Ω;Y ∗ )
P ∈D Q∈D

and

sup 1R kIIR kX ⊗Y
b ∗
R∈D L1 (Rd )
X X
6 ε P DP f ε Q DQ g
Lp (Rd ×Ω;X) Lp0 (Rd ×Ω;Y ∗ )
P ∈D Q∈D
+
6 βp,X kf kLp (Rd ;X) · βp+0 ,Y ∗ kgkLp0 (Rd ;Y ∗ ) .

A combination of the estimates of IR and IIR shows that

khkHmax
1 b ∗)
(Rd ;X ⊗Y

6 sup 1R kIR kX ⊗Y
b ∗ + sup 1R kIR kX ⊗Y
b ∗
R∈D L1 (Rd ) R∈D L1 (Rd )
0 +
6 (pp + βp,X βp+0 ,Y ∗ )kf kLp (Rd ;X) kgkLp0 (Rd ;Y ∗ ) ,

and altogether we have proved the first estimate claimed in the theorem.
+
The final estimate is seen as follows: First, we have βp,X 6 βp,X and
+
βp0 ,Y ∗ 6 βp0 ,Y ∗ = βp,Y by the observation after Proposition 4.2.3, and Propo-
sition 4.2.17(2). Second, denoting p∗ = max(p, p0 ) > 2, we have βp,Z > βp,R =
p∗ − 1 > 21 p∗ by Theorem 4.5.7, and hence pp0 6 (p∗ )2 6 4βp,X βp,Y . 

12.2.d Mei’s counterexample: no simple sufficient conditions

The following theorem shows the impossibility of obtaining simple upper

bounds for operator-valued paraproducts in infinite-dimensional spaces, even
by considering Hilbert spaces only, and even by replacing the bounded mean
oscillation conditions by the stronger L∞ norm.

Theorem 12.2.26 (Mei). Let φ be a function such that

kΠb kL (L2 (R;`2N )) 6 φ(N )kbkL∞ (R;L (`2N )) for all b ∈ L∞ (R; L (`2N )).

Then
1
φ(N ) > k4N kL (L (`2N )) > (log N − 1),
π
where 4N : L (`2N ) → L (`2N ) is the lower triangle projection defined by
(
ei ⊗ ej , if i > j,
4N (ei ⊗ ej ) :=
0, else

and extended by linearity.

146 12 Dyadic operators and the T (1) theorem

Proof. For a ∈ L (`N 2

) and u, v ∈ `2N , we have the tensor product u ⊗ v ∈
L (`N ), and the trace duality ha, u ⊗ vi = hau, vi.
2

Let b ∈ L∞ (R; L (`N 2

)), and f, g ∈ L2 (R; `2N ). We can then write
DX E D X E
hΠb f, gi = hb, hI ihf iI hI , g = b, hf iI ⊗ hhI , gihI =: hb, Π⊗g f i,
I∈D I∈D

where suggestive notation Π⊗g is defined by the last identity. In the two right-
most expressions, the duality is that between L∞ (R; L (`2N )) and its predual
L1 (R; C 1 (`N
2
)). (We recall from Theorem D.2.6 that (C 1 (H))∗ = L (H) for
any Hilbert space H and from Theorem 1.3.10 that (L1 (R; X))∗ = L∞ (R; X ∗ )
when X ∗ has the Radon–Nikodým property, which the finite-dimensional
(hence reflexive) X = L (`2N ) does by Theorem 1.3.21.)
Thus we deduce that
n o
kΠ⊗g f kL1 (R;C 1 (`N
2 )) = sup |hb, Π⊗g f i| : kbkL∞ (R;L (`2N )) 6 1
n o
= sup |hΠb f, gi| : kbkL∞ (R;L (`2N )) 6 1
2 ) kgkL2 (R;`2 ) .
6 φ(N )kf kL2 (R;`N N

We now apply this to a special choice of f, g ∈ L2 (R; `N 2

). Let (ri )N i=1
be the standard realisation of a Rademacher sequence on [0, 1), i.e., ri (t) :=
N
1[0,1) (t) sgn(sin(2i πt)). With u, v ∈ `N2
P
, we take f = i=1 ri hu, ei iei and
PN
g = i=1 ri hv, ei iei , where (ei )i=1 is the standard orthonormal basis of `2N .
N

Then
j−1
N X
X
Π⊗g f (t) = ri (t)hu, ei iei ⊗ rj (t)hv, ej iej
j=1 i=1
 X 
= Dr(t) hu, ei ihv, ej iei ⊗ ej Dr(t)
16i<j6N
 N
X 

= Dr(t) TN hu, ei ihv, ej iei ⊗ ej Dr(t) = Dr(t) TN (u ⊗ v) Dr(t)
i,j=1
PN
where Dr(t) = i=1 ri (t)ei ⊗ ei and 4̃N is the upper triangle projection
defined by (
ei ⊗ ej , if i < j,
4̃N (ei ⊗ ej ) :=
0, else
and extended by linearity. Since Dr(t) is unitary for every t ∈ [0, 1), it follows
that
2 )) = k4̃N (u ⊗ v)kC 1 (`2 ) .
kΠ⊗g f kL1 (R;C 1 (`2N )) = k4̃N (u ⊗ v)kL1 ([0,1);C 1 (`N N

Dy Lemma D.1.1 and the definition of the Schatten class, every s ∈ C 1 (`N
2
)
has a singular value decomposition
 12.3 The T (1) theorem for abstract bilinear forms 147
n
X n
X
s= ak (s)uk ⊗ vk , 2 = kvk k`2 = 1,
kuk k`N N
ak (s) = kskC 1 (`2N )
k=1 k=1

where ak (s) > 0 are the approximation numbers of s. Letting fk , gk ∈

L2 (R; `2N ) of norm one be the functions corresponding to uk , vk , we find that
n
X
k4̃N skC 1 (`2N ) 6 ak (s)k4̃N (uk ⊗ vk )kC 1 (`2N )
k=1
Xn
= ak (s)kΠ⊗gk fk kL1 (R;C 1 (`N
2 ))

k=1
Xn
6 ak (s)φ(N ) = φ(N )kskC 1 (`2N ) .
k=1

Noting that the lower triangle projection 4N on L (`N 2

) = (C 1 (`N
2
))∗ is the
adjoint of the upper triangle projection 4̃N on C (`N )), this implies that
1 2

k4N kL (L (`2N )) = k4̃N kL (C 1 (`N

2 )) 6 φ(N ),

which is the first claimed inequality.

The final bound is essentially Lemma 7.5.12, where a variant
(
ei ⊗ ej , if i > j,
TN (ei ⊗ ej ) :=
0, else

was considered instead. However, the lower bound for the norm of this operator
was achieved by testing with the Hilbert matrix AN = (1{i6=j} (i − j)−1 )Ni,j=1
with vanishing diagonal; hence ∆N (AN ) = TN (AN ), and the same lower
bound follows for ∆N as well. 

12.3 The T (1) theorem for abstract bilinear forms

In Sections 11.2 and 11.3, the leading theme was extrapolating the bound-
edness of a singular integral operator from Lp0 (Rd ; X) to Lp (Rd ; X), with
a different exponent p, or even to Lp (w; X), with a different weight w. A
question that was largely left open in these sections was how to verify the
assumed boundedness on some Lp0 (Rd ; X) to begin with. In the spirit of the
Lp -extension problem discussed in Section 2.1, we here obtain the following
useful answer that allows us to extrapolate the vast existing information about
scalar-valued singular integrals to the UMD-valued situation:
Theorem 12.3.1. Let p0 ∈ (1, ∞), and let T ∈ L (Lp0 (Rd )) be an operator
associated with a Calderón–Zygmund standard kernel K : Ṙ2d → K. Let X
be a UMD space and p ∈ (1, ∞). Then T ⊗ IX extends to a bounded linear
operator on Lp (Rd ; X).
148 12 Dyadic operators and the T (1) theorem

In fact, this result will be obtained as a corollary of general criteria, known as

“T (1) theorems”, for the boundedness of operators associated with Calderón–
Zygmund kernels; and we will also obtain versions dealing with operator-
valued kernels. However, the very statement of these results requires some
preparations that we take up next. Concerning the proofs, we only mention
at this point that the dyadic singular integral operators and paraproducts,
whose boundedness we already studied in Sections 12.1 and 12.2, will play
a significant role; indeed, our general strategy is to decompose a Calderón–
Zygmund operator into a convergent series of dyadic singular integral opera-
tors and paraproducts. Thus, this final section brings together several of the
themes developed in this chapter.

12.3.a Weakly defined bilinear forms

In order to make a non-tautological study of the question of boundedness of

an operator, we need to give a meaning to the notion of an “operator” before
its boundedness has been established. As usual, this will involve postulating
the action of the operator on a dense class of test functions from which we
wish to extend this action to the full space under consideration. For a dyadic
analysis of singular integral operators, it is convenient to adopt the following
framework:

Definition 12.3.2. For a Banach space Z, a Z-valued bilinear form on S(D)

is a bilinear mapping
t : S(D)2 → Z.
If Z = L (X, Y ), we extend the action of such a mapping to

t : S(D; X) × S(D; Y ∗ ) → K

by letting

t(φ ⊗ x, ψ ⊗ y ∗ ) := ht(φ, ψ)x, y ∗ i ∈ K, φ, ψ ∈ S(D), x ∈ X, y ∗ ∈ Y ∗ ,

and extending by bilinearity, observing that S(D; X) = S(D) ⊗ X.

Remark 12.3.3 (S(D) vs. S00 (D) in the definition). Since S00 (D; X) is al-
ready dense in Lp (Rd ; X), in order to construct a bounded bilinear form on
0
Lp (Rd ; X) × Lp (Rd ; Y ), it would be sufficient to have an a priori estimate on
S00 (D; X) × S00 (D; Y ∗ ). However, for the type of theorems that we have in
mind, we also like to make assumptions on the action of our bilinear forms
on functions like 1Q ∈ S(D) \ S00 (D), and hence we need to have our initial
bilinear form defined on the larger product S(D; X) × S(D; Y ∗ ). This gives
rise to the following problem, where we take X = Y = K for simplicity, since
the issue is already present in this case:
Suppose that we have a bilinear form t : S(D)2 → K that satisfies the
estimate
 12.3 The T (1) theorem for abstract bilinear forms 149

|t(f, g)| 6 Ckf kp kgkp0 ∀(f, g) ∈ S00 (D)2 .

Thus there exists T ∈ L (Lp (Rd )) such that t(f, g) = hT f, gi whenever (f, g) ∈
S00 (D)2 . Does it follow that t(f, g) = hT f, gi for all (f, g) ∈ S(D)2 ?
Perhaps unexpectedly, the answer is “no”: Consider the bilinear form
Z Z
t(f, g) := f· g, (f, g) ∈ S(D)2 .
Rd Rd

If (f, g) ∈ S00 (D)2 , we have the a priori bound |t(f, g)| = 0, and hence the
unique operator T ∈ L (Lp (Rd )) is given by T = 0. But of course t is not
identically zero on S(D)2 . It is also clear that there cannot possible be any
T ∈ L (Lp (Rd )) with hT f, gi = t(f, g) for all (f, g) ∈ S(D)2 .
To avoid this problem, we make sure to get our a priori estimates on the
full set S(D; X) × S(D; Y ∗ ).

Definition 12.3.4. A bilinear form t : S(D)2 → L (X, Y ) is said to deter-

mine a bounded operator T ∈ L (Lp (Rd ; X), Lp (Rd ; Y )) provided that this
operator T satisfies
t(f, g) = hT f, gi
for all (f, g) ∈ S(D; X) × S(D; Y ∗ ).

In the case of reflexive spaces, the last-mentioned condition can be charac-

terised by an a priori estimate. Finding sufficient conditions for such an esti-
mate will be our primary concern below. The assumption of reflexivity is not
a serious restriction at this stage, since the deeper related considerations that
we shall encounter below will have much stronger assumptions, anyway.

Lemma 12.3.5. Let X and Y be reflexive Banach spaces, and let X0 ⊆ X

and Y 0 ⊆ Y ∗ be dense. Consider a bilinear form

t : S(D; X0 ) × S(D; Y 0 ) → K.

Let C > 0 be a constant and p ∈ (1, ∞). Then the following conditions, each
to hold for every choice of (f, g) ∈ S(D; X0 ) × S(D; Y 0 ), are equivalent:
(1) There is T ∈ L (Lp (Rd ; X), Lp (Rd ; Y )) of norm at most C such that

hT f, gi = t(f, g).
0 0
(2) There is T ∗ ∈ L (Lp (Rd ; Y ∗ ), Lp (Rd ; X ∗ )) of norm at most C such that

hf, T ∗ gi = t(f, g).

(3) There is a uniform estimate

|t(f, g)| 6 Ckf kLp (Rd ;X0 ) kgkLp0 (Rd ;Y 0 ) .

150 12 Dyadic operators and the T (1) theorem

Proof. (1) ⇒ (3) and (2) ⇒ (3) are immediate.

(3) ⇒ (1): Fix f ∈ Q(Rd ; X0 ). Then g 7→ t(f, g) defines a bounded linear
0 0
functional on a dense subspace of Lp (Rd ; Y ∗ ), and hence on Lp (Rd ; Y ∗ ). Thus
p0 d ∗ ∗
there is Λf ∈ (L (R ; Y )) such that
t(f, g) = hΛf , gi.
Moreover, since Y = Y ∗∗ is reflexive, it has the Radon–Nikodým property
0
by Theorem 1.3.21, and hence Λf ∈ (Lp (Rd ; Y ∗ ))∗ ' Lp (Rd ; Y ) by Theorem
1.3.10.
From the linearity of the left side in f , one deduces that f 7→ Λf is a
linear map from S(D; X) ⊆ Lp (Rd ; X) to Lp (Rd ; Y ), and (3) shows that it is
bounded. Hence there is a bounded extension T ∈ L (Lp (Rd ; X), Lp (Rd ; Y ))
with the required identity for (f, g) ∈ S(D; X0 ) × S(D; Y 0 ).
(3) ⇒ (2): This can be proved either similarly to the previous case, or using
the already proven implication (3) ⇒ (1) and the general existence result of
an adjoint
0 0
T ∗ ∈ L ((Lp (Rd ; Y ))∗ , (Lp (Rd ; X))∗ ) ' L (Lp (Rd ; Y ∗ ), Lp (Rd ; X ∗ )),
where the identification of the spaces was again based on the assumed reflex-
ivity via Theorems 1.3.21 and 1.3.10. By definition, the adjoint satisfies
hf, T ∗ gi = hT f, gi
0
for all (f, g) in Lp (Rd ; X) × Lp (Rd ; Y ∗ ) ⊇ S(D; X0 ) × S(D; Y 0 ). 
The very formulation of the conditions that give rise to the name “T (1) the-
orem” requires us to slightly extend the initial domain of weakly defined sin-
gular integral operators.
Definition 12.3.6. For a bilinear t : S(D)2 → Z, we say that t(hQ
α
, 1) is
well-defined if the series
X
t(hα
Q , 1) := t(hα
Q , 1R )
R∈D
`(R)=`(Q)

α
converges absolutely. We say that t(·, 1) is well-defined if t(hQ , 1) is well-
defined for every Q ∈ D and α ∈ {0, 1}d \ {0}.
We define t(1, hα
Q ) and t(1, ·) analogously.
α
Lemma 12.3.7. If t(hQ , 1) is well-defined, then
(1) for every k ∈ Z with 2−k > `(Q), we have
X
α α
t(hQ , 1) = t(hQ , 1R ),
R∈Dk

where the series converges absolutely in the weak operator topology;

 12.3 The T (1) theorem for abstract bilinear forms 151

(2) for every f ∈ S00 (D), the series

X
t(f, 1) := t(f, 1R )
R∈Dk

converges absolutely at least for all sufficiently negative k ∈ Z; moreover,

the value of the series is independent of k ∈ Z, as long as it converges
absolutely.
The analogous statements hold for t(1, ·).
Proof. (1): Let `(Q) = 2−j . For k = j, the claim of the lemma is just the
definition. For 2−k > 2−j and R ∈ Dk , we have
 X  X
t(hα α
Q , 1R ) = t hQ , 1S = t(hQα
, 1S ).
S∈Dj S∈Dj
S⊆R S⊆R

With f = hαQ , we then have

X X X X
t(f, 1) = t(f, 1S ) = t(f, 1S ) = t(f, 1R ), (12.34)
S∈Dj R∈Dk S∈Dj R∈Dk
S⊆R

where the first equality holds by assumption, and the assumed absolute con-
vergence allows to make the rearrangements and to get the absolute conver-
gence also in the subsequent steps.
(2): Each f ∈ S00 (D) is a linear combination of terms of the form hαQi ,
i

where i ∈ F for some finite index set F . If Q0 ∈ Dj0 is the largest cube
appearing here, then by the previous part of the lemma we know that
X
t(hα
Qi , 1R )
i

R∈Dk

converges absolutely for each k 6 j0 . Hence also

X X X
αi αi
t(f, 1R ) = hf, hQ i
i t(hQ i
, 1R )
R∈Dk i∈F R∈Dk

converges absolutely. If the absolute convergence holds for some j and k, the
equality of the corresponding series follows from (12.34).
The case of t(1, ·) is entirely analogous. 
As we shall see later, the forms t(1, ·) and t(·, 1) are closely related to
paraproducts. Since the boundedness of paraproducts is tricky, it is use-
ful to be able identify situations, when they can be avoided, i.e., when
t(1, ·) = 0 = t(·, 1).
With this goal in mind, we will now discuss an important case of trans-
lation-invariant bilinear forms. We first check that some natural candidates
for the definition are equivalent:
152 12 Dyadic operators and the T (1) theorem

Lemma 12.3.8. Let Z be a Banach space. The following conditions are equiv-
alent for a bilinear form t : S(D)2 → Z:
(1) t(1Q , 1R ) = t(1Q+̇m , 1R+̇m ) for all Q, R ∈ D with `(Q) = `(R), and all
m ∈ Zd , where Q+̇m := Q + m`(Q).
(2) t(f, g) = t(τh f, τh g) for all f, g ∈ S(D) and all dyadic rational vectors
h, i.e., all h of the form h = m2−k for some m ∈ Zd and k ∈ Z, where
τh f (s) := f (s − h).
If Z = L (X, Y ), these are also equivalent to a variant of (2) for all f ∈
S(D; X) and g ∈ S(D; Y ∗ ) instead.

Proof. (2)⇒(1): This is immediate by taking f = 1Q , g = 1R and h = m`(Q),

or f = 1Q ⊗ x, g = 1R ⊗ y ∗ for arbitrary x ∈ X and y ∗ ∈ Y ∗ in the variant
with Z = L (X, Y ).
(1)⇒(2): By definition, each f, g is a linear combination of some indicators
1Q (or 1Q ⊗ x resp. 1Q ⊗ y ∗ ) with Q ∈ D (and x ∈ X, y ∗ ∈ Y ∗ ), and we
have h = mh 2−kh for some mh ∈ Zd and kh ∈ Z. Since any dyadic cube
is an exact union of dyadic cubes of any given smaller size, and h can be
expressed in a similar form h = (2(k−kh ) )2−k for any k > kh , we may assume
that we have Q ∈ Dk and h = m2−k for the same k ∈ Z to begin with.
By bilinearity of both sides of the claim in (2), we thus need to verify that
t(1Q , 1R ) = t(τh 1Q , τh 1R ) = (1Q+̇m , 1R+̇m ) for each Q, R ∈ Dk , but this is
exactly what we assumed in (1). 

Definition 12.3.9. A bilinear form t : S(D)2 → Z is called translation-

invariant, if it satisfies the equivalent conditions of Lemma 12.3.8.

Formally, it is easy to see that t(1, ·) = 0 = t(·, 1) if t is translation invariant.

Namely, if Q ∈ D, and Q1 is the “lower left quadrant” of Q, then
[ X
α α
Q= (Q1 +̇γ), hQ = hhQ iQ1 +̇γ 1Q1 +̇γ ,
γ∈{0,1}d γ∈{0,1}d

−1/2
where the coefficients hhα
Q iQ1 +̇γ are equal to ±|Q| , with equally many of
each sign. Now, formally, we have

“ t(1, 1Q1 +̇γ ) = t(τγ`(Q1 ) 1, τγ`(Q) 1Q1 ) = t(1, 1Q1 ), ”

and hence
X
α
“ (1, hQ )= hhα
Q iQ1 +̇γ t(1, 1Q1 +̇γ )
γ∈{0,1}d
X
α
= hhQ iQ1 +̇γ t(1, 1Q1 ) = 0 · t(1, 1Q1 ) = 0. ”
γ∈{0,1}d

Problems with this computation are:

 12.3 The T (1) theorem for abstract bilinear forms 153

(1) While we defined t(1, hα α

Q ) for cancellative Haar functions hQ , the expres-
sions “t(1, 1Q1 +̇γ )” above need not even be defined; i.e., even if the series
defining the former converges, an analogous series for the latter need not.
(2) The assumption that t is translation invariant was made on the class of
functions S(D) only, and the constant function 1 is not in this class.
Nevertheless, under a mild decay assumption, and some care with limits, we
can bootstrap the above heuristics into a solid argument:

Proposition 12.3.10. Suppose that t : S(D)2 → Z is translation-invariant

and satisfies the decay assumption, for all Q ∈ D and m > MQ , that

kt(1Q , 1Q+̇m )k + kt(1Q+̇m , 1Q )k 6 cQ |m|−d . (12.35)

Then t(1, ·) = 0 = t(·, 1).

Proof. We fix some Q ∈ Dk and α ∈ {0, 1}d \ {0}. By definition, we have

X X
t(1, hα
Q) = t(1Q+̇m , hα
Q ) = lim t(1Q+̇m , hα
Q)
M →∞
m∈Zd m∈Zd
|m|∞ 6M
X X
= lim hhα
Q iQ1 +̇γ t(1Q1 +̇β +̇2m , 1Q1 +̇γ ),
M →∞
β,γ∈{0,1}d m∈Zd
|m|∞ 6M

where rearranging the order of the finite sums inside the limit presents no
issues. Here

t(1Q1 +̇β +̇2m , 1Q1 +̇γ ) = t(1Q1 +̇(β−γ)+̇2m , 1Q1 ),

and hence, noting that β − γ ∈ {−1, 0, 1}d ,

X X
t(1Q1 +̇β +̇2m , 1Q1 +̇γ ) = t(1Q1 +̇n , 1Q1 )
d d
m∈Z n∈Z
|m|∞ 6M n∈[−2M,2M ]d +(β−γ)
 X X 
= + t(1Q1 +̇n , 1Q1 )
d d
n∈Z n∈Z
n∈[−(2M −1),2M −1]d n∈[−2M,2M ]d +(β−γ)
n∈[−(2M
/ −1),2M −1]d
β−γ
=: IM + IIM .
β−γ
In IIM , we note that at least one component ni of n must satisfy |ni | > 2M ,
and hence the decay assumption (12.35) ensures that

kt(1Q1 +̇n , 1Q1 )k 6 cQ1 (1 + 2M )−d .

On the other hand, we have n ∈ [−(2M +1), 2M +1]d \[−(2M −1), (2M −1)]d ,
and the total number of such n ∈ Zd is
154 12 Dyadic operators and the T (1) theorem
(1 + 2(2M + 1))d − (1 + 2(2M − 1))d = (4M + 3)d − (4M − 1)d
6 4d(4M + 3)d−1 ,

and hence
β−γ
kIIM k 6 4d(4M + 3)d−1 × cQ1 (1 + 2M )−d 6 cd cQ1 M −1 .

Substituting back, it follows that

X β−γ
α
t(1, hQ ) = lim hhα
Q iQ1 +̇γ (IM + IIM )
M →∞
β,γ∈{0,1}d
X β−γ
= lim α
hhQ iQ1 +̇γ IIM = lim O(M −1 ) = 0.
M →∞ M →∞
β,γ∈{0,1}d

The computation for t(hα

Q , 1) is entirely similar. 

Remark 12.3.11. It is easy to see from the proof that the decay assumption
(12.35) could be somewhat weakened. We have not strived for maximal gen-
erality at this point, but stated a condition that is both relatively simple to
formulate and easy to verify in our main application to Calderón–Zygmund
singular integrals.

12.3.b The BCR algorithm and Figiel’s decomposition

In order to analyse t(f, g), we will use the auxiliary operators

X X
Ek f = EQ f = hf iQ 1Q , Dk = {Q ∈ D : `(Q) = 2−k }.
Q∈Dk Q∈Dk

X  X  X
Dk f = Ek+1 f − Ek f = EQ 0 f − EQ f = DQ f.
Q∈Dk Q0 ∈ch(Q) Q∈Dk

Our starting point for the analysis of a bilinear form is the following useful
identity:

Lemma 12.3.12 (Beylkin–Coifman–Rokhlin (BCR) algorithm). Let

X, Y be Banach spaces, and let t : S(D)2 → L (X, Y ) be bilinear. Suppose
that f ∈ S(D; X) and g ∈ S(D; Y ∗ ) are constant on all Q ∈ DM . Then for
all integers m < M ,
X X
t(f, g) = t(Dk f, Dk g) + t(Dk f, Ek g)
m6k<M m6k<M
X (12.36)
+ t(Ek f, Dk g) + t(Em f, Em g).
m6k<M
 12.3 The T (1) theorem for abstract bilinear forms 155

Proof. That f is constant on all Q ∈ DM means that f = EM f , and similarly

g = EM g. Thus we have

t(f, g) − t(Em f, Em g) = t(EM f, EM g) − t(Em f, Em g)

X
= (t(Ek+1 f, Ek+1 g) − t(Ek f, Ek g)),
m6k<M

where
t(Ek+1 f, Ek+1 g) = t((Dk + Ek )f, (Dk + Ek )g)
= t(Dk f, Dk g) + t(Dk f, Ek g) + t(Ek f, Dk g) + t(Ek f, Ek g),

and hence
t(Ek+1 f, Ek+1 g) − t(Ek f, Ek g)
= t(Dk f, Dk g) + t(Dk f, Ek g) + t(Ek f, Dk g).

Remark 12.3.13. The upper bound k < M imposed on the summation vari-
ables above is redundant: the condition that f and g are constant on all
Q ∈ DM implies that Dk f = 0 = Dk g for k > M , so that the right side would
remain unchanged if we allow the summations to run to infinity.

The final term in the expansion 12.36 is an error term, and can be controlled
under the following mild conditions, which are obviously necessary for t to
define a bounded operator on Lp :

Definition 12.3.14. We say that a bilinear t : S(D)2 → Z satisfies

(1) the weak boundedness property if

kt(1Q , 1Q )kZ 6 ktkwbp |Q| ∀Q ∈ D;

(2) the adjacent weak boundedness property if

kt(1Q , 1Q+̇n )kZ 6 ktkawbp |Q| ∀Q ∈ D, ∀n ∈ {−1, 0, 1}d . (12.37)

Lemma 12.3.15. Let X, Y be Banach spaces, and let a bilinear t : S(D)2 →

L (X, Y ) satisfy the adjacent weak boundedness property. Then for all f ∈
S(D; X) and g ∈ S(D; Y ), and all negative enough m, we have

|t(Em f, Em g)| 6 2d ktkawbp kEm f kLp (Rd ;X) kEm gkLp0 (Rd ;Y ∗ ) −→ 0.
m→−∞

Proof. We choose m so negative that the (bounded) supports of f ∈ S(D; X)

and g ∈ S(D; Y ∗ ) are both contained in the union of at most 2d cubes Q ∈ Dm
such that any two of them are related by R = Q+̇n for some n ∈ {−1, 0, 1}d .
We then have
156 12 Dyadic operators and the T (1) theorem
X X
t(Em f, Em g) = t(EQ f, ER g) = t(hf iQ 1Q , hgiR 1R ),
Q,R∈Dm Q,R∈Dm

and thus
X
|t(Em f, Em g)| 6 kt(1Q , 1R )kL (X,Y ) khf iQ kX khgiR kY ∗
Q,R∈Dm
X
6 ktkawbp |Q|khf iQ kX khgiR kY ∗
Q,R∈Dm
X X 0
= ktkawbp |Q|1/p khf iQ kX |R|1/p khgiR kY ∗
Q∈Dm R∈Dm

0
 X 1/p  X 0
1/p0
p
6 ktkawbp 2d/p |Q|khf iQ kX 2d/p |R|khgiR kpY ∗
Q∈Dm R∈Dm
d
= 2 ktkawbp kEm f kLp (Rd ;X) kEm gkLp (Rd ;Y ∗ ) ,
which is the claimed bound. 
The other terms in (12.36) can be identified with the various operators that
we have studied in the previous sections:
Definition 12.3.16. Let X, Y be Banach spaces, let t : S(D)2 → L (X, Y )
be a bilinear form, and let t(·, 1) and t(1, ·) be well-defined. We define the
following operators associated with t:
X αγ
Ht := Htα,γ , where Hαγtα,γ
are Haar multipliers (12.3),
0 0
α,γ
X
Tn,t := Tφαγ α,γ ,
n ,tn
where Tφαγ α,γ are Figiel’s operators (12.14)
n ,tn
α,γ∈{0,1}d \{0}
(
φn (Q) := Q+̇n := Q + n`(Q),
with γ
tα,γ α
n (Q) := t(hQ , hQ+̇n ),
X
i
Un,t := Uφα i,α , where Uφα i,α are Figiel’s operators (12.19),
n ,un n ,un
α∈{0,1}d \{0}
(
tn1,α (Q)∗ := t(h0Q+̇n , hα ∗
Q) , i = 1,
with uni,α (Q) :=
t2,α
n (Q) := α
t(hQ 0
, hQ +̇n
), i = 2.
We also define the related paraproducts:
Πt1 := paraproduct with coefficients t(1, hα
Q ),
∗
Πt2 := paraproduct with coefficients t(hα
Q , 1) ,
α,1 α,2
Λt := bi-paraproduct with coefficients πQ = t(1, hα α
Q ) and πQ = t(hQ , 1),

lt := the bilinear form of Λt .

We may drop the subscript t from these notations if it is obvious from the
context.
 12.3 The T (1) theorem for abstract bilinear forms 157

i
Remark 12.3.17. Our indexing of the operators Un,t may appear counterin-
2
tuitive at first sight, as one might like to think of the operators Un,t , which
act on f ∈ L (R ; X) with coefficients t(hQ , hQ+̇n ) ∈ L (X, Y ), as deserv-
p d α 0

1
ing to be the “primary” ones, rather than Un,t , which act on the dual side
p0
Q ) ∈ L (Y , X ). How-
∗ ∗ ∗ ∗
g ∈ L (R ; Y ) with adjoint coefficients t(hQ+̇n , hα
d 0

i
ever, this indexing is chosen, since the operators Un,t naturally arise in par-
allel with the paraproducts Πi of the same index i ∈ {1, 2}—see (12.42) and
(12.43) below—, and it turns out to have some other advantages in the sequel.
With this notation, we can formula Figiel’s decomposition of a bilinear form:
Proposition 12.3.18 (Figiel). Let X, Y be Banach spaces, let t : S(D)2 →
L (X, Y ) be a bilinear form, and let t(·, 1) and t(1, ·) be well-defined. For all
f ∈ S(D; X), g ∈ S(D; Y ∗ ), m ∈ Z,
denoting
u := (I − Em )f ∈ S00 (D; X), v := (I − Em )g ∈ S00 (D; Y ∗ ),
we have the following identity with absolute convergence:
t(f, g) = hHt u, gi + hΠt1 f, vi + hu, Πt2 gi + t(Em f, Em g)+
X n o
1 2
+ hTn,t u, gi + hf, Un,t vi + hUn,t u, gi , (12.38)
d
n∈Z
n6=0

where the operators on the right are as in Definition 12.3.16. If these coeffi-
cients satisfy
α
kt(1, hQ )k, kt(hα
Q , 1)k 6 C|Q|
1/2
, (12.39)
then we have the further identity, with all terms below well defined:
hΠt1 f, vi + hu, Πt2 gi = hΛt f, gi − hm Πt1 f, gi − hf, m Πt2 gi. (12.40)
Remark 12.3.19. Since Hλαγ = Tφαγ 0 ,λ
, we could have incorporated the Haar
multiplier into the second line of (12.38) as hHt u, gi = hT0,t u, gi. But we
prefer to keep it separate, since its treatment will involve some differences
compared to the rest of the Tn,t .
Proof of Proposition 12.3.18. We start with the identity (12.36) of Lemma
12.3.12. Since the sums are finitely nonzero, we are free rearrange as follows,
observing that dyadic cubes Q, R of the same size are necessarily integer (times
side-length) translates of each other:
X X X
t(Dk f, Dk g) = t(DQ f, DR g)
k>m k>m Q,R∈Dk
X X X X X
= t(DQ f, DQ+̇n g) = t(DQ f, DQ+̇n g)
k>m Q∈Dk n∈Zd Q∈D n∈Zd
`(Q)62−m
158 12 Dyadic operators and the T (1) theorem

and we can also switch the order of the last two sums. Observing that u =
(I −Em )f satisfies DQ u = DQ f for `(Q) 6 2−m and DQ u = 0 for `(Q) > 2−m ,
we find that, replacing f by u (and/or g by v) we can drop the restriction
`(Q) 6 2−m in the sum. Moreover, using the convention that summations
over α and γ are always over the set {0, 1}d \ {0},
X XXD γ γ
E
t(DQ u, DQ+̇n g) = t(hα α
Q , hQ+̇n )hhQ , ui, hhQ+̇n , gi
Q∈D α,γ Q∈D
X
= hTφα,γ αγ u, gi = hTn u, gi.
n ,tn
α,γ

Hence
X X X
t(Dk f, Dk g) = hTn u, gi = hHu, gi + hTn u, gi
k>m n∈Zd
(12.41)
n∈Zd
n6=0

For the terms involving Ek , we begin in the same way but then introduce
an additional twist to force some cancellation:
X X X
t(Dk f, Ek g) = t(DQ f, EQ+̇n g)
k>m Q∈D n∈Zd
`(Q)62−m
X X  
= t(DQ f, 1Q+̇n (hgiQ+̇n − hgiQ )) + t(DQ f, 1Q+̇n hgiQ )) .
Q∈D n∈Zd
`(Q)62−m

The assumption that t(·, 1) is well-defined guarantees the absolute convergence

of X
t(DQ f, 1Q+̇n hgiQ ) =: t(DQ f, hgiQ ).
n∈Zd

Recalling that only finitely many DQ f with `(Q) 6 2−m are non-zero, we also
get the absolute convergence of
X X X
t(DQ f, 1Q+̇n hgiQ ) = t(DQ f, hgiQ ) =: pm (f, g),
Q∈D n∈Zd Q∈D
`(Q)62−m `(Q)62−m

and hence, by triangle inequality, that of

X
t(DQ f, 1Q+̇n (hgiQ+̇n − hgiQ )).
n∈Zd

Thus we can make the rearrangements

X X X
t(Dk f, Ek g) = t(DQ f, 1Q+̇n (hgiQ+̇n − hgiQ )) + pm (f, g)
k>m n∈Zd Q∈D
n6=0 `(Q)62−m
 12.3 The T (1) theorem for abstract bilinear forms 159

where adding the summation condition n 6= 0 was for free, since the factor
hgiQ+̇n − hgiQ evidently vanishes when n = 0. Again, replacing f by u allows
us to drop the restrictions to `(Q) 6 2−m both in the sum spelled out above
and in pm (f, g). Moreover,
X
t(DQ u, 1Q+̇n (hgiQ+̇n − hgiQ ))
Q∈D
XXD E
α
= t(hQ α
, h0Q+̇n )hhQ 0
, ui, hhQ 0
+̇n − hQ , gi
α Q∈D
X
= hUφα α,0
2
u, gi = hUn,t u, gi.
n ,tn
α

Directly from the definitions, we also have

X
pm (f, g) = t(DQ u, hgiQ )
Q∈D
X X
= t(hα α
Q hhQ , ui, hgiQ )
Q∈D α∈{0,1}d \{0}
X X D E
= t(hα α
Q , 1)hhQ , ui, hgiQ
Q∈D α∈{0,1}d \{0}
X X D E
= α
hhQ α
, ui, t(hQ , 1)∗ hgiQ = hu, Πt2 gi.
Q∈D α∈{0,1}d \{0}

In the computation above, the fact that u ∈ S00 (D; X) guarantees that all
summations are finite, and the last step is simply the definition of the para-
product via its action of the finitely non-zero Haar expansions in the dual
space. Hence we have verified that
X X
2
t(Dk f, Ek g) = hUn,t u, gi + hu, Πt2 gi, (12.42)
k>m n∈Zd
n6=0

and the proof that

X X
1 1
t(Ek f, Dk g) = hf, Un,t vi + hΠn,t f, vi (12.43)
k>m n∈Z d
n6=0

is entirely analogous. Substituting the previous two identities and (12.41) into
(12.36), we obtain the claimed (12.38).
Under the additional assumption (12.39), we know from Corollary 12.2.12
that hΠt1 f, gi is well-defined and bilinear in (f, g) ∈ S(D; X) × S(D; Y ∗ ), and
hence
hΠt1 f, vi = hΠt1 f, gi − hΠt1 f, Em gi = hΠt1 f, gi − hm Πt1 f, gi.
Similarly, hu, Πt2 gi = hf, Πt2 gi − hf, m Πt2 gi, and the previous two identities
combine to give (12.40), noting that hΠt1 f, gi + hf, Πt2 gi = hΛt f, gi. 
160 12 Dyadic operators and the T (1) theorem

12.3.c Figiel’s T (1) theorem

The previous section culminated in Proposition 12.3.18, which established a

decomposition of a generic bilinear form t : S(D)2 → L (X, Y ) in terms of
various fundamental operators. This is as far as it seems useful to proceed with
identities, and we now turn to conditions that allow us to make meaningful
estimates of the terms in the obtained decomposition. For this purpose, we
introduce a certain family of norms. For a smooth discussion of a couple
of closely related variants, it is convenient to adopt the following general
framework.

Definition 12.3.20. Let Z be a Banach space, and P(Z) the collection of

all subsets of Z. We say that ℘ : P(Z) → [0, ∞] is a good set-bound on Z, if
it satisfies the following properties for all S , T ⊆ Z:
(1) If S ⊆ T , then ℘(S ) 6 ℘(T ).
(2) ℘(S ∪ T ), ℘(S + T ) 6 ℘(S ) + ℘(T ).
(3) If Z ⊆ K, then ℘(Z T ) 6 supz∈Z |z| × ℘(T ).
(4) ℘(T ) = ℘(conv T ) = ℘(abs conv T ).
(5) ℘(T ) = ℘(T ), where T denotes the norm-closure of T .

We primarily have in mind the following three cases:

Lemma 12.3.21. Let X and Y be Banach spaces and p ∈ [1, ∞). Then each
of the following ℘ is a good set-bound on Z = L (X, Y ):
(a) ℘ = U , where U (T ) := sup{kT k : T ∈ T },
(b) ℘ = Rp , the R-bound of order p,
(c) ℘ = Rp∗ , the dual R-bound defined by

Rp∗ (T ) := Rp (T ∗ ), T ∗ := {T ∗ ∈ L (Y ∗ , X ∗ ) : T ∈ T }.

Proof. (a): The verification of the properties is immediate.

(b): Properties (1) and (2) for ℘ = Rp are contained in the items with
same numbers in Proposition 8.1.19. Property (3) follows from

Rp (Z T ) 6 Rp (Z )Rp (T ), Rp (Z ) = sup |z|,

z∈Z

where the first estimate is Proposition 8.1.19(3) and the second is immediate
from Kahane’s contraction principle (cf. the discussion right before Defini-
tion 8.1.1 of R-boundedness). Finally, properties (4) and (5) are contained in
Propositions 8.1.21 and 8.1.22, respectively.
(c): All properties are direct corollaries of the corresponding properties
in (b), since all set operations involved in these properties are well-behaved
under the adjoint operation:
(1) S ⊆ T if and only if S ∗ ⊆ T ∗ ,
 12.3 The T (1) theorem for abstract bilinear forms 161

(2) (S ∪ T )∗ = S ∗ ∪ T ∗ and (S + T )∗ = S ∗ + T ∗ ,
(3) if Z ⊆ K, then (Z T )∗ = Z T ∗ ,
(4) (conv T )∗ = conv(T ∗ ) and (abs conv T )∗ = abs conv(T ∗ ),
(5) ( T )∗ = T ∗ .


Definition 12.3.22 (Figiel norms of a bilinear form). For a bilinear

form t : S(D)2 → L (X, Y ), let tαγn , tn : D → L (X, Y ) be the associated
i,α

functions appearing in Proposition 12.3.18. For s > 0 and a good set-bound ℘

on L (X, Y ), we define
X
ktθ kFigs (℘) := (2 + log2 |n|)s ℘(tθn ), θ ∈ {(α, γ), (i, α)},
n∈Zd
n6=0
X
kt(0) kFigs (℘) := ktαγ kFigs (℘) ,
α,γ∈{0,1}d \{0}
X
kt(i) kFigs (℘) := kti,α kFigs (℘) , i ∈ {1, 2},
α∈{0,1}d \{0}
2
X
ktkFigs (℘) := kt(i) kFigs (℘) .
i=0

When ℘ = U is as in Lemma 12.3.21(a), we write Figs (∞) := Figs (U ).

Remark 12.3.23. Referring to Proposition 12.3.18, one observes that the Figiel
γ
norms impose control on pairings t(hα Q , hQ ), where at least one of the Haar
functions is cancellative, i.e., (α, γ) 6= (0, 0). This is in contrast to the decay
condition (12.35), where α = γ = 0.
∗
Since we also encountered the adjoint function u1,α 1,α
n (Q) := (tn (Q)) , we recall
the following results from the previous volumes:

Proposition 12.3.24. Let X and Y be Banach spaces, T ⊆ L (X, Y ), and

p ∈ (1, ∞). If X is K-convex (resp. a UMD space), then
 
Rp∗0 (T ) 6 Kp,X Rp (T ) 6 βp,X
+
Rp (T ) .

If Y is K-convex (resp. a UMD space), then

 
Rp (T ) 6 Kp,Y Rp∗0 (T ) 6 βp,Y
+
Rp∗0 (T ) .

In particular, if both X and Y are K-convex (resp. UMD spaces), the set-
bounds Rp and Rp∗0 are equivalent on L (X, Y ).

Proof. The first inequalities in both chains are restatements of bounds in

+
Proposition 8.4.1, and we have Kp,Z 6 βp,Z by Proposition 4.3.10. 
162 12 Dyadic operators and the T (1) theorem

Thanks to Proposition 12.3.24, we would not need to distinguish (when work-

ing in UMD spaces) between direct and adjoint R-boundedness conditions, as
such assumptions are actually equivalent. Nevertheless, we choose to do so,
for twofold reasons. First, as far as quantitative conclusions are concerned, we
would lose a constant each time we pass to the dual side, whereas in many
applications, verifying the R-boundedness of concrete operators is just as easy
(or difficult) directly on the dual side, so that applying the general duality
result for R-boundedness is unnecessary. Second, writing the adjoint bounds
explicitly, where they are relevant, will hopefully better clarify the role of the
different assumptions in the estimates.
In the following lemma, we observe that Figiel norm estimates, of the type
we will need to assume any way, will also guarantee the well-definedness of
t(·, 1) and t(1, ·)), which allows us to drop these as separate assumptions in
the sequel.

Lemma 12.3.25. Let X and Y be Banach spaces, and let t : S(D)2 →

L (X, Y ) be a bilinear form. If kt(2) kFig0 (∞) < ∞ (resp. kt(1) kFig0 (∞) < ∞),
then t(·, 1) (resp. t(1, ·)) is well defined, and
α
kt(hQ , 1)k 6 kt2,α kFig0 (∞) |Q|1/2 ,
  (12.44)
kt(1, hα
Q )k 6 kt1,α
kFig 0 (∞) |Q|
1/2
.

Proof. For every Q ∈ D and α ∈ {0, 1}d \ {0}, we have

X X
α
kt(hQ , 1R )k = kt(hα 0
Q , hQ+̇n )k|Q|
1/2

R∈D n∈Zd
`(R)=`(Q)
X
6 ktnα,0 (Q)k|Q|1/2 = kt2,α kFig0 (∞) |Q|1/2 < ∞,
n∈Zd

which shows both that t(·, 1) is well defined and the related bound. The case
of t(1, ·) is analogous. 

Theorem 12.3.26 (T (1) theorem for bilinear forms). Let p ∈ (1, ∞)

and 1 6 ti 6 p 6 qi 6 ∞, i = 0, 1, 2, where q1 = ∞ and t2 = 1. Consider the
following conditions:
(i) X and Y are UMD spaces;
(ii) X has cotype qi and Y has type ti , or one of them has both, for each
i = 0, 1, 2,
(iii) t : S(D)2 → L (X, Y ) is a bilinear form with

X 2
X
DR p (tαγ
0 )+ kt(i) kFigσi (Rp ) < ∞,
α,γ i=0

where σi := 1/ti − 1/qi ,

 12.3 The T (1) theorem for abstract bilinear forms 163

(iv) t satisfies the adjacent weak boundedness property.

Under assumptions (i) through (iv), the bilinear form t − lt defines a bounded
operator T − Λt ∈ L (Lp (Rd ; X), Lp (Rd ; Y )) that satisfies
(a) the norm estimate:
nX
kT − Λt kL (Lp (Rd ;X),Lp (Rd ;Y )) 6 βp,X βp,Y DR p (t0α,γ )+
α,γ
2
X o
+ Ad min C0,i kt(0) kFigσ0 (℘i ) + Bd Ci,i kt(i) kFigσi (℘i )
i=1,2
i=1

where Ad := 6 · (81)d , Bd := 5 200 · (81)d , ℘1 := Rp∗0 , ℘2 := Rp , and

Ci,2 := C(12.15) (X, Y, p, qi , ti ), Ci,1 := C(12.15) (Y ∗ , X ∗ , p0 , t0i , qi0 ),
(b) the representation formula, with absolute convergence for all f ∈ Lp (Rd ; X)
0
and g ∈ Lp (Rd ; Y ∗ ):
X 
1 2
h(T − Λt )f, gi = hHt f, gi + hTn,t f, gi + hf, Un,t gi + hUn,t f, gi ,
d
(12.45)
n∈Z
n6=0

where the operators on the right are as in Definition 12.3.16.

Under assumptions (i) through (iii), the following conditions are equivalent:
(1) t defines a bounded operator T ∈ L (Lp (Rd ; X), Lp (Rd ; Y ));
(2) t satisfies (iv), and lt defines a bounded Λt ∈ L (Lp (Rd ; X), Lp (Rd ; Y )).
Under these equivalent conditions, we have both (a) and (b).
Remark 12.3.27. The assumptions of Theorem 12.3.26 allow a certain trade-
off between the Figiel norms that one imposes on the bilinear form t on the one
hand, and (co)type assumptions (and the size of the related constants) on the
spaces X and Y on the other hand. Indeed, the norms k kFigσi become smaller
with decreasing σi = 1/ti − 1/qi , thus with increasing type ti or decreasing
cotype qi , but at the same time the related constants C(12.15) may increase.
Let 1 6 t 6 p 6 q 6 ∞ and suppose that X has cotype q and Y has
type t, or one of them has both. In Theorem 12.3.26, we will then choose
(t1 , q1 ) = (t, ∞) and (t2 , q2 ) = (1, q); thus σ1 = 1/t and σ2 = 1/q 0 . However,
there are three prominent choices of the exponents t0 and q0 :
(0) With (t0 , q0 ) = (t, q), we have
1 1
σ0 = − 6 min σi ,
t q i=1,2
with strict inequality if both t and q are chosen to be non-trivial (as one
always can for UMD spaces X and Y by Proposition 7.3.15). This shows
that a strictly weaker condition is required on t(0) than on t(i) with i = 1, 2,
but this seems to be largely a curiosity.
164 12 Dyadic operators and the T (1) theorem

(1) With (t0 , q0 ) = (t1 , q1 ) = (t, ∞), we have σ0 = σ1 . Thus, we impose a

stronger norm of t(0) than in case (0), but we achieve the following better
constants in Theorem 12.3.26(a) under this choice:

C0,1 = C(12.15) (Y ∗ , X ∗ , p0 , t0 , 1) = C1,1 ,

while an inspection of (12.15) shows that C0,1 is larger than C1,1 in general.
(2) Similarly, with (t0 , q0 ) = (t2 , q2 ) = (1, q), we get

C0,2 = C(12.15) (X, Y, p, q, 1) = C2,2 .

Using either choice (1) or (2) in Theorem 12.3.26, its key norm estimate admits
the following form, under the assumption (we recall) that X has cotype q and
Y has type t, or one of them has both,
nX
kT − Λt kL (Lp (Rd ;X),Lp (Rd ;Y )) 6 βp,X βp,Y DR p (t0α,γ )+
α,γ
2
X  o
+ Ci Ad kt(0) kFigσi (℘i ) + Bd kt(i) kFigσi (℘i ) ,
i=1

where σ1 = 1/t, σ2 = 1/q 0 , and

C1 := C(12.15) (Y ∗ , X ∗ , p0 , t0 , 1), C2 := C(12.15) (X, Y, p, q, 1).

Proof of Theorem 12.3.26. The core of the proof will consist of establishing
claims (a) and (b) under the full set of assumptions (i) through (iv). Assuming
that this is already done, let us see how to conclude the rest of the proof.
The equivalence of (1) and (2) is asserted under the assumptions (i)
through (iii) only. However, the adjacent weak boundedness property (iv)
is clearly necessary for (1) and it is explicitly assumed in (2), so we can as-
sume that this condition is satisfied in any case, and so we are in fact working
under the full set of assumptions (i) through (iv) also in this remaining part
of the proof. Thus the consequences (a) and (b) of this assumption are valid.
In particular, since the bilinear form t − l defines a bounded operator under
this assumption, it is clear that t defines a bounded operator if and only if l
does.
We then turn to the actual proof of (a) and (b) under the assumptions (i)
through (iv). From Lemma 12.3.25, we get that t(·, 1) and t(1, ·), and hence the
two paraproducts, are well defined, and their coefficients satisfy the bounds
(12.44). For f ∈ S(D; X) and g ∈ S(D; Y ∗ ), we then have both identities
(12.38) and (12.40) provided by Proposition 12.3.18. Combined together, they
read as
t(f, g) = hHum , gi + hΛf, gi + Em (f, g)+
X n o
+ hTn um , gi + hf, Un1 vm i + hUn2 um , gi , (12.46)
d
n∈Z
n6=0
 12.3 The T (1) theorem for abstract bilinear forms 165

where um := (I − Em )f ∈ S00 (D; X), vm := (I − Em )g ∈ S00 (D; Y ∗ ), and the

error term

Em (f, g) = hm Π1 f, gi + hf, m Π2 gi + t(Em f, Em g)

satisfies
 2
X 
|Em (f, g)| 6 cd,p kt(i) kFig0 (∞) + 2d ktkawbp ×
i=1 (12.47)
× kEm f kLp (Rd ;X) kEm gkLp (Rd ;Y ∗ ) −→ 0
m→−∞

by Lemmas 12.2.11 and 12.3.25 for the paraproduct terms and Lemma 12.3.15
for both the final term and the limit.
Directly from Theorem 12.1.11, we deduce that
X
|hHum , gi| 6 |hHtαγ
α,γ um , gi|
0
α,γ
X (12.48)
+
6 βp,X βp+0 ,Y ∗ DR p (tα,γ
0 )kum kp kgkp ,
α,γ

where, and in the rest of the proof, we abbreviate

k kp := k kLp (Rd ;X) , k kp0 := k kLp0 (Rd ;Y ∗ ) .

Note that φn (Q) := Q+̇n satisfies φn (Q) ⊆ 3Q(N ) provided that |n| 6 2N ;
thus in particular for N = dlog+ 2 |n|e; this is relevant in view of applying
Corollary 12.1.27 and Theorem 12.1.28. From Corollary 12.1.27, we deduce
that
X
|hTn um , gi| 6 |hTφαγ α,γ um , gi|
n ,tn
α,γ

6 Ad βp,X βp,Y (2 + log2 |n|)1/t0 −1/q0 min C0,i ℘i (tα,γ

n )kum kp kgkp0
i=1,2

using the notation of the statement of the theorem that we are proving. Hence
X
|hTn um , gi| 6 Ad βp,X βp,Y min C0,i kt(0) kFig1/t0 −1/q0 (℘i ) kum kp kgkp0
i=1,2
n∈Zd
n6=0

Similarly, recalling that t2 := 1, Theorem 12.1.28 guarantees that

X
|hUn2 um , gi| 6 |hUφα ,t2,α um , gi|
n n
α
X
6 Bd βp,X βp,Y (2 + log2 |n|)1/t2 −1/q2 C2,2 ℘2 (t2,α
n )kum kp kgkp0
α

in the notation of the theorem, and hence

166 12 Dyadic operators and the T (1) theorem
X
|hUn2 um , gi| 6 Bd βp,X βp,Y C2,2 kt(2) kFig1/t2 −1/q2 (℘2 ) kum kp kgkp0 .
n∈Zd
n6=0

For the term hf, Un1 vm i, we again apply Theorem 12.1.28 but on the dual
side, with X, Y, p replaced by Y ∗ , X ∗ , p0 . By assumption, Y has type t1 6 p,
and hence Y ∗ has cotype t01 > p0 by Proposition 7.1.13. So we can indeed
apply Theorem 12.1.28 with X, Y, p, q replaced by Y ∗ , X ∗ , p0 , t01 . Recalling
that q1 := ∞, and noting that 1 − 1/t01 = 1/t1 = 1/t1 − 1/q1 , this gives
X
|hf, Un1 vm i| 6 |hf, Uφα ,(t1,α )∗ vm i|
n n
α

6 Bd βp0 ,X ∗ βp0 ,Y ∗ (2 + log2 |n|)1/t1 −1/q1 kf kp kvm kp0 ×

× C(Y ∗ , X ∗ , p0 , t01 )Rp0 ((t1,α ∗
n ) ),

where βp0 ,X ∗ βp0 ,Y ∗ = βp,X βp,Y and

n ) ) = C1,1 Rp0 (tn ) = C1,1 ℘1 (tn )

C(Y ∗ , X ∗ , p0 , t10 )Rp0 ((t1,α ∗ ∗ 1,α 1,α

in the notation of the theorem. Hence

X
|hf, Un1 vm i| 6 Bd βp,X βp,Y C1,1 kt(1) kFig1/t1 −1/q1 (℘1 ) kf kp kvm kp0 .
n∈Zd
n6=0

Noting that kum kp 6 2kf kp and kvm kp0 6 2kgkp0 , and using the assump-
tion about kt(i) kFig1/ti −1/qi (Rp ) (combined with Proposition 12.3.24 in the case
of Rp0 ((t1,α ∗
n ) )), it follows that the series in (12.46) are term-wise and uni-
formly in m dominated by absolutely convergent series. This allows us to pass
to the limit m → −∞ in (12.46) with dominated convergence to deduce that

(t − l)(f, g) = RHS(12.45) ∀ f ∈ S(D; X), g ∈ S(D; Y ∗ ). (12.49)

Taking the same limit in the term-wise bounds above, we obtain

|(t − l)(f, g)| = |t(f, g) − hΛf, gi|
nX
6 βp,X βp,Y DR p (tα,γ
0 ) + Ad min C0,i kt
(0)
kFig1/t0 −1/q0 (℘i )
i=1,2
α,γ (12.50)
2
X o
+ Bd Ci,i kt(i) kFig1/ti −1/qi (℘i ) kf kLp (Rd ;X) kgkLp0 (Rd ;Y ∗ )
i=1

again for all f ∈ S(D; X) and g ∈ S(D; Y ∗ ), where Ad , Bd and Ci are as in

the statement of the Theorem.
This estimate shows that the bilinear form t − l satisfies a relevant a priori
bound, and hence defines an operator T − Λ ∈ L (Lp (Rd ; X), Lp (Rd ; Y )). By
density, it is immediate that (12.50) remains valid with general f ∈ Lp (Rd ; X)
0
and g ∈ Lp (Rd ; Y ∗ ), and this proves the claimed norm bound (a) for T − Λ.
 12.3 The T (1) theorem for abstract bilinear forms 167

We can then replace (t−l)(f, g) by h(T −Λ)f, gi in (12.49). Approximating

0
general f ∈ Lp (Rd ; X) and g ∈ Lp (Rd ; Y ∗ ) by functions as in (12.49), and
using dominated convergence and the term-wise bounds recorded above, this
proves the representation (b). This completes the proof of the claims under
the assumption that t satisfies the adjacent weak boundedness property. 

12.3.d Improved estimates via random dyadic cubes

A feature of Theorem 12.3.26 is that it deals with a bilinear form adapted

to a fixed system of dyadic cubes D. This is an advantage in applications
to questions of intrinsically dyadic nature. But it is also a certain limitation
in view of applications to non-dyadic questions, in that the assumptions of
Theorem 12.3.26 fail to take advantage of possible information about non-
dyadic cubes. For example, with some effort, one could use Theorem 12.3.26
to re-derive the boundedness of the Hilbert transform on Lp (R; X), which
we proved in a different way in Theorem 5.1.13. However, the conclusion
derived from Theorem 12.3.26 would be quantitatively weaker, in terms of
the dependence on the UMD constant βp,X , which was quadratic in Theorem
2
5.1.13. For X = Y , Theorem 12.3.26 also features the explicit factor βp,X , but
there is another βp,X implicit in the constants C(12.15) . On the other hand,
it is evident that, for t(f, g) := hHf, gi, there is no difference in estimating
γ
t(hαI , hK ) for dyadic or non-dyadic intervals I, J. But Theorem 12.3.26, as
formulated, makes no use of this additional information.
We now wish derive to variant of Theorem 12.3.26 to address these issues.
First of all, we need a straightforward generalisation to Rd of the random
dyadic systems that we used in the one-dimensional case in Section 5.1.

Lemma 12.3.28. Let D be a fixed dyadic system on Rd , in the sense of Def-

inition 11.1.6.
(1) For every ω = (ωj )j∈Zd ∈ ({0, 1}d )Z ,

D ω := Q+̇ω : Q ∈ D


is another dyadic system on Rd , where

X
Q+̇ω := Q + `(Q, ω), `(Q, ω) := 2−j ωj .
j:2−j <`(Q)

(2) Conversely, every dyadic system D 0 has this form for some ω ∈ ({0, 1}d )Z .

Proof. Let D 0 be the standard dyadic system, and consider a family of shifts
sj + Dj0 . These clearly satisfy property (i) of Definition 11.1.6. A necessary
and sufficient condition for them to satisfy (ii) of Definition 11.1.6 is that
sj − sj+1 ∈ 2−j−1 Zd .
If D is a dyadic system defined by shifts sj , then D ω is defined by the
shifts sj + ω(j) , where
168 12 Dyadic operators and the T (1) theorem
X
ω(j) := ωk 2−k .
k>j

These satisfy (sj +ω(j) )−(sj+1 −ω(j+1) ) = (sj −sj+1 )+ωj+1 2−j−1 ∈ 2−j−1 Zd ,
and hence D ω is also a dyadic system, as claimed in (1).
Then suppose that D and D 0 are two dyadic systems defined by shifts sj
and s0j , respectively. It is clear that the family Dj = sj + Dj0 only depends on
sj mod 2−j , and hence we may assume without loss of generality that both
sj ∈ [0, 2−j )d and tj := sj0 − sj ∈ [0, 2−j )d . Since both sj − sj+1 ∈ 2−j−1 Zd
and s0j − sj+1
0
∈ 2−j−1 Zd , it follows that also tj − tj+1 ∈ 2−j−1 Zd . Together
with the fact that tj ∈ [0, 2−j )d and tj+1 ∈ [0, 2−j−1 )d , one finds that in
fact tj − tj+1 ∈ 2−j−1 {0, 1}d . Denoting ωj+1 := 2j+1 (tj − tj+1 ) ∈ {0, 1}d , we
obtain X
tj = tj+1 + 2−j−1 ωj+1 = . . . = 2−k ωk = ω(j) ,
k>j

and then
Dj0 = s0j + Dj0 = tj + sj + Dj0 = ω(j) + Dj = Djω ,
as claimed in (2), and this completes the proof. 

Definition 12.3.29. For ω = (ωj )j∈Z ∈ ({0, 1}d )Z , let

jω := sup{j ∈ Z : ωj 6= 0} ∈ Z ∪ {−∞, ∞},

n o
({0, 1}d )Z0 := ω ∈ ({0, 1}d )Z : jω < ∞ .

We say that ω ∈ ({0, 1}d )Z is eventually zero if ω ∈ ({0, 1}d )Z0 .

Lemma 12.3.30. For every ω ∈ ({0, 1}d )Z0 , we have

S(D ω ) = S(D), S0 (D ω ) = S0 (D).

Moreover, there exists an ω ∈ ({0, 1}d )Z0 such that S00 (D ω ) = S0 (D).

Proof. Recall that S(D) is the span of indicators 1Q of Q ∈ D. Since every

Q ∈ Dj can be written as a union of smaller cubes Q0 ∈ Dk , for any S k > j, we
see that, for any given j0 ∈ Z, the space S(D) only depends on j>j0 Dj . On
the other hand, if ω is eventually zero, and jω is as in the definition of this
property, then Djω = Dj for j > jω . The first claimed identity thus follows.
The second identity follows by restricting to functions of vanishing integral
on both sides.
Finally, it is easy to choose ω ∈ ({0, 1}d )Z0 in such a way that D ω contains
an increasing sequence of cubes that exhausts all Rd . Then, given any f ∈
S(D), we can find some Q0 ∈ D ω that contains the support of f . If, in
addition, f ∈ S0 (D) = S0 (D ω ), then f can be expanded in terms of finitely
α
many Haar functions hQ with Q ⊆ Q0 , and thus f ∈ S00 (D ω ). Since this
holds for every f ∈ S0 (D), we obtain the final identity. 
 12.3 The T (1) theorem for abstract bilinear forms 169

Remark 12.3.31. Without the assumption of eventually zero, the conclusion

of Lemma 12.3.30 fails in general. For instance, the indicator of the shifted
dyadic interval 31 + [0, 1) cannot be expressed as a finite linear combination of
standard dyadic intervals.
Thanks to Lemma 12.3.30, any bilinear form t : S(D)2 → Z may also be
regarded as a bilinear form t : S(D ω )2 → Z for every eventually zero ω.
Although the objects in fact coincide, it will be convenient to denote the latter
by tω . This is particularly relevant when considering the various auxiliary
objects derived from the bilinear form. In particular, extending the notation
from Proposition 12.3.18, we have
γ
tω;α,γ
n (R) := t(hα
R , hR+̇n ), R = Q+̇ω ∈ D ω
(
tω;1,α (R)∗ := t(h0R+̇n , hα ∗
R ) , i = 1,
un (R) := nω;2,α
ω;i,α
∗
tn (R) := t(hα 0
R , hR+̇n ) , i = 2.
The advantage of considering several dyadic systems D ω is that this allows
us to dispense with some of the cubes within each D ω .
Definition 12.3.32. For a dyadic system D and k ∈ Z>2 , a cube Q ∈ D is
called k-good (in D) if
1
dist(R, {R(k) ) > `(R(k) ) = 2k−2 `(R),
4
where R(k) is the kth dyadic ancestor of R in D.
Lemma 12.3.33. Consider a random choice of ω ∈ ({0, 1}d )Z>M with respect
to the uniform probability on this space. For every Q ∈ D with `(Q) > 2−M ,
(1) the random set Q+̇ω and the event {Q+̇ω is k-good in D ω } are indepen-
dent;
(2) P(Q+̇ω is k-good in D ω ) = 2−d .
Proof. (1) follows by observing that Q+̇ω depends only on ωj with 2−M 6
2−j < `(Q), whereas {Q+̇ω is k-good in D ω } depends on the relative position
of Q+̇ω with respect to cubes R+̇ω with `(R) = 2k `(Q), which in turn depends
on ωj with `(Q) 6 2−j < 2k `(Q).
(2): When all ωj with `(Q) 6 2−j < 2k `(Q) are independently chosen
from {0, 1}d , it is easy to see that the probability of {Q+̇ω is k-good in D ω }
is equal to the geometric probability (i.e., the relative volume) of the “good
region”
n 1 o 1
Rgood := s ∈ R : dist(s, {R) > `(R) = R̄
4 2
of the D ω -ancestor R of Q, and this is simply
|Rgood | | 1 R̄|
= 2 = 2−d .
|R| |R|

170 12 Dyadic operators and the T (1) theorem

Definition 12.3.34. For θ ∈ {(α, γ), (i, α)}, and n ∈ Zd \ {0}, we define
ω;θ ω;θ
tn,good (R) := 1{R is k(n)-good in D ω } tn (R),

k(n) := 2 + dlog2 |n|e.

good i,good
We define Figiel’s operators Tn,t ω and Um,t ω as in Definition 12.3.16, but
ω;θ θ
with tn,good in place of the respective tn

For n ∈ Zd \ {0}, we have k(n) > 2, and hence the notion of “k(n)-good” is
well-defined. For n = 0 we would formally get k(0) = −∞, and “−∞-good”
reduces to the triviality dist(R, {R) > 0; accordingly, for definiteness, we let
ω;θ
t0,good (R) := t0ω;θ (R).

Replacing all quantities in Definition 12.3.22 by their “good” restrictions, we

have a natural definition of the Figiel norms

ktω;θ
good kFig (℘) ,
s θ ∈ {(α, γ), (i, α)},
ω;(i)
ktgood kFigs (℘) , i = 1, 2, ktω
good kFigs (℘) .

As we are about to see, these good parts will suffice to control a bounded
extension of the form t, and this also allows us to obtain a better dependence
on the UMD constants. Here is the precise statement:

Theorem 12.3.35 (T (1) theorem for bilinear forms, random version).

Let p ∈ (1, ∞) and 1 6 t 6 p 6 q 6 ∞, and consider the conditions:
(i) X and Y are UMD spaces,
(ii) X has cotype q and Y has type t, or one of them has both,
(iii) t : S(D)2 → L (X, Y ) is a bilinear form with

X 2
X
DR p (tω;α,γ
0 ) + min ktω;(0) kFigσi (Rp ) + ktω;(i) kFigσi (Rp ) 6 C,
i=1,2
α,γ∈{0,1}d \{0} i=1

uniformly in ω ∈ ({0, 1}d )0Z , where σ1 = 1/t and σ2 = 1/q 0 .

(iv) the forms tω satisfy the adjacent weak boundedness property ktω kawbp 6 C
uniformly in ω ∈ ({0, 1}d )Z0 ,
Under assumptions (i) through (iii), the following conditions are equivalent:
(1) t defines a bounded linear operator T ∈ L (Lp (Rd ; X), Lp (Rd ; Y ));
(2) t satisfies (iv) and the paraproducts Λtω are uniformly bounded.
Under these equivalent conditions, we have:
 12.3 The T (1) theorem for abstract bilinear forms 171

(a) the norm estimate:

kT kL (Lp (Rd ;X),Lp (Rd ;Y ))

n X
6 sup kΛtω kL (Lp (Rd ;X),Lp (Rd ;Y )) + βp,X βp,Y sup DR p (tω;α,γ
0 )
ω ω
α,γ
 2 o
ω;(0) ω;(i)
X
+ 12 · 2d sup min ci ktgood kFigσi (℘i ) + ci ktgood kFigσi (℘i ) ,
ω i=1,2
i=1

where the suprema are over ω ∈ ({0, 1}d )0Z , and

℘1 := Rp∗0 , ℘2 := Rp , σ1 := 1/t, σ2 := 1/q 0 ,

(12.51)
c1 := min ct0 ,Z ∗ ;p0 , c2 := min cq,Z;p ;
Z=X,Y Z=X,Y

(b) the representation formula

 X n good
hT f, gi = E hHtω f, gi + hΛtω f, gi + 2d hTn,tω f, gi+
n∈Zd
n6=0 (12.52)
o
1,good 2,good
+ hf, Un,t ω gi + hUn,t ω f, gi ,

with absolute convergence for all f ∈ S(D; X) and g ∈ S(D; Y ∗ ), where E

is the expectation over ω ∈ ({0, 1}d )Z6M , and M ∈ Z is any large enough
number such that f and g are constant on all Q ∈ DM .

Proof. We begin by observing that, according to Lemma 12.3.30, assumptions

(i) through (iii) of the present theorem imply assumption (i) through (iii) of
Theorem 12.3.26 uniformly for every ω ∈ ({0, 1}d )Z0 . Thus the qualitative
statement (1)⇔(2) is just an application of Theorem 12.3.26 to each D ω in
place of D, observing the uniformity just mentioned.
The more interesting part consist of the new quantitative conclusions that
we obtain for the implication (2)⇒(1). This requires revisiting some details
of the proof of Theorem 12.3.26.
Let f ∈ S(D; X) and g ∈ S(D; Y ∗ ), and let us specifically assume that
both f and g are constant on all Q ∈ DM for some (in general large) M ∈ Z.
We identify ({0, 1}d )Z6M with {ω = (ωj )j∈Z ∈ ({0, 1}d )Z : ωj = 0 for j > M }.
For each ω ∈ ({0, 1}d )Z6M , we have DMω
= DM , and hence f and g have
the same piecewise constancy property with respect to these dyadic systems.
For each m 6 M and ω ∈ ({0, 1}d )Z6M , we then write an analogue of (12.46),
ω
t(f, g) = hHtω um , gi + ltω (f, g) + Em
ω
(f, g)+
X n o
ω 1 ω 2 ω
+ hTn,tω um , gi + hf, Un,t ω vm i + hUn,tω um , gi , (12.53)
d
n∈Z
n6=0
172 12 Dyadic operators and the T (1) theorem

where all symbols have the same meaning as in (12.46), but with D ω in place
of D. In particular,

uω ω
m = (I − Em )f,
ω
vm ω
= (I − Em )g,
ω ω
where Em = E( |Dm ) satisfy kuω ω
m kp 6 2kf kp and kvm kp0 6 2kgkp0 .
The first and third terms on the right of (12.53) are estimated as in the
proof Theorem 12.3.26. As in (12.47), we have
 2
X 
ω
|Em (f, g)| 6 cd,p ktω;(i) kFig0 (∞) + 2d ktω kawbp kEm
ω ω
f kp kEm gkp0 → 0
i=1

when m → −∞; note that this convergence is bounded by (iii), (iv), and the
ω ω
easy estimates kEm f kp 6 kf kp and kEm gkp0 6 kgkp0 . Then, as in (12.48),
from Theorem 12.1.11 we get
X
|hHtω uω
m , gi| 6 βp,X βp,Y DR p (t0ω;α,γ )kum
ω
kp kgkp0 .
α,γ

The second term on the right of (12.53) is directly estimated by the uniform
boundedness of the paraproducts Λtω .
We then turn to the more interesting part on the second line of (12.53),
where we begin with some observations. Due to the presence of the truncation
parameter m, all dyadic operators in (12.53) involve cubes of side-length at
most 2−m . On the other hand, due to the constancy of f and g on Q ∈ DM =
DM ω
, their martingale differences are non-zero only on cubes of side-length
strictly larger than 2−M . Hence the right-hand side of (12.53) actually depends
on (ωj )m<j6M only, rather than the infinite sequence (ωj )j6M . Nevertheless,
it will be convenient to also refer to this latter sequence, as we are about to
see.
We compute the expectation of (12.53) with respect to the choice of ω ∈
({0, 1}d )Z6M . As we just observed, this is actually just an arithmetic average
over a finite set of 2d(M −m) elements, so no integrability or measurability
issues arise at this point.
We wish to manipulate this average a little. We note that each of the terms
on the second line of (12.53) take the generic form
∗
X
Φ(Q+̇ω),
Q∈D

where
nXD E
γ γ
Φ(R) ∈ t(hα α
R , hR+̇n )hf, hR i, hg, hR+̇n i ,
α,γ
XD E
α
t(hR , h0R+̇n )hf, hα
R i, hg, h 0
R+̇n − h 0
R i ,
α
 12.3 The T (1) theorem for abstract bilinear forms 173
XD γ γ
Eo
0 0 0
t(hR +̇n , hR )hf, hR+̇n − hR i, hg, hR i ,
γ
P∗
and the notation suppresses not only the size condition that 2−M <
−m
`(Q) 6 2 but also an implicit restriction to a fixed finite family of cubes of
each size, depending on the supports of f and g.
Inserting 1 = 2d · E(1{Q+̇ω is k-good} ), it hence follows, using in particular
the independence property established in Lemma 12.3.33(1), that
∗
X ∗
X
E Φ(Q+̇ω) = 2d · E(1{Q+̇ω is k-good} )EΦ(Q+̇ω)
Q∈D Q∈D
∗
X
= 2d


E 1{Q+̇ω is k-good} Φ(Q+̇ω)
Q∈D
∗
X
= 2d · E Φ(Q+̇ω).
Q∈D:
Q+̇ω is k-good

Thus, at the cost of the factor 2d , we can reduce the summation to k-good
cubes only.
Taking the expectation of (12.53) and applying the above observation to
the terms on the second line, with k = k(n) as in Definition 12.3.34, we obtain

t(f, g) = E hHtω um ω
, gi + ltω (f, g) + Emω
(f, g)+
n o
(12.54)
X good ω 1,good ω 2,good ω
+ 2d hTn,t ω um , gi + hf, Un,tω vm i + hUn,t ω um , gi ,
n∈Zd
n6=0

where the various “good” operators are defined in Definition 12.3.34.

When k = k(n) is as in Definition 12.3.34, and R = Q+̇ω is k-good, it
follows directly from Definition 12.3.32 that

dist(R, {R(k,ω) ) > 2k−2 `(R) > |n|`(R),

and hence R+̇n ⊆ R(k,ω) . Thus the operators on the right of (12.54) are in the
scope of the sharper special cases of Figiel’s estimates, Corollary 12.1.27(2)
and Theorem 12.1.28(2).
An application of these estimates to (12.54), in the case of Unω,1 on the dual
side and otherwise directly as in Corollary 12.1.27(2) and Theorem 12.1.28(2),
gives
1,good ω
X ω;1,α
|hf, Un,t ω vm i| 6 6βp,X βp,Y c1 (1 + k(n))σ1 ℘1 (tn,good ω
)kf kp kvm kp 0 ,
α
2,good ω
X
|hUn,t ω um , gi| 6 6βp,X βp,Y c2 (1 + k(n))σ2 ℘2 (tω;2,α ω
n,good )kum kp kgkp ,
0

α
174 12 Dyadic operators and the T (1) theorem
good ω
X ω;α,γ
|hTn,t ω um , gi| 6 3βp,X βp,Y min ci (1 + k(n))σi ℘i (tn,good ω
)kum kp kgkp0 .
i=1,2
α,γ

It follows from Definition 12.3.34 that

k(n) + 1 6 4 + log2 |n| 6 2(2 + log2 |n|),
and hence
ω;(i)
X X
(1 + k(n))σi ℘i (tω;i,α
n,good ) 6 2ktgood kFig i (℘i ) ,
σ

n∈Zd α
n6=0
ω;(0)
X X
(1 + k(n))σi ℘i (tω;α,γ
n,good ) 6 2ktgood kFig i (℘i ) .
σ

n∈Zd α,γ
n6=0

We have thus estimated all terms on the right of (12.54). Let us further
recall that kuω ω p d
m kp 6 2kf kp and um → f in L (R ; X) as m → −∞, with
ω 0 ω
similar results for vm , g and p in place of um , f and p. We can thus pass to
the limit m → −∞ in (12.54) and apply dominated convergence to deduce the
claimed representation formula (12.52). Applying the same estimates above
to (12.52) in place of (12.54), we deduce the claimed norm estimate (a). This
completes the proof of Theorem 12.3.35. 

12.4 The T (1) theorem for singular integrals

A natural question arising from the Theorems 12.3.26 and 12.3.35 above is
whether their assumptions are verified by some familiar operators. In partic-
ular, what is the relation of these conditions to the Calderón–Zygmund oper-
ators discussed in Chapter 11? We will address this question in the present
section. Recall from Definition 11.3.1 that
cK := sup{|s − t|d kK(s, t)k : (s, t) ∈ Ṙ2d }.
Definition 12.4.1 (Weakly defined singular integral operator). Let Z
be a Banach space, and C be a collection of bounded Borel subsets of Rd . We
say that a bilinear form t : S(C )2 → Z is a weakly defined singular integral
with associated kernel K : Ṙ2d → Z, if cK < ∞ and
ZZ
t(1Q , 1R ) = K(s, t)1Q (t)1Q (s) ds dt (12.55)
R2d

whenever Q, R ∈ C are disjoint.

As usual, the main case of interest will be C = D.
The following lemma, which will also play a role later, shows that the
integral in (12.55) is well defined under the assumption that cK < ∞: While
in (12.55) we do not require the cubes to have equal size, we can always
dominate the integral with such a case by passing to a dyadic ancestor of the
smaller cube, if necessary.
 12.4 The T (1) theorem for singular integrals 175

Lemma 12.4.2. For disjoint cubes Q, R ⊆ Rd of equal size `(Q) = `(R), we

have ZZ
1 dvd
d
ds dt 6 (1 + )|Q| < 18 · |Q|,
Q×R |s − t| 2
where vd is the volume of the unit ball in Rd .

Proof. We first write

ZZ ZZ Z ∞
1
d
ds dt = d r−d−1 dr ds dt
Q×R |s − t| Q×R |s−t|
Z ∞
=d |{(s, t) ∈ Q × R : |s − t| < r}|r−d−1 dr.
0

Denoting by vd is the volume of the unit ball in Rd , we have

Z
|{(s, t) ∈ Q × R : |s − t| < r}| = |{t ∈ R : |s − t| < r}| ds
{s∈Q:dist(s,R)<r}
d
vd r |Q| vd rd
6 |{s ∈ Q : dist(s, R) < r}|( ∧ |R|) 6 (r ∧ `(Q)) ( ∧ |R|),
2 `(Q) 2

where we used the geometric observation that, for s ∈ Q ⊆ {R, at least half
of any ball of centre s lies in {R. Hence
`(Q)
|Q| vd rd −d−1
ZZ Z
1
ds dt 6 d r · ·r dr
Q×R |s − t|d 0 `(Q) 2
Z ∞
dvd
+d |Q| · |R|r−d−1 dr = |Q| + |R|,
`(Q) 2

where |R| = |Q|, since `(R) = `(Q).

Finally, dvd /2 = π d/2 /Γ (d/2) =: f (d/2). From the functional equation
Γ (x + 1) = xΓ (x), we find that f (x + 1)/f (x) = π/x, so that max{f (n) :
n ∈ N} = f (4) and max{f (n + 12 ) : n ∈ N} = f (7/2). Computing these two
8 3
values, one checks that max{f (d/2) : d ∈ N} = f (7/2) = 15 π < 17. 

For weakly defined singular integrals, some properties imposed as assumptions

on general bilinear forms are automatically satisfied:

Lemma 12.4.3. Let Z be a Banach space and t : Ṙ2d → Z a weakly defined

singular integral operator with kernel K. Then t satisfies the adjacent weak
boundedness property if and only if it satisfies the weak boundedness property,
and moreover

ktkwbp 6 ktkawbp 6 max{ktkwbp , 18 · cK }.

176 12 Dyadic operators and the T (1) theorem

Proof. The “only if” part is obvious. For “if”, it suffices to estimate t(1Q , 1R )
for R = Q+̇n and n ∈ {−1, 0, 1}d \ {0}. Then Q ∩ R = ∅, so that we have
access to the kernel representation (12.55), and Lemma 12.4.2 provides us
with the bound
ZZ
cK
kt(1Q , 1R )k 6 d
ds dt 6 18 · |Q| · cK .
Q×R |s − t|

Proposition 12.4.4. Let Z be a Banach space and t : Ṙ2d → Z a weakly

defined singular integral operator. If t is translation-invariant (in the sense of
Definition 12.3.9), then t(1, ·) = 0 = t(·, 1).

Proof. By Proposition 12.3.10, it suffices to verify that t satisfies the decay

condition (12.35). Let Q ∈ D and m ∈ Zd \ {−1, 0, 1}d . Then, for s ∈ Q and
t ∈ Q+̇m, and denoting by zQ the centre of Q, we have

|s − t| > |s − t|∞ > |m`(Q)|∞ − |s − zQ |∞ − |t − (zQ + m`(Q))|∞

1 1 1 |m|`(Q)
> |m|∞ `(Q) − `(Q) − `(Q) > |m|∞ `(Q) > √ ,
2 2 2 2 d
and hence
Z Z
cK
kt(1Q , 1Q+̇m )k 6 ds dt
Q |s − t|d
Q+̇m
 2√d d √
2
6 |Q| cK = |Q|cK (2 d)d |m|−d .
|m|`(Q)

This is one half of the decay condition (12.35). The estimate for t(1Q+̇m , 1Q )
is entirely similar. 
Despite the simple observations above, in order to make serious conclusions
about weakly defined singular integrals, we will need the following elaboration
of the earlier Definition 11.3.1:

Definition 12.4.5 (℘–Calderón–Zygmund kernel). Let Z be a Banach

space, ℘ a good set-bound on Z, and K : Ṙ2d → Z. We define the quantities

cK (℘) := ℘({|s − t|d K(s, t) : s 6= t}),

and, for u ∈ [0, 21 ],

n o
1
(℘; u) := ℘ |s − t|d K(s, t) − K(s0 , t) : |s − s0 | 6 u|s − t| ),


ωK (12.56)
n o
2
|s − t|d K(s, t) − K(s, t0 ) : |t − t0 | 6 u|s − t| ).


ωK (℘; u) := ℘ (12.57)
 12.4 The T (1) theorem for singular integrals 177

Remark 12.4.6. (1) We recover Definition 11.3.1 by taking ℘(T ) = U (T ) :=

sup{kT k : T ∈ T }. Our main interest now will be ℘ ∈ {Rp , Rp∗0 }.
(2) In analogy with Lemma 11.3.3, one can check that

i 1
ωK (℘; ) 6 (1 + 2d )cK (℘).
2
(3) If K(s, t) = K(s − t) for some K : Rd \ {0} → Z, then

cK (℘) = ℘({|s|d K(s) : s 6= 0}) =: c̃K (℘),

and, for both i ∈ {1, 2},

n o
i
(℘; u) = ℘ |s|d K(s) − K(s0 ) : |s − s0 | 6 u|s| ) =: ω̃K (℘; u).


ωK

Such a K (or K) is referred to as a convolution kernel.

If t is a weakly defined singular integral with ℘–Calderón–Zygmund kernel K,

the conditions of Definition 12.4.5 only provide control away from the diagonal
s = t. To compensate for this, we also need the following assumption directly
on the bilinear form t:

Definition 12.4.7 (Weak DR p -boundedness property). Letting t :

S(D)2 → L (X, Y ) be a bilinear form, we define
n t(1 , 1 ) o 
Q Q
ktkwbp(DRp ) := DR p
|Q| Q∈D

Our goal in this section will be to use these assumptions to control the Haar
γ
coefficients t(hα
Q , hR ), where R = Q + `(Q)n, in the way that was assumed
in the Theorems 12.3.26 and 12.3.35 on bilinear forms. Using the defining
condition (12.55) and bilinearity (noting that hα Q is a linear combination of
1Q0 for Q0 ∈ ch(Q), and likewise hγR ), we have in particular that
ZZ
γ
t(hαQ , h R ) = K(s, t) ds dt, Q ∩ R = ∅.
Q×R

If K is a ℘-Calderón–Zygmund kernel, we can establish the following esti-

mates:

Lemma 12.4.8. Let Z be a Banach space and ℘ a good set-bound on Z. Let

t : S(D)2 → Z be a weakly defined singular integral with kernel K : Ṙ2d → Z.
Then for all α, γ ∈ {0, 1}d , we have, for all n ∈ Zd \ {0},
n o
γ
℘ t(hαQ , hQ+̇n ) : Q ∈ D 6 18 · 2 · cK (℘),
d
(12.58)

3
√
and, for |n| > 2 d,
178 12 Dyadic operators and the T (1) theorem
3
√
3 d 4 d
n o
γ
℘ t(hα : Q ∈ D 6 ( ) · |n| · ωK (℘;
Q+̇n , hQ )
−d 1
) if γ 6= 0. (12.59)
2 |n|
3
√
3 d 4 d
n o
γ
℘ t(hQ , hQ+̇n ) : Q ∈ D 6 ( ) · |n| · ωK (℘;
α −d 2
) if α 6= 0, (12.60)
2 |n|
Proof. Including momentarily also n = 0 for later use, we have the expansion
γ
X γ
t(hα
Q , hQ+̇n ) = t(1R , 1S )hhα
Q iR hhQ+̇n iS
R∈ch(Q)
S∈ch(Q+̇n)
X γ
= δn,0 t(1R , 1R )hhα
Q iR hhQ+̇n iS
R∈ch(Q)
(12.61)
X γ
+ t(1R , 1S )hhα
Q iR hhQ+̇n iS =: IQ + IIQ .
R∈ch(Q)
S∈ch(Q+̇n)
R6=S

(The summation condition R 6= S in IIQ is automatic for n 6= 0, but it makes

no harm to include it). Since
X γ
X 1 X
|R||hhα
Q iR hhQ+̇n iS | = |R| = 1 = 2d ,
|Q|
R∈ch(Q) R∈ch(Q) S∈ch(Q+̇n)
S∈ch(Q+̇n) S∈ch(Q+̇n)

we see that
n t(1 , 1 ) o
U V
IIQ ∈ 2d abs conv : U, V ∈ D, U ∩ V = ∅, `(U ) = `(V ) ,
|U |

where
ZZ ZZ
ds dt
t(1U , 1V ) = K(s, t) ds dt = |s − t|d K(s, t)
U ×V U ×V |s − t|d
n o
∈ 18 · |U | · abs conv |u − v|d K(u, v) : (u, v) ∈ Ṙ2d ,

by Proposition 1.2.12 and Lemma 12.4.2 in the last step. Combining the above
inclusions with the defining properties of good set-bounds (Definition 12.3.20),
we obtain
℘({IIQ : Q ∈ D}) 6 18 · 2d · cK (℘), (12.62)
which coincides with (12.58) when n 6= 0.
For large values of n, we want to obtain a decay, which is not present
in the uniform estimate just established. In this case we apply the kernel
representation combined with the vanishing mean of hαQ (when α 6= 0), to the
result that
 12.4 The T (1) theorem for singular integrals 179
ZZ
α γ α γ
t(hQ , hQ +̇n
)= K(s, t)hQ (t)hQ +̇n
(s) ds dt
ZZ
γ
= [K(s, t) − K(s, zQ )]hα
Q (t)hQ+̇n (s) ds dt,

where
√ zQ is the centre of Q. For t ∈ Q and s ∈ Q+̇n, we have |t − zQ | 6
1
2 d`(Q), whereas

1√
|s − zQ | > |zQ+̇n − zQ | − |s − zQ+̇n | > (|n| − d)`(Q),
2
and hence √
|t − zQ | 1
2 d 1 3√
6 √ 6 if |n| > d.
|s − zQ | 1
|n| − 2 d 2 2
In this case we have
ZZ
γ 1
α
t(hQ , hQ+̇n ) ∈ |hα (t)hγQ+̇n (s)| ds dt
|s − zQ |d Q
1
√
n d o
× abs conv |u − v|d [K(u, v) − K(u, v 0 )] : |v − v 0 | 6 2
√ |u − v| ,
|n| − 21 d

and hence, by estimate (12.56) of a Calderón–Zygmund kernel (Definition

12.4.5) and the defining properties of good set-bounds (Definition 12.3.20),
we arrive at
n o
γ
℘ t(hα Q , hQ+̇n ) : Q ∈ D

1
ZZ
1  1 √d 
2 2
6 d
ds dt × ωK √
|Q| Q×(Q+̇n) |s − zQ | |n| − 21 d
1  1 √d  3 d  3 √d 
2 2 −d 2
6 √ ωK √ 6 ( ) · |n| ωK 4
(|n| − 21 d)d |n| − 21 d 2 |n|
√
when |n| > 32 d.
γ
The estimate of t(hαQ , hQ+̇n ) with γ 6= 0 is entirely analogous to this, using
regularity in the other variable instead. 
Concerning the diagonal n = 0, which was excluded in Lemma 12.4.8, we have
the following estimate:

Lemma 12.4.9. Let X and Y be Banach spaces and p ∈ (1, ∞). Let t : Ṙ2d →
L (X, Y ) be a weakly defined singular integral with the weak DR p -boundedness
property. Then
γ
DR p ({t(hα d
Q , hQ )}Q∈D ) 6 ktkwbp(DRp ) + 18 · 2 · cK (℘), ℘ ∈ {Rp , Rp∗0 }.
180 12 Dyadic operators and the T (1) theorem

Proof. We use the expansion (12.61) with n = 0,

α γ
t(hQ , hQ ) = IQ + IIQ ,

where we now need to consider also the term IQ . We estimate the expression
in the definition of DR p ({IQ }Q∈D ):
X X X γ
∗ ∗
|Q||hIQ xQ , yQ i| 6 |Q| |ht(1R , 1R )xQ , yQ i||hhα
Q iR hhQ iR |
Q∈D Q∈D R∈ch(Q)
X X
∗
= |ht(1R , 1R )xQ , yQ i|
Q∈D R∈ch(Q)
X
∗
= |ht(1R , 1R )xR(1) , yR (1) i|

R∈D
X
6 ktkwbp(DRp ) εR xR(1) 1R ×
Lp (Ω×Rd ;X)
R∈D
X
∗
× εR yR (1) 1R .
Lp0 (Ω×Rd ;Y ∗ )
R∈D

Using the usual observation that, by Fubini’s theorem and the fact that only
one R ∈ D of each generation is “seen” at each fixed s ∈ Rd , we can replace
the random εR by εn(R) depending on the generation of R only, or further by
the equidistributed sequence of εn(R(1) ) , we have
X X X
εR zR(1) 1R = εn(Q) zQ 1R
Lp (Ω×Rd ;Z) Lp (Ω×Rd ;Z)
R∈D Q∈D R∈ch(Q)
X X
= εn(Q) zQ 1Q = εQ zQ 1Q
Lp (Ω×Rd ;Z) Lp (Ω×Rd ;Z)
Q∈D Q∈D

∗
for both choices of zQ ∈ {xQ , yQ } and Z ∈ {X, Y }. Hence

DR p ({IQ }Q∈D ) 6 ktkwbp(DRp ) ,

and hence, by the obvious triangle inequality for DR p , and its domination by
either ℘ ∈ {Rp , Rp∗0 } according to Lemma 12.1.8, we have

DR p ({t(hQ
α
, hγQ )}Q∈D ) 6 DR p ({IQ }Q∈D ) + DR p ({IIQ }Q∈D )
6 ktkwbp(DRp ) + ℘({IIQ }Q∈D )
6 ktkwbp(DRp ) + 18 · cK (℘)

by (12.62) in the last step. 

We can now give estimates for the Figiel norms featuring in the T (1) theorems
for bilinear forms:
 12.4 The T (1) theorem for singular integrals 181

Lemma 12.4.10. Let Z be a Banach space and ℘ a good set-bound on Z. Let

t : S(D)2 → Z be a weakly defined singular integral with kernel K : Ṙ2d → Z.
Then for all s ∈ [0, 1], we have the estimates

kt(0) kFigs (℘) , kt(i) kFigs (℘) 6 ad cK (℘) + bd kωK

i
(℘)kDinis , i = 1, 2,

where ad , bd depend only on the dimension d, and

Z 1/2
1 s du
kωkDinis := ω(u)(log2 ) . (12.63)
0 u u

Remark 12.4.11. For u ∈ (0, 21 ), we have u1 ∈ (2, ∞), thus log2 1

u ∈ (1, ∞).
Hence (log2 u1 )s and therefore kωkDinis are increasing in s.

Proof of Lemma 12.4.10. From Definition 12.3.22 and Lemma 12.4.8, it fol-
lows that
X
kt(0) kFigs (℘) = ktαγ kFigs (℘) ,
α,γ∈{0,1}d \{0}
X X γ
= Q , hQ+̇n ) : Q ∈ D})
(2 + log2 |n|)s ℘({t(hα
α,γ∈{0,1}d \{0} n∈Zd
n6=0
n X √
6 (2d − 1)2 (2 + log2 (3 d)) · 18 · 2d · cK (℘)+ (12.64)
√
|n|<3 d
√
X 3 d −d i  34 d o
s
+ (2 + log2 |n|) ( ) |n| ωK ℘;
√ 2 |n|
|n|>3 d

=: (2d − 1)2 (I + IIi ) 6 4d (I + IIi ).

Since both α 6= 0 6= γ, one can apply either of the estimates (12.59) or (12.60)
of Lemma 12.4.8, and thus take either i ∈ {1, 2} above. Similarly,
X
kt(2) kFigs (℘) = kt2,α kFigs (℘) ,
α∈{0,1}d \{0}
X X
Q , hQ+̇n ) : Q ∈ D})
(2 + log2 |n|)s ℘({t(hα 0
= (12.65)
α∈{0,1}d \{0} n∈Zd
n6=0

6 (2d − 1)(I + II2 ) 6 4d (I + II2 ),

where we only have access to estimate (12.60), but not (12.59), of Lemma
12.4.8, now that the second Haar function h0Q+̇n is non-cancellative. The very
last step in (12.65) is of course wasteful, but we make it in order to treat the
right-hand sides of both (12.64) and (12.65) at the same time.
Finally, in complete analogy with (12.65), we also have
182 12 Dyadic operators and the T (1) theorem
X
kt(1) kFigs (℘) = kt1,α kFigs (℘) 6 4d (I + II1 ), (12.66)
α∈{0,1}d \{0}

as we now have access to estimate (12.59), but not (12.60), of Lemma 12.4.8.
It is immediate that
X √
4d I = ad · cK (℘), ad := 4d (2 + log2 (3 d)) · 18 · 2d . (12.67)
√
|n|<3 d

For the other term, we partition the summation over dyadic annuli, in which
the summand is roughly a constant:
∞
X X √ √
4d IIi 6 6d (2 + log2 (3 d) + k)s (3 · 2k d)−d ωK
i
(℘; 2−k−2 ).
√
k=0 3·2k d6|n|
√
<3·2k+1 d
√
The unit-cubes Qn with centres n ∈ Zd are √ disjoint, and for |n| < 3 · 2k+1 d,
they are contained in B(0, (3 · 2k+1 + 21 ) d). Thus
X  1 √ d √ d
1 6 vd (3 · 2k+1 + ) d 6 vd (6.5)d 2kd d , (12.68)
√ 2
|n|<3·2k+1 d

where vd is the volume of the unit ball, and hence

∞
X √ d √ √
4d IIi 6 6d vd (6.5)d 2kd d (2 + log2 (3 d) + k)s (3 · 2k d)−d ωK
i
(℘; 2−k−2 )
k=0
√ ∞
X
d
6 (13) vd (2 + log2 (3 d)) (1 + k)s ωK
i
(℘; 2−k−2 ).
k=0

i
Since ωK (℘; u) is non-decreasing, we can finally estimate
Z 2−k−1 i
1 1 s ωK (℘; u) du
(1 + k) s i
ωK (℘; 2−k−2 ) 6 (log2 ) , k = 0, 1, . . . ,
log 2 2−k−2 u log 2 u
and hence
√
d i (13)d vd (2 + log2 (3 d))
4 IIi 6 bd kωK (℘)kDinis , bd := .
log 2
With (12.64), (12.65), (12.66), and (12.67), this concludes the proof. (An
estimate similar to (12.68) could also be used to give a more explicit bound
for the constant ad in (12.67), if desired.) 
We have now everything prepared for proving the following:
Theorem 12.4.12 (T (1) theorem for operator-valued kernels). Let
p ∈ (1, ∞) and 1 6 t 6 p 6 q 6 ∞, and suppose that:
 12.4 The T (1) theorem for singular integrals 183

(i) X and Y are UMD spaces.

(ii) X has cotype q and Y has type t, or one of them has both.
(iii) t : S(D)2 → Z := L (X, Y ) is a weakly defined singular integral and the
kernel K : Ṙ2d → Z of t satisfies the Calderón–Zygmund estimates
1 2
cK (Rp ) + kωK (Rp )kDini1/t + kωK (Rp )kDini1/q0 < ∞. (12.69)

Then the following conditions are equivalent:

(1) t defines a bounded operator T ∈ L (Lp (Rd ; X), Lp (Rd ; Y ));
(2) t satisfies the weak DR p -boundedness property ktkwbp(DRp ) < ∞, and the
associated bi-paraproduct Λt is bounded in L (Lp (Rd ; X), Lp (Rd ; Y ));
(3) each tω satisfies the weak DR p -boundedness property ktω kwbp(DRp ) 6
C, and the associated bi-paraproduct Λtω defines a bounded operator in
L (Lp (Rd ; X), Lp (Rd ; Y )), uniformly in ω ∈ ({0, 1}d )0Z .
Under these equivalent conditions, we have
(a) the first norm estimate:

kT − Λt kL (Lp (Rd ;X),Lp (Rd ;Y ))

n  
6 βp,X βp,Y 4d ktkwbp(DRp ) + cd C1 cK (Rp∗0) + C2 cK (Rp ) +
 o
+ c0d C1 kωK 1
(Rp∗0)kDini1/t + C2 kωK
2
(Rp )kDini1/q0 ,

where cd , cd0 are constants that depend only on d, and

C1 := C(12.15) (Y ∗ , X ∗ , p0 , t0 , 1), C2 := C(12.15) (X, Y, p, q, 1);

(b) the second norm estimate:

kT kL (Lp (Rd ;X),Lp (Rd ;Y )) − sup kΛtω kL (Lp (Rd ;X),Lp (Rd ;Y ))
ω
n  
6 βp,X βp,Y 4 sup kt kwbp(DRp ) + c0d c1 cK (Rp∗0) + c2 cK (Rp ) +
d ω
ω
 o
+ cd c1 kωK (Rp∗0)kDini1/t + c2 kωK
1 1 2
(Rp )kDini1/q0 ,

where the suprema are over ω ∈ ({0, 1}d )Z0 , the constants cd , c0d depend
only on d, and

c1 := min ct0 ,Z ∗ ;p0 , c2 := min cq,Z;p ; (12.70)

Z=X,Y Z=X,Y

(c) the representation formulas (12.45) and (12.52).

Proof. The plan of the proof is to reduce the theorem at hand to Theorems
12.3.26 and 12.3.35 on abstract bilinear forms.
(1)⇔(2): This will be an application of Theorem 12.3.26 (and Remark
12.3.27). Assumption (i) is identical in both theorems. Next, as explained in
184 12 Dyadic operators and the T (1) theorem

Remark 12.3.27, under the (co)type assumption (ii) of Theorem 12.4.12, the
assumption (ii) of Theorem 12.3.26 are satisfied with

(t1 , q1 ) := (t, ∞), (t2 , q2 ) := (1, q),

2
and both choices of (t0 , q0 ) ∈ {(ti , qi )}i=1 . Let σ1 := 1/t and σ2 := 1/q 0 .
Concerning assumption (iii) on the bilinear form t, we need to check that
the kernel assumptions (12.69) of the present theorem imply the assumptions
γ
on the Haar coefficients t(hα Q , hR ) and the related Figiel norms of the bilinear
form t. With the choices of (ti , qi ) as just explained, and recalling that the
set-bounds ℘1 := Rp∗0 and ℘2 := Rp are equivalent in the spaces that we are
considering, the assumption (12.69) can be equivalently written as
i
cK (℘i ) + kωK (℘i )kDiniσi < ∞, i ∈ {1, 2}. (12.71)

By Example 12.1.10, we know that

ktkwbp(DRp ) 6 kT kL (Lp (Rd ;X),Lp (Rd ;Y )) , (12.72)

so in particular the weak DR p -boundedness property is either assumed, or

implied by the assumptions, in each case of Theorem 12.4.12.
From Lemma 12.4.9, we then have
γ
DR p ({t(hα
Q , hQ )}Q∈D ) 6 ktkwbp(DRp ) ,

whereas Lemma 12.4.10 guarantees, for both i ∈ {1, 2}, that

i
kt(0) kFigσi (℘i ) 6 ad cK (℘i ) + bd kωK (℘i )kDiniσi ,
(12.73)
i
kt(i) kFigσi (℘i ) 6 ad cK (℘i ) + bd kωK (℘i )kDiniσi ,

where both right-hand sides of are finite by (12.71). With either choice of
(t0 , q0 ) ∈ {(ti , qi )}2i=1 , the resulting finiteness of the left-hand sides coincides
with the assumption on these quantities in (iii) of Theorem 12.3.26.
Summarising, assumptions (i) through (iii) of Theorem 12.4.12, together
with the weak DR p -boundedness property of t, which is either assumed or
implied by the assumptions of each case of Theorem 12.4.12, imply the corre-
sponding assumptions (i) through (iii) of Theorem 12.3.26. Moreover, the con-
dition of adjacent weak boundedness property appearing in Theorem 12.3.26
also follows from these assumptions by Lemma 12.4.3 and the domination of
uniform bounds by either DR p -bounds or ℘i -bounds:

ktkawbp 6 max{ktkwbp , 18 · cK } 6 max{ktkwbp(DRp ) , 18 · cK (℘i )}.

Hence all assumptions, and thus all conclusions of Theorem 12.3.26 are valid
under the assumptions of Theorem 12.4.12. This proves in particular the qual-
itative equivalence (1)⇔(2).
 12.4 The T (1) theorem for singular integrals 185

(a): For this quantitative estimate, we apply Remark 12.3.27, followed by

(12.72) and (12.73), to get

kT − Λt kL (Lp (Rd ;X),Lp (Rd ;Y ))

nX 2
X  o
α,γ
6 βp,X βp,Y DR p (t0 ) + Ci Ad kt(0) kFigσi (℘i ) + Bd kt(i) kFigσi (℘i )
α,γ i=1
n 2
X  o
6 βp,X βp,Y 4d ktkwbp(DRp ) + Ci cd cK (℘i ) + c0d kωK
i
(℘i )kDiniσi ,
i=1

where cd := (Ad + Bd )ad and c0d := (Ad + Bd )bd . This is readily recognised to
coincide with the bound asserted in (a) of the theorem.
(1)⇔(3): This will be an application of Theorem 12.3.35. Assumptions (i)
and (ii) are identical in both theorems.
Concerning assumption (iii), we need to check that the kernel assumptions
(12.69) of the present theorem imply the estimates on Figiel norms of each
bilinear form tω , uniformly in ω ∈ ({0, 1}d )0Z . We already did this for t = t0
above. However, all the lemmas of this section are stated for an arbitrary
dyadic system D, so we may in particular use them with any D ω in place of
D. Moreover, the constants in these estimates are explicit, and clearly inde-
pendent of the particular ω. This proves the qualitative equivalence (1)⇔(3).
(b): For this quantitative estimate, we apply Theorem 12.3.35(a), followed
by (12.72) and (12.73) with tω and D ω in place of t and D, to get
  1
kT kL (Lp (Rd ;X),Lp (Rd ;Y )) − sup kΛtω kL (Lp (Rd ;X),Lp (Rd ;Y ))
ω βp,X βp,Y
2
ω;(j)
X X X
6 sup DR p (tω;α,γ
0 ) + 12 · 2d sup ci ktgood kFigσi (℘i )
ω ω
α,γ i=1 j∈{0,i}
2
X  
6 4d sup ktω kwbp(DRp ) + sup ci c0d cK (℘i ) + c1d kωK
i
(℘i )kDiniσi ,
ω ω
i=1

where c0d = 24 · 2d · ad and c1d = 24 · 2d · bd . This is readily recognised to coincide

with the bound asserted in (a) of the theorem.
(c): The representation formulas are immediate from Theorems 12.3.26 and
12.3.35, since we already verified that the assumptions of the said theorems
are valid in the present setting. 

12.4.a Consequences of the T (1) theorem

We will now explore various consequences of Theorem 12.4.12 to more par-

ticular classes of operators. While Theorem 12.4.12 gives a complete charac-
terisation of the boundedness of an operator T , a drawback is the fact that
186 12 Dyadic operators and the T (1) theorem

this characterisation involves the boundedness of another operator Λt that is

not necessarily easy to check, as we found in Section 12.2. Thus, the follow-
ing special case, in which these paraproducts are completely avoided, will be
useful:

Corollary 12.4.13 (T (1) theorem for convolution kernels). Let p ∈

(1, ∞) and 1 6 t 6 p 6 q 6 ∞, and suppose that:
(i) X and Y are UMD spaces;
(ii) X has cotype q and Y has type t, or one of them has both;
(iii) t : S(D)2 → Z := L (X, Y ) is a weakly defined singular integral and the
kernel K : Ṙ2d → Z of t has the convolution form K(s, t) = K(s − t) and
satisfies the Calderón–Zygmund estimates
1 1
c̃K (Rp ) + kω̃K (Rp )kDiniσ < ∞, σ := max( , 0 ), (12.74)
t q

where c̃K and ω̃K are as in Remark 12.4.6(3);

(iv) t(1Q , 1Q ) = (1Q+̇m , 1Q+̇m ) for all Q ∈ D and m ∈ Zd .
Then the following conditions are equivalent:
(1) t defines a bounded operator T ∈ L (Lp (Rd ; X), Lp (Rd ; Y ));
(2) t satisfies the weak DR p -boundedness property ktkwbp(DRp ) < ∞;
(3) each tω satisfies the weak DR p -boundedness property ktω kwbp(DRp ) 6 C,
uniformly in ω ∈ ({0, 1}d )Z0 .
Under these equivalent conditions, we have
(a) the norm estimate

kT kL (Lp (Rd ;X),Lp (Rd ;Y ))

n  
6 βp,X βp,Y 4d sup ktω kwbp(DRp ) + c0d c1 c̃K (Rp∗0) + c2 c̃K (Rp ) +
ω
 o
+ cd c1 kω̃K (Rp∗0)kDini1/t + c2 kω̃K (Rp )kDini1/q0 ,
1

where the supremum is over ω ∈ ({0, 1}d )Z0 , the constants cd , c0d depend
only on d, and c1 , c2 are as in (12.70);
(b) the representation formulas (12.45) and (12.52) with Λt = Λtω = 0.

Proof. We will check that t is translation-invariant in the sense of Defini-

tion 12.3.9, i.e., that it satisfies the condition of Lemma 12.3.8(1). The very
assumption (iv) of the corollary already takes care of the case Q = R.
On the other hand, if Q 6= R are dyadic cubes of the same size, then
Q ∩ R = ∅ = (Q+̇m) ∩ (R+̇m), and hence we have access to the kernel
representation
 12.4 The T (1) theorem for singular integrals 187
Z Z Z Z
t(1Q , 1R ) = K(s, t) ds dt = K(s − t) ds dt
R Q R Q
Z Z
= K((s + m) − (t + m)) ds dt
R Q
Z Z
= K(s − t) ds dt = t(1Q+̇m , 1R+̇m ),
R+̇m Q+̇m

which proves the condition of Lemma 12.3.8(1) for arbitrary Q, R ∈ D of

equal size. Thus indeed t is translation-invariant.
Next, we wish to have the same property for tω , for every ω ∈ ({0, 1}d )0Z ,
and requires verifying the identity t(1Q0 , 1R0 ) = t(1Q0 +̇m , 1R0 +̇m ) for each
Q0 , R0 ∈ D ω of equal size. By Lemma 12.3.30, we have S(D ω ) = S(D) when-
ever ω ∈ ({0, 1}d )0Z . If Q0 ∈ D ω , then clearly f = 1Q0 ∈ S(D ω ) = S(D), and
similarly with g = 1R0 where R0 ∈ D ω has the same size. Thus Lemma 12.3.8
guarantees that

t(1Q0 +̇m , 1R0 +̇m ) = t(τm`(Q0 ) f, τm`(Q0 ) g) = t(f, g) = t(1Q0 , 1R0 )

for all Q0 , R0 ∈ D ω of the same size, and hence also tω is translation-invariant.

By Proposition 12.4.4, it then follows that tω (1, ·) = 0 = tω (·, 1), for every
ω ∈ ({0, 1}d )Z0 . Thus the conclusions of the corollary are immediate from
Theorem 12.4.12 by setting all Λt and Λtω to be zero. 

Lemma 12.4.14. Let Z = L (X, Y ) and Φ ∈ Cb ([0, ∞); Z) ∩ C 1 ((0, ∞); Z),
and suppose that
(i) K(u) := 1(0,∞) (u)Φ0 (u) satisfies the Calderón–Zygmund estimate (12.74);
(ii) the range of Φ is R-bounded, Rp (Φ) := Rp ({Φ(u) : u ∈ [0, ∞)}) < ∞;
(iii) a bilinear form t : S(D)2 → Z is defined, for all f, g ∈ S(D), by
ZZ
t(f, g) := lim K(u − v)f (v)g(u) dv du.
ε→0 |u−v|>ε

Then
(1) t is well-defined as a weakly defined singular integral with convolution ker-
nel K(u, v) = K(u − v);
(2) tω satisfies the weak DR p -boundedness property

ktω kwbp(DRp ) 6 kΦ(0)k + min Rp (Φ), Rp∗0 (Φ)};



(3) t(1I , 1I ) = (1I +̇m , 1I +̇m ) for all I ∈ D and m ∈ Z.

Proof. (1): Clearly the integral inside the limit is well-defined, since we are
cutting away the singularity. To show the existence of the limit, let first f = 1I
and g = 1J for some intervals I = [aI , bI ) and J = [aJ , bJ ). Then
188 12 Dyadic operators and the T (1) theorem
Z Z bI
K(u − v)f (v) dv = 1(ε,∞) (u − v)Φ0 (u − v) dv
|u−v|>ε aI
Z bI ∧(u−ε)
= 1(aI +ε,∞) (u) Φ0 (u − v) dv
aI
= 1(aI +ε,∞) (u)[Φ((u − bI ) ∨ ε) − Φ(u − aI )].

Since Φ is continuous on [0, ∞), we have

Z
lim K(u − v)f (v) dv = 1(aI ,∞) (u)[Φ((u − bI )+ ) − Φ(u − aI )],
ε→0 |u−v|>ε

= Φ((u − bI )+ ) − Φ((u − aI )+ ).

Since Φ is bounded on [0, ∞), we can apply dominated convergence to obtain

ZZ
lim K(u − v)1I (v)1J (u) dv du
ε→0 |u−v|>ε
Z (12.75)
= [Φ((u − bI )+ ) − Φ((u − aI )+ )] du.
J

In particular, the limit defining t(f, g) exists for all f, g of the form f = 1I
and g = 1J . By (bi)linearity, it exists for all f, g ∈ S(D).
If f, g ∈ S(D) are disjointly supported, then K(u−v)f (v)g(u) is integrable.
Hence ZZ
t(f, g) = K(u − v)f (v)g(u) dv du

by dominated convergence, and thus t is a weakly defined singular integral

with kernel K(u, v) = K(u − v).
(2): With J = I ∈ D ω , noting that aI 6 u < bI for all u ∈ I, the identity
(12.75) shows that
Z
t(1I , 1I )

= − Φ(0) − Φ(u − aI ) du
|I| I
Z `(I) (12.76)


=− Φ(0) − Φ(u) du ∈ Φ(0) + abco(Φ).
0

Thus, by Lemma 12.1.8, we find that

n t(1 , 1 ) o 
I I
ktω kwbp(DRp ) := DR p
|I| I∈D ω
n t(1 , 1 ) o 
I I
6 min ℘i , ℘0 := Rp , ℘1 := Rp∗0 ,
i=0,1 |I| I∈D ω

6 kΦ(0)k + min ℘i (Φ).

i=0,1

(3): From (12.76) it is evident that t(1I , 1I ) depends only on `(I); since
`(I) = `(I +̇m), it follows that t(1I , 1I ) = t(1I +̇m , 1I +̇m ), as claimed. 
 12.4 The T (1) theorem for singular integrals 189

It often happens that kernels that we encounter satisfy standard Calderón–

Zygmund estimates with the best possible Lipschitz modulus of continuity
ω(u) = O(u) as u → 0, but the implies constant in this estimate can be very
large. At the same time, we also have a trivial bound ω(u) = O(1), where
the implied constant may be much smaller. The following lemma provides a
useful estimate of the Dini norms of ω in such cases, showing that the larger
constant enters the estimates only via its logarithm:
Lemma 12.4.15. Let 0 < A 6 B < ∞ and σ ∈ [0, 1]. If ω(u) 6 min(A, Bu),
then  B
kωkDiniσ 6 3A 1 + logσ+1 .
A
Proof.
Z A/B  Z 1
1 σ  1 σ du
(log 2)σ kωkDiniσ 6 B log du + A log =: I + II,
0 u A/B u u
where
Z A/B A/B B
I 6 −B log u du = −B(u log u − u) = A(log + 1)
0 0 A
and
1
(− log u)σ+1
Z 1
du A  B σ+1
II = A (− log u)σ = −A = log .
A/B u σ+1 A/B σ+1 A

Let G := log(B/A). Since

1 σ
G = (Gσ+1 )1/(σ+1) · 1σ/(σ+1) 6 Gσ+1 + ,
σ+1 σ+1
we obtain
2A σ+1  σ 
I + II 6 G +A 1+ 6 2A(Gσ+1 + 1).
σ+1 σ+1
Since (log 2)−σ 6 (log 2)−1 < 3/2, the claim follows. 
Example 12.4.16. Let ω ∈ [0, π/2], σ ∈ [0, 1], and suppose that
Φ ∈ C([0, ∞), Z) ∩ H ∞ (Σω ; Z)
has an R-bounded range. Then Φ|[0,∞) and K(u) = 1(0,∞) (u)Φ0 (u) satisfy the
assumptions of Lemma 12.4.14 with
℘(Φ) 3℘(Φ)  4 
c̃K (℘) 6 , kω̃K (℘)kDiniσ 6 1 + log1+σ , ℘ ∈ {Rp , Rp∗0 }.
sin ω sin ω sin ω
A particular instance of such a Φ is (the negation of) an R-bounded holomor-
phic semigroup Φ(z) = −e−zA , in which case K(u) = Ae−uA is the kernel of
the so-called maximal regularity operator.
190 12 Dyadic operators and the T (1) theorem

Remark 12.4.17. The role of the parameter σ ∈ [0, 1] in Example 12.4.16 is

relatively insignificant and only recorded for curiosity. First, it only affects the
power of the logarithm. Second, for applying Lemma 12.4.14, it is necessary
to take σ > max(1/t, 1/q 0 ) > 12 , and it is always sufficient to take σ = 1, so
that the power of the logarithm will always be in the range [ 23 , 2].

Proof of Example 12.4.16. Let ℘ ∈ {Rp , Rp∗ }. It is evident that

℘({Φ(u) : u ∈ [0, ∞)}) = ℘({Φ(u) : u ∈ (0, ∞)}) 6 ℘({Φ(z) : z ∈ Σω }).

By Cauchy’s formula, we have

I
j! f (z)
Φ(j) (u) = |dz|, u > 0.
2πi |z−u|=u sin ω (u − z)j+1

Denoting ℘(Φ) := ℘(Φ(z) : z ∈ Σω ), we hence have

tj |dz|
I
j (j) j! j!℘(Φ)
℘(t Φ (t) : t > 0) 6 ℘(Φ) sup j+1
= .
2π t>0 |z−t|=t sin ω (t sin ω) (sin ω)j

With K(u) = 1(0,∞) (u)Φ0 (u), it follows that

℘(Φ)
c̃K (℘) = ℘(|u|K(u) : u 6= 0) = ℘(uΦ0 (u) : u > 0) 6 .
sin ω
Moreover,

ω̃K (℘; s) = ℘(|u|[K(u) − K(u0 )] : |u − u0 | 6 s|u|)

 Z u 
=℘ u Φ00 (v) dv : |u − u0 | 6 su
u0
Z u
2℘(Φ) u
6 sup dv ,
(sin ω) |u−u0 |6su u0 v 2
2

where
u
|u − u0 |
Z
u 1 1 su s
dv = u − = 6 = 6 2s
u0 v2 u u0 u0 (1 − w)u 1−s

for |u − u0 | 6 wu and s ∈ [0, 21 ]. Thus ω̃K (℘; s) 6 4℘(Φ)(sin ω)−2 s.

By Remark 12.4.6(2), we also have ω̃K (℘; s) 6 c̃K (℘) 6 ℘(Φ)(sin ω)−1 .
Thus, an application of Lemma 12.4.15 with 0 < A = ℘(Φ)(sin ω)−1 <
4℘(Φ)(sin ω)−2 = B < ∞, we deduce that

3℘(Φ)  4 
kω̃K kDiniσ 6 1 + log1+σ .
sin ω sin ω
This completes the proof. 
 12.4 The T (1) theorem for singular integrals 191

We proceed to further corollaries of Theorem 12.4.12.

Corollary 12.4.18 (T (1) theorem for antisymmetric kernels). Let p ∈

(1, ∞) and 1 6 t 6 p 6 q 6 ∞, and suppose that:
(i) X and Y are UMD spaces.
(ii) X has cotype q and Y has type t, or one of them has both.
(iii) K : Ṙ2d → Z := L (X, Y ) is an antisymmetric kernel, i.e.,

K(s, t) = −K(t, s) for all (s, t) ∈ Ṙ2d ,

which satisfies the Calderón–Zygmund estimates

1
cK (Rp ) + kωK (Rp )kDinimax(1/t,1/q0 ) < ∞. (12.77)

(iv) A bilinear form t : S(D)2 → Z is defined for all f, g ∈ S(D) by

ZZ
1

t(f, g) := K(s, t) f (t)g(s) − f (s)g(t) dt ds. (12.78)
2
Then t is well-defined as a weakly defined singular integral with kernel K, and
the following conditions are equivalent:
(1) t defines a bounded operator T ∈ L (Lp (Rd ; X), Lp (Rd ; Y ));
(2) Λt defines a bounded operator in L (Lp (Rd ; X), Lp (Rd ; Y ));
(3) each Λtω defines a bounded operator in L (Lp (Rd ; X), Lp (Rd ; Y )), uni-
formly in ω ∈ ({0, 1}d )Z0 .
Under these equivalent conditions, we have
(a) the norm estimates as in parts (a) and (b) of Theorem 12.4.12, with
ktkwbp(DRp ) = ktω kwbp(DRp ) = 0;
(b) the representation formulas (12.45) and (12.52).

Proof. To check that t is well-defined, we need to verify that the integrals

in (12.78) make sense. By linearity, it is enough to consider f = 1Q and
g = 1R for some Q, R ∈ D. If Q ∩ R = ∅, then each of the two terms under
the integral is separately integrable by Lemma 12.4.2, and hence so is their
difference. Otherwise, we may assume by the nestedness of dyadic cubes and
symmetry that, e.g., Q ⊆ R. We can then split

f (t)g(s) − f (s)g(t) = 1Q (t)1R (s) − 1Q (s)1R (t)

= 1Q (t) 1Q (s) + 1R\Q (s) − 1Q (s) 1Q (t) + 1R\Q (t)
= 1Q (t)1R\Q (s) − 1Q (s)1R\Q (t),

observing the cancellation of the two equal terms 1Q (s)1Q (t). We can divide
R \ Q into finitely many cubes P ∈ D of the same size as Q, and then
the integrability of each of the terms on the left against K(s, t) follows from
192 12 Dyadic operators and the T (1) theorem

Lemma 12.4.2. Thus the formula defining t as a bilinear form t : S(D)2 → Z

is meaningful.
To show that t has associated kernel K, let f, g ∈ S(D) be disjointly
supported. As we already observed, in this case both terms under the integral
are separately integrable, and we can write
ZZ
1

t(f, g) = K(s, t) f (t)g(s) − f (s)g(t) dt ds
2
I − II
ZZ ZZ
1 1
= K(s, t)f (t)g(s) dt ds − K(s, t)f (s)g(t) dt ds =: .
2 2 2
Using the antisymmetry of K and interchanging the names of the variables,
and applying Fubini’s theorem, we find that
ZZ ZZ
−II = K(t, s)f (s)g(t) dt ds = K(s, ty)f (t)g(s) ds dt = I.

Hence
I − II
ZZ
t(f, g) = =I= K(s, t)f (t)g(s) dt ds,
2
as required for t to be a weakly defined singular integral with kernel K.
From the defining formula (12.78) it is immediate that t(1Q , 1Q ) = 0, and
hence the quantities featuring in the weak boundedness property of t vanish.
With Q ∈ D ω (which still satisfies 1Q ∈ S(D) for ω ∈ ({0, 1}d )Z0 , by Lemma
12.3.30, the same conclusion extends to tω for all ω ∈ ({0, 1}d )Z0 . The rest of
the corollary is then a direct consequence of Theorem 12.4.12, simply setting
ktkwbp(DRp ) = ktω kwbp(DRp ) = 0. We only need to note that ωK 1
(℘) = ωK 2
(℘)
when K is antisymmetric, which is why a seemingly weaker assumption suffices
in (12.77). 

Corollary 12.4.19 (T (1) theorem for antisymmetric convolutions).

Let p ∈ (1, ∞) and 1 6 t 6 p 6 q 6 ∞, and suppose that:
(i) X and Y are UMD spaces.
(ii) X has cotype q and Y has type t, or one of them has both.
(iii) K : Ṙ2d → Z := L (X, Y ) is an antisymmetric convolution kernel, i.e.,

K(s, t) = K(s − t) = −K(t − s) for all (s, t) ∈ Ṙ2d ,

which satisfies the Calderón–Zygmund estimates (12.74).

(iv) A bilinear form t : S(D)2 → Z is defined for all f, g ∈ S(D) by
ZZ
1

t(f, g) := K(s − t) f (t)g(s) − f (s)g(t) dt ds.
2
Then t is well-defined as a weakly defined singular integral with kernel K,
which defines a bounded operator T ∈ L (Lp (Rd ; X), Lp (Rd ; Y )) and satisfies
 12.4 The T (1) theorem for singular integrals 193

(a) the norm estimate

kT kL (Lp (Rd ;X),Lp (Rd ;Y ))
n  
6 βp,X βp,Y c0d c1 c̃K (Rp∗0) + c2 c̃K (Rp ) +
 o
+ c1d c1 kω̃K (Rp∗0)kDini1/t + c2 kω̃K (Rp )kDini1/q0 ,

where the supremum is over ω ∈ ({0, 1}d )0Z , the constants cd , cd0 depend
only on d, and c1 , c2 are as in (12.70).
(b) the representation formulas (12.45) and (12.52) with Λt = Λtω = 0.
Proof. This is straightforward by combining (the proofs of) Corollaries 12.4.13
and 12.4.18. In particular, in the proof of Corollary 12.4.18 we observed
that any bilinear form defined as in (iv) of the present corollary will satisfy
t(1Q , 1Q ) = 0 for all Q ∈ D, and hence also t(1Q+̇m , 1Q+̇m ) = 0 = t(1Q , 1Q )
for all m ∈ Zd . This is condition (iv) of Corollary 12.4.13 that was not explic-
itly assumed in the corollary that we are proving. 
Remark 12.4.20. As an immediate consequence of Corollary 12.4.13, we obtain
another proof of the essence of Theorem 5.1.13 on the boundedness of the
Hilbert transform H on Lp (R; X) whenever p ∈ (1, ∞) and X is a UMD
space. Indeed, take X = Y , t = 1, and q = ∞, so that the constants in (12.70)
are simply c1 = c2 = 1. Clearly the kernel K(u, v) = π −1 (u − v)−1 of the
Hilbert transform is an antisymmetric convolution kernel, and it is easy to
check the Calderón–Zygmund estimates (12.74) with Dini1 norms. Thus we
obtain the estimate
2
kHkL (Lp (Rd ;X)) 6 c · βp,X ,
with the same quantitative form as (5.24), aside from the unspecified numer-
ical factor above, in contrast to the explicit constant 2 in (5.24). This is quite
natural, considering that (5.24) was obtained by an argument tailored for the
very Hilbert transform, whereas the argument that we just sketched was a
specialisation of a much more general argument to the particular case of H.
The following corollary provides a solution to the Lp extension problem from
Section 2.1 for the important class of Calderón–Zygmund operators:
Theorem 12.4.21 (T (1) theorem for scalar-valued kernels). Let p, s ∈
(1, ∞) and 1 6 t 6 p 6 q 6 ∞, and suppose that:
(i) X is a UMD space with cotype q and type t,
(ii) t : S(D)2 → K is a weakly defined singular integral, whose kernel K :
Ṙ2d → K satisfies the Calderón–Zygmund estimates
2
X
i
cK + kωK kDiniσi < ∞, (12.79)
i=1

where σ1 = 1/t and σ2 = 1/q 0 .

194 12 Dyadic operators and the T (1) theorem

Then the following conditions are equivalent:

(1) t defines a bounded operator T ∈ L (Lp (Rd ; X));
(2) t defines a bounded operator T ∈ L (Ls (Rd ));
(3) ktω kwbp 6 C uniformly in ω ∈ ({0, 1}d )Z0 , and for some bi ∈ BMO(Rd ),

t(1, g) = hb1 , gi, t(f, 1) = hf, b2 i (12.80)

for all f, g ∈ S00 (D ω ) and ω ∈ ({0, 1}d )0Z ;

(4) ktkwbp < ∞, and (12.80) for some bi ∈ BMOD (Rd ) and all f, g ∈ S00 (D).
Under these equivalent conditions, we have
2
kT kL (Lp (Rd ;X)) 6 c̃d βp,X (c1 + c2 )cK +
2
  X  (12.81)
2
+ c̃d βp,X + pp0 βs,X
2
βs,K kT kL (Ls (Rd )) + i
ci kωK kDiniσi ,
i=1

with a dimensional constant c̃d and cotype constants

c1 = ct0 ,X ∗ ;p0 , c2 = cq,X;p .

In particular, every Lp (Rd )-bounded Calderón–Zygmund operator having ker-

nel bounds (12.80) with σ1 = σ2 = 1, extends boundedly to Lp (Rd ; X) for
every UMD space X, and one can take c1 = c2 = 1 in the estimate (12.81).

Proof. (1) ⇒ (2): For s = p, this is evident by restricting the action of the
operator to a one-dimensional subspace of X. The case of general s ∈ (1, ∞)
follows from the Calderón–Zygmund Theorem 11.2.5 (or even just its classical
scalar-valued version).
(2) ⇒ (3): The weak boundedness property follows from Example 12.1.10:

ktω kwbp 6 kT kL (Ls (Rd )) , (12.82)

and we turn to the construction of the functions bi .

The operator T ∈ L (Ls (Rd )) is a Calderón–Zygmund operator with
kernel K that satisfies in particular the Dini conditions in both variables,
and hence both direct and dual (operator-)Hörmander conditions by Lemma
11.3.4. (The qualifier “operator” is redundant for scalar-valued kernels.) By
(just the scalar-valued version of) Theorem 11.2.9, T has an extension Te ∈
L (L∞ (Rd ), BMO(Rd )/K). By Theorem 11.2.9(b), for functions 1 ∈ L∞ (Rd )
∞
and g ∈ S00 (D ω ) ⊆ Lc,0 (Rd ), we have

hTe(1), gi = lim hT (1(1+2M )Q ), gi

M →∞
X
= lim t(1Q+̇m , g) = t(1, g).
M →∞
m∈Zd
|m|∞ 6M
 12.4 The T (1) theorem for singular integrals 195

This is one of the claimed identities with b1 := Te(1) ∈ BMO(Rd ; Y ), and

Theorem 11.2.9, followed by Lemma 11.3.4, provide us with the estimates

kb1 kBMOs (Rd ;Y ) = kTe(1)kBMOs (Rd )

6 (cd kT kL (Ls (Rd ) + kKkHör∗ )k1kL∞ (Rd ) (12.83)
1
6 (cd kT kL (Ls (Rd ) + σd−1 kωK kDini ).

The identity involving b2 := Tf∗ (1), and the estimate

kb2 kBMOs0 (Rd ) = kTf∗ (1)kBMOs0 (Rd )

6 cd kT ∗ kL (Ls0 (Rd )) +
(12.84)
+ k(u, v) 7→ K(u, v)∗ kHör∗ k1kL∞ (Rd )

2
6 (cd kT kL (Ls (Rd )) + σd−1 kωK kDini )

are entirely analogous on the dual side.

(3) ⇒ (4): This is obvious by restricting to ω = 0 and noting that
BMO(Rd ) ⊆ BMOD (Rd ).
(4) ⇒ (1): Under assumption (4), we see that the paraproducts related to
t are in fact Πti = Πbi , where bi ∈ BMOD (Rd ) by assumption. Thus Corollary
12.2.19 guarantees that

kΛt kL (Lp (Rd ;X)) = kΠb1 + Πb∗2 kL (Lp (Rd ;X))

(12.85)
6 64 · 8d · pp0 βs,X
2
s (Rd ) + kb2 k
βs,K (kb1 kBMOD BMOs (Rd ) ).
0
D

Our assumption (4) also involves ktkwbp < ∞, and Corollary 12.1.9 guarantees
that this coincides with the finiteness of ktkwbp(DRp ) = ktkwbp , when t is scalar-
valued. Thus both assumptions ktkwbp(DRp ) < ∞ and kΛt kL (Lp (Rd ;X)) < ∞
of Theorem 12.4.12(2) are satisfied, hence also the equivalent condition of
Theorem 12.4.12(1), and this coincides with condition (1) of the corollary
that we are proving.
The quantitative estimates: While we have already closed the chain of impli-
cations (1) ⇒ (2) ⇒ (3) ⇒ (4) ⇒ (1), the claimed quantitative bounds require
a direct analysis of the implication (3) ⇒ (1), which relates to the implication
(3) ⇒ (1) of Theorem 12.4.12.
As in the proof of “(4) ⇒ (1)”, under assumption (3), we see that
the paraproducts related to tω are in fact Πtiω = Πbωi ; while the function
bi ∈ BMO(Rd ) ⊆ BMOD ω (Rd ) is independent of ω, the superscript of the
paraproduct signifies the fact that the defining series involves Haar functions
and averages related to Q ∈ D ω . Thus, imitating (12.85) and substituting the
bounds (12.83) and (12.84), we obtain, with s1 := s and s2 := s0 ,
196 12 Dyadic operators and the T (1) theorem
kΛtω kL (Lp (Rd ;X)) = kΠbω1 + (Πbω2 )∗ kL (Lp (Rd ;X))
2
X
6 64 · 8d · pp0 βs,X
2
βs,K kbi kBMOsi (Rd )
i=1 (12.86)
2 
X 
6 64 · 8d · pp0 βs,X
2
βs,K i
cd kT kL (Ls (Rd )) + σd−1 kωK kDini .
i=1

where we implicitly dominated kbi kBMOsiω (Rd ) 6 kbi kBMOsi (Rd ) in the first es-
D
timate. We now substitute (12.86) and (12.82) into the second norm estimate
in Theorem 12.4.12(b), noting that all R-bounds and DR p -bounds may be
omitted, since they simply reduce to uniform bounds for scalar-valued func-
tions:

kT kL (Lp (Rd ;X)) 6 sup kΛtω kL (Lp (Rd ;X)) +

ω
n
2
+ βp,X 4d sup ktω kwbp + c0d (c1 + c2 )cK +
ω
 o
1 1 2
+ cd c1 kωK kDini1/t + c2 kωK kDini1/q0 .

This gives the bound asserted in the corollary. 

Remark 12.4.22. If b1 = b2 , the term pp0 βs,X

2
βs,K can be omitted in (12.81).
This applies in particular if T is translation-invariant.

Proof. By inspection of the proof of Theorem 12.4.21, the said term only arises
in the estimate of Λtω in (12.86). Under the assumption that b1 = b2 , we have
Λtω = Λbω1 , and we may replace (12.86) by an application of Theorem 12.2.25:

kΛtω kL (Lp (Rd ;X)) = kΛbω1 kL (Lp (Rd ;X)) 6 30 · 2d · βp,X

2
kb1 kBMO(Rd ) ,

where
1
kb1 kBMO(Rd ) 6 kb1 kBMOs (Rd ) 6 cd kT kL (Ls (Rd )) + σd−1 kωK kDini .

Substituting this alternative estimate into the proof of Theorem 12.4.21, we

obtain the claimed modification of (12.81).
If T is translation-invariant, the paraproduct terms vanish, and hence we
can take b1 = b2 = 0, which is indeed a special case of b1 = b2 . Of course, in
this case, we do not even need to use Theorem 12.2.25. 

12.4.b The dyadic representation theorem

The randomised dyadic representation (12.52) underlying the proof of T (1)

Theorem 12.3.26 can be further reorganised into a form that has proven to be
useful for various extensions. Recalling Definition 12.3.34 of the good parts
of Figiel’s operators, and in particular the quantity k(n) := 2 + dlog2 |n|e, we
 12.4 The T (1) theorem for singular integrals 197

regroup the sum over n ∈ Zd \ {0} in (12.52) according to a constant value of

k(n) as
X ∞
X X
= .
n∈Zd \{0} k=2 n∈Zd
2k−3 <|n|62k−2

We denote by ch(k) (P ) the collection of dyadic descendants of P of generation

k, and define the operators
(k) (k)
X X
DP := DQ , EP := EQ ,
Q∈ch(k) (P ) Q∈ch(k) (P )
k−1
[0,k) (k) (0) (j)
X
DP := EP − EP = DP .
j=0

Lemma 12.4.23. If t : S(D)2 → L (X, Y ) =: Z is a weakly defined singular

integral with kernel K : Ṙ2d → Z, then
X good
Tk := hTn,t ω f, gi = hS
(0,k)
f, gi,
n∈Zd
2k−3 <|n|62k−2
X 1,good
Uk1 := hf, Un,t ω gi = hS (1,k) f, gi,
n∈Zd
2k−3 <|n|62k−2
X 2,good
Uk2 := hUn,t ω f, gi = hS (2,k) f, gi,
n∈Zd
2k−3 <|n|62k−2

where
Z
(i,k) (i,k) (i,k)
X
(i,k)
S f= AP f, AP f (s) = − aP (s, t)f (t) dt,
P ∈D P

and these satisfy the identities

(0,k) (k) (0,k) (k)
AP = DP AP DP ,
(1,k) (k) (0,k) [0,k)
AP = DP AP DP , (12.87)
(2,k) [0,k) (0,k) (k)
AP = DP AP DP .

For i = 1, 2, we have the further splitting

(i,k) (i,k) (i,k)
X
AP f = AP ;P f − AP ;R f
R∈ch(k)
198 12 Dyadic operators and the T (1) theorem

where
Z
(i,k) (i,k)
AP ;R f (s) = − aP ;R (s, u)f (u) du, R ∈ {P } ∪ ch(k) (P ),
R

and these kernels have the bounds

(0,k) (i,k)
℘({aP (s, u), aP,R (s, u) : s, u ∈R ∈ {P } ∪ ch(k) (P ), P ∈ D})
( √
cK (℘), if 2k 6 12 d,
6 cd √ √
i
ωK (℘; 62kd ), if 2k > 12 d,

Proof. By definition, the left-hand side of the claim is equal to

X X D E
γ γ
Tk = t(hα
Q , h Q+̇n
)hf, h α
Q i, hg, h Q+̇n
i ,
n∈Zd Q∈Dk-good
2k−3 <|n|62k−2 α,γ∈{0,1}d \{0}

where the k-goodness of Q guarantees that R := Q+̇n, for |n| 6 2k−2 , shares
with Q the same kth dyadic ancestor R(k) = Q(k) =: P ∈ D. Thus we can
regroup this series under the ancestors P to get
X X D E
γ γ
Tk = tgood (hα
Q , h R )hf, h α
Q i, hg, h R i ,
P ∈D (Q,R)∈Ck (P )
α,γ∈{0,1}d \{0}

where
n 1 1 o
Ck (P ) := (Q, R) : Q, R ∈ ch(k) (P ), `(P ) < |zQ − zR | 6 `(P ) .
8 4
(k)
The subseries under each P ∈ D takes the asserted form hAP f, gi if we define
(0,k)
X γ γ
aP (s, u) := |P | tgood (hα α
Q , hR )hQ (u)hR (s).
(Q,R)∈Ck (P )
α,γ∈{0,1}d \{0}

The cases of Uki are analogous, and lead to representations of the same
form with
(1,k)
X
aP (s, u) := |P | tgood (h0Q , hγR )[h0Q (u) − h0R (u)]hγR (s),
(Q,R)∈Ck (P )
γ∈{0,1}d \{0}

and
(2,k)
X
aP (s, u) := |P | tgood (hα 0 α 0 0
Q , hR )hQ (u)[hR (s) − hQ (s)],
(Q,R)∈Ck (P )
α∈{0,1}d \{0}
 12.4 The T (1) theorem for singular integrals 199

The further splitting is then naturally defined with

(1,k)
X
aP ;P (s, u) := |P | 0
tgood (hQ γ
, hγR )h0Q (u)hR (s),
(Q,R)∈Ck (P )
γ∈{0,1}d \{0}
(1,k)
X
aP ;R (s, u) := |R| tgood (h0Q , hγR )hR
0
(u)hγR (s), R ∈ chk (P ),
Q:(Q,R)∈Ck (P )
γ∈{0,1}d \{0}

where the last summation runs over all relevant Q ∈ chk (P ), for fixed R.
(1,k)
Observe that aP ;R has the factor |R| in front, instead of |P |, due to our
(1,k) R (1,k)
definition of AP ;R f (s) as the average integral −R aP ;R (s, u)f (u) du.
(2,k)
The splitting of aP is entirely analogous; in particular,
(2,k)
X
aP ;Q (s, u) := |Q| α
tgood (hQ α
, h0R )hQ 0
(u)hQ (s), Q ∈ chk (P ).
R:(Q,R)∈Ck (P )
α∈{0,1}d \{0}

It remains to verify that these operators and their kernels satisfy the as-
(0,k) (k) (0,k) (k)
serted properties. The identity AP = DP AP DP is immediate from the
(i,k)
orthogonality of the Haar functions, and the invariance of AP under com-
(k)
position by DP on the side, where the cancellative Haar function appear
(i,k) [0,k)
in aP is justified similarly. Concerning the factors DP , we note that the
are orthogonal projections onto functions supported on P , constant on each
Q ∈ ch(k) (Q), and integrating to zero. Noting the functions hQ 0
− h0R belong
to this class then justifies the remaining parts of the claimed identities.
Concerning the claimed bounds, we note that any given (s, u) ∈ P × P is
contained in exactly one Q × R with Q, R ∈ ch(k) (P ), and moreover,

γ 1Q×R 2kd
|hα
Q ⊗ hR | = = 1Q×R .
|Q|1/2 |R|1/2 |P |
(0,k) (i,k)
The claimed ℘-bounds for aP (s, u), as well as for aP ;P (s, u), then follow
from Lemma 12.4.8, noting that the factor |P | in the definition of these kernels
cancels with the |P1 | above.
(1,k)
For aP ;R with R ∈ ch(k) (P ), all terms in the defining sum are supported
on the same 1R×R , and each individual summand can be estimates by Lemma
12.4.8. We now have the smaller factor |R| in front, but at the same time there
are up to 2kd terms in the sum, all of which accumulate on the same support
now. Since 2kd |R| = |P |, we get the same final bound as before. The case of
(1,k)
aP ;Q with Q ∈ ch(k) (P ) in entirely analogous, and completes the proof. 

Definition 12.4.24. An operator S : S00 (D; X) → S00 (D; Y ) is called a

dyadic shift of type (i, k), where i ∈ {0, 1, 2} and k ∈ {2, 3, . . .}, if
200 12 Dyadic operators and the T (1) theorem
X X
S= AP , AP f (s) = AP ;P f (s) − AP ;Q f (s),
P ∈D Q∈ch(k) (P )

where
Z
AP ;R f (s) = − aP ;R (s, u)f (u) du, R ∈ {P } ∪ ch(k) (P ),
R

supp aP ;R ⊆ R × R,
n o
kSkShift(℘) := ℘ aP ;R (s, u) : s, u ∈ R ∈ {P } ∪ ch(k) (P ), P ∈ D <∞

for ℘ = R2 , and moreover, for every P ∈ D,

(k) (k)
(0) if i = 0, then AP = DK AP DK , and AP ;Q = 0 for all Q ∈ ch(k) (P );
(k) [0,k)
(1) if i = 1, then AP = DK AP DK ;
[0,k) (k)
(2) if i = 2, then AP = DK AP DK .
We say that a shift has type i ∈ {0, 1, 2}, if it has type (i, k) with some k.

Remark 12.4.25. In the language of Definition 12.4.24, the operators S (i,k) of

Lemma 12.4.23 are dyadic shifts of type (i, k), and we may further write
( √
(i,k) cK (℘), √ if 2k 6 12 d,
kS kShift(℘) 6 cd √
i
ωK (℘; 62kd ), if 2k > 12 d.

The key boundedness properties of these dyadic shifts are contained in the
following:

Theorem 12.4.26. Let X and Y be UMD spaces, and p ∈ (1, ∞). Suppose
that X has cotype q and Y has type t for some 1 6 t 6 p 6 q 6 ∞.
Then for all i ∈ {0, 1, 2} and k ∈ {2, 3, . . .}, all dyadic shifts S of type
(i, k) extends to a bounded operator from Lp (Rd ; X) to Lp (Rd ; Y ). Moreover,
they satisfy the norm estimates
(
kSkShift(Rp ) ct0 ,Y ∗ ;p0 · k 1/t , i = 1,
kSkL (Lp (Rd ;X),Lp (Rd ;Y )) 6 4 · βp,X βp,Y × 0
kSkShift(R∗0 ) cq,X;p · k 1/q , i = 2;
p

and the norm of a shift of type (0, k) is bounded by the minimum of these two
bounds, but with 6 in place of 4.

Proof. We divide the proof into case according to the type of the shift under
consideration.
 12.4 The T (1) theorem for singular integrals 201

Shifts of type 1

Let us start with the case i = 1. For f ∈ S00 (Rd ; X) ⊆ Lp (Rd ; X) and
0
g ∈ S00 (Rd ; Y ) ⊆ Lp (Rd ; Y ∗ ), we expand the pairing hSf, gi by separating
the scales according to log2 `(P ) mod k:
k−1 D E
(k) (k)
X X
|hSf, gi| = DP AP f, DP g
j=0 P ∈D
log2 `(P )≡j
mod k
k−1
(k)
X X
6 ε P D P AP f
Lp (Ω×Rd ;Y )
j=0 P ∈D
log2 `(P )≡j
mod k
k−1
(k)
X X
× ε P DP g =: Ij × IIj .
Lp0 (Ω×Rd ;Y ∗ )
P ∈D j=0
log2 `(P )≡j
mod k

(k) P
In Ij , we write out DP = Q∈ch(k) (P ) DQ and note that, in a randomised
sum like here, we are free to replace εQ by εP , since the difference is invisible
to the Lp (Ω; Y ) at a fixed s ∈ Rd . This gives
X X
Ij = ε Q D Q AP f .
Lp (R; Y )
P ∈D Q∈ch(k) (P )
log2 `(P )≡j
mod k

Using the splitting of AP , it then follows that

X X
Ij 6 εQ DQ AP ;P f
Lp (Ω×R; Y )
P ∈D Q∈ch(k) (P )
log2 `(P )≡j
mod k
X X
+ εQ DQ AP ;Q f
Lp (Ω×R; Y )
P ∈D Q∈ch(k) (P )
log2 `(P )≡j
mod k
=: IIIj + IVj .

We first consider IVj . Denoting by Qs the unique dyadic child of Q that

contains a given s ∈ Q, and with the understanding that DQ acts in the s
variable, we have
Z
[0,k)
DQ AP ;Q f (s) = − DQ aP ;Q (s, u)DP f (u) du
Q
Z  
= 1Q (s)− haP ;Q (·, u)iQs − haP ;Q (·, u)iQ (hf iQ − hf iP ) du
Q
202 12 Dyadic operators and the T (1) theorem
[0,k)
=: αP ;Q (s)1Q (s)(hf iQ − hf iP ) = αP ;Q (s)1Q (s)DP (s),

where Z  
αP ;Q (s) := − haP ;Q (·, u)iQs − haP ;Q (·, u)iQ du
Q

belongs to the two-fold multiple of the absolute convex hull of the set appear-
ing in the definition of kSkShift(℘) . Thus

[0,k)
X X
IVj = εQ αP ;Q 1Q DP f
Lp (Ω×R; Y )
P ∈D Q∈ch(k) (P )
log2 `(P )≡j
mod k
[0,k)
X X
6 2kSkShift(Rp ) ε Q 1Q D P f
Lp (Ω×R; X)
P ∈D Q∈ch(k) (P )
log2 `(P )≡j
mod k
[0,k)
X
6 2kSkShift(Rp ) ε P DP f ,
Lp (Ω×R; X)
P ∈D
log2 `(P )≡j
mod k
P
using the identity Q∈ch(k) (P ) 1Q = 1P and the interchangeability of εP and
εQ in the random sum in the last step.
[0,k)
Observing that (DP f )log2 `(P )≡j mod k is a martingale difference decom-
position of f for each j ∈ {0, . . . , k − 1} to deduce directly from the definition
of the UMD constants that
+
IVj 6 2kSkShift(Rp ) βp,X kf kLp (Rd ;X) .

We then turn to term IIIj . By the exchangeability of εP and εQ again,

this can be written as
(k)
X
IIIj = εP DP AP ;P f p ;
,
L (Ω×R Y )
P ∈D
log2 `(P )≡j
mod k

where
Z
(k) (k) [0,k)
DP AP ;P f (s) = − DP aP ;P (s, u)DP f (u) du,
P

(k)
and it is understood that DP acts with respect to the s variable.
We will now make use of the tangent martingale construction as in Corol-
lary 4.4.15 and explained just before the statement of the said result: For
every P ∈ D, let TP be a copy of P equipped with the normalised measure
νP := |P |−1 m|PQ, where m is the Lebesgue measure, and consider the prod-
uct space T := P ∈D TP with probability measure ν := ⊗P ∈D νP . Writing a
typical element of T as t = (tP )P ∈D , we then have
 12.4 The T (1) theorem for singular integrals 203
Z
(k) (k) [0,k)
DP AP ;P f (s) = DP aP ;P (s, tP )DP f (tP ) dν(t).
T

Hence (suppressing, as usual, the dependence of random functions on ω ∈ Ω),

Z
(k) [0,k)
X
IIIj = s 7→ εP DP aP ;P (s, tP )DP f (tP ) dν(t) p d
T L (Ω×R ;Y )
P ∈D
log2 `(P )≡j
mod k
(k) [0,k)
X
6 (s, t) 7→ εP 1P (s)DP aP ;P (s, tP )DP f (tP ) .
Lp (Ω×Rd ×T ;Y )
P ∈D
log2 `(P )≡j
mod k

(k)
Here, DP aP ;P (s, tP ) is the difference of two averages haP ;P (·, tP )iQ , and
hence in twice the absolute convex hull of the set in the definition of kSkShift(℘) .
Thus, the definition of R-boundedness implies that
[0,k)
X
IIIj 6 2kSkShift(Rp ) (s, t) 7→ εP 1P (s)DP f (tP ) p d
.
L (Ω×R ×T ;X)
P ∈D
log2 `(P )≡j
mod k

We are now in a position to apply Corollary 4.4.15. Indeed, the functions

[0,k) [0,k)
DP f are “atoms” in the sense defined before that corollary: DP f is sup-
0 (k)
ported on P , of average 0, and constant on all P ∈ ch (P ), which are the
next smaller cubes in the scales-separated dyadic system {P ∈ D : log2 `(P ) ≡
j mod k}. Thus, a direct application of Corollary 4.4.15 to
[0,k)
X
f= DP f
P ∈D
log2 `(P )≡j
mod k

shows that
[0,k)
X
(s, t) 7→ εP 1P (s)DP f (tP ) 6 βp,X kf kLp (Rd ;X) ,
Lp (Ω×Rd ×T ;X)
P ∈D
log2 `(P )≡j
mod k

and hence
IIIj 6 2kSkShift(Rp ) βp,X kf kLp (Rd ;X) .
Combining this with the estimate for term IVj (and estimating the one-sided
UMD constant by the basic UMD constant), we deduce that

Ij 6 IIIj + IVj 6 4kSkShift(Rp ) βp,X kf kLp (Rd ;X) .

Hence
204 12 Dyadic operators and the T (1) theorem
k−1
X k−1
X
|hSf, gi| 6 Ij × IIj 6 4kSkShift(Rp ) βp,X kf kLp (Rd ;X) IIj ,
j=0 j=0

where
k−1  k−1  10
X X 0 t
IIj 6 k 1/t IIjt
j=0 j=0
 k−1 t0  10
(k)
X X t
= k 1/t ε P DP g
Lp0 (Ω×Rd ;Y ∗ )
j=0 P ∈D
log2 `(P )≡j
mod k
(k)
X
6 k 1/t · ct0 ,Y ∗ ;p0 εP DP g
Lp0 (Ω×Rd ;Y ∗ )
P ∈D

6 k 1/t · ct0 ,Y ∗ ;p0 · βp+0 ,Y ∗ kgkLp0 (Rd ;Y ∗ ) .

Here βp+0 ,Y ∗ 6 βp0 ,Y ∗ = βp,Y by Proposition 4.2.17(2), and ct0 ,Y ∗ ;p0 6 τt,Y ;p by
Proposition 7.1.13 (or its easy extension to deal with the third index in these
constants). This completes the proof for shift of type (1, k).

Shifts of type 2

For a shift of type (2, k), we note that its adjoint S ∗ is a shift of type (1, k),
and hence

kSkL (Lp (Rd ;X),Lp (Rd ;Y )) = kS ∗ kL (Lp0 (Rd ;Y ∗ ),Lp0 (Rd ;X ∗ ))

0
6 4kS ∗ kShift(Rp0 ) βp0 ,Y ∗ βp0 ,X ∗∗ cq,X;p k 1/q
0
= 4kSkShift(R∗0 ) βp,Y βp,X cq,X;p k 1/q ,
p

which is the asserted bound in this case.

Shifts of type 0

Let finally S be a shift of type (0, k). We can then proceed as in the case of
type (1, k) with slight modifications: In view of the eventual application of
the tangent martingale estimate of Corollary 4.4.15, we now separate scales
(k)
by k + 1 levels instead of k, since DP f is only guaranteed to be constant
on Q ∈ ch(k+1) (P ). On the other hand, we now have IVj = 0, and hence
Ij = IIIj .
Following the argument in the case of type (1, k) leads to
(k)
X
IIIj 6 2kSkShift(Rp ) (s, t) 7→ εP 1P (s)DP f (tP ) .
Lp (Ω×Rd ×T ;X)
P ∈D
log2 `(P )≡j
mod k+1
 12.4 The T (1) theorem for singular integrals 205

To complete the estimate, we will need a little additional trick compared to

the previous cases. First, we observe that
(k) (k) [0,k+1)
DP = (I − EP )DP .
Second, we have
(k)
EP f (tP ) = E(f |σ(ch(k+1) (P )))(tP )
 O   
= E t 7→ f (tP ) σ(ch(k+1) (Q)) =: E t 7→ f (tP ) Gk+1 ,
Q∈D

where on the right-hand side we take a conditional expectation with respect

to a product σ-algebra on the product probability space T , of a function that
only depends on the “coordinate” tP of t ∈ T . The importance of this last
formula comes from the fact that only the function inside the conditional
expectation, but not the conditional expectation operator itself, depends on
the dyadic cube P . Using the previous two formulas, it follows that
(k)
X
εP 1P (s)DP f (tP )
Lp (Ω×Rd ×T ;X)
P ∈D
log2 `(P )≡j
mod k+1
[0,k+1)
X
6 εP 1P (s)DP f (tP )
Lp (Ω×Rd ×T ;X)
P ∈D
log2 `(P )≡j
mod k+1
 
[0,k+1)
X
+ E εP 1P (s)DP f (tP ) Gk+1
Lp (Ω×Rd ×T ;X)
P ∈D
log2 `(P )≡j
mod k+1
[0,k+1)
X
62 εP 1P (s)DP f (tP )
Lp (Ω×Rd ×T ;X)
P ∈D
log2 `(P )≡j
mod k+1

by the contractivity of the conditional expectation in the last step. This last
expression has the same form as what we encountered with shifts of type
(1, k), only with k + 1 in place of k. Thus, by an application of the tangent
martingale inequality of Corollary 4.4.15, we have
[0,k+1)
X
εP 1P (s)DP f (tP ) 6 βp,X kf kLp (Rd ;X) .
Lp (Ω×Rd ×T ;X)
P ∈D
log2 `(P )≡j
mod k+1

Thus,
Ij = IIIj 6 4kSkShift(Rp ) βp,X kf kLp (Rd ;X) ,
which is the same bound as for the corresponding terms in the estimate of
shifts of type (1, k). The rest of the argument is exactly the same, only with
k + 1 in place of k, and leads to the conclusion that
206 12 Dyadic operators and the T (1) theorem

kSkL (Lp (Rd ;X),Lp (Rd ;Y )) 6 4kSkShift(Rp ) βp,X βp,Y ct0 ,Y ∗ ;p0 (k + 1)1/t .

Since the adjoint of a shift of type (0, k) is another shift of the same type, we
also obtain
0
kSkL (Lp (Rd ;X),Lp (Rd ;Y )) 6 4kSkShift(R∗0 ) βp,X βp,Y cq,X;p (k + 1)1/q ,
p

and we can take the minimum of the two bounds. Since k > 2, we can also
make the trivial estimates k + 1 6 23 k and 4 · ( 23 )1/v 6 6 for v ∈ {t, q 0 } so that
in case v > 1. 
With the help of the shifts, we can represent any weakly defined singular
integral with appropriate kernel bounds as follows:

Theorem 12.4.27 (Dyadic Representation Theorem). Let p ∈ (1, ∞)

and 1 6 t 6 p 6 q 6 ∞, and suppose that:
(i) X and Y are UMD spaces,
(ii) X has cotype q and Y has type t,
(iii) t : S(D)2 → Z := L (X, Y ) is a weakly defined singular integral and the
kernel K : Ṙ2d → Z of t satisfies the Calderón–Zygmund estimates
1 2
cK (Rp ) + kωK (Rp )kDini1/t + kωK (Rp )kDini1/q0 < ∞.

Then the following conditions are equivalent:

(1) t defines a bounded operator T ∈ L (Lp (Rd ; X), Lp (Rd ; Y ));
(2) each tω satisfies the weak DR p -boundedness property ktω kwbp(DRp ) 6
C, and the associated bi-paraproduct Λtω defines a bounded operator in
L (Lp (Rd ; X), Lp (Rd ; Y )), uniformly in ω ∈ ({0, 1}d )Z0 .
Under these equivalent conditions, we have
(a) the dyadic representation formula
 ∞
X 
hT f, gi = E hHtω f, gi + hΛtω f, gi + hSω(i,k) f, gi
k=2
i∈{0,1,2}

with absolute convergence for all f ∈ S(D; X) and g ∈ S(D; Y ∗ ), where E

is the expectation over ω ∈ ({0, 1}d )Z6M , and M ∈ Z is any large enough
number such that f and g are constant on all Q ∈ DM ; the operators Htω
and Λtω are a Haar multiplier and a paraproduct as in (12.52), and each
(i,k)
Sω is a dyadic shift of type (i, k) (Definition 12.4.24) with respect to the
dyadic system D ω and with shift norms estimated by
( √
(i,k) cK (℘), if 2k 6 12 d,
kSω kShift(℘) 6 cd √ √
ωKi
(℘; 62kd ), if 2k > 12 d;
 12.4 The T (1) theorem for singular integrals 207

(b) the resulting norm estimate:

kT kL (Lp (Rd ;X),Lp (Rd ;Y )) − sup kΛtω kL (Lp (Rd ;X),Lp (Rd ;Y ))
ω
n  
6 βp,X βp,Y 4 sup kt kwbp(DRp ) + c0d cK (Rp ) + cK (Rp∗0 ) +
d ω
ω
 o
1
+ cd ct0 ,Y ∗ ;p0 kωK1
(Rp )kDini1/t + cq,X;p kωK 2
(Rp∗0 )kDini1/q0 ,

where the suprema are over ω ∈ ({0, 1}d )0Z , and the constants cd0 , c1d depend
only on d

Proof. We note that the present assumptions coincide with those of Theorem
12.4.12, except that (ii) of the present theorem is slightly stronger than (ii) of
Theorem 12.4.12. Thus the equivalence of (1) and (2) is just repetition from
Theorem 12.4.12.
The first new claim is the dyadic representation formula (a). To see this,
recall that Theorem 12.4.12 gave the representation formula (12.52), repeated
for convenience as
 X n good
hT f, gi = E hHtω f, gi + hΛtω f, gi + 2d hTn,tω f, gi+
n∈Zd
n6=0
o
1,good 2,good
+ hf, Un,t ω gi + hUn,t ω f, gi ,

where f, g, and E have the same meaning as in the claimed formula (a). On the
other hand, Lemma 12.4.23 and Remark 12.4.25 inform us that the summation
of the three types of terms over n ∈ Zd \ {0} can be rearranged into a sum
over k > 2 and i ∈ {0, 1, 2} exactly as in the assertion.
From the representation (a), we can then estimate

kT kL (Lp (Rd ;X),Lp (Rd ;Y )) − sup kΛtω kL (Lp (Rd ;X),Lp (Rd ;Y ))
ω

6 sup kHtω kL (Lp (Rd ;X),Lp (Rd ;Y ))
ω
∞
X 
+ kSω(i,k) kL (Lp (Rd ;X),Lp (Rd ;Y )) .
k=2
i∈{0,1,2}

The first term here is estimated as in the proof of Theorem 12.4.12 by

4d ktkwbp(DRp ) . For the remaining sum over shifts, we obtain from Theorem
12.4.26 (using this theorem with trivial type t = 1 for small k, and as stated
for large k) that
∞
X
kSω(1,k) kL (Lp (Rd ;X),Lp (Rd ;Y ))
k=2
208 12 Dyadic operators and the T (1) theorem
 X
6 4 · βp,X βp,Y kSω(1,k) kShift(Rp ) · k
√
k:162k 612 d
X 
+ kSω(1,k) kShift(Rp ) ct0 ,Y ∗ ;p0 · k 1/t
√
k:2k >12 d
 X
6 cd βp,X βp,Y cK (Rp )k
√
k:162k 612 d
√
X 6 d 1/t 
1
+ ct0 ,Y ∗ ;p0 ωK (Rp ;
)k
√ 2k
k:2k >12 d
 
6 c0d βp,X βp,Y cK (Rp ) + ct0 ,Y ∗ ;p0 kωK
1
(Rp )kDini1/t

Similarly,
∞
X
kSω(2,k) kL (Lp (Rd ;X),Lp (Rd ;Y ))
k=2
 
6 c0d βp,X βp,Y cK (Rp∗0 ) + cq,X;p kωK
2
(Rp∗0 )kDini1/q0 .

Finally, The sum over shifts of type (0, k) may be estimated by either of the
two bounds above (the different numerical constant in Theorem 12.4.26 is in
any case absorbed into the unspecified dimensional constant). 

Remark 12.4.28. The norm estimate obtained in Theorem 12.4.27 via the rep-
resentation in terms of dyadic shifts is essentially the same as that in Theo-
rem 12.4.12 obtained via Figiel’s representation. While the proof of Theorem
12.4.27 partially relied on the proof of Theorem 12.4.12 to avoid repetition, a
larger part of the machinery behind the proof of Theorem 12.4.12, relying in
particular on Figiel’s Theorems 12.1.25 and 12.1.28 concerning his elementary
operators, was replaced in the proof of Theorem 12.4.27 by Theorem 12.4.26 on
the dyadic shifts, which in turn was based on the tangent martingale bounds
of Corollary 4.4.15.

12.5 Notes

Section 12.1

The Haar multipliers Hλ = Hααλ are special cases of martingale transforms dis-
cussed extensively in Volume I; see in particular Sections 3.5 and 4.2.e. In this
framework, the predictable sequences multiplying the martingale differences
X X
Dα
k f := hf, hα α
Q ihQ , D−α
k f := D−α
Q f
Q∈Dk Q∈Dk
 12.5 Notes 209

are then
X
vkα = λk 1Qk ∈ L∞ (σ(Dk ); L (X, Y )), vk−α ≡ 0.
Q∈Dk

On the other hand, the Haar multipliers Hαγ λ with α 6= γ already take a
departure from the general theory, and this is even more so with the general
operators of Figiel.
(Note that the conventional indices of dyadic analysis and martingale the-
ory are off by one from each other. In martingale theory, it is customary to
emphasise measurability, and hence the indices of martingale differences agree
with those of the σ-algebra that makes them measurable, while predictable
multipliers are measurable with respect to the “previous” σ-algebra with in-
dex k − 1. In dyadic analysis, the emphasis is on the supporting dyadic cubes,
and hence the “kth” martingale difference Dk f is the sum of the local mar-
tingale differences DQ supported, and averaging to zero, on the dyadic cubes
Q ∈ Dk , but then they are actually measurable only with respect to the “next”
σ(Dk+1 ); at the same time, the “predictable” multipliers are then measurable
with respect to the σ-algebra indicated by their index.)
The relaxed R-boundedness notion DR p of Definition 12.1.6 seems to
be new, but the slightly stronger E R p appears implicitly in Di Plinio, Li,
Martikainen, and Vuorinen [2020b, Remark 6.29], where it is shown that
the family |Q|−1 hT 1Q , 1Q i of Example 12.1.10 has this property when T ∈
L (Lp (Rd ; X), Lp (Rd ; Y )) and X and Y are UMD spaces; this also follows by
combining our Example 12.1.10 (on the DR p property of this family) and
Corollary 12.1.17 (the equivalence of DR p and E R p for UMD spaces). An
advantage of the new DR p is that it allows Example 12.1.10 without any
assumptions on the Banach spaces.
The exact characterisation of the boundedness of the Haar multipliers
Hλ in Theorem 12.1.11 is new; by Lemma 12.1.8 and Propositions 12.1.13
and 12.1.14, the characterising condition is strictly more general than the R-
boundedness condition kx 7→ R({λQ : x ∈ Q ∈ D})k∞ < ∞. This seems
at first to contradict Girardi and Weis [2005], where the necessity of uniform
pointwise R-boundedness for operator-valued martingale transforms is estab-
lished. This apparent contradiction is resolved by observing that, in order
to obtain this necessity of R-boundedness, Girardi and Weis [2005] actually
assume that their transforming sequence (vk )k>1 is allowed to multiply any
∞
subsequence (dfnk )k>1 of the martingale difference sequence (df Pk )k=1 , i.e.,
they assume the boundedness
P of the family of operators f 7→ k>1 vk dfnk
instead of just f 7→ k>1 vk dfk . In the case of Haar multipliers, this would
mean that, for a given sequence λ = (λQ )Q∈D we would consider a family of
operators including in particular all
X
α α
f 7→ λQ(k) hf, hQ ihQ ,
Q∈D
210 12 Dyadic operators and the T (1) theorem

where k ∈ N and Q(k) is the k generations larger dyadic ancestor of Q. How-

ever, in particular situations like that of Propositions 12.1.13, each coefficient
λQ is naturally associated to a unique cube Q only.
The underlying ideas of Section 12.1.b come from Figiel [1988], and they
have been developed further by Hytönen [2006], but substantial details of
the present treatment are new. Figiel [1988] also introduced the elementary
operators T and U and proved the first versions of Theorems 12.1.25 and
12.1.28. A novelty of the present treatment, also reflected in the auxiliary
considerations in Section 12.1.b, is to set up the argument in such a way
as to obtain a reasonably efficient dependence of the estimates on the UMD
constants, although we make no claims concerning sharpness. A technical
point was to use the decomposition of Lemma 12.1.22 in such a way that the
parts of the decomposition contribute additively, rather than multiplicatively,
to the operator norms in Theorems 12.1.25 and 12.1.28; while this seems
only natural in retrospect, it was not the case with earlier treatments of the
analogous bounds by Figiel [1988] and Hytönen [2006]. This proof detail only
affects the constants in the final estimates, which was not a concern in these
earlier works.
Besides the “dyadic singular integrals” studied in this section, there are re-
lated classes of operators that might be regarded as “dyadic pseudo-differential
operators”, in that their symbol depends on both the spatial variable s ∈ Rd
and the “dyadic frequency variable” I ∈ D. These are the generalised Haar
multipliers X
Hλ(s) f (s) = λI (s)hf, hI ihI (s),
I∈D

where each coefficient λI (·) is a function. A primary example considered by

Katz and Pereyra [1999] consists of
 w(s) t
λI (s) = wIt (s) := ,
hwiI

where t ∈ R and w is in a (dyadic) Ap or (dyadic) reverse Hölder class.

Given the close relation of their techniques to those of the present section,
it seems likely that some of the results concerning the operators Hλ(·) could
be generalised to functions taking values in a UMD space, but this line of
research seems not to have been pursued so far.

Section 12.2

In analogy with the quote of Stein [1982] on square functions at the beginning
of Chapter 9, also the concept of paraproduct is “not an idea in its pure form,
but rather takes various shapes depending on the uses it is put to”. A friendly
overview to this variety of “shapes and uses” of paraproducts can be found
in Bényi, Maldonado, and Naibo [2010]. Paraproducts were systematically in-
troduced by Bony [1981], but Bényi et al. [2010] convincingly argue that their
 12.5 Notes 211

first version is already implicit in the treatment of commutators of singular

integrals by Calderón [1965].
Our treatment concentrates on dyadic paraproducts. We are uncertain
about the earliest appearance of this notion in the literature but it was
certainly known to Figiel [1990]; according to this paper, the Lp (Rd ; X)-
boundedness of the dyadic paraproduct with a scalar-valued b ∈ BMOD (Rd )
“relies on an estimate due to Jean Bourgain (October 1987, unpublished)”.
This argument was only presented in print much later by Figiel and Woj-
taszczyk [2001]. In particular, Corollary 12.2.19 goes back to these works. The
first results on the boundedness of operator-valued paraproducts on UMD
spaces were obtained by Hytönen and Weis [2006b] for a Fourier-analytic
cousin of the dyadic paraproduct that we have treated. A sufficient condition
similar to Proposition 12.2.16, in terms of a version of the Carleson norm,
was identified there under the name of “Littlewood–Paley–BMO” norm. The
condition of Theorem 12.2.18, in terms of BMO(Rd ; Z) with values in a UMD
subspace Z ,→ L (X, Y ), is also implicit in Hytönen and Weis [2006b], and
explicitly formulated by Hytönen [2006]. However, both Hytönen and Weis
[2006b] and Hytönen [2006] also required an additional R-boundedness con-
dition, most easily formulated by the requirement that the unit ball B̄Z of
Z should be an R-bounded subset of L (X, Y ). This condition was found to
be superfluous by Hytönen [2014] when revising the argument for an exten-
sion to non-doubling measures, a generality that we have not considered here.
The details of the present approach are largely borrowed from Hänninen and
Hytönen [2016], where several simplifications were found when specialising the
considerations back to the case of the Lebesgue measure. A particular novelty
of Hänninen and Hytönen [2016], which we have followed, was to estimate
the vector-valued paraproduct directly in Lp (Rd ; Y ), in contrast to earlier ar-
guments that achieved the Lp bounds only via interpolation from auxiliary
end-point estimates between the Hardy space H 1 (Rd ; X) and L1 (Rd ; Y ) on
the one hand, and between L∞ (Rd ; X) and BMO(Rd ; Y ) on the other hand.
Theorem 12.2.25 on the boundedness of the symmetric paraproduct Λb is
from Hytönen [2021]. The case when p = 2 and X = Y is a Hilbert space was
obtained earlier by Blasco and Pott [2008], and extended to any p ∈ (1, ∞) and
any non-commutative Lp (M) space (with the same p) by Mei [2010]. (Recall
that Lp (M) is a UMD space for p ∈ (1, ∞)—the case of Schatten classes, due
to Bourgain [1986], is treated in Proposition 5.4.2, while the general case can
be found in Berkson et al. [1986b]—so the mentioned result of Mei [2010] is
indeed a special case of Theorem 12.2.25.) The auxiliary material on projective
tensor products is classical; much more on this topic can be found in Ryan
[2002].
Theorem 12.2.26 on the dimensional growth of the norms of operator-
valued paraproducts is from Mei [2006]. The optimal dimensional dependence
in the estimate
kΠb kL (L2 (R;`N
2 )) 6 ψ(N )kbkBMOso (R;L (`2 )) := ψ(N ) sup kb(·)ukBMO (R;`2 ) .
D N D N
u∈B̄`2
N
212 12 Dyadic operators and the T (1) theorem

had been settled some years earlier: Independently, Katz [1997] and Nazarov,
Treil, and Volberg [1997b] proved that ψ(N ) . 1 + log N , and the latter
authors also obtained the preliminary lower bound ψ(N ) & (1 + log N )1/2 .
This was improved to ψ(N ) & 1+log N by Nazarov, Pisier, Treil, and Volberg
[2002a]. For a while, there were hopes in the air of obtaining a dimension-free
estimate with BMOD (R; L (`2N )) in place of BMOD so
(R; L (`N
2
)) on the right.
Some indications that made this plausible are discussed in the introduction
of Mei [2006] who, however, destroyed such hopes were by the main result of
that paper, reproduced as Theorem 12.2.26. In combination with the upper
bound by Katz [1997] and Nazarov et al. [1997b] just mentioned, it shows that
1 + log N is the optimal upper bound for kΠb kL (L2 (R;`2N )) /kbkF (R;L (`2N )) for
any of the choices F ∈ {BMOso ∞
D , BMOD , L }.
Further relations between various BMO-type quantities and the norms
of related transformations in infinite-dimensional Hilbert spaces have been
studied by Blasco and Pott [2008, 2010]. Analogous results in the context
of the operator-valued BMOA space of analytic functions are due to Rydhe
[2017].

Section 12.3

We refer the reader to the Notes of the following section for an account of
the T (1) theorem in its more traditional meaning as a boundedness criterion
for Calderón–Zygmund operators (as in the title of David and Journé [1984]).
The section under discussion presents a rather non-canonical approach to this
theory, introduced and described by Figiel [1990] as follows:
Our approach is indirect in the following sense. Rather than trying
to prove that some “classical” operators are bounded, we start from
considering certain rather new operators, which in our opinion have
a basic nature. (All the “singularities” which can occur in our con-
text are neatly packaged inside the basic operators.) Having estab-
lished precise estimates for the norms of those basic operators, we can
take up the “general case”. We just look at the class of those oper-
ators which can be realised as the sum of an absolutely convergent
(in the operator norm) operator series whose summands are simple
compositions of our basic operators. Then it turns out that the choice
was sufficiently efficient for that class to contain so-called generalised
Calderón–Zygmund operators and much more.
A large part of this section, up to and including T (1) Theorem 12.3.26, is an
updated review of Figiel [1990], incorporating a few elaborations:
• the trade-off between the type and cotype properties of the underlying
spaces and the minimal rate of convergence of the coefficients of the bilinear
form, as in Theorem 12.3.26(ii) (which is implicit in the combination of
Figiel [1988, 1990]);
 12.5 Notes 213

• conditions involving R-boundedness to deal with operator-valued versions

(first introduced into the context of T (1) theorems at large by Hytönen
and Weis [2006b] and into Figiel’s approach by Hytönen [2006]);
• keeping track of, and optimising the argument for, the quantitative de-
pendence on parameters like the UMD constants (which seems new for
this “non-random” version of the T (1) theorem, involving—in contrast to
Theorem 12.3.35—one dyadic system D only).
The decomposition (12.36) of t(f, g) into three one-parameter series, in
contrast to the perhaps more obvious two-parameter decomposition
X
t(f, g) = t(Di f, Dj g),
i,j

was already used by Figiel [1990], but it is frequently referred to as the “BCR
algorithm” after Beylkin, Coifman, and Rokhlin [1991]. They explored its ad-
vantages for the numerical evaluation of singular integrals, also making a con-
nection with the T (1) theorem but apparently independently of Figiel [1990].
A decade later in 2002, when two of the present authors started to investigate
a Banach space valued T (1) theorem (eventually published in Hytönen and
Weis [2006b]), they were also initially unaware of the work of Figiel [1990],
which was first brought to their attention by Hans-Olav Tylli. Ever since, the
approach of Figiel [1990] has been highly influential for the development of
the theory of Banach space valued singular integrals.
The second T (1) Theorem 12.3.35, which makes use of a random choice of
the dyadic system D ω , has a history of its own. This method, referred to by
its inventors as “pulling ourselves by hair”, was introduced by Nazarov, Treil,
and Volberg [1997a] to tackle the difficulties in estimating singular integrals
with respect to a non-doubling measure µ, thus going beyond the established
theory in spaces of homogeneous type due to Coifman and Weiss [1971]. Their
original idea consisted of splitting a function into its “good” and “bad” parts,
according to the “good” and “bad” cubes supporting the martingale differ-
ences DQ f : X X
ω ω
fgood := DQ f, fbad := DQ f,
ω
Q∈Dgood ω
Q∈Dbad

and showing that the latter is small, “on average”, with respect to a random
choice of ω:
ω
Ekfbad kL2 (µ) 6 εkf kL2 (µ) .
As a result, it is enough to estimate (an a priori bounded) operator T of
“good” functions only. Namely, if
ω ω ω ω
|hT fgood , ggood i| 6 Ckfgood k2 kggood k2 6 Ckf k2 kgk2 ,

then
214 12 Dyadic operators and the T (1) theorem
ω ω ω ω ω
|hT f, gi| 6 |hT fgood , ggood i| + |hT fgood , gbad i| + |hT fbad , gi|
ω ω
6 Ckf k2 kgk2 + kT kkf k2 kgbad k2 + kT kkfbad k2 kgk2 .

Taking the expectations of both sides, it follows that

|hT f, gi| 6 Ckf k2 kgk2 + 2εkT kkf k2 kgk2 ,

hence
C
kT k 6 C + 2εkT k, kT k 6 .
1 − 2ε
This method was successfully applied and further developed by Nazarov, Treil,
and Volberg [2002b, 2003]. The latter work was extended to Banach space
valued singular integrals with respect to non-doubling measures by Hytönen
[2014]. The first arXiv version of this paper was posted already in 2008, and
hence it was available to provide the backbone for the proof of the A2 theorem
in Hytönen [2012] (arXiv 2010); see the Notes of Chapter 11 for more on the
latter. It was for the purposes of the A2 theorem that a technical elaboration of
the averaging method of Nazarov, Treil, and Volberg [1997a, 2002b, 2003] had
to be invented: “on average”, the bad part is not only small but completely
absent. This allows the replacement of the estimates above by identities of the
type
ω
hT f, gi = EhTgood f, gi.
The observation that one can combine this averaging method with Figiel’s de-
composition of singular integrals in order to simplify the latter, and thereby
obtain sharper quantitative conclusions (notably, a quadratic dependence on
the UMD constant), was then made in Hytönen [2012] (arXiv 2011), where
a version of Theorem 12.3.35 (for scalar kernels and under vanishing para-
product conditions) was first established. The question of obtaining a linear
dependence on the UMD constant is an outstanding open problem already
in the special case of the Hilbert transform (see Problem O.6); but of course
a possible counterexample could be more feasible within the larger class of
operators covered by Theorem 12.3.35. A positive answer has been obtained
for sufficiently smooth even singular integrals on Lp (R; X) by Pott and Sto-
ica [2014]; their result depends on the same averaging trick and the resulting
dyadic representation theorem, but then applies different techniques to com-
plete the estimate.
While our approach to the “random” T (1) Theorem 12.3.35 took a detour
via the “non-random” T (1) Theorem 12.3.26, we should emphasise that this is
by no means necessary; rather, in many recent extensions of the T (1) theorem,
one starts with the randomised set-up from the beginning, and it is often not
even clear whether this could be avoided. We will say more about some of
these extensions later in these Notes. The reasons that we have chosen to
present also the non-random T (1) Theorem 12.3.26 are (at least) two-fold:
On the one hand, we feel that there is some historical documentary value
in providing (probably) the first detailed exposition of the original Banach
 12.5 Notes 215

space valued T (1) theorem of Figiel [1990], considering also the number of
other results in the literature relying on this in their proofs (although, in
many cases, one could alternatively apply one or several of the more recent
variants). On the other hand, the non-random T (1) Theorem 12.3.26 is not in
all respects subsumed by the random T (1) Theorem 12.3.35, which makes the
first one applicable in some situations where the latter one is not, and it might
hence be useful for the reader to keep the original T (1) Theorem 12.3.26 in
their toolbox.
While we are not aware of many such applications, here is at least one:
Pseudo-localisation of singular integrals refers to estimates of the form

k1{Σf,s T f kLp (Rd ;X) 6 φ(s)kf kLp (Rd ;X) , s ∈ N,

where
(s) (s)
[ X
Σf,s := {9Q : Q ∈ D, DQ f 6≡ 0}, DQ f := DR f,
R∈chs Q

and the point is obtaining a quantitative decay φ(s) → 0 as s → ∞. Case

p = 2 was considered by Parcet [2009] for X = K and by Mei and Parcet
[2009] for a Hilbert space X, with applications to non-commutative Calderón–
Zygmund and Littlewood–Paley theory, respectively. An extension to p ∈
(1, ∞) and a UMD space X was obtained by Hytönen [2011] using a version
of the T (1) Theorem 12.3.26. This leads to studying a bilinear form whose
Haar coefficients satisfy a non-standard estimate of the form
α
|t(hQ , hγQ+̇m )| . |m|−(d+ε) 1(2·2s ,∞) (|m|) + |m|−d 1(4·2s −2,4·2s +2) (|m|).

The first term on the right with decay d + ε is typical, but the second one,
without any ε, is not. However, this term is only supported in a relatively
narrow region of values of the parameter m ∈ Zd , which still allows one to
make favourable estimates of the Figiel norms of t.
A notable aspect of this application is that the construction of the set
Σf,s refers to a fixed dyadic system D, which calls for a Haar expansion of
the operator in terms of this same D, as in the non-random T (1) Theorem
12.3.26, and seems to prevent any effective application of the random systems
D ω , as in the random T (1) Theorem 12.3.35. This suggests that, even after the
successful recent (and very likely future) development of T (1) theorems and
other results based on random dyadic systems, the non-random T (1) Theorem
12.3.26 might not become completely obsolete.

Section 12.4

The classical theory of Calderón and Zygmund [1952] had its focus on con-
volution operators. Their L2 (Rd ) boundedness is amenable to methods of
Fourier analysis, which then serves as a starting point for extrapolation to
216 12 Dyadic operators and the T (1) theorem

other Lp (Rd ) and different function spaces, as discussed at length in Chap-

ter 11. It was observed quite early, notably by Coifman and Weiss [1971],
that these extrapolation aspects of the theory could be extended to much
greater generality, certainly including non-convolution operators on Rd and
much more. On the other hand, the boundedness of some prominent non-
convolution operators was obtained by different methods over the years, in-
cluding the commutators of Calderón [1965, 1977], and the Cauchy integral
on a Lipschitz graph, which we give in the parametrised form
Z ∞
f (t) dt
CA f (s) := p.v. .
−∞ s − t + i(A(s) − A(t))

The boundedness of CA was first established, in the case of a small Lips-

chitz constant kAkLip , by Calderón [1977], and eventually in full generality by
Coifman, McIntosh, and Meyer [1982]. However, a general criterion for verify-
ing the L2 (Rd ) boundedness of any given Calderón–Zygmund operators was
missing.
The first such general criterion was provided by the “T (1) theorem” of
David and Journé [1984]. In its original formulation, this theorem stated that
an operator T : S (Rd ) → S 0 (Rd ), with a Calderón–Zygmund standard ker-
nel, extends to a bounded operator on L2 (Rd ), if and only if it satisfies the
following three conditions, from which the name of the theorem (also intro-
duced by David and Journé [1984] in the title of the first section of their
paper) is derived:
(i) T (1) ∈ BMO(Rd ),
(ii) T ∗ (1) ∈ BMO(Rd ),
(iii) T has the weak boundedness property.
Despite being a complete and elegant characterisation, giving, e.g., the results
of Calderón [1977] as a quick corollary, it turned out that it is not always fea-
sible to use this theorem for some operators. As a prime example, the theorem
of Coifman, McIntosh, and Meyer [1982] could not be directly recovered by
David and Journé [1984], since CA (1) does not admit an expression whose
BMO norm could be easily estimated.
This shortcoming was fixed by the more general “T (b) theorem” of David,
Journé, and Semmes [1985], which replaced (i) and (ii) by the more flexible
conditions
(i’) T (b1 ) ∈ BMO(Rd ),
(ii’) T ∗ (b2 ) ∈ BMO(Rd ),
where one is free to choose the pair of functions bi ∈ L∞ (Rd ) subject only
to the restriction that they be accretive (meaning <bi > δ > 0 almost every-
where) or just para-accretive (a technical generalisation, for which we refer the
interested reader to the original paper). In particular, one can take bi = 1+iA0
for which the computation of (any finite truncations of) CA (1 + iA0 ) is easy.
 12.5 Notes 217

While also this T (b) theorem has been extended to UMD spaces by
Hytönen [2006], the need for this is perhaps not as great as in the scalar-
valued case, at least as far as the extension of the boundedness of scalar-
valued Calderón–Zygmund operators to Lp (Rd ; X) is concerned. The reason
for this is that, while it might be difficult to check the T (1) conditions (i) and
(ii) directly, they can nevertheless be verified by the converse direction of the
T (1) theorem, provided that the L2 (Rd ) boundedness of T is already known
by some other method (such as the T (b) theorem). This is, in essence, the
point of the scalar-kernel T (1) Theorem 12.4.21.

Corollary 12.5.1. Let X be a UMD space, p ∈ (1, ∞), and A : R → R be a

Lipschitz function. Then the Cauchy integral on a Lipschitz graph CA extends
to a bounded operator on L2 (R; X) and

kCA kL (Lp (R;X)) 6 cA pp0 · β2,X

2
,

where cA is a constant that depends on A only.

Sketch of proof. By the theorem of Coifman, McIntosh, and Meyer [1982], the
operator CA is bounded on L2 (R). It is straightforward to verify that the kernel
of CA is a standard kernel, and hence verifies the assumptions of Theorem
12.4.21 with Dini1 conditions (and associated constants depending only on A),
in which case only trivial type and cotype is needed. Thus Theorem 12.4.21,
with s = p = 2, proves the corollary for p = 2. While we could apply Theorem
12.4.21 with s = 2 and any p ∈ (1, ∞), a better quantitative conclusion for
p 6= 2 is obtained by using case p0 = 2 as input to the Calderón–Zygmund
theorem 11.2.5, which then yields the asserted bound for all p ∈ (1, ∞). 
Corollary 12.5.1 seems to have been first stated in Hytönen [2006]; however,
given that it is essentially a concatenation of its scalar case due to Coifman,
McIntosh, and Meyer [1982], and the T (1) theorem of Figiel [1990], it was
probably “known to experts” much earlier. The case when X is a UMD lattice
was established by a different method already by Rubio de Francia [1986].
In a similar way, the extension of the non-homogeneous T (1) theorem of
Nazarov, Treil, and Volberg [2003] to UMD spaces has the following conse-
quence:

Theorem 12.5.2. Let µ be a positive non-atomic Radon measure on C. Then

the following conditions are equivalent
(1) There is a constant c < ∞ such that, for every disk D = D(z, r) ⊆ C, the
measure µ satisfies
(a) the linear growth condition µ(D(z, r)) 6 cr, and
(b) the local curvature condition
ZZZ
dµ(u) dµ(v) dµ(z)
6 cµ(D),
D×D×D R(u, v, z)
218 12 Dyadic operators and the T (1) theorem

where R(u, v, z) is the radius of the circle through u, v, z (understood

as ∞, if the points are collinear).
(2) The Cauchy integral
Z
f (v) dµ(v)
Cµ f (u) :=
C u−v

defines a bounded operator on L2 (µ).

(3) For every UMD space X and every p ∈ (1, ∞), the Cauchy integral Cµ
defines a bounded operator on Lp (µ; X).

Note that CA is (equivalent to) the special case, where µ is the arc-length
measure on the graph {(t, A(t)) : t ∈ R}.
Sketch of proof. The implication (2)⇒(1a) is due to David [1991] and (2)⇒(1b)
due to Melnikov and Verdera [1995] and Mattila, Melnikov, and Verdera
[1996]. The sufficiency of these geometric conditions, (1)⇒(2), was proved
by Tolsa [1999].
The implication ((1a) and (2))⇒(3) follows from an analogue of Theorem
12.4.21 for measures on Rd with the power growth bound µ(B(s, r)) 6 crn
(0 < n 6 d), which is one of the main results of Hytönen [2014]. The implica-
tion (3)⇒(2) is trivial. 
This proof sketch highlights the role of T (1) theorems as a device for extending
deep results about the boundedness of specific operators from scalar-valued to
vector-valued spaces, without the need to revisit the details of the original ar-
guments. Indeed, by using the scalar-valued result (2) as an intermediate step,
the equivalence of (1) and (3) is obtained without ever having to deal with
the local curvature condition (1b) in the context of vector-valued functions!
Our operator-kernel T (1) Theorem 12.4.12 is the outcome of a line of
evolution starting with the first such results obtained by Hytönen and Weis
[2006b] and Hytönen [2006], and continued with several variants and exten-
sions addressing
• non-homogeneous measures (Hytönen [2014] (arXiv 2008), Martikainen
[2012a] (arXiv 2010), Hytönen and Vähäkangas [2015]);
• simplifications of the underlying decomposition of the operator (Hytönen
[2012], Hänninen and Hytönen [2016]);
• sharper conclusions under additional symmetry assumptions (Pott and
Stoica [2014], Hytönen [2021]);
• product-space/multiparameter singularities (Di Plinio and Ou [2018], Hytönen,
Martikainen, and Vuorinen [2019a]);
• multilinear operators (Di Plinio, Li, Martikainen, and Vuorinen [2020b],
Airta, Martikainen, and Vuorinen [2022]).
 12.5 Notes 219

While these papers extend the theory into several directions that we have not
considered here, many of them also provide valuable pieces of insight into the
basic case of linear Calderón–Zygmund operators on Rd with the Lebesgue
measure, which we have tried to incorporate into the present treatment. De-
spite this extensive background material, some aspects of our present T (1)
Theorem 12.4.12 appear to be new:
(1) For the first time, we are able to state an operator-valued T (1) theorem
that gives a characterisation (as in the scalar-valued T (1) theorem of
David and Journé [1984]), and not just a sufficient condition (as in all
operator-valued papers cited above), for the boundedness of a Calderón–
Zygmund operator with an operator-valued kernel. This depends on two
recent ideas, the combination of which appears here for the first time:
(a) Replacing the (sufficient but not necessary) weak R-boundedness
property of most of the previous contributions by the correct weak
DR p -boundedness property. As discussed in the Notes of Section 12.1,
this idea is from Di Plinio, Li, Martikainen, and Vuorinen [2020b].
(b) Treating the bi-paraproduct Λ = ΠT (1) + ΠT∗ ∗ (1) as a single object,
and making its boundedness into a condition in its own right, rather
than trying (in vain) to force it into a form involving some operator-
valued BMO space. This is implicit in Hytönen [2021].
(2) Recording the quantitative dependence of the estimate in terms of both
the UMD and the (co)type constants, and optimising the argument for
what seems to be the best possible bound currently available. This was
available in important special cases (notably in Hytönen [2012]), and ar-
guably implicit in some other works, but seems to be original as an explicit
statement in the present generality.

Consequences of the T (1) theorem

The “T (1) theorem for convolution kernels”, Corollary 12.4.13, is a somewhat

untypical statement, in that convolution kernels have been usually treated by
more traditional Fourier-analytic methods, rather than the T (1) technology.
As such, this very formulation seems to be new. However, essentially the same
class of operators was considered with Fourier methods by Hytönen and Weis
[2007]. (Despite the publication year, this paper was actually the first joint
project of its authors, which they completed and submitted in 8/2002, be-
fore starting their follow-up work on the T (1) theorem, Hytönen and Weis
[2006b], later in the same year.) In place of the combinatorial estimates for
Figiel’s operators from Sections 12.1.b and 12.1.c, this proof employed analo-
gous Fourier-analytic estimates due to Bourgain [1986]. Just like the combi-
natorial details of the T (1) theorem can be simplified with the random dyadic
systems, the proof of the key lemma of Bourgain [1986] was later simplified
in Hytönen [2012] by the same technology.
While the direct comparison of Corollary 12.4.13 with the results of
Hytönen and Weis [2007] is complicated by the presence in Corollary 12.4.13 of
220 12 Dyadic operators and the T (1) theorem

the (untypical in the classical theory) weak boundedness property, Corollary

12.4.19 on antisymmetric kernels comes rather close to some results of Hytönen
and Weis [2007]. Indeed, in this special situation, one can completely avoid
both paraproducts and the weak boundedness property, obtaining a bound-
edness criterion in terms of the Calderón–Zygmund kernel bounds alone.
Corollary 12.4.18 on antisymmetric but non-convolution kernels (where
the weak boundedness is automatic but a paraproduct is present) is probably
new in the operator-valued setting, but a rather straightforward adaptation
of similar statements that are well known in the scalar-valued theory.

On minimal smoothness conditions

As one can see from T (1) Theorem 12.4.12 and its corollaries, the minimal
smoothness of the kernel involves a modulus of continuity kωkDiniσ , where
σ = max(1/t, 1/q 0 ) if X has cotype q and Y has type t, or one of them has
both. In the scalar-valued (or more generally Hilbert space) case, this reduces
to σ = 21 . Incidentally, this appears to be the minimal condition required to
run any known proof of the T (1) theorem, even in the scalar case. As Figiel
[1990] puts it,
it was a nice surprise that such austere methods could in fact lead to
some results which were not less general than their counterparts estab-
lished earlier with no restrictions on the range of admissible methods.
While the original T (1) theorem of David and Journé [1984] and most of its
successors are formulated for Calderón–Zygmund standard kernels, an exten-
sion to Dini-type conditions was obtained shortly after by Yabuta [1985], who
1
proved the theorem under the condition that kω 3 kDini < ∞. It is not obvious
at first sight how this compared to Figiel’s condition kωk 1 < ∞. However,
Dini 2
we may observe that any non-decreasing ω on [0, 1] satisfies
Z 1 Z 1  Z 1 ds α dt
 1 α dt 1 α
ω(t) log = ω(t) 1+α ω(t) 1+α
0 t t 0 t s t
Z 1 Z 1 α
1
 1 ds  dt
6 ω(t) 1+α ω(s) 1+α
0 t s t
Z 1 Z 1
1 dt  1 ds α
6 ω(t) 1+α ω(s) 1+α
0 t 0 s
Z 1 1 ds 1+α

= ω(s) 1+α .
0 s
1 1
With 1+α = 13 , we see that Yabuta’s kω 3 kDini dominates kωkDiniα with α = 2.
(While the Dinis norms were previously defined with log2 in place of log,
and integrating over [0, 21 ] instead of [0, 1], the reader may easily verify that,
extending ω from [0, 21 ] to [0, 1] by ω(t) := ω(min(t, 21 )), these details affect at
most the constants in the final conclusions.)
 12.5 Notes 221

Subsequently, Meyer [1986] (according to Han and Hofmann [1993], but we

have not been able to verify the original reference) relaxed the assumption to
α = 1 (plus a further weakening of the pointwise bounds to integral conditions
rather closer to the Figiel conditions for bilinear forms as in our abstract
T (1) Theorems 12.3.26 and 12.3.35). Han and Hofmann [1993] obtained a
further slight relaxation of the conditions of Meyer [1986], and Yang, Yan, and
Deng [1997] proved the T (1) theorem with assumptions essentially matching
the special case α = 12 of the conditions of Figiel [1990] in the scalar-case.
Later attempts to relax this condition were made by Grau de la Herrán and
Hytönen [2018], who found that the same regularity is sufficient also for the
non-homogeneous T (1) theorem, but did not succeed in relaxing it even in the
standard case. Thus, various different proof strategies all seem to meet this
same threshold.
At the same time, it seems to remain unknown whether even the much
weaker Hörmander conditions of Definition 11.2.1 could in principle be enough
for a T (1) theorem. A positive result in this direction seems out of reach with
the presently available methods, but there does not seem to be any definitive
counterexample to rule out this possibility. As very partial evidence for a
counterexample, Yang, Yan, and Deng [1997] show that the T (1) conditions
for a Hörmander kernel are insufficient to guarantee the boundedness in some
end-point spaces.

The dyadic representation theorem

A dyadic representation formula resembling Theorem 12.4.27 was first ob-

tained by Hytönen [2012] as a key component of the original proof of the A2
Theorem 11.3.26 for all standard Calderón–Zygmund operators in the scalar-
valued case. Subsequent refinements and simplifications of the original rep-
resentation were obtained by Hytönen, Pérez, Treil, and Volberg [2014], and
Hytönen [2017]. The first version of both Theorems 12.4.26 and 12.4.27 for
dyadic shifts and singular integrals on Lp (Rd ; X) with operator-valued kernels
were obtained by Hänninen and Hytönen [2016], by essentially the same tech-
niques (notably, the tangent martingale estimates of Corollary 4.4.15) that
we have followed. In all these contributions, like several other contemporary
ones, the notion of dyadic shift was essentially that of Hytönen [2012], which
is somewhat different from the present Definition 12.4.24. In the shifts of
Hytönen [2012], the components AK take the form
X γ
P
AP f = αQ,R hf, hα
Q ihR ,
Q∈ch(i) (P )
R∈ch(j) (P )

with two independent complexity parameters (i, j) ∈ N2 in place of the single

k > 2 in Theorem 12.4.27. The “new shifts” of Definition 12.4.24 were first
introduced by Grau de la Herrán and Hytönen [2018]. Their Banach space
valued theory, including Theorems 12.4.26 and 12.4.27 in essentially their
222 12 Dyadic operators and the T (1) theorem

present form, as well as multilinear extensions, has been developed by Airta,

Martikainen, and Vuorinen [2022].
As far as proving the T (1) theorem for Calderón–Zygmund operators on
Lp (Rd ; X) is concerned, the advantages of the Dyadic Representation Theo-
rem 12.4.27 over (the randomised version of) Figiel’s representation may be
considered a question of mathematical taste (depending, among other things,
on one’s preference for the tangent martingales methods of Section 4.4 over
the dyadic singular integrals of Section 12.1 or vice versa). However, these
advantages become prominent in extensions of the T (1) theory to other situ-
ations that we have not treated here. Roughly speaking, the decomposition of
Figiel is essentially based on multi-scale versions of translations—reasonably
well-behaved objects as far as translation-invariant spaces like Lp (Rd ; X) are
concerned, but somewhat unstable in more general situations. In contrast, the
basic building block AK of the dyadic shifts are essentially averages, which
are much more stable operations. In particular, the averages f 7→ 1Q hf iQ
over arbitrary cubes Q ⊆ Rd are uniformly bounded on Lp (w) if and only if
w ∈ Ap , which partially explains the usefulness of such objects in the origi-
nal context of proving the A2 theorem. Averages are somewhat well-behaved
even when taken with respect to non-doubling measures, which is the con-
text in which a certain precursor of the dyadic representation of Hytönen
[2012] (arXiv 2010) was established by Hytönen [2014] (arXiv 2008) in order
to extend the non-homogeneous T (1) theorem of Nazarov, Treil, and Volberg
[2003] to the Banach space valued setting. Conversely, after the discovery of
the Dyadic Representation Theorem, it was used by Volberg [2015] to give a
new proof of the non-homogeneous T (1) theorem.
An adaptation of the Dyadic Representation Theorem 12.4.27, by Hytönen,
Li, H., and Vuorinen [2022], was instrumental in extending the T (1) the-
ory to singular integral operators adapted to so-called Zygmund dilations
(x1 , x2 , x3 ) 7→ (sx1 , tx2 , stx3 ), where s, t > 0 are two independent parame-
ters. Variants of the Dyadic Representation Theorem 12.4.27, with the Haar
functions replaced by smoother wavelets, have been explored by Hytönen and
Lappas [2022], Di Plinio, Wick, and Williams [2023c], and Di Plinio, Green,
and Wick [2023b,a].

T (1) theorems on other function spaces

The original T (1) theorem of David and Journé [1984] was a characterisation
of boundedness on L2 (Rd ), while we have dealt with extensions of such re-
sults to Lp (Rd ; X). However, the boundedness of a given (singular integral)
operator is basic question arising in several other function spaces as well, and
the T (1) theorem has served as a model for similar results in other spaces.
(See Chapter 14 for information about the functions spaces appearing in this
s
discussion.) Extensions of the T (1) theorems to Besov spaces Ḃp,q were ob-
s s
tained by Lemarié [1985] and to Triebel–Lizorkin spaces Ḃp,q and Ḟp,q by
Frazier, Han, Jawerth, and Weiss [1989]. In these results, p, q ∈ [1, ∞], and
 12.5 Notes 223

the smoothness parameter s was restricted by the Hölder exponent of the

standard kernel of T . In order to cover a broader range of Besov and Triebel–
Lizorkin spaces, where the smoothness index can take any value s ∈ R, it is
necessary to consider higher order Calderón–Zygmund estimates such as

|∂ α K(s, t)| 6 C|s − t|−d−|α| .

With appropriate assumptions of this type in place, Frazier, Torres, and

Weiss [1988] and Torres [1991] obtained T (1) criteria for the boundedness
s
of Calderón–Zygmund operators on any Triebel–Lizorkin space Ḟp,q , where
s ∈ R and p, q ∈ (0, ∞]. The precise assumptions are necessarily somewhat
technical, and the result splits into three cases, where s < 0; or s > 0 and
p, q ∈ [1, ∞], or s > 0 and min(p, q) ∈ (0, 1).
In a limited range of s again, T (1) theorems on (scalar-)weighted Triebel–
s
Lizorkin spaces Ḟp,q (w) were obtained by Han and Hofmann [1993], and on
s
matrix-weighted Besov spaces Ḃp,q (W ) by Roudenko [2003]. The full scale
s
of both matrix-weighted Besov and Triebel–Lizorkin spaces Ḃp,q (W ) and
s
Ḟp,q (W ) (as well as further generalisations with a fourth index) was covered
by Bu, Hytönen, Yang, and Yuan [2023]. When restricted to the unweighted
case, this last work even slightly simplifies the assumptions of Frazier, Torres,
and Weiss [1988] and Torres [1991].
In all these mentioned works on T (1) theorems beyond Lp spaces, the
focus has been on special T (1) theorems providing sufficient conditions for
boundedness under vanishing paraproduct assumptions. General T (1) the-
orems, providing a characterisation of boundedness on a given space, were
s
obtained on Besov spaces Ḃp,q of positive smoothness s > 0 by Youssfi [1989],
in terms of the weak boundedness property and the boundedness of higher
order paraproducts. A far-reaching extension to Triebel–Lizorkin and other
function spaces, including versions on quite general domains O ⊆ Rd , is due
to Di Plinio, Green, and Wick [2023a].
For Banach space valued functions, special T (1) theorems (i.e., with van-
ishing paraproduct assumptions) on Riesz potential spaces Ḣ s,p (Rd ; X) and
s
Besov spaces Ḃp,q (Rd ; X) were proved by Kaiser [2007, 2009], respectively. The
results in Ḣ s,p (Rd ; X) need the UMD property of X, but those in Ḃp,qs
(Rd ; X)
do not. While we are not going to discuss these specific results in any further
detail, the reader can witness a similar dichotomy—that the UMD property is
needed to obtain results in certain function spaces, but not for analogous re-
sults in certain others—in our discussion of the theory of Banach space valued
function spaces in Chapter 14.
13
The Fourier transform and multipliers

In this chapter, we complement the discussion of three major themes of Fourier

analysis that we have studied in the previous Volumes. The first one is the
Banach space valued Hausdorff–Young inequality

kfbkLp0 (Rd ;X) 6 Ckf kLp (Rd ;X) . (13.1)

As we recall from Section 2.4.b, this is a non-trivial condition, expressed by

saying that the space X have Fourier type p. The basic theory around this
notion was already developed in 2.4.b, but we now turn to the main result on
this topic, Bourgain’s Theorem 13.1.33, which says that (13.1) holds for some
p > 1 if and only if X has some non-trivial type. Section 13.1 is dedicated to
a detailed proof of this deep result.
The second theme is about connecting the Fourier multipliers Tm : f 7→
(mfb)∨ from Chapter 5 and Section 8.3 with the Calderón–Zygmund theory
of Chapter 11. In principle, we have

Tm f = (mfb)∨ = m
b ∗ f = k ∗ f,

where the right-hand side has the formal structure of the operators studied in
Chapter 11, but the question then becomes the correspondence of the condi-
tions on the multiplier m and on the singular convolution kernel k. As we will
see in Section 13.2.a, the function k will be a nice Calderón–Zygmund kernel,
and hence f 7→ k ∗ f will be in the scope of all results of Chapter 11 (notably,
including those dealing with extrapolation of boundedness to the weighted
Lp (w; X) spaces), as soon as m satisfies assumptions like those in the Mihlin
Multiplier Theorem 5.5.10 for sufficiently many derivatives ∂ α m. Moreover,
this result is very general in that it holds for multipliers taking values in arbi-
trary Banach spaces, and then in particular in L (X, Y ) for any Banach spaces
X and Y . However, the required number of derivatives on this level of gener-
ality is higher than that in the Mihlin Multiplier Theorem 5.5.10. Coping only
with the same derivatives as in Mihlin’s theorem turns out to be more deli-
cate and require the use of a Banach space valued Hausdorff–Young inequality
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 225
T. Hytönen et al., Analysis in Banach Spaces, Ergebnisse der Mathematik und ihrer
Grenzgebiete. 3. Folge / A Series of Modern Surveys in Mathematics 76,
https://doi.org/10.1007/978-3-031-46598-7_3
226 13 The Fourier transform and multipliers

(13.1). It will be convenient to know, thanks to Bourgain’s Theorem 13.1.33,

that this estimate is always available in the UMD spaces that we so frequently
deal with (recalling that every UMD space has non-trivial type by Proposition
7.3.15). As we have already seen in a number of occasions (notably, Bourgain’s
Theorem 5.2.10 on the Hilbert transform, and Guerre-Delabrière’s Theorem
10.5.1 on the imaginary powers (−∆)is of the Laplacian), the UMD condition
is often necessary for the theory that we develop.
As the third topic of this chapter, we complement these result by The-
orem 13.3.5 of Geiss, Montgomery-Smith, and Saksman, which significantly
extends the previous examples of Fourier multipliers whose Lp (Rd ; X) bound-
edness implies the UMD condition. As one of its consequences, in Corollary
13.3.9, we are able to compete the characterisation of situations in which there
is a continuous embedding H k,p (Rd ; X) ,→ W k,p (Rd ; X) between two classes
of classical function spaces studied in the previous Volumes. This also pro-
vides a link with the following Chapter 14, where we undertake a systematic
development of the theory of function spaces of Banach space valued functions.
Despite the interconnected themes of the three sections of this chapter, any
of them can be studied independently of the other two by a reader interested
in a particular topic.

13.1 Bourgain’s theorem on Fourier type

Already in Section 2.4.b, we discussed in some detail the notion of Fourier
type, or the extent to which the Hausdorff–Young inequality kfbkp0 6 Ckf kp
remains valid for the Fourier transform of vector-valued functions. In the
Notes of Chapter 2, we also mentioned without proof the main theorem on
this topic, due to Bourgain, stating that non-trivial type implies non-trivial
Fourier type (and hence is equivalent to it, the other direction being a rather
easier Proposition 7.3.6). The aim of this section is to prove this fundamental
result, which will also play a role in the subsequent parts of the book.
We recall from Proposition 2.4.20 that the Fourier type p ∈ [1, 2] of a
Banach space X can be defined by any of the following equivalent conditions,
where moreover any choice of d ∈ Z+ is equivalent by Proposition 2.4.11:
(1) The Fourier transform on Rd , defined on f ∈ L1 (Rd ; X) by
Z
f (ξ) =
b f (x)e−2πix·ξ dξ, ξ ∈ Rd ,
Rd
0
extends to a bounded operator from Lp (Rd ; X) to Lp (Rd ; X).
(2) The Fourier transform on Td , defined on f ∈ L1 (Td ; X) by
Z
f (k) =
b f (t)e−2πit·k dt, k ∈ Zd ,
Td
0
restricts to a bounded operator from Lp (Td ) to `p (Zd ).
 13.1 Bourgain’s theorem on Fourier type 227

(3) The Fourier transform on Zd , defined on x = (xk )k∈Zd ∈ `1 (Zd ; X) by

X
x
b(t) = e−2πik·t xk , t ∈ Td ,
k∈Z

0
extends to a bounded operator from `p (Zd ; X) to Lp (Td ; X).
Denoting the norms of the respective extensions (or restrictions) by ϕp,X (Rd ),
ϕp,X (Td ) and ϕp,X (Zd ), we have:

Proposition 13.1.1. Let X be a Banach space, p ∈ (1, 2] and d ∈ Z+ . Then

d
ϕp,C (Rd−1 )ϕp,X (R) 6 ϕp,X (Rd ) 6 ϕp,X (R) , (13.2)

ϕp,X ∗ (Td ) = ϕp,X (Zd ) ϕp,X (Rd )

 
ϕp,X (Rd ) = ϕp,X ∗ (Rd ) 6 d d 6 . (13.3)
ϕp,X (T ) = ϕp,X ∗ (Z ) ϕp,C (Rd )
0
It is actually known that ϕp,C (Rd ) = (p1/p (p0 )−1/p )d . For the purposes of de-
riving Proposition 13.1.1 with these explicit values, one only needs the easier
0
lower bound ϕp,C (Rd ) > (p1/p (p0 )−1/p )d , which is readily deduced by com-
2
puting the Lp norms of φ(x) = φ(x)b = e−π|x| .
As we shortly recall in more detail, most of the estimates of Proposition
13.1.1 have been proved in Section 2.4.b. To complete the picture with the
final estimate in (13.3) (stated in Proposition 2.4.20 with a weaker constant),
we begin with:

Lemma 13.1.2. Let X be a Banach space and p ∈ (1, ∞). Let f ∈ Lp (Td ; X)
be a trigonometric polynomial, which we identify with its periodic extension
to Rd , and let φ ∈ S (Rd ; X). Then

lim kf (·)φ(ε·)εd/p kLp (Rd ;X) = kf kLp (Td ;X) kφkLp (Rd ) ,
ε↓0

lim kF [f (·)φ(ε·)εd/p ]kLp0 (Rd ;X) = kfbk`p0 (Zd ;X) kφk

b p0 d .
L (R )
ε↓0

Proof. For the Lp norm we have

Z
kf (·)φ(ε·)εd/p kpLp (Rd ;X) = kf (t)φ(εt)kpX εd dt
Rd
Z X 
= kf (t)kpX |φ(ε(t + k))|p εd dt,
Td
k∈Zd

where in parentheses we have a Riemann sum of Rd |φ(t)|p dt.

R
0
For the Lp norm, let us write f (t) = k∈Zd xk ek (t). Then
P

X Z
F [f (·)φ(ε·)εd/p ](ξ) = xk φ(εt)εd/p e2πik·t e−2πiξ·t dt
Rd
k∈Zd
228 13 The Fourier transform and multipliers

b −1 (ξ − k))ε−d/p0
X
= xk φ(ε
k∈Zd

Let us split this into two parts,

b −1 (ξ − k))ε−d/p0 ,
X
I := xk 1Q (ξ − k)φ(ε
k∈Zd
X 0
II := b −1 (ξ − k))ε−d/p ,
xk 1{Q (ξ − k)φ(ε
k∈Zd

where Q = [− 21 , 12 )d . The terms in I are disjointly supported, and hence

X 0 0 0
1/p0
kIkLp0 (Rd ;X) = b −1 (· − k))ε−d/p kp 0
kxk kp k1Q (· − k)φ(ε Lp (Rd )
k∈Zd
X 0 0
1/p0
= b p 0
kxk kp k1Q (ε·)φkLp (Rd )
k∈Zd
X 0 0
1/p0
→ b p 0
kxk kp kφk .
Lp (Rd )
k∈Zd

On the other hand,

b −1 (· − k))ε−d/p0 k p0 d
X
kIIkLp0 (Rd ;X) 6 kxk kk1{Q (· − k)φ(ε L (R )
k∈Zd
X
6 kxk kk1{Q (ε·)φk
b p0 d → 0.
L (R )
k∈Zd

Thus kI + IIkLp0 (Rd ;X) indeed converges to the claimed limit. 

Proof of Proposition 13.1.1. The second bound in (13.2) is contained in Prop-

osition 2.4.11. The first bound is also there, but in a slightly different form,
and the present formulation is obtained by repeating the same proof: Given
f ∈ Lp (R; X) and φ ∈ Lp (Rd−1 ), we have

kfbkLp0 (R;X) kφk

b p0 d−1 = kF (f ⊗ φ)k p0 d
L (R ) L (R ;X)

6 ϕp,X (Rd )kf ⊗ φkLp (Rd ;X) = ϕp,X (Rd )kf kLp (R;X) kφkLp (Rd−1 ) .

Choosing f and φ that (almost) achieve equality in the definition of the con-
stants ϕp,X (R) and ϕp,C (Rd−1 ), we obtain the first bound in (13.2).
The first equality in (13.3) is Proposition 2.4.16. The first pair of inequal-
ities and the two equalities in the middle of in (13.3) are all contained in
Proposition 2.4.20 (either as stated or substituting X ∗ in place of X).
Concerning the last pair of inequalities in (13.3), it suffices to prove that

ϕp,X (Rd )
ϕp,X (Td ) 6 , (13.4)
ϕp,C (Rd )
 13.1 Bourgain’s theorem on Fourier type 229

since the other bound follows with X ∗ in place of X and using the first equality
in (13.3). To this end, it follows from Lemma 13.1.2 that

b p0 d = lim kF [f (·)φ(ε·)εd/p ]k p0 d
kfbk`p0 (Zd ;X) kφk L (R ) L (R ;X)
ε↓0

6 lim ϕp,X (Rd )kf (·)φ(ε·)εd/p kLp (Rd ;X)

ε↓0

= ϕp,X (Rd )kf kLp (Td ;X) kφkLp (Rd ) .

Choosing, again, f and φ that (almost) achieve equality in the definition of

the constants ϕp,X (Td ) and ϕp,C (Rd ), we complete the proof of (13.4), and
hence the Proposition. 
Proposition 13.1.1 at hand, in order to prove that a given Banach space has
Fourier type p, we can pick any of the equivalent conditions amenable to our
analysis. We will eventually achieve our goal with the constant ϕp,X (T), but
a major part of the work will take place on the dual group Z. This has the
advantage of presenting a convenient finite formulation as follows:

Definition 13.1.3. Let X be a Banach space, p, q ∈ [1, ∞] and n ∈ Z+ . Then

(q)
ϕp,X (n) is the smallest admissible constant such that the inequality
n n
X 1/p
(q)
X
ek xk 6 ϕp,X (n) kxk kp , ek (t) := e2πikt (t ∈ T),
Lq (T;X)
k=1 k=1

(p0 )
holds for every choice of x1 , . . . , xn ∈ X. We abbreviate ϕp,X (n) := ϕp,X (n).

Although the case q = p0 is most directly linked with the Hausdorff–Young

inequality on the infinite spaces Rd , Td and Zd , it turns out that our inter-
mediate steps towards this final goal will also need to make use of the more
general definition with “mismatched” exponents. Moreover, we will even need
some further variations of this definition (e.g., involving other index sets F in
place of {1, . . . , n}), but we postpone them until the point where they will be
used. For the moment, we have the fairly obvious

Lemma 13.1.4. Let X be a Banach space and p, q ∈ [1, ∞]. The sequence
(q)
(ϕp,X (n))n>1 is increasing, and

(q) 0
1 6 ϕp,X (n) 6 n1/p , ϕp,X (Z) = lim ϕp,X (n) ∈ [1, ∞].
n→∞

Proof. That the sequence is increasing follows simply by extending a shorter

sequence by additional zeroes. This also shows the existence of a (possibly
infinite) limit limn→∞ ϕp,X (n). The lower bound follows by taking x1 6= 0 =
xk for k > 2, and the upper bound is also simply the triangle and Hölder’s
inequality
230 13 The Fourier transform and multipliers
n n n
X 1/p
X X 0
ek xk 6 kxk k 6 n1/p kxk kp .
Lq (T;X)
k=1 k=1 k=1

n
Given (xk )k=1 , let x = (xk )k∈Z be its zero
Pnextension. The upper bound
ϕp,X (n) 6 ϕp,X (Z) follows by observing that k=1 ek (t)xk is simply x b(−t).
It only remains to check that ϕp,X (Z) 6 limn→∞ ϕp,X (n). Let x = (xk )k∈Z
be finitely supported, i.e., xk = 0 if |k| > N for some finite N . Now

X 2N
X −1 2N
X −1
x
b= e−k xk = e−N +j xN −j = e−N ej xN −j ,
|k|6N −1 j=1 j=1

hence
2N
X −1  2N
X −1 1/p
kb
xkLp0 (T;X) = ej xN −j 6 ϕp,X (2N − 1) kxN −j kp
Lp0 (T;X)
j=1 j=1

= ϕp,X (2N − 1)kxk`p (Z;X) 6 lim ϕp,X (n)kxk`p (Z;X) .

n→∞

By the density of finitely supported sequences in `p (Z; X), this shows that
ϕp,X (Z) 6 limn→∞ ϕp,X (n), and completes the proof. 
The task of proving that a space X has non-trivial Fourier type (assuming
non-trivial type) is hence reduced, in principle, to showing the boundedness
of the sequence (ϕp,X (n))n>1 for some p > 1. Although the proof that we
are about to give is eventually set up slightly differently, this idea serves as a
good motivation for a major part of the subsequent analysis. The proof that
we will present can be roughly divided into the following main steps, treated
in the next four sections:
1. Using type bounds on Sidon sets that partition {1, . . . , n} gives a first mild
improvement ϕ2,X (n) = o(n1/2 ) over the trivial estimate ϕ2,X (n) 6 n1/2 .
2. Comparison with the finite Fourier transform on Zn gives sub-multi-
plicativity and leads to ϕ2,X (n) = O(n1/r−1/2 ) for some r > 1.
3. By a delicate Lemma 13.1.25 of Bourgain, this gives a first uniform bound
(2)
ϕs;X (n) = O(1), but with mismatched exponents s ∈ (1, r) and 2 6= s0 .
4. Standard duality and interpolation, combined with repeating the same
key Lemma 13.1.25 on the dual side, allow us to conclude with p ∈ (1, r).
A thorough reader may recognise some conceptual similarity with the consid-
erations encountered in Section 7.3.b in the context of deducing non-trivial
type (and cotype) from the non-containment of certain subspaces. There we
defined the finite type constant τ2,X (n) as the best constant in the estimate
n
X n
X 1/2
εk x k 6 τ2,X (n) kxk k2 ∀x1 , . . . , xn ∈ X. (13.5)
L2 (Ω;X)
k=1 k=1

These numbers will play a role in the first proof step outlined above.
 13.1 Bourgain’s theorem on Fourier type 231

13.1.a Hinrichs’s inequality: breaking the trivial bound

Recall that our goal is deriving non-trivial Fourier P type from non-trivial type.
Thus, from the knowledge that random sums k εk xk can be dominated P by
k(xk )k`p , we would like to conclude that trigonometric sums e x
k k k can
be similarly dominated (though possibly with a different p). An obvious idea
that suggests itself is to try to dominate the trigonometric sum by the random
sum. Indeed, we know from Section 6.5 that this can be done under particular
circumstances if the trigonometric sum is restricted to a special set called a
Sidon set. This leads to the following strategy: Given the initial sum over k ∈
{1, . . . , N }, we want to partition this into sums over Sidon sets on which we can
make estimates, and this partitioning should be done sufficiently economically
so that it allows us to beat the trivial estimate. To carry out this idea, we
need to be able to
1. efficiently recognise Sidon sets, and
2. decompose arbitrary sets into as few as possible Sidon sets.
We now turn to these tasks. Recall from Section 6.5 that a subset Λ ⊆ Z
is called a Sidon set if the following estimate holds uniformly over all finitely
non-zero sequences (cλ )λ∈Λ of complex numbers:
X X
|cλ | 6 C c λ eλ .
∞
λ∈Λ λ∈Λ

The smallest admissible constant C is called the Sidon constant of Λ and is

denoted by S(Λ). However, this definition in itself is hardly helpful in checking
whether or not a particular set actually satisfies this property. A first sufficient
condition for a set to be a Sidon set was achieved in Proposition 6.5.3, showing
in particular that S({2k : k ∈ N}) 6 4. For the present purposes, we require
a more robust criterion, which is provided in the following:

Definition 13.1.5 (Quasi-independent set). A subset F ⊆ Z \ {0} is

called quasi-independent if αk ≡ 0 is the only finitely non-zero sequence such
that αk ∈ {−1, 0, +1} for all k ∈ F and
X
αk · k = 0.
k∈F

Example
P∞ 13.1.6. The sequence {2k : k ∈ N} is quasi-independent. In fact, if
α
k=0 k 2 k
= 0 for a finitely non-zero sequence (αk )∞
k=1 , then
X X
2k = 2k .
k:αk =+1 k:αk =−1

It follows from the uniqueness of the binary expansion that {k : αk = +1} =

{k : αk = −1}, and this is possible only if both sets are empty. Hence αk ≡ 0.
232 13 The Fourier transform and multipliers

Proposition 13.1.7 (Bourgain). Every quasi-independent set F ⊆ Z \ {0}

is a Sidon set with
S(F ) 6 16.

