Dorina Mitrea PDF

Universitext
Dorina Mitrea
Distributions,
Partial Differential
Equations, and
Harmonic Analysis
Second Edition
Universitext
Universitext
Series editors
Sheldon Axler
San Francisco State University
Carles Casacuberta
Universitat de Barcelona
Angus MacIntyre
Queen Mary University of London
Kenneth Ribet
University of California, Berkeley
Claude Sabbah
École polytechnique, CNRS, Université Paris-Saclay, Palaiseau
Endre Süli
University of Oxford
Wojbor A. Woyczyński
Case Western Reserve University
Universitext is a series of textbooks that presents material from a wide variety of

mathematical disciplines at master’s level and beyond. The books, often well
class-tested by their author, may have an informal, personal even experimental
approach to their subject matter. Some of the most successful and established books
in the series have evolved through several editions, always following the evolution
of teaching curricula, into very polished texts.
Thus as research topics trickle down into graduate-level teaching, first textbooks
written for new, cutting-edge courses may make their way into Universitext.
More information about this series at http://www.springer.com/series/223

Dorina Mitrea
Distributions, Partial
Differential Equations,
and Harmonic Analysis
Second Edition
123
Dorina Mitrea
Department of Mathematics
University of Missouri
Columbia, MO, USA
ISSN 0172-5939 ISSN 2191-6675 (electronic)

Universitext
ISBN 978-3-030-03295-1 ISBN 978-3-030-03296-8 (eBook)
https://doi.org/10.1007/978-3-030-03296-8
Library of Congress Control Number: 2018960188
Mathematics Subject Classification (2010): 35A08, 35A09, 35A20, 35B05, 35B53, 35B65, 35C15,
35D30, 35E05, 35G05, 35G35, 35H10, 35J05, 35J30, 35J47, 42B20, 42B37, 46E35
1st edition: © Springer Science+Business Media New York 2013

2nd edition: © Springer Nature Switzerland AG 2018
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made. 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
Dedicată cu drag lui Diana şi Adrian,
Love, Mom
Preface to the Second Edition
The main additions in the current edition pertain to fundamental solutions and
Sobolev spaces. The list of fundamental solutions from Chapter 7 has been
expanded to include the Helmholtz operator (x7.6), the perturbed Dirac operator
(x7.10), and their iterations (x7.7 and x7.11). Understanding quantitative and
qualitative features of the said fundamental solutions is of paramount importance in
scattering theory, where Helmholtz and perturbed Dirac operators play a prominent
role. Special emphasis is placed on elucidating the nature of their singularity at the
origin and asymptotic behavior at infinity. In this vein, a number of useful results
concerning the behavior of Hankel functions are summarized in x14.10.
The new material concerning Sobolev spaces is contained in Chapter 12. Our
approach builds the theory from ground up and is completely self-contained.
Presently, we limit ourselves to L2-based Sobolev spaces, where the connection
with the Fourier analysis developed earlier in the monograph is most apparent. Such
an approach is also natural from a pedagogical point of view. This being said, many
key results have been given proofs which canonically adapt to more general situ-
ations (such as Lp-based Sobolev spaces, weighted Sobolev spaces, Sobolev spaces
with vanishing traces). The starting point is the treatment of global Sobolev spaces
H s ðRn Þ in Rn of arbitrary smoothness s 2 R, via the Fourier transform (cf. x12.1).
The theory is natural and elegant since the Fourier transform is an isometry on
L2 ðRn Þ. The next step is the consideration of Sobolev spaces in arbitrary open
subsets X of Rn . There are two natural venues to define the latter brand of spaces.
One approach, yielding the scale H s ðXÞ with s 2 R, proceeds via restriction from
the corresponding spaces in Rn (cf. x12.2). For integer amounts of smoothness
m 2 N, one may also introduce Sobolev spaces H m ðXÞ in an intrinsic fashion,
demanding that distributional derivatives up to order m are square-integrable in X
(cf. x12.3).
In relation to Sobolev spaces in an open set X Rn , two basic theorems are
proved in the case when X is a bounded Lipschitz domain. The first is the density of
restrictions to X of smooth compactly supported functions from Rn in the intrinsic
Sobolev space H m ðXÞ. The second is the construction of Calderón’s extension
vii
viii Preface to the Second Edition
operator from H m ðXÞ to H m ðRn Þ. Among other things, these results allow the
identification of the intrinsic Sobolev space H m ðXÞ with the restriction Sobolev
space H m ðXÞ whenever m 2 N and X is a bounded Lipschitz domain.
In x12.4, we treat L2-based Sobolev spaces H 1=2 , of fractional order 1/2, on
boundaries of Lipschitz domains. These are defined as spaces of square-integrable
functions satisfying a finiteness condition involving a suitable Gagliardo–
Slobodeckij semi-norm. In the setting of bounded Lipschitz domains, this study ties
up with the earlier theory via extension and trace results, as explained in x12.5.
Specifically, having first established the density of Lipschitz functions on @X,
where X is a bounded Lipschitz domain, in the space H 1=2 ð@XÞ, we then prove the
existence of a linear and bounded trace operator from H 1 ðXÞ into H 1=2 ð@XÞ, and of
a linear and bounded extension operator from H 1=2 ð@XÞ into H 1 ðXÞ.
The work on this project has been supported in part by the Simons Foundation
grants # 426669 and # 200750 and by a University of Missouri Research Leave
grant. The author wishes to express her gratitude to these institutions.
Columbia, MO, USA Dorina Mitrea

July 2018
Preface to the First Edition
This book has been written from the personal perspective of a mathematician
working at the interface between partial differential equations and harmonic anal-
ysis. Its aim is to offer, in a concise, rigorous, and largely self-contained form, a
rapid introduction to the theory of distributions and its applications to partial dif-
ferential equations and harmonic analysis. This is done in a format suitable for a
graduate course spanning either over one semester, when the focus is primarily on
the foundational aspects, or over a two-semester period that allows for the proper
amount of time to cover all intended applications as well.
Throughout, a special effort has been made to develop the theory of distributions
not as an abstract edifice but rather give the reader a chance to see the rationale
behind various seemingly technical definitions, as well as the opportunity to apply
the newly developed tools (in the natural build-up of the theory) to concrete
problems in partial differential equations and harmonic analysis, at the earliest
opportunity.
In addition to being suitable as a textbook for a graduate course, the monograph
has been designed so that it may also be used for independent study since the
presentation is reader-friendly, mostly self-sufficient (for example, all auxiliary
results originating outside the scope of the present monograph have been carefully
collected and presented in the appendix), and a large number of the suggested
exercises have complete solutions.
Columbia, MO, USA Dorina Mitrea

March 2013
ix
Contents
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv
Common Notational Conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxi
1 Weak Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 The Cauchy Problem for a Vibrating Infinite String . . . . . . . . 1
1.2 Weak Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 The Spaces EðXÞ and DðXÞ . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 Additional Exercises for Chapter 1 . . . . . . . . . . . . . . . . . . . . . 14
2 The Space D0 ðXÞ of Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.1 The Definition of Distributions . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 The Topological Vector Space D0 ðXÞ . . . . . . . . . . . . . . . . . . 27
2.3 Multiplication of a Distribution with a C 1 Function . . . . . . . . 30
2.4 Distributional Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.5 The Support of a Distribution . . . . . . . . . . . . . . . . . . . . . . . . 37
2.6 Compactly Supported Distributions and the Space E0 ðXÞ . . . . . 41
2.7 Tensor Product of Distributions . . . . . . . . . . . . . . . . . . . . . . . 51
2.8 The Convolution of Distributions in Rn . . . . . . . . . . . . . . . . . 64
2.9 Distributions with Higher Order Gradients Continuous or
Bounded . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 78
2.10 Additional Exercises for Chapter 2 . . . . . . . . . . . . . . . . . .... 89
3 The Schwartz Space and the Fourier Transform . . . . . . . . ...... 97
3.1 The Schwartz Space of Rapidly Decreasing Functions . ...... 97
3.2 The Action of the Fourier Transform on the Schwartz
Class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
xi
xii Contents
4 The Space of Tempered Distributions . . . . . . . . . . . . . . . . . . . . . . . 117