By Example 13.1.6, this gives another proof of the fact that {2k : k ∈ N} is a
Sidon set, but with a slightly weaker constant than Proposition 6.5.3.
Proof. This is based on a variant of the Riesz product method also used in
the proof of Proposition 6.5.3, but the details are somewhat different, and we
will provide a self-contained argument. By considering every finite subset of
the original F , we may assume without loss of generality that F is finite to
begin with. Given parameters % ∈ (0, 1] and ξ = (ξk )k∈F ∈ RF , let then
Y

Rξ (t) := 1 + % cos(2π(kt + ξk ))
k∈F
Y %

= 1 + (ek (t)e1 (ξk ) + e−k (t)e−1 (ξk ))
2
k∈F
X  X   X 
= 2−|α| %|α| exp 2πi αk · kt exp 2πi α k ξk ,
α∈{−1,0,+1}F k∈F k∈F
P
where |α| := k∈F |αk | as usual for multi-indices,. (To relax the notation, we
do not explicate the dependence of Rξ on %.)
From the assumption that F is quasi-independent, it follows that
X
αk · k = 0 only if αk ≡ 0,
k∈F

and hence R bξ (0) = 1. It is also clear from the first line of the definition of
Rξ (t) (recalling that % ∈ (0, 1]) that Rξ (t) > 0, and hence
Z 1
kRξ kL1 (T) = Rξ (t) dt = R
bξ (0) = 1.
0

Let us further write

   
(m)
X X X
Rξ (t) := 2−|α| exp 2πi αk · kt exp 2πi α k ξk ,
α∈{−1,0,+1}F k∈F k∈F
|α|=m

so that
#F
(m)
X
Rξ (t) = %m R ξ (t), where
m=0

(0) (1) 1X

Rξ (t) = 1, Rξ (t) = ek (t)e1 (ξk ) + e−k (t)e−1 (ξk ) .
2
k∈F

From the orthogonality of the exponentials, for each j ∈ F , we have

 13.1 Bourgain’s theorem on Fourier type 233
Z 1
(1) 1X
1
Rξ (t)ej (t) dt = δk,−j e1 (ξk ) + δk,j e−1 (ξk ) = e−1 (ξj ),
0 2 2
k∈F

where we observed that k = −j is not possible when k, j belong to the same

quasi-independent set F , since 1·k+1·j = 0 is a direct violation of the defining
R (0)
condition. It is also immediate that Rξ ej = 0 for all j ∈ F ⊆ Z \ {0}.
For X
f= cj ej ,
j∈F

we then conclude that

Z 1 Z 1 #F
X  X 
(m)
Rξ f = %m R ξ cj ej
0 0 m=0 j∈F
Z 1 (13.6)
%X X X (m)
=0+ cj e−1 (ξj ) + cj %m Rξ e j .
2 0
j∈F j∈F m>2

Using again the orthogonality of the exponentials, we have

Z 1  
(m)
X X
Rξ e j = 2−m exp 2πi α k ξk
0
α∈{−1,0,+1}F k∈F
P |α|=m
k∈F αk ·k=−j
Z 1
(m)
X
6 2−m = R0 e j ,
0
α∈{−1,0,+1}F
P |α|=m
k∈F αk ·k=−j

(m) (m)
where R0 is simply Rξ with ξ = 0. It follows that

#F Z 1 #F Z 1 Z 1
(m) (m)
X X
Rξ e j 6 R0 e j = R0 ej 6 kR0 kL1 (T) = 1.
m=0 0 m=0 0 0

The last term in (13.6) can now be estimated by

Z 1 Z 1
(m) (m)
X X X X X
cj %m Rξ e j 6 |cj | %2 Rξ e j 6 |cj |%2 .
j∈F m>2 0 j∈F m>2 0 j∈F

If we now choose ξj so that cj e−1 (ξj ) = |cj |, then (13.6) gives

Z 1 Z 1
%X X X (m)
|cj | = Rξ f − cj %m Rξ e j
2 0 0
j∈F j∈F m>2
X
2
6 kRξ kL1 (T) kf kL∞ (T;X) + % |cj |,
j∈F
234 13 The Fourier transform and multipliers

and hence
% X X
− %2 |cj | 6 kf kL∞ (T;X) = ck ek .
2 L∞ (T;X)
j∈F k∈F

1
Choosing finally % = 4 completes the proof. 
By the previous result, our initial task of decomposing arbitrary sets into
Sidon sets is reduced to decomposing into quasi-independent sets. A first step
in this direction is to know that every set has a quasi-independent subset of
somewhat substantial size.
Lemma 13.1.8. Any finite subset F ⊆ Z\{0} has a quasi-independent subset
F0 ⊆ F of cardinality #F0 > dlog3 #F e.
Proof. Let F0 ⊆ F be a quasi-independent subset of maximal cardinality, and
let nX o
F1 := αk · k : αk ∈ {−1, 0, +1} .
k∈F0

Clearly F1 ⊇ F0 , and we claim that in fact F1 ⊇ F . If not, let k0 ∈ F \ F1 . We

will check that F0 ∪ {k0 } is quasi-independent, contradicting the maximality
of F0 . Namely, suppose that
X
αk · k = 0,
k∈F0 ∪{k0 }

where αk ∈ {−1, 0, +1}. If αk0 = ±1, then

X
k0 = (−αk0 αk ) · k ∈ F1 ,
k∈F0

contradicting k0 ∈
/ F1 . Thus αk0 = 0, but then also αk = 0 for all k ∈ F0 , since
F0 is quasi-independent, and this proves that F0 ∪ {k0 } is quasi-independent.
As explained above, this proves that F1 ⊇ F , and hence
#F 6 #F1 6 3#F0 ,
from which the proposition follows, since #F0 > log3 #F is necessarily an
integer. 
By recursively removing big quasi-independent subsets, we arrive at the de-
sired decomposition of the initial set:
Lemma 13.1.9. For N ∈ Z+ , let
d(N ) := min{k ∈ Z+ : any subset F ⊆ Z \ {0} of size #F 6 N can be
divided into at most k quasi-independent subsets}.
n
2·3
Then d(3n ) 6 for all n ∈ N. For all n > 1, each of the partitioning
n+1
quasi-independent subsets can be chosen to have size at most n.
 13.1 Bourgain’s theorem on Fourier type 235

Proof. Since clearly d(3n ) 6 3n (as each singleton is quasi-independent), the

claim is obvious for n 6 1. For 3 < #F 6 9, Lemma 13.1.8 guarantees a
quasi-independent subset of size 2. Starting from a set of size 9 and repeatedly
extracting 3 quasi-independent subsets of size 2, we are left with a subset of
size 3 that trivially splits into 3 quasi-independent subsets of size 1. Hence
d(32 ) 6 3 + 3 = 6 = 2 · 32 /(2 + 1). We then assume that, for some n > 2,
any set of size 3n can be divided into at most 2 · 3n /(n + 1) quasi-independent
subsets of size at most n, and we prove the same for n + 1.
If #F = 3n+1 , Lemma 13.1.8 guarantees that we can repeatedly extract
quasi-independent subsets Fi (i = 1, . . . , j) of size n + 1, until

3n+1 − j(n + 1) 6 3n < 3n+1 − (j − 1)(n + 1),

thus
2 · 3n 2 · 3n
6j <1+ .
n+1 n+1
The remaining set of size at most 3n can then be divided into at most d(3n )
quasi-independent subsets, and by the induction assumption we have
 2 · 3n  2 · 3n 4 · 3n
d(3n+1 ) 6 j + d(3n ) < 1 + + =1+ .
n+1 n+1 n+1
For n > 2, we have 1/(n + 1) 6 34 /(n + 2) and 3n /(n + 2) > 49 , and hence
4 16  3n 3n 2 · 3n+1
d(3n+1 ) 6 + <6 = ,
9 3 n+2 n+2 (n + 1) + 1
and this completes the induction step. Note that all quasi-independent subsets
that we constructed in the induction step had either size n + 1, or they came
from the induction assumption, in which case their size is at most n. 
In the next remark, we indicate converses to the obtained bounds of Lemmas
13.1.8 and 13.1.9.
Remark 13.1.10. Let F0 ⊆ F := {1, . . . , N } be such that N > 2 and F0 is
quasi-independent. We claim that necessarily #F0 6 2 log2 (N ). Clearly, this
implies d(N ) > 2 logN (N ) . Indeed, write m = #F0 . It suffices to consider m > 2.
2
Let A ⊆ F0 be arbitrary. Then
X X
06 a6 a < N 2.
a∈A a∈F0

Therefore, the number of different values can be estimated by

nX o
# a : A ⊆ F0 6 N 2 .
a∈A
P P
One the other hand, if A, B ⊆ F0 are such that a∈A a = b∈B b, then the
quasi-independence of F0 implies A = B. Therefore,
236 13 The Fourier transform and multipliers
nX o
2m = #{A ⊆ F0 } 6 # a : A ⊆ F0 .
a∈A

We can conclude 2m 6 N 2 and thus the claim follows.

We now possess all the ingredients needed for the first estimate of Fourier type
in terms of type, stated in terms of the finite versions of both properties. The
reader may wish to compare the next proposition to Theorem 7.6.12 which
gives a related inequality for the Walsh system.

Proposition 13.1.11 (Hinrichs’s inequality). For all n > 1 we have

ϕ2,X (3n ) √ τ2,X (n)

√ 6 16 2 · √ .
3n n

Proof. By Lemma 13.1.9, the set {1, . . . , 3n } can be divided into

A 6 2 · 3n /(n + 1)

quasi-independent subsets Fa of size #Fa 6 n. By Proposition 13.1.7, each

quasi-independent Fa is a Sidon set with S(Fa ) 6 16. By Pisier’s Theorem
6.5.5, trigonometric series over a Sidon set is comparable in the Lp norm to the
corresponding Rademacher series, up to the Sidon constant. Chaining these
observations and using the definition of the type constants τ2,X (n) and the
Cauchy–Schwarz inequality, we obtain
n
3
X A X
X A
X X
ek xk = ek xk 6 ek xk
L2 (T;X) L2 (T;X) L2 (T;X)
k=1 a=1 k∈Fa a=1 k∈Fa
A
X X
6 16 εk xk
L2 (Ω;X)
a=1 k∈Fa
A
X  X 1/2
6 16 · τ2,X (#Fa ) kxk k2
a=1 k∈Fa
A X
√ X 1/2
6 16 · max τ2,X (#Fa ) A kxk k2
16a6A
a=1 k∈Fa
r 3n
nX
2·3 1/2
6 16 · τ2,X (n) kxk k2 ,
n+1
k=1

from which the proposition follows. 

The following
√ corollary gives the promised improvement over the trivial bound
ϕ2,X (3n ) 6 3n as soon as n is large enough.
 13.1 Bourgain’s theorem on Fourier type 237

Corollary 13.1.12. Let X be a Banach space of type p ∈ (1, 2]. Then for all
n > 1, we have
ϕ2,X (3n ) √ 0
√ 6 16 2 · τp,X;2 · n−1/p .
3 n

The type constant τp,X;s with a secondary parameter (above s = 2) was

introduced right before Proposition 7.1.4 as the best constant in the inequality
K
X K
X 1/p
εk x k 6 τp,X;s kxk kp , (13.7)
Ls (Ω;X)
k=1 k=1

where x1 , . . . , xK ∈ X and K ∈ Z+ are arbitrary. Recall that τp,X := τp,X;p .

Proof. From the definition of the type constants and Hölder’s inequality, it is
immediate that
τ2,X (n) τp,X;2 · n1/p−1/2 0
√ 6 √ = τp,X;2 · n−1/p .
n n
In combination with Proposition 13.1.11, this gives the result. 

13.1.b The finite Fourier transform and sub-multiplicativity

Note that the improvement of Corollary 13.1.12 over the trivial bound is only
very slight. Our first goal in bootstrapping this initial estimate is to obtain a
power-type bound of the form ϕ2,X (N ) = O(N 1/2−δ ). As the reader can easily
verify (perhaps referring to Lemma 7.3.19), this would readily follow from the
established bound, if in addition we had a sub-multiplicative estimate
?
ϕ2,X (nm) 6 ϕ2,X (n)ϕ2,X (m).
As we do not know whether this is true, we take a detour by comparing
the sequence ϕ2,X (n) with the following discretised variant:
(q)
Definition 13.1.13. Let X be a Banach space and n ∈ Z+ . Then ϕp,X (Zn )
is the best constant in the following inequality with arbitrary x1 , . . . , xn ∈ X:
n
1 X n q 1/q n
X 1/p
(q)
X
ek (h/n)xk 6 ϕp,X (Zn ) kxk kp .
n
h=1 k=1 k=1

(q)
As the notation suggests, ϕp,X (Zn ) has an interpretation as the norm of the
Fourier transform (thus, a Fourier type constant) of functions on the finite
group Zn = Z/nZ, but there is no need to insist too much on this point here.
(q) (q)
The difference of the defining inequalities of ϕp,X (n) and ϕp,X (Zn ) is that
p
the L (T; X) integral norm in the former is replaced by a finite Riemann sum
approximation in the latter. We will next develop some tools for comparing
the two kinds of norms. This will involve elements of some fairly classical
Fourier analysis, and we begin with
238 13 The Fourier transform and multipliers

Definition 13.1.14. The Dirichlet kernel is defined by

X
Dn (t) := ek (t), t ∈ T,
|k|6n

the Fejér kernel by

n
1 X X  |k| 
Fn (t) := Dk (t) = 1− ek (t), t ∈ T,
n+1 n+1
k=0 |k|6n

and the de la Vallée–Poussin kernel by

2n−1
1 X X X  |j| 
Vn (t) := Dk (t) = ej (t) + 2− ej (t), t ∈ T.
n n
k=n |j|6n n<|j|<2n

Lemma 13.1.15. These kernels satisfy the identities

sin(π(2n + 1)t) 1 sin2 (π(n + 1)t)
Dn (t) = , Fn (t) = > 0,
sin(πt) n+1 sin2 (πt)
Vn (t) = 2F2n−1 (t) − Fn−1 (t).
Proof. The formula for Dn is the summation of a geometric series:
X e2πi(2n+1)t − 1 sin(π(2n + 1)t)
Dn (t) := e2πikt = e−2πint 2πit
= .
e −1 sin(πt)
|k|6n

Since
n n
X X  ei2π(n+1)t − 1 
sin(π(2k + 1)t) = = eiπt ei2πkt = = eiπt
ei2πt − 1
k=0 k=0
 sin(π(n + 1)t)  sin2 (π(n + 1)t)
= = eiπ(n+1)t = ,
sin(πt) sin(πt)
we obtain the formula for Fn by summing over the formula for Dk . Finally,
2n−1 2n−1 n−1
1 X 1 X X  1 
Vn = Dk = Dk − Dk = 2nF2n−1 − nFn−1 .
n n n
k=n k=0 k=0


Lemma 13.1.16. If f is a trigonometric polynomial with deg(f ) < n, then
for all s ∈ R we have
Z 1 n
1X
f (t) dt = f (s + h/n),
0 n
h=1

i.e., f can be integrated exactly by uniform Riemann sums of order n.

 13.1 Bourgain’s theorem on Fourier type 239

Proof. It is enough to consider f (t) = ek (t), where |k| < n. We observe that
n
( 2πikn/n
−1
X
2πikh/n e2πik/n ee2πik/n −1 = 0, 0 < |k| < n,
e =
h=1
n, k = 0,

and hence
n n Z 1
1X ek (s) X 2πikh/n
f (s + h/n) = e = ek (s)δk,0 = δk,0 = ek (t) dt.
n n 0
h=1 h=1


On the level of Lp norms, this leads to the following comparison result:

Proposition 13.1.17 (Marcinkiewicz inequality). Let X be a Banach

space and p ∈ [1, ∞). Then for all n ∈ Z+ and x1 , . . . , xn ∈ X, we have
n
1 X n
X p 1/p n
X
ek (h/n)xk 63 ek xk .
n Lp (T;X)
h=1 k=1 k=1

With the usual modification, the result is also true (and entirely trivial) for
p = ∞: of course the supremum over {j/n : j = 1, . . . , n} is dominated by the
supremum over all of T!
Proof. Let
n
X
f (t) := ek (t)xk , m := bn/2c.
k=1

Then (n − 1)/2 6 m 6 n/2 and the function e−(m+1) f is a linear combination

of ek with

−m = 1 − (m + 1) 6 k 6 n − (m + 1) 6 (2m + 1) − (m + 1) = m,

so e−(m+1) f is a trigonometric polynomial of degree m.

Since the de la Vallée–Poussin kernel Vm from Definition 13.1.14 has
Fourier coefficients Vbm (k) = 1 for all values |k| 6 m on which the Fourier
coefficients of e−(m+1) f are supported, we conclude that

Vbm [e−(m+1) f ] b = [e−(m+1) f ] b ,

hence Vm ∗ (e−(m+1) f ) = e−(m+1) f . Thus

kf (t)k = ke−(m+1) (t)f (t)k = kVm ∗ (e−(m+1) f )(t)k

Z
6 |Vm (t − s)|kf (s)k ds
T (13.8)
 Z 1/p0  Z 1/p
6 |Vm (t − s)| ds |Vm (t − s)|kf (s)kp ds
T T
240 13 The Fourier transform and multipliers

By Lemma 13.1.15, we have

|Vm | = |2F2m−1 − Fm−1 | 6 2F2m−1 + Fm−1 , (13.9)

R
where T
Fk (t) dt = Fbk (0) = 1, and hence
Z
|Vm (t − s)| ds 6 3.
T

Substituting into (13.8) and summing, we have

n n
Z X
X 0
kf (h/n)kp 6 3p/p |Vm (h/n − s)|kf (s)kp ds.
h=1 T h=1

Since the right-hand side of (13.9) is a trigonometric polynomial of degree

2m − 1 6 n − 1, Lemma 13.1.16 guarantees that
n
X n
X
|Vm (h/n − s)| 6 (2F2m−1 + Fm−1 )(h/n − s)
h=1 h=1
Z
=n (2F2m−1 + Fm−1 )(u) du = 3n.
T

Substituting back, we conclude that

n Z
1X 0
kf (h/n)kp 6 3p/p 3kf (s)kp ds = 3p kf kpLp (T;X) .
n T
h=1


We now have the desired comparison of the two finite Fourier type constants:
Lemma 13.1.18. For any Banach space X and n ∈ Z+ , we have
(q) (q) (q)
ϕp,X (n) 6 ϕp,X (Zn ) 6 3 · ϕp,X (n).

Proof. Substituting ek (t)xk in place of xk in Definition 13.1.13, we find that

n n q n q/p
1X X (q)
q  X
ek (t + h/n)xk 6 ϕp,X (Zn ) kxk kp
n
h=1 k=1 k=1

Integrating over t ∈ T and using the translation invariance

kf (· + h/n)kLq (T;X) = kf kLq (T;X) ,

we obtain
n Z n q n q/p
1X X (q)
q  X
ek (t)xk dt 6 ϕp,X (Zn ) kxk kp ,
n T
h=1 k=1 k=1
 13.1 Bourgain’s theorem on Fourier type 241
(q) (q)
and hence ϕp,X (n) 6 ϕp,X (Zn ).
The other estimate follows at once from the Marcinkiewicz inequality
(Proposition 13.1.17), which is the first step in
n
1 X n
X q 1/q n
X
ek (h/n)xk 63 ek xk
n Lq (T;X)
h=1 k=1 k=1
n
X 1/p
(q)
6 3ϕp,X (n) kxk kp .
k=1


(q)
The following lemma is our reason for considering the quantities ϕp,X (Zn ):

Lemma 13.1.19. For any Banach space X and m, n ∈ Z+ , we have the sub-
multiplicative estimate
(q) (q) (q)
ϕp,X (Zmn ) 6 ϕp,X (Zm )ϕp,X (Zn ), 1 6 p 6 q 6 ∞;

in particular

ϕp,X (Zmn ) 6 ϕp,X (Zm )ϕp,X (Zn ) ∀p ∈ [1, 2].

Proof. The second estimate is an obvious special case with q = p0 > 2 > p.
For the proof of the general estimate, it is convenient to observe that, by
(q)
simple reindexing and modular arithmetic, the condition defining ϕp,X (Zn )
is unchanged if instead of {1, . . . , n} we take all sums over {0, . . . , n − 1}. In
(q)
the defining condition of the constant ϕp,X (Zmn ), we should then sum over
{0, . . . , mn − 1}, and the key trick of the proof is to use a non-symmetric
reindexing of this range for the h and k sums, namely

h = an + b : a = 0, . . . , m − 1, b = 0, . . . , n − 1,
k = um + v : u = 0, . . . , n − 1, v = 0, . . . , m − 1.

Then
hk = (an + b)(um + v) = aumn + avn + bum + bv,
and hence, noting that e2πiau = 1,

ek (h/mn) = eu (b/n)ev (a/m)ev (b/mn).

Thus
n 1 mn−1
X mn−1
X q o1/q
ek (h/mn)xk
mn
h=0 k=0
n 1 n−1
X 1 m−1
X m−1
X q o1/q
= ev (a/m)yv(b) ,
n m a=0 v=0
b=0
242 13 The Fourier transform and multipliers
n−1
X
yv(b) := ev (b/mn) eu (b/n)xum+v ,
u=0
n 1 n−1  m−1 q/p o1/q
(q)
X X
6 ϕp,X (Zm )q kyv(b) kp
n v=0
b=0
X  1 n−1
n m−1 X n−1 q p/q o1/p
(q)
X
6 ϕp,X (Zm ) eu (b/n)xum+v
v=0
n u=0
b=0

by Minkowski’s inequality for p 6 q,

n m−1 n−1 o1/p

(q) (q)
X X
6 ϕp,X (Zm ) ϕp,X (Zn )p kxum+v kp
v=0 u=0
n mn−1 o1/p
(q) (q)
X
= ϕp,X (Zm )ϕp,X (Zn ) kxk kp ,
k=0

(q)
where we used the defining condition for ϕp,X (Zm ) with the sequences
(b) (q)
(yv )m−1
v=0 for each fixed b = 0, . . . , n − 1, and that for ϕp,X (Zn ) with the
sequences (xmu+v )n−1
u=0 for each fixed v = 0, . . . , m − 1. 
Combining the above results with Corollary 13.1.12 of Hinrichs’s inequal-
ity, we achieve the desired power-type improvement over the trivial estimate
ϕ2,X (N ) 6 N 1/2 . One could try to deduce this from Lemma 7.3.19 applied
to ϕ2,X (Zn ). However, this time that does not work since we do not know
whether ϕ2,X (Zn ) is increasing in n. Therefore, we adapt the proof of the
lemma and use the facts that ϕ2,X (n) is increasing and that ϕ2,X (Zn ) is sub-
multiplicative. Our choice of notation r0 below is indicative of the fact that
this is the Hölder conjugate of a (small) exponent r > 1.

Corollary 13.1.20. Let X be a Banach space of type p ∈ (1, 2]. Then

0
ϕ2,X (N ) 6 C · N 1/2−1/r = C · N 1/r−1/2 ,

where
0 r0
r0 := 3p0 (68 · τp,X;2 )p , C := e 2p0 . (13.10)

Proof. Given N, n ∈ Z+ , let k ∈ Z+ satisfy

3n(k−1) 6 N < 3nk . (13.11)

Then

ϕ2,X (N ) 6 ϕ2,X (3nk ) since ϕ2,X is increasing by Lemma 13.1.4,

6 ϕ2,X (Z3nk ) by the comparison in Lemma 13.1.18,
 13.1 Bourgain’s theorem on Fourier type 243

6 ϕ2,X (Z3n )k by sub-multiplicativity (Lemma 13.1.19),

n k
6 (3 · ϕ2,X (3 )) by the comparison in Lemma 13.1.18.

Therefore, by (13.11) for any s ∈ (1, 2] we find

1 1
N 1/2−1/s ϕ2,X (N ) 6 3n(k−1)( 2 − s ) (3 · ϕ2,X (3n ))k
1 1 1 1
= 3n( s − 2 ) [3n( 2 − s ) · 3 · ϕ2,X (3n )]k

For appropriate n and s, we will show that the term within brackets satisfies
[. . .] 6 1. By Corollary 13.1.12, we can estimate
1 1 √ 0 0 0
[. . .] 6 3n( 2 − s ) · 3 · 16 2 · τp,X;2 · 3n/2 · n−1/p =: 3n/s T n−1/p ,
√ 0
where T := 48 2τp,X;2 . Therefore, setting s0 = (1 + eT p )p0 log(3) and taking
0 0
eT p 6 n < eT p + 1 we find that
0
0 0 0 e−1/p
3n/s T n−1/p 6 e1/p T = 1.
T
From the above we conclude that
1 1 0 0
N 1/2−1/s ϕ2,X (N ) 6 3n( s − 2 ) 6 3n/2 = en log(3)/2 6 es /(2p ) .

The above trivially holds true if we replace s0 by any r0 > s0 . Since 50 6 T 6

68τp,X;2 and p0 > 2, one can check that
0 0 0 0
s0 = (1 + eT p )p0 log(3) = T p (T −p + e)p0 log(3) 6 (68τp,X;2 )p 3p0 =: r0 .

Thus the statement follows. 

Before turning to some of the sophisticated constructions and estimates for
Bourgain’s theorem, we discuss a much simpler situation where one can obtain
ϕ2,X (n) = O(n1/r−1/2 ) with r ∈ (1, 2). It does not play a role in the proof of
Bourgain’s theorem.

Proposition 13.1.21. If X has type p and cotype q, then for all n > 1,

ϕ2,X (n) 6 τ2,X (n)c2,X (n) 6 τp,X;2 cq,X;2 n1/p−1/q .

1 1
Of course the latter bound is nontrivial only if p − q < 12 .

Proof. Let (γh )h>1 be a complex Gaussian sequence (i.e., standard indepen-
dent Gaussian random variables). Also let
n
1 X h
ek = √
γ γh ek ( ), k = 1, . . . , n.
n n
h=1
244 13 The Fourier transform and multipliers

n
Then (eγk )k=1 are also independent standard Gaussian random variables (see
Section E.2). Hence, using the natural Gaussian analogue of the finite type
and cotype constants,
n n 2 n n 2
1X X h 1 γ X X h
xk ek ( ) 6 c2,X (n)2 E γh xk ek ( )
n n n n
h=1 k=1 h=1 k=1
n
X 2
= cγ2,X (n)2 E γ
ek xk
k=1
n
X
6 cγ2,X (n)2 τ2,X
γ
(n)2 kxk k2 .
k=1
γ
Since kγk2 = 1, Proposition 7.1.18 informs us that τ2,X 6 τ2,X and cγ2,X 6
γ
c2,X , and the analogous result for the finite constants τ2,X (n) etc. follows by
1 1
the same argument. Finally, Hölder’s inequality implies τ2,X (n) 6 τp,X;2 n p − 2
1 1
and c2,X (n) 6 cq,X;2 n 2 − q . 

13.1.c Key lemmas for an initial uniform bound

The core of this section consists of two delicate lemmas of Bourgain that
allow us to bootstrap the power-type improvement over the trivial bound on
the growth of ϕ2,X (N ), as given in Corollary 13.1.20, into a uniform estimate
(2)
for the constants ϕs,X (N ) with some s > 1. To streamline the presentation of
the core arguments, we begin with the following classical identity:
fb(j)ej with (fb(j))j∈Z ∈ `1 (Z). Then
P
Lemma 13.1.22. Let f = j∈Z

N
X 1 X
fb(j) = e−1 (nh/N )f (h/N ).
N
j≡n mod N h=1

Proof. We first observe that

N N
1 X X 1 X X
f (h/N ) = fb(j) ej (h/N ) = fb(j),
N N
h=1 j∈Z h=1 j≡0 mod N

which is case n = 0 of the claim.

We apply this with f replaced by
X X
e−n f = fb(j)ej−n = fb(j + n)ej
j∈Z j∈Z

to find that
N
1 X X X
(e−n f )(h/N ) = fb(j + n) = fb(j),
N
h=1 j≡0 mod N j≡n mod N

which is the general case. 

 13.1 Bourgain’s theorem on Fourier type 245

Lemma 13.1.23 (Bourgain). Let F ⊆ Z be a finite subset with #F = N .

Then there exists t0 ∈ T such that at least 81 N of the pairwise disjoint intervals

1 1 1
In := [n − , n + ), n = 1, . . . , N,
N 2 2
satisfy t0 k ∈ In + Z for some k ∈ F .

Proof. We in fact show that this is true for the “average” choice of t0 ∈ T.
For t ∈ T and n = 1, . . . , N , we denote

νn (t) := #{k ∈ F : tk ∈ In + Z},

N (t) := #{n = 1, . . . , N : νn (t) > 0}.

The claim is then that N (t0 ) > 81 N for some t0 ∈ T, and we will prove that
Z 1
1
N (t) dt > N, (13.12)
0 8

which clearly implies the existence of a desired t0 .

The strategy of the proof is as follows. Since each of the N different k ∈ F
satisfies tk ∈ In + Z for exactly one n = 1, . . . , N , we have
N
X X N
X 1/2
N= νn (t) = νn (t) 6 N (t)1/2 νn (t)2 .
n=1 16n6N n=1
νn (t)>0

Integrating and using the Cauchy–Schwarz inequality, we obtain

Z 1 1/2  Z N
1X 1/2
N6 N (t) dt νn (t)2 dt ,
0 0 n=1

and (13.12) follows if we can prove that

Z N
1X
νn (t)2 dt 6 8N. (13.13)
0 n=1

Now X X
νn (t) = 1In +Z (kt) = 1I0 +Z (kt − n/N ).
k∈F k∈F

For the convenience of Fourier analysis, we replace the indicator

1 1
1I0 +Z (t) = 1[− 2N
1 1
, 2N ) (t), t ∈ [− , ),
2 2
by the regularised version given by the 1-periodic extension of the “tent”
246 13 The Fourier transform and multipliers

1 1
s(t) := (1 − N |t|)+ , t ∈ [− , ).
2 2
An elementary computation of the Fourier coefficients shows that
(
1 2 1/N, j = 0,
sb(j) = sinc (πj/N ) = (13.14)
N 0, 0 6= j ≡ 0 mod N.

Note that the first equality above is valid for all j ∈ Z, although in the second
we only consider particular cases. Clearly 0 6 sb(j) = O(j −2 ), so that Lemma
13.1.22 applies to f = s. Since s(h/N ) = 1N Z (h) for h ∈ Z, the conclusion of
the lemma takes a particularly clean form, namely
X 1
sb(j) = ∀n = 1, . . . , N. (13.15)
N
j≡n mod N

We observe that 1I0 +Z (t) 6 2s(t), and hence

X XX
νn (t) 6 2 s(kt − n/N ) = 2 sb(j)ej (kt − n/N )
k∈F k∈F j∈Z
N
X X X
=2 eh (−n/N ) sb(j) ej (kt).
h=1 j≡h k∈F
mod N

Substituting this into (13.13), we can now estimate

Z N
1X Z N
1X N
X X X 2
νn2 6 4 eh (−n/N ) sb(j) ej (kt) dt
0 n=1 0 n=1 h=1 j≡h k∈F
mod N
Z 1 N
X X X 2
=4 N sb(j) ej (kt) dt,
0 h=1 j≡h k∈F
mod N

since the matrix (N −1/2 eh (−n/N ))N

h,n=1 is unitary,
N  X
X X 2
6 4N sb(j) ejk
L2 (T)
h=1 j≡h k∈F
mod N
−1  X
n NX 2  X 2 o
6 4N sb(j)N 1/2 + sb(0)N + sb(j)N 1/2 ,
h=1 j≡h 06=j≡0
mod N mod N
(
X N, j = 0,
since ejk = 1/2
k∈F
L2 (T) N , otherwise,
−1 
n NX 1 1/2 2  1 2 o
= 4N N + N +0 ,
N N
h=1
 13.1 Bourgain’s theorem on Fourier type 247

by (13.15) and (13.14),

= 4N {(N − 1)N −1 + 1} = 4{2N − 1} < 8N = RHS (13.13).

This confirms (13.13) and hence, as explained in the beginning of the proof,
the assertion of the Lemma. 
From the comparison between `pN (X) and `∞
N (X), it is immediate that

(q) (q)
ϕ∞,X (N ) 6 ϕp,X (N ) · N 1/p .

This triviality admits a crucial improvement, where on the left we have a

similar quantity associated to an arbitrary subset F ⊆ Z of size N .
(q)
Definition 13.1.24. Given F ⊆ Z, we denote by ϕ∞,X (F ) be the best con-
stant in the estimate
(q)
X
ek xk q 6 ϕ∞,X (F ) sup kxk k,
L (T;X) k∈F
k∈F

which is to holds for arbitrary families (xk )k∈F in X.

(q) (q)
Clearly the previously considered ϕ∞,X (N ) is the special case ϕ∞,X (N ) =
(q)
ϕ∞,X ({1, . . . , N }) in this notation. In contrast to random sums with indepen-
dent sequences of random variables, the particular choice of the indexing set
F is very relevant here, since the joint distribution of (ek )k∈F can be very
N
different from that of (ek )k=1 .

Lemma 13.1.25 (Bourgain). For any Banach space X and exponents p, q ∈

[1, ∞) we have
(q) (q)
1+1/q
ϕ∞,X (F ) 6 Aϕp,X (N ) · N 1/p , A := 8p(π + 21/q · 3) ,

whenever F ⊆ Z is a subset of size #F = N .

Remark 13.1.26. We only apply Lemma 13.1.25 with p = 2 6 q. In this case

√
3/2
A 6 16(π + 2 · 3) < 1285.

Proof of Lemma 13.1.25. Since

X X
e k xk q 6 kxk k 6 N max kxk k,
L (T;X) k∈F
k∈F k∈F

(q)
and ϕp,X (N ) > 1, we have the trivial estimate

(q) (q) 0 (q) 0

ϕ∞,X (F ) 6 N 6 ϕp,X (N )N 1/p +1/p 6 Aϕp,X (N )N 1/p ∀N 6 Ap .
248 13 The Fourier transform and multipliers
0
Suppose then, for induction, that N > Ap , and moreover that the Lemma
has been verified for all N 0 < N in place of N . For F ⊆ Z+ of size N , we
consider a splitting (with ∅ 6= F0 ( F to be specified shortly)
X X X
ek xk q 6 ek xk q + ek xk q =: I + II.
L (T;X) L (T;X) L (T;X)
k∈F k∈F0 k∈F \F0

(q)
Since F \ F0 ( F is a strictly smaller set and ϕp,X is clearly non-decreasing,
the induction hypothesis applies to show that
(q)
II 6 ϕ∞,X (F \ F0 ) max kxk k
k∈F \F0
(13.16)
(q)
6 Aϕp,X (N )#(F \ F0 )1/p max kxk k.
k∈F

Let us make a specific choice of F0 ( F as follows. By Lemma 13.1.23, there

exist t0 ∈ T and 1 6 n1 < n2 < . . . n` 6 N with ` > 81 #F such that each of
the mutually disjoint sets
1 1 1
I nj + Z = [n − , n + ) + Z, (j = 1, . . . , `),
N 2 2
intersects with the set {kt0 : k ∈ F }. For each j ∈ {1, . . . , `}, we pick a
kj ∈ F such that kj t0 ∈ Inj + Z, and set F0 := {kj : j = 1, . . . , `}. Then
#F0 = ` > 81 #F . The size bound on #F0 shows that (13.16) implies
7 1/p
(q)
II 6 Aϕp,X (N ) N max kxk k. (13.17)
8 k∈F

Let ψ : kj → nj be the corresponding bijection from F0 onto ψ(F0 ) ⊆

{1, . . . , N }. Thus by definition that kj t0 ∈ Iψ(kj ) + Z = Inj + Z for all j =
1, . . . , `. For any h ∈ Z, we then have
X
I= ek (· + ht0 )xk by translation invariance
Lq (T;X)
k∈F0
X ψ(k) X ψ(k)
6 [ek (ht0 ) − eh ( )]ek xk + eh ( )ek xk
N Lq (T;X) N Lq (T;X)
k∈F0 k∈F0

=: I1 (h) + I2 (h),

where (using again the induction hypothesis, now with the smaller set F0 ( F )

(q) ψ(k)
I1 (h) 6 ϕ∞,X (F0 ) max exp(2πikht0 ) − exp(i2πh ) kxk k
k∈F0 N
(q)
 ψ(k) 
6 Aϕp,X (#F0 )1/p max 2π|h| inf kt0 − − j max kxk k (13.18)
k∈F0 j∈Z N k∈F0

(q) π|h|
6 Aϕp,X (N )N 1/p max kxk k,
N k∈F
 13.1 Bourgain’s theorem on Fourier type 249

since kt0 ∈ Iψ(k) + Z.

Having estimated both I1 (h) and II in terms of the induction hypothesis,
the serious work is left with I2 (h), which we first average over a range h =
1, . . . , H 6 N , where a favourable value of H is to be determined. We have
H Z H q
1 X N 1 X X
I2 (h)q = eh (j/N )eψ−1 (j) (t)xψ−1 (j) dt
H H TN
h=1 h=1 j∈ψ(F0 )
Z N N q
N 1 X X
6 eh (j/N )yj (t) dt
H T N
h=1 j=1
(
eψ−1 (j) (t)xψ−1 (j) , j ∈ ψ(F0 ),
yj (t) :=
0, else,
Z  N i1/p q
N (q)
hX
6 ϕp,X (ZN ) kyj (t)kp dt
H T j=1
(q)
by definition of ϕp,X (ZN )
N
q  X q/p
(q)
6 3ϕp,X (N ) kxk kp by Lemma 13.1.18
H
k∈F0
N (q)
q
6 3ϕp,X (N ) (#F0 )q/p max kxk kq .
H k∈F0

Combining the previous bound with (13.17) and (13.18), we have

H
X 1 X
ek xk 6 I + II 6 (I1 (h) + I2 (h)) + II
Lq (T;X) H
k∈F h=1
1 X H 1/q
6 max I1 (h) + I2 (h)q + II (13.19)
16h6H H
h=1
 πH  N 1/q  7 1/p 
(q)
6 A +3 +A N 1/p ϕp,X (N ) max kxk k,
N H 8 k∈F

where N = #F , as we recall. We now choose H so as to essentially equate

the first two terms:

H := bH 0 c, H 0 := A−q/(q+1) N.
0
Since A > 1, we have H 6 H 0 6 N . Recalling that N > Ap , and noting that
p0 > 1 > q/(q + 1), we also observe that H 0 > 1, and hence H > 1. Thus
this choice of H lies in the admissible range considered above. We also have
H 0 6 H + 1 6 2H, and thus

πH  N 1/q πH 0  2N 1/q
A +3 6A +3 = (π + 21/q · 3)A1/(q+1) .
N H N H0
250 13 The Fourier transform and multipliers

We also note that

 7 1/p 1 7  1 7 
− 1 = ξ 1/p−1 −1 6− , for some ξ ∈ ,1 .
8 p 8 8p 8

Substituting into (13.19), we hence have

(q)
h  1  i (q)
ϕ∞,X (F ) 6 (π + 21/q · 3) · A1/(q+1) + 1 − A ϕp,X (N ) · N 1/p .
8p
To complete the induction step, it remains to check that the quantity in
brackets is at most A, which after easy simplification is the same as
1
(π + 21/q · 3) · A1/(q+1) 6 A.
8p
Clearly this is the case with the choice of A stated in the Lemma. 
We are now ready for a first uniform bound on the finite Fourier type con-
stants:

Corollary 13.1.27. Let X be a Banach space, r ∈ (1, 2], and suppose that

ϕ2,X (N ) 6 C · N 1/r−1/2 ∀N ∈ Z+ .

Then for all s ∈ (1, r), we have

(2) Cr
ϕs,X (N ) 6 3500 ∀N ∈ Z+ .
r−s
Proof. By Lemma 13.1.25 and Remark 13.1.26 with p = q = 2, we have
(2)
ϕ∞,X (F ) 6 1285 · ϕ2,X (N ) · N 1/2 6 1285 · C · N 1/r (13.20)

whenever F ⊆ Z has size #F = N .

Let x = (xk )N s
k=1 ∈ `N (X) have norm one. For α ∈ (0, 1) to be chosen, we
denote

Fj := {n ∈ Z : αj < kxn k 6 αj−1 }, x(j) := (1Fj (k) · xk )N

k=1 .

Note that Fj = ∅ and x(j) = 0 for j 6 0, and

#Fj 6 #{n ∈ Z : αj < kxn k} 6 α−js kxk`sN (X) = α−js , j > 1.

Thus
N ∞ ∞
(j) (2) (j)
X X X X
ek xk 6 ek xk 6 ϕ∞,X (Fj ) max kxk k
L2 (T;X) L2 (T;X) k∈Fj
k=1 j=1 k∈Fj j=1
 13.1 Bourgain’s theorem on Fourier type 251
∞
(2)
X
6 1285 · ϕ2,X (#Fj ) · (#Fj )1/2 · αj−1 by (13.20)
j=1
X∞ ∞
X
6 1285 · C(#Fj )1/r αj−1 6 1285 · Cα−js/r αj−1
j=1 j=1
1−s/r
1285 · C α
= .
α 1 − α1−s/r
The choice α = (s/r)r/(r−s) gives

1 α1−s/r α−s/r (r/s)s/(r−s) e

1−s/r
= 1−s/r
= 6
α1−α 1−α 1 − s/r 1 − s/r
by an elementary optimisation in the last step. Substituting back, this gives
N
X e Cr Cr
ek xk 6 1285 · C · = 1285 · e · < 3500 ·
L2 (T;X) 1 − s/r r−s r−s
k=1

for all (xk )N s

k=1 ∈ `N (X) of norm one, which is the claimed bound. 

13.1.d Conclusion via duality and interpolation

With the uniform bound of Corollary 13.1.27, we have already covered the
core of the deep implication from non-trivial type to non-trivial Fourier type.
The rest of the argument depends on the more routine techniques of duality
and interpolation, but is still not entirely straightforward. We now turn our
attention to giving these finishing touches to the proof. At the end of this
section, a statement and proof of Bourgain’s theorem will finally be given.
The first duality that we want to use is most elegantly expressed in terms
of the Fourier type constants on the cyclic group ZN :

Lemma 13.1.28. Let X be a Banach space, N ∈ Z+ and p, q ∈ (1, ∞). Then

(q) 0 (p0 )
N 1/q ϕp,X (ZN ) = N 1/p ϕq0 ,X ∗ (ZN ).

Proof. Since X is norming for X ∗ , Proposition 1.3.1 shows that `pN (X) is
0
norming for `pN (X ∗ ), so that
N
X N
X p0 1/p0
ek (h/N )x∗k
h=1 k=1
N D
nX N
X N
E X 1/p o
= sup xh , ek (h/N )x∗k : kxh kp 61 ,
h=1 k=1 h=1

where, observing the symmetry ek (h/N ) = e2πikh/N = eh (k/N ),

252 13 The Fourier transform and multipliers
N D
X N
X E XN DX
N E
xh , ek (h/N )xk∗ = eh (k/N )xh , x∗k
h=1 k=1 k=1 h=1
N
X N N
q 1/q  X 1/q0
X 0
6 eh (k/N )xh kxk∗ kq
k=1 h=1 k=1
N
X N
1/p  X 1/q0
(q) 0
6 N 1/q ϕp,X (ZN ) kxh kp kxk∗ kq .
h=1 k=1

Substituting back, this proves that

0 (p0 ) (q)
N 1/p ϕq0 ,X ∗ (ZN ) 6 N 1/q ϕp,X (ZN ).

Permuting the names of the exponents and using the isometric embedding of
X into X ∗∗ , it also follows that
(q) (q) 0 (p0 )
N 1/q ϕp,X (ZN ) 6 N 1/q ϕp,X ∗∗ (ZN ) 6 N 1/p ϕq0 ,X ∗ (ZN ),

which proves the claimed equality. 

Corollary 13.1.29. Let X be a Banach space, r ∈ (1, 2], and suppose that

ϕ2,X (N ) 6 C · N 1/r−1/2 ∀N ∈ Z+ .

Then for all s ∈ (1, r) we have

(s0 ) Cr
ϕ∞,X ∗ (F ) 6 1.35 · 107 N 1/s ∀s ∈ (1, r),
r−s
whenever F ⊆ Z is a subset of size #F = N .

Recall from Corollary 13.1.20 that if X has type p ∈ (1, 2], then the assump-
tion is satisfied with C and r as in (13.10).
Proof. By using both estimates of Lemma 13.1.18 with Lemma 13.1.28 in
between, and finally Corollary 13.1.27, we have
0 (s0 ) 0 (s0 ) (2)
N 1/s ϕ2,X ∗ (N ) 6 N 1/s ϕ2,X ∗ (ZN ) = N 1/2 ϕs,X (ZN )
(2) Cr
6 N 1/2 · 3ϕs,X (N ) 6 N 1/2 · 3 · 3500 .
r−s
Then Lemma 13.1.25 and Remark 13.1.26 with p = 2 6 q = s0 show that

(s0 ) (s0 ) Cr 0
ϕ∞,X ∗ (F ) 6 1285 · ϕ2,X ∗ (N ) · N 1/2 < 1.35 · 107 · N 1/2+1/2−1/s .
r−s
whenever F ⊆ Z is a subset of size #F = N . 
 13.1 Bourgain’s theorem on Fourier type 253

We now come to another form of duality, where we pass from the Fourier
transform on Z to that on the circle T, and it is in this latter setting that our
argument will be completed.
Lemma 13.1.30. Let X be a Banach space, 1 6 s 6 ∞, and suppose that
(s0 )
ϕ∞,X ∗ (F ) 6 K · N 1/s

whenever F ⊆ Z is a subset of size #F = N . Then the Fourier transform

Z
F : f ∈ L1 (T; X) 7→ (fb(k))k∈Z , fb(k) = e−k (t)f (t) dt,
T

satisfies the weak-type estimate

kF f k`s0 ,∞ (Z;X) 6 Kkf kLs (T;X) . (13.21)

Proof. Let f ∈ Ls (T; X), let λ > 0, and let F be a finite subset of {k ∈ Z :
kfb(k)k > λ}. (By a periodic analogue of the Riemann–Lebesgue Lemma 2.4.3,
which has essentially the same proof, we could argue that this set is finite to
begin with, but we do not need this here.) Then
1 X b 1 X b
#F 6 kf (k)k = hf (k), x∗−k i
λ λ
k∈F k∈F
for suitable x∗−k ∈ X ∗ of norm one
Z
1 X 
= f (t) e−k (t)x∗−k dt
λ T
k∈F
1 X
6 kf kLs (T;X) e k x k s0
λ L (T;X ∗ )
k∈−F
1 (s0 ) 1
6 kf kLs (T;X) ϕ∞,X ∗ (−F ) 6 kf kLs (T;X) K(#F )1/s ,
λ λ
and hence
λ(#F )1−1/s 6 Kkf kLs (T;X) .
Since this is true for any finite F ⊆ {k ∈ Z : kfb(k)k > λ}, it is also true
for F = {k ∈ Z : kfb(k)k > λ} (showing, a posteriori, the finiteness of this
set). Then the supremum over λ > 0 of the left-hand side is precisely the
0
`s ,∞ (Z; X) norm that we wanted to estimate. 
From (13.21) and the trivial fact that F is bounded from L1 (T; X) →
`∞ (Z; X), it seems apparent that we should conclude that F is bounded
0
from Lp (T; X) to `p (Z; X) by interpolation. However, the version of the
Marcinkiewicz Interpolation Theorem 2.2.3 covered in the text is not suffi-
cient for this purpose, and we would need the generalisation stated in the
Notes as Theorem 2.7.5. We will give a proof of a quantitative version of the
special case relevant for the present application:
254 13 The Fourier transform and multipliers

Lemma 13.1.31. Let X be a Banach space such that (13.21) holds for some
s ∈ (1, 2]. Then

3K
kF f k`t0 (Z;X) 6 kf kLt (T;X) ∀t ∈ (1, s).
(s − t)1/t0

Proof. By homogeneity we may assume that kf kLt (T;X) = 1. We have

Z ∞
0
kF f k` t0 (Z;X) = t0 λt −1 #{k : kfb(k)k > λ} dλ
Z0 ∞
0
6 t0 λt −1 #{k : kfbλ (k)k > θ0 λ} dλ (13.22)
0
Z ∞
0
+ t0 λt −1 #{k : kfbλ (k)k > θ1 λ} dλ,
0

where θ0 + θ1 = 1 and, with parameters A and γ to be chosen shortly,

fλ := f · 1{kf kX 6Aλγ } , f λ := f · 1{kf kX >Aλγ } .

Then
Z
kf λ kL1 (T;X) = kf kX 6 (Aλγ )1−t kf ktLt (T;X) = (Aλγ )1−t
{kf kX >Aλγ }

and hence
kfbλ k`∞ (Z;X) 6 (Aλγ )1−t 6 θ1 λ,
provided that we choose
−1/(t−1)
γ = −1/(t − 1), A = θ1 .

Then the second term on the right of (13.22) vanishes, and subsequently
Z ∞
t0 0
kF f k`t0 (Z;X) 6 t0 λt −1 #{k : kfbλ (k)k > θ0 λ} dλ
Z0 ∞
0 0 0 0
6 t0 λt −1 (θ0 λ)−s K s kfλ ksLs (T;X) dλ
0
by Lemma 13.1.30
 K  s0  Z ∞ 0 0 0
s0 /s0
= t0 λt −s −1 kfλ ksLs (T;X) dλ
θ0 0
 K  s0  Z ∞ 0 0 0
1/s0 s0
6 t0 λt −s −1 kfλ ksX dλ
θ0 0 Ls (T)

by Minkowski’s inequality with exponents s 6 s0

 K s0  Z (A/kf kX )t−1 0 0 0
1/s0 s0
= t0 λt −s −1 kf ksX dλ
θ0 0 Ls (T)
 13.1 Bourgain’s theorem on Fourier type 255

keeping in mind the choice γ = −1/(t − 1)

 K s0  1 h A i(t−1)(t0 −s0 ) 1/s0 s0
s0
= t0 kf k X
θ0 t0 − s0 kf kX Ls (T)
0
 K s 1 0 0
1  0 0 0
 1/s s
s −(t−1)(t −s )
= t0 0 0 kf kX ,
θ0 θ1t −s t0 − s0 Ls (T)

where, observing that tt0 = t + t0 , we have

s0 − (t − 1)(t0 − s0 ) = s0 − (t + t0 − ts0 − t0 + s0 ) = t(s0 − 1),

so that
1/s0 s0 s0
s0 −(t−1)(t0 −s0 ) ts0 /s

t/s
kf kX = kf kX = kf kLt (T;X) = 1.
Ls (T) Ls (T)

0
Taking θ0 = θ1 = 1
2 and using (t0 )1/t 6 e1/e < 23 , we obtain
0 0 0 0 0
2(t0 )1/t K s /t 3K s /t
kF f k`t0 (Z;X) 6 0 6 .
(t0 − s0 )1/t (t0 − s0 )1/t0

Testing (13.21) with a constant function f ≡ x, with Fourier coefficients

0 0
fb(k) = δk,0 x, shows that K > 1 and hence K s /t 6 K. Moreover,

t s t(s − 1) − s(t − 1) s−t

t0 − s 0 = − = = > s − t,
t−1 s−1 (s − 1)(t − 1) (s − 1)(t − 1)
and hence
3K
kF f k`t0 (Z;X) 6 .
(s − t)1/t0


Lemma 13.1.32. Let X be a Banach space, and suppose that there are con-
stants C and r ∈ (1, 2] such that

ϕ2,X (N ) 6 C · N 1/r−1/2

for all N ∈ Z+ . Then for all t ∈ (1, r), we have

109 · C
ϕt,X 6 .
(r − t)1+1/t0

Proof. By Corollary 13.1.29, for all s ∈ (1, r), we then have

(s0 ) Cr
ϕ∞,X ∗ (F ) 6 1.35 · 107 N 1/s =: K · N 1/s
r−s
whenever F ⊆ Z is a subset of size #F = N .
By Lemma 13.1.30, it follows that
256 13 The Fourier transform and multipliers

kF kL (Ls (T;X),`s0 ,∞ (Z;X)) 6 K,

which by Lemma 13.1.31 implies

3K 5 · 107 · C · r
ϕt,X (T) := kF kL (Lt (T;X),`t0 (Z;X)) 6 0 6
(s − t) 1/t (r − s)(s − t)1/t0

for all 1 < t < s < r. Optimising the bound with respect to s in this range,
we choose
t2 + (t − 1)r
s= .
2t − 1
With this choice, a computation shows that

t(r − t) 1 (r − t)(t − 1) 1
r−s= > (r − t), s−t= > (r − t)(t − 1).
2t − 1 3 2t − 1 3
Substituting back,
0
7 31+1/t
ϕt,X (T) 6 5 · 10 · C · r ,
(r − t)1+1/t0 (t − 1)1/t0
0
where r 6 2 and 31+1/t 6 33/2 and, for t ∈ (1, 2),
0
(t − 1)1/t = [(t − 1)t−1 ]1/t > [e−1/e ]1/t > e−1/e .

Thus
33/2 · e1/e 109 · C
ϕt,X (T) 6 108 · C 0 6 .
(r − t)1+1/t (r − t)1+1/t0

We are finally ready for the main theorem:

Theorem 13.1.33 (Bourgain). A Banach space X has non-trivial type if

and only if it has non-trivial Fourier-type. Quantitatively:
(1) If X has Fourier-type t ∈ (1, 2], then it has type t with τt,X 6 ϕt,X (Z).
(2) If X has type p ∈ (1, 2] with related constant τp,X;2 as defined in (13.7),
then it has Fourier-type
1
t=1+
6p0 (68 · τp,X;2 )p0

with constants
0

ϕt,X (R) 6 ϕt,X (T) 6 exp 2(68 · τp,X;2 )p .
 13.1 Bourgain’s theorem on Fourier type 257

Proof. (1): This is contained in Proposition 7.3.6.

(2): This is the main part of the proof, and depends on the results devel-
oped in the section. By Corollary 13.1.20, the assumptions imply that
ϕ2,X (N ) 6 C · N 1/r−1/2 ,
0
where, denoting T := (68 · τp,X;2 )p > 682 > 4 000, we have
r0 3
r0 = 3p0 T, C = e 2p0 = e 2 T .
Thus Lemma 13.1.32 shows that
109 · C
ϕt,X (T) 6 , t ∈ (1, r),
(r − t)1+1/t0
where r > 1 + (3p0 T )−1 . Hence, choosing t := 1 + (6p0 T )−1 ∈ (1, r), we have
0 3
r − t > (6p0 T )−1 , (r − t)1+1/t 6 (6p0 T ) 2 .
Thus, noting that p0 6 p0 log(68τp,X;2 ) = log T , where T > 682 > 4 000,
3 3
ϕt,X (T) 6 109 · e 2 T · (6p0 T ) 2
3 3 3 3
= 109 · 6 2 · (log T ) 2 · T 2 · e 2 T
1 1 1 3
6 e 6 T · e 6 T · e 6 T · e 2 T = e2T .
Finally, ϕt,X (R) 6 ϕt,X (T) is part of Propositions 13.1.1. 
Example 13.1.34. For each r ∈ [2, ∞), the space X = Lr (S) has type 2 with
τ2,X;2 = κr,2,K (the Kahane–Khintchine constant from the scalar-valued case
of Theorem 6.2.4), but Fourier-type t if and only if t ∈ [1, r0 ]. Hence, any esti-
mate of the Fourier-type exponent in terms of the type of X must necessarily
depend not only on the type exponent but also on the type constant of X.
Proof. The estimate τ2,X;2 6 κr,2,K follows from
N
X N
X N
X
εn x n 6 εn x n = εn x n
L2 (Ω;Lr (S)) Lr (Ω;Lr (S)) Lr (S;Lr (Ω))
n=1 n=1 n=1

6 κr,2,K {xn }N
n=1 6 κr,2,K {xn }N
n=1 .
Lr (S;`2N ) `2N (Lr (S))

For the reverse estimate, it suffices to pick some non-zero φ ∈ Lr (S) and
observe that the type inequality for xn = an φ ∈ X implies the Kahane–
Khintchine inequality for an ∈ K.
The fact that X has Fourier-type t if t ∈ [1, r0 ] follows from the scalar-
valued Hausdorff–Young inequality and Minkowski’s inequality:
kfbkLt0 (R;Lr (S)) 6 kfbkLr (S;Lt0 (R)) 6 kf kLr (S;Lt (R)) 6 kf kLt (R;Lr (S))
We indicate two alternative proofs of the “only if” part:
258 13 The Fourier transform and multipliers

(1) In Example 2.1.15, it is verified directly that the Fourier transform is not
0 0 0
bounded from Lp (R; `r ) to Lp (R; `r ) for p ∈ (r0 , 2]. By duality, it is also
0
not bounded from Lp (R; `r ) to Lp (R; `r ).
(2) Proposition 7.3.6 says that if X has Fourier type p, then it has cotype p0 .
But Corollary 7.1.6 says that Lr (S) has cotype p0 only for p0 ∈ [r, ∞].
This concludes the verification of the example. 
We also record the following simpler variant, which is nevertheless sufficient
for many purposes:
1
Proposition 13.1.35. Let X have type p and cotype q, where p − 1q < 12 . Let

1 1 1 1 h1 
:= + − ∈ , 1 .
r 2 p q 2
Then X has every Fourier-type t ∈ (1, r), and
τp,X;2 cq,X;2
ϕt,X (R) 6 ϕt,X (T) 6 109
(r − t)1+1/t0
Proof. By Proposition 13.1.21, we have
1 1 1 1
ϕ2,X (N ) 6 N p − q = N r − 2 , C := τp,X;2 cq,X;2

Thus Lemma 13.1.32 implies the bound for ϕt,X (T), and Proposition 13.1.1
the bound for ϕt,X (R). 

Remark 13.1.36. The assumptions of Proposition 13.1.35 are satisfied by many

“common” spaces of nontrivial type (and hence finite cotype). Namely, such
space often have type or cotype 2, and hence either p1 − 1q = 12 − 1q < 12 or
1 1 1 1 1 1
p − q = p − 2 < 1 − 2 = 2.

13.2 Fourier multipliers as singular integrals

The goal of this section is to see how the results on singular integrals proved
above can be applied to the theory Fourier multipliers developed in Sections
5.3 and 5.5. Given m ∈ bL∞ (Rd ; L (X,bY )), we recall that the operator Tm is
a priori defined as Tm : L1 (Rd ; X) → L1 (Rd ; Y ) by
Z
Tm f (x) = m(ξ)fb(ξ)e2πiξ·x dξ.
Rd

The notation MLp (Rd ; X, Y ) stands for the space of all m ∈ L∞ (Rd ; L (X, Y ))
for which Tm extends to a bounded linear operator from Lp (Rd ; X) to
Lp (Rd ; Y ). The connection of Fourier multipliers to integral operators is par-
ticularly simple in the following special case:
 13.2 Fourier multipliers as singular integrals 259

c (R ; L (X, Y )).
Let X, Y be Banach spaces and m ∈ L∞ d
Proposition 13.2.1. b
1 1 d
Then for all f ∈ L ∩ L (R ; X), we have
Z
Tm f (x) = k(x − y)f (y) dy,
Rd
b b
where k = m ∈ L1 (Rd ; L (X, Y )).

Proof. Under these assumptions, we can make a direct computation

Z
Tm f (x) = m(ξ)fb(ξ)e2πix·ξ dξ
Rd
Z Z 
= m(ξ) f (y)e−2πiy·ξ dy e2πix·ξ dξ
d Rd
ZR  Z  Z b
= m(ξ)e2πi(x−y)·ξ dξ f (y) dy = m(x − y)f (y) dy,
Rd Rd Rd
b
where the first step is the definition of Tm for f ∈ L1 (Rd ; X), the second is
the definition of fb(ξ) for f ∈ L1 (Rd ; X), the third is Fubini’s theorem that
applies since both m ∈ L1 (Rd ; L (X, Y )) and f ∈ L1 (Rd ; X), and the fourth
is the definition of the inverse Fourier transform of m ∈ L1 (Rd ; L (X, Y )). 
The compact support assumption on m in Proposition 13.2.1 is not as restric-
tive as it may seem at first sight, as one can often reduce considerations to
this situation by simple limiting arguments that we shortly explain. Recall
from Definition 5.5.20 that ψ ∈ S (Rd ) is called a smooth Littlewood–Paley
function if
1
(i) ψb is smooth, non-negative, and supported in {ξ ∈ Rd : 2 6 |ξ| 6 2};
X
(ii) b −j ξ) = 1 for all ξ ∈ Rd \ {0}.
ψ(2
j∈Z

Such functions exist by Lemma 5.5.21, whose proof also gives the identity
ψ(ξ) b − ϕ(2ξ)
b = ϕ(ξ) b and hence
X
b −j ξ) = ϕ(2
ψ(2 b −N ξ) − ϕ(2
b −L ξ)
L<j6N

b ∈ D(Rd ) with ϕ(0)

R
for some ϕ b = ϕ = 1. Let

b −j ξ)m(ξ),
mj (ξ) := ψ(2 b −N ξ)m(ξ),
mN (ξ) := ϕ(2
(13.23)
X
mN N L
L (ξ) := m (ξ) − m (ξ) = mj (ξ),
L<j6N

c (R ; L (X, Y )), whereas

and observe that mN ∈ L∞ d

c (R \ {0}; L (X, Y )),

mj , mN ∈ L∞ d
260 13 The Fourier transform and multipliers

i.e., these are supported away from both the origin and infinity. While the
support away from zero is not required by Proposition 13.2.1, it is a conve-
nience for forthcoming considerations due to the special role of the origin in
various multiplier conditions. The next two lemmas describe a precise sense in
which, for many purposes, it is “enough” to study the truncated multipliers
mN .

Lemma 13.2.2. Let X, Y be Banach spaces and m ∈ L∞ (Rd ; L (X, Y )). For
p ∈ (1, ∞), we have m ∈ MLp (Rd ; X, Y ), if and only if mN ∈ MLp (Rd ; X, Y )
uniformly in N , if and only if mN p d
L ∈ ML (R ; X, Y ) uniformly in M and N .

Proof. By the algebra of multipliers (Lemma 5.3.2), we have

b −N ·) f ) = Tm (ϕ2−N ∗ f ),
TmN f = Tm (Tϕ(2

where ϕt (x) = t−d ϕ(t−1 x) and

kϕt ∗ f kp 6 kϕt k1 kf kp = kϕk1 kf kp ,

so that kmN kMLp (Rd ;X,Y ) 6 kϕk1 kmkMLp (Rd ;X,Y ) , and thus

kmN
L kMLp (Rd ;X,Y ) 6 2kϕk1 kmkMLp (Rd ;X,Y ) .

On the other hand, it is evident from property (ii) of Littlewood–Paley

functions that mN (ξ) → m(ξ) as N → ∞ for every ξ ∈ Rd , and mN L (ξ) →
m(ξ) as N → ∞ and L → −∞ for every ξ ∈ Rd \ {0}. In particular, both
limits hold for almost every ξ ∈ Rd . Then Proposition 5.3.16 implies that

kmkMLp (Rd ;X,Y ) 6 lim inf kmN kMLp (Rd ;X,Y ) ,

N →∞
N
kmkMLp (Rd ;X,Y ) 6 lim inf kmL kMLp (Rd ;X,Y ) .
N →∞
L→−∞

13.2.a Smooth multipliers have Calderón–Zygmund kernels

We will be mostly concerned with multipliers satisfying Mihlin-type conditions

of the form
k∂ α m(ξ)k 6 M |ξ|−|α| , ξ ∈ Rd \ {0}, (13.24)
for some set of multi-indices α ∈ Nd . Recall that the Mihlin class, introduced
and used in Definitions 5.3.17 and 5.5.9 and Theorems 5.3.18 and 5.5.10 (in
one and several variables, respectively) to deduce that m ∈ MLp (Rd ; X, Y )
for all p ∈ (1, ∞) without any a priori boundedness assumptions on Tm , fea-
tured stronger R-boundedness versions of such conditions. The difference in
the present context is that we are willing to assume that m ∈ MLp0 (Rd ; X, Y )
for some p0 ∈ (1, ∞) to begin with, and we wish to show that this a priori
 13.2 Fourier multipliers as singular integrals 261

boundedness on one space can then be extrapolated to boundedness on other

function spaces under conditions that are similar to those in Mihlin’s theo-
rems, but without the R-bounded aspects. As a matter of fact, these pointwise
bounds can often be relaxed to weaker integrated versions, which is easily ver-
ified by inspecting the proofs, but for the clarity of the exposition we state
the results under such pointwise assumptions. This is hardly a restriction for
most applications.
The role of the multiplier conditions (13.24) for the kernel estimates is via
careful use of the fundamental relation ∂dj f (ξ) = 2πiξj f (ξ). So as to make most
b
efficient use of the relation, and to unburden the formulae from inessential
constants, we introduce the abbreviation

∂/ := ∂/2πi

so that
/j f (ξ) = ξj fb(ξ).
∂d
The deduction of the kernel estimates is easiest when sufficiently many
derivatives are allowed in (13.24); as it turns out, this is somewhat more
than the collection α ∈ {0, 1}d appearing in Mihlin’s Theorem 5.5.10. We
formulate several results for a generic Banach space Z instead of L (X, Y ), as
the operator structure plays no role here; this also makes the formulae slightly
shorter. We say that a collection A of multi-indices is convex, if α ∈ A implies
β ∈ A whenever 0 6 β 6 α.

Lemma 13.2.3. If m ∈ L∞ (Rd ; Z) satisfies (13.24) for a convex set of multi-

indices α, then each mj ∈ Lc∞ (B(0, 2j+1 ); Z) satisfies
α
k∂/ mj k∞ 6 M 2−j|α|

for the same set of multi-indices, where M is the constant of (13.24).

Proof. By the Leibniz rule, we have

X α 
α b −j ξ)m(ξ)] =
α
∂ mj (ξ) = ∂ [ψ(2 2−j|θ| ∂ θ ψ(2
b −j ξ)∂ α−θ m(ξ),
θ
θ6α

where each ∂ α−θ m also satisfies (13.24) by convexity. Thus

X α 
k∂ α mj (ξ)k 6 2−j|θ| 12j−1 6|ξ|62j+1 M |ξ|−|α−θ|
θ
θ6α
X α 
6 2−j|θ| M (2j−1 )−|α|+|θ|
θ
θ6α
X α 
= M 2−j|α| 2|α| 2|α−θ| · 1|θ|
θ
θ6α
262 13 The Fourier transform and multipliers

= M 2−j|α| 2|α| (2 + 1)|α| = M 2−j|α| 6|α| ,

where the binomial formula was used in the second to last step. The result
follows after dividing both sides by (2π)|α| > 6|α| . 

Lemma 13.2.4. Let Z be a Banach space and f ∈ L∞

c (B(0, A); Z) have dis-
tributional derivatives that satisfy
α
k∂/ f k∞ 6 A−|α|

for some A > 0 and all multi-indices α in some convex set. Then
α
kx 7→ ∂/x [(e2πiy·x − 1)f (x)]k∞ 6 (6 + 2|α| )A|y| · A−|α|

for all y ∈ Rd with |y| 6 A−1 , and for the same set of multi-indices.

Proof. The derivatives are given by

α α α−γ
X
∂/x [(e2πiy·x − 1)f (x)] = (e2πiy·x − 1)∂/ f (x) + y γ e2πiy·x ∂/ f (x),
06=γ6α

and hence
α
X
k∂/x [(e2πiy·x − 1)f (x)]k 6 2π|y|A · A−|α| + |y||γ| A−|α|+|γ|
06=γ6α
 X 
−|α|
6 |y|A · A 2π + (A|y|)|γ|−1 .
06=γ6α

If A|y| 6 1, then (A|y|)|γ|−1 6 1 and 1 = 2|α| − 1.

P
06=γ6α 

Lemma 13.2.5. Let Z be a Banach space and f ∈ L∞

c (B(0, A); Z) have dis-
tributional derivatives that satisfy
α
k∂/ f k∞ 6 A−|α| ∀|α| 6 d + 1

for some A > 0. Then for almost all x, y ∈ Rd with |y| 6 21 |x|, we have

|x|n |fb(x)| 6 cd Ad−n , (13.25)

|x|n |fb(x − y) − fb(x)| 6 cd Ad−n min{A|y|, 1} (13.26)

for all n = 0, 1, . . . , d + 1. In particular, fb ∈ L1 (Rd ; Z) and

kfbk1 6 cd .

Proof. For x ∈ B(0, A), we have

α α
kxα fbk∞ 6 k∂/ f k1 6 k∂/ f k∞ k1B(0,A) k1 6 A−|α| ωd Ad ,
 13.2 Fourier multipliers as singular integrals 263

where ωd is the volume of the unit ball in Rd . With α = nei , this shows that
|xi |n |fb(x)| 6 ωd Ad−n for i = 1, . . . , d, which readily gives (13.25).
We observe that fb(x − y) − fb(x) is the Fourier transform of (e2πix·y −
1)f (x), which satisfies the same assumptions as f for |y| 6 A−1 , except for a
multiplicative factor (6 + 2d )A|y|, by Lemma 13.2.4. An application of (13.25)
to this function in place of f hence gives

|x|n |fb(x − y) − fb(x)| 6 cd Ad−n A|y|

when A|y| 6 1. On the other hand, if A|y| > 1, then we simply estimate
fb(x−y)− fb(x) by (13.25) and the triangle inequality, recalling the assumptions
that |y| 6 21 |x| and n 6 d + 1:

|fb(x − y) − fb(x)| 6 |fb(x − y)| + |fb(x)| 6 cd Ad−n (|x − y|−n + |x|−n )

6 cd Ad−n (2n + 1)|x|−n 6 c0d Ad−n |x|−n .

The last two bounds are both seen to be dominated by the claimed bound in
(13.26).
That fb ∈ L1 (Rd ; Z) is immediate from (13.25) by integrating the estimate
n o
|fb(x)| 6 cd Ad min 1, (A|x|)−d−1 .

Proposition 13.2.6. Let X, Y be Banach spaces and m ∈ L∞ (Rd ; L (X, Y ))

satisfy
k∂ α m(ξ)k 6 M |ξ|−|α| ∀|α| 6 d + 1.
b
N N N
Then each K (x, y) = k (x − y) = m (x − y) is a Calderón–Zygmund kernel
with the following bounds independent of the truncation N :

c 1  |y| 
kk N (x)k 6 , kk N (x − y) − k N (x)k 6 ω ,
|x|d |x|d |x|

for all x, y ∈ Rd with |y| 6 21 |x|, where

 1
c = cd M, ω(t) = cd M · t · 1 + log .
t
Note that the modulus of continuity ω above is slightly “worse” (i.e., with
slower decay as t → 0) than the Lipschitz modulus ω1 (t) = t, but “better”
than any of the Hölder moduli ωδ (t) = tδ for δ ∈ (0, 1).
Proof. By Lemma 13.2.3, the functions mj satisfy the assumptions, and hence
the conclusions, of Lemma 13.2.5 with A = 2j+1 and a multiplicative factor
cd M . Thus,
264 13 The Fourier transform and multipliers

X X cd 2(j+1)(d−h)
|k N (x)| 6 |kj (x)| 6 min M
06h6d+1 |x|h
j6N j∈Z
X X cd 2−(j+1)
6 cd 2(j+1)d M + M 6 c0d |x|−d M.
|x|d+1
j:2j+1 61/|x| j:2j+1 >1/|x|

Similarly, for |y| 6 21 |x|,

X cd 2(j+1)(d−h)
|k N (x − y) − k N (x)| 6 min min{2j+1 |y|, 1}M
06h6d+1 |x|h
j∈Z
X
(j+1)(d+1)
X cd
6 cd 2 |y|M + |y|M
|x|d+1
j:2j+1 61/|x| j+1
j:1/|x|62 61/|y|
−(j+1)
X cd 2
+ M
|x|d+1
j:2j+1 >1/|y|

1 c0d  |x|  c0d

6 c0d d+1
|y|M + d+1
|y| 1 + log M + d+1 |y|M
|x| |x| |y| |x|
c00d  |x| 
6 |y| 1 + log M.
|x|d+1 |y|

This completes the proof. 

With the uniform pointwise bounds of Proposition 13.2.6 at hand, we can
strengthen the sense in which the operator Tm with such bounds is associated
with a Calderón–Zygmund kernel k:

Proposition 13.2.7. Let X, Y be Banach spaces, p ∈ [1, ∞), and m ∈

MLp (Rd ; X, Y ) satisfy

k∂ α m(ξ)k 6 M |ξ|−|α| ∀|α| 6 d + 1.

Then there is a kernel k ∈ C(Rd \{0}; L (X, Y )) that satisfies the same bounds
as k N in Proposition 13.2.6 and such that
Z
Tm f (x) = k(x − y)f (y) dy
Rd

for all f ∈ Lp (Rd ; X) and almost all x ∈ Rd outside the support of f .

Proof. We split the proof into two cases:

Case p ∈ (1, ∞): Let f ∈ Lp (Rd ; X). Using the notation from the proof of
Lemma 13.2.2 and the preceding discussion, we have

TmN
L
f = Tm [(ϕ2−N ∗ f ) − (ϕ2−L ∗ f )],
 13.2 Fourier multipliers as singular integrals 265

where ϕ2−N ∗ f → f in Lp (Rd ; X) as N → ∞ by a standard mollifier re-

sult (e.g., Proposition 1.2.32). We also have kϕR ∗ f kp 6 kϕR kp kf k1 =
0
R−n/p kf k1 → 0 as R → ∞ if f ∈ L1 (Rd ; X) and kϕR ∗ f kp 6 kϕR k1 kf kp =
kf kp uniformly in R. Since (L1 ∩ Lp )(Rd ; X) is dense in Lp (Rd ; X), it follows
that ϕ2−L ∗ f → 0 in Lp (Rd ; X) as L → −∞ for all f ∈ Lp (Rd ; X).
Summarising this discussion, it follows that, for all f ∈ Lp (Rd ; X), we
have the convergence TmLN f → f in Lp (Rd ; X) as N → ∞ and L → −∞. By
passing to a subsequence if needed, we may assume that this convergence also
takes placeb almost everywhere.
If f ∈ L1 (Rd ; X) ∩ L1 (Rd ; X) ⊆ Lp (Rd ; X), then Proposition 13.2.1 shows
that
N
TmLN f = kL ∗ f,

where TmN L
Lp (Rd ; Y ) by Lemma 13.2.2. On
is bounded from Lp (Rd ; X) to b
N
the other hand, kL is a finite sum of kj = mj , where the multipliers mj are
in the scope of Lemma 13.2.5, and hence kL N
∈ L1 (Rd ; L (X, Y )). But then
also f 7→ kN ∗ f is bounded from L (R ; X) to Lp (Rd ; Y ), and the previous
p d

display must remain valid for all f ∈ Lp (Rd ; X) by continuity. Combining

these pieces, we obtain
Z
N
Tm f (x) = lim TmLN f (x) = lim kL (x − y)f (y) dy
N →∞ N →∞ Rd
L→−∞ L→−∞

for all f ∈ Lp (Rd ; X) and almost every x ∈ Rd .

Let us finally consider x ∈ { supp f . Since this set is open, we can pick an
ε > 0 such that B(x, ε) ⊆ { supp f . For such x and any y ∈ supp f , the series
X
N
kj (x − y) = lim kL (x − y)
N →∞
j∈Z L→−∞

converges absolutely by the proof of Proposition 13.2.6. We denote by k(x−y)

the limit. Moreover, the same proposition shows that

N cd M
kkL (x − y)f (y)k 6 kf (y)k,
|x − y|d

which is integrable over y ∈ Rd by Hölder’s inequality, since f ∈ Lp (Rd ; X)

0
and [y 7→ |x − y|−d ] ∈ Lp ({B(x, ε)). Thus
Z Z
N
Tm f (x) = lim kL (x − y)f (y) dy = k(x − y)f (y) dy
N →∞ Rd Rd
L→−∞

N
by dominated convergence. The pointwise estimates of kL are clearly inherited
by k by the pointwise convergence. This completes the proof for p ∈ (1, ∞).
266 13 The Fourier transform and multipliers

Case p = 1: We can still make use of large parts of the preceding consid-
erations, but some details require a modification. The standard mollifier re-
sult (Proposition 1.2.32) still applies to show that ϕ2−N ∗ f → f , and hence
TmN f → Tm f , in L1 (Rd ; X) as N → ∞, but it no longer guaranteed that
ϕ2−L ∗ f should converge to 0 as L → −∞. Hence, we will separately deal
with TmL f . b
For f ∈ L1 (Rd ; X) ∩ L1 (Rd ; X), we have
Z
kTmL f (x)k = ϕ(2−L ξ)m(ξ)fb(ξ) dξ
Rd
Z
6 kmk∞ kfb(ξ)k dξ 6 ωd (2L+1 )d kmk∞ kf k1 .
|ξ|62L+1

Hence TmL extends to a bounded operator from L1 (Rd ; X) to L∞ (Rd ; Y ) of

b ωd 2(L+1)d kmk∞ → 0 as L → −∞.
norm at most
For f ∈ L1 (Rd ; X) ∩ L1 (Rd ; X), we can now write

TmN f = Tm N
L
f + Tm L f = k mN
L
∗ f + TmL f.

Since all of the operators acting on f above are bounded from L1 (Rd ; X) to
L1 (Rd ; Y ) + L∞ (Rd ; Y ), the identity continues to hold for all f ∈ L1 (Rd ; X).
Taking the limits N → ∞ and L → −∞, we have TmN f → Tm f in L1 (Rd ; Y )
and TmL f → 0 in L∞ (Rd ; Y ). Along suitable subsequences, we have both
limits almost everywhere, and hence we arrive at the same pointwise limit
Z
N
Tm f (x) = lim kL (x − y)f (y) dy
N →∞ Rd
L→−∞

as in the case p ∈ (1, ∞). The rest of the proof can then be concluded in
the same way as before. Specifically, let us note that the final application
of dominated convergence is justified simple because the product of [y 7→
|x − y|−d ] ∈ L∞ ({B(x, ε)) and f ∈ L1 (Rd ; X) is integrable. 

Corollary 13.2.8. Let X, Y be Banach spaces and p0 ∈ [1, ∞). Suppose that
m ∈ MLp0 (Rd ; X, Y ) satisfies

k∂ α m(ξ)k 6 M |ξ|−|α| ∀|α| 6 d + 1.

Then Tm extends to a bounded operator from Lp (w; X) to Lp (w; Y ) for every

p ∈ (1, ∞) and every Muckenhoupt weight w ∈ Ap . Moreover,
1
max{1, p−1 }
kTm kL (Lp (w;X),Lp (w;Y )) 6 cd,p (kmkMLp0 (Rd ;X,Y ) + M )[w]Ap .

Proof. By Proposition 13.2.7, the A2 Theorem 11.3.26 applies to such an

operator Tm , and this gives precisely the stated conclusions. 
 13.2 Fourier multipliers as singular integrals 267

Corollary 13.2.9. Let X, Y be UMD spaces. Suppose that m ∈ L∞ (Rd ; X, Y )

satisfies
k∂ α m(ξ)k 6 M |ξ|−|α| ∀|α| 6 d + 1,
and in addition
R |ξ||α| ∂ α m(ξ) : ξ ∈ Rd \ {0} 6 M ∀α ∈ {0, 1}d ,


f

Then Tm extends to a bounded operator from Lp (w; X) to Lp (w; Y ) for every

p ∈ (1, ∞) and every Muckenhoupt weight w ∈ Ap . Moreover,
1
d d max{1, p−1 }
kTm kL (Lp (w;X),Lp (w;Y )) 6 cd,p (min(~p,X , ~p,Y f+M )[w]
)βp,X βp,Y M Ap .
Proof. By Mihlin’s Multiplier Theorem 5.5.10, the assumptions imply that
kmkMLp (Rd ;X,Y ) 6 cd min(~dp,X , ~dp,Y )βp,X βp,Y M
f.

We then conclude with an application of Corollary 13.2.8. 

This proof displays a certain dichotomy between the multiplier conditions
needed to get the boundedness of Tm to begin with, and the conditions needed
to extrapolate this boundedness to other spaces. The former one needs the
stronger R-boundedness assumptions, but only for a smaller number of deriva-
tives, while the latter only needs usual pointwise bounds, but for a larger set of
derivatives. This dichotomy disappears from sight in the following important
special case:
Corollary 13.2.10. Let X be a UMD space. Suppose that a scalar-valued
m ∈ L∞ (Rd ) satisfies
|∂ α m(ξ)| 6 M |ξ|−|α| ∀|α| 6 d + 1.
Then Tm extends to a bounded operator on Lp (w; X) for every p ∈ (1, ∞) and
every Muckenhoupt weight w ∈ Ap . Moreover,
1
max{1, p−1 }
kTm kL (Lp (w;X))) 6 cd,p ~dp,X βp,X
2
M [w]Ap .
Proof. The assumed pointwise bounds coincide with the R-bounds required
by Corollary 13.2.10 in the case of a scalar-valued multiplier m. 

13.2.b Mihlin multipliers have Hörmander kernels

We now turn to the question of kernel estimates assuming only the multiplier
conditions appearing in Mihlin’s Theorem 5.5.10. It turns out that the max-
imal order of d derivatives is just on the border of what we need to make
useful estimates, and in order to cope with this condition, we need to impose
an additional assumption on the underlying Banach space X in terms of the
notion of Fourier type discussed in Section 13.1.
The analogue of Lemma 13.2.5 in the present context is the following
rather more complicated assertion.
268 13 The Fourier transform and multipliers

Lemma 13.2.11. Let X be a Banach space with Fourier type p ∈ (1, 2]. Let
f ∈ L∞ d
c ((−A, A) ; X) satisfy
α
k∂/ f k∞ 6 A−|α| ∀α ∈ {0, 1}d

for some A > 0. Then fb ∈ L1 (Rd ; X) and, denoting

Φp,X := 4p0 (4 + log2+ ϕp,X ),

we have the estimates

kfbk1 6 Φdp,X , (13.27)
4dϕp,X
k1{[−R,R]d fbk1 6 Φdp,X ∀R > 0, (13.28)
(AR)1/p0
kfb(· − y) − fb(·)k1 6 Φdp,X · 4 · 2d A|y| ∀y ∈ Rd , (13.29)
n 8d2 ϕ o
p,X
k1{B(0,3|y|) [fb(· − y) − fb(·)]k1 6 Φdp,X min 2, , 4 · 2d
Ar . (13.30)
(Ar)1/p0
Remark 13.2.12. Thanks to Bourgain’s Theorem 13.1.33, the assumption on
the Banach space X in Lemma 13.2.11 is simply that X has some non-trivial
type r ∈ (1, 2]. Namely, Theorem 13.1.33 guarantees that we can then take
0
p0 = 1 + 6r0 T, ϕp,X 6 e2T , T := (68τr,X;2 )r > 682 > 4 000,

and hence

 2  48 1 2 log 2
Φp,X 6 4 1 + 6r0 T 4+ T = ( + 1)( + 1)r0 T 2
log 2 log 2 6r0 T T
0
6 70 · r0 T 2 = 70r0 (68τr,X;2 )r .

Proof of (13.27). For k ∈ Zd , let

Dk = {x ∈ Rd : xi ∈ [2ki , 2ki +1 ) ∀i = 1, . . . , d}

so that obviously X
kfbk1 = k1Dk f k1 .
k∈Zd

For each k ∈ Zd , we partition 1 = α + β + γ for some α, β, γ ∈ {0, 1}d yet to

be chosen. Then

Dk = Dkα × Dkβ × Dkγ , Dkθ = {(xi )i:θi =1 : xi ∈ [2ki , 2ki +1 )}.

Similarly,

Rd = R α × Rβ × Rγ , Rθ = {(xi )i:θi =1 : xi ∈ R},

and we abbreviate Ls Ltγ := Ls (Rα × Rβ ; Lt (Rγ ; X)).

 13.2 Fourier multipliers as singular integrals 269

For x ∈ Dk , we have |xi | > 2ki , and hence |xβ+γ | > 2k·(β+γ) . We can now
make the following estimate. At a critical point, passing from a norm of the
Fourier transform fb to a norm of f itself, we apply the Fourier type assumption
0 |γ|
to F : Lp (Rγ ; X) → Lp (Rγ ; X), producing the constant ϕp,X (Rγ ) 6 ϕp,X ,
and the trivial boundedness of the Fourier transform F : L1 (Rα+β ; Z) →
L∞ (Rα+β ; Z), with Z = Lq (Rγ ; X) for either q = p or q = p0 , depending on
the (irrelevant) order in which we perform these two steps:

k1Dk fbkL1 6 2−k·(β+γ) k1Dk xβ+γ fbkL1

6 2−k·(β+γ) k1Dk kL1 Lγp kxβ+γ fbkL∞ Lp0
γ

−k·(β+γ) d k·(α+β+γ/p) |γ| β+γ

62 ·2 2 · ϕp,X k∂/ f kL1 Lγp
0 |γ|
6 2d 2k·(α−γ/p ) · ϕp,X A−|β|−|γ| 2d A|α|+|β|+|γ|/p
0 |γ| 0
6 4d 2k·(α−γ/p ) · ϕp,X A|α|−|γ|/p
Y Y Y 0
= 4d × (A2ki ) × 1× (ϕp,X (2ki A)−1/p ).
i:αi =1 i:βi =1 i:γi =1

Since the splitting 1 = α + β + γ is free for us to choose, it is obvious that,

for each i, we choose it to be in the first, second or third category according
to which of the three numbers
0
A2ki , 1, ϕp,X (2ki A)−1/p

is the smallest. This gives us the estimate

X
kfbk1 = k1Dk f k1
k∈Zd
d
XY 0
6 4d min{A2ki , 1, ϕp,X (2ki A)−1/p }
k∈Zd i=1
X 0
d
= 4d min{A2k , 1, ϕp,X (2k A)−1/p }
k∈Z
 X X X 0
d
6 4d A2k + 1+ ϕp,X (2k A)−1/p
0 0
k:A2k 61 k:16A2k 6ϕp k:A2k >ϕp
p,X p,X

p0 0
 0 ϕp,X (ϕp,X )−1/p d
64 d
2 + (1 + log+
2 ϕpp,X ) +
1 − 2−1/p0
0
d
6 4 (3 + p log+
2 ϕp,X + 2p ) 6 (4p0 )d (4 + log+
0 d
2 ϕp,X )
d

0
where we observed that 1 − 2−1/p > 1/(2p0 ), since the function g(u) = u/2 +
2−u satisfies g(u) 6 1 for u = 1/p0 ∈ [0, 21 ], being convex with g(0) = 1 and
g( 12 ) = 1/4 + 2−1/2 < 1. 
270 13 The Fourier transform and multipliers

Proof of (13.28). Making the same decomposition

X
k1{[−R,R]d fbk1 = k1Dk 1{[−R,R]d fbk1
k∈Zd

as in the proof of (13.27), we observe hat 1Dk 1{[−R,R]d is non-zero only if at

least one ki satisfies 2ki +1 > R. Thus
d
X X
k1{[−R,R]d fbk1 6 k1Dk fbk1
i=1 k∈Zd
2ki >R/2
X 0
d−1
6 d · 4d min{A2k , 1, ϕp,X (2k A)−1/p }
k∈Z
 X 0

× min{A2k , 1, ϕp,X (2k A)−1/p } ,
k:2k >R/2

by inspection of the proof of (13.27). The factor raised to power d − 1 is

estimated as in the proof of (13.27) by
X 0
d−1  d−1
min{A2k , 1, ϕp,X (2k A)−1/p } 6 (p0 )d−1 4 + log+
2 ϕ p,X .
k∈Z

On the other hand, we have

X 0
min{A2k , 1, ϕp,X (2k A)−1/p }
k:2k >R/2
X 0
6 ϕp,X (2k A)−1/p
k:2k >R/2
0
ϕp,X (AR/2)−1/p 0
6 −1/p 0 6 4p0 ϕp,X (AR)−1/p ,
1−2
again by recycling some estimates from the proof of (13.27). Collecting the
bounds, the proof of (13.28) is complete. 

Proof of (13.29). We observe that fb(x − y) − fb(x) is the Fourier transform of

f (x)e2πix·y , which verifies the same assumptions as f by Lemma 13.2.4, aside
from the multiplicative factor (6 + 2d )A|y|, provided that A|y| 6 1. Applying
(13.27) to this function gives (13.29) for A|y| 6 1. But for A|y| > 1, (13.29)
is an immediate consequence of (13.27) by the triangle inequality. 
Proof of (13.30). This final bound is a certain synthesis of the other bounds.
The first and third bounds in the minimum are obtained from (13.27) (with
the triangle inequality) and from (13.29), respectively, ignoring the restriction
to {B, which only increases the norm.
 13.2 Fourier multipliers as singular integrals 271

For the second bound, we also use the triangle inequality, but keeping the
restriction to {B. Then
k1{B(0,3|y|) fb(· − y)k1 = k1{B(−y,3|y|) fbk1
6 k1{B(0,2|y|) fbk1 6 k1{(−2r/√d,2r/√d)d ) fbk1 ,

and the same√bound is obvious for fb in place of fb(· − y). Applying (13.28)
with R = 2r/ d produces the required bound. 
Proposition 13.2.13. Let X, Y be Banach spaces, and suppose that m ∈
L∞ (Rd ; L (X, Y )) satisfies

k∂ α m(ξ)k 6 M |ξ|−|α| ∀α ∈ {0, 1}d . (13.31)

b
N
Let K N (t, s) = k N (t−s) = m (x−y) be the kernels related to the Littlewood–
Paley truncations mN of m as in (13.23).
(1) If the space Y has Fourier type p ∈ (1, 2], then the kernels K N satisfy the
Hörmander condition uniformly in N , and quantitatively
Z
k(k N (t − s) − k N (t))xkY dt 6 (2Φp,Y )d+1 M kxkX ∀x ∈ X,
|t|>3|s|

where Φp,Y = 4p0 (4 + log+ 2 ϕp,Y ).

(2) If the space X has Fourier type p ∈ (1, 2], then the kernels K N satisfy the
dual Hörmander condition uniformly in N , and quantitatively
Z
k(k N (t − s)∗ − k N (t)∗ )y ∗ k dt 6 (2Φp,X )d+1 M ky ∗ kY ∗ ∀y ∗ ∈ Y ∗ ,
|t|>3|s|

where Φp,X = 4p0 (4 + log+

2 ϕp,X ).

Proof of (1). From Lemma 13.2.3 it follows that each Littlewood–Paley trun-
cation mj ∈ L∞c (B(0, 2
j+1
); L (X, Y )) satisfies
α
k∂/ mj k∞ 6 2d M 2−(j+1)|α| ,

which is like the condition of Lemma 13.2.11 with A = 2j+1 and an additional
multiplicative constant 2d M .
Moreover, for x ∈ X, the function mj (·)x ∈ L∞ c (B(0, 2
j+1
); Y ) satisfies
d
the same assumption with constant 2 M kxk, and now the range Y also has
Fourier type p ∈ (1, 2], as required to apply Lemma 13.2.11. In particular,
from (13.30), we conclude that
Z
k(kj (t − s) − kj (t))xkY dt
|t|>3|s|
n 8d2 ϕp,Y o
6 Φdp,Y 2d M kxk min 2, j+1 1/p0 , 8 · 2d 2j+1 r .
(2 r)
272 13 The Fourier transform and multipliers
b
N
Since mN ∈ Lc∞ (Rd ; L (X, Y )) ⊆ L1 (Rd ; L (X, Y )), the kernels k N = m ∈
C0 (Rd ; L (X, Y )) are well defined, and we can estimates
Z
k(k N (t − s) − k N (t))xkY dt
|t|>3|s|
XZ
6 k(kj (t − s) − kj (t))xkY dt
j6N {B
 X
d
6 Φp,Y 2d M kxk 8 · 2d 2j+1 r
j:8·2d 2j+1 r62−d−5
X
+ 1
j:2−d−3 62j+1 r6(8d2 ϕ p,Y )p 0
X 8d2 ϕp,Y 
+
(2j+1 r)1/p0
2j+1 r>(8d2 ϕp,Y )p0

p0 1 
6 Φdp,Y 2d M kxk 2 + (log+ 2
2 (8d ϕp,Y ) + d + 4) + −1/p 0
1−2
 
d d + 0
6 Φp,Y 2 M kxk 6 + 3d + log2 ϕp,Y p
6 Φdp,Y 2d M kxk · d · Φp,Y 6 (2Φp,Y )d+1 M kxk.


Proof of (2). We note that (13.31) implies a similar bound for the pointwise
adjoint function m∗ = m(·)∗ ∈ L∞ (Rd ; L (Y ∗ , X ∗ )), while the assumption
that X has Fourier type p ∈ (1, 2] implies that X ∗ has the same Fourier
type with ϕp,X ∗ = ϕp,X (Proposition 2.4.16). Thus case (2) follows from the
already proven case (1) applied to (m∗ , Y ∗ , X ∗ ) in place of (m, X, Y ). 
Corollary 13.2.14. Let X, Y be Banach spaces with non-trivial Fourier type,
let p0 ∈ [1, ∞), and suppose that m ∈ MLp0 (Rd ; X, Y ) satisfies

k∂ α m(ξ)k 6 M |ξ|−|α| ∀α ∈ {0, 1}d .

Then m ∈ MLp (Rd ; X, Y ) for all p ∈ (1, ∞).

Proof. By Lemma 13.2.2, the Littlewood–Paley truncations of m satisfy mN ∈
MLp0 (R d
b ; X, Y ) uniformly in N ∈ Z. By Proposition 13.2.13, the kernels
N
k N = m satisfy both Hörmander and dual Hörmander conditions b uniformly
in N ∈ Z. On the other hand, by Lemma 13.2.11, the kernel kj = mj satisfy
kj (·)x ∈ L1 (Rd ; Y ) for all x ∈ X, uniformly in kxk 6 1, and hence kj ∈
L1so (Rd ; L (X, Y )).
N
It follows that the kernels kL satisfy both Hörmander and dual Hörmander
conditions uniformly in L, N ∈ Z, and they belong to L1so (Rd ; L (X, Y )) (but
N
in general not uniformly). Thus the convolution with kL defines a bounded
p0 d p0 d
operator from L (R ; X) to L (R ; Y ). So does TmN L
, and hence the identity
 13.3 Necessity of UMD for multiplier theorems 273
N
TmN
L
f = kL ∗ f,
b
initially guaranteed by Proposition 13.2.1 for all f ∈ L1 ∩ L1 (Rd ; X), ex-
tends by continuity and density to all f ∈ Lp0 (Rd ; X). Since the operators are
uniformly bounded on this space, and their kernels satisfy both Hörmander
and dual Hörmander conditions uniformly, it follows from the Calderón–
Zygmund Theorem 11.2.5 that they extend boundedly from Lp (Rd ; X) to
Lp (Rd ; Y ) for all p ∈ (1, ∞), again uniformly in L, N ∈ Z. This is the same
N
as mL ∈ MLp0 (Rd ; X, Y ) uniformly in L, N ∈ Z, which, by Lemma 13.2.2,
implies that m ∈ MLp (Rd ; X, Y ). 
The following corollary is just the operator-valued Mihlin Multiplier Theo-
rem 5.5.10 in the special case of Hilbert spaces (in contrast to general UMD
spaces covered by Theorem 5.5.10); we state it here for the sake of pointing
out the alternative approach to this special case via the Calderón–Zygmund
extrapolation theory developed in this chapter.

Corollary 13.2.15. Let H1 , H2 be Hilbert spaces and suppose that m ∈

L∞ (Rd ; L (H1 , H2 )) satisfies

k∂ α m(ξ)k 6 M |ξ|−|α| ∀α ∈ {0, 1}d .

Then m ∈ MLp (Rd ; H1 , H2 ) for all p ∈ (1, ∞).

Proof. By Plancherel’s theorem in both Hilbert spaces, we have

kTm f kL2 (Rd ;H2 ) = kmfbkL2 (Rd ;H2 ) 6 M kfbkL2 (Rd ;H1 ) = M kf kL2 (Rd ;H1 ) ,

and thus kmkML2 (Rd ;H1 ,H2 ) 6 M . Since both Hi have Fourier type 2, Corollary
13.2.14 applies to give that m ∈ MLp (Rd ; H1 , H2 ) for all p ∈ (1, ∞). 

13.3 Necessity of UMD for multiplier theorems

In the previous sections, we have seen Fourier multiplier theorems of roughly
two types:
1. If we already know the boundedness of such an operator on one Lp0 (Rd ; X),
then this boundedness can be extrapolated to other Lp (Rd ; X) spaces un-
der relatively mild (or even no) assumptions on the space X.
2. If we need to prove the boundedness “from scratch”, then the required
assumptions on X tend to be much stronger, and in particular involve the
UMD property.
Let us also recall from the previous volumes that the need of the UMD prop-
erty is not only imposed by the chosen proof strategies, but by the very nature
of things: for prominent examples of multipliers like −i sgn(ξ) corresponding
274 13 The Fourier transform and multipliers

to the Hilbert transform (Theorem 5.2.10), or |ξ|is corresponding to imaginary

powers of the Laplacian (Corollary 10.5.2), the UMD property is indeed nec-
essary. The goal of this section is to continue this list by yet another class of
Fourier multipliers whose boundedness requires UMD, and thereby close the
circle of implications in a number of useful characterisations of UMD spaces.
We start by discussing the types of multipliers that we are going to consider:

Definition 13.3.1. We say that m is constant in the direction of x ∈ Rd \{0}

if m(tx) = m(x) for all t > 0. We say that m is stably constant in the direction
of x ∈ Rd \ {0} if, in addition, we have

lim m(y + tx) = m(x) ∀y ∈ Rd .

t→∞

Note that if m is stably constant in the direction of x, then for every s > 0,

lim m(y + tsx) = lim m(y + tx) = m(x) = m(sx),

t→∞ t→∞

where the last step follows from the assumption (included in the definition of
stably constant) that m is in particular constant in the direction of x.

Example 13.3.2. Suppose that m ∈ C(Rd \{0}) is homogeneous, m(tx) = m(x)

for all t > 0 and x ∈ Rd \ {0}. Then m is stably constant in every direction.
Indeed
lim m(y + tx) = lim m(t−1 y + x) = m(x)
t→∞ t→∞

simply by the continuity of m at x.

Example 13.3.3. Suppose that m ∈ C 1 (Rd \{0}) satisfies the first order Mihlin
condition |∇m(x)| 6 M |x|−1 for all x ∈ Rd \ {0}. If m is constant in the
direction of some x, then m is stably constant in this direction. Indeed
Z 1
|m(y + tx) − m(x)| = |m(y + tx) − m(tx)| = y · ∇m(ys + tx) ds
0
1
M |y|
Z
M ds
6 |y| 6 ,
0 |ys + tx| t|x| − |y|

and clearly this converges to 0 as t → ∞.

Proposition 13.3.4 (Transference from Td to Trd ). Let

m ∈ C(Rd \ {0}; L (X)),

and suppose that it induces a periodic Fourier multiplier

T := Te(m(j))j∈Zn \{0} ∈ L (Lp0 (Td ; X)).

If Tk is the extension of T to Lp0 (Td ; Lp (T(k−1)d ; X)) (Lp (T0 ; X) := X), then
 13.3 Necessity of UMD for multiplier theorems 275
r
X r
X
Tk f k 6 kT kL (Lp0 (Td ;X)) fk
Lp (Trd , dt1 ... dtr ;X) Lp (Trd , dt1 ... dtr ;X)
k=1 k=1

for all fk = fk (t1 , . . . , tk ) ∈ Lp0 (Td , dtk ; Lp (T(k−1)d , dt1 . . . dtk−1 ; X)) that
have non-zero Fourier coefficients with respect to tk only in the directions
where m is stably constant.

Proof. By the density of trigonometric polynomials in Lp , we may assume

that
(k)
X X
fk (t1 , . . . , tk−1 , tk ) = fk (t̄k−1 , tk ) = aj,` ej (t̄k−1 )e` (tk ),
`∈Zn j∈Z(k−1)n
0<|`|6B |j|6B

where
t̄k−1 = (t1 , . . . , tk−1 ) ∈ (Td )k−1 , tk ∈ Td ,
ej (t̄k−1 ) := exp(2πij · t̄k−1 ), e` (tk ) := exp(2πi` · tk ),
and we may choose the same B for all the fk , since there are only finitely many
of them. Then Tk fk has a similar expansion with the (j, `) term multiplied by
m(`).
Let us fix some t̄k := (t̄k−1 , tk ) = (t1 , . . . , tk ) ∈ Tkd for the moment, and

N̄k := (N1 , . . . , Nk−1 , Nk ) = (N̄k−1 , Nk ) ∈ Zk+

to be chosen below.
We will shortly define an auxiliary function of the new variable t ∈ Td . For
this we need to introduce a couple of product-like operations between vectors
of different lengths. We set

N̄k ⊗ t := (N̄k−1 ⊗ t, Nk t) = (N1 t, . . . , Nk t) ∈ (Td )k , N̄k ∈ Zk ,

N̄k−1 j := N1 j1 + . . . + Nk−1 jk−1 ∈ Zd , j = (j1 , . . . , jk−1 ) ∈ (Zd )k−1 .

These operations satisfy the identity

j · (N̄k−1 ⊗ t) = (N̄k−1 j) · t, hence ej (N̄k−1 ⊗ t) = eN̄k−1 j (t),

where · stands for the usual Euclidean scalar product.

The new function is then defined by

f˜k (t) := fk (t̄k + N̄k ⊗ t)

(k)
X X
= aj,` ej (t̄k−1 )e` (tk )eN̄k−1 j+Nk ` (t), (13.32)
`∈Zn j∈Z(k−1)n
0<|`|6B |j|6B

n
k fk : T → X is defined analogously.
The function T]
276 13 The Fourier transform and multipliers

We now want to compare T] ˜

k fk with T fk . They are both multiplier trans-
˜
forms of fk , where in the first one the exponential eN̄k−1 j+Nk ` is multiplied
by m(N̄k−1 j + Nk `), and in the second one by m(`).
By the assumption on fk , we know that m is stably constant in the direc-
(k)
tion of ` whenever aj,` 6= 0, and therefore

lim m(N̄k−1 j + Nk `) = m(`).

Nk →∞

Hence, assuming that N̄k−1 was already chosen, and recalling that j ∈ Z(k−1)d
and ` ∈ Zd with |j|, |`| 6 B take only finitely many different values, we can
choose Nk large enough so that

|m(N̄k−1 j + Nk `) − m(`)| 6 ε

for any preassigned ε > 0 and all relevant values of j and `.

In conclusion, denoting by kgkA the sum of the norms of the Fourier co-
efficients of a trigonometric polynomial g (on a torus of any dimension), we
have
kT] ˜ ˜ ˜
k fk − T fk kp 6 kTk fk − T fk kA 6 εkfk kA 6 εkfk kA .
]
Of course the k · kA norms are finite since the functions above are all trigono-
metric polynomials.
Summing up, it follows that
r
X r
X r
X
T]
k fk 6 T f˜k +ε kfk kA . (13.33)
p p
k=1 k=1 k=1

Here the Lp norms are taken with respect to the variable t ∈ Td , and we
recall that the variables t1 , . . . , tr ∈ Td were kept fixed until now. We now
take the Lp norms of (13.33) with respect to t̄r = (t1 , . . . , tr ) ∈ Trd and use
the triangle inequality to get
Z Z r
X p 1/p
(Tk fk )(t̄r + N̄r ⊗ t) dt dt̄r
Trn Tn X
k=1
Z Z r
X p 1/p
6 kT kL (L0p (Tn ;X)) fk (t̄r + N̄r ⊗ t) dt dt̄r
Trn Tn X
k=1
r
X
+ε kfk kA .
k=1

Exchanging the order of the integrations on Trd and Td , we find by translation

invariance that the dependence on t and N̄r disappears and we are left with
r
X r
X r
X
Tk f k 6 kT kL (Lp0 (Td ;X)) fk +ε kfk kA .
Lp
0 (T
rd ;X) Lp
0 (T
rd ;X)
k=1 k=1 k=1
 13.3 Necessity of UMD for multiplier theorems 277

Since there is no more explicit N̄r dependence, we may take ε → 0, and this
gives the assertion. 

Theorem 13.3.5 (Geiss–Montgomery-Smith–Saksman). Let d > 2 and

m ∈ C(Rd \ {0}) be a multiplier that is stably constant in the directions of
four vectors ±ui , i = 1, 2, where moreover

m(−u1 ) = m(u1 ) 6= m(u2 ) = m(−u2 ).

If m ∈ MLp (Rd ; X), then X is a UMD space and

R
2kmkMLp (Rd ;X) .
βp,X 6 (13.34)
|m(u1 ) − m(u2 )|

To streamline the proof, we recall a transference result that we already ob-

served and used in the proof of Corollary 10.5.2:

Lemma 13.3.6. If m ∈ C(Rd \ {0}) ∩ MLp (Rd ; X), then (m(k))k∈Zd \{0} ∈
MLp0 (Td ; X) and

k(m(k))k∈Zd \{0} kMLp0 (Td ;X) 6 kmkMLp (Rd ;X) .

Proof. This is a slight variant of Proposition 5.7.1, which says that if every
k ∈ Zd is a Lebesgue point of m ∈ L∞ (Rd ), then (m(k))k∈Zd is a Fourier
multiplier on Lp (Td ; X) of at most the norm of the Fourier multiplier m
on Lp (Rd ; X). A slight obstacle is that 0 may fail to be a Lebesgue point
of our m(ξ), no matter how we define m(0). But, if we only consider the
action of these operators on Lp0 (Td ; X), the 0th frequency never shows up,
and one can check that the proof of Proposition 5.7.1 also applies, with trivial
modifications, to the case that each k ∈ Zd \ {0} is a Lebesgue point, giving
exactly what we claimed. 
Proof of Theorem 13.3.5. We begin by essentially the same reduction as in
the proofs of both Theorems 5.2.10 and 10.5.1 (the necessity of UMD for the
boundedness of the Hilbert transform and the imaginary powers of the Lapla-
cian, respectively); but we repeat this short step for the reader’s convenience:
By Theorem 4.2.5 it suffices to estimate the dyadic UMD constant. In order to
most conveniently connect this with Fourier analysis, we choose a model of the
Rademacher system (rk )nk=1 , where the probability space is Tdn = Td1 ×· · ·×Tdn
(each Tdk is simply an indexed copy of Td ), and rk = rk (tk ) is a function of the
kth coordinate tk ∈ Tdk only. Moreover, we are free to choose any instance of
such function, as long as it takes both values ±1 on subsets of Td of measure
1
2 . Then it is sufficient to prove that

n
X n
X
 k fk 6K fk ,
Lp (Tdn ;X) Lp (Tdn ;X)
k=1 k=1
278 13 The Fourier transform and multipliers

where K is the constant on the right of (13.34), for all signs k = ±1, for
all fk of the form fk = φk (r1 , . . . , rk−1 )rk ; these are precisely the martingale
differences of Paley–Walsh martingales (see Proposition 3.1.10). We use the
convention that Lp (T0 ; X) := X.
Let us then observe that, with suitable choice of the invertible matrices
Aj , j = 1, 2, the multipliers mj (ξ) = m(Aj ξ) (of the same multiplier norm
as the original m) are stably constant in the directions of ±ek , k = 1, 2, and
moreover mj (±ek ) = m(u1 ) if j = k and mj (±ek ) = m(u2 ) if j 6= k. Defining
yet another multiplier m0 = 21 (m1 − m2 ) (of at most the same multiplier
norm as m), we find that m0 is also stably constant in the directions of ±ek ,
k = 1, 2, and moreover m0 (±e1 ) = 21 (m(u1 ) − m(u2 )) =: a and m0 (e2 ) = −a.
If we can prove the claim with m0 , e1 , e2 in place of the original m, u1 , u2 , then
the original claim also follows from

2km0 kMLp (Rd ;X) 2kmkMLp (Rd ;X) .

βp,X 6 6
|m0 (e1 ) − m0 (e2 )| |m(u1 ) − m(u2 )|

Dropping the primes, we assume without loss of generality that m(±e1 ) =

a = −m(±e2 ), and m is stably constant in the directions of ±ej , j = 1, 2.
From Proposition 13.3.4 and Lemma 13.3.6 we know that, for suitable
functions fk ,
n
X n
X
Tek fk 6 kmkMLp0 (Td ;X) fk
Lp (Tdn ;X) Lp (Tdn ;X)
k=1 k=1
Xn
6 kmkMLp (Rd ;X) fk ,
Lp (Tdn ;X)
k=1

where Tk is a copy Te(m(j))j∈Zd \{0} acting in the kth Tdk , thus

Tk fk = φk (r1 , . . . , rk−1 )Te(m(j))j∈Zd \{0} rk .

The required condition on fk above is that its Fourier coefficients with respect
to the variable tk should be non-zero only in the directions, where m is stably
constant, i.e., only in the directions ±e1 and ±e2 . Given the product form of
fk , this means more simply that rk should have non-zero Fourier coefficients
only in these directions, which holds in particular if rk is a function of only
the first or only the second coordinate. Note that this gives still (more than)
enough flexibility to make rk equidistributed with a Rademacher variable.
Now, given a sequence (εk )rk=1 , we choose rk to be a function of the first
coordinate if εk = +1, and of the second coordinate if εk = −1. It then follows
that in either case Te(m(j))j∈Zd \{0} rk = aεk rk , and we conclude that
 13.3 Necessity of UMD for multiplier theorems 279
n n
X 1 X
 k fk = Tk f k
Lp (Tdn ;X) |a| Lp (Tdn ;X)
k=1 k=1
n
2 X
6 kmkMLp (Rd ;X) fk ,
|m(e1 ) − m(e2 )| Lp (Tdn ;X)
k=1

which is what we claimed. 

For the sake of precise quantitative conclusions, we also record the follow-
ing variant of Theorem 13.3.5. The assumptions of the next result are much
stronger than those of Theorem 13.3.5, so that the qualitative conclusion that
X is a UMD space is immediate from the previous theorem. The point of this
variant is that under the stronger assumption we can directly estimate the
C
complex UMD constant βp,X of X. The result is not strictly a corollary of
Theorem 13.3.5 itself, but follows by a modification of its proof, as we are
about to see.

Corollary 13.3.7. Let d > 2 and m ∈ C(Rd \ {0}) be an even, homogeneous

multiplier whose range contains the complex unit circle. If m ∈ MLp (Rd ; X),
then X is a UMD space and
C
βp,X 6 kmkMLp (Rd ;X) .

Proof. By the same reductions and notation as in the proof of Theorem 13.3.5,
we now need to check that
n
X n
X
σk f k 6 kmkMLp (Rd ;X) fk ,
Lp (Tdn ;X) Lp (Tdn ;X)
k=1 k=1

for any σk ∈ C with |σk | = 1. By the assumption about the range of m, we

can further write σk = m(uk ) for some uk ∈ C with |uk | = 1.
Consider a large number R > 0. For each k, we can find an integer vector
nk ∈ Zd such that knk −Ruk k`∞ 6 21 . Thus kuk −R−1 nk k`∞ 6 2R1
. Since m is
continuous, by choosing R large enough we ensure that |m(uk )−m(R−1 nk )| 6
δ for each k = 1, . . . , n and any given δ > 0. Thus
n
X n
X
σk fk = m(uk )fk
Lp (Tdn ;X) Lp (Tdn ;X)
k=1 k=1
Xn n
X
6 m(nk )fk + δkfk kLp (Tdn ;X) ,
Lp (Tdn ;X)
k=1 k=1

where we also used the homogeneity m(R−1 nk ) = m(nk ).

We now come to our choice of the Rademachers functions rk appearing in
the martingale differences fk = φk (r1 , . . . , rk−1 )rk . Fixing any Rademacher
function r on T, we take rk (t) := r(nk · t) for t ∈ Td . Substituting nk · t into
the Fourier series of r, we find that
280 13 The Fourier transform and multipliers
X
rk (t) = rb(j)e2πijnk ·t
j∈Z

has a Fourier series involving only frequencies that are multiples of the vector
nk . By the homogeneity of m again, this means that

Te(m(j))j∈Zd \{0} rk = m(nk )rk ,

and thus
n
X n
X
m(nk )fk = Te(m(j))j∈Zd \{0} fk
Lp (Tdn ;X) Lp (Tdn ;X)
k=1 k=1
n
X
6 kmkMLp (Rd ;X) fk .
Lp (Tdn ;X)
k=1

Collecting the estimates, we have checked that

n
X n
X
σk fk 6 kmkMLp (Rd ;X) fk
Lp (Tdn ;X) Lp (Tdn ;X)
k=1 k=1
n
X
+δ kfk kLp (Tdn ;X) ,
k=1

or in other words
n
X
σk rk φk (r1 , . . . , rk−1 )
Lp (Tdn ;X)
k=1
n
X
6 kmkMLp (Rd ;X) rk φk (r1 , . . . , rk−1 )
Lp (Tdn ;X)
k=1
n
X
+δ krk φk (r1 , . . . , rk−1 )kLp (Tdn ;X) .
k=1

While the specific choice of the Rademacher functions rk depended on the

numbers nk , which in turn depended on δ, it is clear that this last bound
is true for any Rademacher sequence (rk )nk=1 , as soon as it is true for one.
Once this observation is made, we see that everything is independent of δ,
and taking the limit δ → 0, we obtain the required bound. 

Corollary 13.3.8. Let X be a Banach space, d > 2 and p ∈ (1, ∞). If any of
the following operators is bounded on Lp (Rd ; X), then X is a UMD space:
(1) a second-order Riesz transform Rj Rk , 1 6 j, k 6 d,
(2) their non-zero difference Rj2 − Rk2 , 1 6 j 6= k 6 d,
(3) the Beurling transform B = (R22 − R12 ) + i2R1 R2 (d = 2).
 13.3 Necessity of UMD for multiplier theorems 281

Moreover, we have the following estimates:

R
(1) βp,X 6 2kRj Rk kL (Lp (Rd ;X)) ,
(2) βp,X 6 kRj2 − Rk2 kL (Lp (Rd ;X)) ,
R
C
(3) βp,X 6 kBkL (Lp (R2 ;X)) .

Proof. These operators correspond to the multipliers

ξj ξk ξj2 − ξk2 ξ1 − iξ2

mRj Rk (ξ) = − , mRj2 −Rk2 (ξ) = − , mB (ξ) = − ,
|ξ|2 |ξ|2 ξ1 + iξ2

each of which is even and homogeneous, in particular stably constant in all

directions.
Writing ξ1 + iξ2 in the polar coordinates as reiθ , it is clear that mB (ξ) =
mB (reiθ ) = −e−i2θ takes all values in the complex unit circle. Hence the
claims concerning B are immediate from Corollary 13.3.7.
For Rj Rk , we observe that mRj2 (ξ) = −ξj2 /|ξ|2 is −1 for ξ = ej and 0 for
ξ = ek , k 6= j, whereas mRj Rk (ξ) = − 21 for ξ = (ej +ek ) and 21 for ξ = (ej −ek )
when k 6= j; in each case we have |m(u1 ) − m(u2 )| = 1 for suitable vectors
ui . For Rj2 − Rk2 , the multiplier is −1 for ξ = ej and +1 for ξ = ek , so that
|m(ej ) − m(ek )| = 2. In each case, the claimed conclusion is immediate from
Theorem 13.3.5. 
Corollary 13.3.8 allows us to complete a characterisation of a function space
embedding that we studied in Section 5.6:

Corollary 13.3.9. Let X be a Banach space, let d, k > 1 and p ∈ (1, ∞).
Then there is a constant C such that

kf kW k,p (Rd ;X) 6 kf kH k,p (Rd ;X) ∀f ∈ S (Rd ; X)

if and only if at least one of the following holds:

(1) d = 1 and k is even, or
(2) X is a UMD space.

Proof. The sufficiency of (1) has been established in Proposition 5.6.10 and
the sufficiency of (2) in Theorem 5.6.11. Moreover, in Theorem 5.6.12, it has
been shown that the UMD property is necessary when k is odd, and that the
boundedness of the second-order Riesz transform R12 is necessary when k is
even and d > 2. By Corollary 13.3.8, the UMD property follows from this,
and hence it is necessary in all cases except (1). 
In our final corollary to Theorem 13.3.5, we dispense with the evenness con-
dition.
282 13 The Fourier transform and multipliers

Corollary 13.3.10. Let d > 1 and m ∈ C(Rd \ {0}) be any positively homo-
geneous multiplier (i.e., m(λξ) = m(ξ) for all ξ ∈ Rd \ {0} and λ > 0) that is
not identically constant. If m ∈ MLp (Rd ; X), then X is a UMD space and

R
4kmkMLp (Rd ;X)
βp,X 6 min ,
∈S d−1
u1 ,u2 |m(u1 ) + m(−u1 ) − m(u2 ) − m(−u2 )|
 2kmkMLp (Rd ;X) 2
R
βp,X 6 (~p,X )2 6 min ,
u∈S d−1 |m(u) − m(−u)|

where at least one of the right-hand sides is finite.

The assumption that m is not identically constant, rather than the perhaps
expected “not identically zero”, is necessary: the Fourier multiplier Tm with
m ≡ c coincides with the scalar multiplication f 7→ c · f , whose boundedness
certainly needs no UMD.
Proof. As pointed out right before Proposition 5.3.7, the assumption that
m ∈ MLp (Rd ; X) implies the same property for the reflected function m(ξ)e :=
m(−ξ). Then, by the triangle inequality, the even and odd parts meven :=
1
e and modd := 21 (m − m)
2 (m + m) e are also positively homogeneous multipliers
of at most the same multiplier norm as m. Since m is not identically constant,
and m = meven + modd , at least one of meven or modd is not identically
constant.
If meven is not identically constant, there are two directions u1 , u2 ∈ S d−1
such that meven (u1 ) 6= meven (u2 ) and hence, by evenness,

meven (−u1 ) = meven (u1 ) 6= meven (u2 ) = meven (−u2 ).

By Example 13.3.2, the homogeneous meven ∈ C(Rd \ {0}) is stably con-

stant in every directions. Hence meven satisfies the assumptions of the Geiss–
Montgomery-Smith–Saksman Theorem 13.3.5, and the said theorem guaran-
tees that, for any such u1 , u2 ∈ S d−1 ,

R
2kmeven kMLp (Rd ;X)
βp,X 6
meven (u1 ) − meven (u2 )
4kmkMLp (Rd ;X)
6 .
m(u1 ) + m(−u1 ) − m(u2 ) − m(−u2 )

(Note that the condition that meven (u1 ) 6= meven (u2 ) is precisely the require-
ment that the denominator is non-zero, and hence can extend the previous
display to all pairs of u1 , u2 ∈ S d−1 ; interpreting 1/0 = ∞, as usual, this only
amounts to adding the triviality βp,X R
6 ∞.)
For the odd part modd , being not identically constant is equivalent to being
not identically zero. If this is the case, there is some direction u ∈ S d−1 such
that m(−u) = −m(u) 6= 0. Writing ξ ∈ Rd as ξ = (ξ · u)u + [ξ − (ξ · u)u], we
consider the invertible linear transformations Aλ ξ = (ξ · u)u + λ[ξ − (ξ · u)u],
 13.4 Notes 283

where λ > 0. By Proposition 5.3.8, each modd ◦ Aλ has the same multiplier
norm as modd . As λ → 0, it is clear that Aλ ξ → (ξ · u)u for all ξ ∈ Rd and
thus, by the continuity of m and hence modd ,

modd ◦ Aλ (ξ) → modd ((ξ · u)u) = sgn(ξ · u)modd (u).

A convergence result for multipliers, Proposition 5.3.16, then implies that

|modd (u)|kξ 7→ sgn(ξ · u)kMLp (Rd ;X) 6 lim inf kmodd ◦ Aλ kMLp (Rd ;X)
λ→0
= kmodd kMLp (Rd ;X) .

By another application of Proposition 5.3.8 with a rotation that sends u to

e1 , it follows that
kξ 7→ sgn(ξ1 )kMLp (Rd ;X) = kξ 7→ sgn(ξ · u)kMLp (Rd ;X)
kmodd kMLp (Rd ;X) 2kmkMLp (Rd ;X)
6 6 .
|modd (u)| |m(u) − m(−u)|

(The bound remains valid for all u ∈ S d−1 , reducing to a triviality if m(u) =
m(−u).) By Fubini’s theorem, we find that

~p,X := kξ 7→ sgn(ξ)kMLp (R;X) kξ 7→ sgn(ξ1 )kMLp (R;X) .

R
The bound between βp,X 6 (~p,X )2 is contained in Corollary 5.2.11. 

13.4 Notes
Section 13.1

The precise quantitative form of the final bound in the comparison of various
Fourier-type constants in Proposition 13.1.1 seems to be new; we were not
aware of this estimate at the time of completing Volume II, where a weaker
0
version was given. The identity ϕp,C (Rd ) = (p1/p (p0 )−1/p )d mentioned below
the said proposition is due to Babenko [1961] in the special case that p0 is an
even integer, and due to Beckner [1975] in full generality.
The main result of this section, Theorem 13.1.33 is from Bourgain [1988a],
with preliminary versions going back to Bourgain [1981, 1982]. The main
theorem of Bourgain [1982] reads as follows: If X is a B-convex Banach space
(which is equivalent to non-trivial type by Proposition 7.6.8), then there are
u, v ∈ (1, ∞) and δ, M ∈ (0, ∞) such that
X 1/u X X 1/v
δ kxγ kuX 6 γxγ 6M kxγ kvX , (13.35)
L2 (G;X)
γ∈Γ γ∈Γ γ∈Γ

whenever {xγ }γ∈Γ is a finitely non-zero sequence of elements of X and Γ is

the spectrum of the compact abelian group G. This is a Hausdorff–Young
284 13 The Fourier transform and multipliers

inequality with mismatched exponents; our Corollary 13.1.27 is the special

case of the right-hand inequality with G = T and Γ = {ek }k∈Z . For these
particular G and Γ , and under the stronger assumption that X be super-
reflexive, (13.35) was proved in Bourgain [1981]. A further predecessor of
such results is due to James [1972], who proved a bound like (13.35) with a
super-reflexive Z in place of both X and L2 (G; X), and zk ∈ Z in place of
∞
both xγ and γxγ , under the assumption that (zk )k=1 is a basic sequence in Z,
i.e.,
XK XL
a k zk 6C a k zk (13.36)
Z Z
k=1 k=1

for all scalars ak and integers K 6 L. Requiring (13.36) for zk = ek xk ∈

Z = L2 (T; X), uniformly in xk ∈ X, is equivalent to the still stronger prop-
erty that X be a UMD space, which is why additional work was required by
Bourgain [1981] to obtain his result for trigonometric series in super-reflexive
spaces. (The estimate (13.36) in the said special case is equivalent to the
L2 (T; X)-boundedness of the periodic Hilbert transform by Proposition 5.2.7,
and this is equivalent to the UMD property by Corollary 5.2.11. UMD spaces
are super-reflexive by Corollary 4.3.8, but the converse is false. Various exam-
ples showing the last point are due to Pisier [1975], Bourgain [1983], Garling
[1990], Geiss [1999], and Qiu [2012]. The example of Qiu [2012] is an infinitely
iterated Lp (Lq ) space, which has been presented in Theorem 4.3.17, but the
super-reflexivity of this space is not treated there.)
As in our treatment in the section under discussion, getting from estimate
(13.35) with mismatched exponents to dual pairs requires further ideas. This
was achieved by Bourgain [1988b], who proved that, for some u1 , v1 ∈ (1, ∞)
and δ1 , M1 ∈ (0, ∞), there further holds
1/u10
u0
X X
δ1 kxγ kX1 6 γxγ
Lu1 (G;X)
γ∈Γ γ∈Γ
X X 1/v1 (13.37)
6 γxγ 0 6 M1 kxγ kvX1 ,
Lv1 (G;X)
γ∈Γ γ∈Γ

when G is either T or the Cantor group {−1, 1}N . For G = T, the leftmost
and rightmost estimates correspond, in our notation, to ϕu1 ,X (T) 6 1/δ1 and
ϕv1 ,X (Z) 6 M1 , respectively. The easy estimate ϕp,X (R) 6 ϕp,X (T) was also
observed by Bourgain [1988b]. In contrast to the case of T, a scaling argument
(substituting f (λ·) in place of f and considering the limit λ → 0 or λ → ∞)
shows that an estimate of the from kfbkLq (R;X) 6 Ckf kLp (R;X) can only hold
for q 0 = p; thus, in order to deduce any Hausdorff–Young inequality on R at
all, the additional steps from the mismatched exponents of Bourgain [1982]
to the dual exponents of Bourgain [1988b] seem to be necessary.
The second half of the argument leading to Bourgain’s Theorem 13.1.33, as
presented in Sections 13.1.c and 13.1.d, is close to the treatment of Bourgain
 13.4 Notes 285

[1988b], although we have also benefited from the exposition of these steps by
Pietsch and Wenzel [1998]. On the other hand, the first half of our treatment,
in Sections 13.1.a and 13.1.b, is also based on Pietsch and Wenzel [1998] but
deviates from the original approach of Bourgain [1982]. The beginning of the
argument, leading to Proposition 13.1.11 on “breaking the trivial bound” is
due to Hinrichs [1996], but it also uses a result of Bourgain [1985], Proposition
13.1.7, on the Sidon property of quasi-independent sets.
We have chosen this approach of Hinrichs [1996] and Pietsch and Wenzel
[1998] due to an independent interest, in our opinion, of some of its intermedi-
ate steps, despite the fact that the original argument of Bourgain [1982, 1988b]
seems slightly more efficient in terms of the final quantitative conclusions. In
any case, the main result says that every Banach space of type p ∈ (1, 2] will
0
have Fourier-type r = 1 + (cτp,X;2 )−p , for some absolute constant c. (The ad-
ditional factor 6p0 in our formulation of Theorem 13.1.33 could obviously be
absorbed by choosing a larger constant c.) The difference is in the numerical
value of c, which is 68 in our formulation (up to the lower order factor just
mentioned) and 17 in Bourgain [1982, 1988b].
In our approach, this√ constant comes from the proof of Corollary 13.1.20,
where the estimate 48 2 (≈ 67.88) 6 68 is made. (Since we are clearly off
Bourgain’s constant at this point already,
√ it would seem pointless to insist in
the decimals here.) The constant 48 2, in turn, is produced as
√ √
48 2 = 16 · 2 · 3, where

(i) 16 is the upper bound of the Sidon constants of quasi-independent sets

from
√ Proposition 13.1.7;
(ii) 2 comes from the factor in front of the upper bound of the number of
quasi-independent sets required to partition a given set in Lemma 13.1.9;
the root is due to the use of this number count after an application of
the Cauchy–Schwarz inequality in the proof of Proposition 13.1.11;
(iii) 3 is the constant from the Marcinkiewicz inequality (Proposition 13.1.17),
which enters into the estimate through an application of the Comparison
Lemma 13.1.18 in the proof of Corollary 13.1.20.
One may speculate that the constant 16 (just below the 17 of Bourgain [1982])
is the heart of the matter, and the other two factors are only produced by
secondary details that should be avoidable by more careful reasoning.
The approach of Bourgain [1982] is based on two abstract results (avoided
in the present treatment) about the collection of tuples of functions
Z Z
O := {ξ = (ξi )ni=1 ∈ L2 (Ω; R)n : kξi k∞ 6 1, ξi = ξi ξj = 0

for all 1 6 i 6= j 6 n}

on a probability space Ω; namely:

(1) The set E of extreme points of O consists of tuples of ±1-valued functions.
286 13 The Fourier transform and multipliers

(2) For each ξ ∈ O, there is a Borel probability measure µ on E such that

Z
ξi = ηi dµ(η), for every i = 1, . . . , n.
E

According to Bourgain [1982], the proof of (1) is “essentially contained in”

Dor [1975], while (2) can be derived from a generalisation of Choquet’s inte-
gral representation theorem due to Edgar [1976]. Combining these abstract
tools with delicate hard analysis, Bourgain [1982] eventually arrives at his key
technical estimate, which in our notation (and exchanging the roles of X and
X ∗ compared to Bourgain [1982]) may be stated as
(2) 0
ϕ∞,X ∗ (F ) 6 K · N 1/t , t0 = (17 · τp,X;2 )p . (13.38)

This is recognised as a close relative of Corollary 13.1.29, where the bound

(s0 ) 0
ϕ∞,X ∗ (F ) 6 K · N 1/s , s0 > r0 = 3p0 (68 · τp,X;2 )p

is obtained. While the left-hand sides are not identical, (13.38) allows Bourgain
[1982] to deduce the Hausdorff–Young inequality with mismatched exponents
as in (13.35) (with X ∗ in place of X) for any v ∈ (1, t), and finally, in Bourgain
[1988b], also the classical Hausdorff–Young inequality (13.37) (again with X ∗
in place of X) with any u1 ∈ (1, v). Since v ∈ (1, t) is arbitrary, one can reach
any u1 ∈ (1, t), and thus in particular the r determined by
0
r0 = (18 · τp,X;2 )p (13.39)

is a Fourier type of X ∗ , and hence of X.

Remark 13.4.1 (A typo in the statement of Bourgain’s theorem in König

[1991]). It seems to be claimed by König [1991] that every space of type
p0
p > 1 would have Fourier-type r with r0 = c · τp,X;2 and c = 18 (forgetting
brackets from (13.39)). As written, this is absurd for any absolute constant c:
It is straightforward to verify that, for every p ∈ (1, 2], the space X = `p
has type p with constant τp,X;2 = 1:
N
X N
X
εn x n 6 εn x n = {xn }N
n=1 2 )
L2 (Ω;`p ) `p (L2 (Ω)) `p (`N
n=1 n=1
N
X 1/p
N p
6 {xn }n=1 p
= kxn kX
`p (`N )
n=1

Thus, were the claim in the beginning of the remark true, all these spaces
c
would have the Fourier-type r = c−1 > 1, which is impossible for p ∈ (1, r)
by Example 2.1.15.
 13.4 Notes 287

Since the numerical constant in (13.38) may be affected by an equivalent

choice of the type constant, we note that Bourgain [1982] is not explicit about
the precise definition of the constant that he denotes by C, but one can see
in the proof of the first step of his Proposition 4 that C = τp,X ∗ ;2 ; recall that
we exchanged the roles of X and X ∗ compared to Bourgain [1982].
More details on quasi-independent sets can be found in the monograph of
Graham and Hare [2013]. Sometimes quasi-independent sets are called disso-
ciate sets, but it seems that in more recent works this terminology is reserved
for the slightly stronger property where one allows αk ∈ {−2, −1, 0, 1, 2} in
Definition 13.1.5. In particular,
√ one can find there that quasi-independent are
Sidon sets with constant 6 6 ≈ 14.70, which is slightly better than the con-
stant 16 in Proposition 13.1.7. The converse bounds of Remark 13.1.10 have
been shown to us by Dion Gijswijt. If one replaces the group Z by another
group it was shown on page 203 in Pietsch and Wenzel [1998] that the bound
of Lemma 13.1.8 is sharp.
The result of Proposition 13.1.21 states that type p and cotype q with
1/p − 1/q < 1/r − 1/2 with r ∈ (1, 2) implies Fourier type r. In the limiting
case of equality it is unknown what happens. However, the result is sharp
in the sense that for every r ∈ (1, 2) and for every p ∈ (r, 2) there exists
a Banach space X such that X has type p, cotype q, and Fourier type r
with p1 − 1q = 1r − 12 , and none of the exponents (p, q, r) can be improved
(see Bourgain [1988a] and Garcı́a-Cuerva, Torrea, and Kazarian [1996]). This
example was also used to show that the dependence on the type constant is
necessary in Theorem 13.1.33. The following improvement was observed in
Garcı́a-Cuerva, Torrea, and Kazarian [1996] for Banach lattices X:

sup{p ∈ (1, 2] :X has Fourier type p}

= sup{p ∈ (1, 2] : X has type p and cotype p0 }.

Section 13.2

In the scalar-valued case, considerations of the kind that we have presented in

this section go back to Hörmander [1960] who used similar methods to rederive
(a variant of) the multiplier theorem of Mihlin [1956, 1957] by transforming
it into a form where the theory of Calderón and Zygmund [1952] could be
applied. The methods of Hörmander [1960] are already very close to the ones
in the Section 13.2.b, the key difference being that he can make use of the
Plancherel theorem to pass between L2 estimates in the space and frequency
variables. For functions taking values in a general Banach space, the only
available substitute is the elementary L1 (Rd ; X)-to-L∞ (Rd ; X) boundedness
of the Fourier transform. This still allows essentially similar conclusions, at
the cost of requiring estimates for a higher number of derivatives as input.
On the other hand, as soon as we start imposing such stronger assumptions,
we can also obtain stronger conclusions, namely, standard Calderón–Zygmund
kernels rather than just Hörmander kernels, as in Section 13.2.a. Scalar-valued
288 13 The Fourier transform and multipliers

versions of such results are again well known; for example, a version of Propo-
sition 13.2.7 with d + 2 derivatives (instead of d + 1 in the said proposition)
appears in the book of Stein [1993]. Under this stronger assumption, Stein
[1993] deduces that k ∈ C 1 (Rd \ {0}), while Proposition 13.2.6 gives the
slightly weaker conclusion that k is just barely below Lipschitz, with a modu-
lus of continuity ω(t) = O(t · log(1 + 1/t)). This is still quite enough to derive
like Corollaries 13.2.8, 13.2.9, and 13.2.10 on the boundedness of Fourier mul-
tipliers on weighted Lp (w; X) spaces. Using the result from Stein [1993] in
place of Proposition 13.2.7, a version of Corollary 13.2.10 assuming d + 2
derivatives was formulated by Meyries and Veraar [2015]. In principle, vari-
ants of Propositions 13.2.6 and 13.2.7 sufficient for Corollaries 13.2.8 through
13.2.10 would only require smoothness of order d + ε, but such statements and
proofs are bound to have additional technicalities due to the very formulation
of fractional order smoothness conditions. Various results in this direction, in-
volving kernel bounds for Fourier multipliers with close-to-critical fractional
smoothness, were explored by Hytönen [2004].
To get rid of the ε > 0 altogether, i.e., to deduce useful b (in view of
Calderón–Zygmund extrapolation) kernel estimates for k = m from just d
derivatives of m, one needs to impose assumptions on the Fourier-type of the
underlying spaces. While we have only dealt with the sufficiency of the Fourier-
type assumption in Section 13.2.b, an early result involving both directions,
in dimension d = 1, is the following:
Theorem 13.4.2 (König [1991]). A Banach space X is K-convex if and
only if every f ∈ C 1 (T, X) has Fourier coefficients (fb(n))n∈Z ∈ `1 (Z; X).
Recall that K-convexity is equivalent to non-trivial type by Pisier’s Theo-
rem 7.4.23, and non-trivial type is equivalent to non-trivial Fourier-type by
Bourgain’s Theorem 13.1.33. The proof of “⇒” in Theorem 13.4.2 is then
straightforward from non-trivial Fourier type. For the converse, König [1991]
starts with a concrete counterexample when X = L1 (T), and approximates
this finite versions that can be represented in `1N , with blow-up in the limit
N → ∞. By the Maurey–Pisier Theorem 7.3.8, if X does not have non-trivial
type, then it contains subspaces isomorphic to `1N uniformly, and hence the
said finite examples can also be represented in X. Finally, the closed graph
theorem guarantees that a sequence of examples with blow-up also guarantees
the existence of a single f ∈ C 1 (T, X) with (fb(n))n∈Z ∈
/ `1 (Z; X).
In our formulation of Proposition 13.2.13, the assumed Fourier-type p ∈
(1, 2] only affects the constant in the estimate. However, by more careful
reasoning, one could show that also the number of the required derivative
∂ α m could be reduced as a function of p; roughly speaking, one needs only
derivatives up to order bd/pc + 1, or more generally fractional smoothness of
order d/p + ε, to obtain the same conclusions. Such results can be found in
Hytönen [2004]. In the more general context of various function spaces, this
phenomenon will be explored further in Chapter 14; see Proposition 14.5.3
and take q = ∞ there.
 13.4 Notes 289

Our focus in the section under discussion has been exploring conditions
that one bneeds to assume on a multiplier m in order that their associated ker-
nel k = m satisfies the assumptions of one of the extrapolation theorems of
Chapter 11 (so that the a priori boundedness of Tm on one Lp0 (Rd ; X) extends
to other spaces), but similar considerations can also be used to reduce the re-
quired smoothness, as a function of the Fourier-type of the underlying spaces,
in results like Mihlin’s Multiplier Theorem 5.5.10 (where the boundedness of
Tm on Lp (Rd ; X) is deduced “from scratch”). Such results were pioneered
by Girardi and Weis [2003b] and further elaborated by Hytönen [2004]. If m
is scalar-valued, it is also possible to replace Fourier-type by quantitatively
weaker assumptions on type or cotype; see Hytönen [2010].

Section 13.3
The main results of this section, notably Proposition 13.3.4, Theorem 13.3.5,
and Corollary 13.3.8, are essentially from Geiss, Montgomery-Smith, and
Saksman [2010], but we have incorporated some improvements, partially in-
spired by unpublished observations of Alex Amenta that he kindly shared
with us.
These results may be seen as successors, in terms of both statement and
proof, of Theorem 5.2.10 of Bourgain [1983] and Theorem 10.5.1 of Guerre-
Delabrière [1991], which deal with the necessity of UMD for the boundedness
of the Hilbert transform and the imaginary powers (−∆)is of the Laplacian,
respectively. However, none of these three results contains any of the other
two.
Certain elaborations of Corollary 13.3.8 are due to Castro and Hytönen
[2016]. Namely, the identity ∂j ∂k u = −Rj Rk ∆u implies that
d
X
k∂j ∂k ukLp (Rd ;X) 6 C kRi2 ukLp (Rd ;X) , (13.40)
i=1

where C 6 kRj Rk kL (Lp (Rd ;X)) , but C could a priori be much smaller. How-
ever, Castro and Hytönen [2016] show that the seemingly weaker inequality
(13.40) still implies the UMD property with the same control
βp,X 6 2C(13.40) (13.41)
as in Corollary 13.3.8 for kRj Rk kL (Lp (Rd ;X)) . More generally, the same paper
proves the necessity of UMD for any member of a family of inequalities of the
form X
k∂ β ukLp (Rd ;X) 6 C k∂ α ukLp (Rd ;X) ,
α∈A
but the relation between the constants is particularly clean in the example
just mentioned.
It could be of interest to identify more general criteria (subsuming previous
related results) for inequalities of classical/harmonic analysis to
290 13 The Fourier transform and multipliers

(1) imply the UMD property of X (as in all mentioned results), or

(2) control the UMD constant βp,X linearly by the constant in the inequality
(as in Theorems 10.5.1 and 13.3.5, but not Theorem 5.2.10).
While we have concentrated, in this section, on lower bounds of multiplier
norms by the UMD constants, Geiss, Montgomery-Smith, and Saksman [2010]
also treat the other direction. In particular, they show that the first two
bounds of Corollary 13.3.8 are actually identities:

k2Rj Rk kL (Lp (Rd ;X)) = kRj2 − Rk2 kL (Lp (Rd ;X)) = βp,X
R
(13.42)

for all 1 6 j 6= k 6 d. The upper bounds for the norms are proved by
representing and estimating the operators by means of stochastic integrals.
Yaroslavstev [2018] obtained further variants of these estimates for related
operators. We plan to detail this in a forthcoming Volume. By (13.41), a
trivial bound, and (13.42), it follows that

βp,X 6 2C(13.40) 6 2kRj Rk kL (Lp (Rd ;X)) = βp,X ,

and hence all these quantities must be equal. In particular, as observed by

Castro and Hytönen [2016], it follows that
1 1
X=R
C(13.40) = βp,R = (max(p, p0 ) − 1),
2 2
using Burkholder’s Theorem 4.5.7 for the last equality. We are not aware of
another method than that of Geiss, Montgomery-Smith, and Saksman [2010]
to determine the exact norms (13.42) or the sharp constant in (13.40), which
highlights the benefits of martingale techniques even for questions of classical
analysis.
In the third case of Corollary 13.3.8 concerning the Beurling–Ahlfors trans-
form, the matching upper bound is an outstanding open problem even for
X = C (see Problems O.1 and O.2).
More generally, Geiss, Montgomery-Smith, and Saksman [2010] prove that
all real, even, and homogeneous (i.e., m(tξ) = m(ξ) ∈ R for all ξ ∈ Rd \ {0}
and t ∈ R \ {0}) multipliers m ∈ C ∞ (Rd \ {0}) satisfy the estimate

kmkMLp (Rd ;X) 6 Cm · βp,X ,

where Cm depends only on m. Note in particular that the estimate is linear in

βp,X , improving on the quadratic estimate provided by T (1) Theorem 12.4.21,
or the still higher order dependence in the Mihlin Multiplier Theorem 5.5.10.
By elaborations of the T (1) technology, linear dependence has also been ob-
tained for a class of even non-convolution operators on Lp (R; X) (but only in
dimension d = 1, as written) by Pott and Stoica [2014], but beyond that the
availability of linear bounds in terms of βp,X remains open. In particular, a
possible linear estimate between βp,X and the norm of the Hilbert transform
~p,X = kHkL (Lp (R;X)) , in either direction, is unknown (see Problem O.6).
 13.4 Notes 291

Certain substitute results related to the latter are due to Domelevo and
Petermichl [2023c,d]. They construct a new dyadic operator and show that
its boundedness is equivalent to that of the Hilbert transform, with linear de-
pendence between the respective norms in both directions. Analogous results
for the Riesz transforms are obtained in Domelevo and Petermichl [2023a,b].
Further estimates between the Hilbert transform (and variants) and de-
coupling constants related to the UMD constant can be found in Osȩkowski
and Yaroslavtsev [2021].
Corollary 13.3.9 characterises situations in which there is a continuous
embedding H k,p (Rd ; X) ,→ W k,p (Rd ; X). Several related results, including
versions on domains O ⊆ Rd , are due to Arendt, Bernhard, and Kreuter
[2020].
14
Function spaces

This chapter presents an in-depth study of several classes of vector-valued

function spaces defined by smoothness conditions. In Volume I we have already
encountered two such classes: the Sobolev spaces W s,p (Rd ; X) for s ∈ N and
s ∈ (0, 1) (Chapter 2) and the Bessel potential spaces H s,p (Rd ; X) for s ∈ R
(Chapter 5). Both classes are parametrised by an integrability parameter p and
smoothness parameter s. The present chapter introduces two related classes of
s
function spaces, the Besov spaces Bp,q (Rd ; X) and the Triebel–Lizorkin spaces
s d
Fp,q (R ; X). From the point of view of applications these spaces play an im-
portant role in the theory of partial differential equations, where they typically
occur as trace spaces associated with initial value problems. What makes these
spaces interesting from a mathematical point of view is the wealth of differ-
ent characterisations of these classes: they can equivalently be introduced via
Littlewood–Paley decompositions, difference norms, and interpolation.
In line with earlier developments, we introduce both Besov spaces and
Triebel–Lizorkin spaces via their Littlewood–Paley decompositions. These
involve a so-called inhomogeneous Littlewood–Paley sequence (ϕk )k>0 of
Schwartz functions on Rd whose Fourier transforms behave, informally speak-
ing, as a dyadic partition of unity radially. In terms of such sequences, the
Besov and Triebel–Lizorkin norms are defined by

kf kBp,q
s (Rd ;X) = (2k ϕk ∗ f )k>0 `q (Lp (Rd ;X))

and
kf kFp,q
s (Rd ;X) := (2ks ϕk ∗ f )k>0 Lp (Rd ;`q (X))
,

in the sense that a tempered distribution f ∈ S 0 (Rd ; X) belongs to either

one space if and only if the respective expression is well defined and finite.
The third parameter q featuring in these definitions is often referred to as the
microscopic parameter.
In both cases, the norms are independent of the Littlewood–Paley se-
quence up to a multiplicative constant independent of f . Accordingly, it will

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 293
T. Hytönen et al., Analysis in Banach Spaces, Ergebnisse der Mathematik und ihrer
Grenzgebiete. 3. Folge / A Series of Modern Surveys in Mathematics 76,
https://doi.org/10.1007/978-3-031-46598-7_4
294 14 Function spaces

be a standing assumption that throughout the chapter we fix a Littlewood–

Paley sequence (ϕk )k>0 once and for all (Convention 14.2.8). Dependence of
constants on this sequence will never be tracked.
Interestingly, the Bessel potential spaces studied in Chapter 5, and whose
study is continued in the present chapter, admit a similar decomposition re-
placing `q -norms by Rademacher norms (Theorem 14.7.5) in case X has UMD:

kf kH s,p (Rd ;X) h (εk 2ks ϕk ∗ f )k>0 εp (Lp (Rd ;X))

= (εk 2ks ϕk ∗ f )k>0 Lp (Rd ;εp (X))

,

using the notation for Rademacher spaces introduced Section 6.3; the equality
of the latter two norms is obtained by repeating the proof of Theorem 9.4.8
for Rademacher sums. Comparing these norms with the previous two, it is
also of interest to note that equivalent norms are obtained if the εp -norm is
replaced by an εq -norm, by the Kahane–Khintchine inequalities.
In view of their very similar definitions, it comes as no surprise that the
theories of Besov and Triebel–Lizorkin spaces largely parallel each other and
resemble the theory of Bessel potential spaces to some extent. There are some
notable differences however, due to the different orders in which the Lp -norm
and `q -norm are taken; as we have already pointed out, the Triebel–Lizorkin
norm is generally speaking more difficult to handle. The main advantage of
the Besov and Triebel–Lizorkin over the Bessel potential spaces is that they
are often easier to work with, and indeed many basic results for these spaces in
the vector-valued setting do not rely on the geometry of the Banach space X.
This is in stark contrast with the theory of Bessel potential spaces, where the
corresponding results often require geometrical properties such as the UMD
property of X or the Radon–Nikodým property of X ∗ , as we have seen in
Chapter 5.
After establishing notation and proving some preliminary results in Sec-
tion 14.1, the class of Besov spaces is introduced in Section 14.4 via their
Littlewood–Paley decompositions. Several basic aspects of these spaces are
discussed, such as their independence of the inhomogeneous Littlewood–Paley
sequence used in the definition, the density of smooth functions, and Sobolev
type embeddings. We continue with several more advanced results, including
a difference norm characterisation, identification the complex and real inter-
polation spaces, and identification of the dual spaces. In Section 14.5 these
results are used to prove embedding theorems for the spaces γ(L2 (Rd ), X)
introduced in Chapter 9 and to prove R-boundedness of the ranges of smooth
operator-valued functions under type and cotype assumptions. In the same
section we discuss Fourier multiplier results for Besov spaces under (co)type
and Fourier type assumptions.
In Section 14.6 the Triebel–Lizorkin spaces are introduced. Proving the
same basic properties as before is more complicated, especially for the impor-
tant endpoint exponent q = 1, and requires the boundedness of the so-called
Peetre maximal function and the boundedness of Fourier multiplier opera-
 14.1 Summary of the main results 295

tors for functions with compact Fourier support in an Lp (Rd ; `q (X))-setting.

Most of the elementary and more advanced results discussed for Besov spaces
have a counterpart for Triebel–Lizorkin spaces and indeed our treatment mir-
rors that of the Besov spaces. Some results, however, have a different flavour,
such as the Sobolev embedding theorem (Theorem 14.6.14), the Gagliardo–
Nirenberg inequalities (Proposition 14.6.15), and the embedding theorem of
Franke and Jawerth (Theorem 14.6.26), all of which have an improvement in
the microscopic parameter q. In some situations this improvement makes it
possible to derive results for general Banach spaces X in an effective way. For
instance, for any Banach space X one has continuous embeddings (here and
below denoted by “,→”)
s
Fp,1 (Rd ; X) ,→ H s,p (Rd ; X) ,→ Fp,∞
s
(Rd ; X) (14.1)
for p ∈ (1, ∞) and s ∈ R. For Hilbert spaces X this can be improved to
H s,p (Rd ; X) = Fp,2
s
(Rd ; X)
with equivalent norms for all p ∈ (1, ∞) and s ∈ R; this identity characterises
Hilbert spaces up to isomorphism (Theorem 14.7.9). The “sandwich result”
(14.1) often makes it possible to prove results about H s,p (Rd ; X) without
conditions on X by factoring through a Triebel–Lizorkin space. At the end
of the section apply some of the obtained result to prove boundedness of
pointwise multiplication by the function 1R+ in Triebel–Lizorkin spaces and
Besov spaces. Such results are non-trivial due to the non-smoothness of 1R+ ,
and are important in applications to interpolation with boundary conditions
of vector-valued function spaces used for evolution equations.
In Section 14.7 we return to the study of Bessel potential spaces and dis-
cuss some basic properties not covered in the earlier volumes. These include
improvements of (14.1) for UMD spaces X under type and cotype assump-
tions, as well as some advanced results on complex interpolation of Bessel
potential spaces (Corollary 14.7.13). At the end of the section we prove the
boundedness of pointwise multiplication by the function 1R+ in Bessel poten-
tial spaces again for UMD spaces.
As we will be using Fourier techniques practically everywhere, it will be
a further standing assumption that throughout the chapter we work over the
complex scalar field. As usually is the case, the case of real Banach spaces
can be treated by standard complexification arguments. In some cases one
can argue directly on real Banach spaces (see Remark 14.2.6). Unless stated
otherwise, X will always denote an arbitrary complex Banach space.

14.1 Summary of the main results

Scattered over this section a wealth of inclusion and interpolation results are
developed. For the convenience of the reader, we include a concise overview
of them here, with pointers to their location.
296 14 Function spaces

In all identities, unless otherwise no geometric restrictions apply to Banach

spaces and the occurring indices are taken in the ranges

p0 , p1 , p ∈ [1, ∞], q0 , q1 , q ∈ [1, ∞], s0 , s1 , s ∈ R, k0 , k1 ∈ N,

or subsets thereof. The interpolation results assume that θ ∈ (0, 1) and where
1 1−θ θ 1 1−θ θ
= + , = +
pθ p0 p1 qθ q0 q1
and

sθ = (1 − θ)s0 + θs1 , kθ = (1 − θ)k0 + θk1 .

The complex and real interpolation spaces of an interpolation couple (X0 , X1 )

of Banach spaces are denoted by

Xθ = [X0 , X1 ]θ , Xθ,p = (X0 , X1 )θ,p

respectively.
Identities. Up to equivalent norms we have the following identifications. If
p ∈ [1, ∞), s ∈ (0, 1), then

W s,p (Rd ; X) = Bp,p

s
(Rd ; X) (Corollary 14.4.25)

and, if s ∈ (0, ∞) \ N,

s
Cub (Rd ; X) = B∞,∞
s
(Rd ; X). (Corollary 14.4.26)

If H is a Hilbert space and p ∈ (1, ∞), s ∈ R, then

H s,p (Rd ; H) = Fp,2

s
(Rd ; H) (Theorem 14.7.9)

and, if p ∈ (1, ∞) and k ∈ N,

W k,p (Rd ; H) = Fp,2

k
(Rd ; H). (Theorem 14.7.9)

If X is a UMD space and p ∈ (1, ∞), k ∈ N, then

W k,r (Rd ; X) = H k,r (Rd ; X). (Theorem 5.6.11)

Embeddings. We have the following continuous embeddings:

S (Rd ; X) ,→ Bp,q
s
(Rd ; X) ,→ S 0 (Rd ; X) (Proposition 14.4.3)
s
Bp,1 (Rd ; X) ,→ Bp,q
s
(Rd ; X) ,→ Bp,∞
s
(Rd ; X) (Proposition 14.4.18)
S (R ; X) ,→
d s
Fp,q (Rd ; X) ,→ S (R ; X)
0 d
(Proposition 14.6.8)
 14.1 Summary of the main results 297
s
Fp,1 s
(Rd ; X) ,→ Fp,q s
(Rd ; X) ,→ Fp,∞ (Rd ; X) (Proposition 14.6.13)
k
Fp,1 (Rd ; X) ,→ W k,p (Rd ; X) ,→ k
Fp,∞ (Rd ; X) (Proposition 14.6.13)
s
Fp,1 (Rd ; X) ,→ H s,p (Rd ; X) ,→ s
Fp,∞ (Rd ; X) (Proposition 14.6.13)

and, if p ∈ [1, ∞),

s
Bp,p∧q (Rd ; X) ,→ Fp,q
s
(Rd ; X) ,→ Bp,p∨q
s
(Rd ; X). (Proposition 14.6.8)

Sobolev embedding theorem I: If (and only if) either one of the following three
conditions holds: p0 = p1 and s0 > s1 ; p0 = p1 and s0 = s1 and q0 6 q1 ;
p0 < p1 and q0 6 q1 and s0 − pd0 > s1 − pd1 ; then

Bps00 ,q0 (Rd ; X) ,→ Bps11 ,q1 (Rd ; X). (Theorem 14.4.19)

Sobolev embedding theorem II: Let p0 , p1 ∈ [1, ∞). If (and only if) either one
of the following three conditions holds: p0 = p1 and s0 > s1 ; p0 = p1 and
s0 = s1 and q0 6 q1 ; p0 < p1 and s0 − pd0 > s1 − pd1 (no condition on q0 , q1 );
then

Fps00,q0 (Rd ; X) ,→ Fps11,q1 (Rd ; X). (Theorem 14.6.14)

Sobolev embedding theorem III: Let p0 , p1 ∈ (1, ∞). If (and only if) either one
of the following three conditions holds: p0 = p1 and s0 > s1 ; p0 < p1 and
s0 − pd0 > s1 − pd1 ; then

H s0 ,p0 (Rd ; X) ,→ H s1 ,p1 (Rd ; X) (Theorem 14.7.1)

and, if in addition s0 , s1 ∈ N, then the same necessary and sufficient conditions

give

W s0 ,p0 (Rd ; X) ,→ W s1 ,p1 (Rd ; X). (Theorem 14.7.1)

For k ∈ N,

k
B∞,1 (Rd ; X) ,→ Cub
k
(Rd ; X) ,→ B∞,∞
k
(Rd ; X). (Proposition 14.4.18)

d
If p0 ∈ [1, ∞] and s0 , s1 > 0 satisfy s0 − p0 > s1 , then

Bps00 ,1 (Rd ; X) ,→ Cub

s1
(Rd ; X) (Proposition 14.4.27)

and, if in addition q ∈ [1, ∞] and s1 6∈ N,

s1
Bps00 ,q (Rd ; X) ,→ Cub (Rd ; X). (Proposition 14.4.27)
298 14 Function spaces

d d
Jawerth–Franke theorem: If p0 < p1 , and s0 − p0 > s1 − p1 , then

Fps00,q (Rd ; X) ,→ Bps11 ,p0 (Rd ; X) (Theorem 14.6.26)

and, if p1 < ∞,

Bps00 ,p1 (Rd ; X) ,→ Fps11,q (Rd ; X). (Theorem 14.6.26)

If k > d, then

k k−d
F1,∞ (Rd ; X) ,→ Cub (Rd ; X). (Corollary 14.6.27)

Embeddings under (co)type assumptions: If (and only if) X has type p ∈ [1, 2],

( 1 − 12 )d
p
Bp,p (Rd ; X) ,→ γ(L2 (Rd ), X). (Theorem 14.5.1)

If (and only if) X has cotype q ∈ [2, ∞],

( 1 − 12 )d
γ(L2 (Rd ), X) ,→ Bq,q
q
(Rd ; X). (Theorem 14.5.1)

If X has type p0 , then for all p ∈ [1, p0 ) we have

1 1
H ( p − 2 )d,p (Rd ; X) ,→ γ(L2 (Rd ), X). (Corollary 14.7.7)

If X has cotype q0 , then for all q ∈ (q0 , ∞) we have

1 1
γ(L2 (Rd ), X) ,→ H ( q − 2 )d,q (Rd ; X). (Corollary 14.7.7)

If X is a UMD Banach space with type p0 ∈ [1, 2] and cotype q0 ∈ [2, ∞], and
if p ∈ (1, ∞), s ∈ R, then

s
Fp,p 0
(Rd ; X) ,→ H s,p (Rd ; X) ,→ Fp,q
s
0
(Rd ; X). (Proposition 14.7.6)

Complex interpolation. Let (X0 , X1 ) be an interpolation couple of Ba-

nach spaces. Let p0 , p1 ∈ [1, ∞] with min{p0 , p1 } < ∞, q0 , q1 ∈ [1, ∞] with
min{q0 , q1 } < ∞, and s0 , s1 ∈ R or k0 , k1 ∈ N. Under these assumptions:

[Bps00 ,q0 (Rd ; X0 ), Bps01 ,q0 (Rd ; X1 )]θ = Bpsθθ ,qθ (Rd ; Xθ ). (Theorem 14.4.30)

If p0 , p1 ∈ (1, ∞) and q0 , q1 ∈ (1, ∞],

s0
[Fp,q (Rd ; X0 ), Fp,q
s1
(Rd ; X1 )]θ = Fp,q
sθ
(Rd ; Xθ ) (Theorem 14.6.23)
 14.1 Summary of the main results 299

and, if in addition X is a UMD space, then

[W k0 ,p0 (Rd ; X0 ), W k1 ,p1 (Rd ; X1 )]θ = H kθ ,pθ (Rd ; Xθ ) (Corollary 14.7.13)

s0 ,p0 d s1 ,p1 d sθ ,pθ d
[H (R ; X0 ), H (R ; X1 )]θ = H (R ; Xθ ). (Theorem 14.7.12)

Real interpolation. Let (X0 , X1 ) be an interpolation couple of Banach

spaces and X be a Banach space. Let p0 , p1 ∈ [1, ∞] with min{p0 , p1 } < ∞,
q0 , q1 ∈ [1, ∞] with min{q0 , q1 } < ∞, s0 , s1 ∈ R, and k0 , k1 ∈ N. Under these
assumptions:

If s0 6= s1 , then

s0
(Bp,q 0
(Rd ; X), Bp,q
s1
1
(Rd ; X))θ,q = Bp,q
sθ
(Rd ; X) (Theorem 14.4.31)
(H s0 ,p (Rd ; X), H s1 ,p (Rd ; X))θ,q = Bp,q
sθ
(Rd ; X). (Theorem 14.4.31)

In addition, if s0 , s1 ∈ N with s0 6= s1 , then

(W s0 ,p (Rd ; X), W s1 ,p (Rd ; X))θ,q = Bp,q

sθ
(Rd ; X) (Theorem 14.4.31)

and if s0 , s1 ∈ (0, 1) with s0 6= s1 and p ∈ [1, ∞), then

(W s0 ,p (Rd ; X), W s1 ,p (Rd ; X))θ,q = Bp,q

sθ
(Rd ; X). (Theorem 14.4.31)

If s0 , s1 ∈ [0, ∞) satisfy s0 6= s1 , then

s0 s1
(Cub (Rd ; X), Cub (Rd ; X))θ,∞ = B∞,∞
sθ
(Rd ; X). (Corollary 14.4.32)

If p ∈ [1, ∞) and s0 6= s1 , then

s0
(Fp,q 0
(Rd ; X), Fp,q
s1
1
(Rd ; X))θ,q = Bp,q
sθ
(Rd ; X). (Proposition 14.6.24)

Duality. With respect to the natural duality pairing of L2 (Rd ; X) and

L2 (Rd ; X ∗ ), for p, q ∈ [1, ∞) and s ∈ R we have, up to equivalent norms,

s
Bp,q (Rd ; X)∗ = Bp−s d ∗
0 ,q 0 (R ; X ) (Theorem 14.4.34)

and, for p, q ∈ (1, ∞) and s ∈ R,

s
Fp,q (Rd ; X)∗ = Fp−s d ∗
0 ,q 0 (R ; X ). (Theorem 14.6.28)

If X ∗ has the Radon-Nikodým property, p ∈ [1, ∞), and s ∈ R, then

0
H s,p (Rd ; X)∗ = H −s,p (Rd ; X ∗ ). (Proposition 5.6.7)
300 14 Function spaces

14.2 Preliminaries
In this section we prepare some, mostly technical, results that will be of use
in our treatments of both Besov and Triebel–Lizorkin spaces.

14.2.a Notation

We start by reviewing some notation that has been introduced in the two
earlier volumes. We use the standard multi-index notation explained in Section
2.5. For the details we refer to the relevant sections (Section 2.4.c for Schwartz
functions, 2.4.d for tempered distributions, 2.5.b and 2.5.d for Sobolev spaces,
and 5.6.a for Bessel potential spaces).
Let X be a Banach space and let d > 1 be an integer. The Schwartz space
S (Rd ; X) is the space of all f ∈ C ∞ (Rd ; X) for which the seminorms

[f ]α,β := sup kxβ ∂ α f (x)k (14.2)

x∈Rd

are finite for all multi-indices α, β ∈ Nd . These seminorms define a locally

convex topology S (Rd ; X) in which sequential convergence fn → f is equiv-
alent to the convergence [f − fn ]α,β → 0 for all multi-indices α, β ∈ Nd . This
topology is metrisable by the metric
X [f − g]α,β
d(f, g) := 2−|α|−|β|
1 + [f − g]α,β
α,β∈Nd

which turns S (Rd ; X) into a complete metric space. Thus S (Rd ; X) has the
structure of a Fréchet space. As a consequence of Lemma 1.2.19 or Lemma
14.2.1, the space Cc∞ (Rd ) ⊗ X is dense in Lp (Rd ; X) for 1 6 p < ∞. We
will prove in Lemma 14.2.1 that Cc∞ (Rd ) ⊗ X is sequentially dense in both
Cc∞ (Rd ; X) and S (Rd ; X).
The space of continuous linear operators

S 0 (Rd ; X) := L (S (Rd ), X)

is called the space of tempered distributions with values in X.

Let D be an open subset of Rd . For 1 6 p 6 ∞ and k ∈ N the Sobolev space
W (D; X) is the space of functions f ∈ Lp (D; X) whose weak derivatives
k,p

∂ α f of order |α| 6 k exist and belong to Lp (D; X). Recall that a function
g ∈ L1loc (D) is said to be the weak derivative of order α of f if
Z Z
f (x)∂ α φ(x) dx = (−1)|α| g(x)φ(x) dx for all φ ∈ Cc∞ (D).
D D

Such a function g, if it exists, is unique. With respect to the norm

X
kf kW k,p (D;X) := k∂ α f kp ,
|α|6k
 14.2 Preliminaries 301

the space W k,p (D; X) is a Banach space. For 1 6 p < ∞ and 0 < s < 1,
the Sobolev–Slobodetskii space W s,p (Rd ; X) is the space of all functions f ∈
Lp (Rd ; X) for which the seminorm
 Z Z kf (x) − f (y)kp 1/p
[f ]W s,p (D;X) := dx dy
D D |x − y|sp+d

is finite. With respect to the norm

kf kW s,p (D;X) := kf kp + [f ]W s,p (D;X) ,

the space W s,p (Rd ; X) is a Banach space. By Theorem 2.5.17, for 1 6 p <∈ ∞
and 0 < s < 1 the real interpolation method gives

(Lp (Rd ; X), W 1,p (Rd ; X))θ,p = W θ,p (Rd ; X)

with equivalent norms.

For 1 6 p 6 ∞ and s ∈ R the Bessel potential space H s,p (Rd ; X) consists
of all u ∈ S 0 (Rd ; X) for which the tempered distribution Js u ∈ S 0 (Rd ; X)
defined by b
2 2 s/2
Js u := ((1 + 4π | · | ) u b)
belongs to Lp (Rd ; X). Recall that the Fourier transform of u is defined by
b(f ) = u(fb) for f ∈ S (Rd ; X), where the Fourier transform of a function
u
f ∈ L1 (Rd ; X) is defined as
Z
fb(ξ) = F f (ξ) := f (x)e−2πix·ξ dx, ξ ∈ Rd .
R

The inverse Fourier transform of a tempered distribution is defined similarly.

With respect to the norm

kukH s,p (Rd ;X) := kJs ukLp (Rd ;X) ,

H s,p (Rd ; X) is a Banach space. The following continuous embeddings hold,

the first being dense if 1 6 p < ∞:

S (Rd ; X) ,→ H s,p (Rd ; X) ,→ S 0 (Rd ; X).

By Theorem 5.6.1, complex interpolation gives

[Lp (Rd ; X), W k,p (Rd ; X)]θ = H θk,p (Rd ; X)

with equivalent norms, provided X is a UMD space, 1 < p < ∞, and k > 1 is
an integer. Under the same assumptions, Theorem 5.6.9 gives

[H s0 ,p (Rd ; X), H s1 ,p (Rd ; X)]θ = H sθ ,p (Rd ; X)

302 14 Function spaces

with equivalent norms, for s0 , s1 ∈ R satisfying s0 < s1 and with sθ = (1 −

θ)s0 + θs1 . Still for UMD spaces X and 1 < p < ∞, by Theorem 5.6.11 for all
integers k > 1 we have

W k,p (Rd ; X) = H k,p (Rd ; X)

with equivalent norms. For k = 0 we have the trivial identities

W 0,p (Rd ; X) = H 0,p (Rd ; X) = Lp (Rd ; X),

valid for all Banach spaces X and 1 6 p < ∞.

For k ∈ N the space Cbk (Rd ; X) consists of all k-times continuously differ-
entiable functions f : Rd → X whose partial derivatives ∂ α f are bounded for
all multi-indices α ∈ Nd satisfying |α| 6 k. With respect to the norm

kf kCbk (Rd ;X) := sup k∂ α f k∞ ,

|α|6k

the space Cbk (Rd ; X) is a Banach space. We denote by Cub k

(Rd ; X) its closed
α
subspace of functions for which ∂ f is uniformly continuous for all |α| 6 k.
For θ ∈ (0, 1) the space of Hölder continuous functions Cbθ (Rd ; X) consists
of all bounded continuous f : Rd → X for which the seminorm

kf (x) − f (y)k
[f ]Cbθ (Rd ;X) := sup
x,y∈Rd ,x6=y |x − y|θ

is finite. With respect to the norm

kf kCbθ (Rd ;X) := kf k∞ + [f ]Cbθ (Rd ;X)

the space Cbθ (Rd ; X) is a Banach space. The Banach space obtained by taking
θ = 1 in these expressions is called the space of Lipschitz continuous functions
and is denoted by Lip(Rd ; X).
For s = k + θ, with k ∈ N and θ ∈ (0, 1), the space Cbs (Rd ; X) is defined as
the space of all f ∈ Cbk (Rd ; X) for which ∂ α f ∈ Cbθ (Rd ; X) for all multi-indices
satisfying |α| 6 k. With the norm

kf kCbs (Rd ;X) := sup k∂ α f kCbθ (Rd ;X) ,

|α|6k

this space is a Banach space. For all s ∈ [0, ∞) we have continuous embeddings

S (Rd ; X) ,→ Cbs (Rd ; X) ,→ S 0 (Rd ; X).

The first embedding is not dense, as non-zero constant functions cannot be

approximated by Schwartz functions. For non-integers s > 0 we will use the
notation
s
Cub (Rd ; X) = Cbs (Rd ; X).
 14.2 Preliminaries 303

14.2.b A density lemma and Young’s inequality

Let U ⊆ Rd be an open set. The elements of the space Cc∞ (U ; X) will be

referred to as X-valued test functions. Sequential convergence in Cc∞ (U ; X) is
defined by insisting that fn → f in Cc∞ (U ; X) if there exists a compact set K
of U containing the support of all fn and k∂ α f − ∂ α f k∞ → 0 for all α ∈ Nd .
Related sequential notions, such as Cauchy sequences, are defined similarly.
Note that if fn → f in Cc∞ (U ; X), then also fn → f in S (Rd ; X), provided
we extend the functions identically zero outside U .

Lemma 14.2.1. The space Cc∞ (Rd ) ⊗ X is sequentially dense in Cc∞ (Rd ; X)
and S (Rd ; X).

Proof. We prove the lemma in two steps.

Step 1 – We first prove that Cc∞ (Rd ; X) is sequentially dense in S (Rd ; X).
Let f ∈ S (Rd ; X). Let ζ ∈ C ∞ (Rd ) satisfy ζ ≡ 1 on {ξ ∈ Rd : |ξ| 6 1}
and ζ ≡ 0 on {ξ ∈ Rd : |ξ| > 2}, and put fn (x) := ζ(x/n)f (x). Then
fn ∈ Cc∞ (Rd ; X). To prove that fn → f in S (Rd ; X) it suffices to check that
for all multi-indices α, β ∈ Nd we have

lim (·)β ∂ α [(1 − ζ(·/n))f ] ∞

= 0.
n→∞

The elementary verification is left to the reader.

Step 2 – Let f ∈ Cc∞ (Rd ; X). Choose bounded open sets O, U, V ⊆ Rd
such that supp(f ) ⊆ U ⊆ U ⊆ V ⊆ V ⊆ O. We first claim that for every
ε > 0 there exists a g ∈ Cc∞ (V ) ⊗ X such that kf − gk∞ 6 ε. Fix ε > 0. Since
f (U ) ⊆ X is compact, it follows that there exist x1 , . . . , xn ∈ X such that
f (U ) ⊆ B(x1 , ε)∪. . . B(xn , ε). The sets U0 = O\U and Uj = f −1 (B(xj , ε))∩V
for j = 1, . . . , n define an open cover (Uj )nj=0 of V . Let (ψj )nj=0 be a smooth
partition
Pn of unity subordinate to thisPcover, i.e., ψj ∈ Cc∞ (Uj ), 0 6 ψj 6 1,
n
and j=0 ψj ≡ 1 on V . Letting g := j=0 ψj ⊗ xj with x0 = 0, for all u ∈ Rd
we have
Xn
kf (u) − g(u)k 6 ψj (u)kf (u) − xj k < ε.
j=0

which proves the claim.

Let φ ∈ Cc∞ (Rd ) satisfy Rd φ(u) du = 1 and put φj (u) := j d φ(ju). By
R

compactness, the exists an index j0 ∈ N such that for all j > j0 and all
g ∈ Cc∞ (V ; X) we have φj ∗g ∈ Cc∞ (O; X) and, for all multi-indices α, β ∈ Nd ,

[φj ∗ g − g]α,β 6 CO,β kφj ∗ ∂ α g − ∂ α gk∞ → 0

as j → ∞, by the uniform continuity of ∂ α g. We conclude that for all such

g and j > 0 we have φj ∗ g → g in S (Rd ; X). In particular, this holds with
g = f . By the claim, we can find a sequence (gk )k>1 in Cc∞ (V ) ⊗ X such that
kf − gk k∞ → 0. Now for each j > j0 the functions gkj := ψj ∗ gk belong to
304 14 Function spaces

Cc∞ (O)⊗X, and by the above we have gkj → gk in S (Rd ; X). For appropriate
jk > j0 we find that gkjk → g in S (Rd ; X). Since gkjk ∈ Cc∞ (O) ⊗ X, this
proves density in Cc∞ (Rd ; X).
To prove density in S (Rd ; X) let f ∈ S (Rd ; X). By Step 1 there exists
a sequence (fn )n>1 in Cc∞ (Rd ; X) such that fn → f in S (Rd ; X). Using
Step 2, for every n > 1 choose a sequence (fn,k )k>1 in Cc∞ (Rd ) ⊗ X such
that fn,k → fn in Cc∞ (Rd ; X). Then in particular, fn,k → fn in S (Rd ; X).
Since convergence in S (Rd ; X) is governed by countably many seminorms, a
standard diagonal argument allows us to find a subsequence such that fn,kn →
f in S (Rd ; X). 
As a corollary to the above lemma we record:

Proposition 14.2.2. For all p ∈ [1, ∞) and s ∈ R the space Cc∞ (Rd ) ⊗ X is
dense in H s,p (Rd ; X).

Proof. By Proposition 5.6.4, for all p ∈ [1, ∞) and s ∈ R we have a dense

embedding S (Rd ; X) ,→ H s,p (Rd ; X). 
We will often make use of the following version of Young’s inequality, which
extends a special case already proven in Lemma 1.2.30.

Lemma 14.2.3 (Young’s inequality). Let p, q, r ∈ [1, ∞] be such that

p + q = 1 + r . If f ∈ L (R ; L (X, Y )) and g ∈ L (R ; X), then f ∗ g ∈
1 1 1 p d q d

Lr (Rd ; Y ) and

kf ∗ gkLr (Rd ;Y ) 6 kf kLp (Rd ;L (X,Y )) kgkLq (Rd ;X) .

Proof. For 1 6 q < ∞, by density it suffices to prove the estimate for g ∈

Cc∞ (Rd ) ⊗ X, and if q = ∞, then p = 1 and r = ∞ and it suffices to prove the
required estimate for f ∈ Cc∞ (Rd ) ⊗ L (X, Y ). In either case, f ∗ g is strongly
measurable and we have the bound kf ∗ gk 6 kf k ∗ kgk. The estimate then
follows from the scalar version of Young’s inequality. 
We recall from Section 1.3 that the variation of an operator-valued measure
Φ : A → L (X, Y ), where (S, A ) is a measurable space, is the measure
kΦk : A → [0, ∞] given by
X
kΦk(A) = sup kΦ(B)k,
π
B∈π

the supremum being taken over all finite disjoint partitions π of the set A ∈ A ;
the is taken in L (X, Y ). We say that Φ has bounded variation if kΦk(S) < ∞.
For a strongly measurable function f : S → X such that
Z
kf (s)k dkΦk(s) < ∞,
S
 14.2 Preliminaries 305

the construction
R of the Bochner integral (see Section 1.2.a) can be repeated
to define S f dΦ as an element of Y satisfying
Z Z
f dΦ 6 kf k dkΦk.
S S

When (S, A , µ) is a measure space, a simple example of an operator-valued

R
measure with bounded variation is obtained by taking Φ(A) := A φ dµ with
φ ∈ L1 (S, µ; L (X, Y )). The total variation of this measure satisfies
kΦk(S) 6 kφkL1 (S,µ;L (X,Y )) .
R
Standard arguments show that S kf kX dkΦk < ∞ if and only if φf ∈
L1 (S; Y ) and in that case
Z Z
f dΦ = φf dµ.
S S

Lemma 14.2.4 (Convolutions with measures). Let Φ : Rd → L (X, Y )

be an operator-valued measure of bounded variation, and let f ∈ Lp (Rd ; X).
d
R
For almost all x ∈ R the integral Rd f (x − y) dΦ(y) is well defined in the
above sense, and the convolution
Z
Φ ∗ f (x) := f (x − y) dΦ(y)
Rd

defines a function Φ ∗ f ∈ Lp (Rd ; Y ) which satisfies

kΦ ∗ f kLp (Rd ;Y ) 6 kΦk(Rd )kf kLp (Rd ;X) .
Proof. For 1 6 p < ∞, Minkowski’s inequality (Proposition 1.2.22) implies
Z Z
x 7→ kf (x − y)k dkΦk(y) 6 x→7 f (x − y) Lp (Rd ;X) dkΦk(y)
Rd Lp (Rd ) Rd
= kf kLp (Rd ;X) kΦk(Rd ).
For p = ∞, the same holds forR trivial reasons. It follows that for almost all
x ∈ Rd the integral Φ ∗ f (x) = Rd f (x − y) dΦ(y) is well defined in Y . By ap-
proximation with simple functions it is seen that Φ ∗ f is strongly measurable,
and since Z
kΦ ∗ f (x)k 6 kf (x − y)k dkΦk(y),
Rd
the required estimate also follows. 

14.2.c Inhomogeneous Littlewood–Paley sequences

We now introduce one of our main technical tools, which allows us to break
up a function spectrally into pieces with control on their frequencies.
Let Φ denote the set of all Schwartz functions ϕ ∈ S (Rd ) with the follow-
ing properties:
306 14 Function spaces

(i) 0 6 ϕ(ξ)
b 6 1, ξ ∈ Rd ,
(ii) ϕ(ξ)
b = 1 if |ξ| 6 1,
(iii) ϕ(ξ)
b = 0 if |ξ| > 23 .
Such functions can be constructed in a similar way as in Lemma 5.5.21.
Remark 14.2.5. If φ ∈ Φ, the function ψ ∈ S (Rd ) given by

ψ(ξ) b − ϕ(2ξ)
b := ϕ(ξ) b

is a smooth Littlewood–Paley function in the sense of Definition 5.5.20, i.e.,

(i) ψb is smooth, non-negative, and supported in {ξ ∈ Rd : 21 6 |ξ| 6 2};
X
(ii) b −k ξ) = 1 for all ξ ∈ Rd \ {0}.
ψ(2
k∈Z

Remark 14.2.6. It is possible to choose the function ϕ is real and even (or
equivalently ϕ
b real and even). In that case it would be possible to use real
Banach spaces in several of the definitions and results of this chapter. For
instance if f ∈ Lp (Rd ; X) or even S 0 (Rd ; X), then ϕ ∗ f can be defined
without using any complex structure.
Definition 14.2.7 (Inhomogeneous Littlewood–Paley sequence). The
inhomogeneous Littlewood–Paley sequence associated with a function ϕ ∈ Φ
is the sequence (ϕk )k>0 in S (Rd ) given by

ϕ
b0 (ξ) := ϕ(ξ),
b k = 0, ξ ∈ Rd ,
(14.3)
ϕ b −k ξ) − ϕ(2
bk (ξ) := ϕ(2 b −k+1 ξ), k > 1, ξ ∈ Rd .

Note the scaling property

ϕ b1 (2−k+1 ξ),
bk (ξ) = ϕ k > 1, (14.4)

and the telescoping properties

n
X X
ϕ b0 (2−n ξ),
bk (ξ) = ϕ ϕ
bk (ξ) = 1. (14.5)
k=0 k>0

We will often use the simple L1 -norm identity

Xn Z Z Z
2πix·ξ −n n
ϕk = e ϕ
b0 (2 ξ) dξ dx = 2 |ϕ0 (2n x)| dx = kϕ0 k1 ,
1 Rd Rd Rd
k=0
(14.6)

which implies
n
X n−1
X
kϕk k1 = ϕk − ϕk 6 2kϕ0 k1 , k > 1. (14.7)
1
k=0 k=0
 14.2 Preliminaries 307

The adjective ‘inhomogeneous’ refers to the special role played by the

function ϕ0 whose support contains an open neighbourhood of the origin.
Inhomogeneous Littlewood–Paley sequences will be used to define the
classes of Besov spaces and Triebel–Lizorkin spaces. Up to equivalent norms,
the definitions of these spaces turn out to be independent of the particular
inhomogeneous Littlewood–Paley sequence chosen. This allows us to fix an
arbitrary such sequence once and for all and always work with that given se-
quence. In order to avoid endless repetitions we therefore make the following
convention.
Convention 14.2.8. Throughout this entire chapter, (ϕk )k∈N denotes the in-
homogeneous Littlewood–Paley sequence associated with a function ϕ ∈ Φ
which we fix once and for all. Whenever this is useful, we extend the index
set of the sequence to include the negative integers by setting
φk ≡ 0, k = −1, −2, . . .
Constants in estimates involving a Littlewood–Paley sequences or spaces de-
fined by using them will often also depend on the generating function ϕ ∈ Φ,
but since it is considered to be fixed we will not express these dependencies
in our estimates.
Let us collect some easy properties of inhomogeneous Littlewood–Paley
sequences. It is immediate to check the Fourier support property
3 k
bk (ξ) ≡ 1 for
ϕ · 2 6 |ξ| 6 2k , k > 1, (14.8)
4
and
bk ⊆ {ξ ∈ Rd : 2k−1 6 |ξ| 6 3 · 2k−1 },
supp ϕ k > 1. (14.9)
In particular we have the disjointness property
bj ∩ supp ϕ
supp ϕ bk = ∅, |j − k| > 2, (14.10)
which implies the orthogonality properties
ϕ bk = 0 and ϕj ∗ ϕk = 0,
bj ϕ |j − k| > 2. (14.11)
From (14.5) and (14.11) we infer
1
X
bk+j ≡ 1 on supp(ϕ
ϕ bk ), k > 0, (14.12)
j=−1

using the convention ϕ−1 = 0 for the case k = 0.

By Proposition 2.4.32, for ψ ∈ S (Rd ) and u ∈ S 0 (Rd ; X) the convolution

ψ ∗ u = F −1 (ψbfb) (14.13)

is well defined as element of C ∞ (Rd ; X) and as such it has at most polynomial

growth. For later use we record the following useful consequence:
308 14 Function spaces

Lemma 14.2.9. Every f ∈ S 0 (Rd ; X) with compact Fourier support belongs

to C ∞ (Rd ; X) and has at most polynomial growth.
Proof. This follows from Proposition 2.4.32 by writing f = f ∗ g with g ∈
S (Rd ) satisfying gb ≡ 1 on supp(f ). 
Returning to the main line of development, by applying (14.13) to the con-
volutions ϕk ∗ u, the latter can be identified with distributions in S 0 (Rd ; X)
and we have the following result:
Lemma 14.2.10. Let E = S (Rd ; X) or E = S 0 (Rd ; X). For all f ∈ E we
have
X X1 X
f= ϕk ∗ f = ϕk+` ∗ ϕk ∗ f
k>0 `=−1 k>0

with convergence of the sums in E.

Proof. The second identity follows by applying the first twice and (14.11). It
thus remains to prove the first identity.
By the second identity in (14.5), (14.13), and the continuity of the
Fourier
P transform on E proved in Proposition 2.4.22, it suffices to show that
k>0 ϕbk g = g for all g ∈ E, with convergence of the sum in E.
First suppose that g ∈ S (Rd ; X). In view of the first identity in (14.5) we
must to show that, for arbitrary multi-indices α, β,

b −n ·))g]
lim (·)β ∂ α [(1 − ϕ(2 ∞
= 0.
n→∞

This is elementary and left to the reader.

Next suppose
P that g ∈ S 0 (Rd ; X). Fix a function ψ ∈ S (Rd ).PWe need to
check that k>0 g(ψ ϕbk ) = g(ψ). For this it suffices to check that k>0 ψ ϕ
bk =
ψ in S (R ), which is the content of the previous case.
d

As a first application of Littlewood–Paley sequence techniques we prove a
lemma that will be useful for establishing Fourier multiplier results in later
subsections. For its proof we recall from Volume I the space
b
L1 (Rd ; X) := {f ∈ L∞ (Rd ; X) : F −1 f ∈ L1 (Rd ; X)},

where the inverse Fourier transforms is viewed as an element of S 0 (Rd ; X).

With respect to the norm

kf kL1 (Rd ;X) = kfbkL1 (Rd ;X) ,

b

b
L1 (Rd ; X) is a Banach space. It enjoys the scaling invariance property

kf (λ·)kL1 (Rd ;X) = kf kL1 (Rd ;X) , λ > 0, (14.14)

b b

which is proved by a simple change of variables.

 14.2 Preliminaries 309

Lemma 14.2.11 (Integrability of Fourier transforms – I). Let f ∈

C d+1 (Rd ; X), and suppose that there exists an ε > 0 such that

Cf,d,ε := max sup (1 + |ξ||α|+ε )k∂ α f (ξ)k < ∞.

|α|6d+1 ξ∈Rd

Then fb ∈ L1 (Rd ; X) and kfbkL1 (Rd ;X) .d,ε Cf,d,ε .

Note that Cf,d,ε is trivially finite (for all ε > 0) if f ∈ C d+1 (Rd ; X) has
compact support.
P
Proof. In view of (14.5) we have kf kL1 (Rd ;X) 6 j>0 kϕ
bj f kL1 (Rd ;X) , and
b b

therefore it is enough to show that for all j > 0 we have

bj f kL1 (Rd ;X) .d 2−(j−1)ε Cf,d,ε .

kϕ (14.15)
b

First we consider indices j > 1. Setting B := {ξ ∈ Rd : |ξ| 6 1}, by (14.4)

and (14.14) we obtain

b1 (·)f (2j−1 ·)kL1 (Rd ;X)

bj f kL1 (Rd ;X) = kϕ
kϕ
b b

b1 (·)f (2j−1 ·))kL1 (B;X) + kF (ϕ

= kF (ϕ b1 (·)f (2j−1 ·))kL1 (Rd \B;X)
=: T1 + T2 .

The first term is easy to handle. Indeed, since kF kL1 →L∞ 6 1 and 0 6 ϕ
b1 6 1,

b1 (·)f (2j−1 ·))k∞

T1 6 |B| kF (ϕ
b1 (·)f (2j−1 ·)kL1 (Rd ;X) 6 |B|kf (2j−1 ·)kL1 (3B\B;X) ,
6 |B|kϕ

b1 is supported in 3B \ B in the last step. Together with the

using that ϕ
assumed bound for f with α = 0, for ξ ∈ 3B \ B we have

Cf,d,ε
kf (2j−1 ξ)k 6 6 2−(j−1)ε Cf,d,ε .
1 + 2(j−1)ε |ξ|ε

Combining this with the previous estimate, this gives the bound T1 6
2−(j−1)ε Cf,d,ε |3B \ B||B|.
For the second term we use the finiteness of Cd := Rd \B |x|−d−1 dx to
R

obtain

T2 6 Cd ξ 7→ |ξ|d+1 F (ϕ
b1 f (2j−1 ·))(ξ) ∞
.

By the estimate |ξ|d+1 .d |α|=d+1 |ξ α | and the identity (2πi)|α| ξ α F (g)(ξ) =

P

F (∂ α g)(ξ), for each ξ ∈ Rd we can further estimate

X
|ξ|d+1 F (ϕb1 f (2j−1 ·))(ξ) X .d (2πξ)α F (ϕb1 f (2j−1 ·))(ξ) X
|α|=d+1
310 14 Function spaces
X
= F (∂ α [ϕ
b1 f (2j−1 ·)])(ξ) X
.
|α|=d+1

Using that ϕ
b1 is compactly supported we obtain

F (∂ α [ϕ
b1 f (2j−1 ·)]) ∞
b1 f (2j−1 ·)]
6 ∂ α [ϕ 1
b1 f (2j−1 ·)]
.d ∂ α [ϕ ∞
.

After an application of the Leibniz rule it remains to estimate terms of the

form ∂ β ϕ
b1 ∂ γ [f (2j−1 ·)] with |β| + |γ| = |α| = d + 1. By the assumptions and
the fact that ϕ b1 is supported in 3B \ B,

b1 ∂ γ [f (2j−1 ·)]k∞ .d
k∂ β ϕ sup k2(j−1)|γ| ∂ γ f (2j−1 ξ)k 6 2−(j−1)ε Cf,d,ε .
16|ξ|63

It follows that T2 .d 2−(j−1)ε Cf,d,ε . This proves (14.15) for j > 1. The case
j = 0 can be shown in a similar way, skipping the dilation step. 
For later reference we state the following consequence of Lemma 14.2.11.

Lemma 14.2.12. Let λ > 0 and suppose that f ∈ C d+1+dλe (Rd ; X) has sup-
port in the ball BR = {ξ ∈ Rd : |ξ| 6 R}. Then (1 + | · |)λ fb(·) ∈ L1 (Rd ; X)
and
k(1 + | · |)λ fb(·)kL1 (Rd ;X) 6 CR,d kf kC d+1+dλe (Rd ;X) .
b

Proof. Upon replacing λ by dλe we may assume that λ ∈ N. By Lemma

14.2.11 we have fb ∈ L1 (Rd ; X). Therefore it suffices to prove the estimate
with (1 + | · |)λ replaced by | · |λ . P
Arguing as before, since |x|λ .d |β|=λ |xβ |,
X
| · |λ fb L1 (Rd ;X)
.d,R βfk 1 d
k∂d L (R ;X) .
|β|=λ

Therefore, the required result follows from Lemma 14.2.11 applied to ∂ β f . 

14.3 Interpolation of Lp -spaces with change of weights

When (S, A , µ) is σ-finite measure space, we call a measurable function w :
S → [0, ∞] a weight if w(x) ∈ (0, ∞) for almost all x ∈ S. On earlier occasions
(e.g., in Appendix J and Chapter 11) we have considered the weighted spaces
n
Lq (w; X) := f : S → X strongly measurable,
Z 1/q o
kf kLq (w;X) := kf (x)kqX w(x) dµ(x) <∞ .
S

For the present purposes, it is more convenient to introduce the variant

 14.3 Interpolation of Lp -spaces with change of weights 311
n
Lqw (S; X) := f : S → X strongly measurable,
Z 1/q o
kf kLw
q
(S;X) := kf (x)w(x)kqX dµ(x) <∞ .
S

For q < ∞, this is just another way of expressing the same spaces with a
different normalisation of the weight, namely Lqw (S; X) = Lq (wq ; X). How-
ever, using the usual modification for q = ∞, the first version reduces to just
L∞ (w; X) = L∞ (S; X) (since dµ and w dµ share the same zero sets), whereas
Lq∞ (S; X) with norm kf kL∞ w (S;X)
= kf wkL∞ (S;X) is a new space with non-
trivial dependence on the weight w.

14.3.a Complex interpolation

Our first main result concerning these spaces is the following:

Theorem 14.3.1 (Stein–Weiss). Let (Y0 , Y1 ) be an interpolation couple of

Banach spaces, let q0 , q1 ∈ [1, ∞] satisfy min{q0 , q1 } < ∞. Let (S, A , µ) be
a σ-finite measure space, let w0 , w1 be two weight functions on S, and let
θ ∈ (0, 1). Then
q
[Lqw00 (S; Y0 ), Lqw11 (S; Y1 )]θ = Lw (S; [Y0 , Y1 ]θ )

isometrically, where
1 1−θ θ
= + , w = w01−θ w1θ .
q q0 q1
We first record the simple:

Lemma 14.3.2. In the setting of Theorem 14.3.1, if fn → f in the norm of

Lqw00 (S; Y0 ) + Lqw11 (S; Y1 ), then a subsequence converges almost everywhere in
the norm of Y0 + Y1 to the same limit function.

Proof. We assume that kfn − f kLqw0 (S;Y0 )+Lqw1 (S;Y1 ) → 0. Hence, for every
0 1
n, there is a decomposition fn − f = fn0 + fn1 , where kfnj kLqwj (S;Yj ) → 0 for
j
j = 0, 1. By the well known version of the Lemma in just one Lp space, a
subsequence of fn0 converges to 0 almost everywhere in the norm of Y0 . By the
same result, a further subsequence of fn1 also converges to 0 almost everywhere
in the norm if Y1 . Thus, along this last subsequence, fn −f = fn0 +fn1 converges
to 0 almost everywhere in the norm of Y0 + Y1 . 
Proof of Theorem 14.3.1. The unweighted version (w0 = w1 = w ≡ 1) of this
result is contained in Theorem 2.2.6. We will reduce the weighted version to
S∞Y := [Y0 , Y1 ]θ . For n ∈ Z+ , we denote
this special case. Let us abbreviate
Sn := {n−1 6 w0 , w1 6 n}. Then n=1 Sn exhausts S, up to a set of measure
zero, by definition of weights.
312 14 Function spaces

q0
Step 1 – Lqw (S; [Y0 , Y1 ]θ ) ⊆ [Lw 0
(S; Y0 ), Lqw11 (S; Y1 )]θ :
q
Let f ∈ Lw (S; Y ), and assume first assume that {f 6= 0} is contained in Sn
for some n ∈ N. Thus

φ := f w ∈ Lq (S; Y ) = [Lq0 (S; Y0 ), Lq1 (S; Y1 )]θ ,

where the equality of space is Theorem 2.2.6, and hence φ = Φ(θ) for some
Φ ∈ H (Lq0 (S; Y0 ), Lq1 (S; Y1 )), where this notation of holomorphic functions
on the unit strip with appropriate boundary behaviour is defined in Section
C.2. The relation φ = Φ(θ) remains valid if we replace Φ(z) by Φ(z)1En , and
hence all the subsequent considerations can be restricted to En . In particular,
multiplication by any power of w0 or w1 is then a bounded operation on any
of the (weighted or not) Lp spaces appearing in this argument. Now
−(1−θ)
f = φw−1 = Φ(θ)w0 w1−θ = F (θ),
−(1−z)
where F (z) := Φ(z)w0 w1−z ∈ H (Lq0 (w0 ; Y0 ), Lq1 (w1 ; Y1 )). Qualita-
tively, the last inclusion is easy from the corresponding relation for Φ and
the restriction of the supports on En , where all multiplications by powers of
wi are bounded. Quantitatively, we have
−(1−j)
kF (j + it)kLqj (wj ;Yj ) = kΦ(j + it)w0 w1−j kLqj (wj ;Yj )
= kΦ(j + it)kLqj (S;Yj ) , j = 0, 1,

thus, recalling that kF kH (X0 ,X1 ) := maxj=0,1 supt∈R kF (j + it)kXj ,

kF kH (Lw
q0
(S;Y0 ),Lw1 (S;Y1 )) = kΦkH (Lq0 (S;Y0 ),Lq1 (S;Y1 )) ,
q (14.16)
0 1

and hence
kf k[Lw
q0
(S;Y0 ),Lw1 (S;Y1 )]θ 6 kF kH (Lw0 (S;Y0 ),Lw1 (S;Y1 ))
q q q
0 1 0 1

= kΦkH (Lq0 (S;Y0 ),Lq1 (S;Y1 )) .

Taking the infimum over all Φ in this space with φ = Φ(θ), we obtain

kf k[Lqw0 (S;Y0 ),Lqw1 (S;Y1 )]θ 6 kφk[Lq0 (S;Y0 ),Lq1 (S;Y1 )]θ
0 1

= kφkLq (S;Y ) = kf kLw

q
(S;Y ) .

Recall that the previous estimate was obtained under the assumption
that f ∈ Lqw (S; Y ) satisfies {f 6= 0} ⊆ Sn . For a general f ∈ Lqw (S; Y ),
this bound holds with either 1Sn f of 1Sn f − 1Sm f in place of f . Since
1Sn f → f in Lqw (S; Y ) by dominated convergence, it follows that 1Sn f
is a Cauchy sequence, and hence convergent, in the interpolation space
q0 q1 q0 q1
[Lw 0
(S; Y0 ), Lw 1
(S; Y1 )]θ and thus in the sum space Lw 0
(S; Y0 ) + Lw 1
(S; Y1 )
by Lemma C.2.5. By Lemma 14.3.2, a subsequence converges almost every-
where to the same limit function. But it is clear that the a.e. limit is f , and
hence
 14.3 Interpolation of Lp -spaces with change of weights 313

kf k[Lqw0 (S;Y0 ),Lqw1 (S;Y1 )]θ = lim k1Sn f k[Lqw0 (S;Y0 ),Lqw1 (S;Y1 )]θ
0 1 n→∞ 0 1

6 lim k1Sn f kLw

q
(S;Y ) = kf kLw
q
(S;Y ) .
n→∞

Step 2 – [Lqw00 (S; Y0 ), Lw

q1
1
(S; Y1 )]θ ⊆ Lqw (S; [Y0 , Y1 ]θ ):
q0
Let f = F (θ) ∈ [Lw 0
(S; Y0 ), Lqw11 (S; Y1 )]θ , where
F ∈ H (Lqw00 (S; Y0 ), Lqw11 (S; Y1 )).
As before, we first assume that {f 6= 0} ⊆ Sn , and then without loss of
generality (multiplying by 1En if necessary) that F (z) has the same property
for every z. We can then reverse the previous reasoning. Defining
(1−z)
Φ(z) := F (z)w0 w1z ,
we check the same relation (14.16), and hence
kf kLq (w;Y ) = kF (θ)wkLq (S;Y ) = kΦ(θ)k[Lq0 (S;Y0 ),Lq1 (S;Y0 )]θ
6 kΦkH (Lq0 (S;Y0 ),Lq1 (S;Y0 )) = kF kH (Lw
q0
(S;Y0 ),Lw1 (S;Y0 )) .
q
0 1

Taking the infimum over the relevant F with F (θ) = f , we get

kf kLqw (S;Y ) 6 kf k[Lw
q0
(S;Y0 ),Lw1 (S;Y0 )]θ ,
q {f 6= 0} ⊆ Sn . (14.17)
0 1

Consider next a general f ∈ [Lqw00 (S; Y0 ), Lw q1

1
(S; Y0 )]θ . Multiplication by
p
1Sn contracts all L spaces, including weighted ones, and hence also the inter-
q0
polation space [Lw 0
(S; Y0 ), Lqw11 (S; Y1 )]θ by Theorem C.2.6. Now (14.17) holds
with 1Sn f in place of f , and hence
k1Sn f kLq (w;Y ) 6 k1Sn f k[Lw
q0
(S;Y0 ),Lw1 (S;Y0 )]θ 6 kf k[Lw0 (S;Y0 ),Lw1 (S;Y0 )]θ .
q q q
0 1 0 1

But then monotone convergence shows that

kf kLq (w;Y ) = lim k1Sn f kLqw (S;Y ) 6 kf k[Lw
q0
(S;Y0 ),Lw1 (S;Y0 )]θ .
q
n→∞ 0 1

This completes the proof. 

For easy reference later in this chapter, we state the special case of the previous
result for sequence space with the weights ws (k) = 2ks on the integers.
q
Proposition 14.3.3 (Complex interpolation of the spaces `w s
(Y )).
Let (Y0 , Y1 ) be an interpolation couple of Banach spaces, let q0 , q1 ∈ [1, ∞]
satisfy min{q0 , q1 } < ∞, and let s0 , s1 ∈ R and θ ∈ (0, 1). Then
[`qw0s0 (Y0 ), `qw1s1 (Y1 )]θ = `qws ([Y0 , Y1 ]θ )
1 1−θ θ
isometrically, where s = (1 − θ)s0 + θs1 and q = q0 + q1 .

Proof. The condition s = (1 − θ)s0 + θs1 is equivalent to ws = ws1−θ

0
wsθ1 ;
whence the Proposition is a special case of Theorem 14.3.1. 
314 14 Function spaces

14.3.b Real interpolation

We next turn to the case of real interpolation. Recall that for a Banach couple
(X0 , X1 ), the real interpolation space (X0 , X1 )θ,p with p ∈ [1, ∞] and θ ∈
(0, 1), was introduced in Section C.3. Also recall from Theorem C.3.14 that if
p0 , p1 ∈ [1, ∞] satisfy p1 = 1−θ θ
p0 + p1 , then (X0 , X1 )θ,p = (X0 , X1 )θ,p0 ,p1 with
equivalent norms, where the latter denotes the Lions–Peetre interpolation of
X0 and X1 (second mean method). The main result of this section is as follows.

Theorem 14.3.4 (Stein–Weiss, real version). Let (Y0 , Y1 ) be an inter-

polation couple of Banach spaces, let q0 , q1 ∈ [1, ∞] satisfy min{q0 , q1 } < ∞.
Let (S, A , µ) be a σ-finite measure space, let w0 , w1 be two weight functions
on S, and let θ ∈ (0, 1). Then

(Lqw00 (S; Y0 ), Lqw11 (S; Y1 ))θ,q0 ,q1 = Lqw (S; (Y0 , Y1 )θ,q0 ,q1 )

isometrically, where
1 1−θ θ
= + , w = w01−θ w1θ .
q q0 q1
In particular,

(Lqw00 (S; Y0 ), Lqw11 (S; Y1 ))θ,q = Lqw (S; (Y0 , Y1 )θ,q ),

with equivalent norms.

Proof. The unweighted version (w0 = w1 = w ≡ 1) of this result is contained

in Theorem 2.2.10. We will reduce the weighted version to this special case.
Let us abbreviate Y := (Y0 , Y1 )θ,q0 ,q1 . As in the proof of Theorem 14.3.1
S∞ we
denote Sn := {n−1 6 w0 , w1 6 n} for each n ∈ Z+ , and observe that n=1 Sn
exhausts S, up to a set of measure zero, by definition of weights.
Step 1 – Lqw (S; Y ) ⊆ (Lqw00 (S; Y0 ), Lqw11 (S; Y1 ))θ,q0 ,q1 :
Let f ∈ Lq (w; Y ), and assume first that {f 6= 0} is contained in Sn for
some n ∈ N. We also make the technical assumption that the weights wj are
discrete, in that they only take values of the form ρk , where ρ > 1 is a fixed
number, and k ∈ Z. This plays a role in the representation (14.18) below. Now

φ := f w ∈ Lq (S; Y ) = (Lq0 (S; Y0 ), Lq1 (S; Y1 ))θ,q0 ,q1 ,

where the equality of spaces is Theorem 2.2.10. Hence, by Definition C.3.10

of the Lions–Peetre interpolation method ( , )θ,q0 ,q1 , for some strongly mea-
surable Φ : (0, ∞) → Lq0 (S; Y0 ) ∩ Lq1 (S; Y1 ), we have
Z ∞
dt
φ= Φ(t) ,
0 t
 14.3 Interpolation of Lp -spaces with change of weights 315

where tj−θ Φ(t) ∈ Lqj ( dt/t; Lqj (S; Yj )) for j = 0, 1, and (as a consequence)
the improper integral converges in Lq0 (S; Y0 )+Lq1 (S; Y1 ). Multiplying by 1Sn
if necessary, we may assume that each Φ(t) is also supported on Sn .
Choosing the auxiliary weight W := w0−1 w1 , we then have
Z ∞ Z ∞ Z ∞
dt dt dt
f = φw−1 = Φ(t)w−1 = Φ(tW )w−1 =: F (t) .
0 t 0 t 0 t
On Sn , both wj are bounded from above and below. Due to the technical
assumption on the discreteness of their ranges, both these weights, and hence
W , only take finitely many possible value on Sn . Hence
K
X
F (t) = Φ(tW )w−1 = 1Ek Φ(tαk )βk−1 (14.18)
k=1

for some αk , βk ∈ (0, ∞) and sets Ek ⊆ Sn , from which it is immediate that

also F : (0, ∞) → Lq0 (S; Y0 ) ∩ Lq1 (S; Y1 ) is strongly measurable. This still
remains true with each Lqj (S; Yj ) replaced by Lqj (wj ; Yj ) since the intersec-
tions of these spaces with functions supported on Sn coincide. With these
qualitative issues out of the way, we make the quantitative observation
Z ∞
q dt
ktj−θ F (t)k jqj
Lwj (S;Yj ) t
0
Z ∞
q dt
= ktj−θ Φ(tW )w−1 wj kLjqj (S;Yj )
t
Z0 ∞ (14.19)
q dt
= kW θ−j w−1 wj tj−θ Φ(t)kLjqj (S;Yj )
t
Z0 ∞
q dt
= ktj−θ Φ(t)kLjqj (S;Yj ) ,
0 t

where in the last step our choice W := w0−1 w1 and the assumption w =
w01−θ w1θ show that W θ−j w−1 wj ≡ 1 for both j = 0, 1 (and indeed having this
identity dictates our choice of the auxiliary W ).
Now, by the Lions–Peetre method, we have

kf k(Lqw0 (S;Y0 ),Lqw1 (S;Y1 ))θ,q 6 sup kt 7→ tj−θ F (t)kLqj ( dt/t;Lqwj (S;Yj ))
0 1 0 ,q1 j
j=0,1

= sup kt 7→ tj−θ Φ(t)kLqj ( dt/t;Lqj (S;Yj )) ,

j=0,1

and taking the infimum over all such Φ shows that

kf k(Lqw0 (S;Y0 ),Lqw1 (S;Y1 ))θ,q 6 kφk(Lq0 (S;Y0 ),Lq1 (S;Y1 ))θ,q0 ,q1
0 1 0 ,q1

= kφkLq (S;Y ) = kf kLqw (S;Y ) .

We proved this assuming that {f 6= 0} ⊆ Sn . For arbitrary f ∈ Lqw (S; Y ),

this is true with either 1Sn f or 1Sn f −1Sm f in place of f . It follows that 1Sn f
316 14 Function spaces

is a Cauchy sequence, and hence convergent, in (Lqw00 (S; Y0 ), Lw q1

1
(S; Y1 ))θ,q0 ,q1 ,
q0 q1
and thus in Lw0 (S; Y0 ) + Lw1 (S; Y1 ) by the very Definition C.3.10 (recall that
f ∈ (X0 , X1 )θ,q0 ,q1 is given by an integral that converges in X0 + X1 ). By
Lemma 14.3.2, a subsequence converges almost everywhere to the same limit,
and hence this limit must be f . Thus f ∈ (Lqw00 (S; Y0 ), Lwq1
1
(S; Y1 ))θ,q0 ,q1 , and

kf k(Lqw0 (S;Y0 ),Lqw1 (S;Y1 ))θ,q = lim k1Sn f k(Lw

q0 q
(S;Y0 ),Lw1 (S;Y1 ))θ,q
0 1 0 ,q1 n→∞ 0 1 0 ,q1

6 lim k1Sn f k q
Lw (S;Y ) = kf k Lqw (S;Y ).
n→∞

We still had the additional hypothesis on the discreteness of the ranges of

both wj . For arbitrary weights wj and ρ > 1, we consider
wj,ρ := sup{ρk : ρk 6 wj , k ∈ Z},
which clearly satisfy the discreteness property, as well as wj,ρ 6 wj 6 ρwj,ρ .
Hence
kf kLw
qj
(S;Yj )
6 ρkf kLqjρ (S;Yj )
j w
j

and Theorem C.3.16 gives the first estimate in

kf k(Lw
q0 q
(S;Y0 ),Lw1 (S;Y1 ))θ,q 6 ρ1−θ ρθ kf k(Lw
q0 q
(S;Y0 ),Lw11,ρ (S;Y1 ))θ,q0 ,q1
0 1 0 ,q1 0,ρ

= ρkf kLq 1−θ (S;Y )

θ
w0,ρ w1,ρ

6 ρkf kLq (w;Y ) .

Taking the limit ρ → 1, we finally deduce
kf k(Lw
q0 q
(S;Y0 ),Lw1 (S;Y1 ))θ,q 6 kf kLw
q
(S;Y )
0 1 0 ,q1

unconditionally.
Step 2 – (Lqw00 (S; Y0 ), Lqw11 (S; Y1 ))θ,q0 ,q1 ⊆ Lwq
(S; Y ):
q0 q1
Let f ∈ (Lw0 (S; Y0 ), Lw1 (S; Y1 ))θ,q0 ,q1 . We make the same initial assumptions
on both
R ∞ f and the weights wj as in the previous part. By definition, we have
q
f = 0 F (t) dtt with t
j−θ
F (t) ∈ Lqj ( dt/t; Lwjj (S; Yj )). Working the previous
computations backwards, we find that
Z ∞ Z ∞ Z ∞
dt −1 dt dt
φ := f w = F (t)w = F (tW )w =: Φ(t) ,
0 t 0 t 0 t
where Φ satisfies the relevant measurability conditions (by the structural as-
sumptions on the weights) and the quantitative relation (14.19). We conclude
that
kφk(Lq0 (S;Y0 ),Lq1 (S;Y1 ))θ,q0 ,q1 6 sup ktj−θ Φ(t)kLqj ( dt/t;Lqj (S;Yj ))
j=0,1

= sup ktj−θ F (t)kLqj ( dt/t;Lqwj (S;Yj ))

,
j=0,1 j
 14.3 Interpolation of Lp -spaces with change of weights 317

and taking the infimum over all relevant F ,

kf kLqw (S;Y ) = kφkLq (S;Y ) = kφk(Lq0 (S;Y0 ),Lq1 (S;Y1 ))θ,q0 ,q1
6 kf k(Lw
q0 q
(S;Y0 ),Lw1 (S;Y1 ))θ,q .
0 1 0 ,q1

For a general f in the interpolation space, applying the previous conclusion

to 1Sn f in place of f , we have

k1Sn f kLw
q
(S;Y ) 6 k1Sn f k(Lqw0 (S;Y0 ),Lqw1 (S;Y1 ))θ,q
0 1 0 ,q1

6 kf k q q
(Lw00 (S;Y0 ),Lw11 (S;Y1 ))θ,q0 ,q1 ,

since multiplication by 1Sn is clearly contractive on each Lqj (wj ; Yj ), and

hence on the interpolation space by Theorem C.3.16. It then follows from
monotone convergence that

kf kLw
q
(S;Y ) = lim k1Sn f kLqw (S;Y ) 6 kf k(Lqw0 (S;Y0 ),Lqw1 (S;Y1 ))θ,q .
n→∞ 0 1 0 ,q1

Finally, the discreteness assumption on the weights can be removed by the

same considerations as in the previous part: For general weights wj and the
auxiliary discrete wj,ρ as in the previous part, we have

kf kLw
q
(S;Y ) = kf kLq 1−θ (S;Y ) 6 ρ(1−θ)+θ kf kLq 1−θ (S;Y )
w0 θ
w1 θ
w0,ρ w1,ρ

6 ρkf k((Lqw0 q
(S;Y0 ),Lw11,ρ (S;Y1 ))θ,q0 ,q1
0,ρ

6 ρkf k((Lqw0 (S;Y0 ),Lw

q1
(S;Y1 ))θ,q ,
0 1 0 ,q1

and taking the limit ρ → 1 completes the proof. 

For applications of the real interpolation theorem to Besov spaces, it is useful
to include a version that is genuine variant, rather than just a special case of
the previous theorem. This version is concerned with the particular case of
S = N or S = Z with the exponential weights ws (k) = 2ks , and restricting
to just one range space Y0 = Y1 = Y . Remarkably, under these circumstances
the condition 1q = 1−θ θ
q0 + q1 of Theorem 14.3.4 can be omitted:

Proposition 14.3.5 (Real interpolation of the spaces `qws (Y )). Let

p, q0 , q1 ∈ [1, ∞], let s0 , s1 ∈ R satisfy s0 6= s1 , let θ ∈ (0, 1), and let
s = (1 − θ)s0 + θs1 . Then

(`qw0s0 (Y ), `qw1s1 (Y ))θ,p = `pws (Y ) with equivalent norms,

with constants in the norm estimates only depending on θ, p, s0 , s1 . Moreover,

for all y ∈ `qw0s0 (Y ) ∩ `qw1s1 (Y ) we have

kyk`pws (Y ) 6 Ckyk1−θ
q
` 0 (Y )
kykθ`qw1 (Y )
,
ws 0 s1

where C only depends on s0 , s1 , θ.

318 14 Function spaces

Proof. We will present the details for S = N, as the case S = Z is proved in

the same way. By interchanging the roles of `qw0s0 (Y ) and `w q1
s1
(Y ) if necessary,
without loss of generality we may assume that s0 > s1 .
Since `wq0
s0
(Y ) ,→ `∞ q1 ∞
ws0 (Y ) and `ws1 (Y ) ,→ `ws1 (Y ) continuously, real inter-
polation (Theorem C.3.3) gives (`ws0 (Y ), `ws1 (Y ))θ,p ,→ (`∞
q0 q1 ∞
ws0 (Y ), `ws1 (Y ))θ,p
q0
continuously. Hence to show that (`w s0
(Y ), `qw1s1 (Y ))θ,p embeds into `pws (Y ) it
suffices to consider the case q0 = q1 = ∞. If y = y (0) +y (1) ∈ `∞ ∞
ws0 (Y )+`ws1 (Y ),
then
(0) (1)
kyk k 6 kyk k + kyk k 6 2−ks0 ky (0) k`∞
ws (Y ) + 2−ks1 ky (1) k`∞
ws (Y ) .
0 1

Multiplying with 2ks0 and taking the infimum over all admissible pairs
(y (0) , y (1) ), we find
2ks0 kyk k 6 K(2k(s0 −s1 ) , y)
using the notation of Section C.3. In combination with the identity θ(s1 −s0 ) =
s − s0 and the fact that the K-functional is non-decreasing, this gives
X 1/p
kyk`w
p
s
6 |2k(s−s0 ) K(2k(s0 −s1 ) , y)|p
k>0
XZ 2(k+1)(s0 −s1 )
dt 1/p
6 C0 |t−θ K(2k(s0 −s1 ) , y)|p
k(s0 −s1 ) t
k>0 2
Z ∞
−θ dt 1/p
6 C0 |t K(t, y)|p = C0 kyk(`qw0s (Y ),`qw1s (Y ))θ,p ,
0 t 0 1

1/p
(θp)
where C0 = (1−2−(s 0 −s)p )1/p
if p < ∞. A simple modification of this argument
gives the same result with C0 = 1 if p = ∞.
To prove the reverse inequality it suffices to consider the case q0 = q1 = 1.
Discretising as before, we find
XZ 2k(s0 −s1 )
dt 1/p
kyk(`1ws (Y ),`1ws (Y ))θ,p 6 |t−θ K(t, y)|p
0 1 (k−1)(s0 −s1 ) t
k>0 2
X 1/p
6 C1 |2−θk(s0 −s1 ) K(2k(s0 −s1 ) , y)|p ,
k>0

(s0 −s)p 1/p

−1)
where C1 = (2 (θp)1/p . If p = ∞ we consider the supremum norm in
the above and take C1 = 2s0 −s . Splitting ym = ym 1{m6k} + ym 1{m>k} , we
estimate
k
X ∞
X
K(2k(s0 −s1 ) , y) 6 2ms0 kym k + 2k(s0 −s1 ) 2s1 m kym k.
m=−∞ m=k+1
 14.4 Besov spaces 319

Therefore, since θ(s1 − s0 ) = s − s0 and (1 − θ)(s1 − s0 ) = s − s1 ,

2−θk(s0 −s1 ) K(2k(s0 −s1 ) , y)

k
X ∞
X
6 2(m−k)(s0 −s) 2ms kym k + 2−(m−k)(s−s1 ) 2ms kym k.
m=−∞ m=k+1

Taking `p -norms in k and using Young’s inequality for convolutions we obtain

X 1/p
|2−θk(s0 −s1 ) K(2k(s0 −s1 ) , y)|p 6 (C2 + C3 )kyk`w
p
s (Y )
.
k>0

P∞ P∞
where C2 = k=0 2−k(s0 −s) and C3 = k=1 2−k(s−s1 ) . This gives the inequal-
ity
kyk(`1ws 1
(Y ),`w s
(Y ))θ,p 6 C1 (C2 + C3 )kyk`w
p
s (Y )
.
0 1

The final assertion is immediate from the first assertion and the log-
convexity inequality (L.2). 