4.1 Definition and Properties of Tempered Distributions . . . . . . . . 117
4.2 The Fourier Transform Acting on Tempered Distributions . . . . 129
4.3 Homogeneous Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . 140
4.4 Principal Value Tempered Distributions . . . . . . . . . . . . . . . . . 148
4.5 The Fourier Transform of Principal Value Distributions . . . . . 153
4.6 Tempered Distributions Associated with jxjn . . . . . . . . . . . . . 160
4.7 A General Jump-Formula in the Class of Tempered
Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
4.8 The Harmonic Poisson Kernel . . . . . . . . . . . . . . . . . . . . . . . . 177
4.9 Singular Integral Operators . . . . . . . . . . . . . . . . . . . . . . . . . . 181
4.10 Derivatives of Volume Potentials . . . . . . . . . . . . . . . . . . . . . . 189
5 The Concept of Fundamental Solution . . . . . . . . . . . . . . . . . . . . . . 203
5.1 Constant Coefficient Linear Differential Operators . . . . . . . . . . 203
5.2 A First Look at Fundamental Solutions . . . . . . . . . . . . . . . . . 205
5.3 The Malgrange–Ehrenpreis Theorem . . . . . . . . . . . . . . . . . . . 209
6 Hypoelliptic Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
6.1 Definition and Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
6.2 Hypoelliptic Operators with Constant Coefficients . . . . . . . . . . 218
6.3 Integral Representation Formulas and Interior Estimates . . . . . 224
7 The Laplacian and Related Operators . . . . . . . . . . . . . . . . . . . . . . 233
7.1 Fundamental Solutions for the Laplace Operator . . . . . . . . . . . 233
7.2 The Poisson Equation and Layer Potential Representation
Formulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
7.3 Fundamental Solutions for the Bi-Laplacian . . . . . . . . . . . . . . 249
7.4 The Poisson Equation for the Bi-Laplacian . . . . . . . . . . . . . . . 254
7.5 Fundamental Solutions for the Poly-harmonic Operator . . . . . . 257
7.6 Fundamental Solutions for the Helmholtz Operator . . . . . . . . . 265
7.7 Fundamental Solutions for the Iterated Helmholtz Operator . . . 279
7.8 Fundamental Solutions for the Cauchy–Riemann Operator . . . . 288
7.9 Fundamental Solutions for the Dirac Operator . . . . . . . . . . . . 293
7.10 Fundamental Solutions for the Perturbed Dirac Operator . . . . . 301
7.11 Fundamental Solutions for the Iterated Perturbed Dirac
Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
7.12 Fundamental Solutions for General Second-Order
Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
7.13 Layer Potential Representation Formulas Revisited . . . . . . . . . 316
Contents xiii
8 The Heat Operator and Related Versions . . . . . . . . . . . . . . . . . . . . 321

8.1 Fundamental Solutions for the Heat Operator . . . . . . . . . . . . . 321
8.2 The Generalized Cauchy Problem for the Heat Operator . . . . . 324
8.3 Fundamental Solutions for General Second-Order Parabolic
Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326
8.4 Fundamental Solution for the Schrödinger Operator . . . . . . . . 330
9 The Wave Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333
9.1 Fundamental Solution for the Wave Operator . . . . . . . . . . . . . 333
9.1.1 The Case n ¼ 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
9.1.2 The Case n ¼ 2p þ 1, p 1 . . . . . . . . . . . . . . . . . . . 338
9.1.3 The Method of Descent . . . . . . . . . . . . . . . . . . . . . . 344
9.1.4 The Case n ¼ 2p, p 1 . . . . . . . . . . . . . . . . . . . . . . 347
9.1.5 Summary for Arbitrary n . . . . . . . . . . . . . . . . . . . . . 350
9.2 The Generalized Cauchy Problem for the Wave Operator . . . . 352
10 The Lamé and Stokes Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . 355
10.1 General Remarks About Vector and Matrix Distributions . . . . 355
10.2 Fundamental Solutions and Regularity for General Systems . . . 360
10.3 Fundamental Solutions for the Lamé Operator . . . . . . . . . . . . 364
10.4 Mean Value Formulas and Interior Estimates for the Lamé
Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
10.5 The Poisson Equation for the Lamé Operator . . . . . . . . . . . . . 378
10.6 Fundamental Solutions for the Stokes Operator . . . . . . . . . . . . 380
10.7 Additional Exercises for Chapter 10 . . . . . . . . . . . . . . . . . . . . 385
11 More on Fundamental Solutions for Systems . . . . . . . . . . . . . . . . . 387
11.1 Computing a Fundamental Solution for the Lamé Operator . . . 387
11.2 Computing a Fundamental Solution for the Stokes
Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
11.3 Fundamental Solutions for Higher Order Systems . . . . . . . . . . 390
11.4 Interior Estimates and Real-Analyticity for Null-Solutions
of Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 406
11.5 Reverse Hölder Estimates for Null-Solutions of Systems . . . . . 411
11.6 Layer Potentials and Jump Relations for Systems . . . . . . . . . . 416
12 Sobolev Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
12.1 Global Sobolev Spaces H s ðRn Þ, s 2 Rn . . . . . . . . . . . . . . . . . 425
12.2 Restriction Sobolev Spaces H s ðXÞ, s 2 R . . . . . . . . . . . . . . . . 438
12.3 Intrinsic Sobolev Spaces H m ðXÞ, m 2 N . . . . . . . . . . . . . . . . 442
12.4 The Space H 1=2 ð@XÞ on Boundaries of Lipschitz Domains . . . 469
12.5 Traces and Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 477
xiv Contents
13 Solutions to Selected Exercises . . ............ . . . . . . . . . . . . . . 497

13.1 Solutions to Exercises from Section 1.4 . . . . . . . . . . . . . . . . . 497
13.2 Solutions to Exercises from Section 2.10 . . . . . . . . . . . . . . . . 505
14 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 543
14.1 Summary of Topological and Functional Analytic Results . . . . 543
14.2 Basic Results from Calculus, Measure Theory,
and Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 552
14.3 Custom-Designing Smooth Cut-off Functions . . . . . . . . . . . . . 560
14.4 Partition of Unity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 562
14.5 The Gamma and Beta Functions . . . . . . . . . . . . . . . . . . . . . . 566
14.6 Surfaces in Rn and Surface Integrals . . . . . . . . . . . . . . . . . . . 567
14.7 Lipschitz Domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 568
14.8 Integration by Parts and Green’s Formula . . . . . . . . . . . . . . . . 575
14.9 Polar Coordinates and Integrals on Spheres . . . . . . . . . . . . . . 576
14.10 Hankel Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 583
14.11 Tables of Fourier Transforms . . . . . . . . . . . . . . . . . . . . . . . . . 588
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 591
Subject Index. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 595
Symbol Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 599
Introduction
It has long been recognized that there is a large overlap and intricate interplay
among Distribution Theory (DT), Partial Differential Equations (PDE), and
Harmonic Analysis (HA). The purpose of this monograph is to guide a reader with a
background in basic real analysis through the journey taking her/him to the stage
when such connections become self-evident.
Another goal of the present book is to convince the reader that traditional
distinctions made among these branches of mathematics are largely artificial and are
often simply a matter of choice in focus. Indeed, given the manner in which they
complement, motivate, and draw inspiration from one another, it is not necessarily a
stretch to attempt to pursue their development virtually simultaneously.
Concerning the triumvirate DT, PDE, HA, while there exist a number of good
reference texts available on the market, they are by and large conceived in such a
way that they either emphasize more one of these topics, typically at the detriment
of the others, or are simply not particularly well suited for a nonspecialist. By way
of contrast, not only is the present text written in a way that brings together and
blends the aforementioned topics in a unified, coherent body of results, but the
resulting exposition is also sufficiently detailed and reader-friendly so that it may be
read independently, outside the formal classroom setting. Indeed, the book is
essentially self-contained, presents a balanced treatment of the topics involved, and
contains a large number of exercises (upwards of two hundred, more than half of
which are accompanied by solutions), which have been carefully chosen to amplify
the effect, and substantiate the power and scope, of the theory discussed here.
While the topics treated are classical, the material is not entirely standard since a
number of results are new even for a seasoned practitioner, and the overall archi-
tectural design of the monograph (including the way in which certain topics are
covered) is original.
Regarding its inception, the present monograph is an expanded version of the
notes I prepared for a course on distribution theory I taught in the Spring of 2007
and the Spring of 2011, at the University of Missouri. My intention was to present
the theory of distributions not as an abstract edifice but rather give the student a
chance to instantaneously see the justification and practical benefits of the multitude
xv
xvi Introduction
of seemingly technical definitions and results, as well as give her/him the oppor-
tunity to immediately see how the newly introduced concepts (in the natural
build-up of the theory) apply to concrete problems in partial differential equations
and harmonic analysis.
Special care has been paid to the pedagogical aspect of the presentation of the
material in the book. For example, a notable feature of the present monograph is the
fact that fundamental solutions for some of the most basic differential operators in
mathematical physics and engineering, including Laplace, heat, wave, poly-
harmonic, Dirac, Lamé, Stokes, and Schrödinger, are systematically deduced
starting from first principles. This stands in contrast with the more common practice
in the literature in which one starts with a certain distribution (whose origins are
fairly obscure) and simply checks that the distribution in question is a fundamental
solution for a given differential operator. Another feature is the emphasis placed on
the interrelations between topics. For example, a clear picture is presented as to how
DT vastly facilitates the computation of fundamental solutions, and the develop-
ment of singular integral operators, tools which, in turn, are used to solve PDE as
well as represent and estimate solutions of PDE.
The presentation is also conceived in such a way as to avoid having to confront
heavy duty topology/functional analysis up front, in the main narrative. For
example, the jargon associated with the multitude of topologies on various spaces
of test functions and distributions is minimized by deferring to an appendix the
technical details while retaining in the main body of the monograph only those
consequences that are most directly relevant to the fluency of the exposition.
While the core material I had in mind deals primarily with the theory of dis-
tributions, the monograph is ultimately devised in such a way as to make the present
material a solid launching pad for a number of subsequent courses, dealing with
allied topics, including:
• Harmonic Analysis
• Partial Differential Equations
• Boundary Integral Methods
• Sobolev Spaces
• Pseudodifferential Operators
For example, the theory of singular integral operators of convolution type in
L2 ðRn Þ is essentially developed here, in full detail, up to the point where more
specialized tools from harmonic analysis (such as the Hardy–Littlewood maximal
operator and the Calderón–Zygmund lemma) are typically involved in order to
further extend this theory (via a weak-(1, 1) estimate, interpolation with L2, and
then duality) to Lp spaces with p 2 ð1; 1Þ, as in [8], [11], [19], [26], [52], [68],
[69], [70], among others.
Regarding connections with partial differential equations, the Poisson problem in
the whole space,
Introduction xvii
Lu ¼ f in D0 ðRn Þ; ð1Þ
is systematically treated here for a variety of differential operators L, including the