14.4 Besov spaces

The various Littlewood–Paley decompositions encountered in Chapter 5 ex-
press the norm of a function f ∈ Lp (Rd ; X) in terms of (sharp or smooth)
dyadic cut-offs in the frequency domain. For instance, in Theorem 5.5.22 we
have seen that if X is a UMD Banach space, p ∈ (1, ∞), and ψ is a smooth
Littlewood–Paley function,
X
kf kLp (Rd ;X) h εk ψk ∗ f p d
, (14.20)
L (Ω×R ;X)
k∈Z

where ψk (x) := 2k ψ(2k x) and (εk )k∈Z is a Rademacher sequence. With an eye
toward the ensuing discussion we also remark that we have an equivalence of
norms
X
kf kLp (Rd ;X) h ε k ϕk ∗ f , (14.21)
Lp (Ω×Rd ;X)
k∈N

where now (ϕk )k∈N is an inhomogeneous Littlewood–Paley sequence as in

(14.20). This follows from Theorem 14.7.5 below, but could already have been
proved in Chapter 5 with the methods presented there.
The idea behind the Littlewood–Paley approach to Besov spaces is to take
this representation as a starting point, introducing an additional smoothness
q
parameter s ∈ R, and trading the norm of the Rademacher sum for an `w s
-
sum. The possibility of having p 6= q presents us with two possible definitions,
q
utilising the spaces `w s
(Lp (Rd ; X)) and Lp (Rd ; `qws (X)) respectively. For p =
320 14 Function spaces

q, these spaces are canonically isometric by Fubini’s theorem. The two choices
lead to the theory of Besov spaces and Triebel–Lizorkin spaces, respectively.
q
The choice `w s
(Z) with the (homogeneous) Littlewood–Paley sequence
(ψk )k∈Z as in (14.20) leads to the so-called homogeneous Besov and Triebel–
Lizorkin spaces. Alternatively, the choice `qws (N) and the use of Littlewood–
Paley sequences (ϕk )k∈N as introduced in Definition 14.2.7 leads to the in-
homogeneous version of these spaces. In what follows we will only present in
the inhomogeneous case. Both classes of spaces are used in applications to
PDE. The advantage of inhomogeneous spaces is that, in the development
of their theory, one can make effective use of Schwartz functions and tem-
pered distributions. The theory of homogeneous spaces is technically more
involved and requires the use of different classes of test functions and equiv-
alence classes of tempered distributions modulo polynomials. Since we have
already encountered Schwartz functions and tempered distributions in many
places, we choose to only develop the theory of inhomogeneous spaces here.
Homogeneous spaces have better scaling properties, and scaling often plays
a crucial role in PDE, but for the purposes of the theory developed here
homogeneous spaces are not essential.
The proofs of (14.20) and (14.21) require the Banach space X to be UMD.
In contrast, in the theory of Besov and Triebel–Lizorkin spaces these norm
equivalences are promoted to definitions, thus eliminating the need of impos-
ing any conditions on X. By taking this approach, most of the fundamental
results in the theory of Besov spaces and Triebel–Lizorkin spaces are true
for arbitrary Banach spaces X. They come with their own versions of the
Mihlin multiplier theorem which does not require the UMD property either,
allowing multipliers without singularities at the origin in case of inhomo-
geneous spaces. The more general multipliers considered in Chapter 5 have
corresponding versions for homogeneous Besov and Triebel–Lizorkin spaces.
Perhaps more surprising is the fact that also for the duality theory of these
spaces no geometrical conditions need to be imposed on X. This contrast
the duality theory for the Bochner spaces, which requires that X ∗ have the
Radon–Nikodým property.

14.4.a Definitions and basic properties

As anticipated in the above discussion, we now introduce scale of Besov spaces

through a Littlewood–Paley decomposition.
s
Definition 14.4.1. Let p, q ∈ [1, ∞] and s ∈ R. The Besov space Bp,q (Rd ; X)
is the space of all f ∈ S 0 (Rd ; X) for which ϕk ∗ f ∈ Lp (Rd ; X) for all k > 0
and for which the quantity

kf kBp,q
s (Rd ;X) := (2ks ϕk ∗ f )k>0 `q (Lp (Rd ;X))

is finite.
 14.4 Besov spaces 321

Here, (ϕk )k>0 is the inhomogeneous Littlewood–Paley sequence that has been
fixed throughout the chapter (see Convention 14.2.8). By the discussion of
(14.13), the tempered distribution ϕk ∗ f is a C ∞ -function of polynomial
growth, so that the Lp -norm in the above definition makes sense.
s (Rd ;X) is indeed a norm, suppose that kf kB s (Rd ;X) = 0.
To see that k·kBp,q p,q

Then ϕ bk f = F (ϕk ∗ f ) = 0 for all k > 0, so ϕk ∗ f = 0 for all k > 0, and

b
therefore f = 0 by Lemma 14.2.10. All other properties of a norm can be
deduced from the fact that k · k`q (Lp (Rd ;X)) is a norm.
It is immediate from Young’s inequality, applied term-wise with respect
to the `q -sum, that ψ ∗ f ∈ Bp,q s
(Rd ; X) whenever ψ ∈ L1 (Rd ) and f ∈
s d
Bp,q (R ; X), and more generally the analogue of Proposition 14.2.3 is valid.
Up to an equivalent norm the above definition is independent on the choice
of the sequence (ϕk )k>0 , as will be shown in Proposition 14.4.2.
From the continuous embedding `q0 ,→ `q1 for 1 6 q0 6 q1 6 ∞ we obtain
the continuous embedding
s
Bp,q 0
s
(Rd ; X) ,→ Bp,q 1
(Rd ; X). (14.22)

For 1 6 q0 , q1 6 ∞ and s0 > s1 we have the continuous embedding

s0
Bp,q 0
s1
(Rd ; X) ,→ Bp,q 1
(Rd ; X). (14.23)

Indeed, for q0 6 q1 this follows from (14.22) and the inequality 2ks0 6 2ks1
for k > 0. For q0 > q1 this follows from Hölder’s inequality with q11 = q10 + 1r
and using that k>0 2−k(s0 −s1 )r < ∞.
P

Proposition 14.4.2. For all p, q ∈ [1, ∞] and s ∈ R, up to an equivalent

s
norm the space Bp,q (Rd ; X) is independent of the choice of inhomogeneous
Littlewood–Paley sequence (ϕk )k>0 .

The proof will give explicit constants depending only on s and ϕ0 (in one
direction), respectively s and ψ0 (in the other direction).
Proof. Suppose (ψk )k>0 is another inhomogeneous Littlewood–Paley seq-
uence. Then the analogues of (14.10) and (14.11) hold with ϕj and ψk ; in
particular for all j, k > 0 with |j − k| > 2 we have ϕk ∗ ψj = 0. Using
(14.12) for the sequence (ψk )k>0 , the triangle inequality, Young’s inequality,
and (14.7), we obtain

(2ks ϕk ∗ f )k>0 `q (Lp (Rd ;X))

1
X
6 (2ks ϕk ∗ ψk+j ∗ f )k>0 `q (Lp (Rd ;X))
j=−1
1
X
6 kϕk k1 2|s| (2(k+j)s ψk+j ∗ f )k>0 `q (Lp (Rd ;X))
j=−1
322 14 Function spaces

6 6kϕ0 k1 2|s| (2ks ψk ∗ f )k>0 `q (Lp (Rd ;X))

,

where we used (14.7). This gives the required estimate in one direction. The
reverse estimate is obtained by reversing the rôles of ϕk and ψk . 

Proposition 14.4.3. For all p, q ∈ [1, ∞] and s ∈ R we have continuous

embeddings
S (Rd ; X) ,→ Bp,q
s
(Rd ; X) ,→ S 0 (Rd ; X).
Moreover, if 1 6 p, q < ∞, then Cc∞ (Rd ) ⊗ X is dense in Bp,q
s
(Rd ; X).

Proof. We split the proof into three steps.

Step 1 – For the first embedding, by (14.22) it is enough to prove that
S (Rd ; X) embeds into Bp,1s
(Rd ; X). For f ∈ S (Rd ; X) and L = Lp,d ∈ N so
2L −1
large that (1 + |2π · | ) ∈ Lp (Rd ) we find
X
kf kBp,1
s (Rd ;X) = 2ks kϕk ∗ f kLp (Rd ;X)
k>0
X
.d,p 2ks (1 + |2π · |2L )ϕk ∗ f L∞ (Rd ;X)
k>0
X
6 2ks (1 + (−∆)L )(ϕ
bk fb) L1 (Rd ;X)
,
k>0

where we used the fact that F −1 maps L1 into L∞ . It remains to estimate

2ks k∂ α (ϕ
bk fb)kL1 (Rd ;X) for multi-indices |α| 6 2L.
First consider k > 1. Then supp ϕk ⊆ Bk := {ξ ∈ Rd : 2k−1 6 |ξ| 6 3 · 2k }
and |Bk | .d 2kd . By Leibniz’ rule and the boundedness on Bk of the functions
∂γ ϕ
bk with |γ| 6 |α| 6 2L = 2Lp,d ,
X
k∂ α (ϕbk fb)kL1 (Rd ;X) .d,p k1Bk ∂ β fbkL1 (Rd ;X) .
|β|6|α|

To estimate the terms on the right-hand side, fix an M ∈ N which is arbitrary

for the moment. Then

k1Bk ∂ β fbkL1 (Rd ;X) 6 k1Bk (1 + | · |2M )−1 kL1 (Rd ) k(1 + | · |2M )∂ β fbkL∞ (Rd ;X)
X
6 |Bk |(1 + 22M (k−1) )−1 [fb]β,δ ,
|δ|62M

using the notation (14.2) for the seminorms defining the Schwartz space. Keep-
ing
P in mind that |Bk | .d 2kd we now choose M = Ms,p,d ∈ N so large that
ks kd 2M (k−1) −1
k>0 2 2 (1 + 2 ) < ∞. With this choice, we obtain the estimate
X
kf kBp,1
s (Rd ;X) .d,p,s [fb]β,δ .
|δ|62M
 14.4 Besov spaces 323

A similar estimate in the case k = 0 can be obtained in a similar, but simpler,

way. Since F is continuous on S (Rd ; X) (see Proposition 2.4.22), this proves
that we have a continuous embedding S (Rd ; X) ,→ Bp,q s
(Rd ; X).
Step 2 – Next we prove that Bp,q s
(Rd ; X) embeds into S 0 (Rd ; X). By
s
(14.22) it is enough to prove that the inclusion mapping Bp,∞ (Rd ; X) ⊆
S 0 (Rd ; X) (by definition Bp,∞
s
(Rd ; X) is contained in S 0 (Rd ; X)) is continu-
ous.
Fix f ∈ Bp,∞s
(Rd ; X) and ψ ∈ S (Rd ), and set fk := ϕk ∗ f and ψk :=
ϕk ∗ ψ. By Lemma 14.2.10 and (14.10) we have

X 1 X
X
f (ψ) = fk (ψ` ) = fk (ψk+` ).
k,`>0 `=−1 k>0

Thus, by (14.13),
1 XZ
X
kf (ψ)k 6 kfk (x)k|ψk+` (x)| dx
`=−1 k>0 Rd

1
X
6 (2ks kfk (·)k)k>0 `∞ (Lp (Rd ;X))
(2−ks ψk+` )k>0 `1 (Lp0 (Rd ))
`=−1

6 3 · 2|s| kf kBp,∞
s (Rd ;X) kψkB −s
0 (Rd ) .
p ,1

Since S (Rd ) ,→ Bp−s d

0 ,1 (R ) continuously by the previous step, the result follows

from this.
Step 3 – To prove density, by Lemma 14.2.1 it suffices to prove the density
of S (Rd ; X) in Bp,q
s
(Rd ; X).
s d
Pn
Fix f ∈ Bp,q (R ; X) and set ζn := k=0 ϕk . By (14.6) we have kζn k1 =
kϕk1 .
s
We will first show that ζn ∗ f → f in Bp,q (Rd ; X). Fix ε > 0 and choose
K ∈ N such that X q
2ksq kϕk ∗ f kL q
p (Rd ;X) < ε .

k>K

By Young’s inequality combined with the identity kζn k1 = kϕ0 k1 we have

ζn ∗ ϕk ∗ f ∈ Lp (Rd ; X) and kζn ∗ ϕk ∗ f kLp (Rd ;X) 6 kϕk1 kϕk ∗ f kLp (Rd ;X) .
s
From this we infer that ζn ∗ f ∈ Bp,q (Rd ; X) and
X q q
2ksq kζn ∗ ϕk ∗ f kL q
p (Rd ;X) < ε kϕk1 .

k>K

Hence by the triangle inequality in `q (Lp (Rd ; X)),

kf − ζn ∗ f kBp,q
s (Rd ;X)
324 14 Function spaces
X 1/q
= 2ksq kϕk ∗ (f − ζn ∗ f )kqLp (Rd ;X)
k>0
K
X 1/q
6 2ksq kϕk ∗ (f − ζn ∗ f )kqLp (Rd ;X) + ε(1 + kϕk1 ).
k=0

The first term in the last expression tends to zero R as n → ∞ by Proposition

1.2.32; here we use that ζn = 2nd ϕ(2n ·) and Rd ϕ dx = ϕ(0) b = 1. This
s
concludes the proof that ζn ∗ f → f in Bp,q (Rd ; X).
It remains to approximate each of the functions fn = ζn ∗ f by elements in
S (Rd ; X). Observe that fn ∈ Lp (Rd ; X) since the functions ϕk ∗ f belong to
Lp (Rd ; X). Let η ∈ S (Rd ) be a functions satisfying η(0) = 1 and supp(b η) ⊆
{ξ ∈ Rd : |ξ| 6 1}. Since F (η(δ·)) = δ −d ηb(δ −1 ·), for all δ ∈ (0, 1) the support
of F (η(δ·)fn ) is contained in a ball of radius 3 · 2n−1 + 1 6 2n+1 ; here we use
the definition of ζn and (14.9). Using (14.11), (14.7), and Young’s inequality,
it follows that
 n+2
X 1/q
kfn − η(δ·)fn kBp,q
s (Rd ;X) = 2ksq kϕk ∗ (fn − η(δ·)fn )kqLp (Rd ;X)
k=0
6 Ckfn − η(δ·)fn kLp (Rd ;X) ,
Pn
where C = Cn,s,q = ( k=0 2ksq )1/q . For each fixed n, the right-hand side
tends to zero as δ ↓ 0 by the dominated convergence theorem. 
s
Next we will prove the completeness of the normed space Bp,q (Rd ; X).
s
Proposition 14.4.4. For p, q ∈ [1, ∞] and s ∈ R the space Bp,q (Rd ; X) is a
Banach space.

The proof requires some preparations. Recall that a sequence (fn )n>1 is said
to converge in S 0 (Rd ; X) if there exists an f ∈ S 0 (Rd ; X) such that fn (φ) →
f (φ) in X for all φ ∈ S (Rd ). Likewise, it is said to be Cauchy in S 0 (Rd ; X)
if (fn (φ))n>1 is a Cauchy sequence in X for all φ ∈ S (Rd ).

Lemma 14.4.5. The space S 0 (Rd ; X) is sequentially complete, i.e., every

Cauchy sequence in S 0 (Rd ; X) is convergent in S 0 (Rd ; X).

Proof. Let (fn )n>1 be a Cauchy sequence in S 0 (Rd ; X). Since X is complete
we may define a linear mapping f : S (Rd ) → X by f (φ) := limn→∞ fn (φ).
We claim that f is continuous. Indeed, for every φ ∈ S (Rd ) the sequence
(fn (φ))n>1 is bounded in X, and therefore the Banach–Steinhaus theorem for
topological vector spaces implies that the sequence (fn )n>1 is equicontinuous.
Hence, given an ε > 0, we can find an open neighbourhood V of 0 in S (Rd )
such that |fn (φ)| 6 ε for all φ ∈ V and n > 1. Taking limits, it follows that
|f (φ)| 6 ε for all φ ∈ V . This means that f is continuous at zero and hence
continuous. 
 14.4 Besov spaces 325

A normed space E ,→ S 0 (Rd ; X) is said to have the Fatou property if for all
sequences (fn )n>1 in E such that

fn → f in S 0 (Rd ; X) and lim inf kfn kE < ∞

n→∞

we have f ∈ E and kf kE 6 lim inf n→∞ kfn kE .

s
Lemma 14.4.6. For all p, q ∈ [1, ∞] and s ∈ R the space Bp,q (Rd ; X) has the
Fatou property.
s
Proof. Choose a sequence (fn )n>1 of functions from Bp,q (Rd ; X) with

fn → f in S 0 (Rd ; X) and lim inf kfn kBp,q

s (Rd ;X) < ∞.
n→∞

Then limn→∞ ϕk ∗ fn = ϕk ∗ f pointwise. In case p < ∞, Fatou’s lemma gives

kϕk ∗ f kLp (Rd ;X) 6 lim inf kϕk ∗ fn kLp (Rd ;X) < ∞.
n→∞

Multiplying with 2ks and taking `q -norms, it follows that we have f ∈

s
Bp,q (Rd ; X) and kf kBp,q
s (Rd ;X) 6 lim inf n→∞ kfn kB s (Rd ;X) (by Fatou’s lemma
p,q
if q < ∞ and directly if q = ∞). For p = ∞ the proof is similar. 

Lemma 14.4.7. Every normed space E ,→ S 0 (Rd ; X) with the Fatou prop-
erty is complete.

Proof. Let (fn )n>1 be a Cauchy sequence in E. Since S 0 (Rd ; X) is se-

quentially complete by Lemma 14.4.5, and E is continuously embedded in
S 0 (Rd ; X), it follows that there exists an f ∈ S 0 (Rd ; X) such that fn → f
in S 0 (Rd ; X). Since (fn )n>1 is a Cauchy sequence in E it is bounded in E.
By the Fatou property of E it follows that f ∈ E. To prove that fn → f in
E we fix an ε > 0 and choose N ∈ N such that for all n, m > N we have
kfm − fn kE < ε. Using the Fatou property once more, we obtain

kf − fn kE 6 lim inf kfm − fn kE 6 ε

m→∞

and the result follows. 

Proof of Proposition 14.4.4. Combine Lemmas 14.4.6 and 14.4.7 and Propo-
sition 14.4.3. 
Coming back to the discussion on homogeneous verses inhomogeneous norms
(see (14.20) and (14.21)), we have the following remark.
Remark 14.4.8. Let p, q ∈ [1, ∞] and s > 0. For f ∈ S 0 (Rd ; X) one has

(2ks ϕk ∗ f )k>0 `q (Lp (Rd ;X))

h (2ks ψk ∗ f )k∈Z `q (Lp (Rd ;X))
+ kf kLp (Rd ;X) ,
326 14 Function spaces

where both expressions are infinite whenever one of them is. Here the (ϕk )k>0
are as in Definition 14.4.1, and thus the left-hand side of the above identity
equals kf kBp,q
s (Rd ;X) . The (ψk )k∈Z are as in (14.20). The first expression on

the right-hand side is equal to the homogeneous Besov norm, which we will
not discuss in detail.
To prove the norm equivalence first recall that ψk = ϕk for k > 1. For “.”
it suffices to observe that by Young’s inequality

ϕ0 ∗ f Lp (Rd ;X)
6 kϕ0 k1 kf kLp (Rd ;X) .

s
Conversely, assume that f ∈ Bp,q (Rd ; X). Since ϕ
b0 = 1 on supp(ψbk ) for k 6 0,
we can write

kψk ∗ f kLp (Rd ;X) = kψk ∗ ϕ0 ∗ f kLp (Rd ;X)

6 kψk k1 kϕ0 ∗ f kLp (Rd ;X) = kψ0 k1 kϕ0 ∗ f kLp (Rd ;X) ,

and thus using that s > 0 we obtain

(2ks ψk ∗ f )k60 `q (Lp (Rd ;X))

6 (2ks ϕ0 ∗ f )k60 `q (Lp (Rd ;X))

6 Cs kϕ0 ∗ f kLp (Rd ;X) .

s
Moreover, since s > 0, from (14.23) Bp,q (Rd ; X) ,→ Bp,1
0
(Rd ; X), and thus by
Lemma 14.2.10
X X
kf kLp (Rd ;X) = ϕk ∗ f p d 6 kϕk ∗ f kLp (Rd ;X)
L (R ;X)
k>0 k>0

= kf kBp,1
0 (Rd ;X) 6 Cs,q kf kB s (Rd ;X)
p,q

14.4.b Fourier multipliers

The goal of this section is to prove a version of the Mihlin multiplier theorem
for operator-valued Fourier multipliers acting on vector-valued Besov spaces.
In contrast to the situation in the Lp -setting (cf. Theorems 5.3.18 and 5.5.10),
where we had to assume the UMD property, a variant of the Mihlin theorem
for Besov spaces holds for arbitrary Banach spaces.
We wish to emphasise that the main result, Theorem 14.4.16 below, is
not applicable to multipliers which are non-smooth or even singular near the
origin. This is due to the presence of the term ϕ0 in the definition of in-
homogeneous Littlewood–Paley sequences, whose support contains the origin
in its interior. For instance, the Fourier multiplier associated to the Hilbert
transform does not satisfy the conditions of the theorem.
Unlike in other chapters, we also include the case p = ∞. In order to avoid
density issues, we define ML∞ (Rd ; X, Y ) as the space of Fourier transforms
of operator-valued measures of bounded variation:
 14.4 Besov spaces 327

Definition 14.4.9. We define

ML∞ (Rd ; X, Y ) := Φ

b : the operator-valued measure
Φ : B(Rd ) → L (X, Y ) is of bounded variation .

b ML∞ (Rd ;X,Y ) = kΦk(Rd ), the space ML∞ (Rd ; X, Y ) is a

With the norm kΦk
Banach space.
For m ∈ ML∞ (Rd ; X, Y ) and f ∈ L∞ (Rd ; X) we define
b
Tm ∗ f := m ∗ f,

recalling that the convolutions with an operator-valued measure of bounded

variation has been introduced in Lemma 14.2.4.

Remark 14.4.10. In the scalar case it can be shown that the space ML∞ (Rd ) =
ML∞ (Rd ; C, C) as defined in Definition 14.4.9 coincides with the space of all
m ∈ L∞ (Rd ) for which the quantity

sup{kTm f k∞ : f ∈ S (Rd ) with kf k∞ 6 1}

is finite, and that this quantity then equals the norm on ML∞ (Rd ) introduced
above. This provides further motivation for Definition 14.4.9.

Various properties discussed in Section 5.3.a extend to p = ∞. Moreover, from

the definition of the Fourier transform one sees that
b L∞ (Rd ;L (X,Y )) 6 kΦk(Rd ).
kΦk

This induces a contractive embedding

ML∞ (Rd ; X, Y ) ,→ L∞ (Rd ; L (X, Y )).

b
For m ∈ ML∞ (Rd ; X, Y ) and f ∈ S (Rd ; X) one can check that mfb = F (m ∗
f ), and by Lemma 14.2.4 for all p ∈ [1, ∞] we have
b b
km ∗ f kLp (Rd ;Y ) 6 kmk(Rd )kf kLp (Rd ;Y ) .

This shows that for all p ∈ [1, ∞] we have a contractive embedding

ML∞ (Rd ; X, Y ) ,→ MLp (Rd ; X, Y ). (14.24)

In the discussion preceding Lemma 14.2.4 it was observed that for any function
φ ∈ L1 (Rd ; L (X, Y )), an operator-valued measure Φ : B(Rd ) → L (X, Y ) of
bounded variation is obtained by setting
Z
Φ(A) := φ dx,
A
328 14 Function spaces

and that its total variation satisfies kΦk(Rd ) 6 kφkL1 (Rd ;L (X,Y )) . In this way
we obtain contractive embeddings
b
L1 (Rd ; L (X, Y )) ,→ ML∞ (Rd ; X, Y ) ,→ MLp (Rd ; X, Y ).

In combination with Lemma 14.2.11 we now obtain the following sufficient

condition on m for membership of MLp (Rd ; X, Y ).
b
Proposition 14.4.11. If the multiplier m ∈ L∞ (Rd ; L (X, Y )) satisfies m ∈
L1 (Rd ; L (X, Y )), then for all p ∈ [1, ∞] we have m ∈ MLp (Rd ; X, Y ) and
b
kmkMLp (Rd ;X,Y ) 6 kmkL1 (Rd ;L (X,Y )) .

In particular, if m ∈ C d+1 (Rd ; L (X, Y )) and there exists an ε > 0 such that

Cm,d,ε := max sup (1 + |ξ||α|+ε )k∂ α m(ξ)k < ∞,

|α|6d+1 ξ∈Rd

then m ∈ MLp (Rd ; X, Y ) and kmkMLp (Rd ;X,Y ) .d,ε Cm,d,ε .

Remark 14.4.12. In applications it can be useful to apply Proposition 14.2.11

to a dilated multiplier m(t·) instead of m(·). The MLp (Rd ; X, Y )-norm is
invariant under dilations, but the expression for Cm,dε is not. A similar remark
applies to Lemma 14.2.11.

Remark 14.4.13. If m ∈ Ccd+1 (Rd ; L (X, Y )) is supported in the ball BR

around the origin, one easily checks that Cm,d,ε .R kmkC d+1 (Rd ;L (X,Y )) .
b
As a consequence we obtain that every m ∈ Ccd+1 (Rd ; L (X, Y )) belongs to
MLp (Rd ; X, Y ) and kmkMLp (Rd ;X,Y ) .d,ε,R kmkC d+1 (Rd ;L (X,Y )) .
b

Remark 14.4.14. Multipliers with singularities in the origin, such as the mul-
tiplier giving rise to the Hilbert transform, are not covered by Proposition
14.4.11.

Before moving to a Mihlin multiplier theorem for Besov space we present an

important result on lifting operators. Recall from Subsection 5.6.a that the
Bessel potential operators are the continuous operators Jσ , σ ∈ R, acting on
S 0 (Rd ; X) by
b
b) , u ∈ S 0 (Rd ; X).
Jσ u := ((1 + 4π 2 | · |2 )σ/2 u

They satisfy J0 = I and Jσ1 +σ2 = Jσ1 ◦ Jσ2 .

Proposition 14.4.15 (Lifting). Let p, q ∈ [1, ∞] and s ∈ R. For all σ ∈ R
we have
s
Jσ : Bp,q s−σ
(Rd ; X) ' Bp,q (Rd ; X) isomorphically.
 14.4 Besov spaces 329

Proof. Noting that Jσ is a bijection from S 0 (Rd ; X) to S 0 (Rd ; X), with

inverse Jσ−1 = J−σ , it suffices to prove that Jσ maps Bp,q s
(Rd ; X) into
s−σ d
Bp,q (R ; X) and is bounded for each σ ∈ R.
We claim that there exists a constant C > 0, independent of k > 0, such
that for all f ∈ S 0 (Rd ; X),

kϕk ∗ Jσ f kLp (Rd ;X) 6 C2kσ kϕk ∗ f kLp (Rd ;X) .

This will imply the result. P1

To prove the claim we use that `=−1 ϕ bk+` ≡ 1 on the support of ϕ
bk to
write
X1
2−kσ Jσ ϕk ∗ f = F −1 (ϕ bk+` fb),
bk mϕ
`=−1
−kσ 2 2 σ/2
where m(ξ) = 2 (1 + 4π |ξ| ) . Using a dilation, Proposition 14.4.11, and
the Fourier support property (14.9), for k > 1 we obtain

kϕk mkMLp (Rd ;X) = kϕ1 (2·)m(2k ·)kMLp (Rd ;X)

.d max sup (1 + |ξ||α|+1 )k∂ α [ϕ1 (2·)m(2k ξ)](ξ)k
|α|6d+1 ξ∈Rd

.d max sup k∂ α [m(2k ·)](ξ)k,

|α|6d+1 1 3
2 6|ξ|6 2

where in the last step we applied the Leibniz rule as before and the Fourier
support properties of ϕ1 given by (14.8) and (14.9). Since m(2k ξ) = (2−2k +
|ξ|2 )σ/2 , it is elementary to check that the latter expression is uniformly
bounded in k > 1. A similar argument shows that ϕ0 m ∈ MLp (Rd ; X). 
The simple multiplier result of Proposition 14.4.11 is already strong enough to
s
prove the version of Mihlin’s multiplier theorem for Besov spaces Bp,q (Rd ; X)
contained in Theorem 14.4.16 below, valid for arbitrary Banach spaces X and
integrability exponents p, q ∈ [1, ∞]. In the statement of the theorem the end-
points p = ∞ and q = ∞ create some technical difficulties, since we cannot use
the density of the Schwartz functions to define Tm . It is for this reason that in
the theorem we assume that the multiplier m is smooth and has derivatives
of polynomial growth. Many interesting multipliers satisfy this condition, and
to proceed with the development of the theory this version suffices for the
time being. A version which avoids this restriction on m will be presented in
Theorem 14.5.6.
When m ∈ C ∞ (Rd ; L (X, Y )) has derivatives of polynomial growth, one
can define the Fourier multiplier Tm as an operator from S 0 (Rd ; X) into
S 0 (Rd ; Y ) by Tm f := F −1 (mfb). To see that this is well-defined it suffices
to note that mfb ∈ S 0 (Rd ; Y ) for f ∈ S 0 (R; X). In the next theorem, Tm is
s
understood to be the restriction of this operator to Bp,q (Rd ; X). The theorem
s
then asserts that, under Mihlin type conditions on m, it maps Bp,q (Rd ; X)
s d
into Bp,q (R ; Y ).
330 14 Function spaces

Theorem 14.4.16 (Mihlin multiplier theorem for Besov spaces). Let

X and Y be Banach spaces and let p, q ∈ [1, ∞] and s ∈ R. Suppose that
m ∈ C ∞ (Rd ; L (X, Y )) has derivatives of polynomial growth, and that

sup sup (1 + |ξ||α| )k∂ α m(ξ)kL (X,Y ) =: Km < ∞. (14.25)

|α|6d+1 ξ∈Rd

Then the Fourier multiplier Tm = F −1 mF restricts to a bounded operator

s
from Bp,q s
(Rd ; X) to Bp,q (Rd ; Y ) of norm kTm k 6 Cs,d Km .

The usual Mihlin condition involves a factor |ξ||α| instead of 1 + |ξ||α| . A

multiplier theorem involving the former can be shown to hold for the scale of
homogeneous Besov spaces.
For finite p and q, the condition m ∈ C ∞ (Rd ; L (X, Y )) can be weakened
to m ∈ C d+1 (Rd ; L (X, Y )). This can be seen by taking f in the dense class
S (Rd ) ⊗ X in the proof below.
s
P1
Proof. For f ∈ Bp,q (Rd ; X) let fn := ϕn ∗ f for n > 0. Since `=−1 ϕ
bk+` ≡ 1
on the support of ϕ bk ,

kTm f kBp,q
s (Rd ;Y ) = (2ks ϕk ∗ F −1 mfb)k>0 `q (Lp (Rd ;Y ))
 1
X 
= 2ks F −1 ϕ
bk m ϕ
bk+` fb
k>0 `q (Lp (Rd ;Y ))
`=−1
1
X
6 2ks F −1 (ϕ
bk mfbk+` )k>0 `q (Lp (Rd ;Y ))
`=−1
1
X
6 sup kϕ
bk mkMLp (Rd ;X,Y ) (2ks fk+` )n>0 `q (Lp (Rd ;Y ))
k>0
`=−1

6 2|s| sup kϕ
bk mkMLp (Rd ;X,Y ) kf kBp,q
s (Rd ;X) .
k>0

To complete the proof we must show that supk>0 kϕ bk mkMLp (Rd ;X,Y ) .d Km .
First consider the case k > 1. Since the multiplier norm is invariant under
dilations by Proposition 5.3.8, it suffices to show that

b1 (·)m(2k−1 ·)kMLp (Rd ;X,Y ) .d Km .

sup kϕ
k>1

By Proposition 14.4.11, it even suffices to show that there exists an ε > 0

such that

max sup (1 + |ξ||α|+ε )k∂ α [ϕ

b1 (·)m(2k−1 ·)](ξ)k .d Km .
|α|6d+1 ξ∈Rd

We will verify this bound for ε = 1. By the Fourier support properties of ϕ1

implied by (14.8) and (14.9), for β 6 α with |α| 6 d + 1 we have
 14.4 Besov spaces 331

sup |(1 + |ξ||α|+1 )∂ β ϕ

b1 (ξ)| 6 Cβ,d .
ξ∈Rd

Hence, by Leibniz’s rule the Mihlin condition on m, and the Fourier support
property of ϕ1 given by (14.9), for all |α| 6 d + 1 we have

sup (1 + |ξ||α|+1 )k∂ α [ϕ

b1 (·)m(2k−1 ·)](ξ)k
ξ∈Rd

= sup (1 + |ξ||α|+1 )k∂ α [ϕ

b1 (·)m(2k−1 ·)](ξ)k
|ξ|>1
X
6 sup (1 + |ξ||α|+1 ) Cα,β |∂ β ϕ1 (ξ)| · 2(k−1)|α−β| |∂ α−β m(2k−1 ξ)|
|ξ|>1 β6α
X
.d sup 2(k−1)|α−β| |∂ α−β m(2k−1 ξ)|
|ξ|>1 β6α
X Km
6 sup 2(k−1)|α−β|
|ξ|>1 β6α 1 + |2k−1 ξ||α−β|
. d Km .
(14.26)

The case k = 0 is proved in similarly, omitting the dilation argument. 

As an application of Theorem 14.4.16, we obtain the following analogue of
Theorem 5.6.11.
Proposition 14.4.17. Let p ∈ [1, ∞), q ∈ [1, ∞], and s ∈ R. For all k ∈ N,
X
|||f |||Bp,q
s (Rd ;X) := k∂ α f kBp,q
s−k
(Rd ;X) (14.27)
|α|6k

s
defines an equivalent norm on Bp,q (Rd ; X)
Proof. As a consequence of Proposition 14.4.15 it suffices to prove the equiv-
alence of (14.27) with kJk f kBp,qs−k
(Rd ;X) . This can be deduced from Theorem
14.4.16 by an argument similar to the one in Theorem 5.6.11. In the present
situation it is important to note that the multipliers in the proof of the propo-
sition also satisfy the more restrictive condition (14.25). Below we present a
simplification of the argument of Theorem 5.6.11 adapted to the Besov space
case. Let hξi = (1 + |2πξ|2 )1/2 .
First we prove the estimate

k∂ α f kBp,q
s−k
(Rd ;X) 6 CkJk f kBp,q
s−k
(Rd ;X) .

Applying the Fourier transform, we have

(2πiξ)α k b
F [∂ α f ](ξ) = (2πiξ)α fb(ξ) = hξi f (ξ) =: mα (ξ)hξik fb(ξ).
hξik
332 14 Function spaces

One checks that mα satisfies the conditions of Theorem 14.4.16, and thus
−1
k∂ α f kBp,q
s−k
(Rd ;X) 6 Cα Cd,p,q kF [h·ik fb]kBp,q
s−k
(Rd ;X)

= Cα Cd,p,q kJk f kBp,q

s−k
(Rd ;X) .

For the reverse estimate it suffices to show

X
kJk f kBp,q
s−k
(Rd ;X) 6 C k∂ α f kBp,q
s−k
(Rd ;X) .
|α|6k

Again we apply Theorem 14.4.16. By induction on k,

X
hξi2k = (1 + |2πξ|2 )k = cα,k (2πiξ)α (2πiξ)α ,
|α|6k

and therefore
hξi2k b X
hξik fb(ξ) = f (ξ) = cα,k mα (ξ)(2πiξ)α fb(ξ)
hξik
|α|6k
X
= α f (ξ),
cα,k mα (ξ)∂d
|α|6k

(2πiξ)α
where mα (ξ) = hξik
. Applying Theorem 14.4.16 to mα now gives

−1
kJk f kBp,q
s−k
(Rd ;X) = kF [h·ik fb]kBp,q
s−k
(Rd ;X)
X
6 |cα,k |kTmα ∂ α f kBp,qs−k
(Rd ;X)
|α|6k
X
6 Cd,p,k k∂ α f kBp,q
s−k
(Rd ;X) .
|α|6k

14.4.c Embedding theorems

We begin by showing that various classes of function spaces lie ‘sandwiched’

between Besov spaces.

Proposition 14.4.18 (Sandwiching with Besov spaces). For all p ∈

[1, ∞], s ∈ R, and m ∈ N, we have continuous embeddings
s
Bp,1 (Rd ; X) ,→ H s,p (Rd ; X) ,→ Bp,∞
s
(Rd ; X), (14.28)
m
Bp,1 (Rd ; X) ,→ W m,p d m
(R ; X) ,→ Bp,∞ (Rd ; X), (14.29)
m
B∞,1 (Rd ; X) ,→ m
Cub (Rd ; X) ,→ B∞,∞
m
(Rd ; X). (14.30)
 14.4 Besov spaces 333

An improvement for p ∈ (1, ∞) will be given in Proposition 14.6.13.

Proof. In order to prove (14.28), by Propositions 5.6.3 and 14.4.15 it suffices
to consider s = m = 0. Similarly, in order to prove (14.29) and (14.30), by
Proposition 14.4.17 it suffices to consider s = m = 0. Therefore, (14.28) and
(14.29) reduce to proving the continuous embeddings
0
Bp,1 (Rd ; X) ,→ Lp (Rd ; X) ,→ Bp,∞
0
(Rd ; X). (14.31)
0
Fix f ∈ Bp,1 (Rd ; X). By definition,
X
kf kBp,1
0 (Rd ;X) = kϕk ∗ f kLp (Rd ;X) .
k>0

In particular, the sum k>0 ϕk ∗ f converges absolutely in Lp (Rd ; X), and

P
the required result follows by Lemma 14.2.10 and the triangle inequality.
To prove the second embedding in (14.31), fix f ∈ Lp (Rd ; X). By Young’s
inequality,

kf kBp,∞
0 (Rd ;X) = sup kϕk ∗ f kLp (Rd ;X)
k>0

6 sup kϕk kL1 (Rd ) kf kLp (Rd ;X) 6 2kϕ0 kL1 (Rd ) kf kLp (Rd ;X) ,
k>0

where the last step uses (14.7). This completes the proof of (14.31).
As we already noted, in order to prove the embeddings in (14.30) it suffices
0
to consider the case m = 0. Fix f ∈ B∞,1 (Rd ; X). As before we see that the
P∞
sum k=0 ϕk ∗ f is absolutely convergent in L∞ (Rd ; X). By Lemma 14.2.10
its sum equals f and
∞
X
kf k∞ 6 kϕk ∗ f k∞ = kf kB∞,1
0 (Rd ;X) .
k=0

To see that f has a uniformly continuous version, we note that by Proposition

2.4.32 we have ϕk ∗ f ∈ C ∞ (Rd ; X) and

k∂j (ϕk ∗ f )k∞ = k(∂j ϕk ) ∗ f k∞ 6 k∂j ϕk k1 kf k∞ 6 k∂j ϕk k1 kf kB∞,1

0 (Rd ;X) .

In particular, each function ϕk ∗f is Lipschitz continuous and hence uniformly

continuous. Therefore f ∈ Cub (Rd ; X) by uniform convergence.
The second embedding in (14.30) follows by combining the embedding
m
Cub (Rd ; X) ,→ W m,∞ (Rd ; X) and (14.29). 
Theorem 14.4.19 (Sobolev embedding for Besov spaces). For given
p0 , p1 , q0 , q1 ∈ [1, ∞], and s0 , s1 ∈ R, we have a continuous embedding

Bps00 ,q0 (Rd ; X) ,→ Bps11 ,q1 (Rd ; X)

if and only if one of the following three conditions holds:

334 14 Function spaces

(i) p0 = p1 and [s0 > s1 or (s0 = s1 and q0 6 q1 )];

(ii) p0 < p1 , q0 6 q1 , and s0 − pd0 = s1 − pd1 ;
(iii) p0 < p1 and s0 − pd0 > s1 − pd1 .

The most interesting cases are (ii) and (iii), since they can be used to change
the integrability parameter from p0 into p1 .
For the proof of the sufficiency of the three conditions we need two lemmas.
The first provides an Lp -estimate for the derivatives under suitable Fourier
support assumptions. Recall from Lemma 14.2.9 that every f ∈ S 0 (Rd ; X)
with compact Fourier support belongs to C ∞ (Rd ; X) and has at most poly-
nomial growth.

Lemma 14.4.20 (Bernstein–Nikolskii inequality). Let p0 , p1 ∈ [1, ∞]

satisfy p0 6 p1 . If f ∈ Lp0 (Rd ; X) satisfies

supp fb ⊆ {ξ ∈ Rd : |ξ| < t}

for some t > 0, then for any multi-index α ∈ Nd there is a constant C =

Cα,d,p0 ,p1 such that
d d
k∂ α f kLp1 (Rd ;X) 6 Ct|α|+ p0 − p1 kf kLp0 (Rd ;X) .

An extension to exponents 0 < p0 6 p1 6 ∞ will be given in Remark 14.6.4.

Proof. By a routine scaling argument it suffices to consider the case t = 1.
Let ψ ∈ S (Rd ) satisfy ψb ≡ 1 on B1 := {x ∈ Rd : |x| < 1} and put
ψα := ∂ α ψ. Then f = ψ ∗ f , and by Young’s inequality with p11 + 1 = p10 + 1q
we obtain

k∂ α f kLp1 (Rd ;X) = k∂ α (ψ ∗ f )kLp1 (Rd ;X)

= kψα ∗ f kLp1 (Rd ;X) 6 kψα kLq (Rd ) kf kLp0 (Rd ;X) .


The next lemma provides shows how the Lp -norm of ϕk ∗ ϕk+j scales with k.

Lemma 14.4.21. For all j ∈ Z there exists a constant Cd,j,p > 0 such that
for all k > 0 and k + ` > 0 we have
0
kϕk+` ∗ ϕk kLp (Rd ) = C`,p,d 2kd/p .

Proof. The identity ϕ b1 (2−k+1 ξ) implies ϕk (x) = 2(k−1)d ϕ1 (2k−1 x)

bk (ξ) = ϕ
and therefore, by a change of variables in x and y,

kϕk+j ∗ ϕk kpLp (Rd )

Z Z p
(k−1)d (k+j−1)d
= 2 2 ϕ1 (2j 2k−1 (x − y))ϕ1 (2k−1 y) dy dx
Rd Rd
 14.4 Besov spaces 335
Z Z p
= 2kd(p−1) 2jdp−d(p−1) ϕ1 (2j (x − y))ϕ1 (y) dy dx
Rd Rd
| {z }
p
=:Cj,p,d

and the result follows. 

Proof of Theorem 14.4.19. For the sufficiency of (i), first consider the case
s0 > s1 and q0 6 q1 . Then the result follows from the fact that for any scalar
sequence (ak )k>0 ,

(2ks1 ak )k>0 `q 1
6 (2ks0 ak )k>0 `q 0
.

If s0 > s1 , the result follows from (14.23).

If (ii) holds, then writing fk := ϕk ∗ f for k > 0, from Lemma 14.4.20 we
infer that
d d
kfk kLp1 (Rd ;X) 6 C2k( p0 − p1 ) kfk kLp0 (Rd ;X) = C2k(s0 −s1 ) kfk kLp0 (Rd ;X) .

It follows that

kf kBps1 ,q (Rd ;X) = (2ks1 fk )k>0 `q1 (Lp1 (Rd ;X))

1 1
ks0
6 C (2 fk )k>0 `q1 (Lp0 (Rd ;X))

= Ckf kBps0 ,q (Rd ;X) 6 Ckf kBps0 ,q (Rd ;X) ,

0 1 0 0

using (14.22) in the last step.

Suppose now that (iii) holds and let t := s0 − pd0 + pd1 . Then t− pd1 = s0 − pd0
and therefore, by the previous step,

kf kBps1 ,q (Rd ;X) 6 Ckf kBpt (Rd ;X) .

1 1 0 ,q1

Since t > s1 , it follows that the conditions (i) are satisfied, and thus

kf kBpt (Rd ;X) 6 Ckf kBps0 ,q (Rd ;X) .

0 ,q1 0 0

Next we move to the necessity of the conditions (i), (ii), and (iii). It suffices
to consider the case X = K.
Suppose that we have the continuous embedding stated in the theorem.
By the closed graph theorem there is a constant C = Cd,p0 ,p1 ,q0 ,q1 ,s0 ,s1 such
that for all f ∈ Bps00 ,q0 (Rd ),

kf kBps1 ,q (Rd ) 6 Ckf kBps0 ,q (Rd ) . (14.32)

1 1 0 0

First we will derive

d d
s0 − > s1 − and p0 6 p1 . (14.33)
p0 p1
336 14 Function spaces

By (14.22), (14.32) also holds (with a possibly different constant) for q0 = 1

and q1 = ∞. The Fourier support properties (14.8) and (14.9) of ϕk then imply

2ks1 kϕk ∗ ϕk kLp1 (Rd ) 6 kϕk kBps1 ,∞ (Rd )

1
1
X
6 Ckϕk kBps0 ,1 (Rd ) 6 C2ks0 kϕk ∗ ϕk+j kLp0 (Rd ) .
0
j=−1

By Lemma 14.4.21 this implies

0 0
2ks1 2kd/p1 6 C̃2ks0 2kd/p0

for some possibly different constant Ce independent of k. Upon letting k → ∞,

d
this gives the inequality s1 − p0 6 s0 − pd0 , or equivalently, s1 − pd1 6 s0 − pd0 .
1 0

Define ft : Rd → C by fbt (x) := ϕ b0 (t−1 ·). Then ϕb0 = 1 and ϕk = 0 for

k > 1 on supp(fbt ) for t > 0 small enough. Therefore,
− pd
t j kf1 kLpj (Rd ) = kft kLpj (Rd ) = kϕ0 ∗ ft kLpj (Rd ) = kft kBpsj ,q (Rd )
j j

Combining this with (14.32) gives

d d
t− p1 kf1 kLp1 (Rd ) 6 Ct− p0 kf1 kLp0 (Rd ) .

Upon letting t ↓ 0, we find that p0 6 p1 . This completes the proof of (14.33).

Now there are two possibilities: (i) p0 < p1 , or (ii) p0 = p1 . First consider
the case (i). If s0 − pd0 > s1 − pd1 , then (iii) follows. Still assuming (i), if
s0 − pd0 = s1 − pd1 , then in order to deduce (ii) it suffices to show that q0 6 q1 .
We claim that for any finite sequence of scalars (ak )nk=1 ,

k(ak )nk=1 k`q1 6 Ck(ak )nk=1 k`q0 , (14.34)

where C is a constant independent of n > 1 and the sequence (ak )nk=1 . Once
established, this claim gives q0 6 q1 .
To prove the claim fix a scalar sequence (ak )nk=1 . Applying (14.32) to the
Pn −3k(s0 + pd0 ) Pn −3k(s1 + pd0 )
function f := k=1 2
0 a ϕ
k 3k = k=1 2
1 a ϕ
k 3k gives the
inequality
n q1 1/q1
−3k(s1 + pd0 )
X X
2ms1 q1 2 1 ak ϕm ∗ ϕ3k
Lp1 (Rd )
m>0 k=1
n
(14.35)
X X −3k(s0 + pd0 ) q0 1/q0
ms0 q0
6C 2 2 0 ak ϕm ∗ ϕ3k .
Lp0 (Rd )
m>0 k=1

Let us analyse the expressions on the left-hand and right-hand sides for general
values of p, q, and s. We have ϕm ∗ ϕ3k 6= 0 only for m = 3k + ` with
 14.4 Besov spaces 337

` ∈ {−1, 0, 1}. This suggests splitting the sum over m into the sums over
m = 3j + ` for ` ∈ {−1, 0, 1}. Using the lemma, they evaluate as
X n q 1/q
−3k(s+ pd0 )
X
2(3j+`)sq 2 ak ϕ3j+` ∗ ϕ3k
Lp (Rd )
j>0 k=1
n
X q 1/q
−3j(s+ pd0 )
= 2(3j+`)sq 2 aj ϕ3j+` ∗ ϕ3j
Lp (Rd )
j=1
n 1/q
−3j(sq+ dq
X 0
)
= C`,p,d 2(3j+`)sq 2 p0 kaj kq 23jdq/p
j=1
n
X 1/q
= 2`s C`,p,d kaj kq .
j=1

We thus find (using the triangle inequality in `q3 for the upper estimate)
X n q 1/q n
X 1/q
−3k(s+ pd0 )
X
2msq 2 ak ϕm ∗ ϕ3k hd,p,s ka` kq .
Lp (Rd )
m>0 k=1 `=1

Inserting this norm equivalence into (14.35) (taking (p, q, s) = (p0 , q0 , s0 ) on

the left and (p, q, s) = (p1 , q1 , s1 ) on the right) we obtain (14.34).
Finally suppose that (ii) holds. Then from s0 − pd0 > s1 − pd1 we see that
s0 > s1 . If s0 = s1 , then by arguing as above it follows that q0 6 q1 and (i)
follows. 

14.4.d Difference norms

In this section we show that Besov spaces with smoothness parameter s > 0
admit a characterisation in terms of difference norms. This characterisation
can be often used to effectively check whether a given concrete function be-
longs to a given Besov space. For example, we check in Corollary 14.4.26
s
that the Besov spaces B∞,∞ (Rd ; X) coincide with certain spaces of s-Hölder
continuous functions.
For functions f : Rd → X and vectors h ∈ Rd , the function ∆h f : Rd → X
is defined by
∆h f (x) := f (x + h) − f (x).
Clearly, the difference operator ∆h thus defined is bounded as an operator on
Lp (Rd ; X) for all 1 6 p 6 ∞, with norm at most 2. We have the following
formula for the powers ∆m m
h = (∆h ) .

Lemma 14.4.22. For all f ∈ L1 (Rd ; X) and h, ξ ∈ Rd we have

m  
X m
∆mh f = (−1)j f (· + (m − j)h).
j=0
j
338 14 Function spaces

Proof. The identity F (f (·+h))(ξ) = e2πih·ξ fb implies F (∆h f )(ξ) = (e2πih·ξ −

1)fb(ξ), from which it follows that
m  
X m
F (∆m h f )(ξ) = (e 2πih·ξ
− 1) mb
f (ξ) = (−1)j e2πih·ξ(m−j) fb(ξ).
j=0
j

Now apply the inverse Fourier transform. 

Definition 14.4.23 (Difference norm for Besov spaces). Let p, q, τ ∈
[1, ∞], s ∈ R, and m ∈ N \ {0}. For functions f ∈ Lp (Rd ; X) we define the
difference norm by setting
Z ∞ Z 1/τ q dt 1/q
(m,τ ) −sq
[f ]B s (Rd ;X) := t − k∆m
h f kτ dh
p,q
0 {|h|6t} Lp (Rd ) t

with obvious modifications for q = ∞ and/or τ = ∞ where the integral with

respect to dt/t and the average are replaced by essential suprema, and
(m,τ ) (m,τ )
|||f |||B s d := kf kLp (Rd ;X) + [f ]B s d .
p,q (R ;X) p,q (R ;X)

Here we used the notation −F := |F1 | F to denote the average over the set F .
R R

In typical applications one takes τ ∈ {1, p, ∞}.

It is clear that τ0 6 τ1 implies
(m,τ0 ) (m,τ1 )
[f ]B s d 6 [f ]B s d . (14.36)
p,q (R ;X) p,q (R ;X)

(m,τ )
The next theorem implies that if s > 0, then each of the norms ||| · |||B s d
p,q (R ;X)
s
with m > s defines an equivalent norm on Bp,q (Rd ; X).
Theorem 14.4.24 (Difference norms for Besov spaces). Let p, q ∈
[1, ∞], s > 0, τ ∈ [1, ∞], and let m > s be an integer. A function
f ∈ Lp (Rd ; X) belongs to Bp,q
s
(Rd ; X) if and only if [f ]m,τ
s (Rd ;X) < ∞, and
Bp,q
the following equivalence of norms holds:
(m,τ )
kf kBp,q
s (Rd ;X) hd,m,s |||f ||| s
B d .
p,q (R ;X)

Before turning to the details of the proof we give some simple applications.
The first two identify the Sobolev–Slobodetskii spaces and the Hölder spaces
(cf. Section 14.1 for the relevant notation) as Besov spaces.
Corollary 14.4.25 (Sobolev–Slobodetskii spaces). Let p ∈ [1, ∞) and
s ∈ (0, 1). Then
s
Bp,p (Rd ; X) = W s,p (Rd ; X)
with equivalent norms. In fact,
(1,p) 1
[f ]B s d = [f ]W s,p (Rd ;X) . (14.37)
p,p (R ;X) (sp + d)1/p |B1 |
 14.4 Besov spaces 339

Proof. By Theorem 14.4.24 it suffices to prove the identity (14.37) for the
seminorms, which follows from Fubini’s theorem and a change of variable:
Z Z Z ∞
(1,p)
p
|B1 |p [f ]B s (Rd ;X) = 1{|h|6t} t−sp−d−1 k∆h f (x)kp dt dh dx
p,p d d
R R 0
Z Z
−1
= (sp + d) |h|−sp−d k∆h f (x)kp dh dx
Rd Rd
p
= (sp + d)−1 [f ]W s,p (Rd ;X) .

Corollary 14.4.26 (Hölder spaces). Let X be a Banach space and let s ∈

(0, ∞) \ N. Then
s
B∞,∞ s
(Rd ; X) = Cub (Rd ; X)
with equivalent norms.

Proof. Let s = k + θ, where k ∈ N and θ ∈ (0, 1). It follows from Proposition

14.4.18 and Theorem 14.4.19 that we have continuous embeddings
s
B∞,∞ k
(Rd ; X) ,→ B∞,1 (Rd ; X) ,→ Cub
k
(Rd ; X).

Therefore there is no loss of generality in assuming that our functions are

k
k-times continuously differentiable. For functions f ∈ Cub (Rd ; X) and multi-
indices |α| 6 k, from Theorem 14.4.24 we infer the equivalences
(1,∞)
k∂ α f kB∞,∞
θ
α
(Rd ;X) hd,θ |||∂ f |||B θ d ;X) = k∂ α f kCub
θ (Rd ;X) ,
∞,∞ (R

where we used the continuous version of ∂ α f to replace the essential supremum

by a supremum. Now the result follows after summation over all multi-indices
|α| 6 k and an application of Proposition 14.4.17. 

Corollary 14.4.27 (Embeddings into Hölder spaces). Let p0 , q ∈ [1, ∞]

and s0 , s1 > 0 satisfy s0 − pd0 > s1 . Then we have the following continuous
embeddings:
s1
(1) Bps00 ,q (Rd ; X) ,→ Cub (Rd ; X) if s1 ∈
/ N;
s0 s1
(2) Bp0 ,1 (R ; X) ,→ Cub (Rd ; X).
d

Proof. (1): By Theorem 14.4.19 and Corollary 14.4.26,

s1
Bps00 ,q (Rd ; X) ,→ B∞,∞
s1
(Rd ; X) = Cub (Rd ; X).

(2): The case s1 ∈

/ N follows from the previous case. If s1 ∈ N, then by
Theorem 14.4.19 and Proposition 14.4.18,

Bps00 ,1 (Rd ; X) ,→ B∞,1

s1 s1
(Rd ; X) ,→ Cub (Rd ; X).


340 14 Function spaces

The proof of Theorem 14.4.24 makes use the following simple lemma. Recall
the Fourier multiplier notation of Subsection 14.4.b.
Lemma 14.4.28. For non-zero ξ, h ∈ Rd let

e2πih·ξ − 1
mh (ξ) := .
2πih · ξ

Then for all p ∈ [1, ∞] we have mh ∈ MLp (Rd ; X) and kmh kMLp (Rd ;X) 6 1.
Proof. By an elementary computation, the associated Fourier multiplier is
given by
Z 1
Tmh f (x) = f (x − ht) dt = µh ∗ f (x), f ∈ Lp (Rd ; X),
0
R1
where µh (A) = 0 1th∈A dt defines a measure by monotone convergence.
Hence the result follows from (14.24). For p < ∞, one can also use the direct
estimate
Z 1
kTmh f kLp (Rd ;X) 6 kf (· − ht)kLp (Rd ;X) dt = kf kLp (Rd ;X) .
0


Proof of Theorem 14.4.24. Let
Z 1/τ
m,τ
Ip (f, k) := − k∆m τ
2−k h f k dh ,
{|h|61} Lp (Rd )

where the integral average has to be replaced by sup|h|61 if τ = ∞. Discretising

the integral over t in the definition of the difference norm (Definition 14.4.23)
and noting that
Z Z Z
1 d
− 6 =2 − ,
{|h|6t} ωd 2−kd {|h|62−k+1 } {|h|62−k+1 }

we obtain
X Z 2−k+1 Z 1/τ q 1/q
(m,τ )
[f ]B s d = t−sq−1 − k∆m τ
h f k dh dt
p,q (R ;X) Lp (Rd )
k∈Z 2−k {|h|6t}
X Z 1/τ q 1/q
6 2d/τ 2ksq − k∆m τ
h f k dh
{|h|62−k+1 } Lp (Rd )
k∈Z
X Z 1/τ q 1/q
= 2d/τ 2(j+1)sq − k∆m τ
2−j h f k dh
{|h|61} Lp (Rd )
j∈Z
X Z 1/τ q 1/q
= 2s+d/τ 2jsq − k∆m τ
2−j h f k dh .
{|h|61} Lp (Rd )
j∈Z
 14.4 Besov spaces 341

Similarly,
X Z 1/τ q 1/q
(m,τ )
[f ]B s d > 2−s−1−d/τ 2jsq − k∆2m−j h f kτ dh .
p,q (R ;X) Lp (Rd )
j∈Z {|h|61}

Hence,
(m,τ )
[f ]B s d hd,s,τ (2ks Ipm,τ (f, k))k∈Z `q (Z)
. (14.38)
p,q (R ;X)

In view of (14.36) and (14.38) it thus suffices to prove the two estimates

kf kLp (Rd ;X) + (2ks Ipm,∞ (f, k))k∈Z `q (Z)

.d,m,s kf kBp,q
s (Rd ;X) , (14.39)
kf kLp (Rd ;X) + (2ks Ipm,1 (f, k))k∈Z `q (Z)
&s,m,d kf kBp,q
s (Rd ;X) . (14.40)

Throughout the proof of (14.39) and (14.40) we will use the standard algebraic
properties of Lp -multipliers discussed in Section 5.3.a.
Put fj := ϕj ∗ f for j > 0. By Hölder’s inequality,
X
kf kLp (Rd ;X) 6 kfj kLp (Rd ;X) 6 (2−js )j>0 `q0 kf kBp,q
q
(Rd ;X) ,
j>0

0
where the assumption s > 0 implies the finiteness of the `q -norm. To prove
(14.39) and (14.40) it therefore remains to estimate Ipm,∞ (f, k) from above
and Ipm,1 (f, k) from below.
Step 1 – We begin with the proof of (14.39). By Lemma 14.2.10 and the
triangle inequality,
1 X
X
Ipm,∞ (f, k) 6 Ipm,∞ (ϕj ∗ fj+` , k),
`=−1 j>0

observing the standing convention ϕ−1 ≡ 0 which implies that f−1 ≡ 0.

Keeping in mind the operator norm inequality k∆h k 6 2 and (14.7), for j > 1
and arbitrary g ∈ Lp (Rd ; X) we have

Ipm,∞ (ϕj ∗ g, k) = sup ∆m

2−k h ϕj ∗ g Lp (Rd ;X)
|h|61| (14.41)
6 2m kϕj ∗ gkp 6 2m+1 kϕk1 kgkp .

b1 (2−(j−1) ξ), we find that

bj (ξ) = ϕ
On the other hand, using that ϕ

Ipm,∞ (ϕj ∗ g, k) 6 sup kF (∆2m−k h ϕj )kMLp (Rd ;X) kgkp

|h|61
−k
6 sup kξ 7→ (e2πi2 h·ξ
b1 (2−(j−1) ·)kMLp (Rd ;X) kgkp .
− 1)m ϕ
|h|61
(14.42)
342 14 Function spaces

By Lemma 14.4.28 and a dilation

−k
kξ 7→ (e2πi2 h·ξ
− 1)m (h · ξ)−m kMLp (Rd ;X) 6 (2π)m 2−km . (14.43)

Moreover, since ϕ1 is a Schwartz function, dilation, and |h| 6 1,

b1 (2−(j−1) ξ)kMLp (Rd ;X)

kξ 7→ (h · ξ)m ϕ
= 2(j−1)m kξ 7→ (h · ξ)m ϕ
b1 (ξ)kMLp (Rd ;X)
(14.44)
X
(j−1)m
62 cα,m kξ 7→ ξ α ϕ
b1 (ξ)kMLp (Rd ;X)
|α|=m
(j−1)m
6 Cm,d 2 ,

where in the last step we used Proposition 14.4.11 with ∂ α ϕ1 ∈ L1 (Rd ). Com-
bining (14.41) with (14.42), estimating the latter using (14.43) and (14.44),
we obtain the estimate

Ipm,∞ (ϕj ∗ g, k) .d,m min{1, 2(j−k)m }kgkp , j > 1.

Similarly one checks that

Ipm,∞ (ϕ0 ∗ g, k) .d,m min{1, 2−km }kgkp

Therefore, with aj,m = min{1, 2jm },

(2ks Ipm,∞ (f, k))k∈Z `q (Z)

1
X  X 
6 2ks Ipm,∞ (ϕj ∗ fj+` , k)
k∈Z `q (Z)
`=−1 j>0
1
X X 
.d,m,s 2−(j−k)s) aj−k,m 2(j+`)s kfj+` kp
k∈Z `q (Z)
`=−1 j>0

.s (2−js aj,m )j>0 `1

(2(j+`)s kfj kp )j>0 `q
.s kf kBp,q
s (Rd ;X) ,

where we applied the discrete version of Young’s inequality and used the
assumption m > s for the finiteness of the `1 norm.
Step 2 – In this step we prove (14.40). For k > 0 let Tk f := 2kd ϕ(2k ·) ∗ f
and Sk f := ϕk ∗ f . By (14.3), for k > 1 we have Sk = Tk − Tk−1 = (I −
Tk−1 ) − (I − Tk ) and therefore

kf kBp,q
s (Rd ;X) = (2ks kSk f kLp (Rd ;X) )k>0 `q
ks
(14.45)
6 kS0 f kLp (Rd ;X) + 2 (2 kTk f − f kLp (Rd ;X) )k>0 `q
.

By Young’s inequality,
 14.4 Besov spaces 343

kS0 f kLp (Rd ;X) 6 kϕ0 k1 kf kLp (Rd ;X) . (14.46)

It remains to estimate the terms with k > 0 by the difference norm.

Choose ψ ∈ S (Rd ) such that ψ(ξ)
b = 1 if |ξ| 6 1 and ψ(ξ)
b = 0 if |ξ| > 3/2.
Let ϕ ∈ S (Rd ) be given by
m−1
X 
m+1 m
ϕ(ξ)
b = (−1) (−1)j ψ(−(m
b − j)ξ)
j=0
j

and define the sequence (ϕk )k>0 as in (14.3). For |ξ| 6 1/m and 0 6 j 6 m−1
we have ψ(−(m
b − j)ξ) = 1 and therefore
m−1
X  m  
m X m 
ϕ(ξ)
b = (−1)m+1 (−1)j = (−1)m+1 (−1)j − (−1)m = 1
j=0
j j=0
j

by the binomial theorem, and for |ξ| > 3/2 we have ϕ(ξ) b = 0. Furthermore
the Fourier supports of ϕj and ϕk are disjoint for |j −k| > Nm , where Nm ∈ N
only depends on m (rather than for |j − k| > 2 as in (14.10) in the case of an
inhomogeneous Littlewood–Paley sequence). Thanks to these properties, the
proof of Proposition 14.4.2 may be repeated to see that this system leads to
s
an equivalent norm on Bp,q (Rd ; X).
Let f ∈ Lp (Rd ; X). We claim that
Z
Tk f (x) − f (x) = (−1)m+1 ∆m2−k y f (x)ψ(y) dy (14.47)
Rd

Indeed, taking Fourier transforms in the x-variable and using Lemma 14.4.22
and the fact that ψ(0)
b = 1, we have

Td b b −k ξ) − 1)fb(ξ)
k f (ξ) − f (ξ) = (ϕ(2
 m−1
X m 
= (−1)m+1 (−1)j ψ(−(m
b − j)2−k ξ) − 1 fb(ξ)
j=0
j
m  
X m
= (−1)m+1 (−1)j ψ(−(m
b − j)2−k ξ)
j=0
j
m   Z
m+1
X m j −k
= (−1) (−1) e2πi(m−j)2 y·ξ ψ(y) dy
j=0
j R d

Z
−k
= (−1)m+1 (e2πi2 y·ξ − 1)m fb(ξ)ψ(y) dy
d
ZR
= (−1) m+1
F (∆2m−k y f )(ξ)ψ(y) dy
Rd

and the claim follows.

344 14 Function spaces

Fix a real number r > 0, the numerical value of which will be fixed in a
moment. Taking norms in (14.47), using that supx∈Rd (1 + |x|r )|ψ(x)| < ∞,
and writing BR := {ξ ∈ Rd : |ξ| 6 R}, it follows that

kf (x) − Tk f (x)k
Z
6 k∆m
2−k y f (x)ψ(y)k dy
R d
Z X Z
−(j+1)r
.ψ k∆m −k
2 y f (x)k dy + 2 k∆2m−k y f (x)k dy
B1 j>0 B2j+1 \B2j
Z X Z
= k∆2m−k y f (x)k dy + 2−(j+1)(r−d) k∆m
2j+1−k h f (x)k dh
B1 j>0 B1 \B 1
2
X Z
−j(r−d)
6 2 k∆m
2j−k h f (x)k dh.
j>0 B1

Taking Lp -norms with respect to x, we obtain the estimate

X
kTk f − f kLp (Rd ;X) .d,ψ 2−j(r−d) Ipm,1 (f, k − j).
j>0

Taking `q -norms with respect to k > 0 and choosing r > d + s, we obtain

(2ks kTk f − f kLp (Rd ;X) )k>0 `q

X 
.d,ψ 2−j(r−d) 2ks Ipm,1 (f, k − j)
k>0 `q
j>0
X 
= 2−j(r−d−s) 2(k−j)s Ipm,1 (f, k − j)
k>0 `q
j>0
X
6 2−j(r−d−s) (2(k−j)s Ipm,1 (f, k − j))k>0 `q
j>0
X
6 2−j(r−d−s) (2ks Ipm,1 (f, k))k∈Z `q
j>0
X
= 2−j(r−d−s) (2ks Ipm,1 (f, k))k∈Z `q
.
j>0

In combination with (14.45) and (14.46) this proves estimate (14.40). 

14.4.e Interpolation

In order to consider interpolation for Besov spaces, we will now introduce the
so-called retraction and co-retraction operators, which allow us to reduce ques-
tions about the interpolation of Besov spaces to the corresponding questions
about the spaces `qws (Lp (Rd ; X)).
 14.4 Besov spaces 345

Lemma 14.4.29. Let p, q ∈ [1, ∞] and s ∈ R. For k > 0 set ψk := ϕk−1 +

ϕk + ϕk+1 . Define the operators

R : `qws (Lp (Rd ; X)) → Bp,q

s
(Rd ; X)
s
S : Bp,q (Rd ; X) → `qws (Lp (Rd ; X))

by
X
R((fk )k>0 ) = ψk ∗ f k , Sf = (ϕk ∗ f )k>0 .
k>0

Then R is bounded of norm 6 60kϕ0 k21 4|s| , S is an isometry, and RS = I.

Proof. It is clear from the definitions of the spaces involved that S is an
isometry. Next we turn to the proof that R is well defined and bounded. By
(14.7) and Young’s inequality, kϕk+` ∗ ψk k1 6 12kϕ0 k21 . Therefore, by another
application of Young’s inequality and (14.11),
X  X 
ψk ∗ fk = ϕj ∗ ψk ∗ fk q
s (Rd ;X)
Bp,q j>0 `ws (Lp (Rd ;X)
k>0 k>0
 X 
= ϕj ∗ ψj+` ∗ fj+` q
j>0 `w p d
s (L (R ;X)
|`|62
X
6 (ϕj ∗ ψj+` ∗ fj+` )j>0 q
`w p d
s (L (R ;X))
|`|62
X
6 12kϕ0 k12 (fj+` )j>0 `qws (Lp (Rd ;X))
|`|62

6 60kϕ0 k12 4|s| k(fj )j>0 k`qws (Lp (Rd ;X)) ,

s
(Rd ; X) being a consequence
P
the convergence of the sum k>0 ψk ∗ fk in Bp,q
of the convergence of the sum j>0 2 fj in L (Rd ; X), for this allows to first
js p
P
perform the same estimates for differences of partial sums.
The identity RS = I follows from Lemma 14.2.10 and the fact that ψbk ≡ 1
on supp(ϕ bk ). 
Now we are ready identify the complex interpolation spaces of Besov spaces
in a very general setting. In contrast to the complex interpolation results for
Sobolev and Bessel potential spaces in Section 5.6, where it was necessary
to impose UMD assumptions, no geometric restrictions on the interpolation
couple (X0 , X1 ) are needed.
Theorem 14.4.30 (Complex interpolation of Besov spaces). Let
(X0 , X1 ) be an interpolation couple of Banach spaces, let p0 , p1 , q0 , q1 ∈ [1, ∞]
satisfy min{p0 , p1 } < ∞ and min{q0 , q1 } < ∞, and let s0 , s1 ∈ R and θ ∈
(0, 1). Furthermore let p1 = 1−θ θ 1 1−θ θ
p0 + p1 , q = q0 + q1 , and s = (1 − θ)s0 + θs1 .
Then
346 14 Function spaces

[Bps00 ,q0 (Rd ; X0 ), Bps11 ,q1 (Rd ; X1 )]θ = Bp,q

s
(Rd ; [X0 , X1 ]θ )

with equivalent norms.

Proof. Let R : `qws (Lp (Rd ; X)) → Bp,qs s

(Rd ; X) and S : Bp,q (Rd ; X) →
`qws (Lp (Rd ; X)) be the retraction and co-retraction operators of Lemma
14.4.29. Set

Ej := `qwjs (Lpj (Rd ; Xj )), Fj := Bpsjj ,qj (Rd ; Xj ), j ∈ {0, 1},

j

and
Eθ := (E0 , E1 )θ , Fθ := (F0 , F1 )θ , Xθ := [X0 , X1 ]θ .
By Theorem 2.2.6 and Proposition 14.3.3, Eθ = `qws (Lp (Rd ; Xθ )) isometrically.
Therefore,

[Bps00 ,q0 (Rd ; X0 ), Bps01 ,q0 (Rd ; X1 )]θ = Fθ = RSFθ ⊆ REθ ⊆ Bp,q
s
(Rd ; Xθ ),

and for all f ∈ Fθ we have

s (Rd ;X ) = kSf k`q (Lp (Rd ;X )) = kSf kE 6 kf kF

kf kBp,q θ ws θ θ θ

using Theorem C.3.3. Conversely, by Theorem C.3.3,

s
Bp,q (Rd ; Xθ ) = RSBp,q
s
(Rd ; Xθ ) ⊆ REθ ⊆ Fθ ,
s
and for all f ∈ Bp,q (Rd ; Xθ ) we have

kf kFθ = kRSf kFθ 6 CkSf kEθ = CkSf k`rw (Lp (Rd ;Xθ )) = Ckf kBp,q
s (Rd ;X ) ,
θ
s

where C = 60kϕ0 k12 4|s| is the constant of Lemma 14.4.29. 

In the next result we identify the Besov spaces as the real interpolation spaces
of Besov spaces, Bessel potential spaces, and Sobolev spaces, allowing only
non-negative integer values of s in the latter case. In contrast to the case of
complex interpolation, the integrability exponent p as well as the range space
X are fixed.

Theorem 14.4.31 (Real interpolation of Besov spaces). Let X be a

Banach space, let p, q, q0 , q1 ∈ [1, ∞], let s0 , s1 ∈ R satisfy s0 6= s1 , and let
θ ∈ (0, 1) and s = (1 − θ)s0 + θs1 . Then
s0
(Bp,q 0
(Rd ; X), Bp,q
s1
1
(Rd ; X))θ,q = Bp,q
s
(Rd ; X), (14.48)
s0 ,p d s1 ,p d s
(H (R ; X), H (R ; X))θ,q = Bp,q (Rd ; X), (14.49)

with equivalent norms. If we additionally assume that s0 , s1 ∈ N, then

(W s0 ,p (Rd ; X), W s1 ,p (Rd ; X))θ,q = Bp,q

s
(Rd ; X) (14.50)
 14.4 Besov spaces 347

with equivalent norms. If instead we additionally assume that p ∈ [1, ∞) and

s0 , s1 ∈ (0, 1), then
(W s0 ,p (Rd ; X), W s1 ,p (Rd ; X))θ,q = Bp,q
s
(Rd ; X) (14.51)
with equivalent norms.
Proof. The identification (14.51) follows from (14.48) and Corollary 14.4.25.
We will give the proof of the remaining identifications in two steps.
Step 1 – If we can prove that (14.48) holds for q0 = q1 ∈ {1, ∞}, then all
s ,p sj sj ,p
remaining cases can be inferred as follows. Let Aqjj ∈ {Bp,q j,H , W sj ,p },
sj ,p sj ,p
where we assume that sj ∈ N if Aqj = W . Then by (14.48), Theorem
C.3.3, (14.22), and Theorem 14.4.18, we have continuous embeddings
s s0 s1
Bp,q (Rd ; X) = (Bp,1 (Rd ; X), Bp,1 (Rd ; X))θ,q
,→ (As0 ,p (Rd ; X), As1 ,p (Rd ; X))θ,q
s0
,→ (Bp,∞ s1
(Rd ; X), Bp,∞ (Rd ; X))θ,q = Bp,q
s
(Rd ; X),
and (14.48), (14.49), (14.50) follow.
Step 2 – It remains to prove (14.48) for r := q0 = q1 ∈ {1, ∞}. The
argument is similar to that of Theorem 14.4.30.
Let R and S be the retraction and co-retraction operators considered in
Lemma 14.4.29. Let
r
Ej := `w s
(Lp (Rd ; X)), sj
Fj := Bp,r (Rd ; X), j ∈ {0, 1},
j

and
Eθ,q := (E0 , E1 )θ,q , Fθ,q := (F0 , F1 )θ,q .
By Proposition 14.3.5, Eθ,q = `qws (Lp (Rd ; X))with equivalent norms, say with
constants C1 , C2 (depending on θ, p, q, s0 , s1 ), i.e.,
C1−1 kgkEθ,q 6 kgk`w
q p d
s (L (R ;X))
6 C2 kgkEθ,q .
From Theorem C.3.3 it follows that
s0
(Bp,r (Rd ; X), Bp,r
s1 s
(Rd ; X))θ,q = Fθ,q = RSFθ,q ⊆ REθ,q ⊆ Bp,q (Rd ; X),
and for all f ∈ Fθ,q we have
s (Rd ;X) = kSf k`q (Lp (Rd ;X)) 6 C2 kSf kE
kf kBp,q ws θ,q
= C2 kf kFθ,q .
In the converse direction, interpolation R and S by Theorem C.3.3,
s
Bp,q s
(Rd ; X) = RSBp,q (Rd ; X) ⊆ REθ,q ⊆ Fθ,q ,
s
and for all f ∈ Bp,q (Rd ; X) we have
kf kFθ,q = kRSf kFθ,q
6 CkSf kEθ,q . CkSf k`qws (Lp (Rd ;X)) = C3 C1 kf kBp,q
s (Rd ;X) ,

where C = 60kϕ0 k12 4|s| is the constant of Lemma 14.4.29. 

348 14 Function spaces

Corollary 14.4.32. Let s0 , s1 ∈ [0, ∞) satisfy s0 6= s1 , let θ ∈ [0, 1], and put
s := (1 − θ)s0 + θs1 . Then
s0 s1
(Cub (Rd ; X), Cub (Rd ; X))θ,∞ = B∞,∞
s
(Rd ; X)
s
with equivalent norms. Moreover, if s ∈
/ N, then B∞,∞ s
(Rd ; X) = Cub (Rd ; X)
with equivalent norms and therefore
s0 s1
(Cub (Rd ; X), Cub (Rd ; X))θ,∞ = (Cub
s
(Rd ; X).

Proof. By Corollary 14.4.26 it suffices to prove the first identity. Since

sj
by Proposition 14.4.18 we have continuous embeddings B∞,1 (Rd ; X) ,→
sj d sj d
Cub (R ; X) ,→ B∞,∞ (R ; X) we can straightforwardly adapt the proof of
Theorem 14.4.31. 
As a simple application we show that multiplication by a smooth function
leads to a bounded operator on Besov spaces.
Example 14.4.33 (Pointwise multiplication by smooth functions – I). Let p, q ∈
[1, ∞] and s > 0, and let k ∈ (s, ∞) ∩ N. If ζ ∈ Cbk (Rd ; L (X, Y )), then
pointwise multiplication
f 7→ ζf
s
defines a bounded operator from Bp,q (Rd ; X) into Bp,q
s
(Rd ; Y ) of norm

s (Rd ;X),B s (Rd ;Y )) .k,s kζkC k (Rd ;L (X,Y )) .

kf 7→ ζf kL (Bp,q p,q b

Indeed, f 7→ ζf is bounded as a mapping from W j,p (Rd ; X) into W j,p (Rd ; Y )

for each j ∈ {0, . . . , k}. Interpolating between the cases j = 0 and j = k by the
real method with parameters ( ks , q) and applying Theorems 14.4.31 and C.3.3,
the desired result is obtained. Alternatively one can prove the boundedness
as a consequence of Theorem 14.4.24.

14.4.f Duality
s
The main result of this section identifies the duals of Besov spaces Bp,q (Rd ; X)
for p, q ∈ [1, ∞). It is interesting that no geometric assumptions are needed on
X. This contrasts with the situation for vector-valued Bochner spaces: recall
that, by Theorem 1.3.10, for σ-finite measures spaces one has Lp (S; X) =
0
Lp (S; X ∗ ) if and only if X ∗ has the Radon–Nikodým property.
We start with the preliminary observation that elements in the duals of
Besov spaces can be naturally identified with tempered distributions. Indeed,
s
if g ∈ Bp,q (Rd ; X)∗ , then for all ϕ ∈ S (Rd ) and x ∈ X we have

|hϕ ⊗ x, gi| 6 kϕ ⊗ xkBp,q

s (Rd ;X) kgkB s (Rd ;X)∗ = kϕkB s (Rd ) kgkB s (Rd ;X)∗ kxk,
p,q p,q p,q

where we used Proposition 14.4.3 to identify the Schwartz function ϕ with

s
an element of Bp,q (Rd ). Thus the mapping x 7→ hϕ ⊗ x, gi defines an element
 14.4 Besov spaces 349

gϕ ∈ X ∗ , of norm kgϕ k 6 kϕkBp,q

s (Rd ) kgkB s (Rd ;X)∗ . By the continuity of the
p,q

embedding S (Rd ) ,→ Bp,qs

(Rd ) (see Proposition 14.4.3), this implies that the
mapping ϕ → gϕ defines an element in S 0 (Rd ; X ∗ ).
PN
In the converse direction, for g ∈ S 0 (Rd ; X ∗ ) and elements f = n=1 ζn ⊗
xn in S (Rd ) ⊗ X, we can define
N
X
g(f ) := hxn , g(ζn )i. (14.52)
n=1

In order to check whether the mapping f 7→ g(f ) defines an element of

s
Bp,q (Rd ; X)∗ , with p, q ∈ [1, ∞), by the density results contained in Lemma
14.2.1 and Proposition 14.6.8, it suffices to check that there is a constant
C > 0 such that

|g(f )| 6 Ckf kBp,q

s (Rd ;X) , f ∈ S (Rd ) ⊗ X. (14.53)

Theorem 14.4.34. Let X be a Banach space and let p, q ∈ [1, ∞) and s ∈ R.

Then every g ∈ Bp−s0 ,q 0 (R ; X ), when viewed as an element of S (R ; X ),
d ∗ 0 d ∗
s d ∗
determines a unique element of Bp,q (R ; X) , and this identification sets up
a natural isomorphism of Banach spaces
s
Bp,q (Rd ; X)∗ ' Bp−s d ∗
0 ,q 0 (R ; X ).

Proof. The second assertion follows from the first, combined with Corollary
14.4.25.
As a preliminary observation to the proof of the first assertion, we recall
Proposition 2.4.32, which asserts that if g ∈ S 0 (Rd ; X ∗ ) and ζ ∈ S (Rd ),
then ζ ∗ g is in C ∞ (Rd ; X ∗ ) and ∂ α g has polynomial growth for any α ∈ Nd .
Moreover, by Lemma 14.2.10, and the support properties (14.11), (14.12), we
have the identity

XZ 1 XZ
X
g(ζ) = hζ(t), gj (t)i dt = hϕj+` ∗ ζ(t), gj (t)i dt, (14.54)
j>0 Rd `=−1 j>0 Rd

where gj := ϕj ∗ g.
We split the proof of the theorem into three steps.
Step 1 – First let g ∈ Bp−s d ∗
0 ,q 0 (R ; X ). Identifying g with an element of

S (R ; X ), in order to prove that g defines an element of Bp,q

0 d ∗ s
(Rd ; X)∗ we
will check that the duality given by (14.52) satisfies the bound (14.53).
By (14.54), if f ∈ S (Rd ) ⊗ X is as in (14.52), then with fj := ϕj ∗ f we
have
X1 XZ
g(f ) = hfj+` (t), gj (t)i dt.
`=−1 j>0 Rd
350 14 Function spaces

By Hölder’s inequality,
1 XZ
X
|g(f )| 6 |hfj+` (t), gj (t)i| dt
`=−1 j>0 Rd

1
X
6 2−`s (2(j+`)s fj+` )j>0 `q (Lp (Rd ;X))
(2−js gj )j>0 `q0 (Lp0 (Rd ;X ∗ ))
`=−1

6 3 · 2|s| kf kBp,q
s (Rd ;X) kgk −s
B 0 (Rd ;X ∗ ) .
p ,q 0

This verifies the bound (14.53).

Step 2 – Suppose next that g ∈ Bp,q s
(Rd ; X)∗ . As explained above, we can
identify g with an element of S (R ; X ∗ ). Let (fj )j>0 be any finitely non-
0 d

zero sequence in S (Rd ) ⊗ X such that k(2js fj )j>0 k`q (Lp (Rd ;X)) 6 1. Put f :=
R(fj )j>0 , where R : `qws (Lp (Rd ; X)) → Bp,q s
(Rd ; X) is the operator considered
in Lemma 14.4.29. Then by (14.54) and the fact that ψbj = ϕ bj−1 +ϕ bj +ϕ
bj+1 = 1
on supp(ϕ bj ) we see that
XZ XZ
g(f ) = hf (t), gj (t)i dt = hfj (t), gj (t)i dt.
j>0 Rd j>0 Rd

Therefore,
XZ
h2js fj (t), 2−js gj (t)i dt = |g(f )| 6 kf kBp,q
s (Rd ;X) kgkB s (Rd ;X)∗
p,q
j>0 Rd

6 kRk kgkBp,q
s (Rd ;X)∗ .

Taking the supremum over all admissible finitely non-zero sequences (fj )j>0 ,
Propositions 1.3.1 and 1.3.3 imply that g belongs to Bp−s d ∗
0 ,q 0 (R ; X ) and

kgkB −s (Rd ;X ∗ ) = (2−js gj )j>0 `q (Lp (Rd ;X ∗ ))

6 kRk kgkBp,q
s (Rd ;X)∗ .
0p ,q 0

Step 3 – Since the identifications in Steps 2 and 3 are inverse to each other,
they set up a bijective correspondence, and the estimates in the above proof
show that this correspondence is bounded in both directions. 
Theorem 14.4.34 permits an extension of Example 14.4.33 to negative smooth-
ness exponents.

Example 14.4.35 (Pointwise multiplication by smooth functions – II). Let X

and Y be Banach spaces, let p ∈ (1, ∞), q ∈ [1, ∞], s 6 0, and let k ∈
(|s|, ∞) ∩ N. For functions ζ ∈ Cbk (Rd ; L (X, Y )), the pointwise multiplication

f 7→ ζf
 14.4 Besov spaces 351

s
defines a bounded operator from Bp,q (Rd ; X) into Bp,q
s
(Rd ; Y ) of norm

kf 7→ ζf kL (Bp,q
s (Rd ;X),B s (Rd ;Y )) .k,s kζkC k (Rd ;L (X,Y )) .
p,q b
(14.55)

To prove this, first assume that q ∈ (1, ∞) and s < 0. From Example 14.4.33
we obtain the boundedness of g 7→ ζ ∗ g from Bp−s d ∗ −s d ∗
0 ,q 0 (R ; Y ) into Bp0 ,q 0 (R ; X ).

Therefore, by Theorem 14.4.34, the adjoint mapping f 7→ ζf is bounded

s
from Bp,q (Rd ; X ∗∗ ) into Bp,q
s
(Rd ; Y ∗∗ ). Restricting to S (Rd ; X) and using
s
density (Proposition 14.4.3) we obtain boundedness from Bp,q (Rd ; X) into
s d
Bp,q (R ; Y ).
Next let q ∈ {1, ∞} and s < 0. Interpolating the inequality (14.55) for
s+ε s−ε
the cases Bp,2 and Bp,2 by the real method with parameters ( 21 , q), and
s+ε s−ε s
using Theorems 14.4.31 to the effect that (Bp,2 , Bp,2 ) 21 ,q = Bp,q we obtain
boundedness in the endpoint cases q ∈ {1, ∞} by Theorem C.3.3.
ε
Finally, if q ∈ [1, ∞] and s = 0, then by interpolating the cases Bp,q and
−ε 1
Bp,q by the real method with parameters ( 2 , q) we obtain the boundedness
also in this case.

As another application of interpolation and duality we present a density re-

sult, which at first sight looks a bit technical. It will be used to derive an
analogues density result for Triebel–Lizorkin spaces (see Proposition 14.6.17)
which will serve to show that several end-point results do not hold (see the
text below Theorem 14.6.32 and Example 14.6.33). Moreover, some of these
density results will be used to prove results on pointwise multiplication by the
non-smooth function 1R+ (see Sections 14.6.h and 14.7.d).
Let
R̈d := (R \ {0}) × Rd−1 .

Proposition 14.4.36 (Density of compactly supported functions). Let

p, q ∈ [1, ∞) and s ∈ R. Then Cc∞ (R̈d ) ⊗ X is dense in Bp,q
s
(Rd ; X) in each
of the following situations:
(1) s < 1/p;
(2) p, q ∈ (1, ∞) and s = 1/p.

Proof. By Proposition 14.4.3 it suffices to show that for every f ∈ Cc∞ (Rd )
there exist fn ∈ Cc∞ (R̈d ) such that fn → f in Bp,q s
(Rd ). Moreover, by the
embedding (14.23) and Theorem 14.4.19 it suffices to prove (2).
In order to prove (2) let fn := ζn f , where ζn (x) = ζ(nx1 , x2 , . . . , xn ) is
multiplication by n in the first coordinate, and where ζ ∈ C ∞ (Rd ) satisfies
ζ = 1 if |x1 | > 2 and ζ = 0 if |x1 | 6 1. Then by Theorem 14.4.31 the following
interpolation inequality holds:
1/p0 1/p
kfn kB 1/p (Rd ) 6 Ckfn kLp (Rd ) kfn kW 1,p (Rd ) .
p,q

Since
352 14 Function spaces

kfn kLp (Rd ) 6 kf k∞ kζn kLp (Rd ) .ζ n−1/p kf k∞

and similarly 0
kfn kW 1,p (Rd ) .ζ n1/p (kf k∞ + k∇f k∞ ),
the interpolation inequality implies that (fn )n>1 is a bounded sequence in
1/p 1/p
Bp,q (Rd ). Using the reflexivity of Bp,q (Rd ) (which follows Theorem 14.4.34)
we find that (fn )n>1 has a weakly convergent subsequence, say fnk → g weakly
1/p
in Bp,q (Rd ). Since also fn → f in S 0 (Rd ), we find that g = f and therefore
w k·k
1/p
fnk → f weakly in Bp,q (Rd ). Therefore, f ∈ Cc∞ (R̈) = Cc∞ (R̈) , where the
1/p
closures are taken in the weak and norm topology of Bp,q (Rd ), respectively.
This completes the proof. 

14.5 Besov spaces, random sums, and multipliers

In the preceding subsections we have proved various results on embedding
Besov spaces into other function spaces and vice versa. In the present sub-
section we take a look at the embeddability of Besov spaces into spaces of
γ-radonifying operators. This question turns out to be intimately connected
with the type and cotype properties of the space X.
The point of departure is provided by Theorems 9.2.10 and 9.7.3, by which
we have the following natural continuous embeddings:
• L2 (S; X) ,→ γ(L2 (S), X) if and only if X has type 2;
• γ(L2 (S), X) ,→ L2 (S; X) if and only if X has cotype 2;
1 1
• W p − 2 ,p (R; X) ,→ γ(L2 (R), X) if and only if X has type p.
In the first two embeddings (S, A , µ) is an arbitrary measure space.
The main result of this section is the following characterisation of type p
and cotype q in terms of embedding properties:
Theorem 14.5.1 (γ-Sobolev embedding – I). Let X be a Banach space
and let p ∈ [1, 2] and q ∈ [2, ∞].
(1) X has type p if and only if the identity mapping on Cc∞ (Rd ) ⊗ X extends
to a continuous embedding
( 1 − 12 )d
p
Bp,p (Rd ; X) ,→ γ(L2 (Rd ), X);

(2) X has cotype q if and only if the identity mapping on Cc∞ (Rd )⊗X extends
to a continuous embedding
( 1 − 12 )d
γ(L2 (Rd ), X) ,→ Bq,q
q
(Rd ; X).

In particular, for any Banach space X we have continuous embeddings

1
d −1d
B1,1
2
(Rd ; X) ,→ γ(L2 (Rd ), X) ,→ B∞,∞
2
(Rd ; X).
 14.5 Besov spaces, random sums, and multipliers 353

The proof of Theorem 14.5.1 provides quantitative estimates for the norms of
these embeddings. It relies on the following Gaussian version of the Bernstein–
Nikolskii inequality (Lemma 14.4.20).

Lemma 14.5.2 (γ-Bernstein–Nikolskii inequality). Let p ∈ [1, 2] and

q ∈ [2, ∞].
(1) Let X have type p. If f ∈ S (Rd ; X) satisfies supp fb ⊆ {ξ ∈ Rd : |ξ| < t},
then for all multi-indices α ∈ Nd we have
d d
γ
k∂ α f kγ(Rd ,X) 6 κ2,p τp,X π |α| t|α|+ p − 2 kf kLp (Rd ;X) .

(2) Let X have cotype q. If f ∈ S (Rd ; X) satisfies supp fb ⊆ {ξ ∈ Rd : |ξ| <

t}, then for all multi-indices α ∈ Nd we have
d d
k∂ α f kLq (Rd ;X) 6 κq,2 cγq,X π |α| t|α|+ 2 − q kf kγ(Rd ;X) .

Here, κ2,p and κq,2 are the Kahane–Khintchine constants introduced in Section
γ
6.2 and τq,X and cγq,X are the Gaussian type and cotype constants of X,
respectively, introduced in Section 7.1.d.
Proof. (1): By a scaling argument it suffices to consider the case t = 21 . By
Example 9.6.5, ∂ α f ∈ γ(Rd ; X) if and only if ξ 7→ ξ α fb ∈ γ(Rd ; X) and in this
case
k∂ α f kγ(Rd ;X) = (2π)|α| kξ 7→ ξ α fb(ξ)kγ(Rd ;X) .

In order to show that ξ 7→ ξ α fb(ξ) ∈ γ(Rd ; X), by Examples 9.1.12 and 9.4.4
it suffices to check fb ∈ γ(Q; X), where Q := [− 12 , 12 ]d ; in that case

(2π)|α| kξ 7→ ξ α fbkγ(Rd ;X) 6 (2π)|α| kξ 7→ ξ α fbkγ(Q;X) 6 π |α| kfbkγ(Q;X) .

The assertion fb ∈ γ(Q; X) is short-hand for the statement that the Pettis
integral operator Ifb : L2 (Q) → X defined by
Z
Ifbg := fb(ξ)g(ξ) dξ, g ∈ L2 (Q),
Q

belongs to γ(L2 (Q), X) (see Section 9.2.a). We will prove the latter by testing
against an orthonormal bases, making use of Theorem 9.1.17.
Let en (ξ) := e2πin·ξ for n ∈ Zd and ξ ∈ Q. These functions define an
orthonormal basis for L2 (Q) and we have
Z
Ifben = fb(ξ)e2πin·ξ dξ = f (n).
Q

By the Kahane–Khintchine inequalities (Theorem 6.2.6) and the type p con-

dition, for any finite subset F ⊆ Zd we have
354 14 Function spaces
X X
γn Ifben = γn f (n)
L2 (Ω;X) L2 (Ω;X)
n∈F n∈F
X
6 κ2,p γn f (n)
Lp (Ω;X)
n∈F
X 1/p
γ
6 κ2,p τp,X kf (n)kp .
n∈F

It follows from Theorem 9.1.17 that fb ∈ γ(Q, X) and, by the above observa-
tions,
 X 1/p
γ
k∂ α f kγ(Rd ;X) 6 π |α| kfbkγ(Q;X) 6 κ2,p τp,X π |α| kf (n)kp .
n∈Zd

To deduce the estimate in the statement of the theorem from it, for h ∈ Q
and s ∈ Rd put fh (s) := f (s + h). Then supp fc
h ⊆ Q and

 X 1/p
γ
k∂ α f kγ(Rd ;X) = k∂ α fh kγ(Rd ;X) 6 κ2,p τp,X π |α| kfh (n)kp .
n∈Zd

Raising both sides to the power p and integrating over h ∈ Q we obtain

Z X 1/p
α γ
k∂ f kγ(Rd ;X) 6 κ2,p τp,X kfh (n)kp dh
Q
n∈Zd
Z 1/p
γ
= κ2,p τp,X kf (s)kp ds .
Rd

(2): This is proved similarly. 

Proof of Theorem 14.5.1. (1): First we prove the ‘only if’ part and assume
that X has type p. Let f ∈ S (Rd ; X), put fk := ϕk ∗ f , and note that
supp fb0 ⊆ {ξ ∈ Rd : |ξ| 6 23 } and

supp fbk ⊆ Sk := {ξ ∈ Rd : 2k−1 6 |ξ| 6 2k+1 }, k > 1.

By Lemma 14.5.2, fk ∈ γ(Rd ; X) and

1 1
γ
kfk kγ(Rd ;X) 6 κ2,p τp,X 2k( p − 2 )d kfk kLp (Rd ;X) .

By Proposition 9.4.13, applied to the decompositions (S2k )k>0 and (S2k+1 )k>0
of Rd \ {0}, for n > m > 0 we obtain
2n n
X 1/p
X 1 1
fk 6 γ
κ2,p τp,X τp,X 22j( p − 2 )pd kf2j kpLp (Rd ;X)
γ(Rd ;X)
k=2m j=m
 14.5 Besov spaces, random sums, and multipliers 355

 n−1
X 1 1
1/p
γ
+ κ2,p τp,X τp,X 2(2j+1)( p − 2 )pd kf2j+1 kpLp (Rd ;X) .
j=m

P2n+1 P2n P2n+1

Sums of the form k=2m , k=2m+1 , and k=2m+1 can be estimated in a
k∈Z fk in S (R ; X) (by Lemma
d
P P
similar way. Since f = k∈Z ϕk ∗ f =
14.2.10) and hence in γ(R ; X) (by the continuous embedding S (Rd ; X) ,→
d

γ(Rd ; X)), it follows that f ∈ γ(Rd ; X) and

X 1 1
1/p
γ
kf kγ(Rd ;X) 6 2κ2,p τp,X τp,X 2j( p − 2 )pd kf2j kpLp (Rd ;X)
j∈Z
γ
= 2κ2,p τp,X τp,X kf k p
( 1 − 1 )d
2
.
Bp,p (Rd ;X)

( 1 − 12 )d
Since S (Rd ; X) is dense in Bp,p
p
(Rd ; X) by Proposition 14.4.3, the identity
( 1 − 1 )d
mapping on S (Rd ; X) extends to a bounded operator from Bp,p p 2
(Rd ; X)
γ
into γ(Rd ; X) of norm at most 2κ2,p τp,X τp,X . The simple proof that this ex-
tension is injective is left to the reader.
Next we prove the ‘if’ part. Since every Banach space has type 1, the ‘if’
part is trivial for p = 1. In the rest of the proof of (1) we may therefore assume
that p ∈ (1, 2]. We will prove the stronger statement that if for some r ∈ (1, ∞]
the identity operator on S (Rd ; X) extends to a bounded operator, say I,
( 1 − 1 )d
from Bp,rp 2
(Rd ; X) into γ(L2 (Rd ), X), X has type r (and then necessarily
r ∈ (1, 2]).
Let ψ ∈ S (Rd ) be such that kψkL2 (Rd ) = 1 and supp(ψ) b ⊆ {ξ ∈ Rd :
d
b1 (ξ) = 1}. For n > 1, let ψn ∈ S(R ) be defined by
ϕ

b −n+1 ξ).
ψbn (ξ) := 2(−n+1)d/2 ψ(2

Then (ψn )n>1 is an orthonormal system in L2 (Rd ). By Proposition 9.1.3, for

PN
any finite sequence (xn )N
n=1 in X we then have, with f := n=1 ψn ⊗ xn ,

N
X 2
kf k2γ(Rd ;X) = E γn xn .
n=1

On the other hand, since ϕk ∗ ψn = δkn ψn (this is seen by taking Fourier

transforms and using the Fourier support properties of ϕk ),
N N
X 1 1 X
kf kq ( 1 − 1 )d = 2( p − 2 )dr kψn krp kxn kr = kψkrp kxn kr .
p 2
Bp,r (Rd ;X) n=1 n=1

By putting things together we see that X has type r, with Gaussian type r
γ
constant τr,X 6 kψkp kIk.
(2): This is proved similarly. 
356 14 Function spaces

14.5.a The Fourier transform on Besov spaces

This section presents some mapping properties of the Fourier transform on

spaces of functions taking values in a Banach space with (co)type or Fourier
type properties. Recall from Section 2.4.b that a Banach space has Fourier type
p ∈ [1, 2] if the Fourier transform, initially defined on L1 (Rd ; X) ∩ Lp (Rd ; X),
0
extends to a bounded operator from Lp (Rd ; X) into Lp (Rd ; X). If that is the
d
case, the norm of this extension is denoted by ϕp,X (R ).

Proposition 14.5.3 (Integrability of Fourier transforms – II). Let p ∈

[1, 2], and suppose that one of the following two conditions holds:
(i) q ∈ [p, ∞] and X has Fourier type p;
(ii) q ∈ [2, ∞] and X has type p and cotype 2.
Let F denote the Fourier transform on S 0 (Rd ; X) and let s := ( p1 − 1q )d.
0
(1) F restricts to a bounded operator from Bp,q
s d q d
0 (R ; X) into L (R ; X);

q0
(2) F restricts to a bounded operator from W bsc+1,p
(R ; X) into L (Rd ; X).
d

The case q = ∞ gives sufficient conditions for the Fourier transform to take
values in L1 (Rd ; X). Different conditions guaranteeing this have been dis-
cussed in Lemma 14.2.11, where growth assumptions on the functions and
their derivatives where imposed.
Proof. We start with case (i). Accordingly, let q ∈ [p, ∞] and let X have
Fourier type p
s d d
(1): Let f ∈ Bp,q 0 (R ; X). Put fk := ϕk ∗ f for k > 0. Let I0 = {ξ ∈ R :
|ξ| < 1} and
In := {ξ ∈ Rd : 2n−1 6 |ξ| < 2n }, n > 1.
The sets In thus defined are pairwise disjoint, we have n>0 In = Rd , and
S

1/q0 1 X 1/q0
0 0
X X
kfbkq0 = k1In fbkqq0 6 k1In fbn+` kqq0 ,
n>0 `=−1 n>0
P
where we used that supp(ϕ bk ) ∩ In = ∅ for |n − k| > 2 and that k>0 ϕ
bk = 1.
1 s 1
By Hölder’s inequality with q0 = d + p0 and the Fourier type p assumption,
for ` ∈ {−1, 0, 1} we have

k1In fbn+` kq0 6 k1In k d kfbn+` kp0

s

6 ϕp,X (Rd )2(n+1)s kfn+` kp0 6 22s ϕp,X (Rd )2(n+`)s kfn+` kp0 .
0 0
Taking `q -norms on both sides we obtain fb ∈ Lq (Rd ; X) and

kfbkq0 6 22s 3ϕp,X (Rd )kf kB s (Rd ;X) .

p,q 0
 14.5 Besov spaces, random sums, and multipliers 357

(2): This follows from (1) since by Proposition 14.4.18 and Theorem
bsc+1
14.4.19 we have the embeddings W bsc+1,p (Rd ; X) ,→ Bp,∞ (Rd ; X) ,→
s d
Bp,1 (R ; X).
Case (ii): Assume now that q ∈ [2, ∞] and that X has type p and cotype 2.
Using the same notation as in case (i), by Hölder’s inequality with q10 = 1r + 12 ,
Theorem 9.2.10, and Lemma 14.5.2 we have

k1In fbn+` kq0 6 k1In kr kfbn+` k2

6 cγ2,X 2d/r(n+1) kfn+` kγ(Rd ;X)
1 1
.d,p c2,X τp,X 2(n+1)d/r 2(n+1)( p − 2 )d kfn+` kp
= c2,X τp,X 2(n+1)s kfn+` kp .

The proof can now be finished as in case (i). 

As an application of Proposition 14.5.3 using the Fourier type of X, we give an

improvement of the Mihlin multiplier theorem for vector-valued Besov spaces
presented in Theorem 14.4.16. Before we do that we derive an immediate
consequence of Propositions 14.4.11.
Corollary 14.5.4 (Fourier multiplier theorem for Lp under Fourier
type). Let p ∈ [1, ∞] and s ∈ R, let X and Y be Banach spaces, and suppose
that one of the following conditions holds:
(i) Y has Fourier type τ ;
(ii) Y has type τ and cotype 2.
Then we have a continuous embedding
d/τ
Bτ,1 (Rd ; L (X, Y )) ,→ MLp (Rd ; X, Y ),

d/τ
i.e., every m ∈ Bτ,1 (Rd ; L (X, Y )) defines a bounded operator Tm from
Lp (Rd ; X) to Lp (Rd ; Y ).
b
Proof. The result is immediate from the fact that m ∈ L1 (Rd ; L (X, Y )) by
Proposition 14.5.3. 

Remark 14.5.5. It is possible to prove a result as in Corollary 14.5.4 under

assumptions on m and m∗ in the strong operator topology if X (equivalently
X ∗ ) has Fourier type τ1 and Y has Fourier type τ2 . Indeed, assume there is a
constant Cm such that

kmxkB d/τ2 (Rd ;Y ) 6 Cm kxk, x ∈ X, (14.56)

τ2 ,1

km∗ y ∗ kB d/τ1 (Rd ;X ∗ ) 6 Cm ky ∗ k, y ∗ ∈ Y ∗ . (14.57)

τ1 ,1

First observe that by (14.56), (14.57) and Proposition 14.5.3,

358 14 Function spaces
b
k mxk L1 (Rd ;Y ) 6 Cτ2 ,Y Cm kxk
b
∗ ∗ (14.58)
km y kL1 (Rd ;X ∗ ) 6 Cτ1 ,X Cm ky ∗ k.
b b
∗
Here mx := F −1 (mx) and m y ∗ := F −1 (m∗ y ∗ ). Therefore, for f ∈ S (Rd ) ⊗
X, by Fubini’s theorem one can write
b Z Z b
km ∗ f kL1 (Rd ;Y ) 6 km(t − s)f (s)k ds dt
d d
ZR ZR b
6 km(r)f (s)k dr ds 6 Cm kf kL1 (Rd ;Y ) .
Rd Rd

This proves that Tm extends uniquely to Tm ∈ L (L1 (Rd ; X), Lb1 (Rd ; Y )).
Since the second line of (14.58) trivially implies that the kernel m satisfies
the dual Hörmander’s condition, it follows from the Calderón–Zygmund ex-
trapolation theorem (Theorem 11.2.5) that Tm extends uniquely to Tm ∈
L (Lp (Rd ; X), Lp (Rd ; Y )) for all p ∈ [1, ∞). By a duality argument a similar
result can be derived for p = ∞.
It is clear from the above proof that we can replace the Fourier type
conditions by the conditions that Y has type τ2 and cotype 2, and X ∗ has
type τ1 and cotype 2.

We continue with an improvement of Theorem 14.4.16 using the Fourier type

or type and cotype Y .
s
Theorem 14.5.6 (Mihlin multiplier theorem for Bp,q (Rd ; X) under
type conditions). Let p, q ∈ [1, ∞] and s ∈ R and X and Y be Banach
spaces and suppose that one of the following conditions holds:
(i) Y has Fourier type τ ;
(ii) Y has type τ and cotype 2.
d
If m ∈ C b τ c+1 (Rd ; L (X, Y )) satisfies

Km := sup sup (1 + |ξ||α| )k∂ α m(ξ)kL (X,Y ) < ∞,

|α|6b τd c+1 ξ∈Rd

s
then there is a bounded operator T : Bp,q (Rd ; X) → Bp,q
s
(Rd ; Y ) with kT k 6
Cd,s,X,Y Km such that T f = F (mfb) for all f ∈ S (R ) ⊗ X.
−1 d

Note that in the case p, q < ∞, one has that T is the unique bounded extension
of Tm : S (Rd ) ⊗ X → S 0 (Rd ; Y ). In the end point case p = ∞ or q = ∞ this
does not make sense since S (Rd ) ⊗ X is not dense in Bp,qs
(Rd ; X). This is the
main reason for the unusual formulation in Theorem 14.5.6.
By a duality argument one can also formulate the (Fourier) (co)type con-
ditions on X ∗ , but the end-point cases require some caution.
s
Proof. For f ∈ Bp,q (Rd ; X) let fk = ϕk ∗ f and mk = ϕ
bk m. Define
 14.5 Besov spaces, random sums, and multipliers 359
1 X
X
Tf = Tmk+` fk . (14.59)
`=−1 k>0

We will check that the series converges in S 0 (Rd ; Y ) and defines an element
s
in Bp,q (Rd ; Y ).
The proof follows the lines of Theorem 14.4.16. First we show that mk
bound MLp (Rd ; X, Y ) with a uniform bound in k > 0. First let k > 1. By
invariance under dilations (see Proposition 5.3.8), Corollary 14.5.4, and the
embeddings (14.23) and (14.29), we have

kmk kMLp (Rd ;X,Y ) = kmk (2k−1 ·)kMLp (Rd ;X,Y )

6 Cτ,Y kmk (2k−1 ·)kB d/τ (Rd ;L (X,Y ))
τ,1
k−1
6 Cτ,Y kmk (2 ·)k d
W b τ c+1,τ (Rd ;L (X,Y ))

Since mk (2k−1 ·) = ϕb1 (·)m(2k−1 ·), by the support properties of ϕb1 is suffices
α
to bound ∂ [ϕ b1 (ξ)m(2k−1 ξ)] for |α| 6 b τd c + 1, uniformly in k > 1 and 1 6
|ξ| 6 3. This can be done in the same way as in (14.26). The case k = 0 can
be proved in the same way without the dilation argument. We can conclude
that

kTmk+` fk kLp (Rd ;Y ) 6 Cd,s,X,Y Km kfk kLp (Rd ;X) (14.60)

Next we check the convergence of the series in (14.59). For ζ ∈ S (Rd ) one
P1
has Tmk+` fk (ζ) = j=−1 Tmk+` fk (ζk+j ), where ζk = ϕk ∗ ζ, and thus

1
X
kTmk+` fk (ζ)kY 6 kTmk+` fk kLp (Rd ;Y ) kζk+j kLp0 (Rd )
j=−1
1
X
6 Cd,s,X,Y Km 2sk kfk kLp (Rd ;Y ) 2|s| 2−s(k+j) kζk+j kLp0 (Rd )
j=−1

Summing over k we see that

X X 1
X
kTmk+` fk (ζ)kY 6 Cd,s,X,Y Km 2sk kfk kLp (Rd ;Y ) 2−sk kζk+j kLp0 (Rd )
k>0 k>0 j=−1
|s|
6 3 · 2 Cd,s,X,Y Km kf kBp,q
s (Rd ;X) kζk −s
B 0 (Rd ) ,
p ,q 0

which gives the required convergence.

By the properties of (ϕn )n>0 we can write
j+1
X 1
X j+1
X 1
X
F (ϕj ∗ T f ) = ϕ
bj ϕ
bk+` mϕ
bk f = ϕ
bj mϕ
bk f = mj fj+` .
k=j−1 `=−1 k=j−1 `−1
360 14 Function spaces

Therefore, the boundedness follows from

1
X
kT f kBp,q
s (Rd ;Y ) 6 k(Tmj fj+` )j>0 k`q (Lp (Rd ;Y ))
`−1
1
X
6 Cd,s,X,Y Km k(fj+` )j>0 k`q (Lp (Rd ;X))
`−1
0
6 Cd,s,X,Y Km kf kBp,q
s (Rd ;X) .

It remains to observe that for f ∈ S (Rd ) ⊗ X, the following identities

hold in S 0 (Rd ; X)
1
X X X
Tcf = bk fb =
bk+` mϕ
ϕ bk fb = mfb.
mϕ
k>0 `=−1 k>0


A further consequence of Proposition 14.5.3 is a Fourier multiplier theorem of
a very different nature, in which the multiplier is non-smooth but the domain
and range spaces have different integrability and smoothness exponents.

Proposition 14.5.7. Let X and Y be Banach spaces with Fourier type p ∈

[1, 2] and let s := ( p1 − p10 )d. Let m : Rd → L (X, Y ) be strongly measur-
able in the strong operator topology and uniformly bounded. Then the Fourier
multiplier Tm = F −1 mF is bounded as an operator from Bp,p s
(Rd ; X) into
0
Lp (Rd ; Y ) with norm
d d
kTm kL (Bp,p
s (Rd ;X),Lp0 (Rd ;Y )) .p ϕp,X (R )ϕp,Y (R ) sup km(ξ)kL (X,Y ) .
ξ∈Rd

Proof. By the Fourier type p of Y ,

kTm f kLp0 (Rd ;Y ) 6 ϕp,Y (Rd )kmfbkLp (Rd ;Y )

6 ϕp,Y (Rd ) sup km(ξ)kL (X,Y ) kfbkLp (Rd ;X) ,
ξ∈Rd

The Fourier type p of X and Proposition 14.5.3, applied with q = p0 , give

kfbkLp (Rd ;X) .p ϕp,X (Rd )kf kBp,p

s (Rd ;X) ,

and the result follows. 

14.5.b Smooth functions have R-bounded ranges

In Chapter 8 we have seen several instances of the general principle that

sufficiently smooth operator-valued functions have R-bounded ranges. The
 14.5 Besov spaces, random sums, and multipliers 361

amount of smoothness needed depends on the geometry of the underlying

Banach spaces. For instance, it was shown in Theorem 8.5.21 that if X has
cotype q and Y has type p, and if T ∈ W s,r (Rd ; L (X, Y )) with ( p1 − 1q )d <
d
r < s < 1, then T has a continuous version whose range is R-bounded.
In the present section we will show that if the Besov scale is used instead
of the Sobolev scale, the analogous result holds for the optimal smoothness
exponent s = ( p1 − 1q )d and the restriction s < 1 can be omitted. The precise
statement reads as follows.

Theorem 14.5.8 (Besov functions with R-bounded range – I). Let

X and Y be Banach spaces, X having cotype q ∈ [2, ∞] and Y having type
d/r
p ∈ [1, 2]. If r ∈ [1, ∞] satisfies 1r > p1 − 1q , then every T ∈ Br,1 (Rd ; L (X, Y ))
has R-bounded range, with R-bound

R(T (t) : t ∈ Rd ) 6 CkT kB d/r (Rd ;L (X,Y )) , (14.61)

r,1

where C is a constant depending on d, p, q, r, X, Y .

d/r
By Theorem 14.4.19, the spaces Br,1 (Rd ; L (X, Y )) increasing in the exponent
r ∈ [1, ∞] and we have continuous embeddings
d/r
Br,1 (Rd ; L (X, Y )) ,→ B∞,1
0
(Rd ; L (X, Y )) ,→ Cub (Rd ; L (X, Y )), (14.62)

the second being a consequence of Proposition 14.4.18. The continuous version

provided by (14.62) is used in the left-hand side of (14.61).
0
In the proof below, we will use the Lorentz space Lr ,σ (Rd ) with σ =
min{ p10 , q1 }
1
+ q1
∈ (0, 1]. Referring to Appendix F, we recall that the Lorentz space
p0
r 0 ,σ
L (R ) is the space of all measurable functions f : Rd → K for which the
d

(quasi-)norm
0
kf kLr0 ,σ (Rd ) := τ 7→ τ 1/r f ∗ (τ ) Lσ (R+ , dτ
τ )

is finite, where

f ∗ (τ ) := inf λ > 0 : |{|f | > λ}| 6 τ ,


τ ∈ R+ ,

is the non-increasing rearrangement of f .

Proof. By the observation before (14.62) it suffices to prove the theorem in
the case 1r = p1 − 1q . In the proof we will only consider r ∈ (1, ∞]; in Theorem
14.5.9 a stronger result is proved which covers the case r = 1 of the present
theorem.
Let us write
X X1 X
T = Tk = ϕk+` ∗ Tk ,
k>0 `=−1 k>0
362 14 Function spaces

where Tk = ϕk ∗ T , and we used (14.12) inPthe second identity. Since T ∈

0
B∞,1 (Rd ; L (X, Y )) (see (14.62)), the series k>0 Tk converges uniformly on
R with respect to the operator norm of L (X, Y ). By Propositions 8.1.19 and
d

8.1.22,
1 X
X
R(T (t) : t ∈ Rd ) 6 R(ϕk+` ∗ Tk (t) : t ∈ Rd ), (14.63)
`=−1 k>0

provided of course that the operator families occurring in the sums are R-
bounded and their R-bounds are summable. Proving this will occupy us in
the remainder of the proof.
Fix an integer n > 1. Starting from the identity ϕn (t) = 2(n−1)d ϕ1 (2n−1 t)
(see (14.4)), it is elementary to check that the non-increasing rearrangements
satisfy ϕ∗n (τ ) = 2(n−1)d ϕ1∗ (2n−1 τ ). Therefore,
0
kϕn kLr0 ,σ (Rd ) = 2(n−1)d kτ 7→ τ 1/r ϕ1∗ (2n−1 τ )kLσ (R+ , dτ )
τ
1/r 0
=2 (n−1)d/r
kτ 7→ τ ϕ∗1 (τ )kLσ (R+ , dτ ) = 2(n−1)d/r kϕ1 kLr0 ,σ (Rd ) ,
τ

the latter being finite since ϕ1 ∈ S (Rd ). A similar calculation can be done
for n = 0.
For t ∈ Rd define ϕn,t ∈ S (Rd ) by ϕn,t (s) := ϕn (t − s). Then ϕn,t is
identically distributed with ϕn . Letting Tk,ϕn,t ∈ L (X, Y ) be the integral
operator from Proposition 8.5.16, i.e.,
Z
Tk,ϕn,t x := ϕn,t (s)Tk (s)x ds,
Rd

it follows from Proposition 8.5.16 with σ = r0 min{ p10 , 1q } and ψ = ϕn that for
all n > 0 and k > 0 the set {ϕn ∗ Tk (t) : t ∈ Rd } is R-bounded, with R-bound

R(ϕn ∗ Tk (t) : t ∈ Rd ) = R(Tk,ϕn,t : t ∈ Rd ) 6 C2nd/r kTk kLr (Rd ;L (X,Y )) .

With (14.63) we conclude that

1 X
X
R(T (t) : t ∈ Rd ) 6 C 2(k+`)d/r kTk kLr (Rd ;L (X,Y ))
`=−1 k>0
d
6 3 · 2 r CkT kB d/r (Rd ;L (X,Y )) .
r,1


We have the following variation of this result for the strong operator topology:
Theorem 14.5.9 (Besov functions with R-bounded range – II). Let X
and Y be Banach spaces and assume that Y has type p ∈ [1, 2]. Suppose that
d/p
T : Rd → L (X, Y ) satisfies T x ∈ Bp,1 (Rd ; Y ) for all x ∈ X and
 14.5 Besov spaces, random sums, and multipliers 363

kT xkB d/p (Rd ;Y ) 6 CT kxk, x ∈ X.

p,1

Then the family {T (t) ∈ L (X, Y ) : t ∈ Rd } is R-bounded, with R-bound

R(T (t) ∈ L (X, Y ) : t ∈ Rd ) 6 CCT ,
where C is a constant depending on p and Y .
Proof. We begin with the case p = 1, which corresponds to the case where Y
is an arbitrary Banach space. By Proposition 14.5.3 we have Tbx ∈ L1 (Rd ; Y )
and
kTbxkL1 (Rd ;Y ) .d kT xkB d/p (Rd ;Y ) 6 CT kxk.
p,1

This implies that we have the integral representation

Z
T (t)x = e2πiξ·t Tb(ξ)x dξ, t ∈ Rd ,
Rd

where the operator-valued kernel is strongly in L1 . Now Theorem 8.5.4 implies

that the family {T (t) : t ∈ Rd } is R-bounded, with R-bound Rp (T (t) : t ∈
Rd ) . d C T .
Next assume that p ∈ (1, 2]. For k > 0 and x ∈ X set Tk (t)x := ϕk ∗ T (t)x.
By Theorem 14.5.1,
1 1
kTk xkγ(L2 (Rd ),Y ) 6 CkTk xk ( 1 − 1 )d
p 2
6 Cd,p,s 2kd( p − 2 ) kTk xkLp (Rd ;Y ) ,
Bp,p (Rd ;Y )
(14.64)

where (setting s = d( p1 − 12 ) for brevity) the second inequality follows from

X
kTk xkpB s (Rd ;Y ) = 2nsp kϕn ∗ ϕk ∗ T xkpLp (Rd ;Y )
p,p
n>0
1
X
= 2(k+`)sp kϕk+` ∗ ϕk ∗ T xkpLp (Rd ;Y )
`=−1
1
X
6 2(k+`)sp kϕk+` kp1 kϕk ∗ T xkpLp (Rd ;Y )
`=−1

6 3 · 2(k+1)sp · 2p kϕkp1 kTk xkpLp (Rd ;Y )

using (14.11) and (14.7).

Choose arbitrary finite sequences (tm )M d M
m=1 in R and (xm )m=1 in X, and
M
let (εm )m=1 be a Rademacher sequence on a probability space (Ω, P). Since
Y has type p > 1 it follows from Theorem 9.6.14 with constant Lp,Y hat
M
X
εm T (tm )xm
L2 (Ω;Y )
m=1
364 14 Function spaces
1
X X M
X
6 εm ϕk+` ∗ Tk (tm )xm
L2 (Ω;Y )
k>0 `=−1 m=1
1
X X M
X Z
= εm Tk (u)xm ϕk+` (tm − u) du
Rd L2 (Ω;Y )
k>0 `=−1 m=1
1
X X M
X
6 Lp,Y kϕk+` kL2 (Rd ) εm T k x m
L2 (Ω;γ(L2 (Rd ),Y ))
k>0 `=−1 m=1

X M
X 
6 Lp,Y Cϕ 2kd/2 Tk εm xm ,
L2 (Ω;γ(L2 (Rd ),Y ))
k>0 m=1

bk+` kL2 (Rd ) 6 Cϕ 2kd/2 .

where we used that (14.9) implies kϕk+` kL2 (Rd ) = kϕ
Applying (14.64) pointwise in Ω, setting C0 := Lp,Y Cϕ Cd,p,s , and using the
Kahane-Khintchine inequalities, we continue estimating

M
X 
X 1 1
6 C0 2kd/2 2k( p − 2 )d Tk εm x m
L2 (Ω;Lp (Rd ;Y ))
k>0 m=1
Z X M
X 
6 C0 κ2,1 2kd/p Tk εm x m dP
Ω k>0 Lp (Rd ;Y )
m=1
Z M
X 
= C0 κ2,1 T εm xm d/p
dP
Ω Bp,1 (Rd ;Y )
m=1
Z M
X
6 C0 κ2,1 CT εm x m dP
Ω X
m=1
M
X
6 C0 κ2,1 CT εm x m .
L2 (Ω;X)
m=1

Putting things together gives the required R-boundedness estimate. 

Remark 14.5.10.
(1) The method of proof for p = 1 in Theorem 14.5.9 could be extended
to p ∈ (1, 2] if Y has Fourier type p. We have not done this, because
Proposition 7.3.6 shows that having type p is weaker than having Fourier
type p.
(2) In the case p = 1 and d = 1, a variation of the argument in Proposition
8.5.7 actually gives a stronger result than Theorem 14.5.9, namely that if
T x ∈ W d,1 (Rd ; L (X, Y )) for all x ∈ X, then the range of T is R-bounded.
 14.6 Triebel–Lizorkin spaces 365

14.6 Triebel–Lizorkin spaces

As we have seen in the preceding sections, the study of Besov spaces is inti-
mately connected with the space `q (Lp (Rd ; X)) through the very definition,
which features the norm
ks
kf kBq,s
p
(Rd ;X) = (2 ϕk ∗ f )k>0 `q (Lp (Rd ;X))
.
s
The class of Triebel–Lizorkin spaces Fp,q (Rd ; X) is obtained upon replacing
`q (Lp (Rd ; X)) by Lp (Rd ; `q (X)), putting

kf kFp,q
s (Rd ;X) = (2ks ϕk ∗ f )k>0 Lp (Rd ;`q (X))
.

The theory of Triebel–Lizorkin spaces is in many respect analogous to the

theory of Besov spaces, but the occurrence of the `q -norm inside the Lp -norm
precludes the use of Young’s inequality to estimate the norm of term-wise
convolutions, a technique that was critically used in our treatment of Besov
spaces. This makes the norm of Triebel–Lizorkin spaces more difficult to deal
with.

14.6.a The Peetre maximal function

The obstruction just noted already makes itself felt if one tries to adapt the
proof that Besov spaces are independent up to an equivalent norm of the in-
homogeneous Littlewood–Paley sequence (ϕk )k>0 to Triebel–Lizorkin spaces.
The encountered difficulty will be resolved by a variant on the Fefferman–
Stein inequality due to Peetre, to which we turn in the present preliminary
subsection.
Throughout this section, unless otherwise stated X is an arbitrary Banach
space. For a strongly measurable function f : Rd → X and r ∈ (0, ∞) we let

Mr f (x) := (M (kf kr )(x))1/r , x ∈ Rd , (14.65)

where M is the Hardy–Littlewood maximal operator introduced in Section

2.3,
Z
1
M f (x) := sup kf (y)k dy,
B3x |B| B

the supremum being taken over all Euclidean balls B in Rd that contain x.
Lemma 14.6.1 (Peetre’s maximal inequality). Fix r, t ∈ (0, ∞) and a
multi-index α ∈ Nd , and let f ∈ S 0 (Rd ; X) satisfy

supp fb ⊆ Bt := {ξ ∈ Rd : |ξ| 6 t}.

Then f ∈ C ∞ (Rd ; X) and there exist constants C1 and C2 , depending only on

|α|, d, r such that for all x ∈ Rd we have
366 14 Function spaces

k∂ α f (x − z)k kf (x − z)k
sup t−|α| d/r
6 C1 sup d/r
6 C2 Mr f (x)
z∈Rd (1 + t|z|) z∈Rd (1 + t|z|)

In particular, taking z = 0, for all x ∈ Rd we have

t−|α| k∂ α f (x)k 6 kf (x)k 6 C2 Mr f (x).

Proof. That the tempered distribution f is represented by a function in

C ∞ (Rd ; X) has already been observed in Lemma 14.2.9. In the remainder
of the proof we assume that this identification has been made.
By an iteration argument it suffices to consider Pd multi-indices satisfying
|α| = 1. The short-hand notation k∇f (x)k = j=1 k∂j f (x)k will be used
throughout the proof. We first consider the case f ∈ S (Rd ; X). Replacing f
by f (t−1 ·), it suffices to prove the result for t = 1.
Step 1 – Choose ψ ∈ S (Rd ) such that ψb ≡ 1 on B1 . Since fb is supported
on B1 , we have f = ψ ∗ f and ∇f = (∇ψ) ∗ f . It follows that for x, z ∈ Rd
and λ > 0,
Z
k∂j f (x − z)k 6 |∂j ψ(x − z − y)| kf (y)k dy
Rd
Z
6 cλ (1 + |x − z − y|)−λ kf (y)k dy,
Rd

where cλ = supy∈Rd (1 + |y|)λ |∂j ψ(y)|. Clearly we have 1 + |x − y| 6 (1 + |x −

z − y|)(1 + |z|), and upon taking λ = d + 1 + d/r we obtain

k∂j f (x − z)k
Z
6 c λ (1 + |x − z − y|)−λ (1 + |z|)−d/r kf (y)k dy
(1 + |z|)d/r Rd
Z
6 cλ (1 + |x − z − y|)−d−1 (1 + |x − y|)−d/r kf (y)k dy
Rd
kf (x − y)k
6 C1 sup ,
y∈Rd (1 + |y|)d/r

where C1 = cλ Rd (1 + |y|)−d−1 dy. This gives the first inequality in the state-
R

ment of the lemma.

Step 2 – Fix ε > 0 and let Qε be the closed cube centred at zero and of
side-length ε. We claim that for all g ∈ C 1 (Qε ; X),

ε Z 1/r
kg(0)k 6 sup k∇g(y)k + − kg(y)kr dy , (14.66)
2 y∈Qε Qε

1
R R
where we write −Q = |Q| Q
for averages. By scaling it suffices prove (14.66)
for ε = 1.
Fix g ∈ C 1 (Q1 ; X). For all y ∈ Q1 we have kyk 6 21 and
 14.6 Triebel–Lizorkin spaces 367
Z 1
g(0) = g(y) + ∇g(ty) · y dt.
0

Therefore, kg(0)k 6 kg(y)k + 21 supy∈Q1 k∇g(y)k. Taking Lr -average over Q1

gives (14.66) for ε = 1.
Step 3 – By Step 2, applied to the function f (x − z − ·),

ε Z 1/r
kf (x − z)k 6 sup k∇f (x − z − y)k + − kf (x − z − y)kr dy .
2 y∈Qε Qε
(14.67)

Now let ε ∈ (0, 1]. It follows from z − Qε ⊆ Q1+|z| that

Z Z
− kf (x − z − y)kr dy = − kf (x − y)kr dy
Qε z−Qε
|Q1+|z| |
Z
6 − kf (x − y)kr dy
|Qε | Q1+|z|
6 ε−d (1 + |z|)d M (kf kr )(x).

Substituting this into (14.67) and dividing by (1 + |z|)d/r , it follows that

kf (x − z)k ε k∇f (x − z − y)k

sup 6 sup sup + ε−d/r Mr f (x)
z∈Rd (1 + |z|)d/r 2 z∈Rd y∈Qε (1 + |z|)d/r
k∇f (x − z)k
6 ε2d/r−1 sup + ε−d/r Mr f (x),
z∈Rd (1 + |z|)d/r

where we used that (1 + |z|) > 21 (1 + |y + z|) for |y| 6 ε 6 1 and performed a
change of variables. Combining this estimate with the first inequality in the
statement of the lemma, and taking ε ∈ (0, 1] small enough, the result follows.
Step 4 – Next let f ∈ S 0 (Rd ; X) and t > 0. Let fδ = ψ(δ·)f , where
ψ ∈ S (Rd ) satisfies ψ(0) = 1, supp ψb ⊆ {ξ ∈ Rd : |ξ| 6 1} and δ ∈
(0, min{1, t}). Recalling that f ∈ C ∞ (Rd ; X), clearly we have fδ ∈ S (Rd ; X),
fbδ has support in B2t and therefore, by the previous steps, the second in-
equality in the statement of the lemma holds if in the two expressions on the
left-hand side f is replaced by fδ and for the right-hand side we note that
Mr fδ (x) 6 kψk∞ Mr f (x). It remains to let δ → 0 on the left-hand side and
note that fδ (x − z) → f (x − z) and similarly for its derivatives. 
Using the pointwise estimate of Lemma 14.6.1, we will now deduce a maximal
inequality in Lp (Rd ; `q ).

Proposition 14.6.2 (Boundedness of Peetre’s maximal function).

Let p ∈ [1, ∞), q ∈ [1, ∞], and let r ∈ (0, min{p, q}). Let f = (fk )k>0 in
Lp (Rd ; `q (X)) be such that supp(fbk ) ⊆ Sk for all k > 0, where Sk ⊆ Rd is a
368 14 Function spaces

compact set with diameter δk > 0. There exists a constant C > 0, depending
only on d, p, q, r, such that
 kfk (· − z)k 
sup 6 Ckf kLp (Rd ;`q (X)) .
z∈Rd (1 + δk |z|)d/r k>0 Lp (Rd ;`q )

∗ ∗
Proof. We use the short-hand notation f = (fk )k>0 and fd/r = (fk,d/r )k>0 ,
where

∗ kfk (x − z)k
fk,d/r (x) = sup , x ∈ Rd . (14.68)
z∈Rd (1 + δk |z|)d/r

Multiplying fk (x) with e2πihk ·x for suitable hk ∈ Rd , we may assume that

each fbk has support in Bk = {ξ ∈ Rd : |ξ| 6 δk } for k > 0.
Let gk (x) := fk (δk−1 x). Then gbk has support in a ball of radius 1 centred
around the origin. Thus by Lemma 14.6.1 there is a constant c, depending
only on d and r, such that for all k > 0 and x ∈ Rd we have

kgk (x − z)k
sup 6 cMr gk (x).
z∈Rd (1 + |z|)d/r

Rewriting this in terms of fk gives

∗ kfk (x − z)k
fk,d/r (x) = sup 6 cMr fk (x).
z∈Rd (1 + δk |z|)d/r

Taking Lp (Rd ; `q ) norms and applying the Fefferman–Stein maximal Theorem

3.2.28 in the space Lp/r (Rd ; `q/r ), we find that

∗ 1/r
kfd/r kLp (Rd ;`q ) 6 ck(Mr fk )k>0 kLp (Rd ;`q ) = c (M (kfk kr ))k>0 Lp/r (Rd ;`q/r )
1/r
.p,q,r c (kfk kr )k>0 Lp/r (Rd ;`q/r )
= ckf kLp (Rd ;`q (X)) .

As a first application we derive a Fourier multiplier theorem for certain func-

tions in Lp (Rd ; `q ) for p ∈ [1, ∞) and q ∈ [1, ∞] which is essential for later con-
siderations about Triebel–Lizorkin spaces. The main difficulty arises if p = 1
or q = 1 since the maximal function is not bounded in these cases. The case
q = 1 turns out to be of particular importance in Section 14.7.a.
The statement of the following theorem, which is needed in the proof of the
Mihlin multiplier theorem for Triebel–Lizorkin spaces (theorem 14.6.11) b is ad-
mittedly somewhat technical. We recall from Subsection 2.4.a that L1 (Rd ; X)
denotes the subspace in L∞ (Rd ; X) of all functions whose inverse Fourier
transform belongs to L1 (Rd ; X).
 14.6 Triebel–Lizorkin spaces 369

Theorem 14.6.3. Let X and Y be Banach spaces and let p ∈ [1, ∞), q ∈
[1, ∞], and r ∈ (0, min{p, q}). Let Sk ⊆ Rd , k > 0, be compact
b sets with
diameter δk > 0. Then for all sequences m = (mk )k>0 in L1 (Rd ; L (X, Y ))
and all f = (fk )k>0 ∈ Lp (Rd ; `q (X)) with supp fbk ⊆ Sk for each k > 0 we
have F −1 mF f ∈ Lp (Rd ; `q (Y )) and

k(F −1 mF f kLp (Rd ;`q (Y ))

6 C sup k(1 + δk | · |)d/r )F −1 mk (·)kL1 (Rd ;L (X,Y )) kf kLp (Rd ;`q (X))
k>0

= C sup k(1 + | · |)d/r )F −1 [mk (δk ·)]kL1 (Rd ;L (X,Y )) kf kLp (Rd ;`q (X))
k>0

where the constant C > 0 depends only on d, p, q, r, provided the supremum

on the right-hand side is finite.

Proof. The kernels Kk := F −1 mk are in L1 (Rd ; L (X, Y )) by assumption.

Therefore, the functions F −1 (mk fbk ) = Kk ∗ fk are well defined in Lp (Rd ; Y )
by Young’s inequality. Let

cm := sup k(1 + δk | · |)d/r Kk (·)kL1 (Rd ;L (X,Y )) .

k>0

Then, using the notation introduced in (14.68),

kfk (y)k
Z
kKk ∗ fk (x)k 6 kKk (x − y)k(1 + δk |x − y|)d/r dy
Rd (1 + δk |x − y|)d/r
Z
∗ ∗
6 fn,d/r (x) kKk (x − y)k(1 + δk |x − y|)d/r dy 6 cm fn,d/r (x).
Rd

The required result follows from this by taking Lp (Rd ; `q )-norms and applying
Proposition 14.6.2.
The final identity of the theorem simply follows by a substitution together
with the dilation property δk−1 (F −1 mk )(δk−1 ·) = F −1 [mk (δk ·)] of the Fourier
transform. 

Remark 14.6.4. Lemma 14.6.1 can be used to extend the Bernstein–Nikolskii

inequality presented in Lemma 14.4.20 to the full range 0 < p0 6 p1 6 ∞.
To this end let ψ be as in the proof of the lemma and note that it suffices to
consider the case that fb has support in the unit ball.
First consider 0 < p0 < p1 6 ∞ and α = 0. If p0 ∈ (0, 1) and p1 = ∞,
then
Z
|f (x)| 6 |ψ(x − y)|kf (y)k dy|
Rd
Z
6 kψk∞ kf (y)k1−p0 kf (y)kp0 dy| 6 kψk∞ kf k1−p
∞
0
kf kpp00
Rd
370 14 Function spaces

and consequently kf k∞ .p0 ,ψ kf kp0 . Since we already knew the result for
p0 > 1, this inequality holds for p0 ∈ (0, ∞). In the remaining case p0 < p1 <
∞, we similarly find that

kf kp1 6 kf k1−p
∞
0 /p1
kf kpp00 /p1 .p0 ,p1 ,ψ kf kp0 .

The case p0 = p1 and α 6= 0 follows by taking Lp1 -norms in the pointwise

estimate k∂ α f (x)k 6 CMr (f )(x) with r ∈ (0, p1 ) (see Lemma 14.6.1) and
using the Lp1 /r -boundedness of the Hardy–Littlewood maximal function, to
conclude that k∂ α f kp1 6 Ckf kp1 .
If p0 < p1 and α 6= 0 combining the previous two cases gives

k∂ α f kp1 6 Ckf kp1 6 C 0 kf kp0 .

14.6.b Definitions and basic properties

We now introduce our main characters. Recall that we have fixed a inhomoge-
neous Littlewood–Paley sequence (ϕk )k>0 in Subsection 14.2.c (see Conven-
tion 14.2.8).

Definition 14.6.5 (Triebel–Lizorkin spaces). Let p ∈ [1, ∞), q ∈ [1, ∞],

s
and s ∈ R. The Triebel–Lizorkin space Fp,q (Rd ; X) is the space of all f ∈
S (R ; X) for which the quantity
0 d

kf kFp,q
s (Rd ;X) := (2ks ϕk ∗ f )k>0 Lp (Rd ;`q (X))

is finite.

We comment on the case p = ∞ and q < ∞ in the Notes, as this exceptional

case behaves differently. Below we will check that the above definition is inde-
pendent on the choice of the Littlewood–Paley sequence up to an equivalent
norm and that the resulting spaces are Banach spaces.
s
It is immediate from Young’s inequality that ψ ∗ f ∈ Fp,q (Rd ; X) when-
1 d s d
ever ψ ∈ L (R ) and f ∈ Fp,q (R ; X), and more generally the analogue of
Proposition 14.2.3 in valid.
By Fubini’s theorem, for all p ∈ [1, ∞) we have
s
Fp,p (Rd ; X) = Bp,p
s
(Rd ; X).

We have continuous embeddings

s
Fp,q 0
s
(Rd ; X) ,→ Fp,q 1
(Rd ; X), 1 6 q0 6 q1 6 ∞, (14.69)

and, by Hölder’s inequality for the `q -norm,

s0
Fp,q 0
s1
(Rd ; X) ,→ Fp,q 1
(Rd ; X), q0 , q1 ∈ [1, ∞], s0 > s1 . (14.70)
 14.6 Triebel–Lizorkin spaces 371

Next we prove that, up to equivalence of norm, the Triebel–Lizorkin spaces

are independent of the choice of the inhomogeneous Littlewood–Paley se-
quence (ϕk )k>0 . The corresponding result for Besov spaces, Proposition 14.4.2,
was rather easy to prove. The case of Triebel–Lizorkin spaces is not so easy
and is based on Proposition 14.6.2. For p > 1 and q > 1 the use of this theo-
rem can be avoided by using instead the estimate kϕk ∗ f k 6 cM f together
with the Fefferman–Stein Theorem 3.2.28.
Proposition 14.6.6. Let p ∈ [1, ∞), q ∈ [1, ∞], and s ∈ R. Up to an equiva-
s
lent norm, the space Fp,q (Rd ; X) is independent of the choice of the inhomo-
geneous Littlewood–Paley sequence (ϕk )k>0 .
Proof. Fix inhomogeneous Littlewood–Paley sequences (ϕk )k>0 and (ψk )k>0 .
0 with |j − k| > 2 we have ψk ∗ ϕj = F −1 (ψbk ϕ
For all j, k >P bj ) = 0. Therefore,
writing f = j>0 fj with fj = ϕj ∗ f ,
1
X
(2ks ψk ∗ f )k>0 Lp (Rd ;`q (X))
6 (2ks ψk ∗ f`+k )k>0 Lp (Rd ;`q (X))
.
`=−1

Fix an arbitrary r ∈ (0, min{p, q}), say r = rp,q = 21 min{p, q}. Applying
Theorem 14.6.3 with δk = 3 · 2k and mk = ψbk to (2ks f`+k )k>0 we obtain
(2ks ψk ∗ f )k>0 Lp (Rd ;`q (X))
6 Cψ,d,p,q,s (2ks f`+k )k>0 Lp (Rd ;`q (X))
0 ks
6 Cψ,d,p,q,s (2 ϕk ∗ f )k>0 Lp (Rd ;`q (X))
.

Since (ψk )k>0 and (ϕk )k>0 were arbitrary, this completes the proof. 
The same argument and (14.5) lead to the following useful estimate.
s
Lemma 14.6.7. Let f ∈ Fp,q (Rd ; X), let (ψk )k>0 be a Littlewood–Paley se-
quence, and set
X n
Sn f := ψk ∗ f, n > 0.
k=0
s
Then Sn f ∈ Fp,q (Rd ; X) and there exists a constant C = C(p, q, d, ψ) such
that
kSn f kFp,q
s (Rd ;X) 6 Ckf kF s (Rd ;X) ,
p,q
n > 0.
We have the following analogue of Proposition 14.4.18 for Triebel–Lizorkin
spaces:
Proposition 14.6.8 (Sandwiching with Besov spaces). For all p ∈
[1, ∞), q ∈ [1, ∞], and s ∈ R, we have the natural continuous embeddings
S (Rd ; X) ,→ Fp,q
s
(Rd ; X) ,→ S 0 (Rd ; X),
the first of which is dense if p, q ∈ [1, ∞), and
s
Bp,p∧q (Rd ; X) ,→ Fp,q
s
(Rd ; X) ,→ Bp,p∨q
s
(Rd ; X).
372 14 Function spaces

By combining the first of these inclusions with Lemma 14.2.1 we see that if
p, q ∈ [1, ∞), then Cc∞ (Rd ) ⊗ X is dense in Fp,q
s
(Rd ; X).
s
Proof. First let p > q. For f ∈ Bp,q (Rd ; X) it follows from the triangle in-
equality in Lp/q (Rd ) that
X
kf kqF s (Rd ;X)
= 2ksq kϕk ∗ f kq
p,q Lp/q (Rd )
k>0
X q q
6 2ksq kϕk ∗ f kL p (Rd ;X) = kf kB s d ;X) .
p,q (R
k>0

This gives the first embedding in the second displayed line of the proposi-
s
tion. The second embedding follows from (14.69), which gives Fp,q (Rd ; X) ,→
s d s d
Fp,p (R ; X) = Bp,p (R ; X) continuously. The case p 6 q is handled similarly.
The continuous embeddings in the first line now follow from the corre-
sponding result for Besov spaces contained in Proposition 14.4.3.
Let us finally show that S (Rd ; X) is dense in Fp,q
s
(Rd ; X). The proof is
s
similar toPStep 3 of the proof of Proposition 14.4.3. Let f ∈ Fp,q (Rd ; X) and
n
set ζn := k=0 ϕk . By (14.6) we have kζn k1 6 kϕ0 k1 .
s
We will first show that ζn ∗ f → f in Fp,q (Rd ; X). Let ε > 0 and choose
K > 0 such that
X 1/q
2ksq kϕk ∗ f kq < ε.
Lp (Rd )
k>K

By Young’s inequality,

ζn ∗ (2ks ϕk ∗ f )k>K < εkϕ0 k1 .

Lp (Rd ;`q (X))

It follows that

kf − ζn ∗ f kFp,q
s (Rd ;X)

K
X 1/q
6 ε(1 + kϕ0 k1 ) + 2ksq kϕk ∗ f − ζn ∗ ϕk ∗ f kq
Lp (Rd )
k=0
K
X
6 ε(1 + kϕ0 k1 ) + 2ks kϕk ∗ f − ζn ∗ ϕk ∗ f kLp (Rd ;X)
k=0

The last term tends to zero as n → ∞ by Proposition 1.2.32.

It remains to approximate each of the functions ζn ∗ f by elements in
S (Rd ; X). This can be done as in Proposition 14.4.3. 
This result enables us to give a quick proof of the completeness of Triebel–
Lizorkin spaces:
 14.6 Triebel–Lizorkin spaces 373

Proposition 14.6.9. For p ∈ [1, ∞), q ∈ [1, ∞], and s ∈ R, the space
s
Fp,q (Rd ; X) is a Banach space.

Proof. As in the Besov case one proves that for all p ∈ [1, ∞), q ∈ [1, ∞],
s
and s ∈ R, the space Fp,q (Rd ; X) has the Fatou property. Since Triebel–
Lizorkin spaces embed into S 0 (Rd ; X) by Proposition 14.6.8, the completeness
s
of Fp,q (Rd ; X) follows from Lemma 14.4.7. 

14.6.c Fourier multipliers

The main result of this subsection is a version of the Mihlin multiplier theorem
for Triebel–Lizorkin spaces. Before we state it we first prove an important
lifting property as we saw in Proposition 14.4.15 for Besov spaces.

Proposition 14.6.10 (Lifting). Let p ∈ [1, ∞), q ∈ [1, ∞], and s ∈ R. Then
for all σ ∈ R,
s
Jσ : Fp,q s−σ
(Rd ; X) ' Fp,q (Rd ; X) isomorphically. (14.71)
s
Proof. As in Proposition 14.4.15 it suffices to show that Jσ maps Fp,q (Rd ; X)
s−σ d
into Fp,q (R ; X) and is bounded for each σ ∈ R We must show that
(2n(s−σ) ϕn ∗ Jσ f )n>0 belongs to Lp (Rd ; `q (X)). This will be done by ap-
plying the multiplier Theorem 14.6.3 to a multiplier m = (mn )n>0 naturally
associated with Jσ .
Write
X 1
2−nσ ϕn ∗ Jσ f = F −1 mn ϕ bn+` fb,
`=−1

where
mn (ξ) = 2−nσ (1 + 4π 2 |ξ|2 )σ/2 ϕ
bn (ξ).
We have mn ∈ C ∞ (Rd ) and, putting δn = 3 · 2n ,
1 1
bn (δn ·) ⊆ ξ ∈ Rd : 6 |ξ| 6

supp ϕ , (n > 1)
6 2
1
b0 (δ0 ·) ⊆ ξ ∈ Rd : |ξ| 6

supp ϕ .
2
Lemma 14.2.12, applied with λ = d + 1 + dd/re with an arbitrary r = rp,q ∈
(0, min{p, q}), gives the estimate

k(1 + | · |)d/r F −1 [mn (δn ·)]kL1 (Rd ;L (X,Y ))

6 Cd kmn (δn ·)kC d+1+dd/re (Rd ;L (X,Y )) 6 Cm,d,r = Cm,d,p,q ,

where the last inequality is elementary to verify.

Since for ` ∈ {−1, 0, 1} we have supp(ϕbn+` fb) ⊆ {ξ ∈ Rd : |ξ| 6 δn } we are
now in a position to apply Theorem 14.6.3 and obtain
374 14 Function spaces

(2n(s−σ) ϕn ∗ Jσ f )n>0 Lp (Rd ;`q (X))

1
X
6 k(F −1 mn 2ns ϕ
bn+` fb)n>0 kLp (Rd ;`q (X))
`=−1
1
X
6 Cm,d,p,q k(2ns ϕn+` ∗ f )n>0 kLp (Rd ;`q (X))
`=−1
0
6 Cm,d,p,q kf kFp,q
s (Rd ;X) .


We continue with the Mihlin multiplier theorem for Triebel–Lizorkin spaces.
Note that the Besov space case was considered in Theorems 14.4.16 and 14.5.6.

Theorem 14.6.11 (Mihlin multiplier theorem for Triebel–Lizorkin

spaces). Let p ∈ [1, ∞), q ∈ [1, ∞], s ∈ R, and X and Y be Banach spaces,
and set N := d + 1 + dmax{ dp , dq }e. If m ∈ C N (Rd ; L (X, Y )) satisfies

Km := sup sup (1 + |ξ||α| )k∂ α m(ξ)kL (X,Y ) < ∞,

|α|6N ξ∈Rd

s
then there is a bounded operator T : Fp,q (Rd ; X) → Fp,q
s
(Rd ; Y ) with kT k 6
Cd,p,q,s,X,Y Km such that T f = F (mfb) for all f ∈ S (Rd ) ⊗ X.
−1

Note that in the case q < ∞, one has that T is the unique bounded extension
of Tm : S (Rd ) ⊗ X → S 0 (Rd ; Y ).
Proof. We define T in the same was as in (14.59) of the proof of Theorem
14.5.6:
1 X
X
Tf = Tmk+` fk ,
`=−1 k>0

s
where f ∈ Fp,q (Rd ; X), fk = ϕk ∗ f and mk = ϕ s
bk m. Since Fp,q (Rd ; X) ⊆
s
Bp,∞ (Rd ; X) it follows from the proof of Theorem 14.5.6 that the above series
converges in S 0 (Rd ; Y ), and that T g = F −1 (mb
g ) for all g ∈ S (Rd ) ⊗ X.
To prove the required boundedness, note that
1
X
kTm f kFp,q
s (Rd ;Y ) 6 2ks F −1 (mϕ
bk+` ϕ
bk fb)k>0 Lp (Rd ;`q (Y ))
.
`=−1

Fix ` ∈ {−1, 0, 1}. Then supp fbk+` ⊆ {|ξ| 6 δk }, where δk = 3 · 2k .

To estimate further it is sufficient to apply Theorem 14.6.3, for which we
choose r = rd,p,q ∈ (0, min{p, q}) such that N = d + 1 + dd/re. To check the
assumptions of the theorem we have to show that
 14.6 Triebel–Lizorkin spaces 375

sup k(1 + | · |)d/r F −1 (ϕ

bk (δk ·)m(δk ·))kL1 (Rd ;L (X,Y )) 6 CKm ,
k>0

where C > 0 is a constant depending only on d and r. Since ϕ bk (δk ·)m(δk ·)

has support in {ξ ∈ Rd : |ξ| 6 1}, the estimate follows from Lemma 14.2.12.

The following result is proved in the same way as Proposition 14.4.17.

Proposition 14.6.12. Let p ∈ [1, ∞), q ∈ [1, ∞], and s ∈ R. For all k ∈ N
the expression
X
|||f |||Fp,q
s (Rd ;X) := k∂ α f kFp,q
s−k
(Rd ;X)
|α|6k

s
defines an equivalent norm on Fp,q (Rd ; X).

14.6.d Embedding theorems

We have already noted the continuous inclusions

S (Rd ; X) ,→ Fp,q
s
(Rd ; X) ,→ S 0 (Rd ; X)

and
s
Bp,p∧q (Rd ; X) ,→ Fp,q
s
(Rd ; X) ,→ Bp,p∨q
s
(Rd ; X)
for s ∈ R, p ∈ [1, ∞) and q ∈ [1, ∞]. Moreover, for any q ∈ [1, ∞], it is
immediate from the definitions that
s
Bp,1 (Rd ; X) ,→ Fp,q
s s
(Rd ; X) ,→ Bp,∞ (Rd ; X). (14.72)

The next result compares Triebel–Lizorkin spaces with the Bessel potential
and Sobolev spaces. It can be improved if X is UMD and has type and cotype
properties (see Proposition 14.7.6 below).

Proposition 14.6.13 (Sandwiching with Triebel–Lizorkin spaces). For

p ∈ (1, ∞), s ∈ R, and m ∈ N, we have the following continuous embeddings:
s
Fp,1 (Rd ; X) ,→ H s,p (Rd ; X) ,→ Fp,∞
s
(Rd ; X), (14.73)
m
Fp,1 m
(Rd ; X) ,→ W m,p (Rd ; X) ,→ Fp,∞ (Rd ; X). (14.74)
s
In view of the embeddings Bp,1 s
(Rd ; X) ,→ Fp,1 s
(Rd ; X) and Fp,∞ (Rd ; X) ,→
s d
Bp,∞ (R ; X), (14.73) and (14.74) improve the corresponding embeddings in
Proposition 14.4.18.
376 14 Function spaces

Proof. For (14.73) and (14.74), by Propositions 5.6.3, 14.6.10 and 14.6.12
it suffices to consider the special case s = m = 0, for which H 0,p (Rd ; X) =
W 0,p (Rd ; X) = Lp (Rd ; X). It thus remains to show the continuous embeddings
0
Fp,1 (Rd ; X) ,→ Lp (Rd ; X) ,→ Fp,∞
0
(Rd ; X). (14.75)

P The first embedding in (14.75) is true for any p ∈ [1, ∞): writing f =
k>0 ϕk ∗ f it follows that
X
kf kLp (Rd ;X) 6 kϕk ∗ f kLp (Rd ;X) = kf kFp,1
0 (Rd ;X) .

k>0

For the second embedding in (14.75) observe that since ϕ ∈ S (Rd ), it has a
radially decreasing majorant which is integrable. Therefore, by Theorem 2.3.8
there is a constant Cd > 0 such that for all k > 0 and almost all x ∈ Rd ,
kϕk ∗ f (x)k 6 Cd M f (x). Therefore, by the Lp -boundedness of the Hardy–
Littlewood maximal function (Theorem 2.3.2),

kf kFp,∞
0 (Rd ;X) = sup kϕk ∗ f k Lp (Rd )
6 Cd kM f kLp (Rd ) .p Cd kf kLp (Rd ;X) .
k>0

This completes the proof. 

We continue with a version of the Sobolev embedding theorem. A surprising
feature is that in case of the Triebel–Lizorkin spaces there is an improvement
in the microscopic parameter q.

Theorem 14.6.14 (Sobolev embedding for Triebel–Lizorkin spaces).

For given p0 , p1 ∈ [1, ∞), q0 , q1 ∈ [1, ∞], and s0 , s1 ∈ R, we have a continuous
embedding
Fps00,q0 (Rd ; X) ,→ Fps11,q1 (Rd ; X)
if and only if one of the following two conditions holds:
(i) p0 = p1 and [s0 > s1 or (s0 = s1 and q0 6 q1 )];
(ii) p0 < p1 and s0 − pd0 > s1 − pd1 .

The main ingredient is a version of the Gagliardo–Nirenberg inequality with

a microscopic improvement.
Proposition 14.6.15 (Gagliardo–Nirenberg inequality for Triebel–
Lizorkin spaces). Let p, p0 , p1 ∈ [1, ∞), q, q0 , q1 ∈ [1, ∞], let s0 , s1 ∈ R with
s0 < s1 , let θ ∈ (0, 1), and assume that p1 = 1−θ θ
p0 + p1 and s = (1 − θ)s0 + θs1 .
For all f ∈ Fps00,q0 (Rd ; X) ∩ Fps11,q1 (Rd ; X) we have f ∈ Fp,q
s
(Rd ; X) and
1−θ
kf kFp,q
s (Rd ;X) 6 Ckf k s0
F d ;X) kf kθFps1,q (Rd ;X)
,
p0 ,q0 (R 1 1

where the constant C > 0 depends only on θ, s0 , s1 .

 14.6 Triebel–Lizorkin spaces 377

Proof. Proposition 14.3.5 (applied with q0 = q1 = ∞) implies that

1−θ θ
(2ks ak )k>0 `q
6 Cs0 ,s1 ,s (2ks0 ak )k>0 `∞
(2ks1 ak )k>0 `∞
(14.76)

for all sequences of scalars (ak )k>0 for which the expression on the right-hand
side is finite.
To prove the desired estimate, by (14.69) it suffices to consider the case
q0 = q1 = ∞. Taking ak (x) = kϕk ∗ f (x)k with x ∈ Rd in (14.76), raising to
the power p and integrating over Rd , by Hölder’s inequality (with exponents
p0 p1
(1−θ)p and θp ) we obtain

1−θ
kf kFp,q
s (Rd ;X) 6 Cs0 ,s1 ,s kf k s0
F d ;X) kf kθFps1,∞ (Rd ;X)
p0 ,∞ (R 1

as required. 
In a similar way one can prove the following variant for the end-point p1 = ∞.

Proposition 14.6.16 (Gagliardo–Nirenberg inequality for Triebel–

Lizorkin spaces – II). Let p, p0 , ∈ [1, ∞), q, q0 ∈ [1, ∞], let s0 , s1 ∈ R
with s0 < s1 , let θ ∈ (0, 1), and assume that p1 = 1−θ
p0 and s = (1 − θ)s0 + θs1 .
s0 d s1 d s
For all f ∈ Fp0 ,q0 (R ; X) ∩ B∞,∞ (R ; X) we have f ∈ Fp,q (Rd ; X) and
1−θ
kf kFp,q
s (Rd ;X) 6 Ckf k s0
F d ;X) kf kθB∞,∞
s1
(Rd ;X)
,
p0 ,q0 (R

where the constant C > 0 depends only on θ, s0 , s1 .

Proof of sufficiency in Theorem 14.6.14. For the sufficiency of (i) first assume
that p0 = p1 , q0 6 q1 , and s0 > s1 . Under these assumptions the result follows
from the fact that

(2ks1 ak )k>0 `q 1
6 (2ks0 ak )k>0 `q 0
.

If p0 = p1 , q0 > q1 , and s0 > s1 , the result follows from (14.23) and (14.72):

Fps00,q0 (Rd ; X) = Fps10,q0 (Rd ; X) ,→ Bps10 ,∞ (Rd ; X)

,→ Bps11 ,∞ (Rd ; X) ,→ Fps11,q1 (Rd ; X).

This completes the proof of (i).

Let us now assume that (ii) holds. By (14.70) it suffices to consider the
case s0 − pd0 = s1 − pd1 . By (14.69) we may furthermore assume that q1 = 1.
First take f ∈ S (Rd ; X). Let θ0 ∈ [0, 1) be such that p11 − 1−θp0
0
= 0.
1 1−θ θ
Choose θ ∈ (θ0 , 1) arbitrary and let r be defined by p1 = p0 + r . Note
that p0 < p1 < ∞ implies r ∈ (p1 , ∞). Let further t ∈ R be defined by
t − dr = s0 − pd0 . Observe that t < s0 and s1 = θt + (1 − θ)s0 (write out
the expression for θt and use the formula for θ/r). Therefore, by Proposition
14.6.15,
378 14 Function spaces

kf kFps1,1 (Rd ;X) 6 Cs0 ,s1 ,θ kf k1−θ

s
F 0 d ;X)
θ
kf kF t d
r,r (R ;X)
. (14.77)
1 p0 ,q0 (R

By the case (ii) in Theorem 14.4.19 (using that r > p1 ),

kf kFr,r
t (Rd ;X) = kf kB t (Rd ;X) 6 Ckf k s1
r,r Bp ,p (Rd ;X)
1 1

= Ckf kFps1,p (Rd ;X) 6 Ckf kFps1,1 (Rd ;X) ,

1 1 1

where in the last step we used (14.69). Substituting the latter estimate into
(14.77), we obtain
1/(1−θ)
kf kFps1,1 (Rd ;X) 6 Cs0 ,s1 ,θ C θ/(1−θ) kf kFps0,q (Rd ;X) . (14.78)
1 0 0

Now if q0 < ∞, then the result follows from the density of S (Rd ; X) in
Fps00,q0 (Rd ; X).
Pn
If q0 = ∞ and f ∈ Fps00,∞ (Rd ; X), we let Sn f = k=0 ϕk ∗ f . Then by
Young’s inequality and the fact that ϕj ∗Sn f = 0 for j > n+1, we have Sn f ∈
Bps00 ,1 (Rd ; X). Thus Theorem 14.4.19 implies Sn f ∈ Bps11 ,1 (Rd ; X). More-
over, by Proposition 14.6.8 and (14.69) we also have Sn f ∈ Fps00,1 (Rd ; X) ,→
Fps00,∞ (Rd ; X) and Sn f ∈ Fps11,1 (Rd ; X). Therefore, by (14.78),

kSn f kFps1,1 (Rd ;X) 6 Cs1/(1−θ)

0 ,s1
C θ/(1−θ) kSn f kFps0,∞ (Rd ;X) 6 C̃kf kFps0,∞ (Rd ;X) ,
1 0 0

where the last estimate follows from Lemma 14.6.7. Since Sn f → f in

S 0 (Rd ; X) by Lemma 14.2.10, the assertion now follows from the fact that
s
Fp,q (Rd ; X) has the Fatou property. 
Proof of necessity in Theorem 14.6.14. By Proposition 14.6.8,

Bps00 ,1 (Rd ; X) ,→ Fps00,q0 (Rd ; X) ,→ Fps11,q1 (Rd ; X) ,→ Bps11 ,∞ (Rd ; X).

Therefore, Theorem 14.4.19 implies that p0 6 p1 . If p0 = p1 , then (i) follows

from (i). If p0 < p1 , then (ii) follows from (iii) and (ii). 
Proposition 14.4.36 has the following analogue for Triebel–Lizorkin spaces:

Proposition 14.6.17 (Density of compactly supported functions). Let

R̈d := R \ {0} × Rd−1 .

Let p, q ∈ [1, ∞) and s ∈ R. Then Cc∞ (R̈d ) ⊗ X is dense in Fp,q

s
(Rd ; X) and
s,p d
H (R ; X) in each of the following situations:
(1) s < 1/p;
(2) p ∈ (1, ∞) and s = 1/p.
 14.6 Triebel–Lizorkin spaces 379

Proof. First consider the Triebel–Lizorkin case. As in the proof of Proposition

14.4.36 (using Propositions 14.6.8) we can reduce to the smooth and scalar-
valued setting. Thus it suffices to show that an arbitrary f ∈ Cc∞ (Rd ) there
1/p
exist functions fn ∈ Cc∞ (R̈d ) such that fn → f in Fp,q (Rd ). By the embedding
(14.70) and Theorem 14.6.14, it suffices to prove this for the case (2). However,
for this case the result follows from Proposition 14.4.36 and the estimate

kf − fn kF 1/p (Rd ) 6 Ckf − fn kF 1/r (Rd ) = Ckf − fn kB 1/r (Rd ) , r ∈ (1, p),
p,q r,r r,r

which follows from Theorem 14.6.14.

The same proof for Bessel potential spaces holds, where we note that
for the reduction to the scalar situation one can use Proposition 5.6.4, and
1/r
the embedding Fr,r (Rd ) ,→ H 1/p,p (Rd ) follows from Proposition 14.6.13 and
Theorem 14.6.14. 
The proof of Theorem 14.5.1 shows that the existence of a continuous embed-
ding
( 1 − 12 )d
p
Bp,r (Rd ; X) ,→ γ(L2 (Rd ), X)
implies that X has type r, and that the existence of a continuous embedding
( 1 − 1 )d
γ(L2 (Rd ), X) ,→ Bq,r q 2
(Rd ; X) implies that X has cotype r. Therefore
in the Besov scale the embeddings of Theorem 14.5.1 cannot be improved
by using the microscopic parameter r. For the Triebel–Lizorkin spaces the
situation is different, as witnessed the following result.

Corollary 14.6.18 (γ-Sobolev embedding – II). Let 1 6 p0 6 2 6 q0 <

∞.
(1) If X has type p0 , then for all p ∈ [1, p0 ) and all r ∈ [1, ∞] we have a
continuous embedding
( 1 − 12 )d
Fp,rp (Rd ; X) ,→ γ(L2 (Rd ), X).

(2) If X has cotype q0 , then for all q ∈ (q0 , ∞) and all r ∈ [1, ∞] we have a
continuous embedding
( 1 − 12 )d
γ(L2 (Rd ), X) ,→ Fq,rq (Rd ; X).

Proof. We give the proof of (1), the proof of (2) being similar. Let 1 6 p < p0 .
Let s0 = ( p10 − 12 )d and s = ( p1 − 12 )d. By Theorem 14.6.14 we have a continuous
embedding
s
Fp,r (Rd ; X) ,→ Fps00,p0 (Rd ; X) = Bps00 ,p0 (Rd ; X).
Now the result follows from Theorem 14.5.1. 
380 14 Function spaces

14.6.e Difference norms

In Section 14.4.d we have discussed a difference norm characterisation for

Besov spaces of positive smoothness. We will now prove a similar result for
the Triebel–Lizorkin spaces. Recall the notation

∆h f (x) = f (x + h) − f (x)

and ∆m m
h = (∆h ) .

Definition 14.6.19 (Difference norm for Triebel–Lizorkin spaces).

Let p ∈ [1, ∞), q ∈ [1, ∞], s > 0, m ∈ N \ {0} and τ ∈ [1, ∞). For f ∈
Lp (Rd ; X) we define the difference norm by setting
Z ∞ Z q/τ dt 1/q
(m,τ )
[f ]F s (Rd ;X) := t−sq − k∆hm f kτX dh ,
p,q
0 {|h|6t} t Lp (Rd )

with obvious modifications if q = ∞, and

(m,τ ) (m,τ )
|||f |||F s d := kf kLp (Rd ;X) + [f ]F s d .
p,q (R ;X) p,q (R ;X)

(m,τ )
It will be shown shortly that each of the norms ||| · |||F s d with m > s and
p,q (R ;X)
d d s
s> min{p,q} − defines an equivalent norm on
τ Fp,q (Rd ; X).
The expression for the seminorm simplifies for τ = q ∈ [1, ∞). Indeed, by
Fubini’s theorem we have
1 Z 1/q
(m,q) −(s+d)q m q
[f ]F s (Rd ;X) = |h| k∆ h f (x)k dh .
p,q (sq + d)1/q |B1 | Rd Lp (Rd )

Theorem 14.6.20 (Difference norms for Triebel–Lizorkin spaces). Let

X be a Banach space and let p, τ ∈ [1, ∞), q ∈ [1, ∞], s > 0, let m > s be an
integer, and suppose that
d d
s> − . (14.79)
min{p, q} τ

Then for all f ∈ S (Rd ; X) the following norm equivalence holds:

(m,τ )
kf kFp,q
s (Rd ;X) hd,m,p,q,s,τ |||f ||| s
F d , (14.80)
p,q (R ;X)

whenever one of these expressions is finite.

Note that the condition (14.79) holds trivially holds if τ 6 min{p, q}, and in
particular if τ = 1. The condition (14.79) is only used in the proof of “&” of
(14.80).
(m,τ )
For the proof we will use a discretised version of |||f |||F s (Rd ;X) . Put
p,q
 14.6 Triebel–Lizorkin spaces 381
Z 1/τ
m,τ
J (f, k)(x) := − k∆2m−k h f (x)kτ dh .
|h|61

As in (14.38) we have
(m,τ )
[f ]F s d hd,s (2ks J m,τ (f, k))k∈Z Lp (Rd ;`q (Z))
.
p,q (R ;X)

Therefore, to obtain (14.80) it suffices to prove the norm equivalence

s (Rd ;X) h kf kLp (Rd ;X) +

kf kFp,q (2ks J m,τ (f, k))k∈Z Lp (Rd ;`q (Z))
, (14.81)

where the implicit constant may depend on d, p, q, m, s, τ . The proof of the

estimate . in (14.81) is similar to Step 2 of the proof of Theorem 14.4.24
except that instead of Proposition 14.4.2 one has to use Proposition 14.6.6
and towards the end of the proof one has to take Lp (Rd ; `q )-norms instead of
`q (Lp (Rd ))-norms.
In the remainder of this subsection we will concentrate on proving the in-
equality & in (14.81). We begin with a lemma involving the maximal function

Mr := (M (kf kr )(x))1/r

introduced in (14.65).

Lemma 14.6.21. Let f ∈ S 0 (Rd ; X) satisfy supp(fb) ⊆ {|ξ| 6 t}. Then f ∈

C ∞ (Rd ; X) and for all r ∈ (0, ∞), m ∈ N, and all x, h ∈ Rd we have

k∆m
h f (x)k .d,m,r (t|h|)
d/r
Mr (f )(x) if |h| > t−1 ; (14.82)
m m
k∆h f (x)k .d,m,r (t|h|) Mr (f )(x) if |h| 6 t−1 . (14.83)

Proof. That f belongs to C ∞ (Rd ; X) follows from Lemma 14.2.9.

Recall that by Lemma 14.6.1

k∂ α f (x + h)k .|α|,d,r t|α| (1 + t|h|)d/r Mr f (x). (14.84)

The estimate (14.82) follows from (14.84) and Lemma 14.4.22, for if |h| > t−1 ,
then
m  
m
X m
k∆h f (x)k 6 kf (x + hj)k
j=0
j

.d,r 2m (1 + t|h|m)d/r Mr f (x) .d,m,r (t|h|)d/r Mr f (x).

To prove (14.83) fix |h| 6 t−1 . Set φ(s) := f (x + sh) for s ∈ R. Then
∆hm f (x) = ∆m 1
1 φ(0). Since for any g ∈ C (R; X) we have k∆1 g(s)k 6
0
supθ∈[s,s+1] kg (θ)k, an induction argument gives

k∆1m φ(s)k 6 sup kφ(m) (θ)k, s ∈ R.

θ∈[s,s+m]
382 14 Function spaces

In particular,
 X 1/2
k∆hm f (x)k 6 sup kφ(m) (θ)k 6 |h|m sup k∂ α f (x + θh)k2 .
θ∈[0,m] θ∈[0,m]
|α|=m

By (14.84) and the fact that t|h| 6 1, for θ ∈ [0, m] we have

k∂ α f (x + θh)k .d,m,r tm (1 + tm|h|)d/r Mr f (x) .d,r,m tm Mr f (x).

Substituting this into the previous estimate gives the required estimate. 
Proof of Theorem 14.6.20. It remains to prove the inequality & in (14.81).
To begin with, from (i) we have inequality
∞
X
kf kLp (Rd ;X) 6 kfj kX 0 (Rd ;X) .d,p,q,s kf kF s (Rd ;X) ,
= kf kFp,1 p,q
Lp (Rd )
j=0

where fj = ϕj ∗ f as always.
To deal with the seminorm, note that from the assumption (14.79) it fol-
lows that we can find r ∈ (0, ∞) and λ ∈ (0, 1] such that

p, q > max{r, λτ } and s > (1 − λ)d/r. (14.85)

Since f = n∈Z fn+k in Lp (Rd ; X) for any k ∈ Z (recall the convention

P
that we set ϕj = 0 for j 6 −1, so that we may put fj = 0 for j 6 1), we have
X
(2ks J m,τ (f, k))k∈Z Lp (Rd ;`q (Z)) 6 (2ks J m,τ (fn+k , k))k∈Z Lp (Rd ;`q (Z)) .
n∈Z

For n 6 0, by (14.83) with t = 2k+n+1 , we have

Z 1/τ
m,τ
J (fn+k , k)(x) = − k∆2m−k h fn+k (x)kτ dh
|h|61
Z 1/τ
.d,m,r − (|2n h|m Mr (fn+k )(x))τ dh
|h|61
nm
62 Mr (fn+k )(x),

and therefore

(2ks J m,τ (fn+k , k)(x))k>0 `q (Z)

.d,m,r (2ks 2nm Mr fn+k (x))k>0 k`q (Z)

= 2n(m−s) (2s(k+n) 2nm Mr fn+k (x))k>0 `q (Z)
.

Since s < m and Mr is bounded on Lp (Rd ; `q ) by the Fefferman–Stein maximal

Theorem 3.2.28, we obtain
 14.6 Triebel–Lizorkin spaces 383
X
(2ks J m,τ (fn+k , k > 0))k Lp (Rd ;`q (Z))
n60
X
.d,m,r 2n(m−s) (2(k+n)s 2nm Mr fn+k )k>0 Lp (Rd ;`q (Z))
n60

.d,m,s kf kFp,q
s (Rd ;X) .

Next take n > 1. Fixing λ ∈ (0, 1] for the moment, we have

J m,τ (fn+k , k)(x)

Z 1/τ
6 sup k∆2m−k h fn+k (x)k1−λ − k∆m
2−k h fn+k (x)k
τλ
dh
|h|61 {|h|61}

=: T1 (x) × T2 (x).

From (14.82) we obtain the pointwise bound

T1 (x) 6 2dn(1−λ)/r Mr (fn+k )(x)1−λ .

Pm
To estimate T2 , we use Lemma 14.4.22 and the inequality ( j=1 |aj |)λτ .λ,m,τ
Pm λτ
j=1 |aj | to obtain
m  
X m
k∆2m−k h fn+k (x)kτ λ .λ,m,τ kfn+k (x)kτ λ + kfn+k (x + 2−k hj)kτ λ .
j=1
j

Estimating both terms by the maximal function, we obtain the pointwise

bound that T2 (x) is less than a constant depending on λ, m, τ times
m  Z 1/τ
 X m
Mτ λ (fn+k )(x)τ λ + − kfn+k (x + 2−k hj)kτ λ dh
j=1
j {|h|61}
m  Z 1/τ

τλ
X m
= Mτ λ (fn+k )(x) + − kfn+k (x + y)kτ λ dy
j=1
j |y|6j2−k

6 (2m + 1)1/τ Mτ λ (fn+k )(x)λ .

Combining the estimates for T1 and T2 , we conclude that

J m,τ (fn+k , 2−k ) .d,λm,r,τ 2dn(1−λ)/r Mr (fn+k )1−λ Mτ λ (fn+k )λ .

Since s > (1−λ)d

r (see (14.85)), by Hölder’s inequality (applied twice) we obtain
X
(2ks J m,τ (fn+k , k))k∈Z Lp (Rd ;`q (Z))
n>1
X (1−λ)d
2−n(s− )
2(n+k)s Mr (fn+k )1−λ Mτ λ (fn+k )λ


. r
k∈Z Lp (Rd ;`q )
n>0
384 14 Function spaces
1−λ λ
2js Mr (fj ) j>0 Lp (Rd ;`q ) 2js Mτ λ (fj ) j>0



.d,λ,r,s Lp (Rd ;`q )

2js fj j>0 Lp (Rd ;`q ) = kf kFp,q

.λ,p,q,r,τ s (Rd ;X) ,

where in the last estimate we used the boundedness of Mr and Mτ λ on and

Lp (Rd ; `q ) thanks to (14.85). 

14.6.f Interpolation

In order to prove interpolation results for the scale of Triebel–Lizorkin spaces

we need the following variation of Lemma 14.4.29.
Lemma 14.6.22. Let s ∈ R, p ∈ (1, ∞) and q ∈ (1, ∞]. For k > 0 set
ψk = ϕk−1 + ϕk + ϕk+1 . Define the operators
q
R : Lp (Rd ; `w s
s
(X)) → Fp,q (Rd ; X)
s
S : Fp,q (Rd ; X) → Lp (Rd ; `w
q
s
(X))

by
X
R(fk )k>0 = ψk ∗ fk , Sf = (ϕk ∗ f )k>0 .
k>0

Then S is an isometry, R is bounded, and RS = I.

Proof. All assertions follow in the sameP way as in Lemma 14.4.29, except for
the boundedness of R. To see that k>0 ψk ∗ fk converges in S 0 (Rd ; X) note
that Lp (Rd ; `w
q
s
(X)) ,→ Lp (Rd ; `w
p
t
(X)) = `wp
t
(Lp (Rd ; X)) for any t < s by
Hölder’s inequality, so the convergence follows from Lemma 14.4.29. To see
that R is bounded, note that since ψbk ≡ 1 on supp(ϕ bk ) we have
X
kR(fk )k>0 kFp,q
s (Rd ;X) 6 (kϕj ∗ ψj+` ∗ fj+` kX )j>0 Lp (Rd ;`q )
ws
|`|62

. sup (M (kfj+` kX ))j>0 q

Lp (Rd ;`w s)
|`|62

.d,p,q sup (kfj+` kX )j>0 q

Lp (Rd ;`w s)
|`|62

6 4|s| (kfj kX )j>0 q

Lp (Rd ;`w
,
s)

where we used Proposition 2.3.9 and the boundedness of the Hardy–Littlewood

maximal function M on Lp (Rd ; `wq
s
), which is an immediate consequence of
the Fefferman–Stein theorem (Theorem 3.2.28); here we use the assumptions
p ∈ (1, ∞) and q ∈ (1, ∞]. 
Using the operators R and S from Lemma 14.6.22 in the same way as in
Theorem 14.4.30, the following theorem identifies the complex interpolation
spaces of Triebel–Lizorkin.
 14.6 Triebel–Lizorkin spaces 385

Theorem 14.6.23 (Complex interpolation of Triebel–Lizorkin spac-

es). Let (X0 , X1 ) be an interpolation couple of Banach spaces and let p0 , p1 ∈
(1, ∞), q0 , q1 ∈ (1, ∞] with min{q0 , q1 } < ∞, s0 , s1 ∈ R and let θ ∈ (0, 1).
Define p1 = 1−θ θ 1 1−θ θ
p0 + p1 , q = q0 + q1 , and s = (1 − θ)s0 + θs1 . Then

[Fps00,q0 (Rd ; X0 ), Fps11,q1 (Rd ; X1 )]θ = Fp,q

s
(Rd ; Xθ ),

isomorphically, where Xθ = [X0 , X1 ]θ .

The following result on the real interpolation of Triebel–Lizorkin spaces can

be derived from the corresponding result for Besov spaces in the same way as
Theorem 14.4.31, but now using the sandwich result of Proposition 14.6.13.

Proposition 14.6.24 (Real interpolation of Triebel–Lizorkin spaces).

Let p ∈ [1, ∞), q0 , q1 , q ∈ [1, ∞], and s0 6= s1 ∈ R. For θ ∈ (0, 1), 1q = 1−θ θ
q0 + q1 ,
and s = (1 − θ)s0 + θs1 we have
s0
(Fp,q 0
(Rd ; X), Fp,q
s1
1
(Rd ; X))θ,q = Bp,q
s
(Rd ; X).

Our next aim is an interpolation result which will be used improve the Sobolev
embedding result of Theorems 14.4.19 and 14.6.14.

Proposition 14.6.25. Let p0 , p1 ∈ (1, ∞), q ∈ (1, ∞], and s ∈ R. For θ ∈

(0, 1) and p1 = 1−θ θ
p0 + p1 we have

(Fps0 ,q (Rd ; X), Fps1 ,q (Rd ; X))θ,p = Fp,q

s
(Rd ; X),
(Fps0 ,1 (Rd ; X), Fps1 ,1 (Rd ; X))θ,p ,→ Fp,1
s
(Rd ; X).

Proof. The first interpolation identity can be proved as in Theorem 14.4.31,

using Lemma 14.6.22 and the isomorphic identification

(Lp0 (Rd ; `qws (X)), Lp1 (Rd ; `qws (X)))θ,p = Lp (Rd ; `qws (X))

which follows from Theorem 2.2.10 and Proposition 14.3.5. The case q = 1
can be deduced from the proof of Theorem 14.4.31 as well. Indeed, since the
operator S of Lemma 14.6.22 is an isometry also for q = 1, we find

kf kFp,1
s (Rd ;X) = kSf kLp (Rd ;`1
w (X))
s

hp,p0 ,p1 ,θ kSf k(Lp0 (Rd ;`1w (X)),Lp1 (Rd ;`1ws (X)))θ,p
s

.p,p0 ,p1 ,θ kf k(Fps (Rd ;X),Fps (Rd ;X))θ,p .

0 ,1 1 ,1


As an application we can prove some further embedding results.
386 14 Function spaces

Theorem 14.6.26 (Jawerth–Franke). Let p0 , p1 , q ∈ [1, ∞] and s0 , s1 ∈ R

satisfy 1 6 p0 < p1 6 ∞ and s0 > s1 . If s0 − pd0 > s1 − pd1 , then we have
continuous embeddings

Bps00 ,p1 (Rd ; X) ,→ Fps11,q (Rd ; X) if p1 < ∞; (14.86)

Fps00,q (Rd ; X) ,→ Bps11 ,p0 (Rd ; X). (14.87)

Since the embedding Fps00,p0 (Rd ; X) ,→ Bps00 ,p1 (Rd ; X) holds trivially, (14.86)
improves the embedding in Theorem 14.6.14. In a similar way one sees that
(14.87) is an improvement of Theorem 14.6.14. Consequently, it follows from
Theorem 14.6.14 that, under the assumption p0 < p1 , the condition s0 − pd0 >
s1 − pd1 is also necessary for both (14.86) and (14.87).

Proof. By the trivial embeddings (14.23) and (14.70), it suffices to consider

s0 − pd0 = s1 − pd1 .
To prove (14.86), assume that p1 < ∞. In view of (14.70) it suffices to
consider q = 1. Fix p0 < r0 < p1 < r1 and θ ∈ (0, 1) such that p11 = 1−θ θ
r0 + r1 .
Let t0 , t1 ∈ R be such that
d d d d
t0 − = s1 − and t1 − = s1 − .
p0 r0 p0 r1

Then (1 − θ)t0 + θt1 = s0 and therefore, using Proposition 14.6.24, Theorem

14.6.14, and Proposition 14.6.25,

Bps00 ,p1 (Rd ; X) = (Fpt00,1 (Rd ; X), Fpt01,1 (Rd ; X))θ,p1

,→ (Frs01,1 (Rd ; X), Frs11,1 (Rd ; X))θ,p1 ,→ Fps11,1 (Rd ; X),

which implies the embedding (14.86).

To prove (14.87) it suffices to consider q = ∞. Moreover, by Theorems
14.4.19 and 14.6.14 we may assume that 1 < p0 < p1 < ∞. Fix 1 < r0 < p0 <
r1 < p1 and θ ∈ (0, 1) such that p10 = 1−θ θ
r0 + r1 . Let t0 , t1 ∈ R be such that

d d d d
t0 − = s0 − and t1 − = s0 − .
p1 r0 p1 r1

Then (1 − θ)t0 + θt1 = s1 . By Proposition 14.6.25, Theorem 14.6.14 and

Proposition 14.6.24,

Fps00,∞ (Rd ; X) = (Frs00,∞ (Rd ; X), Frs10,∞ (Rd ; X))θ,p0

,→ (Fpt10,∞ (Rd ; X), Fpt11,∞ (Rd ; X))θ,p0 = Bps11 ,p0 (Rd ; X).


As an interesting consequence of Theorem 14.6.26 we have the following im-
provement of Corollary 14.4.27 (2), extending it to the case p0 = 1. The result
 14.6 Triebel–Lizorkin spaces 387

d
is false for integrability exponents p0 > 1. Indeed, if s − p0 > 0 and it would
s− pd
hold that Fps0 ,q (Rd ) ,→ Cub (Rd ) for q = ∞, then it would also hold for
0

all q ∈ [1, ∞). However, by Proposition 14.6.17 this would imply that every
function in Fps0 ,q (Rd ) is zero at x1 = 0, which is of course not true.
s s−d
Corollary 14.6.27. If s > d is an integer, then F1,∞ (Rd ; X) ,→ Cub (Rd ; X)
continuously.

The result also holds in the case where s > d is not integer. However, in this
case Corollary 14.4.27 (2) gives a better result.
Proof. By Theorem 14.6.26 and Proposition 14.4.18,
s s−d s−d
F1,∞ (Rd ; X) ,→ B∞,1 (Rd ; X) ,→ Cub (Rd ; X).

14.6.g Duality

The next theorem identifies the duals of vector-valued Triebel–Lizorkin spaces.

Theorem 14.6.28. Let p, q ∈ (1, ∞) and s ∈ R. Then

s
Fp,q (Rd ; X)∗ ' Fp−s d ∗
0 ,q 0 (R ; X )

isomorphically.

The proof is similar to that of Theorem 14.4.34. The restriction p, q > 1 comes
in through Lemma 14.6.22.

s s
14.6.h Pointwise multiplication by 1R+ in Bp,q and Fp,q

In this section we apply the difference norm characterisation of Theorem

14.6.20, as well as the interpolation and duality results proved in this sec-
tion, to study pointwise multiplication in Triebel–Lizorkin spaces with the
non-smooth function 1R+ . The corresponding result for Besov spaces will be
derived afterwards by real interpolation.
As a preparation we first deduce several fractional Hardy inequalities.

Proposition 14.6.29 (Hardy–Young inequality). Let p ∈ [1, ∞] and α ∈

R \ {0}, and let f : R+ → X be strongly measurable and integrable on every
finite interval (0, t). Each of the conditions
Rt
(1) α > 0 and limt→0 1t 0 f (τ ) dτ = 0
Rt
(2) α < 0 and limt→∞ 1t 0 f (τ ) dτ = 0
388 14 Function spaces

implies

kt 7→ t−α f (t)kLp (R+ , dt ;X)

t
 Z t 
−1
6 (1 + |α| ) t 7→ t−α f (t) − − f (τ ) dτ dt
0 Lp (R+ , t ;X)

provided the right-hand side is finite.

Rt
Proof. (1): Let F (t) := f (t) − −0 f (τ ) dτ . Integrating by parts on [t, σ] we
obtain
Z σ
1 s 1 σ 1 t
Z Z Z Z σ
ds
I := 2
f (r) dr ds = − f (r) dr + f (r) dr + f (s) .
t s 0 σ 0 t 0 t s

Therefore,
Z σ Z σ Z σ Z t
ds ds
F (s) = f (s) − I = − f (r) dr − − f (r) dr
t s t s 0 0
Z t (14.88)
= f (σ) − F (σ) − − f (r) dr.
0

Letting t ↓ 0 in (14.88) and taking norms, we obtain the estimate

Z σ
ds
kf (σ)k 6 kF (σ)k + kF (s)k , t > 0.
0 s

e := α − 1 > −1 to
Applying Hardy’s inequality (see Lemma L.3.2(1)) with α
the function s 7→ kF (s)k we obtain

kσ 7→ σ −α f (σ)kLp (R+ , dσ ;X) 6 (1 + α−1 )kσ 7→ σ −α F (σ)kLp (R+ , dσ ;X)

σ σ

which gives the required estimate.

(2): We argue in the same way, but this time we rewrite the right-hand
side of (14.88) as
Z σ Z σ
ds
F (s) = − f (r) dr − f (t) + F (t).
t s 0

Letting σ → ∞ and taking norms, we obtain the estimate

Z ∞
ds
kf (t)k 6 kF (t)k + kF (s)k , t > 0.
t s

Now the proof is finished as before, this time applying Lemma L.3.2(2) with
e := α − 1 < −1.
α 
As an immediate consequence we obtain the following result.
 14.6 Triebel–Lizorkin spaces 389

Proposition 14.6.30 (Fractional Hardy inequality). Let p ∈ [1, ∞) and

β ∈ R, and let f : R+ → X is strongly measurable and integrable on every
finite sub-interval (0, t). Each of the conditions
Rt
(1) β ∈ (1/p, ∞) and limt↓0 −0 kf (τ )k dτ = 0
Rt
(2) β ∈ (−∞, 1/p) and limt→∞ −0 kf (τ )k dτ = 0
implies
Z 
kf kLp (R+ ,t−βp dt;X) 6 C x 7→ x−β − kf (x) − f (x − h)k dh
(0,x) Lp (R+ )
Z
6 C x 7→ sup t−β − kf (x) − f (x − h)k dh
t>0 (0,x∧t) Lp (R+ )

1
with C := 1 + 1 ,
|β− p |
provided the right-hand side is finite.

Proof. By Proposition 14.6.29 with α = β − p1 ,

Z
kf kLp (R+ ,t−βp dt;X) 6 C x 7→ x−β f (x) − − f (τ ) dτ p
(0,x) L (R+ ;X)
Z
6 C x−β − kf (x) − f (x − h)k dhkLp (R+ )
(0,x)
Z
−β
6 C sup t − kf (x) − f (x − h)k dh .
t>0 (0,x∧t) Lp (R+ )

This gives the required estimate in both cases. 

For p ∈ [1, ∞), q ∈ [1, ∞], and s ∈ (1/p, 1) we define the following closed
subspaces of H s,p (R; X) and Fp,q
s
(R; X), respectively:
s,p
0H (R; X) := {f ∈ H s,p (R; X) : f (0) = 0},
s s
0 Fp,q (R; X) := {f ∈ Fp,q (R; X) : f (0) = 0}.

Here we use the bounded continuous version for f (which exists by Corollary
14.4.27 combined with Propositions 14.6.8 and 14.6.13) respectively. The con-
tinuity of the embeddings in Corollary 14.4.27 gives the closedness of these
subspaces.
We can now prove the following fractional Hardy inequality in terms of
s
the spaces Fp,q and H s,p and their analogues 0 Fp,q
s
and 0 H s,p .

Corollary 14.6.31. Let p ∈ [1, ∞) and q ∈ [1, ∞].

s
(1) If s ∈ (1/p, 1), then each of the spaces 0 Fp,q (R; X) and 0 H s,p (R; X) con-
tinuously embeds into Lp (R, |t|−sp dt; X).
s
(2) If s ∈ (0, 1/p), then each of the spaces Fp,q (R; X) and H s,p (R; X) (if
p −sp
p 6= 1) continuously embeds into L (R, |t| dt; X).
390 14 Function spaces

s
Since W s,p (R; X) = Fp,p (R; X) for s ∈ (0, 1), the corollary also covers frac-
tional Sobolev spaces.
Proof. By the embeddings (14.69) and (14.73) it suffices to prove the result
s s
for 0 Fp,∞ (R; X) and Fp,∞ (R; X).
By Proposition Rt 14.6.30, using that for bounded continuousR functions f :
t
R → X we have −0 f (τ ) dτ → f (0) = 0 as t ↓ 0 in case (1) and −0 f (τ ) dτ → 0
as t → ∞ in case (2), we have
Z
k1R+ f kLp (R,|t|−sp dt;X) 6 C x 7→ x−s − kf (x) − f (x − h)k dh p
(0,x) L (R+ )
Z
6 2C x 7→ sup t−s − k∆h f (x)k dh p
t>0 (−t,t) L (R)
(1)
= 2C[f ]F s (R;X) .p,s kf kFp,∞
s (R;X)
p,∞

where in the last step we used Theorem 14.6.20 with m = 1. A similar estimate
holds for f on the negative real axis. 
As a consequence we obtain the following result on pointwise multiplication.
Theorem 14.6.32 (Pointwise multiplication by 1R+ ). Let p ∈ [1, ∞),
q ∈ [1, ∞], and s ∈ (0, 1). Each of the two conditions
s
(1) s ∈ (0, 1/p) and f ∈ Fp,q (R; X)
s
(2) s ∈ (1/p, 1) and f ∈ 0 Fp,q (R; X)
s
implies that 1R+ f ∈ Fp,q (R; X) and

s (R;X) 6 Ckf kF s (R;X) .

k1R+ f kFp,q p,q

Without the condition f (0) = 0, the result is false for s > 1/p. Indeed, this
is clear from the fact that, by combining Corollary 14.4.27 and Proposition
s
14.6.13, we have a continuous embedding Fp,q (R; X) ,→ Cub (R; X). A coun-
terexample to the case s = 1/p will be discussed in Example 14.6.33. It shows
that Propositions 14.6.29, 14.6.30, and Corollary 14.6.31 do not hold for α = 0
and s = 1/p.
Proof. Clearly, k1R+ f kLp (Rd ;X) 6 kf kLp (Rd ;X) . Therefore, using the difference
(1)
norm of Theorem 14.6.20 it remains to estimate [1R+ f ]F s in terms of
p,q (R;X)
(1)
kf kFp,q Fp,q (R;X) . We give the proof for q ∈ [1, ∞); the case q = ∞
s (R;X) and [f ] s

requires the usual obvious modifications.

By the triangle inequality,
(1)
[1R+ f ]F s (R;X)
p,q
Z Z 1 Z q dt p/q 1/p
6 t−sq kf (x + h) − f (x)k dh dx
R+ R+ t (−t,t)∩(−x,∞) t
 14.6 Triebel–Lizorkin spaces 391
Z Z  Z q dt p/q 1/p
−sq 1
+ t kf (x)k dh dx
R+ R+ t (−t,t)∩(−∞,−x) t
Z Z 1 Z q dt p/q 1/p
+ t−sq kf (x + h)k dh dx
(−∞,0) R+ t (−t,t)∩(−x,∞) t
=: (I) + (II) + (III).
(1)
We estimate these three terms separately. Clearly, (I) 6 [f ]F s (R;X) and, with
p,q
p
C = 1 + |sp−1| ,
Z Z ∞
dt p/q 1/p
(II) 6 t−sq
kf (x)kp dx
R+ x t
Z ∞ 1/p
6 (sq)−1/q x−sp kf (x)kp dx
0
.s,p,q kf kFp,q
s (R;X) ,

using Corollary 14.6.31 in the last step.

To estimate (III) fix x ∈ (−∞, 0). By Minkowski’s inequality (Theorem
1.2.22),
Z 1 Z q dt p/q
t−sq kf (x + h)k dh
R+ t (−t,t)∩(−x,∞) t
Z Z dt  1/q
6 t−sq−q 1(h,∞) (t) 1(−x,∞) (h)kf (x + h)k dh
R+ R+ t
Z
= Kq,s h−s−1 1(−x,∞) (h)kf (h + x)k dh
R+
Z
= Kq,s (y − x)−s−1 kf (y)k dy,
R+

where Kq,s = (sq + q)1/q . Setting z = −x and φp (z) = z 1/p (1 + z)−s−1 , (III)
can be estimated using Young’s inequality for convolutions for the multiplica-
tive group R+ with Haar measure dz z :

Z Z p 1/p
(III) 6 Kq,s (y + z)−s−1 kf (y)k dy dz
R+ R+
Z Z 1 dy p dz 1/p
= Kq,s φp (z/y)y −s+ p kf (y)k
R+ R+ y z
Z 1/p
6 Kq,s kφp kL1 (R+ , dz ) y −sp kf (y)kp dy
z
R+

.p,q,s kf kFp,q
s (R;X) ,

using Corollary 14.6.31 as in the estimate for (II). 

392 14 Function spaces

Example 14.6.33. Theorem 14.6.32 is false for s = 1/p even in the scalar-
valued case. Indeed, f ∈ Cc∞ (R) is any function satisfying f ≡ 1 on [−1, 1],
1/p
then for all p ∈ [1, ∞) we have f ∈ Fp,q (R). Let us prove that 1R+ f ∈ /
1/p
Fp,q (R). To this end it suffices to take q = ∞. In case p ∈ (1, ∞) we can use
Theorem 14.6.20 to find
(1)
k1R+ f kF 1/p (Rd ;X) hp k1R+ f k 1/p
p,∞ Fp,∞ (Rd ;X)
Z −x
1
> x 7→ sup t− p −1 |f (x)| dh
t>0 −t Lp (0,1)
1
= x 7→ sup t− p −1 (t − x)
t>x Lp (0,1)
1
−p
&p kx 7→ x kLp (0,1) = ∞.

1 1/r
For p = 1 we note that F1,q (R) ,→ Fr,∞ (R) for all r ∈ (p, ∞) by Theorem
1
14.6.14, and therefore 1R+ f ∈ / F1,q (R).
One could still hope that the boundedness of f 7→ 1R+ f for s = 1/p holds
1/p
on the closure in Fp,q (R) of the smooth functions satisfying f (0) = 0. This
turns out to be false as well. Indeed, in the case q < ∞ the latter space
1/p
coincides with Fp,q (R) by Proposition 14.6.17. If q = ∞, the boundedness is
1/p
also fails, as follows from the previous example and the embedding Fp,∞ (R) ,→
1/r
Fr,r (R) for all r ∈ (p, ∞) contained in Theorem 14.6.14.

By duality and interpolation, we now extend Theorem 14.6.32 to smoothness

exponents s 6 0, which excludes the end-point cases.

Corollary 14.6.34 (Pointwise multiplication by 1R+ ). Let p ∈ (1, ∞),

q ∈ (1, ∞), and s ∈ (−1/p0 , 0]. For all f ∈ Fp,qs
(R; X) we have 1R+ f ∈
s
Fp,q (R; X) and
k1R+ f kFp,q
s (R;X) 6 Ckf kF s (R;X) .
p,q

Proof. By density it suffices to consider f ∈ C ∞ (R \ {0}) ⊗ X. We use duality

result. By Theorems 14.6.28 and 14.6.32 for any g ∈ S (Rd ; X ∗ ) we have

|h1R+ f, gi| = |hf, 1R+ gi| 6 Ckf kFp,q

s (R;X) k1R gk −s
+ F 0 0 (R;X ∗ ) p ,q

6 C 0 kf kFp,q
s (R;X) kgk −s
F 0 0 (R;X ∗ ) .
p ,q

Since S (Rd ; X ∗ ) is dense in Fp−s ∗

0 ,q 0 (R; X ), the result follows by another ap-

plication of Theorem 14.6.28.

The case s = 0 follows by complex interpolation between the cases s and
−s for s > 0 small enough, using Theorems C.2.6 and 14.6.23. 
Applying the real interpolation method instead, we obtain the following for
the Besov scale.
 14.7 Bessel potential spaces 393

Corollary 14.6.35 (Pointwise multiplication by 1R+ ). Let p ∈ (1, ∞),

s
q ∈ [1, ∞], and s ∈ (−1/p, 1/p). For all f ∈ Bp,q (R; X) we have 1R+ f ∈
s
Bp,q (R; X) and
s
s (R;X) 6 Ckf kB s (R;X) ,
k1R+ f kBp,q p,q
f ∈ Bp,q (R; X).

s−ε s+ε s
Proof. First let s > 0. Since (Fp,2 , Fp,2 )1/2,q = Bp,q by Theorem 14.4.31,
the result follows from Theorems 14.6.32 and C.3.3. Here we can allow p = 1
as well.
The result for s < 0 and q ∈ (1, ∞) follows from Theorem 14.4.34 in the
same way as in Corollary 14.6.34. The cases q = 1 and q = ∞ can be obtained
by another real interpolation argument as we did in Example 14.4.35.
The case s = 0 follows by real interpolation between the cases s and −s
for s > 0 small. 

14.7 Bessel potential spaces

In this section we prove Sobolev embeddings and norm estimates for Bessel
potential spaces. Some results will depend on the geometry of X. Real in-
terpolation for H s,p (Rd ; X) has already been considered in Theorem 14.4.31.
Duality for H s,p (Rd ; X) has already been considered in Proposition 5.6.7.

14.7.a General embedding theorems

We begin with the following Sobolev embedding theorem.

Theorem 14.7.1 (Sobolev embedding for Bessel potential spaces and

Sobolev spaces). Let p0 , p1 ∈ (1, ∞) and s0 , s1 ∈ R. We have a continuous
embedding
H s0 ,p0 (Rd ; X) ,→ H s1 ,p1 (Rd ; X)
if and only if one of the following two conditions holds:

p 0 = p1 and s0 > s1 ; (14.89)

d d
p0 < p 1 and s0 − > s1 − . (14.90)
p0 p1
If s0 , s1 ∈ N, then the same necessary and sufficient conditions give the exis-
tence of a continuous embedding

W s0 ,p0 (Rd ; X) ,→ W s1 ,p1 (Rd ; X).

394 14 Function spaces

Proof. We first prove the result for Bessel potential spaces.

‘If’: By Proposition 14.6.13, for p ∈ (1, ∞) and s ∈ R we have continuous
embeddings
s
Fp,1 (Rd ; X) ,→ H s,p (Rd ; X) ,→ Fp,∞
s
(Rd ; X). (14.91)

From Theorem 14.6.14 we see that if either (14.89) or (14.90) holds, then
Fps00,∞ (Rd ; X) ,→ Fps11,1 (Rd ; X). Therefore the required embedding follows from
(14.91) with s = s0 , s1 and p = p0 , p1 .
‘Only if’: If the stated embedding holds, then by (14.91) with s = s0 , s1
and p = p0 , p1 , we also have a continuous embedding Fps00,1 (Rd ; X) ,→
Fps11,∞ (Rd ; X). Therefore, either (14.89) or (14.90) must hold by Theorem
14.6.14.
The corresponding result for Sobolev spaces with integer smoothness can
be proved in the same way, noting that the analogue of (14.91) holds for these
spaces. 

Remark 14.7.2. The embedding of Theorem 14.7.1 for Bessel potential spaces
can be restated as the boundedness of J−(s0 −s1 ) = (1 − ∆)−(s0 −s1 ) from
Lp0 (Rd ; X) into Lp1 (Rd ; X). Since J−(s0 −s1 ) is a positive operator by Propo-
sition 5.6.6, we infer from Theorem 2.1.3 that the boundedness in the scalar
case is actually equivalent to boundedness in the vector-valued situation.

By the same argument as in Theorem 14.7.1, the following result can be

deduced from Proposition 14.6.15.

Proposition 14.7.3 (Gagliardo–Nirenberg inequality for Bessel po-

tential spaces). Let p0 , p1 ∈ (1, ∞), −∞ < s0 < s1 < ∞, and θ ∈ (0, 1),
and let
1 1−θ θ
= + , s = (1 − θ)s0 + θs1 .
p p0 p1
There exists a constant C = Cθ,p0 ,p1 ,s0 ,s1 > 0 such that for all f ∈
H s0 ,p0 (Rd ; X) ∩ H s1 ,p1 (Rd ; X) we have f ∈ H s,p (Rd ; X) and

kf kH s,p (Rd ;X) 6 Ckf k1−θ

H s0 ,p0 (Rd ;X)
kf kθH s1 ,p1 (Rd ;X) .

If, in Proposition 14.7.3, s0 , s1 > 0 are integers and p ∈ (1, ∞), the same argu-
ment gives that f ∈ W s0 ,p0 (Rd ; X) ∩ W s1 ,p1 (Rd ; X) implies f ∈ W s,p (Rd ; X)
and

kf kW s,p (Rd ;X) 6 Ckf k1−θ

W s0 ,p0 (Rd ;X)
kf kθW s1 ,p1 (Rd ;X) . (14.92)

The latter estimate extends to p0 ∈ (1, ∞] and p1 ∈ (1, ∞]. Indeed, if only
one of the exponents is infinite, then (14.92) is a consequence of Proposition
14.6.16 and the sandwich results of Propositions 14.4.18 (see (14.29)) and
 14.7 Bessel potential spaces 395

14.6.13. If p = p0 = p1 ∈ [1, ∞], (14.92) can be deduced from these sandwich

results and real interpolation and (L.2):

(W s0 ,p (Rd ; X), W s1 ,p (Rd ; X))θ,1 ,→ (Bp,∞

s0
(Rd ; X), Bp,∞
s1
(Rd ; X))θ,1
s
= Bp,1 (Rd ; X) (by (14.48))
,→ W s,p (Rd ; X).

Note that this even gives (14.92) for p = p0 = p1 = 1.

The estimate (14.92) self-improves to the following Gagliardo–Nirenberg
type inequality for W s,p (Rd ; X):

Theorem 14.7.4 (Schmeisser–Sickel). Let p0 , p1 , p ∈ (1, ∞], m ∈ N, and

|α| 6 m satisfy
|α| 1 1−θ θ
θ= and = + .
m p p0 p1
There exists a constant C > 0 such that for all f ∈ Lp0 (Rd ; X)∩W m,p1 (Rd ; X)
we have
 X θ
k∂ α f kLp (Rd ;X) 6 Ckf k1−θ
p d
L 0 (R ;X)
k∂ β
f kL p1 (Rd ;X) .
|β|=m

Moreover, the same holds if p = p0 = p1 = 1.

Proof. For θ = |α|

m ∈ {0, 1} there is nothing to prove, so we may assume that
θ ∈ (0, 1). Taking s = |α|, s0 = 0, and s1 = m in (14.92), it follows that
 X θ
k∂ α f kLp (Rd ;X) 6 Ckf k1−θ
Lp0 (Rd ;X)
k∂ β f kLp1 (Rd ;X) .
|β|6m

Applying this to the function f (λ·) for λ > 0, we obtain

d
λ|α|− p k∂ α f kLp (Rd ;X)
d
 X d
θ
6 C(λ− p0 kf kLp0 (Rd ;X) )1−θ λ|β|− p1 k∂ β f kLp1 (Rd ;X) .
|β|6m

d
Now divide both sides by λ|α|− p and pass to the limit λ → ∞. 

14.7.b Embedding theorems under geometric conditions

Littlewood–Paley inequality for Bessel potential spaces

The aim of this paragraph is to prove the following Littlewood–Paley inequal-

ity with smooth cut-offs for H s,p (Rd ; X).
396 14 Function spaces

Theorem 14.7.5 (Littlewood–Paley theorem for Bessel potential

spaces). Let X be a UMD space, p ∈ (1, ∞), and s ∈ R. A tempered distri-
bution f ∈ S 0 (Rd ; X) belongs to H s,p (Rd ; X) if and only if
n
X
|||f |||H s,p (Rd ;X) := sup εk 2ks ϕk ∗ f < ∞.
n>0 Lp (Ω×Rd ;X)
k=0

In this situation the sum k>0 εk 2ks ϕk ∗ f converges, both in Lp (Ω × Rd ; X)

P

and almost surely in Lp (Rd ; X), and we have an equivalence of norms

|||f |||H s,p (R;X) hd,p,s,X kf kH s,p (Rd ;X) .

For s = 0 the above estimate yields an equivalent norm on Lp (Rd ; X) which

is slightly different from the Littlewood–Paley estimate with smooth cut-offs
of Theorem 5.5.22, where the summation was taken over Z and the functions
ψk were of the form 2k ψ(2k ·) for a Littlewood–Paley function ψ in the sense
of Definition 5.5.20.
Proof. ‘Only if’: Fix f ∈ H s,p (Rd ; X). Fix a sequence of signs  = (k )k>0 in
{z ∈ K : |z| = 1}. For integers n > 0, define the function mn ∈ C ∞ (Rd ) by
n
X
mn (ξ) := k 2ks (1 + |ξ|2 )−s/2 ϕ
bk (ξ).
k=0

From the location of the supports of the functions ϕbk one sees three things:
first, that for each ξ ∈ Rd at most three terms in this sum are non-zero
(the sum therefore converges for trivial reasons); second, that k∂ β ϕ
bk k∞ 6
Cβ 2−k|β| ; and third, that

cd = sup sup sup sup |ξ||α| |∂ α mn (ξ)|

 n>0 α∈{0,1}d ξ6=0

is finite, the outer supremum being taken over all sequences of signs  =
(k )k>0 .
By the Mihlin multiplier theorem (Theorem 5.5.10), the Fourier multiplier
operators Tmn associated with mn are bounded on Lp (Rd ; X), with estimates
uniform in n and signs , say sup supn>0 kTmn kL (Lp (Rd ;X)) 6 CX,p,d . Since
n
X
k 2ks ϕk ∗ f = Tmn Js f, (14.93)
k=0

we obtain
n
X
k 2ks ϕk ∗ f 6 CX,p,d kJs f kLp (Rd ;X) = CX,p,d kf kH s,p (Rd ;X) .
Lp (Rd ;X)
k=0
 14.7 Bessel potential spaces 397

Taking k = εk (ω) and passing to the Lp (Ω)-norms, we obtain the estimate

n
X
εk 2ks ϕk ∗ f 6 CX,p,d kf kH s,p (Rd ;X) .
Lp (Ω×Rd ;X)
k=0

‘If’: Assume now that f ∈ S 0 (Rd ; X) satisfies |||f |||H s,p (Rd ;X) < ∞. We
claim that k>0 εk 2ks ϕk ∗ f converges in Lp (Ω; Lp (Rd ; X)) and almost surely
P

in Lp (Rd ; X). Indeed, Lp (Rd ; X) is a UMD space by Proposition 4.2.15, so by

Proposition 4.2.19 it does not contain an isomorphic copy of c0 . The con-
vergence of the sum, in Lp (Ω × Rd ; X) and almost surely in Lp (Rd ; X), now
follows from Corollary 6.4.12. Moreover, by Fatou’s lemma and the Kahane
contraction principle,
X
|||f |||H s,p (Rd ;X) = εk 2ks ϕk ∗ f .
Lp (Ω×Rd ;X)
k>0

For k ∈ {0, 1} choose ψk ∈ Cc∞ (R) such that 0 6 ψbk 6 1, supp ψb0 ⊆ {0 6
|ξ| 6 2} and supp ψb1 ⊆ { 14 6 |ξ| 6 4}, and ψbk ≡ 1 on supp ϕ bk . For k > 2 we
−(k−1)
define ψk := ψ1 (2
b b ·). For ω ∈ Ω put
X X
mω := εj (ω)2−js (1 + | · |2 )s/2 ψbj , gω := εk (ω)2ks ϕk ∗ f.
j>0 k>0

As before,

Cm = sup sup sup |ξ||α| |∂ α mω (ξ)| < ∞.

ω∈Ω α∈{0,1}d ξ6=0

Therefore, by the Mihlin multiplier Theorem 5.5.10,

kTmω gω kLp (Rd ;X) 6 Ckgω kLp (Rd ;X)

for almost every ω ∈ Ω. Considering finite sums first, one checks that ω 7→
Tmω gω is strongly measurable. Since ω 7→ gω belongs to Lp (Ω; Lp (Rd ; X)), it
follows that so does ω 7→ Tmω gω . By the condition ψbk ≡ 1 on supp ϕ
bk , as in
(14.93) we have Z
Tmω gω dP(ω) = Js f.
Ω

By Jensen’s inequality and Fubini’s theorem, f ∈ H s,p (Rd ; X) and

kf kpH s,p (Rd ;X) = kJs f kpLp (Rd ;X)

Z p
= Tmω gω dP(ω)
Lp (Rd ;X)
Z Ω
6 kTmω gω kpLp (Rd ;X) dP(ω)
Ω
398 14 Function spaces
Z
p p
6C kgω kL p (Rd ;X) dP(ω) = C|||f |||H s,p (Rd ;X) .
Ω


We continue with an embedding result under additional geometric assump-
tions on X. The cases p0 = 1 and q0 = ∞ were proved for general Banach
spaces in Propositions 14.4.18 and 14.6.13.

Proposition 14.7.6 (Sandwich theorem under type and cotype). Let

X be a UMD Banach space with type p0 ∈ [1, 2] and cotype q0 ∈ [2, ∞]. For
all p ∈ (1, ∞) and s ∈ R we have continuous embeddings
s
Fp,p 0
(Rd ; X) ,→ H s,p (Rd ; X) ,→ Fp,q
s
0
(Rd ; X).
s
Proof. We only prove Fp,p 0
(Rd ; X) ,→ H s,p (Rd ; X); the other embedding is
proved similarly.
s
Let f ∈ Fp,p 0
(Rd ; X). By Theorem 14.7.5, the Kahane–Khintchine inequal-
ity (Theorem 6.2.4) and the type p0 property of X, we have
n
X
kf kH s,p (Rd ;X) 6 C sup εk 2ks ϕk ∗ f
n>1 Lp (Ω×Rd ;X)
k=0
Z n
X p 1/p
hp C sup εk 2ks ϕk ∗ f dx
n>1 Rd Lp0 (Ω;X)
k=0
Z n
X p/p0 1/p
6 C sup k2ks ϕk ∗ f kp0 dx
n>1 Rd k=0
= Ckf kFp,p
s (Rd ;X) .
0


In combination with Proposition 14.6.13 and Corollary 14.6.18 we obtain:

Corollary 14.7.7 (γ-Sobolev embedding – III). Let p0 ∈ [1, 2] and q0 ∈

[2, ∞].
(1) If X has type p0 , then for all p ∈ [1, p0 ) we have a continuous embedding
1 1
H ( p − 2 )d,p (Rd ; X) ,→ γ(L2 (Rd ), X).

(2) If X has cotype q0 , then for all q ∈ (q0 , ∞) we have a continuous embedding
1 1
γ(L2 (Rd ), X) ,→ H ( q − 2 )d,q (Rd ; X)

By Theorem 9.2.10, for p0 = 2 assertion (1) also holds for p = 2, and for
q0 = 2 assertion (2) also holds for q = 2.
 14.7 Bessel potential spaces 399

Necessity of the type and cotype assumptions

Proposition 14.7.8. Let p ∈ (1, ∞), q ∈ [1, ∞], s ∈ R, and m ∈ N. Then the
following assertions hold with A ∈ {B, F }:
(1) If Asp,q (Rd ; X) ,→ H s,p (Rd ; X) continuously, then X has type q.
(2) If H s,p (Rd ; X) ,→ Asp,q (Rd ; X) continuously, then X has cotype q.
(3) If Akp,q (Rd ; X) ,→ W m,p (Rd ; X) continuously, then X has type q.
(4) If m
W m,p (Rd ; X) ,→ Ap,q (Rd ; X) continuously, then X has cotype q.
Proof. (1): By the lifting properties of Propositions 14.4.15, 14.6.10, and
5.6.3, it suffices to consider s = 0. Fix a finitely non-zero sequence (xn )n>1
in X. Let ψ ∈ S (Rd ) be a non-zero function satisfying supp(ψ) b ⊆ [− 1 , − 1 ]d
4 8
and put X n
f (t, ω) := ψ(t) εn (ω)e2πi2 t1 xn ,
n>1
n
where as always (εn )n>1 is a Rademacher sequence. Since (εn e2πi2 t1
)n>1 is
a Rademacher sequence for each t ∈ Rd , we have
Z p
p
X n
Ekf kLp (Rd ;X) = |ψ(t)|p E εn e2πi2 t1 xn dt
Rd n>1
X p (14.94)
= kψkpLp (Rd ) E εn x n .
n>1

On the other hand, the Fourier support properties of ψ(· b − 2n t1 e1 ) and ϕ bn

(see (14.8) and (14.9)) imply that kf (·, ω) ∗ ϕn kX = |ψ(t)|kxn k and kf (·, ω) ∗
ϕ0 kX = 0. Therefore,

kf (·, ω)kAsp,q (Rd ;X) = kψkLp (Rd ) k(xn )n>1 k`q (X) . (14.95)

Applying the assumption (1) pointwise in Ω, we obtain

X p
kψkpLp (R) E εn xn = Ekf kpLp (Rd ;X)
n>1

6 C p Ekf kpAs d ;X) = C p kψkpLp (Rd ) k(xn )n>1 kp`q (X) .

p,q (R

By the Kahane–Khintchine inequalities, this shows that X has type q.

(2): This follows from the previous proof upon replacing “6” by “>”.
(3): The idea of the proof is the same as in (1), but this case is slightly
more technical. Let (xn )n>1 and ψ be as before and put
X n
f (t, ω) := ψ(t) 2−mn εn (ω)e2πi2 t1 xn =: ψ(t)fm (t, ω).
n>1

By Leibniz’s rule we obtain

400 14 Function spaces
X
∂ α f (t, ω) = cβ,γ ∂ β ψ(t)fm−j (t, ω),
|β|+j=|α|

For j ∈ {0, . . . , m − 1},

k(∂ β ψ)fm−j (·, ω)kLp (Rd ;X) 6 k∂ β ψk∞ kfm−j (·, ω)kLp (Rd ;X)
X
6 k∂ β ψk∞ 2−(m−j)n kxn k 6 k∂ β ψk∞ sup kxn k.
n>1
n>1

For j = m, as in (14.94) we have

X p
Ekf0 kpLp (Rd ;X) = kψkpLp (R) E εn x n .
n>1

By the reverse triangle inequality, this shows that there exists a constant
C = C(d, m, p, ψ) such that
X
kf kLp (Ω;W m,p (Rd ;X)) − kψkLp (R) εn xn Lp (Ω;X) 6 C sup kxn k. (14.96)
n>1
n>1

Stated differently, up a relatively

P small term the norm kf kLp (Ω;W m,p (Rd ;X)) is
equivalent to the norm k n>1 εn xn k of the random sum. As in (14.95) we
see that
kf (·, ω)kAp,q
m (Rd ;X) = kψkLp (Rd ) k(xn )n>1 k`q (X) .

Now from (14.96) and the assumptions, we obtain

X
kψkLp (R) εn x n 6 kf kLp (Ω;W m,p (Rd ;X)) + C sup kxn k
Lp (Ω;X) n>1
n>1

. kf kLp (Ω;Am d
p,q (R ;X))
+ sup kxn k
n>1

. k(xn )n>1 k`q (X) .

(4): This can be proved in the same way as (3). By (14.96) and
P the Kahane
contraction principle, which implies bound supn>1 kxn kp 6 Ek n>1 εn xn kp ,
from the assumption (4) we obtain

kψkLp (Rd ) k(xn )n>1 k`q (X) = kf kLp (Ω;Ap,q

m (Rd ;X))

X
. kf kLp (Ω;W m,p (Rd ;X)) . εn x n .
Lp (Ω;X)
n>1

A Hilbert space characterisation

s
The equality Fp,2 (Rd ; X) = H s,p (Rd ; X) with equivalent norms characterises
Hilbert spaces:
 14.7 Bessel potential spaces 401

Theorem 14.7.9 (Han–Meyer). Let p ∈ (1, ∞), s ∈ R, and m ∈ N. The

following assertions are equivalent:
m
(1) Fp,2 (Rd ; X) = W m,p (Rd ; X) with equivalent norms;
s
(2) Fp,2 (Rd ; X) = H s,p (Rd ; X) with equivalent norms;
(3) X is isomorphic to a Hilbert space.

Proof. (1)⇒(3) and (2)⇒(3): By Proposition 14.7.8, X has type 2 and cotype
2. Therefore X is isomorphic to a Hilbert space by Theorem 7.3.1.
(3)⇒(2): This is immediate from Proposition 14.7.6 and the fact that
Hilbert spaces are UMD (by Theorem 4.2.14) and have type 2 and cotype 2
(by the result of Example 7.1.2).
(3)⇒(1): This is a special case of the previous implication since Theorem
5.6.11 implies W m,p (Rd ; X) = H m,p (Rd ; X) with equivalent norms. 