Laplacian, the biharmonic operator, and the Lamé system. While Sobolev spaces
are not explicitly considered (the interested reader may consult in this regard [2],
[3], [15], [43], [47], [48], [82], to cite just a small fraction of a large body of
literature on this topic), they lurk just beneath the surface as their presence is
implicit in estimates of the form
X
k@ fi ukLp ðRn Þ CðL; pÞk f kLp ðRn Þ ; 1 \ p \ 1; ð2Þ
jfij¼m
where m denotes the order of the elliptic differential operator L.

Such estimates are deduced from an integral representation formula for u,
involving a fundamental solution of L, and estimates for singular integral operators
of convolution type in Rn . In particular, this justifies two features of the present
monograph: (1) the emphasis placed on finding explicit formulas for fundamental
solutions for a large number of operators of basic importance in mathematics,
physics, and engineering; (2) the focus on the theory of singular integral operators
of convolution type, developed alongside the distribution theory. In addition, the
analysis of the Cauchy problem formulated and studied for the heat and wave
operators once more underscores the significance of the fundamental solutions for
the named operators. As a whole, this material is designed to initiate the reader into
the field of partial differential equations. At the same time, it complements, and
works well in tandem with, the treatment of this subject in [5], [14], [24], [33], [34],
[35], [50], [74], [75], [81].
Whenever circumstances permit it, other types of problems are brought into play,
such as the Dirichlet and Neumann problem in the upper half-space for the
Laplacian, as well as more general second order systems. In turn, the latter genre of
boundary value problems motivates introducing and developing boundary integral
methods, and serves as an opportunity to highlight the basic role that layer potential
operators play in this context. References dealing with the latter topic include [37],
[39], [50], [53], [80].
The analysis of the structure of the boundary layer potential operators naturally
intervening in this context also points to the possibility of considering larger classes
of operators where the latter may be composed, inverted, etc., in a stable fashion.
This serves as an excellent motivation for the introduction of such algebras of
operators as pseudodifferential and Fourier integral operators, a direction which the
interested reader may then pursue in, e.g., [27], [73], [75], [81], to name a few
sources.
A brief description of this book’s contents is as follows.
Chapters 1–2 are devoted to the development of the most basic aspects of the
theory of distributions. Starting from the discussion of the Cauchy problem for a
vibrating infinite string as a motivational example, the notion of weak derivative is
xviii Introduction
introduced as a mean of extending the notion of solution to a more general setting,

where the functions involved may lack standard pointwise differentiability prop-
erties. After touching upon classes of test functions, the space of distributions is
then introduced and studied from the perspective of a topological vector space with
various other additional features (such as the concept of support, and a partially
defined convolution product). Chapter 3 contains material pertaining to the
Schwartz space of functions rapidly decaying at infinity and the Fourier transform
in such a setting.
In Chapter 4 the action of the Fourier transform is further extended to the setting
of tempered distributions, and several distinguished subclasses of tempered distri-
butions are introduced and studied (including homogeneous and principal value
distributions). The foundational material developed up to this point already has
significant applications to harmonic analysis and partial differential equations. For
example, a general, higher dimensional jump-formula is deduced in this chapter for
a certain class of tempered distributions (that includes the classical harmonic
Poisson kernel) which is later used as the main tool in deriving information about
the boundary behavior of layer potential operators associated with various partial
differential operators and systems. Also, one witnesses here how singular integral
operators of central importance to harmonic analysis (such as the Riesz transforms)
naturally arise as an extension to L2 of the convolution product of tempered dis-
tributions of principal value type with Schwartz functions.
The first explicit encounter with the notion of fundamental solution takes place
in Chapter 5, where the classical Malgrange–Ehrenpreis theorem is presented.
Subsequently, in Chapter 6, the concept of hypoelliptic operator is introduced and
studied. In particular, here a classical result, due to L. Schwartz, is proved to the
effect that a necessary and sufficient condition for a linear, constant coefficient
differential operator to be hypoelliptic in the entire ambient space is that the named
operator possesses a fundamental solution with singular support consisting of the
origin alone. In Chapter 6 we also prove an integral representation formula and
interior estimates for a subclass of hypoelliptic operators, which are subsequently
used to show that null solutions of these operators are real-analytic.
One of the main goals in Chapter 7 is identifying (starting from first principles)
all fundamental solutions that are tempered distributions for scalar elliptic opera-
tors. While the natural starting point is the Laplacian, this study encompasses a
variety of related operators, such as the bi-Laplacian, the poly-harmonic operator,
the Cauchy–Riemann operator, the Dirac operator, as well as general second-order
constant coefficient strongly elliptic operators. Having accomplished this task then
makes it possible to prove the well-posedness of the Poisson problem (1) (equipped
with a boundary condition at infinity), and derive qualitative/quantitative properties
for the solution such as (2). Along the way, Cauchy-like integral operators are also
introduced and their connections with Hardy spaces is brought to light in the setting
of both complex and Clifford analysis.
Chapter 8 has a twofold aim: determine all fundamental solutions that are
tempered distributions for the heat operator and related versions (including the
Schrödinger operator), then use this as a toll in the solution of the generalized
Introduction xix
Cauchy problem for the heat operator. The same type of program is then carried out
in Chapter 9, this time in connection with the wave operator.
While the analysis up to this point has been largely confined to scalar operators,
the final two chapters in the monograph are devoted to studying systems of dif-
ferential operators. The material in Chapter 10 is centered around two such basic
systems: the Lamé operator arising in the theory of elasticity, and the Stokes
operator arising in hydrodynamics. Among other things, all their fundamental
solutions that are tempered distributions are identified, and the well-posedness
of the Poisson problem for the Lamé system is established. The former issue is then
revisited in the first part of Chapter 11 from a different perspective, and subse-
quently generalized to the case of (homogeneous) constant coefficient systems of
arbitrary order. In Chapter 11 we also show that integral representation formulas
and interior estimates hold for null solutions of homogeneous systems with non-
vanishing full symbol. As a consequence, we prove that such null solutions are
real-analytic and satisfy reverse Hölder estimates. The final topic addressed in
Chapter 11 pertains to layer potentials associated with arbitrary constant coefficient
second-order systems in the upper half-space, and the relevance of these operators
vis-a-vis to the solvability of boundary value problems for such systems in this
setting.
For completeness, a summary of topological and functional analysis results in
reference to the description of the topology and equivalent characterizations of
convergence in spaces of test functions and in spaces of distributions is included in
the appendix (which also contains a variety of foundational results from calculus,
measure theory, and special functions originating outside the scope of this book).
One aspect worth noting in this regard is that the exposition in the main body of the
book may be followed even without being fully familiar with all these details by
alternatively taking, as the starting point, the characterization of convergence in the
various topologies considered here (summarized in the main text under the heading
Fact) as definitions. Such an approach makes the topics covered in the present
monograph accessible to a larger audience while, at the same time, provides a full
treatment of the topological and functional analysis background accompanying the
theory of distributions for the reader interested in a more in-depth treatment.
Finally, each book chapter ends with bibliographical references tailored to its
respective contents under the heading Further Notes, as well as with a number of
additional exercises, selectively solved in Chapter 13.
Common Notational Conventions
Throughout this book the set of natural numbers will be denoted by N, that is
N :¼ f1; 2; . . .g, while N0 :¼ N [ f0g. For each k 2 N set k! :¼ 1 2
ðk 1Þ k, and make the convention that 0! :¼ 1. The letter C will denote the set
of complex numbers, and z denotes the complex conjugate of z 2 C. Also the real
and imaginary parts of a complex number z are denoted by Re z and Im z,
pffiffiffiffiffiffiffi
respectively. The symbol i is reserved for the complex imaginary unit 1 2 C.
The letter R will denote the set of real numbers and its n-fold Cartesian product of
R with itself (where n 2 N) is denoted by Rn . That is,
Rn :¼ fx ¼ ðx1 ; . . .; xn Þ : x1 ; . . .; xn 2 Rg ð3Þ
considered with the usual vector space and inner product structure, i.e.,
P
n
x þ y :¼ ðx1 þ y1 ; . . .; xn þ yn Þ; cx :¼ ðcx1 ; . . .; cxn Þ; x y :¼ xj yj ;
j¼1 ð4Þ
8 x ¼ ðx1 ; . . .; xn Þ 2 Rn ; 8 y ¼ ðy1 ; . . .; yn Þ 2 Rn ; 8 c 2 R:
The standard orthonormal basis of vectors in Rn is denoted by fej g1 j n , where we