14.7.c Interpolation

Real interpolation of vector-valued Bessel potential spaces has already been

considered in Theorem 14.4.31. Complex interpolation was considered in The-
orem 5.6.9, but only in the case p0 = p1 and X0 = X1 . In order to treat a
more general case we need a variant of the complex interpolation results for
`pws (X) of Proposition 14.3.3.
Let (εk )k>0 be a Rademacher sequence on a probability space Ω. Let
p ∈ (1, ∞) and s ∈ R, and let εs,p (X) denote the space of all sequences
(xk )k>0 in X for which
n
X
(xk )k>0 εs,p (X)
:= sup εk 2ks xk < ∞.
n>1 Lp (Ω;X)
k=0

The spaces εp (X) := ε0,p (X) have been introduced in Section 6.3. Clearly
the mapping (xk )k>0 7→ (2ks xk )k>0 defines an isometric isomorphism from
εs,p (X) onto εp (X). For fixed s ∈ R the spaces εs,p (X), 1 < p < ∞, coincide,
with pairwise equivalent norms; this follows from the Kahane–Khintchine in-
equalities as in Proposition 6.3.1. If X does not contain a copy isomorphic
to
P c0 , then Corollary 6.4.12 implies that for any (xk )k>0 in εs,p (X) the sum
ks p
k>0 εk 2 xk converges in L (Ω; X) and almost surely in X, and in this case

X
(xk )k>0 εs,p (X)
= εk 2ks xk .
Lp (Ω;X)
k>0

In particular, the partial sum projections Pn : (xk )k>0 7→ (xk )nk=0 are uni-
formly bounded and strongly convergent to the identity as operators on
εs,p (X).
The next result extends Theorem 7.4.16, which corresponds to the special
case s = 0.
402 14 Function spaces

Lemma 14.7.10. For j ∈ {0, 1} let Xj be a K-convex space and let pj ∈

(1, ∞). For θ ∈ (0, 1) set Xθ := [X0 , X1 ]θ . Then

[εs0 ,p0 (X0 ), εs1 ,p1 (X1 )]θ = εs,p (Xθ ),

1 1−θ θ
where p = p0 + p1 and s = (1 − θ)s0 + θs1 .

Proof. By Proposition 7.4.15, Xθ is K-convex. By Proposition 7.4.5 and

Lemma 7.4.11, Xθ , does not contain an isomorphic copy of c0 , and hence
the partial sum projections Pn on εs,p (Xθ ) are strongly convergent to the
identity.
To prove the required identity one can repeat the argument in Theorem
14.3.1 to reduce the result to the unweighted setting considered in Theorem
7.4.16. 
As a final preparation for the complex interpolation of Bessel potential spaces,
we prove a version of Lemma 14.4.29 for Bessel potential spaces.

Lemma 14.7.11. Let X be a UMD space and let p ∈ [1, ∞], q ∈ [1, ∞], and
s ∈ R. For k > 0 set ψk = ϕk−1 + ϕk + ϕk+1 . The operators

R : εs,p (Lp (Rd ; X)) → H s,p (Rd ; X)

S : H s,p (Rd ; X) → εs,p (Lp (Rd ; X))

defined by X
R(fk )k>0 = ψk ∗ fk , Sf = (ϕk ∗ f )k>0 ,
k>0

are bounded and satisfy RS = I.

Proof. The identity RS = I is proved as in Lemma 14.4.29. The boundedness

of S follows from Theorem 14.7.5. It remains to prove that R is bounded. Let
E := Lp (Ω; Lp (Rd ; X)). By Theorem 14.7.5 and a density argument it suffices
to show that, for all finitely non-zero sequences (f` )`>0 in Lp (Rd ; X),
n
X X X
εk 2ks ϕk ∗ ψj ∗ fj 6C εk 2ks fk , n > 0.
E E
k=0 j>0 k>0

From Theorem 14.7.5 (with s = 0) and Proposition 8.4.6(i) we see that

the sequence {ϕk ∗ : k > 0} is R-bounded in L (Lp (Rd ; X)), with R-bound
at most by Cp,X . Hence also the sequence {ψk ∗ : k > 0} is R-bounded in
this space, with R-bound at most 3Cp,X . Therefore, by the Fourier support
properties (14.8) and (14.9) of ϕk ,
n
X X X n
X
εk 2ks ϕk ∗ ψj ∗ f j 6 εk 2ks ϕk ∗ ψk+` ∗ fk+`
E E
k=0 j>0 |`|62 k=0
 14.7 Bessel potential spaces 403
n
X X
2
6 3Cp,X εk 2ks fk+`
E
|`62 k=0
X
2
6 3Cp,X 4|s| εk 2ks fk ,
E
k>0

where in the last step we used Kahane’s contraction principle. 

Theorem 14.7.12 (Complex interpolation of Bessel potential spaces).
Let (X0 , X1 ) be an interpolation couple of UMD Banach spaces and let
p0 , p1 ∈ (1, ∞), s0 , s1 ∈ R, and θ ∈ (0, 1). Then
[H s0 ,p0 (Rd ; X0 ), H s1 ,p1 (Rd ; X1 )]θ = H s,p (Rd ; Xθ ) with equivalent norms,
1 1−θ θ
where p = p0 + p1 , s = (1 − θ)s0 + θs1 , and Xθ = [X0 , X1 ]θ .
Proof. Let R and S be the operator of Lemma 14.7.11. Let
Ej := εsj ,pj (Lpj (Rd ; Xj )), Fj := H sj ,pj (Rd ; Xj ), j ∈ {0, 1},
and set Eθ := (E0 , E1 )θ,q and Fθ := (F0 , F1 )θ,q . Then, by Theorem 2.2.6 and
Lemma 14.7.10, Eθ = εs,p (Lp (Rd ; Xθ )) isomorphically. Now the proof can be
completed in the same way as in Theorem 14.4.30, replacing `qws by εs,p and
s
Bp,q by H s,p everywhere. 
Theorem 14.7.12 contains several results of Volume I as special cases. To
begin with, it contains Theorem 5.6.9, which asserts that if X is a UMD
space, p ∈ (1, ∞), and s0 < s1 , then
[H s0 ,p (Rd ; X), H s1 ,p (Rd ; X)]θ = H sθ ,p (Rd ; X)
and, if in addition s > 0,
[Lp (Rd ; X), H s,p (Rd ; X)]θ = H θs,p (Rd ; X)
up to equivalent norms. It also contains Theorem 5.6.1, which asserts that if
X is a UMD space, p ∈ (1, ∞), and k > 1 is an integer, then
[Lp (Rd ; X), W k,p (Rd ; X)]θ = H θk,p (Rd ; X)
up to an equivalent norm. This result is obtained by taking X0 = X1 =
X, p0 = p1 = p, s0 = 0, and s1 = k in Theorem 14.7.12 and noting that
H k,p (Rd ; X) = W k,p (Rd ; X) up to equivalent norm by Theorem 5.6.11.
Upon combining Theorem 14.7.12 with Theorem 5.6.11 we obtain another
extension of Theorem 5.6.1:
Corollary 14.7.13 (Complex interpolation for Sobolev spaces). Let
(X0 , X1 ) an interpolation couple of UMD Banach spaces and let p0 , p1 ∈
(1, ∞), k0 , k1 ∈ N, and θ ∈ (0, 1). Then
[W k0 ,p0 (Rd ; X0 ), W s1 ,p1 (Rd ; X1 )]θ = H kθ ,p (Rd ; Xθ ) with equivalent norms,
1 1−θ θ
where p = p0 + p1 , kθ = (1 − θ)k0 + θk1 , and Xθ = [X0 , X1 ]θ .
404 14 Function spaces

As in Examples 14.4.33 and 14.4.35, we can use this corollary to prove bound-
edness of pointwise multiplication by smooth functions:

Example 14.7.14 (Pointwise multiplication by smooth functions – I). Let X

and Y be UMD spaces, let p ∈ [1, ∞] and s ∈ R, and let k ∈ [s, ∞) ∩ N be an
integer. If ζ ∈ Cbk (Rd ; L (X, Y )), then pointwise multiplication

f 7→ ζf

defines a bounded mapping from H s,p (Rd ; X) into H s,p (Rd ; Y ) of norm .k,s
kζkCbk (Rd ;L (X,Y )) .
Indeed, the pointwise multiplier f 7→ ζf is bounded as a mapping from
W j,p (Rd ; X) into W j,p (Rd ; Y ) for each j ∈ {0, . . . , k}. Therefore, for s ∈ N the
result is immediate from Theorem 5.6.11. If −s ∈ N, then the result follows by
the duality result of Proposition 5.6.7 and Theorem 5.6.11. If s ∈ (0, ∞), then
the result follows by interpolation between the cases j = 0 and j = k by the
complex method [·, ·] ks and applying Theorem C.2.6 and Corollary 14.7.13.
Finally, the case s ∈ (−∞, 0) follows by duality again.

14.7.d Pointwise multiplication by 1R+ in H s,p

To conclude this section we present a result on pointwise multiplication by 1R+

for vector-valued Bessel potential spaces. The cases of vector-valued Besov
spaces and Triebel–Lizorkin space have been considered in Section 14.6.h;
in both cases, values in general Banach spaces X could be allowed. In the
Bessel potential case, the proof below requires the UMD property of the range
space X. It seems to be an open problem whether this conditions is actually
necessary.

Theorem 14.7.15 (Pointwise multiplication by 1R+ ). Let p ∈ (1, ∞)

and s ∈ (−1/p0 , 1/p), and let X be a UMD space. For all f ∈ H s,p (R; X) we
have 1R+ f ∈ H s,p (R; X) and

k1R+ f kH s,p (R;X) 6 Ckf kH s,p (R;X) , f ∈ H s,p (R; X).

The UMD property of X will only be used through the following proposition.

Proposition 14.7.16. Let p ∈ (1, ∞) and s > 0, and let X be a UMD space.
(1) The operator (−∆)s : S (Rd ; X) → S 0 (Rd ; X) given by

(−∆)s f = |2π · |s fb

uniquely extends to (−∆)s ∈ L (H s,p (Rd ; X), Lp (Rd ; X)).

(2) For all f ∈ H s,p (Rd ; X) the following norm equivalence holds

kf kH s,p (Rd ;X) hp,X kf kLp (Rd ;X) + k(−∆)s/2 f kLp (Rd ;X) .
 14.7 Bessel potential spaces 405
|2πξ|s
Proof. (1): Let m1 (ξ) = (1+|2πξ|2 )s/2
. Using Mihlin’s multiplier Theorem
5.5.10 one can check that m1 ∈ ML (Rd ; X, Y ). Therefore,
p

k(−∆)s f kp = kTm1 Js f kp 6 km1 kMLp (Rd ;X,Y ) kJs f kp 6 Cp,X kf kH s,p (Rd ;X) .

(2): Note that since s > 0, Proposition 5.6.6 gives that H s,p (Rd ; X) ,→
L (Rd ; X) contractively. This combined with (1) gives the estimate “&”.
p
(1+|2πξ|2 )s/2
The estimate . follows similarly. Let m2 (ξ) = 1+|2πξ|s . Then m2 ∈
MLp (Rd ; X, Y ) as before. Therefore,

kf kH s,p (Rd ;X) = kTm2 (I + (−∆)s/2 )f kp

6 km2 kMLp (Rd ;X,Y ) (kf kp + k(−∆)2/s f kp .


We need two more preparatory results. The first one is a concrete formula for
(−∆)s/2 f as an integral operator.
Lemma 14.7.17. Let s ∈ (0, 1). For f ∈ S (R; X) we have

f (· + h) − f (·)
Z
(−∆)s/2 f = cs dh, x ∈ R,
R |h|1+s

where the integral on the right-hand side converges absolutely pointwise R, and
as a Bochner integral in Lp (R; X) for any p ∈ [1, ∞). Here cs ∈ R \ {0} is a
constant only depending on s.
Proof. The convergence of the integral for |h| > 1 is immediate. The conver-
R1
gence for |h| < 1 follows by writing f (x + h) − f (x) = 0 f 0 (x + th)h dt.
To prove the stated identity we take Fourier transforms on the right-hand
side and use Fubini’s theorem to obtain
Z 2πihξ
f (· + h) − f (·) −1 b
Z
e
F 1+s
dh dx = 1+s
f (ξ) dh = ks |ξ|s fb(ξ),
R |h| R |h|

where from the fact that the odd part of the integral cancels we see that
ks = 2 R+ cos(2πt)−1
R
t1+s dt is in (−∞, 0). This proves the result with constant
cs = ks−1 (2π)s . 

We also need the following inequality.

Lemma 14.7.18 (Hilbert absolute inequality). Let p ∈ (1, ∞). For f ∈

Lp (R+ ) one has

|f (y)|
Z
x 7→ dy p 6 Cp kf kLp (R+ ) .
R+ x + y L (R+ )
406 14 Function spaces
1/p
Proof. Letting ζp (y) = xx+1 , after rewriting the integral, we can use Young’s
inequality for the multiplicative group R+ with Haar measure dx x to obtain

|f (y)|
Z Z Z
dy p dx 1/p
x 7→ dy p = ζp (x/y)y 1/p f (y)
R+ x + y L (R+ ) R+ R+ y x
6 kζp kL1 (R+ , dx ) kf kLp (R+ ) .
x


Proof of Theorem 14.7.15. By Proposition 14.6.17 it suffices to prove the de-
sired estimate for f in the dense class Cc∞ (R \ {0}) ⊗ X. In that case one
actually has g := 1R+ f is in the same class and thus is smooth as well.
We claim that

k(−∆)s/2 gkp 6 k(−∆)s/2 f kp + Cp,s kf kH s,p (R;X) . (14.97)

As soon as we proved the claim, then the result follows. Indeed, applying
Proposition 14.7.16 twice we obtain

kgkH s,p (R;X) hp,X kgkp + k(−∆)s/2 gkp

(14.97)
6 kf kp + k(−∆)s/2 f kp + Cp,s kf kH s,p (R;X) .
hp,X kf kH s,p (R;X) .

To rewrite (−∆)s/2 g in a suitable way, let

S := {(x, h) ∈ R2 : (x > 0 and h < −x) or (x < 0 and h > −x)}.

Then applying Lemma 14.7.17 twice, by elementary considerations we see that

for all x ∈ R,
g(x + h) − g(x)
Z
s/2
(−∆) g(x) = cs dh
|h|1+s
ZR
f (x + h) − f (x)
Z
f (x + h)
= cs 1+s
dh − c s sgn(x) 1S (x, h) dh
R |h| R |h|1+s
Z
f (x + h)
= (−∆)s/2 f (x) − cs sgn(x) 1S (x, h) dh.
R |h|1+s
Taking Lp -norms, we see that (14.97) holds if we can show that
kf (x + h)k
Z
x 7→ 1S (x, h) dh p .p,s kf kH s,p (R;X) . (14.98)
R |h|1+s L (R)

To prove (14.98) we only consider the part Lp (R+ ) as the other one is similar.
By elementary considerations
Z ∞ Z Z ∞  Z −x
kf (x + h)k p kf (x + h)k p
1S (x, h) dh dx = dh dx
0 R |h|1+s 0 −∞ |h|1+s
 14.8 Notes 407
∞ Z ∞
kf (−y)k
Z p
= dh dy
0 0 (y + h)1+s
∞ ∞
y −s kf (−y)k p
Z Z
6 dh dy
0 0 y+h
(i)
6 Cpp ky 7→ |y|−s f (y)kpLp (R;X)
(ii)
6 Cpp Cp,s
p
kf kpH s,p (R;X) ,

where in (i) we applied Lemma 14.7.18 to the function y 7→ y −s kf (−y)k,

and (ii) follows from Corollary 14.6.31(2). This completes the proof of the
remaining estimate (14.98). 

14.8 Notes
Early influential monographs on function spaces are those of Adams [1975] (see
also Adams and Fournier [2003]), Bergh and Löfström [1976], Peetre [1976],
and Triebel [1978]. After these works appeared, a new maximal function argu-
ment was discovered by Peetre [1975] which made it possible to study Besov
and Triebel–Lizorkin spaces in the full range p, q ∈ (0, ∞]. This theory is pre-
sented in detail in the monograph of Triebel [1983] and the more recent works
of Triebel [1992, 2006, 2020, 2013, 2014]; further expositions are due to Ba-
houri, Chemin, and Danchin [2011], Denk and Kaip [2013], Grafakos [2009],
Maz’ya [2011], Runst and Sickel [1996], and Sawano [2018].
Standard references for function spaces in the vector-valued setting in-
clude the works of Amann [1995, 1997, 2019], Triebel [1997], König [1986],
Schmeisser [1987], Schmeisser and Sickel [2001], and Schmeisser and Sickel
[2005]. A unified treatment of Besov and Triebel–Lizorkin spaces and related
classes of function spaces is given by Lindemulder [2021], where the axiomatic
setting of Hedberg and Netrusov [2007] is extended to the vector-valued con-
text. In particular, this covers the weighted and anisotropic settings, and it
allows for Banach function space other than the spaces `q (Lp ) or Lp (`q ) em-
ployed in the construction of the Besov and Triebel–Lizorkin spaces.
The theory of function spaces is a vast topic, and by necessity our
treatment does not cover a number of important topics such as approxi-
mation theory, wavelets, atomic decompositions, weighted spaces, paraprod-
ucts, anisotropic spaces, and typical aspects for bounded domains and man-
ifolds such as traces, extension operators, boundary values, and interpola-
tion with boundary conditions (although some of these topics will be briefly
visited in these notes). Of the omitted themes, we specifically mention the
φ-transform of Frazier and Jawerth [1990], which allows the identification of
Besov and Triebel–Lizorkin spaces with subspaces of appropriate discrete se-
quence spaces. In this identification, the question of boundedness of various
operators on the original function spaces is transformed into the question of
408 14 Function spaces

boundedness of infinite matrices on the corresponding sequence spaces, which

in turn can be deduced from natural almost diagonality estimates of these
matrices, in certain analogy with our proof of the T (1) theorem on Lp (Rd ; X)
spaces through estimates of the matrix coefficients of T with respect to the
Haar basis. This approach lies behind many of the proofs of T (1) theorems in
Besov and Triebel–Lizorkin spaces that we discussed in the Notes of Chapter
12.
The ‘classical’ Besov and Triebel–Lizorkin spaces considered in this chap-
ter are modelled on the gradient ∇ in the setting of Rd . It is possible to
introduce Besov and Triebel–Lizorkin spaces based on different types of sec-
torial operators and to study them in the setting of manifolds; we refer to
Batty and Chen [2020], Haase [2006], Kriegler and Weis [2016], Kunstmann
and Ullmann [2014], Taylor [2011a], Taylor [2011b], Taylor [2011c], Taylor
[1974], and Voigtlaender [2022].

Section 14.2
Lemma 14.2.1 is taken from Amann [1995]. The other results of this section are
standard in the scalar-valued case, and their extensions to the vector-valued
setting are straightforward.

Section 14.3
The complex and real interpolation results for vector-valued and weighted Lq -
spaces of Theorems 14.3.1 and 14.3.4 extend Theorems 2.2.6 and 2.2.10, where
the unweighted case was treated. The scalar-valued case goes back to Stein
and Weiss [1958], and the extension to the vector-valued weighted setting is
well-known, at least for complex interpolation. The case of real interpolation
is included in the work of Kreı̆n, Petunı̄n, and Semënov [1982], and a dif-
ferent approach based on Stein interpolation for the real method is due to
Lindemulder and Lorist [2022]. The interpolation results for q0 = q1 = ∞ are
false in general. Indeed, already Triebel [1978, 1.18.1] gave an example where
[`∞ ∞ ∞ ns
ws0 (X0 ), `ws1 (X1 )]θ 6= `ws ([X0 , X1 ]θ ) with ws (n) = 2 . Propositions 14.3.3
and 14.3.5 are presented by Triebel [1978], who attributes the real case to Pee-
tre [1967]. More generally, Triebel [1978, Section 1.18] identifies the complex
and real interpolation spaces of `p0 ((Xj )j>1 ) and `p1 ((Yj )j>1 ) for p0 , p1 < ∞
and for sequences of interpolation couples (Xj , Yj )j>1 ; here `p ((Zj )j>1 ) is the
space of all sequences (zj )j>1 with zj ∈ Zj such that (kzj kZj )j>1 belongs to
`p , Z ∈ {X, Y }. Proposition 14.3.3 then follows by taking Xj = 2js X and
Xj = 2js Y . It seems that Proposition 14.3.5 can only be stated for a single
space X unless further assumptions on q0 and q1 are made.

Section 14.4
Our introduction of vector-valued Besov spaces is self-contained up to a mod-
est number of prerequisites from earlier chapters. Part of the section follows
 14.8 Notes 409

the presentation by Schmeisser and Sickel [2001]. For the history of Besov
spaces, we refer the reader to Bergh and Löfström [1976] and Triebel [1978,
1983]. Besov spaces appear naturally as real interpolation spaces between Lp
and W k,p (see Theorem 14.4.31). As such, they have important applications
in the theory of evolution equations (see Chapter 18). Moreover, by choosing
the microscopic parameter q suitably, one can often include end-point cases
into the considerations.
In contrast to the theory of the spaces W k,p (Rd ; X) and H s,p (Rd ; X),
where assumptions on the space X such as the Radon–Nikodým property or
the UMD property are often needed, many key results on vector-valued Besov
spaces hold for general Banach spaces X.
Lemma 14.4.5 on the sequential completeness of S 0 (Rd ; X) is a standard
result. It is possible to endow the space Cc∞ (U ; X) with a complete locally
convex topology in such a way that sequential convergence in this topology co-
incides with the ad hoc notion of sequential convergence used here. A detailed
construction is presented by Rudin [1991].

Fourier multipliers

Fourier multipliers for vector-valued Besov spaces have been discussed by

Amann [1997], Weis [1997], Girardi and Weis [2003a], Hytönen [2004], and
Hytönen and Weis [2006a]. In Theorem 14.4.16, we only considered smooth
m, and this restriction was removed in Theorem 14.5.6. The latter result
and related ones can be found in the work of Girardi and Weis [2003a], who
showed that the operator T is a continuous extension (with respect to a weaker
topology) of Tm also if max{p, q} = ∞. Fourier multipliers for vector-valued
Besov spaces have been applied by Weis [1997] to obtain sharp exponential
stability results of C0 -semigroups in spaces with Fourier type p.

Embedding

The sandwich result of Proposition 14.4.18 is very useful in avoiding additional

conditions on the Banach space X. The Sobolev embedding result of Theorem
14.4.19 is standard. Especially the sufficiency is simple to prove via Lemma
14.4.20. For the proof of this lemma and its extension to all 0 < p0 < p1 < ∞
in Remark 14.6.4, we follow Schmeisser and Sickel [2001].

Difference norms

The difference norm characterisation of Besov spaces can be found in many

places. It was already used before the Fourier analytic description of Besov
spaces was given. We refer the reader to Bergh and Löfström [1976], Triebel
[1983], and references therein for historical details. The difference norms have
the advantage that in certain cases one can check by hand whether a given
function belongs to some given Besov space. By choosing the parameter τ in
Theorem 14.4.24 appropriately, the Besov spaces can be identified with other
410 14 Function spaces

classical spaces, as we have done in Corollaries 14.4.25 and 14.4.26 for W s,p
s
and Cub .
In Step 1 of the proof of Theorem 14.4.24 we follow the presentation of
Bergh and Löfström [1976], where the case τ = ∞ was given. Step 2 of the
proof is based on the presentation of Schmeisser and Sickel [2001].

Interpolation

Interpolation of Besov spaces is discussed by Bergh and Löfström [1976],

König [1986], and Triebel [1978, 1983]; further references to the literature
can be found in these works. The method to reduce the proofs to interpola-
tion of `q (Lp )-spaces fits into a more general retraction–co-retraction scheme
explained by [Triebel, 1978, Theorem 1.2.4].
The complex interpolation result of Theorem 14.4.30 is folklore, although
we are not aware of a reference containing the general form with an interpo-
lation couple (X0 , X1 ) presented here. In the special case X = X0 = X1 , the
theorem can be proved in the same way as in the scalar-valued case, and some
end-point results are valid as well. For instance, we have
s
[Bps00 ,q0 (Rd ; X), Bps11 ,q1 (Rd ; X)]θ = Bp,q (Rd ; X), pj , qj ∈ [1, ∞], sj ∈ R,

with equivalent norms, where

1 1−θ θ 1 1−θ θ
= + , = + , s = (1 − θ)s0 + θs1 .
p p0 p1 q q0 q1
The real interpolation result of Theorem 14.4.31 is well known, and the
proof is a simple generalisation of the standard proof for the scalar-valued
case. Several other real interpolation results can be proved with the same
methods. For instance, if min{p0 , p1 } < ∞, min{q0 , q1 } < ∞, and s0 , s1 ∈ R,
then
(Bps00 ,q0 (Rd ; X0 ), Bps11 ,q1 (Rd ; X1 ))θ,p = Bp,p
s
(Rd ; (X0 , X1 )θ,p ),
with equivalent norms, where again p1 = 1−θ θ 1−θ θ
p0 + p1 = q0 + q1 and s =
(1 − θ)s0 + θs1 . This follows Theorem 14.3.4 in a similar way as in Theorem
14.4.30.

Duality
s
In Theorem 14.4.34, we identified the dual of Bp,q (Rd ; X) with respect to
the duality for S (R ; X) and S (R ; X). Unlike in the Lp -setting treated in
d 0 d

Section 1.3, no conditions on X are needed. A result of this type in a more

general abstract setting (including weights and anisotropic function spaces)
is presented by Lindemulder [2021]. The proof that we have given follows
Agresti, Lindemulder, and Veraar [2023].
 14.8 Notes 411

Section 14.5

The characterisations in Theorem 14.5.1 of type and cotype in terms of em-

bedding properties of Besov spaces into spaces of γ-radonifying operators are
due to Kalton, Van Neerven, Veraar, and Weis [2008]. This paper also con-
tains the γ-Bernstein–Nikolskii inequality of Lemma 14.5.2, as well as optimal
embedding results for the smooth spaces γ(H −s,2 (Rd ); X). The consequences
for Bessel potential spaces discussed in Corollary 14.7.7 are taken from Veraar
[2013]. This work also contained the following result:
Theorem 14.8.1. Let X be a Banach lattice, and 1 6 p 6 2 6 q < ∞. If X
is p-convex and q-concave, then
1 1
H ( p − 2 )d,p (Rd ; X) ,→ γ(L2 (Rd ), X),
1 1
γ(L2 (Rd ), X) ,→ H ( q − 2 )d,q (Rd ; X).

It is an open problem to characterise the Banach spaces for which these em-
beddings hold (see Problem Q.14).

Mapping properties of the Fourier transform

The mapping properties of the vector-valued Fourier transform F for Banach

spaces X with Fourier type p contained in Proposition 14.5.3 appear in the pa-
pers by Garcı́a-Cuerva, Kazaryan, Kolyada, and Torrea [1998], König [1991],
and Girardi and Weis [2003a]. Real interpolation of the end-point cases q = p
and q = ∞ in Proposition 14.5.3 gives an alternative proof of some of the
results in the papers just mentioned:
Theorem 14.8.2. Suppose that X has Fourier type p ∈ (1, 2]. Let q ∈ (p, ∞),
r ∈ [1, ∞], and s = dp − dq . Then F is bounded from Bp,r
s
(Rd ; X) into the
0
Lorentz space Lq ,r (Rd ; X).
Proposition 14.5.3 contains a parallel result under the assumption that X
has type p and cotype 2. Recall from Proposition 13.1.35 that, under these
assumptions, X has Fourier type r for any r ∈ [1, p).
The mapping properties of the Fourier transform on vector-valued Lp -
spaces with power weights have been recently studied by Dominguez and
Veraar [2021], who show that a version of the classical Pitt inequalities holds
if and only if X has non-trivial Fourier type. In particular, the following result
was proved:
Theorem 14.8.3. Let X be of Fourier type p0 ∈ (1, 2]. Let 1 < p 6 q < ∞
and β, γ > 0. If
  
1 1 d  1 1
max 0, d + −1 <γ< and β − γ = d 1 − − ,
min{p, p0 } q q p q

then F extends boundedly from Lp (Rd , | · |βp ; X) into Lq (Rd , | · |−γq ; X).
412 14 Function spaces

1
In the limiting case γ = max{0, d( min{p,p 0}
+ 1q − 1)}, the above boundedness
of F still holds true under further restrictions on p and q. Surprisingly, if
X has non-trivial Fourier type (equivalently, by Theorem 13.1.33, non-trivial
type), one can allow p = q = 2 by choosing the weights suitably. A similar
result holds in the periodic setting, but the problem is open for more general
orthogonal systems that have been considered by Stein [1956].

R-boundedness

R-boundedness of smooth operator-valued functions is studied by Girardi and

Weis [2003c] under Fourier type conditions, and by Hytönen and Veraar [2009]
under (co)type conditions; the latter paper contains Theorems 14.5.8 and
14.5.9.

Section 14.6

In this section, we followed part of the presentation of Schmeisser and Sickel

[2001]. For a detailed description of the history of Triebel–Lizorkin spaces, we
refer the reader to Bergh and Löfström [1976], and Triebel [1978, 1983]. Below,
we only discuss those aspects of Triebel–Lizorkin spaces that are specific for
this class of spaces.
s
Triebel–Lizorkin spaces Fp,q were originally introduced as a natural vari-
ant of Besov spaces, with the roles of Lp and `q interchanged in the definition.
s
The special case q = 2 leads to the equality Fp,2 = H s,p with equivalent
norms for p ∈ (1, ∞), and in the early days of the theory the cases q 6= 2
were mostly studied for reasons of mathematical curiosity. The definition of
s
Triebel–Lizorkin spaces given here does not cover the spaces F∞,q . The lat-
ter are known to be connected to BMO spaces, and require a modification of
the definition for which we refer to Triebel [1983]. These spaces are naturally
s s,1/p
contained, as F∞,q = Fp,q for any p ∈ (0, ∞), in the general framework of
s,τ
Triebel–Lizorkin-type spaces Fp,q with a fourth parameter τ ∈ [0, ∞), which
has been introduced by Yang and Yuan [2008] and studied in several subse-
quent works.

Genesis of (vector-valued) Triebel–Lizorkin spaces

Vector-valued Triebel–Lizorkin spaces are needed for the treatment of parab-

olic boundary value problems in the spaces Lp (0, T ; Lq (Rd+ )). Such applica-
tions first appeared in the works of Weidemaier [2002] for q 6 p and scalar
second order equations with inhomogeneous Dirichlet boundary conditions,
and of Denk, Hieber, and Prüss [2007] for p, q ∈ (1, ∞) and more general sys-
tems and boundary conditions. Kunstmann [2015] introduced a new interpo-
s
lation method (·, ·)θ,`q and shows that Fp,q = (Lp , W k,p )s/k,`q with equivalent
norms. This interpolation method fits into the axiomatic setting of discrete
interpolation recently developed by Lindemulder and Lorist [2021].
 14.8 Notes 413

As in the Besov space case, results for vector-valued Triebel–Lizorkin

spaces typically hold without restrictions on the target Banach space X.
Thanks to the sandwich result
s s
Bp,1 ,→ Fp,1 ,→ H s,p ,→ Fp,∞
s s
,→ Bp,∞ ,

one can sometimes deduce results about vector-valued Bessel potential spaces
as well. Within the Triebel–Lizorkin scale, one can get closer to H s,p than in
the Besov scale, which often makes Triebel–Lizorkin spaces more useful. For
instance, the sandwich result can be combined with the Sobolev Embedding
Theorem 14.6.14, which allows arbitrary microscopic improvement for Triebel–
Lizorkin spaces. Further flexibility in sandwiching and embedding theorems
can be built in by introducing weights such as |x|γ or |x1 |γ as was done by
Meyries and Veraar [2012, 2014a].
The boundedness of the Peetre maximal function proved in Proposition
14.6.2 appears in the book of Triebel [1997]. This proposition extends results
of Triebel [1983, Theorem 1.6.3] and Triebel [1997, Formula 15.3(iv)] to the
vector-valued setting.
Theorems 14.6.3 and 14.6.11 are presented by Triebel [1997] for scalar-
valued multipliers m. An operator-valued extension is due to Bu and Kim
[2005].

Gagliardo–Nirenberg inequalities and Sobolev embedding

The Gagliardo–Nirenberg inequalities of Proposition 14.6.15 and 14.6.16 are

taken from Brezis and Mironescu [2001]. Our presentation follows Schmeisser
and Sickel [2001, 2005]. Proposition 14.6.13 and Theorem 14.6.14 can also be
found in these works. Gagliardo–Nirenberg inequalities in the Besov scale can
be found in the paper of Brezis and Mironescu [2018]; they do not allow for a
microscopic improvement.

Difference norms

Difference norm characterisations of Triebel–Lizorkin spaces appear in the

works of Kaljabin [1977, 1979], and Triebel [1983]. Our presentation of Theo-
rem 14.6.20 follows Schmeisser and Sickel [2001], who consider the case τ = 1.

Interpolation and duality

The interpolation and duality results for Triebel–Lizorkin spaces are similar
to their Besov space counterparts. In our presentation, the end-point q = 1 is
excluded, since the Fefferman–Stein inequality for the maximal operator is not
valid in Lp (Rd ; `1 ). This problem can be circumvented by a reduction to in-
terpolation identities for vector-valued Hardy spaces instead of Lp (Rd ; `q (X))
(see Triebel [1983]). The embedding (14.87) of Theorem 14.6.26 is due to
Jawerth [1977], and the one of (14.86) to Franke [1986].
414 14 Function spaces

Fractional Hardy inequalities

The fractional Hardy inequalities of Proposition 14.6.30 and Corollary 14.6.31
are variations of those by Krugljak, Maligranda, and Persson [2000], who
proved the results with a fractional Sobolev norm W s,p on the right-hand
side. The advantage of our formulation is that both the H s,p and the W s,p
s
cases are consequences of the stronger estimate using the space Fp,∞ . Higher-
dimensional versions of fractional Hardy inequalities can be deduced from
the work of Meyries and Veraar [2012], where Sobolev embedding with power
weights are discussed.
Pointwise multiplication by 1R+
Pointwise multiplier results such as the one of Theorem 14.6.32 and Corollaries
14.6.34 and 14.6.35 were proved via paraproducts estimates in more general-
ity by Runst and Sickel [1996]. Some of the results from this monograph were
extended to the weighted vector-valued setting by Meyries and Veraar [2015].
In particular, some of the end-points can be included, and higher dimensional
versions of the results hold. The results of the present section merely serve as
an illustration of how the theory can be applied. Since the work of Grisvard
[1967] and Seeley [1972], it is known that results on pointwise multipliers stand
at the basis of interpolation with boundary conditions. The one-dimensional
1 1
case is useful for evolution equations, since 0 Fp,q (R+ ; X) and 0 Bp,q (R+ ; X)
can be used as the domain of the time-derivative. As in the work of Lin-
demulder, Meyries, and Veraar [2018], one can identity the real and complex
1 0
interpolation spaces between 0 Fp,q (R; X) and Fp,q (R; X) for p, q ∈ (1, ∞) us-
ing the theory of this section, and similarly for Besov spaces for p ∈ (1, ∞)
and q ∈ [1, ∞].

Section 14.7
The Embedding Theorems 14.7.1, 14.7.3, and 14.7.4 are taken from Schmeisser
and Sickel [2001, 2005]. The end-point cases, where min{p0 , p1 } = 1 <
max{p0 , p1 }, are not completely understood; we refer the reader to Brezis
and Mironescu [2018] for a further discussion.
The Littlewood–Paley theorem 14.7.5 is taken from Meyries and Veraar
[2015], who also consider a weighted setting.
The improved embeddings for Besov, Triebel–Lizorkin, and Bessel poten-
tial spaces under UMD and (co)type assumptions stated in Proposition 14.7.6
are due to Veraar [2013]. The converse result presented in Proposition 14.7.8
seems to be new. In the case p = q, Hytönen and Merikoski [2019] have shown
the following more precise result.
Theorem 14.8.4. For k ∈ N and p ∈ [2, ∞), there is a continuous embedding
k
Bq,q (Rd ; X) ,→ W k,q (Rd ; X)
if and only if X has martingale cotype q.
 14.8 Notes 415

In case the embedding constant depend on d in a polynomial way, such results

have applications to quantitative affine approximation in infinite dimensions,
as discussed by Hytönen, Li, and Naor [2016] and Hytönen and Naor [2019].
The proof of Theorem 14.8.4 is based on ideas from these works and results of
Xu [1998] and Martı́nez, Torrea, and Xu [2006] connecting Littlewood–Paley–
Stein inequalities and martingale (co)type. Some of these results have been
extended by Xu [2020]. For open problems related to Theorem 14.8.4, we refer
the reader to Problem Q.13.
Theorem 14.7.9 is due to Han and Meyer [1996], who obtained it as a
consequence of a more general Littlewood–Paley theorem for Lp (Rd ; X). Our
approach is more direct.
The interpolation result of Theorem 14.7.12 was discovered independently
by Amann [2019], Hummel [2019], and Lindemulder and Veraar [2020]. In the
first reference, the anisotropic setting was also covered, and weighted spaces
are included in the latter two references.
Pointwise multipliers
Theorem 14.7.15 is due to Meyries and Veraar [2015], where it appears as
a special case of a general pointwise multiplier theorem for weighted vector-
valued Bessel potential spaces. It is unknown whether the UMD condition
is necessary (see Problem Q.12). The proof presented here is simplified from
that of Lindemulder, Meyries, and Veraar [2018]. Another proof, based on a
difference norm characterisation, is due to Lindemulder [2017]. The scalar case
of Theorem 14.7.15 is due to Shamir [1962] and Strichartz [1967]. Their proof
extends to the vector-valued setting only when the range space is isomorphic
to a Hilbert spaces (see Walker [2003]).
Interpolation with boundary conditions
Applications to complex interpolation with boundary conditions are given by
Lindemulder, Meyries, and Veraar [2018]. Among other things, the domains
of the fractional powers of the first order derivative with Dirichlet boundary
conditions are identified as D(∂ts ) = 0 H s,p (R+ ; X) for s ∈ (0, 1). This extends
a special case of a result of Seeley [1972] to the vector-valued setting. Cer-
tain difficulties in obtaining such identities were overlooked in applications to
evolution equations for several years. The boundedness of pointwise multipli-
cation by indicator functions was proved recently in the anisotropic setting by
Lindemulder [2022]. This solves an open problem of Amann [2019], who used
the boundedness to obtain vector-valued and anisotropic extensions of some
of the results of Seeley [1972] on interpolation with boundary conditions.

Function spaces on domains and extension operators

Function spaces on domains O ⊆ Rd are usually defined by restriction, declar-
ing that f ∈ Asp,q (O) if there exists g ∈ Ap,q
s
(Rd ) such that f = g|O in the dis-
s
tributional sense; the norm on Ap,q (O) is then taken to be the corresponding
416 14 Function spaces

quotient norm. From this definition, it is often complicated to decide whether

a given function belongs to Asp,q (O) and to estimate its norm. Extension oper-
ators help to get a better grip on this problem. Given a domain O ⊆ Rd , an ex-
tension operator for O is a bounded linear operator EO : Asp,q (O) → Asp,q (Rd )
such that
(EO f )|O = f, f ∈ Asp,q (O).
For Lipschitz domains O, Rychkov [1999] constructed a ‘universal’ extension
operator EO which enjoys this property for all s ∈ R, p, q ∈ (0, ∞], and
A ∈ {B, F }. His proof extends to the vector-valued and weighted setting. A
crucial ingredient is the work of Bui, Paluszyński, and Taibleson [1996, 1997],
where the restriction that the Littlewood–Paley function ϕ should have com-
pact Fourier support is relaxed to a moment condition on ϕ and a Tauberian
condition on ϕ.b
Once an extension operator is available, one often tries to obtain an intrin-
sic characterisation of the functions in Asp,q (O), e.g., in terms of differences
and moduli of smoothness. As a consequence of the result of Rychkov [1999], a
s
difference characterisation for Bp,q (O) was obtained in Dispa [2003] for Lips-
s
chitz domains O. A difference norm characterisations for Fp,q (O) was obtained
by Prats [2019] for ε-uniform domains (in particular, for Lipschitz domains).
Other ways to construct extension operators can be found in the books
of Triebel [1983, 1992]. A classical method is to find an extension operator
for W k,p (O), and use real and complex interpolation and duality to obtain an
s
extension operators for Bp,q (O) and H s,p (O) with |s| < k and q ∈ [1, ∞]. This
approach also works for Triebel–Lizorkin spaces if one uses the `q -interpolation
method from Kunstmann [2015] and Lindemulder and Lorist [2021]. These
techniques can also be used for vector-valued function spaces.
Another way to define function spaces on domains is by using wavelets;
see Triebel [2006].

Weighted function spaces

s
Bui [1982] defined and studied theSspaces Bp,q (Rd , w) and Fp,q
s
(Rd , w) for
all weights w in the class A∞ = p>1 Ap , where Ap denotes the class of
Muckenhoupt weights as defined in Appendix J. Crucial to this approach is
the Peetre maximal function and the weighted version of Theorem 3.2.28.
The vector-valued setting was introduced and studied by Meyries and Veraar
[2012, 2015, 2014b], Lindemulder, Meyries, and Veraar [2018], and, from a
more abstract point of view, Lindemulder [2021].
Matrix-weighted Besov spaces have been introduced and investigated by
Roudenko [2003, 2004] for p ∈ [1, ∞), and by Frazier and Roudenko [2004,
0
2008] for p ∈ (0, ∞). The special case Fp,2 (W ) of matrix-weighted Triebel–
Lizorkin spaces, and its identification with Lp (W ), was already considered by
Nazarov and Treil [1996] and Volberg [1997], and more recently by Isralowitz
[2021], but a systematic introduction and study of the full scale of these spaces
 14.8 Notes 417

is only recently due to Frazier and Roudenko [2021]. Matrix-weighted versions

s,τ
Fp,q (W ) of the generalised Triebel–Lizorkin-type spaces of Yang and Yuan
[2008] have been subsequently studied by Bu, Hytönen, Yang, and Yuan [2023].

Two-weight Sobolev embedding

Haroske and Skrzypczak [2008] characterised the validity of the continuous

embedding
Bps00 ,q0 (Rd , w0 ) ,→ Bps11 ,q1 (Rd , w1 )
in terms of the weights w0 , w1 ∈ A∞ , the exponents p0 , p1 , q0 , q1 ∈ (0, ∞],
and the smoothness parameters s0 > s1 . The compactness of this embedding
was characterised as well. A characterisation for Triebel–Lizorkin spaces was
obtained by Meyries and Veraar [2014b] under the additional assumption p0 6
p1 ; as in Theorem 14.6.14, a microscopic improvement occurs. In the vector-
valued setting, the case of power weights is fully understood; see Meyries and
Veraar [2012].

Lp –Lq -multipliers

In the scalar-valued case, Lp –Lq Fourier multiplier theorems for p < q first
appeared in the pioneering work of Hörmander [1960]. The scalar-valued case
has the advantage that one can often factor through an L2 -space and use
Plancherel’s identity. In the Banach space-valued case, this is no longer possi-
ble unless additional conditions on the spaces are imposed. The singularities
in Lp –Lq -multiplier theorems for p < q usually behave in a different way from
the case p = q. Often they are absolutely integrable in some appropriate sense,
and then trivially extend to the vector-valued setting by Proposition 2.1.3. A
typical example where this happens is the classical Hardy–Littlewood–Sobolev
inequality on the Lp –Lq -boundedness of f 7→ | · |−s ∗ f .
For operator-valued Lp –Lq -Fourier multipliers, different phenomena arise.
For details and applications to stability of C0 -semigroups we refer the reader
to Rozendaal and Veraar [2018a, 2017, 2018c,b] and the survey by Rozendaal
[2023]. The homogeneous version of Corollary 14.7.7 implies the following
multiplier result of Rozendaal and Veraar [2018a].

Theorem 14.8.5. Let X be a Banach space with type p0 ∈ (1, 2] and let Y
be a Banach space with cotype q0 ∈ [2, ∞). Let p ∈ (1, p0 ) and q ∈ (q0 , ∞),
where we allow p = 2 if p0 = 2 and q = 2 if q0 = 2. Let r ∈ [1, ∞] satisfy
r = p − q . If m : R \ {0} → L (X, Y ) is a strongly measurable function in
1 1 1 d

the strong operator topology, and such that

{|ξ|d/r m(ξ) : ξ ∈ Rd \ {0}}

is γ-bounded, then Tm uniquely extends to a bounded operator from Lp (Rd ; X)

to Lq (Rd ; Y )).
418 14 Function spaces

The proof of this theorem is based on factorisation through γ(L2 (Rd ), X) and
uses the γ-boundedness of the stated operator family. To obtain a homoge-
neous condition on m, one needs the homogeneous version of the γ-Sobolev
embedding. It is not known whether Theorem 14.8.5 holds for p = p0 and
q = q0 . An exception is the case where X and Y are p-convex and q-concave
Banach lattices, respectively; the result then follows from the homogeneous
version of Theorem 14.8.1. Theorem 14.8.5 was used by Rozendaal [2019] to
obtain boundedness of the H ∞ -calculus on fractional domain spaces for strip
type operators. Rozendaal and Veraar [2018a] also prove the following multi-
plier theorem under Fourier type assumptions.

Theorem 14.8.6. Let X be a Banach space with Fourier type p0 ∈ (1, 2] and
let Y be a Banach space with Fourier type q00 ∈ (1, 2]. Let p ∈ (1, p0 ) and
q ∈ (q0 , ∞), and let r ∈ [1, ∞) satisfy 1r = p1 − 1q . If m : Rd \ {0} → L (X, Y )
is a strongly measurable functions and m ∈ Lr,∞ (Rd ; L (X, Y )), then Tm
uniquely extends to a bounded operator from Lp (Rd ; X) to Lq (Rd ; Y )).

The condition m ∈ Lr,∞ (Rd ; L (X, Y )) allows for singularities of the form
|·|−d/r . The proof in the case Cm,r := kmkL (X,Y ) Lr (Rd ) < ∞ with p10 − q10 =
1
r0 is completely straightforward. Indeed, by Hölder’s inequality,

kTm f kq0 6 ϕq00 ,Y (Rd )kmfbkq00 6 ϕq00 ,Y (Rd )Cm,r kfbkp00

6 ϕq00 ,Y (Rd )ϕp0 ,Y (Rd )Cm,r kf kp0 .

Theorem 14.8.6 can be deduced from this estimate by an interpolation argu-

ment.
The above Fourier multiplier theorems are stated for one specific value
of p and q. However, if the kernel (see Hörmander [1960]) or the multiplier
(see Rozendaal and Veraar [2017]) satisfies certain Hörmander conditions,
boundedness from Lu into Lv can be shown for all u, v ∈ (1, ∞) satisfying
1 1 1 1 1
u − v = p − q =: r . For example, a sufficient condition is

sup |ξ||α|+d/r k∂ α m(ξ)k < ∞, |α| 6 b rd0 c + 1.

ξ6=0

Under Fourier type assumptions on X and Y , the number of derivatives can

be further reduced.
Proposition 14.5.7 can be viewed as a mixed Besov–Lq -Fourier multiplier
theorem in the same spirit as Theorems 14.8.5 and 14.8.6.
15
Extended calculi and powers of operators

In this chapter we address two strongly interwoven topics: How to verify the
boundedness of the H ∞ -calculus of an operator and how to represent and
estimate its fractional powers. For concrete operators such as the Laplace
operator or elliptic partial differential operators, the fractional domain spaces
can often be identified with certain function spaces considered in Chapter 14
and the imaginary powers of the operator are related to singular integral and
pseudo-differential operators treated in Chapters 11 and 13.
Throughout this chapter, unless otherwise stated, we let A be a sectorial
operator on a Banach space X. We work over the complex scalar field.

15.1 Extended calculi

In Chapter 10 we have introduced the Dunford calculus
f 7→ f (A),
defined for functions f ∈ H 1 (Σσ ), the space of holomorphic functions on
Σσ that are integrable with respect to the measure dz z (in the sense of (15.1)
below). We performed a detailed study of the class of operators whose Dunford
calculus, when restricted to H 1 (Σσ )∩H ∞ (Σσ ) extends to a functional calculus
for functions in H ∞ (Σσ ).
In the present section we extend the Dunford calculus of a sectorial op-
erator A to holomorphic functions f of polynomial growth on Σσ . Although
the operators f (A) in this calculus are generally unbounded, the mapping
f 7→ f (A) still shares many properties with bounded functional calculi. This
extended calculus includes all functions in H ∞ (Σσ ), and it agrees with the
H ∞ (Σσ )-calculus of A when this operator has a bounded H ∞ (Σσ )-calculus.
In the next section, it will enable us to define the fractional powers Aα in
terms of the holomorphic functions z α . Sectorial operators A whose imag-
inary powers Ait are bounded are of special interest in view of their close
relationship with a variety of topics studied in these volumes.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 419
T. Hytönen et al., Analysis in Banach Spaces, Ergebnisse der Mathematik und ihrer
Grenzgebiete. 3. Folge / A Series of Modern Surveys in Mathematics 76,
https://doi.org/10.1007/978-3-031-46598-7_5
420 15 Extended calculi and powers of operators

We briefly recall some notation and terminology introduced in Volume II

that will be used throughout this chapter. For 0 < σ < π we denote by

Σσ := {z ∈ C \ {0} : | arg(z)| < σ}

the open sector of angle σ in the complex plane; the argument is taken in
the interval (−π, π). A linear operator (A, D(A)) is sectorial if there exists
σ ∈ (0, π) such that the spectrum σ(A) is contained in Σσ and

Mσ,A := sup kzR(z, A)k < ∞.

z∈{Σσ

Here, for z ∈ %(A), the resolvent set of A, R(z, A) := (z − A)−1 denotes the
resolvent of A. In this situation we say that A is σ-sectorial with constant
Mσ,A . The infimum of all σ ∈ (0, π) such that A is σ-sectorial is called the
angle of sectoriality of A and is denoted by ω(A).
By H 1 (Σσ ) we denote the Banach space of all holomorphic functions f :
Σσ → C for which

kf kH 1 (Σσ ) := sup kt 7→ f (eiν t)kL1 (R+ , dt ) < ∞. (15.1)

t
|ν|<σ

Our objective in this section is to extend the Dunford calculus f 7→ f (A)

to larger classes of functions. This is achieved in two steps: in Subsection
15.1.a we adjoin the constant-one function and the function (1 + z)−1 . Among
other things, this allows us to treat bounded rational functions as well as
bounded functions such as exp(−z). This calculus provides the starting point
for Subsections 15.1.b and 15.1.c, where we extend the calculus to a class of
unbounded functions whose growth at the origin and at infinite is controlled by
a regularising function. Among other things this, extended Dunford calculus
will allow us to define fractional powers of A.

15.1.a The primary calculus

Our first aim is to extend the Dunford calculus f 7→ f (A) of a sectorial

operator A to a slightly larger class of functions f for which one still obtains
bounded operators, while preserving the multiplicativity of the calculus.

Definition 15.1.1. For 0 < σ < π we define E(Σσ ) to be the vector space of
holomorphic functions f : Σσ → C of the form
a
f (z) = f0 (z) + + b,
1+z
where f0 ∈ H 1 (Σσ ) ∩ H ∞ (Σσ ) and a, b ∈ C.

We could, more generally, allow functions f0 ∈ H 1 (Σσ ) here, but not much
is gained by doing so because any such function belongs to H ∞ (Σν ) for all
 15.1 Extended calculi 421

0 < ν < σ (see Proposition H.1.3). This additional generality would in fact
cause some inconvenience in the statement of the multiplicativity rule (Propo-
sition 15.1.4), where one would be forced to switch to slightly smaller angles.
A further advantage of the present definition is that E(Σσ ) is contained in
H ∞ (Σσ ) as a linear subspace.

Lemma 15.1.2. A bounded holomorphic function f : Σσ → C belongs to

E(Σσ ) if and only if it has integrable limits at 0 and ∞, by which we mean
that there exist constants c0 , c∞ ∈ C such that f −c0 and f −c∞ are integrable
with respect to dzz near 0 and ∞, respectively, in the sense that

sup kt 7→ f (eiν t) − c0 kL1 ((0,1), dt ) < ∞

t
|ν|<σ

and
sup kt 7→ f (eiν t) − c∞ kL1 ((1,∞), dt ) < ∞.
t
|ν|<σ

a
Proof. If f = E(Σσ ) is of the form f (z) = f0 (z) + 1+z + b one may take
c0 = a + b and c∞ = b. In the converse direction, if the bounded holomorphic
function f : Σσ → C has integrable limits c0 and c∞ at 0 and ∞, respectively,
−c∞
then f0 (z) := f (z) − c01+z − c∞ belongs to H 1 (Σσ ) ∩ H ∞ (Σσ ). 
The following functions belong to E(Σσ ):

zm
z 7→ for 0 < σ < π and integers n > m > 0;
(1 + z)n
z→
7 exp(−ζz) for 0 < σ < 21 π and ζ ∈ Σ 12 π−σ .

For the first this follows by multiplicativity (proved in Proposition15.1.4 be-

low) and the fact that z 7→ (1+z)−1 and z 7→ z(1+z)−1 = 1−(1+z)−1 belong
to E(Σσ ). For the second this follows by noting that both exp(−ζz)−(1+ζz)−1
and (1 + ζz)−1 − (1 + z)−1 are in H 1 (Σσ ) ∩ H ∞ (Σσ ). Another example will
be encountered in the proof of Theorem 15.2.8.

Definition 15.1.3 (Primary calculus). Let A be a sectorial operator on a

Banach space X and let ω(A) < σ < π. For functions f ∈ E(Σσ ) the bounded
operator f (A) ∈ L (X) is defined by

f (A) := f0 (A) + a(I + A)−1 + bI,

where
a
f (z) = f0 (z) + +b
1+z
with f0 ∈ H 1 (Σσ ) ∩ H ∞ (Σσ ) and a, b ∈ C, and with f0 (A) defined through
the Dunford calculus.
422 15 Extended calculi and powers of operators

Since the constants a and b are uniquely determined by f this is well defined.
For functions in f ∈ H 1 (Σσ ) ∩ H ∞ (Σσ ) the primary calculus of a sectorial
operator A agrees with the Dunford calculus. If A has a bounded H ∞ (Σσ )-
calculus and D(A) ∩ R(A) is dense in X, then for functions f ∈ E(Σσ ) the
definitions of f (A) through the primary calculus agrees with that through the
H ∞ -calculus; this is because in the H ∞ -calculus we have 1+z
1
(A) = (I + A)−1
and 1(A) = I by Theorem 10.2.13.

Proposition 15.1.4. Let A be a sectorial operator on a Banach space X and

let ω(A) < σ < π. For all f, g ∈ E(Σσ ) we have f g ∈ E(Σσ ) and

(f g)(A) = f (A)g(A).

Proof. Let f, g ∈ E(Σσ ) be represented as in Definition 15.1.1. It is clear

that the product f0 g0 belongs to H 1 (Σσ ) ∩ H ∞ (Σσ ) and that the product of
z 7→ (1 + z)−1 with a function in H 1 (Σσ ) ∩ H ∞ (Σσ ) is in H 1 (Σσ ) ∩ H ∞ (Σσ )
again. Finally,
1 1 1 z
· = −
1+z 1+z 1+z (1 + z)2
and the right-hand side is in E(Σσ ). This proves that f g ∈ E(Σσ ).
We have f0 g0 ∈ H 1 (Σσ )∩H ∞ (Σσ ), and the multiplicativity of the Dunford
calculus gives
f0 (A)g0 (A) = (f0 g0 )(A).
Also, with φ(z) = 1/(1 + z) and ζ(z) = z/(1 + z 2 ),

φ(A)2 = (I + A)−2 = φ(A) − ζ(A) = (φ − ζ)(A) = φ2 (A),

where we used Proposition 10.2.3 to see that ζ(A) = A(I + A)−2 in the
Dunford calculus and hence in the primary calculus. Thus it remains to check
that φ(A)f0 (A) = (φf0 )(A). This follows by applying the resolvent identity
and Cauchy’s theorem to the contour integral representation of the Dunford
calculus:
Z
1
φ(A)f0 (A) = f0 (z)(I + A)−1 R(z, A) dz
2πi Γ
Z
1 f0 (z)
= [R(z, A) − R(−1, A)] dz
2πi Γ 1 + z
Z
1 f0 (z)
= R(z, A) dz
2πi Γ 1 + z
= (φf0 )(A).

This completes the proof. 

Example 15.1.5 (Bounded rational functions). As a first application let us

prove that if A is sectorial, then for all integers m > n > 0 we have
 15.1 Extended calculi 423

zm
(A) = Am (I + A)−n ,
(1 + z)n
m
z
noting that z 7→ (1+z) n belongs to E(Σσ ) for all 0 < σ < π.

By Proposition 15.1.4,
zm  z m  1 n−m
(A) = (A) (A)
(1 + z)n 1+z 1+z
= (A(I + A)−1 )m (I + A)m−n = Am (I + A)−n ,

where we used that

z 1
(A) = 1(A) − (A) = I − (I + A)−1 = A(I + A)−1 .
1+z 1+z
Example 15.1.6 (Exponential functions). In this example we assume that A is
sectorial with ω(A) < 21 π. For ω(A) < σ < 21 π and ζ ∈ Σ 12 π−σ define

exp(−ζA) := exp(−ζz)(A),

noting that z 7→ exp(−ζz) belongs to E(Σσ ).

By Proposition 15.1.4,

exp(−ζ1 A) exp(−ζ2 A) = exp(−(ζ1 + ζ2 )z)(A).

Furthermore, for all x ∈ X and n > 1 we have exp(−ζA)x ∈ D(An ) and

(z n exp(−ζz))(A)x = An exp(−ζA)x.

To see this denote the left-hand side by g(A). By Proposition 15.1.4 and
Example 15.1.5,
1  zn 
(I + A)−n g(A) = (A)g(A) = exp(−ζz) (A)
(1 + z)n (1 + z)n
zn
= (A) exp(−ζz)(A) = An (I + A)−n exp(−ζz)(A),
(1 + z)n
from which the claim follows.
The preceding example connects with semigroup theory through Proposition
10.2.7 in Volume II which can be restated in the present language of primary
calculus as follows.
Theorem 15.1.7. Let A be a densely defined sectorial operator on X with
angle ω(A) < 21 π, and let ω(A) < σ < 21 π. Then the bounded holomorphic C0 -
semigroup (S(z))z∈Σ 1 π−σ generated by −A is given by the primary calculus
2
through
S(z) = exp(−zA), z ∈ Σ 21 π−σ ,
where exp(−zA) = exp(−z ·)(A) as in the preceding example.
424 15 Extended calculi and powers of operators

15.1.b The extended Dunford calculus

Throughout this section, A is a sectorial operator on a Banach space X and
we fix ω(A) < σ < π. We proceed to define an extension of the primary
calculus f 7→ f (A) for suitable unbounded functions f . The idea is to use
a regularising function % to “tame” the growth of f near the origin and at
infinity. The resulting operators f (A) are unbounded in general, but they
nevertheless enjoy various good properties. For functions f ∈ H ∞ (Σσ ) and
%(z) = z/(1 + z)2 the construction proposed in the definition has already been
used in Volume II (see (10.14)).
Definition 15.1.8 (Regularisers, extended Dunford calculus). Let A
be a sectorial operator on a Banach space X and let ω(A) < σ < π. Let
f : Σσ → C be holomorphic. A function % ∈ E(Σσ ) is called a regulariser on
Σσ for the pair (f, A) if the following two conditions are met:
• %f ∈ E(Σσ );
• the operator %(A) defined by the primary calculus is injective.
We say that (f, A) is Σσ -regularisable if such a regulariser exists, and in that
case we define

D(f (A)) := x ∈ X : (%f )(A)x ∈ R(%(A)) ,
f (A)x := %(A)−1 (%f )(A)x, x ∈ D(f (A)).
The mapping f 7→ f (A) is referred to as the extended calculus!Dunford of A.
If % is a Σσ -regulariser for the pair (f, A), then so is ρ% for any ρ ∈ H 1 (Σσ ) ∩
H ∞ (Σσ ) such that ρ(A) is injective. Since ρ% ∈ H 1 (Σσ )∩H ∞ (Σσ ), this shows
that regularisers may be assumed to lie in H 1 (Σσ ) ∩ H ∞ (Σσ ) whenever this
is convenient.
In what follows we omit the prefix ‘Σσ -’ whenever the choice of the angle
σ is clear from the context.
A trivial consequence of the first assertion in Proposition 15.1.4 is that if
% ∈ E(Σσ ), then for every function f ∈ E(Σσ ) we have %f ∈ E(Σσ ). If %(A)
is injective, Proposition 15.1.4 implies that for all f ∈ E(Σσ ) the definitions
of f (A) in Definitions 15.1.3 and 15.1.8 agree.
The following proposition shows that the definition of the operator f (A)
is independent of the regulariser.
Proposition 15.1.9 (Well-definedness). Let A be a sectorial operator on
a Banach space X and let ω(A) < σ < π. Let f : Σσ∨τ → C be holomorphic,
where σ, τ ∈ (ω(A), π). If % ∈ E(Σσ ) and ϑ ∈ E(Στ ) are regularisers for
(f, A), then
 
x ∈ X : (%f )(A)x ∈ R(%(A)) = x ∈ X : (ϑf )(A)x ∈ R(ϑ(A))

and, for all x ∈ X belonging to this common set,

%(A)−1 (%f )(A)x = ϑ(A)−1 (ϑf )(A)x.

 15.1 Extended calculi 425

Proof. Replacing σ and τ by σ ∧ τ we may assume that σ = τ . Denote the

domains defined in the statement of the lemma by D% (f (A)) and Dϑ (f (A)). If
x ∈ D% (f (A)), then (%f )(A)x = %(A)y for some y ∈ X. By Proposition 15.1.4
we have ϑ%f ∈ E(Σσ ) and

%(A)(ϑf )(A)x = (ϑ%f )(A)x = ϑ(A)(%f )(A)x = ϑ(A)%(A)y = %(A)ϑ(A)y,

and therefore (ϑf )(A)x = ϑ(A)y by the injectivity of %(A). This shows that
(ϑf )(A)x ∈ R(ϑ(A)), so x ∈ Dϑ (f (A)), and

ϑ(A)−1 (ϑf )(A)x = y = %(A)−1 (%f )(A)x.

Interchanging the roles of % and ϑ, one also sees that if x ∈ Dϑ (f (A)), then
x ∈ D% (f (A)). This concludes the proof. 
The following observation is an immediate consequence of Proposition 15.1.4.
Lemma 15.1.10. Let A be a sectorial operator on a Banach space X and let
ω(A) < σ < π. If f, g : Σσ → C are holomorphic functions and % ∈ E(Σσ )
and ϑ ∈ E(Σσ ) are regularisers for (f, A) and (g, A), respectively, then %ϑ is
a regulariser for both (f, A) and (g, A).
Proposition 15.1.11. Let A be a sectorial operator on a Banach space X
and let ω(A) < σ < π. Let f : Σσ → C be a holomorphic function such that
the pair (f, A) is regularisable.
(1) the operator f (A) is closed;
(2) if % ∈ E(Σσ ) regularises (f, A), then R(%(A)) ⊆ D(f (A)) and

f (A)x = (%f )(A)%(A)−1 x, x ∈ R(%(A)).

Proof. (1): Let xn ∈ D(f (A)) satisfy xn → x and f (A)xn → y in X

as n → ∞. Then (%f )(A)xn → (%f )(A)x since (%f )(A) is bounded, and
%(A)−1 [(%f )(A)xn ] = f (A)xn → y by the definition of f (A). The closedness of
%(A)−1 implies (%f )(A)x ∈ D(%(A)−1 ) = R(%(A)) and %(A)−1 [(%f )(A)x] = y.
By the definition of D(f (A)), this means that x ∈ D(f (A)) and f (A)x = y.
This proves the closedness of f (A).
(2): For x ∈ R(%(A)), say x = %(A)y, we have

(%f )(A)x = (%f )(A)%(A)y = %(A)(%f )(A)y ∈ R(%(A)).

Therefore x ∈ D(f (A)) and

f (A)x = %(A)−1 (%f )(A)x = (%f )(A)y = (%f )(A)%(A)−1 x.


Proposition 15.1.12. Let A be a sectorial operator on a Banach space X
and let ω(A) < σ < π. Let f, g : Σσ → C be holomorphic functions such that
the pairs (f, A) and (g, A) are regularisable.
426 15 Extended calculi and powers of operators

(1) for all a, b ∈ C the pair (af + bg, A) is regularisable, and for all x ∈
D(f (A)) ∩ D(g(A)) we have x ∈ D((af + bg)(A)) and

(af + bg)(A)x = af (A)x + bg(A)x.

(2) the pair (f g, A) is regularisable and

D(f (A)g(A)) = D(g(A)) ∩ D((f g)(A)),

and for all x ∈ X belonging to the common set we have

(f g)(A)x = f (A)g(A)x.

In particular, f (A)g(A)x is closable. If g(A) is bounded, then

(f g)(A) = f (A)g(A)

with equal domains.

(3) if f is zero-free and the pair (1/f, A) is regularisable, then f (A) is injective
and
1

(A) = f (A)−1
f
with equality of domains. In particular, if A is injective and if we set
inv(z) := 1/z, then (inv, A) is regularisable and

inv(A) = A−1 .

Proof. By Lemma 15.1.10 we may select a function % ∈ E(Σσ ) that regularises

both (f, A) and (g, A) (in parts (1) and (2)), respectively both (f, A) and
(1/f, A) (in part (3)).
(1): It is clear that if %f, %g ∈ E(Σσ ), then %(af + bg) ∈ E(Σσ ). The
assumption x ∈ D(f (A))∩D(g(A)) implies that (%f )(A)x and (%g)(A)x belong
to R(%(A)) and therefore we have (%(af + bg))(A)x ∈ R(%(A)). Hence x ∈
D((af + bg)(A)) and

(af + bg)(A)x = %(A)−1 (%(af + bg))(A)x

= a%(A)−1 (%f )(A)x + b%(A)−1 (%g)(A)x = af (A)x + bg(A)x.

(2): By assumption we have %f, %g ∈ E(Σσ ). By Proposition 15.1.4 we

also have %2 f g ∈ E(Σσ ). By multiplicativity we have %2 (A) = (%(A))2 , so
%2 (A) is injective. It follows that the operator (f g)(A) is well defined in the
extended Dunford calculus.
Let x ∈ D(g(A)) ∩ D((f g)(A)). Then, by the definition of g(A)x, multi-
plicativity, and the definition of (f g)(A)x,

(%f )(A)g(A)x = (%f )(A)%(A)−1 (%g)(A)x

= %(A)−1 (%f )(A)(%g)(A)x
 15.1 Extended calculi 427

= %(A)−1 (%2 f g)(A)x

= %(A)%(A)−2 (%2 f g)(A)x
= %(A)(f g)(A)x.

This shows that (%f )(A)g(A)x ∈ R(%(A)) and therefore g(A)x ∈ D(f (A)),
i.e., x ∈ D(f (A)g(A)), and

(f g)(A)x = %(A)−1 (%f )(A)g(A)x = f (A)g(A)x.

In the converse direction, let x ∈ D(f (A)g(A)). Then x ∈ D(g(A)) and

g(A)x ∈ D(f (A)), so (%f )(A)g(A)x ∈ R(%(A)), say (%f )(A)g(A)x = %(A)y.
Then,

(%2 f g)(A)x = (%f )(A)(%g)(A)x = %(A)(%f )(A)g(A)x = %(A)2 y = %2 (A)y.

This shows that (%2 f g)(A)x belongs to R(%2 (A)), so x ∈ D((f g)(A)) by Propo-
sition 15.1.9 and

(f g)(A)x = %2 (A)−1 (%2 f g)(A)x = y = %(A)−1 (%f )(A)g(A)x = f (A)g(A)x.

By part (1) of Proposition 15.1.11 the operator (f g)(A) is closed, and the
above argument shows that it extends f (A)g(A), so f (A)g(A) is closable.
(3): Noting that D((1/f )f )(A) = D(1(A)) = D(I) = X, it follows from
part (2) that if x ∈ D(f (A)), then x ∈ D((1/f )(A)f (A)) and (1/f )(A)f (A)x =
x. Reversing the roles of f and 1/f we also obtain that if x ∈ D((1/f )(A)),
then (1/f )(A)x ∈ D(f (A)) and f (A)(1/f )(A)x = x.
The second assertion follows by considering, e.g., the regulariser %(z) =
z/(1 + z). 
As a consequence of what has been shown in the course of the proof of part
(2), and by applying (2) with f and g interchanged, we find that f (A) and
g(A) commute in the following sense: we have

f (A)g(A)x = g(A)f (A)x = (f g)(A)x

for x ∈ D(f (A)) ∩ D(g(A)) ∩ D(f g(A)).

We continue with a characterisation of the domain of f (A) which, in view
of later applications, we formulate in two versions. For integers n > 1 we write
n n 1
rn (z) := , ζn (z) := − .
n+z n+z 1 + nz
These functions belong to E(Σσ ).

Proposition 15.1.13. Let A be a sectorial operator on a Banach space X

and let ω(A) < σ < π. Let f : Σσ → C be a holomorphic function, and fix an
integer k > 1.
428 15 Extended calculi and powers of operators

(1) If D(A) is dense in X and rnk f ∈ E(Σσ ), then D(Ak ) is densely contained
in D(f (A)), we have
n o
D(f (A)) = x ∈ X : lim (rnk f )(A)x exists in X ,
n→∞

and, for all x ∈ D(f (A)),

f (A)x = lim (rnk f )(A)x.
n→∞

(2) If D(A) ∩ R(A) is dense in X and ζnk f ∈ E(Σσ ), then D(Ak ) ∩ R(Ak ) is
densely contained in D(f (A)), we have
n o
D(f (A)) = x ∈ X : lim (ζnk f )(A)x exists in X ,
n→∞

and, for all x ∈ D(f (A)),

f (A)x = lim (ζnk f )(A)x.
n→∞

In either case, f (A) is densely defined.

Proof. (1): Let %(z) := r1 (z) = (1 + z)−1 . Then %k = rnk ∈ E(Σσ ) and
D(Ak ) = R(%k (A)), so the inclusion R(%k (A)) ⊆ D(f (A)) of Proposition
15.1.11 implies that D(Ak ) ⊆ D(f (A)).
Let x ∈ D(f (A)) and set xn := rnk (A)x. Then xn ∈ D(Ak ) ⊆ D(f (A)),
and by Proposition 10.1.7 we have limn→∞ xn = x (here we use that D(A) is
dense) and
lim f (A)xn = lim f (A)rnk (A)x = lim rnk (A)f (A)x = f (A)x,
n→∞ n→∞ n→∞

where the middle identity follows from the second part of Proposition 15.1.11,
observing that rnk is a regulariser for (f, A). This shows that D(Ak ) is dense
in D(f (A)).
If x ∈ D(f (A)), multiplicativity and the fact that %k rnk f ∈ E(Σσ ) imply
rnk (A)f (A)x = %k (A)−1 rnk (A)(%f )(A)x
= %k (A)−1 (%rnk f )(A)x = (rnk f )(A)x.
Therefore limn→∞ (rnk f )(A)x exists and equals f (A)x.
Conversely, suppose that x ∈ X is such that limn→∞ (rnk f )(A)x =: y exists.
Put zn := rnk (A)(%k f )(A)x. Then zn ∈ D(Ak ), so zn ∈ R(%k (A)). Moreover
zn → (%k f )(A)x, and, by multiplicativity,
%k (A)−1 zn = %k (A)−1 rnk (A)(%k f )(A)x
= %k (A)−1 (%k rnk f )(A)x = (rnk f )(A)x → y.
Since %k (A)−1 is closed it follows that (%k f )(A)x belongs to D(%k (A)−1 ) =
R(%k (A)), and therefore x ∈ D(f (A)).
(2): This is proved in the same way as (1), replacing the use of rn and
Proposition 10.1.7 by ζn and Proposition 10.2.6. 
 15.1 Extended calculi 429

The following result improves Proposition 15.1.12(2) under an additional as-

sumption.

Proposition 15.1.14. Let A be a sectorial operator on a Banach space X

and let ω(A) < σ < π. Let f, g : Σσ → C be holomorphic functions such that
the pairs (f, A) and (g, A) are regularisable. Then f (A)g(A) is closable and

f (A)g(A) = (f g)(A)

in each of the following two cases:

(1) D(A) is dense in X, and f and g are bounded near 0 and have at most
polynomial growth near ∞;
(2) D(A) ∩ R(A) is dense in X, and f and g have at most polynomial growth
near 0 and ∞.

Proof. The closability of f (A)g(A) has already been proved in Proposition

15.1.12. We prove (1); the proof of (2) is entirely similar.
With % := r1 as in the previous proof, the growth assumption implies
that for large enough k > 1 the functions %k f , %k g, and %2k f g belong to
E(Σσ ). Moreover, D(Ak ) = R(%k (A)). The domain D(A2k ) equals R(%2k (A)),
which in turn is contained in D((f g)(A)) by Proposition 15.1.11 applied
with %2k and f g. We also have D(A2k ) ⊆ D(Ak ) ⊆ D(g(A)), and hence
D(A2k ) ⊆ D(f (A)g(A)) by Proposition 15.1.12. Moreover, since D(A) is dense
in X, D(A2k ) is dense in D((f g)(A)) by Proposition 15.1.13. It follows that
D(f (A)g(A)) is dense in D((f g)(A)). 

Theorem 15.1.15 (Composition). Let A be a sectorial operator on a Ba-

nach space X and let ω(A) < σ < π. Let f : Σσ → C be a holomorphic
function such that the pair (f, A) is regularisable, and assume that

f (Σσ ) ⊆ Στ

for some 0 < τ < π. Suppose furthermore that f (A) is sectorial with
ω(f (A)) < τ . If g : Στ → C is a holomorphic function such that the pairs
(g, f (A)) and (g ◦ f, A) are regularisable, then

g(f (A)) = (g ◦ f )(A)

holds under either one of the following additional assumptions:

(i) g ∈ E(Στ );
(ii) (g, f (A)) admits a regulariser ϕ ∈ E(Στ ) such that ϕ ◦ f ∈ E(Σσ ).

The proof depends on the following lemma.

Lemma 15.1.16. Let A be a sectorial operator on a Banach space X and let

ω(A) < σ < π. Let f : Σσ → C be a holomorphic function such that
430 15 Extended calculi and powers of operators

f (Σσ ) ⊆ Στ

for some 0 < τ < π. If % ∈ E(Σσ ) be a regulariser for (f, A) and λ 6∈ Στ ,

then it is a regulariser for ((λ − f )−1 , A) as well, and
1
(A) = R(λ, f (A)).
λ − f (z)

Proof. By assumption, %f ∈ E(Σσ ) and %(A) is injective. By Lemma 15.1.2,

% and %f have integrable limits at 0 and at ∞, say c0 , c∞ and d0 , d∞ , re-
spectively. Putting fλ := 1/(λ − f ), we wish to show that %fλ has integrable
limits at 0 and at ∞; another application of Lemma 15.1.2 then implies that
this function belongs to E(Σσ ), so % regularises (fλ , A). The identity in the
statement of the lemma then follows from Proposition 15.1.12(3).
%(·)
If c∞ = 0, then |f (·) − λ| > δ1 := dist(λ, Στ ) > 0 implies that λ−f (·) has
integrable limit 0 at ∞. Suppose next that c∞ 6= 0. We claim that d∞ /c∞ ∈
Στ . Indeed, otherwise we had
d∞
f (·) − > δ2 := dist(d∞ /c∞ , Στ ) > 0. (15.2)
c∞
d∞
Since both %f and c∞ % have integrable limit d∞ at ∞, the identity

d∞ d∞
%f = %(f − )+ %
c∞ c∞

implies that %(f − dc∞∞

) has integrable limit 0 at ∞. But then (15.2) would
imply that % has integrable limit 0 at ∞, contradicting the assumption that
this integrable limit satisfies c∞ 6= 0. This proves the claim.
With δ := min{δ1 , δ2 } it now follows from

%(z) c2∞
−
λ − f (z) c∞ λ − d∞
%(z) c∞ %(z) c∞ (%(z) − c∞ )
6 − +
λ − f (z) c∞ λ − d∞ c ∞ λ − d∞
c∞ (%(z)f (z) − d∞ ) − d∞ (%(z) − c∞ ) c∞ (%(z) − c∞ )
= +
(λ − f (z))(c∞ λ − d∞ ) c ∞ λ − d∞
1   1
6 |c ∞ ||%(z)f (z) − d ∞ | + |d ∞ ||%(z) − c ∞ | + |%(z) − c∞ |
c∞ δ 2 δ
%(·) c2
that %fλ = λ−f (·) has integrable limit c∞ λ−d∞ at ∞.
∞

Replacing c∞ and d∞ by c0 and d0 , in the same way one sees that %fλ has
c20
integrable limit 0 at 0 if c0 = 0, and integrable limit c0 λ−d 0
at 0 if c0 6= 0. 
Proof of Theorem 15.1.15. We begin with the proof of the theorem under the
additional assumption made in (i), namely, that g ∈ E(Στ ).
 15.1 Extended calculi 431

Step 1 – For g = 1 the theorem is trivial since 1 ◦ f = 1 and 1(f (A)) =

(1 ◦ f )(A) = I. For g(z) = 1/(1 + z) we have g(f (A)) = (I + f (A))−1 and
(g ◦ f )(A) = (1 + f (z))−1 (A) = (I + f (A))−1 , the former by Definition 15.1.3
applied to f (A) and the latter by Lemma 15.1.16.
Step 2 – We now consider a general g ∈ E(Στ ), and write g = g0 +
a/(1 + z) + b with a, b ∈ C and g0 ∈ H 1 (Στ ) ∩ H ∞ (Στ ). Let % ∈ E(Σσ )
be a regulariser for (f, A). By Lemma 15.1.10 we may assume that % also
regularises (g ◦ f, A), and by the observation after Definition 15.1.8 we may
also assume that % ∈ H 1 (Σσ ) ∩ H ∞ (Σσ ). As the proof of Lemma 15.1.16
shows, % also regularises ( λ−f1 (·) , A) for λ 6∈ Στ .
Fix ω(A) < µ < σ and ω(f (A)) < ν < τ . By the Dunford calculus of f (A),
the operator g0 (f (A)) is bounded and for all x ∈ X we have
Z
1
g0 (f (A))x = g0 (z)R(z, f (A))x dz.
2πi ∂Σν

If z ∈ ∂Σν , then by Lemma 15.1.16 for all x ∈ X we have

1
R(z, f (A))x = (A)x. (15.3)
z − f (·)

Using (15.3) and multiplicativity of the primary calculus of A, Fubini’s

theorem, the Cauchy integral theorem, and keeping in mind that % ∈ H 1 (Σσ )∩
H ∞ (Σσ ), we obtain
Z
1
%(A)g0 (f (A))x = g0 (λ)%(A)R(λ, f (A))x dλ
2πi ∂Σν
Z
1 %(·)
= g0 (λ) (A)x dλ
2πi ∂Σν λ − f (·)
 1 2 Z Z %(z) 
= g0 (λ) R(z, A)x dz dλ
2πi ∂Σν ∂Σµ λ − f (z)
 1 2 Z Z g0 (λ) 
= %(z) dλ R(z, A)x dz
2πi ∂Σµ ∂Σν λ − f (z)
Z
1
= %(z)g0 (f (z))R(z, A)x dz
2πi ∂Σµ
= (% · (g0 ◦ f ))(A)x.

Setting h0 (z) := a/(1 + z) + b, by Step 1 we also have h0 (f (A)) = (h0 ◦ f )(A)

and therefore, by Proposition 15.1.4,

%(A)h0 (f (A))x = (% · (h0 ◦ f ))(A)x, x ∈ X.

Adding up, we obtain

%(A)g(f (A))x = (% · (g ◦ f ))(A)x, x ∈ X,

432 15 Extended calculi and powers of operators

the operator g(f (A)) being defined by the primary functional calculus of f (A).
Since % regularises (g ◦ f, A), this implies that every x ∈ X belongs to D((g ◦
f )(A)) and

g(f (A))x = (g ◦ f )(A)x, x ∈ X.

This proves that g(f (A)) = (g ◦ f )(A), and both operators are bounded. This
concludes the proof under the assumption made (i).
For the proof of the theorem under the assumption made in (i), let ϕ ∈
E(Στ ) be a regulariser for the pair (g, f (A)) such that ϕ ◦ f ∈ E(Στ ), and let
ρ be a regulariser for (g ◦ f, A).
We claim that under these circumstances, ρ · (ϕ ◦ f ) regularises (g ◦ f, A).
To this end we must show:
• ρ · (ϕ ◦ f ) · (g ◦ f ) ∈ E(Σσ );
• (ρ · (ϕ ◦ f ))(A) is injective.
The first assertion follows from ρ · (g ◦ f ) ∈ E(Σσ ) (since ρ be a regulariser for
(g ◦ f, A)) and ϕ ◦ f ∈ E(Στ ) (by assumption). For the second assertion we
use the multiplicativity rule of Proposition 15.1.12 (noting that (ϕ ◦ f )(A) =
ϕ(f (A)) by the result of Step 2 and the fact that ϕ ∈ E(Στ )) to see that

(ρ · (ϕ ◦ f ))(A) = ρ(A)(ϕ ◦ f )(A) = ρ(A)ϕ(f (A)).

The right-hand side is the composition of two injective operators; this is be-
cause ρ is a regulariser for (g ◦ f, A) and ϕ is a regulariser for (g, f (A)). This
proves the claim.
In the following computation, in (i) we use the definition of a regulariser,
in (ii) we apply the result of Step 2 to ϕg ∈ E(Στ ), noting that ϕg satis-
fies the conditions of the theorem since g does, (iii) follows from Proposition
15.1.12, noting that ((ϕg) ◦ f ))(A) = (ϕg)(f (A)) is a bounded operator since
ϕg ∈ E(Στ ), (iv) is a simple rewriting, (v) follows from the definition of a
regulariser, noting that ρ · (ϕ ◦ f ) regularises (g ◦ f, A), (vi) follows by another
application of Proposition 15.1.12, and (vii) uses the result of Step 2 once
again:
(i)
ρ(A)ϕ(f (A))g(f (A)) = ρ(A)(ϕg)(f (A))
(ii)
= ρ(A)((ϕg) ◦ f )(A)
(iii)
= (ρ · ((ϕg) ◦ f ))(A)
(iv)
= (ρ · (ϕ ◦ f ) · (g ◦ f ))(A)
(v)
= (ρ · (ϕ ◦ f ))(A)(g ◦ f )(A)
(vi)
= ρ(A)(ϕ ◦ f )(A)(g ◦ f )(A)
(vii)
= ρ(A)ϕ(f (A))(g ◦ f )(A).
 15.1 Extended calculi 433

The identity g(f (A)) = (g ◦ f )(A) follows from this since both ρ(A) and
ϕ(f (A)) are injective. 
Our next aim is to relate the extended Dunford calculus with the H ∞ -calculus.

Theorem 15.1.17 (Boundedness of the extended Dunford calculus).

Let A be a sectorial operator on X with D(A) ∩ R(A) dense in X, and let
ω(A) < σ < π. Then for all functions f ∈ H ∞ (Σσ ) the pair (f, A) is regular-
isable and the following assertions are equivalent:
(1) the operator f (A) defined through the extended Dunford calculus is bounded
for all f ∈ H ∞ (Σσ );
(2) the operator A has bounded H ∞ (Σσ )-calculus.
In this situation the operators f (A) defined through the extended Dunford
calculus and the H ∞ -calculus agree.

Proof. By Proposition 10.1.8, the density of D(A) ∩ R(A) implies that A is

injective. As a consequence, for every f ∈ H ∞ (Σσ ) the function ζ(z) = z/(1+
z)2 is a regulariser for the pair (f, A).
(1)⇒(2): By the boundedness of f (A) and the closedness of ζ(A)−1 , the
identity f (A) = ζ(A)−1 (ζf )(A) (note that D(f (A)) = X) implies that f 7→
f (A) is closed as a linear map from H ∞ (Σσ ) to L (X), and therefore bounded,
by the closed graph theorem. Denoting its norm by M , it follows that

kf (A)k 6 M kf k∞

for all f ∈ H ∞ (Σσ ). In particular, this bound holds for all f ∈ H 1 (Σσ ) ∩
H ∞ (Σσ ). For such functions the extended Dunford calculus agrees with the
Dunford calculus, and therefore the estimate tells us that A has a bounded
H ∞ (Σσ )-calculus.
(2)⇒(1): If x ∈ D(A)∩R(A), then x ∈ R(ζ(A)), say x = ζ(A)y. For the op-
erator f (A) defined through the H ∞ -calculus we have, by the multiplicativity
of the H ∞ -calculus,

f (A)x = ζ(A)f (A)y = (ζf )(A)y,

where the operator on the right-hand side is again defined by the H ∞ -calculus.
We can also define the operator (ζf )(A) through the primary calculus, and
these two definitions agree (they agree for functions in H 1 (Σσ ) ∩ H ∞ (Σσ )
and for the functions (1 + z)−1 and 1). It follows that

f (A)y = ζ(A)−1 (ζf )(A)y.

Since D(A) ∩ R(A) is dense in X, this implies that f (A) = ζ(A)−1 (ζf )(A).
The operator on the right-hand side equals the operator f (A) defined through
the extended Dunford calculus, which is therefore bounded. 
434 15 Extended calculi and powers of operators

We finish this section with a perturbation result that will be useful in connec-
tion with bounded imaginary powers (see the proof of Lemma 15.3.8).
Theorem 15.1.18. Let A be a densely defined sectorial operator on a Banach
space X and let ω(A) < σ < π. Let f ∈ H ∞ (Σσ ) be given. If the operator
f (A), defined through the extended Dunford calculus of A is bounded, then
also the operator f (A + I), defined through the extended Dunford calculus of
A + I, is bounded and we have

kf (A + I)k 6 (1 + Mσ,A )2 (kf (A)k + Cσ kf kH ∞ (Σσ ) ),

where Cσ is a constant depending only on σ and Mσ,A is the sectoriality

constant of A at angle σ.
Proof. Note that ω(A + I) 6 ω(A) and fix ω(A) < ν < σ. The injectivity of
A + I implies that the function ζ(z) = z/(z + 1)2 is a regulariser for (f, A + I)
for any f ∈ H ∞ (Σσ ). Since D(A + I) = D(A) and R(A + I) = X, the second
part of Proposition 15.1.13 implies that f (A + I) is densely defined.
By the extended Dunford calculus of A + I, for all x ∈ D(f (A + I)) we
have
Z
1
f (A + I)x = (ζ(A + I))−1 ζ(z)f (z)R(z, A + I)x dz.
2πi ∂Σν

We have 1/ζ(z) = z + 2 + z −1 , and this easily implies (ζ(A + I))−1 = (A +

I) + 2I + (A + I)−1 . Now,
Z
1
(2I + (A + I)−1 ) ζ(z)f (z)R(z, A + I)x dz
2πi ∂Σν
 1 Z 1 
6 (2 + Mν,A )Mν,A+I kxk 2
|dz| kf kH ∞ (Σσ )
2π ∂Σν |z + 1|

and, noting that R(z, A + I) = R(z, A) + R(z, A + I)R(z, A) by the resolvent

identity,
Z
1
(A + I) ζ(z)f (z)R(z, A + I)x dz
2πi ∂Σν
Z
1
6 (A + I) ζ(z)f (z)R(z, A)x dz
2πi ∂Σν
Z
1
+ ζ(z)f (z)(A + I)R(z, A + I)R(z, A)x dz
2πi ∂Σν
6 k(A + I)ζ(A)kkf (A)kkxk
 1 Z 1 
+ (1 + Mν,A+I )Mν,A kxk |dz| kf kH ∞ (Σσ ) .
2π ∂Σν |z + 1|2

Since kAζ(A + I)k = k(A + I)A(A + I)−2 k 6 Mν,A and Mν,A+I 6 Mν,A , this
proves the estimate
 15.1 Extended calculi 435

kf (A + I)xk 6 (1 + Mν,A )2 kf (A)k + Cν (1 + Mν,A )2 kf kH ∞ (Σσ ) kxk

for x ∈ D(f (A + I)), with Cν = π1 ∂Σν |z+1|

1
R
2 |dz|. Since D(f (A + I)) is dense,

this estimate extends to arbitrary x ∈ X. To conclude the proof we let ν ↑ σ

and note that Mν,A → Mσ,A by an easy estimate based on the resolvent
identity. 

15.1.c Extended calculus via compensation

For functions f ∈ H ∞ (Σσ ) and regulariser %(z) := ζ(z) = z/(1 + z)2 there is
different approach to the extended Dunford calculus via the Cauchy integral
formula, which we outline presently.
Let A be a sectorial operator and let f ∈ H ∞ (Σσ ). For ω(A) < τ < σ 0 < σ,
µ ∈ Σσ0 \ Στ , and x ∈ D(A) ∩ R(A) define
Z
1  1 
f (A)x := f (µ)x + f (z) R(z, A) − x dz. (15.4)
2πi ∂Σσ0 z−µ

Let us check that the integrand converges absolutely. Since x ∈ D(A) ∩ R(A)
we may pick y ∈ D(A) with Ay = x. Then

x (A − µ)R(z, A)x kR(z, A)k

R(z, A)x − = 6 (kAxk + µkxk),
z−µ z−µ |z − µ|

which is of the order O(|z|−2 ) as |z| → ∞ along ∂Στ . Also,

x x kxk
R(z, A)x − = R(z, A)Ay − 6 kR(z, A)Ayk +
z−µ z−µ |z − µ|

which is of the order O(1) as |z| → 0 along ∂Στ , noting that kR(z, A)Ayk =
kR(z, A)[(A − z) + z]yk 6 (1 + kzR(z, A)k)kyk. This establishes the claim.
By an application of Cauchy’s theorem, f (A) is independent of µ ∈ Στ \ Στ 0 .
Since the integrand is an integrable R(A)-valued function, we see that

f (A)x ∈ R(A), x ∈ D(A) ∩ R(A).

Note that if f ∈ H 1 (Σσ )∩H ∞ (Σσ ), the above definition of f (A)x agrees with
(10.7).
We will now check that the definition of f (A)x by (15.4) agrees with the
one via Definition 15.1.1 for the regulariser %(z) = ζ(z) = z/(1 + z)2 . Suppose
that x ∈ D(A) ∩ R(A), say x = ζ(A)y. Starting from the latter definition we
have Z
1 zf (z)
f (A)x = (f ζ)(A)y = R(z, A)y dz.
2πi ∂Στ (1 + z)2
Fix ω(A) < τ 0 < τ and µ ∈ Στ \ Στ 0 . To check that (15.4) agrees with
Definition 15.1.1 we must show that
436 15 Extended calculi and powers of operators

f (µ)A(I + A)−2 y
f (z) h z(z − µ)
Z
1
i
= R(z, A)y − (z − µ)R(z, A) − I x dz.
2πi ∂Στ z − µ (1 + z)2

By Cauchy’s formula, the right-hand side integral evaluates to

h  z(z − µ)
i
= f (z) 2
R(z, A)y − (z − µ)R(z, A) − I x = f (µ)x,
(1 + z) z=µ

as was to be proved. We have thus proved:

Proposition 15.1.19. Let A be a sectorial operator on a Banach space X,

let ω(A) < τ 0 < τ < σ < π. Let f ∈ H ∞ (Σσ ). For all µ ∈ Στ \ Στ 0 and
x ∈ D(A) ∩ R(A) the integral
Z
1  1 
f (A)x := f (µ)x + f (z) R(z, A) − x dz
2πi ∂Στ z−µ

converges absolutely and we have f (A)x = (f ζ)(A)y, in agreement with the

definition of f (A)x through the extended Dunford calculus.

The attentive reader will have noticed that we already used this procedure in
Proposition 10.2.7.

15.2 Fractional powers

In this section, we will apply the extended Dunford calculus to introduce

the fractional powers Aα of a sectorial operator A. Particular instances of
fractional powers such as (−∆)1/2 , the square root of the negative Laplacian,
appear all over in Analysis. On a theoretical level, domains of fractional powers
encode useful smoothness properties of the elements in their domains, and
correspond to (or are closely connected with) interpolation scales between
the domain D(A) and the underlying Banach space X. For example, if the
imaginary powers Ait , t ∈ R, are bounded operators, then for all 0 < α < 1 the
fractional domain D(Aα ) equals the complex interpolation space [X, D(A)]α
as a subspace of X, and as a Banach space up to equivalent norms. As we
have seen in Chapter 4, for the negative Laplacian A = −∆ on X = Lp (Rd ),
the latter can be identified as the Bessel potential space H 2α,p (Rd ).
After introducing fractional powers, we establish several basic algebraic
properties and prove several useful representation formulas. In the next sec-
tion, we then take a closer look at the class of sectorial operators whose
imaginary powers are bounded, and prove a number of non-trivial theorems
connecting this property with (R-, γ-)sectoriality and boundedness of the
H ∞ -calculus.
 15.2 Fractional powers 437

15.2.a Definition and basic properties

In what follows, unless otherwise stated we let A be a sectorial operator acting

in a Banach space X. When additional assumptions are needed, they will
always be stated explicitly.
For α ∈ C it is natural to try to define the fractional power Aα by applying
the extended Dunford calculus to the function

fα (z) := z α := eα log z ,

where we use the branch of the logarithm that is holomorphic in C \ (−∞, 0].
Let 0 < |ν| < σ < π. For z = reiν with r > 0 we have

|fα (z)| = |rα | |eiνα | 6 |z|<α eσ|=α| .

For all integers m, n ∈ N, the function

%m,n (z) := z m (1 + z)−m−n

belongs to E(Σσ ), and

• if <α > 0, then %m,n fα ∈ E(Σσ ) for all integers m > 0, n > <α;
• if <α = 0, then %m,n fα ∈ E(Σσ ) for all integers m, n > 1;
• if <α < 0, then %m,n fα ∈ E(Σσ ) for all integers m > |<α|, n > 0.
The operator %m,n (A) = Am (I + A)−m−n (cf. Example 15.1.5) is injective if
m = 0 or A is injective (or both). This shows:

Proposition 15.2.1. Let A be a sectorial operator on a Banach space X. The

pair (fα , A) is regularisable in each of the following two cases:
• <α > 0
• <α 6 0 and A is injective.
In the first case %0,n (z) = (1 + z)−n with n > <α is a regulariser; in the
second case %n,n (z) = z n (1 + z)−2n with n > |<α| is a regulariser.

In view of these considerations the extended Dunford calculus allows us to

make the following definition.

Definition 15.2.2 (Fractional powers). Let A be a sectorial operator on a

Banach space X. For α ∈ C we define

Aα := fα (A), α ∈ C,

in each of the following two cases:

• <α > 0
• <α 6 0 and A is injective.
438 15 Extended calculi and powers of operators

These operators are closed. Moreover, if <λ > 0 and D(A) is dense, then Aα
is densely defined; if <α 6 0 and D(A) ∩ R(A) is dense, then A is injective and
Aα is densely defined. Using the results of Section 10.1.b, These assertions
follow from Proposition 15.1.11, the domain identifications D(An ) = R(%0,n )
and D(An )∩R(An ) = R(%n,n ), and the fact that D(An ) is dense if D(A) dense,
respectively D(An ) ∩ R(An ) is dense in X if D(A) ∩ R(A) is dense in X.
We begin our study of fractional powers with a consistency check.
Proposition 15.2.3. Let A be a sectorial operator on a Banach space X. For
all n = 0, 1, 2, . . . and fn (z) = z n we have

fn (A) = An with equal domains. (15.5)

If in addition A is injective, this identity extends to all n ∈ Z.

Proof. For n = 0 this reduces to the identity 1(A) = I. For n > 1, consider the
function %n (z) = (1+z)−n and let x ∈ D(An ) = R(%n (A)), say x = (I +A)−n y.
Then

fn (A)x = %n (A)−1 (%n fn )(A)x = An x,

where we used that %n (A) = (I + A)−n in the primary calculus, and that
zn n −n
(%n fn )(A) = (1+z) n (A) = A (I + A) in the primary calculus. This proves
n
that A ⊆ fn (A). In the converse direction, if x ∈ D(fn (A)), then

An (I + A)−n x = (%n fn )(A)x ∈ R(%n (A)) = D(An ),

forcing x ∈ D(An ). This completes the proof of (15.5) for n > 1. For n =
−1, −2, . . . the result follows by applying Proposition 15.1.12(3). 
From the definition of the extended Dunford calculus we immediately deduce
the following result.
Proposition 15.2.4. Let A be a sectorial operator on a Banach space X, and
fix an integer k > 1.
(1) For all x ∈ D(Ak ) the function z 7→ Az x is well defined and holomorphic
on {0 < <z < k}.
(2) If A is injective, then for all x ∈ D(Ak ) ∩ R(Ak ) the function z 7→ Az x is
well defined and holomorphic on {−k < <z < k}.
Theorem 15.2.5. Let A be a sectorial operator on a Banach space X, and
let α, α1 , α2 ∈ C.
(1) If A is injective and α ∈ C, then Aα is injective and

A−α = (Aα )−1 = (A−1 )α with equality of domains.

(2) If <α1 > <α2 > 0, then

D(Aα1 ) ⊆ D(Aα2 ) and R(Aα2 ) ⊇ R(Aα1 ),

 15.2 Fractional powers 439

(3) If A is injective and <α1 < <α2 < 0, then

D(Aα1 ) ⊇ D(Aα2 ) and R(Aα1 ) ⊆ R(Aα2 ).

(4) If <α1 > 0 and <α2 > 0, then

Aα1 +α2 = Aα1 Aα2 with equality of domains.

(5) If A is injective and <α1 < 0 and <α2 < 0, then

Aα1 +α2 = Aα1 Aα2 with equality of domains.

Proof. (1): The injectivity of Aα and the identity A−α = (Aα )−1 follow from
Proposition 15.1.12(3). The identity A−α = (A−1 )α follows from Theorem
15.1.15, noting that A−1 is sectorial with the same angle as A.
(2): We consider the regulariser %k (z) = (1 + z)−k , for which we have
R(%k (A)) = D(Ak ).
Let x ∈ D(Aα1 ) and fix an integer k > max{<α2 , <α1 − <α2 }. In order to
prove that x ∈ D(Aα2 ) we must show that ((1 + z)−k z α2 )(A)x ∈ D(Ak ).
Since 2k > <α1 , by the definition of D(Aα1 ) we have ((1+z)−2k z α1 )(A)x ∈
D(A2k ). Using the multiplicativity of the Dunford calculus, this implies that

z k+α2
Ak (I + A)−2k ((1 + z)−k z α2 )(A)x = (A)x
(1 + z)3k
z k−(α1 −α2 ) z α1
= k
(A) (A)x
(1 + z) (1 + z)2k

belongs to D(A2k ). It follows that (I + A)−2k ((1 + z)−k z α2 )(A)x ∈ D(A3k )

and ((1 + z)−k z α2 )(A)x ∈ D(Ak ) as desired. The opposite inclusion of the
ranges follows from part (4) proved below.
(3): If A is injective and <α1 < <α2 < 0 we can apply parts (2) and (1)
with β1 = −α1 and β2 = α, noting that D(Aαj ) = D(A−βj ) = R(Aβj ) and
R(Aαj ) = R(A−βj ) = D(Aβj ).
(4): Let <α1 > 0 and <α2 > 0. Proposition 15.1.12 implies that Aα1 Aα2 x =
α1 +α2
A x for all x ∈ D(Aα1 Aα2 ) = D(Aα2 ) ∩ D(Aα1 +α2 ). It remains to prove
that D(Aα1 +α2 ) ⊆ D(Aα2 ). But this follows from part (2).
(5): This follows from (1) and (4) by taking inverses. 

Proposition 15.2.6. Let A be a sectorial operator on a Banach space X. Let

c ∈ C \ {0} satisfy |arg c| < π − ω(A). Then:
(1) the operator cA is sectorial with angle ω(cA) 6 ω(A) + | arg(c)|, and for
all ω(A) < σ < π − | arg c| we have Mσ+| arg c|,cA 6 Mσ,A ;
(2) for all α ∈ C, and assuming A to be injective if <α 6 0, we have

(cA)α = cα Aα with equality of domains.

440 15 Extended calculi and powers of operators

Proof. Since (λ−cA)−1 = c−1 (c−1 λ−A)−1 , the condition | arg(c)| < π −ω(A)
guarantees that cA is sectorial with ω(cA) 6 ω(A) + |arg c|. Also, for ω(A) <
σ < π − | arg c| and λ ∈ {Σσ+| arg c| we have c−1 λ ∈ {Σσ and

kλR(λ, cA)k = kc−1 λR(c−1 λ, A)k 6 Mσ,A ,

which gives the bound Mσ+| arg c|,cA 6 Mσ,A .

Choose ω > ω(A) such that ω + |arg c| < π. Fix α ∈ C. Then, for x ∈ X
and k > |<α|,
Z
1
(ρk fα )(cA)x = ρk (z)z α R(z, cA)x dz
2πi ∂Σω+| arg c|

with fα (z) = z α , and ρk (z) := %0,k (z) = (1 + z)−k if <α > 0 and ρk (z) :=
%k,k (z) = z k /(1 + z)2k if <α 6 0. By Cauchy’s theorem we can deform the
path in the above integral to Γ = c · ∂Σω and obtain, by a change of variables,
Z
1
(ρk fα )(cA)x = ρk (z)z α c−1 R(c−1 z, A)x dz
2πi Γ
Z
1
= cα ρk (cz)z α R(z, A)x dz = cα (ρkn (c ·)fα )(A)x.
2πi ∂Σω
(15.6)

If x ∈ D(fα (A)), then (ρk (c ·)fα )(A)x ∈ R(A) (by the definition of D(fα (A)),
since ρk (c ·) is a regulariser for (fα , A)), and (15.6) implies that (ρk fα )(cA)x ∈
R(A) = R(cA). But this implies that x ∈ D(fα (cA)) (by the definition of
D(fα (cA)), since ρk is a regulariser for (fα , cA)). This gives the inclusion
D(fα (A)) ⊆ D(fα (cA)). The same argument in reverse direction gives the
inclusion D(fα (cA)) ⊆ D(fα (A)). Moreover, for any x in this common domain,

fα (cA)x = (ρk (cA))−1 (ρk fα )(cA)x,

cα fα (A)x = cα (ρk (c ·)(A))−1 (ρk (c ·)fα )(A)x = (ρk (c ·)(A))−1 (ρk fα )(cA)x,

the last identity being a consequence of (15.6). Since the right-hand sides are
obviously equal, this gives the result. 
Theorem 15.2.7. Let A be a sectorial operator on a Banach space X. If
0 < α < π/ω(A), then Aα is sectorial, we have

ω(Aα ) = αω(A),

and for all β ∈ C we have

(Aα )β = Aαβ with equality of domains.

If A is R-sectorial and 0 < |α| < π/ωR (A), then Aα is R-sectorial and

ωR (Aα ) = αωR (A).

 15.2 Fractional powers 441

Proof. The proof proceeds in a number of steps.

Step 1 – First consider an arbitrary α > 0. In this step we will prove that
for all µ 6∈ Σαω(A) we have µ ∈ %(Aα ) and

µR(µ, Aα ) = −|µ|1/α R(−|µ|1/α , A) + ψτ (|µ|−1/α A),

where τ = arg µ and

eiτ z + z α
ψτ (z) = .
(eiτ − z α )(1 + z)
Note that ψτ ∈ H 1 (Σσ ) for all σ < |τ |/α.
A straightforward calculation shows

µ |µ|1/α µz + |µ|1/α z α
− = = ψτ (|µ|−1/α z).
µ − zα |µ|1/α + z (µ − z α )(|µ|1/α + z)
Hence
1 1  |µ|1/α −1/α

= + ψ τ (|µ| z) .
µ − zα µ |µ|1/α + z
1
Proposition 15.1.12 implies that ( µ−(·) α )(A) is indeed the inverse of (µ −
α α α
(·) )(A) = µ − A . Thus µ ∈ %(A ) and

1  |µ|1/α −1/α

R(µ, Aα ) = + ψ τ (|µ| z) (A)
µ |µ|1/α + z
(15.7)
1 
= − |µ|1/α R(−|µ|1/α , A) + ψτ (|µ|−1/α A)
µ
1
using that if λ ∈ {Σσ , then λ−· (A)x = R(λ, A)x, and observing that
−1/α
ψτ (|µ| A) is well defined and bounded by the Dunford calculus of A.
Step 2 – Now let 0 < α < π/ω(A). We will prove that the operator Aα is
sectorial, with ω(Aα ) 6 αω(A).
By Step 1, for τ > αω(A) we have µ ∈ %(A) if | arg µ| > τ . Furthermore,
for σ ∈ (ω(A), τ /α) have
Z
−1/α 1 dz
ψτ (|µ| A) = ψτ (|µ|−1/α z)
2πi ∂Σσ z
Z
1 dz
= ψτ (z) .
2πi ∂Σσ z

Hence we may estimate

|dz|
Z
−1/α Mσ,A
kψτ (|µ| A)k 6 |ψτ (z)| .
2π ∂Σσ |z|

Therefore by (15.7) the sectoriality of A implies the sectoriality of Aα with

ω(Aα ) 6 αω(A).
442 15 Extended calculi and powers of operators

Step 3 – Having proved that Aα is sectorial, the identity (Aα )β = Aαβ

follows from the composition rule of Theorem 15.1.15.
Since π/ω(A) > 1 we have 0 < 1/α < π/(αω(A)) 6 π/ω(Aα ). Hence
we may apply the inequality of the angles of sectoriality of Step 2 to Aα to
obtain ω(A) = ω((Aα )1/α ) 6 (1/α)ω(Aα ), the equality A = (Aα )1/α being a
consequence what we just proved. In combination with Step 2, this proves the
equality ω(Aα ) = αω(A).
Step 4 – Using Proposition 10.3.2, the final assertion is proved in the same
way. 
The next theorem shows that α 7→ kAα xk satisfies a useful log–convexity
property.

Theorem 15.2.8 (Interpolation estimate). Let A be a sectorial operator

on a Banach space X. Let α, β, γ ∈ C satisfy

0 < <α < <γ < <β or 0 = α < <γ < <β

and let θ ∈ (0, 1) be such that <γ = (1 − θ)<α + θ<β. Then

D(Aα ) ∩ D(Aβ ) ⊆ D(Aγ ),

and for all x ∈ D(Aα ) ∩ D(Aβ ) and ω(A) < σ < π we have

C
kAγ xk 6 kAα xk1−θ kAβ xkθ ,
θ(1 − θ)

where C is a constant depending only on <β − <α, σ, and A.

Proof. Let m be the smallest integer strictly greater than <β − <α. We will
use
R ∞ the auxiliary function ψ(z) = cz m (1 + z)−2m , where c is chosen so that
ds
0
ψ(s) s = 1. Then the functions
Z 1 Z ∞
ds ds
g(z) := ψ(sz) and h(z) := ψ(sz)
0 s 1 s

are well defined for all z ∈ C and satisfy

Z ∞ Z ∞
ds ds
g(z) + h(z) = ψ(sz) = ψ(s) = 1.
0 s 0 s

We claim that g, h ∈ E(Σσ ). Indeed, we have

1 1
|z|m
Z Z
ds ds
|g(z)| 6 |ψ(sz)| 6 Cσ,m |z|m sm = ,
0 s 0 s m
∞ 1
|z|−m
Z Z
ds ds
|h(z)| 6 |ψ(sz)| 6 Cσ,m |z|−m s−m = .
1 s 0 s m
 15.2 Fractional powers 443

It follows that g an h have integrable limits 0 at 0 and ∞ in the sense of

Lemma 15.1.2, respectively. From g = 1 − h and h = 1 − g we see that g an
h have integrable limits 1 at ∞ and 0, respectively. Therefore Lemma 15.1.2
implies the claim.
For all t > 0, it follows from the claim that

g(tA) + h(tA) = I (15.8)

in the primary calculus of the sectorial operator tA.

Now let x ∈ D(Aα ) ∩ D(Aβ ). Then x ∈ D(Aγ ) and (15.8) implies

Aγ x = g(tA)Aγ x + h(tA)Aγ x. (15.9)

h(z) = z γ−α h(z) belong to E(Σσ ); this

The functions ge(z) = z γ−β g(z) and e
follows from the choice of m and redoing the above computation for ge and eh.
We have
g(tA)Aγ x = ge(tA)(tA)β−γ Aγ x = tβ−γ ge(tA)Aβ x.
Here, the first identity can be justified by viewing ge(t ·) as a regulariser for
(z β−γ , tA) and noting that Aγ x ∈ D(Aβ−γ ); the second identity follows by
first applying Proposition 15.2.6 and then Theorem 15.2.5. Similarly we have

h(tA)Aγ x = tα−γ e
h(tA)Aα x.

From (15.9) it now follows that

h(tA)Aα x.
Aγ x = tβ−γ ge(tA)Aβ x + tα−γ e

Therefore,

kAγ xk 6 t<β−<γ ke
g (tA)kkAβ xk + t<α−<γ ke
h(tA)kkAα xk
6 C t<β−<γ kAβ xk + t<α−<γ kkAα xk ,

where the constant C only depends on <β − <α, σ, and A; we used that from
the definition of the primary calculus for it follows that supt>0 kf (tA)k 6 C <
∞ for f ∈ {e h}, using by (10.9) and the sectoriality of A.
g, e
Optimising the choice of t > 0, we arrive at the estimate
" #
 θ 1−θ  1 − θ θ
γ
kA xk 6 C + kAα xk1−θ kAβ xkθ .
1−θ θ

Since the term in the square brackets is bounded above by 1/(θ(1 − θ)), this
gives the second estimate. 
Remark 15.2.9. It is tempting to believe that
Z 1 Z ∞
ds ds
g(A)x = ψ(sA)x and h(A)x = ψ(sA)x ,
0 s 1 s
444 15 Extended calculi and powers of operators

but these integrals may fail to converges at 0 (the first) and ∞ (the second).
Calderón’s reproducing formula (Proposition 10.2.5) guarantees their conver-
gence (as improper integrals) for elements x ∈ D(A) ∩ R(A) if z 7→ ψ(z)
belongs to H 1 (Σσ ), and for x ∈ D(A) ∩ R(A) if z 7→ ψ(z) log z belongs to
∈ H 1 (Σσ ). The above proof does not depend on these matters; all we needed
there were bounds on the operators g(A) and h(A) that follow directly from
the definitions of these operators through the extended Dunford calculus.

Corollary 15.2.10. Let A be a sectorial operator on a Banach space X with

0 ∈ %(A). Then for all <α > 0 the operator A−α is bounded. Moreover, for
0 < <α < n we have
CMσ,A
kA−α k 6 <α
kA−1 k<α ,
n (1− <α
n )

where C is a universal constant.

Proof. Let 0 < <α < n. By Theorem 15.2.5 and 15.2.8, applied with θ =
1 − <α/n, for all x ∈ X we have

CMσ,A
kA−α xk = kAn−α (A−n x)k 6 <α
kA−n xk<α/n kxk1−<α/n
n (1− <α
n )
CMσ,A
6 <α <α
kA−1 k<α kxk,
n (1 − n )

where C is a universal constant. It follows that A−α is bounded and satisfies

the bound in the statement of the corollary. 

Proposition 15.2.11. Let A be a sectorial operator on a Banach space X;

when considering Aα for <α 6 0 we assume A to be injective. If A has a
bounded H ∞ -calculus and 0 < |α| < π/ωH ∞ (A), then Aα has a bounded H ∞ -
calculus and ωH ∞ (Aα ) = α ωH ∞ (A).

Proof. This follows directly from the identity f (A)x = g(Aα )x for x ∈ D(A) ∩
R(A) and f ∈ H ∞ (Σσ ), with g ∈ H ∞ (Σ|α|σ ) given by f (z) = g(z α ). 
If A is sectorial, then A + ε is sectorial and boundedly invertible. We conclude
this section a some useful result that applies in this situation.

Proposition 15.2.12. Let A be a sectorial operator on a Banach space X. If

D(A) ∩ R(A) is dense in X, then for all α > 0 and ε > 0 we have D(Aα ) =
D((ε + A)α ) with equivalent graph norms.

Proof. The result is clear for α = 1, 2, . . . . Next let α ∈ (0, 1). The functions

(ε + z)α ε + zα
f (z) := − 1, g(z) = −1
ε + zα (ε + z)α
 15.2 Fractional powers 445

belong to H 1 (Σσ ) for all 0 < σ < π. For x ∈ D(Ak ) ∩ R(Ak ) with k large
enough, Proposition 15.1.12 gives

f (A)x = (ε + A)α (ε + Aα )−1 x − x, g(A)x = (ε + Aα )(ε + A)−α )x − x.

Since f (A) and g(A) are bounded, these identities imply D(Aα ) = D((ε+A)α ).
The equivalence of the norms follows from the open mapping theorem.
If β = α + n with n ∈ N and α ∈ (0, 1) then D((ε + A)β ) ⊆ D((ε + A)n )
by Theorem 15.2.5. Thus we obtain

D((ε + A)β ) = D((ε + A)n (ε + A)α )

= {x ∈ D((ε + A)n ) : (ε + A)α x ∈ D((ε + A)n )}
= {x ∈ D(An ) : (ε + A)α x ∈ D(An )}
= {x ∈ D(An ) : Aα x ∈ D(An )}
= D(An Aα ) = D(Aβ ).

Equivalence of norms now follows easily. 

15.2.b Representation formulas

The aim of this section is to prove various integral representations for the
fractional powers of sectorial operators.

Theorem 15.2.13 (Balakrishnan). Let A be a sectorial operator on a Ba-

nach space X and let ω(A) < σ < π. For all 0 < <α < 1 and x ∈ D(A) we
have
sin πα ∞ α−1
Z Z
1
Aα x = z α−1 R(z, A)Ax dz = t (t + A)−1 Ax dt.
2πi ∂Σσ π 0

If in addition A is densely defined and ω(A) < 21 π, then for all x ∈ D(A) we
have
Z ∞
1
α
A x= s−α S(t)Ax dt,
Γ (1 − α) 0

where (S(t))t>0 is the bounded analytic C0 -semigroup generated by −A.

Note that limz↓0 R(z, A)Ax = 0 for x ∈ D(A) by Proposition 10.1.7, so the first
integral is absolutely convergent. By the same reasoning the second integral is
absolutely convergent. The absolute convergence of the third integral follows
near t = 0 from the fact that x ∈ D(A), and near t = ∞ from the bound
kS(t)Axk 6 Ct−1 kxk (see Theorem G.5.3).
Integrating by parts and using with the identity −αΓ (−α) = Γ (1 − α),
the third identity in Balakrishnan’s theorem may equivalently be presented
as
446 15 Extended calculi and powers of operators
Z ∞
1
α
A x= t−α−1 (S(t)x − x) dt, x ∈ D(A).
Γ (−α) 0

The absolute convergence of this integral follows from the bound kS(t)x−xk =
O(t) as t ↓ 0 for x ∈ D(A).
α
z
Proof. For all ε > 0 the function z 7→ z+ε belongs to H 1 (Σσ ) ∩ H ∞ (Σσ ) and
zα
therefore the operator ( z+ε )(A) can be defined by the Dunford calculus and
is bounded. Fix x ∈ D(A). Then x ∈ D(Aα ), and therefore by multiplicativity
of the extended Dunford calculus (Proposition 15.1.12),
 zα 
Aα x = (A)(ε + A)x.
z+ε
Similarly,
 zα   zα 
(A)x = (A)(I + A)x.
z+ε (z + ε)(z + 1)
Combining these identities, we compute
 zα   zα 
Aα x = ε (A)(I + A)x + (A)Ax
(z + ε)(z + 1) z+ε
zα
Z
ε
= R(z, A)(I + A)x dz
2πi ∂Σσ (z + ε)(z + 1)
zα
Z
1
+ R(z, A)Ax dz
2πi ∂Σσ z + ε
= (I) + (II).

Noting that z 7→ z α−1 R(z, A)Ax is integrable along ∂Σσ , the term (I) tends
to 0 as ε ↓ 0 by dominated convergence. Also,
Z Z
1 z 1
(II) = z α−1 R(z, A)Ax dz → z α−1 R(z, A)Ax dz
2πi ∂Σσ z + ε 2πi ∂Σσ

as ε ↓ 0 by dominated convergence. This proves the first identity.

Turning to the second identity, write ∂Σσ = Γσ ∪ Γ−σ where Γ±σ =
{re±iσ ∈ C : r > 0}. It follows from Cauchy’s theorem that
Z
α 1
A x= z α−1 R(z, A)Ax dz
2πi ∂Σσ
Z
1
= z α−1 R(z, A)Ax dz
2πi Γσ
Z
1
+ z α−1 R(z, A)Ax dz
2πi Γ−σ
Z ∞
1
=− (reiσ )α−1 R(reiσ , A)Ax eiσ dr
2πi 0
 15.2 Fractional powers 447
Z ∞
1
+ (re−iσ )α−1 R(re−iσ , A)Ax e−iσ dr
2πi 0
Z ∞
1
→ rα−1 (e−iπ(α−1) − eiπ(α−1) )R(−r, A)Ax dr (as σ → π)
2πi 0
sin απ ∞ α−1
Z
= r (r + A)−1 Ax dr.
π 0

The minus sign in the third identity comes from the fact that Γσ is downwards
oriented. The convergence is a consequence of the dominated convergence
theorem.
To prove the third formula we use the identity just proved together with
the Laplace transform representation of the resolvent (Proposition G.4.1) to
get

sin πα ∞ α−1 ∞ −ts

Z Z
α
A x= t e S(s)Ax ds dt
π
Z 0 ∞ Z ∞ 0
sin πα 
= tα−1 e−ts dt S(s)Ax ds
π 0
Z ∞0
1
= s−α S(s)Ax ds,
Γ (1 − α) 0
sin πα 1
where we used the identity π = Γ (1−α)Γ (α) . 

From this theorem it is rather easy to re-deduce a special case of Theorem

15.2.8 as follows. Let 0 < α < 1. Let M > 0 be such that k(t + A)−1 k 6 M/t
for all t > 0. By Theorem 15.2.13, for all x ∈ D(A) we have
Z ∞
sin πα
α
kA xk 6 tα−1 k(t + A)−1 Axk dt
π 0
Z ρ
sin πα
6 tα−1 k(t + A)−1 Akkxk dt
π 0
Z ∞
sin πα
+ tα−1 k(t + A)−1 kkAxk dt
π ρ
sin πα sin πα
6 (1 + M )ρα kxk + M ρα−1 kAxk
πα π(1 − α)

with absolute convergence of all integrals. Up to this point we have assumed

that x ∈ D(A). The estimate extends to general x ∈ D(Aα ) by approximation
as in the proof of that theorem. The estimate of Theorem 15.2.8 is obtained
by optimising over ρ as in the proof of the theorem.

Corollary 15.2.14. Let A be a sectorial operator on a Banach space X and

let ω(A) < σ < π. Let 0 < α < 1 and λ ∈ {Σσ .
(1) The operator Aα R(λ, A) is bounded and
448 15 Extended calculi and powers of operators

kAα R(λ, A)k 6 Cα Mσ,A (Mσ,A + 1)|λ|α−1 ,

where Cα = sin(πα) 1
πα 1−α .
(2) If, in addition, A is densely defined and ω(A) < 21 π, and (S(t))t>0 denotes
the bounded analytic C0 -semigroup generated by −A, then for all t > 0 the
operator Aα S(t) is bounded and

kAα S(t)k 6 Cα MA t−α ,

R∞
where Cα = 1
Γ (1−α) 0
τ −α (1 + τ )−1 kxk dτ and MA = supt>0 tkAS(t)k.

From Theorem G.5.3 we recall that supt>0 tkAS(t)k < ∞.

Proof. For the first assertion, fix λ ∈ {Σσ . The boundedness of Aα R(λ, A)
is evident from the inclusion D(A) ⊆ D(Aα ). For all x ∈ D(A), by Theorem
15.2.13 we have
sin(πα) ∞ α−1
Z
Aα R(λ, A)x = t (t + A)−1 R(λ, A)Ax dt.
π 0

We split the integral on the right into two parts and estimate them separately.
First, writing A = (A + t) − t,
Z |λ| Z |λ|
tα−1 (t + A)−1 R(λ, A)Ax dt 6 tα−1 k[I − t(t + A)−1 ]R(λ, A)xk dt
0 0
Z |λ|
−1
6 |λ| tα−1 (1 + M )kλR(λ, A)xk dt
0
M (M + 1) α−1
6 |λ| kxk.
α
Similarly, but now writing A = (A − λ) + λ,
Z ∞ Z ∞
−1
t α−1
(t + A) R(λ, A)Ax dt 6 (1 + M )kxk tα−2 kt(t + A)−1 k dt
|λ| |λ|
M (M + 1) α−1
6 |λ| kxk.
1−α
Turning to the second assertion, by analyticity the operators S(t) map
X into D(A) and supt>0 tkAS(t)k < ∞. The boundedness of the operators
Aα S(t) follows from the boundedness of AS(t) and the inclusion D(A) ⊆
D(Aα ). To prove the estimate, note that for all x ∈ X we have
Z ∞
1
Aα S(t)x = s−α AS(t + s)x ds,
Γ (1 − α) 0

so, for t > 0,

 15.2 Fractional powers 449
Z ∞
C
kAα S(t)xk 6 s−α (t + s)−1 kxk ds
Γ (1 − α) 0
Z ∞
Ct−α
= τ −α (1 + τ )−1 kxk dτ.
Γ (1 − α) 0

As a corollary to Theorem 15.2.13 we have the following representation for-
mula for the negative fractional powers of A.
Corollary 15.2.15. Let A be an injective sectorial operator on a Banach
space X and let ω(A) < σ < π. For all 0 < <α < 1 and x ∈ R(A) we
have
sin πα ∞ −α
Z Z
−α 1 −α
A x= z R(z, A)x dz = t (t + A)−1 x dt.
2πi ∂Σσ π 0

If, in addition, A is densely defined and ω(A) < 21 π, and if (S(t))t>0 denotes
the bounded analytic C0 -semigroup generated by −A, then for all x ∈ R(A)
we have
Z ∞
−α 1
A x= t−α S(t)x dt.
Γ (α) 0
Note that if x = Ay with y ∈ D(A), then R(z, A)x = −y + zR(z, A)y, so the
first integral is absolutely convergent. In the same way it is checked that the
second integral is absolutely convergent. From kS(t)xk = kAS(t)yk = O(1/t)
as t → ∞ (by Theorem G.5.3) we see that the third integral is absolutely
convergent.
Proof. Writing x = Ay with y ∈ D(A) we have A−α x = A1−α y, and Theorem
15.2.13 gives
Z Z
1 1
A−α x = A1−α y = z −α R(z, A)Ay dz = z −α R(z, A)x dz.
2πi ∂Σσ 2πi ∂Σσ
The second identity is proved in the same way. The third follows from the
second by following the lines of the proof of Theorem 15.2.13. 
When A boundedly invertible, the identities in the corollary hold for arbitrary
x ∈ X. If in addition A is densely defined, the result extends to arbitrary
<α > 0 as follows:
Theorem 15.2.16. Let A be a densely defined sectorial operator on a Banach
space X with 0 ∈ %(A), and let ω(A) < σ < π. Then for all <α > 0 we have
Z
−α 1
A x= z −α R(z, A)x dz, x ∈ X,
2πi ∂(Σσ \Bε )

with Bε := {z ∈ C : |z| < ε}, where ε > 0 is so small that Bε ⊆ %(A).

450 15 Extended calculi and powers of operators

Proof. First let x ∈ D(Ak ) with k > <α and set y = ζ(A)−k x = (I +
A)2k A−k x, where ζ(z) = z/(z + 1)2 . The integral
Z
1
Tα x := z −α R(z, A)x dz
2πi ∂(Σσ \Bε )

is absolutely convergent and defines a bounded operator Tα . We may now

repeat the proof of the multiplicativity of the Dunford calculus (Theorem
10.2.2) to obtain, with ω(A) < ν < σ,
Z
1
Tα x = (Tα ◦ ζ k (A))y = Tα ◦ ζ(z)k R(z, A)y dz
2πi ∂(Σν \Bε/2 )
Z
1
= z −α ζ(z)k R(z, A)y dz
2πi ∂(Σν \Bε/2 )
Z
1
= z −α ζ(z)k R(z, A)y dz.
2πi ∂Σν

In the last step, the assumption k > |<λ| was used to justify the change of
contour by Cauchy’s theorem. By the definition of A−α x via the extended
Dunford calculus, the right hand side equals A−α x. This proves the first iden-
tity for x ∈ D(Ak ). Using the second part of Proposition 15.1.13, the general
case follows from it by approximation, noting that Tα is a bounded operator
on X. 

Theorem 15.2.17. Let −A be the generator of a bounded C0 -semigroup

(S(t))t>0 on X. Then A is densely defined and sectorial of angle ω(A) 6 21 π,
for all 0 < α < 1 the operator Aα is densely defined and sectorial of angle
ω(Aα ) 6 21 πα, and the bounded analytic C0 -semigroup generated by −Aα is
given by Z ∞
Sα (t)x = fα,t (s)S(s)x ds, t > 0, x ∈ X,
0
where, for t > 0,
Z c+i∞
1 α
fα,t (s) := esz−tz dz, s > 0,
2πi c−i∞

is a non-negative function which is independent of c > 0 and satisfies

Z ∞
fα,t (s) ds = 1.
0

Proof. By generalities from semigroup theorem (see Section G.2), the assump-
tions imply that A is densely defined and sectorial with ω(A) 6 21 π. By
Proposition 15.2.7, Aα is densely defined and sectorial of angle 21 πα and con-
sequently −Aα generates a bounded analytic C0 -semigroup by Theorem G.5.2.
 15.2 Fractional powers 451

By Example 15.1.6 we furthermore have Sα (t) = et (Aα ), where et (z) = e−tz .

Hence by the composition rule of Theorem 15.1.15 we have

Sα (t) = gα,t (A),

α
where gα,t (z) = e−tz .
Let 21 π < ν < σ < min{ 21 π/α, π}. By the Phillips calculus (Proposition
10.7.2(2)), Z ∞
Sα (t)(A)x = fα,t (s)S(s)x ds, x ∈ X,
0

where fα,t ∈ L1 (R+ ) is given (with Bε = {z ∈ C : |z| = ε}) by

Z
1 α
fα,t (s) = − esz−tz dz
2πi ∂(Σν \Bε )
Z Z c+i∞
1 sz−tz α 1 α
=− e dz = esz−tz dz
2πi ∂Σν 2πi c−i∞

for c > 0. The second and third identity follow from Cauchy’s formula, the
use of which is justified by noting that for z = reiσu with u > 0 we have
α
|esz−tz | = exp(sr cos σ − t<eα(ln r+iσ) )
= exp(sr cos σ − trα cos(ασ)),
α
from which it follows that z 7→ esz−tz is integrable along ∂Σν . In its stated
α
form, Proposition 10.7.2(2) requires gα,t = e−tz to be in H 1 (Σσ ), which is
not the case. The reader may check, however, that the proof still works in
the present situation if we replace integration over ∂Σν by integration over
∂(Σν \ Bε ). For λ > 0 we have
Z ∞ Z c+i∞ Z ∞
−λs 1 α
e fα,t (s) ds = e−λs esz−tz ds dz
0 2πi c−i∞ 0
Z c+i∞ −tzα (15.10)
1 e −tλα
=− dz = e .
2πi c−i∞ z − λ
Using
R∞ the non-negativity of fα,t , upon passing to the limit λ ↓ 0 gives
0
f α,t (s) ds = 1.
Finally, the fact that fα,t is non-negative follows from (15.10), the fact
α
that λ 7→ e−tλ is completely monotone and the Post–Widder real inversion
theorem for the Laplace transform. We refer the reader to the Notes for further
details. 
We finish with two examples.
Example 15.2.18 (Fractional derivatives). For 1 < p < ∞, the operator A =
d/dt with domain D(A) = {f ∈ W 1,p (0, T ; X) : f (0) = 0} is sectorial on
Lp (0, T ; X) of angle 21 π and for all <α > 0 and f ∈ Lp (0, T ; X) we have
452 15 Extended calculi and powers of operators
Z x
−α 1
A f (x) = (x − y)α−1 f (y) dy for almost all x ∈ R.
Γ (α) 0

The operators A−α are called the (Liouville) fractional derivatives. In partic-
ular, Z x
A−1 f (x) = f (y) dy
0
−1
The operator V := A is called the Volterra operator. These formulas are
special cases of Theorem 15.2.16 once we note that −A is the generator of the
C0 -semigroup on Lp (0, T ; X) given by
(
f (s − t), s ∈ [0, T ], s > t,
S(t)f (s) =
0, otherwise.

To see that the generator of this semigroup is indeed −A, let us denote the
generator by B for the moment. It is clear that Y := {f ∈ C 1 ([0, T ]; X) :
f (0) = 0} is contained in D(B) and Bf = −f 0 = −Af for all f ∈ Y . Since Y
is also invariant under the semigroup, Y is dense in D(B) by Lemma G.2.4.
But A is a closed operator and Y is also dense in D(A), and therefore B = −A
with equal domains.

Example 15.2.19 (Poisson semigroup). Let A be the Laplace operator on

Lp (Rd ; X), where 1 6 p < ∞ is fixed and X is a Banach space. This operator
has been introduced in Section 5.5 by declaring

D(A) := H 2,p (Rd ; X),

Af := ∆f, f ∈ D(A),

where H 2,p (Rd ; X) is the Banach space of all f ∈ Lp (Rd ; X) admitting a

weak Laplacian ∆f in Lp (Rd ; X) (see (5.44)). As was noted in Lemma 5.5.5,
Cc∞ (Rd ; X) is dense in D(A), and consequently A can be equivalently defined
as the closure of the operator f 7→ ∆f acting in Cc∞ (Rd ; X), where ∆f is
now defined in terms of the classical second order derivatives of f . For UMD
spaces X and exponents 1 < p < ∞, Proposition 5.5.4 shows that

H 2,p (Rd ; X) = W 2,p (Rd ; X),

and Theorem 5.6.11 establishes a Fourier analytic characterisation of these

spaces as the Banach space of all tempered distributions u ∈ S 0 (Rd ; X) such
that the tempered distribution
b
((1 + 4π 2 | · |2 )b
u)

belongs to Lp (Rd ; X).

Let us now return to the general situation where 1 6 p < ∞ and X is
a general Banach space. From this point on, we will simply write ∆ for the
 15.2 Fractional powers 453

Laplace operator in Lp (Rd ; X). As was shown in Example G.5.6, −∆ is the

generator of a C0 -semigroup of contractions (H(t))t>0 on Lp (Rd ; X), the heat
semigroup, given by H(0) = I and

H(t)f := kt ∗ f, t > 0,
2
where kt (x) = (4πt)−d/2 e−|x| /(4t) is the heat kernel. It was shown in the same
example that this semigroup extends analytically to {z ∈ C : <z > 0} by the
formula
H(z)f = kz ∗ f, <z > 0,
and that this extension is uniformly bounded and strongly continuous on
every sector Σω with 0 < ω < 21 π. As a consequence, −∆ is a densely defined
sectorial operator of angle ω(∆) = 0.
By Theorem 15.2.17, the operator (−∆)1/2 is densely defined and sectorial
of angle 0 and generates a bounded analytic C0 -semigroup (P (z))z∈Σω for
every 0 < ω < 21 π on Lp (Rd ; X), the so-called Poisson semigroup. By Theorem
15.1.7, in the primary calculus of (−∆)1/2 this semigroup is given by

P (z)f = exp(−z∆1/2 ), z ∈ Σω , f ∈ Lp (Rd ; X).

An explicit representation is obtained from Theorem 15.2.17, from which it

follows that
Z ∞
P (t)f = kt (s)H(s)x ds, t > 0, f ∈ Lp (Rd ; X),
0

where, for t > 0,

Z c+i∞
1 1/2
kt (s) := esz−tz dz, s > 0.
2πi c−i∞

is a non-negative function which is independent of c > 0 and satisfies

Z ∞
fα,t (s) ds = 1.
0

We wish to prove here that

P (t)f = pt ∗ f, t > 0,

where
Γ ( 21 (d + 1)) t
pt (x) = 1 1
π 2 (d+1) (t2 + |x|2 ) 2 (d+1)
is the Poisson kernel. For d = 1 it takes the simpler form
1 t
pt (x) = .
π t + x2
2
454 15 Extended calculi and powers of operators

By Theorem 15.2.5 we have ((−∆)1/2 )2 f = −∆f for f ∈ D(∆) and there-

fore, by the composition rule of Theorem 15.1.15,

exp(t(−∆)1/2 )f = φt (∆)f, f ∈ D(∆),

1/2
with φt (z) = e−tz . It follows from Proposition 15.1.13 that, for f ∈ D(∆),

P (t)f = φt (∆)f = lim φt (∆)ψn (∆)f,

n→∞

where
n
ψn (z) =
, n > 1.
n+z
The remainder of the proof will be devoted to proving the identity

φt (∆)φn (∆)f = pt ∗ ψn (∆)f. (15.11)

These functions are regularisers for (exp(−t ·), ∆). Once this has been shown
the identity
P (t)f = pt ∗ f, f ∈ Lp (Rd ; X),
follows from Proposition 10.1.7 by passing to the limit n → ∞ in (15.11).
Fixing f ∈ D(∆) ∩ R(∆) and t > 0. Below we will show that
∞ 2
te−t /4s −zs
Z
1/2
e−tz = e ds. (15.12)
0 2π 1/2 s3/2
Assuming this identity for the moment, by Fubini’s theorem and Example
15.1.6 we have
φt (∆)ψn (∆)f = (φt ψn )(∆)f
Z
1 1/2
= e−tz ψn (z)R(z, ∆)f dz
2πi ∂Σσ
Z ∞ 2
te−t /4s −zs
Z
1
= e ψn (z)R(z, ∆)f ds dz
2πi ∂Σσ 0 2π 1/2 s3/2 (15.13)
Z ∞ 2
te−t /4s 1
Z
= e−zs ψn (z)R(z, ∆)f dz ds
0 2π 1/2 s3/2 2πi ∂Σσ
Z ∞ 2
te−t /4s
= exp(−s∆)ψn (∆)f ds.
0 2π 1/2 s3/2
On the other hand,
Z ∞
(∗) 1 t 1 ds
pt (x) = 1 1 s− 2 (d+1) e−1/4s
(4π) 2 (d+1) (t2 + |x|2 ) 2 (d+1) 0 s
Z ∞
t 1 2
+|x|2 )/4s
= 1 s− 2 (d+3) e−(t ds
(4π) 2 (d+1) 0
 15.2 Fractional powers 455
∞ 2
te−t /4s
Z
= kt (x) ds,
0 2π 1/2 s3/2
2
where ks (x) = (4πs)−d/2 e−|x| /4s denotes the heat kernel associated with ∆
and (∗) follows from
Z ∞ Z ∞
− 21 (d+1) −1/4s ds 1 1 du 1 1
s e =4 2 (d+1)
u 2 (d+1) e−u = 4 2 (d+1) Γ ( (d + 1)).
0 s 0 u 2

Now Fubini’s theorem implies

pt ∗ ψn (∆)f
Z ∞Z 2 ∞
te−t /4s
Z
1
= ks (· − y) ψn (z)R(z, ∆)f (y) ds dz dy
−∞ 0 (4π)1/2 s3/2 2πi ∂Σσ
Z ∞ 2 Z ∞
te−t /4s 1
Z
= ψn (z) ks (· − y)R(z, ∆)f (y) dy dz ds
0 2π 1/2 s3/2 2πi ∂Σσ −∞
Z ∞ 2
te−t /4s 1
Z
= ψn (z) exp(−s∆)R(z, ∆)f dz ds
0 2π 1/2 s3/2 2πi ∂Σσ
Z ∞ 2
te−t /4s
= exp(−s∆)ψn (∆)f ds.
0 2π 1/2 s3/2
(15.14)

The identity (15.11) is obtained by combining (15.13) and (15.14).

It remains to prove (15.12). First, the substitution u = c/t gives
Z ∞ Z ∞
−( ct −t)2 c −( c −u)2
e dt = 2
e u du.
0 0 u

Renaming the second integration variable and adding the two formulas, the
substitution s = uc − u gives
Z ∞
1 ∞ 1 ∞ −s2
Z Z
c −( c −u)2 c  −( c −u)2 1
2
e u du = 1 + 2
e u du = e ds = π 1/2 .
0 u 2 0 u 2 −∞ 2

We will apply this identity with c = 21 tz 1/2 . Completing squares and changing
variables twice, we obtain
∞ 2 ∞
te−t /4s −zs t −(t/2√s−z1/2 √s)2 ds
Z Z
1/2
etz e ds = e √
0 s3/2 s 2 s
Z0 ∞
t −(t/2u−z1/2 u)2
= e du = π 1/2 ,
0 u2

and this is the formula (15.12) we wanted to prove.

456 15 Extended calculi and powers of operators

15.3 Bounded imaginary powers

A special role is played by sectorial operators whose purely imaginary frac-
tional powers Ait are bounded. As their definition requires that D(A) ∩ R(A)
be dense it will be convenient to introduce the following terminology.

Definition 15.3.1 (Standard sectorial operators). A standard sectorial

operator is a sectorial operator A with the property that D(A) ∩ R(A) is dense
in X.

The following proposition recalls some results proved in Proposition 10.1.8.

Proposition 15.3.2. Let A be a sectorial operator on a Banach space X.

Then:
(1) if A is standard, then A is injective;
(2) A is standard if and only if it is densely defined and has dense range;
(3) if X is reflexive, the following assertions are equivalent:
(i) A is standard sectorial;
(ii) A is injective;
(iii) A has dense range.

In view of (1), the fractional powers Aα of a standard sectorial operator A

are well defined for all α ∈ C, and all results from the previous section are
applicable to A.
In applications, standardness is hardly a restrictive assumption. In most
situations the Banach space will be reflexive and even UMD, and in such
spaces for a sectorial operator A we have the direct sum decomposition

X = N(A) ⊕ R(A).

By (2), the part of A in R(A) is standard sectorial, and the part of A in N(A)
is identically zero.

Example 15.3.3 (Standardness of the Laplacian on Lp (Rd ; X)). Let us con-

sider the Laplace operator ∆ on Lp (Rd ; X), where 1 < p < ∞ and X is a
UMD space, with domain D(∆) = H 2,p (R; X). It is shown in Example 10.1.5
that −∆ is sectorial of angle 0 on Lp (Rd ; X) for all 1 6 p < 1, and standard
sectorial if and only if 1 < p < ∞.

For standard sectorial operators A on a Banach space X, the next diagram

summarises the main results of this section.
 15.3 Bounded imaginary powers 457

(1)
γ-bounded bounded
H ∞ -calculus H ∞ -calculus
(3)
(2) (4)

γ-BIP with (5)

BIP with
ωγ−BIP (A) < π ωBIP (A) < π

(6) (7)
(8)
almost
γ-sectorial
γ-sectorial

The implications (1), (2), (4), (5), and (8) are trivial. The implication (3)
follows from Theorem 15.3.21, where it is also shown that equivalence holds
when X has Pisier’s contraction property. The implications (1)–(5) are equiv-
alence when X is a Hilbert space. The implication (6) follows from Theorem
15.3.19, and the implication (7) is Theorem 15.3.16.

15.3.a Definition and basic properties

For t ∈ R consider the function

ft (z) := z it := exp(it log z),

where we use the branch of the logarithm that is holomorphic on C \ (−∞, 0].
From |ft (z)| = exp(−t arg(z)) it follows that ft ∈ H ∞ (Σσ ) for each 0 < σ < π
and
kft kH ∞ (Σσ ) 6 exp(σ|t|).
Thus if A is a standard sectorial operator with a bounded H ∞ (Σσ )-calculus,
the operators
Ait := ft (A)
are well defined as bounded operators on X. Some examples of operators with
bounded imaginary powers will be discussed in Subsection 15.3.h.
When A is merely standard sectorial, we may use the extended Dunford
calculus to define the operators Ait , t ∈ R, as closed and densely defined
operators in X. This suggests the following definition.
Definition 15.3.4 (BIP). A linear operator A acting in a Banach space X
is said to have bounded imaginary powers (briefly, A has bounded imaginary
powers) if A is standard sectorial and the operators Ait are bounded for all
t ∈ R.
Examples of operators with bounded imaginary powers will be given in Section
15.3.h.
458 15 Extended calculi and powers of operators

Proposition 15.3.5. If A has bounded imaginary powers, then the family

(Ait )t∈R is a C0 -group.

Proof. It is evident from the definition through the extended Dunford calculus
that t 7→ Ait x is strongly measurable for all x ∈ X. We have already seen that
Ai0 x = 1(A)x = x for all x ∈ D(A)∩R(A), so Ai0 = I. The identity Ais Ait x =
Ai(s+t) x follows from Proposition 15.1.12. Proposition G.2.7 implies that t 7→
Ait x is continuous for all x ∈ X. 
When A has bounded imaginary powers, then by the above result and the
general theory of C0 -(semi)groups, there exist constants M > 1 and ω ∈ R
such that
kAit k 6 M eω|t| .
This allows us to define the abscissa

ωBIP (A) := inf ω ∈ R : sup e−ω|t| kAit k < ∞ .


t∈R

We have the following improvement of Corollary 15.2.10 in the presence

of bounded imaginary powers. The point of the estimate in part (1) is that
boundedness of the imaginary powers permits us to obtain an estimate that
is uniform all the way up the imaginary axis.

Proposition 15.3.6. If A has bounded imaginary powers and 0 ∈ %(A), and

if kAit k 6 M e−ω|t| for all t ∈ R, then:
(1) the operator A−z is bounded for every <z > 0, and

kA−z k 6 CA M eω|=z| kA−1 k<z , <z > 0,

where CA depends only on MA := supt>0 (1 + t)k(t + A)−1 k and kA−1 k;

(2) for all <z1 > 0 and <z2 > 0 we have A−z1 A−z2 = A−(z1 +z2 ) ;
(3) for all x ∈ X the mapping z 7→ A−z x is continuous on {<z > 0} and
holomorphic on {<z > 0}.

Proof. (1) and (2): By assumption for all t ∈ R the operators Ait are bounded,
and for <z > 0 the operators A−z are bounded by Corollary 15.2.10. For <z1 >
0 and <z2 > 0 the identity A−z1 A−z2 = A−(z1 +z2 ) follows from Proposition
15.1.12, noting that all operators occurring in this identity are bounded.
We next prove the norm estimate. We begin by noting that
0
CA := sup kA−s k < ∞
s∈[0,1]

by Corollary 15.2.15, with a constant CA depending only on the constant MA

(which is finite since A is boundedly invertible).
By writing z = s + it with s ∈ [0, 1], it follows from the identity A−z =
−s −it
A A that
 15.3 Bounded imaginary powers 459
0
sup kA−z k 6 CA sup kA−it k 6 CA
0
M e−ω|t| 6 CA M e−ω|t| kA−1 k<z ,
06<z61 t∈R

0
where CA = CA / max{1, kAk−1 }. This gives the desired bound in (1) for
0 6 <z 6 1.
For z = z 0 + n with n > 1 and 0 6 <z 0 < 1, the estimate in (1) now follows
from
0 0 0
kA−z k = kA−z −n k 6 kA−z kkA−n k 6 CA M −ω|t| kA−1 k<z kA−1 kn
= CA M −ω|t| kA−1 k<z .

(3): Fix an arbitrary integer k > 1 and fix an element x ∈ D(Ak ) ∩ R(Ak ).
We have already seen that z 7→ A−z x is holomorphic on {|<z| < k}; in
particular z 7→ A−z x is continuous on {0 6 <z < k}. The holomorphy on
{|<z| < k} and continuity on {0 6 <z < k} of z 7→ A−z x for general x ∈ X
follows by approximation xn → x with xn ∈ D(Ak ) ∩ R(Ak ), noting that the
above norm estimate implies that the convergence A−z xn → A−z x is locally
uniform on {0 6 <z < k}. 

15.3.b Identification of fractional domain spaces

An important justification for studying boundedness of imaginary powers

comes from Theorem 15.3.9 below, which states that boundedness of the
imaginary powers implies the coincidence of the fractional power scale and
the complex interpolation scale. For the proof of this result we need some
lemmas. The first extends the relation Aα Aβ = Aα+β , which has been proved
so for only for α, β satisfying <α · <β > 0.

Lemma 15.3.7. If A has bounded imaginary powers, then for all α ∈ C and
t ∈ R we have
Aα Ait = Ait Aα = Aα+it
with equality of domains.

Proof. Since Ait is bounded it is clear that D(Aα ) = D(Ait Aα ). From Propo-
sition 15.1.12(2) we already know the inclusion D(Ait Aα ) ⊆ D(Aα+it ) with
Ait Aα x = Aα+it x for all x ∈ D(Ait Aα ), as well as the equality D(Ait Aα ) =
D(Aα+it )∩D(Aα ). Combining these results, we obtain Ait Aα = Aα with equal
domains. 
The second lemma considers bounded imaginary powers for shifted operators:

Lemma 15.3.8. If A has bounded imaginary powers, then A + ε has bounded

imaginary powers for all ε > 0. If kAit k 6 M eω|t| and ω(A) < σ < π, then

k(A + ε)it k 6 M 0 e(ω∨σ)|t| ,

for some constant M 0 independent of ε > 0.

460 15 Extended calculi and powers of operators

Proof. It is immediate from Theorem 15.1.18 applied to ε−1 A that A + ε has

bounded imaginary powers. By Proposition 15.2.6 we have (ε−1 A)it = ε−it Ait
and (A + ε)it = εit (ε−1 A + I)it with equal domains in both cases. Hence, by
the estimates provided by Theorem 15.1.18 and Proposition 15.2.6, for any
fixed ω(A) < σ < π we have

k(A + ε)it k = k(ε−1 A + I)it k

6 (1 + Mσ,ε−1 A )2 (kAit k + Cσ kz 7→ z it kH ∞ (Σσ ) )
6 (1 + Mσ,A )2 (M eω|t| + Cσ eσ|t| ).

Theorem 15.3.9 (Fractional powers through complex interpolation).

If A has bounded imaginary powers, then for all α > 0 and 0 < θ < 1,

D(Aαθ ) = [X, D(Aα )]θ with equivalent norms.

Proof. By Proposition 15.2.12 and Lemma 15.3.8 we may replace A by A + I

if necessary, and thereby assume without loss of generality that 0 ∈ %(A).
This allows us to use the results of Proposition 15.3.6.
Choose M > 1 and ω ∈ R such that kAit k 6 M eω|t| for all t ∈ R. We
begin by proving the inclusion D(Aαθ ) ⊆ [X, D(Aα )]θ . Fix 0 < θ < 1 and
x ∈ D(Aαθ ), and put
2
f (z) := e(z−θ) A−αz Aαθ x, z ∈ S,

where S = {z ∈ C : 0 < <z < 1} is the unit strip in the complex plane.
Then f is holomorphic as an X-valued function on S and satisfies f (θ) = x.
Moreover, by Proposition 15.3.6, f is continuous and uniformly bounded on S.
Using the notation introduced in Appendix C, to prove that x ∈ [X, D(Aα )]θ
we must check that f ∈ H (X, D(Aα )). For this it remains to be checked that
t 7→ f (it) belongs to Cb (R; X) and t 7→ f (1+it) belongs to Cb (R; D(Aα )). The
former follows from what has already been said, and for the latter we write
kf (1 + it)kD(Aα ) = kf (1 + it)k + kAα f (1 + it)k. Again by what has already
been said, the function t 7→ f (1 + it) belongs to Cb (R; X). The second term
can be estimated as follows:
2
kAα f (1 + it)k = ke(1+it−θ) Aα A−α(1+it) Aαθ xk
2 2
−t2
= ke(1+it−θ) A−iαt Aαθ xk 6 e(1−θ) M eαω|t| kAαθ xk,

and this is a bounded function of t ∈ R. Here we used Lemma 15.3.7, which

implies that D(Aα ) = D(A−α(1+it) ) and A−α(1+it) y = A−α Ait y for y ∈ X.
To prove the reverse inclusion [X, D(Aα )]θ ⊆ D(Aαθ ) we will use the results
and notation of Appendix C. Fix x ∈ [X, D(Aα )]θ and let f ∈ H (X, D(Aα ))
satisfy f (θ) = x. By Corollary C.2.8 there is a sequence of functions fn ∈
H0 (X, D(Aα ); D(Aα )) such that fn (θ) =: xn → x in [X, D(Aα )]θ .
 15.3 Bounded imaginary powers 461

Since D(Aα ) ⊆ D(Aαz ) for z ∈ S and fn takes values in D(Aα ) we may

define 2
gn (z) := e(z−θ) Aαz fn (z), z ∈ S.
With respect to the norm of X, each function gn is bounded on S. By the
three lines lemma,
n o
kxn k = kfn (θ)k 6 max sup kfn (it)k, sup kfn (1 + it)k ,
t∈R t∈R
n o
αθ
kA xn k = kgn (θ)k 6 max sup kgn (it)k, sup kgn (1 + it)k .
t∈R t∈R

Moreover, for all t ∈ R,

2
−t2 2
−t2
kgn (it)k 6 eθ kAiαt fn (it)k 6 eθ M eαω|t| kfn (it)k,
2
−t2
kgn (1 + it)k 6 e(1−θ) kAiαt Aα fn (1 + it)k
2
−t2
6 eθ M eαω|t| kfn (1 + it)kD(Aα ) .

Here we used Lemma 15.3.7, which implies that D(Aα ) = D(Aα+iαt ) and
Aα+iαt y = Aiαt Aα y for y ∈ D(Aα ).
It follows from these estimates that kxn k . kfn kH (X,X) 6 kfn kH (X,D(Aα ))
and kAαθ xn k . kfn kH (X,D(Aα )) , and therefore kxn kD(Aαθ ) . kfn kH (X,D(Aα )) .
Replacing xn by xn − xm in the above argument, we find that the se-
quence (xn )n>1 is Cauchy in D(Aαθ ) and therefore converges to a limit. Since
xn → x in X, this limit must be x. This proves that x ∈ D(Aαθ ) and that
kxkD(Aαθ ) . kf kH (X,D(Aα )) . Taking the infimum with respect to f it follows
that kxkD(Aαθ ) . kxk[X,D(Aα )]θ . 

This theorem self-improves in an obvious manner. Upon replacing X by D(Aβ )

and using that D(Aγ ) = D(Aγ+it ) we arrive at the following more general
result.

Corollary 15.3.10. If A has bounded imaginary powers, then for all α, β ∈ C

with 0 6 α < β < ∞ we have

D(A(1−θ)α+θβ ) = [D(Aα ), D(Aβ )]θ

with equivalent norms.

Let us revisit the Laplace operator ∆ on Lp (Rd ; X), where 1 < p < ∞ and
X is a UMD space, with domain D(∆) = H 2,p (R; X). It was already noted
above that −∆ is standard sectorial of angle 0 on Lp (Rd ; X) for all 1 6 p <
1, and by Theorem 10.2.25 it has a bounded H ∞ -calculus of angle 0. As a
consequence, −∆ has bounded imaginary powers. Applying Theorem 15.3.9,
for all 0 < θ < 1 we obtain

D((−∆)θ ) = [Lp (Rd ; X), H 2,p (Rd ; X)]θ with equivalent norms.
462 15 Extended calculi and powers of operators

In Chapter 5 we have proved Seeley’s theorem (Theorem 5.6.9), from which

it follows that if X is a UMD space and 1 < p < ∞, then for all 0 < θ < 1 we
have

[Lp (Rd ; X), H 2,p (Rd ; X)]θ = H 2θ,p (Rd ; X) with equivalent norms.

Thus we obtain the following result.

Theorem 15.3.11 (Laplacian on Lp (Rd ; X)). Consider the Laplace oper-

ator ∆ on Lp (Rd ; X), where 1 < p < ∞ and X is a UMD space, with domain
D(∆) = H 2,p (R; X). Then for all 0 < θ < 1 we have

D((−∆)θ ) = H 2θ,p (Rd ; X) with equivalent norms.

15.3.c Connections with sectoriality

It is part of the definition that an operator with bounded imaginary powers

is standard sectorial, but there is no obvious a priori relation between the
abscissa ωBIP (A) and the angle of sectoriality ω(A). The main result of this
section is the following result, which says that ω(A) 6 ωBIP (A). Moreover, if
X is a UMD space, then A is R-sectorial of angle ωR (A) 6 ωBIP (A).

Theorem 15.3.12 (Clément–Prüss). Let A be an operator with bounded

imaginary powers on a Banach space X, and assume that ωBIP (A) < π.
(1) A is sectorial of angle ω(A) 6 ωBIP (A).
(2) If X is a UMD space, then A is R-sectorial of angle ωR (A) 6 ωBIP (A).

The key lemma is the following representation formula. It expresses the re-
solvent of A in terms of the imaginary powers Ait , and a such it provides the
key insight behind the Clément–Prüss theorem.

Lemma 15.3.13 (Prüss–Sohr). Let A be an operator with bounded imagi-

nary powers on a Banach space X, and assume that ωBIP (A) < π. Let λ = reiθ
with r > 0 and |θ| < π − ωBIP (A). Then for all x ∈ D(A) ∩ R(A) we have
Z ∞
1 1 π
(I + λA)−1 x = x + p.v. λ−it A−it x dt,
2 2πi −∞ sinh(πt)

the convergence of the principal value integral on the right-hand side being
part of the assertion. Furthermore, for all 0 < s < 1,
Z
1 1
λs As (1 + λA)−1 x = λit Ait x dt. (15.15)
2 R sin(π(s − it))

Proof. We begin with the proof of the first identity. It proceeds in three steps.
Step 1 – First take r = 1 and θ = 0. In this step, for all x ∈ X we will
prove that
 15.3 Bounded imaginary powers 463
Z ∞
1 1 π
x+ p.v. A−is x ds
2 2πi −∞ sinh(πs)
Z c+i∞
1 π
= lim A−z x dz,
c↓0 2πi c−i∞ sin(πz)

the convergence of the principal value integral being part of the assertion. Note
that the integrals occurring on right-hand side converge absolutely thanks to
the estimates
| sinh(π(c + it))| = O(eπ|t| ) as t → ±∞
and
kA−c−it xk 6 M eω|t| kA−c xk, t ∈ R,
for all ωBIP (A) < ω < π, with M > 1 a constant depending on ω.
By Cauchy’s theorem we have
Z c+i∞ Z
1 π 1 π
A−z x dz = A−z x dz,
2πi c−i∞ sin(πz) 2πi Γc sin(πz)

where Γc is the (upwards oriented) contour consisting of the union of the two
(1) (3)
half-lines Γc = {is : s 6 −c} and Γc = {is : s > c} and the semi-circle
(2) iϑ 1 1
Γc = {ce : ϑ ∈ [− 2 π, 2 π]}. As c ↓ 0, the contributions along the two
half-lines converge to the principal value integral and the contribution along
the semi-circle converges to 21 x. The latter follows by noting that A−z x → x
as z → 0 in the closed right-half plane, by the continuity of z 7→ A−z x on
that set (see Proposition 15.3.6). Hence
1
2π πceiϕ
Z Z
1 π 1 1
lim dz = lim iϕ
dϕ =
c↓0 2πi (2)
Γc sin(πz) c↓0 2π − 12 π sin(πce ) 2

(since sin(πceiϕ ) = πceiϕ + O(c3 ) as c ↓ 0.

Step 2 – In this step we will prove the lemma for r = 1 and θ = 0 with
x ∈ D(A) ∩ R(A), i.e., we show that
Z ∞
1 1 π
(I + A)−1 x = x + p.v. A−is x ds
2 2πi −∞ sinh(πs)

for all x ∈ D(A)∩R(A). (Note that I +A is boundedly invertible as part of the

definition of bounded imaginary powers, since A is assumed to be standard
sectorial).
Let y := (I + A)x. Then
464 15 Extended calculi and powers of operators
Z ∞
1 1 π
y+ p.v. A−is y ds
2 2πi −∞ sinh(πs)
Z ∞
1 1 π
= (I + A)x + p.v. A−is x ds
2 2πi −∞ sinh(πs)
Z ∞
1 π
+ p.v. A1−is x ds (15.16)
2πi −∞ sinh(πs)
Z
1 1 π
= (I + A)x + p.v. A−z y dx
2 2πi Γc sin(πz)
Z
1 π
+ p.v. A1−z y dx.
2πi Γc sin(πz)

In view of sin(πz) = − sin(π(1 − z)), after a change of variable in the last

integral the contributions over all four half-lines cancel and we are left with
Z Z
1 1 π −z 1 π
(I + A)x + p.v. A x dz − p.v. A−z x dz,
2 2πi (2)
Γc sin(πz) 2πi Γ
ec(2) sin(πz)
(2)
where Γec = {1 − ceiϑ : ϑ ∈ [− 21 π, 21 π]}. As c ↓ 0, the first integral tends
to 21 y and the second to − 21 Ax. In the limit c ↓ 0 the three terms on the
right-hand side of (15.16) therefore add up to x. This proves the identity
Z ∞
1 1 π
x= y+ p.v. A−is y ds.
2 2πi −∞ sinh(πs)

Multiplying on both sides with (I + A)−1 gives the desired result.

Step 3 – The general case follows by applying the result of Step 2 to the
operator λA, which by Proposition 15.2.6 has bounded imaginary powers and
satisfies (λA)−is = λ−is A−is . This completes the proof of the first identity.
Using it, and fixing 0 < s < 1, for x ∈ D(A) ∩ R(A) we obtain
Z ∞
1 1 π
λs As (I + λA)−1 x = λs As x + p.v. λs−it As−it x dt
2 2πi −∞ sinh(πt)
1 ∞−is
Z
1 s s 1
= λ A x− λit Ait x dt
2 2 −∞−is sin(π(s − it))
1 ∞
Z
1
= λit Ait x dt
2 −∞ sin(π(s − it))

by the Cauchy theorem, noting that As Ait = As+it by Theorem 15.2.5 in the
first step. This gives the second identity. 

Remark 15.3.14. A more direct proof of the second identity can be given as
follows. Starting from the identity
Z ∞
e2πitξ 2e2πsξ
dt = , 0 < s < 1, ξ ∈ R, (15.17)
−∞ sin(π(s − it)) 1 + e2πξ
 15.3 Bounded imaginary powers 465

the substitution z = e2πξ gives

Z ∞
z it 2z s
dt = , 0 < s < 1, z ∈ R+ . (15.18)
−∞ sin(π(s − it)) 1+z

By analytic continuation this extends to all z ∈ C with | arg(z)| < π.

Let λ ∈ C \ {0} as in the statement of the lemma. For x ∈ D(A) ∩ R(A) it
follows from Proposition 15.1.19 that Ait x is given by the Bochner integral
Z
1  1 
Ait x = µit x + z it R(z, A) − x dz,
2πi Γν z−µ

where ω(A) < | arg µ| < ν. Substituting this identity into the right-hand
side of (15.15), a short computation involving Fubini’s theorem, (15.18), and
Cauchy’s theorem gives the result. At the expense of some additional compu-
tations, instead of invoking Proposition 15.1.19 one may also directly use the
definition for Ait x as given in Definition 15.1.8.

Proof of Theorem 15.3.12. (1): First let λ = reiθ with r > 0 and |θ| <
π − ω(A). By subtraction we obtain the identity
Z ∞
−1 −1 1 π
(I + λA) x = (I + A) x + p.v. (λ−is − 1)A−is x ds
2πi −∞ sinh(πs)

for x ∈ D(A2 ) ∩ R(A2 ). The crux is that the term λ−is − 1 is of the order
O(|s|) near s = 0 and can therefore be estimated as |λ−is − 1| . |s| ∧ 1.
Similarly, | sinh(s)| . (|s| ∧ 1)eπ|s| . Therefore the principal value integral is
actually absolutely convergent and bounded in x. As a consequence of this,
the identity extends to arbitrary x ∈ X.
The proof is completed by observing that the integral in the right hand
side of the identity
Z ∞
1 π
(I + λA)−1 x = (I + A)−1 x + (λ−is − 1)A−is x ds
2πi −∞ sinh(πs)

is absolutely convergent for any λ = reiθ with r > 0 and |θ| < π − ωBIP (A).
Indeed, recalling the estimates for λ−is − 1 and sinh(s) mentioned earlier,
choosing ωBIP (A) < ω < π so that |θ| < π − ω we estimate
Z ∞ Z
π −is −is
(λ − 1)A x ds . πe−π|s| Mω eω|s| kxk ds
−∞ sinh(πs) |s|>1

with a constant independent of x. The right-hand side defines a holomorphic

extension of the function λ 7→ (I +λA)−1 x to the open sector Σπ−ωBIP (A) . As a
consequence the spectrum of A must be contained in the closure of ΣωBIP (A) .
Finally, the sectoriality estimate on the complement of this closure follows
from the estimate.
466 15 Extended calculi and powers of operators

(2): Fix ωBIP (A) < ω < ν < π and choose numbers λn = rn eiθn with
rn > 0 and |θn | < π − ν, as well as vectors xn ∈ X; n = 1, . . . , N . We wish
to show that there exists a constant C, independent of the choices just made,
such that
N
X N
X
εn (I + λn A)−1 xn 6C εn x n ,
L2 (Ω;X) L2 (Ω;X)
n=1 n=1

N
where (εn )n=1 is a Rademacher sequence defined on a probability space (Ω, P).
By a simple approximation argument, there is no loss of generality in assuming
that xn ∈ D(A) ∩ R(A) for all n = 1 . . . , N .
Since ω(A) 6 ωBIP (A) by the Clément–Prüss theorem, Lemma 15.3.13
(with λ = 1), the representation formulas of Lemma 15.3.13 hold for λ = reiθ
with r > 0 and |θ| < π − ν, with x ∈ D(A2 ) ∩ R(A2 ), and
Z ∞
−1 1 1
iθ
(I + re A) x = x + ψθ (s)r−is A−is x ds
2 2πi −∞
Z ∞
1
+ η(s)r−is A−is x ds,
2πi −∞
Z 1
1 ds
+ p.v. r−is A−is x
2πi −1 s
1
=: x + Tr,θ x + Sr x + Rr x,
2
where
π π 1(−1,1) (s)
ψθ (s) = (eθs − 1), η(s) := − .
sinh(πs) sinh(πs) s
Applying this to λ = λn we obtain
N
X
εn (I + λn A)−1 xn
L2 (Ω;X)
n=1
N N
1 X X
6 εn x n + εn Trn ,θn xn
2 n=1 L2 (Ω;X)
n=1
L2 (Ω;X)

N
X N
X
+ ε n Sr n x n + εn R rn x n .
L2 (Ω;X) L2 (Ω;X)
n=1 n=1

We will estimate the last three expressions separately.

To start with the first, we note that |ψθn (s)| . e(θn −π)|s| 6 e−ν|s| . There-
fore, by the Kahane contraction principle and the bound kAis k 6 M eω|s| ,
N Z ∞ N
X 1 X
εn Trn ,θn xn 6 A−is εn ψθn (s)xn ds
n=1
L2 (Ω;X) 2π −∞ n=1
L2 (Ω;X)
 15.3 Bounded imaginary powers 467
Z ∞ N
1 X
. M eω|s| εn ψθn (s)xn ds
2π −∞ n=1
L2 (Ω;X)

Z ∞ N
1 (ω−ν)|s|
X
. Me εn x n ds
2π −∞ n=1
L2 (Ω;X)

N
X
= CA,ν εn x n .
L2 (Ω;X)
n=1

The second term is treated similarly, now using that |η(s)| . e−π|s| :
N Z ∞ N
X 1 X
ε n Sr n x n 6 A−is εn η(s)xn ds
n=1
L2 (Ω;X) 2π −∞ n=1
L2 (Ω;X)

Z ∞ N
1 X
. M eω|s| εn η(s)xn ds
2π −∞ n=1
L2 (Ω;X)

Z ∞ N
1 X
. M e(ω−π)|s| εn x n ds
2π −∞ n=1
L2 (Ω;X)

N
X
0
= CA,ν εn x n .
L2 (Ω;X)
n=1

For estimating the third term we use the UMD property of X through the
boundedness of the Hilbert transform on L2 (R; X).
We begin with a preliminary observation. Let us set Un (s) = (rn A)−is =
−is −is
rn A for brevity. Then by the Kahane contraction principle, for all s ∈ R
we have
N
X N
X
εn Un (s)xn 6 εn A−is xn
L2 (Ω;X) L2 (Ω;X)
n=1 n=1
(15.19)
N
X
ω|s|
6 Me εn x n .
L2 (Ω;X)
n=1

Fix 0 < δ < 1 and t ∈ [− 12 , 21 ]. Then

N Z N Z
X ds X ds
εn Un (s)xn = εn Un (t) Un (s − t)xn
n=1 δ<|s|<1 s n=1 δ<|s|<1 s
N Z
X ds
= εn Un (t) ϕn (t − s)
n=1 |s|>δ s
N Z 1+t
X ds
− εn Un (s)xn
n=1 1 s
468 15 Extended calculi and powers of operators
N Z −1+t
X ds
+ εn Un (s)xn ,
n=1 −1 s

where ϕn (τ ) = 1(−1,1) (τ )Un (−τ )xn . Integrating over t ∈ (− 21 , 12 ), we obtain

N Z N Z 12 Z
X ds X ds
εn Un (s)xn = εn Un (t) ϕn (t − s) dt
n=1 δ<|s|<1 s n=1 − 12 |s|>δ s
N Z 12 Z 1+t
X ds
− εn Un (s)xn dt
n=1
1
−2 1 s
N Z 12 Z −1+t
X ds
+ εn Un (s)xn dt.
n=1
1
− 2 −1 s

Since X is UMD and φn ∈ L2 (R; X), the limit

Z Z
ds ds
lim ϕn (· − s) = lim ϕn (· − s)
δ↓0 |s|>δ s δ↓0 δ<|s|<R s
R→∞

exists in L2 (R; X) by Theorem 5.1.1 and equals πHφn , where H is the Hilbert
transform. As a result we obtain
N N Z 1
X X ds
ε n Rr n x n = εn p.v. Un (s)xn
n=1 n=1 −1 s
N Z
X ds
= εn lim Un (s)xn
n=1
δ↓0 δ<|s|<1 s
N Z 1
X 2
=π εn Un (t)Hϕn (t) dt
n=1 − 12
N Z 1 Z 1+t
X 2 ds
− εn Un (s)xn dt
n=1 − 12 1 s
N Z 1 Z −1+t
X 2 ds
+ εn Un (s)xn dt
n=1 − 12 −1 s
=: (I) + (II) + (III).

It remains to estimate the three terms on the right-hand side. For estimating
+ −
(I) we use that kHkL (L2 (R;X)) 6 2β2,X β2,X (see Theorem 5.1.13). Applying
the Kahane–Khintchine inequality, this gives
N
X Z 1/2
εn Un (t)Hϕn (t) dt
−1/2 L2 (Ω;X)
n=1
 15.3 Bounded imaginary powers 469
N
X Z 1/2
h εn Un (t)Hϕn (t) dt
−1/2 L1 (Ω;X)
n=1
Z 1/2 N
X
= εn Un (t)H[1(−1,1) (·)Un (− · )xn ](t) dt
−1/2 n=1 L1 (Ω;X)

Z 1/2 N
X
6 εn Un (t)H[1(−1,1) (·)Un (− · )xn ](t) dt
−1/2 L1 (Ω;X)
n=1
Z 1/2 N
X
6M εn H[1(−1,1) (·)Un (− · )xn ](t) dt
−1/2 L1 (Ω;X)
n=1
Z 1/2 h N
X i
= M1/2 E H 1(−1,1) (·) εn Un (− · )xn (t) dt
−1/2 n=1
h N
X i
6 M1/2 E H 1(−1,1) (·) εn Un (− · )xn
L2 (R;X)
n=1
N
X
+ −
6 2β2,X β2,X M1/2 E εn 1(−1,1) (·)Un (− · )xn
L2 (R;X)
n=1

and, by (15.19),
N
X
E εn 1(−1,1) (·)Un (− · )xn .
L2 (R;X)
n=1
N
X
=E εn (·)Un (− · )xn
L2 (−1,1;X)
n=1
N
X
6 εn (·)Un (− · )xn
L2 (Ω;L2 (−1,1;X))
n=1
N
X
= εn (·)Un (− · )xn
L2 (−1,1;L2 (Ω;X))
n=1
N
X
6 M1/2 εn x n
L2 (−1,1;L2 (Ω;X))
n=1
N
X
= M1/2 εn x n
L2 (Ω;X))
n=1

where M1/2 := sup|t|61/2 kA−it k.

To estimate (II) we use (15.19) again:
N Z 1/2 Z 1+t
X ds
εn Un (s)xn dt
n=1 −1/2 1 s L2 (Ω;X)
470 15 Extended calculi and powers of operators
Z 1/2 Z 1+t N
X ds
6 εn Un (s)xn dt
−1/2 1 n=1
L2 (Ω;X) s
Z 1/2 Z 1+t N
X ds
6 M2 εn x n dt
−1/2 1 n=1
L2 (Ω;X) s
N
X
6 M2 εn x n
L2 (Ω;X)
n=1

where M2 := sup|t|62 kA−it k.

The estimation of (III) is entirely similar. 

15.3.d Connections with almost γ-sectoriality

We have consistently limited our treatment of the H ∞ -calculus and related

topics to sectorial operators. It is of some interest to consider the wider class
of so-called almost sectorial operators, defined as follows.

Definition 15.3.15 (Almost sectorial operators). Let σ ∈ (0, π). A linear

operator A acting in a Banach space X is called:
(i) σ-almost sectorial if σ(A) ⊆ Σσ and the set

λAR(λ, A)2 : λ ∈ C \ Σ σ


is uniformly bounded;
(ii) σ-almost γ-sectorial if σ(A) ⊆ Σσ and the set

λAR(λ, A)2 : λ ∈ C \ Σ σ


is γ-bounded.
The operator A is called almost sectorial, respectively almost γ-sectorial if it
is σ-almost sectorial, respectively σ-almost γ-sectorial, for some σ ∈ (0, π).

Almost R-sectorial operators are defined similarly, replacing γ-boundedness

by R-boundedness.
For an almost sectorial, respectively an almost γ-sectorial operator A, we
define

e (A) := inf σ ∈ (0, π) : A is σ-almost sectorial ,
ω

ωeγ (A) := inf σ ∈ (0, π) : A is σ-almost γ-sectorial .

The identity
λAR(λ, A)2 = [λR(λ, A)]2 − λR(λ, A)
shows that every (γ-)sectorial operator is almost (γ-)sectorial and
 15.3 Bounded imaginary powers 471

e (A) 6 ω(A), respectively ω

ω eγ (A) 6 ωγ (A).

The above definitions may appear somewhat ad hoc at first sight, but the
motivation to introduce them is as follows. The operators λR(λ, A) used in
the definition of sectoriality can be represented in the primary calculus of A
as
λ
λR(λ, A) = rλ (A) with Rλ (z) = .
λ−z
Indeed, the functions rλ belong to the class E(Σσ ) introduced in Section
15.1.a as long as 0 < σ < |<λ|. They do not belong to H 1 (Σσ ), however, and
this fact is responsible for some of the technical issues encountered in several
proofs. In contrast, the operators λAR(λ, A)2 used in the definition of almost
sectoriality can be represented in the Dunford calculus of A, for we have
λz
λAR(λ, A)2 = reλ (A) with reλ (z) = .
(λ − z)2

Indeed, the functions reλ belong to H 1 (Σσ ) for 0 < σ < |<λ|. Further motiva-
tion will be given in the Notes at the end of the chapter.
The following result gives a version of the (second part of) Clément–Prüss
theorem (Theorem 15.3.12) holds without making any assumptions on the Ba-
nach space X. The price to pay is that only almost γ-sectoriality is obtained:

Theorem 15.3.16. Let A be an operator with bounded imaginary powers on

a Banach space X. Then A is almost γ-sectorial of angle ω
eγ (A) 6 ωBIP (A).

Proof. Fix ωBIP (A) < θ0 < θ < π and suppose that λ1 , . . . , λn ∈ C are non-
zero and satisfy | arg(λk )| > θ. Note that | arg(µk )| 6 π − θ. Set µk := −1/λk .
Then for all choices x1 , . . . , xn ∈ X we have, by Lemma 15.3.13,
n
1/2
X
E γk λk A1/2 R(λk , A)xk
k=1
n
1/2
X
=E γk µk A1/2 (1 + µk A)−1 xk
k=1
Z n
1 1 X
6 E γk µit it
k A xk dt
2 R | sin(π( 12 − it))| k=1
(∗)
Z (π−(θ−θ 0 ))|t| n
1 e −(π−(θ−θ 0 ))|t| it

it
 X
6 e A sup |µ k | E γk xk dt
2 R | sin(π( 12 − it))| 16k6n
k=1
0 n
(∗∗) e(π−(θ−θ ))|t|
Z
1 −θ 0 |t| it
X
6 dt sup e A E γ k xk
2 R | sin(π( 12 − it))| t∈R k=1
n
X
= CE γ k xk ,
k=1
472 15 Extended calculi and powers of operators

where in (∗) we used the contraction principle and in (∗∗) the fact that for
| arg(µ)| 6 π − θ and t ∈ R we have
0 0 0
e−(π−(θ−θ ))|t| Ait |µit | = e−(π−(θ−θ ))|t| e− arg(µ)t Ait 6 e−θ |t| Ait 6 C 0 ,
0
where C 0 := supt∈R ke−θ |t| Ait k is finite since ωBIP (A) < θ0 , and where
0
C0 e(π−(θ−θ ))|t|
Z
C := dt.
2 R | sin(π( 12 − it))|

We have shown that the family

 1/2 1/2
λ A R(λ, A) : | arg(λ)| > θ

is γ-bounded. Taking squares, it follows that the family

λAR(λ, A)2 : | arg(λ)| > θ



is γ-bounded as well. Moreover we see that ωeγ (A) 6 θ. This being true for all
ωBIP (A) < θ < π, it follows that ω
eγ (A) 6 ωBIP (A). 

15.3.e Connections with γ-sectoriality

We start with a definition.

Definition 15.3.17 (γ-bounded imaginary powers). An operator A is

said to have γ-bounded imaginary powers (briefly, A has γ-BIP) if it has
bounded imaginary powers and the family

{Ait : |t| 6 1}

is γ-bounded.

If A has γ-bounded imaginary powers, the group property Ais Ait = Ai(s+t)
combined with Proposition 8.1.20 (or rather, the elementary bound in the
discussion preceding it) implies that set

{e−ω|t| Ait : t ∈ R}

is γ-bounded for large enough ω > 0. Thus it makes sense to define the abscissa

ωγ -BIP (A) := inf ω > 0 : {e−ω|t| Ait : t ∈ R} is γ-bounded .



Replacing γ-boundedness by R-boundedness, we may similarly introduce op-

erators A with R-BIP along with their abscissa ωR-BIP (A) Since finite co-
type implies equivalence of Rademacher sums and Gaussian sums (Corollary
7.2.10), an operator A on a Banach space with finite cotype has R-bounded
imaginary powers if and only A has γ-bounded imaginary powers. As the
 15.3 Bounded imaginary powers 473

ensuing proofs will make clear, operators with γ-bounded imaginary powers
can be effectively studied using the continuous square functions introduced in
Section 10.4.b. It is for this reason that our results will be stated for operators
with γ-bounded imaginary powers. The analogous results for operators with
R-bounded imaginary powers automatically follow if the underlying Banach
space has finite cotype.

Proposition 15.3.18. If A has γ-bounded imaginary powers, then ωBIP (A) =

ωγ-BIP (A).

Proof. Let ωBIP (A) < ν < θ. For each n ∈ Z the singleton {Ain } is γ-bounded,
with γ-bound γ({Ain } = kAin k 6 M eν|n| , where M is a constant independent
of n ∈ Z. By Proposition 8.1.20 (with p = 1 and q = ∞), the set

{e−θ|n| Ain : n ∈ Z}

is γ-bounded. Combined with the fact that {Ais : s ∈ [−1, 1]} is γ-bounded,
by Proposition 8.1.19(3) we obtain that ωγ -BIP (A) < θ. 
We have seen in Theorem 15.3.12 that bounded imaginary powers imply sec-
toriality with angle ω(A) 6 ωBIP (A). The next theorem provides the analogue
for γ-bounded imaginary powers.

Theorem 15.3.19. If A has γ-bounded imaginary powers with ωγ -BIP < π,

then A is γ-sectorial with ωγ (A) 6 ωγ -BIP (A).

Proof. The proof proceeds in three steps.

Step 1 – In Steps 2 and 3 we will prove that each of the families of operators
1
Γs := {ts As (1 + tA)−1 : t > 0}, where 0 < s < ,
2
is γ-bounded, uniformly with respect to the parameter s ∈ (0, 21 ). In the
present step we show how the theorem follows from this.
For x ∈ D(A) ∩ R(A) we have

lim ts As (I + tA)−1 x = (I + tA)−1 x.

s↓0

Hence by Fatou’s lemma, for all finite sequences x1 , . . . , xn ∈ D(A) ∩ R(A)

and t1 , . . . , tn > 0 we have
n
X 2
E γk (I + tk A)−1 xk
k=1
n
X 2 n
X 2
6 lim inf E γk tsk As (I + tk A)−1 xk 6 CE γ k xk ,
s↓0
k=1 k=1
474 15 Extended calculi and powers of operators

where C is any finite upper bound for the γ-bounds of the families Γs , s ∈
(0, 21 ). This proves that the set {(I + tA)−1 : t > 0} is γ-bounded.
Applying this reasoning to operators eiθ with 0 < |θ| < π − ωγ-BIP (and
noting that the identity (e±iθ A)it = e∓θ Ait implies that these operators still
have γ-bounded imaginary powers) and using Proposition 8.5.8 to extrapolate
γ-boundedness from the boundary of a sector to the full sector, it follows that
A is γ-sectorial and ωγ (A) 6 ωγ -BIP (A).
Step 2 – We now turn to the proof of the γ-boundedness of the families Γs
uniformly with respect to s ∈ (0, 12 ). We claim that it suffices to prove that
for all f ∈ S (R; X) we have
Z
t 7→ ks (t − u)Ai(t−u) f (u) du 6 Ckf kγ(R;X) , (15.20)
R γ(R;X)

1
where the constant C is independent of 0 < s < 2 and
1
ks (t) := , t ∈ R.
2 sin(π(s − it))
Indeed, suppose that (15.20) has been proved. By Fubini’s theorem and the
second identity of Lemma 15.3.13, for all ξ ∈ R we have
Z Z
ks (t − u)Ai(t−u) f (u) e−2πitξ du dt
R R
Z Z
= ks (t)Ait e−2πitξ dt f (u)e−2πiuξ du = e−2πξs As (1 + e−2πξ A)−1 fb(ξ).
R R

Hence by (15.20) and the fact, observed in Example 9.6.5, that the Fourier
transform extends to an isometry on γ(R; X), we obtain

kξ 7→ e−2πsξ As (1 + e−2πξ A)−1 fb(ξ)kγ(R;X) 6 Ckf kγ(R;X) = Ckfbkγ(R;X) .

Since the Fourier transform maps S (R; X) onto itself and this space is dense
in γ(R; X), this estimate extends to all strongly measurable function g : S →
X representing an element of γ(R; X) by density. Then converse to the γ-
multiplier theorem (Proposition 9.5.6) implies that Γs is γ-bounded, with
γ-bound at most C.
Step 3 – To complete the proof of the theorem it remains to prove the
bound (15.20) with a uniform constant C independent of s ∈ (0, 21 ). We start
with the observation that by (15.18) we have

e−sξ
kbs (ξ) = ,
1 + e−ξ

which implies that kbs ∈ L∞ (R) uniformly in s ∈ (0, 21 ).

Fix s ∈ (0, 12 ). For n ∈ Z, set In := [2n − 1, 2n + 1) and define, for
ϕ ∈ Cc (R),
 15.3 Bounded imaginary powers 475
Z
Ts(n) ϕ(u) := Ks(n) (u, v)ϕ(v) dv, u ∈ R,
R

where
X
Ks(n) (u, v) := ks (u − v)1Ij (u)1Ij+n (v).
j∈Z

This sum trivially converges pointwise in (t, v), since each such point is con-
tained in at most one rectangle Ij × Ij+n . We wish to show that the opera-
(n)
tor Ts thus defined extends to a bounded operator on L2 (R), uniformly in
s ∈ (0, 21 ).
For ϕ ∈ Cc (R) we have, by the disjointness of the intervals Ij , monotone
convergence, and a change of variables,
Z X Z 2
kTs(n) ϕk22 = 1Ij (u) ks (u − v)1Ij+n (v)ϕ(v) dv du
R j∈Z R
Z X Z 2
= 1Ij (u) ks (u − v)1Ij+n (v)ϕ(v) dv du
R j∈Z R
XZ Z 2
6 ks (u − v)1Ij+n (v)ϕ(v) dv du
j∈Z R R
X X
2 2
= kkbs 1\
Ij+n ϕk2 6 k1\
Ij+n ϕk2
j∈Z j∈Z
XZ
= |ϕ(u)| du = kϕk22 .
2

j∈Z Ij+n

(n)
This shows that Ts extends to a bounded operator on L2 (R). Moreover,
since

|Ks(n) (u, v)| 6 |ks (u − v)|1{|u−v|>2(|n|−1)}

and
1
|ks (u)| 6 . e−π|u| , |u| > 1,
2| sinh(πu)|
by Young’s inequality we have
Z
(n) (n)
kTs k 6 kKs k1 . e−π|u| du . e−2π|n| , |n| > 2.
{|u|>2(|n|−1)}

(n)
By the γ-extension theorem (Theorem 9.6.1), the operators Ts extend to
bounded operators on γ(R; X) and

kTs(n) kL (γ(R:X)) 6 C0 e−2π|n| .

476 15 Extended calculi and powers of operators

for some absolute constant C0 > 0.

Define p : R → Z by p(t) := 2j when t ∈ Ij . Then |p(t) − t| 6 1 for all
t ∈ R. Let ωγ-BIP (A) < θ < π and let C1 , C2 > 0 be such that

γ{Ais : s ∈ [−1, 1]} 6 C1 , γ{Ais : s ∈ R} 6 C2 eθ|s| .

We may of course relate these constants, but that would only complicate the
estimate below a bit. Fix a Schwartz function f ∈ S (R; X) and an integer
(n)
n ∈ Z. If (u, t) belongs to the support of Ks , then u ∈ Ij and v ∈ Ij+n for
some j ∈ Z, from which it follows that p(u) = p(v) − 2n. Therefore we may
estimate
Z
u 7→ Ks(n) (u, v)Ai(u−v) f (v) dv
R γ(R;X)
Z
= u 7→ Ks(n) (u, v)Ai(u−p(u)+p(v)−v−2n) f (u) du
R γ(R;X)
Z
6 C1 u 7→ Ks(n) (u, v)Ai(p(v)−v−2n) f (v) du
R γ(R;X)
−2π|n| i(p(u)−u−2n)
6 C0 C1 e u 7→ A f (u) γ(R;X)

6 C0 C12 C2 e−2(π−θ)|n| kf kγ(R;X)

using the γ-multiplier theorem (Theorem 9.5.1) in the second and fourth step.
Since
X
ks (u − v) = Ks(n) (u, v), u, v ∈ R,
n∈Z

the bound (15.20) now follows from the triangle inequality. 

15.3.f Connections with boundedness of the H ∞ -calculus

It has already been observed that standard sectorial operators with a bounded
H ∞ -calculus have bounded imaginary powers and ωBIP (A) 6 ωH ∞ (A), the
angle of the H ∞ -calculus of A (see Definition 10.2.10). The following theorem
gives a more precise version of this result.

Theorem 15.3.20 (Cowling–Doust–McIntosh–Yagi). If A is a standard

sectorial operator with a bounded H ∞ (Σσ )-calculus for some ω(A) < σ < π,
then A has bounded imaginary powers and

ωBIP (A) = ωH ∞ (A).

Moreover,
∞ σ|t|
kAit k 6 Mσ,A e , t ∈ R,
∞
where Mσ,A is the boundedness constant of the H ∞ (Σσ )-calculus of A.
 15.3 Bounded imaginary powers 477

Proof. It remains to prove the inequality ωH ∞ (A) 6 ωBIP (A). In view of

the Clément–Prüss theorem (Theorem 15.3.12), which asserts that ω(A) 6
ωBIP (A), it suffices to prove that if ω(A) < µ < ν 6 σ with kAit k 6 M eµ|t|
for all t ∈ R, then A has a bounded H ∞ (Σν )-calculus.
To this end let f ∈ H ∞ (Σν ). We will show that
X
f (z) = z ik fk (z), z ∈ Σν , (15.21)
k∈Z

for suitable functions fk ∈ H ∞ (Σσ ) satisfying

X
eµ|k| kfk kH ∞ (Σσ ) 6 Ckf kH ∞ (Σν ) (15.22)
k∈Z

with constant C > 0 independent of f . Once this has been shown, we may set
X
f (A) := Aik fk (A),
k∈Z

with convergence in the norm of L (X); here, the operators fk (A) are defined
through the H ∞ (Σσ )-calculus of A. The bound (15.22) implies that

kf (A)k 6 CM kf k∞ . (15.23)

To complete the proof that A admits a bounded H ∞ (Σν )-calculus, we will

show that for f ∈ H 1 (Σν ) ∩ H ∞ (Σν ) the operator f (A) thus defined agrees
with the Dunford calculus of A.
Step 1 – In this step we prove everything up to and including (15.23).
Using the change of variables z = ew we transform sectors to horizontal strips
and must show that every g ∈ H ∞ (Sν ) can be expressed as
X
g(w) = eikw gk (w), w ∈ Sν ,
k∈Z

where Sθ = {z ∈ C : |=(z)| < θ} and the functions gk ∈ H ∞ (Sσ ) satisfy

X
eµ|k| kgk kH ∞ (Sσ ) 6 CkgkH ∞ (Sν ) .
k∈Z

Let φ ∈ Cc (R) satisfy

(i) 0P6 φ(x) ∈ 1(−1,1) (ξ) for all ξ ∈ R,
(ii) k∈Z φ(ξ − k) = 1 for all ξ ∈ R,

and set Z b
gk (w) := φ(w − t)g(t)e−ikt dt, w ∈ Sσ .
R b
By the Paley–Wiener theorem, φ is an entire function with sufficient decay
to ensure the convergence of the integral for every w ∈ C. Fixing w ∈ Sσ
478 15 Extended calculi and powers of operators

and k ∈ Z, and sing Cauchy’s theorem to shift the path of integration, for
 ∈ {−1, 1} we may write
Z b
gk (w) := φ(w − t − iν)g(t + iν)e−ik(t+iν) dt, w ∈ Sσ .
R

Taking  = − sgn(k) gives the bound

kgk k∞ 6 Cσ,ν e−ν|k| kgk∞ , k ∈ Z,

with Z b
Cσ,ν = sup |φ(x + iy)| dx < ∞.
|y|<σ+ν R

Setting hk (ξ) := φ(ξ − k)b

g (ξ), a simple calculation gives
c Z b
hk (w) = φ(w − t)g(t)e−ik(w−t) dt = eikw gk (w), w ∈ Sσ .
R
P
Since k∈Z hk (ξ) = gb(ξ) for all ξ ∈ R, the result follows by taking inverse
Fourier transforms.
Step 2 – It remains to show that for f ∈ H 1 (Σν ) ∩ H ∞ (Σν ) the operator
f (A) defined by (15.21) agrees with the Dunford calculus. For this it suffices
to observe that for such functions f , the functions gk constructed in Step 1
belong to H 1 (Sσ ) ∩ H ∞ (Sσ ) and

kgk kH 1 (Sσ ) 6 Cσ,ν e−ν|k| kgkH 1 (Sν ) , k ∈ Z,

with Cσ,ν as before. It follows that the sum defining f (A) also converges in
H 1 (Σν ). The required consistency now follows by interchanging summation
and integration, along with the fact that (z 7→ z ik fk (z))(A) = Aik fk (A) in the
extended Dunford calculus, hence a posteriori also in the Dunford calculus. 
With Theorem 15.3.19 at our disposal we will now investigate the connection
between the γ-boundedness of the imaginary powers Ait and the boundedness
of the H ∞ -calculus of A. In preparation of the next result, it is useful to point
out that in some of these results in Chapter 10 the finite cotype assumption
can be dropped if one defines discrete square functions in terms of Gaussian
sums instead of using Rademacher sums. To be explicit, assuming Definition
10.4.1 to have been modified in this way, the finite cotype assumption can be
dropped in the following results:
• Proposition 10.4.15(2). Indeed, the proof uses the finite cotype assumption
only to pass from Gaussian sums to the Rademacher sums used in the
definition of discrete square functions.
• Theorem 10.4.16(1). Indeed, the finite cotype assumption was only used
to apply Proposition 10.4.15(2).
 15.3 Bounded imaginary powers 479

• Proposition 10.4.20. Indeed, the finite cotype assumption was only used
to apply Proposition 10.4.15(2).
The next theorem establishes the connection between γ-bounded imaginary
powers and boundedness of the H ∞ -calculus.

Theorem 15.3.21 (Bounded H ∞ -calculus ⇔ γ-BIP). Let A be stan-

dard sectorial on a Banach space X.
(1) If A has γ-bounded imaginary powers with ωγ-BIP (A) < π, then A has a
bounded H ∞ -calculus and

ωH ∞ (A) 6 ωγ-BIP (A).

(2) If A has a bounded H ∞ -calculus and X has Pisier’s contraction principle,

then A has γ-bounded imaginary powers and

ωγ-BIP (A) 6 ωH ∞ (A).

Since Pisier’s contraction principle implies finite cotype (Corollary 7.5.13),

a sectorial operator A acting in a Banach space with this property has γ-
bounded imaginary powers if and only A has R-bounded imaginary powers,
and in that case
ωγ -BIP (A) = ωR-BIP (A).
Before turning to the proof of the theorem, we isolate a lemma which
is essentially contained in the proof of Theorem 10.4.16. For the reader’s
convenience we repeat the argument here.

Lemma 15.3.22. Let A be standard sectorial, let ω(A) < σ < π, and suppose
that there is a constant C > 0 such that for all ψ ∈ H 1 (Σσ ) and x∗ ∈
D(A∗ ) ∩ R(A∗ ) we have

t 7→ ψ(tA∗ )x∗ γ(R+ , dt ∗ 6 CkψkH 1 (Σσ ) kx∗ k.

t ;X )

Then for all non-zero φ ∈ H 1 (Σσ ) there is a constant cφ > 0 such that for all
x ∈ D(A) ∩ R(A) we have

kxk 6 2Ccφ Mσ,A t 7→ φ(tA)x γ(R+ , dt

.
t ;X)

Note that the assumptions on x, x∗ , φ, and ψ imply that t 7→ φ(tA)x ∈

∗ ∗ ∗
γ(R+ , dt dt
t ; X) and t 7→ ψ(tA )x ∈ γ(R+ , t ; X ) by Lemma 10.4.14 (which
only assumes sectoriality and can therefore be applied to both A and A∗ ).
Proof.
R∞ Fix a non-zero φ ∈ H 1 (Σσ ) and fix an arbitrary ψ ∈ H 1 (Σσ ) such that
dt ∗ ∗ ∗
0
φ(t)ψ(t) t = 1. For all x ∈ D(A) ∩ R(A) and x ∈ D(A ) ∩ R(A ), from
the reproducing formula of Proposition 10.2.5, the trace duality inequality of
Theorem 9.2.14, and our assumption we obtain
480 15 Extended calculi and powers of operators
Z ∞
∗ dt
|hx, x i| = hφ(tA)ψ(tA)x, x∗ i
0 t
6 t 7→ φ(tA)x γ(R+ , dt ;X) t 7→ ψ(tA)∗ x∗ γ(R+ , dt ∗
t t ;X )
∗
6 CkψkH 1 (Σσ ) t 7→ φ(tA)x γ(R+ , dt
kx k,
t ;X)

where we used that ψ(tA)∗ = ψ(tA∗ ). Taking the supremum over all x∗ ∈
D(A∗ ) ∩ R(A∗ ) of normR 6 1, the result now follows from Lemma 10.2.19, with
∞
cφ = inf{kψkH 1 (Σσ ) : 0 φ(t)ψ(t) dtt = 1}. 
Proof of Theorem 15.3.21. (1): Fix ωγ-BIP (A) < σ < π. Then the set
{e−σ|t| Ait : t ∈ R} is γ-bounded.
Step 1 – In this step we prove that for all ϑ > σ and x ∈ X the function
t 7→ e−ϑ|t| Ait x belongs to γ(R; X).
By the result of Example 9.4.12 (taking H = C), the function t 7→
e−(ϑ−σ)|t| ⊗ x belongs to γ(R; X) and
1
t 7→ e−(ϑ−σ)|t| ⊗ x γ(R;X)
= kt 7→ e−(ϑ−σ)|t| kL2 (R) kxk h kxk.
(ϑ − σ)1/2

Hence by the γ-multiplier theorem (Theorem 9.5.1), t 7→ e−ϑ|t| Ait x belongs

to γ∞ (R; X) and
1
t 7→ e−ϑ|t| Ait x γ∞ (R;X)
. γ({e−σ|t| Ait : t ∈ R}). (15.24)
(ϑ − σ)1/2

We claim that the functions t 7→ e−ϑ|t| Ait x actually belong to the closed
subspace γ(R; X) of γ∞ (R; X). To prove this, let B be the generator of the C0 -
group (Ait )t∈R . For all x ∈ D(B) and all 0 < a < b < ∞ and −∞ < a < b < 0
the function t 7→ e−ϑ|t| Ait x belongs to C 1 ([a, b]; X), and hence to γ(a, b; X)
by Proposition 9.7.1. Since D(B) is dense in X, the claim now follows from
Corollary 9.5.2.
Step 2 – The formula

cosh(πt) ∞ − 1 +it
Z
1
a− 2 +it = u 2 (u + a)−1 du, a > 0, t ∈ R, (15.25)
π 0

may be proved by a contour integration argument. Alternatively, it can be

obtained from a standard identity for the Mellin transform of the function
(1 + t)−1 and some substitutions.
Set θ := π − ϑ. By analytic continuation, the identity (15.25) extends to
complex a ∈ C \ (−∞, 0]. For z ∈ Σπ−θ we may substitute a = e−iθ z to
obtain, after a bit of algebra,

cosh(πt) 1 iθ ∞ − 1 +it 1/2 iθ

Z
θt it
e z = e 2 u 2 z (e u + z)−1 du.
π 0
 15.3 Bounded imaginary powers 481

Since ω(A) < π − θ (this is because ω(A) 6 ωBIP (A) = ωγ-BIP (A) by the
Clément–Prüss theorem and Proposition 15.3.18, and ωγ-BIP (A) < ϑ = π − θ
by assumption), we can apply Lemma 10.2.17 (with p = 1) to this identity
and obtain, for all x ∈ X,

cosh(πt) ∞ 1 iθ − 1 +it 1/2 iθ

Z
θt it
e A x= e 2 u 2 A (e u + A)−1 x du
π
Z0 (15.26)
cosh(πt) ∞ itv 1 v+ 1 iθ 1/2 iθ v
= e e 2 2 A (e e + A)−1 x dv,
π −∞

where the second identity results from the substitution u = ev .

By Step 1, the function t 7→ e−ϑ|t| Ait x belongs to γ(R; X). Since cosh(πt) ∼
πt
e and π − θ = ϑ, this implies that the function

e−θ|t| Ait x
t 7→
cosh(πt)

belongs to γ(R; X).

By Theorem 9.6.1, the γ-extension of Fourier–Plancherel transform is an
isometry from γ(R; X) onto itself. Dividing both sides of (15.26) by cosh(πt)
and applying this isometry, it follows that the function
1
v 7→ eπv+ 2 iθ A1/2 (eiθ e2πv + A)−1 x

belongs to γ(R; X) and

1 eθt
v 7→ eπv+ 2 iθ A1/2 (eiθ e2πv + A)−1 x γ(R;X)
h t 7→ Ait x
cosh(πt) γ(R;X)
−ϑ|t| it
h t 7→ e A x γ(R;X)
1
.A kxk,
ϑ−σ
using (15.24) in the last step. Substituting back ev = u and leaving out terms
of modulus one since they do not affect the γ-norms,
1
u 7→ u1/2 A1/2 (eiθ u + A)−1 x γ(R, du
.A kxk.
u ;X) ϑ−σ
The term in the norm on the left-hand side is of the form φ(u−1 A) with φ(z) =
z 1/2 (eiθ + z)−1 . This function belongs to H 1 (Σϑ0 ) for all 0 < ϑ0 < ϑ = π − θ,
and the estimate can be interpreted as giving the square function estimate

t 7→ φ(tA)x γ(R+ , dt
. kxk, x ∈ X.
t ;X)

Note that up to this point we only have used that A is sectorial and has
bounded imaginary powers (the γ-sectoriality assumption will only be used
482 15 Extended calculi and powers of operators

towards the end of the proof). Because of this, we can apply the same reason-
ing to the part A of A∗ in X := D(A∗ ). Indeed, this operator is sectorial and
has bounded imaginary powers on X and (A )it x∗ = (Ait )∗ x∗ for x∗ ∈ X ;
we leave the easy verification as an exercise to the reader. Together with the
identity φ(tA)∗ x∗ = φ(tA )x∗ , which is equally easy to verify, this gives the
dual square function estimate

t 7→ φ(tA)∗ x∗ γ(R+ , dt ∗ . kx∗ k, x∗ ∈ X = D(A∗ ).

t ;X )

Hence by Lemma 15.3.22,

kxk . t 7→ φ(tA)x γ(R+ , dt

, x ∈ D(A) ∩ R(A),
t ;X)

with an implied constant independent of x. We may now apply Theorem

10.4.19 (noting that thanks to Theorem 15.3.19 we have ωγ (A) 6 ωγ -BIP (A))
to conclude that A has a bounded H ∞ (Σϑ0 )-calculus for all ωγ-BIP (A) < ϑ0 <
θ. This completes the proof.
(2): Let A have a bounded H ∞ (Σσ )-calculus for some ω(A) < σ < π,
and let ϑ > σ. Recalling the bound |z it | 6 e|t|| arg(z)| , the R-boundedness (and
hence the γ-boundedness, as the Pisier contraction property implies finite
cotype) of the set {e−θ|t| Ait } follows from Theorem 10.3.4(3). This shows
that A has γ-bounded imaginary powers and ωγ-BIP (A) 6 ϑ. 

15.3.g The Hilbert space case

The last main result of this chapter is the following characterisation of sectorial
operators on Hilbert spaces with bounded imaginary powers.

Theorem 15.3.23. For any standard sectorial operator A on a Hilbert space

H the following assertions are equivalent:
(1) A has a bounded H ∞ -calculus;
(2) A has bounded imaginary powers.
In this situation we have

ωH ∞ (A) = ωBIP (A).

If in addition we have 0 ∈ %(A), then the above conditions are equivalent to

(3) D(A1/2 ) = (H, D(A)) 21 ,2 with equivalent norms.

In view of the equivalence of uniform boundedness and γ-boundedness for

families of Hilbert space operators, the equivalence of (1) and (2) is a special
case of the results in the preceding subsection. A version of the equivalence of
these conditions with (3) for general Banach spaces will be discussed in the
Notes at the end of the chapter.
 15.3 Bounded imaginary powers 483

Proof. It remains to prove the implications (2)⇒(3)⇒(1) under the additional

assumption 0 ∈ %(A). As a preliminary observation we point out that this
assumption implies that we have equivalences of norms
kxkD(A1/2 ) h kA1/2 xk (15.27)
and
kxk(H,D(A)) 1 ,2 h λ 7→ λ1/2 A(λ + A)−1 x dt . (15.28)
2 L2 (R+ , t ;H)

Indeed, (15.27) follows by writing x = A−1/2 A1/2 x and using Corollary 15.2.10
to get
kA1/2 xk 6 kxk + kA1/2 xk 6 (kA−1/2 k + 1)kA1/2 xk.
The equivalence (15.28) follows from Proposition K.4.1.
(2)⇒(3): The equality D(A1/2 ) = (H, D(A)) 21 ,2 is an immediate conse-
quence of Peetre’s theorem (Theorem C.4.1), which in the present situation
implies that for each θ ∈ (0, 1) we have
(H, D(A))θ,2 = [H, D(A)]θ with equivalent norms,
and Theorem 15.3.9, which identifies [H, D(A)]1/2 as the fractional domain
space D(A1/2 ) up to an equivalent norm.
(3)⇒(1): On H define
|||x||| := λ 7→ λ1/2 A1/2 (λ + A)−1 x L2 (R+ , dt
.
t ;H)

In view of (15.27) and (15.28) and the assumption in (3), we have the norm
equivalences
|||x||| h kA−1/2 xk(H,D(A)) 1 ,2 h kA−1/2 xkD(A1/2 ) h kxk.
2

Consequently, ||| · ||| defines an equivalent Hilbertian norm on H. Recalling that

γ(L2 (R+ dt 2 dt
t ), H) = L (R+ , t ; H) isometrically, the implication now follows
from Theorem 10.4.21. 

15.3.h Examples

It has already been noted that every standard sectorial operator A with a
bounded H ∞ (Σσ )-calculus for some 0 < σ < π has bounded imaginary pow-
ers. Here we wish to highlight two examples:
Example 15.3.24 (Laplacian). Let 1 < p < ∞ and let X be a Banach space. It
was already noted in the discussion preceding Theorem 15.3.11 that if X is a
UMD space, then the negative of the Laplace operator ∆ on Lp (Rd ; X) with
domain D(∆) = H 2,p (Rd ; X) has bounded imaginary powers. In the converse
direction, it was shown in Section 10.5 that if −∆ has bounded imaginary
powers on Lp (Rd ; X), then X is a UMD space.
484 15 Extended calculi and powers of operators

Example 15.3.25 (First derivative). Let 1 < p < ∞ and let X be a UMD
space.
(1) The operator A = d/dx on Lp (R; X) with domain D(A) = W 1,p (R; X)
has bounded imaginary powers with angle 21 π.
(2) The operator A = d/dt on Lp (R+ ; X) with domain D(A) = {f ∈
W 1,p (R+ ; X) : f (0) = 0} has bounded imaginary powers with angle 21 π.
(3) The operator A = d/dt on Lp (0, T ; X) with domain D(A) = {f ∈
W 1,p (0, T ; X) : f (0) = 0} has bounded imaginary powers with angle 21 π
and, more precisely, we have the estimate
1
kAis k .T (1 + s2 )e 2 π|s| , s ∈ R.

For the proofs of (1), (2), and the first part of (3) one may observe that in
each of these three cases A is standard sectorial.
In the case (1), −A generates the translation group on Lp (R; X), and in
the other two cases −A is the generator of the C0 -semigroup on Lp (I; X)
(with I = R+ resp. (0, T )) given by
(
f (s − t), s ∈ I, s > t,
S(t)f (s) =
0, otherwise.

All three semigroups are contractive and, in the scalar-valued case, positive. It
follows that we can apply the Hieber–Prüss theorem (Theorem 10.7.12), which
gives that each of these operators has a bounded H ∞ -calculus of angle 21 π.
It then follows from Theorem 15.3.20 that each of the operators has bounded
imaginary powers.

15.4 Strip type operators

It has already been noted in Volume II that the theories of Hardy spaces over
a sector and a strip large rather similar. This similarity can be lifted to the
operator level by introducing the ‘strip’ version of sectorial operators. Such
operator admit again a Dunford calculus, a primary calculus, and an extended
calculus, and one may ask about the boundedness of their H ∞ -calculus. Since
this topic is somewhat peripheral to the mainstream of these volumes, we will
not embark on a systematic exploration of strip type operator, but rather
concentrate on the relationship between sectorial operators and strip type
operators. We have already seen several examples, both in Volume II and
the present volume, where the relationship between sectorial operators and
bisectorial operators (the mediating function being z 7→ z 2 ) can be exploited
in the study of sectorial operators. Likewise the connection with strip type
operators (the mediating function being z 7→ ez ) can sometimes be exploited.
At the end of this section we demonstrate this by giving a proof of the Dore–
Venni theorem by using the properties of strip type operators.
 15.4 Strip type operators 485

15.4.a Nollau’s theorem

For ϑ > 0 let

Sϑ := {z ∈ C : |=z| < ϑ}
be the strip of height ϑ. From Appendix H we recall the definition of the
Hardy space H p (Sϑ ), 1 6 p 6 ∞, as the Banach space of all holomorphic
functions f : Sϑ → C for which the norm

kf kH p (Sϑ ) := sup kt 7→ f (t + iy)kLp (R)

|y|<ϑ

is finite.

Definition 15.4.1. A linear operator A acting in a Banach space X is said

to be of strip type ω > 0 if σ(A) ⊆ Sω and

sup (|=z| − ω)kR(z, A)k < ∞.

z6∈Sω

It is said to be of standard strip type ω > 0 if it is strip type ω > 0 and

D(A) ∩ R(A) is dense in X.

The operator A is said to be of (standard) strip type if it is of (standard) strip

type ω for some ω > 0. The number

ω S (A) := inf{ω > 0 : A is of strip type ω}

is called the height of A.

Example 15.4.2. By the easy part of the Hille–Yosida theorem, if iA is the

generator of a C0 -group (U (t))t∈R satisfying kU (t)k 6 M eω|t| for all t ∈ R
and certain M > 1 and ω > 0, then A is of strip type ω.

Theorem 15.4.3 (Nollau). If A is standard sectorial, then log(A) is of

standard strip type with ω S (A) 6 ω(A), and the following Poisson type formula
holds:
Z ∞
1
R(z, log(A)) = − (t + A)−1 dt, |=z| > π.
0 (z − log t)2 + π 2

Proof. We proceed in two steps.

Step 1 – First we assume in addition that A is bounded and invertible. Let
ω(A) < ν 0 < ν < σ < π and fix λ ∈ C with |=λ| > π and µ ∈ Σν \ Σν 0 . The
function z 7→ 1/(λ − log z) is holomorphic and bounded on Σσ . Let x ∈ X.
Then by Proposition 15.1.19,
1
(A)x
λ − log
486 15 Extended calculi and powers of operators
Z
1 1 1  1 
= x+ R(z, A) − x dz
λ − log µ 2πi ∂Σν λ − log z z−µ
Z ∞
1 1 eiν  1 
= x− R(teiν , A) − iν x dt
λ − log µ 2πi 0 λ − iν − log t te − µ
Z ∞
1 e−iν  1 
+ R(e−iν , A) − −iν x dt.
2πi 0 λ + iν − log t te −µ
By dominated convergence we may pass to the limit ν → π and obtain
1
(A)x
λ − log
Z ∞
1 1 1  1 
= x− (t + A)−1 − x dt
λ − log µ 2πi 0 λ − iπ − log t t+µ
Z ∞
1 1  1 
+ (t + A)−1 − x dt
2πi 0 λ + iπ − log t t+µ
Z ∞
1 1 
−1 1 
= x− (t + A) − x dt
λ − log µ 0 (λ − log t)2 + π 2 t+µ
Z ∞
1
=− 2 + π2
(t + A)−1 x dt,
0 (λ − log t)
where we used that
Z ∞ Z
1 1 1 1 1
2 + π2 t + µ
dt = − lim dz
0 (λ − log t) ν→π 2πi ∂Σν λ − log z z − µ
1
=− .
λ − log µ
1
By the multiplicativity of the extended calculus, λ−log (A) is inverse to λ −
log(A). This gives λ ∈ %(log(A)) as well as the identity for the resolvent.
The resolvent estimate follows from the following estimates, where we write
z = x + iy and set M := supt>0 kt(t + A)−1 k:
Z ∞
1 dt
kR(z, log(A))k 6 M 2 2
|(z − log t) + π | t
Z0 ∞
1
6M 2 + π2 |
ds
−∞ |(z − s)
Z ∞
1
=M 2 2 2 2 2 1/2
ds
−∞ (((x − s) − y + π ) + (2(x − s)y) )
Z ∞
1
=M 2 2 2 2 2 2 1/2
dr
−∞ ((r − y + π ) + 4r y )
Z ∞
1
6M 2 + y2 − π2
dr
−∞ r
Mπ
= 2
(y − π 2 )1/2
 15.4 Strip type operators 487

Mπ
6 ,
|y| − π
where we used the elementary inequalities

(r2 − y 2 + π 2 )2 + 4r2 y 2 = r4 + y 4 + π 4 + 2r2 y 2 + 2π 2 r2 − 2π 2 y 2

> r4 + y 4 + π 4 + 2r2 y 2 − 2π 2 r2 − 2π 2 y 2
= (r2 + y 2 − π 2 )2

and (keeping in mind that |y| > π, so 2|y| − π > π)

y 2 − π 2 > y 2 − π(2|y| − π) = (|y| − π)2 .

This proves the theorem under the additional assumption that A is bounded
and has bounded inverse.
Step 2 – To deduce the general case, for ε > 0 we consider the operators

Aε = (A + ε)(I + εA)−1 .

For λ > 0 we have

λ + Aε = λ(I + εA)(I + εA)−1 + (A + ε)(I + εA)−1

= (λ + ε + (λε + 1)A)(I + εA)−1 ,

and therefore λ + Aε is invertible. For λ > 0 we estimate

k(λ + Aε )−1 k = k(I + εA)(λ + ε + (λε + 1)A)−1 k

ε 1 λ+ε
= ( + A)( + A)−1
λε + 1 ε λε + 1
ε 1 λ+ε λ+ε
= I +( − )( + A)−1
λε + 1 ε λε + 1 λε + 1
1 ε 1 λ + ε  λε + 1
6 + − MA
λ λε + 1 ε λε + 1 λ + ε
1 1 − ε2
= + MA
λ (λ + ε)(λε + 1)
1 + MA
6 ,
λ
where MA = supλ>0 kλ(λ + A)−1 . It follows that
 
sup sup kλ(λ + Aε )−1 k 6 1 + MA ,
ε>0 λ>0

and therefore the operators Aε are uniformly sectorial. In particular the results
of Step 1 apply to Aε , with bounded that are uniform in ε > 0.
Step 3 – Take 0 < ν < π close enough to π so that ∂Σδ is contained in
the resolvent set of each Aε . Noting that R(z, Aε )x → R(z, A)x uniformly on
488 15 Extended calculi and powers of operators

∂Σδ (this follows from uniform sectoriality and similar estimates as above),
we may pass to the limit ε ↓ 0 and obtain
Z
1 1 1 1  1 
(A)x = x+ R(z, A) − x dz
λ − log λ − log µ 2πi ∂Σν λ − log z z−µ
Z
1 1 1  1 
= lim x+ R(z, Aε ) − x dz
ε↓0 λ − log µ 2πi ∂Σν λ − log z z−µ
Z ∞
1
= lim − (t + Aε )−1 x dt
ε↓0 0 (λ − log t)2 + π 2
Z ∞
1
=− 2 + π2
(t + A)−1 x dt.
0 (λ − log t)
1
Next we show that λ ∈ %(log(A)) and λ−log (A) is a two-sided inverse for
λ−log(A). For x ∈ D(A2 )∩R(A2 ), say x = ζ 2 (A)y, by the general properties of
1
the extended Dunford calculus we have λ−log (A)x ∈ D(A)∩R(A) ⊆ D(log(A))
and
1 ζ
(λ − log(A)) (A)x = ζ(A)(λ − log(A)) (A)y
λ − log λ − log
ζ
= (ζ(λ − log))(A) (A)y
λ − log
 ζ 
= ζ(λ − log) (A)y
λ − log
= ζ 2 (A)y = x
1
and similarly λ−log (A)(λ − log(A))x = x. By density and closedness, these
identities extend to general x ∈ X and x ∈ D(log(A)), respectively.
Finally, the strip type estimate for A follows from the corresponding esti-
mate for Aε proved above, by letting ε ↓ 0 and using dominated convergence
once more. 
One may set up a Dunford calculus and extended Dunford calculus for strip
type operators in much the same way as we did for sectorial operators as
follows. For an operator A of strip type and f ∈ H 1 (Sσ ), where σ > ω S (A),
the Dunford integral
Z
1
f (A)x := f (z)R(z, A)x dx,
2πi ∂Sν
defines a bounded operator f (A) on X. The defining integral converges abso-
lutely and by Cauchy’s theorem it is independent of the choice of ν. Moreover,
Z
1 C 1 C
kf (A)k 6 lim sup |f (z)| |dz| 6 kf kH 1 (Sσ ) .
ν↓ω S (A) 2π ν − ω |=z|=ν π σ − ω S (A)

The elementary properties of the extended Dunford calculus extend to the

strip case.
 15.4 Strip type operators 489

15.4.b Monniaux’s theorem

We have seen (Proposition 15.3.5) that if an operator B in a Banach space

X has bounded imaginary powers, then the bounded operators B it form a
C0 -group on X. In this subsection we will show that if X is a UMD space,
then conversely every C0 -group on X of growth type less than π with injective
generator is of the form U (t) = B it for some operator B in X with bounded
imaginary powers:
Theorem 15.4.4 (Monniaux). Let (U (t))t∈R be a C0 -group on a UMD
space X satisfying kU (t)k 6 M eω|t| for all t ∈ R and some M > 1 and
0 6 ω < π. Assume furthermore that its generator iA is injective. Then there
exists an operator B in X with bounded imaginary powers, given by

B it = U (t), t ∈ R.

Moreover, we have A = log(B) with equal domains.

Intuitively, one has B = eA ; the identity B it = U (t) then corresponds to the
intuition that (eA )it = eitA .
The proof of the theorem relies on several ingredients. The first is the
following lemma.
Lemma 15.4.5. Let (U (t))t∈R be a C0 -group on a UMD space X. If, for some
M > 1 and ω ∈ R, we have kU (t)k 6 M eω|t| for all t ∈ R, then for all x ∈ X
the principal value integral
Z 1
dt
p.v. U (t)x
−1 t
converges in X and has norm
Z 1
dt
p.v. U (t)x 6 6C 2 ~2,X kxk,
−1 t

where ~2,X := kHkL (L2 (R;X)) , and C := sup|t|62 kU (t)k.

Proof. All we need to do is stripping the Rademacher sums from the esti-
mates in the last part of the proof of Theorem 15.3.12(2). For the reader’s
convenience we include the proof that results from this.
Fix 0 < δ < 1, s ∈ [− 21 , 12 ], and x ∈ X. Then
Z Z
dt ds
U (t)x = U (s) U (t − s)x
δ<|t|<1 t δ<|t|<1 s
Z
dt
= U (s) ϕx (s − t)
|t|>δ t
Z 1+s Z −1+s
dt dt
− U (t)x + U (t)x ,
1 t −1 t
490 15 Extended calculi and powers of operators

where ϕx (τ ) = 1(−1,1) (τ )U (−τ )x. Integrating over [− 21 , 12 ], we obtain

Z Z 1 Z
dt 2 dt
U (t)x = U (s) ϕx (s − t)
ds
δ<|t|<1 t − 12 |t|>δ t
Z 1 Z 1+s Z 12 Z −1+s
2 dt dt
− U (t)x ds + U (t)x ds.
− 12 1 t 1
− 2 −1 t

Since X is UMD and φx ∈ L2 (R; X), the limit

Z Z
dt dt
lim ϕx (· − t) = lim ϕx (· − t)
δ↓0 |t|>δ t δ↓0 δ<|t|<R t
R→∞

exists in L2 (R; X) by Theorem 5.1.1 and equals πHφx , where H denotes the
Hilbert transform. As a result we obtain
Z 1 Z
dt dt
p.v. U (t)x = lim U (t)x
−1 t δ↓0 δ<|t|<1 t
Z 12
=π U (s)Hϕx (s) ds
− 12
1 1
Z Z 1+s Z Z −1+s
2 dt 2 dt
− U (t)x ds + U (t)x ds
− 12 1 t − 12 −1 t
=: I + II + III.

With constants C := sup|t|62 kU (t)k and ~2,X := kHkL (L2 (R;X)) , we have
1
Z 2  21 √
kIk 6 πC kHϕx (s)k2 ds 6 πC~2,X kϕx kL2 (R;X) 6 2πC 2 ~2,X kxk.
− 12

The other two terms are elementary with

1
Z Z 1+s
2 dt
kIIk 6 C kxk ds 6 C log 2kxk,
− 12 1 t

√ in exactly the same way. Note that both C > kU (0)k = 1

and III can bounded
and ~2,X > 1, and 2π + 2 log 2 < 6. 
We will use this lemma for the second ingredient for the proof of Theorem
15.4.4, a primary calculus for strip type operators. We work under the as-
sumptions of Theorem 15.4.4 and let ω < σ < π. For functions

g ∈ L1ω (R) = g ∈ L1loc (R) : t 7→ eω|t| g(t) ∈ L1 (R)



we define the bounded operator gb(A) := Φg (A) by the Phillips calculus (see
Section 10.7.a):
 15.4 Strip type operators 491
Z ∞
gb(A)x := g(t)U (t)x dt, x ∈ X.
0
Obviously,
g (A)kL (X) 6 M kgkLω1 (R) ,
kb
where kgkLω1 (R) := kt 7→ eω|t| g(t)kL1 (R) and M is in Theorem 15.4.4. The
following lemma enables us to enrich this calculus with certain bounded func-
tions in H ∞ (Sσ ) which have limits for <z → ±∞.

Lemma 15.4.6. The Fourier transform of the distribution

Z
1  dt 
h(t) := p. v. 1(−1,1) (t) hh, φi := lim φ(t) ∀φ ∈ S (R) ,
t ε↓0 ε<|t|<1 t

equals Z 1 Z 1
dt dt
h(ξ) =
b e−2πitξ =2 sin(2πtξ)
−1 t 0 t
and its analytic continuation to Sσ satisfies

h(z) = ±π.
lim b
|=z|<σ
<z→∞

Proof. The first assertion follows by elementary computation and the second
from the standard improper integral
Z ∞
dt π
sin t =
0 t 2

and a change of variables. 

For small ε > 0 let hε (t) := t−1 1(−1,−ε]∪[ε,1) (t). Applying the Phillips calculus,
we obtain Z 1
h
cε (A)x = hε (t)U (t)x dt
−1

and therefore the principal value integral

Z π
dt
h(A)x := p.v.
b U (t)x
−π t

exists and satisfies

h(A)xk 6 6C 2 ~2,X kxk
kb
with constants as in Lemma 15.4.5.
Now we are in a position to define our primary calculus:
492 15 Extended calculi and powers of operators

Definition 15.4.7 (Primary calculus). Let A be a strip type operator and

let ω S (A) < ω < σ. For functions f : Sσ → C of the form

f = gb + ab
h+b
1
with a, b ∈ C, g ∈ Lω (R), and h as above, we set

f (A) := gb(A) + ab
h(A) + bI.

The condition on f is satisfied if f 0 is bounded and there exists an α > 1 such

that
f 0 (z) = O(|z|α ) as |<z| → ∞. (15.29)
The primary calculus enjoys similar properties as the one for sectorial
operators; in particular it is multiplicative and consistent with the Dunford
calculus. The proof is elementary but a bit tedious and it is therefore left to
the reader.
For every r ∈ R the primary calculus can be applied to A + r in place of
A, noting that i(A + r) generates the C0 -group (eirt U (t))t∈R . It is immediate
from the above constructions that the estimates are uniform with respect to
r, i.e., for all f : Sσ → C of the above form we have

sup kf (A + r)k < ∞. (15.30)

r∈R

We will now exploit the fact that, for 0 6 |ω| < σ < π, the exponential
function z 7→ ez maps the line =z = ω bijectively onto the ray arg(z) = ω.
Thus, it maps the strip Sσ bijectively onto the sector Σσ . From this, we infer
that if λ ∈ {Σσ , then the function
1
rλ (z) := (15.31)
λ − ez
is bounded and holomorphic on Sσ . What is more, this function is of the form
discussed above and therefore rλ (A) is well defined in the primary calculus
(as its derivative satisfies (15.29)).

Remark 15.4.8. In hindsight, one could have introduced the primary calculus
using the functions rλ instead of h. This would restore the symmetry with the
definition of the primary calculus for sectorial operator. Hoever, this would
require an independent construction of the operators rλ (A) = R(λ, B) by
different means.

By the algebraic properties of the functions rλ and the multiplicativity

properties of the calculus, the operators Rλ form a pseudo-resolvent in the
sense of the following proposition.
 15.4 Strip type operators 493

Lemma 15.4.9 (Pseudo-resolvents). Let V ⊆ C be a non-empty con-

nected open set and let (Rλ )λ∈V be a family of bounded operators on a Banach
space X satisfying the resolvent identity

Rλ − Rµ = (µ − λ)Rλ Rµ , λ, µ ∈ V.

If Rλ0 is injective for some λ0 ∈ V , then there exists a unique closed operator
B on X such that V ⊆ %(B) and Rλ = (λ − B)−1 for all λ ∈ V .

Proof. The resolvent identity implies that any two Rλ and Rµ commute.
If Rλ x = 0, the identity Rλ − Rλ0 = (λ0 − λ)Rλ0 Rλ implies that Rλ0 x = 0
and therefore x = 0. It follows that Rλ is injective. If y ∈ R(Rλ0 ), there is a
unique x ∈ X such that y = Rλ0 x. Then y = Rλ (I − (λ0 − λ)Rλ0 )x ∈ R(Rλ ).
This shows that N := N(Rλ ) = {0} and R := R(Rλ ) are independent of
λ ∈ V . Hence if y ∈ R, then for all λ ∈ V there is a unique xλ ∈ X such that
y = Rλ xλ . Then, by the resolvent identity,

(µ − λ)Rλ Rµ y = Rλ y − Rµ y = Rλ Rµ xµ − Rµ Rλ xλ = Rλ Rµ (xµ − xλ ).

It follows that (µ − λ)y = (xµ − xλ ). This implies that µy − xµ = λy − xλ

is independent of λ, µ ∈ V . Denoting this element by By, we obtain a linear
operator B : y 7→ λy − xλ with domain D(B) = R(Rλ0 ). It satisfies

Rλ (λ − B)y = Rλ xλ = y,

so Rλ is a left inverse to λ−B. Applying this to Rλ y instead of y we also obtain

Rλ (λ−B)Rλ y = Rλ y, and the injectivity of Rλ therefore gives (λ−B)Rλ y = y,
so Rλ is a right inverse to λ − B. This proves that λ ∈ %(B) and (λ − B)−1 =
Rλ . That B is closed follows from the fact that its resolvent set contains the
non-empty set V . 
This construction gives us a rigorous way to construct the operator eA as the
closed operator B given by the lemma.
Proof of Theorem 15.4.4. By Lemma 15.4.9 there exists a unique closed op-
erator B on X such that Rλ = rλ (A) = (λ − B)−1 for all λ ∈ V := {Σσ . We
note that
λR(λ, B) = (I − λ−1 B)−1 = r1 (A − log λ),
and therefore (15.30) implies that B is σ-sectorial.
Since sectorial operators on reflexive Banach spaces are densely defined
(see Proposition 10.1.9), the operator B is densely defined. We claim that B
in injective. If λ, µ ∈ {Σσ , then
ez
(λ − B)−1 (µ − B)−1 B ⊆ fλ,µ (B) with fλ,µ (z) = .
(λ − ez )(µ − ez )

The operator fλ,µ (B) is injective in view if the identity

494 15 Extended calculi and powers of operators

1 µ−1
fλ,µ (B) = − (B) (B).
λ − ez µ−1 − e−z
This proves the claim. As a consequence of Proposition 15.3.2, we obtain that
B is standard sectorial.
The identity B it = U (t) follows by writing out the definition of B it in
the extended calculus of B. This results in an expression involving a Dunford
integral containing the resolvent of B. This resolvent can be expressed, via the
definition of the primary calculus, in terms of the Phillips calculus of the C0 -
group U . The details are laborious and are left to the reader. From this, and
the general properties of the extended calculus, it follows that A = log(B). 

15.4.c The Dore–Venni theorem

In this section we apply Monniaux’s theorem (Theorem 15.4.4) to prove the

celebrated Dore–Venni theorem on the closedness of sums of closed operators.
We base our proof on the following lemma. It uses the fact that if iG is the
generator of a bounded C0 -group, then G is bisectorial of angle 0 (see Example
10.6.3). In what follows we use that the extended Dunford calculus can be
developed also for bisectorial operators. When we cite results from Section
15.1, it is understood that we actually refer to their bisectorial counterparts.
We leave it to the reader to verify that these counterparts do indeed hold.

Lemma 15.4.10. Let (U (t))t∈R and (V (t))t∈R be commuting C0 -groups on a

Banach space X with generators −iA and −iB, respectively, such that

kU (t)k 6 MA eωA |t| and kV (t)k 6 MB eωB |t|

for all t ∈ R, where MA , MB > 1 and ωA , ωB > 0 satisfy ωA + ωB 6 π. Let

−iC denote the generator of the C0 -group W (t) = U (t)V (t), t ∈ R. Then for
all x ∈ D(eA eB ) we have x ∈ D(eC ) and

eA eB x = eC x.

Proof. We begin by noting that

kU (t)k 6 MA MB e(ωA +ωB )|t| , t ∈ R.

In what follows we fix ωA + ωB < σ < π.

Fix 0 < ϑ < 21 π and consider the functions f, g ∈ H 1 (Σϑ ) given by
(
e2z (1 + ez )−3 , z ∈ Σϑ+ ,
f (z) =
e−z (1 + e−z )−3 , z ∈ Σϑ− ,

and
 15.4 Strip type operators 495
(
ez (1 + ez )−3 , z ∈ Σϑ+ ,
g(z) =
e−2z (1 + e−z )−3 , z ∈ Σϑ− .

For G ∈ {A, B, C} the operators f (G) and g(G) are well defined and bounded
in the bisectorial Dunford calculus. Moreover, by (the bisectorial counterpart
of) Proposition 15.1.12 these operators are injective and ez = f (z)/g(z) im-
plies
eG = g(G)−1 f (G).
Our aim is to prove that

f (A)f (B)g(C) = g(A)g(B)f (C). (15.32)

Once we have shown this, from f (A)g(B)−1 ⊆ g(B)−1 f (A) (this follows by
observing that D(g(B)−1 ) = R(g(B)) and using that f (A) and g(B) commute,
we obtain

eA eB = g(A)−1 f (A)g(B)−1 f (B)

⊆ g(A)−1 g(B)−1 f (A))f (B)
= (g(A)−1 g(B)−1 g(C)−1 )(g(C)f (A))f (B))
= (g(C)−1 g(A)−1 g(B)−1 )(g(A)g(B)f (C))
= (g(C)−1 g(A)−1 g(B)−1 )(g(B)g(A)f (C))
= g(C)−1 f (C) = eC ,

using that g(A) and g(B) commute in the penultimate equality.

The proof of (15.32) relies on the properties of the Phillips calculus (see
Section 10.7.a). We recall from Proposition 10.7.2 and Lemma 10.7.3 that if
iG generates a bounded C0 -group (W (t))t∈R on X, then for 0 < ϑ < 21 π and
h ∈ H 1 (Σϑbi ) one has
Z ∞
h(G)x = φh (t)W (t)x dt, x ∈ X,
−∞

where φh ∈ L1 (R) is given by

φ
ch (ξ) = h(−2πξ), ξ ∈ R.

Applying this to G ∈ {−A, −B, −C} and h ∈ {f, g}, and keeping in mind
that −iA, −iB, and −iC generate the groups U (t), V (t), and U (t)V (t), re-
spectively, the identity (15.32) takes the form
Z Z Z
φg (r)φf (t)φf (s)U (t + r)V (s + r) dr ds dt
R R R
Z Z Z
= φf (r)φg (t)φg (s)U (t + r)V (s + r) dr ds dt,
R R R
496 15 Extended calculi and powers of operators

or equivalently,
Z Z Z Z
φg (r)F (t, s)U (t)V (s) ds dt = φg (r)G(t, s)U (t)V (s) ds dt
R R R R

with
Z Z
F (t, s) = φg (r)φf (t−r)φf (s−r) dr, G(t, s) = φf (r)φg (t−r)φg (s−r) dr.
R R

Taking Fourier transforms, we obtain

Z Z
F (x, y) =
b F (t, s)e−2πi(tx+sy) dt ds
R R
Z Z Z
= φg (r)φf (t − r)φf (s − r)e−2πi(tx+sy) dr dt ds
ZR ZR ZR
= φg (r)φf (t − r)φf (s − r)e−2πitx e−2πisy e−2πir(x+y) dr dt ds
R R R

=φcg (x + y)φ
cf (x)φ
cf (y)
= g(−2π(x + y))f (−2πx)f (−2πy)

and similarly
b y) = f (−2π(x + y))g(−2πx)g(−2πy).
G(x,

It is evident from the definitions of f and g that the two right-hand sides
are equal. Therefore F = G be the uniqueness of Fourier transforms. This
completes the proof of (15.32). 

Theorem 15.4.11 (Dore–Venni). Suppose that A and B are resolvent

commuting standard sectorial operators on a UMD Banach space X. If both
A and B have bounded imaginary powers with

ωBIP (A) + ωBIP (B) < π,

then there exists a constant K > 0 such that

kAxk + kBxk 6 Kk(A + B)xk, x ∈ D(A) ∩ D(B).

As a consequence A + B, with its natural domain D(A + B) = D(A) ∩ D(B),

is closed.

Proof. Fix ωBIP (A) + ωBIP (B) < ω < π. Since A and B resolvent commute,
the operators UA (s) = Ais and UB (t) = B it commute for all s, t ∈ R and
U (t) := Ait B −it is a C0 -group satisfying kU (t)k 6 Keω|t| for all t ∈ R and
some K > 1. Hence by Monniaux’s Theorem 15.4.4 we have U (t) = C it ,
t ∈ R, for some standard sectorial operator C having bounded imaginary
 15.5 The bisectorial H ∞ -calculus revisited 497

powers. The generators of the C0 -groups equal i log A, −i log B, and i log C.
Since I + C is invertible there is a constant M > 0 such that have

kxk 6 M kx + Cxk, x ∈ D(C).

By Lemma 15.4.10 applied to i log A and −i log B, for all x ∈ D(AB −1 )

we have x ∈ D(C) and AB −1 x = Cx. Hence for all x ∈ D(A) ∩ D(B) we have
Bx ∈ D(AB −1 ) ⊆ D(C), and therefore

kBxk 6 M kBx + CBxk = M kBx + Axk.

The same argument with the roles of A and B interchanged gives the inequal-
ity
kAxk 6 M kBx + CBxk = M kBx + Axk.
Together, these two inequalities imply the inequality in the statement of the
theorem. This implies the closedness of A + B by a routine argument. 

15.5 The bisectorial H ∞ -calculus revisited

The bisectorial H ∞ -calculus has been introduced and studied in Section 10.6.
The purpose of the present section is to study in more detail the spectral
projections associated with the left and right halves of the bisector. These
can be thought of as abstract Riesz projections. The main result is Theorem
15.5.2, which establishes that if A is a standard bisectorial operator with a
bounded H ∞ (Σσbi )-calculus, then A2 is a standard 2σ-sectorial operator and

D((A2 )1/2 ) = D(A),

with equivalence of norms

k(A2 )1/2 xk h kAxk.

We use the notation introduced in Section 10.6. Specifically, for 0 < ω <
1
2π we define Σω+ := Σω and Σω− := −Σω , and denote by

Σωbi := Σω+ ∪ Σω−

the bisector of angle ω. Recall that a linear operator A on a Banach space X

is said to be bisectorial if there exists an ω ∈ (0, 21 π) such that the spectrum
σ(A) is contained in Σωbi and
bi
Mω,A := sup kzR(z, A)k < ∞.
bi
z∈{Σω

In this situation we say that A is ω-bisectorial. The infimum of all ω ∈ (0, 21 π)

such that A is ω-bisectorial is called the angle of bisectoriality of A and is
denoted by ω bi (A).
498 15 Extended calculi and powers of operators

15.5.a Spectral projections

What distinguishes the theory of bisectorial operators from its sectorial coun-
terpart is the possibility to consider the functions 1Σ + and 1Σ − , both of which
are bounded and holomorphic as functions on the bisector Σ bi = Σ + ∪ Σ − .
If a bisectorial operator A has a bounded H ∞ (Σ bi )-calculus, the operators
1Σ + (A) and 1Σ − (A) are well defined as bounded operators on D(A) ∩ R(A)
and take the role of “spectral projections” associated with the sectors Σ +
and Σ − . (The reason for writing quotations is that one has to be a bit care-
ful here since 0 may belong to the spectrum of A.) From the multiplicativity
of the H ∞ -calculus it follows that the operators 1Σ + and 1Σ − are indeed
projections, and that they are mutually orthogonal in the sense that

1Σ + (A)1Σ − (A) = 1Σ − (A)1Σ + (A) = 0.

The injectivity of A on D(A) ∩ R(A), which follows from Proposition 10.1.8,

will be seen to imply the identity

1Σ − (A) + 1Σ − (A) = I.

The importance of the operators 1Σ + (A) and 1Σ − (A) stems from their anal-
ogy to the Riesz projections; in particular, their difference

1Σ + (A) − 1Σ − (A) =: sgn(A)

may be thought of as an abstract analogue of the Hilbert transform.

Proposition 15.5.1. If A is a bisectorial operator on a Banach space X with
a bounded H ∞ (Σσbi )-calculus for some ω bi (A) < σ < 21 π, then the operators

P + := 1Σσ+ (A), P − := 1Σσ− (A),

are well defined as bounded projections on D(A) ∩ R(A). As such they are
mutually orthogonal in the sense that

P + P − = P − P + = 0,

and complementary in the sense that

P + + P − = I.

Denoting the parts of A in X + := R(P + ) and X − := R(P − ) by A+ and A−

respectively, then both A+ and −A− are sectorial on X + and X − and have
bounded H ∞ (Σσ+ )-calculus on these spaces. We have

σ(A± ) ⊆ σ(A) ∩ Σσ± ,

and if D(A) ∩ R(A) is dense we also have

(σ(A) ∩ Σσ± ) \ {0} ⊆ σ(A± ) \ {0}.

 15.5 The bisectorial H ∞ -calculus revisited 499

There is some abuse of notation in writing X ± for the range of P ± , as these

operators are projections defined on D(A) ∩ R(A) only, but we do not want
to overburden the notation.
Proof. That the operators P + and P − are mutually orthogonal projections
on D(A) ∩ R(A) and add up to the identity follows from the general properties
of the H ∞ -calculus. It is also clear that both projections commute with the
resolvent of A. Thus the spaces X + and X − are invariant under the resolvent
of A. Denote by A+ and A− the parts of A to X + and X − . It suffices to
prove that A+ has the asserted properties; the result for A− then follows by
applying it to the bisectorial operator −A.
Step 1a – In this step we prove that σ(A+ ) ⊆ σ(A) ∩ Σσ+ .
Let us write p+ (z) = 1Σσ+ (z) (so p+ (A)x = P + x) and rµ (z) = (µ − z)−1
for brevity.
The crux of the argument is the observation that if µ ∈ {Σσ+ , then rµ p+
is a bounded holomorphic function on the bisector Σσbi even when µ ∈ Σσ− .
Therefore the operator (rµ p+ )(A) is well defined by the H ∞ (Σσbi )-calculus of
A.
We shall prove next that the restriction of this operator to X + is a two-
sided inverse of µ − A+ . This will show that inclusion σ(A+ ) ⊆ Σσ+ . By
general spectral considerations we also have σ(A+ ) ⊆ σ(A), and together
these inclusions prove
σ(A+ ) ⊆ σ(A) ∩ Σσ+ .
First we prove that the restriction of (rµ p+ )(A) to X + is a two-sided
inverse of µ − A+ for µ ∈ {Σσbi . Fix x ∈ D(A) ∩ R(A). We have rµ (A)x =
R(µ, A)x by the properties of the H ∞ (Σσbi )-calculus. The multiplicativity of
this calculus then implies

(rµ p+ )(A)x = rµ (A)p+ (A)x = R(µ, A)P + x,

(15.33)
(rµ p+ )(A)x = (p+ rµ )(A)x = P + R(µ, A)x.

It follows that X + is invariant under R(µ, A). Multiplying the first identity on
the left and the second on the right by µ − A+ we see (rµ p+ )(A) is a two-sided
inverse to µ − A+ . It follows that µ ∈ %(A+ ) and

R(µ, A+ ) = (rµ p+ )(A)|X + . (15.34)

We now consider the case of a general µ ∈ {Σσ , which will be handled

by the resolvent identity. To this end fix an arbitrary λ ∈ {Σσbi . The scalar
identity
1 1 λ−µ
= +
µ−z λ−z (λ − z)(µ − z)
translates, after multiplying with p+ (z) and using the additivity and multi-
plicativity of the H ∞ -calculus, into the identity
500 15 Extended calculi and powers of operators

(rµ p+ )(A)x = (rλ p+ )(A)x + (λ − µ)(rλ p+ )(A)(rµ p+ )(A)x, (15.35)

still for x ∈ D(A) ∩ R(A). Among other things it implies that X + is invariant
under (rµ p+ )(A), since we have just proved that (rλ p+ )(A) = P + R(µ, A)
maps into X + . By (15.33) and (15.34) (applied with λ in place of µ), the
right-hand side of (15.35) (hence also the left-hand side) belongs to D(A),
and for x ∈ X + we obtain

(µ − A)(rµ p+ )(A)x = (µ − λ)(rµ p+ )(A)x + (λ − A)(rµ p+ )(A)x

h i
= (µ − λ)(rµ p+ )(A)x + x + (λ − µ)(rµ p+ )(A)x
= x.

It follows that (rµ p+ )(A)x ∈ D(A+ ) and (rµ p+ )(A) is a right inverse of µ−A+
on X + . Also, using (15.34) (again with λ in place of µ) and the fact that
(rλ p+ )(A) and (rµ p+ )(A) in (15.35) commute, for x ∈ D(A+ ) we obtain

(rµ p+ )(A)(µ − A+ )x = (rµ p+ )(A)(µ − λ)x + (rµ p+ )(A)(λ − A+ )]x

= (µ − λ)(rµ p+ )(A)x + [x + (λ − µ)(rµ p+ )(A)x] = x,

and therefore (rµ p+ )(A) is also a right inverse.

Step 1b – We now prove that if µ 6= 0 belongs to σ(A)∩Σσ+ and D(A)∩R(A)
is dense, then µ ∈ σ(A+ ).
To this end let µ ∈ Σσ+ . Since µ 6= 0, we have µ ∈ {Σσ− and therefore, by
the version of Step 1a for A− , µ ∈ %(A− ). Now if we also had µ ∈ %(A+ ),
then along the decomposition X = D(A) ∩ R(A) = X + ⊕ X − , the operator
R(µ, A+ )⊕R(µ, A− ) would be a two-sided inverse for µ−A and it would follow
that µ ∈ %(A). Hence for µ 6= 0 this proves the inclusion σ(A) ∩ Σσ+ ⊆ σ(A+ ).
Step 2 – We next establish the resolvent bound for A+ . The uniform
bi
boundedness of zR(z, A+ ) on {Σσ follows from the uniform bounded-
ness of zR(z, A) on this set by taking restrictions. In particular, for any
1 + ±iϑ
2 π < ϑ < π − σ, zR(z, A ) is uniformly bounded on the two rays re ,
and then the sectorial version of the three lines lemma implies the uniform
boundedness of zR(z, A+ ) on Σϑ− .
Step 3 – The prescription f (A+ ) := (p+ f )(A) defines a bounded linear
multiplicative mapping from H ∞ (Σσ+ ) into L (D(A) ∩ R(A)) that agrees with
the Dunford calculus of A+ for functions f ∈ H 1 (Σσ+ ) ∩ H ∞ (Σσ+ ). This im-
mediately implies that A+ has a bounded H ∞ (Σσ+ )-calculus. 

15.5.b Sectoriality versus bisectoriality

The next theorem establishes a relationship between a bisectorial operator A

and the square root of the sectorial operator A2 . In the proof we shall use the
fact, which can be routinely checked by redoing the arguments of Section 15.1,
 15.5 The bisectorial H ∞ -calculus revisited 501

that an extended Dunford calculus can be set up for bisectorial operators and
that it enjoys similar properties as in the sectorial case.
Theorem 15.5.2. If A is a standard bisectorial operator with a bounded
H ∞ (Σσbi )-calculus on a Banach space X, then A2 is a standard 2σ-sectorial
operator and
D((A2 )1/2 ) = D(A).
For all x in this common domain we have
k(A2 )1/2 xk h kAxk
with constants independent of x.
Proof. The function a(z) := (z 2 )1/2 is holomorphic on Σσbi and equals z on Σσ+
and −z on Σσ− . Likewise, the function sgn(z) := z/(z 2 )1/2 is holomorphic on
Σσbi and equals 1 on Σσ+ and −1 on Σσ− . Thus, sgn(z)a(z) = z for all z ∈ Σσbi .
By the multiplicativity of the extended functional calculus (cf. Proposition
15.1.12 for the sectorial case), it follows that, if x ∈ D((A2 )1/2 ), then x ∈ D(A)
and Ax = sgn(A)(A2 )1/2 x. Taking norms, we find that
kAxk 6 M k(A2 )1/2 xk,
where M is the boundedness constant of the H ∞ (Σσbi )-calculus of A.
In the same way, the identity a(z) = sgn(z)z implies that, if x ∈
D((A2 )1/2 ), then x ∈ D(A) and (A2 )1/2 x = sgn(A)Ax. Taking norms gives
k(A2 )1/2 xk 6 M kAxk.

It is of some interest to interpret this theorem for the Hodge–Dirac oper-
ator D on L2 (Rd ) ⊕ L2 (Rd ; Cd ) of Example 10.6.5,
0 ∇∗
 
D= .
∇ 0
where ∇∗ = −div is the adjoint of ∇. Its square is of the form
 
−∆ 0
D2 = ,
0 ∗

where ∗ equals (at least formally) −∇◦div. Taking g(z) = sgn(z) := z/(z 2 )1/2 ,
then (at least formally)
(−∆)−1/2 0
     
0 −div 0 ∗
g(D) = D(D2 )−1/2 = · = .
∇ 0 0 ∗ ∇/(−∆)1/2 0
Thus we see the Riesz transform
R = ∇(−∆)1/2
appear as an entry in the functional calculus of D.
502 15 Extended calculi and powers of operators

15.6 Notes
Section 15.1

The idea to use regularising functions to extend the functional calculus to suit-
able classes of unbounded functions goes back to McIntosh [1986]. A compre-
hensive discussion of extended functional calculi is presented by Haase [2006];
see also Haase [2020]. Our treatment in Sections 15.1 and 15.2 is based on
Haase [2006] and Kunstmann and Weis [2004]. The proof of Theorem 15.1.18
is taken from the former reference.

Section 15.2

The first unified account of the theory of fractional powers was undertaken
by Komatsu in a series of papers starting with Komatsu [1966]. This paper
contains the results presented here and much more. Some earlier works on the
subject are due to Balakrishnan [1960], Hille and Phillips [1957], Kato [1960,
1961], Krasnosel0 skiı̆ and Sobolevskiı̆ [1959], Phillips [1952], Watanabe [1961],
and Yosida [1960]. Modern accounts include Albrecht, Duong, and McIntosh
[1996], Denk, Hieber, and Prüss [2003], Dore [1999, 2001], Haase [2006], and
Martı́nez Carracedo and Sanz Alix [2001]. Our treatment barely scratches the
surface of this rich and vast subject, and we have only included the most basic
results needed for the treatment of bounded imaginary powers. Our approach
based on the extended Dunford calculus has the advantage of keeping the
technical details at a minimum, but the price to be paid is that we must make
somewhat restrictive assumptions on the operator A.
Theorem 15.2.8 is from Komatsu [1966], but the proof presented here is
taken from Haase [2006]. Theorem 15.2.13, 15.2.17, and 15.2.16 are due to
Balakrishnan [1960] (see also Yosida [1980]). Some authors take one of the
formulas in the first and third theorem as the definition of the fractional
powers. For further information on the classical approach to fractional powers
via integral representations, we refer the reader to the monographs Butzer
and Berens [1967] and Martı́nez Carracedo and Sanz Alix [2001]. A complete
proof of the non-negativity of the function fα,t in Theorem 15.2.17 can be
found in Yosida [1980, Proposition IX.11.2].

Section 15.3

A detailed account of the theory of bounded imaginary powers is presented

by Haase [2006]; see also Amann [1995] and Prüss and Simonett [2016].
Example 15.3.25 is from Dore and Venni [1987]. Lemma 15.3.8 is from
Prüss and Sohr [1990], where a different proof based on properties of the
Mellin transform is given. Alternative proofs were obtained subsequently by
Monniaux [1997] and Uiterdijk [1998]. The elementary proof presented here,
based on the perturbation result of Theorem 15.1.18, is taken from Haase
 15.6 Notes 503

[2006]. Theorem 15.3.9, identifying domains of fractional powers with complex

interpolation spaces in the presence of bounded imaginary powers, is due to
Seeley [1971]; see Triebel [1978] for references to earlier results in this direction.
Theorem 15.3.12, connecting the angles (R-)sectoriality and bounded
imaginary powers, is due to Prüss and Sohr [1990] (ω(A) 6 ωBIP (A)) and
Clément and Prüss [2001] (ωR (A) 6 ωBIP (A), in a UMD space). The es-
timation of the three terms in the last part of the proof are patterned af-
ter the proof of Lemma 15.4.5, which is taken from Monniaux [1999] and
extends earlier results of Zsidó [1983] and Berkson, Gillespie, and Muhly
[1986a]. Lemma 15.3.13 is from Prüss and Sohr [1990]. In Remark 15.3.14, the
identity (15.17) can be equivalently stated in terms of the Mellin transform;
see, e.g., Titchmarsh [1986]. The two theorems about (almost) γ-sectoriality
and (γ-)bounded imaginary powers, Theorem 15.3.16 (e ωγ (A) 6 ωBIP (A)) and
its proof, as well as Theorem 15.3.19 (ωγ (A) 6 ωγ -BIP (A)), are taken from
Kalton, Lorist, and Weis [2023]. Theorem 15.3.20 (ωBIP (A) = ωH ∞ (A)) is
from Cowling, Doust, McIntosh, and Yagi [1996], whose proof we follow. The
result proved in this paper shows that ωH ∞ (A) = max{ω(A), ωBIP (A)}; as
was noted in the main text, this improves to ωH ∞ (A) = ωBIP (A) by virtue
of the Clément–Prüss Theorem 15.3.12. A different proof of Theorem 15.3.20,
based on the theory of Euclidean structures, is presented by Kalton, Lorist,
and Weis [2023]. Theorem 15.3.21, on the equivalence of bounded H ∞ -calculus
and γ-bounded imaginary powers, is taken from Kalton and Weis [2016]. An
alternative proof is presented by Kalton, Lorist, and Weis [2023, Theorem
4.5.6 and Corollary 4.5.7]
An example of a sectorial operator on the space c0 without bounded
imaginary powers was was given by Komatsu [1966]. Hilbert space examples
were constructed subsequently by McIntosh and Yagi [1990] and Baillon and
Clément [1991], where a general way to construct such examples using con-
ditional bases was invented. Venni [1993] showed that, in any Banach space
with a Schauder basis, there are densely defined sectorial operators A for
which some, but not all, imaginary powers are bounded. More precisely, it
can be arranged that Aikπ = I if k is an even integer and Aikπ is unbounded
if k is an odd integer. Hieber [1996] constructed an example of a pseudo-
differential operator acting in Lp (R), 1 < p < ∞, p 6= 2, that is sectorial but
does not admit bounded imaginary powers. Examples of operators in Lp (S),
p 6= 2, with bounded imaginary powers but without a bounded H ∞ -calculus
can be found in Cowling, Doust, McIntosh, and Yagi [1996]. This reference
also contains the proof of the inequality ωH ∞ (A) 6 ωBIP (A).
Theorem 15.3.23 is due to McIntosh [1986]. In this connection, it is also
of interest to mention the result of Yagi [1984, Theorem B] that an invertible
sectorial operator A on a Hilbert space has bounded imaginary powers if
D(Aα ) = D(A?α ) for all α ∈ [0, ).
504 15 Extended calculi and powers of operators

Around almost sectoriality: further results

The next result, a proof of which is given by Kalton, Lorist, and Weis [2023,
Proposition 4.2.4], connects the (almost) γ-sectoriality of A to γ-boundedness
of the associated semigroup.

Proposition 15.6.1. Let A be a sectorial operator on X with ω(A) < 21 π and

let ω(A) < σ < 21 π. Then
(i) A is γ-sectorial with ωγ (A) 6 σ if and only if the set
 −zA
e : z ∈ Σν

is γ-bounded for all 0 < ν < 21 π − σ;

(ii) A is almost γ-sectorial with ω
eγ (A) 6 σ if and only if the set

zAe−zA : z ∈ Σν


is γ-bounded for all 0 < ν < 21 π − σ.

For operators A that are diagonal with respect to a Schauder basis, the notion
of γ-almost sectoriality captures a natural property of the basis. In order to
formulate this in the form of a proposition, we first recall that a sequence
P basis of X if every x ∈ X
(xn )n∈Z in a Banach space X is called a Schauder
has a unique representation of the form x = n∈Z cn xn . Associated with a
Schauder basis is its sequence of coordinate projections (Pn )n∈Z on X, defined
by X
Pn cj xj := xn , n ∈ Z,
j∈Z

and the sequence of partial sum projections (Sn )n∈N , defined by

X n
X
Sn xj := xj , n ∈ N,
j∈Z j=−n

Pn
that is, Sn = k=−n Pk . For any Schauder basis, the set of partial sum
projections is uniformly bounded, and by taking differences the same is seen
to be true for the set of coordinate projections.
On a Banach space X with a Schauder basis (xn )n∈Z , we may consider
the diagonal operator A defined by

Axn := 2n xn , xn ∈ Xn , n ∈ Z,

with its natural maximal domain. It was shown in Proposition 10.2.28 that A
is sectorial of angle ω(A) = 0, and that A has a bounded H ∞ -calculus if and
only if (xn )n∈Z is unconditional. The following result is due to Kalton, Lorist,
and Weis [2023, Proposition 6.1.3].
 15.6 Notes 505

Proposition 15.6.2. For the operator A just defined, the following is true:
(1) A is γ-sectorial if and only if the sequence (Sn )n∈N is γ-bounded;
(2) A is almost γ-sectorial if and only if the sequence (Pn )n∈Z is γ-bounded.

We revisit this result in the Notes of Chapter 17 in connection with the prob-
lem of finding examples of sectorial operators that are not R-sectorial.

Around the Hilbert space characterisation: the γ-interpolation method

Theorem 15.3.23 asserts that a standard sectorial operator A on a Hilbert

space H has a bounded H ∞ -calculus if and only if it has bounded imaginary
powers. This equivalence is nothing but the specialisation to Hilbert spaces of
the equivalence, for any standard sectorial operator A on a Banach space X, of
having a bounded H ∞ -calculus and having γ-bounded imaginary powers, as
stated in Theorem 15.3.21. The aim of this section is to explain that also the
third equivalence in Theorem 15.3.23 is the specialisation to Hilbert spaces of a
corresponding statement for Banach spaces. Recall that this third equivalence
states, under the additional assumptions that 0 ∈ %(A) and ω(A) < 21 π,
that bounded H ∞ -calculus and boundedness of imaginary powers for A are
equivalent to the equality

D(A1/2 ) = (X, D(A))1/2,2 with equivalent norms. (15.36)

The Banach space version of this equivalence, due to Kalton, Lorist, and Weis
[2023, Corollary 5.3.9], replaces (15.36) with the condition

D(A1/2 ) = (X, D(A))γ1/2 with equivalent norms,

the interpolation space on the right-hand side being obtained via the so-called
γ-interpolation method which we briefly outline next.
A discrete version of the γ-interpolation method was already considered
by Kalton, Kunstmann, and Weis [2006], where Rademacher variables were
used instead of Gaussian variables. In that paper, the method was used to
study perturbations of the H ∞ -calculus for various differential operators. The
continuous version of the Gaussian method was introduced by Suárez and
Weis [2006, 2009], where Gaussian interpolation of Bochner spaces Lp (S; X)
and square function spaces γ(S; X), as well as a Gaussian version of ab-
stract Stein interpolation, was studied. An abstract framework covering the
γ-interpolation method, as well as the real and complex interpolation methods,
has been recently developed by Lindemulder and Lorist [2021]. The present
treatment follows the memoir of Kalton, Lorist, and Weis [2023]; theorem
numbers in brackets refer to this memoir. As was pointed out in this refer-
ence, the results in Kalton, Kunstmann, and Weis [2006] and Suárez and Weis
[2006, 2009] were based on a draft version of the memoir.
Let (X0 , X1 ) be an interpolation couple of Banach spaces and let θ ∈ (0, 1).
We call an operator
506 15 Extended calculi and powers of operators

T : L2 (R) + L2 (R, e−2t dt) → X0 + X1

admissible, and write T ∈ A , if T ∈ γ(L2 (R, e−2jt dt), Xj ) for j = 0, 1. For

such operators we define

kT kA := max kTj kγ(L2 (R,e−2jt dt),Xj ) ,

j=0,1

where Tj denotes the operator T from L2 (R, e−2jt dt) into Xj . It is routine to
check that A is complete with respect to this norm.
Denoting eθ : t 7→ eθt , we define (X0 , X1 )θγ as the space of all x ∈ X0 + X1
for which the quantity

kxk(X0 ,X1 )θγ := inf kT kA : T ∈ A , T (eθ ) = x



is finite. This space is a quotient space of A , and as such it is a Banach space.

By [Proposition 3.3.2], the set of finite rank operators T is dense in A ; in
particular, we have:

Proposition 15.6.3. X0 ∩ X1 is dense in (X0 , X1 )γθ .

[Proposition 3.3.1] gives the following interpolation result.

Theorem 15.6.4 (γ-interpolation of operators). Suppose that (X0 , X1 )

and (Y0 , Y1 ) are interpolation couples of Banach spaces. Let S : X0 + X1 →
Y0 + Y1 be a bounded operator such that S(X0 ) ⊆ Y0 and S(X1 ) ⊆ Y1 . Then
S maps (X0 , X1 )θγ to (Y0 , Y1 )γθ boundedly, with norm
1−θ θ
kSkL ((X0 ,X1 )γθ ,(Y0 ,Y1 )γθ ) 6 kSkL (X0 ,Y0 ) kSkL (X1 ,Y1 ) .

By [Proposition 3.4.1], the norm of (X0 , X1 )γθ can be equivalently expressed

as follows.

Proposition 15.6.5. Let A• be the set of all strongly measurable functions

f : R+ → X0 ∩ X1 such that t 7→ tj f (t) belongs to γ(R+ , dt
t ; Xj ) for j = 0, 1.
For f ∈ A• , define

kf kA• := max t 7→ tj f (t) γ(R+ , dt .

j=0,1 t ;Xj )

Then for all x ∈ (X0 , X1 )γθ we have

n Z ∞ o
kxk (X0 ,X1 )γ
θ
= inf kf kA• : f ∈ A• , tθ f (t) dt = x ,
0

where the integral converges in the Bochner sense in X0 + X1 .

[Theorem 3.4.4] contains the following relationship of the γ-interpolation

method with the real and complex methods.
 15.6 Notes 507

Theorem 15.6.6 (Relationship with the real and complex method).

Let (X0 , X1 ) be an interpolation couple of Banach spaces, and let 0 < θ < 1.
(1) If X0 and X1 have type p0 , p1 ∈ [1, 2] and cotype q0 , q1 ∈ [2, ∞] respectively,
then we have the continuous embeddings

(X0 , X1 )θ,p ,→ (X0 , X1 )θγ ,→ (X0 , X1 )θ,q

where p1 = 1−θ θ 1 1−θ θ

p0 + p1 and q = q0 + q1 .
(2) If X0 and X1 have type 2, then we have the continuous embedding

[X0 , X1 ]θ ,→ (X0 , X1 )γθ .

If X0 and X1 have cotype 2, then we have the continuous embedding

(X0 , X1 )γθ ,→ [X0 , X1 ]θ .

Since a Banach space X is isomorphic to a Hilbert space if and only if X has

type 2 and cotype 2 (by Kwapień’s Theorem 7.3.1), we obtain the corollary
(cf. the Corollary of Peetre’s Theorem C.4.1 for the first equivalence):
Corollary 15.6.7. Let (H0 , H1 ) be an interpolation couple of Hilbert spaces,
and let 0 < θ < 1. Then

(H0 , H1 )θ,2 = [H0 , H1 ]θ = (H0 , H1 )γθ

with equivalent norms.

Section 15.4

Much of the theory developed in the first three sections of this chapter has an
analogue for strip type operators. The general theory of this class of operators
is developed in the book of Haase [2006], which also treats their connections
with logarithms of sectorial operators. Analogues of the results of Sections
10.3 and 10.4 are presented by Kalton and Weis [2016].
Theorem 15.4.3, on the strip type property and integral representation of
the logarithm log(A) of a standard sectorial A, is due to Nollau [1969]. Our
proof is a variation of the presentation by Haase [2006]. The converse problem
to Theorem 15.4.3, whether the exponent of a striptal operator is sectorial, is
subtle; we refer to Haase [2006] for a counterexample. Theorem 15.4.4, on the
identification of C0 -groups on a UMD space as bounded imaginary powers
of a standard sectorial operator, can be viewed as a partial result in the
positive direction. It was first proved by Monniaux [1999] with a very different
argument based on the notion of analytic generator. The proof presented here
is essentially that of Haase [2009]. Another proof can be found in Haase [2007].
Theorem 15.4.11 about the sum of operators, both of which have bounded
imaginary powers, is due to Dore and Venni [1987] under the slightly stronger
assumption on A and B that they satisfy the resolvent bounds
508 15 Extended calculi and powers of operators

M
k(t + A)−1 k, k(t + B)−1 k 6 , t > 0.
1+t
In its present form, where it is only assumed that
M
k(t + A)−1 k, k(t + B)−1 k 6 , t > 0,
t
the result was obtained by Prüss and Sohr [1990]. The original proof of Dore
and Venni [1987] is ingenious and relatively short, and has been sketched in
the Notes of Chapter 5. It depends on a representation formula for (A + B)−1
in terms of fractional powers of A and B. The refinement of these arguments
by Prüss and Sohr [1990], to obtain the more general case, depends on subtle
approximation arguments for operators A with bounded imaginary powers
which, like the proof presented here, use the functional calculus associated
with the C0 -group (Ait )t∈R and Mellin transform techniques.
The beautiful proof of the Dore–Venni Theorem 15.4.11 presented here
is due to Haase [2007] and fits well in the mainstream of ideas developed in
this volume. This paper also contains our proof of Theorem 15.4.4, which is
originally due to Monniaux [1999] with a different proof based on the notion
of an analytic generator. Our presentation uses some ideas of Haase [2006,
Section 4.2], where a detailed presentation of the theory of strip type operators
if given. With these methods, the operator B = etA can also be defined using
the extended Dunford calculus.
The importance of the Dore–Venni Theorem 15.4.11 is mostly historical,
and the more recent sum-of-operator theorems proved in the next chapter
have turned out to be more versatile in their usage. It is for this reason that
we have contented ourselves with a somewhat sketchy presentation, leaving a
few tedious details to the reader.

Section 15.5

The results of this section follow Duelli [2005] and Duelli and Weis [2005],
where Theorem 15.5.2 (k(A2 )1/2 xk h kAxk) is proved. By the Hieber–Prüss
Theorem 10.7.10, it covers the case where iA generates a bounded C0 -group.
A version of Theorem 15.5.2 (with inhomogeneous estimates) for the case that
iA generates a C0 -group of exponential growth type ω > 0 is due to Haase
[2007].
The spectral projections P ± of Proposition 15.5.1 are studied in more de-
tail by Arendt and Zamboni [2010], Duelli [2005], and Duelli and Weis [2005].
In particular, Arendt and Zamboni [2010] show that, if A is an invertible bi-
sectorial operator and δ > 0 is so small that the closure of the ball B(0, δ)
belongs to the resolvent set of A, then for x ∈ D(A), these projections are
given by
Z
1 dz
P ±x = R(z, A)Ax ,
2πi ∂(Σν± \B(0,δ)) z
 15.6 Notes 509

where ω bi (A) < ν < σ is arbitrary. The extended Dunford calculus for bisec-
torial operators, in particular the analogue of Proposition 15.1.12, which was
used in the proof of Theorem 15.5.2, has been studied by Duelli [2005].

The Kato square root problem

A long-standing question about fractional powers of operators, and a major

motivation for the development of the theory of H ∞ -calculus at large, was the
square root problem of Kato [1961]. To present this problem, we recall that a
linear operator A in a Hilbert space H with inner product ( | ) is called
• accretive, if <(Au|u) > 0 for all u ∈ D(A);
• maximal accretive, if an extension Ae ⊇ A is accretive exactly when Ae = A;
• regularly accretive, if A is maximally accretive and there is an associated
sesquilinear form a in H such that <a(v, v) > 0 for all v ∈ D(a), and

(Au|v) = a(u, v) for all u ∈ D(A) ⊆ D(a) and all v ∈ D(a).

For a regularly accretive operator, Kato [1961] defines its real part <A as the
unique maximal accretive operator associated with the sesquilinear form
1
<a : (u, v) 7→ [a(u, v) + a(v, u)]
2
in the above sense. He then proceeds to show that

D(Aα ) = D((A∗ )α ), if A is maximal accretive and α ∈ [0, 21 ),

= D((<A)α ), if A is regularly accretive and α ∈ [0, 21 ),

and that these identities can fail for α > 21 , “but it is not known whether or
not α = 21 can be included”. Kato [1961, Remark 1] writes:
This is perhaps not true in general. But the question is open even
when A is regularly accretive. In this case it appears reasonable to
1 1 1
suppose that both D(A 2 ) and D((A∗ ) 2 ) coincide with D((<A) 2 ) =
D(a), where <A is the real part of A and a is the regular sesquilinear
form which defines A.
As suspected by Kato [1961], a counterexample to the general case of maximal
accretive operators was found shortly after by Lions [1962], but the regularly
accretive case was only disproved a decade later by McIntosh [1972].
What came to be known as Kato’s square root problem was subse-
quently redefined by McIntosh [1982], making the case that, what Kato [1961]
“really had in mind”, was differential operators A = − div B(x)∇, where
B ∈ L∞ (Rd ; Cd ) is such that <(B(x)e|e) > δ > 0 for a.e. x ∈ Rd and all
e ∈ Cd of norm one. For such A, the associated sesquilinear form is
Z
a(u, v) = (B(x)∇u(x)|∇v(x)) dx
Rd
510 15 Extended calculi and powers of operators

with domain D(a) = D(∇) = W 1,2 (Rd ), and the problem thus takes the form
√ ?
k − div B∇ukL2 (Rd ) h k∇ukL2 (Rd ;Cd ) . (15.37)

McIntosh [1982] further suggested that this question was related to Calderón’s
problem about the L2 -boundedness of the Cauchy integral on a Lipschitz
graph (discussed in the Notes of Chapter 12). As pointed out by Alan McIn-
tosh in several oral communications with the authors of this book over the
first decade of this century (the quote within the first sentence of this para-
graph is also from this oral source), before his making this connection, Kato’s
question was generally regarded as being a soft one, several levels easier than
the problem of Calderón, which everyone agreed to be hard. Nevertheless,
the connection suggested by McIntosh [1982] turned out to be a fruitful one,
and the combined efforts of Coifman, McIntosh, and Meyer [1982] led to a
proof of both the boundedness of the Cauchy integral and, what turned out
to be equivalent, McIntosh’s reformulation of Kato’s square root problem in
dimension d = 1.
After this, the status of the redefined square root problem increased sub-
stantially, and important progress was made by Coifman, Deng, and Meyer
[1983], Fabes, Jerison, and Kenig [1985], McIntosh [1985], Alexopoulos [1991],
Journé [1991], Auscher and Tchamitchian [1998], and Auscher, Hofmann,
Lewis, and Tchamitchian [2001], but it took two decades from the one-
dimensional result of Coifman, McIntosh, and Meyer [1982] until a complete
solution was achieved by Hofmann, Lacey, and McIntosh [2002] in dimension
d = 2 and then by Auscher, Hofmann, Lacey, McIntosh, and Tchamitchian
[2002] in all dimensions.
While heavily building on ideas and results about functional calculus of
the second-order operator A = − div B∇, the original solution of the square
root problem was not quite a “pure” functional calculus theorem in the sense
that the gradient featuring in (15.37) does not have the form f (A) of objects
in the functional calculus of A. This “issue” was fixed by a new approach
developed by Axelsson, Keith, and McIntosh [2006] which, in contrast to the
sectorial calculus of second-order operators employed by Auscher et al. [2002],
was based on the bi-sectorial calculus of first-order differential operators, and
promoted the relevance of bi-sectorial operators and bi-sectorial H ∞ -calculus
in subsequent research. Quoting the MathSciNet review of Axelsson et al.
[2006] by Ian Doust, this paper provided “a remarkable consolidation of many
of the ideas that have arisen in the so-called Calderón program”, not only
reproving the square root theorem of Auscher et al. [2002] and several other
results by a unified approach, but also obtaining new geometric applications
to the behaviour of the Hodge–Dirac operator on a Riemannian manifold
under measurable perturbations of the Riemannian metric. In fact, the very
framework of Axelsson et al. [2006] is based on a general notion of perturbed
Hodge–Dirac operators; in the application to the Kato square root problem,
these take the form
 15.6 Notes 511
 
0 − div B
DB := , D(DB ) = D(∇) ⊕ D(div B)
∇ 0

so that, at least formally,

  12  1 
2 21 − div B∇ 0 (− div B∇) 2 0
(DB ) = = 1 ,
0 −∇ div B 0 (−∇ div B) 2
1 1 1
2 2
D((DB ) ) = D((− div B∇) 2 ) ⊕ D((−∇ div B) 2 ).

Hence, if one can establish bounded bi-sectorial H ∞ -calculus of DB , as Ax-

elsson et al. [2006] do, the Kato conjecture (15.37) will be an immediate con-
2 21
sequence of the estimate k(DB ) uk h kDB uk provided by Theorem 15.5.2.
This first-order approach of Axelsson, Keith, and McIntosh [2006] has been
influential for several subsequent developments. Of particular interest for the
themes of the present volumes is a version of the Kato square root theorem
in Lp (Rd ; X). This was obtained by Hytönen, McIntosh, and Portal [2008]
by a Banach space extension of the methods of Axelsson et al. [2006]. In
the language of the original operator A = − div B∇, the result of Hytönen,
McIntosh, and Portal [2008] can be stated as follows:
Theorem 15.6.8. Let X be a UMD space, and suppose that both X and X ∗
have the RMF property (Definition 3.6.10). Let B, B −1 ∈ L∞ (Rd ; L (Cd )),
where B −1 (x) := B(x)−1 is the pointwise inverse of the matrix-valued function
B. Let 1 6 p− < p+ 6 ∞, and suppose that A := − div B∇ is sectorial in
Lp (Rd ; X) for all p ∈ J := (p− , p+ ). Then the following are equivalent:
(1) For all p ∈ J, the following four sets are R-bounded in Lp (Rd ; X):
n √ √ o
Aa,b := (t −∆)a (I + t2 A)−1 (t −∆)b : t > 0 , a, b ∈ {0, 1}.

(2) For all p ∈ J, A has a bounded H ∞ -calculus in Lp (Rd ; X), and

√
k AukLp (Rd ;X) h k∇ukLp (Rd ;X)d

Remark 15.6.9. Several remarks concerning Theorem 15.6.8 are in order:

(a) While (2) contains a full analogue of (15.37) in Lp (Rd ; X), along with the
bounded H ∞ -calculus of independent interest, the characterising condition
(1) is less satisfactory than in the scalar-valued case, as it involves non-
trivial R-boundedness properties of operators on Lp (Rd ; X). However, note
that the R-boundedness of A0,0 is simply the R-sectoriality of A which,
by Theorem 10.3.4(2), is a general necessary condition for the bounded
H ∞ -calculus contained in (2). When X = C and p = 2, verifying (1)
from easy-to-check pointwise conditions on B is straightforward operator
theory, and the difficult harmonic analysis enters in passing from (1) to (2).
Curiously, in the Banach space valued generality, the easy part becomes
unavailable but the difficult part may still be pushed through.
512 15 Extended calculi and powers of operators

(b) Another shortcoming of Theorem 15.6.8 compared to the scalar-valued L2

case is that, in order to get (2) for a given p, one needs to verify (1) for an
open range of exponents (p − ε, p + ε). However, it was subsequently shown
by Hytönen and McIntosh [2010], and later with a different argument by
Auscher and Stahlhut [2013], that conditions of type (1) self-improve from
one exponent p to a small range around it. This allows one to obtain a
version of Theorem 15.6.8 for a fixed p in place of the range of p as stated.
(c) The RMF property (Definition 3.6.10) and the related Rademacher max-
imal function (Definition 3.6.8) were first introduced by Hytönen, McIn-
tosh, and Portal [2008] for the very needs of running the argument to prove
Theorem 15.6.8, but these notions (or their extensions) have proven to be
useful in other contexts, notably in the study of Banach space valued mul-
tilinear operators by Di Plinio and Ou [2018], Di Plinio, Li, Martikainen,
and Vuorinen [2020b], and Amenta and Uraltsev [2020].
(d) For Lp -estimates related to Kato’s square root problem in the scalar-valued
case, there is an alternative approach based on a generalisation of the
Calderón–Zygmund theory discussed in Chapter 11, extrapolating from
the L2 -results of Auscher, Hofmann, Lacey, McIntosh, and Tchamitchian
[2002]. The operators now under consideration are far less regular than
those covered in Chapter 11, and the extrapolation yields their bounded-
ness, in general, only in some subinterval (p− , p+ ) ⊆ (1, ∞) determined by
the details of the operator in question. Based on an extrapolation theory
for “non-integral operators” developed by Blunck and Kunstmann [2003],
a systematic investigation of the maximal ranges of p for various Lp esti-
mates related to the Kato square root problem is carried out by Auscher
[2007].
(e) Yet another approach to the scalar-valued Lp theory is due to Frey, McIn-
tosh, and Portal [2018]. As in the approach of Hytönen, McIntosh, and Por-
tal [2008] and in contrast to that of Auscher [2007], they work directly in
Lp instead of extrapolating from L2 ; also their “proof shows that the heart
of the harmonic analysis in L2 extends to Lp for all p ∈ (1, ∞), while the
restrictions in p come from the operator-theoretic part of the L2 proof”.
A novelty in their approach is using the theory of tent spaces; on the
side of the results, this allows them to dispense with the R-boundedness
assumptions required by Hytönen, McIntosh, and Portal [2008].

Given the focus of these volumes on analysis in Banach spaces, we have

not covered, in the discussion above, the rich literature of extensions and
applications of the machinery of Auscher, Hofmann, Lacey, McIntosh, and
Tchamitchian [2002] and Axelsson, Keith, and McIntosh [2006] in the L2 the-
ory of partial differential operators and equations; for this, we refer the reader
to the numerous papers citing these pioneering contributions. The first-order
approach to the Kato square root problem of Axelsson, Keith, and McIntosh
[2006] has been adapted in Maas and Van Neerven [2009] to the Gaussian set-
ting to prove a nonsymmetric version of the Meyer inequalities in Malliavin
 15.6 Notes 513

calculus. This work belongs to a circle of ideas that will be treated in Volume
IV.
16
Perturbations and sums of operators

In this chapter we address a couple of topics in the theory of H ∞ -calculus

centering around the question what can be said about an operator of the form
A+B when A and B have certain “good” properties such as being (R-)sectorial
or admitting a bounded H ∞ -calculus. The chapter is divided into two sections.
The first considers the case where B is “smaller” than A in certain ways, and
the second considers the case where A and B are essentially on equal footing.
The results of this chapter play an important role in applications as well in
the further development of the abstract theory and will be needed in our
treatment, in the next to chapter, of the maximal regularity problem.

16.1 Sums of unbounded operators

In general it is a rather delicate problem to give a meaning to the operator

sum A + B when A and B are unbounded operators acting in a Banach space
X. The simplest approach is to define

D(A + B) := D(A) ∩ D(B),

(16.1)
(A + B)x := Ax + Bx, x ∈ D(A + B),

but in concrete cases this definition may be vacuous due to the possibility
that D(A) ∩ D(B) could be unreasonably small or even trivial, i.e., equal to
{0}. Various methods to deal with this problem have been developed, such
as the method of forms. In the context of evolution equations, the two prime
applications one has in mind are cases where either A is the linear operator
governing the equation, e.g., a linear differential operator in the space vari-
ables, and B is the derivative with respect to time, or both A and B are
differential operators in the space variable, typically with B being of lower
order than A. In both of these cases, the resolvent operators R(λ, A) and
R(µ, B) commute and D(A) ∩ D(B) is “large”, in that it contains all elements
of the form R(λ, A)R(µ, B)x with x ∈ X. In fact we have the following result.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 515
T. Hytönen et al., Analysis in Banach Spaces, Ergebnisse der Mathematik und ihrer
Grenzgebiete. 3. Folge / A Series of Modern Surveys in Mathematics 76,
https://doi.org/10.1007/978-3-031-46598-7_6
516 16 Perturbations and sums of operators

Proposition 16.1.1. If A and B are sectorial operators acting in X whose

resolvents commute, then D(A) ∩ D(B) is dense in both D(A) and D(B)

Proof. As a consequence of the resolvent commutation and Proposition 10.1.7

we have
lim λR(λ, A)R(µ, B)x = R(µ, B)x in D(B)
λ→−∞

for all µ ∈ %(B) and

lim µR(λ, A)R(µ, B)x = R(λ, A)x in D(A)

µ→−∞

for all λ ∈ %(A) 

It is for this reason that we will stick to the somewhat naive approach em-
bodied in (16.1); the operator sum A + B will always be understood as given
in this way.
Let us briefly clarify the meaning of the term ‘resolvent commutation’ used
in the above proposition. If commutation identity

R(λ, A)R(µ, B) = R(µ, B)R(λ, A)

holds for some λ ∈ %(A) and µ ∈ %(B), then it holds for all λ0 ∈ %(A)
and µ0 ∈ %(B) in the connected components of %(A) containing λ and µ,
respectively. This is an easy consequence of the Taylor series identities
∞
X
0
R(λ , A) = (λ − λ0 )n R(λ, A)n+1 ,
n=0
X∞
R(µ0 , B) = (µ − µ0 )n R(µ, B)n+1 ,
n=0

which follow from repeated application of the resolvent identity (see Section
10.1.b). The following definition then suggests itself naturally:

Definition 16.1.2 (Resolvent commutation). The sectorial operators A

and B are said to resolvent commute when

R(λ, A)R(µ, B) = R(µ, B)R(λ, A)

holds for some (or equivalently, all) λ, µ in the connected set {Σσ ∩ {Στ for
some (or equivalently, all) ω(A) < σ < π and ω(B) < τ < π.

16.2 Perturbation theorems

When it comes to checking the boundedness of the H ∞ -calculus of concrete
operators, in particular elliptic differential operators, perturbation theorems
 16.2 Perturbation theorems 517

are often the method of choice. Perturbation arguments compare a “compli-

cated” operator with a more “basic” operator such as the Laplace operator or
an elliptic operator with constant coefficients. In order to cover a multitude of
concrete situations we phrase these perturbation arguments in the framework
of sectorial operators and their scale of fractional domain spaces. The case of
lower-order perturbations of the form
L = A + B with B : D(Aα ) → X for some 0 < α < 1
(Theorem 16.2.7) is readily obtained from the corresponding theorem about
relatively bounded perturbations of the form
L = A + B with kBxk 6 δkAxk for small δ > 0
(Theorem 16.2.3). In contrast to sectoriality, boundedness of the H ∞ -funct-
ional calculus is not preserved under small relatively bounded perturbations,
unless additional relative boundedness assumptions are made with respect
to the fractional domains (Example 16.2.10 and Theorem 16.2.8). Analogous
perturbation theorems for R-sectorial operators are proved as well.
Because of their importance in applications, in particular for the study
of non-linear evolution equations, the literature on perturbation theorems is
extensive. We can present only a representative selection of such theorems
and some model applications serving as illustrations. Variants and extensions
of these results, in particular to elliptic operators and pseudo-differential op-
erators, will be discussed in the Notes.
We next introduce some notation which will be used throughout this chap-
ter and the next ones. Recalling from Definition 10.1.1 that an operator A is
called σ-sectorial if the set {λ 6= 0, | arg(λ)| > σ} is contained in the resolvent
set %(A) and
sup kλR(λ, A)k < ∞,
λ6=0, | arg(λ)|>σ

we define
Mσ,A := sup{kλR(λ, A)k : λ 6= 0, | arg(λ)| > σ},
fσ,A := sup{kAR(λ, A)k : λ 6= 0, | arg(λ)| > σ}.
M
When A is σ-R-sectorial (the definition being similar), for p ∈ [1, ∞) we set
fRp := Rp ({λR(λ, A) : λ 6= 0, | arg(λ)| > σ}),
M σ,A
fRp := Rp ({AR(λ, A) : λ 6= 0, | arg(λ)| > σ}),
M σ,A

where Rp (T ) denote the R-bound with exponent p (see Remark 8.1.2).

16.2.a Perturbations of sectorial operators

To set the stage for the results to follow, we begin with an elementary per-
turbation result for sectorial operators.
518 16 Perturbations and sums of operators

Proposition 16.2.1. If A is an σ-sectorial operator on X and B ∈ L (X) is

bounded, then for all λ0 > M kBk the operator λ0 + A + B is σ-sectorial.
Proof. Set M := Mσ,A for brevity. Fix a non-zero λ ∈ C with | arg(λ)| > σ.
Then λ ∈ %(A) and kR(λ, A)k 6 M/|λ|. Because
(λ − (A + B)) = (I − BR(λ, A))(λ − A)
and kBR(λ, A)k 6 M kBk/|λ|, for |λ| > M kBk the operator I − BR(λ, A) is
invertible. For such λ it follows that λ ∈ %(A + B) and
∞
X
R(λ, A + B) = R(λ, A) [BR(λ, A)]n
n=0

by the Neumann series. This gives the bound

M 1 M
kR(λ, A + B)k 6 = ,
|λ| 1 − M kBk/|λ| |λ| − M kBk
valid for non-zero λ ∈ C satisfying | arg(λ)| > σ and |λ| > M kBk. Shifting
A + B over λ0 > M kBk, the result follows from this. 
The following lemma describes a useful technique that will enable us to deal
with lower-order and relatively bounded perturbations.
Lemma 16.2.2 (The method of continuity). Let E and F be Banach
spaces. Let (Lt )t∈[0,1] be a family of bounded linear operators from E into F
such that t 7→ Lt is continuous from [0, 1] into L (E, F ). Suppose furthermore
that there exists a constant C > 0 such that for all t ∈ [0, 1] and all x ∈ E we
have
kxk 6 CkLt xk.
Then L0 is surjective if and only if L1 is surjective.
Proof. Since [0, 1] is compact, t 7→ Lt is uniformly continuous. Therefore we
ε
can find δ > 0 such that |t − s| < δ implies kLt − Ls k 6 2C .
The assumption of the lemma imply that the operators Lt are injective.
Now suppose that Ls is invertible for a given s ∈ [0, 1]. We will show that Lt
is invertible for all t ∈ [0, 1] satisfying |t − s| < δ. Clearly, this implies the
required result by an iteration argument.
Fix f ∈ F and let T : E → E be the mapping given by T (x) = y, where
y ∈ E is the unique solution to Ls y = f + Ls x − Lt x. We claim that T is a
uniform contraction. Indeed, by the assumed a priori estimate,
kT (x1 ) − T (x2 )k = ky1 − y2 k 6 CkLs y1 − Ls y2 k
Since Ls y1 − Ls y2 = (Ls − Lt )(x1 − x2 ) we obtain
1
kT (x1 ) − T (x2 )k 6 CkLs − Lt k kx1 − x2 k 6 kx1 − x2 k.
2
This proves the claim. By the Banach fixed point theorem, T has a unique
fixed point x. It follows that Ls x = f + Ls x − Lt x, and hence Lt x = f . 
 16.2 Perturbation theorems 519

As a first application of this lemma we prove the following result on relatively

bounded perturbations of sectorial operators.

Theorem 16.2.3 (Relatively bounded perturbations of sectorial op-

erators). Let A be an σ-sectorial operator, and let B : D(A) → X be a linear
operator that satisfies

kBxk 6 δkAxk + Kkxk, x ∈ D(A), (16.2)

where K > 0 and δ ∈ (0, 1) satisfies δ M

fσ,A < 1. Then the operator A + B
with domain D(A + B) := D(A) is closed, and the following assertions hold:
(1) For all λ ∈ R large enough, λ + A + B is σ-sectorial.
(2) If (16.2) holds with K = 0, then A + B is σ-sectorial.

Proof. Observe that for all x ∈ D(A),

kAxk 6 k(A + B)xk + kBxk 6 k(A + B)xk + δkAxk + Kkxk. (16.3)

Therefore, (1 − δ)kAxk 6 k(A + B)xk + Kkxk. By a routine argument, (16.2)

and (16.3) imply that A + B is closed.
We will prove both assertions at the same time by showing that λ0 +A+B
is sectorial for any fixed λ0 > 0 large enough, permitting λ0 = 0 if (16.2) holds
with K = 0.
Fix λ ∈ λ0 + Σσ . We will apply Lemma 16.2.2 to E = D(A), F = X, and
the operators Lt : D(A) → X given by

Lt x := (λ + A + tB)x, t ∈ [0, 1],

where D(A) will be equipped with the equivalent norm

|||x||| = k(λ − λ0 )xk + kAxk.

We first prove the following a priori estimate: For all λ0 > 0 large enough
there exists a constant C > 0 such that

|||x||| 6 CkLt xk, x ∈ D(A), t ∈ [0, 1]. (16.4)

Let x ∈ D(A) and set y := Lt x. Then (λ + A)x = y − tBx. Multiplying this

identity with A(λ + A)−1 on both sides and using (16.2), we obtain

fσ,A kyk + M
kAxk 6 M fσ,A kyk + M
fσ,A kBxk 6 M fσ,A δkAxk + M
fσ,A Kkxk.

fσ,A δ < 1, it follows that

Since M

kAxk 6 C0 kyk + C0 Kkxk, (16.5)

where C0 = M fσ,A δ)−1 . To estimate kxk, writing λx = y − tBx − Ax

fσ,A (1 − M
we find that
520 16 Perturbations and sums of operators

|λ|kxk 6 kyk + kBxk + kAxk

6 kyk + (δ + 1)kAxk + Kkxk 6 C1 kyk + C2 Kkxk

where C1 := 1 + (δ + 1)C0 and C2 := (δ + 1)C0 + 1, so that

C1
kxk 6 kyk =: Dkyk,
|λ| − C2 K

provided we take λ0 > C2 K sufficiently large (in order that |λ| > C2 K). Such
choices of λ0 imply that |λ − λ0 | 6 Cσ |λ| and, together with (16.5),

|||x||| = kAxk + k(λ − λ0 )xk 6 C0 kyk + C0 Kkxk + |λ − λ0 |kxk

6 C0 kyk + C0 Kkxk + Cσ (C1 kyk + C2 Kkxk)
6 Ckyk = CkLt xk

where C := (C0 + Cσ C1 ) + (C0 + Cσ C2 )DK, which is (16.4). Scrutinising the

proof, we see that λ0 = 0 can be allowed if (16.2) holds with K = 0.
Since L0 = λ + A is surjective, Lemma 16.2.2 gives that L1 = λ + A + B
is surjective, and hence boundedly invertible by (16.4). Also by (16.4), for all
y ∈ X and λ ∈ λ0 + Σπ−σ (where we may take λ0 = 0 if K = 0),

k(λ − λ0 )(λ + A + B)−1 yk 6 kλ(λ + A + B)−1 yk 6 Ckyk,

which proves λ0 + A + B is σ-sectorial. 

Theorem 16.2.4 (Relatively bounded perturbations of R-sectorial

operators). Let A be σ-R-sectorial, and suppose that B : D(A) → X is
a linear operator which satisfies

kBxk 6 δkAxk + Kkxk, x ∈ D(A), (16.6)

fRp < 1 for some p ∈ [1, ∞). Then the

where K > 0 and δ ∈ (0, 1] satisfies δ M σ,A
operator A + B with domain D(A + B) := D(A) is closed, and the following
assertions hold:
(1) For all λ ∈ R large enough, λ + A + B is σ-R-sectorial.
(2) If (16.6) holds with K = 0, then A + B is σ-R-sectorial.

Proof. The method of proof is similar to that of Theorem 16.2.3. Again we

will prove both assertions at the same time. Let (Ω, P) be a probability space
supporting a Rademacher sequence (εn )n>1 . For notational convenience we
write k · kp = k · kLp (Ω;X) . We will show that λ0 + A + B is R-sectorial for all
λ0 > 0 large enough, and that we may take λ0 = 0 if (16.6) holds with K = 0.
The assumptions of the theorem imply those of Theorem 16.2.3, and there-
fore A + B satisfies its conclusions. It remains to prove the R-boundedness of
the set
(λ − λ0 )(λ + A + B)−1 : λ 6= 0, λ ∈ λ0 + Σπ−σ .

 16.2 Perturbation theorems 521

To this end let n > 1, non-zero λ1 , . . . , λn ∈ λ0 + Σπ−σ , and y1 , . . . , yn ∈ X be

arbitrary and fixed. Let xj ∈ X be the unique solution to (λj + A + B)xj = yj
for each j ∈ {1, . . . , n}. It suffices to show that there is a constant C > 0 such
that
n
X n
X
εj (λj − λ0 )xj 6C ε j yj .
p p
j=1 j=1

Since Axj = A(λj + A)−1 [yj − Bxj ], the R-sectoriality of A gives

n
X n
X n
X
εj Axj 6M ε j yj +M εj Bxj
p p p
j=1 j=1 j=1
Xn Xn n
X
6M ε j yj + Mδ εj Axj + MK εj xj ,
p p p
j=1 j=1 j=1

fRp for brevity. Therefore,

where M := M σ,A

n
X n
X n
X
εj Axj 6 C0 ε j yj + C0 K εj x j ,
p p p
j=1 j=1 j=1

where C0 = CM (1 − δM )−1 . Since λj xj = yj − Bxj − Axj , we also find

n
X n
X n
X n
X
εj λj x j 6 ε j yj + εj Bxj + εj Axj
p p p p
j=1 j=1 j=1 j=1
Xn Xn n
X
6 ε j yj +K εj x j + (1 + δ) εj Axj
p p p
j=1 j=1 j=1
n
X n
X
6 C1 ε j yj + C1 K εj x j ,
p p
j=1 j=1
(16.7)

where C1 = 1 + (1 + δ)C0 .
Next we claim that there exist D > 0 and λ0 > 0 such that

|λj − λ0 | 6 D(|λj | − 2C1 K), (16.8)

Writing λj = λ0 + reiφ with |φ| < π − σ, (16.8) can be equivalently written as

(r + 2DC1 K)2 6 D2 (λ02 + r2 + 2λ0 r cos φ).

√
If |φ| 6 21 π, then cos φ > 0 and the estimate holds with D = 2 and λ0 =
C1 K. If 21 π < |φ| < π, set δ := 1 + cos φ and note that δ ∈ (0, 1). It then
follows that
522 16 Perturbations and sums of operators

λ20 + r2 + 2λ0 r cos φ = λ20 + r2 − 2λ0 r(1 − δ)

= δ(λ20 + r) + (1 − δ)(λ0 − r)2 > δ(λ20 + r2 )
p
and the estimate holds with D = 8/δ and λ0 = DC1 K. This proves the
claim.
The claim implies |(λj − λ0 )/λj | 6 D and 2C1 DK 6 D|λj |, and therefore
the Kahane contraction principle (see Theorem 6.1.13) implies
n n n
X X λj − λ0 X
εj (λj − λ0 )xj = εj λj x j 6D εj λj x j
j=1
p
j=1
λj p
j=1
p

and
n
X n
X n
X
2C1 DK εj x j 6D εj |λj |xj =D εj λj x j .
p p p
j=1 j=1 j=1

Taking the averages of the last two estimates we obtain

n n n
1 X X X
εj (λj − λ0 )xj 6D εj λj x j − C1 DK εj x j
2 j=1 p
j=1
p
j=1
p

n
X
6 C1 D ε j yj ,
p
j=1

where in the last step we applied (16.7). 

As a simple corollary to the above results we show that the smallness condi-
tions on the constants can be lifted in the case of lower order perturbations.
The notation is as in Appendix C.
Corollary 16.2.5 (Lower order perturbations of (R-)sectorial opera-
tors). Let A be sectorial (resp. R-sectorial) and let θ ∈ (0, 1). If
B : D(Aθ ) → X
is a bounded linear operator, then for all large enough λ ∈ R the operator
λ + A + B is sectorial (resp. R-sectorial) with ω(λ + A + B) 6 ω(A) (resp.
ωR (λ + A + B) 6 ωR (A)).
Proof. It suffices to check the conditions of Theorems 16.2.3 and 16.2.4. For
x ∈ D(A), by the interpolation estimate of Theorem 15.2.8 we obtain
kBxk 6 kBk kxkD(Aθ ) 6 kBk kxk1−θ kxkθD(A) .

Using the inequality a1−θ bθ 6 (1 − θ)a + θb, for all ε > 0 we obtain
1 1
kxk1−θ kxkθD(A) 6 (1 − θ)ε− 1−θ kxk + ε θ kxkD(A) .
The result now follows by combining the estimates and choosing ε > 0 small
enough. 
 16.2 Perturbation theorems 523

The same proof works if one assumes that B : (X, D(A))θ,p → X is a bounded
operator for some θ ∈ (0, 1) and p ∈ [1, ∞], or that B : [X, D(A)]θ → X is a
bounded operator for some θ ∈ (0, 1). A similar remark applies to Theorem
16.2.8 below.

16.2.b Perturbations of the H ∞ -calculus

Having studied perturbations of sectorial and R-sectorial operators, we now

turn to perturbation of the H ∞ -calculus. The first proposition addresses shifts
by a positive multiple of the identity. In certain applications it enables one to
improve “for sufficiently large ν > 0” to “for all ν > 0”.

Proposition 16.2.6 (Perturbation by a multiple of the identity). Let

A be a sectorial operator on X.
(1) If A has a bounded H ∞ (Σσ )-calculus, then A+νI has a bounded H ∞ (Σσ )-
∞ ∞
calculus for all ν > 0, and Mσ,A+ν 6 Mσ,A .
∞
(2) If A + ν0 I has a bounded H (Σσ )-calculus for some ν0 > 0, then A + νI
has a bounded H ∞ (Σσ )-calculus for all ν > 0.

Proof. Assertion (1) is obtained by applying the bounded H ∞ -calculus of

A to the function fν (z) = f (z + ν), noting that fν (A) = f (A + ν); since
kf (·+ν)kH ∞ (Σσ ) 6 kf kH ∞ (Σσ ) , this also gives the bound for the boundedness
constants of the H ∞ -calculi.
For the proof of assertion (2) we fix ν > 0. Writing A+ν = (A+ε)+(ν −ε)
we see that there is no loss of generality in assuming that A is invertible. We
also may assume that 0 < ν < δ, where δ > 0 is to be specified later, for
once we have the converse for such ν the general case follows by repeated
application of the first part of the proposition.
For f ∈ H 1 (Σσ ) ∩ H ∞ (Σσ ) consider
Z Z
1 1
f (λ)R(λ, A + ν0 ) dλ − f (λ)R(λ, A + ν) dλ
2πi Γ 2πi Γ
Z
1
= (ν0 − ν) f (λ)R(λ, A + ν)R(λ, A + ν0 ) dλ
2πi Γ
Z
1
= (ν0 − ν) f (λ)R(λ, A + ν)R(0, A + ν0 ) dλ
2πi Γ
Z
1
+ (ν0 − ν) f (λ)R(λ, A + ν)[R(λ, A + ν0 ) − R(0, A + ν0 )] dλ
2πi Γ
Z
1
= (ν0 − ν)R(−ν0 , A) f (λ)R(λ, A + ν) dλ
2πi Γ
Z
1
− (ν0 − ν) λf (λ)R(λ, A + ν)R(λ − ν0 , A)R(−ν0 , A) dλ.
2πi Γ

If we call the last integral I(f ), we have

524 16 Perturbations and sums of operators
Z
1
f (λ)R(λ, A + ν0 ) dλ
2πi Γ
Z
1
= [I + (ν0 − ν)R(−ν0 , A)] f (λ)R(λ, A + ν) dλ − (ν0 − ν)I(f ).
2πi Γ

For invertible A the operator −AR(−ν0 , A) = I +ν0 R(−ν0 , A) is invertible

as well. Since the set of invertible operators is open in L (X), there exists an
r > 0 so small that I +(ν0 −ν)R(−ν0 , A) is invertible if νkR(−ν0 , A)k < r, i.e.,
if ν < δ := r/kR(−ν0 , A)k. Under this assumption we have the representation
f (A + ν) = [I + (ν0 − ν)R(−ν0 , A)]−1 [f (A + ν0 ) + (ν0 − ν)I(f )].
Hence
kf (A + ν)k 6 [I − (ν0 − ν)R(−ν0 , A)]−1


kf (A + ν0 )k + (ν0 − ν)kI(f )k .
(16.9)
By the assumptions we have kf (A + ν0 )k 6 Ckf kH ∞ (Σσ ) . We estimate the
integral I(f ) by splitting it into Γ1 = Γ ∩ {|λ| 6 1} and Γ2 = Γ ∩ {|λ| > 1}
and using
R(λ − ν0 , A) = (λ − ν0 )−1 [R(λ − ν0 , A)A + I].
This gives
Z
1
I(f ) = λf (λ)R(λ, A + ν)R(λ − ν0 , A) dλ [R(−ν0 , A)]
2πi Γ1
Z
1
+ λf (λ)R(λ, A + ν)(λ − ν0 )−1 R(λ − ν0 , A) dλ [AR(−ν0 , A)]
2πi Γ2
Z
1
+ λf (λ)R(λ, A + ν)R(−ν0 , A)(λ − ν0 )−1 dλ
2πi Γ2
= (I) + (II) + (III).
The integrals (I) and (II) can be estimated by Ckf kH ∞ (Σσ ) with constant C
only depending on A, ν0 , σ. The third can be rewritten with the help of the
resolvent identity and Cauchy’s formula:
Z Z
1 f (λ) 1 f (λ)
(III) = R(−ν0 , A) dλ − R(λ, A + ν) dλ
2πi Γ2 λ − ν0 2πi Γ2 λ − ν0
Z
1 f (λ)
= f (ν0 )R(−ν, A) − R(−ν0 , A) dλ
2πi Γ1 λ − ν0
Z
1 f (λ)
− R(λ, A + ν) dλ.
2πi Γ2 λ − ν0
The two remaining integrals can again be estimated by Ckf kH ∞ (Σσ ) with
constant C only depending on A, ν0 , σ. With (16.9) we arrive at
kf (A + ν)k 6 C 0 kf kH ∞ (Σσ ) , f ∈ H 1 (Σσ ) ∩ H ∞ (Σσ ),
thus completing the proof. 
 16.2 Perturbation theorems 525

We continue with the following result for lower order perturbations.

Theorem 16.2.7 (Lower order perturbations of the H ∞ -calculus). Let

A be a sectorial operator and suppose that B is linear operator in X satisfying

D(Aα ) ⊆ D(B)

and
kBxk 6 akAα xk + bkxk, x ∈ D(A),
for suitable real numbers a, b > 0 and α ∈ (0, 1). If A has a bounded H ∞ (Σσ )-
calculus in X for some ω(A) < σ < π, then A+B +ν has a bounded H ∞ (Σσ )-
calculus in X for all sufficiently large ν > 0.

Proof. By Proposition 16.2.3, for large enough ν > 0 the operator A + B + ν

is sectorial and ω(A + B + ν) 6 ω(A). By taking ν larger if necessary, we may
assume that 0 ∈ %(A + B + ν).
We start from the identity

R(λ, A + B + ν) = R(λ, A + ν) + R(λ, A + B + ν)BR(λ, A + ν)

= R(λ, A + ν) + M (λ),

which may be verified by applying λ − (A + B + ν) on both sides, and where

M (λ) = R(λ, A + B + ν)[B(A + ν)−α ](A + ν)α R(λ, A + ν).

For functions f ∈ H 1 (Σσ ) ∩ H ∞ (Σσ ) this gives the Dunford integral

Z
1
f (A + B) = f (A) + f (λ)M (λ) dλ,
2πi Γη

where the contour Γη = ∂Ση with ω(A) < η < σ is chosen as usual. Near the
origin, the integrand is bounded since we assumed that 0 ∈ %(A + B + ν). For
large values of |λ| the integrand may be estimated pointwise by

kf (λ)M (λ)k 6 M1 M2 |λ|−α kB(A + ν)−α kkf kH ∞ (Σσ ) ,

since
M1 := sup{kλR(λ, A + B + ν)k : | arg λ| = η}
is finite by sectoriality of A + B + ν and

M2 := sup{kλ1−α (A + ν)α R(λ, A + ν)k : | arg λ| = η}

is finite by sectoriality of A + ν and Corollary 15.2.14. It follows that the

integral converges absolutely and its norm is bounded by a constant times
kf kH ∞ (Σσ ) . This completes the proof. 
526 16 Perturbations and sums of operators

Our main perturbation theorem asserts that the H ∞ -calculus of an R-sectorial

operator is preserved under relatively bounded perturbations of the H ∞ -
calculus if we add an additional relative boundedness assumption in the frac-
tional domains scale.
Theorem 16.2.8 (Relatively bounded perturbations of the H ∞ -cal-
culus). Let A be a densely defined sectorial operator with a bounded H ∞ (Σσ )-
calculus and let B be a densely defined τ -R-sectorial operator on X, with
D(A) ⊆ D(B) and 0 ∈ %(A)
and satisfying the relative bound
(i) kBxk 6 C0 kAxk for all x ∈ D(A).
Suppose that at least one of the following two additional relative bounds is also
satisfied:
(ii) there exists an α ∈ (0, 1) such that B maps D(A1+α ) into D(Aα ) and
kAα Bxk 6 C1 kA1+α xk, x ∈ D(A1+α );
(iii) there exists an α ∈ (0, 1) such that
kA−α Bxk 6 C1 kA1−α xk, x ∈ D(A1−α ).
Then, given the constant C1 in (ii) or (iii), there is a small enough constant
C > 0 so that if (i) holds with 0 6 C0 6 C, then A + B has a bounded
H ∞ (Σσ∨τ )-calculus.
fσ∨τ,A , then the condition 0 ∈ %(A) may
If in (ii) or (iii) we have C1 < 1/M
be replaced by the weaker condition that A be injective and B maps D(A1−α )
into D(A−α ).

In the last line of the statement of the theorem, recall the notation M fθ,A =
sup{kAR(λ, A)k : λ 6= 0, k arg(λ)| > θ}.
If X has the triangular contraction property, in particular if X is a UMD
space, then by Theorem 10.3.4 we have ωR (A) 6 ωH ∞ (A) and therefore the
theorem applies.
At the end of the section, an example will be presented which shows that
the additional assumptions (ii) and (iii) cannot be omitted.
We will reduce the theorem to the following technical lemma.
Lemma 16.2.9. Let A be a densely defined sectorial operator with a bounded
H ∞ (Σσ )-calculus and let B a densely defined R-sectorial operator on X. Let
ω(A) < σ < π and ωR (B) < τ < π, and set µ := max{σ, τ }. Suppose there
exists a holomorphic function M : {| arg(λ)| > µ} → L (X) with R-bounded
range and a β ∈ (0, 1) such that
R(λ, B) = R(λ, A) + Aβ R(λ, A)M (λ)A1−β R(λ, A), | arg(λ)| > µ. (16.10)
Then B has a bounded H ∞ (Σµ )-calculus.
 16.2 Perturbation theorems 527

Proof. Our aim is to prove that there exists a function φ ∈ H 1 (Σµ ) and a
constant C > 0 such that for all integers N > 1, all scalars −N , . . . , N of
modulus one, and all t > 0 we have
X
n φ(t2n B) 6 C.
|n|6N

Once we have this, it follows from Proposition 10.4.11 (and tracking angles in
its proof) that B has a bounded H ∞ (Σµ )-calculus.
Let µ < ν < π and consider the function ψν ∈ H 1 (Σµ ) given by

z 1/2
ψν (z) = , z ∈ Σµ ,
(eiν − z)1/2 (2eiν − z)1/2

so that φν := ψν2 satisfies

z 1 2
φν (z) = = iν − iν .
(eiν − z)(2eiν − z) e −z 2e − z

By (16.10),

R(λ, t2n B) = t−1 2−n R(t−1 2−n λ, B)

= t−1 2−n R(t−1 2−n λ, A)
+ t−1 2−n Aβ R(t−1 2−n λ, A)M (t−1 2−n λ)A1−β R(t−1 2−n λ, A)
= R(λ, t2n A)
+ t2n Aβ R(λ, t2n A)M (t−1 2−n λ)A1−β R(λ, t2n A).

By Corollary 15.2.14, the right-hand side has decay of order |λ|−1 as |λ| → ∞
in the complement of Σµ . Hence, by Cauchy’s theorem and taking µ < τ < ν,
Z
1
φν (t2n B) = φν (λ)R(λ, t2n B) dλ
2πi ∂Στ
= φν (t2n A) + t2n Aβ R(eiν , t2n A)M (t−1 2−n eiν )A1−β R(eiν , t2n A)
− t2n+1 Aβ R(2eiν , t2n A)M (t−1 2−n 2eiν )A1−β R(2eiν , t2n A)
= φν (t2n A) + t2n Aβ R(eiν , t2n A)M (t−1 2−n eiν )A1−β R(eiν , t2n A)
− t2n−1 Aβ R(eiν , t2n−1 A)M (t−1 2−(n−1) eiν )A1−β R(eiν , t2n−1 A)
= φν (t2n A) + φβ,ν (t2n A)M (t−1 2−n eiν )φ1−β,ν (t2n A)
− φβ,ν (t2n−1 A)M (t−1 2−(n−1) eiν )φ1−β,ν (t2n−1 A)
= (I) + (II) + (III),
(16.11)

where for α > 0 we define φα,ν ∈ H 1 (Σµ ) by

528 16 Perturbations and sums of operators

zα
φα,ν (z) := .
eiν − z
In the penultimate identity of (16.11) we used the identity

φα,ν (τ A) = τ α Aα R(eiν , τ A),

which follows from Propositions 15.1.12 and 15.2.6.

We estimate the terms (I)–(III) separately. We begin with (II). Fixing
x ∈ X, by randomisation with a Rademacher sequence (εn )n∈Z ,
X
n φβ,ν (t2n A)M (t−1 2−n eiν )φ1−β,ν (t2n A)x
|n|6N
X
= sup n hM (t−1 2−n eiν )φ1−β,ν (t2n A)x, φβ,ν (t2n A)∗ x∗ i
kx∗ k61
|n|6N
D X X E
= sup E n εn M (t−1 2−n eiν )φ1−β,ν (t2n A)x, εn φβ,ν (t2n A∗ )x∗
kx∗ k61
|n|6N |n|6N
X
−1 −n iν n
6E n εn M (t 2 e )φ1−β,ν (t2 A)x
|n|6N
X
× sup E εn φβ,ν (t2n A∗ )x∗
kx∗ k61
|n|6N
X X
.M E n
εn φ1−β,ν (t2 A)x sup E εn φβ,ν (t2n A∗ )x∗ ,
kx∗ k61
|n|6N |n|6N

where the implicit constant in the last step is the R-boundedness constant of
M . Similarly, shifting the index by one and using the contraction principle,
we estimate (III) as follows:
X
n φβ,ν (t2n A)M (t−1 2−n eiν )φ1−β,ν (t2n A)x
|n|6N
X X
.M E εn φ1−β,ν (t2n−1 A)x sup εn φβ,ν (t2n−1 A∗ )x∗
kx∗ k61
|n|6N |n|6N
X X
6E n
εn φ1−β,ν (t2 A)x sup E εn φβ,ν (t2n A∗ )x∗ .
kx∗ k61
|n|6N +1 |n|6N +1

By the same argument, for (I) we obtain

X X
n φν (t2n A)x 6 E εn ψν (t2n A)x .
|n|6N |n|6N

Taking the supremum over N > 1 and t > 0, this proves the square function
bound
 16.2 Perturbation theorems 529
X
sup sup n φν (t2n B)x 6 Ckxk + 2C 0 sup kxkφ1−β ,A sup kx∗ kφβ ,A∗
|n|>N t>0 t>0 kx∗ k61
|n|6N

6 C 00 kxk,

where the estimate in the last step follows from the boundedness of the
H ∞ (Σσ )-calculus of A through Theorem 10.4.4. 
Proof of Theorem 16.2.8. By the second part of Theorem 16.2.4, assump-
tion (i) implies that A + B is σ-R-sectorial operator provided the small-
ness condition on C0 in (i) holds. Moreover, if we impose C0 < 1, then
for all x ∈ D(A + B) = D(A) we have kAxk 6 k(A + B)xk + kBxk 6
k(A + B)xk + C0 kAxk and therefore kAxk 6 (1 − C0 )−1 k(A + B)xk, while at
the same time k(A + B)xk 6 kAxk + kBxk 6 (1 + C0 )kAx. We conclude that

kAxk hC0 k(A + B)xk, x ∈ D(A + B) = D(A).

Furthermore, for λ ∈ {Σσ∨τ we have λ ∈ %(A + B) and the resolvent operator

is represented by the perturbation formula of Proposition 16.2.1,
∞
X
R(λ, A + B) = R(λ, A) [BR(λ, A)]n , | arg λ| > σ ∨ τ, (16.12)
n=0

again provided C0 is small enough, for then kBR(λ, A)k 6 C0 kAR(λ, A)k 6
C0 kλR(λ, A) − Ik 6 C0 (1 + Mσ∨τ,A ) < 1 and the series converges absolutely.
First we assume that (i) and (iii) hold. For the time being, we do not
assume that 0 ∈ %(A) (in which case A−1 is bounded by Corollary 15.2.10),
but only assume that A is invertible and B maps D(A1−α ) into D(A−α ). Then
U := A−α BAα−1 is bounded on X of norm kU k = C1 and we have

R(λ, A)BR(λ, A) = R(λ, A)Aα U A1−α R(λ, A).

f−1 , then the sum

If C1 < M σ∨τ,A
X
M (λ) := [U AR(λ, A)]k U
k>0

converges in operator norm and defines a holomorphic function for | arg λ| >
σ ∨ τ . We then can rewrite (16.12) in the form
X
R(λ, A + B) = R(λ, A) + Aα R(λ, A) [U AR(λ, A)]k U A1−α R(λ, A)
k>0

= R(λ, A) + Aα R(λ, A)M (λ)A1−α R(λ, A).

By the R-sectoriality of A and Proposition 8.1.24, the set {M (λ) : | arg(λ)| >
σ} is R-bounded. Thus we derived the representation required in Lemma
16.2.9 and we can conclude that A+B also has a bounded H ∞ (Σσ∨τ -calculus.
530 16 Perturbations and sums of operators

It remains to prove that the smallness assumption C1 < M f−1

σ∨τ,A can be
removed if 0 ∈ %(A). Under this assumption, let (i) and (iii) hold, but not
necessarily the smallness condition C1 < M f−1 .
σ∨τ,A
Using the scale of homogeneous fractional domain spaces Xα := D(Aα )
with norms kxkXα := kAα xk (recall that we are assuming 0 ∈ %(A)) we can
restate our assumptions as stating that B extends to a bounded operator
from X1 to X and from X1−α to X−α , with norm at most C0 and C1 respec-
tively. By complex interpolation B acts as a bounded operator [X1 , X1−α ]θ
to [X, X−α ]θ with norm 6 C01−θ C1θ , 0 < θ < 1.
Since A has a bounded H ∞ -calculus and therefore bounded imaginary
powers, by Corollary 15.3.10 we have

[X1 , X1−α ]θ = X1−θα , [X, X−α ]θ = X−θα

with equivalent norms, with equivalence constants which may be chosen in-
dependent of θ ∈ (0, 1). Thus we obtain that B acts as a bounded operator
from X1−θα to X−θα with norm . C01−θ C1θ , 0 < θ < 1.
We can choose θ so small that B satisfies (iii) for α0 = θα with C10 <
−1
M
f
σ∨τ,A no matter how big C1 was. This completes the proof of the case (iii).
Finally assume that (ii) holds for some α ∈ (0, 1). By Proposition 15.1.12
we have Aα−1 ⊆ Aα A−1 and A1+α ⊆ Aα A (in fact we have equality in the
second case by Theorem 15.2.5), and therefore

kAα−1 Bxk = kAα A−1 Bxk 6 C1 kA1+α A−1 Bxk = C1 kAα Bxk

implies that (iii) holds for the exponent 1 − α ∈ (0, 1) and x ∈ D(A1+α ). Since
D(A1+α ) is dense in D(Aα ) by Proposition 15.1.13, (iii) holds for the exponent
1 − α ∈ (0, 1) and x ∈ D(Aα ). 
We conclude this section with an example, due to McIntosh and Yagi, shows
that boundedness of the H ∞ -calculus is not preserved by small relatively
bounded perturbations even when X is a Hilbert space. This shows that the
additional assumptions (ii) or (iii) in Theorem 16.2.7 cannot be omitted, no
matter how small the constant on (i) is chosen.

Example 16.2.10. We construct a bisectorial operator A on Hilbert space H

admitting a bounded bisectorial H ∞ -calculus with ωH bi
∞ (A) = 0, such that

for any given ε > 0, an operator Bε on H exists which is relatively bounded

with respect to A, with kBε xk 6 εkAxk, and such that A + Bε fails to have
a bounded bisectorial H ∞ -calculus. This operator moreover satisfies (A +
Bε )2 = A2 + Cε , where C is relatively bounded with respect to A2 , with
kCε xk 6 2εkA2 xk.
By the first part of Theorem 10.6.7, the operator A2 has a bounded H ∞ -
calculus with ωH ∞ (A) = 0. If A2 + Cε = (A + Bε )2 had a bounded H ∞ -
calculus, then by the second part of Theorem 10.6.7 A + Bε would have a
 16.2 Perturbation theorems 531

bounded bisectorial H ∞ -calculus, and this is not the case. We conclude that
A2 + Cε does not have a bounded H ∞ -calculus.
Let us proceed to the construction of the operators A and Bε . Fix ε > 0.
Omitting subscripts ε in what follows, for n = 1, 2, . . . we will construct
bounded operators An and Bn on a finite-dimensional Hn with the following
properties for any 0 < σ < 21 π:
• An and An + Bn are σ-bisectorial with kBn k 6 εkAn k;
• A2n and (An + Bn )2 = A2n + Cn with kCn k 6 2εkA2n k;
• the spectra of An and An + Bn is contained in (−∞, 1] ∪ [1, ∞);
• the resolvents of An and An + Bn satisfy

kR(λ, An )k 6 1/=(λ), kR(λ, An + Bn )k 6 (1 + ε)/=(λ),

for all λ ∈ C \ R;
• An and An + Bn have contractive, respectively bounded, H ∞ (Σσ± )-calculi;
• the spectral projections 1Σσ± (An + Bn ) have norm > n.
The counterexample with the stated properties is obtained by taking
M M M M
H= Hn , A := Tn , B := Bn , C := Cn .
n>1 n>1 n>1 n>1

The operator A has a contractive H ∞ (Σσ± )-calculus. Furthermore, the in-

equalities kBn k 6 εkAn k imply that D(A) ⊆ D(B) and B is relatively bounded
with respect to A, with relative bound 6 ε. The operator A + B with domain
D(A + B) = D(A) doesn’t have a bounded H ∞ (Σσbi )-calculus: for if it had,
then the associated spectral projections would be bounded; but if they were,
then their restrictions to Hn would be uniformly bounded in n; but these
restrictions have norm > n.
We now turn to the details of the construction. Choose Nn > 1 so large
(n)
2ε
that 3π log(Nn + 1). On CNn +1 consider the matrices Tn = (tjk )N j,k=0 and
n

(n)
Sn = (sjk )N
j,k=0 given by
n

(n) (n) ε
tjk = 2j δjk , sjk := δj6=k .
π(k − j)
Then Tn is self-adjoint and Sn Tn is skew-adjoint. The self-adjoint matrix iSn
is the Nn × Nn Toeplitz matrix with generating function εθ/π, θ ∈ (−π, π),
that is, we have
sjk = fbj−k , j, k = 0, . . . , Nn .
Since the norm of a Toeplitz matrix with bounded real-valued generating
function f is bounded by kf kL∞ (T) , we see that kSn k 6 ε.
(n)
The matrix Zn = (zjk )N
j,k=0 given by
n

(n) 2k ε
zjk = δj6=k
π(k − j)(2j + 2k )
532 16 Perturbations and sums of operators

has norm
Nn −1
2N n ε X 1
kZn k > kZeNn k =
π j=0 (Nn − j)(2j + 2Nn )
ε 2Nn 1 1  (16.13)
> N −1 N
+ + · · · + 1
π (2 n + 2 n ) N n Nn − 1
2ε
> log(Nn + 1) > n,
3π
where (en )N
n=0 denote the standard unit vectors in C
n Nn +1
, and it satisfies

T n Z n + Z n T n = Sn T n . (16.14)

On Hn := C Nn +1 × C Nn +1 define the operators

       
T 0 0 Sn T n I Zn 0 −Zn
An := n , Bn := , Pn+ := , Pn− := .
0 −Tn 0 0 0 0 0 I

One checks that

    
0 Sn Tn 0 0 Sn
Bn = = An ,
0 0 0 −Tn 0 0
so
kBn k 6 kSn kkAn k 6 εkAn k.
Also, using that Sn Tn = −(Tn Sn )∗ , we have
 2
Tn Tn Sn Tn − Sn Tn2 0 Tn Sn Tn − Sn Tn2
  
2 2
(An +Bn ) = = An + =: A2n +Cn
0 Tn2 0 0

with
kCn k 6 kTn Sn Tn − Sn Tn2 k 6 2kSn kkTn k2 6 2εkA2n k,
where we used that Tn∗ = −Tn , so Tn is normal and therefore kTn k2 = kT 2 k.
Furthermore, one checks that σ(An ) = σ(An + Bn ) and
 
R(λ, Tn ) R(λ, Tn )Sn Tn R(λ, −Tn )
R(λ, An + Bn ) = (16.15)
0 R(λ, −Tn )

for all λ ∈ %(An ) = %(An + Bn ). In particular,

σ(An + Bn ) = σ(An )
= σ(Tn ) ∪ σ(−Tn ) = {1, 2, 4, . . . , 2Nn } ∪ {−1, −2, −4, . . . , −2Nn }

By self-adjointness, for λ 6∈ R we have kAn k 6 |=(λ)|−1 , so An is σ-bisectorial

for all 0 < σ < 21 π. By (16.15), for λ 6∈ R we have λ ∈ %(An + Bn ), and for
λ 6∈ Σσ
 16.3 Sum-of-operator theorems 533

kR(λ, An + Bn )k 6 |=(λ)|−1 + kR(λ, Tn )Sn Tn R(λ, Tn )k .σ (1 + ε)|=(λ)|−1 .

It follows that An + Bn is σ-bisectorial for all 0 < σ < 21 π.

The operators Pn+ and Pn− are projections,

Pn+ + Pn− = I, Pn+ Pn− = Pn− Pn+ = 0,

and by (16.13) their norms satisfy

kPn+ k > kZn k > n, kPn− k > kZn k > n.

To complete the construction we will show that

Pn± = 1Σσ± (An + Bn ).

Indeed, using (16.14) and (16.15), for 0 < ν < σ we formally compute

1Σσ± (An + Bn )
Z
1
= R(z, An + Bn ) dz
2πi ∂Σν±
Z  
1 R(z, Tn ) R(z, Tn )Sn Tn R(z, −Tn )
= dz
2πi ∂Σν± 0 R(z, −Tn )
Z  
1 R(z, Tn ) R(z, Tn )(Tn Zn + Zn Tn )R(z, −Tn )
= dz
2πi ∂Σν± 0 R(z, −Tn )
Z  
1 R(z, Tn ) R(z, Tn )Zn + Zn R(z, −Tn )
= dz
2πi ∂Σν± 0 R(z, −Tn )x
 
(∗) I Zn
= = Pn+ ,
0 0

where (∗) is a consequence of Cauchy’s theorem, which gives

Z Z
1 1
R(z, Tn ) dz = I, R(z, −Tn ) dz = 0,
2πi ∂Σν± 2πi ∂Σν±

noting that σ(Tn ) = {1, 2, 4, . . . , 2Nn } is contained in Σσ+ . To make the

computation rigorous, one brings in additional terms ζk (Tn ), where ζk (z) =
k 1
k+z − 1+kz as in Proposition 10.2.6, to be able to work with the Dunford cal-
culus for functions in H 1 (Σσ ) ∩ H ∞ (Σσ ) throughout; one passes to the limit
k → ∞ at the end. The proof that 1Σσ± (An + Bn ) = Pn− is entirely similar.

16.3 Sum-of-operator theorems

The perturbations B studied in Section 16.2 have the property that D(B) is
contained in D(A), so that the sum the sum A + B may be defined unam-
biguously by the prescription (A + B)x := Ax + Bx. In all these cases, B is
534 16 Perturbations and sums of operators

“small” in comparison with A. Under a resolvent commutation assumption,

in the present section we treat A and B on a more equal footing.
We begin with a general result (Theorem 16.3.2) which says that the sum
A + B of two resolvent commuting sectorial operators A and B satisfying
ω(A) + ω(B) < π always has a sectorial extension, and that this extension is
the closure of A + B if both A and B are densely defined. In applications to
maximal regularity of solution of evolution equations – the topic of the last
two chapters of this book – more is needed, namely, that A + B is closed and
the following inequality holds:

kAxk + kBxk 6 Ck(A + B)xk, x ∈ D(A) ∩ D(B) (16.16)

with a constant C independent of x ∈ D(A) ∩ D(B). For later use we record

the simple fact that this inequality in fact implies closedness:

Proposition 16.3.1. If A and B are closed operators satisfying (16.16), then

the operator A+B with its natural domain D(A+B) = D(A)+D(B) is closed.

Proof. The proof is immediate: if xn → x and (A + B)xn → y, then (16.16)

implies that the sequences (Axn )n>1 and (Bxn )n>1 are Cauchy. The closed-
ness of A and B implies that x ∈ D(A) ∩ D(B) and y = limn→∞ (A + B)xn =
Ax + Bx = limn→∞ Axn + limn→∞ Bxn = Ax + Bx = (A + B)x. 
As it turns out, the inequality (16.16) is rather delicate, and it only holds under
additional assumptions on A, B, and X. We have already encountered one
such situation: the Dore–Venni theorem (Theorem 15.4.11), which assumes
that A and B resolvent commute and have bounded imaginary powers, with
ωBIP (A) + ωBIP (B) < π, and the underlying Banach space X is a UMD space.
In applications, however, one if often confronted with the situation where
one of the operator is only (R-)sectorial, whilst the other operator has better
properties such as a bounded H ∞ -calculus. In the present section, for resolvent
commuting sectorial operators A and B acting in a Banach X we will prove
the following results:
• If A and B are densely defined, A has a bounded H ∞ -calculus and B is
R-sectorial, and if ωH ∞ (A) + ωR (B) < π, then A + B is densely defined
and sectorial, with

ω(A + B) 6 max{ωH ∞ (A), ωR (B)}

and the reverse triangle inequality (16.16) holds. If in addition X has the
triangular contraction property, then A + B is R-sectorial and

ωR (A + B) 6 max{ωH ∞ (A), ωR (B)}.

(Theorem 16.3.6).
 16.3 Sum-of-operator theorems 535

• If A and B are densely defined and have bounded H ∞ -calculi with

ωH ∞ (A) + ωH ∞ (B) < π and X has Pisier’s contraction property, then
A + B has a bounded H ∞ -calculus with

ωH ∞ (A + B) 6 max{ωH ∞ (A), ωH ∞ (B)}

and the reverse triangle inequality (16.16) holds (Theorem 16.3.10).

• If A has an absolute calculus with

ωabs (A) + ω(B) < π,

then the reverse triangle inequality (16.16) holds (Theorem 16.3.14). The
same conclusion holds if X is a Hilbert space, A has bounded imaginary
powers and B is densely defined, and ωBIP (A) + ω(B) < π (Theorem
16.3.15).
To conclude this section we provide an example of the type of applications
that will be studied in depth in the next two chapters and which indeed have
motivated the development of the abstract approach to sums of operators
presented here.
Suppose that −A generates a C0 -semigroup on a Banach space X and
consider the inhomogeneous abstract Cauchy problem
(
u0 (t) + Au(t) = f (t), t ∈ [0, T ],
(ACP)
u(0) = 0.

As we will explain in the next chapter, a thorough understanding of this

problem is of paramount importance to the study of more general classes of
nonlinear, possibly time-dependent, evolution equations. In order to connect
(ACP) with operator sums we consider the weak derivative

Du := u0

viewed as a closed operator on Lp (0, T ; X) (with 1 6 p 6 ∞) with domain

D(D) := 0 W 1,p (0, T ; X) = u ∈ W 1,p (0, T ; X) : u(0) = 0



It will be checked in the next chapter (see Section 17.3.c) that this operator
is sectorial of angle 21 π. Using this operator, we can rewrite (ACP) as the
abstract operator equation
(D + A)u
e =f

in Lp (0, T ; X), where A e is the natural extension of A to a closed operator

p e := Lp (0, T ; D(A)) by
acting in X e := L (0, T ; X), defined on D(A)

(Af
e )(t) := A(f (t)), t ∈ (0, T ).

In the next chapter (see Propositions 17.3.14 and 17.3.15) we prove that the
following assertions are equivalent:
536 16 Perturbations and sums of operators

(1) the inverse triangle inequality (16.16) holds, i.e., there is a constant C > 0
such that

kAuk
e p + kDukp 6 Ck(A
e + D)ukp , e ∩ D(D);
u ∈ D(A)

(2) A
e + D is closed;
(3) e + D boundedly invertible;
A
(4) A has maximal Lp -regularity on (0, T ).
For the problem (ACP), maximal Lp -regularity means that the unique mild
solution of the problem, which is given in terms of the semigroup S generated
by −A as Z t
u(t) = S(t − s)f (s) ds
0

belongs to Lp (0, T ; D(A)) ∩ 0 W 1,p (0, T ; X) = D(A)

e ∩ D(D). As we will see
in the next chapter, the bounded invertibility of A e + D corresponds to the
existence and uniqueness of mild solutions for (ACP). Maximal Lp -regularity
will be studied in depth in the next chapter, where also a version of the above
equivalences with (0, T ) replaced by R+ will be proved.

16.3.a The sum of two sectorial operators

We begin with a general result about sums of resolvent commuting operators.

It is not quite as useful as the deeper sums-of-operator theorems proved in the
next sections, but its virtue lies in the generality of its assumptions, namely,
it is only required that A and B are sectorial with ω(A) + ω(B) < π. The
price to be paid is that we do not obtain sectoriality, or even closedness, of
A + B, but only the weaker result that A + B has a sectorial extension. A
second reason to present this result in fair detail is that some techniques that
go into the proof will resurface in later proofs.
Theorem 16.3.2 (Sums of sectorial operators). If A and B are resolvent
commuting sectorial operators satisfying

ω(A) + ω(B) < π

then the operator A + B with its natural domain D(A + B) = D(A) + D(B)
has a closed extension to a sectorial operator C which satisfies

ω(C) 6 max{ω(A), ω(B)}.

Furthermore,
(1) If A or B is injective, then C is injective;
(2) If A and B are densely defined, then C is densely defined;
(3) If A and B are densely defined and A or B is standard sectorial, then C
is standard sectorial.
 16.3 Sum-of-operator theorems 537

If (2) holds (and hence if (3) holds), then C equals the closure of A + B.

The proof of this theorem will be given shortly. We first pause a brief moment
to explain why the condition

ω(A) + ω(B) < π

enters naturally in this theorem. Variants of this condition appear in all sum-
of-operator theorems we are about to encounter. Arguing naively, one would
like to realise the operator sum A + B through a ‘bivariate’ extended Dunford
calculus as (z +w)(A+B), where z +w is short-hand for the function (z, w) 7→
z + w. With this notation, to prove sectoriality of A + B one must estimate
the norms of
λ
λR(λ, A + B) = (A, B)
λ − (z + w)
for all λ ∈ C in the complement of a sector Σω containing all sums z + w with
z ∈ Σσ and w ∈ Στ , where ω(A) < σ < π and ω(B) < τ < π as usual. But
the algebraic sum Σσ + Στ is a sector only if σ + τ 6 π! Under this condition,
Σσ + Στ = Σmax{σ,τ } . In contrast, when σ + τ > π the reader may check that
Σσ + Στ = C. Clearly, the condition σ + τ 6 π forces ω(A) + ω(B) < π, and
in that case we may replace σ and τ by slightly smaller values to arrange that
σ + τ < π. Incidentally, this heuristic argument also shows that the inequality
ω(A + B) 6 max{ω(A), ω(B)} is natural to expect.
Let us now turn to the proof Theorem 16.3.2. Let A and B be resolvent
commuting sectorial operators in X satisfying ω(A) + ω(B) < π, and let
ω(A) < σ < π and ω(B) < τ < π be such that σ + τ < π. The construction of
the sectorial operator C extending A+B is based on the following observation,
which makes use of the primary calculus involving the spaces E(Σ) introduced
in Section 15.1.a. For a holomorphic function h ∈ E(Σσ ) ⊗ E(Στ ) of the form
N
X
h(z, w) = fn (z)gn (w)
n=1

with all fn ∈ E(Σσ ) and gn ∈ E(Στ ), we may define

N
X
h(A, B) := fn (A)gn (B).
n=1

It is not difficult that the operator h(A, B) is well defined, in the sense that it
does not depend on the particular representation of h. We now observe that
z+w  1  1 1  1 
h(z, w) := = 1− + 1− .
(1 + z)(1 + w) 1+z 1+w 1+z 1+w

This identifies the left-hand side as an element of E(Σσ ) ⊗ E(Στ ). Thinking

of
538 16 Perturbations and sums of operators

1
ρ(z, w) :=
(1 + z)(1 + w)
as a regulariser for the function (z, w) 7→ z + w, we define

C := (I + A)(I + B)h(A, B)

with domain

D(C) := x ∈ X : h(A, B)x ∈ R((I + B)−1 (I + A)−1 .



A bit of algebra reveals that

• x ∈ D(C) ⇐⇒ (A + B)(I + B)−1 (I + A)−1 x ∈ R((I + B)−1 (I + A)−1 )

and, for x ∈ D(C),

Cx = (I + A)(I + B)(A + B)(I + B)−1 (I + A)−1 x.

From this equivalence, by a standard argument one deduces that

• C is closed.
Proof of Theorem 16.3.2. We will prove that C defined by the above proce-
dure has the required properties.
It is immediate from the definition that D(A) ∩ D(B) is contained in D(C);
this is the same as saying that C is an extension of A + B. In fact, a moment’s
reflection shows that

D(A) ∩ D(C) = D(A) ∩ D(B) = D(C) ∩ D(B). (16.17)

Choose ω(A) < σ < π and ω(B) < τ < π in such a way that σ + τ < π.
As was already observed above, the condition σ + τ < π implies that

Σσ + Στ := {z + w : z ∈ Σσ , w ∈ Στ } = Σmax{σ,τ } .

For z ∈ Σσ , w ∈ Στ , and λ ∈ C with max{σ, τ } < | arg(λ)| < π, we write

λ λ2 λzw
= + .
λ − (z + w) (λ − z)(λ − w) (λ − (z + w))(λ − z)(λ − w)
These functions are holomorphic on Σσ × Στ and one may check that

λR(λ, C) = λ2 R(λ, A)R(λ, B) + fλ (A, B), (16.18)

where fλ (A, B) can be defined in terms of the function

λzw
fλ (z, w) :=
(λ − (z + w))(λ − z)(λ − w)
as the absolutely convergent Dunford integral
 16.3 Sum-of-operator theorems 539
Z Z
1
fλ (A, B) := fλ (z, w)R(z, A)R(w, B) dz dw.
(2πi)2 ∂Στ ∂Σσ

With µ := λ/|λ| we then have

Z Z
1
kfλ (A, B)k .σ,τ,A,B |dz| |dw| (16.19)
∂Στ ∂Σσ |µ − (z + w)||µ − z||µ − w|

and the sectoriality of C with angle ω(C) 6 ω now easily follows from the
fact that value of the integral on the right-hand side of (16.19) is uniformly
bounded with respect to µ on the arc {|µ| = 1, | arg(µ)| > max{σ, τ }.
Since the choices ω(A) < σ < π and ω(B) < τ < π and max{σ, τ } < ω < π
were arbitrary, it follows that ω(C) 6 max{ω(A), ω(B)}.
It remains to prove the assertions (1)–(3).
(1): Suppose that is injective and let x ∈ D(C) be such that Cx = 0. By
the definition of C, this means that h(A, B)x = (I + B)−1 (I + A)−1 y for some
y ∈ X and Cx = (I + A)(I + B)h(A, B)x = y = 0. Consider the function
z
g(z, w) = , z, w ∈ Σσ .
(1 + z)2 (z + w)
By the primary calculus, for fixed z ∈ Σσ we have

g(z, B) = z(1 + z)−2 (z + B)−1 .

Borrowing some terminology from the next subsection, this function belongs
to H 1 (Σσ ; A ), where A is the set of operators in L (X) commuting with the
resolvent of A, and we may define a bounded operator g(A, B) through the
Dunford integral
Z
1
g(A, B) := z(1 + z)−2 (z + B)−1 (z − A)−1 dz.
2πi ∂Σν
In view of

(z + B)−1 (z − A)−1 C
= (z + B)−1 (z − A)−1 [(I + A)(I + B)h(A, B)]
= [(1 + z)(z − A)−1 − I][(1 − z)(z + B)−1 + I]h(A, B)
= [(1 + z)(z − A)−1 − I][(1 − z)(z + B)−1 + I](A + B)(I + A)−1 (I + B)−1
= [(1 + z)(z − A)−1 − I][I − (I + A)−1 ][(1 − z)(z + B)−1 + I](I + B)−1
+ [(1 + z)(z − A)−1 − I][(1 − z)(z + B)−1 + I][I − (I + B)−1 ](I + A)−1
= (z − A)−1 − (z + B)−1 ,

where the last line follows by the resolvent identity. By Cauchy’s theorem,
Z Z
z 
−1 −1
 z
2
(z − A) − (z + B) dz = 2
(z − A)−1 dz
∂Σν (1 + z) ∂Σν (1 + z)
540 16 Perturbations and sums of operators

= A(I + A)−2 x

It follows that

0 = g(A, B)Cx = A(I + A)−2 x.

Since A and (I − A)−1 are injective, this forces x = 0.

If B is injective and Cx = 0, the same argument (with the roles of A and
B reversed) again shows that x = 0.
(2): If A and B are densely defined, then so is C by (16.17). If x ∈ D(A),
then the vectors xn := n2 (n+A)−1 (n+B)−1 x belong to D(A+B) and converge
to x in X as n → ∞. Similarly, the vectors Cxn = n2 (n + A)−1 (n + B)−1 Cx
converge to Cx in X as n → ∞. This shows that D(A + B) is dense in D(C)
with respect to the graph norm.
(3): Suppose now that A and B are standard sectorial. Then D(A) and
D(B) are dense, and therefore D(C) is dense by (2). Furthermore, arguing
as in part (1) we see that for all x ∈ X we have g(A, B)x ∈ D(C) and
Cg(A, B)x = A(I + A)−2 x. Since R(A(I + A)−2 ) = D(A) ∩ R(A), it follows
that D(A) ∩ R(A) ⊆ R(C) and therefore R(C) is dense. By Proposition 10.1.8,
this implies that D(C) ∩ R(C) is dense, i.e., C is standard sectorial. 
In the next proposition we assume that A and B are sectorial operators in X
satisfying ω(A) + ω(B) < π, and choose ω(A) < νA < σA < π, ω(B) < νB <
σA < π, and max{νA , νB } < ν < σ < π. The operator C is as in Theorem
16.3.2.

Proposition 16.3.3. Every λ 6∈ Σmax{σ,τ } belongs to %(C) and

Z Z
1 %(z)%(w)
%(A)R(λ, C)%(B) = R(z, A)R(w, B) dw dz.
(2πi)2 ∂ΣνA ∂ΣνB λ − (z + w)

In its stated form, the proposition will be useful in the proof of Theorem
16.3.10. It is clear from the proof that the proposition could be stated with
%(A), %(B), and R(λ, C) replaced by more general operators φ(A), ψ(B), and
f (A) under suitable conditions on the functions φ, ψ, and f . We leave the
details to the interested reader.
Proof. It has already been observed that every λ 6∈ Σmax{νA ,νB } belongs to
the resolvent set of C, and by (16.18) (using the notation introduced there)
we have R(λ, C) = λR(λ, A)R(λ, B) + gλ (A, B), where

zw 1 λ
gλ (z, w) := = − .
(λ − (z + w))(λ − z)(λ − w) (λ − (z + w)) (λ − z)(λ − w)
(16.20)

Inserting this into the Dunford integral

 16.3 Sum-of-operator theorems 541

%(A)R(λ, C)%(B)
Z Z
1
= %(z)%(w)R(z, A)R(λ, C)R(w, B) dw dz,
(2πi)2 ∂ΣνA ∂ΣνB

we see that this results in the sum of three integrals, where (I) corresponds
to the contribution λR(λ, A)R(λ, B), and (II) and (III) correspond to the
splitting of gλ by (16.20). By a simple computation involving Fubini’s theorem
and Cauchy’s theorem, the integrals (I) and (III) cancel, and the integral (II)
equals the one in the statement of the lemma. 

16.3.b Operator-valued H ∞ -calculus and closed sums

In this section we extend the Dunford calculus of a sectorial operator A to an

operator-valued Dunford calculus and study the question when this calculus
is bounded with respect to the H ∞ -norm. The idea is to obtain (16.16) from
the boundedness of the operator f (A, B) in terms of the function f (λ, B) =
B(λ + B)−1 in the operator-valued calculus. Loosely speaking, this gives a
way to define an operator “A(A + B)−1 ” even when A + B fails to be bounded
invertible. With the operator at hand, it is possible to run a rigorous version
of the estimate

kAxk = kA(A + B)−1 (A + B)xk 6 Ck(A + B)xk

with C = kA(A + B)−1 k. From this one also obtains the estimate

kBxk 6 k(A + B)xk + kAxk 6 (1 + C)k(A + B)xk,

and together these estimates give (16.16), with implied constant 1 + 2C.
In what follows, A always denotes a sectorial operator on a Banach space
X, and we fix ω(A) < σ < π. Let A be a closed sub-algebra of L (X) resolvent
commuting with A, i.e.,

T R(z, A) = R(z, A)T for all T ∈ A and z ∈ %(A).

We then denote by H 1 (Σσ ; A ) the space of all holomorphic functions F :

Σσ → A for which
Z
dt
kF kH 1 (Σσ ;A ) := sup kF (eiν t)k
|ν|<σ R+ t

is finite. It is easily checked that, which respect to this norm, H 1 (Σσ ; A )

is a Banach space. For functions F ∈ H 1 (Σσ ; A ) we can define a bounded
operator F (A) ∈ A by means of the operator-valued Dunford integral
Z
1
F (A) = F (z)R(z, A) dz,
2πi ∂Σν
542 16 Perturbations and sums of operators

where ω(A) < ν < σ. The resulting operator is independent of the particular
choice of ν, and it satisfies
Mσ,A
kf (A)k 6 kF kH 1 (Σσ ;A ) ,
π
where Mσ,A = supλ∈{Σσ kλR(λ, A)k.
As in Proposition 10.2.2, this calculus is multiplicative and satisfies the
following convergence property: if Fn , F ∈ H 1 (Σσ ; A ) are uniformly bounded
and satisfy Fn (z)x → F (z)x for all z ∈ Σσ and x ∈ X, then for all g ∈ H 1 (Σσ )
we have

lim (fn g)(A)x = (f g)(A)x, x ∈ X.

n→∞

Denote by H ∞ (Σσ ; A ) the space of all holomorphic functions F : Σσ → A

for which the set {F (z) : z ∈ Σσ } is uniformly bounded. Endowed with the
norm
kF kH ∞ (Σσ ;A ) := sup{kF (z)k : z ∈ Σσ },
this space is easily seen to be Banach space. In the same way one defines
RH ∞ (Σν ; A ) as the space of all holomorphic functions F : Σν → A for
which the set {F (z) : z ∈ Σν } is R-bounded. Endowed with the norm

kF kRH ∞ (Σσ ;A ) := R({F (z) : z ∈ Σσ })

(the R-bound of {F (z) : z ∈ Σσ }), this space is a Banach space.

The main result of this section is the following theorem.

Theorem 16.3.4. Let A be a sectorial operator on a Banach space X, let

ω(A) < σ < π, and suppose that A has a bounded H ∞ (Σσ )-calculus. Then
there exists a unique bounded linear mapping F 7→ F (A) from RH ∞ (Σσ ; A )
into L (D(A) ∩ R(A)) with the following properties:
(1) For every function F ∈ RH ∞ (Σσ ; A ) ∩ H 1 (Σσ ; A ) the operator F (A)
coincides with the one defined by the Dunford integral;
(2) For all F, G ∈ RH ∞ (Σσ ; A ) we have F G ∈ RH ∞ (Σσ ; A ) and

(F G)(A) = F (A)G(A) = G(A)F (A);

(3) Whenever the functions Fn , F ∈ RH ∞ (Σσ ; A ) are uniformly bounded and

satisfy Fn → F pointwise on Σσ , then limn→∞ Fn (A)x = F (A)x for all
x ∈ D(A) ∩ R(A).
Furthermore, if X has Pisier’s contraction property and T is an R-bounded
subset of A , then for all 0 < σ < ν < π the family

F (A) : F ∈ RH ∞ (Σν ; A ), F (z) ∈ T for all z ∈ Σν



is R-bounded.
 16.3 Sum-of-operator theorems 543

Parts (2) and (3) are analogues of the corresponding results in Theorem
10.2.13 and the proofs are similar. The proof of (1), which is the non-trivial
part of the theorem, is based on an extension of Lemma 10.3.13, which states
that if A is a sectorial operator on a Banach space X and F ∈ H 1 (Σσ ; A ) is
given, with ω(A) < σ < π, then for all ω(A) < ν < σ we have
Z
1 dz
F (A) = z 1/2 F (z)φz (A) , (16.21)
2πi ∂Σν z

where φz (λ) := λ1/2 /(z − λ). The proof is identical to that of Lemma 10.3.13;
all one needs to do is to replace H 1 (Σσ ) by H 1 (Σσ ; A ) throughout, and
so is the justification of the well-definedness of the operators φz (A) and the
convergence of integral on the right-hand side of (16.21).
We also need the following strengthening of Lemma 10.3.8:

Lemma 16.3.5. Let A be a sectorial operator on a Banach space X with a

bounded H ∞ -calculus, and let ωH ∞ (A) < ν < σ < π. Suppose φ, ψ ∈ H 1 (Σσ ),
and let T ⊆ A be R-bounded. Then for all finite subsets F ⊆ Z, all scalars
|aj | 6 1 and operators Tj ∈ T (j ∈ F ), and all x ∈ D(A) ∩ R(A),
X
sup aj Tj φ(2j tA)ψ(2j tA)x 6 CkφkH 1 (Σσ ) kψkH 1 (Σσ ) kxk,
t>0
j∈F

where C is a constant depending only on ν, σ, and A.

Proof. Let A0 denote the part of A in X0 := D(A) ∩ R(A). This operator is

standard sectorial and has a bounded H ∞ -calculus, with the same bounds,
and the same holds for its adjoint A∗0 . Let (εj )j∈Z be a Rademacher sequence.
For norm one vectors x ∈ X0 and x∗ ∈ X0∗ , and for any fixed t > 0 and finite
subset F ⊆ Z we may estimate
DX E
aj Tj φ(2j tA)ψ(2j tA)x, x∗
j∈F
X
= aj hTj ψ(2j tA0 )x, φ(2j tA∗0 )x∗ i
j∈F
DX X E
= E εj aj Tj ψ(2j tA0 )x, εk φ(2k tA∗0 )x∗
j∈F k∈F
 X 2 1/2  X 2 1/2
6 E εj aj Tj ψ(2j tA0 )x E εk φ(2k tA∗0 )x∗
j∈F k∈F
∞ 2
6 Kσ−ν (Mν,A ) R(T )kφkH 1 (Σσ ) kψkH 1 (Σσ ) kxkkx∗ k

using R-boundedness, the Kahane contraction principle, and Theorem 10.4.4

(and its notation) in the last step. The result now follows by taking the supre-
mum over all x∗ ∈ X0∗ of norm at most 1. 
544 16 Perturbations and sums of operators

Proof of Theorem 16.3.4. Let ω(A) < ν < σ < π and let F be as in (1). As in
the proof of Theorem 10.3.4(3), for x ∈ D(A) ∩ R(A) and F ∈ H 1 (Σσ ; A ) ∩
RH ∞ (Σσ ; A ) we find
X X 1 Z 2
−iν/2 dt
F (A)x = e F (e−iν 2j t)φe−iν (t−1 2−j A)x ,
=±1
2πi 1 t
j∈Z

where φz (λ) = λ1/2 /(z − λ). Then, with aj () = e−iν/2 ,

1 X
kF (A)xk 6 sup sup sup aj ()F (e−iν 2j t)φe−iν (t−1 2−j A)x .
π =±1 k>1 t>0
|j|6k

Now we choose T to be the R-bounded range of F , and we let φ = ψ =

(φe−iν )1/2 if  = 1 and φ = ψ = (φeiν )1/2 if  = −1. Applying the lemma
twice, we obtain
2
kF (A)xk 6 C max kφ −iν kH 1 (Σσ ) kxk,
π =±1 e
where C is the constant of the lemma.
The proofs of multiplicativity and the convergence property proceed as in
Theorem 10.2.13.
Regarding the final assertion, we may adapt the proof of Theorem 10.3.4(3),
replacing the scalar functions fn and f by A -valued functions Fn and F . 
As an application of the operator-valued calculus we prove a useful variant
of the Dore–Venni theorem (Theorem 15.4.11). In that theorem, both A and
B were assumed to have bounded imaginary powers and act in a UMD Ba-
nach space X. In the present theorem, we weaken the assumption on A and
strengthen the assumption on B.
Theorem 16.3.6 (The sum of an R-sectorial operator and an oper-
ator with bounded H ∞ -calculus). Let A and B be resolvent commuting
densely defined (respectively, standard) sectorial operators on a Banach space
X. Assume that A has a bounded H ∞ -calculus, B is R-sectorial, and

ωH ∞ (A) + ωR (B) < π.

Then A+B is a densely defined (respectively, standard) sectorial operator and

ω(A + B) 6 max{ωH ∞ (A), ωR (B)}.

Moreover, there exists a constant C > 0 such that

kAxk + kBxk 6 Ck(A + B)xk, x ∈ D(A) ∩ D(B). (16.22)

If X has the triangular contraction property, then A + B is R-sectorial with

ωR (A + B) 6 max{ωR (A), ωH ∞ (B)}.

 16.3 Sum-of-operator theorems 545

Proof. The idea is to define A(A+B)−1 and B(A+B)−1 as bounded operators

on D(A) ∩ R(A) using the operator-valued functional calculus for A.
Step 1 – We first assume that A and B are standard sectorial. Let A
denote the closed sub-algebra of L (X) comprised of all operators resolvent
commuting with A. Choose ωH ∞ (A) < σ < π and ωR (B) < τ < π and such
that σ + τ < π. We wish to apply the operator-valued calculus of A to the
function
F (z) := B(z + B)−1 = I − zR(z, −B).
This function belongs to RH ∞ (Σσ ; A ) since the spectrum of −B is contained
in the closure of −Στ = {z ∈ C : | arg(z)| > π − τ } and σ < π − τ . Further-
more, the function
G(z) := ζn (z)2 (z + B)ζn (B)2 ,
n 1
with ζn (z) = n+z − 1+nz as in Proposition 10.2.6, is easily seen to belong to
H (Σσ ; A ) ∩ RH (Σσ ; A ) by R-sectoriality. We have
1 ∞

(F G)(z) = F (z)G(z) = ζn (z)2 Bζn (B)2 ,

and in the operator-valued Dunford calculus the operators G(A) and (F G)(A)
are given by

G(A) = Bζn (B)2 ζn (A)2 + Aζn (A)2 ζn (B)2 ,

(F G)(A) = Bζn (B)2 ζn (A)2 ,

using resolvent commutation to do some rewriting. By the multiplicativity of

the operator-valued H ∞ -calculus of A we have

Bζn (B)2 ζn (A)2 = (F G)(A) = F (A)G(A)

= F (A) Bζn (B)2 ζn (A)2 + Aζn (A)2 ζn (B)2 .

The boundedness of the operator-valued H ∞ -calculus of A then gives, for

x ∈ D(A) ∩ D(B),

kζn (A)2 ζn (B)2 Bxk .σ,A ζn (A)2 ζn (B)2 Bx + ζn (B)2 ζn (A)2 Ax

Letting n → ∞ and using A and B are standard sectorial, we obtain the

inequality
kBxk .σ,A k(A + B)xk.
From this we also obtain

kAxk 6 kBxk + kAx + Bxk .σ,A k(A + B)xk.

We have already observed in Proposition 16.3.1 that (16.22) implies the closed-
ness of A + B. The standard sectoriality of A + B now follows from Theorem
16.3.2.
546 16 Perturbations and sums of operators

Step 2 – We now assume that A and B are densely defined, but not nec-
essarily standard sectorial. Then the operators Aε := A + ε and Bε : +B + ε
are standard sectorial, and we may apply the above reasoning with Fε (z) :=
Bε (z +Bε )−1 and Gε (z) := ζn (z)2 (z +Bε )ζn (Bε )2 . This results in the estimate

kAε xk + kBε xk .σ,A k(Aε + Bε )xk

with an implied constant that is uniform in ε > 0 and independent of x. The

estimate
kAxk + kBxk .σ,A k(A + B)xk
follows from this by letting ε ↓ 0.
Step 3 – Suppose finally that X has the triangular contraction property.
From the proof of Theorem 16.3.2 (and keeping in mind that A + B equals
the operator C of that theorem by what we have already proved) we recall
the identity

λR(λ, A + B)
Z Z
1
= λ2 R(λ, A)R(λ, B) + fλ (z, w)R(z, A)R(w, B) dz dw.
(2πi)2 ∂Στ ∂Σσ

Outside the closure of Σσ+τ the operators λ2 R(λ, A)R(λ, B) are R-bounded,
by the R-sectoriality of A (which follows from the second part of Theorem
10.3.4) and B (by assumption). The operators corresponding to the Dunford
integral with fλ are R-bounded by Theorem 8.5.2; the integrability properties
required to apply theorem have already been observed in the proof of Theorem
16.3.2 (see (16.19)). 
Since standard sectorial operators with a bounded H ∞ -calculus on a Banach
space with the triangular contraction property are R-sectorial with ωR (A) =
ωH ∞ (A) (see Corollary 10.4.10), we have the following corollary.
Corollary 16.3.7. Let A and B be resolvent commuting densely defined (re-
spectively, standard) sectorial operators with bounded H ∞ -calculi satisfying
ωH ∞ (A) + ωH ∞ (B) < π on a Banach space with the triangular contraction
property. Then A + B is a densely defined (respectively, standard) sectorial
operator and (16.22) holds.

16.3.c The joint H ∞ -calculus

As a first application of the operator-valued functional calculus we construct

the joint functional calculus of resolvent commuting standard sectorial oper-
ators.
Denote by H 1 (Σσ1 × · · · × Σσn ) the space of holomorphic functions on
Σσ1 × · · · × Σσn which obey the obvious integrability estimate extending the
case n = 1. For functions f ∈ H 1 (Σσ1 ×· · ·× Σσn ) we define the joint Dunford
calculus by
 16.3 Sum-of-operator theorems 547
 1 n Z Z Yn
f (A1 , . . . , An ) = ... f (λ1 , . . . , λn ) R(λj , Aj ) dλ1 . . . dλn
2πi ∂Σνn ∂Σν1 j=1
(16.23)

where ω(Ai ) < νi < σi for i = 1, . . . , n.

If n = 2, by Fubini’s theorem we can formally rewrite (16.23) as
Z  1 Z
1 
f (A1 , A2 ) = f (λ1 , λ2 )R(λ2 , A2 ) dλ2 R(λ1 , A1 ) dλ1
2πi ∂Σν1 2πi ∂Σν2
Z
1
= f (λ1 , A2 )R(λ1 , A1 ) dλ1
2πi ∂Σν1
= Φ1 (f (·, A2 ))(A1 ),

where Φ1 : g 7→ g(A1 ) denotes the operator-valued calculus of A1 , provided of

course that all terms are well defined. This indicates the way how to extend
(16.23) to H ∞ (Σσ1 × · · · × Σσn ) using induction where each of the operator
Aj has a bounded H ∞ -calculus. Here, H ∞ (Σσ1 × · · · × Σσn ) denotes the space
of bounded holomorphic functions on Σσ1 × · · · × Σσn .
The following straightforward extension of Lemma 10.2.17 will be useful.
As before, by A we denote the set of bounded operators commuting with the
resolvent of A.

Lemma 16.3.8. Let A have a bounded H ∞ (Σσ )-calculus on X. Suppose that

f : [a, b] × Σσ → A is a measurable function with the following properties:
(i) z 7→ f (s, z) belongs to RH ∞ (Σσ ; A ) for all s ∈ [a, b];
Z bZ ∞
dt
(ii) sup kf (s, eiν t)k ds < ∞.
|ν|<σ a 0 t
Rb
Then the function g(z) = a f (s, z) ds belongs to H ∞ (Σσ ; A ) and
Z b
g(A)x = f (s, A)x ds, x ∈ X.
a

The straightforward proof is left to the reader.

We will now apply the operator-valued calculus to the sum-of-operators
problem next.

Theorem 16.3.9. Let A1 , . . . , An be densely defined resolvent commuting sec-

torial operators on a Banach space X with the Pisier contraction princi-
ple, and assume that Aj has a bounded H ∞ (Σσj )-calculus, j = 1, . . . , n.
Then for σj < νj < π, (16.23) extends to an algebra homomorphism
Φ : H ∞ (Σν1 × · · · × Σνn ) → L (D(A) ∩ R(A)) with the following convergence
property:
548 16 Perturbations and sums of operators

If the functions fm , f are uniformly bounded in H ∞ (Σν1 × · · · ×

Σνn ) and limm→∞ fm = f pointwise on Σν1 × · · · × Σνn , then
limm→∞ Φ(fm )x = Φ(f )x for all x ∈ D(A) ∩ R(A).
Moreover, the set of operators

Φ(f ) : f ∈ H ∞ (Σσ1 × · · · × Σσn ), kf k∞ 6 1



is R-bounded.

Notation. In place of Φ(f ) we shall write f (A1 , . . . , An ).

Proof. By Aj we denote the sub-algebra of all operators in L (X) that com-
mute with Aj and put A := A1 ∩ · · · ∩ An . Note that R(λ, Aj ) ∈ A for
all j = 1, . . . , n. The case n = 1 follows from the general properties of the
H ∞ -calculus. Assume now that A2 , . . . , An have a joint functional calculus
Ψ : H ∞ (Σσ2 × · · · × Σσn ) → L (X) with the required properties. Since X has
Pisier’s contraction property, the set

T = g(A2 , . . . , An ) : g ∈ H ∞ (Σσ2 × · · · × Σσn ), kgkH ∞ 6 1



is an R-bounded subset of A ⊆ A1 by Theorem 10.3.4(3). By Φ1 we denote

the operator-valued functional calculus of A1 defined on RH ∞ (Σσ1 ; A1 ) as
constructed in Theorem 16.3.4.
Given a function f ∈ H 1 (Σσ1 × · · · × Σσn ) ∩ H ∞ (Σσ1 × · · · × Σσn ) with
kf kH ∞ 6 1, the set
{f (λ1 , ·, . . . , ·) : λ1 ∈ Σσ1 }
is uniformly bounded in H ∞ (Σσ2 × · · · × Σσn ). Hence

Ψ f (λ1 , ·, . . . , ·) = f (λ1 , A2 , . . . , An ) ∈ T
 
for all λ1 ∈ Σσ1 .

Furthermore, the function

λ1 7→ f (λ1 , A2 , . . . , An )
 1 n−1 Z Z Yn
= ... f (λ1 , . . . , λn )R(λj , Aj ) dλ2 . . . dλn
2πi ∂Σν2 ∂Σνn j=2

is holomorphic on Σσ1 . Again by Theorem 10.3.4(3), f (·, A2 , . . . , An ) ∈

RH ∞ (Σσ1 ; A1 ). Consequently we can define


Φ(f ) = Φ1 f (·, A2 , . . . , An )

using Theorem 16.3.13. We can extend this definition to arbitrary f ∈

H ∞ (Σσ1 × · · · × Σσn ). The required properties of Φ now follow from the corre-
sponding properties of Ψ and Φ1 . For instance, Φ extends (16.23) by Fubini’s
theorem. To check the multiplicativity, choose f, g ∈ H ∞ (Σσ1 × · · · × Σσn ).
Then
 16.3 Sum-of-operator theorems 549


Φ(f · g) = Φ1 (f · g)(·, A2 , . . . , An )


= Φ1 f (·, A2 , . . . , An ) · g(·, A2 , . . . , An )



= Φ1 f (·, A2 , . . . , An ) Φ1 g(·, A2 , . . . , An ) = Φ(f )Φ(g).

Also, if fm , f are bounded in H ∞ (Σσ1 × · · · × Σσn ) and fm → f pointwise as

m → ∞, then by the convergence property of the operator-valued calculus of
A1 we have

lim fm (λ1 , A2 , . . . , An )x = f (λ1 , A2 , . . . , An )x

m→∞

for every fixed λ1 ∈ Σσ1 and all x ∈ X. Now apply the convergence property
of Φ1 to Fm (λ) = fm (λ, A2 , . . . , An ) and F (λ) = f (λ, A2 , . . . , An ).
The final R-boundedness assertion follows directly from the final assertion
of Theorem 16.3.4. 
As an application we have the following variant of Corollary 16.3.7. This is
result is actually true for Banach space X with the triangular contraction
property; we refer to the Notes for a discussion of this fact.

Theorem 16.3.10. Let A and B be resolvent commuting standard sectorial

operators with bounded H ∞ -calculi satisfying ωH ∞ (A) + ωH ∞ (B) < π on a
Banach space X with Pisier’s contraction property. Then A + B admits a
bounded H ∞ -calculus with

ωH ∞ (A, B) 6 max{ωH ∞ (A), ωH ∞ (B)}.

For the proof we need a technical proposition. For the sake of its formulation,
the joint Dunford calculus of two resolvent commuting sectorial operators A
and B will be denoted by ΦA,B : f 7→ f (A, B), for functions f ∈ H 1 (Σσ ) ×
H 1 (Στ ). Likewise, the operator-valued Dunford calculus of A will be denoted
by ΦA : F 7→ F (A), for operator-valued functions F ∈ H 1 (Σσ ; A ) where A
is set of operators resolvent commuting with A.

Proposition 16.3.11. Let A and B be resolvent commuting sectorial op-

erators acting in a Banach space X satisfying ω(A) + ω(B) < π, and let
max{ω(A), ω(B)} < σ < π. Let C denote the operator sum of A and B as
constructed above. Then for all f ∈ H 1 (Σσ ) we have

%(A)f (C)%(B) = ΦA,B ((z, w) 7→ %(z)f (z + w)%(w))

= ΦA (z 7→ %(z)f (z + B)%(B)),

where %(z) = z/(1 + z)2 .

Proof. By the definition of the joint Dunford calculus,

ΦA,B ((z, w) 7→ %(z)(λ − (z + w))%(w))

550 16 Perturbations and sums of operators
Z Z
1 %(z)%(w)
= R(z, A)R(w, B) dw dz.
(2πi)2 ∂ΣνA ∂ΣνB λ − (z + w)
On the other hand, by Proposition 16.3.3 and Cauchy’s theorem,
%(A)f (C)%(B)
Z
1
= f (λ)%(A)R(λ, C)%(B) dλ
2πi ∂Σν
Z Z Z
1 %(z)%(w)
= f (λ) R(z, A)R(w, B) dw dz dλ
(2πi)3 ∂Σν ∂ΣνA ∂ΣνB λ − (z + w)
Z Z
1
= %(z)%(w)
(2πi)2 ∂ΣνA ∂ΣνB
 1 Z f (λ) 
× dλ R(z, A)R(w, B) dw dz
2πi ∂Σν λ − (z + w)
Z Z
1
= %(z)%(w)f (z + w)R(z, A)R(w, B) dw dz
(2πi)2 ∂ΣνA ∂ΣνB

which equals

= ΦA,B ((z, w) 7→ %(z)f (z + w)%(w))

but also
Z  1 Z
1 
= %(z) %(w)f (z + w)R(w, B) dw R(z, A) dz
2πi ∂ΣνA 2πi ∂ΣνB
= ΦA (z 7→ %(z)f (z + B)%(B)).

Proof of Theorem 16.3.10. Since X has Pisier’s contraction property, A and
B admit R-bounded operator-valued H ∞ -calculi by Theorem 10.3.4. Choose
ω(A) < σA < π and ω(B) < σB < π such that σA + σB < π, and let
max{σA , σB } < σ < π. For f ∈ H ∞ (Σσ ), the function
F (z, w) := f (z + w)
belongs to H ∞ (ΣσA ) × H ∞ (ΣσB ). Since B has an R-bounded H ∞ -calculus,
the set {F (z, B) : z ∈ ΣσA } is an R-bounded subset of the set A of bounded
operators resolvent commuting with A. Applying the operator-valued calculus
of A we obtain a bounded operator F (A, B) on D(A) ∩ R(A). By Proposition
16.3.11,
%(A)f (A + B)%(B) = %(A)F (A, B)%(B),
where %(z) = z/(1 + z)2 . Since %(A) is injective and %(B) has dense range (we
assumed that A and B are standard sectorial), we conclude that f (A + B) =
F (A, B) is a bounded operator on D(A) ∩ R(A). The bound kf (A, B)k .
kf kH ∞ (Σσ ) follows by tracing the steps of the proof. 
 16.3 Sum-of-operator theorems 551

Inspection of the proof shows that the ‘standard’ assumption on A may be

weakened to ‘densely defined and injective’. In reflexive spaces, however, these
conditions imply standardness (see Proposition 10.1.9).

16.3.d The absolute calculus and closed sums

The main result of this section (Theorem 16.3.13) provides a version of The-
orem 16.3.6 in which no assumption on the Banach space is needed, the as-
sumption on a A is weakened to sectoriality, and the assumption on B is
strengthened to having an absolute calculus.

Definition 16.3.12 (Absolute functional calculus). Let A be a sectorial

operator acting in a Banach space X, and let ω(A) < σ < π. We say that A
admits an absolute calculus on Σσ if there exist a constant M > 0 and g, h ∈
H 1 (Σσ ) ∩ H ∞ (Σσ ), with khkH 1 (Σσ ) = 1, such that for all x, y ∈ D(A) ∩ R(A)
the validity of the estimate

kh(tA)g(tA)xk 6 kg(tA)yk for all t>0

implies kxk 6 M kyk.

We denote

ωabs (A) := inf σ ∈ (0, π) : A admits an absolute calculus on Σσ .

Examples of classes of operators with an absolute calculus will be given in the

next subsection.

Theorem 16.3.13 (Absolute calculus implies operator-valued H ∞ -

calculus). Let A be a sectorial operator in a Banach space X, let ω(A) <
σ < π, and suppose that A admits an absolute calculus on Σσ . Then A ad-
mits a bounded operator-valued H ∞ (Σσ )-calculus. In particular, A admits a
bounded H ∞ (Σσ )-calculus.

Proof. Let g, h be as in Definition 16.3.12, and let F ∈ H 1 (Σσ ; A ) ∩

H ∞ (Σσ ; A ). Choose ω(A) < ν < σ. For z ∈ D(A) ∩ R(A) we estimate
Z
1
kh(tA)F (A)zk 6 |h(tλ)F (λ)|kR(λ, A)zk |dλ|
2π ∂Σν
|dλ|
Z
Mν,A
6 kF k∞ |h(tλ)| kzk
2π ∂Σν |λ|
Mν,A
6 kF k∞ kzk,
π
where Mν,A = supλ∈{Σν kλR(λ, A)k is finite by sectoriality. Now, given y ∈
D(A) ∩ R(A), we let x := F (A)y, note that x ∈ D(A) ∩ R(A), and apply the
estimate with z := g(tA)y to obtain
552 16 Perturbations and sums of operators

Mν,A
kh(tA)g(tA)xk 6 kF k∞ kkg(tA)yk.
π
By the absolute calculus, this implies
Mν,A Mν,A
kF (A)yk = kxk 6 M kF k∞ kyk.
π π
By the same argument as in the proof of Theorem 10.2.13, this proves the
first assertion.
The second assertion follows by taking F (λ) = f (λ)IX . 

Theorem 16.3.14 (Closedness from the absolute calculus). Let A and

B be resolvent commuting densely defined sectorial operators. If A has an
absolute calculus on Σσ and σ + ω(B) < π, then the operator A + B with
domain D(A + B) = D(A) ∩ D(B) is closed and

kAxk + kBxk 6 C(A + B)xk, x ∈ D(A) ∩ R(A).

Proof. By Theorem 16.3.13, A admits a bounded operator-valued H ∞ (Σσ )-

calculus. Since B is τ -sectorial, the family f (z, B) = −zR(−z, B), z ∈ Σπ−τ ,
is uniformly bounded and commutes with the resolvent of A. Now Theorem
16.3.13 implies that f (A, B) is well defined in the operator-valued calculus as
a bounded operator on X. The reverse Hölder inequality

kAxk + kBxk 6 Ck(A + B)xk, x ∈ D(A) ∩ D(B),

is obtained by same argument as in Theorem 16.3.6. It was already observed

that the closedness of A + B follows from it. 
We will prove in the next section (see Theorem 16.3.18) that a standard
sectorial operator on a Hilbert space has an absolute calculus if and only if it
has a bounded imaginary powers. Taking this for granted for now, as a special
case of Theorem 16.3.14 we recover the following classical result.

Theorem 16.3.15 (Da Prato–Grisvard). Let A and B be resolvent com-

muting sectorial operators in a Hilbert space H. If A has bounded imaginary
powers, B is densely defined, and ωBIP (A) + ω(B) < π, then the operator
A + B with domain D(A + B) = D(A) ∩ D(B) is closed and we have

kAxk + kBxk 6 Ck(A + B)xk, x ∈ D(A) ∩ D(B),

with C a constant independent of x.

Comparing this result with the Dore–Venni theorem, where both A and B are
assumed to have bounded imaginary powers, we observe that here, bounded-
ness of the imaginary powers is imposed only on A.
 16.3 Sum-of-operator theorems 553

16.3.e The absolute calculus and real interpolation

In this subsection and the next, we show connect the absolute calculus with
the theory of real interpolation. The crucial observation is contained in the
following theorem.
Theorem 16.3.16 (Lp -bounds imply absolute calculus). Let A be a sec-
torial operator in a Banach space X and let ω(A) < σ < π. Let 1 6 p 6 ∞,
and suppose that there exist φ ∈ H 1 (Σσ ) ∩ H ∞ (Σσ ) such that
kxkφ,p := t 7→ φ(tA)x Lp (R+ , dt
, x ∈ D(A) ∩ R(A),
t ;X)

induces an equivalent norm on D(A) ∩ R(A), the finiteness of the norms on

the right-hand side being part of the assumptions. Then A has an absolute
calculus on Σσ .
The proof depends on the following lemma.
Lemma 16.3.17. Let A be a sectorial operator acting in X and let ω(A) <
σ < π. If for some p ∈ [1, ∞] and some ψ ∈ H 1 (Σσ ) ∩ H ∞ (Σσ ) one has t 7→
ψ(tA)x ∈ Lp (R+ , dt 1
t ; X) for all x ∈ D(A) ∩ R(A), then for all φ ∈ H (Σσ ) ∩
H ∞ (Σσ ) one has t 7→ φ(tA)x ∈ Lp (R+ , dt
t ; X) for all x ∈ D(A) ∩ R(A) and
we have the equivalence of norms
φ(·A)x Lp (R+ , dt
hφ,ψ,σ,A ψ(·A)x Lp (R+ , dt
t ;X) t ;X)

with implied constants independent of x.

Proof. Let ψ ∈ H 1 (Σσ ) ∩ H ∞ (Σσ ) have the properties as stated, and let
φ ∈ H 1 (Σσ ) ∩ H ∞ (Σσ ) be arbitrary and fixed. Choose an auxiliary function
g ∈ H 1 (Σσ ) such that Z ∞
dt
g(t)ψ(t) = 1.
0 t
First we let x ∈ D(A) ∩ R(A). By the Calderón reproducing formula
(Proposition 10.2.5) and the multiplicativity of the Dunford calculus,
Z ∞
dt
g(tA)ψ(tA)x =x (16.24)
0 t
with improper convergence of the left-hand side integral. Fix ω(A) < ν < σ.
For all s > 0 and 0 < r < R < ∞, by Fubini’s theorem and multiplicativity
we have
Z R Z R Z
dt 1  dt
φ(sA)g(tA)ψ(tA)x = φ(sλ)g(tλ)R(λ, A) dλ ψ(tA)x
r t r 2πi ∂Σν t
Z Z R
1  dt 
= φ(sλ) g(tλ)R(λ, A)ψ(tA)x dλ.
2πi ∂Σν r t
(16.25)
554 16 Perturbations and sums of operators

By (16.24) (with x replaced by φ(sA)x), upon passing to the limits r ↓ 0 and

R → ∞ in (16.25) (using dominated convergence to deal with the right-hand
side) we obtain
Z ∞
dt
φ(sA)x = φ(sA)g(tA)ψ(tA)x
t
Z0 ∞  Z
1  dt
= φ(sλ)g(tλ)R(λ, A) dλ ψ(tA)x
0 2πi ∂Σν t
Z
1
= φ(sλ)G(λ)x dλ
2πi ∂Σν

with Z ∞
dt
G(λ) := g(tλ)R(λ, A)ψ(tA) .
0 t
Applying Young’s inequality for L1 (R+ , dt
t ) twice (after parametrising ∂Σν
−1
and substituting s 7→ s ), we obtain that φ(·A)x ∈ Lp (R+ , dt
t ; X) and

Mν,A
φ(·A)x Lp (R+ , dt
6 kφkH 1 (Σσ ) kgkH 1 (Σσ ) ψ(·A)x Lp (R+ , dt
.
t ;X) π t ;X)

Now that we know that t 7→ φ(tA)x belongs to Lp (R+ , dt t ; X), the opposite
norm estimate is obtained by reversing the roles of φ and ψ. This proves the
theorem for x ∈ D(A) ∩ R(A).
For x ∈ D(A) ∩ R(A) the result follows by approximation, noting that
ψ(tA)xkLp (R+ , dt ;X) 6 Ckx by a closed graph argument. 
t

Proof of Theorem 16.3.16. Fix functions g, h ∈ H 1 (Σσ ) ∩ H ∞ (Σσ ), with

khkH 1 (Σσ ) = 1 for some ω(A) < ν < σ, and assume that x, y ∈ D(A) ∩ R(A)
satisfy
kh(tA)g(tA)xk 6 kg(tA)yk, t > 0.
Let 1 6 p 6 ∞ and f ∈ Lp (R+ ; dt t ; X) be as in the assumptions in the
theorem. Then, by Lemma 16.3.17, applied to φ = g and ψ = g · h, the
functions t 7→ g(tA)x, t 7→ h(tA)g(tA)x = (hg)(tA)x, and t 7→ g(tA)y belong
to Lp (R, dt
t ; X) and

kxk h f (·A)x Lp (R+ , dt

h h(·A)g(·A)x Lp (R+ , dt
t ;X) t ;X)

6 g(·A)y Lp (R+ , dt
h f (·A)y Lp (R+ , dt
h kyk
t ;X) t ;X)

with implied constants independent of x and y. Hence, g and h satisfy the

condition in the definition of the absolute calculus. 
An immediate application if the following characterisation of the absolute
calculus in Hilbert spaces.
 16.3 Sum-of-operator theorems 555

Theorem 16.3.18 (Hilbert space case). For a standard sectorial operator

A acting in a Hilbert space H, the following assertions are equivalent:
(1) A has a bounded H ∞ -calculus;
(2) A has an absolute calculus;
(3) A has bounded imaginary powers.
In this situation we have

ωH ∞ (A) = ωabs (A) = ωBIP (A).

Further equivalences are obtained in Theorems 10.4.21 (square function es-

timates) and 10.4.22 (generation of a contraction semigroup with respect to
some equivalent Hilbertian norm).
Proof. The implication (2)⇒(1) has already been proved in Theorem 16.3.13.
The implication (1)⇒(2) follows from the same theorem, because the bound-
edness of the H ∞ -calculus of a sectorial operator on H implies the square
function bounds
Z ∞ dt 1/2
kxkH h kt 7→ g(tA)xkγ(L2 (R+ , dt ),H) = kg(tA)xk2
t
0 t

by Theorem 10.4.16 and Proposition 9.2.9.

For standard sectorial operators A, the equivalence (1)⇔(3) is contained
in Theorem 15.3.23. 
The main result of this section is Theorem 16.3.20 which asserts that invertible
sectorial operators have a bounded H ∞ -calculus on the real interpolation
spaces (X, D(A))θ,p . We begin with a general result which describes these
interpolation spaces in terms of the Dunford calculus of A.

Theorem 16.3.19 (Real interpolation spaces between X and D(A)).

Let 0 < θ < 1 and p ∈ [1, ∞], and let A be a sectorial operator on X. Let
ω(A) < σ < π and suppose that 0 6= φ ∈ H 1 (Σσ ) ∩ H ∞ (Σσ ) is such that the
function z 7→ z −1 φ(z) belongs to H 1 (Σσ ) ∩ H ∞ (Σσ ) as well. Then
n o
(X, D(A))θ,p = x ∈ X : t 7→ t−θ φ(tA)x ∈ Lp (R+ , dt t ; X)

with equivalence of norms

kxk(X,D(A))θ,p h kxk + t 7→ t−θ φ(tA)x dt ,

Lp (R+ , t ;X)

where the implied constants only depend on σ, A, and φ. If 0 ∈ %(A), we also

have equivalence of homogeneous norms

kxk(X,D(A))θ,p h t 7→ t−θ φ(tA)x dt .

Lp (R+ , t ;X)
556 16 Perturbations and sums of operators

The theorem should be compared with the first part of Proposition K.4.1,
which asserts that If A is a sectorial operator in X, then
n o
(X, D(A))θ,p = x ∈ X : λ 7→ λθ kGR(λ, G)xk ∈ Lp (R+ , dλ
λ )

with equivalence of norms

kxk(X,D(G))θ,p h kxk + λ 7→ λθ kG(λ + G)−1 xk dλ .

Lp (R+ , λ )

In the E(Σσ )-calculus of A we have G(λ+G)−1 = φ(λ−1 G) with φ(z) = z/(z+

1). The case treated in Theorem L.2.4 corresponds to the choice φ(z) = ze−z .
Proof. ‘⊆’: For t > 0 and x = x0 + x1 with x0 ∈ X and x1 ∈ D(A), write

φ(tA)x = φ(tA)x0 + φ(tA)x1

and note that φ(tA)x1 ∈ D(A) with Aφ(tA)x1 = φ(tA)Ax1 . Furthermore

write φ0 (z) := φ(z) and φ1 (z) := z −1 φ(z). Then

kφ(tA)xk 6 kφ(tA)xk + tk(tA)−1 φ(tA)Axk

= kφ0 (tA)xk + tkφ1 (tA)Axk


6 Cσ,A kφ0 kH 1 (Σσ ) kxk + tkφ1 kH 1 (Σσ ) kxkD(A) ,

where Cσ,A is a constant only depending on σ and A. Taking the infimum

over all such decompositions, we obtain

kφ(tA)xk 6 Cσ,A max{kφ0 kH 1 (Σσ ) , kφ1 kH 1 (Σσ ) }K(t, x; X, D(A)).

It follows that if x ∈ (X, D(A))θ,p , then t 7→ t−θ φ(t, A)x ∈ Lp (R+ , dt

t ; X) and

kt 7→ t−θ φ(t, A)xkLp (R+ , dt ;X)

t

6 Cσ,A max{kφ0 kH 1 (Σσ ) , kφ1 kH 1 (Σσ ) }kxk(X,D(A))θ,p .

‘⊇’: Let x ∈ X be such that t 7→ t−θ φ(tA)x belongs to Lp (R+ , dt t ; X).

Choose f ∈ H 1 (Σσ ) ∩ H ∞ (Σσ ) in such a wayR∞ that f 1 (z) := zf (z) belongs
to H 1 (Σσ ) and the normalisation condition 0 f (s)φ(s) ds s = 1 is satisfied.
Noting that f φ ∈ H 1 (Σσ ), for z ∈ Σσ we define
Z 1 Z ∞
ds ds
h(z) := f (sz)φ(sz) , g(z) := f (sz)φ(sz) .
0 s 1 s

By substitution, g(z) + h(z) = 1 for all z ∈ Σσ .

The assumption t−θ φ(tA)x ∈ Lp (R+ , dt t ; X) implies that t
−θ
(f φ)(tA)x ∈
p dt
L (R+ , t ; X) since k(f φ)(tA)xk = kf (tA)φ(tA)xk .σ,A kf kH 1 (Σσ ) kφ(tA)xk
by multiplicativity. Hölder’s inequality therefore implies that (f φ)(tA)x ∈
 16.3 Sum-of-operator theorems 557

Lp ((0, 1), dt
t ; X) for every t > 0. Hence, as in the proof of Proposition 10.2.5
we have
Z 1
ds
h(tA)x = (f φ)(stA) . (16.26)
0 s
Next, noting the identities
Z ∞ Z ∞
ds ds
zg(z) = zf (sz)φ(sz) = s−1 f1 (sz)φ(sz) ,
1 s 1 s

reasoning similarly as for h we have g(A)x ∈ D(A) and, for t > 0,

Z ∞ Z ∞
ds ds
Ag(tA)x = (st)−1 (f1 φ)(stA)x = s−1 (f1 φ)(sA)x .
1 s t s

Accordingly, for the decomposition x = h(tA)x + g(tA)x ∈ X + D(A) we

obtain
K(t, x; X, D(A)) 6 kh(tA)xk + tkg(tA)kD(A)
= kh(tA)xk + tkg(tA)xk + tkAg(tA)xk
Z ∞
ds
6 kh(tA)xk + tkg(tA)xk + t s−1 (f1 φ)(sA)x .
t s
(16.27)

We have tkg(tA)xk .σ,A,φ tkgkH 1 (σσ ) kxk. Trivially, we also have

K(t, x; X, D(A)) 6 kxk,

By taking the minimum of these estimates, it follows that

Z ∞
ds
K(t, x; X, D(A)) .σ,A,φ kh(tA)xk + min{1, t}kxk + t s−1 (f1 φ)(sA)x
t s
=: (I) + (II) + (III).

We will estimate (t−θ ×) (I), (II), (III) separately.

For term (II) it is immediately clear that t−θ min{1, t}kxk belongs to
L (R+ , dt
p
t ).
By (16.26) we may estimate term (I) by
Z 1 Z t
ds ds
kh(tA)xk = (f φ)(stA) 6 k(f φ)(sA)xk =: σx (t).
0 s 0 s

We can now apply the first part of Hardy’s inequality (Lemma L.3.2) to obtain
that t 7→ t−θ σx (t) ∈ Lp (R+ , dt
t ; X) and

1
kt 7→ t−θ σx (t)kLp (R+ , dt ;X) .σ,A kt 7→ t−θ (f φ)(tA)xkLp (R+ , dt ;X)
t θ t
558 16 Perturbations and sums of operators

1
.σ,A kf kH 1 (Σσ ) kt 7→ t−θ φ(tA)xkLp (R+ , dt ;X) .
θ t

To estimate term (III), we note that

Z ∞ Z ∞
ds ds
t s−1 (f1 φ)(sA)x 6 kf1 kH 1 (Σσ ) kφ(sA)xk
t s t s
by multiplicativity and since s > t on the domain of integration, Therefore,
by the second part of Lemma L.3.2,
Z ∞
−θ ds
t 7→ t ts−1 (f1 φ)(sA)x
t s Lp (R+ , dtt ;X)
1
.σ,A kf1 kH 1 (Σσ ) kt 7→ t−θ φ(tA)xkLp (R+ , dt ;X) .
θ t

Combining things, we have shown that t−θ K(t, x; X, D(A)) belongs to

L (R+ , dt
p
t ) and

kt−θ K(t, x; X, D(A))kLp (R+ , dt ) .σ,A kxk + kt−θ φ(tA)xkL1 (R+ , dt ;X) .
t t

This is the same as saying that x ∈ (X, D(A))θ,p and

kxk(X,D(A))θ,p .σ,A kcxk + kt−θ φ(tA)xkL1 (R+ , dt ;X) .

t

Finally, if 0 ∈ %(A) we may endow D(A) with the equivalent norm

x 7→ kAxk. In doing so, the term (II) disappears and the first equivalence
of homogeneous norms is obtained.
Suppose next that A is invertible, in (16.27) we can estimate

kg(tA)xk 6 kA−1 kkg(tA)Axk

and therefore the second term can be estimated in the same way at the third
term appearing in the second line of (16.27). 

Theorem 16.3.20 (Absolute calculus on real interpolation spaces).

Let A be a densely defined sectorial operator with 0 ∈ %(A). Then for all
1 6 p 6 ∞ and 0 < θ < 1, the part Aθ,p of A in the real interpolation space
(X, D(A))θ,p has an absolute calculus.

The proof of this theorem depends on the following lemma.

Lemma 16.3.21. Under the assumptions of the theorem, for every α < 1 − θ
the norm
Z ∞
p dt 1/p zα
kxkα = t−θ ζα (tA)xk with ζα (z) = , (16.28)
0 t (1 + z)2α

is an equivalent norm on (X, D(A))θ,p .

 16.3 Sum-of-operator theorems 559

Proof. This is an immediate consequence of Theorem 16.3.20, as the condition

α > 1 − θ ensures that the conditions of the theorem hold for ζα . 
Proof of Theorem 16.3.20. We write Y := (X, D(A))θ,p for brevity.
Step 1 – We begin by preparing two helpful estimates.
First, for all s, t > 0 we have
  s t α
kζα (sA)ζα (tA)k 6 CA min , . (16.29)
t s
Indeed, by multiplicativity of the E(Σσ )-functional calculus and the identity

T : = (tA)α (1 + tA)−2α (sA)α (1 + sA)−2α

s
α  2α −1  
= A (t + A)−2α (s−1 )2α (s−1 + A)−2α
t
we obtain
s
α s
α
kT k 6 sup kA(t−1 + A)−1 k2α · sup ks−1 (s−1 + A)−1 k2α 6 CA .
t t>0 s>0 t

Since the same estimate holds with s and t interchanged, this gives (16.29).
Second, for all s > 0, t ∈ [s, 2s], and x ∈ X we have

kζα (sA)xk = kζα ((s − t/2)A)ζ((t/2)A)xk 6 Ckζα ((t/2)A)xk, (16.30)

where C = supr>0 kζα (rA)k is finite by (10.9).

Step 2 – Now we turn to the actual proof of the theorem.
With the notation introduced above, we take g := ζ2α and h := ζδ , where
α ∈ N satisfies α > 1 − θ and δ > 0. For x, y ∈ Y assume that

kh(tA)g(tA)xkY 6 kg(tA)ykY , t > 0. (16.31)

Then for all s > 0,

Z 2s
 dt 
kζ3α (sA)xk . sθp t−θp kζ3α+δ (sA)xkp
s t
Z 2s
dt
.s θp
t−θp kζα ((t/2)A)ζ2α+δ (sA)xkp (by (16.30))
s t
Z 4s
dt
= (2s) θp
t−θp kζα (tA)ζ2α+δ (sA)xkp
2s t
. (2s)θp kζ2α+δ (sA)xkpY (by (16.28))
= (2s) kh(sA)g(sA)xkpY
θp

6 (2s)θp kg(sA)ykpY (by (16.31))

Z ∞
dt
. (2s) θp
t−θp kζα (tA)ζα (sA)ζα (sA)ykp (by (16.28))
0 t
560 16 Perturbations and sums of operators
Z s
 t
αp dt
. kζα (sA)yk p
(2s) θp
t−θp
0 s t
Z ∞
s
αp dt 
+ (2s)θp t−θp (by (16.29))
s t t
.θ,p kζα (sA)ykp

with implied constants depending on A, σ, θ, p, and α. Integrating the left-

and right-hand sides in this estimate with respect to s−θp ds
s and using (16.28)
twice, we see that kxkY . kykY with implied constant independent of x and
y. This proves that Aθ,p has an absolute calculus on Y . 

Corollary 16.3.22 (Dore). If A is a densely defined sectorial operator on

X, with 0 ∈ %(A), then for all 1 6 p 6 ∞ and 0 < θ < 1, the part Aθ,p of A
in the real interpolation space (X, D(A))θ,p has a bounded H ∞ -calculus.

The invertibility assumption cannot be dropped in the corollary, and hence in

the theorem. Indeed, let A be a bounded sectorial operator without a bounded
H ∞ -calculus (such operators exist, even on a separable Hilbert space, by
Corollary 10.2.29). Then D(A) = X and therefore (X, D(A))θ,p = X for all
0 < θ < 1 and p ∈ [1, ∞]. By assumption, A doesn’t have a bounded H ∞ -
calculus on this space.

16.4 Notes

Section 16.1

The problem of defining the sum of two unbounded operators A and B can
be approached from various angles. Besides the direct approach of defining
A + B as (the closure of) the operator given on D(A) ∩ D(B) by (A + B)x =
Ax + Bx, which works well if A and B have commuting resolvents, various
other approaches can be taken. When A and B generate uniformly bounded
C0 -semigroups S and T respectively, conditions can be formulated in order
that the limit in the Trotter product formula

V (t)x := lim (S(t/n)T (t/n))n x

n→∞

exist for all x ∈ X, and defines a C0 -semigroup whose generator C is the

closure of the operator A + B initially defined on D(A) ∩ D(B) by (A + B)x =
Ax+Bx Engel and Nagel [2000]; resolvent commutation is not needed in these
results. A different approach is the form method, suitable when A and B are
defined on a Hilbert space with inner product (·|·). This method in provides
conditions under which the (closure of the) sum c := a + b of the sesquilinear
forms
a(x, y) := (Ax|y), b(x, y) = (Bx|y)
 16.4 Notes 561

is associated with a closed operator C satisfying

c(x, y) = (Cx|y).

Like always, subtle domain questions have to be taken care of. A detailed treat-
ment is given in Kato [1995]; for a gentle introduction see, e.g., Van Neerven
[2022].

Section 16.2

In this section some classical perturbation theorems for sectorial operators

are extended to R-sectorial operators. Theorem 16.2.4 on relatively bounded
perturbations of R-sectorial operators is basically from Kunstmann and Weis
[2001], with some improvements of constants. More sophisticated perturbation
results using real interpolation are contained in Haak, Haase, and Kunstmann
[2006].
Proposition 16.2.6 on perturbations of the H ∞ by multiples of the identity
and the main theorem of this section, Theorem 16.2.8 on relatively bounded
perturbations of the H ∞ -calculus, are from Kalton, Kunstmann, and Weis
[2006]. This paper contains a number of variants of the relative bounded-
ness conditions, some of them modelled after form perturbations. Part (iii) of
Theorem 16.2.8 was proved independently by Denk, Dore, Hieber, Prüss, and
Venni [2004]. The perturbation theorem 16.2.7 for lower order perturbations
of the H ∞ -calculus is due to Amann, Hieber, and Simonett [1994].
Example 16.2.10 is due to McIntosh and Yagi [1990]; a proof of the fact that
the norm of a Toeplitz matrix with bounded real-valued generating function
f is bounded by kf kL∞ (T) can be found in Garoni and Serra-Capizzano [2017,
Theorem 6.1]. A further example can be found in Kalton [2007]; see also the
review paper Batty [2009].
The philosophy behind some of these perturbation theorems is that the
boundedness of the H ∞ -calculus is encoded in the fractional domain spaces of
the operator in the following sense (see Kalton, Kunstmann, and Weis [2006]):
If A and B are two standard sectorial operators on a reflexive Banach space
X, and if for some 0 < α1 < α2 < 32 and j = {0, 1} we have

D(Aαj ) = D(B αj ) and kAαj xk h kB αj xk (16.32)

for all x in this common domain, then if one of the operators has a bounded
H ∞ -calculus, then so does the other. Notice that there are no smallness as-
sumptions here.
The basic idea of the proofs of Theorem 16.2.8, the comparison theorem
just quoted, and their variants is to use the relative bounded or the equivalence
of norms of (16.32) to show the equivalence of the discrete square function
norms X
x 7→ εj φ(2j A)x
j∈Z
562 16 Perturbations and sums of operators

with the corresponding ones for B. Here, φ is usually of the form φ(z) =
z α (1 + z)n with α < n. The two conditions (ii) and (iii) of Theorem 16.2.8
and (16.32) correspond to the two sides of the square function estimate.
In Kalton, Kunstmann, and Weis [2006] it is also explained how to use
these perturbation theorems to establish the boundedness of the H ∞ -calculus
for rather general classes of elliptic operators on H s,p (Rd ) or H s,p (D) for
smooth domains D ⊆ Rd with Lopatinskii–Shapiro boundary conditions. The
idea is to compare them to constant coefficients. A related approach is used in
Denk, Hieber, and Prüss [2003]. For more recent results the reader is referred
to the Notes of Chapter 17.
Let us mention two further topics related to the H ∞ -calculus and its per-
turbations.

Extrapolation of the H ∞ -calculus in the Lp -scale

We have seen in Chapter 11 that for singular integral it is often a successful

strategy to first prove a Hilbert space result, then prove a weaker result on
Lp0 or some endpoint of the Lp scale, and then extend the Hilbert space result
to Lp -spaces by interpolation between 2 and p0 . This idea also proves fruitful
for perturbation theorems; see Kunstmann and Weis [2017]. As in the case
of classical singular integral operators, the Littlewood–Paley theory and so-
called off-diagonal estimates play a crucial role. Here, the “Littlewood–Paley
decomposition” of a standard sectorial operator B on a space Lp (S) with a
bounded H ∞ -calculus is expressed as the equivalence of norms
X 1/2
kxkLp (S) h |φ(2j B)x|2
Lp (S)
j∈Z

for all x ∈ Lp (S). Given a second standard sectorial operator A on Lp (S), the
following R-boundedness condition expresses that A is “close” to B in terms
of the “ Littlewood–Paley pieces” φ(2j A) and φ(2j B):
 
R φ(s2j+k A)ψ(t2k B) : j ∈ Z . 2−βk (16.33)

for some β > 0 and all k ∈ Z and s, t ∈ [1, 2]. Kunstmann and Weis [2017]
contains the following theorem:

Theorem 16.4.1. Let A and B be standard sectorial operators consistently

defined on L2 (S) and Lp0 (S), with p0 ∈ (1, ∞) \ {2}, and assume that A is
R-sectorial on Lp0 (S) and B has a bounded H ∞ -calculus on both L2 (S) and
Lp0 (A). If (16.33) holds for p = 2, then A has a bounded H ∞ -calculus on
both L2 (S) and all spaces Lp (S) with p between 2 and p0 .

This theorem can be extended to the case where A is defined on a comple-

mented subspace of Lp and A is a “retract” of B in a suitable sense. In this
 16.4 Notes 563

way one can, for example, derive the boundedness of the H ∞ -calculus of the
Stokes operator on the Helmholtz space Lp0 (D) from the boundedness of the
H ∞ -calculus of the Laplace operator on Lp (D), for bounded Lipschitz do-
mains D ⊆ Rd with d > 3 and | 21 − p1 | < 2d 1
(see Kunstmann and Weis
[2017]).

Scales of fractional domain spaces and interpolation

Let A be a standard sectorial operator and denote by Ẋα the completion of

D(Aα ) with respect to the norm kxkα := kAα xk for α ∈ R. The methods
of Section 15.3.b show that a Hilbert space operator A has a bounded H ∞ -
calculus if and only if these fractional domain spaces can be identified with
the complex interpolation spaces

[Ẋα , Ẋβ ]θ = Ẋγ

with (1 − θ)α + θβ = γ for α 6= β and θ ∈ (0, 1). This is not true anymore in
Banach spaces, where complex interpolation is related to boundedness of the
imaginary powers, rather than the boundedness of the H ∞ -calculus. However,
such an identification is possible with the help of the γ-interpolation method
introduced in the Notes to Section 15.3. It is shown in Kalton, Kunstmann,
and Weis [2006, Section 5.3] that a standard γ-sectorial operator A on a
Banach space X with non-trivial type has a bounded H ∞ -calculus if and only
if
(Ẋα , Ẋβ )γθ = Ẋδ
with (1 − θ)α + θβ = δ for α 6= β and θ ∈ (0, 1). Even when A does not
have a bounded H ∞ -calculus, the spaces (Ẋα , Ẋβ )γθ can be identified with
γ
certain square function spaces Hs,A which are defined as the completion of
m m
D(A ) ∩ R(A ), m > |s| + 1, with respect to the norm

t 7→ t−s φ(tA)x γ(R+ , dt

, x ∈ D(Am ) ∩ R(Am )
t ;X)

for some φ ∈ H 1 (σσ ) such that z 7→ z −s φ(z) still belongs to H 1 (σσ ). Complete
proofs can be found in Kalton, Lorist, and Weis [2023].

Section 16.3

The operator-sum method as a purely functional analytic approach to evo-

lution equations goes back to Da Prato and Grisvard [1975], where already
Theorem 16.3.15 is proved. Our proof of Theorem 16.3.2 follows Haase [2006],
where further properties of the operator C extending A + B are discussed.
In the setting of Hilbert spaces, the operator-valued H ∞ -calculus was in-
troduced in Albrecht, Franks, and McIntosh [1998]. Theorem 16.3.4 is taken
from Kalton and Weis [2001]. It is implicit in Lancien and Le Merdy [1998]
564 16 Perturbations and sums of operators

(see also Lancien, Lancien, and Le Merdy [1998, Remark 6.5] and Albrecht,
Franks, and McIntosh [1998]) that any sectorial operator on a Hilbert space
with a bounded H ∞ -calculus has a bounded operator-valued H ∞ -calculus.
In these papers the “right” method of proof for Theorem 16.3.4 was already
found, but the crucial ingredient of R-boundedness was still missing.
Theorem 16.3.6 is due to Kalton and Weis [2001]. In the next chapter, the
connections of Theorem 16.3.6 with maximal Lp -regularity will be discussed
in detail. As we will see in Volume IV, the operator-valued functional calculus
of Theorem 16.3.4 can be used to give a short proof for stochastic maximal Lp -
regularity; see Van Neerven, Veraar, and Weis [2015b]. In Clément and Prüss
[2001] it is shown that if A is an injective operator generating a bounded C0 -
group on a UMD space X, and B is an invertible closed linear operator in X
resolvent commuting with A such that ±iB is R-sectorial, then the operator
A + B with domain D(A) ∩ D(B) is closed and invertible. If B is also sectorial
with angle ω(B) < 12 π, then A + B is sectorial as well, and ω(A + B) < 21 π.
Theorem 16.3.9 is due to Lancien, Lancien, and Le Merdy [1998], Lancien
and Le Merdy [1998], who extend an earlier result of Albrecht [1994] on Lp -
spaces with 1 < p < ∞.
That Theorem 16.3.10 holds more generally for Banach spaces with the
triangular contraction property was shown by Le Merdy [2003].
The absolute functional calculus was introduced in Kalton and Kucherenko
[2010], where Theorems 16.3.13, 16.3.14, and 16.3.20 were proved. The defi-
nition of the absolute calculus may be a little off-putting if one is accustomed
to thinking in terms of spectral theory, but the benefits of this notion is con-
siderable:
• It implies an operator-valued H ∞ -calculus and sum-of-operators theorem
without the complexities of R-boundedness, just as in Hilbert spaces (see
Theorems 16.3.13 and 16.3.14).
• It leads to a simple sufficient condition for the abstract functional calculus
of a sectorial operator A in terms of the equivalence
Z ∞ dt 1/p
kxk h kφ(tA)xkp (16.34)
0 t

(see Theorem 16.3.16, implicit in Kalton and Kucherenko [2010]). The

criterion already suffices for the common applications and it shows that
in L1 , L2 , and C(K) spaces the absolute functional calculus is equivalent
to the H ∞ -calculus; see Kalton and Kucherenko [2010]. It also shows that
the absolute calculus is mainly a tool for real interpolation spaces and non-
UMD spaces. Indeed, for every operator A with a bounded H ∞ -calculus
on Lp with 1 < p < ∞ we have (cf. Section 10.4)
Z ∞ dt 1/2
kxkp h t 7→ φ(tA)x γ(R , dt ;X) h |φ(tA)x|2 ,
+ t
0 t p
 16.4 Notes 565

which is a norm decidedly different from (16.34) when p 6= 2. However, in

this setting it provides a unified approach to many results of Dore and Da
Prato–Grisvard.
• The absolute functional calculus can be characterised in terms of gen-
eralised real interpolation spaces, where the role of Lp (R+ , dt
t ) is taken
over by more general Banach function spaces E over R+ . Essentially, a
standard sectorial operator A, acting on an intermediate spaces X for
a couple (X0 , X), where X0 and X1 are appropriate fractional domain
spaces of X, has an absolute calculus if and only if X = (X0 , X1 )θ,E
for some θ ∈ (0, 1) and an appropriate choice of such a Banach function
space E; see Kalton and Kucherenko [2010], where a precise statement
of the result an be found. This characterisation of the absolute calcu-
lus compares nicely with the characterisation of the H ∞ -calculus through
the γ-interpolation method and the close relationship of bounded imag-
inary powers with the complex interpolation method described in the
previous chapter. A necessary condition for the existence of a Banach
function space E with X = (X0 , X1 )θ,E is the monotonicity of the K-
functional for the couple (X0 , X1 ), in the sense that it has the property
that K(t, x; X0 , X1 ) 6 K(t, y; X0 , X1 ) for some x ∈ X0 + X1 and y ∈ X,
and all t > 0, then x ∈ X and kxkX 6 ckykX . In fact, this is where the
definition of the absolute calculus has its origin.
The proof of Theorem 16.3.19 follows Haase [2005], where some additional
details have been written out. The theorem also admits a homogeneous ver-
sion, which is presented in [Haase, 2006, Section 6.4]. A by-product of Theorem
16.3.19 is the equivalence of norms
kxk + kt 7→ t−θ φ(tA)xkLp (R+ , dt ;X) h kxk + kt 7→ t−θ ψ(tA)xkLp (R+ , dt ;X)
t t

1 ∞
for functions φ, ψ ∈ H (Σσ ) ∩ H (Σσ ) satisfying the conditions of Theorem
16.3.19. Interestingly, this equivalence of norms remains true under somewhat
weaker conditions on φ and ψ; see Haase [2006, Theorem 6.4.2]. The proof
follows the lines of the equivalence of continuous square functions in Chapter
10, with simplifications due to the fact that various subtleties in the handling
of γ-norms can now be avoided. This more general version of the equivalence
of norms covers the function φ(z) = z/(z + 1) which is implicit in the first
part of Proposition K.4.1.
Corollary 16.3.22 is a classical result due to Dore [1999]. This result was
subsequently generalised to standard sectorial operators in Dore [2001], where
it was shown that such operators have a bounded H ∞ -calculus on (X, D(A) ∩
R(A))θ,p ; see also Kalton and Kucherenko [2010], who establish their absolute
functional calculus.

Sums of non-commuting operators

In Section 16.3 we studied the closedness (and further properties) of sums
of operator A + B under the assumption that A and B are resolvent com-
566 16 Perturbations and sums of operators

muting. In this paragraph, we briefly comment on the closedness of sums

of non-commuting operators, provided suitable condition bounds and, some-
times, domain compatibility assumptions, are imposed on the commutator
[A, B] = AB − BA. The first such result was obtained by Da Prato and Gris-
vard [1975], who proved that the closure of A + B is invertible and sectorial
under commutator conditions. Closedness of A + B itself was proved under
further conditions in the case that X is a Hilbert space. Labbas and Terreni
[1987] obtained similar results under a different type of commutator condi-
tions, and Monniaux and Prüss [1997] proved a Dore–Veni type theorem for
non-commuting operators under the commutator condition of Labbas–Terreni.
In his PhD thesis, Štrkalj [2000] proved a version of Kalton and Weis [2001]
for non-commuting operators under the same condition as Labbas–Terreni in
the case that X is a B-convex space, and Prüss and Simonett [2007] proved
a similar result under either one of the above commutator conditions without
any restrictions on the space X. Moreover, under the condition that A and
B has a bounded H ∞ -calculus, with one of them R-bounded, it was shown
in this paper that A + B has a bounded H ∞ -calculus. Similar results were
proved under a different commutator conditions in Roidos [2018]. Products
of non-commuting operators have been considered in Štrkalj [2000], Haller-
Dintelmann and Hieber [2005].
Typical applications of these results include parabolic PDE on wedge or
cone domains, where an elliptic operator C can be split into two space direc-
tions to obtain simpler operators A and B. This type of application was worked
out in detail by Prüss and Simonett [2007], Prüss and Simonett [2006] and con-
tinued in Nau and Saal [2012], Maier and Saal [2014], Köhne, Saal, and West-
ermann [2021]. For another typical application to non-autonomous parabolic
problems, in which case ∂t and A(t) are non-commuting on Lp (0, T ; X), the
reader is referred to Di Giorgio, Lunardi, and Schnaubelt [2005a]. Applications
to hyperbolic problems appear in Alouini and Goubet [2014].
An interesting class of operator sums for non-commuting operators arises
in connection with (an abstract version of) the Weyl commutation relation
for position and momentum operators. The general theory of such operators
has an altogether different flavour due to its connections with the Heisenberg
group; the reader is referred to Putnam [1967] for a general overview. In
connection with the topics treated in this volume, the following result is worth
mentioning. Suppose two d-tuples of operators A = (A1 , . . . , Ad ) and B =
(B1 , . . . , Bd ) acting on a Banach space X are given such that iA1 , . . . , iAd and
iB1 , . . . , iBd generate bounded C0 -groups satisfying the Weyl commutation
relations
eisAj eitAk = eitAk eisAj , eisBj eitBk = eitBk eisBj
eisAj eitBk = e−istδjk eitBk eisAj .
Here, for clarity of exposition, we use exponential notation for the C0 -groups
involved. Under this condition, the operator sum
 16.4 Notes 567
d
1 2 1X 2
(A + B 2 ) = A + Bj2
2 2 j=1 j

is the abstract counterpart of the quantum harmonic oscillator. Under the

assumption that X is a UMD Banach lattice, it is shown in Van Neerven and
Portal [2020] (under an additional boundedness assumption of the Weyl cal-
culus associated with the pair (A, B)) and Van Neerven, Portal, and Sharma
[2023] that the operator 21 (A2 + B 2 ) − 12 d is R-sectoriality and has a bounded
of the H ∞ -calculus.
17
Maximal regularity

The present chapter is one of the central ones of this book project. Maximal
regularity provides a link between the general theory of operator-valued sin-
gular integrals and the theory of H ∞ -functional calculus with the regularity
theory for evolution equations. In some cases progress on these topics was
even motivated by applications to maximal regularity.
Maximal Lp -regularity for the Cauchy problem
(
u0 (t) + Au(t) = f (t), t ∈ I,
(17.1)
u(0) = 0,

with A an unbounded linear operator A on a Banach space and I = (0, T ) or

R+ , requires that if the right-hand side f of (17.1) belongs to Lp (I; X), then
(17.1) has a unique strong solution u that has the best regularity possible in
this situation. Namely, u0 and Au should both belong to Lp (I; X) as well, and

ku0 kLp (I;X) + kAukLp (I;X) 6 Cku0 + AukLp (I;X) = Ckf kLp (I;X) . (17.2)

This a priori estimate is often crucial to fixed point methods for more general
classes of non-linear partial differential equations with A as their linear main
part. This will be explained in more detail in Chapter 18.
Unfortunately, maximal Lp -regularity for A may fail even if −A is the
generator of an analytic semigroup on Lq (0, 1) (see Section 17.4.c). Proving
maximal Lp -regularity usually requires quite sophisticated methods. In earlier
chapters, we have prepared for three approaches to maximal regularity:
• Operator-valued singular integral operators:
Formally using the variation of constants formula for the solution of (17.2),
we have
Z t
Au(t) = Ae−(t−s)A f (s) ds.
0

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 569
T. Hytönen et al., Analysis in Banach Spaces, Ergebnisse der Mathematik und ihrer
Grenzgebiete. 3. Folge / A Series of Modern Surveys in Mathematics 76,
https://doi.org/10.1007/978-3-031-46598-7_7
570 17 Maximal regularity

Since kAe−tA k . 1t (and often even h), it is clear that we have to deal
with singular integrals to obtain maximal Lp -regularity. The details of this
approach will be presented in Section 17.3.a.

• The operator sum method:

This method refers to the inverse triangle inequality stated in (17.2). We
show that (17.2) can be obtained by applying the sum-of-operator theorems
d
of Chapter 16 with one operator chosen as D = dt on Lp (I; X) (see Section
17.3.c).
With both methods one can show that a sectorial operator A on a UMD
space X has maximal Lp -regularity if and only if A is R-sectorial of angle
ωR (A) < π/2. The notion of R-sectoriality was introduced in Chapter 10.
There is a third method which simplifies the problem somewhat:
• Interpolation theory:
Here, the real interpolation method is of particular interest because the prop-
erties of sectorial operators often drastically improve when considered on the
real interpolation space Lp (I; (X, D(A))θ,q ) or on a Besov space Bp,q s
(I; X)
as “regularity” space. In the first case, the part of A on (X, D(A))θ,q auto-
matically has a bounded H ∞ -calculus if 0 ∈ %(A) (see Corollary 16.3.22). In
the second case, since Besov spaces are real interpolation spaces a unified ap-
proach can be given via the absolute functional calculus, which was introduced
in Section 16.3.d, This approach will be presented in Section 17.3.c.
In Sections 17.1 and 17.2 we discuss the essential properties and variants
of maximal Lp -regularity which are useful in applications to evolution equa-
tions. For instance, we show that for an additional initial value u(0) = x in
(17.1), maximal Lp -regularity with 1 < p < ∞ is preserved if and only if
x ∈ (X, D(A))1− p1 ,p (see Section 17.2.b). In Section 17.2.e we discuss extrap-
olation of integrability and extrapolation to power weights; this allows the
treatment of a wider class of initial values. Perturbation results for maximal
Lp -regularity appear in Sections 17.2.g and 17.3.

17.1 The abstract Cauchy problem

Throughout this section we assume that A is a linear operator with domain
D(A) in a Banach space X. We are interested in finding solutions, in an
appropriate sense, to the inhomogeneous abstract Cauchy problem
(
u0 (t) + Au(t) = f (t), t ∈ I,
(ACPx )
u(0) = x,

where either I = R+ = (0, ∞) or I = (0, T ) is a bounded interval, and the

initial value x is taken from X. We will use the notation
 17.1 The abstract Cauchy problem 571

L1loc (I; X)

for the space of strongly measurable functions on I which are integrable on

every bounded interval contained in I. If I is a bounded interval, then of
course L1loc (I; X) = L1 (I; X).
Definition 17.1.1 (Strong solutions). A strongly measurable function u :
I → X is called a strong solution to the problem (ACPx ) associated with a
given f ∈ L1loc (I; X) if
(i) u takes values in D(A) almost everywhere and Au belongs to L1loc (I; X);
(ii) u solves the integrated version of (ACPx ), that is, for almost all t ∈ I
we have Z t Z t
u(t) + Au(s) ds = x + f (s) ds.
0 0

It is clear from condition (ii) in the above definition that a strong solution u
is equal almost everywhere to a continuous function on I. In what follows, we
will always work with this version, which satisfies the identity of (ii) for all
t ∈ I and satisfies
u(0) = x.
If (0, T0 ) is a bounded interval contained in I and u is an strong solution of
(ACPx ) on I, then by taking the supremum over [0, T0 ] in (ii) we obtain

kukC([0,T0 ];X) 6 kxk + kAukL1 (0,T0 ;X) + kf kL1 (0,T0 ;X) .

By Lemma 2.5.8 and Proposition 2.5.9, indefinite integrals of functions in

L1 (0, T ; X) belong to W 1,1 (0, T ; X) and are differentiable almost everywhere,
with derivative equal to the original function almost everywhere. In the present
setting we may apply this to Au and f to obtain that strong solutions u on I
are differentiable almost everywhere, and satisfy

u0 + Au = f almost everywhere on I (17.3)

The definition of a mild solution given in Volume II can be extended to

locally bounded strongly measurable semigroups as discussed in Appendix K:
Definition 17.1.2 (Mild solutions). Suppose that −A generates a locally
bounded strongly measurable semigroup S on X. For f ∈ L1loc (I; X) and x ∈
D(A), the continuous function u ∈ L1loc (I; X) defined by
Z t
u(t) := S(t)x + S ∗ f (t) := S(t)x + S(t − s)f (s) ds, t ∈ I,
0

is called the mild solution of the problem (ACPx ),

(
u0 (t) + Au(t) = f (t), t ∈ I,
u(0) = x.
572 17 Maximal regularity

Since we assume x ∈ D(A), by Proposition K.1.9 the mild solution u extends

continuously to I and satisfies u(0) = x.
In the case that A is densely defined (by Proposition 10.1.9, this is auto-
matic if X is reflexive), Theorem G.3.2 shows that the mild solution is also
the unique weak solution of (ACPx ). In the setting considered presently, we
do not assume A to be densely defined and weak solutions cannot be de-
fined. Instead, the next proposition states the equivalence of strong and mild
solutions. It uses the following terminology.
In line with Definition K.1.2 (and keeping in mind Remark K.1.4), a linear
operator (G, D(G)) in X is said to be the generator of a locally bounded
strongly measurable semigroup S satisfying kS(t)k 6 M eωt for all t > 0, if
the following conditions are satisfied:
(i) {λ ∈ C : <λ > ω} ⊆ %(G);
(ii) for all <λ > ω and x ∈ X we have
Z ∞
R(λ, G)x = e−λt S(t)x dt.
0

By strong measurability and exponential decay of the integrand, the integral

in (ii) converges as a Bochner integral in X. It follows from Proposition K.1.7
Rt
that for all x ∈ X and t > 0 we have 0 S(s)x ds ∈ D(G) and
Z t
G S(s)x ds = S(t)x − x;
0
Rt
if moreover x ∈ D(G), then both expressions are equal to 0
S(t)Gx ds. This
fact will be used repeatedly below.

Proposition 17.1.3. Let −A generate a locally bounded strongly measurable

semigroup S on a Banach space X. Let f ∈ Lp (I; X) with 1 6 p 6 ∞, and
let x ∈ D(A). Then for any function u ∈ C(I; X), the following assertions are
equivalent:
(1) u is a strong solution on I of (ACPx );
(2) u is the mild solution of (ACPx ) on I, u is differentiable almost every-
where, and u0 ∈ L1loc (I; X);
(3) u is the mild solution of (ACPx ), u takes values in D(A) almost every-
where, and Au ∈ L1loc (I; X).
In particular, a strong solution, if it exists, is unique, and it then equals the
mild solution.

Proof. Fix λ ∈ %(−A).

(1)⇒(2): We have already seen that u is differentiable almost everywhere
and weakly differentiable with u0 ∈ L1loc (I; X). It remains to prove that u is
the mild solution. Fix t ∈ I and define v : [0, t] → X by
 17.1 The abstract Cauchy problem 573

v(s) := (λ + A)−1 S(t − s)u(s) = S(t − s)(λ + A)−1 u(s).

Then v ∈ C([0, t]; X) by the strong continuity of S on D(A).

By Proposition K.1.9, for all x0 ∈ D(A) the function vx0 (s) := S(t−s)(λ+
A)−1 x0 is continuously differentiable on (0, t). The product rule, which may
be applied in the points where u is differentiable, gives that v is differentiable
almost everywhere on (0, t) with derivative

v 0 (s) = (λ + A)−1 S(t − s)u0 (s) + A(λ + A)−1 S(t − s)u(s).

Since u is assumed to be a strong solution, it takes values in D(A) almost

everywhere and therefore

v 0 (s) = (λ + A)−1 S(t − s)(u0 (s) + Au(s)) = (λ + A)−1 S(t − s)f (s)

for almost all s ∈ (0, t). This implies v ∈ W 1,p (0, t; X) and, by Proposition
2.5.9,
Z t
−1
v(t) − (λ + A) S(t)x = v(t) − v(0) = v 0 (s) ds
0
Z t
−1
= (λ + A) S(t − s)f (s) ds.
0

This gives u(t) = (λ + A)v(t) = S(t)x + S ∗ f (t) as required.

To prove the remaining implications, we first derive a formula in case u is
the mild solution. Integrating A(λ + A)−1 u over [0, t] gives
Z t Z tZ s
A(λ + A)−1 (u(s) − S(s)x) ds = A(λ + A)−1 S(s − r)f (r) dr ds
0 0 0
Z tZ t
= A(λ + A)−1 S(s − r)f (r) ds dr
0 r
Z t Z t
= A S(s − r)(λ + A)−1 f (r) ds dr
0 r
Z t
=− (λ + A)−1 [S(t − r)f (r) − f (r)] dr
0
 Z t 
= −(λ + A)−1 u(t) − S(t)x − f (r) dr ,
0
(17.4)

where Rwe used the definition of a mild solution and Hille’s theorem; the iden-
t
tity A r S(s − r)x ds = −[S(t − r)x − x] which as used in this computation
is justified by Proposition K.1.7.
(2)⇒(3): Since u is almost everywhere differentiable with u0 ∈ L1loc (I; X),
and (λ + A)−1 S(t)x = S(t)(λ + A)−1 x is differentiable with derivative A(λ +
A)−1 S(t)x, (17.4) and Proposition 2.5.9 imply
574 17 Maximal regularity

A(λ + A)−1 u(t) = −(λ + A)−1 [u0 (t) − f (t)] almost all t ∈ I.
Combining this with A(λ + A)−1 u(t) = u(t) − λ(λ + A)−1 u(t) we find that
u ∈ D(A) almost everywhere and Au = −u0 + f ∈ L1loc (I; X).
(3)⇒(1): Since u ∈ D(A) almost everywhere and Au ∈ L1loc (I; X), (17.4)
implies
Z t Z t
−1 −1
(λ + A) Au(s) ds − A(λ + A) S(s)x ds
0 0
h Z t i
= −(λ + A)−1 u(t) − S(t)x − f (s) ds .
0
−1
Rt
It follows from this that A(λ+A) 0
S(s)x ds ∈ D(A), which in turn implies
Rt
that 0 S(s)x ds ∈ D(A). Applying λ + A on both sides, we obtain
Z t Z t Z t
Au(s) ds + A S(s)x ds = −u(t) + S(t)x + f (s) ds.
0 0 0
Rt
Since A 0
S(s)x ds = S(t)x − x, this shows that u is a strong solution. 
The following proposition provides a large class of functions f : [0, T ] → X
for which the mild solution u = S ∗ f to (ACP0 ) is a strong solution.
Proposition 17.1.4. Let A be sectorial of angle < 21 π and let S be the ana-
lytic semigroup generated by −A. Then for all f ∈ C α ([0, T ]; X) with α > 0,
the mild solution u = S ∗ f to (ACP0 ) satisfies
u ∈ C([0, T ]; X) ∩ L∞ (0, T ; D(A)).
In particular, u is a strong solution to (ACP0 ).
Proof. In the discussion following Definition 17.1.2 we have already seen that
u ∈ C([0, T ]; X). In the remainder of the proof there is no loss of generality
in assuming that 0 < α < 1. For t ∈ [0, T ] we can write
Z t Z t
u(t) = S(t − r)(f (r) − f (t)) dr + S(t − r)f (t) ds =: u1 (t) + u2 (t).
0 0

Since by (K.4) we have sups∈(0,T ] ksAS(s)k < ∞, we find that u1 (t) ∈ D(A)
and
Z t
kAu1 (t)k 6 kAS(t − r)(f (r) − f (t))k dr
0
Z T
6 [f ]C α ([0,T ];X) rα−1 kAS(r)k dr.
0

From Proposition K.1.11 we see that u2 (t) ∈ D(A) and Au2 (t) = (I−S(t))f (t),
and thus
kAu2 (t)k 6 sup kI − S(t)k sup kf (t)k.
t∈[0,T ] t∈[0,T ]

 17.2 Maximal Lp -regularity 575

17.2 Maximal Lp -regularity

After these preparation on the Cauchy problem, we will now turn to the main
topic of the chapter.

17.2.a Definition and basic properties

We now discuss solutions, in an appropriate Lp sense, to the inhomogeneous

abstract Cauchy problem (ACPx ).

Definition 17.2.1 (Lp -solutions). A strong solution u to (ACPx ) associated

with a function f ∈ Lp (I; X) is called an Lp -solution if Au ∈ Lp (I; X).

To connect this to Definition 17.1.1 of strong solutions, note that functions

in Lp (I; X) belong to L1loc (I; X). A strong solution on I is an L1 -solution on
every bounded subinterval contained in I. From this point onwards, in the
context of Lp -solutions it will always be understood that the inhomogeneities
f belong to Lp (I; X) even when this is not explicitly mentioned.
We will now concentrate, for a while, on the special initial value x = 0,
i.e., the Cauchy problem
(
u0 (t) + Au(t) = f (t), t ∈ I,
(ACP0 )
u(0) = 0.

If (0, T ) is a bounded interval contained in I and u is an Lp -solution of (ACP0 )

on I, then by using Hölder’s inequality and taking the supremum over t ∈ [0, T ]
we obtain
0
kukC([0,T ];X) 6 T 1/p (kAukLp (0,T ;X) + kf kLp (0,T ;X) ), (17.5)

where p1 + p10 = 1. By another application of Hölder’s inequality it follows that

u ∈ Lp (0, T ; X) and

kukLp (0,T ;X) 6 T (kAukLp (0,T ;X) + kf kLp (0,T ;X) ). (17.6)

In contrast to (17.6), an Lp -solution on I = R+ need not belong to Lp (R+ ; X).

In fact, we will see in Proposition 17.2.8 and Corollary 17.2.25 that the fol-
lowing equivalence holds: (ACP0 ) admits an Lp -solution u ∈ Lp (R+ ; X) for
every f ∈ Lp (R+ ; X) if and only if 0 ∈ %(A).
Concrete examples of Lp -solutions that do not belong to Lp (R+ ; X) will
be presented in Section 17.4. It is true, however, that if u is an Lp -solution
on R+ , then for every ε > 0 the rescaled function t 7→ e−εt u(t) belongs to
Lp (R+ ; X). Indeed, by (17.5),
X
kt 7→ e−εt u(t)kLp (R+ ;X) 6 e−ε(k−1) kukLp (k−1,k;X)
k>1
576 17 Maximal regularity
X 0
6 k 1/p e−ε(k−1) (kAukLp (R+ ;X) + kf kLp (R+ ;X) ).
k>1

This proves the claim.

By (17.3) we have u0 + Au = f almost everywhere on I and therefore
ku0 kLp (I;X) 6 kAukLp (I;X) + kf kLp (I;X) . (17.7)
As a result we have:
Proposition 17.2.2. If u is an Lp -solution of (ACP0 ) on I, then
(1) u ∈ 0 ẆA1,p (I; X) ⊆ 0 Ẇ 1,p (I; X), where
1,p ¯
0 Ẇ
1,p
(I; X) := {v ∈ Wloc (I; X) : v 0 ∈ Lp (I; X), v(0) = 0},
1,p 1,p ¯
0 ẆA (I; X) := {v ∈ 0 Ẇ (I; X) : v(·) ∈ D(A) a.e. on I, Av ∈ Lp (I; X)}
are normed spaces with
kvk0 Ẇ 1,p (I;X) := kv 0 kLp (I;X) ,
kvk0 Ẇ 1,p (I;X) := max kv 0 kLp (I;X) , kAvkLp (I;X) ;

A

(2) for all bounded subintervals (0, T ) ⊆ I we have

1,p
0 ẆA (I; X) ,→ Lp (0, T ; D(A)) ∩ W 1,p (0, T ; X)
with

max kvkLp (0,T ;D(A)) , kvkW 1,p (0,T ;X) 6 (T + 1)kvk0 Ẇ 1,p (I;X) . (17.8)
A

In particular, the Lp -solution u belongs to these intersections.

Proof. (1): By Definitions 17.2.1 and 17.1.1, u(0) = 0 and u takes values in
D(A) a.e., and Au ∈ Lp (I; X). By the discussion following Definition 17.1.1, it
¯
also follows that u ∈ C([0, T ]; X) ⊆ Lp ([0, T ]; X) for every bounded [0, T ] ⊆ I,
1,1 ¯ 0 p
as well as u ∈ Wloc (I; X). From (17.7), we get u ∈ L (I; X), and thus
u ∈ 0 ẆA1,p (I; X), as claimed.
The only axiom of a norm that might not be entirely evident is to check
that kvk0 Ẇ 1,p (I;X) = 0 only if v = 0. However, by Proposition 2.5.9, every
Rt
v ∈ 0 Ẇ 1,p (I; X) satisfies v(t) = 0 v 0 (s) ds almost everywhere, and hence
indeed v = 0 if kvk0 Ẇ 1,p (I;X) = kv 0 kLp (I;X) = 0.
(2): Recall that D(A) ⊆ X is normed by kxkD(A) := kxkX + kAxkX . Thus
kvkLp (0,T ;D(A)) 6 kvkLp (0,T ;X) + kAvkLp (0,T ;X)
6 T kv 0 kLp (I;X) + kAvkLp (I;X) 6 (T + 1)kvk0 Ẇ 1,p (I;X) .
A

Similarly,
kvkW 1,p (0,T ;D(A)) 6 kvkLp (0,T ;X) + kv 0 kLp (0,T ;X)
6 T kv 0 kLp (I;X) + kv 0 kLp (I;X) 6 (T + 1)kvk0 Ẇ 1,p (I;X) .
A

Thus the maximum of the two left-hand sides has this same upper bound. 
 17.2 Maximal Lp -regularity 577

Proposition 17.2.3. If A : D(A) → X is a closed operator, then 0 ẆA1,p (I; X)

is a Banach space.

Proof. If un ∈ 0 Ẇ 1,p (I; X) is a Cauchy sequence, then un0 ∈ Lp (I; X) have a

Rt
limit v in this space by the completeness of Lp (I; X). Then u(t) := 0 v(s) ds
belongs to 0 Ẇ 1,p (I; X), and clearly un → u in 0 Ẇ 1,p (I; X).
Let then un ∈ 0 ẆA1,p (I; X) be a Cauchy sequence, and u ∈ 0 Ẇ 1,p (I; X)
its limit in this larger space. By Proposition 2.5.9, we have
0 0
kun − ukC([0,T ];X) 6 T 1/p kun0 − u0 kLp (I;X) = T 1/p kun − uk0 Ẇ 1,p (I;X)

for every bounded interval [0, T ] ⊆ I. ¯ Hence un (t) → u(t) for t ∈ [0, T ]
¯
and, [0, T ] ⊆ I being arbitrary, for t ∈ I. (Here we are thinking of con-
tinuous versions of the Sobolev functions in question.) On the other hand,
Aun ∈ Lp (I; X) is a Cauchy sequence, hence convergent to some limit
w ∈ Lp (I; X), and thus a subsequence converges almost everywhere. If t ∈ I
is such a point, then un (t) → u(t) while Aun (t) → w(t). Since A is closed, it
follows that u(t) ∈ D(A) and Au = w ∈ Lp (I; X). This shows that the limit
u ∈ 0 Ẇ 1,p (I; X) in fact belongs to 0 ẆA1,p (I; X). Since un → u in 0 Ẇ 1,p (I; X)
and Aun → w = Au in Lp (I; X), it follows that un → u ∈ 0 ẆA1,p (I; X). 
We now come to the principal notion to be studied in this chapter. Here,
and in the remainder of this subsection, we assume that 1 6 p 6 ∞ unless
explicitly stated otherwise.

Definition 17.2.4 (Maximal Lp -regularity). A linear operator A has max-

imal Lp -regularity on I if there exists a constant C > 0 such that for all
f ∈ Lp (I; X) problem (ACP0 ) admits a unique Lp -solution uf on I and

kAuf kLp (I;X) 6 Ckf kLp (I;X) . (17.9)

The least admissible constant in this definition will be called the maximal
reg
Lp -regularity constant of A on I and will be denoted by Mp,A (I).
p
If A has maximal L -regularity on I, then for bounded subintervals
(0, T ) ⊆ I, the estimate (17.5) implies that if uf is an Lp -solution of (ACP0 )
with f ∈ Lp (I; X), then its restriction to (0, T ) satisfies
reg
kuf kLp (0,T ;X) 6 T 1/p kuf kC([0,T ];X) 6 T (Mp,A (I) + 1)kf kLp (I;X) . (17.10)

Moreover, (17.7) implies that if u is a Lp -solution to (ACP0 ) Unfortunately, we

cannot yet conclude that A has maximal Lp -regularity on (0, T ) ⊆ I, because
uniqueness of the solution on (0, T ) is unclear. We come back to this issue in
Lemma 17.2.16.
Let us immediately observe that maximal Lp -regularity implies closedness:

Proposition 17.2.5 (Closedness). If a linear operator A has maximal Lp -

regularity on I, then A is a closed operator.
578 17 Maximal regularity

Proof. Let xn → x in X, with xn ∈ D(A), and Axn → y in X. Choose a non-

zero non-negative function φ ∈ Cc1 (I) and put un := φ ⊗ xn and u := φ ⊗ x.
Set fn := u0n + Aun = φ0 ⊗ xn + φ ⊗ Axn and f := φ0 ⊗ x + φ ⊗ y. Then un
is the unique Lp -solution associated with fn . Let v be the unique Lp -solution
associated with f . Since fn → f in Lp (I; X), it follows from (17.10) (applied
with fn − f ) that un → v in C([0, T ]; X) for all [0, T ] ⊆ I. Since un → u in
Cb (I; X) it follows that u = v. Therefore u is an Lp -solution associated with
f . This forces x ∈ D(A), which implies that φ ⊗ Ax = Au = f − u0 = φ ⊗ y,
so Ax = y. This proves that A is closed. 

Corollary 17.2.6. Let A be a linear operator with maximal Lp -regularity

on I. For f ∈ Lp (I; X) let uf denote the corresponding Lp -solution in
Lp (I; D(A)) ∩ W 1,p (I; X). Then the mapping

M : Lp (I; X) → 0 ẆA1,p (I; X), M f = uf ,

is an isomorphism with
1 reg
kf kLp (I;X) 6 kM f k0 Ẇ 1,p (I;X) 6 (Mp,A (I) + 1)kf kLp (I;X)
2 A

Proof. If A has maximal Lp -regularity on I, then (17.7) implies that

reg
ku0f kLp (I;X) 6 (Mp,A (I) + 1)kf kLp (I;X) , (17.11)

and hence, in combination with (17.9), that

reg
kuf k0 Ẇ 1,p (I;X) 6 (Mp,A (I) + 1)kf kLp (I;X) . (17.12)
A

The existence and uniqueness of the Lp -solution uf implies that the mapping
f 7→ uf is linear from Lp (I; X) into 0 ẆA1,p (I; X).
Conversely, given u ∈ 0 ẆA1,p (I; X), we have f := u0 + Au ∈ Lp (I; X) with

kf kLp (I;X) 6 ku0 kLp (I;X) + kAukLp (I;X) 6 2kuk0 Ẇ 1,p (I;X) ,
A

and it is evident that u is an Lp -solution of (ACP0 ) with datum f . Thus, by

uniqueness, u = uf is in the range of M , so this operator is onto. 
We recall from Section C that for a Banach couple (X, Y ) the intersection
X ∩ Y is a Banach space with respect to the norm defined by

kxkX∩Y := max kxkX , kxkY .

Combining the estimate in the definition of maximal Lp -regularity with

(17.11) and (17.10) we obtain:

Proposition 17.2.7. Suppose that A has maximal Lp -regularity on I. If u

is the unique Lp -solution on I of (ACP0 ), then for all bounded subintervals
(0, T ) ⊆ I we have u ∈ Lp (0, T ; D(A)) ∩ W 1,p (0, T ; X) and
 17.2 Maximal Lp -regularity 579
reg
kukLp (0,T ;D(A))∩W 1,p (0,T ;X) 6 (1 + T )(Mp,A (I) + 1)kf kLp (I;X) .

Moreover, if I = (0, T ), then

1,p
0 ẆA (I; X) ' Lp (0, T ; D(A)) ∩ W 1,p (0, T ; X)

and for all v in this space, we have

kvk0 Ẇ 1,p (I;X) 6 kvkLp (0,T ;D(A))∩W 1,p (0,T ;X) 6 (T + 1)kvk0 Ẇ 1,p (I;X) .
A A

Proof. This first bound is immediate by combining (17.8) and (17.12). For
the comparison of the spaces, the first estimate is clear, and the second one
is (17.8). 
The next result gives another sufficient condition, also allowing infinite inter-
vals I, that Lp -solutions u do in fact belong to Lp (I; X).

Proposition 17.2.8. Let A be a linear operator with maximal Lp -regularity

on I. If 0 ∈ %(A), then
1,p
0 ẆA (I; X) ' Lp (I; D(A)) ∩ 0 W 1,p (I; X)

and, for all u in this space,

kuk0 Ẇ 1,p (I;X) 6 kukLp (I;D(A))∩0 W 1,p (I;X) 6 (kA−1 k + 1)kuk0 Ẇ 1,p (I;X) .
A A

p p
In this situation, for all f ∈ L (I; X), the unique L -solution uf to (ACP0 )
belongs to this space and satisfies
reg
kuf kLp (I;D(A))∩W 1,p (I;X) 6 (kA−1 k + 1)(Mp,A (I) + 1)kf kLp (I;X) .

Proof. The boundary value v(0) = 0 is assumed in both spaces, and the first
norm estimate is evident. The second estimate follows by taking the maximum
of

kukW 1,p (R+ ;X) 6 kukLp (R+ ;X) + ku0 kLp (R+ ;X)
6 kA−1 k kAukLp (R+ ;X) + ku0 kLp (R+ ;X)
6 (kA−1 k + 1)kuk0 Ẇ 1,p (R+ ;X)
A

and

kukLp (R+ ;D(A)) 6 kukLp (R+ ;X) + kAukLp (R+ ;X)

6 (kA−1 k + 1)kAukLp (R+ ;X) 6 (kA−1 k + 1)kuk0 Ẇ 1,p (R+ ;X) .
A

From these bounds and (17.12), the final claim concerning uf follows. 

Corollary 17.2.9. Let A be a linear operator with maximal Lp -regularity on

I, where we assume that at least one of the following conditions holds:
580 17 Maximal regularity

• I = (0, T ) is bounded;
• 0 ∈ %(A).
For f ∈ Lp (I; X) let uf denote the corresponding Lp -solution in Lp (I; D(A))∩
W 1,p (I; X). Then
1,p
0 ẆA (I; X) ' Lp (I; D(A)) ∩ 0 W 1,p (I; X);

for any v in this space, we have

kvk0 Ẇ 1,p (I;X) 6 kvkLp (I;D(A))∩W 1,p (I;X) 6 (min{kA−1 k, T } + 1)kvk0 Ẇ 1,p (I;X) ;
A A

and M : f 7→ uf is an isomorphism from Lp (I; X) into this space.

Proof. This follows by combining Corollary 17.2.6 with either Proposition

17.2.7 for I = (0, T ), or Proposition 17.2.8 for 0 ∈ %(A). (If I = (0, T ) and
0 ∈ %(A), we could apply either Proposition 17.2.7 or Proposition 17.2.8
for the same qualitative conclusion, but one or the other may give a better
quantitative bound, depending on the relative size of T and kA−1 k.) 
By Proposition 17.2.5, maximal Lp -regularity implies closedness. If the op-
erator A is assumed closed to begin with, the next two results provide two
different ways of apparently relaxing Definition 17.2.4, yet actually defining
the same property in the end. Namely, we may either require
• the existence of a unique Lp -solution with estimate (17.9) only on a dense
subspaces of functions f ∈ Lp (I; X) (Proposition 17.2.10), or
• the existence of a unique Lp -solution for all f ∈ Lp (I; X), but without
postulating any estimate on their size (Proposition 17.2.11).

Proposition 17.2.10. Let A be a closed linear operator, and let F be a dense

subspace of Lp (I; X). Suppose that for all f ∈ F there exists a unique Lp -
solution uf to (ACP0 ), and that this solution satisfies (17.9), that is,

kAuf kLp (I;X) 6 Ckf kLp (I;X) ,

with a constant C independent of f ∈ F . Then A has maximal Lp -regularity

reg
on I with Mp,A (I) 6 C.

Proof. Let f ∈ Lp (I; X) be arbitrary, and suppose that u1 and u2 are Lp -

solutions to (ACP0 ) on I. Then u = u1 − u2 is an Lp -solution to the homo-
geneous problem u0 + Au = 0 on I and u(0) = 0. Therefore, the uniqueness
assumption gives that u = 0, and therefore u1 = u2 .
Let again f ∈ Lp (I; X) be arbitrary, choose functions fn ∈ F converging
to f in Lp (I; X), and denote the unique Lp -solutions associated with fn by
un . By the assumed estimate, the functions Aun form a Cauchy sequence
in Lp (I; X) and therefore converge to a limit v in Lp (I; X). By (17.10) the
functions un converge uniformly on every bounded interval [0, T ] contained
 17.2 Maximal Lp -regularity 581

in I to a continuous function u on [0, T ]. As a result, by taking limits in the

definition of a strong solution we obtain
Z t Z t
u(t) + v(s) ds = f (s) ds.
0 0

Since A is closed, a standard subsequence argument furthermore gives that

v takes values in D(A) almost surely and v = Au in Lp (I; X). It follows
that u is an Lp -solution of (ACP0 ). Moreover, taking limits we also see that
kAukLp (I;X) 6 Ckf kLp (I;X) . 

Proposition 17.2.11. If A is closed and for all f ∈ Lp (I; X) there exists

a unique Lp -solution uf of (ACP0 ), then for all f ∈ Lp (I; X) the estimate
(17.9) holds, that is,

kAuf kLp (I;X) 6 Ckf kLp (I;X) ,

In particular, A has maximal regularity on I.