have set ej :¼ ð0; . . .; 0; 1; 0; . . .; 0Þ 2 Rn with the only nonzero component on the
j-th slot. We shall also consider the two canonical (open) half-spaces of Rn , denoted
by Rn :¼ fx ¼ ðx1 ; . . .; xn Þ 2 Rn : xn [ 0g. Hence, Rnþ ¼ Rn1 ð0; 1Þ and
Rn ¼ Rn1 ð1; 0Þ.
Given a multi-index fi ¼ ðfi1 ; . . .; fin Þ 2 Nn0 , we set
supp fi :¼ fj 2 f1; . . .; ng : fij 6¼ 0g; ð5Þ
X
n
fi! :¼ fi1 !fi2 ! fin ! and jfij :¼ fij ; ð6Þ
j¼1
xxi
xxii Common Notational Conventions
Y fij @
@ fi :¼ @j where @j :¼ for j ¼ 1; . . .; n; ð7Þ
j2supp fi
@xj
Y fi
xfi :¼ xj j for every x ¼ ðx1 ; . . .; xn Þ 2 Cn ; ð8Þ
j2supp fi
with the convention that @ ð0;...;0Þ is the identity operator and xð0;...;0Þ :¼ 1. Also if
fl ¼ ðfl1 ; . . .; fln Þ 2 Nn0 is another multi-index we shall write fl fi provided fl j fij
for each j 2 f1; . . .; ng, in which case we set fi fl :¼ ðfi1 fl 1 ; . . .; fin fln Þ. We
shall also say that fl\fi if fl fi and fl 6¼ fi. Recall that the Kronecker symbol is
defined by djk :¼ 1 if j ¼ k and djk :¼ 0 if j 6¼ k.
All functions in this monograph are assumed to be complex-valued unless
otherwise indicated. Derivatives of a function f defined on the real line are going to
k
be denoted using f 0 , f 00 , etc., or f ðkÞ , or ddxfk .
Throughout the book, X denotes an arbitrary open subset of Rn . If A is an
arbitrary subset of Rn , then A, A, and @A denote its interior, its closure, and its
boundary, respectively. In addition, if B is another arbitrary subset of Rn , then their
set theoretic difference is denoted by AnB :¼ fx 2 A : x 62 Bg. In particular, the
complement of A is Ac :¼ Rn nA. For any E Rn we let vE stand for the charac-
teristic function of the set E (i.e., vE ðxÞ ¼ 1 if x 2 E and vE ðxÞ ¼ 0 if x 2 Rn nE).
For k 2 N0 [ f1g, we will work with the following classes of functions that are
vector spaces over C:
C k ðXÞ :¼ fu : X ! C : @ fi u continuous 8 fi 2 Nn0 ; jfij kg; ð9Þ
Ck ðXÞ :¼ fujX : u 2 Ck ðUÞ; U Rn open set containing X g; ð10Þ
C0k ðXÞ :¼ fu 2 C k ðXÞ : supp u compact subset of Xg: ð11Þ
As usual, for any Lebesgue-measurable (complex-valued) function f defined on a

Lebesgue-measurable set E Rn and any p 2 ½1; 1
we write
( R
ð E j f jp dxÞ1=p if 1 p \ 1;
k f kLp ðEÞ :¼ ð12Þ
ess-sup E j f j if p ¼ 1;
and denote by Lp ðEÞ the Banach space of (equivalence classes of) Lebesgue-
measurable functions f on E satisfying k f kLp ðEÞ \1. Also, we will work with
locally integrable functions and with compactly supported integrable functions. For
p 2 ½1; 1
these are defined as
Common Notational Conventions xxiii
Lploc ðXÞ :¼ ff : X ! C : f Lebesgue measurable such that

ð13Þ
k f kLp ðKÞ \ 1; 8K X compact setg;
and, respectively, as
Lpcomp ðXÞ :¼ f f 2 Lp ðXÞ : supp f compact subset of Xg; ð14Þ
where supp f is defined in (2.5.12).

Given a measure space ðX; lÞ, a measurable set A X with 0\lðAÞ\1, and a
function f 2 L1 ðA; lÞ, we define the integral average of f over A by
Z Z
1
– f dl : ¼ f dl: ð15Þ
A lðAÞ A
If E is a Lebesgue-measurable subset of Rn , the Lebesgue measure of E is

denoted by jEj. If x 2 Rn and radius r [ 0, set Bðx; rÞ :¼ fy 2 Rn : jy xj \ rg
for the ball of center x and radius R and its boundary is denoted by @Bð0; rÞ. The
unit sphere in Rn centered at zero is S n1 :¼ fx 2 Rn : jxj ¼ 1g ¼ @Bð0; 1Þ and its
surface measure is denoted by xn1 .
For n; m 2 N and R an arbitrary commutative ring with multiplicative unit we
denote by Mnm ðRÞ the collection of all n m matrices with entries from R. If
B 2 Mnm ðRÞ, then BT denotes its transpose and if n ¼ m, then det B denotes the
determinant of the matrix B, while Inn denotes the identity matrix.
Regarding semi-orthodox notational conventions, A :¼ B stands for “A is
defined as being B”, while A ¼: B stands for “B is defined as being A”. Also, the
letter C when used as a multiplicative constant in various inequalities, is allowed to
vary from line to line. Whenever necessary, its dependence on the other parameters
a; b; . . . implicit in the estimate in question is stressed by writing Cða; b; . . .Þ or
Ca;b;... in place of just C.
Chapter 1
Weak Derivatives
Abstract Starting from the discussion of the Cauchy problem for a vibrating infinite
string as a motivational example, the notion of a weak derivative is introduced as
a mean of extending the notion of solution to a more general setting, where the
functions involved may lack standard pointwise differentiability properties. Here
two classes of test functions are also defined and discussed.
1.1 The Cauchy Problem for a Vibrating Infinite String
The partial differential equation
∂21 u − ∂22 u = 0 in R2 (1.1.1)
was derived by Jean d’Alembert in 1747 to describe the displacement u(x1 , x2 ) of a

violin string as a function of time and distance along the string. Assuming that the
string is infinite and that at time x2 = 0 the displacement is given by some function
ϕ0 ∈ C 2 (R) leads to the following global Cauchy problem
⎧
⎪
⎪
⎪ u ∈ C 2 (R2 ),
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎨∂1 u − ∂2 u = 0 in R ,
⎪ 2 2 2
⎪
⎪
⎪
(1.1.2)
⎪
⎪
⎪ u(·, 0) = ϕ in R,
⎪
⎪
⎪
0
⎪
⎪
⎩(∂2 u)(·, 0) = 0 in R.
Thanks to the regularity assumption on ϕ0 , it may be checked without difficulty that

the function
u(x1 , x2 ) := 12 [ϕ0 (x1 + x2 ) + ϕ0 (x1 − x2 )], for every (x1 , x2 ) ∈ R2 , (1.1.3)
© Springer Nature Switzerland AG 2018 1