Proof. By Proposition 17.2.2, we have uf ∈ 0 ẆA1,p (I; X). The existence and
uniqueness of uf for every f ∈ Lp (I; X) guarantees that M : f 7→ uf is a
linear mapping from Lp (I; X) to 0 ẆA1,p (I; X). Since A is closed, 0 ẆA1,p (I; X)
is a Banach space by Proposition 17.2.3.
To prove that M is bounded, by the closed graph theorem, it suffices to
check that it is closed. To this end, let fn → f in Lp (I; X) and un := M fn → v
in 0 ẆA1,p (I; X). Since u0n +Aun = fn , taking limits gives v 0 +Av = f . It follows
that v is an Lp -solution corresponding to f . The uniqueness of Lp -solutions
therefore gives v = M f .
Thus M is closed, hence bounded, and therefore

kAuf kLp (I;X) 6 kuf k0 Ẇ 1,p (I;X) 6 Ckf kLp (I;X) ,

A

which is the claimed maximal regularity estimate. 

We end this section by showing a commutation relation between A and the
solution operator. It will only be presented in the case I = R+ , because that
is the case we need in the proof of Theorem 17.2.15 below. It also holds for
bounded intervals; this can easily be deduced from Theorem 17.2.19.

Lemma 17.2.12. Suppose that A has maximal Lp -regularity on R+ , and

denote by M : f 7→ uf the solution operator that assigns to a function
f ∈ Lp (R+ ; X) the corresponding Lp -solution uf of (ACP0 ). Then for ev-
ery f ∈ Lp (R+ ; D(A)) one has M f ∈ Lp (R+ ; D(A)) and

AM f = M Af

as functions in Lp (R+ ; X).

582 17 Maximal regularity

Here, for functions g ∈ Lp (R+ ; D(A)) we define Ag ∈ Lp (R+ ; X) pointwise

almost everywhere by (Ag)(t) := A(g(t)).
Proof. Let D : 0 Ẇ 1,p (I; X) → Lp (I; X) be given by Df := f 0 and C :
Lp (I; X) → 0 Ẇ 1,p (I; X) by
Z t Z t
Cf (t) := e−s f (t − s) ds = e−(t−s) f (s) ds.
0 0

This is well defined, since by Young’s inequality (see Lemma 14.2.3)

Z t
kCf kLp (I;X) 6 t 7→ e−s (1I f )(t − s) ds
0 Lp (R+ ;X)
−s
6 ks 7→ e kL1 (R+ ) kf kLp (I;X) = kf kLp (I;X) .

Moreover, for all f ∈ Lp (I; x) we have

(I + D)Cf = f.

Indeed, for f ∈ Lp (I; X) the product rule and an integration by parts give
h Z t i
DCf (t) = ∂t e−t es f (s) ds
0
Z t
−t
= −e es f (s) ds + f (t) = −Cf (t) + f (t).
0

Similarly, C(I +D)f = f for f ∈ 0 Ẇ 1,p (X) by integration by parts. Therefore,

C = (I + D)−1 .
Recall that M is bounded as an operator from Lp (I; X) to 0 ẆA1,p (I; X).
Before we give the final argument we state some identities which will be
needed. A key identity is

(A + D)M g = g = M (A + D)g, g ∈ 0 ẆA1,p (I; X),

where the first identity (valid more generally for all g ∈ Lp (I; X)) is a re-
statement of the equation that a solution M g with datum g must satisfy,
and the second one follows by noting that g ∈ 0 ẆA1,p (I; X) has the correct
boundary value g(0) = 0 by definition, and evidently solves the equation with
datum (A + D)g, so it must be the Lp -solution by uniqueness. Therefore, if
we can prove that M and D commute in a suitable sense, the desired result
will follow.
First observe that, by Hille’s theorem (Theorem 1.2.4),

A(I + D)−1 f = (I + D)−1 Af, 1

f ∈ Lloc ¯ D(A)).
(I;

Next we check that BM f = M Bf for f ∈ Lp (R+ ; X) ⊆ E(X) and for

B ∈ {(I + D)−1 , D(I + D)−1 }. Clearly, it suffices to consider the case B =
(I + D)−1 . Let u := M f and v := (I + D)−1 u. Then,
 17.2 Maximal Lp -regularity 583
Z t Z t
v 0 (t) + Av(t) = e−s u0 (t − s) ds + e−s Au(t − s) ds
0 0
Z t
= e−s f (t − s) ds = (I + D)−1 f (t),
0

from which it follows that v = M (I + D)−1 f . This gives the required result.
With these preparations out of the way, we can turn to the proof of the
lemma. It remains to show that for f ∈ Lp (R+ ; D(A)) one has M Af = AM f .
By the previous observations,

(1 + D)−1 M Af = M (1 + D)−1 Af
= M A(1 + D)−1 f
= M (A + D)(1 + D)−1 f − M D(1 + D)−1 f
= (1 + D)−1 f − D(1 + D)−1 M f
= (1 + D)−1 (f − DM f )
= (1 + D)−1 (AM f ).

The required identity is obtained by applying (1 + D) on both sides. 

17.2.b The initial value problem

Up to this point we have considered existence and uniqueness of Lp -solutions

for the inhomogeneous problem (ACP0 ) with zero initial condition. As we
will show presently, by using maximal Lp -regularity it is possible to obtain
necessary and sufficient conditions for the existence and uniqueness of Lp -
solutions for the problem (ACPx ) with non-zero initial conditions,
(
u0 (t) + Au(t) = f (t), t ∈ I,
u(0) = x,

1
where f ∈ Lloc (I; X) and x ∈ X; as before, I is either a bounded interval
(0, T ) or R+ = (0, ∞). An obvious necessary condition for the existence of
an Lp -solution u is that f should be in Lp (I; X); this is immediate from the
definition of an Lp -solution, which imply that u0 and Au belong to Lp (I; X),
and we then have the trivial bound

kf kLp (I;X) 6 ku0 kLp (I;X) + kAukLp (I;X) .

In view of this observation, in what follows it will always be assumed that

f ∈ Lp (I; X) even when this is not stated explicitly. We also maintain the
standing assumption 1 6 p 6 ∞ which is in force unless explicitly stated
otherwise.
We begin with a necessary condition on the initial condition x for the
existence of an Lp -solution. As a preliminary observation we note that as
584 17 Maximal regularity

in Proposition 17.2.2 one shows that if u is an Lp -solution to (ACPx ), then

u ∈ Lp (0, T ; D(A)) ∩ W 1,p (0, T ; X) for all bounded intervals (0, T ) ⊆ I.
In the next two results we interpret (X, D(A))0,∞ and (X, D(A))1,∞ as X
and D(A), respectively. It will be necessary to use the equivalent norms

k · k(X,D(A))1− 1 ,p h k · kTr
(X,D(A))
p 1− 1 ,p
p

of Theorem L.2.3.

Proposition 17.2.13 (Necessary condition on the initial condition).

Let 1 6 p < ∞. If for some f ∈ Lp (I; X), the problem (ACPx ) admits an
Lp -solution on I, then x ∈ (X, D(A))1− p1 ,p . If (0, T ) ⊆ I is a bounded interval,
then

kxkTr

(X,D(A)) 1 6 (1 + 1/T ) max kukLp (0,T ;D(A)) , kukW 1,p (0,T ;X) ;
1− ,p
p

if I = R+ and u ∈ Lp (R+ ; X), then

kxkTr 6 max kukLp (R+ ;D(A)) , ku0 kLp (R+ ;X) .


(X,D(A)) 1 1− ,p
p

Proof. We will prove that x ∈ (X, D(A))Tr

1− 1 ,p
, and make the estimates related
p
to the corresponding norm. That x ∈ (X, D(A))1− p1 ,p is then immediate using
the equivalence of the spaces from Theorem L.2.3.
By Definition L.2.1 (and the discussion after it), this involves identifying x
1,1
with the initial value x = v(0) of some v ∈ Wloc ([0, ∞); X) (now that X0 = X
and X1 = D(A), the sum space is simple X0 + X1 = X) such that
1
kt 7→ t1− p v 0 (t)kLp (R+ , dt ;X) = kv 0 kLp (R+ ;X)
t

and
1
kt 7→ t1− p v(t)kLp (R+ , dt ;D(A)) = kvkLp (R+ ;D(A))
t

1,p ¯
are finite. By assumption, we know that x = u(0), where u = uf ∈ Wloc (I; X);
however, it might be that I is only a finite interval, and even if I = R+ , the
norm kukLp (R+ ;X) 6 kukLp (R+ ;D(A)) need not be finite.
To fix both problems at once, let (0, T ) ⊆ I be finite, and let φ(t) :=
(1 − t/T )+ . Then v(t) := φ(t)u(t) is defined on all R+ , and it still satisfies
v(0) = u(0) = x. Since v 0 = φ0 u + φu0 , where |φ0 | 6 T1 1(0,T ) and |φ| 6 1(0,T ) ,
we obtain
kvkLp (R+ ;D(A)) 6 kukLp (0,T ;D(A))
6 kukLp (0,T ;X) + kAukLp (0,T ;X)

and
 17.2 Maximal Lp -regularity 585

1 1 
kv 0 kLp (R+ ;X) 6 kukLp (0,T ;X) + ku0 kLp (0,T ;X) 6 + 1 kukW 1,p (0,T ;X)
T T
The Lp -norms of Au and u0 above are clearly finite for an Lp -solution to
(ACPx ), and so is the Lp -norm of u on a finite (0, T ) ⊆ I by
Z t
kukLp (0,T ;X) = t 7→ x + u0 (s) ds
0 Lp (0,T ;X)

6 T 1/p kxkX + T ku0 kLp (0,T ;X) .

Thus we have checked that x ∈ (X, D(A))Tr

1− 1 ,p
. Referring to Definition
p
L.2.1 again, we also have the bound

kxk(X,D(A))Tr 6 max{kv 0 kLp (R+ ;X) , kvkLp (R+ ;D(A)) }.

1− 1 ,p
p

The claimed bound dealing with a bounded (0, T ) ⊆ I follows from this at
once. The case of I = R+ follows by taking the limit T → ∞, noting that the
term T1 kukLp (0,T ;X) → 0 in this case. 

Proposition 17.2.14 (Sufficient condition on the initial condition).

If A has maximal Lp -regularity on I, then for all f ∈ Lp (I; X) and x ∈
(X, D(A))1− p1 ,p the problem (ACPx ) admits a unique Lp -solution u. Moreover,
for 1 < p 6 ∞,
reg reg
kAukLp (I;X) 6 3Mp,A (I)kxkTr
(X,D(A)) + Mp,A (I)kf kLp (I;X) .
1− 1 ,p
p

For p = 1, the constant 3 needs to be replaced by a constant depending on A.

We first prove this for 1 < p 6 ∞; the case p = 1 will be established below
Lemma 17.2.22; there, also the precise estimate for kAukL1 (I;X) is stated (as it
turns out, the constant 3 gets replaced by the supremum over I of the norms
of the semigroup generated by −A).
Proof for 1 < p 6 ∞. Uniqueness of Lp -solutions follows from maximal Lp -
regularity. For the proof of existence we first consider the case 1 < p < ∞.
By Theorem L.2.3, for every ε > 0 we can find a function g ∈ Lp (R+ ; D(A)) ∩
W 1,p (R+ ; X) such that g(0) = x and

max kgkLp (R+ ;D(A)) , kgkW 1,p (R+ ;X) 6 (1 + ε)kxkTr


(X,D(A)) 1 . 1− ,p
p

Let v be the unique Lp -solution of the problem

v 0 + Av = f + g 0 + Ag, v(0) = 0.

The idea of this construction is that we use g to reduce matters to a problem

with zero initial condition, which has been studied in the previous section. By
586 17 Maximal regularity

maximal Lp -regularity, the above problem has a unique Lp -solution v, and it

satisfies
reg
kAvkLp (I;X) 6 Mp,A (I)kf + g 0 + AgkLp (I;X)
reg reg Tr
6 Mp,A (I)kf kLp (I;X) + 2(1 + ε)Mp,A (I)kxk(X,D(A)) .
1− 1 ,p
p

Now u := v + g is an Lp -solution to (ACPx ), and it satisfies

kAukLp (I;X) 6 kAvkLp (I;X) + kAgkLp (I;X)

reg reg
6 Mp,A (I)kf kLp (I;X) + 3(1 + ε)Mp,A (I)kxkTr
(X,D(A)) .
1− 1 ,p
p

Since ε > 0 was arbitrary, this proves the result.

If p = ∞, the above argument can be repeated with g ≡ x. 
By combining the necessary and sufficient conditions on the data x and f ,
for bounded intervals I we obtain that maximal Lp -regularity implies the
following norm isomorphism between the data (x, f ) on the on hand, and the
solution u = ux,f on the other hand, embodied by the mapping

X1− p1 , p1 × Lp (I; X) → Lp (I; D(A)) ∩ W 1,p (I; X)

which maps the data (x, f ) to the solution u; this remains correct for I = R+
if 0 ∈ %(A). This extends Corollary 17.2.9 to the case of non-zero initial data.

17.2.c The role of semigroups

The aim of this section is to prove that if A has maximal Lp -regularity on

a bounded interval (0, T ), then −A generates an analytic semigroup on X,
and that this semigroup is bounded analytic if A has maximal Lp -regularity
on R+ . This will enable us to characterise maximal regularity in terms of the
boundedness of a convolution operator related to mild solutions.

Maximal regularity implies sectoriality

The main result of this paragraph reads as follows.

Theorem 17.2.15 (Dore). Let A be a linear operator on a Banach space

X, and let 1 6 p 6 ∞ be fixed. Then:
(1) if A has maximal Lp -regularity on a bounded interval (0, T ), then −A
generates an analytic semigroup on X, and λ + A is sectorial of angle
< π/2 for all λ ∈ R large enough. Moreover, for <λ large enough,
reg
kA(λ + A)−1 k 6 2Mp,A (R+ ). (17.13)
 17.2 Maximal Lp -regularity 587

(2) if A has maximal Lp -regularity on R+ , then −A generates a bounded an-

alytic semigroup on X, and A is sectorial of angle < π/2. Moreover, for
all <λ > 0,
reg
kA(λ + A)−1 k 6 Mp,A (R+ ). (17.14)

In Theorem 17.3.1 we will prove an R-bounded version of (17.14) for exponents

p ∈ [1, ∞).
Proof of Theorem 17.2.15(1). Proposition 17.2.5 shows that A is closed. Let
I = (0, T ) and denote by
MI : f 7→ uf
the operator that assigns to a function f ∈ Lp (I; X) the corresponding Lp -
solution uf ∈ Lp (I; X) to the problem (ACP0 ) with zero initial condition.
Step 1 – Our first aim is to prove the injectivity of λ + A for <λ > 0 large
enough. Fix x ∈ D(A) and <λ > 0 and set

u(t) := eλt x − x,
f (t) := u0 (t) + Au(t) = (λ + A)eλt x − Ax.

Since f ∈ Lp (I; X) and u = uf , by maximal Lp -regularity we have

kAukLp (I;X) 6 Ckf kLp (I;X) ,

reg
where C = Mp,A (I). Substituting the definitions of f and u, this implies

keλ(·) − 1kLp (I) kAxk 6 C keλ(·) kLp (I) k(λ + A)xk + T 1/p kAxk .
 
(17.15)

By scaling and using keλ(·) − 1kLp (I) > keλ(·) kLp (I) − k1kLp (I) , we find that

(keλT (·) kLp (0,1) − 1)kAxk 6 C keλT (·) kLp (0,1) k(λ + A)xk + kAxk . (17.16)
 

Let r0 > 0 be such that ker0 (·) kLp (0,1) − 1 > 2C. For <λ > r0 /T , we obtain

kAxk 6 keλT (·) kLp (0,1) k(λ + A)xk.

Now if (λ + A)x = 0, it follows that Ax = 0, λx = 0, and thus x = 0, which

proves the injectivity of λ + A. This proves the asserted injectivity of λ + A
for <λ > r0 /T . To arrive at this conclusion, the detour via (17.16) is not
necessary, as we could have directly worked with (17.15), but (17.16) will be
useful later.
Step 2 – Next we construct a right-inverse of λ + A for <λ > 0 sufficiently
large. Fix λ ∈ C with <λ > 0 and define fλ ∈ Lp (I) by fλ (t) := e−λt . For all
x ∈ X we have
1
kfλ ⊗ xkLp (I;X) 6 kxk (17.17)
(<λ)1/p
588 17 Maximal regularity

Put
Z
Rλ x := 2<λ e−λt MI (fλ ⊗ x)(t) dt. (17.18)
I

By maximal Lp -regularity, the function MI (fλ ⊗ x) belongs to Lp (I; D(A)).

Interpreting the integral in (17.18) as a Bochner integral in D(A) and viewing
A as a bounded operator from D(A) to X, we find that Rλ x ∈ D(A) and
Z
ARλ x = 2<λ e−λt AMI (fλ ⊗ x)(t) dt. (17.19)
I

By Hölder’s inequality, (17.9), and (17.17),

kARλ xk 6 2|<λ|ke−λ(·) kLp0 (I) kAMI (fλ ⊗ x)kLp (I;X)

(17.20)
6 2C|<λ|ke−λ(·) kLp0 (I) kfλ ⊗ xkLp (I;X) 6 2Ckxk.
reg
with C = Mp,A (I). Once the remainder of the proof has been finished, (17.13)
follows from this estimate.
By (17.3) and an integration by parts (which is justified by the fact that
Lp -solutions are in W 1,p (I; X)),
Z
ARλ x = −2<λ e−λt [(MI (fλ ⊗ x))0 (t) − (fλ ⊗ x)(t)] dt
I
Z T
(17.21)
= −2<λ e−λT (MI (fλ ⊗ x))(T ) − λRλ x + 2<λ e−2<λt x dt
0
= −2<λ[e−λT (MI (fλ ⊗ x))(T )] − λRλ x + (1 − e−2<λT )x.

If we could let T → ∞ then this would give ARλ x = −λRλ x + x. This is not
possible, however, since I = (0, T ) is fixed. Instead, we will take <λ so large
that the remainder term has norm 6 1/2. To this end let

Qλ := 2<λ[e−λT (MI (fλ ⊗ x))(T )] + e−2<λT x. (17.22)

Suppose that <λ > 1/T . By (17.10), (17.17), and the fact that for any 0 <
α 6 1 the function t 7→ tα e−t/2 is bounded above on R+ by (2α/e)α 6 1,
0
kQλ xk 6 2<λ e−<λT T 1/p (C + 1)kfλ ⊗ xkp + e−2<λT kxk
0
6 2(<λT )1/p e−<λ T (C + 1)kxk + e−2 kxk (17.23)
− 21 <λ T
6 2e (C + 1)kxk + e−2 kxk.

Therefore, if <λ > r1 /T with r1 := 2 log(4(C + 1)), then

1
kQλ k 6 (17.24)
2
 17.2 Maximal Lp -regularity 589

and I − Qλ is invertible with k(I − Qλ )−1 k 6 2. Since by (17.21) and (17.22)

imply (λ + A)Rλ x = (I − Qλ )x, it follows that Rλ (I − Qλ )−1 is a right inverse
to λ + A.
Step 3 – Next we will show that the above right-inverse is actually an
inverse for <λ large enough and obtain the required resolvent estimate. Let
rA = max{r0 , r1 }. By Steps 1 and 2, λ + A is injective and surjective, hence
invertible, for <λ > rA /T . Therefore, for each <λ > rA /T , λ ∈ %(−A) and
(λ + A)−1 = Rλ (I − Qλ )−1 coincides with the right-inverse constructed in
Step 2. By (17.20), (17.24), and the Neumann series expansion,

kλ(λ + A)−1 k 6 1 + kA(λ + A)−1 k 6 1 + kARλ (I − Qλ )−1 k

6 1 + 2Ck(I − Qλ )−1 k 6 1 + 4C.

In combination with Lemma G.1.4, this proves that large enough translates
of −A are sectorial of angle < 21 π. The results of Appendix K imply that
these translates, and hence −A itself, generate analytic semigroups on X in
the sense of Definition K.1.2. 
Proof of Theorem 17.2.15(2) with non-sharp constant. The above proof for
I = (0, T ) can be repeated for I = R+ , with MI replaced by MR+ , up to
equation (17.22) which now reads

Qλ := 2<λ[e−λT (MR+ (fλ ⊗ x))(T )] + e−2<λT x.

We can still apply (17.10) to this equation, and in combination with (17.17)
we arrive at the bound given in (17.23). This time we can pass to the limit
T → ∞ and arrive at the identity ARλ x = −λRλ x + x. Therefore, Rλ is a
right-inverse of λ + A. To see that it is a left-inverse it suffices to check that
ARλ = Rλ A. The latter follows from (17.19) and Lemma 17.2.12.
Sectoriality follows from (17.20) with a slightly worse bound than stated
in (17.14). The bound will be further improved below. 
Having analytic semigroups available, we can now deduce maximal regularity
on sub-intervals by using the equivalence result for strong and mild solutions
described in Proposition 17.1.3. Recall from Definition 17.1.2 that if S =
(S(t))t>0 is a locally bounded strongly measurable semigroup S on X, then
for x ∈ D(A) the continuous function u : [0, ∞) → X defined by
Z t
u(t) := S(t)x + S(t − s)f (s) ds, t ∈ I,
0

is called the mild solution of the problem (ACPx ). Continuity of t 7→ S(t)x

for elements for x ∈ D(A) follows from Proposition K.1.5, but in the present
situation there is no need to refer to this result since we consider the initial
value x = 0.
590 17 Maximal regularity

Lemma 17.2.16 (Maximal Lp -regularity on subintervals). Let 1 6

p 6 ∞ and let A be a linear operator on a Banach space X with maximal
Lp -regularity on I. Then A has maximal Lp -regularity on every subinterval
(0, T ) ⊆ I and
reg reg
Mp,A (0, T ) 6 Mp,A (I). (17.25)
Proof. The first issue to deal with is the uniqueness of strong solutions on
(0, T ). Let u ∈ Lp (0, T ; D(A)) be an Lp -solution to (ACP0 ) on (0, T ) with
inhomogeneity f = 0. Since A has maximal Lp -regularity on I, it generates
an analytic semigroup by Dore’s Theorem 17.2.15. This allows us to apply
Proposition 17.1.3, which guarantees that u equals the mild solution on (0, T ),
which is obviously identically zero. It follows that u = 0.
Now let f ∈ Lp (0, T ; X) be arbitrary, and let u be the Lp -solution to
(ACP0 ) on I with inhomogeneity f , where f is the zero extension of f to I.
Then u|(0,T ) is a strong solution of (ACP0 ) on (0, T ), and
kAu|(0,T ) kLp (0,T ;X) 6 kAukLp (I;X)
reg reg
6 Mp,A (I)kf kLp (I;X) = Mp,A (I)kf kLp (0,T ;X) .
By uniqueness, u|(0,T ) is the only possible strong solution, we conclude that
A has maximal Lp -regularity on (0, T ) and that (17.25) holds. 
Remark 17.2.17. For 1 < p < ∞ the following alternative argument can be
used to obtain uniqueness on subintervals. Suppose that A has maximal Lp -
regularity on I, and (0, T ) ⊆ I be a bounded subinterval. Let u be an Lp -
solution to the problem u0 + Au = 0 on (0, T ) with initial condition u(0) = 0.
We claim that u = 0. By Proposition 17.2.2 we have u ∈ Lp (0, T ; D(A)) ∩
W 1,p (0, T ; X) and thus, by repeating the argument of the proof of Proposition
17.2.13 for ũ := u(T − ·) ∈ Lp (0, T ; D(A)) ∩ W 1,p (0, T ; X), we obtain u(T ) =
ũ(0) ∈ (X, D(A))1− p1 ,p . By Proposition 17.2.14, there exists a unique Lp -
solution v of v 0 + Av = 0 on I and v(0) = u(T ). Let u(t) := u for t ∈ [0, T ]
and u(t) := v(t − T ) for t ∈ (T, ∞) ∩ I. Then u is an Lp -solution to the
problem u0 + Au = 0 on I with initial condition u(0) = 0. Therefore, u = 0
by maximal Lp -regularity, and thus u = 0 on (0, T ).
We are now in a position complete the proof of Theorem 17.2.15 by showing
the estimate (17.14) for the sectoriality constant.
Proof of the sectoriality bound (17.14). We have already proved that A is sec-
torial of angle < π/2. It remains to prove the bound (17.14). By Lemma
17.2.16 we know that maximal Lp -regularity holds on (0, T ) with estimate
reg
(17.25). Replacing the constant C in Step 1 by the larger constant Mp,A (R+ ),
λT ·
and passing to the limit T → ∞ using that ke kLp (0,1) → ∞ in (17.16), it
follows that
kAxk 6 Ck(λ + A)xk.
Since λ + A is invertible, this gives (17.14). 
 17.2 Maximal Lp -regularity 591

The next result gives a sufficient condition for maximal Lp -regularity on R+

by establishing uniform estimates on bounded intervals.

Proposition 17.2.18. Let 1 6 p 6 ∞ and let A be a linear operator. The

following are equivalent:
(1) For each T ∈ (0, ∞), the operator A has maximal Lp -regularity on (0, T )
reg
and supT >0 Mp,A (0, T ) < ∞;
(2) The operator A has maximal Lp -regularity on R+ .
reg reg
In this case we have Mp,A (R+ ) = supT >0 Mp,A (0, T ).

Proof. (2) ⇒ (1) is clear from Lemma 17.2.16, and the estimate “>” follows
from (17.25).
(1) ⇒ (2): Let f ∈ Lp (R+ ; X). For every 0 < T < ∞ there exists a unique
p
L -solution uT to (ACP0 ) on (0, T ). By uniqueness, uT |[0,S] = uS for every
S 6 T . Therefore, we can construct a function u : [0, ∞) → X such that
u|[0,T ] = uT for every 0 < T < ∞. Then u is a strong solution to (ACP0 ) on
R+ . Moreover, uniqueness follows by taking restrictions to [0, T ]. Since

kAukLp (0,T ;X) 6 Ckf kLp (0,T ;X) 6 Ckf kLp (R+ ;X) ,
reg
where C = supT >0 Mp,A (0, T ), the desired result follows from the monotone
convergence theorem by passing to the limit T → ∞. 
Now that we know that maximal Lp -regularity of A implies that −A generates
an analytic semigroup it is of interest to relate Lp -solutions to mild solutions.

Theorem 17.2.19. Let −A generate an analytic semigroup on a Banach

space X and let 1 6 p 6 ∞. Let F be any dense subspace of Lp (I; X). Then
A has maximal Lp -regularity on I if and only if the mapping f → V f , defined
for functions f ∈ F by
Z t
V f (t) := A S(t − s)f (s) ds, t ∈ I,
0

p
is well defined, maps F into L (I; X), and there is a constant C > 0 such
that
kV f kLp (I;X) 6 Ckf kLp (I;X) , f ∈ F.
In this situation, V uniquely extends to a bounded operator on Lp (I; X) such
that
V f = Auf ,
where uf is the mild and Lp -solution to (ACP0 ) associated with f , and the
reg
least admissible constant C in the above inequality coincides with Mp,A (I).
592 17 Maximal regularity

Remark 17.2.20 (On the choice of the space F ). In the above formulation it
is implicit that one should have
Z t
S(t − s)f (s) ds ∈ D(A)
0

for almost all t ∈ I. By Proposition 17.1.4 this holds for instance for F =
C α (I; X) with α > 0 (which is dense if p ∈ [1, ∞)). In the important case
when D(A) = X, it can be useful to take F = Lp (I; D(A)) (which is dense
if p ∈ [1, ∞)); in this case the operator A can even be pulled through the
integral. This choice doesn’t work for p = ∞ even when D(A) = X, because
the space L∞ (I; D(A)) in general fails to be dense in L∞ (I; X). As we will
see in Section 17.2.f, this problem can be circumvented by considering the
equivalent notion of maximal C-regularity.
Remark 17.2.21 (Connection with singular integrals). For analytic semigroups
S with generator −A, the bound

kAS(t)k = O(1/t) as t ↓ 0

(apply (K.4) to λ + A for λ large enough) shows that in those cases where
it is possible to pull A through the integral in the definition of V , a singular
convolution integral is obtained. This aspect of the theory will be further
studied at later stage in this chapter (see, for example, Theorems 17.2.31,
17.2.39, and Section 17.3).
Proof of Theorem 17.2.19. ‘Only if’: Every f ∈ Lp (I; X) gives rise to a unique
Lp -solution uf which, by Proposition 17.1.3, equals the mild solution. There-
fore, the mapping f 7→ Auf = V f extends to a bounded operator on Lp (I; X)
by maximal Lp -regularity.
‘If’: Let uf be the mild solution associated with f ∈ F . The assumptions
on F imply that uf takes values in D(A) and that Au ∈ Lp (I; X), so uf is an
Lp -solution by Proposition 17.1.3. Since V f = Auf for all f ∈ F , maximal Lp -
regularity follows from the boundedness of V , closedness of A, and Proposition
17.2.10. 
We continue with a useful special property of maximal L1 -regularity which will
be needed several times below: it leads to a characterisation of maximal L1 -
regularity in Theorem 17.3.11, and it allows us to give a proof of Proposition
17.2.14 for p = 1, and it will be used in Proposition 17.2.32.
Lemma 17.2.22. Let A have maximal L1 -regularity on I, and let S be the
analytic semigroup generated by −A. Then
Z
reg
kAS(t)xk dt 6 M1,A (I)CI,S kxk, x ∈ X, (17.26)
I

where CI,S = supt∈I kS(t)k.

 17.2 Maximal Lp -regularity 593

Proof. Recall that −A generates an analytic semigroup S by Theorem 17.2.15.

Let CI,S := supt∈I kS(t)k. This number is finite: if I = R+ , this follows from
the fact that A is sectorial in that case and the semigroup generated by A is
uniformly bounded; if I = (0, T ) is a bounded interval, this follows from the
fact that λ + A is sectorial for <λ large enough, and then t 7→ e−λt S(t) is
uniformly bounded.
Let ε ∈ (0, 1) be arbitrary and set Iε := I ∩(ε, 1/ε) and φ(t) := min{t/ε, 1}.
Then φ is weakly differentiable and kφ0 kL1 (I) = 1. For x ∈ X set

u(t) := φ(t)S(t)x and f (t) := u0 (t) + Au(t) = φ0 (t)S(t)x.

Since A has maximal L1 -regularity we obtain that

Z Z
reg reg
kAS(t)xk dt = kAu(t)k dt 6 M1,A (I)kf kL1 (I;X) 6 M1,A (I)CI,S kxk.
Iε Iε

Letting ε ↓ 0 we obtain (17.26). 

Remark 17.2.23. A little more can be said. For t > s > 0 with t ∈ I we have
Z t Z t Z t
kS(s)x − S(t)xk = A S(r)x dr = AS(r)x dr 6 kAS(r)xk dr
s s s

By this estimate and the Bochner integrability of AS(·), the limit S(0+)x :=
limt↓0 S(t)x exists and it satisfies
Z t
kS(0+)x − S(t)xk 6 kAS(r)xk dr.
0

In the special case that X is a dual space and A is the adjoint of a C0 -

semigroup on the predual of X, the weak∗ -continuity of S allows us to identify
the limit as S(0+)x = x. Thus we obtain the following result: Let G generate
a C0 -semigroup on Banach space Y . If G∗ has maximal L1 -regularity, then
the adjoint semigroup is strongly continuous on Y ∗ .

We are now able to prove Proposition 17.2.14 for p = 1:

Proof of Proposition 17.2.14 for p = 1. Since x ∈ (X, D(A))1− p1 ,p by assump-
tion, Corollary C.3.15 implies that x ∈ D(A), and therefore t 7→ S(t)x is
strongly continuous by Proposition K.1.5. The function u = Sx + S ∗ f is an
L1 -solution of (ACPx ), and Theorem 17.2.19 and Lemma 17.2.22 imply that

kAukL1 (I;X) 6 kASxkL1 (I;X) + kAS ∗ f kL1 (I;X)

reg reg
6 M1,A (I)CI,S kxk + M1,A (I)kf kL1 (I;X) .


594 17 Maximal regularity

In order to prepare for some results in the next sections, we conclude the
present section by introducing a constant related to the maximal regularity
constant with an additional parameter λ; in applications, this additional flex-
ibility can sometimes be exploited.
From Theorem 17.2.19 we recall that if A is a linear operator in X such
that −A generates an analytic semigroup S on a Banach space X, then A has
maximal Lp -regularity on some bounded interval (0, T ) if and only if
Z t
V f (t) := A S(t − s)f (s) ds, t ∈ [0, T ],
0

is well defined for all functions f in some dense subspace of Lp (I; X) and
extends to a bounded operator V on Lp (I; X). When I = (0, T ) is a bounded
interval, in these circumstances for all λ ∈ C the operator Vλ defined by
Z t
Vλ f (t) := A Sλ (t − s)f (s) ds, t ∈ [0, T ],
0

has the same properties, where Sλ (t) := e−λt S(t) is the rescaled semigroup.
Indeed, in this case we have

kVλ f kLp (I;X) 6 max{1, e−<λT }kV kL (Lp (I;X)) ks 7→ e−λs f (s)kLp (I;X)

where I = (0, T ). We will come back to this in Proposition 17.2.27.

Suppose now that A is a linear operator in X such that −A generates an
analytic semigroup S on a Banach space X; we make R t no maximal regularity
assumptions. If f ∈ Lp (I; X) is a function such that 0 Sλ (t−s)f (s) ds ∈ D(A)
for almost all t ∈ I, we may define the function Vλ f as above, that is,
Z t
Vλ f (t) := A Sλ (t − s)f (s) ds, t ∈ I. (17.27)
0

Note that Vλ f = Auλ almost everywhere on I, where uλ is the mild solution

to the problem
(
u0 (t) + (λ + A)u(t) = f (t), t ∈ I,
(17.28)
u(0) = 0.

If the mild solution uλ exists and takes values in D(A) for all for all f ∈ F ,
where F is some dense subspace of Lp (I; X), then the mapping f 7→ Vλ f
defined by (17.27) uniquely extends to a bounded linear mapping on Lp (I; X)
and we may define
reg
Mp,A,λ (I) := kVλ kL (Lp (I;X)) .
reg reg
Thus if A has maximal Lp -regularity on I, then Mp,A,0 (I) = Mp,A (I). For
other values of λ, we will derive various bounds in Theorem 17.2.24 and Propo-
reg reg
sition 17.2.27. We should emphasise that, in general, Mp,A,λ (I) 6= Mp,A+λ (I);
 17.2 Maximal Lp -regularity 595
Rt
the latter would be the norm the operator Ṽλ f (t) := (A+λ) 0 Sλ (t−s)f (s) ds,
while in (17.27), only the semigroup is rescaled, but not the operator A.
If α, β ∈ R and Vα extends to a bounded operator on Lp (I; X), then so
does Vα+iβ and

kVα+iβ f kLp (I;X) = kVα eiβ· f kLp (I;X) , f ∈ Lp (I; X). (17.29)
reg reg
Clearly, this implies Mp,A,α+iβ (I) = Mp,A,α (I) for any α, β ∈ R.

17.2.d Uniformly exponentially stable semigroups

In the previous subsection we have seen that maximal Lp -regularity on R+

implies maximal Lp -regularity on every bounded interval (0, T ). To state a
result in the converse direction we need some preparations.
Recall from Appendix K that for a locally bounded semigroup (S(t))t>0
there exist M > 1 and µ ∈ R such that kS(t)k 6 M eµt for all t > 0. The
semigroup is said to be uniformly exponentially stable if µ can be taken strictly
negative.
If S is an analytic semigroup, then there exist constants M > 1 and µ ∈ R
such that

kS(t)k 6 M eµt and ktAS(t)k 6 M eµt , t > 0, (17.30)

and if such a semigroup S is uniformly exponentially stable we can take µ < 0

in (17.30) (this follows from the proof of the second of these inequalities).
The main result of this section states, among other things, that if A has
maximal Lp -regularity on a bounded interval and the analytic semigroup gen-
erated by −A is uniformly exponentially stable, then A has maximal Lp -
regularity on R+ .
reg
Recall that the constants Mp,A,λ (R+ ) have been introduced at the end of
the previous section.

Theorem 17.2.24 (Dore–Kato). Let 1 6 p 6 ∞ be fixed, and suppose that

the linear operator A has maximal Lp -regularity on some bounded interval
I = (0, T ). If the analytic semigroup S generated by −A satisfies (17.30) with
constants M > 1 and µ ∈ R, then for all <λ > µ we have

reg reg 2M
Mp,A,λ (R+ ) 6 2 max{e<λT /2 , e−<λT }Mp,A (I) + . (17.31)
T (<λ − µ)

In particular, if S is uniformly exponentially stable with ω := −µ > 0, then

A has maximal Lp -regularity on R+ and

reg reg 2M
Mp,A (R+ ) 6 2Mp,A (I) + . (17.32)
Tω
596 17 Maximal regularity

Two remarks are in order. First of all, in (17.31) no claim is made with
regard to maximal Lp -regularity on R+ ; this is done only in (17.32). Secondly,
in situations where A has maximal Lp -regularity on every bounded interval
(0, T ) one can optimise the choice of T in (17.31) by letting it depend on the
parameter λ (see, for instance, the proof of Proposition 17.2.27).
Proof. Let Sλ (t) = e−λt S(t), where λ = α + iβ ∈ C with α > µ. For f ∈
Lp (R+ ; X) the expression for the mild solution

uf,λ := Sλ ∗ f

defines a continuous function uf,λ : [0, ∞) → X. We will show that uf,λ takes
values in D(A) almost everywhere and that

kAuf,λ kLp (R+ ;X) 6 Kkf kLp (R+ ;X) (17.33)

reg
where K = 2e<λT /2 Mp,A (I)+2M/T (α−µ). Once we have shown this, Propo-
sition 17.1.3 implies that uf,λ is an Lp -solution to (ACP0 ) with A replaced by
λ + A, and (17.31) follows from (17.33). The second assertion of the theorem
follows by taking λ = 0.
In view of (17.29), in order to prove (17.33) it suffices to consider β = 0,
and thus λ ∈ R. For j = 0, 1, 2, . . . put tj := 21 jT . For t ∈ (tj , tj+1 ) with j > 1
we write
Z t Z tj−1
uf,λ (t) = Sλ (t − s)f (s) ds + Sλ (t − s)f (s) ds
tj−1 0
Z t−tj−1 Z tj−1
= Sλ (t − tj−1 − s)f (s + tj−1 ) ds + Sλ (t − s)f (s) ds
0 0
(j) (j)
=: uf,λ (t) + vf,λ (t).
(j)
By maximal Lp -regularity and the discussion below (17.29), uf,λ ∈ D(A)
almost everywhere on (tj−1 , tj+1 ), and
(j)
kAuf,λ kLp (tj ,tj+1 ;X)
Z t−tj−1
= t 7→ A Sλ (t − tj−1 − s)f (s + tj−1 ) ds
0 Lp (tj ,tj+1 ;X)
Z τ
= τ 7→ e−λτ A S(τ − s)eλs f (s + tj−1 ) ds
0 Lp (T /2,T ;X)
reg
6 max{e−<λT /2 , e−<λT }Mp,A (I) max{1, e<λT }kf (· + tj−1 )kLp (I;X)
reg
= max{e<λT /2 , e−<λT }Mp,A (I)kf kLp (tj−1 ,tj+1 ;X)
6 bj + bj+1 ,
reg
where bj := max{eλT /2 , e−λT }Mp,A (I)kf kLp (tj−1 ,tj ;X) . A similar computation
with t−1 := 0 shows that
 17.2 Maximal Lp -regularity 597

kAuf,λ kLp (0,T /2;X) 6 b1 .

(j)
To estimate the norm of Avf,λ we recall from (17.30) that ktASλ (t)k 6
M e−(λ−µ)t . Since t − s > T /2 for t ∈ (tj , tj+1 ) and s ∈ (0, tj−1 ), we can
estimate
Z tj−1
(j)
kAvf,λ (t)k 6 kASλ (t − s)f (s)k ds
0
Z t
−1
6 2M T e−(λ−µ)(t−s) kf (s)k ds = 2M T −1 (φµ ∗ kf (·)k)(t),
0

where φµ (s) = e−(λ−µ)s 1R+ (s). Therefore,

(j)
kAvf,λ kLp (tj ,tj+1 ;X) 6 2M T −1 φµ ∗ kf (·)k =: cj .
Lp (tj ,tj+1 )

This proves that for all j > 1,

kAuf,λ kLp (tj ,tj+1 ;X) 6 bj + bj+1 + cj . (17.34)

As observed above, the estimate (17.34) also holds for j = 0 if we set b0 = 0.

Taking `p -norms over j > 0 in (17.34), we obtain

kAuf,λ kLp (R+ ;X)

= (kAuf,λ kLp (tj ,tj+1 ;X) )j>0 `p
6 2k(bj )j>0 k`p + k(cj )j>0 k`p
reg
= 2 max{e<λT /2 , e−<λT }Mp,A (I)kf kLp (R+ ;X) + 2M T −1 φµ ∗ kf (·)k Lp (R+ )
 
reg
6 2 max{e<λT /2 , e−<λT }Mp,A (I) + 2M (T (λ − µ))−1 kf kLp (R+ ;X) ,

where in the last step we applied Young’s inequality. This proves (17.33). 
We conclude this section by showing the necessity of the invertibility assump-
tion in Proposition 17.2.8.

Corollary 17.2.25. Suppose that the linear operator A has maximal Lp -

regularity on R+ for a given 1 6 p 6 ∞. If there exists a constant C > 0
such that for all f ∈ Lp (R+ ; X) the Lp -solution u of (ACP0 ) satisfies

kukLp (R+ ;D(A)) 6 Ckf kLp (R+ ;X) ,

then −A generates an analytic semigroup that is uniformly exponentially stable

and, in particular, 0 ∈ %(A).

Proof. From Theorem 17.2.15, we see that −A generates an analytic semi-

group. Finally, the uniformly exponential stability follows from Proposition
K.2.2. 
598 17 Maximal regularity

17.2.e Permanence properties

We have already encountered several permanence properties for maximal Lp -
regularity: if A has maximal Lp -regularity on R+ , then it has maximal Lp -
regularity on every bounded interval (0, T ), and the converse holds if the
semigroup generated by −A is uniformly exponentially stable. Further per-
manence properties are summarised in the following qualitative theorem. The
proofs of the corresponding quantitative versions will be spread out over the
remainder of this section.
As before, I denotes a bounded interval (0, T ) or R+ = (0, ∞).
Theorem 17.2.26. Let A be a linear operator on a Banach space X and let
1 6 p 6 ∞. If A has maximal Lp -regularity on I, the following assertions
hold:
(1) (Translation) λ+A has maximal Lp -regularity on I in each of the following
two situations:
(i) I = R+ and <λ > 0;
(ii) I = (0, T ) is a bounded interval and λ ∈ C.
(2) (Change of interval) A has maximal Lp -regularity on every bounded inter-
val (0, T 0 ).
(3) (Scalar multiples) λA has maximal Lp -regularity on I for all λ > 0.
(4) (Extrapolation of exponent) A has maximal Lq -regularity on I for all
q ∈ (1, ∞).
0
(5) (Duality) If D(A) = X, then A∗ has maximal Lp -regularity on I.
reg
We start preparations for proving (1) by estimating the constants Mp,A,λ (I)
reg reg
introduced after (17.28). Recalling that Mp,A,λ (I) 6= Mp,A+λ (I), note that
this will not yet constitute a proof of (1) itself; we only turn to this task after
establishing the next proposition.
Proposition 17.2.27 (Translation I). Let 1 6 p 6 ∞ and let A be a linear
operator on a Banach space X with maximal Lp -regularity on I, and consider
the problem
(
u0 (t) + (λ + A)u(t) = f (t), t ∈ I,
u(0) = 0.
(1) For all <λ > 0 and f ∈ Lp (I; X) the above problem admits a unique
Lp -solution u, and it satisfies the estimate
reg
kAukLp (I;X) 6 Mp,A (I)kf kLp (I;X) .
reg reg
In particular, Mp,A,λ (I) 6 Mp,A (I).
(2) If I = (0, T ) is a bounded interval, then for all λ ∈ C and f ∈ Lp (I; X) the
above problem admits a unique Lp -solution u, and it satisfies the estimate
reg
kAukLp (I;X) 6 Mp,A (I)e|<λ|T kf kLp (I;X) .
reg reg
In particular, Mp,A,λ (I) 6 e|<λ|T Mp,A (I).
 17.2 Maximal Lp -regularity 599

If I = (0, T ) is bounded and <λ > 0, we can apply either (1) or (2), but the
first estimate, which is independent of λ, will be sharper. Both cases can be
combined into
reg
kAukLp (I;X) 6 Mp,A (I)e(<λ)− T kf kLp (I;X) ,

where x− := max(−x, 0) is the negative part of number.

Proof. We first prove that (2) holds for arbitrary λ ∈ C. Accordingly let
I = (0, T ) and fix λ ∈ C and f ∈ Lp (0, T ; X), and consider the function
fλ ∈ Lp (0, T ; X) defined by fλ (t) := e−λt f (t). If u denotes the Lp -solution of
the problem u0 + Au = f , then v(t) := e−λt u(t) defines an Lp -solution of the
problem v 0 +(λ+A)v = fλ . Put L± = max{e±<λT , 1}, so that L+ L− = e|<λ|T .
Using (17.6) and (17.9) we have

kAvkLp (0,T ;X) = kAe−λ(·) ukLp (0,T ;X)

6 L− kAukLp (0,T ;X)
reg
6 Mp,A (I)L− kf kLp (0,T ;X)
reg
6 Mp,A (I)L+ L− kfλ kLp (0,T ;X) .
reg reg
This gives the result with Mp,A,λ (I) 6 e|<λ|T Mp,A (I) for any λ ∈ C.
(1): Step 1 – First we prove a non-optimal estimate in the case I = R+ .
By Theorem 17.2.15, −A generates a bounded analytic semigroup (S(t))t>0 ,
and by Proposition K.1.11 and (K.4) we can find a constant M > 1 such that
kS(t)k 6 M and ktAS(t)k 6 M for all t > 0.
Fix an arbitrary T ∈ (0, ∞). Recall that A has maximal Lp -regularity on
reg reg
IT = (0, T ) with Mp,A (IT ) 6 Mp,A (R+ ) (see (17.25)). By Theorem 17.2.24,
for all <λ > 0 one has
reg reg
Mp,A,λ (R+ ) 6 2e<λT /2 Mp,A (R+ ) + 2M (T <λ)−1 .

2
If <λ > 0, the choice T = <λ gives the bound
reg reg
Mp,A,λ (R+ ) 6 2eMp,A (R+ ) + M =: K, <λ > 0. (17.35)

Now the idea is to improve this bound by a maximum principle applied to

the operator Vλ defined after (17.28). We first consider the case of bounded
intervals, and take limits afterwards.
Step 2 – First consider the case where I = (0, T ) is a bounded interval. By
Theorem 17.2.15, −A generates an analytic semigroup (S(t))t>0 . Therefore,
as before we can find µ, M > 0 such that kS(t)k 6 M eµt for all t > 0. By
(K.2), this implies
Z ∞
k(µ + 1 + A)−1 k 6 e−(µ+1) kS(t)k dt 6 M,
0
600 17 Maximal regularity

and hence

kA(µ + 1 + A)−1 k 6 1 + k(µ + 1)(µ + 1 + A)−1 k 6 1 + M (µ + 1) =: Cµ ,

and hence kAxk 6 Cµ k(µ + 1 + A)xk for x ∈ D(A). Therefore, using the
λ-analogue of (17.25) and (17.35), for all λ ∈ C with <λ > µ + 1 we obtain
reg reg
Mp,A,λ (I) 6 Mp,A,λ (R+ )
reg
6 Cµ Mp,A+µ+1,λ−µ−1 (R+ )
reg
6 Cµ (2eMp,A+µ+1 (R+ ) + M ) =: Kµ .

If 0 6 <λ < µ + 1, then by the proof of (2) we find that

reg reg reg
Mp,A,λ (I) 6 e<λT Mp,A (I) 6 e(µ+1)T Mp,A (I).

Combining both cases we see that Vλ is uniformly bounded on Lp (I; X) on

{<λ > 0}. One can check that Vλ is holomorphic on C+ .
We claim that λ 7→ Vλ is continuous on C+ . Indeed, let λ1 , λ2 ∈ C+ and
f ∈ Lp (I; X). Then Vλj f = Auj , where uj is the Lp -solution to u0j + (λj +
A)uj = f and uj (0) = 0. By Proposition 17.1.3 we can write uj = Sλj ∗ f ,
and thus by Young’s inequality

kuj kLp (I;X) 6 kSλj kL1 (0,T ;L (X)) kf kLp (I;X) , j ∈ {1, 2}.

Subtracting the equations for u1 and u2 one sees that u = u1 − u2 satisfies

u0 + (λ1 + A)u = (λ2 − λ1 )u2 . Therefore, by maximal Lp -regularity,

k(Vλ1 − Vλ2 )f kLp (I;X) = kAukLp (I;X)

reg
6 Mp,A,λ 1
(I)|λ2 − λ1 |ku2 kLp (I;X) .
reg
6 Mp,A,λ 1
(I)|λ2 − λ1 |kSλ2 kL1 (0,T ;L (X)) kf kLp (I;X) .

This gives the required continuity of λ 7→ Vλ .

Since λ 7→ Vλ is holomorphic and uniformly bounded on the open half
plane, and continuous on the closed half plane, the Phragmén–Lindelöf prin-
ciple implies that for all <λ > 0,
reg reg
Mp,A,λ (I) = kVλ k 6 sup kViβ k = Mp,A (I),
β∈R

where the last identity follows from (17.29).

Step 3 – In case I = R+ , we can just let T → ∞ in the previous esti-
mate, and apply the analogue of Proposition 17.2.18 for the problem with the
additional λ. 
Proof of Theorem 17.2.26 (1)–(3). We begin with the proof of (1). First let
I = (0, T ) be a bounded interval and fix λ ∈ C. By Propositions 17.1.3 and
 17.2 Maximal Lp -regularity 601

17.2.27 the mild and Lp -solution to u0 + (λ + A)u = f with u(0) = 0 is given

by u = Lλ f := Sλ ∗ f , where Sλ (t) = e−λt S(t). It is clear that Lλ is bounded
on Lp (I; X), and

k(λ + A)ukLp (I;X) 6 kλukLp (I;X) + kAukLp (I;X)

reg

6 kλLλ k + Mp,A,λ (I) kf kLp (I;X) .

Using (the comments right after)Proposition 17.2.27, it follows that

reg reg
Mp,λ+A (I) 6 kλLλ k + e(<λ)− T Mp,A (I).

Next let I = R+ and <λ > 0. The previous considerations apply with
(<λ)− = 0; the boundedness of Lλ now follows from Young’s inequality, which
gives the estimate
M |λ|
kλLλ k 6 kt 7→ λM e−tλ kL1 (R+ ) =
<λ
where kS(t)k 6 M for all t > 0. This gives the estimate

reg M |λ| reg

Mp,λ+A (R+ ) 6 + Mp,A (R+ ). (17.36)
<λ
reg
(2): If 0 < T 0 < |I|, the result follows from (17.25) with Mp,A (0, T 0 ) 6
reg 0
Mp,A (I). It remains to consider the case 0 < |I| < T < ∞. By The-
orem 17.2.15, the operator −A generates an analytic semigroup (S(t))t>0 .
Let M > 1 and ω > 0 be such that kS(t)k 6 M eωt for all t > 0 and fix
a number λ > ω. By case (1) of the theorem, which we already proved,
the translated operator λ + A has maximal Lp -regularity on I. Its gen-
erated semigroup Sλ (t) = e−λt S(t) is uniformly exponentially stable with
kSλ (t)k 6 M e−(λ−ω)t , where λ − ω > 0 Therefore, by the Dore–Kato theorem
17.2.24, A + λ has maximal Lp -regularity on R+ , and then by restriction also
on (0, T 0 ). Another application of the translation property (case (1) of the
theorem) shows that A has maximal Lp -regularity on (0, T 0 ).
(3): Let Iλ = λ1 I and fix an arbitrary f ∈ Lp (Iλ ; X). Note that v is the
unique Lp -solution to v 0 + Av = λ−1 f (λ−1 ·) on I if and only if u := v(λ·)
is the unique Lp -solution to u0 + λAu = f on Iλ , where we use zero initial
conditions for both problems. Moreover,
1
kλAukLp (Iλ ;X) = λ1− p kAvkLp (I;X)
1
6 λ− p Mp,A
reg reg
(I)kf (λ−1 ·)kLp (I;X) = Mp,A (I)kf kLp (Iλ ;X) .
reg
It follows that λA has maximal Lp -regularity on Iλ with Mp,λA (Iλ ) 6
reg
Mp,A (I). If I = R+ we are done; if I is bounded we apply (2). 
reg
Remark 17.2.28. The estimate (17.36) shows that the constant Mp,λ+A (R+ )
1 1
is uniformly bounded in λ ∈ Σφ for any φ ∈ (− 2 π, 2 π).
602 17 Maximal regularity

Next we present a converse to the translation result of Proposition 17.2.27,

which can be used to reduce the question whether A has maximal Lp -regularity
on R+ to the case where A is replaced by λ + A with λ > 0.

Proposition 17.2.29 (Translation II). Let 1 6 p 6 ∞ and let A be a linear

operator on a Banach space X. Suppose that there is a constant C > 0 such
that for all 0 < T < ∞ and 0 < λ < ∞ the operator λ + A has maximal
Lp -regularity on (0, T ), and that for all f ∈ Lp (0, T ; X) the Lp -solution u to
(
u0 + (λ + A)u = f on (0, T ),
(17.37)
u(0) = 0,

satisfies

kAukLp (0,T ;X) 6 Ckf kLp (0,T ;X) . (17.38)

reg
Then A has maximal Lp -regularity on R+ and Mp,A (R+ ) = CA , where CA is
the infimum of all admissible constants C in (17.38).

Proof. Fix f ∈ Lp (R+ ; X). By Theorem 17.2.26(1ii), applied to A + λ in place

of A and −λ in place of λ, the operator A = (A + λ) − λ has maximal Lp -
regularity on bounded intervals (0, T ). Therefore, as in Proposition 17.2.18 it
follows that there exists a function u : R+ → X such that for any T > 0 its
restriction to (0, T ) is an Lp -solution to (17.37) with λ = 0. Fix λ > 0. Since
u0 + (λ + A)u = f + λu on (0, T ), we obtain

kAukLp (0,T,X) 6 CA kf + λukLp (0,T ;X) .

reg
Passing to the limit λ ↓ 0 we obtain Mp,A (0, T ) 6 CA . As T was arbitrary,
Proposition 17.2.18 implies that A has maximal Lp -regularity on R+ with
reg reg
Mp,A (R+ ) 6 CA . The converse estimate CA 6 Mp,A (R+ ) follows from Propo-
sition 17.2.27(1). 

Extrapolation of integrability

We now set out to prove that maximal Lp -regularity of a linear operator A

is independent of the integrability exponent p. This depends on a circle of
ideas of a different flavour and will be deduced from the Calderón–Zygmund
Theorem 11.2.5. The key point is to interpret maximal Lp -regularity as the Lp -
boundedness of the singular integral operator V with kernel t 7→ 1I (t)AS(t),
where S is the analytic semigroup generated by −A (see Theorem 17.2.19). It
is assumed throughout that I = (0, T ) is a bounded interval or I = R+ .
First we present a lemma in which we check that the kernel t 7→ 1I (t)AS(t)
and the operator V satisfy the conditions introduced in Definition 11.2.1. For
later purposes we also check the conditions of Definition 11.3.1. We use the
terminology and notation introduced in these two definitions.
 17.2 Maximal Lp -regularity 603

Lemma 17.2.30. Let A be a linear operator with maximal Lp0 -regularity on

I for some p0 ∈ [1, ∞]. Let S denote the analytic semigroup generated by −A,
and set C := supt∈I ktAS(t)k. Let k : R → L (X) be given by

AS(t), t > 0;
k(t) :=
0, t 6 0,

and define K : I × I → L (X) by

K(t, s) := k(t − s), s, t ∈ I.

Then the following assertions hold:

(1) the operator V defined in Theorem 17.2.19 has kernel K in the sense of
Definition 11.2.1;
(2) the kernel K is C 1 -Calderón–Zygmund kernel in the sense of Definition
1 2
11.3.1 with cK = C and cK = cK = 4C 2 ; a Dini kernel in the sense of the
same definition with

kωK kDini 6 4C(1 + log+ 4C);

and an operator-Hörmander kernel in the sense of Definition 11.2.1 with

kKkHörop (I) = kKkHör∗op (I) 6 8C(1 + log+ 4C).

(3) the kernel K is an operator-Hörmander kernel in the sense of Definition

11.2.1 with kKkHörop (I) = kKkHör∗op (I) 6 4C 2 ;
(4) the kernel K is a Dini kernel in the sense of Definition 11.3.1 with cK = C
and kωkDini = 4C 2 .

Note that we here specifically apply the Definition 11.2.1(4) with I = (0, T )
or I = R+ ; these are a one-dimensional cube and a quadrant, respectively.
By Theorem 17.2.15, −A generates an analytic semigroup (S(t))t>0 , which
is bounded if I = R+ . By (K.4), C := supt∈I ktAS(t)k is a finite quantity.
By Proposition K.3.1, C > 1/e unless A is bounded, hence 4C > 1, and log+
may be replaced by log in the Dini and Hörmander bounds of the lemma.
Proof. (1): By Theorem 17.2.19 the operator V extends to a bounded operator
on Lp0 (I; X). To show that V has kernel K, let f ∈ Lp0 (I; X) be compactly
supported and let F = supp(f ) be its support. Fix t ∈ I such that δ :=
dist(t, F ) > 0. Then
Z Z
kK(t, s)f (s)k ds 6 sup kAS(t − s)k kf (s)k ds < ∞.
I s∈[0,t−δ] F

Moreover, by Hille’s Theorem 1.2.4,

Z Z Z
V f (t) = A S(t − s)f (s) ds = AS(t − s)f (s) ds = K(t, s)f (s) ds.
F F I
604 17 Maximal regularity

(2): It is clear that for s, t ∈ I, kK(s, t)k 6 C/|s − t|, thus we can take
cK = C in Definition 11.3.1. Moreover, for all s, t ∈ I and r ∈ {s, t}, we have

k(s − t)2 ∂r K(s, t)k = 1R+ (s − t)k(s − t)2 A2 S(s − t)k

2
6 1R+ (s − t)4 s−t s−t
2 AS( 2 ) 6 4C 2 .

This verifies the claimed estimates of a C 1 -Calderón–Zygmund kernel. (3)

and (4): These are immediate from Lemma 11.3.4 in dimension d = 1; hence
2d+1 = 4, and σd−1 = 2 (the unit sphere is now {−1, 1} with (d − 1) = 0-
dimensional measure 2). 
As an application of Theorem 11.2.5 we obtain the following extrapolation
result for maximal Lp -regularity. It gives the assertion (4) on extrapolation of
Theorem 17.2.26.

Theorem 17.2.31 (Extrapolation of the exponent). Let A have max-

imal Lp0 -regularity on I for some p0 ∈ [1, ∞]. Then A has maximal Lp -
regularity on I for all p ∈ (1, ∞), with
 
reg
Mp,A (I) 6 c · pp0 Mpreg
0 ,A
(I) + C(1 + log+ C) , (17.39)

where c is an absolute constant, and C := supt∈I tkAS(t)k, where S is the

analytic semigroup generated by A under the assumptions.

Proof. By Lemma 17.2.30 the operator V has kernel K, where K is as in the

lemma, and this kernel satisfies the operator-Hörmander conditions with the
bounds stated in the said lemma. By Theorem 11.2.5, the operator V extends
to a bounded operator on Lp (I; X) with norm bounded by the right-hand side
of (17.39). Therefore, by Theorem 17.2.19, A has maximal Lp -regularity on I
for all p ∈ (1, ∞), with the same bound. 

Duality

The following duality result for maximal Lp -regularity implies assertion (5) of
Theorem 17.2.26. As before, we assume that I = (0, T ) is a bounded interval
or I = R+ .

Proposition 17.2.32 (Duality). Let A be a densely defined closed operator

acting in a Banach space X, and let 1 6 p 6 ∞. The following assertions are
equivalent:
(1) A has maximal Lp -regularity on I;
0
(2) A∗ has maximal Lp -regularity on I.
reg
In this case we have Mp,A (I) = Mpreg
0 ,A∗ (I).
 17.2 Maximal Lp -regularity 605

For the proof of this theorem we need the following lemma, which is a vari-
ant of Corollary 1.3.2 except for the fact that we interchange the roles of X
and X ∗ for the sake of the application we have in mind; the corresponding
version without this interchange is true as well, with the same proof (or as an
application, by passing through the bidual).

Lemma 17.2.33. Let I ⊆ R be a (bounded or unbounded) interval. If the

subspace Y ⊆ X is norming for X ∗ , then

{f ∈ Cb (I; X) : f takes values in Y } (17.40)

is norming for L1 (I; X ∗ ) with respect to the duality

Z
hf, gi = hf (t), g(t)i dt, f ∈ Cb (I; Y ), g ∈ L1 (I; X ∗ ).
I

In (17.40), we deliberately avoid denoting the space by Cb (I; Y ) for the fol-
lowing reason: if Y happens to have a norm k kY of its own, the notation
Cb (I; Y ) refers to functions f : I → Y that are continuous with respect to
k kY , while in (17.40), we think of continuity relative to the norm k kX of the
ambient space X.
Proof. We will show that kgkL1 (I;X ∗ ) = supf |hf, gi|, where the supremum is
taken over all f in (17.40) of norm 6 1. The estimate “>” is clear. To prove
the converse estimate, the density of simple functions (Lemma 1.2.19) and the
regularity of the LebesgueP measure imply that it suffices to consider simple
n
functions of the form g = j=1 1Ij x∗j , where I1 , . . . , In are disjoint bounded
intervals contained in I, and x∗1 , . . . , x∗n are elements of X ∗ . Let ε ∈ (0, 1)
R arbitrary and choose functions ϕj ∈ Cc (I) such that 0 6 ϕj 6 1Ij and
be
ϕ (t) dt > (1 − ε)[IJ | for each j. Choose yj ∈ Y of norm kyj k 6 1 such
Ij j
Pn
that kx∗j k 6 (1 + ε)hyj , x∗j i. Then f := j=1 ϕj yn has norm 6 1 and
n
X Z
hf, gi = hyj , x∗j i ϕj (t) dt
j=1 I
n
1−εX 1−ε
> |Ij |kx∗j k = kgkL1 (I;X ∗ ) .
1 + ε j=1 1+ε


Proof of Proposition 17.2.32. (1)⇒(2) for bounded intervals I = (0, T ): Let
A have maximal Lp -regularity on I = (0, T ). By Theorem 17.2.15 A is sec-
torial of angle < π/2 and −A generates an analytic semigroup (S(t))t>0 . As
a consequence (see Remark K.1.12), −A∗ generates an analytic semigroup as
well, and it is given by (S ∗ (t))t>0 , where S ∗ (t) := (S(t))∗ . To prove that A∗
0
has maximal Lp -regularity on I we will use Theorem 17.2.19.
606 17 Maximal regularity

We begin with the case p ∈ (1, ∞). Fix f ∈ Lp (0, T ; D(A)) (where, to be
sure, we think of D(A) as a Banach space with respect to the graph norm
kxkD(A) := kxkX + kAxkX ) and g ∈ C 1 ([0, T ]; X ∗ ), and write fe(t) := f (T − t)
and ge(s) = g(T − s) for s, t ∈ I. Then the function S ∗ fe takes values in D(A),
and by Proposition 17.1.4 we have S ∗ ∗ g ∈ C([0, T ]; X) ∩ L∞ (0, T ; D(A)). By
Fubini’s theorem and suitable substitutions,
Z TD Z t E
∗ ∗
hf, A S ∗ gi = Af (t), S ∗ (t − s)g(s) ds dt
0 0
Z T D Z s E
= A S(s − t)fe(t) dt, ge(s) ds = hAS ∗ fe, gei,
0 0

where the brackets in the first and last term refer to the duality between
0
Lp (I; X) and Lp (I; X ∗ ), the latter viewed as a closed subspace of the dual of
p
L (I; X) (see Proposition 1.3.1). Therefore,

|hf, A∗ S ∗ ∗ gi| = |hAS ∗ fe, gei| 6 kAS ∗ fekLp (I;X) ke

g kLp0 (I;X ∗ )
reg
6 Mp,A (I)kf kLp (I;X) kgkLp0 (I;X ∗ ) .

Since the embedding Lp (I; D(A)) ,→ Lp (I; X) is dense (recall that D(A) was
assumed to be densely defined in Proposition 17.2.32 that we are proving),
this estimate extends to all f ∈ Lp (I; X) by density. Thus by Proposition
1.3.1
reg
kA∗ S ∗ ∗ gkLp0 (I;X ∗ ) 6 Mp,A (I)kgkLp0 (I;X ∗ ) .
0
Therefore, A∗ has maximal Lp -regularity, and Mpreg reg
0 ,A∗ (I) 6 Mp,A (I) by The-

1 ∗ p0 ∗
orem 17.2.19 and the density of C ([0, T ]; X ) in L ([0, T ]; X ).
In the case p = ∞, the above proof can be modified by using function from
the subspace {f ∈ C([0, T ]; X) : f takes values in D(A)}, which is norming
0
for Lp (0, T ; X ∗ ) = L1 (0, T ; X ∗ ) by Lemma 17.2.33.
Finally, let p = 1. Let g ∈ L∞ (I; X ∗ ) and f ∈ L1 (I; D(A)). By Theorem
17.2.31, A has maximal Lq -regularity on I for all q ∈ (1, ∞). Therefore, by the
previous proof, A∗ has maximal Lq -regularity for all q ∈ (1, ∞). In particular,
Theorem 17.2.19 implies that S ∗ ∗ g takes values in D(A∗ ) almost everywhere
on I. As before one can check that hf, A∗ S ∗ ∗ gi = hAS ∗ fe, gei, and thus
reg
|hf, A∗ S ∗ ∗ gi| 6 M1,A (I)kf kL1 (I;X) kgkL∞ (I;X ∗ ) .

Therefore, as before we obtain that A∗ has maximal L∞ -regularity, with

reg reg
M∞,A ∗ (I) 6 M1,A (I). This completes the proof of the “only if” part in the

case I = (0, T ) is a bounded interval.

0
(2)⇒(1) for bounded intervals I = (0, T ): Let A∗ have maximal Lp -
regularity on I = (0, T ). By Theorem 17.2.15, A∗ is sectorial of angle < π/2,
and hence A is sectorial of angle < π/2 as well, and therefore the densely
defined (by assumption) A generates an analytic C0 -semigroup (S(t))t>0 . To
 17.2 Maximal Lp -regularity 607

complete the proof, first let p ∈ [1, ∞). Fix functions f ∈ Lp (I; D(A)) and
0 0
g ∈ Lp (I; X ∗ ), and define fe and ge as before. By the maximal Lp -regularity
of A∗ on I, the convolution S ∗ ∗ ge takes values in D(A∗ ) almost everywhere
on (0, T ). Using Fubini’s theorem and the same substitutions as before, we
obtain
Z TZ t
hAS ∗ f, gi = hS(t − s)Af (s), g(t)i ds dt
0 0
Z T Z s
= g (t)i dt ds = hf, A∗ S ∗ ∗ gei.
hAfe(s), S ∗ (s − t)e
0 0

reg
Therefore, the maximal Lp -regularity of A and the bound Mp,A (I) 6 Mpreg
0 ,A∗ (I)

follow as before.
If p = ∞, one can take f ∈ L∞ (I; X) and g ∈ L1 (I; D(A∗ )). The density
of D(A∗ ) in X ∗ follows from Lemma 17.2.22 and Remark 17.2.23. Therefore,
we can argue similarly as in the p = 1 case of the implication (1)⇒(2) by
reversing the roles of (A, p, f ) and (A∗ , p0 , g).
(2)⇔(1) for I = R+ : The case I = R+ follows from the previous cases,
and the following identities from Proposition 17.2.18:

Mpreg reg
0 ,A (R+ ) = sup Mp0 ,A (0, T ),
reg
Mp,A reg
∗ (R+ ) = sup Mp,A∗ (0, T ).
T >0 T >0

Extrapolation to weighted spaces

Earlier in this section, we have seen that maximal Lp0 -regularity for one ex-
ponent p0 ∈ [1, ∞] extrapolates to maximal Lp -regularity for all exponents
p ∈ (1, ∞). We shall now consider several extrapolation results to weighted
spaces. Recall that a weight is a measurable function w : (0, ∞) → (0, ∞) such
that w ∈ L1loc [0, ∞). Here we use the notation L1loc (I; X) introduced at the
beginning of this chapter, taking I = (0, ∞). To be explicit, the local integra-
bility condition asks for w to be integrable on every bounded subinterval of
(0, ∞).
Let I = (0, T ) be a bounded interval or I = R+ = (0, ∞).
Given a weight w on (0, ∞) and an exponent p ∈ [1, ∞), the space
Lp (I, w; X) can be defined as the Banach space of strongly measurable func-
tions f : I → X such that t 7→ w1/p (t)f (t) belongs to Lp (I; X). With respect
to the natural norm on this space, for functions belonging to this space we
have the identity
kf kLp (I,w;X) = kw1/p f kLp (I;X) .
In order to also cover the exponent p = ∞ it is useful to proceed slightly
differently.
608 17 Maximal regularity
p
Definition 17.2.34. For weights w ∈ Lloc ([0, ∞)) we define Lpw (I; X) as the
Banach space of strongly measurable functions f : I → X such that t 7→
w(t)f (t) belongs to Lp (I; X). With respect to the natural norm on this space,
we then have the identity

kf kLw
p
(I;X) = kwf kLp (I;X) .

In what follows we will always make the assumption that the weight w is
chosen in such a way that, as sets, we have the inclusion
p 1
Lw (I; X) ⊆ Lloc (I; X), (17.41)

using again the notation introduced at the beginning of the chapter. For p ∈
0
[1, ∞], the inclusion (17.41) holds if and only w−1 ∈ Lploc (I).
With the spaces Lpw (I; X) at hand, we may define the notion of maximal
Lw -regularity in the obvious manner, replacing all occurrences of Lp in Defini-
p

tion 17.2.4 by Lpw . We leave it to the reader to check that several basic results
on maximal Lp -regularity extend mutatis mutandis to maximal Lpw -regularity;
this includes Propositions 17.2.5, 17.2.10, and Theorem 17.2.19.
The present section will be concerned with identifying situations in which
maximal Lpw -regularity can be inferred from maximal Lq -regularity (where
the exponents p and q are possibly different). We start with two elementary
results in this direction for power weights tα and exponential weights e−λt .
From the point of view of applications to evolution equations, these suffice in
most cases; some of them are also valid in some of the end-point cases p = 1
and p = ∞. After that, we consider the more complicated case of Ap -weights,
restricting ourselves to p ∈ (1, ∞) in that case.
In order to treat extrapolation to power weights, we need the following
lemma. At the expense of additional technicalities, the implicit assumption of
strong measurability with respect to the uniform operator norm can be relaxed
somewhat, but in the applications we have in mind, it is always fulfilled.
Lemma 17.2.35. Let X and Y be a Banach spaces, and let p ∈ (1, ∞] and
α ∈ R be fixed. Let k : R+ × R+ → [0, ∞) be given by
|1 − (t/s)α |
k(t, s) = 1t6=s K(t, s),
|t − s|

where K ∈ L∞ (R2+ ; L (X, Y )) is a given function, and let D ⊆ R+ × R+ be a

measurable set. Then the operator Tk : Lp (R+ ; X) → Lp (R+ ; Y ), defined by
Z
Tk f (t) = 1D (t, s)k(t, s)f (s) ds, t ∈ R+ ,
R+

is a bounded of norm at most Cα kKk∞ , where Cα,p is a constant depending

only on α and p, in each of the following cases (with p1 + p10 = 1):
(1) D = {(t, s) : t > s} and α ∈ (−∞, 1/p0 ).
 17.2 Maximal Lp -regularity 609

(2) D = R+ × R+ and α ∈ (−1/p, 1/p0 ).

Proof. Clearly, it suffices to consider α 6= 0. Let c : R+ → R+ be given by

|1 − tα |
c(t) = 1J (t)t1/p ,
|t − 1|

where J = (1, ∞) in case (1) and J = (0, ∞) \ {1} in case (2). Then
Z
kTk f (t)k 6 1D (t, s)kk(t, s)kkf (s)k ds
R+
|1 − (t/s)α | 1/p
Z
−1/p ds
6 kKk∞ t 1J (t/s)(t/s)1/p s kf (s)k .
R+ |(t/s) − 1| s
−1/p
= kKk∞ t c ? g(t),

where g(s) = s1/p kf (s)k and we used the convolution product ? of the group
(R+ , ·) with Haar measure ds/s. Therefore, Young’s inequality implies

kTk f kLp (R+ ;X) 6 kKk∞ kc ? gkLp (R+ , dt )

t

6 kKk∞ kckL1 (R+ , dt ) kgkLp (R+ , dt )

t t

= kKk∞ kckL1 (R+ , dt ) kf kLp (R+ ;X) .

t

It remains to check that kckL1 (R+ , dt ) < ∞. We do this by splitting R+ into

t
the three parts (0, 1/2) ∪ (1/2, 3/2) ∪ (3/2, ∞). Using that α < 1/p0 and p > 1,
we find
Z ∞ Z ∞
dt 1
c(t) 63 t p −2 max{tα , 1} dt < ∞.
3/2 t 3/2

On the interval (1/2, 3/2), c is uniformly bounded by the mean value theorem
R 3/2
and therefore 1/2 c(t) dtt < ∞. On the interval (0, 1/2), c is only non-zero in
case (2), and in this situation α > −1/p implies
Z 1/2 Z 1/2
dt 1
c(t) 62 t p −1 max{1, tα } dt < ∞.
0 t 0

Using Lemma 17.2.35 we can prove the equivalence of weighted and un-
weighted maximal Lp -regularity in case of power weights. The reader may
check that the proof below can be extended to various other classes of integral
operators. Actually, a similar argument has already been used in Proposition
10.2.31 for proving weighted Lp -boundedness of the Hilbert transform.
610 17 Maximal regularity

Proposition 17.2.36 (Extrapolation with power weights). Let A be a

linear operator on a Banach space X, let I = (0, T ) or I = R+ , let p1 + p10 = 1,
and consider the weight

wα (t) = tα with α ∈ (−1/p, 1/p0 ).

Then the following assertions hold:

(1) If p ∈ (1, ∞], then A has maximal Lp -regularity on I if and only if A has
maximal Lpwα -regularity on I.
(2) If A has maximal L1 -regularity on I, then A has maximal L1wα -regularity
on I.

We do not know about a possible converse of (2), but it does not seem to be
particularly relevant for applications.
Proof. (1): Using the scaling properties of the weights wα , one can check that
Theorem 17.2.15 extends to the weighted setting. Therefore, −A generates an
analytic semigroup, which is bounded in case I = R+ . In view of Theorem
17.2.19 and its weighted extension, it suffices to prove that the operator V
introduced in this theorem is bounded on Lp (I; X) if and only if it is bounded
on Lpwα (I; X).
The boundedness of V on Lpwα (I; X) is equivalent to the boundedness of
the operator Vα on Lp (I; X), where
Z
Vα f (t) := A 1t>s tα S(t − s)s−α f (s) ds.
I

For λ ∈ %(A),

(λ − A)−1 Vα f (t)−(λ − A)−1 V f (t)

Z
= 1t>s A(λ − A)−1 S(t − s)((t/s)α − 1)f (s) ds
I
Z
= (λ − A)−1 k(t, s)f (s) ds,
I

where k(t, s) is as in Lemma 17.2.35(1) with K(t, s) = 1t>s R· (t − s)AS(t − s).

Since K ∈ L∞ (R2+ ; L (X)), the lemma gives that f 7→ I k(·, s)f (s) ds is
bounded on Lp (I; X). Applying λ − A on both sides of the above identity, we
conclude that the boundedness of V and Vα are equivalent.
(2): By Lemma 17.2.22 there is a constant C such that
Z
kAS(t)xk dt 6 Ckxk, x ∈ X.
I

Let f ∈ L1wα (I; X) and set T = ∞ if I = R+ . Then, since α 6 0,

 17.2 Maximal Lp -regularity 611
Z Z t Z Z T
kAS(t − s)f (s)k ds wα (t) dt = kAS(t − s)f (s)ktα dt ds
I 0
ZI Zs
6 kAS(t)f (s)ksα dt ds
I I
6 Ckf kLw
1 (I;X) .
α

1
This implies that V is bounded on Lw α
(I; X) with norm at most C. 
The proof actually shows that, in case −A generates a bounded analytic semi-
group, we could also allow any α 6 −1/p in Proposition 17.2.36. Since this
is unimportant for applications to evolution equations and since such weights
are not locally integrable near zero (which would lead to problems in Theorem
17.2.15), we do not elaborate on this any further.
We will apply the above result to the inhomogeneous problem
(
u0 (t) + Au(t) = f, t ∈ I,
(17.42)
u(0) = x,
where I = (0, T ) or I = R+ . The aim is to extend Proposition 17.2.14 to
the weighted setting. In the absence of weights, this proposition was proved
using the trace method. The proof presented here is different and uses the
semigroup generated by −A.
We fix p, p0 ∈ [1, ∞] satisfying p1 + p10 = 1, and consider the weight
wα (t) = tα for α ∈ (−1/p, 1/p0 ) ∪ {0}.
We define
X1− p1 −α,p := (X, D(A))1− p1 −α,p for α ∈ (−1/p, 1/p0 )

and set X0,1 := X and X1,∞ := D(A).

Corollary 17.2.37. Let A be a linear operator acting in a Banach space X.
If A has maximal Lp -regularity on I and f ∈ Lpwα (I; X), then for all x ∈
X1− p1 −α,p there is a unique strong solution u of (17.42) such that Au, u0 ∈
Lpwα (I; X), and there is a constant C > 0, independent of f and x, such that
kAukLpwα (I;X) + ku0 kLpwα (I;X) 6 C(kf kLpwα (I;X) + kxkX1− 1 −α,p ). (17.43)
p

As in Proposition 17.2.13 one checks that the conditions on the data (f, x)
are also necessary if α + p1 ∈ (0, 1).
By choosing α appropriately, we can allow any u(0) = x ∈ (X, D(A))θ,p for
θ ∈ (0, 1). On the other hand, if p ∈ (1, ∞), we can apply Theorem L.4.1 to see
that u ∈ C(I; X1− p1 ,p ) (i.e, the solution regularises instantaneously from its
initial value u(0) into X1− p1 ,p . This agrees with the behaviour of Lp -solutions
without weights. Note that an estimate for kukLpwα (I;X) is not included in
(17.43), but can be obtained in the case of a bounded interval I = (0, T ) by
the same method as in (17.5).
612 17 Maximal regularity

Proof. By Dore’s Theorem 17.2.15, −A generates an analytic semigroup S,

and this semigroup is uniformly bounded if I = R+ . Proposition 17.1.3 then
gives the uniqueness of strong solutions.
p
By Proposition 17.2.36, A has maximal Lw α
-regularity on I. To derive exis-
tence and the estimate (17.42) we write u = uf +ux , where uf ∈ Lpwα (I; D(A))
is the strong solution to (17.42) with x = 0, and ux = S(·)x. By maximal Lpwα -
regularity we obtain

kAuf kLpwα (I;X) 6 Ckf kLw

p
α (I;X)
.

Also, by Theorem L.2.4,

kAux kLw
p
α (I;X)
6 CkxkX1− 1 −α,p .
p

Note that the extremal case α = 0 and p = ∞ is trivial. For the extremal case
α = 0 and p = 1 we use Lemma 17.2.22.
Clearly, u = uf + ux is a strong solution to (17.42), and combining the
estimates gives (17.43). 
For later use we observe that, by restriction and uniqueness, the optimal
constant C in the estimate is monotone with respect to this interval I.
Next we consider the case of exponential weights. Here, only the case
I = R+ is of interest, as t 7→ eλt is uniformly bounded from below and above
on bounded subsets of R+ .

Proposition 17.2.38 (Extrapolation with exponential weights). Sup-

pose that A is a linear operator in X that has maximal Lp -regularity on R+
for some p ∈ [1, ∞). Then for all λ > 0, the operator A has maximal Lpe−λ(·) -
reg
regularity on R+ with constant at most Mp,A (R+ ).

Proof. Since A has maximal Lp -regularity on R+ , there exists a unique strong

solution to the problem (17.42) with initial value x = 0. Therefore, it suffices
to prove Lpe−λ(·) (R+ ; X)-estimates for the solution.
For f ∈ Lpe−λ(·) (R+ ; X) we write fλ = e−tλ f . By Proposition 17.2.27 there
exists a unique Lp -solution v of v 0 + (A + λ)v = fλ with initial condition
v(0) = 0, and
reg reg
kAvkLp (R+ ;X) 6 Mp,A (R+ )kfλ kLp (R+ ;X) = Mp,A (R+ )kf kLp−λ(·) (R+ ;X) .
e

It follows that the function u defined by u(t) = eλt v(t) is the unique strong
solution to (17.42) with initial condition x = 0 and by the previous estimate
we obtain
reg
kAukLp−λ(·) (R+ ;X) = kAvkLp (R+ ;X) 6 Mp,A (R+ )kf kLp−λ(·) (R+ ;X) .
e e


 17.2 Maximal Lp -regularity 613

We continue with a result that allows us to extrapolate to weighted spaces

Lp (I, w; X), with weights w taken from the Muckenhoupt Ap class (see Ap-
pendix J). Here we will rely on the A2 Theorem 11.3.26. In order to make a
connection with Propositions 17.2.36 and 17.2.38 we record two simple obser-
vations: (i) | · |αp ∈ Ap if and only if α ∈ (−1/p, 1/p0 ), and (ii) the exponential
weights eλ(·) belong to Ap only in the trivial case λ = 0. Both assertions follow
from elementary calculations.
Theorem 17.2.39 (Extrapolation with Ap -weights). Suppose that A is a
linear operator in X that has maximal Lp0 -regularity on an interval I for some
p0 ∈ [1, ∞]. Then for all p ∈ (1, ∞) and weights w ∈ Ap (I) the operator A has
maximal Lp (w)-regularity on I in the sense that for all f ∈ Lp (I, w; X) there
exists a unique strong solution u of (ACP0 ), which satisfies Au ∈ Lp (I, w; X)
and
  max(1, 1 )
kAukLp (I,w;X) 6 cp Mpreg0 ,A
(I) + C(1 + log + C) [w]Ap (I) p−1 kf kLp (I,w;X) ,

where cp depends only on p, and C := supt∈I tkAS(t)k, where S is the analytic

semigroup generated by A under the assumptions.
Proof. Fix p ∈ (1, ∞) and w ∈ Ap (I). Theorem 17.2.19 extends to the
weighted setting. Accordingly, it suffices to check that the operator V defined
in that theorem extends to a bounded operator on Lp (I, w; X). By Lemma
17.2.30, the operator V has kernel K, which is a Dini kernel with cK 6 C and
kωK kDini 6 4C(1 + log+ (4C)). Therefore, by Theorem 11.3.26, applied in the
case of a one-dimensional cube I = (0, T ) or quadrant I = R+ , the operator
V extends to a bounded operator on Lp (I, w; X) of norm
  max(1, 1 )
kV kL (Lp (I,w;X)) 6 cp kV kL (Lp0 (I;X)) + cK + kωK kDini [w]Ap (I) p−1
  max(1, 1 )
6 c0p Mpreg
0 ,A
(I) + C(1 + log + C) [w]Ap (I) p−1 .

As in Theorem 17.2.19 we see that for all f ∈ Lp (I, w; X) there exists a unique
Lp (w)-solution u of (ACP0 ), which satisfies Au = V f , and hence the claimed
estimate for solutions follows by combining the previous operator norm bound
with

kAukLp (I,w;X) 6 kV kL (Lp (I,w;X)) kf kLp (I,w;X) .

17.2.f Maximal continuous regularity

In this section, we introduce the notion of maximal continuous regularity, or,

more briefly, maximal C-regularity, and show that for densely defined opera-
tors it is equivalent to maximal L∞ -regularity. Once this has been proved, it
is no longer necessary to distinguish between these notions.
614 17 Maximal regularity

Throughout this section we let I = (0, T ) be a bounded interval or I = R+ .

As before we consider the inhomogeneous abstract Cauchy problem
(
u0 (t) + Au(t) = f (t), t ∈ I,
(ACP0 )
u(0) = 0,

Definition 17.2.40. A linear operator A acting in a Banach space X has

maximal C-regularity on I if there exists a constant C > 0 such that for all
f ∈ Cb (I; X) the problem (ACP0 ) admits a unique strong solution uf on I,
the almost everywhere defined function Auf is equal almost everywhere to a
function in Cb (I; X), and this version satisfies

kAuf kCb (I;X) 6 Ckf kCb (I;X) . (17.44)

The least admissible constant in this definition will be called the maximal
reg
C-regularity constant of A on I and will be denoted by Mcont,A (I).

Proposition 17.2.41. If A has maximal C-regularity on I, then:

(1) A is closed;
(2) for all f ∈ Cb (I; X), the strong solution uf takes values in D(A) and
defines a continuous function from I to D(A).

Remark 17.2.42. As a consequence of this proposition, an equivalent definition

of maximal C-regularity is obtained if these properties are built into the def-
inition. It is with this equivalent version of the definition that we will always
work.

Proof. (1): This follows by repeating the proof of the corresponding result for
maximal Lp -regularity (Proposition 17.2.5) verbatim.
(2): Fix f ∈ Cb (I; X). It follows from the definition of a strong solution
that uf (t) ∈ D(A) for almost all t ∈ I and uf ∈ Cb (I; X). By the definition
of maximal C-regularity, the almost everywhere defined function Auf is equal
almost everywhere, say outside the null set Nf , to a function F ∈ Cb (I; X).
Now let t ∈ I be arbitrary and choose a sequence tn → t, with every tn in
I \ Nf . Then uf (tn ) → uf (t) in X by continuity, and Auf (tn ) → F (t) in X.
Hence by closedness, uf (t) ∈ D(A) and similarly Auf (t) = F (t). This proves
the first assertion. The second is now evident from the fact that both uf and
Auf are well defined pointwise and continuous as X-valued functions. 
As in the case of maximal Lp -regularity, in order to check maximal C-
regularity for a given closed operator A it suffices to prove existence, unique-
ness, and the estimate (17.44) for strong solutions corresponding to inhomo-
geneities f in a dense subspace F on Cb (I; X). The reader may check that
all results in Section 17.2.a have an analogue for maximal C-regularity, with
similar proofs. In particular we record the following analogue of Proposition
17.2.7:
 17.2 Maximal Lp -regularity 615

Proposition 17.2.43. Suppose that A has maximal C-regularity on I. If uf

is the unique strong solution on I of (ACP0 ), then for all bounded subintervals
(0, T ) ⊆ I we have uf ∈ C([0, T ]; D(A)) ∩ C 1 ([0, T ]; X) and
reg
kuf kC([0,T ];D(A))∩C 1 ([0,T ];X) 6 (1 + T )(Mcont,A (I) + 1)kf kCb (I;X) .

The following is the analogue of Dore’s Theorem 17.2.15 and can be proved
in exactly the same way.

Theorem 17.2.44. Let A be a linear operator on a Banach space X. Then:

(1) if A has maximal C-regularity on a bounded interval (0, T ), then −A gen-
erates an analytic semigroup on X, and λ + A is sectorial of angle < π/2
for all λ ∈ R large enough;
(2) if A has maximal C-regularity on R+ , then −A generates a bounded ana-
lytic semigroup on X, and A is sectorial of angle < π/2. Moreover,
reg
kA(λ + A)−1 k 6 Mcont,A (R+ ), λ ∈ C+ . (17.45)

As a consequence, we obtain analogues for maximal C-regularity of Lemma

17.2.16, Proposition 17.2.18, Theorem 17.2.19, Theorem 17.2.24, Corollary
17.2.25, and Theorem 17.2.26(1)–(3).
By Proposition 17.1.3, for every f ∈ C(I; X) the strong solution u provided
by Definition 17.2.40 coincides with the mild solution:
Z t
u(t) = S(t − s)f (s) ds, t ∈ I.
0

This will be used in the proof of Theorem 17.2.46.

It is important to observe that Cb (I; D(A)) does not need to be dense in
Cb (I; X) if I is unbounded, even when D(A) is dense in X. (While it is easy to
approximate f ∈ Cb (I; X) by functions with values in D(A), the difficulty is
ensuring that the approximating function would also be bounded with respect
to the graph norm kxkD(A) = kxkX + kAxkX of D(A).) However, we do have
the following result.

Lemma 17.2.45 (Density). Suppose that A is densely defined and −A gen-

erates a C0 -semigroup S. Let I = R+ or I = (0, T ). If there exists a constant
C > 0 such that for all f ∈ Cb (I; D(A)) we have

kAS ∗ f (t)k 6 Ckf kCb (I;D(A)) , t ∈ I,

reg
then A has maximal C-regularity on I and Mcont,A (I) 6 C.

Under the assumptions stated, S ∗ f is continuous with values in D(A), and

therefore AS ∗ f (t) is well defined for all t ∈ I.
616 17 Maximal regularity

Proof. If I = (0, T ), the result is clear from the density of Cb (I; D(A))
in Cb (I; X). Next let I = R+ . We already observed that u := S ∗ f
is continuous with values in D(A). Let f ∈ Cb (I; X) and fix T > 0.
Choose fn ∈ Cb ([0, T ]; D(A)) such that fn → f |[0,T ] in Cb ([0, T ]; X). Let
f (t) = f (T ∧ t) and f n (t) = fn (T ∧ t). Then f n → f in Cb ([0, ∞); X).
It follows that un := S ∗ f n is a Cauchy sequence and hence convergent in
Cb ([0, T ]; D(A)). Since un → u := S ∗ f in X pointwise on [0, T ], it follows
that u := S ∗ f ∈ Cb ([0, T ]; D(A)) and

kAukCb ([0,T ];X) = lim kAun kCb ([0,T ];X)

n→∞
6 lim Ckf n kCb ([0,T ];X)
n→∞
6 lim Ckf n kCb ([0,∞);X) 6 Ckf kCb ([0,∞);X)
n→∞

Since T was arbitrary this proves that u = S ∗ f satisfies Au ∈ Cb ([0, ∞); X)

and kAukCb ([0,∞);X) 6 Ckf kCb ([0,∞);X) . 
Our next aim is to show that maximal C-regularity coincides with maximal
L∞ -regularity.
Theorem 17.2.46 (Continuous versus L∞ -type maximal regularity).
A densely defined linear operator A has maximal C-regularity if and only if it
reg reg
has maximal L∞ -regularity, and in that case Mcont,A (I) = M∞,A (I).
Proof. Each of the two conditions implies that −A generates a C0 -semigroup
S on the Banach space X in which A acts.
Suppose first that A has maximal L∞ -regularity. Let f ∈ Cb (I; D(A)).
Then u = S ∗ f is the unique L∞ -solution to (ACP0 ). Since u is continuous
with values in D(A) it follows that
reg
sup kAu(t)k = kAukL∞ (I;X) 6 M∞,A (I)kf kL∞ (I;X) .
t∈I

This implies that A has maximal C-regularity. This part of the proof does not
use the density of D(A).
Next suppose that A is densely defined and has maximal C-regularity.
As in the proof of Proposition 17.2.32 one shows that A∗ has maximal L1 -
reg
regularity with constant at most Mcont,A (I). Now Proposition 17.2.32 implies
reg reg reg
that A has maximal L∞ -regularity with M∞,A (I) = M1,A ∗ (I) 6 Mcont,A (I).


Conditions for maximal L∞ and C-regularity will be given in Section 17.3.c.
For later applications we also state a version of Proposition 17.2.36 for
maximal C-regularity with power weights. For this purpose, for α ∈ R we let
wα (t) := tα and consider the Banach space
n o
Cwα (I; X) := f ∈ C(I \ {0}; X) : sup ktα f (t)k < ∞
t∈I
 17.2 Maximal Lp -regularity 617

with norm

kf kCwα (I;X) := sup ktα f (t)k.

t∈I

Replacing Cb (I; X) by Cwα (I; X) in the definition of maximal C-regularity,

we obtain the definition of maximal Cwα -regularity.

Proposition 17.2.47 (Extrapolation with power weights). Let A be a

linear operator A in X. For any α ∈ [0, 1), the following assertions are equiv-
alent:
(1) A has maximal C-regularity on I;
(2) A has maximal Cb,α -regularity on I.

Proof. Fixing α ∈ [0, 1), we can repeat the argument of Proposition 17.2.36,
using the additional fact that in Lemma 17.2.35(1) the operator Tk maps
L∞ (I; X) into Cb (I; Y ) if K is continuous (using notation introduced in the
lemma). To check the latter, let

|1 − (t/s)α |
c(t, s) := 1t>s
|t − s|

and recall from the proof of Lemma 17.2.35 that c(t, ·) ∈ L1 (I). We have
Z
kTk f (t + h) − Tk f (t)kY 6 kf k∞ c(t, s)kK(t + h, s) − K(t, s)k ds
I
Z
+ kf k∞ kKk∞ |c(t + h, s) − c(t, s)| ds.
I

The first term tends to zero as h → 0 by dominated convergence. For the

second term, note that, if h > 0, one can write
Z Z t Z t+h
|c(t + h, s) − c(t, s)| ds = |c(t + h, s) − c(t, s)| ds + c(t + h, s) ds.
I 0 t

To estimate the second term here, the mean value theorem applied to xα − 1
with x > 1, gives
Z t+h Z t+h t+h Z t+h
s −1 α
c(t + h, s) ds 6 α ds = ds → 0
t t t+h−s t s
as h ↓ 0. For the first term, the triangle inequality and the mean value theorem
lead to
(t + h)α − tα 1 1
|c(t + h, s) − c(t, s)| 6 + ((t/s)α − 1) −
sα (t + h − s) t+h−s t−s
tα−1 h h
6α α + ((t/s)α − 1)
s (t + h − s) (t + h − s)(t − s)
618 17 Maximal regularity

=: I + II.
α−1 α−1
Since sα t(t+h−s)
h
6 t sα is integrable, part I can be handled via dominated
convergence. For part II, we estimate
Z t/2
4h t/2
Z
II ds 6 2 ((t/s)α − 1) ds → 0,
0 t 0
and, the mean value theorem and dominated convergence theorem,
Z t Z t  Z t
t  h αh
II ds 6 α −1 ds = ds → 0.
t/2 t/2 s (t + h − s)(t − s) t/2 s(t + h − s)
The case h < 0 can be treated in a similar way. 
Next we present a version of Corollary 17.2.37 for maximal C-regularity. In
Chapter 18 we also need the variants of the on subspaces of Cwα , which we
first introduce. For α ∈ [0, 1), I = (0, T ] or I = (0, ∞) let
Cwα ,0 (I; X) := {u ∈ Cwα (I; X) : lim tα ku(t)k = 0},
t↓0
1
Cw α
(I; X) := {u ∈ Cwα (I; X) : u0 ∈ Cwα (I; X)}.

kukCw1 α (I;X) = sup tα ku(t)k + sup tα ku0 (t)k

t∈I t∈I
1 1
for u ∈ Cw α
(I; X).
Then one can check that both Cwα (I; X) and Cw α
(I; X)
are Banach spaces, and Cwα ,0 (I; X) is a closed subspace of Cwα .
Corollary 17.2.48. Let A be a linear operator acting in a Banach space X,
and suppose that A has maximal C-regularity on I = (0, T ] or I = R+ . For
α ∈ (0, 1) let X1−α,∞ denote the closure of D(A) in the real interpolation
space (X, D(A))1−α,∞ . Then for all f ∈ Cwα (I; X) and all x ∈ X1−α,∞ there
exists a unique strong solution u of (17.42) such that Au, u0 ∈ Cwα (I; X), and
there is a constant C > 0, independent of f and x, such that
kAukCwα (I;X) + ku0 kCwα (I;X) 6 C(kf kCwα (I;X) + kxkX1−α,p ).
If additionally f ∈ Cwα ,0 (I; X), then Au, u0 ∈ Cwα ,0 (I; X).
Proof. We argue analogously as in Corollary 17.2.37, where the Lpwα -setting
was discussed. For the initial value part, a different argument is needed. Let
ux = S(·)x. We claim that Aux ∈ Cwα ,0 (I; X). For x ∈ D(A), the claim is
clear, and moreover by Corollary 17.2.37 and continuity
kAux kCwα (I;X) + ku0x kCwα (I;X) 6 CkxkX1−α,p .
Therefore, by a density argument the latter extends to all x ∈ X1−α,p .
It remains to check the final assertion for f ∈ Cwα ,0 (I; X). By density it
suffices to consider f : I → X such that f = 0 in a neighbourhood (0, δ) of
zero. By the equivalence with mild solutions (see Proposition 17.1.3), one sees
that u = 0 on (0, δ). Therefore, Au, u0 ∈ Cwα ,0 (I; X) as required. 
 17.2 Maximal Lp -regularity 619

17.2.g Perturbation and time-dependent problems

In this section we will prove several perturbation results for maximal Lp -

regularity with p ∈ [1, ∞]. Later on, in Corollary 17.3.10, we will improve
these results for UMD spaces X and p ∈ (1, ∞).
We will use the method of continuity introduced in the preceding chapter
(see Lemma 16.2.2). For the reader’s convenience, we recall its formulation:
Let (Lt )t∈[0,1] be a family of bounded linear operators from a Banach space E
into another Banach space F such that t 7→ Lt is continuous from [0, 1] into
L (E, F ). Suppose furthermore that there exists a constant C > 0 such that
for all t ∈ [0, 1] and all x ∈ E we have

kxk 6 CkLt xk.

Then L0 is surjective if and only if L1 is surjective.

We begin by proving an additive perturbation result for relatively A-
bounded perturbations B in the sense of (17.46) below. The proof uses a
variation of the argument in Theorem 16.2.3, as well as the flexibility in the
estimates with the additional parameter λ as discussed in Proposition 17.2.27.

Proposition 17.2.49 (Perturbation). Let A be a linear operator with max-

imal Lp -regularity on I for some p ∈ [1, ∞]. Suppose that B : D(A) → X is a
linear operator that satisfies

kBxk 6 δkAxk + Kkxk, x ∈ D(A), (17.46)

reg
where δ ∈ (0, 1) satisfies δMp,A (I) < 1 and K > 0. Then the operator A + B
with domain D(A + B) := D(A) is closed, and the following assertions hold:
(1) For all λ ∈ R large enough, λ + A + B has maximal Lp -regularity on I.
(2) If I = (0, T ) is bounded, then A + B has maximal Lp -regularity on I.
(3) If I = R+ and K = 0, then A + B has maximal Lp -regularity on I.

By the p-independence of maximal Lp -regularity for p ∈ (1, ∞), in applying

reg
this proposition, one can choose the value of p for which Mp,A (I) is minimal, in
reg
order to allow the maximal range of δ ∈ (0, 1) in the condition δMp,A (I) < 1.
Proof. Closedness of A + B can be proved by repeating the corresponding
argument in the proof of Theorem 16.2.3. For notational convenience, in the
reg
remainder of the proof we will use the short-hand notation C := Mp,A (I) and
kukp = kukLp (I;X) .
(1): By Dore’s Theorem 17.2.15, −A generates an analytic semigroup
(S(t))t>0 on X. Let M > 0 and ω ∈ R be such that kS(t)k 6 M eωt for
all t > 0, and that 0 ∈ %(λ + A) for all λ > ω, where the latter is possible by
Definition K.1.2(i).
We will first show that λ + A + B has maximal Lp -regularity for λ > ω
large enough. In order to prove this result, we will apply Lemma 16.2.2 with
620 17 Maximal regularity

E = {u ∈ W 1,p (I; X) ∩ Lp (I; D(A)) : u(0) = 0},

and F = Lp (I; X) with Lθ u = u0 + (λ + A + θB)u for θ ∈ [0, 1]. The space E

is a Banach space with respect to the norm

kukE = (λ − ω)kukp + kAukp + ku0 kp .

We first prove the following a priori estimate: for all λ > ω large enough there
exists a constant D > 0 such that

kukE 6 DkLθ ukp , u ∈ E. (17.47)

Let u ∈ E and set f := Lθ u. Then u0 + (λ + A)u = f − θBu, and therefore,

by Proposition 17.2.27,

kAukp 6 Ckf kp + CkBukp 6 Ckf kp + CδkAukp + CKkukp .

Since Cδ < 1, we obtain

kAukp 6 C0 kf kp + C0 Kkukp , (17.48)

where C0 = C(1 − Cδ)−1 . To estimate (λ − ω)kukp , we let fe and u

e denote the
zero extensions of and f and u to all of R+ (to deal with the case I = (0, T ))
and use the assumption u(0) = 0 along with Young’s inequality to estimate

ukLp (R+ ) 6 Sλ ∗ (fe − θBe

kukLp (I) = ke u) Lp (R )
+
Z ∞
6 kSλ (t)k dt · k(f − θBu)kLp (I) ,
0

−λt
where Sλ (t) = e S(t) is the semigroup generated by −A − λ. Hence, for
λ > ω,
Z ∞
(λ − ω)kukp 6 M (λ − ω)e−(λ−ω)t dt · (kf kp + kBukp )
0
6 M (kf kp + δkAukp + Kkukp ) 6 C1 kf kp + C1 Kkukp ,

where in the last step we used (17.48) and took C1 = M (δC0 + 1). For λ >
ω + C1 K this gives

kukp 6 C1 (λ − ω − C1 K)−1 kf kp . (17.49)

Substituting the latter in (17.48) gives

kAukp 6 C2 kf kp , (17.50)

where C2 = C0 + C0 KC1 (λ − ω − C1 K)−1 . Combining the estimates (17.49)

and (17.50) with the equation for u we obtain
 17.2 Maximal Lp -regularity 621

ku0 kp 6 λkukp + kAukp + kBukp + kf kp 6 C3 kf kp , (17.51)

where C3 is a constant depending on C, δ, K, M , and λ. Combining the

estimates (17.49), (17.50), and (17.51), and recalling that f = Lθ u, we obtain
(17.47).
By Theorem 17.2.26(1), λ + A has maximal Lp -regularity. Hence the op-
1,p
erator L0 : u 7→ u0 + (λ + A)u from 0 Ẇλ+A (I; X) to Lp (I; X) is surjective.
For the “large enough” λ that we consider, we also have 0 ∈ %(λ + A) by
Definition K.1.2(i). Under this condition, Proposition 17.2.8 applies to λ + A
in place of A, and guarantees that
1,p
0 Ẇλ+A (I; X) ' Lp (I; D(λ + A)) ∩ 0 W 1,p (I; X)
' Lp (I; D(A)) ∩ 0 W 1,p (I; X) ' E,

where E is the space introduced earlier in this proof; we used the easy obser-
vation that the domain of λ + A and A coincide as sets, and the associated
graph norms are equivalent. Thus L0 : E → F = Lp (I; X) is surjective.
From Lemma 16.2.2 we deduce that L1 is surjective, and hence invertible
by (17.47). Therefore, for each f ∈ Lp (I; X) there exists a unique u ∈ E with
u0 + (λ + A + B)u = f and (17.47) we obtain

k(λ + A + B)ukp 6 λkukp + kAukp + kBukp

e k,
6 (λ + K)kukp + (1 + δ)kAukp 6 Dkf

where D e = (λ + K)(λ − ω)−1 D + (1 + δ)D. This proves that λ + A + B has

maximal Lp -regularity.
(2): By (1) and Theorem 17.2.26(1), A + B has maximal Lp -regularity.
(3): Let f ∈ Lp (R+ ; X). It follows from (2) that u0 + (A + B)u = f with
u(0) = 0 has a unique Lp -solution on every interval (0, T 0 ) with T 0 < ∞,
and therefore by uniqueness, we can construct a strong solution u on R+ . It
remains to estimate k(A + B)ukp . Since A has maximal Lp -regularity, writing
u0 + Au = f − Bu it follows that

kAukp 6 M kf kp + M kBukp 6 M kf kp + M δkAukp .

This implies kAukp 6 M (1 − M δ)−1 kf kp . Moreover, kBukp 6 δkAukp 6

δM (1 − M δ)−1 kf kp . Therefore A + B has maximal Lp -regularity on R+ with
reg
constant Mp,A+B (R+ ) 6 (1 + δ)M (1 − M δ)−1 . 
As an immediate consequence, maximal Lp -regularity is stable under lower
order perturbations:

Corollary 17.2.50. Let A be a linear operator with maximal Lp -regularity on

I for some p ∈ [1, ∞] and T ∈ (0, ∞]. Suppose that B : (X, D(A))θ,1 → X is
a bounded linear operator for some θ ∈ (0, 1).
622 17 Maximal regularity

(1) For all λ ∈ R large enough, λ + A + B has maximal Lp -regularity on I.

(2) If I = (0, T ) is bounded, then A + B has maximal Lp -regularity on I.

Proof. Under the stated assumptions, the relative smallness condition (17.46)
on B was checked in the proof of Corollary 16.2.5. Thus the claims are imme-
diate from Proposition 17.2.49. 
As an application of the previous perturbation results, we prove well-posedness
for a non-autonomous Cauchy problem. This result can be seen as an extension
of Corollary 17.2.37 to the time-dependent setting. Let X0 and X1 be Banach
spaces such that X1 ,→ X0 with a continuous embedding, and let A : [0, T ] →
L (X1 , X0 ) be strongly measurable in the strong operator topology. Consider
the time-dependent inhomogeneous problem
(
u0 (t) + A(t)u(t) = f (t), t ∈ (0, T )
(17.52)
u(0) = x.

Mutatis mutandis, the notion of a strong solution can be extended to (17.52).

In the next result we interpret X0,1 = (X0 , X1 )0,1 as X0 ; this case arises
if p = 1 and α = 0. A version for maximal C-regularity instead can be proved
in a similar way.

Theorem 17.2.51 (Maximal Lp -regularity for time-dependent A). Let

X0 and X1 be Banach spaces with continuous embedding X1 ,→ X0 . For a
given p ∈ [1, ∞) and
(
0 [0, 1/p0 ), if p ∈ (1, ∞),
α ∈ [0, 1/p ) ∪ {0} =
{0}, if p = 1,

consider the space

X1− p1 −α,p := (X0 , X1 )1− p1 −α,p .
Let A ∈ C([0, T ]; L (X1 , X0 )) be a mapping with the following two properties:
(i) there exists a constant L > 0 such that for all t ∈ [0, T ],

L−1 kxkX1 6 kA(t)xkX0 + kxkX0 6 LkxkX1 , x ∈ X1 ; (17.53)

(ii) for all t ∈ [0, T ] the operator A(t), viewed as an operator acting in X0
with domain D(A(t)) = X1 , has maximal Lp -regularity on (0, T ) with
reg
M := sup Mp,A(t) (0, T ) < ∞.
t∈[0,T ]

Set wα (t) := tα . Under these assumptions, for all x ∈ (X0 , X1 )1− p1 −α,p and
f ∈ Lpwα (0, T ; X0 ) there exists a unique strong solution

u ∈ Lpwα (0, T ; X1 ) ∩ Ww1,p

α
(0, T ; X0 ) ∩ C([0, T ]; X1− p1 −α,p ) ∩ C((0, T ]; X1− p1 ,p ),
 17.2 Maximal Lp -regularity 623

of (17.52), and there exists a constant C > 0, only depending on A, p, α, and

T , such that

kukLw
p
α (0,T ;X1 )
+ kukWw1,p (0,T ;X0 )
α

+ kukC([0,T ];X1− 1 −α,p ) + sup tα ku(t)kX1− 1 ,p

p t∈(0,T ] p (17.54)
6 Ckf kLpwα (0,T ;X1 ) + CkxkX1− 1 −α,p .
p

Assumption (i) says that the graph norms of A(t), viewed as an operator
acting in X0 with domain D(A(t)) = X1 , are equivalent to the norm of X1 ,
uniformly with respect to t ∈ [0, T ]; Assumption (ii) says that these operators
have maximal Lp -regularity on (0, T ), uniformly with respect to t ∈ [0, T ].
The fact that u(t), for t > 0, takes values in X1− p1 ,p , which is a smaller
space than the initial value space X1− p1 −α,p if α > 0, is referred to by saying
that the solution instantaneously regularises if α > 0.
The proof of the theorem uses an iteration argument, in which the interval
[0, T ] is subdivided into smaller intervals [tn , tn+1 ]. On each of these, the
problem (17.52) is solved by a fixed point argument, taking u(tn ) as the initial
value for solving (17.52) on [tn , tn+1 ]. For this to work, it is crucial that
u(tn ) belong to the correct trace space; the bookkeeping needed to achieve
this is made possible by maximal Lp -regularity and is effected through a
judicious choice of the fixed point space. In the process, we ascertain that
upon completion of the (n + 1)th step, a suitable a priori inequality that
encodes all this information is being extended from [0, tn ] to [0, tn+1 ].
Implicit in this reasoning are two observations, the easy proofs of which
we leave to the reader:
• Gluing strong solutions on subintervals gives a strong solution on the full
interval. This follows by iterating the definition of a strong solution.
• Gluing Sobolev functions on subintervals gives a Sobolev function on the
full interval. This can be seen, e.g., through the characterisation of W 1,p -
functions as indefinite integrals of Lp -functions (see Section 2.5.c).
Proof. For each τ ∈ (0, T ] let Mτ : Lp (0, T ; X0 ) → Lp (0, T ; X1 ), f 7→ uf be
the solution map defined in Corollary 17.2.9, associated with A(τ ) in place of
A. (Note that the subscript τ of Mτ refers only to the operator A(τ ), whereas
the time interval under consideration is the same (0, T ) for all τ .)
Since A(τ ) has maximal Lp -regularity on (0, T ) with constant M , it follows
from (17.53) and (17.10) that
 
kuf kLp (0,T ;X1 ) 6 L kA(τ )uf kLp (0,T ;X0 ) + kuf kLp (0,T ;X0 )
 
6 L M kf kLp (0,T ;X0 ) + T (M + 1)kf kLp (0,T ;X0 )
6 L(1 + T )(M + 1)kf kLp (0,T ;X0 ) ,
624 17 Maximal regularity

and hence
kMτ k 6 (1 + T )(M + 1)L =: M
f1 .
By the extrapolation of maximal regularity with power weights (Proposition
17.2.36), A(0) has maximal Lpwα -regularity on (0, T ) with a certain constant
M0 and therefore in a similar way as before we see that the associated mapping
M0 : L wp
α
(0, T ; X0 ) → Lw p
α
(0, T ; X1 ) is bounded by with constant M
f0 .
Let M f = max{M f1 }. By uniform continuity, we can choose δ > 0 such
f0 , M
that, for all s, t ∈ [0, T ],
1
|t − s| < δ ⇒ kA(t) − A(s)kL (X1 ,X0 ) < .
2M
f

Let N > 1 be an integer such that T /N < δ. Set tn := nT N for n ∈ {0, . . . , N },

and set t−1 := 0.
Let n ∈ {0, . . . , N − 1} be fixed. Suppose that a unique strong solution u
on (0, tn ) exists, and (17.54) holds on [0, tn ] instead of [0, T ] with constant Cn
instead of C. (Note that this assumption is vacuous for n = 0.) We will show
how to obtain a unique strong solution and the estimate (17.54) on [0, tn+1 ].
Let β := α if n = 0, and β := 0 if n > 1.
If p ∈ (1, ∞), we can apply Theorem L.4.1, to find that there is a constant
C > 0 such that

kukC([0,tn+1 ];X1− 1 −β,p ) + sup tβ ku(t)kX1− 1 ,p

p t∈(0,tn+1 ] p
(17.55)
6 CkukLpwβ (0,tn+1 ;X1 ) + CkukWw1,p (0,tn+1 ;X0 ) .
β

The case p = 1 is not covered by Theorem L.4.1, and we need to argue

differently. Then α = 0 by assumption, and thus β = 0 in all cases. Hence
the right-hand side of (17.55) involves unweighted norms only, and the two
terms on the left of (17.55) coincide in this case. Moreover, X0,1 = X0 by the
convention that we made right before the statement of Theorem 17.2.51 that
we are proving. Thus (17.55) for p = 1 follows simply from the continuity of the
embedding W 1,1 (0, tn+1 ; X0 ) ,→ C([0, tn+1 ]; X0 ); the first term on the right
of (17.55) is not even needed in this case. Summarising, we have established
(17.55) for all p ∈ [1, ∞).
By the validity of the estimate (17.54) on [0, tn ], and the fact that wα (t) =
tα is bounded away from zero for t ∈ [t1 , T ] (allowing us to pass from the
weight wα used on [0, t1 ] to the trivial weight w0 = 1 on the intervals [tn , tn+1 ]
for n > 1), it is therefore enough to prove the estimate

kukLpwβ (tn ,tn+1 ;X1 ) + kukWw1,p (tn ,t+1 ;X0 )

β

6 Ckf kLpwβ (tn ,tn+1 ;X1 ) + Cku(tn )kX1− 1 −β,p.

p

Finally, since u0 = −Au + f it is enough to estimate the first term on the left
with the right-hand side.
 17.2 Maximal Lp -regularity 625

To shorten the notation we write

Zj := Lpwβ (tn , tn+1 ; Xj ), j ∈ {0, 1}.
By the extrapolation of maximal regularity with power weights (Proposition
p
17.2.36), the operator A(tn ) has maximal Lw β
-regularity. Thus, we can define
a new operator Φ : Z1 → Z1 by Φ(v) := y, where y ∈ Z1 is the unique strong
solution to
(
y 0 (t) + A(tn )y(t) = (A(tn ) − A(t))v(t) + f (t), t ∈ (tn , tn+1 )
y(tn ) = u(tn ).
Note that existence and uniqueness of the strong solution follows by maximal
Lpwβ -regularity, Corollary 17.2.37 concerning the related initial value problem,
the fact that the initial value satisfies u(tn ) ∈ X1− p1 −β,p , and an easy time
shift between the interval (tn , tn+1 ) above and the interval (0, tn+1 − tn ) of
the form considered elsewhere in this chapter.
Since y = Φ(v1 ) − Φ(v2 ) is a strong solution of
(
y 0 (t) + A(tn )y(t) = (A(tn ) − A(t))(v1 (t) − v2 (t)), t ∈ (tn , tn+1 )
y(tn ) = 0,
it follows that Φ(v1 ) − Φ(v2 ) = Mtn (1(tn ,tn+1 ) (A(tn ) − A(·))(v1 − v2 )) on
[tn , tn+1 ], and thus
1
kΦ(v1 ) − Φ(v2 )kZ1 6 M
fk(A(tn ) − A(·))(v1 − v2 )kZ 6
0
kv1 − v2 kZ1 .
2
Therefore, the Banach fix point theorem implies that Φ has a unique fix point
y ∈ Z1 . Extending u to [0, tn+1 ] as u|(tn ,tn+1 ] = y, one can check that u turns
into a strong solution on (0, tn+1 ).
Finally, we prove the required a priori estimate. It suffices to estimate u
on [tn , tn+1 ]. By the above we obtain
1
kukZ1 6 kΦ(u) − Φ(0)kZ1 + kΦ(0)kZ1 6 kukZ1 + kΦ(0)kZ1 .
2
Therefore, kukZ1 6 2kΦ(0)kZ1 . To estimate the latter, we note that Φ(0) is
the solution z of the initial value problem
(
z 0 (t) + A(tn )z(t) = f (t), t ∈ (tn , tn+1 ),
z(tn ) = u(tn ).
Thus, we again apply Corollary 17.2.37 concerning such initial value problems,
together with (17.54) on [0, tn ], to obtain
kΦ(0)kZ1 6 Kkf kZ0 + Kku(tn )kX1− 1 −β,p
p

6 Kkf kZ0 + Cn Kkf kLpwα (0,tn ;X0 )

6 (Ktβ−α
n + Cn K)kf kLpwα (0,tn+1 ;X0 ) ,
where of course 00 = 1, as usual. 
626 17 Maximal regularity

17.3 Characterisations of maximal Lp -regularity

So far, we were mostly dealing with situations where maximal Lp -regularity in
some form was assumed, with the goal of deriving various consequences of it.
we now turn to the problem of finding conditions for maximal Lp -regularity in
terms of certain properties of the operator A. We already have seen a necessary
condition for maximal Lp -regularity on R+ in Theorem 17.2.15, namely, that
−A generate a bounded analytic semigroup on X, or equivalently, that A be
sectorial of angle strictly less than 21 π.

17.3.a Fourier multiplier approach

The main theorem of this section, Theorem 17.3.1, provides a necessary and
sufficient condition for maximal Lp -regularity on R+ in the setting of UMD
Banach spaces.

Theorem 17.3.1 (Maximal Lp -regularity and R-sectoriality). Suppose

that A is a linear operator on a Banach space X and p ∈ [1, ∞].
reg
(1) If A has maximal Lp -regularity on R+ with constant Mp,A (R+ ), then A is
R-sectorial with angle ωR (A) < 21 π, and if p ∈ [1, ∞) the following holds:
reg
Rp ({AR(λ, A) : λ ∈ C+ \ {0}}) 6 Mp,A (R+ ).

(2) If X is a UMD space, p ∈ (1, ∞) and A is R-sectorial with angle ωR (A) <
1 p
2 π, then A has maximal L -regularity on R+ with

reg 2
Mp,A (R+ ) 6 400~p,X βp,X (M + 1)2 ,

where M = Rp ({AR(it, A) : t ∈ R \ {0}}).

(3) If X is a UMD space, p ∈ (1, ∞), and −A generates an analytic semigroup
(S(t))t>0 such that the sets {S(t) : t > 0} and {tAS(t) : t > 0} are R-
bounded, then A has maximal Lp -regularity on R+ .

Part (2) of this theorem extends to maximal Lp -regularity on bounded in-

tervals if one considers R-sectoriality of A + λ for λ large enough. Similarly,
part (3) extends to bounded intervals [0, T ] if one considers R-boundedness
of {S(t) : t ∈ (0, T ]} and {tAS(t) : t ∈ (0, T ]}. Both assertions follow by
combining Theorem 17.3.1 with Theorem 17.2.26(1).

Remark 17.3.2. If A 6= 0, then lim inf t→0 kA(it + A)−1 k > 1, and in particular
the constant in Theorem 17.3.1(2) satisfies M > 1. To see this, let x ∈ D(A)
be so that both Ax 6= 0. Thus yt := (it + A)x is non-zero when |t| is small
enough. Now

A(it + A)−1 yt = Ax = yt − itx,

 17.3 Characterisations of maximal Lp -regularity 627

hence kA(it + A)−1 yt k > kyt k − |t|kxk, and therefore

|t|kxk
kA(it + A)−1 k > 1 − .
kyt k

Here kyt k → kAxk =

6 0 as t → 0, and thus the right-hand side tends to 1.
The assumption that A 6= 0 is obviously necessary, for otherwise A(it +
A)−1 = 0.

To prove Theorem 17.3.1 we use the reformulation of maximal Lp -regularity of

Theorem 17.2.19 and the following lemma, where we extend the convolution
operator V introduced in Theorem 17.2.19 to an operator on Lp (R+ ; X) or
Lp (R; X). After that, we can reformulate the boundedness of the convolution
operator in terms of Fourier multipliers.

Lemma 17.3.3. Let −A be a linear operator that generates an analytic semi-

group (S(t))t>0 on a Banach space X. Let p ∈ [1, ∞) and let F (R) =
Cc∞ (R) × X, and let F (R+ ) = {f |R+ : f ∈ F (R)}. The following assertions
are equivalent:
(1) the mapping f → V f , defined for functions f ∈ F (R+ ) by
Z t
V f (t) := A S(t − s)f (s) ds,
0

is well defined, takes values in Lp (R+ ; X), and extends to a bounded op-
erator on Lp (R+ ; X);
(2) the mapping f → V f , defined for functions f ∈ F (R) by
Z ∞
V f (t) := A S(t − s)f (s) ds,
−∞

where we set S(−t) = 0 for t > 0, is well defined, takes values in Lp (R; X),
and extends to a bounded operator on Lp (R; X).
reg
In this case we have kV k = kV k = Mp,A (R+ ). Furthermore, the equivalence
of (1) and (2) and norm identity hold for p = ∞ if we additionally assume
that S be uniformly exponentially stable and set F (I) = L∞ (I; X).

In the above result, it is part of the assumptions that the convolutions take
values in D(A) almost everywhere. If f is sufficiently regular, one can pull
the operator A inside the integral by Hille’s theorem (Theorem 1.2.4). By
Theorem 17.2.19, (1) holds if and only if A has maximal Lp -regularity on R+ .
Proof. (2)⇒(1): This implication hold trivially, and so does the bound kV k 6
kV k.
(1)⇒(2) for p ∈ [1, ∞): Without loss of generality we can assume that
F (I) are the functions in Lp (I; X) with support in a bounded set.
628 17 Maximal regularity

Fix f ∈ F (R) and pick r > 0 so that the support of f is in (−r, ∞). Let
fr : R+ → X be defined by fr (s) := f (s − r). Then for all t > −r we have
Z ∞ Z t+r
V f (t) = AS(t+r−s)f (s−r) ds = AS(t+r−s)fr (s) ds = V fr (t+r).
−∞ 0

whereas V f (t) = 0 for t < −r since the support of f is in (−r, ∞) and S(t) = 0
for t < 0. Hence
Z ∞ Z ∞
p p
kV f kLp (R;X) = kV fr (t + r)k dt = kV fr (t)kp dt
−r 0
= kV fr kpLp (R+ ;X) 6 kV kp kfr kpLp (R+ ;X) = kV kp kf kpLp (R;X) .

It follows that V is bounded on Lp (R; X) and kV k 6 kV k.

(1)⇒(2) for p = ∞, under the additional assumption that (S(t))t>0 is an
uniformly exponentially stable analytic semigroup: By uniform exponential
stability there exist M > 1 and ω > 0 such that ktAS(t)k 6 M e−ωt for all
t > 0 (see Theorems 17.2.15 and G.5.3). It suffices to prove that for any fixed
f ∈ L∞ (R; X), and R 6 −1

reg CeωR/2 
kV f kL∞ (R/2,∞;X) 6 M∞,A (R+ ) + kf kL∞ (R;X) .
−R
Indeed, by letting R → −∞, the desired estimate then follows. We split the
convolution into integrals over (−∞, R) and (R, ∞). On the first interval,
bringing the norms inside, for t ∈ (R/2, ∞) we can estimate
Z R Z R
M
kAS(t − s)f (s)k ds 6 e−ω(t−s) dskf kL∞ (R;X)
−∞ −∞ t − s
−ω(t−R)
Me
6 kf kL∞ (R;X) .
(t − R)ω
2M eωR/2
6 kf kL∞ (R;X) .
−Rω

R ∞the integral over (R, ∞)

For R ∞ let g(s) = f (s + R). Then, by the assumption
that R S(t − s)f (s) ds = 0 S(t − R − s)g(s) ds takes vales in D(A) for
almost all t ∈ (R/2, ∞) and
Z ∞ Z ∞
A S(t − s)f (s) ds = A S(t − R − s)g(s) ds
R 0
reg
6 M∞,A (R+ )kgkL∞ (R+ ;X)
reg
6 M∞,A (R+ )kf kL∞ (R;X) .

The result follows by combining both estimates. 

 17.3 Characterisations of maximal Lp -regularity 629

The space S 0 (Rd ; X) of tempered distributions has been introduced in Section

2.4.d. In the next lemma, we deal with the case of L (X) in place of X.

Lemma 17.3.4. Let A be sectorial of angle < 21 π, and let S be the bounded
analytic semigroup generated by −A. In S 0 (R; L (X)) the following identity
holds:
F (A(1 + A)−1 1R+ S) = A(1 + A)−1 (2πi · +A)−1 .

Somewhat informally the lemma states that F (1R+ S) = (2πi · +A)−1 . The
terms A(1 + A)−1 cannot be left out, however, since AS and (2πi · +A)−1
need not be locally integrable.
Proof. For δ > 0 and t ∈ R let kδ (t) := 1(0,∞) (t)e−δt S(t). The function
kδ belongs to L∞ (R; L (X)), which we view as continuously embedded in
S 0 (R; L (X)). Then, for all Schwartz functions φ ∈ S (R; X),
Z
−1 b
A(1 + A) k0 (φ) = A(1 + A)−1 k0 (t)φ(t)
b dt
R
Z
(i)
= lim A(1 + A)−1 kδ (t)φ(t) b dt
δ↓0 R
= lim A(1 + A)−1 kbδ (φ)
δ↓0
Z
(ii)
= lim A(1 + A)−1 (2πiξ + δ + A)−1 φ(ξ) dξ
δ↓0 R
Z
(iii)
= A(1 + A)−1 (2πiξ + A)−1 φ(ξ) dξ,
R

where (i) and (iii) follow by dominated convergence and (ii) by the definition
of a generator (Definition G.2.1). 
b
We recall from Section 2.4.a that L1 (R; X) denotes the space of functions that
are the inverse Fourier transform of a function in L1 (R; X).

Proposition 17.3.5. Let A be a sectorial operator of angle ω(A) < 21 π, and

let 1 6 p < ∞. The following assertions are equivalent:
(1) A has maximal Lp -regularity on R+ ;
(2) the Fourier multiplier operator
b b b
Tm : L1 (R; X) → L1 (R; X), f 7→ (mfb)

associated with the function

m(ξ) = A(2πiξ + A)−1

extends to a bounded operator on Lp (R; X).

630 17 Maximal regularity
reg
In this situation, we have V = Tm and Mp,A (R+ ) = kTm kL (Lp (R;X)) , where
V is the operator defined in Lemma 17.3.3.

Proof. By the sectoriality assumption, −A generates a bounded analytic semi-

group (S(t))t>0 . Moreover, m belongs to L∞ (R; L (X)) and hence defines an
element of S 0 (R; L (X)). By Theorem 17.2.19 and Lemma 17.3.3, condition
reg
(1) is equivalent to the boundedness of V and we have Mp,A (R+ ) = kV k.
Therefore, to prove the proposition it is enough to show Tm = V . Clearly, it
suffices to show that for all f ∈ Cc∞ (R) ⊗ X

(1 + A)−1 Tm f = (1 + A)−1 V f (17.56)

as elements of S 0 (R; X). Indeed, after applying 1 + A it remains to use the

density of Cc∞ (R) ⊗ X in Lp (R; X).
Fix f ∈ Cc∞ (R)⊗X and consider the function Let t 7→ k(t) = 1(0,∞) (t)S(t)
in L∞ (R; L (X)), viewed as an element of S 0 (R; L (X)). Then, in S 0 (R; X),

(1 + A)−1 V f = A(1 + A)−1 k ∗ f.

Taking Fourier transforms in S 0 (R; X), we obtain

F ((1 + A)−1 V f ) = A(1 + A)−1 F (k ∗ f )

= A(1 + A)−1 b
k fb
(∗)
= (1 + A)−1 mfb
= F (T(1+A)−1 m f ) = F ((1 + A)−1 Tm f ),

where in (∗) we used Lemma 17.3.4. Hence (17.56) follows. 

Now that we have connected the maximal Lp -regularity of an operator A to
the theory of operator-valued Fourier multipliers, we can use the theory of
Chapter 5 to prove Theorem 17.3.1.
Proof of Theorem 17.3.1. (1): First consider p ∈ [1, ∞). If A has maximal Lp -
regularity on R+ , then by Theorem 17.2.15, A is sectorial of angle < 21 π. By
Proposition 17.3.5, the Fourier multiplier Tm , where m(ξ) = A(i2πξ + A)−1 ,
is bounded on Lp (R; X). The Clément–Prüss Theorem 5.3.15 then gives the
R-boundedness of the set {m(ξ) : ξ ∈ R \ {0}}, with bound

Rp ({m(ξ) : R \ {0}}) 6 kTm kL (Lp (R;X)) .

By the observed sectoriality of A, the function λ 7→ A(λ+A)−1 is bounded

and holomorphic in a sector Σσ with some σ ∈ ( 21 π, π), and we just noted
that it is R-bounded on the boundary (minus the origin) of the smaller sector
Σπ/2 = C+ . Hence, by Proposition 8.5.8, this function is R-bounded on C+ \
{0}, with the same R-bound. This in turn implies the R-bound
 17.3 Characterisations of maximal Lp -regularity 631

Rp ({λ(λ + A)−1 : λ ∈ C+ \ {0}}) 6 1 + kTm kL (Lp (R;X)) .

By Proposition 10.3.3 (of which the proof extends to the case of non-dense
D(A)), the R-boundedness of {λ(λ + A)−1 : λ ∈ C+ \ {0}} implies the R-
sectoriality of A with angle ωR (A) < 21 π.
If p = ∞, then by Theorem 17.2.31, A has maximal Lq -regularity on R+
for any q ∈ (1, ∞), and R-sectoriality follows from the previous case.
(2): If A is R-sectorial of angle ωR (A) < 12 π, then the function m(ξ) =
A(i2πξ + A)−1 has R-bounded range with R-bound M as in the assumptions.
Moreover, m ∈ C 1 (R \ {0}; L (X)), and for all ξ ∈ R \ {0} we have

ξm0 (ξ) = −2πiξA(2πiξ + A)−2 = −m(ξ)(I − m(ξ)) = m2 (ξ) − m(ξ).

This function also has R-bounded range, with R-bound at most M 2 + M .

Thus m belongs to Mihlin’s class (Definition 5.3.17) with norm

kmkMp (R;X,Y ) 6 Rp ({m(ξ) : ξ ∈ R \ {0}}) + Rp ({ξm0 (ξ) : ξ ∈ R \ {0}})

6 M + (M 2 + M ) 6 (M + 1)2 .

The boundedness of Tm on Lp (R; X) now follows from the operator-valued

Mihlin Multiplier Theorem 5.3.18, with norm estimate
2 2
kTm kL (Lp (Rd ;X)) 6 400~p,X βp,X kmkMp (R;X,Y ) 6 400~p,X βp,X (M + 1)2 .

Now the required result follows from Proposition 17.3.5.

(3): Again this follows from the extension of Proposition 10.3.3 to the
non-densely defined setting. Explicit bounds can be obtained from that proof.


Corollary 17.3.6. If A has a bounded H ∞ -calculus of angle ωH ∞ (A) < 21 π

on a UMD space X, then A has maximal Lp -regularity on R+ for all 1 < p <
∞.

Proof. Since X is a UMD space it also satisfies the triangular contraction

property by Theorem 7.5.9. Therefore, by Theorem 10.3.4(2), A is R-sectorial
of angle ωR (A) 6 ωH ∞ (A) < 21 π. Applying Theorem 17.3.1 gives the result.

The above results, when combined with various results proved in Volume II,
subsume several classical maximal regularity criteria.
Recall from Theorem 15.3.20 that every standard sectorial (Definition
15.3.1) operator A with a bounded H ∞ -calculus has bounded imaginary pow-
ers and ωBIP (A) = ωH ∞ (A); thus, at least for standard sectorial operators,
the existence of bounded imaginary powers is less restrictive than having a
bounded H ∞ -calculus. Nevertheless, it turns out that bounded imaginary
powers (with a condition on their angle) already suffice to imply maximal
regularity:
632 17 Maximal regularity

Corollary 17.3.7 (Dore–Venni, Prüss–Sohr). If A has bounded imagi-

nary powers with ωBIP < 21 π on a UMD space X, then A has maximal Lp -
regularity on R+ for all 1 < p < ∞.
Proof. By Theorem 15.3.12, such an A is R-sectorial of angle ωR (A) 6
ωBIP (A) < 21 π. Now apply Theorem 17.3.1. 
An alternative approach to this result via sums of operators will be presented
in the next section.
In Theorem 17.2.15, we have seen that maximal Lp -regularity of a linear
operator A implies that −A generates an analytic semigroup. Moreover, in
the case that we have maximal Lp -regularity on the full half-line R+ , this
semigroup will have to be bounded. In Hilbert spaces the following converse
holds:
Corollary 17.3.8 (De Simon). Let A be a linear operator on a Hilbert
space.
(1) If −A generates a bounded analytic C0 -semigroup, then A has maximal
Lp -regularity on R+ for all p ∈ (1, ∞).
(2) If −A generates an analytic C0 -semigroup, then A has maximal Lp -
regularity on (0, T ) for all p ∈ (1, ∞) and T ∈ (0, ∞).
Proof. We give two proofs of (1).
(i): By Proposition G.2.3(4), the generator −A of a C0 -semigroup is a
closed and densely defined operator. For such an operator, Theorem G.5.2
guarantees that generating an analytic semigroup implies sectoriality of angle
< 21 π. In a Hilbert space, every uniformly bounded family of operators is R-
bounded by Theorem 8.1.3(1). Thus the sectoriality and R-sectoriality angles
coincide, and A is even R-sectorial of angle ωR (A) < 21 π. Hence Theorem
17.3.1(2), together with the fact that Hilbert spaces are UMD (Proposition
4.2.14), implies that A has maximal Lp -regularity on R+ for all 1 < p < ∞.
(ii): By Plancherel’s theorem, the operator Tm of Proposition 17.3.5 on
L2 (R; X) satisfies
kTm kL (L2 (Rd ;X)) 6 kmkL∞ (R;L (X)) = sup kA(is + A)−1 k =: M,
s∈R
2 reg
and A has maximal L -regularity on R+ with constant M2,A (R+ ) 6 M by
the same proposition. Therefore, by Theorem 17.2.31, A has maximal Lp -
regularity for all p ∈ (1, ∞).
(2): Since −(λ + A) generates a bounded analytic C0 -semigroup, it has
maximal Lp -regularity on R+ by (1). Therefore, the result follows from the
first and second permanence property of Theorem 17.2.26. 
Corollary 17.3.9. Let −A be the generator of a positive C0 -contraction semi-
group S = (S(t))t>0 on a space Lq (T ), where (T, B, ν) is a measure space and
1 < q < ∞. If S extends to a bounded analytic C0 -semigroup on Lq (T ), then
A has maximal Lp -regularity for all 1 < p < ∞.
 17.3 Characterisations of maximal Lp -regularity 633

Proof. By Theorem 10.7.13, A is R-sectorial of angle ωR (A) < 21 π, and there-

fore the result follows from Theorem 17.3.1. 

Consequences for perturbation of maximal Lp -regularity

Another consequence of Theorem 17.3.1 is an improvement of the perturbation

result of Proposition 17.2.49. For R-sectorial operators A and for ωR (A) <
ω < π and p ∈ [1, ∞), we have defined (see Section 16.2)
Rp
:= Rp AR(z, A) : z ∈ {Σω .


Mω,A

Corollary 17.3.10 (Perturbation). Let A be a linear operator, acting in

a UMD space X, with maximal Lp -regularity on R+ for some p ∈ (1, ∞).
Suppose that B : D(A) → X is a linear operator satisfying

kBxk 6 δkAxk + Kkxk, x ∈ D(A),

p R
where δ ∈ (0, 1) is such that δM 1 π,A (R+ ) < 1, and where K > 0. Then the
2
operator A + B with domain D(A + B) := D(A) is closed, and the following
assertions hold:
(1) for all λ ∈ R large enough, λ + A + B has maximal Lp -regularity on R+ ;
(2) if I = (0, T ) is bounded, then A + B has maximal Lp -regularity on I;
(3) if K = 0, then A + B has maximal Lp -regularity on R+ .
R p reg
By Theorem 17.3.1, we have M 1 π,A (R+ ) 6 Mp,A (R+ ). Therefore, in the case
2
of strict inequality, the above result improves the perturbation result of Propo-
sition 17.2.49 for UMD spaces X.
Proof. (1): By Theorem 17.3.1 it is enough to check that λ + A + B is
R-sectorial of angle < 12 π, and this is immediate from the stability of R-
sectoriality under relatively small perturbations (Theorem 16.2.4).
(2): This follows from (1) by subtracting λ (see Theorem 17.2.26(1)).
(3): This follows from the proof of (1) by taking λ = 0. 

17.3.b The end-point cases p = 1 and p = ∞

In Section 17.3.a we characterised maximal Lp -regularity for p ∈ (1, ∞) in

terms of R-sectoriality of the resolvent along the imaginary axis in case the
underlying Banach space has the UMD property. In the present section we
provide a characterisation of maximal Lp -regularity for the end-point cases
p = 1 and p = ∞.
Theorem 17.3.11 (Kalton–Portal, maximal L1 -regularity). Let I =
(0, T ) or I = R+ . Let −A be the generator of an analytic semigroup (S(t))t>0
on a Banach space X, and suppose that CI,S := supt∈I kS(t)k < ∞. Then the
following are equivalent:
634 17 Maximal regularity

(1) A has maximal L1 -regularity on I;

(2) there exists a constant C > 0 such that
Z
kAS(t)xk dt 6 Ckxk, x ∈ X.
I
−1 reg
In this case CI,S CA 6 M1,A (I) 6 CA , where CA is the infimum of all admis-
sible C in the above estimate. Moreover, for all f ∈ L1 (I; X), the function
s ∈ (0, t) 7→ S(t − s)f (s) ∈ D(A) is Bochner integrable and
Z Z t
kAS(t − s)f (s)k ds dt 6 CA CI,S kf kL1 (I;X) .
I 0
reg
Proof. The implication (1)⇒(2), with estimate CA 6 M1,A (I)CI,S , has al-
ready been established Lemma 17.2.22.
(2)⇒(1): Let f ∈ L1 (I; X) and set T = ∞ if I = R+ . Then
Z Z t Z Z T
kAS(t − s)f (s)k ds dt = kAS(t − s)f (s)k dt ds
I 0 I s
Z Z
6 kAS(t)f (s)k dt ds 6 Ckf kL1 (I;X) .
I I

This proves the convolution estimate kAS ∗ f kL1 (I;X) 6 Ckf kL1 (I;X) , which
completes the proof of (1) and provides the Bochner integrability of s 7→
AS(t − s)f (s) on (0, t). 
Turning to case p = ∞, we point out that we now impose a stronger as-
sumption of the analytic semigroup, involving the prefix ‘C0 -’, compared to
Theorem 17.3.11.
Theorem 17.3.12 (Kalton–Portal, maximal L∞ -regularity). Let I =
(0, T ) or I = R+ . Let −A be the generator of an analytic C0 -semigroup
(S(t))t>0 on a Banach space X, and suppose that supt∈I kS(t)k < ∞. Then
the following assertions are equivalent:
(1) A has maximal L∞ -regularity on I;
(2) A has maximal C-regularity on I;
(3) there exists a constant C > 0 such that
kxk 6 lim sup kS(t)xk + C sup ktAS(t)xk, x ∈ X,
t→T t∈I

where we set T = ∞ if I = R+ .
If these conditions are satisfied, and if we denote
MI,k := sup k(tA)k S(t)k, k = 0, 1, 2,
t∈I
reg
then (3) holds with C = (1 + 4M∞,A (I)MI,0 ), and
reg reg
M∞,A (I) = Mcont,A (I) 6 MI,1 log(2) + CMI,2 .
 17.3 Characterisations of maximal Lp -regularity 635

Remark 17.3.13.
(1) If T < ∞, we have lim supt→T kS(t)xk = kS(T )xk, while if T = ∞, then
for all x ∈ R(A) we have lim supt→∞ kS(t)xk = 0.
(2) The quantities MI,k are finite under the assumptions of Theorem 17.3.12,
although only case k = 0 is explicitly postulated:
For k = 1, let first I = R+ . Then we are assuming that (S(t))t>0 is
a bounded analytic C0 -semigroup, and the finiteness of MI,1 is an ap-
plication of Theorem G.5.3. If I = (0, T ) with T < ∞, then (S(t))t>0
is only assumed to be an analytic C0 -semigroup, but not necessarily
bounded. (The finiteness of MI,0 is automatic for any semigroup.) How-
ever, Sλ (t) = e−λt S(t) will be a bounded analytic C0 -semigroup, if λ > 0
is sufficiently large. Then Theorem G.5.3 applies to show that

Mλ := sup kt(λ + A)Sλ (t)k < ∞,

t>0

and then
 
MI,1 6 sup eλt kt(λ + A)Sλ (t)k + ktλS(t)k 6 eλT Mλ + T λMI,0 < ∞.
t∈I

Finally, for k = 2, we can use the semigroup property to deduce that

MI,2 = sup k(tA)2 S( 21 t)2 k 6 4 sup k( 21 tA)S( 12 t)k2 6 4MI,1

2
t∈I t∈I

Proof of Theorem 17.3.12. The equivalence of (1) and (2) was already ob-
tained in Theorem 17.2.46. Note that MI,k < ∞ for each k > 0. This is clear
if T < ∞ and follows from the uniform boundedness of the semigroup if k > 1.
(3)⇒(2): Let f ∈ Cb (I; D(A)) and fix t ∈ I. Since u := S ∗ f takes values
in D(A), we can use the assumption to obtain

kAu(t)k 6 lim sup kS(r)Au(t)k + C sup krA2 S(r)u(t)k. (17.57)

r→T r∈I

For the first part of (17.57), the bound kAS(σ)k 6 MI,1 /σ gives
Z t
kS(r)Au(t)k 6 kAS(t − s + r)f (s)k ds
0
Z t
1
6 MI,1 dskf kCb (I;X)
0 t−s+r
t

= MI,1 log 1 + kf kCb (I;X) .
r
As r ↑ T , the logarithm tends to log(1 + Tt ), which vanishes if T = ∞. For
the second part of (17.57), the bound kA2 S(σ)k 6 MI,2 /σ 2 gives
Z t
2
krA S(r)u(t)k 6 krA2 S(t − s + r)f (s)k ds
0
636 17 Maximal regularity
Z t
r
6 MI,2 dskf kCb (I;X)
0 (t − s + r)2
t
= MI,2 kf kCb (I;X)
t+r
Therefore, (17.57) becomes
 t

kAu(t)k 6 MI,1 log 1 + + CMI,2 kf kCb (I;X) .
T
Now Lemma 17.2.45 implies that A has maximal C-regularity.
(1)⇒(3): To obtain the desired estimate, it suffices to consider I = (0, T )
with T < ∞. The case I = R+ can then be deduced from the relation of max-
imal regularity on finite intervals and R+ established in Proposition 17.2.18.
Moreover, the generator −A of a C0 -semigroup has a dense domain by Propo-
sition G.2.3(4), and hence it suffices to consider x ∈ D(A).
Proposition 17.2.32 implies that A∗ has maximal L1 -regularity with con-
reg
stant at most M∞,A (I). Therefore, Theorem 17.3.11 gives
Z
reg
kA∗ S(t)∗ x∗ k dt 6 M∞,A (I)MI,1 kx∗ k, x∗ ∈ X. (17.58)
I

Since x ∈ D(A), an integration by parts implies

Z T Z T
2
tA S(t)x dt = −T AS(T )x + AS(t)x dt = −T AS(T )x + S(T )x − x.
0 0
∗ ∗
Therefore, for all x ∈ X ,
Z T
hx, x∗ i = hS(T )x, x∗ i − hT AS(T )x, x∗ i − htAS(t/2)x, A∗ S(t/2)∗ x∗ i dt.
0

By (17.58),
Z T
reg
htAS(t/2)x, A∗ S(t/2)∗ x∗ i dt 6 4M∞,A (I)MI,1 sup ktAS(t)xk kx∗ k,
0 t∈(0,T )

and hence
reg
|hx, x∗ i| 6 kS(T )xk kx∗ k + (1 + 4M∞,A (I)MI,1 ) sup ktAS(t)xk kx∗ k.
t∈(0,T )

This implies the required result. 

17.3.c Sum-of-operators approach

It follows from the discussion in Section 17.2 that a closed operator A on

a Banach space X has maximal Lp -regularity on I if and only if for every
f ∈ Lp (I; X) the inhomogeneous problem
 17.3 Characterisations of maximal Lp -regularity 637
(
u0 (t) + Au(t) = f (t), t ∈ I,
(17.59)
u(0) = 0,

admits a unique Lp -solution u of (17.59) such that Au, u0 ∈ Lp (I; X). Here,
I = (0, T ) or I = R+ , and u0 is the weak derivative of u on I. Let
1,p
(I; X) := u ∈ W 1,p (I; X) : u(0) = 0 ,

0W

recalling that functions in W 1,p (I; X) admit a unique version that is contin-
uous on I.
Let D denote the weak derivative, viewed as a closed operator on X e =
Lp (I; X) with domain
D(D) := 0 W 1,p (IX).
We will also consider D as an operator on Xe = Cb (I; X) with domain

D(D) := u ∈ Cb1 (I; X) : u(0) = 0 .



In Proposition 17.3.16 we investigate various properties of the operator D.

By Ae we denote the closed operator (Ay)(t)
e = A(y(t)) on Lp (I; X) with
p
domain D(A) := L (I; D(A)).
e
Maximal Lp -regularity can be characterised in terms of the operator sum
A + D as follows.
e

Proposition 17.3.14. Let −A be the generator of a C0 -semigroup (S(t))t>0

on a Banach space X, let 1 6 p 6 ∞, and assume that at least one of the
following conditions holds:
• I = (0, T ) is bounded;
• (S(t))t>0 is uniformly exponentially stable.
The following assertions are equivalent:
(1) A has maximal Lp -regularity on I;
(2) Ae + D boundedly invertible;
(3) there is a constant M > 0 such that

kAuk
e p + kDukp 6 M k(A
e + D)ukp , u ∈ D(A)
e ∩ D(D);

(4) A
e + D is closed.

The same equivalences holds if in (1) we consider maximal C-regularity and in

e = Cb (I; D(A)) and D(D) = {u ∈ C 1 (I; X) : u(0) = 0}.
(2)–(4) we take D(A) b

Proof. Let X e := Lp (I; X) when considering maximal Lp -regularity with p ∈

[1, ∞], and let X
e = Cb (I; X) when considering maximal C-regularity.
The equivalence of maximal L∞ -regularity with (3)–(4) will be considered
at the end of the proof. We first deal with maximal Lp -regularity for 1 6 p <
∞ and with maximal C-regularity.
638 17 Maximal regularity

(1)⇒(2): This follows from the definition of maximal Lp (resp. C)-

regularity and Corollary 17.2.9.
(2)⇒(3): This is trivial.
(3)⇒(4): This was already observed in Proposition 16.3.1.
(4)⇒(1): First consider the Lp -setting with p < ∞. The closedness as-
sumption implies that D(A) e ∩ D(D) is a Banach space with respect to the
e set uf (t) = t S(t − s)f (s) ds. Then
R
graph norm of A + D. For f ∈ D(A))
e
0
uf ∈ D(A)e ∩ D(D) and Aue f + Duf = f , and therefore

kuf kD(A)∩D(D)
e
e + D)uf kp = kuf kp + kf kp 6 (M + 1)kf kp ,
= kuf kp + k(A

where M is the norm of the operator f 7→ uf as a bounded operator on X. e

The density of D(A) in X implies that f 7→ uf extends uniquely to a bounded
e e
operator from X e ∩ D(D). Therefore, (1) holds by Theorem 17.2.19.
e into D(A)
In the continuous setting, the above proof can be repeated and instead of
the density argument we use Lemma 17.2.45.
In the L∞ -setting, (4) implies the C-version of (4), and the C-version of
(1) holds. Therefore, A has maximal L∞ -regularity by Theorem 17.2.46. 
Next we present a version of Proposition 17.3.14 in the case where I = R+ and
−A generates a bounded C0 -semigroup, but without the condition 0 ∈ %(A).
For I = R+ the equivalence with (2) no longer holds (see Corollary 17.2.25).
Also, the equivalence with (4) does not hold. Indeed, in Section 17.4.c we will
see that there exist bounded operators A on Lq such that −A generates a
bounded analytic C0 -semigroup, but A does not have maximal Lp -regularity
on R+ unless q = 2. The boundedness of A clearly implies that A e + D is
closed.
In the proposition we also allow I = (0, T ), in which case some information
on the uniformity of the constants with respect to T can be deduced.

Proposition 17.3.15. Let −A be the generator of a C0 -semigroup on a Ba-

nach space X, let 1 6 p 6 ∞, and I = (0, T ) or I = R+ , and suppose that
MS (I) := supt∈I kS(t)k < ∞. The following assertions are equivalent:
(1) A has maximal Lp -regularity on I;
(2) there is a constant M > 0 such that

e p + kDukp 6 M k(A
kAuk e + D)ukp , u ∈ D(A)
e ∩ D(D);

Moreover, letting M0 (I) denote the least admissible constant in (2), one has
reg
M0 (I) 6 2Mp,A (I) + 1,
reg reg
Mp,A (0, T ) 6 M0 (0, T ), and Mp,A (R+ ) 6 M0 (R+ ) + 2MS (I).
 17.3 Characterisations of maximal Lp -regularity 639

The same equivalence and estimates hold if in (1) we consider maximal C-

regularity and in (2) we take D(A)e = Cb ([0, ∞); D(A)) and D(D) = {u ∈
1
Cb ([0, ∞); X) : u(0) = 0}.
Furthermore, in the above situation A
e + D is closed and injective.

Proof. (1)⇒(2): For every f ∈ Lp (R+ ; X) there exists a unique Lp -solution

(resp C-solution) uf satisfying kAue f kp 6 M reg (I)kf kp , where M reg (I) is the
p,A p,A
constant of Definition 17.2.4. Then kDuf kp 6 k(A e + D)uf kp + kAu e f kp 6
reg
(1 + Mp,A (I))kf kp , and therefore

e f kp + kDuf kp 6 (1 + 2M )kf kp = (1 + 2M reg (I))k(A

kAu e + D)uf kp .
p,A

On the other hand, given u ∈ D(A) e ∩ D(D) we may take f := (A e + D)u. Then
uf = u, and substituting this in the above inequality gives the desired result.
(2)⇒(1): First consider the Lp -setting with 1 6 p < ∞. For f ∈ D(A))
e set
Rt
uf (t) = 0 S(t − s)f (s) ds. Note that Auf + Duf = f . In the case I = (0, T ),
one has uf ∈ D(A)e ∩ D(D) and (2) implies

e f kp 6 M k(A
kAuf kp = kAu e + D)uf kp = kf kp ,

and conclude the required result by density of D(A))e in Lp (0, T ; X). In the
p
case I = R+ it need not be true that uf ∈ L (R+ ; X), and thus we cannot
apply the estimate of (2). In order to get around this problem, fix T > 0 and
define vT : [0, ∞) → X by vT := φT uf , where φT : [0, ∞) → [0, 1] is the
piecewise linear function satisfying connecting the points (0, 1), (T, 1), and
(2T, 0), and which is zero on the interval (2T, ∞).
We have vT ∈ D(A) e ∩ D(D) and

e + D)uf + φT0 uf ,
e + D)vT = φT (A
(A

and therefore

kAuf kLp (0,T ;X) = kAvT kLp (0,T ;X)

6 kAvT kLp (R+ ;X)
6 M k(A
e + D)vT kLp (0,∞;X)
6 M kf kLp (R+ ;X) + M T −1 kuf kLp (T,2T ;X) .

Writing MS := supt>0 kS(t)k, Hölder’s inequality gives

2T
kuf kLp (T,2T ;X)
Z 1/p
MS 
6 tp−1 dt kf kLp (R+ ;X) = MS cp kf kLp (R+ ;X) ,
T T T

2p −1
where cpp = p 6 2. Combining the estimates and letting T → ∞, we obtain
640 17 Maximal regularity

kAuf kLp (R+ ;X) 6 M kf kLp (R+ ;X) + MS cp kf kLp (R+ ;X) .

The density of D(A) e in Lp (R+ ; X) implies that f 7→ Auf extends uniquely to

a bounded operator from Lp (R+ ; X) into D(A) e ∩ D(D). Therefore, (1) holds
by Theorem 17.2.19.
In the continuous setting, the above proof can be repeated with a C 1 -
adjustment of the cut-off function φT . Instead of the density argument we
can use Lemma 17.2.45.
In the L∞ -setting, if (2) holds, then the estimate holds for all u ∈
Cb (I; D(A)) ∩ Cb1 (I; X). Therefore, A has maximal C-regularity on I by the
preceding argument. Now Theorem 17.2.46 implies that A has maximal L∞ -
regularity on I.
The final assertion on closedness has already been observed in Proposition
16.3.1. To prove injectivity, suppose that (A+D)u
e = 0. Now (2) gives Du = 0,
so u is constant. Since u(0) = 0, we see that u = 0. 

The operator sum A e + D of Propositions 17.3.14 and 17.3.15 actually falls in

the setting of Theorem 16.3.2. In the case I = (0, T ), Theorem 17.2.15 implies
that there exists a λ ∈ R such that A + λ (and thus A e + λ) is sectorial of angle
< π/2. Moreover, D − λ is sectorial of angle 6 π/2 (see Proposition 17.3.16
below). Therefore, A e + D = (A e + λ) + (D − λ) is as in Theorem 16.3.2. For
I = R+ , the assumptions in Proposition 17.3.15 imply that A is sectorial of
angle < π/2. Hence also in this case the sum falls in the setting of Theorem
16.3.2.
Theorem 16.3.6 will now be applied to derive sufficient conditions for max-
imal Lp -regularity in terms of the operator A and the space X (see Theorem
17.3.19). This will lead to a different proof of Theorem 17.3.1(2). We first col-
lect some properties of the operator D on Lp (I; X) introduced before Propo-
sition 17.3.14. We will only state the next proposition for p ∈ [1, ∞), since D
is not densely defined if p = ∞.
Proposition 17.3.16. Let p ∈ [1, ∞) and let X be a Banach space. Consider
the derivative operator D on Lp (0, T ; X) subject to the one-sided Dirichlet
boundary condition u(0) = 0, as introduced before Proposition 17.3.14. Then
the following hold:
(1) σ(D) = ∅;
(2) for all λ ∈ C the operator D + λ is densely defined, sectorial of angle
6 π/2, and invertible on Lp (0, T ; X);
(3) if X is a UMD space and p ∈ (1, ∞), then D has a bounded H ∞ -calculus
on Lp (0, T ; X) of angle ωH ∞ (D) 6 π/2, and for all σ ∈ (π/2, π) and
f ∈ H 1 (Σσ ) ∩ H ∞ (Σσ ) we have
kf (D)k 6 Cσ Cp,X kf kH ∞ (Σσ ) ;
(4) if X is a UMD space and p ∈ (1, ∞), then for all λ ∈ R the operator D + λ
has a bounded H ∞ -calculus on Lp (0, T ; X) of angle ωH ∞ (D + λ) 6 π/2.
 17.3 Characterisations of maximal Lp -regularity 641

In (3) and (4), the angles are in fact equal to π/2. This will not be needed
later on.
Proof. We begin by observing that the operator −D generates the C0 -
semigroup S of right-shifts on Lp (0, T ; X),

f (s − t), if s − t ∈ I;
S(t)f (s) =
0, if s − t ∈
/ I.

This semigroup is the tensor extension of the positive contraction semigroup

corresponding to the scalar-valued case.
(1): The semigroup S is nilpotent in the sense that S(t) = 0 for all t > T .
In particular, for all ω ∈ R there exists a constant M > 1 such that kS(t)k 6
M eωt for all t > 0. Hence by Proposition G.4.1, for all ω ∈ R the half-plane
{<λ > ω} is contained in the resolvent set of −D.
(2): It is clear that D is densely defined. We claim that D is sectorial of
angle 6 π/2. Indeed, since S is nilpotent we have σ(D) = ∅, and for fixed
λ ∈ C and all <(µ) > 0 we have
T
max{e−<(λ) , 1}
Z
kR(µ, −λ − D)f kp 6 e−(<(µ+λ)t kf kp dt 6 kf kp .
0 <(µ)

This implies the asserted sectoriality. The surjectivity of λ + D follows from

the fact that every λ ∈ C belongs to the resolvent set %(D) by part (1).
(3): This is immediate from Theorem 10.7.12, with constant Cp,X =
2
βp,X ~p,X and Cσ = 200 cos1 σ .
(4): For λ > 0 this follows from Proposition 16.2.6(1) applied to the op-
erator D. For λ < 0, this follows from Proposition 16.2.6(2) applied to the
sectorial operator D + λ − 1. 
For I = R+ , arguing as in the above proof one obtains the following.
Proposition 17.3.17. Let p ∈ [1, ∞) and let X be a Banach space. Consider
the derivative operator D on Lp (R+ ; X) subject to the one-sided Dirichlet
boundary condition u(0) = 0, as introduced before Proposition 17.3.14. Then
the following hold:
(1) the operator D + λ is densely defined, sectorial of angle 6 π/2 on
Lp (R+ ; X);
(2) if X is a UMD space and p ∈ (1, ∞), then D has a bounded H ∞ -calculus
on Lp (R+; X) of angle ωH ∞ (D) 6 π/2, and for all σ ∈ (π/2, π) and
f ∈ H 1 (Σσ ) ∩ H ∞ (Σσ ) we have

kf (D)k 6 Cσ Cp,X kf kH ∞ (Σσ ) ;

In the following we give a different proof of Theorem 17.3.1(2) based on H ∞ -

calculus techniques.
642 17 Maximal regularity

Theorem 17.3.18 (Maximal Lp -regularity through the H ∞ -calculus).

Suppose that A is a linear operator on a UMD space X and p ∈ (1, ∞).
Suppose at least one of the following conditions holds:
(1) I = (0, T ) and there exists a λ ∈ R such that A + λ is R-sectorial of angle
ωR (A + λ) < π/2;
(2) I = R+ , A is R-sectorial of angle ωR (A + λ) < π/2.
Then A has maximal Lp -regularity on I.

Proof. It is straightforward to check that A and D are resolvent commuting.

Moreover, since X is reflexive (see Theorem 4.3.3), it follows from Proposition
10.1.9 that D(A) is dense in X.
(1): By the assumption we can find a (possibly different) λ ∈ R such that
A + λ (and thus A e + λ) is invertible and standard R-sectorial. By Proposi-
tion 17.3.16, D has a bounded H ∞ -calculus and is standard sectorial. From
Theorem 16.3.6, applied with A e + λ and D, we see that

e p + kDukp 6 Cλ,T k(λ + A

k(λ + A)uk e ∩ D(D).
e + D)ukp , u ∈ D(A)

Therefore, Proposition 17.3.14 implies that λ + A has maximal Lp -regularity

on (0, T ), and by Theorem 17.2.26(1) A has maximal Lp -regularity as well.
(2): By Proposition 17.3.17(1) and (2), D is densely defined and sectorial
and has a bounded H ∞ -calculus on Lp (R+ ; X). Therefore, Theorem 16.3.6
implies
e + D)ukp , u ∈ D(A)
e p + kDukp 6 Ck(A
kAuk e ∩ D(D).

Therefore, the result follows from Proposition 17.3.15. 

The following theorem is valid for arbitrary Banach spaces X and exponents
1 6 p 6 ∞. We let I = (0, T ) or I = R+ , and on Lp (I; X) we define the
operator Ae by
(Au)(t)
e := Au(t), t ∈ I,
e := Lp (I; D(A)).
with domain D(A)

Theorem 17.3.19 (Maximal Lp -regularity through the absolute H ∞ -

calculus). Suppose that A is a densely defined linear operator on a Banach
space X, let 1 6 p < ∞, and suppose that at least one of the following
conditions holds:
(1) I = (0, T ) and there exists a λ ∈ C such that A e + λ has an absolute
H ∞ (Σσ )-calculus on Lp (I; X) with σ < π/2;
(2) I = R+ and A e has an absolute H ∞ (Σσ )-calculus on Lp (I; X) with σ <
π/2.
Then A has maximal Lq -regularity on I for all q ∈ [1, ∞) ∪ {p}.
 17.3 Characterisations of maximal Lp -regularity 643

Proof. (1): We use a variation of the argument in Theorem 17.3.18(1). By the

assumption we can find a (possibly different) λ ∈ C such that A e + λ has an
∞
operator-valued H (Σσ )-functional calculus with σ < π/2, is invertible, and
since p < ∞ it is standard sectorial. By Proposition 17.3.16, D is sectorial of
angle 6 π/2. From Theorem 16.3.14 we see that

e p + kDukp 6 Cλ,T k(λ + A

k(λ + A)uk e + D)ukp , u ∈ D(A)
e ∩ D(D).

Therefore, Proposition 17.3.14 implies that λ+A has maximal Lq -regularity on

(0, T ), and thus Theorem 17.2.26(1) implies that A has maximal Lq -regularity,
for all q ∈ [1, ∞) ∪ {p}.
(2): One can repeat the proof of Theorem 17.3.18(2) with a similar varia-
tion as in the above proof. 
As a consequence of Theorem 16.3.20 we obtain the following result.

Corollary 17.3.20 (Da Prato–Grisvard for 1 6 p < ∞). Let −A generate

an analytic C0 -semigroup (S(t))t>0 on a Banach space X, and suppose that
at least one of the following conditions holds:
(1) I = (0, T );
(2) I = R+ and S is uniformly exponentially stable.
Then for all 1 6 p < ∞ and 0 < θ < 1, the part Aθ,p of A in the real
interpolation space (X, D(A))θ,p has maximal Lq -regularity for all q ∈ (1, ∞)∪
{p}.

Proof. By the extrapolation result of Theorem 17.2.26(5), it suffices to con-

sider q = p below.
(1): We first prove the result for 1 6 p < ∞; the case p = ∞ is considered
at the end of the proof. By the Stein–Weis theorem (Theorem 14.3.4),

(X, e θ,p = (Lp (I; X), Lp (I; D(A)))θ,p = Lp (I; (X, D(A))θ,p ).
e D(A))

Therefore, Theorem 16.3.20 implies that for λ ∈ R large enough λ + A eθ,p has
p
an absolute functional calculus on L (I; (X, D(A))θ,p ) of angle < π/2. Since
Proposition 17.3.16 gives that D is sectorial of angle π2 , we can apply Theorem
16.3.14 and Proposition 17.3.14 to conclude the required result.
(2): For 1 6 p < ∞ the result follows from (1) and Theorem 17.2.24. Now
let p = ∞. We use the equivalent norm of Theorem L.2.4 for (X, D(A))θ,∞ .
For all t > 0,
Z t
kAS ∗ f (t)kX 6 kAS(t − s)f (s)kX ds
0
Z t
.A,θ k(t − s)θ−1 dskf kL∞ (0,T ;(X,D(A))θ,∞ )
0
644 17 Maximal regularity

6 Cθ kf kL∞ (0,T ;(X,D(A))θ,∞ ) .

Moreover,
kAS ∗ f (t)k(X,D(A))θ,∞
hθ,A sup r1−θ kAS(r)AS ∗ f (t)k
r>0

6 r1−θ kAS(r)AS ∗ f (t)k

Z t
6 sup r1−θ kAS((t + r − s)/2)k kAS((t + r − s)/2)f (s)k ds
r>0 0
Z t
. sup r1−θ (t + r − s)θ−2 dskf kL∞ (0,T ;(X,D(A))θ,∞ )
r>0 0
. Cθ kf kL∞ (0,T ;(X,D(A))θ,∞ ) .
Taking the supremum over t > 0, we infer that Aθ,∞ has maximal L∞ -
regularity.
(1) for p = ∞: By Theorem 17.2.26(1), the result follows from the case
p = ∞ of (2). 
Remark 17.3.21. In the proof of the case p = ∞ we did not use that (S(t))t>0
is strongly continuous. Also for p < ∞ this assumption can be avoided by
using a direct proof (see the notes).
As a consequence of the case p = ∞, by a density argument we obtain a
similar result for maximal C-regularity using the spaces C(I; (X, D(A))0θ,∞ ),
where (X, D(A))0θ,∞ is defined as the closure of D(A) in (X, D(A))θ,∞ .

17.3.d Maximal Lp -regularity on the real line

Up to this point we have been concerned with maximal regularity on the
half-line and on bounded intervals. In the present section we discuss maximal
Lp -regularity on the real line. We have already briefly touched upon it in
Section 17.3.a, which in hindsight states that, if −A generates an analytic
semigroup, then A has maximal Lp -regularity on the half-line if and only if it
has maximal Lp -regularity on the real line.
It is worth to considering the case of the real line separately for two reasons:
(1) there are operators A which have maximal Lp -regularity on R for which
−A does not generate a semigroup (see below Definition 17.3.23);
(2) in certain situations, it is possible to derive maximal regularity on the
half-line by checking maximal Lp -regularity on the real line (see Theorem
17.3.32).
For a given function f ∈ L1loc (R; X), consider the problem
u0 (t) + Au(t) = f (t), t ∈ R. (17.60)
Notice that there is no initial value as time runs over all of R.
 17.3 Characterisations of maximal Lp -regularity 645

Definition 17.3.22 (Strong solutions and Lp -solutions). A strongly mea-

surable function u : R → X is called a strong solution of (17.60) associated
1
with a given f ∈ Lloc (R; X) if
(i) u takes values in D(A) almost everywhere and Au belongs to L1loc (R; X);
(ii) for almost all s, t ∈ R we have
Z t Z t
u(t) − u(s) + Au(r) dr = f (r) dr.
s s

A strong solution u associated with a function f ∈ Lp (R; X) is called an

Lp -solution if Au ∈ Lp (R; X).

As before one shows that if u : R → X is a strong solution of (17.60), then u

is differentiable almost everywhere, u takes values in D(A) almost everywhere,
both u0 and Au belong to L1loc (R; X), and (17.60) holds for almost all t ∈ R.
Moreover, strong solutions have a continuous version. If u is an Lp -solution
of (17.60), then for all real numbers a < b we have
0
ku(b) − u(a)k 6 (b − a)1/p (kAukLp (a,b;X) + kf kLp (a,b;X) )

as in (17.5), and as in (17.6) it follows that u0 ∈ Lp (R; X) and

ku0 kLp (R;X) 6 kAukLp (R;X) + kf kLp (R;X) .

The definition of maximal Lp -regularity on R is as in Definition 17.2.4,

except that now we take I = R. However, to avoid technical problems involving
uniqueness and approximation we furthermore ask for an a priori estimate in
Lp (R; D(A)) instead of just the norm of Au.

Definition 17.3.23 (Maximal Lp -regularity). A linear operator A has

maximal Lp -regularity on R if for every f ∈ Lp (R; X) there exists a unique
Lp -solution u of (17.60) on R, and there is a constant C > 0 independent of
f such that

kukLp (R;D(A)) 6 Ckf kLp (R;X) . (17.61)

The least admissible constant C will be called the maximal Lp -regularity con-
reg
stant and will be denoted by Mp,A (R).
reg
Unlike in the definition of Mp,A (I) for bounded intervals I = (0, T ) or R+ , an
a priori bound on kukLp (R;X) is also included in the estimate (17.61) through
the use of the graph norm. Also, unlike in the case I = (0, T ) or I = R+ , in the
above definition the uniqueness is assumed a priori. In fact, when uniqueness
holds, then A is injective. Indeed, suppose that Ax = 0. Then, setting u = x,
one has u0 + Au = 0, and therefore u = 0 by uniqueness. Conversely, if A is
injective, then uniqueness is immediate from the estimate (17.61).
646 17 Maximal regularity

Example 17.3.24. The operator A = 0 does not have maximal Lp -regularity

on R, but of course it has maximal Lp -regularity on R+ .
The analogue of Lemma 17.2.16 for the real line does not hold: it is not
true that maximal Lp -regularity on R implies maximal Lp -regularity on all
subintervals of R. This can be seen as follows. By reflection, the following
observation is immediate:
Proposition 17.3.25. Let 1 6 p 6 ∞ be fixed. A linear operator A has
maximal Lp -regularity on R if and only if −A has maximal Lp -regularity on
reg reg
R, and in this case we have Mp,A (R) = Mp,−A (R).
Now, if maximal Lp -regularity on R would always imply maximal Lp -regularity
on R+ , we could apply this fact to both A and −A and obtain from Theorem
17.2.15 that A generates an analytic group. It is easy to see that this im-
plies that A is bounded. Indeed, denoting the group by (S(t))t∈R , the identity
A = S(−1)(AS(1)) exhibits A as the product of two bounded operators (the
latter being bounded by Theorem G.5.3).
Example 17.3.26. Let A be an unbounded linear operator on X with maximal
Lp -regularity on R. Then the operator on X ×X given by A(x, y) = (Ax, −Ay)
has maximal Lp -regularity on R. If A had maximal Lp -regularity on (0, T ),
then A would generate an analytic semigroup by Theorem 17.2.15, and thus
both A and −A generate analytic semigroups. This contradicts the unbound-
edness of A.
Remark 17.3.27. Examples 17.3.24 and 17.3.26 show that the two properties
of maximal Lp -regularity on R and on R+ are incomparable in both directions;
neither implies the other one.
Remark 17.3.28. The reader may check that the following results extend to
maximal Lp -regularity on R:
(1) maximal Lp -regularity of A implies closedness of A (Proposition 17.2.5);
(2) it is enough to verify maximal Lp -regularity on a dense subspace (Propo-
sition 17.2.10);
(3) under closedness of A and unique solvability for all f , maximal Lp -
regularity follows (Proposition 17.2.11).
Indeed, by using the fact that maximal Lp -regularity on R is formulated with
the more restrictive estimate (17.61), it is straightforward to extend the pre-
vious proofs.

Maximal regularity on R and bisectoriality

In the next paragraph we will discuss the relationship between maximal Lp -

regularity on R+ and R. Prerequisite to this discussion is the following result.
It is a version for the real line of Dore’s Theorem 17.2.15), which gave the
sectoriality of operators with maximal Lp -regularity on R+ .
 17.3 Characterisations of maximal Lp -regularity 647

Theorem 17.3.29 (Mielke). Let X be a Banach space and let 1 6 p 6 ∞.

Let λ ∈ C and set λ0 = |λ| if λ 6= 0 and λ0 = 1 if λ = 0. Let A be a linear
operator in X such that λ + A has maximal Lp -regularity on R, and suppose
that there exists a constant C such that for all f ∈ Lp (R; X) the Lp -solution
u of

u0 + (λ + A)u = f (17.62)

satisfies

λ0 kukX + kAukX 6 Ckf kLp (R;X) . (17.63)

Lp (R)

Then λ + A is bisectorial, 0 ∈ %(λ + A), and for all ξ ∈ R we have

λ0 k(λ + iξ + A)−1 k + kA(λ + iξ + A)−1 k 6 C.

In the special case λ = 0, the left-hand side of (17.63) coincides with

reg
kukLp (R;D(A)) and therefore (17.63) holds with C = Mp,A (R).
To prove the theorem, we first derive a maximal regularity result with
exponential weights, related to Proposition 17.2.38. Given a non-negative lo-
cally integrable function w on R, upon replacing the Lp -spaces in Definition
p
17.3.23 by the spaces Lw of function u such that wu ∈ Lp (not to be confused
with L (w), where the condition is w1/p u ∈ Lp ), we obtain the definition of
p
p
maximal Lw -regularity on R. Propositions 17.2.5, 17.2.10, and 17.2.11 extend
to this setting.
Proposition 17.3.30. Suppose that the conditions of Theorem 17.3.29 hold,
and let α ∈ R satisfy |α| < λ0 /C. Then the operator λ+A has maximal Lpe−α|·| -
regularity on R. Moreover, for all f : R → X such that e−α|·| f ∈ Lp (R) the
Lep−α|·| -solution u to (17.62) satisfies

C
λ0 e−α|·| kukX + e−α|·| kAukX 6 ke−α|·| f kLp (R;X) .
Lp (R) 1 − |α|λ−1
0 C

Proof. As in the proof of extrapolation with exponential weights on R+

(Proposition 17.2.38), one checks that if e−α|·| f ∈ Lp (R; X), then a locally
integrable function u : R → X is an Lpe−α|·| -solution to (17.60) if and only if
v = e−α|·| u is an Lp -solution to

v 0 + Av + α sgn(·)v = fα ,

where fα = e−α|·| f . Therefore it will be enough to analyse the equation for v,

which will be done by a fixed point argument.
Let Y := Lp (R; D(A)) with norm

kvkY = λ0 kvkX + kAvkX . (17.64)

Lp (R)
648 17 Maximal regularity

For g ∈ Y let T (g) := v, where v is the unique Lp -solution to

v 0 + (λ + A)v = fα − α sgn(·)g.

The assumptions imply that T maps Y into itself and

|α|C
kT (g1 ) − T (g2 )kY 6 |α|Ckg1 − g2 kLp (R;X) 6 kg1 − g2 kY , (17.65)
λ0
where the constant on the right is < 1 by our assumptions. Therefore, T has
a unique fixed point v ∈ Y , and this is our required Lp -solution. Moreover,
by (17.65) and the assumed estimate of Theorem 17.3.29,

|α|C
kvkY 6 kT (v) − T (0)kY + kT (0)kY 6 kvkY + Ckfα kLp (R;X) .
λ0
C
It follows that kvkY 6 Kkfα kLp (R;X) with K = 1−|α|Cλ−1
, and therefore
0

λ0 e−α|·| kukX + e−α|·| kAukX = kvkY 6 Kkfα kLp (R;X)

Lp (R)

= Kke−α|·| f kLp (R;X) ,

as claimed. 
Proof of Theorem 17.3.29. Fix α ∈ (0, λ0 /C), and let Y = Lp (R; D(A)) be
normed as in (17.64).
Let x ∈ X, s ∈ R, and ξ ∈ R be fixed, and for t ∈ R set fs (t) := eiξ(t+s) x.
Then e−α|·| fs ∈ Lp (R; X), and by Proposition 17.3.30 there exists a unique
Lpe−α|·| -solution us to
u0s + (λ + A)us = fs ,
and we have
C
ke−α|·| us kY 6 ke−α|·| fs kLp (R;X) . (17.66)
1 − αCλ−1
0

Since fs (t) = f0 (s+t) = eiξs f0 (t), it follows that us (t) = u0 (s+t) = eiξs u0 (t).
Therefore u0 (s) = eiξs y with y := u0 (0). Since e−α|·| u0 ∈ Y , we must have
y ∈ D(A), and consequently

eiξs (iξ + λ + A)y = u00 (s) + (λ + A)u0 (s) = f0 (s) = eiξs x.

This shows that (λ + iξ + A)y = x. Moreover, by (17.66) with s = 0, upon

letting α ↓ 0 we find that λ0 kyk + kAyk 6 Ckxk.
It remains to prove that λ + iξ + A is injective. Let y ∈ D(A) be such that
(λ + iξ + A)y = 0. Then u(t) := eitξ y satisfies u0 + (λ + A)u = 0, and therefore
u = 0 by uniqueness of the Lpe−α|·| -solution. 
 17.3 Characterisations of maximal Lp -regularity 649

Connections between maximal regularity on R and R+

The following two results connect maximal Lp -regularity on R+ and R under

suitable additional assumptions on A. This provides a way to check maximal
Lp -regularity on R+ via the case R; in fact, this is one of our main motivations
to consider the case R.
When A has maximal Lp -regularity on R we define the operator MR :
p
L (R; X) → Lp (R; D(A)) by

MR f := uf , (17.67)

where uf is the unique Lp -solution of (17.60). The norm of this operator is

reg
given by kMR k = Mp,A (R).
Proposition 17.3.31 (Maximal regularity: R+ versus R – I). Let p ∈
[1, ∞]. For a linear operator A, the following assertions are equivalent:
(1) A has maximal Lp -regularity on R+ and 0 ∈ %(A);
(2) A has maximal Lp -regularity on R and −A generates a bounded analytic
semigroup.
In this case, one has
reg
Mp,A (R+ ) = kV kL (Lp (R;X)) = kAMR kL (Lp (R;X)) ,

and the Lp -solution u to (17.60) satisfies Au = V f = AMR f , where V is as

in Lemma 17.3.3.
Proof. We start with some general observations. Let (S(t))t>0 be a bounded
analytic semigroup with generator −A, and suppose that 0 ∈ %(A). By Propo-
sition K.2.3, (S(t))t>0 is a uniformly exponentially stable analytic semigroup.
Define the operator T on Lp (R; X) by
Z t
T f := S(t − s)f (s) ds.
−∞

This operator is well defined and bounded by Young’s inequality. For f ∈

Lp (R; X) let uf := T f . As in the proof of Proposition 17.1.3 one can check
that for all s < t one has
Z t  Z t 
A(1 + A)−1 uf (r) dr = −(1 + A)−1 uf (t) − uf (s) + f (r) dr .
s s
(17.68)

Therefore, (1 + A)−1 uf is an Lp -solution to (17.60) with right-hand side (1 +

A)−1 f .
(1)⇒(2): From Theorem 17.2.15 it follows that −A generates a bounded
analytic semigroup (S(t))t>0 and, since 0 ∈ %(A), it is uniformly exponen-
tially stable by Proposition K.2.3. Moreover, by Theorem 17.2.19 and Lemma
650 17 Maximal regularity

17.3.3, V extends to a bounded operator on Lp (R; X) and kV f kLp (R;X) 6

reg
Mp,A (R+ )kf kLp (R;X) for all f ∈ Lp (R; X). Let f ∈ Lp (R; X) and set
reg
uf := A−1 V f = T f . Clearly, kAuf kLp (R;X) 6 Mp,A (R+ )kf kLp (R;X) by the
boundedness of V . It remains to check that uf is an Lp -solution to (17.60).
From (17.68) and the fact that Auf ∈ Lp (R; X) we find that
Z t h Z t i
−1 −1
(1 + A) Auf (r) dr = −(1 + A) uf (t) − uf (s) + f (r) dr .
s s

Applying 1 + A to both sides, it follows that uf is an Lp -solution to (17.60).

(2)⇒(1): By Theorem 17.3.29, we have 0 ∈ %(A), and thus (S(t))t>0 is
uniformly exponentially stable by Proposition K.2.3. For f ∈ Lp (R; X), let
uf := T f . By the observation below (17.68) and by uniqueness, we find that

(1 + A)−1 T f = (1 + A)−1 uf = u(1+A)−1 f = (1 + A)−1 uf = (1 + A)−1 MR f,

where MR is as in (17.67). Applying 1 + A on both sides, it follows that

T f = MR f . Therefore, V = AT = AMR are all bounded on Lp (R; X). Now
Theorem 17.2.19 and Lemma 17.3.3 imply that A has maximal Lp -regularity
on R+ . 
In situations where one does not know a priori whether −A generates an
analytic semigroup, the following result can often be used. In particular, this
is useful in proving maximal Lp -regularity in the case that A is a differential
operator and D(A) includes a boundary condition.
Theorem 17.3.32 (Maximal regularity: R+ versus R – part II). Let A
be a linear operator and 1 6 p 6 ∞. Suppose that, for all λ > 0, the operator
λ + A has maximal Lp -regularity on R, and suppose that there exist a constant
C > 0 and an integer n ∈ N such that, for all λ > 0 and all f ∈ Lp (R; X),
the Lp -solution uλ of

u0 + (λ + A)u = f (17.69)

satisfies

λkuλ kX + kAuλ kX 6 C(1 + λ)n kf kLp (R;X) . (17.70)

Lp (R)

Under these assumptions,

(1) the estimate (17.70) holds with n = 0;
(2) −A generates a bounded analytic semigroup;
(3) for all λ > 0, the operator λ + A has maximal Lp -regularity on R+ ;
(4) for all λ > 0 and f ∈ Lp (R+ ; X), there exists a unique Lp -solution vλ to
(
v 0 (t) + (λ + A)v(t) = f (t), t ∈ R+ ,
(17.71)
v(0) = 0;
 17.3 Characterisations of maximal Lp -regularity 651

(5) this vλ satisfies

λkvλ kX + kAvλ kX 6 Ckf kLp (R+ ;X) . (17.72)

Lp (R+ )

reg
In particular, Mp,A (R+ ) 6 C.

Proof. Fix λ > 0. By assumption (17.70) and Mielke’s Theorem 17.3.29, we

have iξ ∈ %(λ + A) for all ξ ∈ R and

λk(λ + iξ + A)−1 k + kA(λ + iξ + A)−1 k 6 C(1 + λ)n , ξ ∈ R. (17.73)

As a consequence also k(λ + iξ)(λ + iξ + A)−1 k 6 C(1 + λ)n + 1. As the right-

hand side remains bounded for λ ↓ 0, it follows by a power series expansion
that C+ \ {0} ⊆ %(−A) and (17.73) extends to all λ + iξ ∈ C+ \ {0}. Fix
z
ε > 0. Setting ζε (z) := z+ε A(z + A)−1 for z ∈ C+ \ {0} and ζε (0) = 0, upon
rewriting (17.73) with z = λ + iξ, we see that

kζε (z)k 6 C(1 + <z)n , z ∈ C+ .

Hence, we can apply the Phragmén–Lindelöf principle to obtain

sup kζε (z)k 6 sup kζε (z)k 6 sup kA(iξ + A)−1 k 6 C,

z∈C+ z∈∂C+ ξ∈R\{0}

where the last estimate follows by taking λ = 0 in (17.73). Letting ε ↓ 0, it

follows that A is sectorial of angle < 21 π, and accordingly −A generates a
bounded analytic semigroup.
By Proposition 17.3.31, λ + A has maximal Lp -regularity on R+ for all
λ > 0. Moreover, this result and the fact that 0 ∈ %(λ + A) also imply that,
for all λ > 0, the solution uλ to (17.69) satisfies
Z t
uλ (t) = e−λ(t−s) S(t − s)f (s) ds,
−∞

for f ∈ L (R; X). Thus f 7→ uλ defines a bounded mapping MλR : Lp (R; X) →

p

Lp (R; D(A)) and by (17.70) we have

kλMλR f kX + kAMλR f kX 6 C(1 + λ)n kf kLp (R;X) .

Lp (R)

Reasoning as after (17.28), for all λ > 0 and ξ ∈ R we have

kMRλ+iξ f kLp (R;X) = kMRλ f kLp (R;X) .

Therefore, in a similar way as before, the Phragmén–Lindelöf principle applied

to the mapping z ∈ λ + C+ 7→ (zMzR f, AMzR f ) gives that, for all z ∈ λ + C+ ,

kzMzR f kX + kAMzR f kX 6 Ckf kLp (R;X) .

Lp (R)
652 17 Maximal regularity

Letting λ ↓ 0, shows that we can take n = 0 in (17.70).

Now fix λ > 0. By Theorem 17.2.19, the solution vλ to (17.71) satisfies
Z t
vλ = e−λ(t−s) S(t − s)f (s) ds
0

whenever f ∈ Lp (R+ ; X). Therefore, we find that vλ = uλ on R+ , where

uλ = MRλ (1R+ f ), and thus

λkvλ kX + kAvλ kX 6 λkuλ kX + kAuλ kX

Lp (R+ ) Lp (R)

6 Ckf kLp (R+ ;X) .

This proves (17.72) for λ > 0. The extension to λ = 0 is a consequence of

Proposition 17.2.29. 
Theorem 17.3.32 admits the following converse.

Proposition 17.3.33 (Maximal regularity: R+ versus R – III). Let X

be a Banach space, 1 6 p 6 ∞, and let A be a linear operator with maximal
Lp -regularity on R+ . Then for all λ > 0, the operator λ + A has maximal
Lp -regularity on R, and for all f ∈ Lp (R; X), the Lp -solution uλ of

u0 + (λ + A)u = f on R (17.74)

satisfies
reg
λkuλ kLp (R;X) + kAuλ kLp (R;X) 6 (M + Mp,A (R+ ))kf kLp (R;X) , (17.75)

where M is such that kS(t)k 6 M for all t > 0, with (S(t))t>0 the bounded
analytic semigroup generated by A.

Proof. Fix λ > 0. By Dore’s Theorem 17.2.15, −A generates a bounded ana-

lytic semigroup S = (S(t))t>0 . Thus one can take ω = 0 in the growth bound
(K.1), and hence λ + A is invertible for λ > 0 by Definition K.1.2 of a genera-
tor. It follows from Theorem 17.2.26(2) that λ + A has maximal Lp -regularity
on R+ . Since we also have 0 ∈ %(λ + A), Proposition 17.3.31 guarantees that
λ + A has maximal Lp -regularity on R as well.
Next we prove the estimate (17.75). Define the operator MRλ : Lp (R; X) →
p
L (R; D(A)) by
MRλ f = uλ ,
where uλ is the unique solution of (17.74). By Proposition 17.3.31,
Z t
MRλ f (t) = e−λ(t−s) S(t − s)f (s) ds, t > 0.
−∞

Therefore, by Young’s inequality

 17.3 Characterisations of maximal Lp -regularity 653

λkMRλ f kLp (R;X) 6 M kf kLp (R;X) .

reg
It remains to prove the bound kAMRλ kL (Lp (R;X)) 6 Mp,A (R+ ). By Theo-
p p
rem 17.2.19, the L -solution vλ to (17.71) with f ∈ L (R+ ; X) satisfies
Z t
vλ (t) = MRλ+ f (t) := e−λ(t−s) S(t − s)f (s) ds, t > 0.
0

Therefore, by Proposition 17.2.27(1)

reg
kAMRλ+ kL (Lp (R+ ;X)) 6 Mp,A (R+ ).

The required bound for AMRλ can be deduced from this by the same argument
as in Lemma 17.3.3. 
From Theorem 17.3.32 and Proposition 17.3.33, it follows that, in order to
establish maximal Lp -regularity of A on R+ , it suffices to prove maximal
Lp -regularity of λ + A on R for all λ > 0, along with the estimate (17.70)
for some n ∈ N (or equivalently, for n = 0). The advantage of considering
problems on R is that one can apply the Fourier transform in the time variable,
and initial value conditions do not play any role. In particular, this leads to
simplifications in studying evolution equation with inhomogeneous boundary
values and initial values. This will be further explained in the Notes at the
end of the chapter.

Fourier multipliers related to maximal Lp -regularity on R

The remainder of this section is devoted to extending some of the Fourier

multiplier connections of maximal Lp -regularity on R+ to the real line. A
variation of Proposition 17.3.5 is needed for this, the proof of which has to be
adapted since the operators V were defined via the bounded analytic semi-
group generated by −A (see Lemma 17.3.3). In the present situation, where
we assume that A has maximal Lp -regularity on R, Mielke’s Theorem 17.3.29
implies that A is bisectorial rather than sectorial, as was the case with Dore’s
Theorem 17.2.15 on R+ . Below we will only consider the case 0 ∈ %(A) (which
is a necessary assumption by Theorem 17.3.29). The multiplier below will be
(2πiξ + A)−1 , instead of the multiplier A(2πiξ + A)−1 used in Proposition
17.3.5.

Proposition 17.3.34. Let A be a bisectorial operator in X with 0 ∈ %(A),

and let 1 6 p < ∞. The following assertions are equivalent:
(1) A has maximal Lp -regularity on R;
(2) the Fourier multiplier operator
b b b
Tm : L1 (R; X) → L1 (R; D(A)), f 7→ (mfb)

associated with the function

654 17 Maximal regularity

m(ξ) = (2πiξ + A)−1

extends to a bounded operator from Lp (R; X) to Lp (R; D(A)).

(3) the Fourier multiplier operators
b b b
Tm : L (R; X) → L1 (R; X), f 7→ (mk fb)
k
1

associated with the functions m0 (ξ) := m(ξ) and m1 (ξ) := Am(ξ) extend
to a bounded operators from Lp (R; X) to Lp (R; X).
In this situation, the Lp -solution u to (17.60) satisfies u = Tm f , and we have
reg
max{N0 , N1 } 6 Mp,A (R) = kTm kL (Lp (R;X),Lp (R;D(A)) 6 N0 + N1 ,

where Nk := kTmk kL (Lp (R;X)) for k = 0, 1.

Proof. We start by making some observations that will be needed in the proofs
of both implications. Let f ∈ S (R; X) be fixed. Then the function ξ 7→
(2πiξ+A)−1 fb(ξ) belongs to S (R; D(A)). Thus we may define u ∈ S (R; D(A))
by

b := (2πi · +A)−1 fb.

u (17.76)

This gives the identity Tm f = u. Moreover, since F (u0 ) = (2πi·)b

u, one has

F (u0 + Au) = F (u0 ) + AF (u) = (2πi · +A)b

u = fb,

and thus u0 + Au = f in S (R; X).

Next we show that any Lp -solution u to (17.60) satisfies u = Tm f . Since
u is an Lp -solution we have u, Au, u0 ∈ Lp (R; X) and

u and F (Au) = Ab
F (u0 ) = (2πi·)b u

in S 0 (R; X). It follows that

u = F (u0 + Au) = fb.

((2πi·) + A)b

We conclude that u b = ((2πi·) + A)−1 fb = mfb, and thus u = Tm f .

(1)⇒(2): Let f ∈ S (R; X) be given, and let u be the Lp -solution to
(17.60). By the preliminary observations we have u = Tm f , and by maximal
Lp -regularity we have the estimate
reg
kTm f kLp (R;D(A)) = kukLp (R;D(A)) 6 Mp,A (R)kf kLp (R;X) .

Now the bounded extension is obtained by the density of S (R; X) in Lp (R; X).
(2)⇒(1): By the preliminary observations, for any given f ∈ S (R; X) the
function u defined by (17.76) satisfies u0 + Au = f in S (R; X), and u = Tm f
is in S (R; X). Therefore, u is an Lp -solution to (17.60) and it satisfies
 17.3 Characterisations of maximal Lp -regularity 655

kukLp (R;D(A)) = kTm f kLp (R;D(A)) 6 kTm kL (Lp (R;X),Lp (R;D(A))) kf kLp (R;X) .
If v is another Lp -solution to (17.60), these observations also imply that v =
Tm f = u. Therefore, any Lp -solution satisfies the required a priori estimate.
reg
The equality Mp,A (R) = kTm kL (Lp (R;X),Lp (R;D(A))) follows from a com-
bination of the bounds established in the course of proving (1)⇒(2) and
(2)⇒(1).
(2)⇔(3): Since kxkD(A) := kxkX + kAxkX , it follows that
kTm0 f kLp (R;X) = kTm f kLp (R;X) 6 kTm f kLp (R;D(A)) ,
kTm1 f kLp (R;X) = kATm f kLp (R;X) 6 kTm f kLp (R;D(A)) ,
kTm f kLp (R;D(A)) 6 kTm f kLp (R;X) + kATm f kLp (R;X) ,
which readily implies both implications and the related estimates. 
The next theorem presents an extrapolation result for maximal Lp -regularity
on R.
Theorem 17.3.35 (Extrapolation). Let A be a densely defined bisectorial
operator that has maximal Lp0 -regularity on R for some p0 ∈ [1, ∞). Then for
all p ∈ (1, ∞) and Muckenhoupt weights w ∈ Ap , the operator A has maximal
Lp (w)-regularity on R. Moreover, the strong solution of (17.60) satisfies
1
max(1, p−1 )
kukLp (R,w;D(A)) 6 Cp0 ,p (M 2 + Mpreg
0 ,A
(R))[w]Ap kf kLp (R,w;X) ,

where M = supξ∈R\{0} k(iξ + A)−1 kL (X,D(A)) .

Remark 17.3.36. Note that
1 = kIkL (X) = k(iξ + A)(iξ + A)−1 kL (X)
6 |ξ|k(iξ + A)−1 kL (X) + kA(iξ + A)−1 kL (X)
6 (|ξ| + 1)k(iξ + A)−1 kL (X,D(A)) 6 (|ξ| + 1)M ;
hence, with ξ → 0, we find that M > 1.
Proof of Theorem 17.3.35. By Mielke’s Theorem 17.3.29 (applied to λ = 0),
the assumed maximal Lp0 -regularity of A on R implies that 0 ∈ %(A).
Thus Proposition 17.3.34 applies to show that the Fourier multiplier op-
erator Tm associated with the function m(ξ) = (2πiξ + A)−1 , ξ ∈ R, is
bounded from Lp0 (R; X) to Lp (R; D(A)). Clearly, km(ξ)kL (X,D(A)) 6 M
and kξm0 (ξ)kL (X,D(A)) 6 M 2 + M for all ξ ∈ R \ {0}. Therefore, Corol-
lary 13.2.8 implies that Tm is bounded from Lp (R, w; X) to Lp (R, w; D(A))
for all p ∈ (1, ∞) and all w ∈ Ap , with norm estimate
1
max(1, p−1 )
kTm kL (Lp0 (R;X),Lp0 (R;D(A))) 6 cp (M 2 + Mpreg
0 ,A
(R))[w]Ap ,

where we absorbed the lower order terms to M 2 by Remark 17.3.36. By Propo-

sition 17.3.34, Tm f is the Lp -solution to (17.60) for all f ∈ S (R; X). Since
this space is dense in Lp (R, w; X), the result follows. 
656 17 Maximal regularity

Next we present a duality result for maximal regularity on R. Note that unlike
in Proposition 17.2.32 we do not consider the end-points p = 1 and p = ∞.
Proposition 17.3.37 (Duality). Let X be a Banach space, and let 1 < p <
∞. Suppose that A is a closed and densely defined linear operator in X. Then
A has maximal Lp -regularity on R if and only if its adjoint A∗ has maximal
0
Lp -regularity on R, and in this case
1 reg
M (R) 6 Mpreg reg
0 ,A∗ (R) 6 2Mp,A (R).
2 p,A
Proof. By Proposition 17.3.34, maximal Lp -regularity of A on R is equivalent
to the boundedness of the two multipliers mk (ξ) = Ak (i2πξ + A)−1 (k = 0, 1),
0
on Lp (R; X). Similarly, maximal Lp -regularity of A∗ on R is equivalent to
the boundedness of the two multipliers (A∗ )k (i2πξ + A∗ )−1 (k = 0, 1) on
0
Lp (R; X ∗ ). It is evident that these are the pointwise adjoints mk (ξ)∗ of the
mk (ξ). By Proposition 5.3.7, if m ∈ MLp (R; X, X), then its pointwise adjoint
0
satisfies m∗ ∈ MLp (R; X ∗ , X ∗ ) with

kTm∗ kL (Lp0 (R;X ∗ )) 6 kTm kL (Lp (R;X)) ,

and one can extract from the short proof that the converse implication and
estimate are also true. A combination of these results proves the proposition
at hand, and it is also easy to get the quantitative statement
1
X 1
X
Mpreg
0 ,A∗ (R) 6 kTm∗k kL (Lp0 (R;X ∗ )) = kTmk kL (Lp (R;X))
k=0 k=0
reg
6 2 max kTmk kL (Lp (R;X)) 6 2Mp,A (R).
k=0,1

The converse is proved similarly. 

As in Theorem 17.3.1, we can characterise maximal Lp -regularity on R by
using R-bisectoriality.
Theorem 17.3.38 (Maximal Lp -regularity and R-bisectoriality). Let
A be a densely defined bisectorial operator on a Banach space X with 0 ∈ %(A).
(1) If A has maximal Lp -regularity on R for some p ∈ (1, ∞) with constant
reg
Mp,A (R), then A is R-bisectorial with angle ωR (A) < 21 π, and
reg
M := Rp ({(iξ + A)−1 ∈ L (X, D(A)) : ξ ∈ R \ {0}}) 6 Mp,A (R).

(2) If X is a UMD space, p ∈ (1, ∞), and A is R-bisectorial, then A has

maximal Lp -regularity on R with
reg 2
Mp,A (R) 6 400~p,X βp,X (M + 1)2 .

The proof of Theorem 17.3.38 is completely analogous to Theorem 17.3.1 and

we therefore omit the details.
 17.4 Examples and counterexamples 657

17.4 Examples and counterexamples

In this section we will first show that the Laplace and Poisson operator on
Lq (Rd ; X) have maximal Lp -regularity if and only if X is a UMD space.
Moreover, several end points cases are studied. In the final Subsection 17.4.c
we will construct sectorial operators on Lq -spaces with q ∈ (1, ∞) \ {2}, which
fail maximal Lp -regularity on finite time intervals.

17.4.a The heat semigroup and the Poisson semigroup

By Example 10.1.5, the Laplacian −∆ is sectorial of angle 0 on Lq (Rd ; X),

1 6 q < ∞, and therefore by Theorem 15.2.7 its fractional powers (−∆)α/2
are sectorial of angle 0 for all α > 0. In particular, every operator (−∆)α/2
generates an analytic C0 -semigroup on Lq (Rd ; X). Of special interest are the
cases α = 1 and α = 2, which correspond to the Poisson semigroup and the
heat semigroup, respectively.
The aim of this section is to characterise when (−∆)α/2 has maximal Lp -
regularity on Lq (Rd ; X). We start with exponents q ∈ (1, ∞); the case q = 1
will be considered in Example 17.4.2.

Theorem 17.4.1. Let X be a Banach space and let 1 < p, q < ∞ and α > 0,
let λ > 0, and consider the operator A = λ + (−∆)α/2 as a closed and densely
defined operator on Lq (Rd ; X). Let I = R+ or I = (0, T ) with T ∈ (0, ∞).
The following assertions are equivalent:
(1) A has maximal Lp -regularity on I;
(2) X is a UMD space.
R reg
In this situation we have βp,X 6 2Mp,A (R+ ).

By inspection of the proof, the reader can also track a quantitative estimate
in the other direction. However, the present method is sub-optimal for this
purpose, and we therefore refrain from being more explicit about the result.
Proof. (2)⇒(1): By Theorem 10.2.25 (which, in turn, was proved with the help
of the Mihlin Multiplier Theorem 5.5.10), −∆ has a bounded H ∞ -calculus of
angle 0. Moreover, the proof of Theorem 10.2.25 shows that this functional
calculus is explicitly given by Fourier multiplier through the formula

f (−∆) = Tf (4π2 |·|2 ) .

By Proposition 15.2.11, the bounded H ∞ -calculus is inherited by the powers

(−∆)α/2 , and the proof of Proposition 15.2.11 provides the representation

f ((−∆)α/2 ) = f ((·)α/2 )(−∆) = Tf (2α πα |·|α ) .

Finally, the translates (−∆)α/2 + λ also inherit the bounded H ∞ -calculus,

and again we have a representation
658 17 Maximal regularity

f ((−∆)α/2 + λ) = f (· + λ)((−∆)α/2 ) = Tf (2α πα |·|α +λ) .

This bounded H ∞ -calculus implies maximal Lp -regularity of (∆)α/2 on R+

by Corollary 17.3.6. The permanence properties in Theorem 17.2.26 show that
it extends to bounded subintervals (0, T ) of R+ .
(1)⇒(2): By Theorem 17.2.26, we may assume that p = q and λ > 0.
α/2
Since the semigroup S(t) := e−tA = e−t(−∆) e−λt is uniformly exponentially
stable, A has maximal Lp -regularity on I = R+ by Theorem 17.2.24.
Recall from Theorem 4.2.5 that the UMD property of X is equivalent
(with the same constant) to its dyadic version, in which the defining in-
equality involves only finitely many vectors of X at a time. From this, it
is immediate that the UMD property of X is equivalent to the property that
all finite-dimensional subspaces of X have the UMD property, with a uni-
form upper bound for their UMD constants. On the other hand, the operator
A = (−∆)α/2 + λ on Lp (Rd ; X) is the tensor extension of a scalar-valued op-
erator on Lp (Rd ). Thus, if Y ⊆ X is a subspace, then Lp (Rd ; Y ) ⊆ Lp (Rd ; X)
is an invariant subspace for A, and the maximal Lp -regularity of A on
Lp (Rd ; X) implies its maximal Lp -regularity on Lp (Rd ; Y ), with at most
the same constant. Combining these observations, it suffices to shows that
reg
βp,Y 6 Mp,A| p d
(R+ ) for all finite-dimensional subspaces Y ⊆ X. The
L (R ;Y )
advantage of this reduction is that we then already know that Y is a UMD
space, qualitatively, so that the results of the “(2)⇒(1)” part of the proof are
available to us, but of course it still remains to prove the good quantitative
estimate for βp,Y .
By Proposition 17.3.5 (with Lp (Rd ; Y ) in place of X), the function m(η) :=
A(2πiη + A)−1 defines a bounded Fourier multiplier Tm on Lp (R; Lp (Rd ; Y ))
reg
of norm kTm kL (Lp (R;Lp (Rd ;Y ))) = Mp,A (R+ ). For each fixed η ∈ R \ {0}, it is
−1
evident that A(2πiη + A) = fη (A), where fη (z) = z(2πiη + z)−1 . Thus, by
part “(2)⇒(1)”, we have

m(η) = fη (A) = Tfη (2α πα |·|α +λ) =: TM (η,·) .

For a function of the form φ ⊗ ψ with φ ∈ S (R), ψ ∈ S (Rd ; Y ), we have

Z
i2πηt
Tm (φ ⊗ ψ)(t, s) = m(η)(φ(η)ψ)(s)e
b dη
R
Z
i2πηt
= TM (η,·) (φ(η)ψ)(s)e
b dη
R
Z Z
i2πξ·s
= M (η, ξ)φ(η)
b ψ(ξ)e
b dξei2πηt dη
R Rd
Z
= M (η, ξ)φ \ ⊗ ψ(η, ξ)ei2π(η,ξ)·(t,s) d(η, ξ)
R×Rd
= TM (φ ⊗ ψ)(t, s),
 17.4 Examples and counterexamples 659

where Tm on the left is the Fourier multiplier on Lp (R; Lp (Rd ; Y )) with

operator-valued symbol m : R → L (Lp (Rd ; Y )), while TM on the right is the
Fourier multiplier on Lp (R×Rd ; Y ) with scalar-valued symbol M : Rd+1 → C.
Since linear combinations of φ ⊗ ψ as above are dense in Lp (R × Rd ; Y ) '
Lp (R; Lp (Rd ; Y )), we obtain
reg
kM kMLp (R×Rd ;Y ) = kTm kL (Lp (R;Lp (Rd ;Y )) 6 Mp,A (R+ ),

where
λ + |2πξ|α
M (η, ξ) = .
2πiη + λ + |2πξ|α
Invoking the same scaling argument as in Proposition 5.5.2, we find that
the multiplier M2 defined by
|ξ|α
M2 (η, ξ) =
iη + |ξ|α
reg
satisfies kM2 kMLp (R×Rd ;X) 6 Mp,A (R+ ). Clearly M2 satisfies M2 (η, 0) = 0
and M (0, ξ) = 1 for every η 6= 0 and ξ 6= 0. Thus Theorem 13.3.5 applies to
show that
R reg
βp,Y 6 2kM2 kMLp (R×Rd ;X) 6 2Mp,A (R+ ),

and this concludes the proof. 

The next example deals with the endpoint q = 1.

Example 17.4.2. The aim of this example is to prove that the operator −∆,
viewed as a closed and densely defined operator on L1 (Rd ), fails to have
maximal Lp -regularity on any bounded interval (0, T ) for any p ∈ [1, ∞].
Suppose, for a contradiction, that −∆ has maximal Lp -regularity on (0, T ).
By Theorem 17.2.24, for large enough λ > 0, the operator λ − ∆ has maximal
Lp -regularity on R+ . Then, by Theorem 17.3.1(1), λ−∆ is R-sectorial of angle
< π/2. By Proposition 10.3.3, this implies that the family {e−(λ−∆)t : t > 0}
is R-bounded, and by the contraction principle this implies that the family
{et∆ : t ∈ [0, 1]} is R-bounded. We will now show that the latter is not the
case.
σ 2 |y|2
Set fσ (y) := σ d e− 2 . Completing squares, one obtains

σd
Z
|x−y|2 σ 2 |y|2
et∆ fσ (x) = d/2
e− 2t e− 2 dy
(2πt) Rd
d |σx|2
Z 2
σ − 2(σ2 t+1) − σ t+1 |y− σ2x |2
= d/2
e e 2t t+1 dy
(2πt) Rd
σd −
|σx|2
= 2 d/2
e 2(σ2 t+1) .
(σ t + 1)
660 17 Maximal regularity

By Fubini’s theorem and Khintchine’s inequality,

N
X N
X 1/2
tn ∆
εn e fσ > κ1,2 |etn ∆ fσ |2 =: kF kL1 (Rd )
L1 (Ω;L1 (Rd )) L1 (Rd )
n=0 n=0

Setting σ = 2N and tn = 22n−2N , we obtain

N
σ 2d |σx|2 1/2
Z X
−
kF kL1 (Rd ) = e 22n +1 dx
Rd n=0
(22n + 1) d

N
|x|2 1/2
Z X 1 − 22n +1
= e dx
Rd n=0
(22n + 1)d

For m > 0 consider the disjoint annuli Am = {x : 2m < |x| < 2m+1 }. Splitting
the integral and estimating, we obtain
N Z |x|2
X 1 − 2(22m +1)
kF kL1 (Rd ) > e dx
m=0 Am (22m + 1)d/2
N
2md |2m x|2
Z
−
X
= e 2(22m +1) dx > (N + 1)cd ,
m=0 A0 (22m + 1) d/2

where cd > 0 depends only on the dimension d. On the other hand, by Fubini’s
theorem and Hölder’s inequality in Ω, we find
N
X √ √
εn f σ 6 N kfσ kL1 (Rd ) = N (2π)d/2 .
L1 (Ω;L1 (Rd ))
n=0

Therefore, the set {et∆ : t ∈ [0, 1]} fails to be R-bounded as claimed.

Example 17.4.3. One can check that −∆ is sectorial of angle 0 on C0 (Rd ).

However, the operator −∆ on C0 (Rd ) does not have maximal Lp -regularity
on (0, T ) for any p ∈ [1, ∞]. Indeed, since L1 (Rd ) is isometrically embedded
in the dual of C0 (Rd ), this easily follows by combining the duality result of
Proposition 17.2.32, with Example 17.4.2.
As a consequence −∆ also does not have maximal Lp -regularity on
any space which contains C0 (Rd ) isomorphically (e.g. Cub (Rd ), Cb (Rd ), and
L∞ (Rd )).

17.4.b End-point maximal regularity versus containment of c0

In Section 17.3.b we have characterised maximal L1 - and L∞ -regularity, and

in Corollary 17.3.20 we have seen that these properties hold on the real in-
terpolation spaces (X, D(A))θ,1 and (X, D(A))θ,∞ , respectively. Below we will
 17.4 Examples and counterexamples 661

obtain several results which show that positive results for p = 1 and p = ∞
are rare, and geometric restrictions on the underlying spaces are required.
From Theorem 17.2.46 we recall that maximal L∞ -regularity coincides
with maximal C-regularity in the case that A is densely defined, and therefore
all results in this section pertaining to maximal L∞ -regularity also hold for
maximal C-regularity.
We begin with a theorem due to Baillon, which implies that in Banach
spaces without an isomorphic copy of c0 , only bounded operators A can have
maximal L∞ -regularity. Examples of Banach spaces without an isomorphic
copy of c0 are reflexive Banach spaces and L1 -spaces. To see the latter, one
may for instance note that L1 -spaces have cotype 2 (Proposition 7.1.4), but
c0 does not (Corollary 7.1.10).

Theorem 17.4.4 (Baillon). Let A be a linear operator acting in a Banach

space X not containing a closed subspace isomorphic to c0 . If A has maximal
L∞ -regularity on I, then A is bounded.

Proof. It suffices to consider the case that I = (0, T ) with T ∈ (0, ∞). Reason-
ing by contradiction, we assume that A is an unbounded linear operator with
maximal L∞ -regularity and construct an isomorphic copy of c0 inside X. This
will be done by means of the Bessaga–Pelczyński theorem (Theorem 1.2.40),
which asserts that if (xn )n>1 is a sequence in a Banach space X satisfying
inf n>1 kxn k > 0 and
Xk
 j xj 6 C
j=1

for all k > 1 and all signs 1 , . . . , k ∈ {−1, 1}, with a uniform constant C,
then the closed linear span of (xn )n>1 contains a subspace isomorphic to c0 .
By Theorem 17.2.15, −A generates an exponentially bounded analytic
semigroup (S(t))t>0 . By Proposition K.3.1,
1
lim sup tkAS(t)k > ,
t↓0 e

and hence there exists a sequence T > t1 > t2 > . . . ↓ 0 and a sequence
(yj )j>1 in X of norm one vectors such that

1
tj kAS(tj )yj k > , j > 1.
2e
By restriction to [0, t1 ] and passing to a subsequence, we may assume that
tj
t1 = T and tj+1 6 for all j > 1.
2j
Setting M := supt∈[0,T ] kS(t)k and fj (s) := S(s)yj , we have
662 17 Maximal regularity
Z tj
tj kAS(tj )yj k = A S(tj − s)fj (s) ds 6 C sup kfj (t)k 6 CM,
0 t∈[0,T ]

reg
where C = M∞,A (I).
Set xj := tj AS(tj )yj for j > 0. By the preceding estimates,
1
6 inf kxj k 6 sup kxj k 6 CM.
2e j>0 j>0

To check the second condition of the Bessaga–Pelczyński theorem, fix k > 1

and signs 1 , . . . , k ∈ {−1, 1}. Define f : [0, T ] → X by
(
j S(t − T + tj )yj , t ∈ Ij := [T − tj , T − tj+1 ), j = 1, . . . , k,
f (t) :=
0, t ∈ [T − tk+1 , T ].

Then for all s ∈ [T − tk+1 , T ],

k Z
X
(S ∗ f )(s) = S(s − t)f (t) dt
j=1 Ij

k
X Z k
X
= j S(s − T + tj )yj dt = (tj − tj+1 )j S(s − T + tj )yj .
j=1 Ij j=1

The function s 7→ AS ∗ f (s) is continuous on [T − tk+1 , T ], so we can evaluate

it at s = T to obtain
k
X tj+1
(AS ∗ f )(T ) = (1 − )j xj ,
j=0
tj

where we used the definition of the vectors xj . Therefore,

k
X k
X
 j xj 6 j xj − (AS ∗ f )(T ) + k(AS ∗ f )(T )k
j=1 j=1
k k
X tj+1 X 1
6 j xj + Ckf kL∞ (0,T ;X) 6 CM + CM 6 2CM,
j=1
tj j=1
2j

where we used that tj+1 /tj 6 1/2j .

Having checked the conditions of the Bessaga–Pelczyński theorem, it fol-
lows that X contains an isomorphic copy of c0 . 
We continue with dual version of Baillon’s theorem. This time, the condition
is that X ∗ does not contain a closed subspace isomorphic to c0 . Examples of
such spaces are reflexive space, as well as the space L∞ (S), C0 (K), Cub (K)
and Cb (K), where S is a measure space and K is a locally compact Hausdorff
space. This follows from the fact that their duals have cotype 2; see the Notes.
 17.4 Examples and counterexamples 663

Corollary 17.4.5 (Guerre-Delabriere). Let A be a densely defined linear

operator acting in a Banach space X whose dual does not contain a closed
subspace isomorphic to c0 . If A has maximal L1 -regularity on I, then A is
bounded.

Proof. It suffices to consider I = (0, T ). By Proposition 17.2.32 A∗ has max-

imal L∞ -regularity on (0, T ), and the result follows by applying Baillon’s
theorem to this operator. 
The condition that X or X ∗ contain no isomorphic copy of c0 cannot be
omitted in Theorem 17.4.4 and Corollary 17.4.5, respectively:

Example 17.4.6. On X = c0 consider the unbounded operator A : (yn )n>1 7→

(nyn )n>1 with maximal domain

D(A) = {(xn )n>1 ∈ c0 : (nxn )n>1 ∈ c0 }.

Then A is closed and densely defined. The adjoint operator A∗ on `1 is given

by the same multiplication operator and has maximal domain

D(A∗ ) = {(xn )n>1 ∈ `1 : (nxn )n>1 ∈ `1 },

and is again closed and densely defined. Let I = R+ or I = (0, T ) with

T ∈ (0, ∞) and p ∈ [1, ∞]. Then:
(1) A has maximal Lp -regularity on I if and only if p ∈ (1, ∞].
(2) A∗ has maximal Lp -regularity on I if and only if p ∈ [1, ∞).
Assertion (1) follows from (2) and the duality result of Proposition 17.2.32.
To prove (2), let F be the dense subspace of all f ∈ L1 (R+ ; `1 ) for which
there exists an N > 1 such that for all n > N for all t ∈ I we have fn (t) = 0.
Referring to Theorem 17.2.19, it suffices to show that for every f ∈ F the
mild solution uf of the problem u0 + Au = f , u(0) = 0, takes values in D(A∗ )
and
kA∗ uf kL1 (R+ ;`1 ) 6 kf kL1 (R+ ;`1 ) .
We begin by noting that uf is given by
Z t
(uf (t))n = e−n(t−s) fn (s) ds.
0

Since fn vanishes for n > N it is clear that A∗ uf (t) ∈ D(A∗ ), and by Young’s
inequality
Z ∞X Z t
kA∗ uf kL1 (R+ ;`1 ) = ne−n(t−s) fn (s) ds dt
0 n>1 0

XZ ∞ Z t
= ne−n(t−s) fn (s) ds dt
n>1 0 0
664 17 Maximal regularity
X
6 kne−n· kL1 (R+ ) kfn kL1 (R+ )
n>1
X
= kfn kL1 (R+ ) 6 kf kL1 (R+ ;`1 ) ,
n>1

interchanging sum and integral in the last step. From the extrapolation result
of Theorem 17.2.31 we find that A∗ has maximal Lp -regularity for all p ∈
[1, ∞). Finally, we note that by Theorem 17.4.4, A∗ cannot have maximal
L∞ -regularity, for `1 does not contain a closed subspace isomorphic to c0 .

It is a classical result due to Sobczyk that if a closed subspace of a separable

Banach space X is isomorphic to c0 , then this subspace is complemented (see
the Notes). Therefore, Example 17.4.6(1) can be extended to any separable
Banach space containing an isomorphic copy of c0 . In this sense, the example
shows that for separable Banach spaces Theorem 17.4.4 is optimal. By another
classical result, due to Bessaga and Pelczyński, the dual a Banach space X
contains an isomorphic copy of c0 if and only if X contains a complemented
copy of `1 (see the Notes). In this sense, the Example 17.4.6(2) shows that
Corollary 17.4.5 is optimal.
While the previous results show that maximal L∞ - and L1 -regularity of
unbounded operators
• cannot occur in the absence of c0 (Theorem 17.4.4, Corollary 17.4.5),
• can occur in the presence of c0 (Example 17.4.6),
it is worth noting that, even in the latter case, such end-point regularity is
not something to be typically expected. In particular, Examples 17.4.2 and
17.4.3 show that the Laplacian −∆ fails both maximal L1 -regularity and
L∞ -regularity on X = L1 (Rd ) and X = C0 (Rd ), respectively. Moreover, by
Theorem 17.2.46, −∆ also does not have maximal C-regularity.
We complete our discussion of Example 17.4.6 by also considering the same
operator as in this example on the space `q with q ∈ (1, ∞).

Example 17.4.7. On `q with q ∈ (1, ∞), we consider the closed densely defined
operator A : (yn )n>1 7→ (nyn )n>1 with maximal domain

D(A) = {(xn )n>1 ∈ `q : (nxn )n>1 ∈ `q }.

Let I = R+ or I = (0, T ) with T ∈ (0, ∞) and let p ∈ [1, ∞]. We will show
that A has maximal Lp -regularity if and only if p ∈ (1, ∞). Indeed, the fact
that maximal Lp -regularity fails for p = ∞ and p = 1 follows from Theorem
17.4.4 and Corollary 17.4.5. For p = q we argue as in Example 17.4.6. For all
f ∈ Lq (R+ ; `q ), by Young’s convolution inequality we have
Z ∞ XZ t q XZ ∞ Z t q
ne−n(t−s) fn (s) ds dt = ne−n(t−s) fn (s) ds dt
0 n>1 0 n>1 0 0
 17.4 Examples and counterexamples 665
X q q
6 kne−n(·) kL 1 (R ) kfn kLq (R )
+ +
n>1

6 kf kqLq (R+ ;`q ) .

The case p ∈ (1, ∞) \ {q} follows by extrapolation, using Theorem 17.2.31.

17.4.c Analytic semigroups may fail maximal regularity

The aim of the present section is to prove the following theorem.

Theorem 17.4.8 (Kalton–Lancien, Fackler). For each 1 < q < ∞ the
space Lq (0, 1) admits a densely defined invertible sectorial operator Aq , with
ω(Aq ) = 0, such that the following assertions hold:
(1) the operator Aq is R-sectorial if and only if q = 2;
(2) we have Sq (t) = Sr (t) for all 1 < q, r < ∞ and t > 0, where Sq and Sr
denote the bounded analytic C0 -semigroups generated by −Aq and −Ar ,
respectively.
Theorem 17.4.8 has a number of interesting consequences:
• In Theorem 17.3.1, one cannot replace ‘R-sectorial’ by ‘sectorial’.
• Combining the theorem with Theorem 17.2.24 and the first part of The-
orem 17.3.1, it follows that maximal Lp -regularity on R+ and (0, T ) may
fail for densely defined invertible sectorial operators A of zero angle on
Lq (0, 1) unless q = 2.
• It follows that maximal Lp -regularity on R+ may fail for bounded sectorial
operators A of zero angle on Lq (0, 1) unless q = 2. Indeed, consider A−1
from the previous bullet. It is clearly bounded, and sectorial of angle zero
since λR(λ, A) = A−1 R(1/λ, A−1 ). However, A−1 cannot be R-sectorial,
since this identity would then imply that A is R-sectorial.
• Since A2 has maximal Lp -regularity, the second assertion of the theorem
shows that maximal Lp -regularity does not extrapolate from L2 (0, 1) to
Lq (0, 1) in the case of consistent bounded analytic C0 -semigroups on these
spaces.
• Theorem 10.7.13 (where the second occurrence of the word ‘contraction’
should be deleted from the formulation of the theorem as printed), ‘con-
tractive for t > 0 and positive’ cannot be replaced by ‘bounded for t > 0’.
The operator Aq constructed in the proof of the theorem is a diagonal operator
with respect to Schauder basis for a complemented subspace of Lq (0, 1) that
is unconditional and non-homogeneous in the sense of the following definition;
on the complement of this subspace we take A = 0.
Definition 17.4.9. Let X be a Banach space.
(1) A sequence (xn )n>1 in X is said to be a basic sequence if it is a Schauder
basis for its own closed linear span.
666 17 Maximal regularity

(2) Two basic sequences (xn )n>1 and (yn )n>1 in X are said to be equivalent
if for any scalar sequence (cn )n>1 it is true that
X X
cn xn converges if and only if cn yn converges.
n>1 n>1

(3) A Schauder basis (xn )n>1 in X is said to be homogeneous if every two

disjoint subsequences of (xn )n>1 are equivalent as basic sequences.
If two basic sequences (xn )n>1 and (yn )n>1 are equivalent, their closed linear
spans [(xn )n>1 ] and [(yn )n>1 ] are isomorphic as Banach spaces. Indeed, the
correspondence X X
bn xn 7→ bn yn
n>1 n>1

gives a bijection between [(xn )n>1 ] and [(yn )n>1 ]. Since the coordinate projec-
tions of a Schauder basis are bounded, this bijection is closed, and therefore
bounded by the closed graph theorem. Applying the same argument to its
inverse, it follows that this bijection is an isomorphism.
The following lemma isolates the crucial property of Lq (0, 1) that will be
needed in the proof of Theorem 17.4.8.
Lemma 17.4.10. The Haar basis of Lq (0, 1) is unconditional for every 1 <
q < ∞, and non-homogeneous for q 6= 2.
Using the notation of Section 9.1.h, the (L2 -normalised) Haar basis is given
as (hn )n>1 , where h1 ≡ 1 and hn := φj,k for n > 2, where n = 2j + k with
j = 0, 1, 2, . . . and k = 1, . . . , 2j , and

φj,k := 2j/2 1( k−1 , k−1/2 ) − 2j/2 1( k−1/2 , k ) .

2j 2j 2j 2j

Note that φj,k is supported on the interval ( k−1 k

2j , 2j ).

Proof of Lemma 17.4.10. Unconditionality for 1 < q < ∞ of the Haar basis
is proved in Corollaries 4.5.8 and 4.5.16.
Consider the subsequence (φn,2 )n>2 . The supports of the functions gn
are disjoint, contained in (0, 21 ), and therefore they span a closed sub-
space isomorphic to `q . On the other hand, the functions in the sequence
(φj,k )j=1,2,3,...; k=2j−1 +1,...2j have support in ( 21 , 1) and their linear span con-
P 2j
tains the ‘( 21 , 1)-Rademacher’ functions rj = k=2j−1 +1 φj,k , j > 1, whose
closed span is isomorphic to `2 by the Khintchine inequalities. Since it can be
shown (a proof is included in the final paragraph of this section) that `q does
not contain a closed subspace isomorphic to `2 unless q = 2, it follows that
for 1 < q < ∞ with q 6= 2, the Haar basis of Lq (0, 1) is unconditional and
non-homogeneous. 
With this lemma at hand, Theorem 17.4.8 is seen to be a special case of the
following more general result:
 17.4 Examples and counterexamples 667

Theorem 17.4.11. Every Banach space with a non-homogeneous uncondi-

tional basis admits a densely defined invertible sectorial operator A, with
ω(A) = 0, which is not R-sectorial.

By invoking additional results from the theory of Schauder bases, Kalton and
Lancien proved an even stronger result, namely, that if X is a Banach space
with an unconditional basis (xn )n>1 with the property every sectorial operator
of angle zero on X is R-sectorial, then X is isomorphic to `2 . We will comment
on this result Notes at the end of the chapter.
Proof of Theorem 17.4.11. Let (en )n>1 be a non-homogeneous unconditional
basis for X. By renorming, this basis can be made into an 1-unconditional
basis, i.e., an unconditional basis satisfying
X X
λn cn en 6 kλk`∞ cn en
n>1 n>1

∈ X and λ = (λn )n>1 ∈ `∞ . Indeed, the equivalent

P
for all x = m>1 cn en
norm X X
cn en := sup λn c n e n
n>1 kλk∞ 61 n>1

will do. This renorming preserves non-symmetry of the basis, and will be used
in the proof below to eliminate inessential unconditionality constants from
several estimates. As a side-remark, a Banach space with a 1-unconditional
basis is a Banach lattice with respect to the coordinate-wise ordering.
P (emj )j>1 and (enj )j>1P
Choose disjoint subsequences , as well as a scalar
sequence (aj )j>1 , such that j>1 aj emj converges and j>1 aj enj does not
converge.
The construction below will produce a sectorial operator A on the closed
linear span Y of the two sequences (emj )j>1 and (enj )j>1 that has all the
stated properties. The desired example on the whole space X is then obtained
by taking the direct sum with the identity operator on the closed linear span
Z of the remaining basis vectors (note that X = Y ⊕ Z with contractive
projections thanks to the 1-unconditionality).
Since the subsequences (emj )j>1 and (enj )j>1 do not overlap, the first
be relabelled as (e2j )j>1 and the second as (e2j−1 )j>1 . Define the sequence
(fj )j>1 in X by
(
ej (j odd)
fj =
ej + ej−1 (j even)

so that e2j = f2j − f2j−1 . It is elementary to check that (fj )j>1 is a Schauder
basis for its closed linear span. Consider the diagonal operator
X X
A: cj fj 7→ 2j cj fj
j>1 j>1
668 17 Maximal regularity

with its natural domain. By Lemma 10.2.27 (restricting to the positive in-
tegers as index set) this operator is sectorial of zero angle, and moreover
0 ∈ %(A). The analytic C0 -semigroup (S(t))t>0 generated by −A is bounded
and strongly continuous on every sector of angle strictly less than 21 π.
Assume, for a contradiction, that A is R-sectorial. Let λ = (λj )j>1 be a
sequence of positive real numbers to be chosen shortly. Since e2j = f2j −f2j−1 ,
we obtain
`
X
εj aj λj (λj + A)−1 (f2j − f2j−1 )
j=k
`  λ a
X j j λj a j 
= εj f 2j − f 2j−1
λj + 22j λj + 22j−1
j=k
`  λ a
X j j λj a j 
= εj (e 2j + e 2j−1 ) − e 2j−1
λj + 22j λj + 22j−1
j=k
`  λ a  λ
X j j j λj  
= εj e 2j + a j − e 2j−1 .
λj + 22j λj + 22j λj + 22j−1
j=k

Now take λj := 22j . The above identity then takes the form
` ` 1
X X 1 
εj aj 22j (22j + A)−1 (f2j − f2j−1 ) = εj aj e2j + aj e2j−1 .
2 6
j=k j=k

The 1-conditionality of (ej )j>1 and the assumed R-boundedness imply that
`  `
X 1 1  X 1 1 
aj e2j − aj e2j−1 = εj aj e2j + aj e2j−1
2 6 2 6 L2 (Ω;X)
j=k j=k
`
X
6 R({λ(λ + A)−1 : λ > 0}) εj aj e2j
L2 (Ω;X)
j=k
`
X
= R({λ(λ + A)−1 : λ > 0}) aj e2j ,
j=k

where R({λ(λ + A)−1 : λ > 0}) denotes the R-bound of the set {λ(λ +
A)−1 : λ > 0}. Since the right-hand side tends to P
0 as k, ` → ∞, so does
the left-hand side. By the convergence of the sum j>1 aj e2j , this in turn
implies
X`
lim lim aj e2j−1 = 0.
k,`→∞ k,`→∞
j=k
P`
This contradicts the fact that j=k aj e2j−1 fails to converge. 
 17.4 Examples and counterexamples 669

On closed subspaces of `q

In this paragraph we complete the proof that the Haar basis of Lq (0, 1) is
non-homogeneous for q ∈ [1, ∞) unless q = 2 (Lemma 17.4.10). The missing
piece of information was the following result.

Proposition 17.4.12. Let 1 6 q < ∞. The space `q contains a closed sub-

space isomorphic to `2 (if and) only if q = 2.

Since `q and `2 are isomorphic (if and) only if q = 2 (e.g., by looking at

their type and cotype), this proposition is an immediate consequence of the
following result.

Proposition 17.4.13. Every infinite-dimensional closed subspace of `q con-

tains a closed subspace isomorphic to `q .

The proof of this proposition is based on two lemmas, for which we need to
introduce some terminology. Recall that a sequence in a Banach space X is
said to be a basic sequence if it is a Schauder basis for its closed linear span,
and that two basic sequences (xn )n>1 and (yn )n>1 are said to be equivalent
if for any scalar sequence (cn )n>1 it is true that
X X
cn xn converges if and only if cn yn converges.
n>1 n>1

The basis constant of a basic sequence (xn )n>1 is defined as the (finite) number
supN >1 kPN k, where PN is the projection in the closed span [(xn )n>1 ] of
(xn )n>1 defined by
X XN
PN cn xn := c n xn .
n>1 n=1

A sequence (yn )n>1 in X is said to be a block sequence of the basic sequence

(xn )n>1 in X if there exist a strictly increasing sequence of positive integers
(Nn )n>1 and a scalar sequence (cn )n>1 such that for all m > 1 we have

m −1
NX
ym = c n xn ,
m=Nm−1

where N0 := 1. Clearly, a block sequence of a basic sequence is a basic se-

quence.

Lemma 17.4.14. Let (xn )n>1 be a basic sequence in a Banach space X and
let (yn )n>1 be a sequence in X which satisfies
X 1
kxn − yn k 6 ,
3K
n>1
670 17 Maximal regularity

where K is the basis constant of (xn )n>1 . Then (yn )n>1 is a basic sequence
equivalent to (xn )n>1 , and the mapping
X X
cn xn 7→ c n yn
n>1 n>1

sets up an isomorphism of their closed linear spans.

P P
Proof. The mapping T : n>1 cn xn → n>1 cn yn is well defined from the
closed linear span [(xn )n>1 ] to the closed linear span [(yn )n>1 ], because
M
X M
X M
X
c n yn 6 c n xn + |cn |kyn − xn k
n=m n=m n=m
M
X M
X
6 c n xn + 2Kkxk · kyn − xn k
n=m n=m
P
implies that the partial
P sums of n>1 cn yn form a Cauchy sequence, and
therefore the sum n>1 cn yn converges. Moreover,
X 1 2
kx − T xk 6 |cn |kxn − yn k 6 2Kkxk · = kxk.
2K 3
n>1

This implies that kT xk > 13 kxk for all x ∈ [(xn )n>1 ], and therefore T is an
isomorphism from [(xn )n>1 ] onto its range in [(yn )n>1 ], which is dense since
it contains all finite linear combinations of the yn . It follows that T is an
isomorphism from [(xn )n>1 ] onto [(yn )n>1 ].
Finally, let y ∈ [(yn )n>1 ] and let x := T −1 y. Since (xn )P
n>1 is a Schauder
basisPfor [(xn )n>1 ] we have a unique representation x = n>1 cn xn . Then
y = n>1 cn yn and this representation is unique. This proves that (yn )n>1 is
a Schauder basis for [(yn )n>1 ], i.e., (yn )n>1 is a basic sequence. 

Lemma 17.4.15. Every infinite-dimensional closed subspace Y of a Banach

space X with a Schauder basis (xn )n>1 has a closed subspace Z with a nor-
malised Schauder basis (yn )n>1 , equivalent to a block sequence (un )n>1 of
(xn )n>1 , and with [(yn )n>1 ] ' [(un )n>1 isomorphically.

Proof. Since Y is infinite-dimensional,

P for every k > 1 it is possible to find
elements in Y of the form n>k an xn . P
Let y1 ∈ Y be an arbitrary norm one vector in Y , say y1 = n>1 an xn ∈
PN1 −1
Y . Choose N1 > 1 so large that ky1 − u1 k < 1/(4K), where u1 = n=1 an xn
and K is the basisPconstant of (xn )n>1 . Next let y2 ∈ Y be a norm one vector
of the form y2 = n>N1 an xn and choose N2 > 1 so large that ky2 − u2 k <
PN2 −1
1/(42 K), where u2 = n=N 1
an xn . Continuing in this way, we obtain a block
sequence (un )n>1 of (xn )n>1 for which we have
 17.5 Notes 671
X 1 X 1 1
kyn − un k 6 = .
K 4n 3K
n>1 n>1

By Lemma 17.4.14, (yn )n>1 is a normalised basic sequence equivalent to the

block sequence (un )n>1 and [(yn )n>1 ] ' [(un )n>1 isomorphically. The closed
linear span Z of (yn )n>1 has the desired properties. 
Proof of Proposition 17.4.12. Let Y be an infinite-dimensional closed sub-
space of `q . By Lemma 17.4.15, Y has a closed subspace Z with a normalised
Schauder basis (yn )n>1 equivalent to a block sequence (un )n>1 of (xn )n>1 and
with [(yn )n>1 ] ' [(un )n>1 isomorphically.
If (cn )n>1 is any scalar sequence, then
k k Nj −1
X q X X q
c n un = cj an en
`q `q
n=j n=j n=Nj−1
k Nj −1 k k
X X (∗) X X q
= |cj an |q h |cj |q = cn en ,
`q
n=j n=Nj−1 n=j n=j

where (∗) follows from the fact that 43 6 kun k 6 54 for all n > 1. This
computation shows that [(un )n>1 ] ' [(en )n>1 ] = `q isomorphically, and since
also [(yn )n>1 ] ' [(un )n>1 ] isomorphically we conclude that Z := [(yn )n>1 ] '
`q isomorphically. 

17.5 Notes
Parabolic maximal regularity estimates of the form ku0f kp + kAuf kp 6 Ckf kp
can be traced to back to the Ladyženskaja, Solonnikov, and Ural0 ceva [1968].
Early contributions to the abstract operator-theoretic framework include
Sobolevskiı̆ [1964], De [1964], Grisvard [1969], Da Prato and Grisvard [1975].
More recent expositions can be found in Amann [1995], Kunstmann and Weis
[2004], Denk, Hieber, and Prüss [2003], Prüss and Simonett [2016].
There is an extensive literature of applications of maximal Lp -regularity
to non-linear parabolic problems; we refer to Prüss and Simonett [2016] and
references therein. An abstract approach to quasi-linear and semi-linear evo-
lution equations based on maximal Lp -regularity will be presented in Chapter
18. A small sample of the applications in the literature, which counts several
hundreds of papers, is given in the notes of that chapter.

Section 17.1

Most of the semigroups in this chapter are analytic semigroups which are not
necessarily strongly continuous. These are widely used in the literature, and
an extensive treatment can be found in Lunardi [1995].
672 17 Maximal regularity

Our treatment of the initial value problem

u0 (t) + Au(t) = f (t), t ∈ [0, T ],

u(0) = x,

for non-densely defined operators A in a Banach space X is focussed on the

existence, uniqueness, and regularity of Lp -solutions. For continuous functions
f : [0, T ] → X, the problem is studied by Da Prato and Sinestrari [1987]
where the following result concerning classical solutions is established: If (i)
A satisfies the resolvent estimates of the Hille–Yosida theorem, (ii) f (t) =
Rt
f (0) + 0 g(s) ds for some Bochner integrable function g : (0, T ) → X, and
(iii) x ∈ D(A) satisfies
−Ax + f (0) ∈ D(A),
then there exists a unique u ∈ C 1 ([0, T ]; X) ∩ C([0, T ]; D(A)) satisfying the
initial value problem pointwise in t ∈ [0, T ]. A proof using the theory of locally
Lipschitz integrated semigroups was given by Kellerman and Hieber [1989].

Section 17.2

In our definition of maximal Lp -regularity, we only impose that the Lp -solution

should satisfy Au ∈ Lp (I; X). In the early literature it was often assumed
that in addition to this one also has u ∈ Lp (I; X). This imposes a genuine
restriction if I = R+ and makes proofs often simpler. The present definition
was proposed by Weis [2001b].
Several results in Section 17.2 are based on the important paper by Dore
[2000]. In particular this applies to Propositions 17.2.5, 17.2.7 17.2.10, 17.2.11,
Lemma 17.2.12, Theorems 17.2.15 and 17.2.19, the second part of Theorem
17.2.24, and Theorem 17.2.26(1)–(4).
The results of Section 17.2.b are standard applications of the trace method
for real interpolation as discussed in Appendix L. An exception is the case
p = 1, which is treated by an ad hoc method here.
The proof of Theorem 17.2.15 follows Dore [2000], but injectivity is shown
by an argument from Prüss and Simonett [2016]. Under the a priori assump-
tion that −A generates a C0 -semigroup, the result was stated in Sobolevskiı̆
reg
[1964]. The sectoriality estimate kA(λ + A)−1 k 6 Mp,A (R+ ) of (17.14) ap-
pears to be new. Propositions 17.2.18 and 17.2.29 seem to be new as well, and
turn out to be quite useful in approximation arguments for proving maximal
Lp -regularity on R+ .
In Dore [2000], the proof of the second part of Theorem 17.2.24 is at-
tributed to Kato. The technical estimate (17.31) for the operator λ + A seems
to be new, and is crucial in the proof of a new type of maximum principle for
maximal regularity presented in Proposition 17.2.27(1).
Theorems 17.2.19, 17.2.26(1)–(4) are due to Dore [2000], although the
proof of the extrapolation of the integrability parameter of part (4) is cred-
ited to Kato, who reduced the proof to the extrapolation result Benedek,
 17.5 Notes 673

Calderón, and Panzone [1962, Theorem 2 ]. This result is not immediately ap-
plicable however, since it is phrased for kernels which are locally integrable on
Rd . In the present context one needs a corresponding result for kernels which
are locally integrable on Rd \{0}. The approach taken in the main text circum-
vents this problem by using the more general extrapolation result of Chapter
11. Under the assumption that −A generate an analytic C0 -semigroup, the
extrapolation result of part (4) of Theorem 17.2.31 is due to Cannarsa and
Vespri [1986] and Coulhon and Lamberton [1986]; see also Sobolevskiı̆ [1964];
the quantitative bound appears to be new. An extrapolation result for maxi-
mal Lp -regularity to the vector-valued Hardy space H 1 (I; X) is presented in
Hytönen [2005].
The duality result of Theorem 17.2.265 is due to Kalton and Portal [2008],
who proved it for maximal Lp -regularity on R+ via discrete maximal regular-
ity. The precise bounds for arbitrary intervals of the duality result obtained
in Proposition 17.2.32 seem to be new.

Weighted extrapolation

Extrapolation to weighted estimates is a standard method in harmonic anal-

ysis, and we presented it from a modern viewpoint in Chapter 11. Extrap-
olation of Lp -boundedness of singular integral operators to Lp -spaces with
power weights goes back a long way. It was already considered by Hardy and
Littlewood [1936] and Babenko [1948] in the one-dimension setting. The d-
dimensional case was considered in Stein [1957]. These results do not require
any smoothness of the kernel.
Part of Proposition 17.2.47 on extrapolation of maximal C-regularity for
weights wα (t) = tα appeared in Clément and Simonett [2001]. In Prüss and
Simonett [2004], the extrapolation result of Proposition 17.2.36 for the weights
wα was proved for p ∈ (1, ∞). A proof based on Schur’s lemma appears in
Auscher and Axelsson [2011]. Our proof is based on Young’s inequality for the
multiplicative group R+ . Some of the results in the end-points and maximal
C-regularity (see Proposition 17.2.47) may be new. The extrapolation result
for exponential weights presented in Proposition 17.2.38 is folklore.
The non-quantitative version of Theorem 17.2.39 follows from the weighted
Calderón–Zygmund theory in Rubio de Francia, Ruiz, and Torrea [1986].
Extensions to weighted Banach function spaces and applications to non-
autonomous maximal Lp -regularity have been given in Chill and Fiorenza
[2014]. The class of admissible weights for singular integral operators on the
half-line was analysed by Chill and Król [2018].

Maximal C-regularity

Maximal continuous regularity was first shown on real interpolation spaces

by Da Prato and Grisvard [1975, 1979], where it was used to study quasilin-
ear evolution equations. This paper also contains one implication of Theorem
17.2.46, as well as the result that positive self-adjoint operators on Hilbert
674 17 Maximal regularity

spaces with compact resolvent fail maximal C-regularity. Weighted analogues

and simpler proofs of the results in Da Prato and Grisvard [1975, 1979] were
presented by Angenent [1990a]. This paper also contains far-reaching applica-
tions to fully non-linear equations. Further results, and applications to non-
linear equations and their stability, can be found in Amann [1995], Clément
and Simonett [2001], Da Prato and Lunardi [1988], LeCrone and Simonett
[2020], Lunardi [1995], Shao and Simonett [2014], Travis [1981] and many
other papers. Some applications of maximal C-regularity to quasi-linear evo-
lution equations will be presented in Chapter 18. A variant of the results by
Dore [2000] for maximal C-regularity can be found in LeCrone and Simon-
ett [2011]. Versions of Corollary 17.3.20 for p = ∞ also hold for (weighted)
maximal C-regularity, and can be proved via the same method or derived
afterwards (see Remark 17.3.21 and Proposition 17.2.47). The equivalence of
maximal L∞ and C-regularity may be known, but we are not aware of a ref-
erence stating this result. We also do not know whether the condition that
D(A) be dense in X can be removed.
A different version of the perturbation result in Proposition 17.2.49 ap-
pears in Dore [2000]. Our proof uses a less restrictive relative bound.
The application to maximal Lp -regularity of continuous time-dependent
operators A : I → L (X1 , X0 ) as in Theorem 17.2.51 is standard. We will
discuss time-dependent problems in more detail below.

Section 17.3

According to Coulhon and Lamberton [1986], the following question was asked
by Brezis:
“à quelles conditions sur l’espace de Banach E a-t-on la régularité Lp
pour tout A générateur d’un semi-groupe analytique borné sur E ?”
[Under which conditions on the Banach space E does every bounded analytic
semigroup generator have maximal Lp -regularity?]
Corollary 17.3.8, which is due to De [1964], asserts maximal Lp -regularity
for negative generators of bounded analytic C0 -semigroups on Hilbert spaces.
In Coulhon and Lamberton [1986] it was shown that the Poisson semigroup
has maximal Lp -regularity on Lq (Rd ; X) if and only if X is a UMD space (see
Theorem 17.4.1 for the precise formulation). Lamberton [1987] subsequently
showed that if −A generates a bounded analytic C0 -semigroup S on a space
L2 (Ω) such that the operators S(t) extend to contractions on Lq (Ω) for every
q ∈ [1, ∞], then A has maximal Lp -regularity for all p ∈ (1, ∞).
A breakthrough was made by Dore and Venni [1987], where it was shown
that every operator A in a UMD Banach space X with bounded imaginary
powers of angle ωBIP (A) < π/2 has maximal Lp -regularity for p ∈ (1, ∞). The
approach was based on their preliminary version of (what is nowadays called)
the Dore–Venni theorem (Theorem 15.4.11), in which the additional condition
that the operators be invertible was made. This invertibility condition was
 17.5 Notes 675

removed in Prüss and Sohr [1990], and Corollary 17.3.7 follows from this
paper.
It took approximately ten years Brezis’s question was settled definitively:
• Kalton and Lancien [2000] showed that Brezis’s question, in the form
stated, has a negative answer. For every q ∈ (1, ∞) \ {2}, they gave an
example of an operator −A generating a bounded analytic C0 -semigroup
on Lq (0, 1) such that A does not have maximal Lp -regularity on finite time
intervals (see Theorem 17.4.8).
• It was shown in Weis [2001a,b] that a sectorial operator A on a UMD
Banach space X has maximal Lp -regularity if and only if A is R-sectorial
(see Theorem 17.3.1).
The negative general answer of Kalton and Lancien [2000] will be further dis-
cussed below. The characterisation of maximal Lp -regularity in UMD spaces of
Weis [2001a,b] is based on the operator-valued Fourier multiplier theorem for
UMD-valued multipliers (Theorem 5.3.18), which was proved in the same pa-
per. This work led to many follow-up studies in which maximal Lp -regularity
was proved by checking R-sectoriality, some of which will be discussed be-
low. In particular, Weis’s result implies the sufficient conditions for maximal
Lp -regularity stated in Corollaries 17.3.7 and 17.3.9. Corollary 17.3.9 resem-
bles the previously mentioned result of Lamberton [1987], which instead of
positivity, assumes contractivity on Lq for all q ∈ [1, ∞]. One can check that
under the latter condition the argument in the proof of Corollary 17.3.9 can
be repeated: the R-sectoriality with ωR (A) < π/2 proved in Theorem 10.7.13
was based on Akcoglu’s maximal ergodic Theorem 10.7.14. The latter was
stated and proved for semigroups which are positive and contractive for some
q ∈ (1, ∞), but is also valid if contractivity is assumed for all q ∈ [1, ∞] (see
Dunford and Schwartz [1958, Theorem VIII.7.7]).
The perturbation result of Corollary 17.3.10 improves Proposition 17.2.49
for UMD spaces, since the condition of the relative bound is weaker. A similar
result under a different condition can be found in Kunstmann and Weis [2001].
The characterisations of maximal Lp -regularity for the endpoints p = 1
and p = ∞ of Theorems 17.3.11 and 17.3.12 are to due to Kalton and Portal
[2008], who proved these results via discrete versions of maximal Lp -regularity.
It seems that their proof of Theorem 17.3.12 also assumes that the operator
A be densely defined.
The alternative proof of Theorem 17.3.1 presented in Theorem 17.3.18 is
due to Kalton and Weis [2001], and is based on their sum-of-operator method
explained in Theorem 16.3.6. Proposition 17.3.14 seems to be folklore. The
fact that only the weaker result of Proposition 17.3.15 holds in the case I = R+
seems to be less known. The properties of the derivative operator Du = u0 on
Lp (I; X) with a Dirichlet condition at the left end-point collected in Proposi-
tion 17.3.16 are standard. Extensions to weighted spaces Lp (R+ , tγ d, t; X) can
be found in Prüss and Simonett [2004] and Meyries and Schnaubelt [2012b]. A
direct proof based on the Mihlin multiplier theorem and the duality between
676 17 Maximal regularity

−D and D can be found in Lindemulder, Meyries, and Veraar [2018], where

also the complex interpolation spaces of D(D) are identified.

Differential operators with maximal Lp -regularity

The characterisation of Weis [2001a,b] presented in Theorem 17.3.1, led to

many follow-up works in which R-sectoriality of angle ωR A < π/2 was
checked for concrete differential operators. Moreover, due to the estimate
ωH ∞ (A) 6 ωR (A) (see Theorem 10.4.9), these R-sectoriality estimate also
led to new examples of operators with a bounded H ∞ -calculus of angle
ωH ∞ (A) < π/2. Before the characterisation was known usually BIP-conditions
on A were checked which for elliptic differential operators often required α-
Hölder regularity of the coefficients and Poisson estimates on the kernels.
Details can be found in Duong and Robinson [1996], Hieber and Prüss [1997],
Coulhon and Duong [2000] and references therein.
First of all Corollary 17.3.9 and the mentioned result of Lamberton [1987]
provides a wide class of examples with maximal Lp -regularity on Lq -spaces.
The contractivity condition can however be a serious constraint when dealing
with systems or differential operators of order four or higher (see Langer and
Maz0 ya [1999]).
In their influential memoir, Denk, Hieber, and Prüss [2003] proved max-
imal Lp -regularity for a large class of uniformly elliptic systems under the
following conditions
• domain equals Rd , R+
d
, or Ω ⊆ Rd which is open, bounded, and smooth;
• non-divergence form with bounded and uniformly continuous coefficients;
• boundary conditions of Lopatinskii-Shapiro type.
Piasecki, Shibata, and Zatorska [2020] proved maximal Lp -regularity without
the need of uniform Lopatinskii-Shapiro conditions. Moreover, on Rd Haller-
Dintelmann, Heck, and Hieber [2006], Heck and Hieber [2003] managed to
prove maximal Lp -regularity under VMO (vanishing mean oscillation) condi-
tions on the coefficients, but do not consider any boundary conditions.
In Clément and Prüss [2001] and Weis [2001a] it is explained that often
R-sectoriality can be checked by a simple domination by maximal functions,
leading to an effective approach to prove maximal Lp -regularity. These ideas
emerged into a very flexible framework in a series of papers by Blunck and
Kunstmann [2002, 2003, 2004, 2005] (see also Kunstmann and Weis [2004]),
who proved R-sectoriality under off-diagonal bounds on the semigroup. They
apply their results to elliptic operators in divergence form. Here no symmetry
of the coefficients is required and thus contractivity cannot be expected on Lq
with q 6= 2, and thus Corollary 17.3.9 is not applicable.
Kunstmann [2008] proved maximal Lp -regularity a class of elliptic differ-
ential operators A in non-divergence form only assuming that −A generates
an analytic semigroup in Lq (Rd ). This can be seen as a positive answer to
Brezis’s question for a special class of elliptic differential operators on Lq (Rd ).
 17.5 Notes 677

The following theorem is immediate from Theorem 17.3.1, and the Ap -

weighted characterisation of R-boundedness given in Theorem 8.2.6, which
trivially extends to open subsets of Rd , if one uses restrictions of Ap -weight
on Rd . This method was first used by Fröhlich [2001], Fröhlich [2007] and
Haller, Heck, and Hieber [2003] to prove maximal Lp -regularity.
Theorem 17.5.1. Let Ω ⊆ Rd be an open set. Let r ∈ (1, ∞) and let φr :
[1, ∞) → [0, ∞) be a non-decreasing function. Let A be a θ-sectorial operator
on Lr (Ω) with θ ∈ (0, π/2), and suppose that for all w ∈ Ar and λ ∈ {Σθ , the
resolvent R(λ, A) extends to an element in L (Lr (Ω, w)) and satisfies

kλR(λ, A)kL (Lr (Ω,w)) 6 φ([w]Ar ).

Then A has maximal Lp -regularity on Lq (Ω, w) for any p, q ∈ (1, ∞) and any
w ∈ Aq .
The above result can be seen as another way to answer to Brezis question on
Lr -spaces: if the semigroup is bounded analytic on all Lr (Ω, w) with w ∈ Ar ,
then one has maximal Lp -regularity. Theorem 17.5.1 was used by Haller, Heck,
and Hieber [2003] to prove maximal Lp -regularity for a large class of elliptic
systems on Rd .

Maximal Lp -regularity for linear operators arising in fluid dynamics

Fröhlich [2007] used Theorem 17.5.1 to prove maximal Lp -regularity for the
Stokes operator on C 1,1 -domains, and extends the papers Farwig and Sohr
[1997], Fröhlich [2003] where Rd and the half-space were considered, respec-
tively. Currently, such results under the weakest conditions on the domain are
due to Kunstmann and Weis [2017], and state that for any θ ∈ (0, π/2) the
Stokes operator has a bounded H ∞ -calculus of angle < π/2 (and thus maxi-
mal Lp -regularity) on bounded Lipschitz domains for all q ∈ (1, ∞) satisfying
| 1q − 12 | < 2d
1
+ ε Here ε > 0 only depends on the Lipschitz domain, and the
angle θ. The condition on q comes from the sectoriality result proved by Shen
[2012].
There is a large number of maximal Lp -regularity (or even H ∞ -calculus)
results for operators of Stokes type arising in fluid dynamics. Some of the
recent ones include Choudhury, Hussein, and Tolksdorf [2018], Giga, Gries,
Hieber, Hussein, and Kashiwabara [2017], Hieber and Prüss [2020], Prüss
[2018], Shibata [2020], Shibata and Shimizu [2008], Simonett and Wilke
[2022a], Tolksdorf [2018, 2020], Tolksdorf and Watanabe [2020], Watanabe
[2022, 2023]. Applications of maximal Lp -regularity to equations of Navier–
Stokes type are too numerous to list here, but some are mentioned in Chapter
18.

Inhomogeneous boundary conditions

A Fourier analytic approach to maximal Lp -regularity for inhomogeneous

boundary value problems in Lq was developed by Weidemaier [2002], Denk,
678 17 Maximal regularity

Hieber, and Prüss [2007]. Extensions to boundary conditions of relaxation type

were obtained in Denk, Prüss, and Zacher [2008a] in the case p = q. Weight in
time were added in Meyries and Schnaubelt [2012a], and the spatial weights
dist(·, ∂Ω)γ were added in Lindemulder [2020]. A further generalisation was
obtained in Hummel and Lindemulder [2022], where the solution space is taken
as Lp (0, T ; As,q ) and As,q is a weighted Bessel potential, Besov, or Triebel–
Lizorkin space and the smoothness s > 2 is restricted by the order of the
boundary operators, and the integrability parameter q is in (1, ∞). Weights
in space have influence on which regularity is needed on the boundary data in
a similar way as weights in time do. A maximal Lp -regularity theory for the
heat equation with inhomogeneous boundary conditions and power weights
outside the Ap -range was obtained by Lindemulder and Veraar [2020]. It is
not well-understood which weights one can take for more general equations.
Another type of generalisation was obtained by Dong and Gallarati [2018,
2020] who consider the setting of Denk, Hieber, and Prüss [2007] but only
require VMO conditions in time and space.
Maximal Lp -regularity on R vs. R+
In the study of parabolic PDE the equation u0 + Au = f is often considered
on the full real line R instead of R+ ; see for instance the monograph Krylov
[2008] and the papers Dong and Kim [2011, 2018].
For problems in which A depends on time it is often easier to first deal
with the full real line, since then Fourier analytic techniques are available.
Such techniques also plays a role in Auscher and Egert [2016, 2023], Dier and
Zacher [2017] in the variational setting of Lions where the operator A is given
by a time-dependent bilinear form.
For evolution equation with inhomogeneous boundary conditions, it is of-
ten simpler to first deal with the (inhomogeneous) boundary conditions in
space, without having to deal with compatibility conditions coming from the
initial conditions by working on the time line R. The initial conditions are
then introduced in the last step. This approach is for instance followed in
Lindemulder and Veraar [2020] for the heat equation, and in Hummel and
Lindemulder [2022] for elliptic systems under Lopatinskii-Shapiro conditions.
Theorem 17.3.29 is due to Mielke [1987]. The equivalence between max-
imal Lp -regularity on R and R+ is due to Hummel and Lindemulder [2022].
Proposition 17.3.31, Theorem 17.3.32, and Proposition 17.3.33 are extensions
of their results. It would be interesting to find versions of these results for time-
dependent operators A. Proposition 17.3.34 and Theorem 17.3.38 appear in
Arendt and Duelli [2006], where also applications to quasilinear equations are
given. Proposition 17.3.37 and Theorem 17.3.35 are standard variants of their
corresponding results on R+ .
Maximal regularity on interpolation spaces
Theorem 17.3.19 is based on the ideas in Kalton and Kucherenko [2010]. With
a different proof, Corollary 17.3.20 is due to Da Prato and Grisvard [1975],
 17.5 Notes 679

where the result is stated for analytic semigroup which are not necessarily
continuous. A direct proof for 1 6 p < ∞ can be found in Prüss and Si-
monett [2016]. The proof for p = ∞ is taken from Lunardi [1995]. Examples
of operators with maximal L1 -regularity on real interpolation spaces of the
form (X, D(A))θ,1 can be obtained from Corollary 17.3.20 or by more direct
arguments. This often leads to maximal regularity results for operators acting
s s
the Besov spaces Bq,1 or their homogeneous counterparts Ḃq,1 . These spaces
appear naturally in fluid dynamics; see, e.g., Danchin, Hieber, Mucha, and
Tolksdorf [2020], Ogawa and Shimizu [2016, 2021], Xu [2022], and references
therein.
Our proof of Corollary 17.3.20 uses the fact that Lp (I; (X, D(A))θ,p ) equals
the real interpolation space (Lp (I; X), Lp (I; D(A)))θ,p up to an equivalent
norm. Alternative results can be obtained if one replaces Lp (I; X) by a real
s
interpolation space such as the Besov space Bp,q (I; X), or even by other spaces
s
such as the Triebel–Lizorkin spaces Fp,q (I; X). To prove maximal regularity
results in such spaces at least two methods are available which will be dis-
cussed briefly. We will only provide the details in the case I = R in order to
avoid difficulties with extension operators and initial value conditions.
The first method is due to Kalton and Kucherenko [2010], and is simi-
lar as in Corollary 17.3.20. However, this time the absolute calculus is used
for Du = u0 instead of A. It follows from Proposition 14.4.17 that D is
s s+1
a closed operator on Bp,q (R; X) with domain Bp,q (R; X) for any s ∈ R,
p, q ∈ [1, ∞]. An application of the Mihlin multiplier theorem (Theorem
14.4.16) for Besov spaces shows that for all µ > 0 the operator µ + D is an
invertible sectorial operator of angle 6 π/2. Moreover, by Theorem 14.4.31,
s s0 s1
Bp,q (R; X) = (Bp,q (R; X), Bp,q (R; X))θ,q if (1 − θ)s0 + θs1 = s. Therefore,
Theorem 16.3.14 implies that D has an absolute function calculus of an-
gle 6 π/2. From Theorem 16.3.20 we see that, at least for p, q < ∞, for
any densely defined sectorial operator A such that A − µ is also sectorial,
the operator A + D = (A − µ) + (D + µ) is invertible as an operator from
s s+1
Bp,q (R; D(A)) ∩ Bp,q (R; X). This implies a maximal regularity result for A
s
in Bp,q (R; X).
The second method is due to Amann [1997], Girardi and Weis [2003a], and
Bu and Kim [2005], and uses a Fourier multiplier Mihlin’s multiplier theorem
for Besov and Triebel Lizorkin spaces (Theorems 14.4.16 and 14.6.11 respec-
tively). Indeed, one can repeat the argument of Theorem 17.3.1(2), except
that one needs 0 ∈ %(A) to avoid problems with the Mihlin condition at zero
in Theorems 14.4.16 and 14.6.11. An extension of these results to Besov–Orlicz
spaces was recently given in Ondreját and Veraar [2020], where it was used
to study temporal regularity of stochastic convolutions.

Section 17.4

The special case of λ = 0 and α = 1 of Theorem 17.4.1 is due to Coulhon and

Lamberton [1986], where maximal Lp -regularity of (−∆)1/2 on Lp (R; X) was
680 17 Maximal regularity

characterised in terms of the UMD property for X. A quantitative version

appears in Hytönen [2015], where it was shown that
1 R reg R
max{βp,X , ~p,X } 6 Mp,(−∆) 1/2 (R+ ) 6 βp,X + ~p,X . (17.77)
2
R
Here, βp,X denotes the (real) UMD constant of X and ~p,X denotes the norm
of the Hilbert transform. In Theorem 17.4.1 we obtained the lower bound
1 R reg
2 βp,X 6 Mp,λ+(−∆)α/2 (R+ ) for all dimensions and all λ > 0 and α > 0 as in
the statement of the theorem. It is unknown whether the second estimate of
(17.77) extends to this setting. For this one would need an extension of the
results in Geiss, Montgomery-Smith, and Saksman [2010] to multipliers m :
Rd+1 → C which satisfy the anisotropic homogeneity condition m(tη, t1/α ξ) =
m(η, ξ) for η ∈ R, ξ ∈ Rd and t > 0. Dimension free upper estimates for
reg
Mp,−∆ (R+ ) can be obtained from Krylov and Priola [2017] in the scalar case,
which provides another potential method to extend (17.77) to other values of
α.
For non-UMD spaces X, Theorem 17.4.1 provides examples of sectorial
operators without maximal Lp -regularity. For Banach function spaces the fol-
lowing improvement holds:

Theorem 17.5.2 (Kalton, Lorist, and Weis [2023]). For an order con-
tinuous Banach function space X the following assertions are equivalent:
(1) {λ(λ − ∆)−1 : λ > 0} is R-bounded on Lp (Rd ; X);
(2) X is a UMD space.

Combining their proof with the dilation argument in Theorem 17.4.1, one can
show that the same equivalence holds when the negative Laplacian −∆ is
replaced by A = λ + (−∆)α/2 . Notice that condition (1) is equivalent to R-
sectoriality. It is unknown whether Theorem 17.5.2 holds for general Banach
spaces X.

Baillon’s result on maximal C-regularity

Theorem 17.4.4 is the main result of Baillon [1980]. Our presentation combines
ideas of Baillon [1980] and Eberhardt and Greiner [1992] (in the latter paper
we do not understand how the closed graph theorem is applied, for the space
of piecewise continuous functions is not complete). Example 17.4.6(1), which
shows that the conclusion of Baillon’s theorem fails for c0 , is taken from Dore
[2000] and seems to be an example due to Kato (See the MathSciNet review of
the paper Baillon [1980]). Corollary 17.4.5 and Examples 17.4.6(2) and 17.4.7
are due to Guerre-Delabrière [1995].
Before the statement of Corollary 17.4.5, we mentioned that the duals of
L∞ (S) and certain C(K)-spaces have cotype 2. This follows from the fact
that these duals are abstract L1 -spaces and thus isometric to an L1 -space, see
 17.5 Notes 681

Lindenstrauss and Tzafriri [1979, Theorem 1.b.2 and the proof of Theorem
1.b.6].
A simple proof of the result, due to Sobczyk [1941], that isomorphic copies
of c0 in separable Banach spaces are always complemented can be found in
Lindenstrauss and Tzafriri [1977, Theorem 2.f.5]. That X contains a comple-
mented copy of `1 if X ∗ contains a copy of c0 is due to Bessaga and Pelczyński
[1958]; see also Lindenstrauss and Tzafriri [1977, Proposition 2.e.8]. Variations
of Baillon’s theorem and connections to control theory can be found in Jacob,
Schwenninger, and Wintermayr [2022], where it is shown that if −A generates
an analytic C0 -semigroup S, then the boundedness and well-definedness of the
operator ΦT : L∞ (0, T ; X) → X given by ΦT (f ) = AS ∗ f (T ) characterises
boundedness of A. This shows that one cannot replace the essential supremum
by a supremum in the definition of maximal L∞ -regularity. From the proof of
Corollary 17.3.20 for p = ∞ one sees that in certain situations ΦT is bounded
when S is not strongly continuous on (X, D(A))θ,∞ .

Counterexamples to maximal Lp -regularity

After partial results by several authors, Brezis’s question was finally solved
to the negative in Kalton and Lancien [2000], who constructed counterexam-
ples in separable Banach lattices not isomorphic to a Hilbert space, and in
particular in Lq -spaces with 1 < q < ∞, q 6= 2. This paper implies part (1)
of Theorem 17.4.8. More general counterexamples were constructed in Kalton
and Lancien [2002]. The construction presented here is due to Fackler [2014,
2016] and has the additional merits of being explicit and solving to the nega-
tive the extrapolation problem for maximal Lp -regularity; in particular, this
work implies part (2) of Theorem 17.4.8. Our proof of Theorem 17.4.11 is a
simplified version of the proof in Fackler [2016]. The result in this paper is
stated with ‘non-homogeneous’ replaced by ‘non-symmetric’, and then invokes
a result from Singer [1970] to reduce matters to the non-homogeneous setting
up to a permutation of one of the two subsequences. To deduce from this
the isomorphic characterisation of `2 mentioned after the statement of The-
orem 17.4.11, one uses a result of Lindenstrauss and Zippin [1969] (see also
McArthur [1972, Theorem 7.6], Fackler [2016, Proposition 5.5]) which states
that if X is a Banach space with an unconditional basis, and if X is not isomor-
phic to c0 , `1 , or `2 , then X has a normalised unconditional, non-symmetric
basis. The spaces c0 and `1 can be excluded, because on these spaces it is
possible to give simple direct constructions of sectorial operators that are
not R-sectorial (Kalton and Lancien [2000], Fackler [2016, Propositions 4.2
and 4.3]). For more on the theory of unconditional bases in Banach spaces
the reader may consult Lindenstrauss and Tzafriri [1977], Singer [1970]. Our
presentation of Proposition 17.4.12 follows Lindenstrauss and Tzafriri [1977].
In connection with Theorem 17.4.11 it is of interest to observe that the
diagonal operator A featuring in the proof is R-sectorial (respectively, almost
R-sectorial) if and only if the partial sum projections (respectively, the coor-
682 17 Maximal regularity

dinate projections) of the basis on which it acts are R-bounded; see Kalton,
Lorist, and Weis [2023, Proposition 6.1.3].
Several counterexamples to maximal Lp -regularity were presented by
Le Merdy [1999] in case X = L1 (T) and X = C(T). These are connected to
Examples 17.4.2 and 17.4.3 on L1 (Rd ) and C0 (Rd ) which are motivated by this
paper. The mentioned paper also contains a counterexample in X = K (`2 ),
the space of compact operators on `2 , of a densely defined invertible secto-
rial operator A of angle ω(A) < 21 π admitting bounded imaginary powers
with kAit k = 1 for all t ∈ R but without maximal Lp -regularity on bounded
intervals for any 1 < p < ∞. Some further counterexamples and standard
constructions for sectorial and Ritt operators can be found in Arnold and
Le Merdy [2019].
Abstract results showing that a large class of weakly compact C0 -semi-
groups on C(K) and L1 (Ω) fail to be R-bounded (and thus do not have max-
imal Lp -regularity) can be found in Hoffmann, Kalton, and Kucherenko [2004].
Further extension are given in Kalton and Kucherenko [2008], Kucherenko and
Weis [2005].

Maximal Lp -regularity for non-autonomous equations

Linear non-autonomous evolution equations are of the form

(
u0 (t) + A(t)u(t) = f (t), t ∈ I,
u(0) = x,

where (A(t))t∈I is a family of unbounded operators on X. One can define

maximal Lp -regularity in a similar way as in the autonomous case. The the-
ory distinguishes between the setting where the domains D(A(t)) are constant
in time and the setting in which they varies in time. Constant domains typi-
cally appear for equations in non-divergence form without boundary, or with
Dirichlet boundary conditions. The varying domain case arises in the case of
time-dependent boundary conditions and is much harder to treat. We will
discuss these cases separately.

Constant domains

Let X0 and X1 be Banach spaces with a continuous and dense embedding

X1 ,→ X0 , and suppose that A : I → L (X1 , X0 ) is a mapping with the
property that the operators A(t) are sectorial of angle ω(A(t)) < π/2, with
uniform estimates with respect to t ∈ I. There are several approaches to
evolution equations governed by time-dependent operators of this form. Stan-
dard references are Amann [1995], Lunardi [1995], Tanabe [1979], where it is
assumed that A ∈ C α (I; L (X1 , X0 )) for some α > 0.
Maximal Lp -regularity for continuous time-dependent operators A : I →
L (X1 , X0 ) was studied in details by Prüss and Schnaubelt [2001], where also
 17.5 Notes 683

the corresponding evolution family generated by (A(t))t∈I is studied. Theorem

17.2.51 can be seen as the weighted analogue of their result. on maximal Lp -
regularity. They also show that the maximal Lp -regularity of A(t0 ) for each
fixed t0 ∈ I is necessary for maximal Lp -regularity for the time-dependent
case. The continuity condition on A was further weakened in Amann [2004]
and Arendt, Chill, Fornaro, and Poupaud [2007].
It is an open problem to find a characterisation (or a sharp sufficient con-
dition) on A of maximal Lp -regularity in the case A is only measurable in
time (in the uniform operator topology). There are counterexamples avail-
able in the literature, which show the limits of possible general theory. One
counterexample due to Pierre and Schmitt [1997] goes as follows:

0
u (t, x) = a(t, x)∆u(t, x) + f (t, x), t ∈ [0, T ], x ∈ B,

u(t, x) = 0 t ∈ [0, T ], x ∈ ∂B, (17.78)

u(0) = 0,


where B is the unit ball in Rd . It follows from their results that there exist
constants m, M > 0 and a measurable a : [0, T ] × Ω → R with m 6 a 6 M
on [0, T ] × B, such that for q > 1 close to 1 there is no maximal Lp regularity
for (17.78) on X0 = Lq (B). Another counterexample for a similar equations
but on Ω = R, can be obtained from Krylov [2016], where also the range of q
is analysed for which maximal Lq -regularity in Lq (R) does hold.
For differential operators with coefficients that are measurable in time and
VMO in space (with uniform estimates in time), Krylov [2008, Theorem 4.3.7
and Chapter 7] shows that the usual conditions such as uniform ellipticity
are sufficient for maximal Lp -regularity in Lq (Rd ) with p > q. The proof is
based on several sophisticated maximal function techniques. By a duality ar-
gument, the condition p > q can be removed in case of constant coefficients in
space. Generalisations to higher order elliptic systems (and certain boundary
conditions) have been obtained in Dong and Kim [2011] for p = q.
A more abstract operator-theoretic approach was taken by Gallarati and
Veraar [2017a,b], where arbitrary p, q ∈ (1, ∞) were considered and certain
R-boundedness conditions on the evolution family and commutator condi-
tions were assumed. In the special case of elliptic operators with coefficients
that are measurable in time, they use their abstract results to prove maximal
Lp -regularity with Ap -weights in time (and possibly also in space) in Lq (Rd ).
Afterwards they use the method of “freezing the coefficients” to extend to co-
efficients which are measurable in time and continuous in space (uniformly in
time). Using Rubio de Francia’s extrapolation theory (see Theorem J.2.1) they
then obtain maximal Lp -regularity on Lq (Rd ) for all p, q ∈ (1, ∞). Dong and
Kim [2018] unified the above results and proved estimates involving parabolic
weights and coefficients which are measurable in time and VMO in space. In
Dong and Krylov [2019] and Krylov [2020], Rubio de Francia’s extrapolation
theory was used to obtain regularity estimates in mixed Lp -norms for fully
684 17 Maximal regularity

non-linear elliptic and parabolic equations with measurable dependence in

time and VMO in space.

Non-constant domains and the Acquistapace–Terreni conditions

Standard references for parabolic equations with time-dependent domains are

Acquistapace and Terreni [1987, 1992], Yagi [1991, 1990] and the monographs
Amann [1995], Tanabe [1979, 1997], Yagi [2010]. An extensive theory of max-
imal C α -regularity (weighted and unweighted) has been developed under the
conditions introduced in Acquistapace and Terreni [1987]. A special case of
their assumptions read as follows:
• there exists a ω ∈ (0, π/2) such that every A(t) is ω-sectorial, with uniform
estimates in t ∈ I, and 0 ∈ %(A);
• there exist ω ∈ (0, π/2), L > 0, and µ, ν ∈ (0, 1] with µ + ν > 1, such that
for all ω 6 | arg(λ)| 6 π one has

|λ|ν kA(t)R(λ, A(t))(A(t)−1 − A(s)−1 )k 6 L|t − s|µ .

The second condition can be understood as a µ-Hölder continuity condition.

Under the above conditions, Hieber and Monniaux [2000a,b] derive max-
imal Lp -regularity via kernel bounds, and Štrkalj [2000], Portal and Štrkalj
[2006] derive maximal Lp -regularity via R-sectoriality. Extensions to maxi-
mal Lp -regularity on R are discussed in Di Giorgio, Lunardi, and Schnaubelt
[2005b].
Fackler [2018] proved maximal Lp -regularity in case under fractional
Sobolev conditions on A, assuming a further technical condition on cer-
tain intermediate spaces between X and D(A(t)). When specialised to the
Hilbert space setting, this paper provides sufficient conditions for maximal
L2 -regularity, which is related to Lions’s problem for non-autonomous evolu-
tion equations (see Lions [1961, p. 68]). The reader is referred to the survey
Arendt, Dier, and Fackler [2017] for details on recent progress on this problem,
and for remaining open questions.
Maximal Lp -regularity results for elliptic operators in divergence form un-
der mixed smoothness conditions in space and time can be found Dier and
Zacher [2017] for X = L2 , and more recently in Bechtel and Gabel [2022] for
X = Lq .

Maximal γ-regularity

In this paragraph we briefly discuss the Gaussian counterpart of maximal Lp -

regularity, namely, maximal γ-regularity, which was introduced and studied in
Van Neerven, Veraar, and Weis [2015a]. It turns out that, in any Banach space,
a densely defined sectorial operator A has maximal γ-regularity if and only
it is γ-sectorial. Combining this result with Theorem 17.5.4, as an immediate
corollary one obtains that, in UMD Banach spaces, the notions of maximal
 17.5 Notes 685

Lp -regularity and maximal γ-regularity are equivalent. This points the way
to using maximal regularity techniques beyond the UMD setting.
In what follows, the reader is assumed to be familiar with the basic theory
of γ-radonifying operators as presented in Section 9.1, whose notation and
terminology we follow here. As in the definition of maximal Lp -regularity, our
starting point is the inhomogeneous abstract Cauchy problem
(
u0 (t) + Au(t) = f (t), t ∈ I,
(ACP0 )
u(0) = 0,

where either I = (0, T ) or I = (0, ∞) = R+ .

Definition 17.5.3 (Maximal γ-regularity). A linear operator A has max-

imal γ-regularity on I if for every f ∈ Cc1 (I) ⊗ D(A) there exists a unique
strongly measurable function u : I → X with the following properties:
(i) u takes values in D(A) almost everywhere and Au belongs to γ(L2 (I); X);
(ii) u solves the integrated version of (ACP0 ), i.e., for almost all t ∈ I we
have
Z t Z t
u(t) + Au(s) ds = f (s) ds.
0 0

(iii) we have the estimate

kAukγ(L2 (I),X) 6 Ckf kγ(L2 (I);X) ,

A strongly measurable function u : I → X is called an γ-solution of (ACP0 )

associated with a given f ∈ γ(L2 (I), X) if it satisfies conditions (i) and (ii).

In (ii), for an operator f ∈ γ(L2 (I); X) and a measurable subset F ⊆ I of

finite measure we define
Z
f dµ := f (1F ).
F

Note that
Z t
f (r) dr 6 (t − s)1/2 kf kγ(a,b;X) , a 6 s 6 t 6 b.
s
Rt
It follows that t 7→ a
f (s) ds ∈ C([a, b]; X) and
Z t
t 7→ f (s) ds 6 (b − a)1/2 kf kγ(a,b;X) .
a C([a,b];X)

Rt
It follows from Theorem 9.6.1 that t 7→ a f (s) ds also belongs to γ(a, b; X).
Indeed, it is trivial to check that for functions f ∈ L2 (a, b) ⊗ X this mapping
686 17 Maximal regularity

coincides with the γ-extension of the indefinite integral, viewed as a bounded

operator from L2 (a, b) into itself. Since L2 (a, b) ⊗ X is dense in γ(a, b; X) this
proves the claim. The theorem also gives the norm estimate
Z t
t 7→ f (s) ds 6 (b − a)1/2 kf kγ(a,b;X) ,
a γ(a,b;X)

observing that the indefinite integral, as an operator on L2 (a, b), has norm
(b − a)1/2 .
In analogy with Theorem 17.2.15, whose proof can be repeated almost
verbatim, we have:

Theorem 17.5.4. If A is a closed linear operator with maximal γ-regularity

on (0, T ), then −A generates an analytic C0 -semigroup. If A has maximal
γ-regularity on R+ , this semigroup is bounded.

Let A be a densely defined γ-sectorial operator, and let (S(t))t>0 be the

bounded analytic C0 -semigroup generated by −A. By the preceding remarks,
for any f ∈ γ(L2 (R+ ); X) we may define the mild solution to (ACP0 ) by
Z t
uf (t) := S(t − s)f (s) ds, t > 0, t ∈ I.
0

Theorem 17.5.5 (Maximal γ-regularity for the problem (ACP0 )). Let
A be densely defined γ-sectorial on a Banach space X with ωγ (A) < 21 π, and
let (S(t))t>0 be the semigroup on X generated by −A. Then A has maximal
γ-regularity on I if and only if for every f ∈ γ(L2 (I), X) the mild solution
uf takes values in D(A) almost everywhere and Au ∈ γ(L2 (I); X). In this
situation, mild solutions and γ-solutions agree.

The following theorem shows that the UMD assumption of Theorem 17.3.1
can be lifted if maximal Lp -regularity is replaced by maximal γ-regularity.

Theorem 17.5.6 (Characterising maximal γ-regularity). Let A be a

densely defined sectorial operator on a Banach space X. Then:
(1) if A is γ-sectorial and ωγ (A) < 21 π, then A has maximal γ-regularity on R+ ;
(2) if A has maximal γ-regularity on R+ , then A is γ-sectorial;
(3) if X is a UMD Banach space and 0 ∈ %(A), then A has maximal γ-
regularity on R+ if and only if A has maximal Lp -regularity for some/all
p ∈ (1, ∞).

The final result gives a sufficient condition for pointwise regularity of γ-

solutions with a uniform bound in time.

Theorem 17.5.7. Suppose that A is a densely defined γ-sectorial operator

with a bounded H ∞ -calculus of angle ωH ∞ (A) < 21 π on a Banach space X
with finite cotype. If 0 ∈ %(A), then for all f ∈ γ(L2 (R+ ), X) the associated
 17.5 Notes 687

γ-solution u(t) takes values in D(A1/2 ) for all t > 0, the resulting function
uf : R+ → D(A1/2 ) is uniformly continuous and satisfies

sup ku(t)kD(A1/2 )) 6 Ckf kγ(L2 (R+ ),X) ,

t∈R+

for some constant C independent of f .

Theorem 17.5.7 indicates an important difference with the theory of maximal
Lp -regularity consists in terms of the trace space involved. Whereas maxi-
mal Lp -regularity allows for the treatment of non-linear problems with ini-
tial values in the space real interpolation space (X, D(A))1− p1 ,p , in the pres-
ence of maximal γ-regularity initial values in the complex interpolation space
[X, D(A)] 21 can be allowed. For a fuller discussion of this point the reader is
referred to Van Neerven, Veraar, and Weis [2015a].

Reversing the integrability in time and space

In the special case that X = Lq (S), maximal γ-regularity is equivalent to

maximal regularity estimates in the spaces Lq (S; L2 (I)). Thus compared to
the results of this chapter the order of time and space are interchanged. In
the special setting of X = Lq (S), one can even study maximal regularity
estimates in the spaces Lq (S; Lp (I)). A detailed study of the latter is given in
Antoni [2017]. In some cases one can also obtain such bounds via Lp (S; Lp (I))-
estimates if one can add Ap -weights in S, and afterwards apply Rubio de
Francia’s extrapolation theory (see Theorem J.2.1).
Another subtle way to interchange time and space is to formulate maximal
regularity in so-called tent spaces. The definition and some historical details on
these spaces can be found in the notes of Chapter 10. For details on maximal
regularity in tent spaces the reader is referred to Auscher, Monniaux, and
Portal [2012b], Auscher, Kriegler, Monniaux, and Portal [2012a], Auscher,
Monniaux, and Portal [2019].

Miscellaneous topics

The discrete-time version of maximal Lp -regularity, usually referred to as

maximal `p -regularity, was studied in detail in Blunck [2001a,b]. In numer-
ical analysis, it can be used to show stability and convergence of numerical
schemes. There is a large body of subsequent work on this topic; the reader
is referred to Ashyralyev, Piskarev, and Weis [2002], Akrivis, Li, and Lubich
[2017], Blunck and Kunstmann [2002], Geissert [2006], Kalton and Portal
[2008], Kemmochi [2016], Kovács, Li, and Lubich [2016], Lizama [2015], Por-
tal [2003, 2005], and references therein. In some of these works, connections
with the theory of Ritt operators play a role. A discussion on these operators
is contained in the Notes of Chapter 10.
Maximal Lp -regularity for evolution equations which are of second order
equations in time, are introduced from a modern point of view in Chill and
688 17 Maximal regularity

Srivastava [2005]. In this paper, among other things, characterisations of max-

imal Lp -regularity in terms of Fourier multipliers are given and applications
to quasi-linear evolution equations are presented. Subsequent work on second
order equations includes Arendt, Chill, Fornaro, and Poupaud [2007], Batty,
Chill, and Srivastava [2008], Chill and Srivastava [2008], Denk and Schnaubelt
[2015], Poblete [2009].
Parabolic mixed order systems do not fall within the setting of analytic
semigroup theory. Maximal Lp -regularity for such types of systems can still
be developed via operator-valued Fourier multiplier theory; this was done by
Denk, Saal, and Seiler [2008b], Denk and Volevich [2008], Denk and Kaip
[2013].
In Zacher [2005], R-sectoriality was used to prove maximal Lp -regularity
for Volterra equations. The weighted setting in time was subsequently intro-
duced in Prüss [2019].
18
Nonlinear parabolic evolution equations in
critical spaces

As we have seen in the preceding sections, in the context of inhomogeneous

linear evolution equations, maximal regularity enables one to set up an iso-
morphism between the space of data (initial value and inhomogeneity) and
the solution space. In the present section we will show how this idea can be
used to study to non-linear evolution equations. Specifically, we consider a
class of quasi-linear evolution equations of the form
(
u0 (t) + A(u(t))u(t) = F (u(t)), t ∈ (0, T ),
u(0) = u0 .

The setting is as follows. We are given a pair of Banach spaces (X0 , X1 )

along with a continuous embedding X1 ,→ X0 . The initial value u0 is taken
in a suitable interpolation space of X0 and X1 , and for each v0 in some
neighbourhood Y of u0 the operator A(v0 ) is a linear operator in X0 with
domain D(A(v0 )) = X1 . As such, we interpret A(v0 ) as a bounded linear
operator in L (X1 , X0 ). The mapping F is defined on an interpolation space
of X0 and X1 , takes values in X0 , and is assumed to satisfy suitable local
Lipschitz conditions; the precise assumptions will be formulated later. Our
aim is to present several local well-posedness results, and to discuss a blow-up
criterion which can be used to derive global well-posedness.
Before we start with this programme, we first explain the difference be-
tween semi-linear and quasi-linear evolution equations. In the quasi-linear
case, the typical situation is that

A : Y → L (X1 , X0 ) and F : Z → X0

are Lipschitz continuous on bounded subsets of Y and Z, where Y and Z are

(subsets of) suitable interpolation spaces between X0 and X1 . In the semi-
linear case one typically has that

A ∈ L (X1 , X0 ) and F : Z → X0 ,

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 689
T. Hytönen et al., Analysis in Banach Spaces, Ergebnisse der Mathematik und ihrer
Grenzgebiete. 3. Folge / A Series of Modern Surveys in Mathematics 76,
https://doi.org/10.1007/978-3-031-46598-7_8
690 18 Nonlinear parabolic evolution equations in critical spaces

where A is a fixed operator in L (X1 , X0 ) and F is a locally Lipschitz contin-

uous mapping on bounded subsets of Z, where Z is as before. Clearly, every
semi-linear equation is quasi-linear, but the converse is not true. In principle
one can allow A and F to be also time-dependent, but in order to keep the
presentation within reasonable bound we will not consider this additional level
of generality.
When analysing evolution equations with maximal Lp -regularity methods,
one usually takes Y equal to (or a subset of) the real interpolation space
(X0 , X1 )1− p1 ,p , at least in the absence of weights. The reason for taking Y of
this form is that one has a continuous embedding (see Corollary L.4.6)

Lp (0, T ; X1 ) ∩ W 1,p (0, T ; X0 ) ,→ C([0, T ]; (X0 , X1 )1− p1 ,p ). (18.1)

The space on the left-hand side is the usual space in which solutions lie when
maximal Lp -regularity techniques are applicable. For the space Z one can take
either take Y , or more generally (X0 , X1 )β,1 with β ∈ [1 − p1 , 1), the latter
requiring polynomial growth restrictions on F . In practice, we often split F
into two parts F = FTr + Fc , where

FTr : Y → X0 , Fc : Z → X0 (18.2)

with Y and Z as before. Here the subscript Tr stands for trace space and c
stands for critical. The word critical is also used in the title of the chapter.
In Section 18.2 we will give a definition of a criticality using only evolution
equation terminology. Surprisingly, this often coincides with criticality from a
PDE perspective.
The following simple example explains why the additional flexibility in
choosing Z may be expected to be useful.
Example 18.0.1. On Rd consider the equation
(
∂t u − a(u)∆u = −u3 ,
(18.3)
u(0) = u0 ,

where a : R → [0, ∞) is a given locally Lipschitz continuous function. If a is

non-constant, then (18.3) leads to a quasi-linear evolution equation, and if a is
constant it leads to a semi-linear evolution equation. In both cases, the spaces
X0 and X1 need to be chosen as function spaces relative to which the defini-
tions A(u)v := a(u)∆v and F (u) := −u3 admit meaningful interpretations. A
possible choice is to take

X0 := Lq (Rd ), X1 := W 2,q (Rd ).

2− 2
With these choices, Y := (X0 , X1 )1− p1 ,p equals the Besov space Bq,p p (Rd )
2 d
(see Theorem 14.4.31). If we assume that 2 − p − q > 0, then we have the
continuous embedding (see Corollary 14.4.27)
 18 Nonlinear parabolic evolution equations in critical spaces 691
2− 2
Bq,p p (Rd ) ,→ Cb (Rd ).

The space Y then consists of bounded continuous functions, and consequently

for u ∈ Y we can interpret a(u) as a bounded continuous function. The op-
erator A(u) is then well defined as an element of L (X1 , X0 ). On the other
hand, since F (u) = −u3 belongs to X0 if and only if u ∈ L3q (Rd ), for u ∈ Y
we can interpret F (u) as an element of X0 as soon as we have a continuous
embedding
2− 2
Bq,p p (Rd ) ,→ L3q (Rd ).
This embedding holds under the (much weaker) condition 2 − p2 − dq > − 3q d
,
where even equality is allowed if p 6 3q (see (14.22), Proposition 14.6.13, and
Theorem 14.6.26). To optimally exploit this fact in situations where the more
stringent condition 2 − p2 − dq > 0 mentioned earlier is not needed, e.g., in the
semi-linear case arising when a ≡ 1, we may admit functions F defined on a
space Z that is larger than Y . Even more flexibility is created if we take time
integrability into account as by maximal Lp -regularity methods we actually
only need

W 1,p (0, T ; X0 ) ∩ Lp (0, T ; X1 ) ,→ L3p (0, T ; L3q (Rd ))

in order to define F (u). Conditions for this are given by Corollary L.4.7 which
2
in the current situation α = 0, h = 3 and thus θ = 1 − 3p lead to the
2θ,q 3q d 2
requirement H ,→ L , which holds if and only if q + p 6 3, which is even
weaker than what we saw before. We will come back to this point in Examples
18.1.3 and 18.3.1.
In Section 18.1 we start with the study of local existence and uniqueness
for semi-linear equations, where the function F is defined on the trace space
Y = (X0 , X1 )1− p1 with p ∈ (1, ∞). Here we can admit initial values u0 which
belong to the space Y . We present this setting separately, as it allows us to
introduce some important techniques in the simplest possible setting.
In Sections 18.2 we turn to the study of local well-posedness in the tech-
nically more demanding quasi-linear setting. At the same time, we improve
on the assumptions needed to make things work: it is possible to allow expo-
nents p ∈ [1, ∞] and functions F of the form FTr + Fc as in (18.2), with FTr
defined on Y as before and Fc on a larger space Z. Furthermore, we work in
a weighted setting in time. This has at least three important advantages:
(i) it allows initial data u0 belonging to the space (X0 , X1 )1−α− p1 , where
α > 0 is a parameter associated with the weight;
(ii) global existence of solutions can be proved under milder blow-up criteria;
(iii) it allows the inclusion of the endpoint p = ∞ (inclusion of the endpoint
p = 1 is possible for different reasons).
Blow-up criteria will be discussed in Section 18.2.d. After presenting an illus-
trating example in Section 18.3, the final Section 18.4 presents long term and
692 18 Nonlinear parabolic evolution equations in critical spaces

even global well-posedness results for small initial data in the case F = Fc
(i.e., FTr = 0).

18.1 Semi-linear evolution equations with F = FTr

In this section we study local well-posedness of semi-linear evolution equations
of the form
(
u0 (t) + Au(t) = F (u(t)), t ∈ (0, T ),
(18.4)
u(0) = u0 ,

where T > 0 is fixed. Parabolic partial differential equations of evolution type

can often be cast into this form. Some examples will be encountered below.
Our standing assumptions are as follows. We let X0 and X1 be Banach
spaces, with X0 continuously embedded into X0 , we fix p ∈ (1, ∞), and make
the following assumptions:
(1) A : X1 → X0 is a bounded linear operator;
(2) F : (X0 , X1 )1− p1 ,p → X0 is a locally Lipschitz function;
(3) u0 belongs to (X0 , X1 )1− p1 ,p .

For the sake of brevity, in what follows we will write

X1− p1 ,p := (X0 , X1 )1− p1 ,p

and refer to this space as the trace space associated with the problem (18.4).
The following definition extends the notion of Lp -solutions to the present
setting.

Definition 18.1.1. A function u ∈ Lp (0, T ; X1 ) ∩ W 1,p (0, T ; X0 ) is called an

Lp -solution to (18.4) if for all t ∈ [0, T ] we have
Z t Z t
u(t) − u0 + Au(s) ds = F (u(s)) ds.
0 0

The assumptions imply that Au belongs to Lp (0, T ; X0 ), and therefore the

first integral is well defined as a Bochner integral in X0 . To prove the Bochner
integrability of F (u) : s 7→ F (u(s)) in X0 , we note that the assumptions and
(18.1) imply that u ∈ C([0, T ]; X1− p1 ,p ). Consequently, F (u) is well defined as
a function in C([0, T ]; X0 ).
In order to get acquainted with the type of arguments involved, we begin
by proving a preliminary local existence and uniqueness result. Later on, in
Section 18.2, this result will be further improved in several ways, and contin-
uous dependence on the initial data and conditions for global well-posedness
will be discussed.
 18.1 Semi-linear evolution equations with F = FTr 693

Theorem 18.1.2 (Local well-posedness for semi-linear problems). Let

X1 ,→ X0 as stated, let p ∈ (1, ∞), and assume the conditions (1), (2), (3) to
be satisfied. Suppose that
(1) The operator A, viewed as a linear operator in X0 with domain D(A) = X1 ,
has maximal Lp -regularity on bounded time intervals;
(2) There exists an non-decreasing function ψ : (0, ∞) → (0, ∞) such that for
all r > 0 and all x, y ∈ X1− p1 ,p satisfying

kxkX1− 1 ,p 6 r and kykX1− 1 ,p 6 r

p p

one has

kF (x) − F (y)kX0 6 ψ(r)kx − ykX1− 1 ,p . (18.5)

p

Then for all R > 0 there exists a T > 0 such that for all u0 ∈ X1− p1 ,p satisfying
ku0 kX1− 1 ,p 6 R the problem (18.4) has a unique Lp -solution u. Moreover,
p

u ∈ Lp (0, T ; X1 ) ∩ W 1,p (0, T ; X0 ).

The bound (18.5) is a quantified local Lipschitz assumption, where “local”

refers to bounded subsets of X1− p1 ,p . We note that by (18.1) the Lp -solution
u satisfies

u ∈ C([0, T ]; X1− p1 ,p ).

As a preparation for the proof, we first explain how the maximal Lp -

regularity of A will be used to prove the theorem. By maximal Lp -regularity,
for all f ∈ Lp (0, T ; X0 ) the problem
(
u0 + Au = f on (0, T )
u(0) = u0

admits a unique Lp -solution. Moreover, there exists a constant Cp,A,T , inde-

pendent of f and u0 , such that

kukLp (0,T ;X1 )∩W 1,p (0,T ;X0 ) 6 Cp,A,T (kf kLp (0,T ;X0 ) + ku0 kX1− 1 ,p ). (18.6)
p

This follows from Proposition 17.2.14 and a repetition of the argument in

(17.10) and (17.11). For the optimal choice of these constants, using (17.25)
one can check that Cp,A,T 6 Cp,A,T 0 whenever T 6 T 0 < ∞.
Proof of Theorem 18.1.2. The theorem will be established by applying the
Banach fixed point theorem on a suitable bounded closed subset of the max-
imal regularity space
694 18 Nonlinear parabolic evolution equations in critical spaces

MRp (0, T ) := Lp (0, T ; X1 ) ∩ W 1,p (0, T ; X0 ).

The boundedness of the embedding MRp (0, T ) ,→ C([0, T ]; (X0 , X1 )1− p1 ,p )

(see (18.1)) enables us to use the local Lipschitz assumption on F .
Let R > 0 and fix an u0 ∈ X1− p1 ,p satisfying ku0 kX1− 1 ,p 6 R. In order to
p
define a suitable subset of MRp (0, T ) on which the fixed point argument can be
performed, we introduce the reference solution zu0 as the unique Lp -solution
to
(
z 0 (t) + Az(t) = 0, t > 0,
z(0) = u0 .

Note that zu0 ∈ MRp (0, T ) for every T < ∞. By (18.1) (and its proof) and
(18.6) (with f = 0) we have

sup kzu0 (t)kX1− 1 ,p 6 Cp,T kzu0 kMRp (0,1)

t∈[0,1] p

6 Cp,T Cp,A,T ku0 kX1− 1 ,p 6 Cp,T Cp,A,T R =: MR .

p

Fix an arbitrary T ∈ (0, 1], and consider the closed ball

B1T (u0 ) := {u ∈ MRp (0, T ) : u(0) = u0 , ku − zu0 kMRp (0,T ) 6 1}.

Let Φ : B1T (u0 ) → MRp (0, T ) be defined by Φ(v) := u, where u is the unique
Lp -solution to the problem
(
u0 + Au = F (v),
u(0) = u0 .

This unique solution exists by the discussion preceding the proof; note that
F (v) ∈ C([0, T ]; X0 ) by the continuity of F and (18.1).
For later purpose we observe that for all v1 , v2 ∈ B1T (u0 ), Corollary L.4.6
(using v1 (0) − v2 (0) = 0 to get T -independent constants) implies that

kv1 − v2 kC([0,T ];X1− 1 ,p ) 6 Cp kv1 − v2 kMRp (0,T ) 6 2Cp . (18.7)

p

In particular, since T 6 1, upon taking v1 = v ∈ B1T (u0 ) and v2 := zu0 ∈

B1T (u0 ), we find that

kvkC([0,T ];X1− 1 ,p ) 6 kv − zu0 kC([0,T ];X1− 1 ,p ) + kzu0 kC([0,T ];X1− 1 ,p )

p p p
(18.8)
6 2Cp + MR =: NR .

To be able to apply the Banach fixed point theorem to Φ, we need to check

that Φ maps the closed ball B1T (u0 ) into itself and is uniformly contractive on
it. For both assertions it will be necessary to choose T ∈ (0, 1] small enough.
 18.1 Semi-linear evolution equations with F = FTr 695

First we check that Φ maps B1T (u0 ) into itself. For all v ∈ B1T (u0 ) one has

kF (v(t))kX0 6 kF (v(t)) − F (zu0 (t))kX0 + kF (zu0 (t)) − F (0)kX0 + kF (0)kX0

6 ψ(NR )(kv(t) − zu0 (t)kX1− 1 ,p + kzu0 (t)kX1− 1 ,p ) + kF (0)kX0
p p

6 ψ(NR )(2Cp + MR ) + kF (0)kX0 =: CR,F ,

where we used (18.5) and (18.8). Thus

kF (v)kLp (0,T ;X0 ) 6 T 1/p CR,F .

Therefore, letting u = Φ(v) and using the maximal Lp -regularity estimate

(18.6) for the equation which u − zu0 satisfies, we find that

ku − zu0 kMRp (0,T ) 6 Cp,A,T kF (v)kLp (0,T ;X0 ) 6 Cp,A,1 T 1/p CR,F .

Therefore, for 0 < T 6 (Cp,A,1 CR,F )−p ∧ 1 we find that u ∈ B1T (u0 ).
To check that Φ is a uniform contraction, let vi ∈ B1T (u0 ) for i ∈ {1, 2}.
Using the maximal Lp -regularity estimate (18.6) for the equation which u1 −u2
satisfies, and (18.5), we find that

kΦ(v1 ) − Φ(v2 )kMRp (0,T ) 6 Cp,A,T kF (v1 ) − F (v2 )kLp (0,T ;X0 )
6 Cp,A,1 T 1/p ψ(NR )kv1 − v2 kC([0,T ];X1− 1 ,p )
p

1/p
6 Cp,A,1 T ψ(NR )Cp kv1 − v2 kMRp (0,T ) ,

where in the last step we used (18.7). Therefore, combining both conditions
on T it follows that for T = 21 ((Cp,A,1 ψ(NR )Cp )−p ∧ (Cp,A,1 CR,F )−p ∧ 1) the
mapping Φ is a uniform contraction on B1T (u0 ) with
1
kΦ(v1 ) − Φ(v2 )kMRp (0,T ) 6 kv1 − v2 kMRp (0,T ) .
2
By the Banach fixed point theorem, the restriction of Φ to B1T (u0 ) has a
unique fixed point u ∈ B1T (u0 ). From the definition of Φ, it is immediate that
u is an Lp -solution to (18.4).
It remains to prove the uniqueness. Uniqueness is clear on B1T (u0 ), but
we still need to prove uniqueness in the larger set MRp (0, T ). Let u1 , u2 ∈
MRp (0, T ) be Lp -solutions to (18.4). Then for every t ∈ [0, T ], by Corollary
L.4.6 (with t-independent constant), (18.6), and the remarks below it, and
(18.5),

ku1 (t) − u2 (t)kX1− 1 ,p 6 Cp ku1 − u2 kMRp (0,t)

p

6 Cp Cp,A,T kF (u1 ) − F (u2 )kLp (0,t;X0 )

6 Cp Cp,A,T ψ(N )ku1 − u2 kLp (0,t;X1− 1 ,p ) ,
p
696 18 Nonlinear parabolic evolution equations in critical spaces

where N is such that kui kC([0,T ];X1− 1 ,p ) 6 N for i ∈ {1, 2}. Therefore, apply-
p
ing Gronwall’s inequality to ku1 (t) − u2 (t)kpX , we find that u1 ≡ u2 on
1− 1 ,p
p
[0, T ]. 
Here is a simple example to which Theorem 18.1.2 can be applied. Further
examples will be given in Section 18.3.
Example 18.1.3. Let A ∈ L (X1 , X0 ), where
X0 = H s,q (Rd ) and X1 = H s+2,q (Rd )
with s ∈ (−2, 0] and q ∈ (1, ∞). In the present situation, Theorem 14.4.31
s+2− 2
shows that X1− p1 ,p = Bq,p p (Rd ). In order to have a concrete equation in
mind note that one for instance could take A to be a second order differential
operator such as −∆, and we could consider the PDE
(
∂t u − ∆u = f (u),
u(0) = u0 ,

where f : R → R is a given locally Lipschitz function satisfying f (0) = 0.

Suppose now that s + 2 − p2 − dq > 0. Then, by the Sobolev embedding in
Corollary 14.4.27, we have a continuous embedding
X1− p1 ,p ,→ Cb (Rd ). (18.9)

We claim that the so-called Nemitskii map

F (u)(x) = f (u(x)), x ∈ Rd ,
is well defined and locally Lipschitz as a mapping from X1− p1 ,p into X0 . To
prove this, fix N > 0 and elements u, v ∈ X1− p1 ,p satisfying kukX1− 1 ,p 6 N
p
and kvkX1− 1 ,p 6 N . Then, we obtain that for some constant L depending on
p
f , N , and the embedding constant of (18.9),
Z 1/q
kF (u) − F (v)kX0 6 |f (u(x)) − f (v(x))|q dx
Rd
Z 1/q
6L |u(x) − v(x)|q dx
Rd
6 LCku − vkX1− 1 ,p ,
p

where in the last step we used (14.22) and Proposition 14.4.18. Taking v = 0
and using the assumption f (0) = 0, one also obtains
kF (u)kX0 6 LCkukX1− 1 ,p .
p

These two estimates prove the claim.

 18.2 Local well-posedness for quasi-linear evolution equations 697

The above estimates on F are not optimal and the condition on the exponents,
namely, s+2− p2 − dq > 0 turns out to be far from sharp. We also notice that in
s+2− 2
the example we can only treat rather smooth initial values u0 ∈ Bq,p p (Rd )
(in particular they need to be Hölder continuous). This turns out to be far
from sharp. Both these sharpness issues will be addressed in the next section.

18.2 Local well-posedness for quasi-linear evolution

equations
In the present section and the next, we will study local well-posedness for
quasi-linear evolution equations of the form introduced at the beginning of
Chapter 18,
(
u0 (t) + A(u(t))u(t) = F (u(t)), t ∈ (0, T ),
u(0) = u0 .

We will make several changes to the simple setting considered in Section 18.1.
Besides the fact that the operator A now depends on the solution u, the
changes are as follows:
• The non-linearity is of the form

F = FTr + Fc ,

where FTr plays a similar role as in Section 18.1, and Fc is the so-called
critical part of F . We assume that both FTr and Fc are defined on a suitable
subset of Xσ,p (see (18.10) below) with σ ∈ [0, 1 − p1 ], and that Fc satisfies
a suitable polynomial growth condition.
• Weights in time are added (see Corollaries 17.2.37 and 17.2.48). This will
enable us to reduce the smoothness conditions on the initial data. At the
same time, this makes it possible to formulate flexible conditions for global
existence.
• The full range p ∈ [1, ∞] will be considered.
In Example 18.3.1 we will see that the new setting takes care of the issues
raised in the discussion after Example 18.1.3.

18.2.a Setting

Turning to the details, as before we make the standing assumption that we

have a continuous embedding of Banach spaces

X1 ,→ X0 .

Without loss of generality we will always assume that the constant in the
embedding is > 1.
698 18 Nonlinear parabolic evolution equations in critical spaces

We further fix
p ∈ [1, ∞]
and
α ∈ [0, p10 ) ∪ {0},

where p1 + p10 = 1; we take α > 0 if p = ∞. The exponent α enters into the

weight wα (t) := tα that will be used later. In applications, the choice of α
determines which initial condition u0 can be allowed; larger values of α permit
initial conditions with less smoothness. The exponent
1
σ := 1 − α −
p
has already been encountered in Corollary 17.2.37, and will occur frequently
in what follows.
For the sake of notational brevity, we will use the conventions that

(X0 , X1 )0,r := X0 for r ∈ [1, ∞],

(X0 ,X1 )θ,r
Xθ,r := X1 for θ ∈ (0, 1) and r ∈ [1, ∞], (18.10)
Xθ := Xθ,1 for θ ∈ (0, 1).

Note that Xθ,r = (X0 , X1 )θ,r if θ ∈ (0, 1) and r ∈ [1, ∞), because in these
ranges X0 ∩ X1 = X1 is dense in (X0 , X1 )θ,r by Corollary C.3.15. For θ = 0,
Xθ,r = (X0 , X1 )θ,r holds for all r ∈ [1, ∞] by definition.
Remark 18.2.1. There is some flexibility with regard to the choice of the spaces
Xθ in (18.10). These spaces will appear only in the assumptions on the non-
linearity Fc through (18.11) below. The only requirement needed is that Xθ,1
continuously embeds into this space. In the above definition one could for
instance take Xθ to be Xθ,r , [X0 , X1 ]θ , or D((ω + A(u0 ))θ ) for ω ∈ R large
enough.
In addition to the above-stated assumptions on the spaces X0 , X1 and the
parameters p, α, σ, we make the following structural assumptions on the
operator A and the non-linearity F .

Assumption 18.2.2. We fix an open set Oσ,p ⊆ Xσ,p and assume:

(1) the initial condition u0 belongs to Oσ,p ;
(2) there exists a constant L > 0 such that the mapping A : Oσ,p → L (X1 , X0 )
satisfies

kA(u) − A(v)kL (X1 ,X0 ) 6 Lku − vkXσ,p , u, v ∈ Oσ,p ;

(3) the mapping F : X1 ∩ Oσ,p → X0 admits a decomposition F = FTr + Fc ,

where
 18.2 Local well-posedness for quasi-linear evolution equations 699

(i) FTr : X1 ∩ Oσ,p → X0 and there exists an LTr > 0 such that

kFTr (u) − FTr (v)kX0 6 LTr ku − vkXσ,p , u, v ∈ X1 ∩ Oσ,p ;

(ii) Fc : X1 ∩ Oσ,p → X0 and there exist m > 1, βj ∈ (σ, 1), ρj > 0 for
j ∈ {1, . . . , m}, and Lc > 0 such that
m
ρ ρ
X
kFc (u) − Fc (v)kX0 6 Lc (1 + kukXjβ + kvkXjβ )ku − vkXβj
j j
j=1
(18.11)

for all u, v ∈ X1 ∩ Oσ,p , and where

1 + ρj σ
βj 6 , j ∈ {1, . . . , m}. (18.12)
1 + ρj

Several clarifying comments are in order.

In typical applications, the set Oσ,p is a bounded subset of Xσ,p . In sit-
uations where A, FTr , and Fc are defined on all of Xσ,p (in case of A and
FTr ) and X1 (in case of Fc ), the constants L, LTr and Lc will increase with
Oσ,p . Thus, although some of the above Lipschitz estimates are formulated
as global Lipschitz conditions on Oσ,p , they should actually be thought of as
local Lipschitz conditions on Xσ,p .
The quasi-linear operator A is Lipschitz on the same space as FTr . In the
semi-linear case the operator A can be taken constant on Oσ,p .
The assumptions on the non-linearity FTr are very similar to the ones in
Theorem 18.1.2 in case α = 0 and p ∈ (1, ∞), but for simplicity we chose
to let FTr be defined on the full space (X0 , X1 )1− p1 ,p in that result. Taking
larger values of α leads to more restrictive conditions on FTr . However, at the
same time it will lead to less conditions on the initial data. The mapping FTr
uniquely extends to a continuous mapping on Oσ,p .
A central role is played by the non-linear mapping Fc , where c stands
for “critical”. Let β = maxj∈{1,...,m} βj . By (18.11) and the density of X1
in Xβ we find that Fc uniquely extends to a locally Lipschitz function Fc :
Xβ
Oσ,p ∩ X1 → X0 .
The restriction (18.12) should be seen as a balance between the polynomial
growth rate ρj of the local Lipschitz constant and the regularity exponent βj .
The larger ρj is, the smaller βj needs to be. The case of equality plays a
special role:

Definition 18.2.3 (Criticality). Let Assumption 18.2.2 hold. The space

Xσ,p and the parameter σ are called critical if equality holds in (18.12)
for some j ∈ {1, . . . , m}. In case of strict inequality in (18.12) for all
j ∈ {1, . . . , m}, the space Xσ,p and the parameter σ are called sub-critical.
700 18 Nonlinear parabolic evolution equations in critical spaces

In applications to concrete non-linear PDEs, the parameters βj and ρj are

determined by spatial Sobolev embedding and the growth order of the poly-
nomial non-linearity. Often, one can choose a minimal σ for which at least
one of the inequalities becomes an equality. After σ has been determined, one
can choose α and p such that σ = 1 − α − p1 holds. Here, p is usually chosen
large (and thus α ∈ [0, 1/p0 ) ∪ {0} is close to 1 − σ), as this leads to the best
time regularity results. Quite often, the critical space (for the initial values)
Xσ,p has some scaling behaviour which fits well to the scaling behaviour of
solutions to the corresponding PDE. If p = ∞, the unweighted case α = 0
cannot be considered due to a technical reason: in the proofs below we need
α + p1 > 0.
In the case p = 1, which is allowed in our setting, the other assumptions
enforce α = 0 and σ = 0. In particular, there is not much flexibility for
the function FTr and it needs to be locally Lipschitz on X0 . On the other
hand, (18.11) is still quite flexible: for instance if m = 1 one can allow Fc :
X1/(1+ρ) → X0 where growth of power ρ is allowed for the Lipschitz constant.

Remark 18.2.4 (Time-dependent and inhomogeneous settings). It is possible

to extend the above setting to time-dependent mappings A : [0, T ] × Oσ,p →
L (X1 , X0 ) and F : [0, T ] × X1 ∩ Oσ,p → X0 . This does not lead to any
major changes as long as the mapping properties of A and F and estimates
are uniform in t ∈ [0, T ] (or the constants in the estimates satisfy a suitable
integrability condition). Usually, A is assumed to be continuous in time, so
that maximal regularity of A(0, u0 ) can be used in local well-posedness results
in a similar way as we did in Theorem 17.2.51. Continuity in time can be
avoided by introducing a suitable notion of maximal regularity for the case of
time-dependent A.
One can also allow a further inhomogeneity by allowing non-linearities of
the form F = FTr + Fc + f , where f : (0, T ) → X0 satisfies appropriate
integrability assumptions.

We will now proceed to the main theorems on local well-posedness for the
quasi-linear problem
(
u0 + A(u)u = F (u), on (0, T ),
(18.13)
u(0) = v0 ,

where v0 ∈ Oσ,p can be taken as the given u0 or close to u0 in Xσ,p -norm.

Allowing v0 to be taken from a neighbourhood of u0 will be important as we
will also give prove continuous dependence on the initial data. Moreover, it
will be used to obtain criteria for global well-posedness.
Define
 p
p Lwα (0, T ; X1 ) ∩ Ww1,p (0, T ; X0 ) if p < ∞;
MRα (0, T ) := α
1
Cwα ,0 ((0, T ]; X1 ) ∩ Cw α ,0
((0, T ]; X 0 ) if p = ∞,
 18.2 Local well-posedness for quasi-linear evolution equations 701

where we recall that the Banach space Cwα ,0 ((0, T ]; X1 ) was defined before
∞
Corollary 17.2.48, and is a closed subspace of Lw α
(0, T ; X1 ). Similar assertions
1
hold for its C -variant.
p
Definition 18.2.5. Let Assumption 18.2.2 hold. A function u ∈ MRα (0, T )
p
is called a Lwα -solution to (18.13) on (0, T ) if u takes values in Oσ,p ,
A(u)u, F (u) ∈ L1 (0, T ; X0 ), and for all t ∈ [0, T ] we have
Z t Z t
u(t) − v0 + A(u(s))u(s) ds = F (u(s)) ds.
0 0

Later on in Lemma 18.2.8, we will see that the integrability assumptions on

A(u)u and F (u) are actually redundant, and that one even has A(u)u, F (u) ∈
p
Lw α
(0, T ; X0 ).

18.2.b Main local well-posedness result

The main result of this section is the following local well-posedness for quasi-
linear equations.
Theorem 18.2.6 (Local well-posedness for quasi-linear problems).
Let Assumption 18.2.2 hold. If, for some u0 ∈ Oσ,p , the operator A(u0 ) has
maximal Lp -regularity (maximal C-regularity if p = ∞) on finite time inter-
vals, then there exist T > 0 and ε > 0 such that for all

v0 ∈ BXσ,p (u0 , ε) ⊆ Oσ,p

the problem (18.13) has a unique Lw p

α
-solution uv0 ∈ MRpα (0, T ). Moreover,
there exists a constant C > 0 such that for all v0 , v1 ∈ BXσ,p (u0 , ε) we have

kuv0 − uv1 kMRpα (0,T ) 6 Ckv0 − v1 kXσ,p . (18.14)

From Corollary L.4.6 we additionally see that

uv0 ∈ C([0, T ]; X1−α− p1 ,p ) ∩ C((0, T ]; X1− p1 ,p ). (18.15)

This shows that for α > 0, the solution u instantaneously (that is, for
t ∈ (0, T ]) regularises from X1−α− p1 ,p to X1− p1 ,p . By similar arguments, an
analogous continuous dependence as in (18.14) holds in C([0, T ]; X1−α− p1 ,p )
and in the weighted space Cwα ((0, T ]; X1− p1 ,p ).
The parameters T , ε, and C in Theorem 18.2.6 depend on the choice
of u0 in general. The parameters T and ε need to be small enough for the
conclusions of the theorem to hold. This has several reasons. First of all, ε
must be small because we need BXσ,p (u0 , ε) to be contained in Oσ,p . More
importantly, the proof uses the maximal regularity of A(u0 ) to obtain local
well-posedness of (18.13) with initial value v0 , via a perturbation argument
involving the smallness of ku0 − v0 kXσ,p .
702 18 Nonlinear parabolic evolution equations in critical spaces

The time T must be small for two reasons. First of all, we need to assure
that u maps [0, T ] to Oσ,p . Secondly, in the proof of the theorem we also
need T to be small in order to be able to use fixed point arguments. This
is hardly surprising: already in the familiar setting of ordinary differential
equations, blow-up can occur in the presence of locally Lipschitz continuous
non-linearities F . Theorem 18.2.15 will provide conditions under which one
can extend the time interval of existence and uniqueness to the full interval
[0, ∞). For a special class of semi-linear equations, Theorem 18.2.17 will give
large-time well-posedness for small initial data.

18.2.c Proof of the main result

The proof Theorem 18.2.6 uses a fixed point argument similar to the one of
Theorem 18.1.2. However, the proof is technically more demanding due to the
quasi-linear structure of the problem, the presence of the additional term Fc ,
the use of weights, and the admission of the full range p ∈ [1, ∞]; some new
ideas are needed to deal with these difficulties.
We will use the following abbreviations to keep the formulas at a reasonable
length. For k ∈ {0, 1} and j ∈ {1, . . . , m} we let, with notation introduced
earlier,
 p
Lwα (0, T ; Xk ) if p < ∞
Ek :=
Cwα ,0 ((0, T ]; Xk ) if p = ∞
(
(ρj +1)p
Lwα/(ρ j +1)
(0, T ; Xβj∗ ) if p < ∞
Yj :=
Cwα/(ρj +1) ,0 ((0, T ]; Xβj∗ ) if p = ∞,

1+ρj σ
where βj∗ := 1+ρj . Assumption 18.2.2 implies that βj 6 βj∗ .

Lemma 18.2.7. Let Assumption 18.2.2 hold. Then for all T > 0 we have
continuous embeddings

MRpα (0, T ) ,→ C([0, T ]; Xσ,p ),

MRpα (0, T ) ,→ Yj , j ∈ {1, . . . , m},

and there exists a constant M1,T > 0 such that for all u ∈ MRpα (0, T ) and
j ∈ {1, . . . , m} we have

kukC([0,T ];Xσ,p ) + kukYj 6 M1,T kukMRpα (0,T ) . (18.16)

These constants may be chosen so that supT >1 M1,T < ∞. For functions
u ∈ MRpα (0, T ) satisfying u(0) = 0, the constants M1,T can be replaced by a
constant M1 independent of T > 0.
Proof. For p ∈ [1, ∞), the embeddings and estimates follow from Corollar-
ies L.4.6 and L.4.7, where for p = 1 we additionally use Remark L.4.2 and
Proposition L.4.5.
 18.2 Local well-posedness for quasi-linear evolution equations 703

For p = ∞, the same can be done if Yj is replaced by Cwα/(ρj +1) ((0, T ]; Xβj∗ ).
∞
To get the embedding into its closed subspace Yj , recall u ∈ MRα (0, T ) =
1
Cwα,0 ((0, T ]; X1 ) ∩ Cwα,0 ((0, T ]; X0 ). Then, for all j ∈ {1, . . . , m}, by (L.19),

ku(t)kXβ∗ 6 (1 − σ)−1 ku(t)kX

λ
σ,∞
1−λ
ku(t)kX 1
,
j

1−βj∗ ρj
where λ = α = 1+ρj . Hence

tα/(ρj +1) ku(t)kXβ∗ 6 (1 − σ)−1 kukλC([0,T ];Xσ,∞ ) (tα ku(t)kX1 )1−λ ,

j

and the latter tends to zero as t ↓ 0 since u ∈ E1 . This shows that u ∈ Yj . 

Lemma 18.2.8. Let Assumption 18.2.2 hold. Let u, v, z ∈ MRpα (0, T ) be

given, and assume that u and v take values in Oσ,p . Then we have A(u)z ∈ E0 ,
Fc (u) ∈ E0 , and FTr (u) ∈ C([0, T ]; X0 ). Moreover,

kA(u)z − A(v)zkE0 6 Lku − vkC([0,T ];Xσ,p ) kzkE1

and
kFTr (u) − FTr (v)kC([0,T ];X0 ) 6 LTr ku − vkC([0,T ];Xσ,p ) ,
m
ρ ρ ρ 
X
Cβjj ,X Lc T δj + kukYjj + kvkYjj ku − vkYj ,

kFc (u) − Fc (v)kE0 6
j=1
(18.17)
αρj ρj
where δj = 1+ρj + (1+ρj )p .

This lemma asserts in particular that the integrability assumptions on A(u)u

and F (u) in Definition 18.2.5 are automatically satisfied for functions u ∈
MRpα (0, T ).
Proof. First consider the case p < ∞. By Assumption 18.2.2(2),

kA(u(t))z(t) − A(v(t))z(t)kX0 6 Lku(t) − v(t)kXσ,p kz(t)kX1 . (18.18)

This gives the required estimate for A. Taking v ≡ u0 ∈ Xσ,p fixed, one also
sees that the function t 7→ A(u(t))z(t) belongs to E0 .
By Assumption 18.2.2(3),

kFTr (u(t)) − FTr (v(t))kX0 6 LTr ku(t) − v(t)kXσ,p .

This implies the estimate for FTr in (18.17); the assumptions on FTr and the
continuity of u : [0, T ] → Xσ,p (see Lemma 18.2.7) imply that t 7→ FTr (u(t))
belongs to C([0, T ]; X0 ).
Next, we have u, v ∈ Yj by Lemma 18.2.7. Moreover, for all j ∈ {1, . . . , m},
704 18 Nonlinear parabolic evolution equations in critical spaces

ρ ρ
(1+kukXjβ∗ + kvkXjβ∗ )ku − vkXβ∗
j j j Lp
wα (0,T )

(i)
ρ ρ
6 (1 + kukXjβ∗ + kvkXjβ∗ )k j(ρ +1)p/ρj ku − vk j (ρ +1)p
j j Lwαρ (0,T ) Lwα/(ρ (0,T ;Xβ ∗ )
j /(ρj +1) j +1) j

(ii)
ρ ρ 
6 T δj +

kukYjj + kvkYjj ku − vkYj
1 j ρ
where in (i) we applied Hölder’s inequality with (1+ρ j)
+ (1+ρ j)
= 1 and
in (ii) the definition of Yj and the triangle inequality. The estimate for Fc
in (18.17) now follows from Assumption 18.2.2 and the inequality kxkXβj 6
Cβj ,X kxkXβ∗ (see Proposition L.1.1(2)), where we used that the embedding
j
constant in X1 ,→ X0 is > 1, and thus Cβj ,X > 1.
The estimate for FTr immediately extends to p = ∞. The estimates for
A and Fc also extend to p = ∞ if we replace E0 = Cwα ,0 ((0, T ]; X0 ) by
L∞wα (0, T ; X0 ). In order to obtain the estimates in the E0 -norm, it remains
to prove that t 7→ A(u(t))z(t) and t 7→ Fc (u(t)) are continuous on (0, T ] and
tα kA(u(t))z(t)kX0 and tα kFc (u(t))kX0 are bounded and tend to zero as t ↓ 0.
To prove continuity for A, we observe that for s, t ∈ (0, T ]
kA(u(t))z(t) − A(u(s))z(s)kX0
6 k(A(u(t)) − A(u(s))z(t)kX0 + kA(u(s))(z(t) − z(s))kX0
6 Lku(t) − u(s)kXσ,∞ kz(t)kX1 + kA(u(s))kL (X1 ,X0 ) k(z(t) − z(s))kX1
L
6 α ku(t) − u(s)kXσ,∞ kzkE1 + kA(u(s))kL (X1 ,X0 ) k(z(t) − z(s))kX1 .
t
The latter tends to zero if t → s, and the desired continuity follows. To prove
the bound and convergence of tα kA(u(t))z(t)kX0 , we observe that by (18.18),
applied with v ≡ x ∈ X1 ∩ Oσ,p ,
kA(u(t))z(t)kX0 6 kA(x)z(t)kX0 + kA(u(t))z(t) − A(x)z(t)kX0
6 kA(x)kL (X1 ,X0 ) kz(t)kX1 + Lku − xkC([0,T ];Xσ,p ) kz(t)kX1 .
Since z ∈ Cwα ,0 ((0, T ]; X1 ), this implies the desired boundedness and conver-
gence.
To prove continuity for Fc , note that by Assumption 18.2.2, for s, t ∈ (0, T ]
we have
m
ρ ρ
X
kFc (u(t)) − Fc (u(s))kX0 6 Lc (1 + ku(t)kXjβ + ku(s)kXjβ )ku(t) − u(s)kXβj.
j j
j=1

The latter tends to zero as t → s. Indeed, since u ∈ Yj , u : (0, T ] → Xβj∗ ,→

Xβj is continuous for each j ∈ {1, . . . , m}. To prove the bound and the con-
vergence for Fc , note that as already mentioned, we have
m
ρ ρ ρ
X
kFc (u) − Fc (v)kCwα ((0,T ];X0 ) 6 Lc Cβjj ,X (T δj + kukYjj + kvkYjj )ku − vkYj .
j=1
 18.2 Local well-posedness for quasi-linear evolution equations 705

From the definitions of δj and Yj it follows that T δj , kukYj , kvkYj → 0 as

T ↓ 0. In particular, the estimate implies that tα kFc (u(t)) − Fc (v(t))kX0 → 0
as t ↓ 0. For v ≡ x with x ∈ X1 ∩ Oσ,p it is clear that tα kFc (v)kX0 → 0 as
t ↓ 0. Therefore, tα kF (u(t))kX0 → 0 as t ↓ 0, and hence F (u) ∈ E0 . 
For each v0 ∈ Xσ,p and T > 0, we define the reference solution zv0 ∈
MRpα (0, T ) as the Lpwα -solution to the following linear problem (see Corol-
laries 17.2.37 and 17.2.48):
(
u0 + A(u0 )u = 0, on R+ ,
u(0) = v0 .

Clearly, the mapping v0 7→ zv0 is linear.

Let ε, r, T > 0 be fixed for the moment; these parameters will be chosen
small enough shortly. For v0 ∈ BXσ,p (u0 , ε) ⊆ Oσ,p consider the following
p
subset of MRα (0, T ):

BrT (v0 ) = {v ∈ MRpα (0, T ) : v(0) = v0 , kv − zu0 kMRpα (0,T ) 6 r}. (18.19)

Note that BrT (v0 ) is a closed subset of MRpα (0, T ) by the continuity of the
trace at zero (see Lemma 18.2.7).
To prove local well-posedness for (18.13), we will apply the Banach fixed
point theorem to the mapping Φv0 : BrT (v0 ) → BrT (v0 ) defined by Φv0 (v) = u,
p
where u is the Lw α
-solution to
(
u0 + A(u0 )u = (A(u0 ) − A(v))v + F (v), on (0, T ),
(18.20)
u(0) = v0 .

Below we will first ensure that BrT (v0 ) ⊆ Oσ,p for ε, r > 0 small enough, so
that A(v) and F (v) are well-defined. Then from its definition, it is clear that
Φv0 maps BrT (v0 ) to MRpα (0, T ). Below we will check that for ε, r, T > 0 small
enough, Φv0 is well defined as a mapping from BrT (v0 ) to itself by using the
maximal regularity assumption on A(u0 ) and the mapping properties of A
p
and F . Note that a function u is an Lw α
-solution to (18.13) if and only if u
p
is an Lwα -solution (Cwα -solution if p = ∞) to (18.20) with u = v. Before we
turn to the fixed point argument we need several preparatory lemmas.
Choose ε0 > 0 such that BXσ,p (u0 , ε0 ) ⊆ Oσ,p . Fix T1 > 0 such that

kzu0 − u0 kC([0,T1 ];Xσ,p ) < ε0 /3. (18.21)

By Corollaries 17.2.37 and 17.2.48, there is a constant CT1 such that for every
v0 ∈ Xσ,p we have

kzv0 kMRpα (0,T1 ) 6 CT1 kv0 kXσ,p . (18.22)

The constant CT1 will depend on T1 in general, but this will not create prob-
lems since T1 is fixed.
706 18 Nonlinear parabolic evolution equations in critical spaces

In order to show that A(v) and F (v) in (18.20) are well defined, we need
to check that v(t) ∈ Oσ,p for all t ∈ (0, T ) when ε ∈ (0, ε0 ), r ∈ (0, 1], and
T ∈ (0, T1 ] are small enough. This is taken care of in the next lemma.

Lemma 18.2.9. Let Assumption 18.2.2 hold, and let ε0 > 0 be chosen as
before (18.21). For small enough r ∈ (0, 1] and ε ∈ (0, ε0 ) the following holds:
For all v0 ∈ BXσ,p (u0 , ε), all T ∈ (0, T1 ], and all v ∈ BrT (v0 ), one has kv −
u0 kC([0,T ];Xσ,p ) < ε0 , and thus v(t) ∈ Oσ,p for all t ∈ [0, T ].

Proof. For notational convenience we write k · k∞,T = k · kC([0,T ];Xσ,p ) . For all
v ∈ BrT (v0 ),

kv−zv0 k∞,T 6 M1 kv − zv0 kMRpα (0,T ) (by (18.16))

6 M1 kv − zu0 kMRpα (0,T ) + M1 kzu0 − zv0 kMRpα (0,T1 )
6 M1 r + M1 CT1 ku0 − v0 kXσ,p (by (18.22)).

Therefore, by (18.16), (18.21), and (18.22)

kv − u0 k∞,T 6 kv − zv0 k∞,T + kzv0 − zu0 k∞,T1 + kzu0 − u0 k∞,T1

6 M1 r + CT1 (M1 + M1,T1 )ku0 − v0 kXσ,p + kzu0 − u0 k∞,T1
6 M1 r + CT1 (M1 + M1,T1 )ε + ε0 /3.
(18.23)

This implies the required result for all r, ε > 0 small enough. 
In the next lemma we collect some estimates for A, FTr , and Fc , which will
be used to ensure that Φv0 maps BrT (v0 ) to itself.

Lemma 18.2.10 (Smallness). Let Assumption 18.2.2 hold. Fix T ∈ (0, T1 ]

and let ε ∈ (0, ε0 ) and r ∈ (0, 1] be as in Lemma 18.2.9. Then for all v0 ∈
BXσ,p (u0 , ε) and v ∈ BrT (v0 ) we have

k(A(v) − A(u0 ))vkE0

6 M1 r + M2,T1 ε + kzu0 − u0 kC([0,T ];Xσ,p ) (r + kzu0 kE1 ),
1
6 T α+ p LTr ε0 + kFTr (u0 )kX0 ,


kFTr (v)kE0
kFc (v)kE0 6 Cε,r,T (u0 )r + Cε,T (u0 ),

where Cε,r,T (u0 ) and Cε,T (u0 ) are independent of v0 and v, Cε,r,T (u0 ) and
Cε,T (u0 ) are non-decreasing in each of the variables ε, r, and T , and satisfy
Cε,r,T (u0 ), Cε,T (u0 ) → 0 as ε, r, T ↓ 0.

Proof. We use the short-hand notation k · k∞,T := k · kC([0,T ];Xσ,p ) .

As in (18.23), one sees that

kv − u0 k∞,T 6 M1 r + CT1 (M1 + M1,T1 )ε + kzu0 − u0 k∞,T .

 18.2 Local well-posedness for quasi-linear evolution equations 707

Therefore, by Lemma 18.2.8,

k(A(v) − A(u0 ))vkE0

6 Lkv − u0 k∞,T kvkE1


6 M1 r + CT1 (M1 + M1,T1 )ε + kzu0 − u0 k∞,T kvkE1 .


6 M1 r + CT1 (M1 + M1,T1 )ε + kzu0 − u0 k∞,T (r + kzu0 kE1 ).

For FTr we have pointwise estimate

kFTr (v)kX0 6 kFTr (v) − FTr (u0 )kX0 + kFTr (u0 )kX0
6 LTr kv − u0 kXσ,p + kFTr (u0 )kX0
6 LTr ε0 + kFTr (u0 )kX0 ,

where in the last step we used Lemma 18.2.9. Taking Lpwα -norms, we obtain
1
kFTr (v)kE0 6 T α+ p (LTr ε0 + kFTr (u0 )kX0 ).

The estimate for Fc is more difficult to obtain. By the second estimate in

(18.17),
m
ρ ρ ρ

X
kFc (v) − Fc (zu0 )kE0 6 Cβjj ,X Lc T δj + kvkYjj + kzu0 kYjj kv − zu0 kYj .
j=1

It remains to estimate kvkYj and kv − zu0 kYj . By (18.16),

kv − zu0 kYj 6 kv − zv0 kYj + kzv0 − zu0 kYj

6 M1 kv − zv0 kMRpα (0,T ) + M1,T1 kzv0 − zu0 kMRpα (0,T1 )
6 M1 kv − zu0 kMRpα (0,T ) + (M1 + M1,T1 )kzv0 − zu0 kMRpα (0,T1 )
6 M1 r + (M1 + M1,T1 )CT1 kv0 − u0 kXσ,p
6 M1 r + (M1 + M1,T1 )CT1 ε,

applying (18.22) in the penultimate estimate. Similarly,

kvkYj 6 kv − zu0 kYj + kzu0 kYj 6 M1 r + (M1 + M1,T1 )CT1 ε + kzu0 kYj .

Combining things, we obtain the estimate

kFc (v)kE0
6 kFc (v) − Fc (zu0 )kE0 + kFc (zu0 )kE0
m
ρ
X
Cβjj ,X Lc T δj + 2(M1 r + C
eT ε + kj,T (u0 ))ρj (M1 r + C


6 eT ε) + kc,T (u0 ),
1 1
j=1

where have set C eT = (M1 + M1,T )CT , kj,T (u0 ) = kzu kY , and kc,T (u0 ) =
1 1 1 0 j
kFc (zu0 )kE0 . Note that kj,T (u0 ) → 0 and kc,T (u0 ) → 0 as T ↓ 0 since zu0 ∈
MRpα (0, T ) ⊆ Yj and since Fc (zu0 ) ∈ E0 by Lemma 18.2.8.
708 18 Nonlinear parabolic evolution equations in critical spaces

The estimate kFc (v)kE0 6 Cε,r,T (u0 )r + Cε,T (u0 ) in the statement of the
lemma now follows, with constants
m
ρ
X
Cβjj ,X Lc T δj + 2(M1 r + C
eT ε + kj,T (u0 ))ρj M1


Cε,r,T (u0 ) = 1
j=1
eT εM −1 Cε,1,T (u0 ) + kc,T (u0 ),
Cε,T (u0 ) = C 1 1

where we used that r ∈ (0, 1]. 

Remark 18.2.11. In the last part of the proof one does not have kj,T (u0 ) → 0
∞
and kc,T (u0 ) → 0 as T ↓ 0 if one were to use maximal Lw α
-regularity or data
u0 in (X0 , X1 )σ,∞ rather than in the closed subspace Xσ,∞ . This is one of
the reasons for working with maximal Cwα -regularity and data in Xσ,∞ . It is
also clear from the above proof that α = 0 leads to difficulties if p = ∞. For
1
example, the estimate for FTr (v) in Lemma 18.2.10 contains a factor T α+ p
which does not vanish in the limit T ↓ 0 if α = 0 and p = ∞.

The final lemma contains Lipschitz variations of the above estimates, which
will be used to show that Φv0 is a uniform contraction.

Lemma 18.2.12 (Lipschitz estimates). Let Assumption 18.2.2 hold. Fix

T ∈ (0, T1 ] and let ε ∈ (0, ε0 ) and r ∈ (0, 1] be as in Lemma 18.2.9. Then for
all v1,0 , v2,0 ∈ BXσ,p (u0 , ε), and all v1 ∈ BrT (v1,0 ) and v2 ∈ BrT (v2,0 ),

k(A(v1 ) − A(v2 ))v1 kE0 + k(A(u0 ) − A(v2 ))(v1 − v2 )kE0

+ kFTr (v1 ) − FTr (v2 )kE0 + kFc (v1 ) − Fc (v2 )kE0

can be estimated from above by

Lε,r,T (u0 )(kv1 − v2 kMRpα (0,T ) + kv1,0 − v2,0 kXσ,p ),

where Lε,r,T (u0 ) is a constant independent of v1,0 , v2,0 , v1 , v2 , non-decreasing

in each of the variables ε, r, T , and satisfying Lε,r,T (u0 ) → 0 as ε, r, T ↓ 0.

Proof. We use the short-hand notation k · k∞,T := k · kC([0,T ];Xσ,p ) .

First we provide an estimate for kvk∞,T and kv − u0 k∞,T for v ∈ BrT (v0 )
and v0 ∈ BXσ,p (u0 , ε). By (18.22) and (18.16),

kvk∞,T 6 kzv0 k∞,T + kv − zv0 k∞,T

6 M1,T1 CT1 kv0 kXσ,p + M1 kv − zv0 kMRpα (0,T ) (18.24)
6 (M1,T1 + M1 )CT1 kv0 kXσ,p + M1 kvkMRpα (0,T ) .

Similarly, setting kT (u0 ) := kzu0 − u0 k∞,T ,

 18.2 Local well-posedness for quasi-linear evolution equations 709

kv − u0 k∞,T
6 kv − zv0 k∞,T + kzv0 − zu0 k∞,T + kT (u0 )
6 M1 kv − zv0 kMRαp (0,T ) + M1,T1 kzv0 − zu0 kMRαp (0,T1 ) + kT (u0 )
6 M1 kv − zu0 kMRαp (0,T ) + (M1 + M1,T1 )kzv0 − zu0 kMRαp (0,T1 ) + kT (u0 )
6 M1 r + (M1 + M1,T1 )CT1 ε + kT (u0 ).
(18.25)

To estimate the first A-term, by Lemma 18.2.8 we obtain

k(A(v1 ) − A(v2 ))v1 kE0 6 Lkv1 − v2 k∞,T kv1 kE1

6 Lkv1 − v2 k∞,T (r + kzu0 kE1 ).

Therefore, the required estimate follows from (18.24) with v0 = 0.

For the second A-term, we again use Lemma 18.2.8 and obtain

k(A(u0 ) − A(v2 ))(v1 − v2 )kE0 6 Lku0 − v2 k∞,T kv1 − v2 kE1 .

Therefore, the required estimate follows from (18.25).

For the FTr -term, we use Lemma 18.2.8 to obtain
1
kFTr (v1 ) − FTr (v2 )kE0 6 T α+ p LTr kv1 − v2 k∞,T .

Therefore, the estimate follows from (18.24) again.

The Fc -term is more difficult to estimate. In the same way as in (18.24)
and (18.25) one shows that

kvkYj 6 (M1,T1 + M1 )CT1 kv0 kXσ,p + M1 kvkMRpα (0,T ) (18.26)

and

kvkYj 6 M1 r + (M1 + M1,T1 )CT1 ε + kzu0 kYj . (18.27)

By the second estimate in (18.17),

m
ρ ρ ρ

X
kFc (v1 ) − Fc (v2 )kE0 6 Cβjj ,X Lc T δj + kv1 kYjj + kv2 kYjj kv1 − v2 kYj .
j=1

Using (18.26), we find

kv1 − v2 kYj 6 (M1,T1 + M1 )CT1 kv1,0 − v2,0 kXσ,p + M1 kv1 − v2 kMRpα (0,T ) .

The required estimate for Fc now follows by applying (18.27) to estimate

ρ ρ
kv1 kYjj and kv2 kYjj . 
710 18 Nonlinear parabolic evolution equations in critical spaces

After these preparations we are ready to turn to the proof of Theorem 18.2.6.
It will be useful to recall the maximal regularity estimate which follows from
Corollaries 17.2.37 and 17.2.48: for all f ∈ E0 and v0 ∈ Xσ,p there exists a
unique Lpwα -solution (Cwα -solution if p = ∞) to the problem
(
u0 + A(u0 )u = f on (0, T ),
u(0) = v0 ,

and there exists a constant CT > 0, independent of f and v0 , such that

kukMRpα (0,T ) 6 CT kf kE0 + CT kv0 kXσ,p . (18.28)

This constant CT also depends on A(u0 ) and p, but we can choose it in such
a way that CT 6 CT1 whenever T < T1 ; this follows from a weighted version
of (17.25).
Proof of Theorem 18.2.6. Fix ε ∈ (0, ε0 ) and r ∈ (0, 1] be as in Lemma
18.2.9, and let T ∈ (0, T1 ]. Let BrT (v0 ) be as in (18.19). Let Φv0 : BrT (v0 ) →
p
MRα (0, T ) be defined by Φv0 (v) := u, where u is the Lpwα -solution (Cwα -
solution if p = ∞) to the problem
(
u0 + A(u0 )u = (A(u0 ) − A(v))v + F (v),
(18.29)
u(0) = v0 .

Then v takes values in Oσ,p by Lemma 18.2.9, and we have (A(v) − A(u0 ))v ∈
E0 and F (v) ∈ E0 by Lemma 18.2.10. Below Theorem 18.2.6 we have already
observed that local existence and uniqueness follow if we can show that Φv0
has a unique fixed point.
Since u − zu0 satisfies (18.29) with v0 replaced by v0 − u0 , by the maximal
regularity estimate (18.28) applied on (0, T1 ) (see (18.21) for the definition of
T1 ) we have


ku − zu0 kMRpα (0,T ) 6 CA,T1 ku0 − v0 kXσ,p + k(A(u0 ) − A(v))v + F (v)kE0
eε,r,T r + C
6 CA,T1 ε + C eε,T ,

applying Lemma 18.2.10 in the last step, and where C eε,r,T and Ceε,T are con-
stants such that Cε,r,T → 0 as ε, r, T ↓ 0. Therefore, for r, ε, T > 0 small
e
enough we obtain ku − zu0 kMRpα (0,T ) 6 r, and thus u ∈ BrT (v0 ).
Next, fix vj,0 ∈ BXσ,p (u0 , ε) and vj ∈ BrT (vj,0 ) for j ∈ {1, 2}. Then u =
Φv1,0 (v1 ) − Φv2,0 (v2 ) solves the problem
(
u0 + A(u0 )u = (A(u0 ) − A(v1 ))v1 − (A(u0 ) − A(v2 ))v2 + F (v1 ) − F (v2 ),
u(0) = v1,0 − v2,0 .

Therefore, by the maximal regularity estimate (18.28),

 18.2 Local well-posedness for quasi-linear evolution equations 711

kukMRpα (0,T ) 6 CA,T1 (RA + RF ) + CA,T1 kv1,0 − v2,0 kXσ,p ,

where

RA : = k(A(u0 ) − A(v1 ))v1 − (A(u0 ) − A(v2 ))v2 kE0

6 k(A(v1 ) − A(v2 ))v1 kE0 + k(A(u0 ) − A(v2 ))(v1 − v2 )kE0

and

RF : = kF (v1 ) − F (v2 )kE0

6 kFTr (v1 ) − FTr (v2 )kE0 + kFc (v1 ) − Fc (v2 )kE0 .

From Lemma 18.2.12 we deduce that

kukMRαp (0,T ) 6 CA,T1 Lε,r,T (u0 )kv1 − v2 kMRαp (0,T )

+ CA,T1 (Lε,r,T (u0 ) + 1)kv1,0 − v2,0 kXσ,p .

Choosing ε > 0, r > 0, and T > 0 so small that CA,T1 Lε,r,T (u0 ) 6 1/2, we
obtain
1
kΦv1,0 (v1 ) − Φv2,0 (v2 )kMRαp (0,T ) 6 kv1 − v2 kMRpα (0,T )
2 (18.30)
+ (CA,T1 + 1)kv1,0 − v2,0 kXσ,p .

The estimate (18.30) allows us to finish the proof of local well-posedness.

By (18.30), Φv0 : BrT (v0 ) → BrT (v0 ) is a uniform contraction, and thus it has
a unique fixed point uv0 ∈ BrT (v0 ). This is the required solution to (18.13).
Moreover, (18.30) implies that for all v1,0 , v2,0 ∈ BXσ,p (u0 , ε),

1
kuv1,0 − uv2,0 kMRpα (0,T ) 6 kuv1,0 − uv2,0 kMRpα (0,T )
2
+ (CA,T1 + 1)kv1,0 − v2,0 kXσ,p

which implies

kuv1,0 − uv2,0 kMRpα (0,T ) 6 2(CA,T1 + 1)kv1,0 − v2,0 kXσ,p .

This gives (18.14).

It remains to prove uniqueness. Uniqueness does hold if we only consider
solutions in BrT (v0 ). In order to derive uniqueness for the larger set MRpα (0, T ),
we will replace ε and T by suitable smaller values εe and Te. The above estimates
then show that Φv0 : BrT (v0 ) → BrT (v0 ) and (18.30) holds with T replaced by
e e

Te.
r

Let εe := min ε, 8(CA,T +1) and set
1

n ro
Te := inf t ∈ [0, T ] : kuu0 − zu0 kMRpα (0,t) > ,
2
712 18 Nonlinear parabolic evolution equations in critical spaces

where inf ∅ := T . Then, for all v0 ∈ BXσ,p (u0 , εe),

r
kuv0 − uu0 kMRpα (0,T ) 6 2(CA,T1 + 1)kv0 − u0 kXσ,p 6 .
4
In particular,
3r
kuv0 − zu0 kMRpα (0,Te) 6 kuv0 − uu0 kMRpα (0,Te) + kuu0 − zu0 kMRpα (0,Te) 6 .
4

We claim that for every v0 ∈ BXσ,p (u0 , εe), the element uv0 ∈ MRpα (0, Te)
is the unique Lpwα -solution to (18.13). To show this, we will prove the slightly
stronger result (which will play a key role in the construction of the maximal
solution in Section 18.2.d) that, for an τ > 0, if v ∈ MRpα (0, τ ) is an Lpwα -
solution to (18.13), then v ≡ uv0 on [0, Te ∧ τ ]. This will give the theorem for
Te instead of T .
Let
τv := inf{t ∈ [0, Te ∧ τ ] : kv − zu0 kMRpα (0,t) > r},

setting inf ∅ := Te ∧ τ . Then v|[0,τv ] belongs to Brτv (v0 ), and since τv 6 T it

follows that v|[0,τv ] = uv0 |[0,τv ] by uniqueness of the fixed point in Brτv (v0 ).
Thus we obtain

kv − zu0 kMRpα (0,τv ) = kuv0 − zu0 kMRpα (0,τv ) 6 kuv0 − zu0 kMRpα (0,Te) < r,

and therefore τv = Te ∧ τ . This gives the claimed result. 

18.2.d Maximal solutions

Having established local well-posedness in Theorem 18.2.6, we will now extend

the time interval on which the solution exists to a maximal time interval
[0, Tmax (v0 )).
Definition 18.2.13. Let Assumption 18.2.2 hold and assume that v0 ∈ Oσ,p .
A pair (v, Tmax (v0 )) is called a maximal Lpwα -solution to (18.13) if Tmax (v0 ) ∈
(0, ∞] and v : [0, Tmax (v0 )) → X0 are such that
• for all T ∈ (0, Tmax (v0 )), v|(0,T ) belongs to MRpα (0, T ) and is an Lpwα -
solution to (18.13) on (0, T );
• whenever u ∈ MRpα (0, T ) is a unique Lpwα -solution to (18.13) for some
T > 0, one has T 6 Tmax (v0 ) and u ≡ v on (0, T ).
Note that maximal Lpwα -solutions are unique. An even stronger uniqueness
assertion will be derived in Remark 18.2.16 under further restrictions. We will
now show that the solution to (18.13) provided by Theorem 18.2.6 can be
extended to a maximal Lpwα -solution.
 18.2 Local well-posedness for quasi-linear evolution equations 713

Theorem 18.2.14 (Maximal solutions). Let Assumption 18.2.2 hold, let

u0 ∈ Oσ,p , and suppose that A(u0 ) has maximal Lp -regularity (C-regularity
if p = ∞) on finite time intervals. Let ε > 0 be as in Theorem 18.2.6, and
let v0 ∈ Oσ,p be such that ku0 − v0 kXσ,p < ε. Then there exists a maximal
Lpwα -solution (u, Tmax (v0 )) to (18.13).

Proof. Let us say that an Lpwα -solution v to (18.13) on (0, T ) has the unique-
p
ness property if for any τ > 0 and any Lw α
-solution u to (18.13) on (0, τ ), we
have v ≡ u on [0, T ∧ τ ]. Let Tmax (v0 ) be the supremum of all T > 0 such
p
that there exists an Lw α
-solution to (18.13) on (0, T ) with the uniqueness
property. Then Tmax (v0 ) > 0 by Theorem 18.2.6. Note that the uniqueness
property was established as part of the uniqueness proof. It follows that there
exists a maximal Lpwα -solution u : [0, Tmax (v0 )) → X0 to (18.13). 

Theorem 18.2.15 (Global well-posedness for quasi-linear equations).

Let Assumption 18.2.2 hold, and suppose that for all u0 ∈ Oσ,p the operator
A(u0 ) has maximal Lp -regularity (C-regularity if p = ∞) on finite time in-
tervals. Let v0 ∈ Oσ,p and let v : [0, Tmax (v0 )) → X0 be the maximal solution
provided by Theorem 18.2.14. If Tmax (v0 ) < ∞, then either
• limt↑Tmax (v0 ) v(t) does not exist in Xσ,p , or
• v∗ := limt↑Tmax (v0 ) v(t) exist in Xσ,p , but v∗ ∈
/ Oσ,p .

The final assertion in the theorem is called a blow-up criterion. Blow-up cri-
teria can be used to prove global well-posedness. In typical applications, as-
suming Tmax (v0 ) < ∞, energy estimates can be used to show that v∗ :=
limt↑Tmax (v0 ) v(t) exists in Oσ,p . This contradicts Theorem 18.2.15 and thus
leads to Tmax (v0 ) = ∞, i.e., global existence. Further blow-up criteria are
discussed in the Notes.
Proof. Assuming that T0 := Tmax (v0 ) < ∞ and that v∗ := limt↑T0 v(t) exists
in Xσ,p with v∗ ∈ Oσ,p , a contradiction will be derived.
The idea is to restart the problem at time T0 with initial value v∗ and apply
Theorem 18.2.6 to extend v to a larger time interval [0, T0 + δ]. However, it is
not self-evident that v ∈ MRpα (0, T0 + δ). This problem will be overcome by
using a compactness argument.
From the continuity of v and the assumption that the limit v∗ at t = T0
exist, it follows that the set

K := {v(t) : t ∈ [0, T0 )} ∪ {v∗ }

is compact in Xσ,p . By Theorem 18.2.6, for all x ∈ K there exists an open ball
p
B(x, εx ) ⊆ Oσ,p such that for initial values from B(x, εx ) we can find an Lw α
-
p
solution in MRα (0, tx ) for some tx > 0. Since K is compact, the open cover
{B(x, εx ) : x ∈ K} has a finite sub-cover {B(xn , εxn ) : n = 1, . . . N }. Let
δ := minn=1,...,N txn . Then for all x ∈ K there exists a unique Lpwα -solution
ux ∈ MRpα (0, δ) to the problem
714 18 Nonlinear parabolic evolution equations in critical spaces
(
u0 + A(u)u = F (u),
(18.31)
u(0) = x.

Now we are ready to define a suitable extension of v. Let x := v(T0 − 21 δ),

and let ux ∈ MRpα (0, δ) be as above. Then t 7→ v(T0 − 21 δ + t) belongs to
MRpα (0, γ) for all γ ∈ (0, δ) and is an Lpwα -solution to (18.31). Therefore,
uniqueness gives that v(T0 − 21 δ + t) = ux (t) for all t ∈ [0, δ/2). Now one can
check that the function vext : [0, T0 + δ/2] → Xσ,p defined by

v(t), t ∈ [0, T0 );
vext (t) =
ux (t − T0 + 21 δ), t ∈ [T0 − 21 δ, T0 + 21 δ].

is well defined, belongs to MRpα (0, T + 21 δ), and is an Lpwα -solution to (18.13)
on (0, T0 + 21 δ). This contradicts the maximality of T0 . 
Under the conditions of Theorem 18.2.15, one can leave out the uniqueness
from the second bullet in Definition 18.2.13. This excludes the existence of an
(non-unique) Lpwα -solution u ∈ MRpα (0, T ) which extends v.
Remark 18.2.16. Let Assumption 18.2.2 hold, and suppose that for all u0 ∈
Oσ,p the operator A(u0 ) has maximal Lp -regularity (C-regularity if p = ∞)
on finite time intervals. Let v0 ∈ Oσ,p and let v : [0, Tmax (v0 )) → X0 be
the maximal solution provided by Theorem 18.2.14. Now suppose that u ∈
MRpα (0, T ) is an Lpwα -solution to (18.13) for some T > 0. We claim that
T 6 Tmax (v0 ) and u ≡ v on (0, T ). To see this, first note that by the uniqueness
property of the proof of Theorem 18.2.14 one has u = v on [0, T ∧ Tmax (v0 )).
Thus it remains to show T 6 Tmax (v0 ). Suppose that T > Tmax (v0 ). Since
u ∈ MRpα (0, T ), it follows from Lemma 18.2.7 that

v∗ := lim v(t) = lim u(t) = u(Tmax (v0 )) exists in Xσ,p ,

t↑Tmax (v0 ) t↑Tmax (v0 )

and v∗ ∈ Oσ,p . This contradicts Theorem 18.2.15 and thus the claim follows.
As a consequence of Theorem 18.2.15 we obtain the following criteria for global
well-posedness for (18.13) in the semi-linear case.

Theorem 18.2.17 (Global well-posedness for semi-linear equations).

Let Assumption 18.2.2 hold for any bounded open set Oσ,p , and that A ∈
L (X1 , X0 ) has maximal Lp -regularity (maximal C-regularity if p = ∞) on
finite time intervals. Then for every v0 ∈ Xσ,p there exists a maximal Lpwα -
solution (v, Tmax (v0 )) to (18.13) with Tmax (v0 ) > 0. Moreover, if either one
of the following holds:
(1) p < ∞, sup kv(t)kXσ,p + kvkLpwα (0,Tmax (v0 );X1 ) < ∞;
t∈[0,Tmax (v0 ))
(2) p = ∞, sup kv(t)kXσ,∞ + tα kv(t)kX1 < ∞;
t∈[0,Tmax (v0 ))
 18.2 Local well-posedness for quasi-linear evolution equations 715

(3) sup kv(t)kXσ,p < ∞ and Assumption 18.2.2 holds in the sub-
t∈[0,Tmax (v0 ))
critical case,
then Tmax (v0 ) = ∞, and thus the Lpwα -solution v exists globally.

Proof. The existence of the maximal solution has already been observed in
Theorem 18.2.14.
We start with a preliminary observation. Fix ρ > 0 and T ∈ (0, ∞),
and set β ∗ := 1 − (α + p1 )(1 − ρ+1 1
). We claim that for all β ∈ (σ, β ∗ ] and
u ∈ L∞ (0, T ; Xσ,p ) ∩ Lpwα (0, T ; X1 ) we have

kukLhp
w (0,T ;(X0 ,X1 )β,1 ) 6 CT kukλL∞ (0,T ;Xσ,p ) kuk1−λ
Lp
w (0,T ;X1 )
, (18.32)
α/h α

1−β
where h = ρ + 1, λ ∈ (0, 1) is given by λ = 1
α+ p
, and where CT also depends
on α, h, p and is non-decreasing in T . From the assumption on β it follows
1
that λ ∈ [1 − 1+ρ , 1). Moreover, if β < β ∗ , one even has λ > 1 − ρ+1
1
. To prove
(18.32), note that by (C.6), Theorem L.3.1, and (L.2),

ku(t)kβ,1 6 Cku(t)k((X0 ,X1 )σ,p ,X1 )1−λ,1

6 Cku(t)kλσ,p ku(t)k1−λ
X1 ,

with the understanding that ku(t)kσ,p needs to be replaced by ku(t)kX0 in the

case p = 1. Taking Lhp
wα/h (0, T )-norms on both sides gives

kukLhp (0,T ;(X0 ,X1 )β,1 ) 6 CkukλL∞ (0,T ;Xσ,p ) kuk1−λ

hp(1−λ)
w α/h Lwα/(hp(1−λ)) (0,T ;X1 )

6 CT kukλL∞ (0,T ;Xσ,p ) kuk1−λ

Lp
wα (0,T ;X1 )
,

where we used h(1 − λ) = (1 + ρ)(1 − λ) 6 1.

(1): Suppose, for a contradiction, that Tmax (v0 ) < ∞. Let Oσ,p ⊆ Xσ,p
be a bounded open set such that v([0, Tmax (v0 ))) ⊆ Oσ,p . Taking β = βj∗
and h = ρj + 1 in (18.32), we obtain u ∈ Yj for every j, and thus Fc (v) ∈
Lpwα (0, Tmax (v0 ); X0 ) by Lemma 18.2.8. Since FTr : Oσ,p → X0 has linear
growth, it is straightforward to check that

FTr (v) ∈ L∞ (0, Tmax (v0 ); X0 ) ⊆ Lpwα (0, Tmax (v0 ); X0 ).

Therefore, maximal Lp -regularity of A implies that v ∈ MRpα (0, Tmax (v0 )). In
particular, limt↑Tmax (v0 ) v(t) exists in Xσ,p (see Lemma 18.2.7). This contra-
dicts Theorem 18.2.14. It follows that Tmax (v0 ) = ∞.
(2): This can be proved similarly, this time using maximal C-regularity.
(3): Suppose, for a contradiction, that Tmax (v0 ) < ∞. Let Oσ,p be as in
the proof of (1), and let T ∈ (0, Tmax (v0 )). As before, it suffices to prove
v ∈ MRp (0, Tmax (v0 )).
716 18 Nonlinear parabolic evolution equations in critical spaces

By maximal regularity (see Corollaries 17.2.37 and 17.2.48) we can esti-

mate

kvkMRp (0,T ) 6 C(kv0 kXσ,p + kFc (v)kLpwα (0,T ;X0 ) + kFTr (v)kLw
p
α (0,T ;X0 )
),
(18.33)

where the constant C depends on Tmax (v0 ), but not on T . As before,

kFTr (v)kLw
p
α (0,T ;X0 )
can be estimated above by K(1 + kvkL∞ (0,Tmax (v0 );Xσ,p ) ).
The Fc -term is more complicated to handle; this is where the subcriticality
enters. Set (
(ρj +1)p
Lwα/(ρ j +1)
(0, T ; Xβj ) if p < ∞;
Y j :=
Cwα/(ρj +1) ,0 ((0, T ]; Xβj ) if p = ∞.
Fix x ∈ Oσ,p ∩ X1 . Repeating the proof of the second estimate in (18.17) with
βj∗ replaced by βj , we obtain

kFc (v)kLpwα (0,T ;X0 ) 6 kFc (v) − Fc (x)kLpwα (0,T ;X0 ) + kFc (x)kLpwα (0,T ;X0 )
m
ρ ρ
ρ
X
6 Lc T δj + kvkYej + kxkYej kv − xkYej
j j j
j=1
m
ρ +1
X
6 Lc Cj,x + kvkYej ,
j
j=1

where in the last step we used Young’s inequality in the form aρ b 6 aρ+1 +
bρ+1 , and the constant Cj,x depends on Tmax (v0 ) but not on T . Let

M := sup kv(t)kXσ,p .
t∈[0,Tmax (v0 ))

1−βj
By (18.32) with h = ρj + 1, β = βj , and λj = 1 ,
α+ p
we find that

ρ +1 ρ +1 (1−λj )(ρj +1)

kvkYej 6 CT j M λj (ρj +1) kvkLpw
j α (0,T ;X1 )
ρ +1
6 CT j Cj,ε M (λj (ρj +1))/(1−βj ) + εkvkLpwα (0,T ;X1 ) ,

where we used βj = (1 − λj )(ρj + 1) ∈ (0, 1) by subcriticality, and we used

βj −βj /(1−βj ) 1/(1−βj )
Young’s inequality
Pm in the form ab 6 ε a + εb for arbitrary
ε > 0. Taking j=1 this results in the estimate

kFc (v)kLpwα (0,T ;X0 ) 6 CM,ε + Lc mεkvkLpwα (0,T ;X1 ) .

Combining this estimate with (18.33), we obtain

(1 − CεLc m)kvkMRp (0,T ) 6 C(kv0 kXσ,p + kFTr (v)kLpwα (0,T ;X0 ) ).

Setting ε = (2CLc m)−1 and letting T tend to Tmax (v0 ), it follows that v ∈
MRp (0, Tmax (v0 )). 
 18.3 Examples and comparison 717

18.3 Examples and comparison

In order to understand the assumptions on the non-linearity Fc in Assump-
tion 18.2.2, we will now discuss in detail a standard situation, and make a
comparison with Example 18.1.3 which involved only the non-linearity FTr .

Example 18.3.1 (Critical spaces and non-linearities). Let

X0 := H s,q (Rd ), X1 := H s+2,q (Rd )

with s ∈ (−2, 0] and q ∈ (1, ∞). Since s + 2 > 0, powers of functions in X1

are well defined. Notice that X1 features two more derivatives than X0 ; this
is the typical situation encountered in applications to PDEs with a leading
term of second order. Note that (see Theorem 5.6.9)

[X0 , X1 ]β = H s+2β,q (Rd )

and
s+2σ
Xσ,p = Bq,p (Rd ),
where p ∈ (1, ∞) (extensions to the end-points are possible, but not considered
here for simplicity) and σ ∈ (0, 1/p0 ] are arbitrary but fixed for the moment.
Suppose now that f ∈ C 1 (R) satisfies

f (0) = 0 and |f 0 (t)| 6 `|t|ρ , t ∈ R, (18.34)

for a suitable exponent ρ > 0 and constant ` > 0. Let Fc : X1 → X0 be given

by
(Fc (u))(x) := f (u(x)), x ∈ Rd .
Then Fc is well-defined and Lipschitz on bounded subsets of Xβ under suitable
conditions. Indeed, for all u, v ∈ X1 ,

kFc (u)−Fc (v)kX0

= kf (u) − f (v)kH s,q
6 Ckf (u) − f (v)kr (Sobolev embedding)
6 C`k(|u|ρ + |v|ρ )(u − v)kr (mean value theorem)
6 C`(kukρ(ρ+1)r + kvkρ(ρ+1)r )ku − vk(ρ+1)r (Hölder inequality)
6 C`(kukρXβ + kvkρXβ )ku − vkXβ (Sobolev embedding),

provided we impose some further restrictions in order to justify the application

of the Sobolev embeddings. Specifically, the first Sobolev embedding can be
applied if − dr = s − dq and 1 < r 6 q, which leads to the condition

d
s>− . (18.35)
q0
718 18 Nonlinear parabolic evolution equations in critical spaces

The second Sobolev embedding can be applied if

d d
s + 2β − =− , and q 6 (ρ + 1)r,
q (ρ + 1)r

which after substitution of the identity − dr = s − d

q leads to the condition

d 1  d dq(ρ + 1)
s + 2β − = s− and q 6 .
q ρ+1 q d − qs
Thus we arrive at the conditions
ρ d  dρ
β= − s and s > − . (18.36)
2(ρ + 1) q q
Sobolev embeddings can also be applied in sub-optimal cases, but here we wish
to demonstrate certain optimality and scaling behaviour which is present only
if all Sobolev embeddings are sharp.
Combining (18.36) with the (sub)criticality condition (18.12), we obtain
d 
ρ − s 6 2 + 2ρσ,
q
and criticality holds if
1d  1
σ= −s − .
2 q ρ
Since σ ∈ (0, 1/p0 ] we arrive the following condition on (q, s) to obtain a
critical setting:
1d  1 1
0< − s − 6 0. (18.37)
2 q ρ p
If (18.37) holds for some p, then it also holds for all larger values of p,
and one can take the limit p → ∞. Thus (18.35), (18.36), (18.37), and the
assumption s ∈ (−2, 0] imply
n d 2 do d 2 dρ
max −2+ − , −2, − 0 < s < − , and − 6 s 6 0. (18.38)
q ρ q q ρ q
In the converse direction, if (18.38) holds, then (18.37) holds for large enough
p, so the existence of a triple (p, q, s) satisfying the aforementioned conditions
is equivalent to (18.38).
Elementary computations show that we can find pairs (s, q) satisfying these
conditions holds if and only if
2 2 d
ρ> and < . (18.39)
d ρ(ρ + 1) q
In this case, the corresponding critical space for the initial data is given by
 18.3 Examples and comparison 719
d
s+2σ −2
Xσ,p = Bq,p (Rd ) = Bq,p
q ρ
(Rd ). (18.40)

An interesting feature of (18.40) is that the parameter s does not appear in

the critical space Xσ,p and the smoothness parameter is independent of p.
d
−2
Remark 18.3.2. The homogeneous variant of Bq,p q ρ
(Rd ) scales as ku(λ·)k h
2
λ ρ kuk. It follows from this that if u is a solution to a PDE with leading
second order differential operator in the space variables, with non-linearity
f (u) = k|u|ρ+1 (or similar scaling behaviour), and with initial data u0 , then
2
(t, x) 7→ λ ρ u(λ2 t, λx) is a solution to the same equation with initial data
2
λ ρ u0 (λ·). This shows that the scaling of the space we encountered in (18.40)
is the correct one (up to being a inhomogeneous Besov space).
Specialising to the case dq − ρ2 = 0 and taking p large enough, we also see
that one can consider initial data from Lq (Rd ), as this spaces embeds into
0
Bq,p (Rd ). This space has the same scaling behaviour as just discussed.

In (18.40) the limiting case where q = 21 dρ(ρ + 1) shows that we can ‘almost’
−2/(ρ+1)
treat initial data from the space Bq,p (Rd ). The less important so-called
microscopical tuning parameter p in (18.40) needs to be so large that (18.37)
holds.
Unlike in Example 18.1.3, it now becomes possible to take the special
structure of f into account. The space of initial data which we could consider
s+2− 2
in the example was Bq,p p (Rd ) with s ∈ (−2, 0] and s + 2 − p2 − dq > 0. Under
these restrictions, the smoothness parameter satisfies s+2− p2 > dq , which leads
to a much smaller class of initial data than considered in (18.40). Introducing
weights in the set-up of Example 18.1.3, does not change anything.

Remark 18.3.3. When Rd is replaced by a bounded domain, the condition

(18.34) on f in Example 18.3.1 can be weakened to

|f 0 (t)| 6 `(1 + |t|ρ ), t ∈ R.

Indeed, the step where Hölder’s inequality is used can then be replaced by

k(1 + |u|ρ + |v|ρ )(u − v)kr 6 C`(1 + kukρ(ρ+1)r + kvkρ(ρ+1)r )ku − vk(ρ+1)r .

Similarly, one can check that f (0) is allowed to be non-zero.

We finish this section with an example illustrating how Theorems 18.2.6 and
18.2.15 can be applied to obtain local and global well-posedness for certain
concrete PDEs.

Example 18.3.4 (Local well-posedness for the Allen-Cahn equation). On Rd

with d > 2 (the case d = 1 can be included by making subcritical choices) we
consider the so-called Allen-Cahn equation
720 18 Nonlinear parabolic evolution equations in critical spaces
(
∂t u − ∆u = −u3 + u,
(18.41)
u(0) = u0 .

This equation fits into the setting discussed in Example 18.3.1 with X0 =
H s,q (Rd ) and X1 = H s+2,q (Rd ) for suitable (q, s). Indeed, taking ρ = 2, one
checks that (18.39) holds if 1 < q < 3d. Let s ∈ (−2, 0] be such that (18.38)
holds with ρ = 2, and set σ := 21 ( dq − s) − 12 . Choose p ∈ (1, ∞) so large that
(18.37) holds. Then, by Example 18.3.1, F (u) = −u3 satisfies the Assumption
18.2.2. We choose to include the linear part of −u3 + u into the operator A.
Another possibility would be to put it into F as well, and consider ρ1 = 2 and
ρ2 > 0 arbitrary small.
From Example G.5.6 it follows that for s = 0 the operator Au = −∆u − u
on X0 , with domain X1 , is sectorial of angle zero. Moreover, by Theorems
17.4.1 and 17.2.26, A has maximal Lp -regularity on finite time intervals for
all p ∈ (1, ∞). Since the Bessel potentials (1 − ∆)t/2 commute with ∆, the
maximal Lp -regularity extends to the full range s ∈ R.
From now on we view (18.41) as an abstract problem of the form (18.13).
In particular, we say that (18.41) admits a (maximal) (p, q, s, σ)-solution if
(18.13) has a (maximal) Lpwα -solution. Applying Theorems 18.2.6 and 18.2.14,
d
−1
q
it follows that for every u0 ∈ Oσ,p = Xσ,p = Bq,p (Rd ) (see (18.40)), the prob-
lem (18.41) admits a maximal (p, q, s, σ)-solution (u, Tmax (u0 )). Moreover,

u ∈ Ww1,p
α
(0, T ; H s,q (Rd )) ∩ Lpwα (0, T ; H s+2,q (Rd ))
d
−1 2
s+2− p
(18.42)
q
∩ C([0, T ]; Bq,p (Rd )) ∩ C([τ, T ]; Bq,p (Rd ))

for all 0 < τ < T < Tmax (u0 ), where we used the instantaneous regularisation
stated in (18.15).
Global well-posedness can often be obtained via Theorem 18.2.17, but to
apply it to the rough initial data considered in the above example requires
first performing a (weighted) bootstrap argument to obtain enough regular-
ity in space and time. After that, suitable energy estimate can be applied.
Bootstrapping regularity will not be discussed here (a concise discussion of
this technique is included in the Notes). Instead, we will only prove global
well-posedness for sufficiently smooth initial data. This is done in the next
example. In particular, all initial data u0 ∈ Lq (Rd ) for q ∈ (d, 2d) are covered
if d ∈ {2, 3, 4, 5, 6}.
Example 18.3.5 (Global well-posedness for the Allen-Cahn equation). Consider
again the problem (18.41) in dimension d > 2. In order to obtain that u
takes values in H 1,q (Rd ), the smallest value of s which we can allow (without
bootstrapping) is s = −1. Let q ∈ ( d2 , 2d) and p ∈ (2, ∞) are such that
d 2 d 1
q + p 6 2 (see (18.37)), and set ρ := 2, σ := 2q , and α := 1 − p − σ. These
choices form a special case of Example 18.3.4, and in particular they lead to
a critical setting.
 18.3 Examples and comparison 721
d
−1
Let u0 ∈ Bq,p q
(Rd ); note that this space contains Lq (Rd ) if q > d. By
the result of Example 18.3.4, the problem (18.41) admits a (unique) maximal
(p, q, s, σ)-solution, and for all 0 < τ < T < Tmax (u0 ) we have
1− 2
u ∈ Lp (τ, T ; H 1,q (Rd )) ∩ C([τ, T ]; Bq,p p (Rd )).

We will show global existence, i.e., Tmax (u0 ) = ∞, under the more restrictive
conditions
2q
max{d, 2d − 6} < q < 2d and 2 < p 6 . (18.43)
2d − q

For d = 2 we can take q ∈ (2, 4) and p ∈ (2, 2q/(4 − q)]. For d = 3 we can
take q ∈ (3, 6) and p ∈ (2, 2q/(6 − q)]. We do not claim this is optimal, and we
expect that by further bootstrapping some of these conditions can be omitted.
Step 1 – Assuming that Tmax (u0 ) < ∞, we will derive a contradiction
with Theorem 18.2.17(1). For the latter it suffices to use Step 2 below. How-
ever, we prefer to show the techniques to check Theorem 18.2.17(1) since
this can be useful for other situations. This boils down to showing that
u ∈ Lpwα (0, T ; H 1,q (Rd )) and

sup ku(t)kq
d −1 < ∞.
t∈[0,Tmax (v0 )) Bq,p (Rd )

By (18.42), both assertions are clear on [0, τ ] for any τ < Tmax (v0 ). Thus it
suffices to show that, for some τ > 0,

u ∈ Lp (τ, Tmax (u0 ); H 1,q (Rd )) and sup ku(t)k q

d −1 < ∞.
t∈[τ,Tmax (u0 )) Bq,p (Rd )
(18.44)

Step 2 – We show the second part of (18.44). Since dq − 1 < 0, by the easy
embeddings of (14.23) and Proposition 14.4.18, it is enough to show that

sup ku(t)kLq (Rd ) < ∞.

t∈[τ,Tmax (u0 ))

The idea will be to apply the chain rule of Lemma 18.3.6 below. For this we
need that u3 ∈ L1 (τ, T ; Lq ) for 0 < τ < T < Tmax (u0 ). To see this, note that
by Sobolev embedding with θ − dq = − 3q d
and interpolation,

3(1−θ)
ku3 kLq = kuk3L3q 6 Ckuk3H θ,q 6 C 0 kukLq kuk3θ
H 1,q .

As observed before, the Lq -norm of u is uniformly bounded on [τ, T ]. Thus for

the integrability of ku3 kLq in time it remains to note that u ∈ Lp (τ, T ; H 1,q ) ,→
L3θ (τ, T ; H 1,q ) since p > 2 > 2d
q = 3θ.
Applying the chain rule to the identity
722 18 Nonlinear parabolic evolution equations in critical spaces
Z t Z t
u(t) − u(τ ) = ∆u(r)dr + −u3 (r) + u(r) dr, t ∈ [τ, T ],
τ τ

we see that
Z tZ
q q
ku(t)kL q (Rd ) = ku(τ )kLq (Rd )
− q(q − 1) |u|q−2 |∇u|2 dx dr
τ R d
Z tZ
+q |u|q−2 (−u4 + u2 ) dx dr (18.45)
τ Rd
Z t
q q
6 ku(τ )kLq (Rd ) + q ku(r)kL q (Rd ) dr.
τ
q
Therefore, by Gronwall’s lemma applied to t 7→ ku(t)kL q (Rd ) ,

q q q(t−τ )
ku(t)kL q (Rd ) 6 ku(τ )kLq (Rd ) e .

Since we assumed Tmax (u0 ) < ∞, this implies the desired bound
q q qTmax (u0 )
N := sup ku(t)kL q (Rd ) 6 ku(τ )kLq (Rd ) e < ∞. (18.46)
t∈[τ,Tmax (u0 ))

As a consequence of (18.45), we also find that

Z Tmax (u0 ) Z
|u|q−2 |∇u|2 dx dr 6 Cq,Tmax (u0 ) ku(τ )kqLq (Rd ) , (18.47)
τ Rd

(1+qTmax (u0 )) q(Tmax (u0 ))

where Cq,Tmax (u0 ) = q(q+1) e .
1− 2
Step 3 – By (18.42) we have u(τ ) ∈ Bq,p p = (X0 , X1 )1− p1 ,p . Therefore,
if we can show that −u3 + u belongs to Lp (τ, Tmax (u0 ); H −1,q (Rd )), the first
part of (18.44) follows from maximal Lp -regularity applied on the interval
(τ, Tmax (u0 )) with inhomogeneity u − u3 .
It is clear from Step 2 that u has the required regularity, so it remains to
consider the term u3 . By Sobolev embedding,

ku3 kLp (τ,Tmax (u0 );H −1,q (Rd )) 6 Cku3 k qd

Lp (τ,Tmax (u0 );L q+d (Rd ))

= Ckuk3L3p (τ,Tmax (u0 );Lq0 (Rd )) ,

3qd
where q0 = q+d . To prove that the latter is finite, note that by Sobolev
d dq
embedding with θ − 2 = − 2q 0
(then θ ∈ (0, 1] by (18.43) and 2q0 /q > 2 since
q < 2d),
q/2
kukLq0 (Rd ) = k|u|q/2 kL2q0 /q
6 C0 k|u|q/2 kH θ,2
 18.3 Examples and comparison 723
1−θ
6 C1 k|u|q/2 kL 2 (Rd ) k|u|
q/2 θ
kW 1,2 (Rd )
6 C2 k|u| kL2 (Rd ) + k|u|q/2 k1−θ
q/2
k∇|u|q/2 kL
θ
 
2 (Rd )
L2 (Rd )
θZ θ/2 i
q(1−θ)/2 q
h
q/2
= C2 kukLq (Rd ) + kukLq (Rd ) θ |u|q−2 |∇u|2 dx .
2 Rd
qθ 
h Z θ/2 i
6 C2 N 1/2 + N (1−θ)/2 θ |u|q−2 |∇u|2 dx ,
2 Rd

where we used (18.46). Therefore, u ∈ L3p (τ, Tmax (u0 ); Lq0 (Rd )) follows if we
can check that
Z Tmax (u0 ) Z 3pθ/q
|u|q−2 |∇u|2 dx dt < ∞.
τ Rd
q
The latter follows from (18.47) since our choice of θ satisfies θ 6 3p , which
follows from (18.43).

The following chain rule was used in Example 18.3.5.

Lemma 18.3.6 (Chain rule in the weak setting). Let q ∈ [2, ∞) and
p ∈ (1, ∞). Suppose that u ∈ C([τ, T ]; Lq (Rd )) ∩ Lp (τ, T ; H 1,q (Rd )), G ∈
0
Lp (τ, T ; Lq (Rd ; Rd )), and g ∈ L1 (τ, T ; Lq (Rd )) are such that for all t ∈ [τ, T ]
Z t Z t
u(t) = u(τ ) + ∇ · G(s) ds + g(s) ds, (18.48)
τ τ

where the equality is meant in the space H −1,q (Rd ). Then

Z t
q q
ku(t)kLq (Rd ) = ku(τ )kLq (Rd ) −q(q − 1) hG(s), |u(s)|q−2 ∇u(s)i ds
τ
Z t (18.49)
q−2
+q hg(s), |u(s)| u(s)i ds,
τ
0
where the duality pairing is in (Lq , Lq ) in both cases.

In view of the Mihlin multiplier theorem (see Theorem 5.5.10),

2πiξ
k∇ · GkH −1,q (Rd ) = F −1 [ξ 7→ · G(ξ)]
b 6 Cp,d kGkLq (Rd ;Rd ) ,
(1 + |ξ|2 )1/2 Lq (Rd )

and therefore the integral of ∇ · G exists as a Bochner integral in H −1,q (Rd ).

Proof. Without loss of generality we may assume that τ = 0. First we es-
tablish some boundedness properties which also show the well-definedness
of the integrals appearing in (18.49). For all v ∈ L∞ (0, T ; Lq (Rd )) and
w ∈ Lp (0, T ; Lq (Rd )), by Hölder’s inequality in the space variables with
1 q−2 1 1 1
q + q + q = 1, and subsequently in the time variable with p + p0 = 1,
724 18 Nonlinear parabolic evolution equations in critical spaces
Z T Z
|G| |v|q−2 |w| dx ds
0 Rd
Z T
q−2 (18.50)
6 kGkLq (Rd ;Rd ) kvkL q (Rd ) kwkLq (Rd ) ds
0
6 kvkq−2
L∞ (0,T ;Lq (Rd ))
kGkLp0 (0,T ;Lq (Rd ;Rd )) kwkLp (0,T ;Lq (Rd )) .

In a similar way one proves that

Z T Z
q−1
|g| |v|q−1 dx ds 6 kgkL1 (0,T ;Lq (Rd )) kvkL ∞ (0,T ;Lq (Rd )) .
0 Rd

Let ϕ ∈ Cc∞ (Rd ) be such that Rd ϕ dx = 1, and let ϕn := nd ϕ(n·). By

R
1
Theorem 2.3.8, for all f ∈ Lloc (Rd ) we have

ϕn ∗ f → f and |ϕn ∗ f | 6 M f almost everywhere, (18.51)

where M denotes the Hardy–Littlewood maximal operator.

Taking convolutions in (18.48), we obtain
Z t Z t
un (t) − un (0) = ∇ · Gn (s) ds + gn (s) ds,
0 0

where un = ϕn ∗ u, ∇Gn = ∇ · (ϕn ∗ G) = ϕn ∗ (∇ · G), and gn = ϕn ∗ g.

Fix x ∈ Rd and let R > 0 be so large that |u(s, x)| 6 R for all s ∈ [0, T ]. Let
ζ ∈ Cc2 (R) be such that ζ(y) = |y|q for |y| 6 R. Note that ζ 0 (y) = |y|q−2 y
and ζ 00 (y) = |y|q−2 for |y| 6 R. Applying the chain rule for weak derivatives
in time to the function t 7→ ζ(u(t, x)), we obtain
Z t
q q
|un (t, x)| = |un (0, x)| + q |un (s, x)|q−2 un (s, x)∇ · Gn (s, x) ds
0
Z t
+q q|un (s, x)|q−2 un (s, x)gn (s, x) ds.
0

Integrating over Rd and using Fubini’s theorem and integrating by parts, we

obtain
Z t
q q
kun (t)kL q (Rd ) = ku n (0)k Lq (Rd )
− q(q − 1) hGn (s), |un (s)|q−2 ∇un (s)i ds
0
Z t
+ qhgn (s), |un (s)|q−2 un (s)i ds.
0

From the observation (18.51) we deduce that un → u in Lq (Rd ) pointwise in

0
[0, T ], un → u in Lp (0, T ; H 1,q (Rd )), Gn → G in Lp (0, T ; Lq (Rd )), and gn → g
in L1 (0, T ; Lq (Rd )). Thus it remains to let n → ∞ in the above identity and
use the boundedness/continuity properties from the beginning of the proof
 18.4 Long-time existence for small initial data and F = Fc 725

to obtain convergence. Indeed, after extracting almost everywhere convergent

subsequences and relabelling, convergence follows by dominated convergence.
For instance, for the first term,
Z TZ Z TZ
|un |q−2 Gn · ∇un dx ds → |u|q−2 G · ∇u dx ds
0 Rd 0 Rd
q−2
follows since |un | Gn · ∇un is dominated by |M u|q−2 M G M |∇u|, which is
an integrable function by (18.50) and the boundedness of M on Lq (Rd ). 

18.4 Long-time existence for small initial data and

F = Fc
Short-time existence and uniqueness has been proved in Theorems 18.1.2 and
18.2.6 In this section we prove that, under suitable conditions, for initial val-
ues with small norm in Xσ,p one can obtain well-posedness on arbitrary long
time intervals [0, T ]. This result is typical for the semi-linear setting. The as-
sumptions on F will be similar to the ones of Section 18.2. However, we will
assume that F = Fc , F (0) = 0, and replace (18.11) by the slightly more re-
strictive condition (18.52) below. Moreover, we assume A ∈ L (X1 , X0 ), and
thus we only consider the semi-linear setting.
Theorem 18.4.1 (Semi-linear equations with small initial data). Let
p ∈ [1, ∞] and α ∈ [0, p10 ) ∪ {0}, where we take α > 0 if p = ∞. Let σ =
1 − α − p1 ∈ [0, 1/p0 ] ∩ [0, 1). Let X0 and X1 be Banach spaces such that X1 ,→
X0 with embedding constant CX > 1. Let Oσ,p ⊆ Xσ,p be an open set and
suppose that 0 ∈ Oσ,p . Let A ∈ L (X1 , X0 ) and suppose that A has maximal
Lp -regularity (C-regularity if p = ∞) on finite time intervals. Suppose that
Fc : X1 ∩ Oσ,p → X0 is such that Fc (0) = 0 and
m
ρ ρ
X
kFc (u) − Fc (v)kX0 6 Lc (kukXjβ + kvkXjβ )ku − vkXβj (18.52)
j j
j=1

1+ρ σ
for all u, v ∈ X1 ∩ Oσ,p , where βj ∈ (σ, 1), ρj > 0 are such that βj 6 1+ρjj
for j ∈ {1, . . . , m}. Then for every T ∈ (0, ∞) there exist ε > 0 such that for
each kv0 kXσ,p 6 ε, the problem
(
u0 + Au = F (u), on (0, T ),
(18.53)
u(0) = v0 ,
has a unique Lpwα -solution uv0 ∈ MRpα (0, T ). Moreover, there is a C > 0 such
that for all kv0 kXσ,p , kv1 kXσ,p 6 ε,
kuv0 − uv1 kMRpα (0,T ) 6 Ckv0 − v1 kXσ,p . (18.54)
If additionally, A has maximal Lp -regularity (C-regularity if p = ∞) on R+
and 0 ∈ %(A), then the above holds with (0, T ) replaced by R+ .
726 18 Nonlinear parabolic evolution equations in critical spaces

p
Proof. In the proof we use the notation Ej = Lw α
(0, T ; Xj ). Let u0 = 0 and
set T1 = T . Without loss of generality we may assume T > 1 and r 6 1. Let
Φv0 : BrT (v0 ) → MRpα (0, T ) be defined by Φv0 (v) := u, where u is the unique
Lpwα -solution to
(
u0 + Au = F (v),
u(0) = v0 .

Note that for r ∈ (0, 1] and ε > 0 small enough, v takes values in Oσ,p by
Lemma 18.2.9, and by Lemma 18.2.10 we have F (v) ∈ E0 . Below Theorem
18.2.6, it has already been observed that local existence and uniqueness follow
if we can show that Φv0 has a unique fixed point.
p
By the maximal regularity estimate (18.28) we have u ∈ MRα (0, T ), u(0) =
v0 , and

kukMRαp (0,T ) 6 CA,T kv0 kXσ,p + CA,T kF (v)kE0

m
ρj +1
X
6 CA,T ε + CA,T CLc kvkMR p
α (0,T )
j=1
m
X
6 CA,T ε + CA,T CLc rρj +1 ,
j=1

where the estimate for F (v) follows from Lemmas 18.2.7 and 18.2.8, the con-
stant C can be taken T -independent since T > 1, and we used (18.19) with
u0 = 0 and zu0 = 0. Note that the terms T δj can be avoided due to the more
restrictive condition (18.52). The above estimate shows that for r, ε > 0 small
enough, kukMRpα (0,T ) 6 r, and thus u ∈ BrT (v0 ).
Next, fix vj,0 ∈ BXσ,p (u0 , ε) and vj ∈ BrT (vj,0 ) for j ∈ {1, 2}. Then u =
Φv1,0 (v1 ) − Φv2,0 (v2 ) solves the problem
(
u0 + Au = F (v1 ) − F (v2 ),
u(0) = v1,0 − v2,0 .

Therefore, by the maximal regularity estimate (18.28),

kukMRαp (0,T ) 6 CA,T kF (v1 ) − F (v2 )kE0 + CA,T kv1,0 − v2,0 kXσ,p ,

From Lemmas 18.2.7 and 18.2.8 we obtain that

kF (v1 ) − F (v2 )kE0

m
ρj ρj
X  
6 CLc kv1 kMR p
α (0,T )
+ kv2 kMR p
α (0,T )
kv1 − v2 kMRαp (0,T )
j=1
Xm
6 2CLc rρj kv1 − v2 kMRpα (0,T ) .
j=1
 18.5 Notes 727

Therefore, by choosing r > 0 small enough,

1
kΦv1,0 (v1 ) − Φv2,0 (v2 )kMRαp (0,T ) 6 kv1 − v2 kMRpα (0,T )
2 (18.55)
+ CA,T kv1,0 − v2,0 kXσ,p .

By (18.55), Φv0 : BrT (v0 ) → BrT (v0 ) is a strict contraction, and thus it has
a unique fixed point uv0 ∈ BrT (v0 ). This is the required solution to (18.53).
Moreover, (18.55) implies that for v1,0 , v2,0 ∈ BXσ,p (u0 , ε)

1
kuv1,0 − uv2,0 kMRpα (0,T ) 6 kuv1,0 − uv2,0 kMRpα (0,T )
2
+ CA,T kv1,0 − v2,0 kXσ,p ,

and thus
kuv1,0 − uv2,0 kMRpα (0,T ) 6 2CA,T kv1,0 − v2,0 kXσ,p .
which gives (18.54).
In case A has maximal regularity on R+ and 0 ∈ %(A), then (18.28) holds
with (0, T ) replaced by R+ . Moreover, one can check that Lemma 18.2.9 holds
with (0, T1 ) replaced by R+ . Therefore, one can repeat the above argument
on the half line. 

18.5 Notes

The theory of abstract non-linear parabolic evolution equations has a long

history going back to the work of the Japanese school in the 1960s, with con-
tributions of Fujita, Kato, Tanabe, and others. Excellent monographs on the
subject are available, including Amann [1995], Friedman [1969], Henry [1981],
Lunardi [1995], Lions [1969], Pazy [1983], Prüss and Simonett [2016], Tanabe
[1979], Yagi [2010]. For the purpose of this chapter we chose to limit ourselves
to the maximal Lp -regularity approach to quasi-linear evolution equations,
mostly focussing on local well-posedness. Other approaches, including max-
imal Hölder regularity, the so-called Kato approach, and the theory mono-
tone operators, are treated in some of the references just mentioned. Maximal
regularity techniques have important applications to a number of topics not
covered in this volume, such as linearised stability, semi-flows, higher order
regularity, sharp conditions for global well-posedness, numerical analysis, and
applications to concrete PDEs. Maximal Lp -regularity for stochastic evolution
equations will be covered in Volume IV.
The maximal Lp -regularity approach to quasi-linear evolution equations
was initiated by the influential paper Clément and Li [1993/94], and further
investigated and extended in Prüss [2002] and Amann [2005]. The semi-linear
setting of Theorem 18.1.2 is a special case of the results in these works, and
is presented here as a warm-up to the later results.
728 18 Nonlinear parabolic evolution equations in critical spaces

A local well-posedness result under the assumption of maximal C-regularity

was obtained by Clément and Simonett [2001]. The use of weights in time
seems essential in the latter (see Remark 18.2.11). Based on the weighted
maximal Lp -regularity result of Prüss and Simonett [2004], the Lp -setting was
extended to a weighted setting in time by Köhne, Prüss, and Wilke [2010].
The use of weights has several advantages:
• well-posedness in case of rough initial data
• instantaneous regularisation
• compactness properties of orbits
and has become standard in the theory of evolution equations. All of the above
works have found applications to concrete quasi- and semi-linear PDEs, many
of which are collected and mentioned in the influential monograph Prüss and
Simonett [2016]. Since the number of applications is too large to discuss here,
we will mainly focus on applications of the theory of critical spaces in these
notes.
For parabolic equations it is often possible to bootstrap regularity in time
and space. Sometimes one can even derive real analyticity via use the so-
called parameter trick of Angenent [1990b,a]; see also Prüss and Simonett
[2016, Section 5.2] for a presentation in the setting of abstract quasi-linear
evolution equations. Applications of maximal Lp -regularity techniques to the
study of linearised stability for non-linear parabolic evolution equation can
be found in Lunardi [1995], Prüss [2002], Prüss, Simonett, and Zacher [2009],
Matioc and Walker [2020], and references therein.

Critical spaces

In the present abstract evolution equations framework, the splitting F =

FTr +Fc was first introduced in LeCrone et al. [2014]. In this paper, local well-
posedness in the subcritical case was proved using maximal Lp -regularity for
1 < p < ∞. Shortly afterwards, it was realised in Prüss and Wilke [2017] that
under additional conditions on A and (X0 , X1 ), local well-posedness can even
be obtained in the critical case. Consequences for the Navier–Stokes equations
were discussed in Prüss and Wilke [2018]. Further results and applications to
concrete and abstract problems were given in Prüss, Simonett, and Wilke
[2018]. In particular, this paper discusses the relationship between scaling
invariance and criticality for several concrete PDEs. It is remarkable that an
abstract definition for criticality can be given which leads to new insights for
many concrete PDEs. In the same paper, by way of an example it is shown that
the sub-criticality condition (18.12) cannot be improved. The Lp -framework
was extended to maximal C-regularity in LeCrone and Simonett [2020].
Theorem 18.2.6 unifies and extends several of the results mentioned in
the preceding discussion. For simplicity, here we only considered the case
where A and F are time-independent, but this restriction can be avoided
easily (see Remark 18.2.4). The unification lies in the fact that one proof is
 18.5 Notes 729

presented which works for all p ∈ [1, ∞] and all admissible weights, with p = ∞
corresponding to maximal C-regularity. Moreover, we do not need geometric
conditions on X0 such as the UMD property, or further conditions on A(u0 )
besides maximal Lp - or C-regularity. In part of the existing literature, the
spaces Xβj appearing in (18.11) are taken as the complex interpolation spaces
[X0 , X1 ]βj . Taking the real interpolation spaces (X0 , X1 )βj ,1 leads to a less
restrictive condition on Fc and is easier to work with in the proofs.
The case p = 1 of Theorem 18.2.6 seems to be new. It is important to
observe that for p = 1 one is forced to take σ = α = 0, which in turn forces
the X0 -valued trace part FTr to be defined on an open subset Oσ,p of the
same space X0 . For non-linearities of the form F = FTr , this requirement rules
out many interesting examples of non-linearities. However, by allowing non-
linearities with a critical part, i.e., non-linearities of the form F = FTr + Fc ,
many interesting examples can be covered even when p = 1, the point being
that it suffices to have Fc locally Lipschitz with respect to the norms of the
smaller spaces X1/(1+ρj ) (with the ρj ’s as in Assumption 18.2.2). On the other
hand, according to Theorem 17.4.5, operators with maximal L1 -regularity are
rare. An exception is the case where X0 itself is a real interpolation space in
which case the Da Prato–Grisvard theorem applies (see Corollary 17.3.20).
It should be observed that a more flexible condition on Fc could be used
in (18.11), namely
m
ρ ρ
X
kFc (u) − Fc (v)kX0 6 Lc (1 + kukXjϕ + kvkXjϕ )ku − vkXβj , (18.56)
j j
j=1

with ϕj ∈ (σ, 1), βj ∈ (σ, ϕj ], along with the subcriticality condition

ρj (ϕj − σ) + βj 6 1, j ∈ {1, . . . , m}. (18.57)

The formulation (18.56) allows for different space regularity for u, v, and u−v
on the right-hand side (see Agresti and Veraar [2022a] and Prüss, Simonett,
and Wilke [2018]). However, in all known examples, it suffices to take ϕj = βj
(as we do in the main text) in order to obtain the sharpest results. Note that
by taking ϕj = βj , (18.57) reduces to the sub-criticality condition (18.12).

Global well-posedness and blow-up criteria

The existence of a maximal time interval in Theorem 18.2.14 is a standard re-

sult. Often it is only stated and proved under the more restrictive assumption
that A(v0 ) have maximal Lp - or C-regularity for all v0 ∈ Oσ,p . The present for-
mulation only uses maximal regularity of A(u0 ). In a slightly different set-up
it appears in Agresti and Veraar [2022a].
The global well-posedness result of Theorem 18.2.15 is also standard. The
statement and proof closely follow Prüss and Simonett [2016, Corollary 5.1.2].
The weight tα can be helpful in proving global well-posedness, as estimates in
the space Xσ,p are easier to obtain for smaller values of σ (i.e., for higher values
730 18 Nonlinear parabolic evolution equations in critical spaces

of α). In the semi-linear case, the blow-up criteria can be further weakened as
was done in Theorem 18.2.17. In case of semi-linear functions F of quadratic
type, blow-up criteria appear in Prüss, Simonett, and Wilke [2018, Section
2.1]. Some of these were extended, for a more general class of semi-linearities
F , in to a stochastic setting in Agresti and Veraar [2022b, Theorem 4.11].
Simplifying this to the deterministic setting, one arrives at the following result:

Theorem 18.5.1 (Serrin-type blow-up criteria). Let p ∈ (1, ∞), suppose

that Assumption 18.2.2 holds, and let A ∈ L (X1 , X0 ) have maximal Lp -
regularity on finite time intervals. Let (u, Tmax (u0 )) denote the maximal Lpwα -
solution to
(
u0 + Au = F (u), on (0, T ),
u(0) = u0 .

Suppose that for each j ∈ {1, . . . , m} we have

ρj < 1 + αp or (α = 0 and ρj 6 1).

Then the following assertions hold:

• for all T < Tmax (u0 ) one has kukLp (0,T ;X1−α ) < ∞;
• if Tmax (u0 ) < ∞, then kukLp (0,Tmax (u0 );X1−α ) = ∞.

In the case of (sub-)quadratic semi-linearity F , one has ρj 6 1 and the above

condition always holds.

Applications

The theory of quasi-linear evolution equations in critical spaces as presented

in this chapter has been applied to models in several scientific areas which
include fluid dynamics, chemistry, neuroscience, free boundary problems, and
differential geometry. For details we refer to the founding papers and books
LeCrone, Prüss, and Wilke [2014], Prüss and Simonett [2016], Prüss and Wilke
[2018], Prüss, Simonett, and Wilke [2018], LeCrone and Simonett [2020], and
for further applications to the more recent papers Hieber and Prüss [2018],
Mazzone, Prüss, and Simonett [2019a,b], Simonett and Prüss [2019], Binz,
Hieber, Hussein, and Saal [2020], Giga, Gries, Hieber, Hussein, and Kashi-
wabara [2020], Hieber, Hussein, and Saal [2023], Hieber, Kress, and Stinner
[2021], Mazzone [2021], Prüss, Simonett, and Wilke [2021], Court and Ku-
nisch [2022], Simonett and Wilke [2022b]. This list is likely to expand in the
near future, as the splitting F = FTr + Fc of Theorem 18.2.6 has proved to
be very powerful in applications to concrete non-linear parabolic equations of
semi- and quasi-linear type. It leads to new insights for many PDEs to which
the original framework of Clément and Li [1993/94] was applicable. More-
over, some of the new blow-up criteria can make it possible to obtain global
well-posedness results.
 18.5 Notes 731

The examples considered in Section 18.3 are very basic, and local/global
well-posedness is well known for a broad class of initial values. The examples
are merely chosen to demonstrate the abstract theorems of Section 18.2 in a
simple setting. The method to check the blow criteria in Example 18.3.5 is
taken from Agresti and Veraar [2023a], where these techniques are used in
several examples.
An extension of the results of Section 18.2 to stochastic quasi-linear evo-
lution equations in critical spaces was recently obtained in Agresti and Ver-
aar [2022a,b], where completely new proofs where required. Applications to
stochastic PDE can be found in these works, as well as in Agresti and Ver-
aar [2021, 2022c, 2023b,a], Agresti [2022], Agresti, Hieber, Hussein, and Saal
[2022a,b].
Q
Questions

Calderón–Zygmund operators
The extrapolation theory for the boundedness of Calderón–Zygmund opera-
tors developed in Chapter 11 is in many ways analogous and parallel to the
extrapolation of Lp inequalities for martingale transforms that we discussed
in Section 3.5. Specifically, the quantitative statement of Calderón–Zygmund
Theorem 11.2.5(3) is analogous to the estimate (3.48) of Martingale Extrapo-
lation Theorem 3.5.4; in both cases, the Lp norm of an operator is controlled
by the Lq norm multiplied by the factor pp0 = p+p0 , which exhibits the correct
blow up of these norms as p → 1 or p → ∞. However, in the case of martingale
transforms with respect to a Paley–Walsh filtration, (3.49) gives a more pre-
0
cise estimate with the factor pq + pq0 . While this has the same rate of blow up
as p → 1 or p → ∞, it gives a better estimate if the “starting point” q is either
large or close to 1. (For instance, think of the case that q is large and p = 2q.)
Given the well-behaved nature of the Lebesgue measure with respect to which
the Calderón–Zygmund singular integrals are integrated, it seems reasonable
to expect that the behaviour of these operators should be as good as that of
martingale transforms with respect to a Paley–Walsh filtration; however, this
is not reflected in the quantitative estimates of Calderón–Zygmund Theorem
11.2.5. We therefore pose the question:
Problem Q.1. Under the assumptions of Calderón–Zygmund Theorem 11.2.5
(or even just for more regular operators with a standard kernel), is there an
estimate
?  p p0  
kT kL (Lp (Rd ;X),Lp (Rd ;Y )) 6 cd + 0 kT kL (Lq (Rd ;X),Lq (Rd ;Y )) + CK ,
q q
for all p, q ∈ (1, ∞), where cd depends only on the dimension and CK only on
the kernel K of the operator T , either via the quantities kKkHör and kKkHör∗
from Definition 11.2.1 of a Hörmander kernel, or cK and ωK from Defini-
tion 11.3.1 of a Calderón–Zygmund kernel? In particular, does the Hilbert
transform satisfy an estimate
© Springer Nature Switzerland AG 2023 733
T. Hytönen et al., Analysis in Banach Spaces, Ergebnisse der Mathematik und ihrer
Grenzgebiete. 3. Folge / A Series of Modern Surveys in Mathematics 76,
https://doi.org/10.1007/978-3-031-46598-7
734 Q Questions
? p p0 
~p,X 6 C + ~q,X , ~p,X := kHkL (Lp (R;X)) (Q.1)
q q0
with some universal constant C?
The prospective estimate (Q.1) would be an analogue of Theorem 4.2.7, which
gives a similar bound for the UMD constants βp,X and βq,X in place of ~p,X
and ~q,X .
Recall that Problem O.6 asks about a possible linear dependence between
βp,X and ~p,X for p = 2, which remains wide open. Note that if this linear
relation was true for all p ∈ (1, ∞), and moreover with constants indepen-
dent of p, then (Q.1) would immediately follow via this linear relation from
Theorem 4.2.7 for the UMD constants. Thus, (Q.1) could be thought of as a
simpler model problem related to the presumably difficult Problem O.6.

Sparse domination and weighted inequalities

As we have seen in Chapter 11, and many more examples can be found in
the literature, sparse domination of operators is very efficient in capturing
essential information about their boundedness properties on various function
spaces, particularly in view of sharp weighted norm inequalities. As discussed
in the Notes of that Chapter, convex body domination provides a useful elab-
oration in view of applications like matrix-weighted inequalities and commu-
tator estimates. The proofs of existing convex body domination results follow
the same broad outline as their sparse domination counterparts, but require
elaborations at critical points of the argument. Rather than reworking each
sparse domination proof for the convex body improvement, it would be useful
to have a general statement guaranteeing that one implies the other—or, to
know that such a statement is impossible, justifying the need of case-by-case
study. This raises the question:
Problem Q.2. Does sparse domination imply convex body domination?
More precisely, let 1 6 p < r < q 6 ∞ and
T ∈ L (Lr (Rd ; X), Lr (Rd ; Y )).
For each N ∈ Z+ , consider the following property:
For some constants ε ∈ (0, 1) and α, C ∈ [1, ∞), for all f = (fn )N n=1 ∈
L∞c (R d
; X) N
and g = (g ) N
n n=1 ∈ L ∞
c (R d
; Y ∗ N
) (or some other suitable
test function spaces), there is an ε-sparse collection S ⊆ D such that
|hT f, gi|
N
(Q.2)
X C ZZ X
6 sup hfn (s), φ(s)ihψ(t), gn (t)i ds dt,
|Q| φ,ψ αQ×αQ n=1
Q∈S
0
where the supremum is over all φ ∈ Lp (αQ; X ∗ ) and ψ ∈ Lq (αQ; Y )
such that
 Q Questions 735
Z Z
p0
− kφ(s)kX ∗ ds 6 1, − kψ(t)kqY dt 6 1.
αQ αQ

If this property holds for N = 1, does it follow (a) for N = 2, or even (b) for
all N ∈ Z+ ? If not, what is a counterexample? Does the implication hold
• under additional assumptions on the operator T ?
• with a relaxed conclusion with p + ε and q − ε in place of p and q?

For N = 1, the estimate (Q.2) reduces, after elementary duality results, to

sparse domination in the form
X Z 1/p Z 0
1/q0
|hT f, gi| 6 C 0 |Q| − kfn (s)kpX dt − kgn (t)kqY ∗ dt .
Q∈S αQ αQ

For N > 1, the estimate (Q.2) corresponds to convex body domination of

Banach space valued functions in the sense of Hytönen [2023]. While we have
stated Problem Q.2 in this generality, it is open even for X = Y = R.
Recall that Rubio de Francia’s extrapolation Theorem J.2.1 gives a way
to extend the boundedness of T on Lp (Rd , w) for all w ∈ Ap with p ∈ (1, ∞)
fixed, to boundedness on Lq (Rd , w) for all w ∈ Aq and all q ∈ (1, ∞).
The following problem has been wide open for many years.
Problem Q.3 (Rubio de Francia extrapolation in UMD spaces). Let
X be a UMD Banach space and p ∈ (1, ∞). Suppose T ∈ L (Lp (w)) for any
w ∈ Ap and there exists an increasing function φ : [1, ∞) → [1, ∞) such
that kT kL (Lp (w)) 6 φ([w]Ap ). Does the tensor extension T ⊗ IX extend to an
element in L (Lp (Rd ; X)).
The tensor extension is indeed bounded if X is a UMD Banach function spaces,
and this was proved by Bourgain [1984] and Rubio de Francia [1986] by using
a variation of the Rubio de Francia algorithm. Weighted and rescaled exten-
sions have been obtained in Amenta, Lorist, and Veraar [2019]. An extension
to the multilinear limited range setting was obtain in Lorist and Nieraeth
[2019]. A solution to the stated problem needs different techniques, because
the known proofs in UMD Banach function spaces eventually boil down to a
scalar estimate.
If more structure is assumed on a single operator T , then often the UMD
property of X is enough to deduce the boundedness of T ⊗ IX from the
boundedness of T . Typical examples are:
• T = Tm , where m : Rd \ {0} → C satisfies Mihlin’s conditions;
• T = t is an element in L (Ls (Rd )) in Theorem 12.4.21 with s ∈ (1, ∞)
(and say t = 1 and q = ∞).
Recently, Lorist and Nieraeth [2022] showed that for UMD Banach func-
tion spaces X, sparse domination of T implies, sparse domination of T ⊗ IX .
736 Q Questions

Since sparse domination implies (weighted) boundedness, this leads to the fol-
lowing natural question whether their results remains true for general UMD
spaces. In particular this would provide a subclass of operators for which we
can answer Problem Q.3.
Problem Q.4. Let X be a UMD space and let p ∈ (1, ∞). Let ε > 0, and
T ∈ L (Lp (Rd )) be an operator such that for any f, g ∈ Lc∞ (Rd ) there exists
a ε-sparse collection of cubes S such that
Z XZ Z
|T f | · |g| dx 6 − |f | dx− |g| dx |S|.
Rd S∈S S S

Does the tensor extension T ⊗ IX extend to an element of L (Lp (Rd ; X))?

Moreover, does sparse domination hold for T ⊗ IX ? More precisely, does
there exist a δ > 0 and constant CX such that for all simple functions
f ∈ L∞c (R , X) and g ∈ Lc (R ) there exists a δ-sparse collection of cubes S
d ∞ d

such that
Z XZ Z
kT ⊗ IX f kX · |g| dx 6 CX − kf k dx− |g| dx |S|?
Rd S∈S S S

Paraproducts

In Corollary 12.2.19, we have proved in particular (we now consider dimension

d = 1 for simplicity) that the paraproduct
X
Πb f := hb, hI ihf iI hI ,
I∈D

with a scalar-valued symbol b ∈ BMO(R), extends to a bounded operator

on Lp (R; X) for all p ∈ (1, ∞), if X is a UMD space. (As discussed in the
Notes of Chapter 12, Figiel [1990] attributes this result to unpublished work
of J. Bourgain.) Although we have not treated it in the text, one can show
by adapting the same argument (which we leave as an exercise to the reader)
that the same conclusion is also valid if X has martingale type 2 (see Section
3.5.d). Since this class is incomparable with UMD, this raises the natural
question:

Problem Q.5. What is the largest class of Banach spaces X such that

kΠb f kL2 (R;X) 6 CkbkBMO(R) kf kL2 (R;X)

for all b ∈ BMO(R) and f ∈ L2 (R; X), where C is independent of b and f ?

By what we just discussed, this class should at least include all UMD spaces
and all spaces of martingale type 2.
 Q Questions 737

T (1) theorems

In the various T (1) theorems proved in Sections 12.3 and 12.4 and discussed
in the related Notes, we have seen the following:
(1) Under assumptions on the Haar coefficient of a bilinear form with respect
to a fixed dyadic system, the induced operator satisfies Lp (Rd ; X) bounds
that are cubic in the UMD constant βp,X (see Theorem 12.3.26).
(2) Under assumptions on the Haar coefficient of a bilinear form with re-
spect to an ensemble of dyadic system, the induced operator satisfies
Lp (Rd ; X) bounds that are quadratic in the UMD constant βp,X (see The-
orem 12.3.35). Such assumptions are verified by weakly defined Calderón–
Zygmund operators with standard kernels (see Theorem 12.4.12), and in
particular by extensions of scalar-valued Calderón–Zygmund operators on
Lp (Rd ) (see Theorem 12.4.21).
(3) Under additional symmetry and smoothness assumptions, the induced
operator satisfies Lp (Rd ; X) bounds that are linear in the UMD constant
βp,X . These results were not covered in the present volume, but can be
found in the works of Geiss et al. [2010] for a class of Fourier multipliers,
and of Pott and Stoica [2014] for a class of Calderón–Zygmund operators
in dimension d = 1.
A major problem is the following:

Problem Q.6. Is there an upper bound that is linear in the UMD constant for
all operators in the scope of Theorems 12.3.26, 12.3.35, 12.4.12, and 12.4.21?
If not, what is a counterexample?

Answering Problem Q.6 in the positive is expected to be very hard, given

that it remains open even for the prototype case of the Hilbert transform (see
Problem O.6). On the other hand, Problem Q.6 allows more flexibility for a
possible counterexample than Problem O.6. Such a counterexample should
still be of interest in indicating the limits of how much one can ever hope to
prove with linear dependence on the UMD constant.
We also formulate a more open-ended related task:

Problem Q.7. Can any of the bounds in (1) through (3) be improved? In par-
ticular, does T (1) Theorem 12.3.26 with a single dyadic system allow bounds
that are quadratic in βp,X ? Do Figiel’s elementary operators from Theorems
12.1.25 and 12.1.28 admit such bounds? If not, what is a counterexample?

Given the relatively wide scope of estimates that one can prove with bounds
quadratic in the UMD constant (in contrast to the somewhat restricted class of
linear estimates currently available), Problem Q.7 would appear to be more
approachable than the presumably very hard Problem Q.6. In view of the
currently different quantitative bounds in (1) and (2) above, Problem Q.7 has
a “philosophical” dimension concerning the role of random dyadic systems
738 Q Questions

in the estimation of singular integral operators: Are the methods based on

random cubes genuinely stronger, or could one also recover the same results
without resorting to random cubes?
While Problems Q.6 and Q.7 deal with the role of the UMD condition in
the boundedness of singular integral operators, one can ask related questions
about the roles of type and cotype:

Problem Q.8. Is the interdependence between the required decay of the Haar
coefficients (resp. modulus of smoothness of the kernel) and type and cotype
of the spaces in T (1) Theorems 12.3.26 and 12.3.35 (resp. 12.4.12 and 12.4.21)
sharp? If not, what are the sharp forms of these theorems? What are exam-
ples of bilinear forms or operators satisfying weaker forms of the assumptions
and failing the conclusions of these theorems concerning boundedness of the
induced operator?

This problem is relevant and open even in the scalar-valued case with type
and cotype 2, in which case the existing T (1) theorems require Figiel or Dini
norms of order 21 . Nevertheless, the fact that the Banach space valued theory
“interprets” the seemingly arbitrary number 21 as max( 1t , q10 ) with the type
and cotype exponents t and q could suggest the construction of possible coun-
terexamples through extremal situations for the type and cotype inequalities
even in the scalar-valued case.
On the side of the open-ended Problem Q.8 we pose the more provocative:

Problem Q.9. Is any of the T (1) theorems valid for all classical Dini kernels?
If not, what is an example of an operator with a Dini kernel, satisfying all
assumptions of a T (1) theorem yet failing to be bounded on L2 (Rd ) or on
L2 (Rd ; X) for some UMD space X?

In the scalar-valued T (1) theory, this is related to the investigation of the

minimal assumptions in the line of research pursued by Yabuta [1985], Meyer
[1986], Han and Hofmann [1993], Yang, Yan, and Deng [1997], and Grau de la
Herrán and Hytönen [2018].

Fourier type

According to Proposition 13.1.35, a Banach space of type p and cotype q

“close enough”, in that p1 − 1q < 12 , has Fourier type t for every t ∈ (1, r),
where 1r = p1 − 1q + 12 < 1.
By adapting an example of Bourgain [1988a], it has been shown by Garcı́a-
Cuerva, Torrea, and Kazarian [1996] that the conclusion is false, in general,
for every t > r. This is based on the bfollowing construction, for every r ∈ (1, 2)
0
and θ ∈ (0, 1). Let Xr,θ := [Lr (R), Lr (R)]θ , where
b b
0
Lr (R) := {g : g ∈ Lr (R)} ⊆ Lr (R)
 Q Questions 739
b
is equipped with the norm kgkLr (R) := kgkLr (R) , and the inclusion above
b

follows from the classical Hausdorff–Young inequality.

Garcı́a-Cuerva et al. [1996] check that Xr,θ has Fourier-type r as well as
type p and cotype q determined by
1 1−θ θ 1 1−θ θ
= + , = + ,
p 2 r q r0 2
and that none of these numbers can be improved. Here the exponents satisfy
1 1 1 1 1 1 1 1 1 1 1 1
− + = (1 − θ) − 0 +θ − + = − + = ,
p q 2 2 r r 2 2 r 2 2 r
which is exactly the relation in Proposition 13.1.35. The fact that none of these
can be improved shows in particular that Xr,θ does not have any Fourier
type t > r. On the other hand, the situation is slightly better than in the
conclusions of Proposition 13.1.35 in that Xr,θ also has Fourier type r, not
only Fourier type t ∈ (1, r). This motivates the question about the end-point
case of Proposition 13.1.35 (see Pietsch [2007, 6.1.8.6]):

Problem Q.10. If a Banach space X has type p and cotype q with p1 − 1q < 12 ,
does it follow that X has Fourier type r determined by 1r = p1 − 1q + 12 ? If not,
what is a counterexample?

A particular case of a positive answer to Problem Q.10 would be that every

space of type p and cotype 2 would have Fourier type p. If this was the
case, then case (i) of Proposition 14.5.3 (and its consequences) would already
contain case (ii), making the latter redundant. The fact that one can derive
the same conclusions under the seemingly alternative sets of assumptions in
Proposition 14.5.3 may be taken as an indication that one of these conditions
is actually contained in the other one, supporting a positive answer to Problem
Q.10.
Another general open problem is how Fourier type p and Walsh type p
are related. The definition of Walsh (co)type 2 can be found before Theorem
7.6.12, and extends to p in the obvious way. As for Fourier type, spaces X have
Walsh type p if and only if they have Walsh cotype p0 (see Garcı́a-Cuerva,
Kazaryan, Kolyada, and Torrea [1998]). Moreover, Walsh type p implies type
p and cotype p0 . In particular, spaces with Walsh type 2 are isomorphic to
a Hilbert space. The result of Bourgain [1988a] shows that nontrivial type
implies Fourier and Walsh type p for some p ∈ (1, 2]. Another result for the
corresponding operator version of the relation between Fourier and Walsh
type for p = 2 can be found in Hinrichs [2001].
Problem Q.11. Let X be a Banach space and p ∈ (1, 2). Prove or disprove
X has Fourier type p if and only if X has Walsh type p.
740 Q Questions

Function spaces and geometry of Banach spaces

In Theorem 14.7.15 we have seen that for UMD spaces 1R+ d acts as pointwise

s,p 0
multiplier on H (R; X) for p ∈ (1, ∞) and −1/p < s < 1/p. The same result
holds for general dimension d, and there exist at least three different proofs
of this fact (see Meyries and Veraar [2015], Lindemulder [2017], Lindemulder,
Meyries, and Veraar [2018]). However, it is not known whether the UMD
property or any other condition on X is necessary.
Problem Q.12. Let p ∈ (1, ∞) and s ∈ (0, 1/p). Characterise those Banach
spaces X for which 1Rd+ acts as pointwise multiplier on H s,p (Rd ; X).
k
Theorem 14.8.4, due to Hytönen and Merikoski [2019], states that Bp,p (Rd ; X)
k,p d
embeds continuously into W (R ; X) if and only if X has martingale cotype
p (see Section 3.5.d).
Problem Q.13. Let p ∈ (1, 2), k ∈ N and s ∈ R. Characterise those Banach
spaces X for which one has continuous embeddings
k
W k,p (Rd ; X) ,→ Bp,p (Rd ; X) (Q.3)
s,p d s
H (R ; X) ,→ Bp,p (Rd ; X). (Q.4)
s
A similar question can be asked with Bp,p (Rd ; X) replaced by Fp,qs
(Rd ; X) or
s d
Bp,q (R ; X). Moreover, by Proposition 14.7.8, each of the embeddings (Q.3)
and (Q.4) implies that X has type p. In case of UMD Banach spaces X, type
p is also sufficient for these embeddings. Therefore, it is natural to conjecture
that (Q.3) and (Q.4) are both equivalent to martingale type p.
Theorem 14.5.1, due to Kalton, Van Neerven, Veraar, and Weis [2008],
characterises the Banach spaces X for which the Sobolev embedding
( 1 − 12 )d
p
Bp,p (Rd ; X) ,→ γ(L2 (Rd ), X)

holds as the Banach space that have type p. A corresponding result for the
converse embedding, with ‘p’ replaced by ‘q’, characterises Banach spaces with
cotype q. It is natural to ask for similar embeddings for Bessel potential spaces
and Triebel-Lizorkin spaces.
Problem Q.14. Let p ∈ (1, 2) and r ∈ (p, ∞). Characterise those Banach
spaces X for which one has continuous embeddings
1 1
H ( p − 2 )d,p (Rd ; X) ,→ γ(L2 (Rd ), X) (Q.5)
1
(p − 12 )d
Fp,r (Rd ; X) ,→ γ(L2 (Rd ), X). (Q.6)

In Corollaries 14.6.18 and 14.7.7 it was shown that having type r > p is
sufficient. Moreover, it is known that each of the embeddings implies type p,
that and the embedding (Q.5) holds for p-convex Banach lattices; see Veraar
 Q Questions 741

[2013]. It is also known that the embedding (Q.6) for r = ∞ characterises

stable type p; see Van Neerven, Veraar, and Weis [2007] and Veraar [2013].
The validity of the embedding (Q.5) for Banach spaces X with type p would
have consequences for the Lp -Lq -multiplier theory developed in Rozendaal
and Veraar [2018a], which in turn has applications to stability of semigroups
of operators.

Functional calculus

Several open problems related to the H ∞ -functional calculus have been stated
in Volume II. Theorem 15.3.9 is a classical result of Seeley [1971] and states
that for a sectorial operator A with bounded imaginary powers, the domain
of fractional powers D(Aθ ) coincides with the complex interpolation space
[X, D(A)]θ . A converse holds for Hilbert spaces X, where boundedness of imag-
inary powers characterises the boundedness of the H ∞ -calculus. For Banach
spaces, boundedness of imaginary powers does not imply the boundedness of
the H ∞ -calculus (see Example 10.2.32).

Problem Q.15. Let A be a sectorial operator on a Banach space X, and let

θ ∈ (0, 1). Under what conditions on X does the validity of the domain iden-
tification D(Aθ ) = [X, D(A)]θ imply that A has bounded imaginary powers?

Maximal regularity

The characterisation, in Theorem 17.3.1, of maximal Lp -regularity in terms

of R-sectoriality was stated for the class of UMD spaces. The necessity of R-
sectoriality holds without conditions on the Banach space. Therefore it seems
natural to pose the following problem:
Problem Q.16. Characterise those Banach spaces X with the property that
all R-sectorial operators A in X have maximal Lp -regularity.
Theorem 17.4.1 shows that maximal Lp -regularity of −∆ implies that
R reg
βp,X 6 2Mp,−∆ (R+ ), (Q.7)

where −∆ is considered on Lp (Rd ; X). A converse bound was obtained as

well, but since we applied the Mihlin multiplier theorem in the proof, the
bound seems far from optimal. Krylov and Priola [2017] showed that in the
reg
scalar case dimension independent bounds hold for Mp,−∆ (R+ ), and their
proof seems to extend to the vector-valued setting to improve the bound we
obtained. Still, powers of the UMD constant will be needed if one applies the
Mihlin multiplier theorem for d = 1. In case of the Poisson operator (−∆)1/2
the following two-sides bound was proved in Hytönen [2015]:
1 R reg R
max{βp,X , ~p,X } 6 Mp,(−∆) 1/2 (R+ ) 6 βp,X + ~p,X (Q.8)
2
742 Q Questions

via the results in Geiss, Montgomery-Smith, and Saksman [2010]. Our lower
estimate (Q.7) was obtained via an anisotropic extension of part of their
result. It seems an interesting problem to try to extend the techniques in the
latter paper to find a full analogue of the estimates (Q.8) for the maximal
Lp -regularity constant of the Laplace operator:
reg
Problem Q.17. Let Mp,−∆ (R+ ) be the maximal Lp -regularity constant of
−∆ on Lp (Rd ; X) on R+ , where X is a UMD space. We ask if exist universal
constants 0 < c 6 C < ∞ such that
reg
c max{βp,X , ~p,X } 6 Mp,−∆ (R+ ) 6 C(βp,X + ~p,X ).

In Theorem 17.4.8 we have seen that there are operators A such that −A
generates an analytic semigroup on Lq , but A does not have maximal Lp -
regularity on finite time intervals. This provided a negative answer to Brezis’s
question as explained in the notes of Chapter 17. The question still remains
open for differential operators.
Problem Q.18. Let O ⊆ Rd be open, and let A be a differential operator
on Lp (O) with p ∈ (1, ∞) such that −A generates an analytic semigroup on
Lp (O). Does A have maximal Lp -regularity?
Some evidence in favor of having maximal Lp -regularity can be found in
Blunck and Kunstmann [2002], Kunstmann [2008] for operators in divergence
and non-divergence form respectively.
In Theorem 17.2.15 we have seen that maximal Lp -regularity of A implies
that −A generates an analytic semigroup. Such a result seems unavailable for
time-dependent operators A without any further conditions. Because of the
time-dependence, generation of a semigroup has to be replaced by generation
of an evolution family. A family of bounded operators (S(t, s))06s6t6T on a
Banach space X is called an evolution family if
(1) S(t, t) = I for all t ∈ [0, T ];
(2) S(t, s)S(s, r) = S(t, r) for all 0 6 r 6 s 6 t 6 T ;
It is called strongly continuous if S : {(t, s) ∈ [0, T ]2 : s 6 t} → L (X) is
strongly continuous. References for the theory of evolution families include
Amann [1995], Engel and Nagel [2000], Lunardi [1995], Pazy [1983], Tanabe
[1979] and Yagi [2010].
Problem Q.19. Let X0 and X1 be Banach spaces, with X1 continuously
and densely embedded in X0 , and let p ∈ [1, ∞]. Suppose that A : [0, T ] →
L (X1 , X0 ) is strongly measurable in the uniform operator topology, and that
there exist a constant 0 < C < ∞ such that

C −1 kxkX1 6 kA(t)xkX0 + kxkX0 6 CkxkX1 , x ∈ X1 , t ∈ [0, T ].

Suppose further that A has maximal Lp -regularity in the sense that for all
f ∈ Lp (0, T ; X0 ) there exists a unique strong solution u ∈ W 1,p (0, T ; X0 ) ∩
 Q Questions 743

Lp (0, T ; X1 ) with u(0) = 0 to the problem u0 (t) + A(t)u = f (t) for t ∈ [0, T ].
Does there exists a strongly continuous evolution family (S(t, s))06s6t6T on
X0 such that for all f ∈ Lp (0, T ; X0 ) the strong solution u is equals
Z t
u(t) = S(t, s)f (s) ds, t ∈ [0, T ]?
0

The problem is also open if additionally one assumes that for each for fixed
t0 ∈ [0, T ] the operator −A(t0 ) generates an analytic semigroup on X0 . By a
result of Prüss and Schnaubelt [2001], for A ∈ C([0, T ]; L (X1 , X0 )) with the
property that −A(t0 ) generates an analytic semigroup group in X0 for each
fixed t0 ∈ [0, T ], an associated evolution family (S(t, s))06s6t6T .
The following converse to Problem Q.19 is also open.
Problem Q.20. Let X0 and X1 be UMD spaces, with X1 continuously and
densely embedded in X0 . Suppose that A : [0, T ] → L (X1 , X0 ) is strongly
measurable in the uniform operator topology, and that there exist a constant
0 < C < ∞ such that

C −1 kxkX1 6 kA(t)xkX0 + kxkX0 6 CkxkX1 , x ∈ X1 , t ∈ [0, T ].

Let (S(t, s))06s6t6T be a strongly continuous evolution family on X0 , and

suppose that for each x ∈ X1 , u = S(·, s)x is the unique strong solution to
the problem
(
u0 (t) + A(t)u(t) = 0, t ∈ [s, T ],
u(s) = x.

Suppose that for all 0 6 s < t 6 T we have S(t, s) ∈ L (X0 , X1 ) and the
families

{S(t, s) : 0 6 s 6 t 6 T }
{(t − s)A(r)S(t, s) : 0 6 s 6 t 6 T, r ∈ [0, T ]}

are R-bounded as subsets of L (X0 ). Does it follows that A has maximal

Lp -regularity?
The problem is also open if additionally one assumes that there exists a θ ∈
(0, π/2) such that for all t ∈ [0, T ], A(t) is θ-sectorial and the family

{λR(λ, A(t)) : λ ∈ {Σθ , t ∈ [0, T ]}

is R-bounded. In Gallarati et al. [2016], Gallarati and Veraar [2017a,b], a

maximal Lp -regularity result is proved under hypotheses related to those in
Problem Q.20, but with additional commutativity conditions. In applications
to concrete PDE, this means that the coefficients of elliptic operators need to
be space-independent.
744 Q Questions

Sufficient conditions for maximal L2 -regularity for elliptic second order

operators in divergence form with measurable coefficients in time and space
on X0 = H −1,2 can be obtained via the theory of Lions [1961, 1969] (also
see Auscher and Egert [2016], Dier and Zacher [2017]). Disser, ter Elst, and
Rehberg [2017] extended some of these results to maximal Lp -regularity on
X0 = H −1,q for q and p in an small interval around 2 via Sneiberg ex-
trapolation. Interesting counterexamples to maximal Lp -regularity for time-
dependent operators A appear in Pierre and Schmitt [1997].
It would be interesting to see if Problem Q.20 can be understood from the
point of singular integrals on T (1) or T (b)-theorems (see Chapter 12). For this
one could use the kernel K(t, s) = 1s<t A(0)S(t, s) ∈ L (X0 ).
K
Measurable semigroups

This appendix extends some results of Appendix G to the setting of semi-

groups that are not necessarily strongly continuous. In applications, such semi-
groups arise in several natural ways. For instance, the adjoint semigroup on
X ∗ associated with a C0 -semigroup on a Banach space X is weak∗ -continuous
but not necessarily strongly continuous, and if A is a sectorial operator of
angle less than 12 π, then −A generates a semigroup of operators which is a
C0 -semigroup if and only if A is densely defined. The framework introduced
here treats all such examples in a unified way.

K.1 Measurable semigroups

In what follows we will always assume that X is a Banach space.
Definition K.1.1 (Measurable semigroups). A family S = {S(t)}t>0
of bounded linear operators acting on X is called a semigroup if S(t)S(s) =
S(t + s) for all t, s > 0. A semigroup S is said to be:
• weakly Y -measurable, where Y is a given subspace of X ∗ , if for all x ∈ X
and x∗ ∈ Y the mapping t 7→ hS(t)x, x∗ i is measurable;
• weakly measurable, if it is weakly X ∗ -measurable;
• strongly measurable, if for all x ∈ X the mapping t 7→ S(t)x is strongly
measurable.
A semigroup S is locally bounded if the mapping t 7→ S(t) is bounded on
every bounded interval.
It has been shown in Proposition G.2.7 that every strongly measurable semi-
group is strongly continuous for t > 0, that is, for all x ∈ X the mapping
t 7→ S(t)x is continuous for t > 0. The Pettis measurability theorem (The-
orem 1.1.20) implies that if Y is a weak∗ -dense subspace of the dual of a
separable Banach space X, then every weakly Y -measurable semigroup on X
is strongly measurable, and hence strongly continuous for t > 0.
© Springer Nature Switzerland AG 2023 745
T. Hytönen et al., Analysis in Banach Spaces, Ergebnisse der Mathematik und ihrer
Grenzgebiete. 3. Folge / A Series of Modern Surveys in Mathematics 76,
https://doi.org/10.1007/978-3-031-46598-7
746 K Measurable semigroups

Every locally bounded semigroup S satisfies an exponential bound

kS(t)k 6 M eωt , t > 0, (K.1)
for suitable M > 1 and ω ∈ R; this is proved as in Proposition G.2.2.
Definition K.1.2 (Generators). Let Y be a weak∗ -dense subspace of X ∗ ,
and let S be a locally bounded weakly Y -measurable semigroup on X satisfying
(K.1). A linear operator (G, D(G)) in X is said to be the Y -generator of S if
the following conditions are satisfied:
(i) {λ ∈ C : <λ > ω} ⊆ %(G);
(ii) for all <λ > ω and t > 0 we have S(t)R(λ, G) = R(λ, G)S(t);
(iii) for all <λ > ω, x ∈ X, and x∗ ∈ Y we have
Z ∞
hR(λ, G)x, x∗ i = e−λt hS(t)x, x∗ i dt. (K.2)
0

If a Y -generator exists, it is unique. Also, by the injectivity of the Fourier

transform, two locally bounded strongly measurable semigroups agree if they
have the same Y -generator. If two locally bounded weakly measurable semi-
groups S and Se have the same Y -generator, then for all x ∈ X and x∗ ∈ Y
we have hS(t)x, x∗ i = hS(t)x,
e x∗ i for almost all t > 0. In the latter case, more
cannot be asserted in general: the identically zero semigroup and the left and
right translation semigroups on `2 (R) have generator 0.
Remark K.1.3. R ∞Suppose that the semigroup S is strongly measurable. Then
the integrals 0 e−λt S(t)x dt as in (K.2) exist in the strong sense as Bochner
integrals in X and define an X-valued holomorphic function for <λ > ω. By
analytic continuation, and using the fact that the resolvents diverge in norm
near the spectrum, it follows that s(G) 6 ω0 (S), where

s(G) = inf ω ∈ R : {<λ > ω} ⊆ %(G)

is the spectral abscissa of G, and

ω0 (S) = inf ω ∈ R : ∃M > 1 : kS(t)k 6 M eωt for all t > 0



is the exponential growth bound of S.

In what follows, whenever there is no risk of confusion, we understand the
weak∗ -dense subspace Y to be given and drop the prefix ‘Y -’.
Remark K.1.4. Assumption (ii) is implied by (i) and (iii) in each of the fol-
lowing two cases:
• S is strongly measurable: for then one has
Z ∞
R(λ, G)x = e−λs S(s)x ds
0

for every x ∈ X as a Bochner integral, and one may use that bounded
operators can be pulled through such integrals;
 K.1 Measurable semigroups 747

• S ∗ (t)Y ⊆ Y for all t > 0: for then one has, for all x∗ ∈ Y ,

hS(t)R(λ, A)x, x∗ i = hR(λ, A)x, S ∗ (t)x∗ i

Z ∞
= e−λt hS(s)x, S ∗ (t)x∗ i ds
Z0 ∞
= e−λt hS(s + t)x, x∗ i ds = hR(λ, A)S(t)x, x∗ i.
0

Proposition K.1.5. Let G generate the locally bounded weakly Y -measurable

semigroup S on X. Then:
(1) for all x ∈ D(G) we have S(t)x ∈ D(G) and GS(t)x = S(t)Gx;
(2) for all x ∈ D(G) and x∗ ∈ Y the function hS(·)x, x∗ i is differentiable on
(0, ∞) and dtd
hS(t)x, x∗ i = hS(t)Gx, x∗ i;
(3) if Y is norming, then

D(G) = x ∈ X : lim kS(t)x − xk = 0
t↓0

= x ∈ X : lim kλR(λ, G)x − xk = 0
λ→∞

and S restricts to a C0 -semigroup on D(G) whose generator equals the

part of G in D(G).
Proof. (1): Applying the resolvent commutation assumption to y = (λ − G)x
with x ∈ D(G) gives S(t)x = R(λ, G)S(t)(λ − G)x. Since the right-hand
side is in D(G), so is the left-hand side. Applying λ − G on both sides gives
(λ − G)S(t)x = S(t)(λ − G)x and the result follows.
(2): If x ∈ D(G) and λ > ω, we may write x = R(λ, G)y for some y ∈ X.
Then, for all s, t > 0 and x∗ ∈ Y ,

hS(t)x − S(s)x, x∗ i
= hR(λ, G)S(t)y − R(λ, G)S(s)y, x∗ i
Z ∞ Z ∞
−λr ∗
= e hS(t + r)y, x i dr − e−λr hS(s + r)y, x∗ i dr
0 0
Z ∞ Z ∞
= eλt e−λr hS(r)y, x∗ i dr − eλs e−λr hS(r)y, x∗ i dr
t s
Z ∞ Z t
−λr ∗
λt
= (e − e ) λs
e hS(r)y, x i dr − eλs e−λr hS(r)y, x∗ i dr.
t s

Fox fixed t > 0, the right-hand side is a continuous function of s > 0. Hence
this is also true for the left-hand side, from which we infer that s 7→ hS(s)x, x∗ i
is continuous. Dividing by t − s and letting s → t, by the continuity just
observed we obtain
D S(t)x − S(s)x E Z ∞
∗
lim , x = λe λt
e−λr hS(r)y, x∗ i dr − hS(t)y, x∗ i
s→t t−s t
748 K Measurable semigroups

= λhS(t)R(λ, G)y, x∗ i − hS(t)y, x∗ i

= hS(t)GR(λ, G)y, x∗ i
= hS(t)Gx, x∗ i.

(3): Repeating the argument of (2) with S(s)x replaced by x, and using
that Y is norming, the estimate (K.1) implies, for λ > ω and t > 0,

kS(t)x − xk = sup |hS(t)x − x, x∗ i|

x∗ ∈Y
kx∗ k61
 Z ∞ Z t 
6 M kyk |eλt − 1| e−(λ−ω)r dr + e−(λ−ω)r dr ,
t 0

with x = R(λ, G)y as before. As t ↓ 0, the right-hand side tends to 0. This

proves strong continuity for x ∈ D(G). Strong continuity for x ∈ D(G) follows
from this by local boundedness. It is clear that the restriction of S to D(G)
is a C0 -semigroup on this subspace.
In the converse direction, if x ∈ X is such that limt↓0 S(t)x = x strongly
in X, the semigroup property implies that t 7→ S(t)x is continuous in [0, ∞)
(with the convention S(0)x = x), and for <λ > ω we obtain
Z ∞
λR(λ, G)x = λe−λt S(t)x dt,
0

where the integral converges as a Bochner integral in X. By strong continuity,

passing to the limit λ → ∞ gives limλ→∞ λR(λ, G)x = x. This shows that
x ∈ D(G). This proves the first identity as well as the inclusion ‘⊆’ for the
second identity. Conversely, if limλ→∞ λR(λ, G)x = x, then x ∈ D(G). This
gives the inclusion ‘⊇’ for the second identity.
Let us denote the generator of S|D(G) by G0 and the part of G in D(G) by
G1 . If x ∈ D(G0 ), then
Z ∞
R(λ, G)(λ − G0 )x = e−λt S(t)(λ − G0 )x dt = R(λ, G0 )(λ − G0 )x = x,
0

where the first identity follows from the fact that G generates S and the second
from the fact that G0 generates S0 := S|D(G) , both in the sense of Definition
K.1.2; by strong continuity, there is no need to evaluate against functionals
in Y . Since the left-hand side belongs to D(G), so does the right-hand side x.
Applying λ − G to both sides, we obtain the identity (λ − G0 )x = (λ − G)x.
Since the former belongs to D(G), so does the latter. This proves that x ∈
D(G1 ) and G0 x = G1 x.
In the converse direction, if x ∈ D(G1 ), then writing x = R(λ, G)y with
y ∈ X gives, for all x∗ ∈ Y ,
Z ∞
∗
hx, x i = e−λt hS(t)y, x∗ i dt
0
 K.1 Measurable semigroups 749
Z ∞
= e−λt hS(t) (λ − G)x, x∗ i dt
0 | {z }
∈D(G)
Z ∞
= e−λt hS0 (t)(λ − G)x, x∗ i dt = hR(λ, G0 )(λ − G)x, x∗ i,
0

where the last step follows from the fact that G0 generates S0 = S|D(G) in the
sense of Definition K.1.2. It follows that x = R(λ, G0 )(λ − G)x ∈ D(G0 ). 

Proposition K.1.6 (Rescaling). Let G be generate the locally bounded

weakly Y -measurable semigroup S = (S(t))t>0 on X. Then for any µ ∈ C
the operator Gµ := G − µ generates the locally bounded scalarly measurable
semigroup Sµ = (e−µt S(t))t>0 .

Proof. We have kSµ (t)k 6 M e(ω−<µ)t , {λ ∈ C : <λ > ω − <µ} ⊆ %(Gµ ), and
if <λ > ω − <µ, then <(λ + µ) > ω, and therefore for all x ∈ X and x∗ ∈ Y
we have
Z ∞ Z ∞
−λt ∗
e hSµ (t)x, x i dt = e−(λ+µ)t hS(t)x, x∗ i dt
0 0
= hR(λ + µ, G)x, x∗ i = hR(λ, Gµ )x, x∗ i.


The following proposition provides an analogue of Proposition G.2.3(3), which
states that if S is aR C0 -semigroup on X with generator A, then for all x ∈ X
t
and t > 0 one has 0 S(s)x ds ∈ D(A) and
Z t
A S(s)x ds = S(s)x − x,
0

and that if x ∈ D(A), then also

Z t Z t
A S(s)x ds = S(s)Ax ds.
0 0

The difficulty in the present set-up is that the integrals of the semigroup orbits
make no a priori sense. Establishing that the integrals do indeed exist in X
as “weak Y -integrals” is part of our task in proving the proposition. In the
strongly measurable case, all this poses no problems and in the proposition
below one can simply take
Z t
0
xt = S(s)x ds
0

as a Bochner integral in X and redo the proof of Proposition G.2.3.

750 K Measurable semigroups

Proposition K.1.7 (Main theorem of calculus). Let G be generate the

locally bounded weakly Y -measurable semigroup S on X, and assume that Y
is norming. For all t > 0 and x ∈ X there exists a unique xt0 ∈ D(G) such
that for all x∗ ∈ Y we have
Z t
0 ∗
hxt , x i = hS(s)x, x∗ i ds, x∗ ∈ Y,
0

and
Gxt0 = S(t)x − x.
If in addition x ∈ D(G), we furthermore have
Z t
hGx0t , x∗ i = hS(s)Gx, x∗ i ds, x∗ ∈ Y.
0

Proof. Let x ∈ X and t > 0 be fixed.