D. Mitrea, Distributions, Partial Differential Equations, and Harmonic Analysis,
Universitext, https://doi.org/10.1007/978-3-030-03296-8 1
2 1 Weak Derivatives
is a solution of (1.1.2). This being said, the expression of u in (1.1.3) continues to

be meaningful under much less restrictive assumptions on ϕ0 . For example, u is a
well-defined continuous function in R2 whenever ϕ0 ∈ C 0 (R). While in this case
expression (1.1.3) is no longer a classical solution of (1.1.2), it is natural to ask
whether it is possible to identify a new (and possibly weaker) sense in which (1.1.3)
would continue to satisfy ∂21 u − ∂22 u = 0.
To answer this question, fix a function u ∈ C 2 (R2 ) satisfying ∂21 u − ∂22 u = 0
pointwise in R2 . If ϕ ∈ C0∞ (R2 ) is an arbitrary function and R ∈ (0, ∞) is a number
such that supp ϕ ⊆ (−R, R) × (−R, R), then integration by parts gives

0= (∂21 u − ∂22 u)ϕ dx
R2
R R
= ∂21 u(x1 , x2 )ϕ(x1 , x2 ) dx1 dx2
−R −R
R R
− ∂22 u(x1 , x2 )ϕ(x1 , x2 ) dx1 dx2
−R −R

= (∂21 ϕ − ∂22 ϕ)u dx. (1.1.4)
R2

Note that the condition R2 (∂21 ϕ − ∂22 ϕ)u dx = 0 for all ϕ ∈ C0∞ (R2 ) is meaningful
even if u ∈ C 0 (R2 ), which suggests the following definition.
Definition 1.1. A function u ∈ C 0 (R2 ) is called a weak (generalized) solution

of the equation ∂21 u − ∂22 u = 0 in R2 if

(∂21 ϕ − ∂22 ϕ)u dx = 0 for all ϕ ∈ C0∞ (R2 ). (1.1.5)
R2
Returning to (1.1.3), let us now check that, under the assumption ϕ0 ∈ C 0 (R),
the function u defined in (1.1.3) is a generalized solution of ∂21 u − ∂22 u = 0 in R2 .
Concretely, fix ϕ ∈ C0∞ (R2 ) and write

(∂21 ϕ − ∂22 ϕ)u dx (1.1.6)
R2

1
= ∂21 ϕ(x1 , x2 ) − ∂22 ϕ(x1 , x2 ) ϕ0 (x1 + x2 ) + ϕ0 (x1 − x2 ) dx1 dx2

2 R R
y + y y − y
1 1 2 1 2
= (∂21 ϕ) , ϕ0 (y1 ) + ϕ0 (y2 ) dy1 dy2

4 R R 2 2

1 y1 + y2 y1 − y2
− (∂22 ϕ) , ϕ0 (y1 ) + ϕ0 (y2 ) dy1 dy2 ,

4 R R 2 2
where for the last equality in (1.1.6) we have made the following change of vari-
ables: y1 = x1 + x2 and y2 = x1 − x2 . If we now let ψ(y1 , y2 ) := ϕ y1 +y2 y1 −y2
2 , 2 for
1.1 The Cauchy Problem for a Vibrating Infinite String 3
(y1 , y2 ) ∈ R2 , then
∂1 ψ(y1 , y2 ) (1.1.7)
1 y + y y − y 1 y + y y − y
1 2 1 2 1 2 1 2
= (∂1 ϕ) , + (∂2 ϕ) ,
2 2 2 2 2 2
and
∂2 ∂1 ψ(y1 , y2 ) (1.1.8)
1 2 y1 + y2 y1 − y2 1 y + y y − y
1 2 1 2
= (∂1 ϕ) , − (∂2 ∂1 ϕ) ,
4 2 2 4 2 2
1 y + y y − y 1 y + y y − y
1 2 1 2 1 2 1 2
+ (∂1 ∂2 ϕ) , − (∂22 ϕ) ,
4 2 2 4 2 2
which, when used in (1.1.6), give

(∂21 ϕ − ∂22 ϕ)u dx = ∂1 ∂2 ψ(y1 , y2 )[ϕ0 (y1 ) + ϕ0 (y2 )] dy1 dy2 . (1.1.9)
R2 R R
Let R ∈ (0, ∞) be such that supp ϕ ⊂ (−R, R) × (−R, R). Then the support of ψ is
contained in the set of points (y1 , y2 ) ∈ R2 satisfying −2R ≤ y1 + y2 ≤ 2R and
−2R ≤ y1 − y2 ≤ 2R. Hence, if R > 2R we have supp ψ ⊂ (−R , R ) × (−R , R ) and
integration by parts yields

(∂21 ϕ − ∂22 ϕ)u dx
R2
R R
= ϕ0 (y1 ) ∂2 ∂1 ψ(y1 , y2 ) dy2 dy1
−R −R
R R
+ ϕ0 (y2 ) ∂1 ∂2 ψ(y1 , y2 ) dy1 dy2
−R −R
R

= ϕ0 (y1 ) ∂1 ψ(y1 , R ) − ∂1 ψ(y1 , −R ) dy1
−R
R

+ ϕ0 (y2 ) ∂2 ψ(R , y2 ) − ∂2 ψ(−R , y2 ) dy2 = 0. (1.1.10)
−R
In summary, this proves that
for every ϕ0 ∈ C 0 (R) the function u defined as in (1.1.3)

(1.1.11)
is a weak solution of the equation ∂21 u − ∂22 u = 0 in R2 .
We emphasize that in general, there is no reason to expect that u has any pointwise
differentiability properties if ϕ0 is merely continuous.
1.2 Weak Derivatives
Convention. Unless otherwise specified, Ω denotes an arbitrary open subset of Rn .

Before proceeding with the definition of weak (Sobolev) derivatives we discuss
a phenomenon that serves as motivation for the definition. Let f ∈ C m (Ω), where
m ∈ N0 . Given an arbitrary ϕ ∈ C0∞ (Ω), consider a function F ∈ C0m (Rn ) that agrees
with f in a neighborhood of supp ϕ. For example, we may take F = ψ f , where
tilde denotes the extension by zero outside Ω and ψ ∈ C0∞ (Rn ) is identically one
in a neighborhood of supp ϕ (see Proposition 14.34 for the construction of such
a function). Also, pick R > 0 large enough so that supp ϕ ⊂ B(0, R). Then for
each α ∈ Nn0 with |α| ≤ m, integration by parts (cf. Theorem 14.60 for a precise
formulation) and support considerations yield

(∂α f )ϕ dx = (∂α F)ϕ dx
Ω B(0,R)

= (−1)|α| F ∂α ϕ dx = (−1)|α| f ∂α ϕ dx. (1.2.1)
B(0,R) Ω
This computation suggests the following definition.

Definition 1.2. If f ∈ Lloc
1
(Ω) and α ∈ Nn0 , we say that ∂α f belongs to Lloc
1
(Ω) in a
weak (Sobolev) sense provided there exists some g ∈ Lloc (Ω) with the property
1
that
|α|
gϕ dx = (−1) f ∂α ϕ dx for every ϕ ∈ C0∞ (Ω). (1.2.2)
Ω Ω
Whenever this happens, we shall write ∂α f = g and call g the weak derivative
of order α of f .
The fact that the concept of weak derivative is unambiguously defined is then
ensured by the next theorem.

Theorem 1.3. If g ∈ Lloc
1
(Ω) and Ω gϕ dx = 0 for each ϕ ∈ C0∞ (Ω) then g = 0
almost everywhere on Ω.
Proof. Consider a function φ satisfying (see (14.3.3) for a concrete example)
φ ∈ C0∞ (Rn ), φ ≥ 0, supp φ ⊆ B(0, 1),

(1.2.3)
and φ(x) dx = 1,
Rn
and for each ε > 0 define

1 x
φε (x) := φ for each x ∈ Rn . (1.2.4)
εn ε
Then for each ε > 0 we have
1.2 Weak Derivatives 5
φε ∈ C0∞ (Rn ), φε ≥ 0, supp φε ⊆ B(0, ε),

(1.2.5)
and φε (x) dx = 1.
Rn

Fix now x ∈ Ω and ε ∈ 0, dist(x, ∂Ω) . Then B(x, ε) ⊆ Ω which, in light of (1.2.5),
implies φε (x − ·)Ω ∈ C0∞ (Ω). Consequently, under the current assumptions on g we
have
g(y)φε (x − y) dy = 0. (1.2.6)
Ω
In particular, if we also assume that x is a Lebesgue point for g (i.e., the limit
(14.2.11) holds for this x with f replaced by g), then (1.2.6) combined with the
properties of φε allow us to write

|g(x)| = g(x)φε (x − y) dy − g(y)φε (x − y) dy
Ω Ω
y
1
≤ |g(x) − g(y)|φ dy
εn B(x,ε) ε

c
≤ |g(x) − g(y)| dy −→+ 0, (1.2.7)
|B(x, ε)| B(x,ε) ε→0
where c := ωn−1
n
φ
L (R ) . The convergence in (1.2.7) is due to Lebesgue’s Differen-
∞ n
tiation Theorem (cf. Theorem 14.14). This proves that g(x) = 0 for every x ∈ Ω that
is a Lebesgue point for g, hence g = 0 almost everywhere in Ω.
Next we present a few examples related to the notion of weak (Sobolev) deriva-
tive.
Example 1.4. Consider the function