Step 1 – We begin with uniqueness. If
Z t
S(t)x − x = 0 and hS(s)x, x∗ i ds = 0, x∗ ∈ Y,
0

then integrating the first identity over [0, t] with t > 0, applying x∗ , and
subtracting the second identity, we are left with the identity thx, x∗ i = 0,
valid for all x∗ ∈ Y , and therefore x = 0. But for x = 0, uniqueness is clear.
Step 2 – For the proof of existence, fix an arbitrary λ > ω and consider
the rescaled semigroup Sλ generated by Gλ as in Proposition K.1.6. This
semigroup is uniformly exponentially stable, by which we mean that it satisfies
(K.1) with a negative exponent. In particular, Gλ is boundedly invertible. The
0
element xt,λ := Gλ−1 (Sλ (t)x − x) belongs to D(Gλ ) = D(G) and satisfies
0
Gλ xt,λ = Sλ (t)x − x.

Applying the definition of a generator to Gλ and performing a change of

variables, for all x∗ ∈ Y we obtain

hx0t,λ , x∗ i = hGλ−1 (Sλ (t)x − x), x∗ i

Z ∞ Z t
∗
=− hSλ (s)(Sλ (t)x − x), x i ds = hSλ (s)x, x∗ i ds.
0 0

This proves the proposition for the semigroup Sλ .

Step 3 – We wish to use the results of Steps 1 and 2 to derive the propo-
sition for the original semigroup S. To guess the formula for xt0 , we formally
Rt
write xt0 = 0 S(s)x ds (not worrying about measurability and integrability
issues) and integrate by parts to arrive at
 K.1 Measurable semigroups 751
Z t Z t
x0t = S(s)x ds = eλs Sλ (s)x ds
0 0
Z t Z t Z s
= eλt Sλ (r)x dr − λ eλs Sλ (r)x dr ds (K.3)
0 0 0
Z t
0 0
= eλt xt,λ −λ eλs xs,λ ds,
0
Rs
with x0s,λ = Gλ−1 [Sλ (s)x
− x] = 0 Sλ (r)x dr as in Step 2, the second identity
being again formal. The last integral appearing in (K.3) is well defined as a
0
Bochner integral in X because s 7→ xs,λ = S(s)Gλ−1 x − Gλ−1 x is continuous as
an X-valued function, S being strongly continuous on D(G) = D(Gλ ).
We now take the right-hand side of (K.3) as the definition of x0t . The strong
continuity of S on D(G) = D(Gλ ) implies that
Z t Z t
eλs x0s,λ ds = eλs (S(s)G−1 −1
λ x − Gλ x) ds
0 0

belongs to D(G). Indeed, using the notation used in the proof of Proposition
K.1.5 it belongs to D(G0 ), and we have D(G0 ) ⊆ D(G). Since also x0t,λ ∈ D(G)
(by Step 2), we conclude from the representation (K.3) that x0t belongs to
D(G), and we have
Z t
Gx0t = eλt Gx0t,λ − λG eλs x0s,λ ds
0
Z t
−1 λt
= GGλ (S(t)x − e x) − λG eλs (Sλ (s)G−1 −1
λ x − Gλ x) ds
0
Z t
−1 λt
= (I + λGλ )(S(t)x − e x) − λG S(s)G−1
λ x ds
0
+ (eλt − 1)GG−1
λ x
(∗)
= (I + λG−1 λt −1 −1
λ )(S(t)x − e x) − λ(S(s)Gλ x − Gλ x)
+ (eλt − 1)(I + λG−1
λ )x
= S(t)x − x,

where (∗) uses the reasoning involving strong continuity of S on D(G) =

D(Gλ ) once more. This proves that x0t has the required properties.
If x ∈ D(G), then by Proposition K.1.5(2), for all x∗ ∈ Y we have
Z t Z t
d
hGx0t , x∗ i = hS(t)x − x, x∗ i ds = hS(s)x, x∗ i ds = hS(r)Gx, x∗ i dr;
0 ds 0

to justify the second identity we note that s 7→ hS(s)x, x∗ i is differentiable on

(0, t], with bounded derivative.
752 K Measurable semigroups

Finally, applying this with G−1 x in place of x and noting that (G−1 x)0t =
G−1 x0t , we obtain
Z t
hx0t , x∗ i = hS(s)x, x∗ i ds.
0

This completes the proof. 

To finish the abstract treatment, let us write
Z t
x0 =: S(s)x ds
0

for the unique element x0 ∈ D(G) satisfying the conclusions of the proposition.
With this notation, we have the following result.
Proposition K.1.8. Under the assumptions of Proposition K.1.7, for all
s, t > 0 and x ∈ X we have
Z t Z t
S(s) S(r)x dr = S(r + s)x dr,
0 0
R t+s Rs
where the right-hand side is shorthand for 0
S(r)x dr − 0
S(r)x dr.
Proof. As in the proof of Proposition K.1.7, this is easy for the rescaled semi-
group Sλ . Indeed, for all ∈ X and x∗ ∈ Y we have
0 0
Gλ Sλ (s)xt,λ = Sλ (s)Gλ xt,λ
0 0
= Sλ (t)[Sλ (s)x] − [Sλ (s)x] = Gλ xt+s,λ − Gλ xs,λ .

Applying Gλ−1 we obtain the identity

0 0
Sλ (s)xt,λ = xt+s,λ − x0s,λ .

Applying x∗ ∈ Y to both sides, this can be rewritten as

Z t+s Z s
0
hSλ (s)xt,λ , x∗ i = hSλ (r)x, x∗ i dr − hSλ (r)x, x∗ i dr
0 0
Z t
= hSλ (r)[Sλ (s)x], x∗ i dr.
0

These identities imply that S(s)x0t,λ satisfies the two properties of Proposition
K.1.7 with x replaced by S(s)x. By uniqueness, in the notation introduced
above this gives the desired identity
Z t Z t
Sλ (s) Sλ (r)x dr = Sλ (r + s)x dr.
0 0

The general case can again be deduced from the rescaled case, by similar
arguments as before. We leave the details to the reader. 
 K.1 Measurable semigroups 753

We proceed with some important examples. The first demonstrates the con-
sistency of Definition K.1.2 with the corresponding Definition G.2.1 for C0 -
semigroups.

Proposition K.1.9 (C0 -semigroups). If G generates a C0 -semigroup S in

the sense of Definition K.1.2, then G = G, e is the generator of S in
e where G
the sense of Definition G.2.1.
e satisfy (K.2), which
Proof. This is immediate from the fact that both G and G
implies that their resolvents coincide. 

Proposition K.1.10 (Adjoints of C0 -semigroups). If G generates the ad-

joint S ∗ of a C0 -semigroup S on X in the sense of Definition K.1.2, then
G = A∗ , where A is the generator of S in the sense of Definition G.2.1.

This is proved in the same way (take Y = X this time).

Proposition K.1.11 (Sectorial operators of angle less than 21 π). Let

A be a sectorial operator in X of angle ω(A) < 21 π, and for t > 0 define the
bounded operators S(t) on X by
Z
1
S(t)x := e−zt R(z, A)x dz, x ∈ X,
2πi Γ

where Γ is the downwards oriented boundary of Σσ ∪ B, for any ω(A) <

σ < 21 π and any closed ball B centred at the origin. Then S = (S(t))t>0
is a bounded and strongly measurable semigroup on X, and the operator −A
generates S in the sense of Definition K.1.2.

Proof. Boundedness of S is proved by repeating the argument of Theorem

G.5.2, and strong measurability is evident. The semigroup property is proved
in the same way as the multiplicativity of the Dunford calculus (Theorem
10.2.2). For the reader’s convenience we present the details. Fix ζ, ζ 0 ∈ Ση
and choose contours Γ and Γ 0 as above, with Γ to the right of Γ 0. Then, by
the resolvent identity, Cauchy’s theorem, Fubini’s theorem, and the Cauchy
integral formula,
Z Z
1 0
S(z 0 )S(z)x = eλz+µz R(λ, A)R(µ, A)x dλ dµ
(2πi)2 Γ Γ 0
0 R(λ, A)x − R(µ, A)x
Z Z
1
= eλz+µz dλ dµ
(2πi)2 Γ Γ 0 µ−λ
Z Z
1 0 R(λ, A)x
= 2
eλz+µz dµ dλ
(2πi) Γ Γ 0 µ−λ
Z
1 0
= eλz+λz R(λ, A)x dλ = S(z + z 0 )x.
2πi Γ
754 K Measurable semigroups

Next we check that −A is the generator of S in the sense of Definition K.1.2.

Resolvent commutations is clear. For all x∗ ∈ X ∗ , by Fubini’s theorem and
Cauchy’s theorem we have
Z ∞ Z Z ∞
1 
e−λr hS(r)x, x∗ i dr = e−(λ+z)r dr hR(z, A)x, x∗ i dz
0 2πi Γ 0
Z
1 1
= hR(z, A)x, x∗ i dz
2πi Γ λ + z
= hR(−λ, A)x, x∗ i = hR(λ, −A)x, x∗ i.


Remark K.1.12. Suppose that A is a densely defined sectorial operator of an-
gle ω(A) < 21 π. By Theorem G.5.2, −A generates a bounded analytic C0 -
semigroup S in the sense of Definition G.2.1. Since the adjoint operator A∗ is
sectorial and has the same angle, combination of the above two results shows
that −A∗ generates the bounded analytic (but not necessarily C0 ) semigroup
S ∗ in the sense of Definition K.1.2.

K.2 Uniform exponential stability

In this section we state two propositions providing elementary characterisa-
tions of uniform exponential stability. A semigroup S on a Banach space X
is said to be uniformly exponentially stable if there exist M > 1 and ω > 0
such that kS(t)k 6 M e−ωt for all t > 0. A locally bounded semigroup S is
uniformly exponentially stable if and only if

kS(t0 )k < 1 for some t0 > 0.

The ‘only if’ part is obvious. To prove the ‘if’ part, we show that kS(t)k 6
M e−ωt with

M = eωt0 sup kS(s)k, ω = log(kS(t0 )k−1/t0 ).

s∈(0,t0 ]

Indeed, let t > 0 be fixed and write t = (n + θ)t0 with θ ∈ [0, 1) and n ∈ N.
Then, with the above choices of ω and M ,

kS(t)k 6 kS(θt0 )k kS(t0 )kn 6 M e−ωt0 e−ωt0 n = M e−ωt0 (n+1) 6 M e−ωt .

Proposition K.2.1 (Datko–Pazy). Let S be a strongly measurable semi-

group on a Banach space X and let 1 6 p < ∞. The following assertions are
equivalent:
(1) the semigroup S is uniformly exponentially stable;
(2) the semigroup S is locally bounded and for all x ∈ X the orbit t 7→ S(t)x
belongs to Lp (R+ ; X).
 K.2 Uniform exponential stability 755

Proof. The implication (1)⇒(2) part is clear. To prove the converse implica-
tion (2)⇒(1), by the closed graph theorem there exists a constant C > 0 such
that kS(·)xkLp (R+ ;X) 6 Ckxk for all x ∈ X. Let M > 1 and ω > 0 be such
that kS(t)k 6 M eωt for all t > 0. Then
t
1 − e−ωpt
Z
kS(t)xkp = e−wp(t−s) kS(t − s)S(s)xkp ds 6 M p C p kxkp .
ωp 0

It follows that supt>0 kS(t)k =: K < ∞ and, for t > 0,

Z t
tkS(t)xkp = kS(t − s)S(s)xkp ds 6 K p C p kxkp .
0

Therefore kS(t)k 6 CKt−1/p and the result follows from the preliminary
observation. 

Proposition K.2.2. Let S be a strongly measurable semigroup and let 1 6

p 6 ∞. The following assertions are equivalent:
(1) the semigroup S is uniformly exponentially stable;
(2) the semigroup S is locally bounded and for all f ∈ Lp (R+ ; X) we have
S ∗ f ∈ Lp (R+ ; X).

Below we will see that S ∗ f can often be interpreted as the so-called mild
solution to an abstract Cauchy problem.
Proof. (1)⇒(2): If kS(t)k 6 M e−ωt for all t > 0, with ω > 0, then kS∗f (t)k 6
φ ∗ f (t) for all t > 0, where φ(s) = M e−ωs 1R+ (s). Taking Lp -norms, Young’s
inequality gives S ∗ f ∈ Lp (R+ ; X) and

kS ∗ f kLp (R+ ;X) 6 kφkL1 (R+ ) kf kLp (R+ ;X) 6 M ω −1 kf kLp (R+ ;X) .

(2)⇒(1): By the closed graph theorem there exists a constant C > 0 such
that
kS ∗ f kLp (R+ ;X) 6 Ckf kLp (R+ ;X) , f ∈ Lp (R+ ; X).
First consider p ∈ [1, ∞). Let M > 1 and µ > 0 be such that kS(t)k 6 M eµt
for all t > 0. Fix ε > 0 and x ∈ X, and set f (t) := e−(µ+ε)t S(t)x for t > 0.
Then
M
kf kLp (R+ ;X) 6 kxk
(εp)1/p
and
t
1 − e−(µ+ε)t
Z
S ∗ f (t) = S(t − s)f (s) ds = S(t)x.
0 µ+ε
Fixing any τ > 0 we obtain

kt 7→ S(t)xkLp (R+ ;X) 6 kt 7→ S(t)xkLp (0,τ ;X) + kt 7→ S(t)xkLp (τ,∞;X)

756 K Measurable semigroups
µ+ε
6 M eµτ τ 1/p kxk + kS ∗ f kLp (τ,∞;X)
1 − e−(µ+ε)τ
µ+ε
6 M eµτ τ 1/p kxk + Ckf kLp (R+ ;X)
1 − e−(µ+ε)τ
µ+ε CM
6 M eµτ τ 1/p kxk + −(µ+ε)τ
kxk,
1−e (εp)1/p

using that kf (t)k 6 M e−εt kxk in the last inequality. Proposition K.2.1 now
gives the required uniformly exponential stability.
For p = ∞ the above argument can be repeated to give the bound

sup kS(t)xk 6 kt 7→ S(t)xkL∞ (0,τ ;X) + kt 7→ S(t)xkL∞ (τ,∞;X)

t>0
µ+ε
6 M eµτ kxk + kS ∗ f kL∞ (R+ ;X)
1 − e−(µ+ε)τ
µ+ε
6 M eµτ kxk + Ckf kL∞ (R+ ;X)
1 − e−(µ+ε)τ
µ+ε
6 M eµτ kxk + CM kxk =: Kkxk.
1 − e−(µ+ε)τ
This shows that for every x ∈ X the function t 7→ fx (t) := S(t)x belongs to
L∞ (R+ ; X). Since S ∗ fx (t) = tS(t)x for all t > 0, we find that

kS(t)xk = t−1 kS ∗ fx (t)k 6 t−1 Ckfx kL∞ (R+ ;X) 6 t−1 Kkxk.

We now apply the preliminary observation preceding Proposition K.2.1. 

For analytic semigroups we have the following simple necessary and sufficient
condition for uniform exponential stability:
Proposition K.2.3. Let A be a sectorial operator of angle ω(A) < 21 π. The
analytic semigroup generated by −A is uniformly exponentially stable if and
only if 0 ∈ %(A).
Proof. If the semigroup (S(t))t>0 generated by −A is uniformly exponentially
stable, the observation of Remark K.1.3 implies that %(−A) contains the right-
half plane <λ > −δ} for some δ > 0. In particular, 0 ∈ %(A).
In the converse direction, if 0 ∈ %(A), then A−δ is sectorial with ω(A−δ) <
1
2 π for any δ > 0 so small that the disc {|z| < 2δ} is contained in %(A).
But that means that the semigroup (eδt S(t))t>0 is bounded, so (S(t))t>0 is
uniformly exponentially stable. 

K.3 Analytic semigroups

Let A be a sectorial operator of angle ω(A) < π/2. As we have seen in Proposi-
tion K.1.11, −A generates the bounded strongly measurable semigroup given
by the formula
 K.3 Analytic semigroups 757
Z
1
S(t)x = e−zt R(z, A)x dz, t > 0, x ∈ X,
2πi Γ

where Γ is as described in the proposition. By the same formula, this semi-

group extends to complex times z ∈ Σ 21 π−ω(A) . For any ω(A) < σ < 21 π, the
resulting family S = S(z))z∈Σ 1 π−σ is bounded and holomorphic.
2
As in the case of analytic C0 -semigroups, one proves that if A be sectorial
of angle < 21 π and S is the analytic semigroup generated by −A, then for all
x ∈ X we have S(t)x ∈ D(A) and

MA := sup ktAS(t)k < ∞. (K.4)

t>0

As a consequence, for any integer n > 1, S(t) maps X into D(An ) and

sup ktn An S(t)k 6 sup ktAS(t/n)kn 6 (nMA )n .

t>0 t>0

The following result shows that the quantity (K.4) cannot be arbitrarily
small unless the operator A is bounded. This fact plays a role in a construction
of a counterexample in Chapter 17 (see Theorem 17.4.4).

Proposition K.3.1. Let A be sectorial of angle < 12 π and let S be the locally
bounded strongly measurable semigroup generated by −A. If
1
lim sup tkAS(t)k < ,
t↓0 e

then A is bounded.

Proof. For all t > 0 and x ∈ X we have −AS(t)x = S 0 (t)x. To see this, we first
use Hille’s theorem to move A into the integral, then we write A = (A − z) + z
and use that Z
1
e−zt x dz = 0
2πi Γ
by Cauchy’s theorem, and note that
Z
1
ze−zt R(z, A)x dz = S 0 (t)x
2πi Γ
by differentiation under the integral sign. As a consequence, for all fixed t > 0
the assumption of the lemma implies
t 0 t
1
lim sup S < .
n→∞ n n e

By induction, S (n) (t) = (−A)n S(t) = (−AS( nt ))n x = (S 0 ( nt ))n . The in-
n
equality nn /n! 6 en implies that lim supt→∞ nn! k(S 0 ( nt ))n k < 1, and therefore
there exists a δ > 0 such that for every t > 0 the series
758 K Measurable semigroups
∞ ∞
X 1 X (z − t)n nn  t 0 t
n
S(z)
e := (z − t)n S (n) (t) = S
n=0
n! n=0
tn n! n n

converges absolutely and uniformly on the set Bt := {z ∈ C : |z −t| < (1+δ)t}

and defines an analytic function, and we have Se = S on Bt ∩ (0, ∞). Since
every Bt contains the origin and S(0)
e = I, it follows that

lim kS(s) − Ik = lim kS(s)

e − Ik = 0. (K.5)
s↓0 s↓0

The remainder of the proof is devoted to showing that if G generates an

locally bounded strongly measurable semigroup S such that (K.5) holds, then
G is bounded. We follow the standard proof of the corresponding result for
C0 -semigroups, while keeping an eye on the fact that in the present situation
we use a different generator concept.
First of all, we note that (K.5) implies that S extends to a continuous
mapping from [0, ∞) to L (X). By Proposition K.1.9, the generator −A
coincides with the generator of S as a C0 -semigroup in the sense of Def-
inition G.2.1.R t In particular, A is densely defined. ChooseR tt > 0 so small
that kI − 1t 0 S(s) dsk < 1, using the shorthand notation 0 S(s) ds for the
Rt Rt
bounded operator x 7→ 0 S(s)x ds. Then the operator 1t 0 S(s) ds is invert-
ible and therefore the identity
Z t
1  t+h
Z Z h
1 
(S(h) − I) S(s) ds = S(s) ds − S(s) ds
h 0 h t 0

implies that
Z t+h Z h  1 Z t −1
1 1
(S(h) − I) = S(s) ds − S(s) ds S(s) ds .
h h t 0 t 0

We may now let h ↓ 0 in operator norm and conclude that

1 Z t −1
−A = (S(t) − I) S(s) ds .
t 0
This shows that on its dense domain, A coincides with a bounded operator.
Since A is also closed, this forces D(A) = X and consequently A is bounded.


K.4 An interpolation result

For the proof of Theorem 15.3.23 we need a characterisation of the real in-
terpolation space (X, D(A)) 21 ,2 for sectorial operators A, which will be given
presently. A further characterisation, valid in the case where −A is the gen-
erator of an analytic semigroup, is given in the next appendix (see Theorem
L.2.4).
 K.4 An interpolation result 759

Proposition K.4.1. Let θ ∈ (0, 1) and 1 6 p 6 ∞.

(1) If A is a sectorial operator in X, then
n o
(X, D(A))θ,p = x ∈ X : λ 7→ λθ A(λ + A)−1 x ∈ Lp (R+ , dλ
λ ; X)

with equivalence of norms

kxk(X,D(A))θ,p h kxk + λ 7→ λθ A(λ + A)−1 x dλ .

Lp (R+ , λ ;X)

If 0 ∈ %(A), we also have equivalence of homogeneous norms

kxk(X,D(A))θ,p h λ 7→ λθ A(λ + A)−1 x dλ .

Lp (R+ , λ ;X)

(2) If −A generates a C0 -semigroup (S(t))t>0 on X, then

n o
(X, D(A))θ,p = x ∈ X : t 7→ t−θ (S(t)x − x) ∈ Lp (R+ , dt
t ; X)

with equivalence of norms

kxk(X,D(A))θ,p h kxk + t 7→ t−θ (S(t)x − x) dt .

Lp (R+ , t ;X)

If the semigroup is uniformly exponentially stable, we also have equivalence

of homogeneous norms

kxk(X,D(A))θ,p h t 7→ t−θ (S(t)x − x) dt .

Lp (R+ , t ;X)

(1)
Proof. Denote the two sets on the right-hand side in (1) and (2) by Xθ,p
(2)
and Xθ,p , respectively. We will write K(t, x) := K(t, x; X, D(A)) for the K-
functional of the real interpolation method (see Appendix C).
For (1) we will prove continuous inclusions
(1)
(X, D(A))θ,p ⊆ Xθ,p ⊆ (X, D(A))θ,p ,

and for (2) it will then suffice to prove continuous inclusions

(2) (1)
(X, D(A))θ,p ⊆ Xθ,p ⊆ Xθ,p .

The norm equivalences are proved along the way.

(1)
(1): We start with the inclusion (X, D(A))θ,p ⊆ Xθ,p . Let x ∈ (X, D(A))θ,p
be given. If x = x0 + x1 with x0 ∈ X and x1 ∈ D(A), then for λ > 0 we have,
with M := supµ>0 kµ(µ + A)−1 k,

λθ kA(λ + A)−1 xk 6 λθ kA(λ + A)−1 x0 k + λθ kA(λ + A)−1 x1 k

6 (M + 1)λθ kx0 k + M λθ−1 kAx1 k
760 K Measurable semigroups

6 (M + 1)λθ (kx0 k + λ−1 kAx1 k).

It follows that

λθ kA(λ + A)−1 xk 6 (M + 1)λθ K(λ−1 , x).

By a change of variables the right-hand side is seen to belong to Lp (R+ , dt

t ),
with norm equal to (M + 1)kxk(X,D(A))θ,p . It follows that λθ A(λ + A)−1 x
(1)
belongs to Lp (R+ , dt
t ; X), hence x ∈ Xθ,p , and

λθ A(λ + A)−1 x Lp (R+ , dt

6 (M + 1)kxk(X,D(A))θ,p .
t ;X)

(1)
We next prove the inclusion Xθ,p ⊆ (X, D(A))θ,p . Suppose x ∈ X is such
that λ 7→ λθ A(λ + A)−1 x belongs to Lp (R+ , dt
t ; X). Then, using the decom-
position x = A(λ + A)−1 x + λ(λ + A)−1 x ∈ X + D(A), we obtain

λθ K(λ−1 , x) 6 λθ (kA(λ + A)−1 xk + λ−1 kλ(λ + A)−1 xkD(A) )

= λθ (2kA(λ + A)−1 xk + k(λ + A)−1 xk).

Since kλ(λ + A)−1 k 6 M for λ > 0, the right-hand is in Lp ((1, ∞), dλ

λ ), with
norm at most
 1 1/p
2kxk∗θ,p + M kxk if 1 6 p < ∞;
(1 − θ)p
2kxk∗θ,p + M kxk if p = ∞,

where kxk∗θ,p := kλ 7→ λθ (λ + A)−1 xkLp (R+ , dt ;X) . It follows that t 7→

t
t−θ K(t, x) belongs to Lp ((1, ∞), dt
t ). Also, since D(A) ⊆ X, we have K(t, x) 6
kxk and therefore t 7→ t−θ K(t, x) is in Lp ((0, 1), dt
t ). Together, these facts im-
−θ p dt
ply that t 7→ t K(t, x) is in L (R+ , t ) and therefore x ∈ (X, D(A))θ,p . Our
estimates moreover show that

kxk(X,D(A))θ,p = t 7→ t−θ K(t, x) Lp (R+ , dt

.θ,p kxk∗θ,p + kxk.
t )

Assume next that 0 ∈ %(A). Since limλ↓0 (λ + A)−1 = A−1 in L (X) we

may choose r = rA > 0 so small that k(λ + A)−1 − A−1 k 6 12 kA−1 k for all
0 < λ < r. It follows that for all x ∈ X we have k(λ + A)−1 xk > 12 kA−1 xk,
and therefore, for all x ∈ D(A), kA(λ + A)−1 xk > 12 kxk. It follows that for all
x ∈ D(A),

kλθ A(λ + A)−1 xkLp (R+ , dt ;X) > kλ 7→ λθ A(λ + A)−1 xkLp ((0,r), dt ;X)
t t

1
> kxkkλ 7→ λθ kLp ((0,r), dt ) =: Cθ,p,A kxk.
2 t
 K.4 An interpolation result 761

This estimate extends to general x ∈ (X, D(A))θ,p by applying it to µR(µ, A)x

and letting µ → ∞ on both sides. This gives the equivalence with the homo-
geneous norm.
(2)
(2): We begin with the inclusion (X, D(A))θ,p ⊆ Xθ,p . Fix an arbitrary
element x ∈ (X, D(A))θ,p . If x = x0 + x1 with x0 ∈ X and x1 ∈ D(A), then
for all t > 0 we have

t−θ kS(t)x − xk 6 t−θ (kS(t)x0 − x0 k + kS(t)x1 − x1 k)

6 t−θ ((M + 1)kx0 k + tM kAx1 k) 6 (M + 1)t−θ K(t, x),

where M := supt>0 kS(t)k; to obtain the estimate for x1 we used that

Rt
S(t)x1 − x1 = 0 S(s)Ax1 ds. It follows that t 7→ t−θ (S(t)x − x) belongs
to Lp (R+ , dt
t ; X) and

t 7→ t−θ kS(t)x − xk 6 (M + 1)kxk(X,D(A))θ,p .

Lp (R+ , dt
t )

(2) (2)
To prove the inclusion Xθ,p ⊆ Xθ,p , suppose that x ∈ X is such that
t 7→ t−θ (S(t)x − x) ∈ Lp (R+ , dtt ; X). Using the identity A(λ + A)
−1
=
−1
λ(λ + A) x − x and the Laplace transform representation of the resolvent
(Proposition G.4.1), for λ > 0 we have
Z ∞
−1
A(λ + A) = λe−λt (S(t)x − x) dt.
0

First let 1 6 p < ∞. By Jensen’s inequality we estimate

Z ∞ Z ∞ p dλ
−1 p
θ
λ 7→ λ A(λ + A) x Lp (R+ , dλ ;X) = λ θp
λe−λt (S(t)x − x) dt
λ λ
Z0 ∞ Z ∞0
dλ
6 λθp λe−λt kS(t)x − xkp dt
λ
Z0 ∞Z ∞0
dλ 
= λθp+1 e−λt kS(t)x − xkp dt
0 λ
Z 0∞
= Cθ,p t−θp−1 kS(t)x − xkp dt
0
p
= Cθ,p t 7→ t−θ (S(t)x − x) Lp (R+ , dt
,
t ;X)

R∞
where Cθ,p = 0 µθp+1 e−µ dµ
µ = Γ (θp+1). This gives the desired inclusion for
R ∞ θ+1 −λt θ
1 6 p < ∞. For p = ∞ we note that for all λ > 0 we have 0 λ e t dt =
R ∞ θ −s
0
s e ds = Γ (1 + θ), and therefore

λ 7→ λθ A(λ + A)−1 x L∞ (R+ , dλ ;X)

λ
Z ∞
θ+1 −λt
= sup λ e (S(t)x − x) dt
λ>0 0
762 K Measurable semigroups
Z ∞
6 sup λθ+1 e−λt tθ dt · sup t−θ kS(t)x − xk
λ>0 0 t>0
−θ
6 Cθ t 7→ t (S(t)x − x) L∞ (R+ , dt
,
t ;X)

with Cθ = Γ (1 + θ).
Assume finally that kS(t)k 6 M e−ωt for all t > 0, with M > 1 and
ω > 0. Choose R = RM,ω > 0 so large that kS(t)k 6 21 for t > R. Then
kS(t)x − xk > 21 kxk for t > R. If 1 6 p < ∞, it follows that
Z ∞
−θ p dt
t 7→ t (S(t)x − x) p dt > t−θp (kxk/2)p =: CR,θ,p kxkp .
L (R+ , t ;X) R t

This gives the equivalence with the homogeneous norm for 1 6 p < ∞. For
p = ∞ we simply note that

t 7→ t−θ (S(t)x − x) dt > t 7→ t−θ (S(t)x − x) dt

L∞ (R+ , t ;X) L∞ ((R,∞), t ;X)
1 −θ
> R kxk.
2


K.5 Notes

Definition K.1.2 is in the spirit of the corresponding definition in the theory of

integrated semigroups; see, e.g., Arendt [1987], Kellerman and Hieber [1989].
The proof of Proposition K.2.1 is taken from Engel and Nagel [2000, Theo-
rem V.1.8]; a systematic treatment is given in Van Neerven [1996]. Analytic
semigroups with non-densely defined generators have been studied in detail
in Lunardi [1995]. The proofs of Propositions K.3.1 and K.4.1 are taken from
Pazy [1983] and Lunardi [2009], respectively.
L
The trace method for real interpolation

This appendix is a continuation of Appendix C in Volume I, where the real

and complex methods were introduced. After recalling some basic definitions
(Section L.1), we treat several additional topics including the trace method
(Section L.2), reiteration theorems (Section L.3), and Sobolev embedding the-
orems and mixed derivatives (Section L.4).

L.1 Preliminaries

We begin by recalling some definitions from Appendix C.

Throughout this appendix, (X0 , X1 ) is an interpolation couple, that is, an
ordered pair of Banach spaces, both of which are continuously embedded in
a Hausdorff topological vector space X . For t > 0 and x ∈ X0 + X1 := {x ∈
X : x = x0 + x1 with x0 ∈ X0 and x1 ∈ X1 }, the K-functional is defined by

K(t, x; X0 , X1 ) := inf{kx0 kX0 + tkx1 kX1 : x0 ∈ X0 , x1 ∈ X1 , x = x0 + x1 }.

The function t 7→ K(t, x; X0 , X1 ) is non-decreasing and continuous, and the

function x 7→ K(t, x; X0 , X1 ) is sub-additive. Moreover,

tK(t−1 , x; X0 , X1 ) = K(t, x; X1 , X0 ).

From now on, when no confusion is likely to occur we abbreviate

K(t, x) = K(t, x; X0 , X1 ).

For 0 < θ < 1 and 1 6 p 6 ∞, the real interpolation space (X0 , X1 )θ,p is the
Banach space defined by

(X0 , X1 )θ,p := x ∈ X0 + X1 : kxkθ,p < ∞ ,

with norm

© Springer Nature Switzerland AG 2023 763

T. Hytönen et al., Analysis in Banach Spaces, Ergebnisse der Mathematik und ihrer
Grenzgebiete. 3. Folge / A Series of Modern Surveys in Mathematics 76,
https://doi.org/10.1007/978-3-031-46598-7
764 L The trace method for real interpolation
 Z ∞
dt 1/p

 [t−θ K(t, x)]p if 1 6 p < ∞;
kxkθ,p := 0 t
ess sup t−θ K(t, x)
 if p = ∞.
t>0

One has (X, X)θ,p = X and (X0 , X1 )θ,p = (X1 , X0 )1−θ,p with identical norms.
By Lemma C.3.12, X0 ∩ X1 is dense in (X0 , X1 )θ,p whenever 0 < θ < 1 and
1 6 p < ∞. In what follows we let
(X0 ,X1 )θ,∞
Xp,∞ := X0 ∩ X1 (L.1)

denote the closure of X0 ∩ X1 in (X0 , X1 )θ,∞ .

Fix an x ∈ X0 ∩ X1 and consider the mapping c 7→ cx as operators from
K into X0 and into X1 , respectively. Real interpolation of these operators
(Theorem C.3.3) with θ ∈ (0, 1) and p ∈ [1, ∞] gives the estimate

kxkθ,p 6 kxk1−θ θ
X0 kxkX1 , x ∈ X0 ∩ X1 . (L.2)

In (C.6) we have seen that for all 0 < θ < 1 and 1 6 p0 6 p1 6 ∞ we have
the continuous inclusion

(X0 , X1 )θ,p0 ,→ (X0 , X1 )θ,p1 . (L.3)

The next result shows that more can be said in the special case when X1 ⊆ X0
with continuous inclusion mapping (we write X1 ,→ X0 in this situation).

Proposition L.1.1. If we have a continuous embedding X1 ,→ X0 with norm

C, the following assertions hold:
(1) For all 0 < θ < 1 and 1 6 p 6 ∞ we have continuous embeddings

X1 ,→ (X0 , X1 )θ,p ,→ X0

with
kxkX0 6 C θ kxkθ,p 6 CkxkX1 , x ∈ X1 .
(2) For all 0 < θ0 < θ1 < 1, and all 1 6 p0 , p1 , p 6 ∞ we have a continuous
embedding

(X0 , X1 )θ1 ,p1 ,→ (X0 , X1 )θ0 ,p0

with

kxkθ0 ,p0 6 C θ1 −θ0 (θ1 − θ0 )−1 + θ0−1 kxkθ1 ,p1 ,

x ∈ (X0 , X1 )θ1 ,p1 .

Proof. (1): Let x = x0 + x1 with x0 ∈ X0 and x1 ∈ X1 . Then

min{1, t}kxkX0 6 kx0 kX0 + tkx1 kX0 6 kx0 kX0 + Ctkx1 kX1 .

Therefore, min{1, t}kxkX0 6 K(Ct, x), and we conclude that

 L.2 The trace method 765

kxkX0 6 sup t−θ K(Ct, x) = C θ kxkθ,∞ 6 C θ kxkθ,p

t>0

using the contractivity of the inclusion (L.3). Also, by (L.2) and the inequality
just proved,
1−θ 1−θ
kxkθ,p 6 kxkX 0
kxkθX1 6 C θ(1−θ) kxkθ,p kxkθX1 .

This implies the remaining estimate kxkθ,p 6 C 1−θ kxkX1 .

(2): By the continuity of the inclusion (L.3), it suffices to consider the
case p0 = 1 and p1 = ∞. Fix x ∈ X1 . Since K(t, x) 6 kxkX0 for all t > 0,
Z C Z ∞
−θ0 dt dt
kxkθ0 ,1 = t K(t, x) + t−θ0 K(t, x)
0 t C t
Z C Z ∞
dt dt
6 sup t−θ1 K(t, x) tθ1 −θ0 + kxkX0 t−θ0
t>0 0 t C t
C θ1 −θ0 C −θ0
= kxkθ1 ,∞ + kxkX0
θ1 − θ0 θ0
C θ1 −θ0 C θ1 −θ0
6 kxkθ1 ,∞ + kxkXθ1 ,∞ ,
θ1 − θ0 θ0
where in the last step we applied the estimate of (1). This gives the first
estimate. 

L.2 The trace method

In this section we introduce another interpolation method, the so-called trace
method, and prove its equivalence with the real interpolation method.

Definition L.2.1 (Trace method). For p ∈ [1, ∞] and θ ∈ (0, 1), the space

(X0 , X1 )Tr
θ,p

is defined as the set of all x ∈ X0 + X1 for which there exists a strongly

measurable function u : (0, ∞) → X0 + X1 which is weakly differentiable, and
satisfies the following three properties:
(i) t 7→ t1−θ u0 (t) belongs to Lp (R+ , dt
t ; X0 );
7 t1−θ u(t) belongs to Lp (R+ , dt
(ii) t → t ; X1 );
(iii) u(0) = x.
On (X0 , X1 )Tr
θ,p we define a norm by

kxk(X0 ,X1 )Tr

θ,p
n o (L.4)
:= max kt 7→ t1−θ u0 (t)kLp (R+ , dt ;X0 ) , kt 7→ t1−θ u(t)kLp (R+ , dt ;X1 ) ,
t t
766 L The trace method for real interpolation

where the infimum over u extends over all strongly measurable functions
u : (0, ∞) → X0 + X1 with the above three properties. Note that if
u : (0, ∞) → X0 + X1 is strongly measurable and satisfies (i) and (ii), then
u ∈ W 1,1 ((0, T ); X0 + X1 ) for all 0 < T < ∞; in particular, u is equal al-
most everywhere to a (uniquely defined) continuous function from [0, ∞) to
X0 + X1 . In condition (iii), we refer to this version when imposing u(0) = x.
We continue with a technical result which shows that, in the definition of
the trace method, we may restrict ourselves to functions u ∈ C 1 ((0, ∞); X0 ∩
X1 ) without changing the norm defined by (L.4). Indeed, this is due to the
fact that the constant ε > 0 can be taken arbitrarily small in both inequalities
in (L.6).
Proposition L.2.2. For ε > 0 let gε : [0, ∞) → R be the ‘tent shaped’ piece-
wise linear function which is identically zero on [0, 1] and [1 + 2ε, ∞) and
whose graph connects the points (1, 0), (1 + ε, ε−1 ) and (1 + 2ε, 0) linearly. Let
ϕε (t) := tgε (t). Let u : (0, ∞) → X0 + X1 be strongly measurable and satisfy
conditions (i) and (ii) of Definition L.2.1, and define uε : (0, ∞) → X0 + X1
by
Z ∞ Z ∞
dτ dτ
uε (t) := ϕε (t/τ )u(τ ) = ϕε (τ )u(t/τ ) . (L.5)
0 τ 0 τ
Then uε ∈ C([0, ∞); X0 + X1 ) ∩ C 1 ((0, ∞); X0 ∩ X1 ), we have uε (0) = u(0),
and for all 0 < θ < 1,

kt 7→ t1−θ u0ε (t)kLp (R+ , dt ;X0 ) 6 kt 7→ t1−θ u0 (t)kLp (R+ , dt ;X0 ) ,

t t
(L.6)
kt 7→ t1−θ uε (t)kLp (R+ , dt ;X1 ) 6 (1 + 2)kt 7→ t1−θ u(t)kLp (R+ , dt ;X1 ) ,
t t

and
2(1 + 2ε)2
kt 7→ t2−θ u0ε (t)kLp (R+ , dt ;X1 ) 6 kt 7→ t1−θ u(t)kLp (R+ , dt ;X1 ) .
t ε t

Proof. Observe that

Z ∞ Z ∞
dt
ϕε (t) = gε (t) dt = 1, (L.7)
0 t 0

and that the function uε is defined by convolution on the multiplicative

group (0, ∞) with Haar measure dτ /τ . From this it is easy to see that
u ∈ C([0, ∞); X0 + X1 ) implies uε ∈ C([0, ∞); X0 + X1 ) and uε (0) = x (use
the last expression in (L.5)). It also implies that uε ∈ C 1 ((0, ∞); X0 ∩ X1 )
and Z ∞
dτ
u0ε (t) = ϕ(τ )u0 (t/τ ) 2 .
0 τ
Writing Z ∞
dτ
t1−θ u0ε (t) = τ −θ ϕ(τ ) · (t/τ )1−θ u0 (t/τ ) ,
0 τ
 L.2 The trace method 767

Young’s inequality implies

kt 7→ t1−θ u0ε (t)kLp (R+ , dt ;X0 )

t

6 kt 7→ t−θ ϕε (t)kL1 (R+ , dt ) kt 7→ t1−θ u0 (t)kLp (R+ , dt ;X0 )

t t

6 kt 7→ t1−θ u0 (t)kLp (R+ , dt ;X0 ) ,

t

−θ
where we used (L.7) and the fact that t 6 1 on the support of ϕε . Similarly,

kt 7→ t1−θ uε (t)kLp (R+ , dt ;X1 )

t

6 kt 7→ t1−θ ϕε kL1 (R+ , dt ) kt 7→ t1−θ u(t)kLp (R+ , dt ;X1 )

t t

6 (1 + 2ε)kt 7→ t1−θ u(t)kLp (R+ , dt ;X1 ) .

t

Finally, in view of
Z ∞
dτ
t2−θ u0ε (t) = (t/τ )2−θ ϕ0ε (t/τ ) · τ 1−θ u(τ ) ,
0 τ

another application of Young’s inequality gives

kt 7→ t2−θ u0ε (t)kLp (R+ , dt ;X1 )

t

6 kt 7→ t2−θ ϕ0ε (t)kL1 (R+ , dt ) kt 7→ t1−θ u(t)kLp (R+ , dt ;X1 )

t t

6 (1 + 2ε)kt 7→ gε (t) + tgε0 (t)kL1 (R+ , dt) kt 7→ t 1−θ

u(t)kLp (R+ , dt ;X1 )
t
7
2+ 2ε
= (1 + 2ε) kt 7→ t1−θ u(t)kLp (R+ , dt ;X1 ) ,
ε t

where the last equality follows by exact computation of the L1 -norm in the
preceding line. 

Theorem L.2.3 (Lions). For all 0 < θ < 1 and 1 6 p 6 ∞ we have

(X0 , X1 )Tr
θ,p = (X0 , X1 )θ,p

with equivalent norms satisfying

1 12
kxkθ,p 6 kxk(X0 ,X1 )Tr 6 kxkθ,p ,
64Cθ θ,p 1−θ

where Cθ = max{θ−1 , (1 − θ)−1 }.

Proof. Let x ∈ (X0 , X1 )Trθ,p , choose u such that Definition L.2.1(i)-(iii) hold,
and let uε with 0 < ε < 1 be as in Proposition L.2.2.
Setting v(t) = tu0ε (t), for j ∈ {0, 1} we find

kt 7→ tj−θ v(t)kLp (R+ , dt ;Xj ) = kt 7→ t1+j−θ u0ε (t)kLp (R+ , dt ;Xj )

t t
768 L The trace method for real interpolation

6 Cε kt 7→ t1−θ u(j) (t)kLp (R+ , dt ;Xj )

t

with Cε = 2(1 + 2ε)2 /ε. Therefore,

kvkLp (R+ , dt ;X0 +X1 )

t

6 kvkLp (0,1, dt ;X0 ) + kvkLp (1,∞, dt ;X1 )

t t

6 kt 7→ t−θ v(t)kLp (R+ , dt ;X0 ) + kt 7→ t1−θ v(t)kLp (R+ , dt ;X1 )

t t

6 Cε kt 7→ t1−θ u(t)kLp (R+ , dt ;X0 ) + kt 7→ t1−θ u0 (t)kLp (R+ , dt ;X1 ) .

t t

The right-hand side is finite because u satisfies

R ∞ conditions (i) and (ii) of Defi-
nition L.2.1. As a consequence, the integral 0 v(t) dt t converges as a Bochner
integral in X0 + X1 (see (C.17)), and
Z ∞ Z ∞
dt
v(t) = u0ε (t) dt = lim uε (t) − lim uε (t). (L.8)
0 t 0 t→∞ t→0

The existence of the limits follows from the convergence of the integral. From
Proposition L.2.2 we see that limt→∞ uε (t) = 0 in X1 . From the definition of
uε we obtain limt→0 uε (t) = uε (0) = u(0) = x.
From Theorem C.3.14 it follows that x ∈ (X0 , X1 )θ,p and, using the nota-
tion of the theorem,

kxkθ,p 6 4Cθ kxkθ,p,p

6 4Cθ max kt 7→ tj−θ v(t)kLp (R+ , dt ;Xj )
j∈{0,1} t

2
(1 + 2ε)
6 8Cθ max kt 7→ t1−θ u(j) (t)kLp (R+ , dt ;Xj ) .
ε j∈{0,1} t

With ε = 1/2, taking the infimum over all admissible functions u gives the
bound
kxkθ,p 6 64Cθ kxk(X0 ,X1 )Tr
θ,p
.
In the converse direction, suppose that x ∈ (X0 , X1 )θ,p and fix ε > 0. By
R ∞ C.3.14 there exists a strongly measurable v : (0, ∞) → X0 ∩ X1 such
Theorem
that 0 v(t) dt
t = x and for j ∈ {0, 1},

kt 7→ tj−θ v(t)kLp (R+ , dt ;Xj ) 6 (1 + ε)kxkθ,p,p 6 12(1 + ε)kxkθ,p .

t

Let u : (0, ∞) → X0 + X1 be given by

Z ∞ Z 1
dτ dτ
u(t) = v(τ ) = v(t/τ ) .
t τ 0 τ
It follows that
Z 1
1−θ dτ
kt 7→ t u(t)kLp (R+ , dt ;X1 ) 6 kt 7→ t1−θ v(t/τ )kLp (R+ , dt ;X1 )
t
0
t τ
 L.2 The trace method 769
Z 1
= τ −θ kt 7→ t1−θ v(t)kLp (R+ , dt ;X1 ) dτ
t
0
6 12(1 + ε)(1 − θ)−1 kxkθ,p .

Similarly, since u0 = −v(t)/t we obtain

kt 7→ t1−θ u0 (t)kLp (R+ , dt ;X0 ) 6 kt 7→ t−θ v(t)kLp (R+ , dt ;X0 ) 6 12(1 + ε)kxkθ,p .
t t

Since also u(0) = x, we conclude that

n o
θ 0
kxkTr θ
θ,p 6 max kt 7→ t u(t)kLp (R+ , dt ;X0 ) , kt 7→ t u (t)kLp (R+ , dt ;X1 )
t t

6 12(1 + ε)(1 − θ)−1 kxkθ,p .

Since ε > 0 was arbitrary, this proves the result. 

As an application we obtain the following characterisation of real interpolation
spaces using analytic semigroups. We note that for I = R+ , the theorem is a
special case of Theorem 16.3.19 in the main text.

Theorem L.2.4. Let −A generate an analytic semigroup (S(t))t>0 on a Ba-

nach space X. Let either I = R+ and assume that S is uniformly bounded, or
let I = (0, T ). Let p ∈ [1, ∞] and θ ∈ (0, 1). Then
n o
(X, D(A))θ,p = x ∈ X : kt 7→ t1−θ AS(t)xkLp (I, dt ;X)
t

with equivalence of norms

kxk(X,D(A))θ,p h kxk + kt 7→ t1−θ AS(t)xkLp (I, dt ;X) ,

t

with implied constants only depending on θ and A. If 0 ∈ %(A) and I = R+ ,

we also have equivalence of homogeneous norms

kxk(X,D(A))θ,p h kt 7→ t1−θ AS(t)xkLp (R+ , dt ;X) .

t

Proof. In the proof below we use the shorthand notation

|||x|||I := kt 7→ t1−θ AS(t)xkLp (I, dt ;X) .

t

First we consider the case I = R+ under the additional assumption that

S is uniformly bounded. Let M1 , M2 > 0 be such that kS(t)k 6 M1 and
ktAS(t)k 6 M2 for all t > 0.
Suppose first that x ∈ X satisfies |||x|||R+ < ∞. We will prove that x ∈
(X, D(A))θ,p . Let u(t) = e−t S(t)x. Then u0 (t) = −e−t S(t)x − e−t AS(t)x, and

kt 7→ t1−θ u0 (t)kLp (R+ , dt ;X) 6 Kθ,p M1 kxk + |||x|||R+ ,

t
770 L The trace method for real interpolation

kt 7→ t1−θ u(t)kLp (R+ , dt ;D(A)) 6 Kθ,p M1 kxk + |||x|||R+ ,

t

where Kθ,p = kt 7→ t1−θ e−t kLp (R+ , dt ) . Therefore, by Theorem L.2.3, x ∈

t
(X, D(A))θ,p and

kxk(X,D(A))θ,p 6 64Cθ (Kθ,p M1 kxk + |||x|||R+ ).

If, in addition to the assumptions already made, we also assume that 0 ∈ %(A),
then S is uniformly exponentially stable (because A − ε is sectorial in that
case and hence generates a bounded analytic semigroup), and we can use
u(t) = S(t)x in the above proof instead. This gives

kt 7→ t1−θ u0 (t)kLp (R+ , dt ;X) = |||x|||R+ ,

t

kt 7→ t1−θ u(t)kLp (R+ , dt ;D(A)) 6 max{1, kA−1 k}|||x|||R+ ,

t

and consequently

kxk(X,D(A))θ,p 6 64Cθ max{1, kA−1 k}|||x|||R+ .

In the conversely direction, suppose that x ∈ (X, D(A))θ,p . By the defini-

tion of the K-method t 7→ t−θ K(t, x) belongs to Lp (R+ , dt
t ), where

K(t, x) = inf kx0 kX + tkx1 kD(A) : x0 ∈ X, x1 ∈ D(A), x0 + x1 = x .

Let x0 ∈ X and x1 ∈ D(A) satisfy x0 + x1 = x. Then

kAS(t)xk 6 kAS(t)x0 k + kAS(t)x1 k 6 t−1 (M2 kx0 k + tM1 kx1 kD(A) ).

Therefore, kAS(t)xk 6 t−1 max{M1 , M2 }K(t, x), and

|||x|||R+ 6 max{M1 , M2 }kt 7→ t−θ K(t, x)kLp (R+ , dt ;X)

t

= max{M1 , M2 }kxk(X,D(A))θ,p .

Since also kxk 6 Ckxk(X,D(A))θ,p , this concludes the proof for the case I = R+ .
In case I = (0, T ), we use a simple scaling argument. For this let M > 0
and ω > 0 be such that e−ωt kS(t)k 6 M e−t and e−ωt ktAS(t)k 6 M e−t for
all t > 0, and set Sω (t) = e−ωt S(t). In both of the implications below we will
use that (X, D(ω − A))1− p1 ,p = (X, D(A))1− p1 ,p with equivalent norms.
First suppose that |||x|||(0,T ) := kt1−θ AS(·)xkLp ((0,T ), dt ;X) < ∞. In order
t
to show that x ∈ (X, D(A))θ,p with the desired norm estimate, note

kt 7→ t1−θ ASω (t)xkLp (R+ , dt ;X) 6 |||x|||(0,T ) + M kt 7→ t−θ e−t kLp ((T,∞), dt ) kxk
t t
1
6 |||x|||(0,T ) + M T −θ− p kxk.

Similarly, we obtain
 L.3 Reiteration 771

kt 7→ t1−θ Sω (t)xkLp (R+ , dt ;X) 6 M kt 7→ t1−θ e−t kLp (R+ , dt ) kxk =: M Lθ,p kxk.
t t

Therefore, from the case I = R+ and the above observation we see that
x ∈ (X, D(A))θ,p and

ckxk(X,D(ω−A))θ,p 6 kt 7→ t1−θ (ω − A)Sω (t)xkLp (R+ , dt ;X) + kxk

t
1
−θ− p
6 |||x|||(0,T ) + (M T + M Lθ,p + 1)kxk.

In the converse direction, let x ∈ (X, D(A))θ,p . Then by the case I = R+

and the above estimates,

e−ωT |||x|||(0,T ) 6 kt 7→ t1−θ (ω − A)Sω (t)xkLp (R+ , dt ;X)

t

+ ωkt 7→ t1−θ Sω (t)xkLp (R+ , dt ;X)

t

6 Ckxk(X,D(ω−A))θ,p + ωM Lθ,p kxk,

which implies the required result. 

L.3 Reiteration
Next we will prove the reiteration theorem.

Theorem L.3.1 (Reiteration). Let p ∈ [1, ∞], and fix 0 6 θ0 < θ1 6 1 and
λ ∈ (0, 1). Suppose that Y0 , Y1 are Banach spaces with continuous embeddings

(X0 , X1 )θj ,1 ,→ Yj ,→ (X0 , X1 )θj ,∞ , j ∈ {0, 1},

with norm estimates

A−1 kxkθj ,∞ 6 kxkYj 6 Bkxkθj ,1 , x ∈ Xθj ,1 , j ∈ {0, 1},

where we set Y0 = X0 if θ0 = 0 and Y1 = X1 if θ1 = 1. Then

(Y0 , Y1 )λ,p = (X0 , X1 )θ,p with θ = (1 − λ)θ0 + λθ1 ,

with equivalent norms satisfying

θ1 − θ0 1024B n1 1 o
kxkθ,p 6 kxk(Y0 ,Y1 )λ,p 6 max , kxkθ,p .
A 1−λ λ 1−λ
As a special case we have


(X0 , X1 )θ0 ,p0 , (X0 , X1 )θ1 ,p1 λ,p
= (X0 , X1 )θ,p with θ = (1 − λ)θ0 + λθ1
(L.9)

with equivalent norms, for all choices p0 , p1 , p ∈ [1, ∞].

772 L The trace method for real interpolation

Proof. First let x ∈ (Y0 , Y1 )λ,p . Suppose that x0 ∈ Y0 and x1 ∈ Y1 are such
that x = x0 + x1 . Then

K(t, x; X0 , X1 ) 6 K(t, x0 ; X0 , X1 ) + K(t, x1 ; X0 , X1 )

6 tθ0 kx0 kθ0 ,∞ + tθ1 kx1 kθ1 ,∞
6 tθ0 A(kx0 kY0 + tθ1 −θ0 kx1 kY1 ),

Therefore, K(t, x; X0 , X1 ) 6 tθ0 AK(tθ1 −θ0 , x; Y0 , Y1 ), and consequently

kxkθ,p = kt 7→ t−θ K(t, x; X0 , X1 )kLp (R+ , dt )

t

6 Akt 7→ tθ0 −θ K(tθ1 −θ0 , x; Y0 , Y1 )kLp (R+ , dt )

t
−λ(θ1 −θ0 ) θ1 −θ0
= Akt 7→ t K(t , x; Y0 , Y1 )kLp (R+ , dt )
t
−1 −λ
6 A(θ1 − θ0 ) ks 7→ s K(s, x; Y0 , Y1 )kLp (R+ , ds )
s

= A(θ1 − θ0 )−1 kxk(Y0 ,Y1 )λ,p .

In the converse direction, suppose that x ∈ (X0 , X1 )θ,p and fix ε > 0. Let
u and uε be as in Definition L.2.1 and Proposition L.2.2, and set vε (t) :=
uε (t1/(θ1 −θ0 ) ). By the substitution s = t1/(θ1 −θ0 ) , the estimate

kykYi 6 Bkykθi ,1 6 Bkyk1−θ θi

X0 kykX1 , y ∈ X0 ∩ X1 ,
i
(L.10)

Hölder’s inequality (with exponents 1/(1 − θ0 ) and 1/θ0 ), we obtain

R0 : = t 7→ t1−λ vε0 (t) Lp (R+ , dt

t ;Y0 )

1
= t 7→ t−λ+1/(θ1 −θ0 ) u0ε (t1/(θ1 −θ0 ) ) Lp (R+ , dt
θ1 − θ0 t ;Y0 )

= s 7→ s1−θ+θ0 u0ε (s) Lp (R+ , ds ;Y0 )

s

6 B s 7→ (s 1−θ
ku0ε (s)k)1−θ
X0 (s
0 2−θ
ku0ε (s)kX1 )θ0
Lp (R+ , ds
s )
1−θ0 θ0
6 B s 7→ s1−θ u0ε (s) Lp (R+ , ds
s 7→ s2−θ u0ε (s) Lp (R+ , ds
s ;X0 ) s ;X1 )
1−θ0 θ0
6 BCε s 7→ s1−θ u0 (s) Lp (R+ , ds ;X0 ) s 7→ s1−θ u(s) Lp (R+ , ds ;X1 )
s s
n o
1−θ0 θ0
6 BCε max s 7→ s1−θ u0 (s) Lp (R+ , ds ;X0 ) , s 7→ s1−θ u(s) Lp (R+ , ds ;X1 )
s s

2
2(1+2ε)
applying Proposition L.2.2 in the penultimate
R ∞ 0 step, with Cε := ε . By
same substitution and writing uε (s) = s uε (r) dr (see (L.8)), it follows from
Hardy’s inequality (see Lemma L.3.2(2) below) and (L.10) that

R1 : = t 7→ t1−λ vε (s) Lp (R+ , dt

t ;Y1 )
1−λ 1/(θ1 −θ0 )
= t 7→ t uε (t ) Lp (R+ , dt
t ;Y1 )
 L.3 Reiteration 773

= (θ1 − θ0 ) s 7→ sθ1 −θ uε (s) Lp (R+ , ds

s ;Y1 )

θ1 − θ0
6 ks 7→ s1−θ+θ1 u0ε (s)kLp (R+ , ds ;Y1 )
θ1 − θ s

B
6 ks 7→ (s1−θ ku0ε (s)kX0 )1−θ1 (s2−θ ku0ε (s)kX1 )θ1 kLp (R+ , ds )
1−λ s

B
6 ks 7→ s1−θ u0ε (s)kLp (R1 , ds ;X ) ks 7→ s2−θ u0ε (s)kL1p (R , ds ;X )
1−θ θ
1−λ + s 0 + s 1

BCε
6 ks 7→ s1−θ u0 (s)k1−θ 1
Lp (R+ , ds
ks 7→ s1−θ u(s)kθL1p (R , ds ;X )
1−λ s ;X0 ) + s 1

BCε n
1−θ0 θ0
o
6 max s 7→ s1−θ u0 (s) Lp (R+ , ds ;X0 ) , s 7→ s1−θ u(s) Lp (R+ , ds ;X1 )
1−λ s s

2
with Cε := 2(1+2ε)
ε as before.
Combining these estimates and taking the infimum over all admissible
functions u, we obtain the bound
BCε
kxk(Y0 ,Y1 )Tr 6 kxk(X0 ,X1 )Tr .
λ,p 1−λ θ,p

1
Setting ε = 2 gives Cε = 16, and using the estimate from Theorem L.2.3, we
obtain
1 o
n1
kxk(Y0 ,Y1 )λ,p 6 64 max kxk(Y0 ,Y1 )Tr
,
λ 1−λ λ,p

1024B n1 1 o
6 max , kxk(X0 ,X1 )Tr .
1−λ λ 1−λ θ,p


In the above proof we used Hardy’s inequality.
Lemma L.3.2 (Hardy’s inequality). Let p ∈ [1, ∞].
(1) If α > −1 and f : R+ → [0, ∞] is measurable, then
Z s
−α 1 1
s 7→ s f (t) dt p 6 kf kLp (R+ ,t−αp dt
s 0 ds
L (R+ , s ) |α + 1| t )

(2) If α < −1 and f : R+ → [0, ∞] is measurable, then

1 ∞
Z
1
s 7→ s−α f (t) dt 6 kf kLp (R+ ,t−αp dt
s s Lp (R+ , ds ) |α + 1| t )
s

Proof. (1): Without loss of generality

Rs we may assume that f is integrable on
finite intervals. Set u(s) := 1s 0 f (t) dt and note that
Z 1
−α
s u(s) = s−α f (θs) dθ.
0
774 L The trace method for real interpolation

Taking Lp -norms and applying Minkowski’s inequality, we obtain

Z 1
−α
ks 7→ s u(s)kLp (R+ , ds ) 6 ks 7→ s−α f (θs)kLp (R+ , ds ) dθ
s s
0
Z 1
= θα kt 7→ t−α f (t)kLp (R+ , dt ) dθ
t
0
1
= kt 7→ t−α f (t)kLp (R+ , dt )
α+1 t

∞
(2): This is proved similarly, this time setting v(s) := 1s s f (t) dt and
R

noting that Z ∞
s−α v(s) = s−α f (θs) dθ.
1


L.4 Mixed derivatives and Sobolev embedding

In this section we prove various weighted embedding results relating to the

trace method. This is done first for functions spaces over R+ ; the case of
bounded intervals is treated afterwards by using extension techniques.
We begin by introducing some terminology. For α ∈ [0, 1) we consider the
weight
wα (t) = tα .
Let I ⊆ R+ be an interval. By definition, a strongly measurable function
u : I → X belongs to Lpwα (I; X) if t 7→ tα u(t) belongs to Lp (I; X), and in this
case we set
kukLpwα (I;X) := kwα ukLp (I;X) .
We furthermore define Ww1,p α
(I; X) to be the space of all f ∈ Lpwα (I; X) ad-
0
mitting a weak derivative f belonging to Lpwα (I; X). As in the unweighted
case, this space is a Banach space with respect to the norm

kf kWw1,p (I;X) := kf kLpwα (I;X) + kf 0 kLpwα (I;X) .

α

For α ∈ [0, 1/p0 ), functions in Ww1,p

α
(I; X) always have a version that is con-
tinuous on I, and we will always use this version.
As an application of the trace method, for interpolation couples (X0 , X1 )
we obtain some mixed derivative result for the spaces

Ww1,p
α
(I; X0 ) ∩ Lpwα (I; X1 ).
 L.4 Mixed derivatives and Sobolev embedding 775

L.4.a Results for the half-line

Theorem L.4.1 (Mixed derivatives and traces). Let p ∈ (1, ∞] and α ∈

[0, 1/p0 ), where we take α 6= 0 if p = ∞, and let θ := 1 − α − p1 . Then for all
u ∈ Ww1,p α
(R+ ; X0 ) ∩ Lpwα (R+ ; X1 ) the estimate

sup ku(t)k(X0 ,X1 )θ,p 6 Kθ ku0 k1−θ

Lp
kukθLpw (L.11)
w (R+ ;X0 )
α α (R+ ;X1 )
t>0

holds with Kθ = 64 max{θ−1 , (1 − θ)−1 }. If p ∈ (1, ∞), we moreover have a

continuous embedding

Ww1,p
α
(R+ ; X0 ) ∩ Lpwα (R+ ; X1 ) ,→ Cb ([0, ∞); (X0 , X1 )θ,p ) (L.12)

and
1/p 1−1/p
sup tα ku(t)k(X0 ,X1 )1− 1 ,p 6 K1− p1 ku0 kLpw kukLpw . (L.13)
t>0 p α (R+ ;X0 ) α (R+ ;X1 )

The above result is often applied in the setting where we have a continuous
embedding X1 ,→ X0 . In that case, (L.13) shows instantaneous regularisation
in case α > 0.
Remark L.4.2. The estimate (L.11) extends to the case p = 1 and α = 0
(in whichR case we have θ = 0, Xθ,1 = X0 , and w0 ≡ 1). Indeed, writing
∞
u(t) = − t u0 (s) ds and using density of compactly supported functions, we
see that W 1,1 (R; X0 ) ,→ C([0, ∞); X0 ) continuously, and

sup ku(t)kX0 6 ku0 kLp (R+ ;X0 ) .

t>0

Proof of Theorem L.4.1. Let u ∈ Ww1,p α

(R+ ; X0 ) ∩ Lpwα (R+ ; X1 ) be arbitrary
and fixed.
Step 1 – The assumptions imply that for j ∈ {0, 1} we have u(1−j) ∈
Lwα (R+ ; Xj ). Using that α = 1 − θ − p1 , we have
p

1
ku(1−j) kLpwα (R+ ;Xj ) = kt 7→ t1−θ− p u(1−j) (t)kLp (R+ ;Xj )
= kt 7→ t1−θ u(1−j) (t)kLp (R+ , dt ;Xj ) .
t

By the argument after Definition L.2.1, this implies that u ∈ W 1.1 (0, T ); X0 +
X1 ) for all T > 0, and therefore u has a version belonging to C([0, ∞; X0 +X1 ).
We will denote this version by u again. By Definition L.2.1, applied to x :=
u(0), we obtain that u(0) ∈ (X0 , X1 )θ,p and, reversing the above identities,

ku(0)k(X0 ,X1 )θ,p 6 Kθ ku(0)k(X0 ,X1 )Tr

θ,p

= Kθ max kt 7→ t1−θ u(1−j) (t)kLp (R+ , dt ;Xj )

j∈{0,1} t (L.14)
= Kθ max ku(1−j) kLpwα (R+ ;Xj ) ,
j∈{0,1}
776 L The trace method for real interpolation

Step 2 – In this step we prove the bound (L.11). Applying (L.14) to uλ :=

u(λ·), we obtain

ku(0)k(X0 ,X1 )θ,p 6 Kθ max λθ−j ku(1−j) kLw

p
α (R+ ;Xj )
.
j∈{0,1}

Optimising over λ > 0, we obtain

ku(0)k(X0 ,X1 )θ,p 6 Kθ ku0 kL

1−θ
p
θ
kukL p . (L.15)
w (R+ ;X0 )
α w α (R+ ;X1 )

Applying (L.15) to u(· + t) for fixed t > 0, we obtain

ku(t)k(X0 ,X1 )θ,p 6 Kθ ku0 (· + t)kL

1−θ
p ku(· + t)kθLw
p
w (R+ ;X0 )
α α (R+ ;X1 )

6 Kθ ku0 k1−θ
p kukθLw
p ,
Lw α (R+ ;X0 ) α (R+ ;X1 )

which is (L.11).
Step 3 – In the remainder of the proof we assume that p ∈ (1, ∞). In the
present step we prove the continuous embedding (L.12). Applying (L.15) to
u(· + t) − u(· + s) with 0 6 s 6 t, we obtain

ku(t) − u(s)k(X0 ,X1 )θ,p

6 Kθ ku0 (t + ·) − u0 (s + ·)kL
1−θ
p
θ
ku(t + ·) − u(s + ·)kL p
w (R+ ;X0 )
α w α (R+ ;X1 )
0
6 Kθ ku (t − s + ·) − u0 kL
1−θ
p
wα (R+ ;X0 )
ku(t − s + ·) − θ
ukL p
wα (R+ ;X1 )
.
(L.16)

The latter tends to zero as t − s → 0 by the continuity of translations in

p
Lw α
(R+ ; Xi ) (here we use that p ∈ (1, ∞)).
Step 4 – Still under the assumption that p ∈ (1, ∞), we will now prove
(L.13). Fix an arbitrary t > 0. For j ∈ {0, 1},

tα ku(1−j) (· + t)kLp (R+ ;Xj ) = tα ku(1−j) kLp (t,∞;Xj ) 6 ku(1−j) kLw

p
α (R+ ;Xj )
.
(L.17)
1
Applying (L.15) with α = 0 and θ = 1 − p to the function u(· + t), we obtain
the estimate
1/p 1−1/p
ku(t)k(X0 ,X1 )1− 1 ,p 6 K1− p1 ku0 (· + t)kLp (R+ ;X0 ) ku(· + t)kLp (R+ ;X1 ) .
p

Multiplying on both sides with tα (where now α can be allowed to take any
value in [0, 1/p0 )) and using (L.17), we obtain
1/p 1−1/p
tα ku(t)k(X0 ,X1 )1− 1 ,p 6 K1− p1 tα ku0 (· + t)kLp (R+ ;X0 ) ku(· + t)kLp (R+ ;X1 )
p

1/p 1−1/p
6 K1− p1 ku0 kLpw kukLpw ,
α (R+ ;X0 ) α (R+ ;X1 )

and (L.13) follows. 

 L.4 Mixed derivatives and Sobolev embedding 777

We also need a variant of Theorem L.4.1 in the special case of weighted spaces
of continuous functions. For α ∈ [0, 1), and I = (0, T ] or I = (0, ∞), define

Cwα (I; X) := {u ∈ C(I; X) : sup tα ku(t)k < ∞},

t∈I
Cwα ,0 (I; X) := {u ∈ Cwα (I; X) : lim tα ku(t)k = 0},
t↓0
1
Cw α
(I; X) := {u ∈ Cwα (I; X) : u0 ∈ Cwα (I; X)}.

These are Banach spaces with respect to the norms

kukCwα (I;X) := sup tα ku(t)k,

t∈I
kukCw1 α (I;X) := sup tα ku(t)k + sup tα ku0 (t)k,
t∈I t∈I

and Cwα ,0 (I; X) is a closed subspace of Cwα (I; X).

Corollary L.4.3 (Mixed derivatives and traces). Let α ∈ (0, 1), set θ :=
1 − α. Then we have a continuous embedding
1
Cw α
((0, ∞); X0 ) ∩ Cwα ,0 ((0, ∞); X1 ) ,→ Cb ([0, ∞))
1
Proof. Fix an arbitrary Cw α
((0, ∞); X0 ) ∩ Cwα ,0 ((0, ∞); X1 ). It is clear that
u : (0, ∞) → X0 ∩ X1 is continuous, and we have already seen in Theorem
L.4.1 that u : [0, ∞) → (X0 , X1 )θ,p is continuous. It follows that on (0, ∞), u
takes values in Xp,∞ . Also, by (L.11) (applied with p = ∞),

sup ku(t)k(X0 ,X1 )θ,p 6 Kθ ku0 k1−θ θ

Cw (R+ ;X0 ) kukCwα (R+ ;X1 ) . (L.18)
α
t>0

It thus remains to show that u(τ ) → u(0) in (X0 , X1 )θ,p as τ ↓ 0; this in

particular implies that u(0) ∈ Xθ,p . (Note that this cannot be extracted from
(L.12), as the case p = ∞ was excluded there).
Replacing u by φu, where φ ∈ C 1 ([0, ∞)) is such that φ = 1 on [0, 1/2] and
φ = 0 on [1, ∞), we may assume that u ≡ 0 on [1, ∞). Let τ > 0. Applying
(L.18) to u(· + τ ) − u(·) and using the estimate

ku0 − u0 (· + τ )kCwα (R+ ;X0 ) 6 2ku0 kCwα (R+ ;X0 ) ,

from (L.16) we find

ku(τ ) − u(0)k(X0 ,X1 )θ,p 6 21−θ ku0 kC

1−θ
w (R+ ;X0 )
θ
ku(· + τ ) − ukC wα (R+ ;X1 )
.
α

To complete the proof, it thus suffices to show that ku(·+τ )−ukCwα (R+ ;X1 ) →
0 as τ ↓ 0. To this end let ε > 0, and choose δ0 ∈ (0, 1) such that for all
t ∈ (0, 2δ0 ] we have tα ku(t)kX1 < ε. By the continuity of u on (0, ∞) as an
X0 ∩ X1 -valued function and the support condition, u is uniformly continuous
on [δ0 , ∞) as an X0 ∩X1 -valued function, hence also as an X1 -valued function.
778 L The trace method for real interpolation

Therefore, we can find δ1 > 0 such that for all τ ∈ (0, δ1 ) and t > δ0 we have
ku(t + τ ) − u(t)kX1 < ε. Setting δ = min{δ0 , δ1 }, we find that for all τ ∈ (0, δ),

sup tα ku(t + τ ) − u(t)kX1 6 sup 2tα ku(t)kX1 < ε,

t∈(0,δ0 ) t∈(0,2δ0 )

sup tα ku(t + τ ) − u(t)kX1 6 sup ku(t + τ ) − u(t)kX1 < ε.

t∈[δ0 ,∞) t∈[δ0 ,1]

This gives the claim and completes the proof. 

As a consequence we obtain the following result. It reduces to Theorem L.4.1
upon letting h → ∞.
Theorem L.4.4 (Mixed derivatives and Sobolev embedding). Let p ∈
[1, ∞] and α ∈ [0, 1/p0 ) ∪ {0} be such that α + p1 > 0. Let h ∈ (1, ∞), and set
r = hp and θ = 1 − (α + p1 )(1 − h1 ). Then we have a continuous embedding

Ww1,p
α
(R+ ; X0 ) ∩ Lpwα (R+ ; X1 ) ,→ Lrwα/h (R+ ; (X0 , X1 )θ,1 ),

and there exists a constant Cp,α,h , depending only on (p, α, h), such that

kukLrwα (R+ ;(X0 ,X1 )θ,1 ) 6 Cp,α,h ku0 k1−θ

Lp
kukθLpw .
w (R+ ;X0 ) α α (R+ ;X1 )

Proof. First assume that p > 1. Let σ = 1 − α − p1 . (L.9) (taking θ0 = σ and

θ1 = 1, and A the norm constant in the embedding (X0 , X1 )σ,p ,→ (X0 , X1 )σ,∞
in (C.6)) and (L.2),

kxkθ,1 6 Ckxk((X0 ,X1 )σ,p ,X1 )1−λ,1 6 Ckxkλσ,p kxk1−λ

X1 (L.19)
1−θ 1
if λ ∈ (0, 1) satisfies σλ + (1 − λ) = θ, that is, if λ = 1
α+ p
= 1− h. In
particular, r(1 − λ) = p, and therefore,

kukLrw (R+ ;(X0 ,X1 )θ,1 ) 6 CkukλL∞ (R+ ;Xσ,p ) kuk1−λ

r(1−λ)
α/h L wα (R+ ;X1 )
(L.20)
= CkukλL∞ (R+ ;Xσ,p ) kuk1−λ
Lp
wα (R+ ;X1 )
,

where C depends on (p, α, h). On the other hand, by Theorem L.4.1,

kukL∞ (R+ ;Xσ,p ) 6 Kσ ku0 k1−σ

Lp
kukσLpw .
w (R+ ;X0 ) α α (R+ ;X1 )

We conclude that
λ(1−σ)
kukLrw (R+ ;(X0 ,X1 )θ,1 ) 6 C 0 Kσλ ku0 kLpw kuk1−λ+λσ
Lp
α/h α (R+ ;X0 ) wα (R+ ;X1 )
1
1− h
= C 0 Kσ ku0 k1−θ
Lp
kukθLpw ,
w (R+ ;X0 )
α α (R+ ;X1 )

where C 0 depends on (p, α, h).

If p = 1, then α = σ = 0, and (L.20) still holds if we replace Xσ,p by X0 .
Combining this with Remark L.4.2, this gives the desired embedding. 
 L.4 Mixed derivatives and Sobolev embedding 779

L.4.b Extension operators

By using extension operators, we can obtain versions of the previous results

on bounded intervals.

Proposition L.4.5 (Extension operators). Let p ∈ [1, ∞], and let α ∈

[0, 1/p0 ) ∪ {0}. There exists a linear mapping ET : L0 (0, T ; X) → L0 (R+ ; X)
with the following properties:
(1) For all u ∈ L0 (0, T ; X) we have ET u|(0,T ) = u;
(2) For all u ∈ Lpwα (0, T ; X),

kET ukLpwα (R+ ;X) 6 2kukLpwα (0,T ;X) ;

(3) For all u ∈ Ww1,p

α
(0, T ; X),

kET ukWw1,p (R+ ;X) 6 (2 + 3T −1 )kukWw1,p (0,T ;X) ;

α α

(4) For all u ∈ Ww1,p

α
(0, T ; X) with u(0) = 0,

kET ukWw1,p (R+ ;X) 6 Cα,p kukWw1,p (0,T ;X) ,

α α

0
where Cα,p = 2 + 3(1 − αp0 )1/p if p > 1 and Cα,p = 5 if p = 1.
p
The assertions (2), (3), and (4) also hold with Lw α
(0, T ; X) and Ww1,p
α
(0, T ; X)
1
replaced by Cwα ((0, T ]; X) and Cwα ((0, T ]; X), respectively.

Often we will need (4) in order to ensure that the bounds are T -independent
(note that the bound in (3) involves th term T −1 ). This is not possible without
the condition u(0) = 0 if p < ∞. Indeed, let u ∈ Ww1,p α
(0, T ; X) be such that
ku(0)k = 1. Then by Sobolev embedding

1 = ku(0)kX 6 CkET ukWw1,p (R+ ;X) 6 CkET k kukWw1,p (0,T ;X) .

α α

Now since kukWw1,p (0,T ;X) → 0 as T ↓ 0, we see that kET k 1,p → ∞ as

α Wwα (0,T ;X)

T ↓ 0.
Proof of Proposition L.4.5. (1), (2), and (3): First let T = 1. Let φ : [1, ∞) →
[0, 1] be given by φ(t) = max{3 − 2t, 0}. Set

u(t)
 on (0, 1),
E1 u(t) := φ(t)u(2 − t) on (1, 23 ),
on [ 23 , ∞).

0


If u ∈ Lpwα (0, 1; X), then E1 u ∈ Lpwα (R+ ; X), and if u ∈ Ww1,p α

(0, 1; X), then
α 0
E1 u ∈ Ww1,p
α
(R + ; X), and since t α
φ(t) 6 1 and t φ (t) 6 3 we obtain

kE1 ukLpwα (R+ ;X) 6 2kukLpwα (0,1;X) , if u ∈ u ∈ Lpwα (0, 1; X),

780 L The trace method for real interpolation

k(E1 u)0 kLpwα (R+ ;X) 6 2ku0 kLpwα (0,1;X) + 3kukLpwα (0,1;X) if u ∈ Ww1,p
α
(0, 1; X).
p
This implies the result for T = 1. In the general case let u ∈ Lw α
(0, T ; X) or
1,p
u ∈ Wwα (0, T ; X), and set uT (t) := u(tT ). Define the extension operator by

ET u(t) := E1 (uT )(t/T ).

For u ∈ Lpwα (0, T ; X) this leads to

1
kET ukLpwα (R+ ;X) = T α+ p kE1 uT kLpwα (R+ ;X)
1
6 2T α+ p kuT kLpwα (0,1;X) = 2kukLpwα (0,T ;X) .

Similarly, for u ∈ Ww1,p

α
(0, T ; X),
1
k(ET u)0 kLpwα (R+ ;X) = T α+ p −1 k(E1 uT )0 kLpwα (R+ ;X)
1 1
6 2T α+ p −1 ku0T kLpwα (0,1;X) + 3T α+ p −1 kuT kLpwα (0,1;X)
6 2ku0 kLpwα (0,T ;X) + 3T −1 kukLpwα (0,T ;X) .

(4): We claim that

T −1 kukLpwα (0,T ;X) 6 cα,p ku0 kLpwα (0,T ;X) ,

0
where cα,p = (1 − αp0 )1/p if p > 1 and cα,p = 1 if p = 1. Combined with the
proof of (3) the claim gives the desired result.
To prove the claim, we restrict ourselves to p ∈ (1, ∞). the simpler cases
p = 1 and p = ∞ are left to the reader. By Hölder’s inequality,
Z t Z t 1/p0
0
ku(t)k 6 ku0 (s)k ds 6 ku0 kLpwα (0,T ;X) s−αp ds
0 0
1
= cα,p ku kLpwα (0,T ;X) t1− p −α ,
0

Taking Lpwα -norms, we obtain

Z T 1/p
0
kuk Lp
wα (0,T ;X)
6 cα,p ku k Lp
wα (0,T ;X)
tp−1 dt 6 cα,p T ku0 kLpwα (0,T ;X) .
0

To prove the final assertion, let ε > 0. We can repeat the proof of (3)
and (4) with a slight deformation of φ, worsening the constant by at most an
additive term ε. It then remains to let ε ↓ 0. 

L.4.c Results for bounded intervals

As an application of the results of the foregoing sections we obtain the follow-

ing mixed derivative type result on bounded intervals. Recall the definition of
Xp,∞ in (L.1).
 L.4 Mixed derivatives and Sobolev embedding 781

Corollary L.4.6 (Mixed derivatives and traces). Let p ∈ (1, ∞] and

α ∈ [0, 1/p0 ), where we take α 6= 0 if p = ∞. Set θ := 1 − α − p1 . Then for all
0 < T < ∞ we have the (continuous) inclusions
1
Cw α
((0, T ]; X0 ) ∩ C0,wα ((0, T ]; X1 ) ,→ C([0, T ]; Xθ,∞ );
1
Cw α
((0, T ]; X0 ) ∩ C0,wα ((0, T ]; X1 ) ⊆ C((0, T ]; X0 ∩ X1 );

and, if p < ∞,

Ww1,p
α
(0, T ; X0 ) ∩ Lpwα (0, T ; X1 ) ,→ C([0, T ]; (X0 , X1 )θ,p ),
Ww1,p
α
(0, T ; X0 ) ∩ Lpwα (0, T ; X1 ) ⊆ C((0, T ); (X0 , X1 )1− p1 ,p ).

Moreover, for all u ∈ Ww1,p

α
(0, T ; X0 ) ∩ Lpwα (0, T ; X1 ) the following estimates
hold:

sup ku(t)k(X0 ,X1 )θ,p 6 Kθ CT kuk1−θ

W 1,p (0,T ;X )
kukθLpw ,
wα 0 α (R+ ;X1 )
t∈[0,T ]

sup tα ku(t)k(X0 ,X1 )1− 1 ,p 6 K1− p1 CT kuk1−θ

W 1,p (0,T ;X )
kukθLpw ,
p wα 0 α (0,T ;X1 )
t∈(0,T )

where Kθ = 64 max{θ−1 , (1 − θ)−1 } and CT = 2 + 3T −1 . For functions satis-

fying u(0) = 0, the above constants CT can be replaced by the T -independent
constant Cα,p defined in Proposition L.4.5.

By an argument similar to that after the statement of Proposition L.4.5,

without the condition u(0) it is not possible to obtain estimates with T -
independent constants for small T .
Proof. Let u ∈ Ww1,p
α
(0, T ; X0 )∩Lpwα (0, T ; X1 ). Combining Theorem L.4.1 and
Proposition L.4.5, we obtain that u ∈ C([0, T ]; (X0 , X1 )θ,p ) and

kukC([0,T ];(X0 ,X1 )θ,p ) 6 kET ukC([0,∞);(X0 ,X1 )θ,p )

6 Kθ kET uk1−θ
W 1,p (R
kET ukθLpw
wα + ;X0 ) α (R+ ;X1 )

−1
6 Kθ (2 + 3T )kuk1−θ
Ww1,p kukθLpw (0,T ;X1 ) .
α (0,T ;X0 ) α

From this estimate it is clear that if addition u(0) = 0 holds, then the constant
2 + 3T −1 can be replaced by the constant Cα,p of Proposition L.4.5.
The remaining cases are proved in the same way, using in addition Corol-
lary L.4.3 and Theorem L.4.4. 
In the same way we obtain the following result from Theorem L.4.4.

Corollary L.4.7 (Mixed derivatives and Sobolev embedding). Let p ∈

[1, ∞] and α ∈ [0, 1/p0 ) ∪ {0} satisfy α + p1 > 0. Let h ∈ (1, ∞), and set r = hp
and θ = 1 − (α + p1 )(1 − h1 ). Then we have a continuous embedding
782 L The trace method for real interpolation

Ww1,p
α
(0, T ; X0 ) ∩ Lpwα (0, T ; X1 ) ,→ Lw
r
α/h
(0, T ; (X0 , X1 )θ,1 ),

and there is a constant Cp,α,h only depending on (p, α, h) such that

kukLrwα (0,T ;(X0 ,X1 )θ,1 ) 6 Cp,α,h CT kuk1−θ

W 1,p (0,T ;X )
kukθLpw ,
wα 0 α (0,T ;X1 )

where Cp,α,h is as in Theorem L.4.4 and CT = 2 + 3T −1 . For functions satis-

fying u(0) = 0, the above constants CT can be replaced by the T -independent
constant Cα,p defined in Proposition L.4.5.

L.5 Notes
The trace method of Section L.2 is due to J. L. Lions in a classical series
of papers. We follow the presentation in Lunardi [2009] and Triebel [1978],
where a detailed historical account is given. A version of the trace method
with fractional smoothness was recently obtained in Agresti, Lindemulder,
and Veraar [2023]. The reiteration Theorem L.3.1 for real interpolation is due
to Lions and Peetre [1964]. A unified presentation of the reiteration method,
which covers many interpolation methods (including the real and complex
method), can be found in Lindemulder and Lorist [2021].
The mixed derivative result of Theorem L.4.1 is a standard consequence
of the trace method. The end-point case of Corollary L.4.3 is less standard,
but important in evolution equations, and can be found in Lunardi [1995].
The mixed derivative result of Theorem L.4.4 is also a simple consequence
of the trace method, although the simple proof presented here may be new.
Fractional versions have been proved in Agresti and Veraar [2022a].
The presentation of Proposition L.4.5, which is a standard result on ex-
tension operators for bounded intervals (0, T ), follows this reference. Other
constructions can be found in Meyries and Schnaubelt [2012b]. A discussion
on extension operators for more general domains in Rd can be found in the
notes of Chapter 14. In Corollaries L.4.6 and L.4.7 the extension operators are
used to obtain versions of the mixed derivative results on bounded intervals.
In Chapter 18 it is important that the embedding constants can be taken
independent of the size of the interval if one works with functions vanishing
at zero.
References

Acquistapace, P. and B. Terreni. A unified approach to abstract linear nonau-

tonomous parabolic equations. Rend. Sem. Mat. Univ. Padova, 78:47–107,
1987. 684
Acquistapace, P and B. Terreni. Regularity properties of the evolution opera-
tor for abstract linear parabolic equations. Differential Integral Equations,
5(5):1151–1184, 1992. 684
Adams, R. A. Sobolev spaces, volume Vol. 65 of Pure and Applied Mathemat-
ics. Academic Press, New York-London, 1975. 407
Adams, R. A. and J. J. F. Fournier. Sobolev spaces, volume 140 of Pure
and Applied Mathematics. Elsevier/Academic Press, Amsterdam, second
edition, 2003. 407
Agresti, A. Delayed blow-up and enhanced diffusion by transport noise for
systems of reaction-diffusion equations. arXiv:2207.08293, 2022. 731
Agresti, A. and M. C. Veraar. Stochastic Navier–Stokes equations for turbu-
lent flows in critical spaces. arXiv:2107.03953, 2021. 731
Agresti, A. and M. C. Veraar. Nonlinear parabolic stochastic evolution equa-
tions in critical spaces part I. Stochastic maximal regularity and local exis-
tence. Nonlinearity, 35(8):4100–4210, 2022a. 729, 731, 782
Agresti, A. and M. C. Veraar. Nonlinear parabolic stochastic evolution equa-
tions in critical spaces part II: Blow-up criteria and instataneous regular-
ization. J. Evol. Equ., 22(2):Paper No. 56, 96, 2022b. 730, 731
Agresti, A. and M. C. Veraar. The critical variational setting for stochastic
evolution equations. arXiv:2206.00230, 2022c. 731
Agresti, A. and M. C. Veraar. Reaction-diffusion equations with transport
noise and critical superlinear diffusion: Global well-posedness of weakly dis-
sipative systems. arXiv:2301.06897, 2023a. 731
Agresti, A. and M. C. Veraar. Reaction-diffusion equations with transport
noise and critical superlinear diffusion: Local well-posedness and positivity.
J. Differential Equations, 368:247–300, 2023b. 731

© Springer Nature Switzerland AG 2023 783

T. Hytönen et al., Analysis in Banach Spaces, Ergebnisse der Mathematik und ihrer
Grenzgebiete. 3. Folge / A Series of Modern Surveys in Mathematics 76,
https://doi.org/10.1007/978-3-031-46598-7
784 REFERENCES

Agresti, A., M. Hieber, A. Hussein, and M. Saal. The stochastic primitive

equations with transport noise and turbulent pressure. Stoch. Partial Differ.
Equ. Anal. Comput., pages 1–81, 2022a. 731
Agresti, A., M. Hieber, A. Hussein, and M. Saal. The stochastic primi-
tive equations with non-isothermal turbulent pressure. arXiv:2210.05973,
2022b. 731
Agresti, A., N. Lindemulder, and M. C. Veraar. Real interpolation by the
trace method for fractional smoothness. 2023. Unpublished manuscript.
410, 782
Aimar, H., L. Forzani, and F. J. Martı́n-Reyes. On weighted inequalities for
singular integrals. Proc. Amer. Math. Soc., 125(7):2057–2064, 1997. 80
Airta, E., H. Martikainen, and E. Vuorinen. UMD-extensions of Calderón–
Zygmund operators with mild kernel regularity. J. Fourier Anal. Appl., 28
(4):Paper No. 64, 40, 2022. 218, 222
Akrivis, G., B. Li, and C. Lubich. Combining maximal regularity and energy
estimates for time discretizations of quasilinear parabolic equations. Math.
Comp., 86(306):1527–1552, 2017. 687
Albrecht, D. Functional calculi of commuting unbounded operators. PhD
thesis, Monash University, 1994. 564
Albrecht, D., X. Duong, and A. McIntosh. Operator theory and harmonic
analysis. In Instructional workshop on analysis and geometry, Part III
(Canberra, 1995), volume 34 of Proc. Centre Math. Appl. Austral. Nat.
Univ., pages 77–136. Austral. Nat. Univ., Canberra, 1996. 502
Albrecht, D., E. Franks, and A. McIntosh. Holomorphic functional calculi
and sums of commuting operators. Bull. Austral. Math. Soc., 58(2):291–
305, 1998. 563, 564
Alexopoulos, G. La conjecture de Kato pour des opérateurs différentiels ellip-
tiques, à coefficients périodiques. C. R. Acad. Sci. Paris Sér. I Math., 312
(3):263–266, 1991. 510
Almeida, V., J. J. Betancor, J. C. Fariña, and L. Rodrı́guez-Mesa. Quan-
titative weighted estimates for harmonic analysis operators in the Bessel
setting by using sparse domination. Anal. Math. Phys., 12(4):Paper No.
101, 46, 2022. 76
Alouini, B. and O. Goubet. Regularity of the attractor for a Bose–Einstein
equation in a two dimensional unbounded domain. Discrete Contin. Dyn.
Syst. Ser. B, 19(3):651–677, 2014. 566
Amann, H. Linear and quasilinear parabolic problems. Vol. I: Abstract linear
theory, volume 89 of Monographs in Mathematics. Birkhäuser Boston Inc.,
Boston, MA, 1995. 407, 408, 502, 671, 674, 682, 684, 727, 742
Amann, H. Operator-valued Fourier multipliers, vector-valued Besov spaces,
and applications. Math. Nachr., 186:5–56, 1997. 407, 409, 679
Amann, H. Maximal regularity for nonautonomous evolution equations. Adv.
Nonlinear Stud., 4(4):417–430, 2004. 683
Amann, H. Quasilinear parabolic problems via maximal regularity. Adv.
Differential Equations, 10(10):1081–1110, 2005. 727
 REFERENCES 785

Amann, H. Linear and quasilinear parabolic problems. Vol. II: Function

spaces, volume 106 of Monographs in Mathematics. Birkhäuser/Springer,
Cham, 2019. 407, 415
Amann, H., M. Hieber, and G. Simonett. Bounded H∞ -calculus for elliptic
operators. Differ. Integral Equ., 7(3-4):613–653, 1994. 561
Amenta, A. and G. Uraltsev. Banach-valued modulation invariant Carleson
embeddings and outer-Lp spaces: the Walsh case. J. Fourier Anal. Appl.,
26(4):Paper No. 53, 54, 2020. 512
Amenta, A., E. Lorist, and M. C. Veraar. Rescaled extrapolation for vector-
valued functions. Publ. Mat., 63(1):155–182, 2019. 735
Anderson, T. C., B. Hu, and J. Roos. Sparse bounds for discrete singular
Radon transforms. Colloq. Math., 165(2):199–217, 2021. 79
Angenent, S. B. Nonlinear analytic semiflows. Proc. Roy. Soc. Edinburgh
Sect. A, 115(1-2):91–107, 1990a. 674, 728
Angenent, S. B. Parabolic equations for curves on surfaces. I. Curves with
p-integrable curvature. Ann. of Math. (2), 132(3):451–483, 1990b. 728
Antoni, M. Regular random field solutions for stochastic evolution equations.
PhD thesis, Karlsruher Institut für Technologie (KIT), 2017. 687
Arendt, W. Resolvent positive operators. Proc. London Math. Soc. (3), 54
(2):321–349, 1987. 762
Arendt, W. and M. M. Duelli. Maximal Lp -regularity for parabolic and elliptic
equations on the line. J. Evol. Equ., 6(4):773–790, 2006. 678
Arendt, W. and A. Zamboni. Decomposing and twisting bisectorial operators.
Studia Math., 197(3):205–227, 2010. 508
Arendt, W., R. Chill, S. Fornaro, and C. Poupaud. Lp -maximal regularity
for non-autonomous evolution equations. J. Differential Equations, 237(1):
1–26, 2007. 683, 688
Arendt, W., D. Dier, and S. Fackler. J. L. Lions’ problem on maximal regu-
larity. Arch. Math. (Basel), 109(1):59–72, 2017. 684
Arendt, W., M. Bernhard, and M. Kreuter. Elliptic problems and holomorphic
functions in Banach spaces. Illinois J. Math., 64(3):331–347, 2020. 291
Arnold, L. and C. Le Merdy. New counterexamples on Ritt operators, secto-
rial operators and R-boundedness. Bull. Aust. Math. Soc., 100(3):498–506,
2019. 682
Ashyralyev, A., S. Piskarev, and L. Weis. On well-posedness of difference
schemes for abstract parabolic equations in Lp ([0, T ]; E) spaces. Numer.
Funct. Anal. Optim., 23(7-8):669–693, 2002. 687
Astala, K. Area distortion of quasiconformal mappings. Acta Math., 173(1):
37–60, 1994. 72
Astala, K., T. Iwaniec, and E. Saksman. Beltrami operators in the plane.
Duke Math. J., 107(1):27–56, 2001. 72, 73
Auscher, P. On necessary and sufficient conditions for Lp -estimates of Riesz
transforms associated to elliptic operators on Rn and related estimates.
Mem. Amer. Math. Soc., 186(871):xviii+75, 2007. 512
786 REFERENCES

Auscher, P. and A. Axelsson. Remarks on maximal regularity. In Parabolic

problems, volume 80 of Progr. Nonlinear Differential Equations Appl., pages
45–55. Birkhäuser/Springer Basel AG, Basel, 2011. 673
Auscher, P. and M. Egert. On non-autonomous maximal regularity for elliptic
operators in divergence form. Arch. Math. (Basel), 107(3):271–284, 2016.
678, 744
Auscher, P. and M. Egert. A universal variational framework for parabolic
equations and systems. hal-03879193v2, 2023. 678
Auscher, P. and T. P. Hytönen. Orthonormal bases of regular wavelets in
spaces of homogeneous type. Appl. Comput. Harmon. Anal., 34(2):266–
296, 2013. 70
Auscher, P. and S. Stahlhut. Remarks on functional calculus for perturbed
first order Dirac operators. In Operator theory in harmonic and non-
commutative analysis, pages 31–43. Springer-Verlag, 2013. 512
Auscher, P. and P. Tchamitchian. Square root problem for divergence opera-
tors and related topics. Astérisque, (249):viii+172, 1998. 510
Auscher, P., S. Hofmann, J. L. Lewis, and P. Tchamitchian. Extrapolation
of Carleson measures and the analyticity of Kato’s square-root operators.
Acta Math., 187(2):161–190, 2001. 510
Auscher, P., S. Hofmann, M. Lacey, A. McIntosh, and Ph. Tchamitchian. The
solution of the Kato square root problem for second order elliptic operators
on Rn . Ann. of Math. (2), 156(2):633–654, 2002. 510, 512
Auscher, P., C. Kriegler, S. Monniaux, and Pierre Portal. Singular integral
operators on tent spaces. J. Evol. Equ., 12(4):741–765, 2012a. 687
Auscher, P., S. Monniaux, and P. Portal. The maximal regularity operator on
tent spaces. Commun. Pure Appl. Anal., 11(6):2213–2219, 2012b. 687
Auscher, P., S. Monniaux, and P. Portal. On existence and uniqueness for
non-autonomous parabolic Cauchy problems with rough coefficients. Ann.
Sc. Norm. Super. Pisa Cl. Sci. (5), 19(2):387–471, 2019. 687
Axelsson, A., S. Keith, and A. McIntosh. Quadratic estimates and functional
calculi of perturbed Dirac operators. Invent. Math., 163(3):455–497, 2006.
510, 511, 512
Babenko, K. I. On conjugate functions. Doklady Akad. Nauk SSSR (N.S.),
62:157–160, 1948. 673
Babenko, K. I. An inequality in the theory of Fourier integrals. Izv. Akad.
Nauk SSSR Ser. Mat., 25:531–542, 1961. 283
Bagchi, S., S. Hait, L. Roncal, and S. Thangavelu. On the maximal function
associated to the spherical means on the Heisenberg group. New York J.
Math., 27:631–675, 2021. 79
Bahouri, H., J.-Y. Chemin, and R. Danchin. Fourier analysis and nonlinear
partial differential equations, volume 343 of Grundlehren der Mathematis-
chen Wissenschaften. Springer, Heidelberg, 2011. 407
Baillon, J.-B. Caractère borné de certains générateurs de semi-groupes
linéaires dans les espaces de Banach. C. R. Acad. Sci. Paris Sér. A-B,
290(16):A757–A760, 1980. 680
 REFERENCES 787

Baillon, J.-B. and Ph. Clément. Examples of unbounded imaginary powers of

operators. J. Funct. Anal., 100(2):419–434, 1991. 503
Balakrishnan, A. V. Fractional powers of closed operators and the semigroups
generated by them. Pacific J. Math., 10:419–437, 1960. 502
Barron, A. and J. Pipher. Sparse domination for bi-parameter operators using
square functions. arXiv:1709.05009, 2017. 79
Barron, A., J. M. Conde-Alonso, Y. Ou, and G. Rey. Sparse domination and
the strong maximal function. Adv. Math., 345:1–26, 2019. 79
Batty, C. J. K. Unbounded operators: functional calculus, generation, pertur-
bations. Extracta Math., 24(2):99–133, 2009. 561
Batty, C. J. K. and C. Chen. Besov spaces associated with non-negative
operators on Banach spaces. arXiv:2006.07008, 2020. 408
Batty, C. J. K., R. Chill, and S. Srivastava. Maximal regularity for second
order non-autonomous Cauchy problems. Studia Math., 189(3):205–223,
2008. 688
Bechtel, S. and F. Gabel. Non-autonomous Lq (Lp ) maximal regular-
ity for complex systems under mixed regularity in space and time.
arXiv:2208.02527, 2022. 684
Beckner, W. Inequalities in Fourier analysis. Ann. of Math. (2), 102(1):
159–182, 1975. 283
Beltran, D. and L. Cladek. Sparse bounds for pseudodifferential operators. J.
Anal. Math., 140(1):89–116, 2020. 79
Beltran, D., R. Oberlin, L. Roncal, A. Seeger, and B. Stovall. Variation bounds
for spherical averages. Math. Ann., 382(1-2):459–512, 2022a. 79
Beltran, D., J. Roos, and A. Seeger. Multi-scale sparse domination.
arXiv:2009.00227, 2022b. Mem. Amer. Math. Soc., to appear. 79
Benea, C., F. Bernicot, and T. Luque. Sparse bilinear forms for Bochner Riesz
multipliers and applications. Trans. London Math. Soc., 4(1):110–128, 2017.
78
Benedek, A., A.-P. Calderón, and R. Panzone. Convolution operators on
Banach space valued functions. Proc. Nat. Acad. Sci. U.S.A., 48:356–365,
1962. 71, 672
Bényi, Á., D. Maldonado, and V. Naibo. What is . . . a paraproduct? Notices
Amer. Math. Soc., 57(7):858–860, 2010. 210
Bergh, J. and J. Löfström. Interpolation spaces. An introduction, volume
223 of Grundlehren der Mathematischen Wissenschaften. Springer-Verlag,
Berlin, 1976. 407, 409, 410, 412
Berkson, E., T. A. Gillespie, and P. S. Muhly. Théorie spectrale dans les
espaces UMD. C. R. Acad. Sci. Paris Sér. I Math., 302(4):155–158, 1986a.
503
Berkson, E., T. A. Gillespie, and P. S. Muhly. Abstract spectral decomposi-
tions guaranteed by the Hilbert transform. Proc. London Math. Soc. (3),
53(3):489–517, 1986b. 211
Bernicot, F., D. Frey, and S. Petermichl. Sharp weighted norm estimates
beyond Calderón–Zygmund theory. Anal. PDE, 9(5):1079–1113, 2016. 78
788 REFERENCES

Bessaga, C. and A. Pelczyński. On bases and unconditional convergence of

series in Banach spaces. Studia Math., 17:151–164, 1958. 681
Beylkin, G., R. Coifman, and V. Rokhlin. Fast wavelet transforms and nu-
merical algorithms. I. Comm. Pure Appl. Math., 44(2):141–183, 1991. 213
Binz, T., M. Hieber, A. Hussein, and M. Saal. The primitive equations with
stochastic wind driven boundary conditions. arXiv:2009.09449, 2020. 730
Blasco, O. Hardy spaces of vector-valued functions: duality. Trans. Amer.
Math. Soc., 308(2):495–507, 1988. 70
Blasco, O. and S. Pott. Embeddings between operator-valued dyadic BMO
spaces. Illinois J. Math., 52(3):799–814, 2008. 211, 212
Blasco, O. and S. Pott. Operator-valued dyadic BMO spaces. J. Operator
Theory, 63(2):333–347, 2010. 212
Blunck, S. Maximal regularity of discrete and continuous time evolution equa-
tions. Studia Math., 146(2):157–176, 2001a. 687
Blunck, S. Analyticity and discrete maximal regularity on Lp -spaces. J. Funct.
Anal., 183(1):211–230, 2001b. 687
Blunck, S. and P. C. Kunstmann. Weighted norm estimates and maximal
regularity. Adv. Differential Equations, 7(12):1513–1532, 2002. 676, 687,
742
Blunck, S. and P. C. Kunstmann. Calderón–Zygmund theory for non-integral
operators and the H ∞ functional calculus. Rev. Mat. Iberoam., 19(3):919–
942, 2003. 512, 676
Blunck, S. and P. C. Kunstmann. Weak type (p, p) estimates for Riesz trans-
forms. Math. Z., 247(1):137–148, 2004. 676
Blunck, S. and P. C. Kunstmann. Generalized Gaussian estimates and the
Legendre transform. J. Operator Theory, 53(2):351–365, 2005. 676
Bony, J.-M. Calcul symbolique et propagation des singularités pour les
équations aux dérivées partielles non linéaires. Ann. Sci. École Norm. Sup.
(4), 14(2):209–246, 1981. 210
Borges, T., B. Foster, Y. Ou, J. Pipher, and Z. Zhou. Sparse bounds for
the bilinear spherical maximal function. J. Lond. Math. Soc. (2), 107(4):
1409–1449, 2023. 79
Bourgain, J. On trigonometric series in super reflexive spaces. J. London
Math. Soc. (2), 24(1):165–174, 1981. 283, 284
Bourgain, J. A Hausdorff–Young inequality for B-convex Banach spaces.
Pacific J. Math., 101(2):255–262, 1982. 283, 284, 285, 286, 287
Bourgain, J. Some remarks on Banach spaces in which martingale difference
sequences are unconditional. Ark. Mat., 21(2):163–168, 1983. 284, 289
Bourgain, J. Extension of a result of Benedek, Calderón and Panzone. Ark.
Mat., 22(1):91–95, 1984. 735
∞
Bourgain, J. Subspaces of lN , arithmetical diameter and Sidon sets. In Prob-
ability in Banach spaces, V (Medford, Mass., 1984), volume 1153 of Lecture
Notes in Math., pages 96–127. Springer, Berlin, 1985. 285
Bourgain, J. Vector-valued singular integrals and the H 1 -BMO duality. In
Probability theory and harmonic analysis (Cleveland, Ohio, 1983), vol-
 REFERENCES 789

ume 98 of Monogr. Textbooks Pure Appl. Math., pages 1–19. Dekker, New
York, 1986. 70, 211, 219
Bourgain, J. Vector-valued Hausdorff–Young inequalities and applications. In
Geometric aspects of functional analysis (1986/87), volume 1317 of Lecture
Notes in Math., pages 239–249. Springer-Verlag, Berlin, 1988a. 283, 287,
738, 739
Bourgain, J. Vector-valued Hausdorff–Young inequalities and applications. In
Geometric aspects of functional analysis (1986/87), volume 1317 of Lecture
Notes in Math., pages 239–249. Springer, Berlin, 1988b. 284, 285, 286
Bownik, M. and D. Cruz-Uribe. Extrapolation and factorization of matrix
weights. arXiv:2210.09443, 2022. 83
Brezis, H. and P. Mironescu. Gagliardo–Nirenberg, composition and products
in fractional Sobolev spaces. J. Evol. Equ., 1(4):387–404, 2001. 413
Brezis, H. and P. Mironescu. Gagliardo–Nirenberg inequalities and non-
inequalities: the full story. Ann. Inst. H. Poincaré C Anal. Non Linéaire,
35(5):1355–1376, 2018. 413, 414
Bu, F., T. P. Hytönen, D. Yang, and W. Yuan. Matrix-weighted Besov-type
and Triebel–Lizorkin-type spaces. arXiv:2304.00292, 2023. 223, 417
Bu, S. and J.-M. Kim. Operator-valued Fourier multiplier theorems on Triebel
spaces. Acta Math. Sci. Ser. B (Engl. Ed.), 25(4):599–609, 2005. 413, 679
Buckley, S. M. Estimates for operator norms on weighted spaces and reverse
Jensen inequalities. Trans. Amer. Math. Soc., 340(1):253–272, 1993. 72, 76
Bui, H.-Q. Weighted Besov and Triebel spaces: interpolation by the real
method. Hiroshima Math. J., 12(3):581–605, 1982. 416
Bui, H.-Q., M. Paluszyński, and M. H. Taibleson. A maximal function charac-
terization of weighted Besov–Lipschitz and Triebel–Lizorkin spaces. Studia
Math., 119(3):219–246, 1996. 416
Bui, H.-Q., M. Paluszyński, and M. Taibleson. Characterization of the Besov–
Lipschitz and Triebel–Lizorkin spaces. The case q < 1. In Proceedings of the
conference dedicated to Professor Miguel de Guzmán (El Escorial, 1996),
volume 3, pages 837–846, 1997. 416
Burkholder, D. L. A geometric condition that implies the existence of cer-
tain singular integrals of Banach-space-valued functions. In Conference on
harmonic analysis in honor of Antoni Zygmund, Vol. I, II (Chicago, Ill.,
1981), Wadsworth Math. Ser., pages 270–286. Wadsworth, Belmont, CA,
1983. 83
Butzer, P. L. and H. Berens. Semi-groups of operators and approximation,
volume 145 of Grundlehren der Mathematischen Wissenschaften. Springer-
Verlag New York Inc., New York, 1967. 502
Calderón, A.-P. Commutators of singular integral operators. Proc. Nat. Acad.
Sci. U.S.A., 53:1092–1099, 1965. 211, 216
Calderón, A.-P. Cauchy integrals on Lipschitz curves and related operators.
Proc. Nat. Acad. Sci. U.S.A., 74(4):1324–1327, 1977. 216
Calderón, A.-P. and A. Zygmund. On the existence of certain singular inte-
grals. Acta Math., 88:85–139, 1952. 71, 215, 287
790 REFERENCES

Cannarsa, P. and V. Vespri. On maximal Lp regularity for the abstract Cauchy

problem. Boll. Un. Mat. Ital. B (6), 5(1):165–175, 1986. 673
Castro, A. J. and T. P. Hytönen. Bounds for partial derivatives: necessity of
UMD and sharp constants. Math. Z., 282(3-4):635–650, 2016. 289, 290
Chanillo, S. and R. L. Wheeden. Some weighted norm inequalities for the
area integral. Indiana Univ. Math. J., 36(2):277–294, 1987. 84
Chen, W., R. Han, and M. T. Lacey. Weighted estimates for one-sided mar-
tingale transforms. Proc. Amer. Math. Soc., 148(1):235–245, 2020. 80
Chill, R. and A. Fiorenza. Singular integral operators with operator-valued
kernels, and extrapolation of maximal regularity into rearrangement invari-
ant Banach function spaces. J. Evol. Equ., 14(4-5):795–828, 2014. 673
Chill, R. and S. Król. Weighted inequalities for singular integral operators on
the half-line. Studia Math., 243(2):171–206, 2018. 80, 673
Chill, R. and S. Srivastava. Lp -maximal regularity for second order Cauchy
problems. Math. Z., 251(4):751–781, 2005. 687
Chill, R. and S. Srivastava. Lp maximal regularity for second order Cauchy
problems is independent of p. Boll. Unione Mat. Ital. (9), 1(1):147–157,
2008. 688
Choudhury, A. P., A. Hussein, and P. Tolksdorf. Nematic liquid crystals in
Lipschitz domains. SIAM J. Math. Anal., 50(4):4282–4310, 2018. 677
Christ, M. A T (b) theorem with remarks on analytic capacity and the Cauchy
integral. Colloq. Math., 60/61(2):601–628, 1990. 70
Cladek, L. and Y. Ou. Sparse domination of Hilbert transforms along curves.
Math. Res. Lett., 25(2):415–436, 2018. 79
Clément, Ph. and S. Li. Abstract parabolic quasilinear equations and appli-
cation to a groundwater flow problem. Adv. Math. Sci. Appl., 3(Special
Issue):17–32, 1993/94. 727, 730
Clément, Ph. and J. Prüss. An operator-valued transference principle and
maximal regularity on vector-valued Lp -spaces. In Evolution equations and
their applications in physical and life sciences (Bad Herrenalb, 1998), vol-
ume 215 of Lecture Notes in Pure and Appl. Math., pages 67–87. Dekker,
New York, 2001. 503, 564, 676
Clément, Ph. and G. Simonett. Maximal regularity in continuous interpolation
spaces and quasilinear parabolic equations. J. Evol. Equ., 1(1):39–67, 2001.
673, 674, 728
Coifman, R. R. and C. Fefferman. Weighted norm inequalities for maximal
functions and singular integrals. Studia Math., 51:241–250, 1974. 72
Coifman, R. R. and G. Weiss. Analyse harmonique non-commutative sur cer-
tains espaces homogènes. Lecture Notes in Mathematics, Vol. 242. Springer-
Verlag, Berlin-New York, 1971. 70, 213, 216
Coifman, R. R., A. McIntosh, and Y. Meyer. L’intégrale de Cauchy définit
un opérateur borné sur L2 pour les courbes lipschitziennes. Ann. of Math.
(2), 116(2):361–387, 1982. 216, 217, 510
 REFERENCES 791

Coifman, R. R., D. G. Deng, and Y. Meyer. Domaine de la racine carrée de

certains opérateurs différentiels accrétifs. Ann. Inst. Fourier (Grenoble), 33
(2):x, 123–134, 1983. 510
Conde-Alonso, J. M. and G. Rey. A pointwise estimate for positive dyadic
shifts and some applications. Math. Ann., 365(3-4):1111–1135, 2016. 76
Conde-Alonso, J. M., A. Culiuc, F. Di Plinio, and Y. Ou. A sparse domination
principle for rough singular integrals. Anal. PDE, 10(5):1255–1284, 2017.
78
Coulhon, T. and X. T. Duong. Maximal regularity and kernel bounds: obser-
vations on a theorem by Hieber and Prüss. Adv. Differential Equations, 5
(1-3):343–368, 2000. 676
Coulhon, T. and D. Lamberton. Régularité Lp pour les équations d’évolution.
In Séminaire d’Analyse Fonctionelle 1984/1985, volume 26 of Publ. Math.
Univ. Paris VII, pages 155–165. Univ. Paris VII, Paris, 1986. 673, 674, 679
Court, S. and K. Kunisch. Design of the monodomain model by artificial
neural networks. Discrete Contin. Dyn. Syst., 42(12):6031–6061, 2022. 730
Cowling, M., I. Doust, A. McIntosh, and A. Yagi. Banach space operators
with a bounded H ∞ functional calculus. J. Austral. Math. Soc. Ser. A, 60
(1):51–89, 1996. 503
Cruz-Uribe, D., J. Isralowitz, and K. Moen. Two weight bump conditions for
matrix weights. Integral Equations Operator Theory, 90(3):Paper No. 36,
31, 2018. 83
Cruz-Uribe, D. V., J. M. Martell, and C. Pérez. Sharp weighted estimates for
approximating dyadic operators. Electron. Res. Announc. Math. Sci., 17:
12–19, 2010. 73, 74
Culiuc, A., F. Di Plinio, and Y. Ou. Domination of multilinear singular in-
tegrals by positive sparse forms. J. Lond. Math. Soc. (2), 98(2):369–392,
2018a. 78
Culiuc, A., F. Di Plinio, and Y. Ou. Uniform sparse domination of singular
integrals via dyadic shifts. Math. Res. Lett., 25(1):21–42, 2018b. 82
Culiuc, A., R. Kesler, and M. T. Lacey. Sparse bounds for the discrete cubic
Hilbert transform. Anal. PDE, 12(5):1259–1272, 2019. 79
Da Prato, G. and P. Grisvard. Sommes d’opérateurs linéaires et équations
différentielles opérationnelles. J. Math. Pures Appl. (9), 54(3):305–387,
1975. 563, 566, 671, 673, 674, 678
Da Prato, G. and P. Grisvard. Equations d’évolution abstraites non linéaires
de type parabolique. Ann. Mat. Pura Appl. (4), 120:329–396, 1979. 673,
674
Da Prato, G. and A. Lunardi. Stability, instability and center manifold the-
orem for fully nonlinear autonomous parabolic equations in Banach space.
Arch. Rational Mech. Anal., 101(2):115–141, 1988. 674
Da Prato, G. and E. Sinestrari. Differential operators with nondense domain.
Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4), 14(2):285–344 (1988), 1987. 672
Danchin, R., M. Hieber, P. B. Mucha, and P. Tolksdorf. Free boundary prob-
lems via Da Prato–Grisvard theory. arXiv:2011.07918, 2020. 679
792 REFERENCES

David, G. Wavelets and singular integrals on curves and surfaces, volume

1465 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1991. 218
David, G. and J.-L. Journé. A boundedness criterion for generalized Calderón–
Zygmund operators. Ann. of Math. (2), 120(2):371–397, 1984. 212, 216,
219, 220, 222
David, G., J.-L. Journé, and S. Semmes. Opérateurs de Calderón–Zygmund,
fonctions para-accrétives et interpolation. Rev. Mat. Iberoam., 1(4):1–56,
1985. 216
Denk, R. and M. Kaip. General parabolic mixed order systems in Lp and
applications, volume 239 of Operator Theory: Advances and Applications.
Birkhäuser/Springer, Cham, 2013. 407, 688
Denk, R. and R. Schnaubelt. A structurally damped plate equation with
Dirichlet–Neumann boundary conditions. J. Differential Equations, 259(4):
1323–1353, 2015. 688
Denk, R. and L. R. Volevich. Parabolic boundary value problems connected
with Newton’s polygon and some problems of crystallization. J. Evol. Equ.,
8(3):523–556, 2008. 688
Denk, R., M. Hieber, and J. Prüss. R-boundedness, Fourier multipliers and
problems of elliptic and parabolic type. Mem. Amer. Math. Soc., 166(788),
2003. 502, 562, 671, 676
Denk, R., G. Dore, M. Hieber, J. Prüss, and A. Venni. New thoughts on old
results of R. T. Seeley. Math. Ann., 328(4):545–583, 2004. 561
Denk, R., M. Hieber, and J. Prüss. Optimal Lp -Lq -estimates for parabolic
boundary value problems with inhomogeneous data. Math. Z., 257(1):193–
224, 2007. 412, 677, 678
Denk, R., J. Prüss, and R. Zacher. Maximal Lp -regularity of parabolic prob-
lems with boundary dynamics of relaxation type. J. Funct. Anal., 255(11):
3149–3187, 2008a. 678
Denk, R., J. Saal, and J. Seiler. Inhomogeneous symbols, the Newton polygon,
and maximal Lp -regularity. Russ. J. Math. Phys., 15(2):171–191, 2008b. 688
Di Giorgio, D., A. Lunardi, and R. Schnaubelt. Optimal regularity and Fred-
holm properties of abstract parabolic operators in Lp spaces on the real
line. Proc. London Math. Soc. (3), 91(3):703–737, 2005a. 566
Di Giorgio, D., A. Lunardi, and R. Schnaubelt. Optimal regularity and Fred-
holm properties of abstract parabolic operators in Lp spaces on the real
line. Proc. London Math. Soc. (3), 91(3):703–737, 2005b. 684
Di Plinio, F. and Y. Ou. Banach-valued multilinear singular integrals. Indiana
Univ. Math. J., 67(5):1711–1763, 2018. 218, 512
Di Plinio, F., Y. Q. Do, and G. N. Uraltsev. Positive sparse domination of
variational Carleson operators. Ann. Sc. Norm. Super. Pisa Cl. Sci. (5),
18(4):1443–1458, 2018. 78
Di Plinio, F., T. P. Hytönen, and K. Li. Sparse bounds for maximal rough
singular integrals via the Fourier transform. Ann. Inst. Fourier (Grenoble),
70(5):1871–1902, 2020a. 78, 83
 REFERENCES 793

Di Plinio, F., K. Li, H. Martikainen, and E. Vuorinen. Multilinear operator-

valued Calderón–Zygmund theory. J. Funct. Anal., 279(8):108666, 62,
2020b. 209, 218, 219, 512
Di Plinio, F., A. W. Green, and B. D. Wick. Wavelet resolution and Sobolev
regularity of Calderón–Zygmund operators on domains. arXiv:2304.13909,
2023a. 222, 223
Di Plinio, F., W. Green, and B. D. Wick. Bilinear wavelet representation of
Calderón-Zygmund forms. Pure Appl. Anal., 5(1):47–83, 2023b. 222
Di Plinio, F., B. D. Wick, and T. Williams. Wavelet representation of singular
integral operators. Math. Ann., 386(3-4):1829–1889, 2023c. 222
Dier, D. and R. Zacher. Non-autonomous maximal regularity in Hilbert
spaces. J. Evol. Equ., 17(3):883–907, 2017. 678, 684, 744
Dispa, S. Intrinsic characterizations of Besov spaces on Lipschitz domains.
Math. Nachr., 260:21–33, 2003. 416
Disser, K., A. F. M. ter Elst, and J. Rehberg. On maximal parabolic regularity
for non-autonomous parabolic operators. J. Differential Equations, 262(3):
2039–2072, 2017. 744
Domelevo, K. and S. Petermichl. The dyadic Riesz vector I. arXiv:2309.02802,
2023a. 291
Domelevo, K. and S. Petermichl. The dyadic Riesz vector II.
arXiv:2309.02803, 2023b. 291
Domelevo, K. and S. Petermichl. The dyadic and the continuous Hilbert
transforms with values in Banach spaces. arXiv:2212.00090, 2023c. 291
Domelevo, K. and S. Petermichl. The dyadic and the continuous Hilbert
transforms with values in Banach spaces. Part 2. arXiv:2303.12480, 2023d.
291
Domingo-Salazar, C., M. Lacey, and G. Rey. Borderline weak-type estimates
for singular integrals and square functions. Bull. Lond. Math. Soc., 48(1):
63–73, 2016. 84
Dominguez, O. and M. C. Veraar. Extensions of the vector-valued Hausdorff-
Young inequalities. Math. Z., 299(1-2):373–425, 2021. 411
Dong, H. and C. Gallarati. Higher-order elliptic and parabolic equations with
VMO assumptions and general boundary conditions. J. Funct. Anal., 274
(7):1993–2038, 2018. 678
Dong, H. and C. Gallarati. Higher-order parabolic equations with VMO as-
sumptions and general boundary conditions with variable leading coeffi-
cients. Int. Math. Res. Not. IMRN, (7):2114–2144, 2020. 678
Dong, H. and D. Kim. On the Lp -solvability of higher order parabolic and
elliptic systems with BMO coefficients. Arch. Ration. Mech. Anal., 199(3):
889–941, 2011. 678, 683
Dong, H. and D. Kim. On Lp -estimates for elliptic and parabolic equations
with Ap weights. Trans. Amer. Math. Soc., 370(7):5081–5130, 2018. 678,
683
794 REFERENCES

Dong, H. and N. V. Krylov. Fully nonlinear elliptic and parabolic equations

in weighted and mixed-norm Sobolev spaces. Calc. Var. Partial Differential
Equations, 58(4):Paper No. 145, 32, 2019. 683
Dor, L. E. On projections in L1 . Ann. of Math. (2), 102(3):463–474, 1975.
286
Dore, G. H ∞ functional calculus in real interpolation spaces. Stud. Math.,
137(2):161–167, 1999. 502, 565
Dore, G. Maximal regularity in Lp spaces for an abstract Cauchy problem.
Adv. Differential Equations, 5(1-3):293–322, 2000. 672, 674, 680
Dore, G. H ∞ functional calculus in real interpolation spaces. II. Stud. Math.,
145(1):75–83, 2001. 502, 565
Dore, G. and A. Venni. On the closedness of the sum of two closed operators.
Math. Z., 196(2):189–201, 1987. 502, 507, 508, 674
Dragičević, O. Weighted estimates for powers and the Ahlfors–Beurling op-
erator. Proc. Amer. Math. Soc., 139(6):2113–2120, 2011. 73
Dragičević, O. and A. Volberg. Sharp estimate of the Ahlfors–Beurling opera-
tor via averaging martingale transforms. Michigan Math. J., 51(2):415–435,
2003. 73
Dragičević, O., L. Grafakos, M. C. Pereyra, and S. Petermichl. Extrapola-
tion and sharp norm estimates for classical operators on weighted Lebesgue
spaces. Publ. Mat., 49(1):73–91, 2005. 73
Duelli, M. M. Functional calculus for bisectorial operators and applications
to linear and non-linear evolution equations. PhD thesis, University of
Karlsruhe, 2005. 508, 509
Duelli, M. M. and L. W. Weis. Spectral projections, Riesz transforms and
H ∞ -calculus for bisectorial operators. In Nonlinear elliptic and parabolic
problems, volume 64 of Progr. Nonlinear Differential Equations Appl., pages
99–111. Birkhäuser, Basel, 2005. 508
Dunford, N. and J. T. Schwartz. Linear operators. I: General theory. Pure
and Applied Mathematics, Vol. 7. Interscience Publishers, Inc., New York;
Interscience Publishers, Ltd., London, 1958. 675
Duoandikoetxea, J. and J. L. Rubio de Francia. Maximal and singular integral
operators via Fourier transform estimates. Invent. Math., 84(3):541–561,
1986. 77
Duong, X. T. and D. W. Robinson. Semigroup kernels, Poisson bounds, and
holomorphic functional calculus. J. Funct. Anal., 142(1):89–128, 1996. 676
Eberhardt, B. and G. Greiner. Baillon’s theorem on maximal regularity. Acta
Appl. Math., 27(1-2):47–54, 1992. 680
Edgar, G. A. Extremal integral representations. J. Funct. Anal., 23(2):145–
161, 1976. 286
Engel, K.-J. and R. Nagel. One-parameter semigroups for linear evolution
equations, volume 194 of Graduate Texts in Mathematics. Springer-Verlag,
New York, 2000. 560, 742, 762
Fabes, E. B., D. S. Jerison, and C. E. Kenig. Multilinear square functions and
partial differential equations. Amer. J. Math., 107(6):1325–1368, 1985. 510
 REFERENCES 795

Fackler, S. The Kalton–Lancien theorem revisited: maximal regularity does

not extrapolate. J. Funct. Anal., 266(1):121–138, 2014. 681
Fackler, S. Maximal regularity: positive counterexamples on UMD-Banach
lattices and exact intervals for the negative solution of the extrapolation
problem. Proc. Amer. Math. Soc., 144(5):2015–2028, 2016. 681
Fackler, S. Nonautonomous maximal Lp -regularity under fractional Sobolev
regularity in time. Anal. PDE, 11(5):1143–1169, 2018. 684
Farwig, R. and H. Sohr. Weighted Lq -theory for the Stokes resolvent in exte-
rior domains. J. Math. Soc. Japan, 49(2):251–288, 1997. 677
Fefferman, R. Harmonic analysis on product spaces. Ann. of Math. (2), 126
(1):109–130, 1987. 79
Fefferman, R. Ap weights and singular integrals. Amer. J. Math., 110(5):
975–987, 1988. 79
Fefferman, R. and J. Pipher. Multiparameter operators and sharp weighted
inequalities. Amer. J. Math., 119(2):337–369, 1997. 77
Fefferman, R. and E. M. Stein. Singular integrals on product spaces. Adv. in
Math., 45(2):117–143, 1982. 79
Figiel, T. On equivalence of some bases to the Haar system in spaces of
vector-valued functions. Bull. Polish Acad. Sci. Math., 36(3-4):119–131
(1989), 1988. 210, 212
Figiel, T. Singular integral operators: a martingale approach. In Geometry
of Banach spaces (Strobl, 1989), volume 158 of London Math. Soc. Lecture
Note Ser., pages 95–110. Cambridge Univ. Press, Cambridge, 1990. 211,
212, 213, 215, 217, 220, 221, 736
Figiel, T. and P. Wojtaszczyk. Special bases in function spaces. In Handbook
of the geometry of Banach spaces, Vol. I, pages 561–597. North-Holland,
Amsterdam, 2001. 211
s
Franke, J. On the spaces Fpq of Triebel–Lizorkin type: pointwise multipliers
and spaces on domains. Math. Nachr., 125:29–68, 1986. 413
Frazier, M. and B. Jawerth. A discrete transform and decompositions of
distribution spaces. J. Funct. Anal., 93(1):34–170, 1990. 407
Frazier, M. and S. Roudenko. Matrix-weighted Besov spaces and conditions
of Ap type for 0 < p 6 1. Indiana Univ. Math. J., 53(5):1225–1254, 2004.
416
Frazier, M. and S. Roudenko. Traces and extensions of matrix-weighted Besov
spaces. Bull. Lond. Math. Soc., 40(2):181–192, 2008. 416
Frazier, M. and S. Roudenko. Littlewood-Paley theory for matrix-weighted
function spaces. Math. Ann., 380(1-2):487–537, 2021. 417
Frazier, M., R. Torres, and G. Weiss. The boundedness of Calderón–Zygmund
operators on the spaces Ḟpα,q . Rev. Mat. Iberoamericana, 4(1):41–72, 1988.
223
Frazier, M., Y.-S. Han, B. Jawerth, and G. Weiss. The T 1 theorem for Triebel–
Lizorkin spaces. In Harmonic analysis and partial differential equations (El
Escorial, 1987), volume 1384 of Lecture Notes in Math., pages 168–181.
Springer, Berlin, 1989. 222
796 REFERENCES

Frey, D., A. McIntosh, and P. Portal. Conical square function estimates and
functional calculi for perturbed Hodge–Dirac operators in LP . J. Anal.
Math., 134(2):399–453, 2018. 512
Friedman, A. Partial differential equations. Holt, Rinehart and Winston, Inc.,
New York-Montreal, Que.-London, 1969. 727
Fröhlich, A. Stokes- und Navier–Stokes-Gleichungen in gewichteten Funk-
tionenräumen. PhD thesis, TU Darmstadt, 2001. 677
Fröhlich, A. The Stokes operator in weighted Lq -spaces. I. Weighted estimates
for the Stokes resolvent problem in a half space. J. Math. Fluid Mech., 5
(2):166–199, 2003. 677
Fröhlich, A. The Stokes operator in weighted Lq -spaces. II. Weighted resolvent
estimates and maximal Lp -regularity. Math. Ann., 339(2):287–316, 2007.
677
Gallarati, C. and M. C. Veraar. Maximal regularity for non-autonomous equa-
tions with measurable dependence on time. Potential Anal., 46(3):527–567,
2017a. 683, 743
Gallarati, C. and M. C. Veraar. Evolution families and maximal regularity
for systems of parabolic equations. Adv. Differential Equations, 22(3-4):
169–190, 2017b. 683, 743
Gallarati, C., E. Lorist, and M. C. Veraar. On the `s -boundedness of a family
of integral operators. Rev. Mat. Iberoam., 32(4):1277–1294, 2016. 743
Ganguly, P. and S. Thangavelu. On the lacunary spherical maximal function
on the Heisenberg group. J. Funct. Anal., 280(3):Paper No. 108832, 32,
2021. 79
Garcı́a-Cuerva, J. and J. L. Rubio de Francia. Weighted norm inequalities and
related topics, volume 116 of North-Holland Mathematics Studies. North-
Holland Publishing Co., Amsterdam, 1985. 71, 73
Garcı́a-Cuerva, J., J. L. Torrea, and K. S. Kazarian. On the Fourier type of
Banach lattices. In Interaction between functional analysis, harmonic anal-
ysis, and probability (Columbia, MO, 1994), volume 175 of Lecture Notes in
Pure and Appl. Math., pages 169–179. Dekker, New York, 1996. 287, 738,
739
Garcı́a-Cuerva, J., K. S. Kazaryan, V. I. Kolyada, and J. L. Torrea. The
Hausdorff–Young inequality with vector-valued coefficients and applica-
tions. Uspekhi Mat. Nauk, 53(3(321)):3–84, 1998. translation in Russian
Math. Surveys 53 (1998), no. 3, 435–513. 411, 739
Garling, D. J. H. Random martingale transform inequalities. In Probability
in Banach spaces 6 (Sandbjerg, 1986), volume 20 of Progr. Probab., pages
101–119. Birkhäuser Boston, Boston, MA, 1990. 284
Garoni, C. and S. Serra-Capizzano. Generalized locally Toeplitz sequences:
theory and applications. Vol. I. Springer, Cham, 2017. 561
Geiss, S. A counterexample concerning the relation between decoupling con-
stants and UMD-constants. Trans. Amer. Math. Soc., 351(4):1355–1375,
1999. 284
 REFERENCES 797

Geiss, S., S. Montgomery-Smith, and E. Saksman. On singular integral and

martingale transforms. Trans. Amer. Math. Soc., 362(2):553–575, 2010.
289, 290, 680, 737, 742
Geissert, M. Maximal Lp regularity for parabolic difference equations. Math.
Nachr., 279(16):1787–1796, 2006. 687
Giga, Y., M. Gries, M. Hieber, A. Hussein, and T. Kashiwabara. Bounded
H ∞ -calculus for the hydrostatic Stokes operator on Lp -spaces and applica-
tions. Proc. Amer. Math. Soc., 145(9):3865–3876, 2017. 677
Giga, Y., M. Gries, M. Hieber, A. Hussein, and T. Kashiwabara. Analyticity
of solutions to the primitive equations. Math. Nachr., 293(2):284–304, 2020.
730
Girardi, M. and L. W. Weis. Operator-valued Fourier multiplier theorems on
Besov spaces. Math. Nachr., 251:34–51, 2003a. 409, 411, 679
Girardi, M. and L. W. Weis. Operator-valued Fourier multiplier theorems on
Lp (X) and geometry of Banach spaces. J. Funct. Anal., 204(2):320–354,
2003b. 289
Girardi, M. and L. W. Weis. Criteria for R-boundedness of operator families.
In Evolution equations, volume 234 of Lecture Notes in Pure and Appl.
Math., pages 203–221. Dekker, New York, 2003c. 412
Girardi, M. and L. W. Weis. Integral operators with operator-valued kernels.
J. Math. Anal. Appl., 290(1):190–212, 2004. 71
Girardi, M. and L. W. Weis. Operator-valued martingale transforms and
R-boundedness. Illinois J. Math., 49(2):487–516 (electronic), 2005. 209
Grafakos, L. Classical Fourier analysis, volume 249 of Graduate Texts in
Mathematics. Springer, New York, second edition, 2008. 71
Grafakos, L. Modern Fourier analysis, volume 250 of Graduate Texts in Math-
ematics. Springer, New York, second edition, 2009. 407
Graham, C.C. and K.E. Hare. Interpolation and Sidon sets for compact
groups. CMS Books in Mathematics/Ouvrages de Mathématiques de la
SMC. Springer, New York, 2013. 287
Grau de la Herrán, A. and T. P. Hytönen. Dyadic representation and bound-
edness of nonhomogeneous Calderón–Zygmund operators with mild kernel
regularity. Michigan Math. J., 67(4):757–786, 2018. 221, 738
Grisvard, P. Caractérisation de quelques espaces d’interpolation. Arch. Ra-
tional Mech. Anal., 25:40–63, 1967. 414
Grisvard, P. Équations différentielles abstraites. Ann. Sci. École Norm. Sup.
(4), 2:311–395, 1969. 671
Guerre-Delabrière, S. Some remarks on complex powers of (−∆) and UMD
spaces. Illinois J. Math., 35(3):401–407, 1991. 289
Guerre-Delabrière, S. Lp -regularity of the Cauchy problem and the geometry
of Banach spaces. Illinois J. Math., 39(4):556–566, 1995. 680
Haak, B. H., M. H. A. Haase, and P. C. Kunstmann. Perturbation, interpo-
lation, and maximal regularity. Adv. Differential Equations, 11(2):201–240,
2006. 561
798 REFERENCES

Haase, M. H. A. A functional calculus description of real interpolation spaces

for sectorial operators. Studia Math., 171(2):177–195, 2005. 565
Haase, M. H. A. The functional calculus for sectorial operators, volume 169
of Operator Theory: Advances and Applications. Birkhäuser Verlag, Basel,
2006. 408, 502, 507, 508, 563, 565
Haase, M. H. A. Functional calculus for groups and applications to evolution
equations. J. Evol. Equ., 7(3):529–554, 2007. 507, 508
Haase, M. H. A. A transference principle for general groups and functional
calculus on UMD spaces. Math. Ann., 345(2):245–265, 2009. 507
Haase, M. H. A. On the fundamental principles of unbounded functional
calculi. arXiv:2009.04733, 2020. 502
Haller-Dintelmann, R. and M. Hieber. H ∞ -calculus for products of non-
commuting operators. Math. Z., 251(1):85–100, 2005. 566
Haller-Dintelmann, R., H. Heck, and M. Hieber. Lp -Lq estimates for parabolic
systems in non-divergence form with VMO coefficients. J. London Math.
Soc. (2), 74(3):717–736, 2006. 676
Haller, R., H. Heck, and M. Hieber. Muckenhoupt weights and maximal Lp -
regularity. Arch. Math. (Basel), 81(4):422–430, 2003. 677
Han, Y.-S. and S. Hofmann. T1 theorems for Besov and Triebel–Lizorkin
spaces. Trans. Amer. Math. Soc., 337(2):839–853, 1993. 221, 223, 738
Han, Y.-S. and Y. Meyer. A characterization of Hilbert spaces and the vector-
valued Littlewood–Paley theorem. Methods Appl. Anal., 3(2):228–234, 1996.
415
Hänninen, T. S. and T. P. Hytönen. The A2 theorem and the local oscillation
decomposition for Banach space valued functions. J. Operator Theory, 72
(1):193–218, 2014. 70, 75, 76
Hänninen, T. S. and T. P. Hytönen. Operator-valued dyadic shifts and the
T (1) theorem. Monatsh. Math., pages 1–41, 2016. 70, 211, 218, 221
Hardy, G. H. and J. E. Littlewood. Some more theorems concerning Fourier
series and Fourier power series. Duke Math. J., 2(2):354–382, 1936. 673
Haroske, D. D. and L. Skrzypczak. Entropy and approximation numbers of
embeddings of function spaces with Muckenhoupt weights. I. Rev. Mat.
Complut., 21(1):135–177, 2008. 417
Heck, H. and M. Hieber. Maximal Lp -regularity for elliptic operators with
VMO-coefficients. J. Evol. Equ., 3(2):332–359, 2003. 676
Hedberg, L. I. and Y. Netrusov. An axiomatic approach to function spaces,
spectral synthesis, and Luzin approximation. Mem. Amer. Math. Soc., 188
(882):vi+97, 2007. 407
Henry, D. Geometric theory of semilinear parabolic equations, volume 840
of Lecture Notes in Mathematics. Springer-Verlag, Berlin-New York, 1981.
727
Hieber, M. Examples of pseudo-differential operators in Lp spaces with un-
bounded imaginary powers. Arch. Math. (Basel), 66(2):126–131, 1996. 503
 REFERENCES 799

Hieber, M. and S. Monniaux. Pseudo-differential operators and maximal reg-

ularity results for non-autonomous parabolic equations. Proc. Amer. Math.
Soc., 128(4):1047–1053, 2000a. 684
Hieber, M. and S. Monniaux. Heat-kernels and maximal Lp -Lq -estimates: the
non-autonomous case. J. Fourier Anal. Appl., 6(5):467–481, 2000b. 684
Hieber, M. and J. Prüss. Heat kernels and maximal Lp -Lq estimates for
parabolic evolution equations. Comm. Partial Differential Equations, 22
(9-10):1647–1669, 1997. 676
Hieber, M. and J. Prüss. On the bidomain problem with FitzHugh–Nagumo
transport. Arch. Math. (Basel), 111(3):313–327, 2018. 730
Hieber, M. and J. Prüss. Bounded H ∞ -calculus for a class of nonlocal op-
erators: the bidomain operator in the Lq -setting. Math. Ann., 378(3-4):
1095–1127, 2020. 677
Hieber, M., K. Kress, and C. Stinner. The Keller–Segel system on bounded
convex domains in critical spaces. Partial Differ. Equ. Appl., 2(3):Paper
No. 38, 14, 2021. 730
Hieber, M., A. Hussein, and M. Saal. Global strong well-posedness of the
stochastic bidomain equations with FitzHugh–Nagumo transport. SIAM J.
Math. Anal., 55(4):4140–4161, 2023. 730
Hille, E. and R. S. Phillips. Functional analysis and semi-groups. American
Mathematical Society Colloquium Publications, vol. 31. American Mathe-
matical Society, Providence, R.I., 1957. rev. ed. 502
Hinrichs, A. On the type constants with respect to systems of characters of a
compact abelian group. Studia Math., 118(3):231–243, 1996. 285
Hinrichs, A. Simulating the Walsh system with trigonometric polynomials
and Fourier type of operators. Arch. Math. (Basel), 77(4):323–327, 2001.
739
Hoffmann, M., N. J. Kalton, and T. Kucherenko. R-bounded approximating
sequences and applications to semigroups. J. Math. Anal. Appl., 294(2):
373–386, 2004. 682
Hofmann, S., M. Lacey, and A. McIntosh. The solution of the Kato problem
for divergence form elliptic operators with Gaussian heat kernel bounds.
Ann. of Math. (2), 156(2):623–631, 2002. 510
Honzı́k, P. An example of a singular integral and a weight. Int. Math. Res.
Not. IMRN, (9):7391–7398, 2023. 78
Hörmander, L. Estimates for translation invariant operators in Lp spaces.
Acta Math., 104:93–140, 1960. 71, 287, 417, 418
Hummel, F. B. Stochastic transmission and boundary value problems. PhD
thesis, University of Konstanz, 2019. 415
Hummel, F. B. and N. Lindemulder. Elliptic and parabolic boundary value
problems in weighted function spaces. Potential Anal., 57(4):601–669, 2022.
678
Hunt, R., B. Muckenhoupt, and R. Wheeden. Weighted norm inequalities for
the conjugate function and Hilbert transform. Trans. Amer. Math. Soc.,
176:227–251, 1973. 72, 81
800 REFERENCES

Hytönen, T. P. Fourier embeddings and Mihlin-type multiplier theorems.

Math. Nachr., 274/275:74–103, 2004. 288, 289, 409
Hytönen, T. P. Convolutions, multipliers and maximal regularity on vector-
valued Hardy spaces. J. Evol. Equ., 5(2):205–225, 2005. 673
Hytönen, T. P. An operator-valued T b theorem. J. Funct. Anal., 234(2):
420–463, 2006. 210, 211, 213, 217, 218
Hytönen, T. P. New thoughts on the vector-valued Mihlin–Hörmander mul-
tiplier theorem. Proc. Amer. Math. Soc., 138(7):2553–2560, 2010. 289
Hytönen, T. P. Pseudo-localisation of singular integrals in Lp . Rev. Mat.
Iberoam., 27(2):557–584, 2011. 215
Hytönen, T. P. Vector-valued singular integrals revisited—with random
dyadic cubes. Bull. Pol. Acad. Sci. Math., 60(3):269–283, 2012. 214, 218,
219
Hytönen, T. P. The sharp weighted bound for general Calderón–Zygmund
operators. Ann. of Math. (2), 175(3):1473–1506, 2012. 74, 77, 84, 214, 221,
222
Hytönen, T. P. The vector-valued nonhomogeneous T b theorem. Int. Math.
Res. Not. IMRN, (2):451–511, 2014. 211, 214, 218, 222
Hytönen, T. P. A quantitative Coulhon–Lamberton theorem. In Oper-
ator semigroups meet complex analysis, harmonic analysis and mathe-
matical physics, volume 250 of Oper. Theory Adv. Appl., pages 273–279.
Birkhäuser/Springer, Cham, 2015. 680, 741
Hytönen, T. P. Representation of singular integrals by dyadic operators, and
the A2 theorem. Expo. Math., 35(2):166–205, 2017. 74, 221
Hytönen, T. P. An operator-valued T (1) theorem for symmetric singular
integrals in UMD spaces. J. Funct. Anal., 280(7):Paper No. 108933, 17,
2021. 211, 218, 219
Hytönen, T. P. Some remarks on convex body domination. arXiv:2301.00617,
2023. 83, 735
Hytönen, T. P. and A. Kairema. Systems of dyadic cubes in a doubling metric
space. Colloq. Math., 126(1):1–33, 2012. 70
Hytönen, T. P. and M. T. Lacey. The Ap -A∞ inequality for general Calderón–
Zygmund operators. Indiana Univ. Math. J., 61(6):2041–2092, 2012. 74,
75
Hytönen, T. P. and S. Lappas. The dyadic representation theorem using
smooth wavelets with compact support. J. Fourier Anal. Appl., 28(4):
Paper No. 66, 24, 2022. 222
Hytönen, T. P. and K. Li. Weak and strong Ap -A∞ estimates for square
functions and related operators. Proc. Amer. Math. Soc., 146(6):2497–2507,
2018. 84
Hytönen, T. P. and H. Martikainen. Non-homogeneous T b theorem and ran-
dom dyadic cubes on metric measure spaces. J. Geom. Anal., 22(4):1071–
1107, 2012. 70
Hytönen, T. P. and A. McIntosh. Stability in p of the H ∞ -calculus of first-
order systems in Lp . In The AMSI-ANU Workshop on spectral theory and
 REFERENCES 801

harmonic analysis, volume 44 of Proc. Centre Math. Appl. Austral. Nat.

Univ., pages 167–181. Austral. Nat. Univ., Canberra, 2010. 512
Hytönen, T. P. and J. Merikoski. Vector-valued local approximation spaces.
J. Fourier Anal. Appl., 25(2):299–320, 2019. 414, 740
Hytönen, T. P. and A. Naor. Heat flow and quantitative differentiation. J.
Eur. Math. Soc. (JEMS), 21(11):3415–3466, 2019. 415
Hytönen, T. P. and C. Pérez. Sharp weighted bounds involving A∞ . Anal.
PDE, 6(4):777–818, 2013. 74
Hytönen, T. P. and A. Rosén. Causal sparse domination of Beurling maximal
regularity operators. J. Anal. Math., pages 1–28, 2023. 80
Hytönen, T. P. and O. Tapiola. Almost Lipschitz-continuous wavelets in met-
ric spaces via a new randomization of dyadic cubes. J. Approx. Theory,
185:12–30, 2014. 70
Hytönen, T. P. and A. V. Vähäkangas. The local non-homogeneous T b theo-
rem for vector-valued functions. Glasg. Math. J., 57(1):17–82, 2015. 218
Hytönen, T. P. and M. C. Veraar. R-boundedness of smooth operator-valued
functions. Integral Equations Operator Theory, 63(3):373–402, 2009. 412
Hytönen, T. P. and L. W. Weis. Singular integrals on Besov spaces. Math.
Nachr., 279(5-6):581–598, 2006a. 409
Hytönen, T. P. and L. W. Weis. A T 1 theorem for integral transformations
with operator-valued kernel. J. Reine Angew. Math., 599:155–200, 2006b.
211, 213, 218, 219
Hytönen, T. P. and L. W. Weis. Singular convolution integrals with operator-
valued kernel. Math. Z., 255(2):393–425, 2007. 219, 220
Hytönen, T. P., A. McIntosh, and P. Portal. Kato’s square root problem in
Banach spaces. J. Funct. Anal., 254(3):675–726, 2008. 511, 512
Hytönen, T. P., M. T. Lacey, H. Martikainen, T. Orponen, M. C. Reguera,
E. T. Sawyer, and I. Uriarte-Tuero. Weak and strong type estimates
for maximal truncations of Calderón–Zygmund operators on Ap weighted
spaces. J. Anal. Math., 118(1):177–220, 2012. 74, 84
Hytönen, T. P., M. T. Lacey, and C. Pérez. Sharp weighted bounds for the q-
variation of singular integrals. Bull. Lond. Math. Soc., 45(3):529–540, 2013.
75
Hytönen, T. P., C. Pérez, S. Treil, and A. Volberg. Sharp weighted estimates
for dyadic shifts and the A2 conjecture. J. Reine Angew. Math., 687:43–86,
2014. 74, 221
Hytönen, T. P., S. Li, and A. Naor. Quantitative affine approximation for
UMD targets. Discrete Anal., 2016. Paper No. 6, 37. 415
Hytönen, T. P., L. Roncal, and O. Tapiola. Quantitative weighted estimates
for rough homogeneous singular integrals. Israel J. Math., 218(1):133–164,
2017. 76, 77, 78
Hytönen, T. P., H. Martikainen, and E. Vuorinen. Multi-parameter estimates
via operator-valued shifts. Proc. Lond. Math. Soc. (3), 119(6):1560–1597,
2019a. 79, 218
802 REFERENCES

Hytönen, T. P., S. Petermichl, and A. Volberg. The sharp square function

estimate with matrix weight. Discrete Anal., (Paper No. 2, 8), 2019b. 85
Hytönen, T. P., K. Li, Martikainen H., and E. Vuorinen. Multiresolution
analysis and Zygmund dilations. 2022. arXiv:2203.15777. 222
Isralowitz, J. Sharp matrix weighted strong type inequalities for the dyadic
square function. Potential Anal., 53(4):1529–1540, 2020. 85
Isralowitz, J. Matrix weighted Triebel–Lizorkin bounds: a simple stopping
time proof. Proc. Amer. Math. Soc., 149(10):4145–4158, 2021. 416
Isralowitz, J., S. Pott, and I. P. Rivera-Rı́os. Sharp A1 weighted estimates for
vector-valued operators. J. Geom. Anal., 31(3):3085–3116, 2021. 83
Isralowitz, J., S. Pott, and S. Treil. Commutators in the two scalar and matrix
weighted setting. J. Lond. Math. Soc. (2), 106(1):1–26, 2022. 83
Ivanisvili, P. and A. Volberg. Martingale transform and square function:
some weak and restricted weak sharp weighted estimates. arXiv:1711.10578,
2018. 84
Iwaniec, T. and G. Martin. Quasiregular mappings in even dimensions. Acta
Math., 170(1):29–81, 1993. 72
Jacob, B., F. L. Schwenninger, and J. Wintermayr. A refinement of Baillon’s
theorem on maximal regularity. Stud. Math., 263:141–158, 2022. 681
James, R. C. Super-reflexive spaces with bases. Pacific J. Math., 41:409–419,
1972. 284
Jawerth, B. Some observations on Besov and Lizorkin–Triebel spaces. Math.
Scand., 40(1):94–104, 1977. 413
John, F. Extremum problems with inequalities as subsidiary conditions. In
Studies and Essays Presented to R. Courant on his 60th Birthday, January
8, 1948, pages 187–204. Interscience Publishers, New York, 1948. 81, 82
Journé, J.-L. Remarks on Kato’s square-root problem. volume 35, pages
299–321. 1991. Conference on Mathematical Analysis (El Escorial, 1989).
510
Kaiser, C. On an operator-valued T (1) theorem by Hytönen and Weis. In
CMA/AMSI Research symposium asymptotic geometric analysis, harmonic
analysis, and related Topics, volume 42 of Proc. Centre Math. Appl. Austral.
Nat. Univ., pages 66–80. Austral. Nat. Univ., Canberra, 2007. 223
Kaiser, C. Calderón–Zygmund operators with operator-valued kernel on ho-
mogeneous Besov spaces. Math. Nachr., 282(1):69–85, 2009. 223
Kaljabin, G. A. Characterization of functions from spaces of Besov–Lizorkin–
Triebel type. Dokl. Akad. Nauk SSSR, 236(5):1056–1059, 1977. 413
Kaljabin, G. A. Description of traces for anisotropic spaces of Triebel–Lizorkin
type. Trudy Mat. Inst. Steklov., 150:160–173, 322, 1979. 413
Kalton, N. J. Perturbations of the H ∞ -calculus. Collect. Math., 58(3):291–
325, 2007. 561
Kalton, N. J. and T. Kucherenko. Rademacher bounded families of operators
on L1 . Proc. Amer. Math. Soc., 136(1):263–272 (electronic), 2008. 682
Kalton, N. J. and T. Kucherenko. Operators with an absolute functional
calculus. Math. Ann., 346(2):259–306, 2010. 564, 565, 678, 679
 REFERENCES 803

Kalton, N. J. and G. Lancien. A solution to the problem of Lp -maximal

regularity. Math. Z., 235(3):559–568, 2000. 675, 681
Kalton, N. J. and G. Lancien. Lp -maximal regularity on Banach spaces with
a Schauder basis. Arch. Math. (Basel), 78(5):397–408, 2002. 681
Kalton, N. J. and P. Portal. Remarks on `1 and `∞ -maximal regularity for
power-bounded operators. J. Aust. Math. Soc., 84(3):345–365, 2008. 673,
675, 687
Kalton, N. J. and L. W. Weis. The H ∞ -calculus and sums of closed operators.
Math. Ann., 321(2):319–345, 2001. 563, 564, 566, 675
Kalton, N. J. and L. W. Weis. The H ∞ -calculus and square function estimates.
In Nigel J. Kalton Selecta, Vol. 1, pages 715–764. Springer-Verlag, 2016.
503, 507
Kalton, N. J., P. Kunstmann, and L. W. Weis. Perturbation and interpolation
theorems for the H ∞ -calculus with applications to differential operators.
Math. Ann., 336(4):747–801, 2006. 505, 561, 562, 563
Kalton, N. J., J. M. A. M. van Neerven, M. C. Veraar, and L. W. Weis. Em-
bedding vector-valued Besov spaces into spaces of γ-radonifying operators.
Math. Nachr., 281:238–252, 2008. 411, 740
Kalton, N. J., E. Lorist, and L. W. Weis. Euclidean structures and operator
theory in Banach spaces. Memoirs of the AMS, 288(1433), 2023. 503, 504,
505, 563, 680, 682
Kato, T. Note on fractional powers of linear operators. Proc. Japan Acad.,
36:94–96, 1960. 502
Kato, T. Fractional powers of dissipative operators. J. Math. Soc. Japan, 13:
246–274, 1961. 502, 509
Kato, T. Perturbation theory for linear operators. Classics in Mathematics.
Springer-Verlag, Berlin, 1995. Reprint of the 1980 edition. 561
Katz, N. H. Matrix valued paraproducts. J. Fourier Anal. Appl., 3:913–921,
1997. 212
Katz, N. H. and M. C. Pereyra. Haar multipliers, paraproducts, and weighted
inequalities. In Analysis of divergence (Orono, ME, 1997), Appl. Numer.
Harmon. Anal., pages 145–170. Birkhäuser Boston, Boston, MA, 1999. 70,
210
Kellerman, H. and M. Hieber. Integrated semigroups. J. Funct. Anal., 84(1):
160–180, 1989. 672, 762
Kemmochi, T. Discrete maximal regularity for abstract Cauchy problems.
Studia Math., 234(3):241–263, 2016. 687
Köhne, M., J. Prüss, and M. Wilke. On quasilinear parabolic evolution equa-
tions in weighted Lp -spaces. J. Evol. Equ., 10(2):443–463, 2010. 728
Köhne, M., Jürgen Saal, and L. Westermann. Optimal Sobolev regularity for
the Stokes equations on a 2D wedge domain. Math. Ann., 379(1-2):377–413,
2021. 566
Komatsu, H. Fractional powers of operators. Pacific J. Math., 19:285–346,
1966. 502, 503
804 REFERENCES

König, H. Eigenvalue distribution of compact operators, volume 16 of Operator

Theory: Advances and Applications. Birkhäuser Verlag, Basel, 1986. 407,
410
König, H. On the Fourier-coefficients of vector-valued functions. Math.
Nachr., 152:215–227, 1991. 286, 288, 411
Kovács, B., B. Li, and C. Lubich. A-stable time discretizations preserve max-
imal parabolic regularity. SIAM J. Numer. Anal., 54(6):3600–3624, 2016.
687
Krasnosel0 skiı̆, M. A. and P. E. Sobolevskiı̆. Fractional powers of operators
acting in Banach spaces. Dokl. Akad. Nauk SSSR, 129:499–502, 1959. 502
Krause, B., M. Lacey, and M. Wierdl. On convergence of oscillatory ergodic
Hilbert transforms. Indiana Univ. Math. J., 68(2):641–662, 2019. 78
Kreı̆n, S. G., Yu. Ī. Petunı̄n, and E. M. Semënov. Interpolation of linear op-
erators, volume 54 of Translations of Mathematical Monographs. American
Mathematical Society, Providence, R.I., 1982. 408
Kriegler, C. and L. W. Weis. Paley–Littlewood decomposition for sectorial
operators and interpolation spaces. Math. Nachr., 289(11-12):1488–1525,
2016. 408
Krugljak, N., L. Maligranda, and Lars E. Persson. On an elementary approach
to the fractional Hardy inequality. Proc. Amer. Math. Soc., 128(3):727–734,
2000. 414
Krylov, N. V. Lectures on elliptic and parabolic equations in Sobolev spaces,
volume 96 of Graduate Studies in Mathematics. American Mathematical
Society, Providence, RI, 2008. 678, 683
Krylov, N. V. On parabolic equations in one space dimension. Comm. Partial
Differential Equations, 41(4):644–664, 2016. 683
Krylov, N. V. Rubio de Francia extrapolation theorem and related topics in
the theory of elliptic and parabolic equations. A survey. Algebra i Analiz,
32(3):5–38, 2020. 683
Krylov, N. V. and E. Priola. Poisson stochastic process and basic Schauder
and Sobolev estimates in the theory of parabolic equations. Arch. Ration.
Mech. Anal., 225(3):1089–1126, 2017. 680, 741
Kucherenko, T. and L. W. Weis. Real interpolation of domains of sectorial
operators on Lp -spaces. J. Math. Anal. Appl., 310(1):278–285, 2005. 682
Kunstmann, P. C. On maximal regularity of type Lp -Lq under minimal as-
sumptions for elliptic non-divergence operators. J. Funct. Anal., 255(10):
2732–2759, 2008. 676, 742
Kunstmann, P. C. A new interpolation approach to spaces of Triebel–Lizorkin
type. Illinois J. Math., 59(1):1–19, 2015. 412, 416
Kunstmann, P. C. and A. Ullmann. Rs -sectorial operators and generalized
Triebel–Lizorkin spaces. J. Fourier Anal. Appl., 20(1):135–185, 2014. 408
Kunstmann, P. C. and L. W. Weis. Perturbation theorems for maximal Lp -
regularity. Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4), 30(2):415–435, 2001.
561, 675
 REFERENCES 805

Kunstmann, P. C. and L. W. Weis. Maximal Lp -regularity for parabolic

equations, Fourier multiplier theorems and H ∞ -functional calculus. In
Functional analytic methods for evolution equations, volume 1855 of Lec-
ture Notes in Math., pages 65–311. Springer, Berlin, 2004. 502, 671, 676
Kunstmann, P. C. and L. W. Weis. New criteria for the H ∞ -calculus and the
Stokes operator on bounded Lipschitz domains. J. Evol. Equ., 17:387–409,
2017. 562, 563, 677
Labbas, R. and B. Terreni. Somme d’opérateurs linéaires de type parabolique.
I. Boll. Un. Mat. Ital. B (7), 1(2):545–569, 1987. 566
Lacey, M. and C. Thiele. A proof of boundedness of the Carleson operator.
Math. Res. Lett., 7(4):361–370, 2000. 74
Lacey, M. T. An elementary proof of the A2 bound. Israel J. Math., 217(1):
181–195, 2017. 76
Lacey, M. T. Sparse bounds for spherical maximal functions. J. Anal. Math.,
139(2):613–635, 2019. 79
Lacey, M. T. and J. Scurry. Weighted weak type estimates for square func-
tions. arXiv:1211.4219, 2012. 84
Lacey, M. T. and S. Spencer. Sparse bounds for oscillatory and random
singular integrals. New York J. Math., 23:119–131, 2017. 78
Lacey, M. T., S. Petermichl, and M. C. Reguera. Sharp A2 inequality for Haar
shift operators. Math. Ann., 348(1):127–141, 2010. 73
Lacey, M. T., D. Mena, and M. C. Reguera. Sparse bounds for Bochner–Riesz
multipliers. J. Fourier Anal. Appl., 25(2):523–537, 2019. 78
Ladyženskaja, O. A., V. A. Solonnikov, and N. N. Ural0 ceva. Linear and
quasilinear equations of parabolic type. Translated from the Russian by
S. Smith. Translations of Mathematical Monographs, Vol. 23. American
Mathematical Society, Providence, R.I., 1968. 671
Lamberton, D. Équations d’évolution linéaires associées à des semi-groupes
de contractions dans les espaces Lp . J. Funct. Anal., 72(2):252–262, 1987.
674, 675, 676
Lancien, F., G. Lancien, and C. Le Merdy. A joint functional calculus for
sectorial operators with commuting resolvents. Proc. London Math. Soc.
(3), 77(2):387–414, 1998. 564
Lancien, G. and C. Le Merdy. A generalized H ∞ functional calculus for
operators on subspaces of Lp and application to maximal regularity. Illinois
J. Math., 42(3):470–480, 1998. 563, 564
Langer, M. and V. Maz0 ya. On Lp -contractivity of semigroups generated by
linear partial differential operators. J. Funct. Anal., 164(1):73–109, 1999.
676
Le Merdy, C. Counterexamples on Lp -maximal regularity. Math. Z., 230(1):
47–62, 1999. 682
Le Merdy, C. Two results about H ∞ functional calculus on analytic UMD
Banach spaces. J. Aust. Math. Soc., 74(3):351–378, 2003. 564
LeCrone, J. and G. Simonett. Continuous maximal regularity and analytic
semigroups. Discrete Contin. Dyn. Syst., (Dynamical systems, differential
806 REFERENCES

equations and applications. 8th AIMS Conference. Suppl. Vol. II):963–970,

2011. 674
LeCrone, J. and G. Simonett. On quasilinear parabolic equations and con-
tinuous maximal regularity. Evol. Equ. Control Theory, 9(1):61–86, 2020.
674, 728, 730
LeCrone, J., J. Prüss, and M. Wilke. On quasilinear parabolic evolution
equations in weighted Lp -spaces II. J. Evol. Equ., 14(3):509–533, 2014.
728, 730
Lemarié, P. G. Continuité sur les espaces de Besov des opérateurs définis par
des intégrales singulières. Ann. Inst. Fourier (Grenoble), 35(4):175–187,
1985. 222
Lerner, A. K. A pointwise estimate for the local sharp maximal function with
applications to singular integrals. Bull. Lond. Math. Soc., 42(5):843–856,
2010. 70
Lerner, A. K. Sharp weighted norm inequalities for Littlewood–Paley opera-
tors and singular integrals. Adv. Math., 226(5):3912–3926, 2011. 84
Lerner, A. K. On an estimate of Calderón–Zygmund operators by dyadic
positive operators. J. Anal. Math., 121:141–161, 2013a. 75
Lerner, A. K. A simple proof of the A2 conjecture. Int. Math. Res. Not.
IMRN, (14):3159–3170, 2013b. 75
Lerner, A. K. On pointwise estimates involving sparse operators. New York
J. Math., 22:341–349, 2016. 76
Lerner, A. K. and F. Nazarov. Intuitive dyadic calculus: the basics. Expo.
Math., 37(3):225–265, 2019. 76
Lerner, A. K. and S. Ombrosi. Some remarks on the pointwise sparse domi-
nation. J. Geom. Anal., 30(1):1011–1027, 2020. 76
Lerner, A. K., S. Ombrosi, and C. Pérez. A1 bounds for Calderón–Zygmund
operators related to a problem of Muckenhoupt and Wheeden. Math. Res.
Lett., 16(1):149–156, 2009. 84
Lerner, A. K., F. Nazarov, and S. Ombrosi. On the sharp upper bound related
to the weak Muckenhoupt–Wheeden conjecture. Anal. PDE, 13(6):1939–
1954, 2020. 84
Lerner, A. K., E. Lorist, and S. Ombrosi. Operator-free sparse domination.
Forum Math. Sigma, 10:Paper No. e15, 28, 2022. 76
Lindemulder, N. Difference norms for vector-valued Bessel potential spaces
with an application to pointwise multipliers. J. Funct. Anal., 272(4):1435–
1476, 2017. 415, 740
Lindemulder, N. Maximal regularity with weights for parabolic problems with
inhomogeneous boundary conditions. J. Evol. Equ., 20(1):59–108, 2020. 678
Lindemulder, N. An intersection representation for a class of anisotropic
vector-valued function spaces. J. Approx. Theory, 264:Paper No. 105519,
61, 2021. 407, 410, 416
Lindemulder, N. Pointwise multiplication by the characteristic function of
the half-space on anisotropic vector-valued function spaces. Indag. Math.
(N.S.), 33(4):885–904, 2022. 415
 REFERENCES 807

Lindemulder, N. and E. Lorist. A discrete framework for the interpolation of

Banach spaces. arXiv:2105.08373, 2021. 412, 416, 505, 782
Lindemulder, N. and E. Lorist. Stein interpolation for the real interpolation
method. Banach J. Math. Anal., 16(1):Paper No. 7, 18, 2022. 408
Lindemulder, N. and M. C. Veraar. The heat equation with rough boundary
conditions and holomorphic functional calculus. J. Differential Equations,
269(7):5832–5899, 2020. 415, 678
Lindemulder, N., M. Meyries, and M. C. Veraar. Complex interpolation with
Dirichlet boundary conditions on the half line. Math. Nachr., 291(16):2435–
2456, 2018. 414, 415, 416, 676, 740
Lindenstrauss, J. and L. Tzafriri. Classical Banach spaces. Vol. I: Sequence
spaces, volume 92 of Ergebnisse der Mathematik und ihrer Grenzgebiete.
Springer-Verlag, Berlin-New York, 1977. 681
Lindenstrauss, J. and L. Tzafriri. Classical Banach spaces, Vol. II: Function
spaces, volume 97 of Ergebnisse der Mathematik und ihrer Grenzgebiete.
Springer-Verlag, Berlin, 1979. 681
Lindenstrauss, J. and M. Zippin. Banach spaces with a unique unconditional
basis. J. Funct. Anal., 3:115–125, 1969. 681
Lions, J.-L. Équations différentielles opérationnelles et problèmes aux limites,
volume Band 111 of Die Grundlehren der mathematischen Wissenschaften.
Springer-Verlag, Berlin-Göttingen-Heidelberg, 1961. 684, 744
Lions, J.-L. Espaces d’interpolation et domaines de puissances fractionnaires
d’opérateurs. J. Math. Soc. Japan, 14:233–241, 1962. 509
Lions, J.-L. Quelques méthodes de résolution des problèmes aux limites non
linéaires. Dunod, Paris; Gauthier-Villars, Paris, 1969. 727, 744
Lions, J.-L. and J. Peetre. Sur une classe d’espaces d’interpolation. Inst.
Hautes Études Sci. Publ. Math., (19):5–68, 1964. 782
Lizama, C. lp -maximal regularity for fractional difference equations on UMD
spaces. Math. Nachr., 288(17-18):2079–2092, 2015. 687
Lorist, E. On pointwise `r -sparse domination in a space of homogeneous type.
J. Geom. Anal., 31(9):9366–9405, 2021. 76
Lorist, E. and Z. Nieraeth. Vector-valued extensions of operators through
multilinear limited range extrapolation. J. Fourier Anal. Appl., 25(5):2608–
2634, 2019. 735
Lorist, E. and Z. Nieraeth. Sparse domination implies vector-valued sparse
domination. Math. Z., 301(1):1107–1141, 2022. 735
Lunardi, A. Analytic semigroups and optimal regularity in parabolic problems.
Progress in Nonlinear Differential Equations and their Applications, 16.
Birkhäuser Verlag, Basel, 1995. 671, 674, 679, 682, 727, 728, 742, 762, 782
Lunardi, A. Interpolation theory. Appunti. Scuola Normale Superiore di Pisa
(Nuova Serie). Edizioni della Normale, Pisa, second edition, 2009. 762, 782
Luque, T., C. Pérez, and E. Rela. Optimal exponents in weighted estimates
without examples. Math. Res. Lett., 22(1):183–201, 2015. 76
808 REFERENCES

Maas, J. and J. M. A. M. van Neerven. Boundedness of Riesz transforms

for elliptic operators on abstract Wiener spaces. J. Funct. Anal., 257(8):
2410–2475, 2009. 512
Maier, S. and J. Saal. Stokes and Navier–Stokes equations with perfect slip
on wedge type domains. Discr. Contin. Dyn. Syst., Ser. S, 7(5):1045–1063,
2014. 566
Martikainen, H. Vector-valued non-homogeneous T b theorem on metric mea-
sure spaces. Rev. Mat. Iberoam., 28(4):961–998, 2012a. 218
Martikainen, H. Representation of bi-parameter singular integrals by dyadic
operators. Adv. Math., 229(3):1734–1761, 2012b. 79
Martı́nez Carracedo, C. and M. Sanz Alix. The theory of fractional powers
of operators, volume 187 of North-Holland Mathematics Studies. North-
Holland Publishing Co., Amsterdam, 2001. 502
Martı́nez, T., J. L. Torrea, and Q. Xu. Vector-valued Littlewood–Paley–Stein
theory for semigroups. Adv. Math., 203(2):430–475, 2006. 415
Matioc, B.-V. and C. Walker. On the principle of linearized stability in inter-
polation spaces for quasilinear evolution equations. Monatsh. Math., 191
(3):615–634, 2020. 728
Mattila, P., M. S. Melnikov, and J. Verdera. The Cauchy integral, analytic
capacity, and uniform rectifiability. Ann. of Math. (2), 144(1):127–136,
1996. 218
Maz’ya, V. Sobolev spaces with applications to elliptic partial differential
equations, volume 342 of Grundlehren der Mathematischen Wissenschaften.
Springer, Heidelberg, augmented edition, 2011. 407
Mazzone, G. On the free rotations of rigid bodies with a liquid-filled gap. J.
Math. Anal. Appl., 496(2):Paper No. 124826, 37, 2021. 730
Mazzone, G., J. Prüss, and G. Simonett. On the motion of a fluid-filled rigid
body with Navier boundary conditions. SIAM J. Math. Anal., 51(3):1582–
1606, 2019a. 730
Mazzone, G., J. Prüss, and G. Simonett. A maximal regularity approach to
the study of motion of a rigid body with a fluid-filled cavity. J. Math. Fluid
Mech., 21(3):Paper No. 44, 20, 2019b. 730
McArthur, C. W. Developments in Schauder basis theory. Bull. Amer. Math.
Soc., 78:877–908, 1972. 681
McIntosh, A. On the comparability of A1/2 and A∗1/2 . Proc. Amer. Math.
Soc., 32:430–434, 1972. 509
McIntosh, A. On representing closed accretive sesquilinear forms as
(a1/2 u, a∗1/2 v). In Nonlinear partial differential equations and their ap-
plications, Coll. de France Semin., Vol. III, Res. Notes Math. 70, 252-267
(1982). 1982. 509, 510
McIntosh, A. Square roots of elliptic operators. J. Funct. Anal., 61(3):307–
327, 1985. 510
McIntosh, A. Operators which have an H∞ functional calculus. In Minicon-
ference on operator theory and partial differential equations (North Ryde,
 REFERENCES 809

1986), volume 14 of Proc. Centre Math. Anal. Austral. Nat. Univ., pages
210–231. Austral. Nat. Univ., Canberra, 1986. 502, 503
McIntosh, A. and A. Yagi. Operators of type ω without a bounded H∞
functional calculus. In Miniconference on operators in analysis (Sydney,
1989), volume 24 of Proc. Centre Math. Anal. Austral. Nat. Univ., pages
159–172. Austral. Nat. Univ., Canberra, 1990. 503, 561
Mei, T. Notes on matrix valued paraproducts. Indiana Univ. Math. J., 55(2):
747–760, 2006. 211, 212
Mei, T. An extrapolation of operator-valued dyadic paraproducts. J. Lond.
Math. Soc. (2), 81(3):650–662, 2010. 211
Mei, T. and J. Parcet. Pseudo-localization of singular integrals and noncom-
mutative Littlewood–Paley inequalities. Int. Math. Res. Not. IMRN, (8):
1433–1487, 2009. 215
Melnikov, M. S. and J. Verdera. A geometric proof of the L2 boundedness of
the Cauchy integral on Lipschitz graphs. Int. Math. Res. Not. IMRN, (7):
325–331, 1995. 218
Meyer, Y. La minimalité de le espace de Besov Ḃ00,1 et la continuité de
opérateurs définis par des intégrales singulières, volume 4 of Monografı́as
de Matemáticas. Univ. Autónoma de Madrid, 1986. 221, 738
Meyries, M. and R. Schnaubelt. Maximal regularity with temporal weights
for parabolic problems with inhomogeneous boundary conditions. Math.
Nachr., 285(8-9):1032–1051, 2012a. 678
Meyries, M. and R. Schnaubelt. Interpolation, embeddings and traces of
anisotropic fractional Sobolev spaces with temporal weights. J. Funct.
Anal., 262(3):1200–1229, 2012b. 675, 782
Meyries, M. and M. C. Veraar. Sharp embedding results for spaces of smooth
functions with power weights. Studia Math., 208(3):257–293, 2012. 413,
414, 416, 417
Meyries, M. and M. C. Veraar. Traces and embeddings of anisotropic function
spaces. Math. Ann., 360(3-4):571–606, 2014a. 413
Meyries, M. and M. C. Veraar. Characterization of a class of embeddings for
function spaces with Muckenhoupt weights. Arch. Math. (Basel), 103(5):
435–449, 2014b. 416, 417
Meyries, M. and M. C. Veraar. Pointwise multiplication on vector-valued
function spaces with power weights. J. Fourier Anal. Appl., 21(1):95–136,
2015. 288, 414, 415, 416, 740
Mielke, A. Über maximale Lp -Regularität für Differentialgleichungen in
Banach- und Hilbert-Räumen. Math. Ann., 277(1):121–133, 1987. 678
Mihlin, S. G. On the multipliers of Fourier integrals. Dokl. Akad. Nauk SSSR
(N.S.), 109:701–703, 1956. 287
Mihlin, S. G. Fourier integrals and multiple singular integrals. Vestn. Leningr.
Univ. (Ser. Mat. Mekh. Astron.), 12(7):143–155, 1957. 287
Monniaux, S. A perturbation result for bounded imaginary powers. Arch.
Math. (Basel), 68(5):407–417, 1997. 502
810 REFERENCES

Monniaux, S. A new approach to the Dore–Venni theorem. Math. Nachr.,

204:163–183, 1999. 503, 507, 508
Monniaux, S. and J. Prüss. A theorem of the Dore–Venni type for noncom-
muting operators. Trans. Amer. Math. Soc., 349(12):4787–4814, 1997. 566
Muller, P. A. and I. P. Rivera-Rı́os. Quantitative matrix weighted estimates
for certain singular integral operators. J. Math. Anal. Appl., 509(1):Paper
No. 125939, 38, 2022. 83
Nau, T. and J. Saal. H ∞ -calculus for cylindrical boundary value problems.
Adv. Differential Equations, 17(7-8):767–800, 2012. 566
Nazarov, F. and A. Volberg. A simple sharp weighted estimate of the dyadic
shifts on metric spaces with geometric doubling. Int. Math. Res. Not.
IMRN, (16):3771–3789, 2013. 74
Nazarov, F., S. Treil, and A. Volberg. Cauchy integral and Calderón–Zygmund
operators on nonhomogeneous spaces. Internat. Math. Res. Notices, (15):
703–726, 1997a. 213, 214
Nazarov, F., S. Treil, and A. Volberg. Counterexample to the infinite-
dimensional Carleson embedding theorem. C. R. Acad. Sci. Paris Sér.
I Math., 325(4):383–388, 1997b. 212
Nazarov, F., G. Pisier, S. Treil, and A. Volberg. Sharp estimates in vector
Carleson imbedding theorem and for vector paraproducts. J. Reine Angew.
Math., 542:147–171, 2002a. 212
Nazarov, F., S. Treil, and A. Volberg. Accretive system T b-theorems on non-
homogeneous spaces. Duke Math. J., 113(2):259–312, 2002b. 214
Nazarov, F., S. Treil, and A. Volberg. The T b-theorem on non-homogeneous
spaces. Acta Math., 190(2):151–239, 2003. 214, 217, 222
Nazarov, F., S. Treil, and A. Volberg. Two weight inequalities for individual
Haar multipliers and other well localized operators. Math. Res. Lett., 15
(3):583–597, 2008. 73
Nazarov, F., A. Reznikov, and A. Volberg. The proof of A2 conjecture in a
geometrically doubling metric space. Indiana Univ. Math. J., 62(5):1503–
1533, 2013. 70, 74
Nazarov, F., S. Petermichl, S. Treil, and A. Volberg. Convex body domination
and weighted estimates with matrix weights. Adv. Math., 318:279–306, 2017.
82
Nazarov, F. L. and S. Treil. The hunt for a Bellman function: applications to
estimates for singular integral operators and to other classical problems of
harmonic analysis. Algebra i Analiz, 8(5):32–162, 1996. 81, 416
Neerven, J. M. A. M. van. The asymptotic behaviour of semigroups of lin-
ear operators, volume 88 of Operator Theory: Advances and Applications.
Birkhäuser Verlag, Basel, 1996. 762
Neerven, J. M. A. M. van. Functional analysis, volume 201 of Cambridge
Studies in Advanced Mathematics. Cambridge University Press, Cambridge,
2022. 561
 REFERENCES 811

Neerven, J. M. A. M. van and P. Portal. The Weyl calculus for group gener-
ators satisfying the canonical commutation relations. J. Operator Theory,
83(2):253–298, 2020. 567
Neerven, J. M. A. M. van, M. C. Veraar, and L. W. Weis. Conditions for
stochastic integrability in UMD Banach spaces. In Banach spaces and their
applications in analysis, pages 125–146. Walter de Gruyter, Berlin, 2007.
741
Neerven, J. M. A. M. van, M. C. Veraar, and L. W. Weis. Maximal γ-
regularity. J. Evol. Equ., 15(2):361–402, 2015a. 684, 687
Neerven, J. M. A. M. van, M. C. Veraar, and L. W. Weis. Stochastic integra-
tion in Banach spaces – a survey. In Stochastic analysis: A series of lectures,
volume 68 of Progress in Probability. Birkhäuser Verlag, 2015b. 564
Neerven, J. M. A. M. van, P. Portal, and H. Sharma. Spectral multiplier
theorems for abstract harmonic oscillators on UMD lattices. C. R. Math.
Acad. Sci. Paris, 361:835–846, 2023. 567
Nollau, V. Über den Logarithmus abgeschlossener Operatoren in Banachschen
Räumen. Acta Sci. Math. (Szeged), 30:161–174, 1969. 507
Oberlin, R. Sparse bounds for a prototypical singular Radon transform.
Canad. Math. Bull., 62(2):405–415, 2019. 79
Ogawa, T. and S. Shimizu. End-point maximal L1 -regularity for the Cauchy
problem to a parabolic equation with variable coefficients. Math. Ann., 365
(1-2):661–705, 2016. 679
Ogawa, T. and S. Shimizu. Maximal L1 -regularity of the heat equation and
application to a free boundary problem of the Navier–Stokes equations near
the half-space. J. Elliptic Parabol. Equ., 7(2):509–535, 2021. 679
Ondreját, M. and M. C. Veraar. On temporal regularity of stochastic con-
volutions in 2-smooth Banach spaces. Ann. Inst. Henri Poincaré Probab.
Stat., 56(3):1792–1808, 2020. 679
Orponen, T. On the Haar shift representations of Calderón–Zygmund opera-
tors. Proc. Amer. Math. Soc., 141(8):2693–2698, 2013. 77
Osȩkowski, A. and I. Yaroslavtsev. The Hilbert transform and orthogonal
martingales in Banach spaces. Int. Math. Res. Not. IMRN, (15):11670–
11730, 2021. 291
Ou, Y. Multi-parameter singular integral operators and representation theo-
rem. Rev. Mat. Iberoam., 33(1):325–350, 2017. 79
Parcet, J. Pseudo-localization of singular integrals and noncommutative
Calderón–Zygmund theory. J. Funct. Anal., 256(2):509–593, 2009. 215
Pazy, A. Semigroups of linear operators and applications to partial differential
equations, volume 44 of Applied Mathematical Sciences. Springer-Verlag,
New York, 1983. 727, 742, 762
Peetre, J. Sur les espaces de Besov. C. R. Acad. Sci. Paris Sér. A-B, 264:
A281–A283, 1967. 408
Peetre, J. On spaces of Triebel–Lizorkin type. Ark. Mat., 13:123–130, 1975.
407
812 REFERENCES

Peetre, J. New thoughts on Besov spaces. Mathematics Department, Duke

University, Durham, N.C., 1976. Duke University Mathematics Series, No.
1. 407
Pérez, C., S. Treil, and A. Volberg. On A2 conjecture and corona decomposi-
tion of weights. arXiv:1006.2630, 2010. 74
Petermichl, S. Dyadic shifts and a logarithmic estimate for Hankel operators
with matrix symbol. C. R. Acad. Sci. Paris Sér. I Math., 330(6):455–460,
2000. 73
Petermichl, S. The sharp bound for the Hilbert transform on weighted
Lebesgue spaces in terms of the classical Ap characteristic. Amer. J. Math.,
129(5):1355–1375, 2007. 73
Petermichl, S. The sharp weighted bound for the Riesz transforms. Proc.
Amer. Math. Soc., 136(4):1237–1249, 2008. 73
Petermichl, S. and A. Volberg. Heating of the Ahlfors–Beurling operator:
weakly quasiregular maps on the plane are quasiregular. Duke Math. J.,
112(2):281–305, 2002. 73
Phillips, R. S. On the generation of semi-groups of linear operators. Pacific
J. Math., 2:343–369, 1952. 502
Piasecki, T., Y. Shibata, and E. Zatorska. On the maximal Lp -Lq regularity
of solutions to a general linear parabolic system. J. Differential Equations,
268(7):3332–3369, 2020. 676
Pierre, M. and D. Schmitt. Blowup in reaction-diffusion systems with dissi-
pation of mass. SIAM J. Math. Anal., 28(2):259–269, 1997. 683, 744
Pietsch, A. History of Banach spaces and linear operators. Birkhäuser Boston
Inc., Boston, MA, 2007. 739
Pietsch, A. and J. Wenzel. Orthonormal systems and Banach space geometry,
volume 70 of Encyclopedia of Mathematics and its Applications. Cambridge
University Press, Cambridge, 1998. 285, 287
Pisier, G. Un exemple concernant la super-réflexivité. In Séminaire Maurey-
Schwartz 1974–1975: Espaces Lp , applications radonifiantes et géométrie
des espaces de Banach, Annexe No. 2, page 12. Centre Math. École Poly-
tech., Paris, 1975. 284
Pisier, G. Martingales in Banach spaces, volume 155 of Cambridge Studies in
Advanced Mathematics. Cambridge University Press, 2016. 70
Poblete, V. Maximal regularity of second-order equations with delay. J.
Differential Equations, 246(1):261–276, 2009. 688
Portal, P. Discrete time analytic semigroups and the geometry of Banach
spaces. Semigroup Forum, 67(1):125–144, 2003. 687
Portal, P. Maximal regularity of evolution equations on discrete time scales.
J. Math. Anal. Appl., 304(1):1–12, 2005. 687
Portal, P. and Ž. Štrkalj. Pseudodifferential operators on Bochner spaces and
an application. Math. Z., 253(4):805–819, 2006. 684
Pott, S. and A. Stoica. Linear bounds for Calderón–Zygmund operators with
even kernel on UMD spaces. J. Funct. Anal., 266(5):3303–3319, 2014. 214,
218, 290, 737
 REFERENCES 813

Prats, M. Measuring Triebel–Lizorkin fractional smoothness on domains in

terms of first-order differences. J. Lond. Math. Soc. (2), 100(2):692–716,
2019. 416
Prüss, J. Maximal regularity for evolution equations in Lp -spaces. Conf.
Semin. Mat. Univ. Bari, 285:1–39 (2003), 2002. 727, 728
Prüss, J. H ∞ -calculus for generalized Stokes operators. J. Evol. Equ., 18(3):
1543–1574, 2018. 677
Prüss, J. Vector-valued Fourier multipliers in Lp -spaces with power weights.
Studia Math., 247(2):155–173, 2019. 688
Prüss, J. and R. Schnaubelt. Solvability and maximal regularity of parabolic
evolution equations with coefficients continuous in time. J. Math. Anal.
Appl., 256(2):405–430, 2001. 682, 743
Prüss, J. and G. Simonett. Maximal regularity for evolution equations in
weighted Lp -spaces. Arch. Math. (Basel), 82(5):415–431, 2004. 673, 675,
728
Prüss, J. and G. Simonett. Operator-valued symbols for elliptic and parabolic
problems on wedges. In Partial differential equations and functional anal-
ysis, volume 168 of Oper. Theory Adv. Appl., pages 189–208. Birkhäuser,
Basel, 2006. 566
Prüss, J. and G. Simonett. H ∞ -calculus for the sum of non-commuting op-
erators. Trans. Amer. Math. Soc., 359(8):3549–3565, 2007. 566
Prüss, J. and G. Simonett. Moving interfaces and quasilinear parabolic evo-
lution equations, volume 105 of Monographs in Mathematics. Birkhäuser,
Basel, 2016. 502, 671, 672, 679, 727, 728, 729, 730
Prüss, J. and H. Sohr. On operators with bounded imaginary powers in
Banach spaces. Math. Z., 203(3):429–452, 1990. 502, 503, 508, 675
Prüss, J. and M. Wilke. Addendum to the paper “On quasilinear parabolic
evolution equations in weighted Lp -spaces II”. J. Evol. Equ., 17(4):1381–
1388, 2017. 728
Prüss, J. and M. Wilke. On critical spaces for the Navier–Stokes equations.
J. Math. Fluid Mech., 20(2):733–755, 2018. 728, 730
Prüss, J., G. Simonett, and R. Zacher. On convergence of solutions to equilib-
ria for quasilinear parabolic problems. J. Differential Equations, 246(10):
3902–3931, 2009. 728
Prüss, J., G. Simonett, and M. Wilke. Critical spaces for quasilinear parabolic
evolution equations and applications. J. Differential Equations, 264(3):
2028–2074, 2018. 728, 729, 730
Prüss, J., G. Simonett, and M. Wilke. On the Navier–Stokes equations on
surfaces. J. Evol. Equ., 21(3):3153–3179, 2021. 730
Putnam, C. R. Commutation properties of Hilbert space operators and re-
lated topics. Ergebnisse der Mathematik und ihrer Grenzgebiete [Results
in Mathematics and Related Areas], Band 36. Springer-Verlag New York,
Inc., New York, 1967. 566
Qiu, Y. On the UMD constants for a class of iterated Lp (Lq ) spaces. J. Funct.
Anal., 263(8):2409–2429, 2012. 284
814 REFERENCES

Reguera, M. C. On Muckenhoupt–Wheeden conjecture. Adv. Math., 227(4):

1436–1450, 2011. 84
Reguera, M. C. and C. Thiele. The Hilbert transform does not map L1 (M w)
to L1,∞ (w). Math. Res. Lett., 19(1):1–7, 2012. 84
Roidos, N. Closedness and invertibility for the sum of two closed operators.
Adv. Oper. Theory, 3(3):582–605, 2018. 566
Roudenko, S. Matrix-weighted Besov spaces. Trans. Amer. Math. Soc., 355
(1):273–314, 2003. 223, 416
Roudenko, S. Duality of matrix-weighted Besov spaces. Studia Math., 160(2):
129–156, 2004. 416
Rozendaal, J. Functional calculus for C0 -groups using type and cotype. Q. J.
Math., 70(1):17–47, 2019. 418
Rozendaal, J. Operator-valued (Lp , Lq ) Fourier multipliers and stability the-
ory for evolution equations. Indag. Math. (N.S.), 34(1):1–36, 2023. 417
Rozendaal, J. and M. C. Veraar. Fourier multiplier theorems on Besov spaces
under type and cotype conditions. Banach J. Math. Anal., 11(4):713–743,
2017. 417, 418
Rozendaal, J. and M. C. Veraar. Fourier multiplier theorems involving type
and cotype. J. Fourier Anal. Appl., 24(2):583–619, 2018a. 417, 418, 741
Rozendaal, J. and M. C. Veraar. Sharp growth rates for semigroups using
resolvent bounds. J. Evol. Equ., 18(4):1721–1744, 2018b. 417
Rozendaal, J. and M. C. Veraar. Stability theory for semigroups using (Lp , Lq )
Fourier multipliers. J. Funct. Anal., 275(10):2845–2894, 2018c. 417
Rubio de Francia, J. L. Martingale and integral transforms of Banach space
valued functions. In Probability and Banach spaces (Zaragoza, 1985), vol-
ume 1221 of Lecture Notes in Math., pages 195–222. Springer, Berlin, 1986.
217, 735
Rubio de Francia, J. L., F. J. Ruiz, and J. L. Torrea. Calderón–Zygmund
theory for operator-valued kernels. Adv. in Math., 62(1):7–48, 1986. 71,
673
Rudin, W. Functional analysis. International Series in Pure and Applied
Mathematics. McGraw-Hill Inc., New York, second edition, 1991. 409
Runst, Th. and W. Sickel. Sobolev spaces of fractional order, Nemytskij oper-
ators, and nonlinear partial differential equations, volume 3 of de Gruyter
Series in Nonlinear Analysis and Applications. Walter de Gruyter & Co.,
Berlin, 1996. 407, 414
Ryan, R. A. Introduction to tensor products of Banach spaces. Springer
Monographs in Mathematics. Springer-Verlag London, Ltd., London, 2002.
211
Rychkov, V. S. On restrictions and extensions of the Besov and Triebel–
Lizorkin spaces with respect to Lipschitz domains. J. London Math. Soc.
(2), 60(1):237–257, 1999. 416
Rydhe, E. Vectorial Hankel operators, Carleson embeddings, and notions of
BMOA. Geom. Funct. Anal., 27(2):427–451, 2017. 212
 REFERENCES 815

Sawano, Y. Theory of Besov spaces, volume 56 of Developments in Mathe-

matics. Springer, Singapore, 2018. 407
Schmeisser, H.-J. Vector-valued Sobolev and Besov spaces. In Seminar
analysis of the Karl-Weierstrass-Institute of Mathematics 1985/86 (Berlin,
1985/86), volume 96 of Teubner-Texte Math., pages 4–44. Teubner, Leipzig,
1987. 407
Schmeisser, H.-J. and W. Sickel. Traces, Gagliardo–Nirenberg inequalities and
Sobolev type embeddings for vector-valued function spaces, Jenaer Schriften
Math. Inf., Math/Inf/24/01, Jena, 2001. 407, 409, 410, 412, 413, 414
Schmeisser, H.-J. and W. Sickel. Vector-valued Sobolev spaces and Gagliardo–
Nirenberg inequalities. In Nonlinear elliptic and parabolic problems, vol-
ume 64 of Progr. Nonlinear Differential Equations Appl., pages 463–472.
Birkhäuser, Basel, 2005. 407, 413, 414
Schwartz, J. A remark on inequalities of Calderón–Zygmund type for vector-
valued functions. Comm. Pure Appl. Math., 14:785–799, 1961. 71
Seeley, R. Norms and domains of the complex powers AzB . Amer. J. Math.,
93:299–309, 1971. 503, 741
Seeley, R. Interpolation in Lp with boundary conditions. Studia Math., 44:
47–60, 1972. 414, 415
Shamir, E. Une propriété des espaces H s, p . C. R. Acad. Sci. Paris, 255:
448–449, 1962. 415
Shao, Y. and G. Simonett. Continuous maximal regularity on uniformly reg-
ular Riemannian manifolds. J. Evol. Equ., 14(1):211–248, 2014. 674
Shen, Z. Resolvent estimates in Lp for the Stokes operator in Lipschitz do-
mains. Arch. Ration. Mech. Anal., 205(2):395–424, 2012. 677
Shibata, Y. R boundedness, maximal regularity and free boundary problems
for the Navier Stokes equations. In Mathematical analysis of the Navier–
Stokes equations, volume 2254 of Lecture Notes in Math., pages 193–462.
Springer, Cham, 2020. 677
Shibata, Y. and S. Shimizu. On the Lp -Lq maximal regularity of the Neumann
problem for the Stokes equations in a bounded domain. J. Reine Angew.
Math., 615:157–209, 2008. 677
Simon, L. de. Un’applicazione della teoria degli integrali singolari allo studio
delle equazioni differenziali lineari astratte del primo ordine. Rend. Sem.
Mat. Univ. Padova, 34:205–223, 1964. 671, 674
Simonett, G. and J. Prüss. On the microscopic bidomain problem with
FitzHugh–Nagumo ionic transport. J. Elliptic Parabol. Equ., 5(1):25–45,
2019. 730
Simonett, G. and M. Wilke. H ∞ -calculus for the surface Stokes operator and
applications. J. Math. Fluid Mech., 24(4):Paper No. 109, 23, 2022a. 677
Simonett, G. and M. Wilke. H ∞ -calculus for the surface Stokes operator and
applications. J. Math. Fluid Mech., 24(4):Paper No. 109, 23, 2022b. 730
Singer, I. Bases in Banach spaces. I, volume 154 of Grundlehren der Mathe-
matischen Wissenschaften. Springer-Verlag, New York-Berlin, 1970. 681
816 REFERENCES

Sobczyk, A. Projection of the space (m) on its subspace (c0 ). Bull. Amer.
Math. Soc., 47:938–947, 1941. 681
Sobolevskiı̆, P. E. Coerciveness inequalities for abstract parabolic equations.
Dokl. Akad. Nauk SSSR, 157:52–55, 1964. Translation in Soviet Math. Dokl.
5 (1964), 894–897. 671, 672, 673
Stein, E. M. Interpolation of linear operators. Trans. Amer. Math. Soc., 83:
482–492, 1956. 412
Stein, E. M. Note on singular integrals. Proc. Amer. Math. Soc., 8:250–254,
1957. 673
Stein, E. M. The development of square functions in the work of A. Zygmund.
Bull. Amer. Math. Soc. (N.S.), 7(2):359–376, 1982. 210
Stein, E. M. Harmonic analysis: real-variable methods, orthogonality, and
oscillatory integrals, volume 43 of Princeton Mathematical Series. Princeton
University Press, Princeton, NJ, 1993. 288
Stein, E. M. and G. Weiss. Interpolation of operators with change of measures.
Trans. Amer. Math. Soc., 87:159–172, 1958. 408
Strichartz, R. S. Multipliers on fractional Sobolev spaces. J. Math. Mech.,
16:1031–1060, 1967. 415
Štrkalj, Ž. R-Beschränktheit, Summensätze abgeschlossener Operatoren und
operatorwertige Pseudodifferentialoperatoren. PhD thesis, Universität Karl-
sruhe, 2000. 566, 684
Suárez, J. and L. W. Weis. Interpolation of Banach spaces by the γ-method. In
Methods in Banach space theory, volume 337 of London Math. Soc. Lecture
Note Ser., pages 293–306. Cambridge Univ. Press, Cambridge, 2006. 505
Suárez, J. and L. W. Weis. Addendum to “Interpolation of Banach spaces by
the γ-method”. Extracta Math., 24(3):265–269, 2009. 505
Tanabe, H. Equations of evolution, volume 6 of Monographs and Studies
in Mathematics. Pitman (Advanced Publishing Program), Boston, Mass.,
1979. 682, 684, 727, 742
Tanabe, H. Functional analytic methods for partial differential equations,
volume 204 of Monographs and Textbooks in Pure and Applied Mathematics.
Marcel Dekker Inc., New York, 1997. 684
Taylor, M. E. Pseudo differential operators, volume Vol. 416 of Lecture Notes
in Mathematics. Springer-Verlag, Berlin-New York, 1974. 408
Taylor, M. E. Partial differential equations I. Basic theory, volume 115 of
Applied Mathematical Sciences. Springer, New York, second edition, 2011a.
408
Taylor, M. E. Partial differential equations II. Qualitative studies of linear
equations, volume 116 of Applied Mathematical Sciences. Springer, New
York, second edition, 2011b. 408
Taylor, M. E. Partial differential equations III. Nonlinear equations, volume
117 of Applied Mathematical Sciences. Springer, New York, second edition,
2011c. 408
Titchmarsh, E. C. Introduction to the theory of Fourier integrals. Chelsea
Publishing Co., New York, third edition, 1986. 503
 REFERENCES 817

Tolksdorf, P. On the Lp -theory of the Navier–Stokes equations on three-

dimensional bounded Lipschitz domains. Math. Ann., 371(1-2):445–460,
2018. 677
Tolksdorf, P. The Stokes resolvent problem: optimal pressure estimates and
remarks on resolvent estimates in convex domains. Calc. Var. Partial Dif-
ferential Equations, 59(5):Paper No. 154, 40, 2020. 677
Tolksdorf, P. and K. Watanabe. The Navier–Stokes equations in exterior
Lipschitz domains: Lp -theory. J. Differential Equations, 269(7):5765–5801,
2020. 677
Tolsa, X. L2 -boundedness of the Cauchy integral operator for continuous
measures. Duke Math. J., 98(2):269–304, 1999. 218
Torres, R. H. Boundedness results for operators with singular kernels on
distribution spaces. Mem. Amer. Math. Soc., 90(442):viii+172, 1991. 223
Travis, C. C. Differentiability of weak solutions to an abstract inhomogeneous
differential equation. Proc. Amer. Math. Soc., 82(3):425–430, 1981. 674
Treil, S. Sharp A2 estimates of Haar shifts via Bellman function. In Recent
trends in analysis, volume 16 of Theta Ser. Adv. Math., pages 187–208.
Theta, Bucharest, 2013. 74
Treil, S. and A. Volberg. Wavelets and the angle between past and future. J.
Funct. Anal., 143(2):269–308, 1997. 81, 83
Triebel, H. Interpolation theory, function spaces, differential operators, vol-
ume 18 of North-Holland Mathematical Library. North-Holland Publishing
Co., Amsterdam, 1978. 407, 408, 409, 410, 412, 503, 782
Triebel, H. Theory of function spaces, volume 78 of Monographs in Mathe-
matics. Birkhäuser Verlag, Basel, 1983. 407, 409, 410, 412, 413, 416
Triebel, H. Theory of function spaces. II, volume 84 of Monographs in Math-
ematics. Birkhäuser Verlag, Basel, 1992. 407, 416
Triebel, H. Fractals and spectra, volume 91 of Monographs in Mathematics.
Birkhäuser Verlag, Basel, 1997. 407, 413
Triebel, H. Theory of function spaces. III, volume 100 of Monographs in
Mathematics. Birkhäuser Verlag, Basel, 2006. 407, 416
Triebel, H. Local function spaces, heat and Navier–Stokes equations, volume 20
of EMS Tracts in Mathematics. European Mathematical Society (EMS),
Zürich, 2013. 407
Triebel, H. Hybrid function spaces, heat and Navier–Stokes equations, vol-
ume 24 of EMS Tracts in Mathematics. European Mathematical Society
(EMS), Zürich, 2014. 407
Triebel, H. Theory of function spaces IV, volume 107 of Monographs in Math-
ematics. Birkhäuser/Springer, Cham, 2020. 407
Uiterdijk, M. On functional calculi for closed linear operators. PhD thesis,
Delft University of Technology, 1998. 502
Vagharshakyan, A. Recovering singular integrals from Haar shifts. Proc.
Amer. Math. Soc., 138(12):4303–4309, 2010. 73
818 REFERENCES

Venni, A. A counterexample concerning imaginary powers of linear operators.

In Functional analysis and related topics, 1991 (Kyoto), volume 1540 of
Lecture Notes in Math., pages 381–387. Springer-Verlag, Berlin, 1993. 503
Veraar, M. C. Embedding results for γ-spaces. In Recent trends in analysis,
volume 16 of Theta Ser. Adv. Math., pages 209–219. Theta, Bucharest,
2013. 411, 414, 740, 741
Voigtlaender, F. Structured, compactly supported Banach frame decomposi-
tions of decomposition spaces. Dissertationes Math., 575:1–179, 2022. 408
Volberg, A. Matrix Ap weights via S-functions. J. Amer. Math. Soc., 10(2):
445–466, 1997. 81, 416
Volberg, A. The proof of the nonhomogeneous T 1 theorem via averaging of
dyadic shifts. Algebra i Analiz, 27(3):75–94, 2015. 222
Walker, C. On diffusive and non-diffusive coalescence and breakage processes.
PhD thesis, University of Zürich, 2003. 415
Watanabe, J. On some properties of fractional powers of linear operators.
Proc. Japan Acad., 37:273–275, 1961. 502
Watanabe, K. Strong time-periodic solutions to chemotaxis-Navier–Stokes
equations on bounded domains. Discrete Contin. Dyn. Syst., 42(11):5577–
5590, 2022. 677
Watanabe, K. Decay estimates of gradient of the Stokes semigroup in exterior
Lipschitz domains. J. Differential Equations, 346:277–312, 2023. 677
Watson, D. K. Weighted estimates for singular integrals via Fourier transform
estimates. Duke Math. J., 60(2):389–399, 1990. 78
Weidemaier, P. Maximal regularity for parabolic equations with inhomoge-
neous boundary conditions in Sobolev spaces with mixed Lp -norm. Electron.
Res. Announc. Amer. Math. Soc., 8:47–51, 2002. 412, 677
Weis, L. W. Stability theorems for semi-groups via multiplier theorems. In Dif-
ferential equations, asymptotic analysis, and mathematical physics (Pots-
dam, 1996), volume 100 of Math. Res., pages 407–411. Akademie Verlag,
Berlin, 1997. 409
Weis, L. W. A new approach to maximal Lp -regularity. In Evolution equa-
tions and their applications in physical and life sciences (Bad Herrenalb,
1998), volume 215 of Lecture Notes in Pure and Appl. Math., pages 195–
214. Dekker, New York, 2001a. 675, 676
Weis, L. W. Operator-valued Fourier multiplier theorems and maximal Lp -
regularity. Math. Ann., 319(4):735–758, 2001b. 672, 675, 676
Wiener, N. and P. Masani. The prediction theory of multivariate stochastic
processes. II. The linear predictor. Acta Math., 99:93–137, 1958. 81
Wilson, M. The intrinsic square function. Rev. Mat. Iberoam., 23(3):771–791,
2007. 84
Wilson, M. Weighted Littlewood–Paley theory and exponential-square integra-
bility, volume 1924 of Lecture Notes in Mathematics. Springer, Berlin, 2008.
84
Xu, H. Maximal L1 regularity for solutions to inhomogeneous incompressible
Navier–Stokes equations. J. Differential Equations, 335:1–42, 2022. 679
 REFERENCES 819

Xu, Q. Littlewood–Paley theory for functions with values in uniformly convex

spaces. J. Reine Angew. Math., 504:195–226, 1998. 415
Xu, Q. Vector-valued Littlewood–Paley–Stein theory for semigroups ii. Int.
Math. Res. Not. IMRN, 2020(21):7769–7791, 2020. 415
Yabuta, K. Generalizations of Calderón–Zygmund operators. Studia Math.,
82(1):17–31, 1985. 220, 738
Yagi, A. Coı̈ncidence entre des espaces d’interpolation et des domaines de
puissances fractionnaires d’opérateurs. C. R. Acad. Sci. Paris Sér. I Math.,
299(6):173–176, 1984. 503
Yagi, A. Parabolic evolution equations in which the coefficients are the gen-
erators of infinitely differentiable semigroups. II. Funkcial. Ekvac., 33(1):
139–150, 1990. 684
Yagi, A. Abstract quasilinear evolution equations of parabolic type in Banach
spaces. Boll. Un. Mat. Ital. B (7), 5(2):341–368, 1991. 684
Yagi, A. Abstract parabolic evolution equations and their applications. Springer
Monographs in Mathematics. Springer-Verlag, Berlin, 2010. 684, 727, 742
Yang, D. and W. Yuan. A new class of function spaces connecting Triebel–
Lizorkin spaces and Q spaces. J. Funct. Anal., 255(10):2760–2809, 2008.
412, 417
Yang, Q., L. Yan, and D. Deng. On Hörmander condition. Chinese Sci. Bull.,
42(16):1341–1345, 1997. 221, 738
Yaroslavstev, I. S. Fourier multipliers and weak differential subordination of
martingales in UMD Banach spaces. Studia Math., 243(3):269–301, 2018.
290
Yosida, K. Fractional powers of infinitesimal generators and the analyticity
of the semi-groups generated by them. Proc. Japan Acad., 36:86–89, 1960.
502
Yosida, K. Functional analysis, volume 123 of Grundlehren der Mathematis-
chen Wissenschaften. Springer-Verlag, Berlin, sixth edition, 1980. 502
Youssfi, A. Continuité-Besov des opérateurs définis par des intégrales sin-
gulières. Manuscripta Math., 65(3):289–310, 1989. 223
Zacher, R. Maximal regularity of type Lp for abstract parabolic Volterra
equations. J. Evol. Equ., 5(1):79–103, 2005. 688
Zsidó, L. Spectral properties of the analytic generator and singular integrals.
Atti Accad. Naz. Lincei Mem. Cl. Sci. Fis. Mat. Natur. Sez. Ia (8), 17(4):
i–ii and 105–134, 1983. 503
Index

A + B, 515 MD , 6
s
Bp,q (Rd ; X), 320 Mσ,A , 420
∞
Cc (U ; X), 303 M
fσ,A , 517
Cbθ (Rd ; X), 302 fRp , 517
M σ,A
Cbk (Rd ; X), 302 fRp , 517
Cubs
(Rd ; X), 302 M σ,A
Cub (Rd ; X), 302
k ℘ (good set-bound), 160
E(Σσ ), 420 R̈d , 351
s
Fp,q (Rd ; X), 370 %(A), 420
H (Σσ ; A ), 542
∞ RH ∞ (Σσ ; A ), 542
H p (Sϑ ), 485 S 0 (Rd ; X), 300
s,p d S (Rd ; X), 300
RH (R ; X), 301 Sϑ , 485
−, 6
Lpw (I, w; X), 607 σ(A), 420
b pw (I; X), 608
L Σω , 420
Σωbi , 497
L1 (Rd ; X), 308
W k,p (D; X), 300
Lip(Rd ; X), 302
W s,p (Rd ; X), 301
L∞ d
fin (R ; X), 24 1,p
reg 0 ẆA , 576
Mcont,A (I), 614
reg [w]Ap , 57
Mp,A,λ (I), 594 (X0 , X1 )Tr
reg θ,p , 765
Mp,A (I), 577 Xθ,∞ , 764
MRp , 693 ω(A), 420
#
M0,λ , 13 ω bi (A), 497

© Springer Nature Switzerland AG 2023 820

T. Hytönen et al., Analysis in Banach Spaces, Ergebnisse der Mathematik und ihrer
Grenzgebiete. 3. Folge / A Series of Modern Surveys in Mathematics 76,
https://doi.org/10.1007/978-3-031-46598-7
 INDEX 821

ωH ∞ (A), 476 decomposition, 25

ωBIP (A), 458 kernel, 41
ωabs (A), 551 R-bounded version, 176
,→, 295, 764 theorem, 24, 45
c̃K (℘) (kernel bound), 177
A1 conjecture, 84 cK (℘) (kernel bound), 176
A2 conjecture, 73 complex interpolation
A2 theorem, 64 and fractional domains, 460
abscissa of Besov spaces, 345
BIP, 458 of Bessel potential spaces, 403
absolute calculus, 551 of Lorentz spaces, 313
and Lp -bounds, 553 of Sobolev spaces, 403
implies operator-valued calculus, 551 of Triebel–Lizorkin spaces, 385
algorithm of weighted Lq , 311
BCR, 154, 213 condition
Allen-Cahn equation, 719 Dini, 42
almost γ-sectorial, 470 with logarithm, 181
almost sectorial, 470 Hörmander, 23
angle conjecture
of bisectoriality, 497 A2 , 73
of sectoriality, 420 Astala–Iwaniec–Saksman, 72
atom (of the Hardy space), 17 Kato’s square root, 509
Muckenhoupt–Wheeden, 84
BCR algorithm, 154, 213
convolution kernel, 177
Besov space, 320
critical
Bessel potential operator, 328
part, 697
Bessel potential space, 301
space, 699
BIP, 457
⇒ R-sectoriality under UMD, 462 de la Vallée–Poussin kernel, 238
⇒ ω(A) 6 ωBIP(A) , 462 decomposition
⇒ almost γ-sectoriality, 471 Calderón–Zygmund, 25
associated C0 -group, 458 Figiel’s, 157, 163
diagram of implications, 457 random version, 171
γ-, 472 derivative
Hilbert space case, 482 fractional, 452
R-, 472 diagram
bisector, 497 H ∞ -calculus vs. BIP vs. sectorial,
bisectorial operator, 497 457
block sequence, 669 difference norm
blow-up criterion, 713 for Besov spaces, 338
BMO (bounded mean oscillation), 16 for Triebel–Lizorkin spaces, 380
duality with the Hardy space, 20 difference operator, 337
vs. singular integrals on L∞ , 34 dilation
bounded imaginary powers, 457 of a dyadic cube, 51
Calderón–Zygmund Dini condition
822 INDEX

with logarithm, 181 and complex interpolation, 460

in the A2 theorem, 76 fractional powers, 437
Dini condition/kernel, 42 interpolation estimate, 442
Dirichlet kernel, 238 functional calculus
dissociate set, 287 joint, 546
DR p -bounded, 92 operator-valued, 542
dual Hörmander
condition/kernel, 23 γ-BIP, 472
dual weight, 57 ⇔ H ∞ -calculus, 479
trick, 57 ⇒ γ-sectoriality, 473
Dunford calculus diagram of implications, 457
joint, 546 γ-interpolation method, 505
operator-valued, 541 γ-sectoriality
dyadic cube ⇐ γ-BIP, 473
dilated, 51 ⇐ BIP under UMD, 462
good, 169 and maximal γ-regularity, 686
random, 169 Gaussian
shifted, 52 random variable, xix
dyadic representation theorem, 206 generator, 746
dyadic shift, 199 good
boundedness on Lp (Rd ; X), 200 dyadic cube, 169
dyadic square function, 84 set-bound, 160
dyadic system, 4
standard, 5 H ∞ -calculus
⇔ γ-BIP, 479
E R p -bounded, 92 ⇒ ωBIP (A) = ωH ∞ (A) , 476
evolution family, 742 bisectorial, 497
example diagram of implications, 457
Kalton–Lancien, 665 joint, 547
McIntoch–Yagi, 530 on real interpolation spaces, 558
extended calculus operator-valued, 542
and H ∞ -calculus, 433 vs. bounded extended calculus, 433
via compensation, 435 Haar multipliers, 95
extension problem Hardy space, 17
Calderón–Zygmund operators, 193 atomic ⇔ maximal, 18
paraproducts, 139 duality with BMO, 20
Hardy–Littlewood
Fejér kernel, 238 dyadic maximal function, 6
Figiel’s decomposition, 157, 163 height
random version, 171 of a strip type operator, 485
Fourier transform, 301 Hodge–Dirac operator, 501
Fourier type, 356 Hörmander condition/kernel, 23
⇐ non-trivial type, 256
fractional derivative, 451, 452 imaginary powers
fractional domains γ-bounded, 472
 INDEX 823

bounded, 457 Lipschitz continuous, 302

inequality Littlewood–Paley sequence
γ-Bernstein–Nikolskii, 353 inhomogeneous, 306
Bernstein–Nikolskii, 334 lower order perturbation
Gagliardo–Nirenberg of the H ∞ -calculus, 525
for Bessel potential spaces, 394
matrix weight, 80
for Triebel–Lizorkin spaces, 376,
377 maximal γ-regularity, 684
Hardy, 773 maximal L1 -regularity
Hardy, fractional, 389 and copies of c0 in the dual, 663
Hardy–Young, 387 maximal L∞ -regularity
Peetre maximal, 365 and copies of c0 , 661
reverse Hölder, 78 maximal Lp -regularity, 577
Young, 304 R+ versus (0, T ), 591
instantaneous regularisation, 611, 623 R+ versus R, 649, 650, 652
integrable limits, 421 and duality, 598
interpolation and passing to subintervals, 590
γ-method, 505 and regularity of mild solutions, 591
trace method, 765 and UMD, 657
extrapolation of the exponent, 604
John–Strömberg maximal operator, 13 implies analyticity, 586
joint functional calculus, 546 implies closedness, 577
implied by absolute calculus, 551 on R, 645
on R, and bisectoriality, 647
Kato square root problem, 509 permanence properties, 598
kernel weighted, on R, 647
Calderón–Zygmund, 41 with exponential weights, 612, 647
R-bounded version, 176 with weights, 613
convolution, 177 maximal function
de la Vallée-Poussin, 238 dyadic Hardy–Littlewood, 6
Dini, 42 maximal inequality
Dirichlet, 238 Peetre, 365
dual Hörmander, 23 maximal operator
Fejér, 238 #
John–Strömberg M0,λ , 13
Hörmander, 23 Lerner’s, 45
of an integral operator, 22 strong, 79
Poisson, 453 maximal regularity
standard, 42 Lp -, 577
γ-, 685
Lp -solution space, 576 continuous, 614
Laplacian weighted Lp -, 608
in Lp (Rd ; X), 456, 462 weighted continuous, 617
lemma maximal regularity constant, 594
Prüss–Sohr, 462 maximal solution, 712
Lerner’s formula, 7 microscopic parameter, 293
824 INDEX

Muckenhoupt weight, 57 by a relatively bounded operator:

Muckenhoupt–Wheeden conjecture, 84 counterexample, 530
Poisson
Nemitskii map, 696 kernel, 453
non-trivial type semigroup, 452
⇒ Fourier type, 256 primary calculus, 421
pseudo-localisation, 215
off-diagonal estimates, 562 pseudo-resolvent, 493
i
ωK (℘) (modulus of continuity), 176
ω̃K (℘) (modulus of continuity), 177 quadrant, 5
Operator quasi-independent set, 231
Ritt, 687 vs. dissociate set, 287
operator
almost γ-sectorial, 470 R-bounded
almost sectorial, 470 range of smooth functions, 361
bisectorial, 497 Rademacher variable, xix
and spectral projections, 498 random variable
vs. sectorial, via squares, 501 Gaussian, xix
Hodge–Dirac, 501 Rademacher, xix
sectorial, 420 real interpolation
standard, 456 of Besov spaces, 346
sparse, 50 of Triebel–Lizorkin spaces, 385
strip type, 485 of weighted Lq , 314
Volterra, 452 weighted `q (X), 317
oscλ (f ; Q), 2 real interpolation spaces
bounded H ∞ -calculus on, 558
paraproduct, 122 reference solution, 694
conditions for boundedness, 138 regularisable, 424
symmetric case, 143 regularisation
Mei’s counterexample, 145 instantaneous, 611, 623
perturbation regulariser, 424
of a sectorial operator relatively bounded, 619
by a bounded operator, 518 resolvent commutation, 516
by a lower order operator, 522 reverse Hölder inequality, 78
by a relatively bounded operator, Riesz projections
519 abstract, 497
of an R-sectorial operator Ritt operators, 687
by a lower order operator, 522
by a relatively bounded operator, Schauder basis
520 1-unconditional, 667
of the H ∞ -calculus Schwartz space, 300
by a lower order operator, 525 sector, 420
by a multiple of the identity, 523 sectorial operator, 420
by a relatively bounded operator, almost, 470
526 almost γ-, 470
 INDEX 825

bi-, 497 Kalton–Weis, 544

standard, 456
semigroup, 745 tangent martingales
generated by a sectorial operator, to estimate dyadic shifts, 202
753 tempered distributions, 300
heat, 453 tent spaces, 687
locally bounded, 745 test functions, 303
Poisson, 452, 453 for dyadic operators, 88
strongly measurable, 745 theorem
uniformly exponentially stable, 595, γ-Sobolev embedding
754 for Besov spaces, 352
weakly measurable, 745 for Bessel potential spaces, 398
shift (dyadic), 199 A2 -, 64
shifted dyadic cube, 52 Baillon, 661
Sidon set, 231 Balakrishnan, 445
singular integral Bessaga–Pelczyński, 38
weakly defined, 174 boundedness of
Sobolev space, 300 dyadic shifts, 200
Sobolev–Slobodetskii space, 301 Figiel’s T operators, 108
solution Figiel’s U operators, 116
Lp -, 575, 645, 692 Haar multipliers, 95
Lpwα -, 701 paraproducts, 138
γ-, 685 symmetric paraproducts, 143
mild, 571, 589, 686 Bourgain, on Fourier type, 256
strong, 571, 645 Calderón–Zygmund, 24, 45
weak, 572 Clément–Prüss, 462, 471
sparse Cowling–Doust–McIntosh–Yagi, 476
collection of sets, 6 Cruz-Uribe–Martell–Pérez, 57
domination Da Prato–Grisvard, 552
of abstract operators, 45, 53 Datko–Pazy, 754
of singular integrals, 54 De Simon, 632
operator, 50 Dore, 586
spectral projections Dore–Kato, 595
of a bisectorial operator, 498 Dore–Venni, 496, 632
square function dyadic representation, 206
dyadic, 84 equivalence of maximal γ-regularity
square root problem of Kato, 509 and γ-sectoriality, 686
standard kernel, 42 equivalence of maximal Lp -regularity
standard sectorial operator, 456 and R-sectoriality in UMD spaces,
Steinhaus variable, xix 626, 656
strong maximal operator, 79 Figiel, 108, 116
sums of operators, 515 Han–Meyer, 401
Dore–Venni, 496 Jawerth–Franke, 386
for non-commuting operators, 566 König, 288
for Weyl pairs, 566 Kalton–Lancien, 665
826 INDEX

Lerner, 7 uniformly exponentially stable, 595

Mei, 145
Mielke, 647 variation, 304
Mihlin VMO, 676
for Besov spaces, 330
for Triebel–Lizorkin spaces, 374 weak boundedness property, 155
under Fourier type, 358 adjacent, 155
Monniaux, 489 DR p -, 177
Nollau, 485 weak solution, 572
Peetre maximal, 367 weakly defined singular integral, 174
pointwise multipliers weight, 56
on Besov spaces, 393 and maximal regularity, 607
on Bessel potential spaces, 404 dual, 57
on Triebel–Lizorkin spaces, 390, interpolation with change of, 310
392 matrix-, 80
Prüss–Sohr, 632 Weyl commutation relations, 566
Schmeisser–Sickel, 395
Sobolev embedding
for Besov spaces, 333
for Triebel–Lizorkin spaces, 376,
379
sparse domination
abstract version, 45, 53
of singular integrals, 54
T (1)
bilinear forms, 162, 170
operator-valued kernels, 182
scalar-valued kernels, 193
trace method, 765
trace space, 692
translation-invariant
bilinear form, 152
trick
dual weight, 57
Triebel–Lizorkin space, 370
Trotter product formula, 560
type
⇒ Fourier type, 256

UMD
and R-boundedness of λ(λ − ∆)−1 ,
680
and maximal Lp -regularity, 657
necessity for multipliers, 277
unconditional, 1-, 667