⎧
⎪
⎪
⎪ x > 0,
⎨ x,
f : R −→ R, f (x) := ⎪
⎪ ∀ x ∈ R. (1.2.8)
⎪
⎩0, x ≤ 0,
Note that f is continuous on R but not differentiable at 0. Nonetheless, f has a weak

derivative of order one that is equal to the Heaviside function
⎧
⎪
⎪
⎨1, x > 0,
⎪
H : R −→ R, H(x) := ⎪ ⎪ ∀ x ∈ R. (1.2.9)
⎪
⎩0, x ≤ 0,
Indeed, if ϕ ∈ C0∞ (R), then integration by parts yields

∞ ∞ ∞ ∞
− f (x)ϕ (x) dx = − xϕ (x) dx = −xϕ(x) + ϕ(x) dx
−∞ 0 0 0
∞
= H(x)ϕ(x) dx, (1.2.10)
−∞
which shows that

Hϕ dx = − f ϕ dx, ∀ ϕ ∈ C0∞ (R). (1.2.11)
R R
Note that H ∈ Lloc

1
(R) hence H is the weak (or Sobolev) derivative of order one of
the function f in R.
Example 1.5. Does there exist a function g ∈ Lloc

1
(R) such that g is a weak derivative
(of order one) of the Heaviside function? To answer this question first observe that
for each ϕ ∈ C0∞ (R) we have
∞ ∞
− H(x)ϕ (x) dx = − ϕ (x) dx = ϕ(0). (1.2.12)
−∞ 0
Suppose that H has a weak derivative of order one, and call this g ∈ Lloc
1
(R). Then,
∞
by Definition 1.2 and (1.2.12), we have R g ϕ dx = ϕ(0) for all ϕ ∈ C0 (R). This
∞
forces 0 gϕ dx = 0 for all ϕ ∈ C0∞ (R \ {0}). In concert with Theorem 1.3, the latter
yields g = 0 almost everywhere on R \ {0}. When combined with (1.2.12), this gives
that 0 = R gϕ dx = ϕ(0) for all ϕ ∈ C0∞ (R), leading to a contradiction (as there are
functions ϕ ∈ C0∞ (R) with ϕ(0) 0). Thus, a weak (Sobolev) derivative of order
one of H does not exist.
Having defined the notion of weak (Sobolev) derivatives for locally integrable
functions, we return to the notion of weak solution considered in Definition 1.1 in a
particular case, and extend this to more general partial differential equations. To set
the stage, let P(x, ∂) be a linear partial differential operator of order m ∈ N of the
form

P(x, ∂) := aα (x)∂α , aα ∈ C |α| (Ω), α ∈ Nn0 , |α| ≤ m. (1.2.13)
|α|≤m
Also, suppose f ∈ C 0 (Ω) is a given function and that u is a classical solution of the
partial differential equation P(x, ∂)u = f in Ω. That is, assume u ∈ C m (Ω) and the
equation holds pointwise in Ω. Then, for each ϕ ∈ C0∞ (Ω), integration by parts gives

f ϕ dx = ϕ aα (∂α u) dx = (−1)|α| ∂α (aα ϕ) u dx. (1.2.14)
Ω |α|≤m Ω Ω |α|≤m
Hence, if we define
P (x, ∂)ϕ := (−1)|α| ∂α (aα ϕ), (1.2.15)
|α|≤m
and call it the transpose of the operator P(x, ∂), the resulting equation from
(1.2.14) becomes

f ϕ dx = [P (x, ∂)ϕ]u dx, ∀ ϕ ∈ C0∞ (Rn ). (1.2.16)
Ω Ω
1.3 The Spaces E(Ω) and D(Ω) 7
Thus, any classical solution u of P(x, ∂)u = f in Ω satisfies (1.2.16). On the other
hand, there might exist functions u ∈ Lloc
1
(Ω) that satisfy (1.2.16) but are not clas-
sical solutions of the given equation. Such a scenario has been already encountered
in (1.1.11) (cf. also the subsequent comment). This motivates the following general
definition (compare with Definition 1.1 corresponding to P(x, ∂) = ∂21 − ∂22 ).
Definition 1.6. Let u, f ∈ Lloc

1
(Ω) be given and assume that P(x, ∂) is as in (1.2.13).
Then P(x, ∂)u = f is said to hold in the weak (or Sobolev) sense if

f ϕ dx = [P (x, ∂)ϕ]u dx, ∀ ϕ ∈ C0∞ (Ω). (1.2.17)
Ω Ω
From the comments in the preamble to Definition 1.6 we know that if u is a

classical solution of P(x, ∂)u = f in Ω for some f ∈ C 0 (Ω), then u is also a weak
(Sobolev) solution of the same equation. Conversely, if u ∈ C m (Ω) is a weak solution
of the partial differential equation P(x, ∂)u = f for some given function f ∈ C 0 (Ω),
then by Definition 1.6 and integration by parts we obtain

f ϕ dx = [P (x, ∂)ϕ]u dx = ϕ P(x, ∂)u dx. (1.2.18)
Ω Ω Ω
Since ϕ ∈ C0∞ (Ω) is arbitrary, Theorem 1.3 then forces f = P(x, ∂)u almost every-
where in Ω, hence ultimately everywhere in Ω, since the functions in question are
continuous.
In summary, the above discussion shows that the notion of weak solution of a
partial differential equation is a natural, unambiguous, and genuine generalization
of the concept of classical solution, in the following precise sense:
• any classical solution is a weak solution,
• any sufficiently regular weak solution is classical,
• weak solutions may exist even in the absence of classical ones.
1.3 The Spaces E(Ω) and D(Ω)
A major drawback of Definition 1.2 is that while the right-hand side of (1.2.2) is
always meaningful, it cannot always be written it in the form given by the left-
hand side of (1.2.2). In addition, it might be the case that some locally integrable
function in Ω may admit weak (Sobolev) derivatives of a certain order and not of
some intermediate lower order (see the example in Exercise 1.29). The remedy is
to focus on the portion of (1.2.2) that always makes sense. Specifically, given f ∈
1
Lloc (Ω) and α ∈ Nn0 , define the mapping

gα : C0∞ (Ω) → C, gα (ϕ) := (−1)|α| f (∂α ϕ) dx, ∀ ϕ ∈ C0∞ (Ω). (1.3.1)
Ω
The functional in (1.3.1) has the following properties.

1. gα is linear, i.e., gα (λ1 ϕ1 + λ2 ϕ2 ) = λ1 gα (ϕ1 ) + λ2 gα (ϕ2 ) for every λ1 , λ2 ∈ C, and
every ϕ1 , ϕ2 ∈ C0∞ (Ω).
2. For each ϕ ∈ C0∞ (Ω) we may estimate

|gα (ϕ)| ≤ | f ||∂α ϕ| dx ≤ | f | dx sup |∂α ϕ(x)|. (1.3.2)
Ω supp ϕ x∈suppϕ

The fact that the term supp ϕ | f | dx in (1.3.2) depends on ϕ is inconvenient if we want
to consider the continuity of gα in some sense. Nonetheless, if a priori a compact
set K ⊂ Ω is fixed and the requirement supp ϕ ⊆ K is imposed, then (1.3.2) becomes

|gα (ϕ)| ≤ | f | dx sup |∂α ϕ(x)| (1.3.3)
K x∈K

and, this time, K | f | dx is a constant independent of ϕ. This observation motivates
considering an appropriate topology τ on C ∞ (Ω). For the exact definition of this
topology see Section 14.1. We will not elaborate here more on this subject other
than highlighting those key features of τ that are particularly important for our future
investigations. To record the precise statements of these features, introduce

E(Ω) := C ∞ (Ω), τ , (1.3.4)
a notation which emphasizes that E(Ω) is the vector space C ∞ (Ω) equipped with the
topology τ. We then have:
Fact 1.7 A sequence {ϕ j } j∈N ⊂ C ∞ (Ω) converges in E(Ω) to some ϕ ∈ C ∞ (Ω) as
j → ∞ if and only if

∀ K ⊂ Ω compact, ∀ α ∈ Nn0 , we have lim sup ∂α (ϕ j − ϕ)(x) = 0, (1.3.5)
j→∞ x∈K
E(Ω)
in which case we use the notation ϕ j −−−−→ ϕ.
j→∞
Fact 1.8 E(Ω) is a locally convex, metrizable, and complete topological vector
space over C.
It is easy to see that as a consequence of Fact 1.7 we have the following result.
Remark 1.9. A sequence {ϕ j } j∈N ⊂ C ∞ (Ω) converges in E(Ω) to a function ϕ ∈
C ∞ (Ω) as j → ∞, if and only if for any compact set K ⊂ Ω and any m ∈ N0 one has

lim sup sup ∂α (ϕ j − ϕ)(x) = 0. (1.3.6)
j→∞ α∈Nn , |α|≤m x∈K
0
E(Ω)
Exercise 1.10. Prove that if ϕ j −−−−→ ϕ then the following also hold:
j→∞
E(Ω)
(1) ∂α ϕ j −−−−→ ∂α ϕ for each α ∈ Nn0 ;
j→∞
E(Ω)
(2) a ϕ j −−−−→ a ϕ for each a ∈ C0∞ (Ω).
j→∞
A standard way of constructing a sequence of smooth functions in Rn that con-

verges in E(Rn ) to a given f ∈ C ∞ (Rn ) is by taking the convolution of f with
dilations of a function as in (1.2.3). This construction is discussed in detail next.
Example 1.11. Let f ∈ C ∞ (Rn ) be given. Then a sequence of functions from C ∞ (Rn )
that converges to f in E(Rn ) may be constructed as follows. Recall φ from (1.2.3)
and define
φ j (x) := jn φ( jx) for x ∈ Rn and each j ∈ N. (1.3.7)
Clearly, for each j ∈ N we have

φj ∈ C0∞ (Rn ), supp φ j ⊆ B(0, 1/ j), and φ j dx = 1. (1.3.8)
Rn
Now if we further set for each j ∈ N

f j (x) := f (x − y)φ j (y) dy
Rn

= f (x − z/ j)φ(z) dz for each x ∈ Rn , (1.3.9)
B(0,1)
then f j ∈ C ∞ (Rn ). Also, if K is an arbitrary compact set in Rn and α ∈ Nn0 , then

α α
|∂ f j (x) − ∂ f (x)| ≤ |∂α f (x − z/ j) − ∂α f (x)| φ(z) dz
B(0,1)
1
≤ max
∂β f
L∞ (K)
∀ x ∈ K, (1.3.10)
j |β|=|α|+1
E(Rn )
:= {x ∈ Rn : dist (x, K) ≤ 1}. Hence f j −−−−→ f , as desired.
where K
j→∞
The previous approximation result may be further strengthened as indicated in

the next two exercises.
Exercise 1.12. Prove that C0∞ (Rn ) is sequentially dense in E(Rn ). That is, show that
for every f ∈ C ∞ (Rn ) there exists a sequence of functions { f j } j∈N from C0∞ (Rn ) with
E(Rn )
the property that f j −−−−→ f .
j→∞
Hint: Let ψ ∈ C0∞ (Rn ) be such that ψ(x) = 1 whenever |x| < 1. Then given f ∈
C ∞ (Rn ) define f j (x) := ψ(x/ j) f (x), for every x ∈ Rn and every j ∈ N.
Exercise 1.13. Prove that C0∞ (Ω) is sequentially dense in E(Ω).

Hint: Consider the sequence of compacts

K j := x ∈ Ω : dist(x, ∂Ω) ≥ 1j ∩ B(0, j), ∀ j ∈ N. (1.3.11)

Then K j = Ω and K j ⊂ K̊ j+1 for every j ∈ N. For each j ∈ N pick a func-
j∈N
tion ψ j ∈ C0∞ (Ω) with ψ j ≡ 1 in a neighborhood of K j and supp ψ j ⊆ K j+1 (cf.
Proposition 14.34). If f ∈ C ∞ (Ω), define f j := ψ j f for every j ∈ N.
Moving on, we focus on defining a topology on C0∞ (Ω) that suits the purposes we
have in mind. Since C0∞ (Ω) ⊂ E(Ω), one option would be to consider the topology
induced by this larger ambient on C0∞ (Ω). However, this topology has the distinct
drawback of not preserving the property of being compactly supported under con-
vergence. Here is an example to that effect.
Example 1.14. Consider the function
⎧
⎪
⎪
⎪ 1
⎪ x− 1 − 1
2
⎪
⎨ 2 4
ϕ(x) := ⎪
⎪ e if 0 < x < 1, for each x ∈ R. (1.3.12)
⎪
⎪
⎪
⎩0 if x ≤ 0 or x > 1,
Note that ϕ ∈ C ∞ (R), supp ϕ = [0, 1], and ϕ > 0 in (0, 1). For each j ∈ N define
ϕ j (x) := ϕ(x − 1) + 12 ϕ(x − 2) + · · · + 1j ϕ(x − j), ∀ x ∈ R. (1.3.13)
E(R)
Then ϕ j ∈ C ∞ (R), supp ϕ j = [1, j + 1], and ϕ j −−−−→ ϕ where
j→∞

∞
1
ϕ(x) := jϕ x −j for each x ∈ R. (1.3.14)

j=1
Clearly this limit function does not have compact support.

The flaw just highlighted is remedied by introducing a different topology on
C0∞ (Ω) that is finer than the one inherited from E(Ω). First, for each K ⊂ Ω compact,
denote by DK (Ω) the vector space consisting of functions from C ∞ (Ω) supported in
K endowed with the topology induced by E(Ω). Second, consider on C0∞ (Ω) the

inductive limit topology of the spaces DK (Ω) K⊂Ω and denote the resulting topo-
compact
logical vector space by D(Ω). For precise definitions see Section 14.1. The topology
induced on {ϕ ∈ C0∞ (Ω) : supp ϕ ⊆ K} by this inductive limit topology coincides
with the topology on DK (Ω). Two features that are going to be particularly impor-
tant for our analysis are singled out below (see Section 14.1.0.4 in this regard).
Fact 1.15 D(Ω) is a locally convex and complete topological vector space over C.
Fact 1.16 A sequence {ϕ j } j∈N ⊂ C0∞ (Ω) converges in D(Ω) to some ϕ ∈ C0∞ (Ω) as
j → ∞ if and only if the following two conditions are satisfied:
(1) there exists a compact set K ⊂ Ω such that supp ϕ j ⊆ K for all j ∈ N and
supp ϕ ⊆ K;
(2) for any α ∈ Nn0 we have lim sup x∈K ∂α (ϕ j − ϕ)(x) = 0.
j→∞
D(Ω)
We abbreviate (1)–(2) by simply writing ϕ j −−−−→ ϕ.
j→∞
In view of Fact 1.7 one obtains the following consequence of Fact 1.16.
D(Ω)
Remark 1.17. ϕ j −−−−→ ϕ if and only if
j→∞
(1) there exists a compact set K ⊂ Ω such that supp ϕ j ⊆ K for all j ∈ N, and
E(Ω)
(2) ϕ j −−−−→ ϕ.
j→∞
If one now considers the identity map from D(Ω) into E(Ω), a combination of
Remark 1.17, and Theorem 14.6 yields that this map is continuous. Hence, if we
also take into account Exercise 1.12, it follows that
D(Ω) is continuously and densely embedded into E(Ω). (1.3.15)
Exercise 1.18. Suppose ω is an open subset of Ω and consider the map

⎧
⎪
⎪
∞ ∞ ⎨ϕ on ω,
⎪
ι : C0 (ω) → C0 (Ω), ι(ϕ) := ⎪
⎪ ∀ ϕ ∈ C0∞ (ω). (1.3.16)
⎪
⎩0 on Ω \ ω,
D(ω) D(Ω)
Prove that if ϕ j −−−−→ ϕ then ι(ϕ j ) −−−−→ ι(ϕ). Use Theorem 14.6 to conclude that
j→∞ j→∞
ι : D(ω) → D(Ω) is continuous.
Exercise 1.19. Let x0 ∈ Rn and consider the translation by x0 map defined as
t x0 : D(Rn ) −→ D(Rn )
(1.3.17)
t x0 (ϕ) := ϕ(· − x0 ), ∀ ϕ ∈ C0∞ (Rn ).
Prove that t x0 is linear and continuous.
Hint: Use Theorem 14.6.

D(Ω)
Exercise 1.20. Prove that if ϕ j −−−−→ ϕ then the following also hold:
j→∞
D(Ω)
(1) ∂α ϕ j −−−−→ ∂α ϕ for each α ∈ Nn0 ;
j→∞
D(Ω)
(2) a ϕ j −−−−→ a ϕ for each a ∈ C ∞ (Ω).
j→∞
Exercise 1.21. Prove that the map D(Ω) ϕ → aϕ ∈ D(Ω) is linear and continuous
for every a ∈ C ∞ (Ω).
Hint: Use Exercise 1.20 and Theorem 14.6.

E(Ω) D(Ω)
Exercise 1.22. Prove that if ϕ j −−−−→ ϕ then a ϕ j −−−−→ a ϕ for each a ∈ C0∞ (Ω).
j→∞ j→∞
E(Ω)
Also show that if ϕ j −−−−→ ϕ and a ∈ C0∞ (ω) for some open subset ω of Ω, then
j→∞
D(ω)
a ϕ j −−−−→ a ϕ.
j→∞
As a consequence of Remark 1.17, we see that the topology D(Ω) is finer than
the topology C0∞ (Ω) inherits from E(Ω), and an example of a sequence of smooth,
compactly supported functions in Ω convergent in E(Ω) to a limit which does not be-
long to D(Ω) has been given in Example 1.14. The example below shows that even
if the limit function is in D(Ω), one should still not expect that convergence in E(Ω)
of a sequence of smooth, compactly supported functions in Ω implies convergence
in D(Ω).
Example 1.23. Let ϕ be as in (1.3.12) and for each j ∈ N set ϕ j (x) := ϕ(x− j), x ∈ R.
Clearly, ϕ j ∈ C ∞ (R) and supp ϕ j = [ j, j + 1] for all j ∈ N. If K ⊂ R is compact,
then there exists j0 ∈ N such that K ⊆ [− j0 , j0 ]. Consequently, supp ϕ j ∩ K = ∅

for j ≥ j . Thus, trivially, sup ϕ(k) (x) = 0 if j ≥ j which shows that ϕ −−−−→ 0.
E(R)
0 j 0 j
x∈K j→∞
Consider next the issue whether {ϕ j } j∈N converge in D(R). If this were to be the
case, there would exist r ∈ (0, ∞) such that supp ϕ j ⊆ [−r, r] for every j. However,
∞
supp ϕ j = [1, ∞) which leads to a contradiction. Thus, {ϕ j } j∈N does not converge
j=1
in D(R).
For n, m ∈ N, denote by Mn×m (R) the collection of all n × m matrices with

entries in R. Recall that a map L : Rm → Rn is linear if and only if there exists a
matrix A ∈ Mn×m (R) such that L(x) = Ax for every x ∈ Rm , where Ax denotes the
multiplication of the matrix A with the vector x viewed as an element in Mm×1 (R).
Moreover, such a matrix is unique. In the sequel, we follow the standard practice
of denoting by A the linear map associated with a matrix A. It is well-known that
a linear mapping A ∈ Mn×n (R) is invertible if and only if it is open. The following
lemma shows that the composition to the right with an invertible matrix defines a
linear and continuous mapping from D(Rn ) into itself.
Lemma 1.24. Suppose A ∈ Mn×n (R) is such that det A 0. Then the composition
mapping
D(Rn ) ϕ → ϕ ◦ A ∈ D(Rn )
(1.3.18)
is well defined, linear and continuous.
Proof. Let ϕ ∈ C0∞ (Rn ). By the Chain Rule we have ϕ ◦ A ∈ C ∞ (Rn ). We claim that

supp(ϕ ◦ A) = x ∈ Rn : Ax ∈ supp ϕ . (1.3.19)
Indeed, if x ∈ Rn and x supp(ϕ ◦ A) then there exists r > 0 such that ϕ ◦ A = 0

on B(x, r). Hence, the open set O := A(B(x, r)) contains Ax and ϕ = 0 on O, which
implies O ∩ supp ϕ = ∅. In particular x does not belong to the set in the right-
hand side of (1.3.19). Conversely, if x ∈ Rn is such that Ax supp ϕ, then there
exists r > 0 such that ϕ = 0 on B(Ax, r). Since A−1 (B(Ax, r)) is open, contains
x, and ϕ ◦ A vanishes identically on it, we conclude that x does not belong to the
set in the left-hand side of (1.3.19). The proof of the claim is finished. Moreover,

x ∈ Rn : Ax ∈ supp ϕ = A−1 (supp ϕ), so (1.3.19) may be ultimately recast as
supp(ϕ ◦ A) = A−1 (supp ϕ), ∀ ϕ ∈ C0∞ (Rn ). (1.3.20)
Given that ϕ has compact support and A−1 is continuous, from (1.3.20) we see that
ϕ ◦ A has compact support. This proves that the map in (1.3.18) is well defined,
while its linearity is clear. To show that this map is also continuous, by Fact 1.15
and Theorem 14.6, matters are reduced to proving sequential continuity at 0. The
latter is now a consequence of Fact 1.16, (1.3.20), the continuity of A−1 , and the
Chain Rule.
Convention. In what follows, we will often identify a function f ∈ C0∞ (Ω) with
its extension by zero outside its support, which makes such an extension belong to
C0∞ (Rn ).
Exercise 1.25. Suppose Ω1 , Ω2 are open sets in Rn and let F : Ω1 → Ω2 be a C ∞

diffeomorphism.
(1) Prove that ϕ ◦ F ∈ C0∞ (Ω1 ) for all ϕ ∈ C0∞ (Ω2 ).
D(Ω2 ) D(Ω1 )
(2) Prove that if ϕ j −−−−→ ϕ then ϕ j ◦ F −−−−→ ϕ ◦ F.
j→∞ j→∞
(3) Prove that the map D(Ω2 ) ϕ −→ ϕ ◦ F ∈ D(Ω1 ) is linear and continuous.
(4) Prove that the map D(Ω2 ) ϕ −→ | det(DF)| ϕ ◦ F ∈ D(Ω1 ) is linear and
continuous, where DF denotes the Jacobian matrix of F.
Hint: For (1) show that supp(ϕ ◦ F) = F −1 (supp ϕ), for (2) you may use Fact 1.16
and the Chain Rule. Given Fact 1.15, (2), and Exercise 1.20, in order to prove (3)
and (4) you may apply Theorem 14.6.
Further Notes for Chapter 1. The concept of weak derivative goes back to the pioneering work of
the Soviet mathematician Sergei Lvovich Sobolev (1908–1989). Although we shall later extend the
scope of taking derivatives in a generalized sense to the larger class of distributions, a significant
portion of partial differential equations may be developed solely based on the notion of weak
derivative. For example, this is the approach adopted in [14], where distributions are avoided alto-
gether. A good reference to the topological aspects that are most pertinent to the spaces of test
functions considered here is [76], though there are many other monographs dealing with these
issues. The interested reader may consult [12], [65], [77], and the references therein.
1.4 Additional Exercises for Chapter 1
Exercise 1.26. Given ϕ0 ∈ C 2 (R), ϕ1 ∈ C 1 (R), and F ∈ C 1 (R2 ), show that the
function u : R2 → R defined by
x1 +x2

u(x1 , x2 ) := 12 (ϕ0 (x1 + x2 ) + ϕ0 (x1 − x2 ) + 12 ϕ1 (t) dt
x1 −x2
x2 x1 +(x2 −t)
− 1
2 F(ξ, t) dξ dt, ∀ (x1 , x2 ) ∈ R2 , (1.4.1)
0 x1 −(x2 −t)
is a classical solution of the problem

⎧
⎪
⎪
⎪ u ∈ C 2 (R2 ),
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎨∂1 u − ∂2 u = F in R ,
⎪ 2 2 2
⎪
⎪
⎪
(1.4.2)
⎪
⎪
⎪u(·, 0) = ϕ0 in R,
⎪
⎪
⎪
⎪
⎪
⎩(∂2 u)(·, 0) = ϕ1 in R.
Exercise 1.27. Determine

the values of a ∈ R for which the function f : R → R
x, if x ≥ a,
defined by f (x) := for each x ∈ R, has a weak derivative.
0, if x < a,

1, if x ≥ 2,
Exercise 1.28. Consider f : R → R defined by f (x) := for each
0, if x < 2,
x ∈ R. Does the function f have a weak derivative?
Exercise 1.29. Let f : R2 → R be defined by f (x, y) := H(x)+H(y) for each (x, y) ∈
R2 (where H is the Heaviside function from (1.2.9)). Prove that for α = (1, 1) and
β = (1, 0) the weak derivatives ∂α f and ∂α+β f exist, while the weak derivative ∂β f
does not exist.
Exercise 1.30. Compute the√ weak derivative of order one of f : (−1, 1) → R
defined by f (x) := sgn(x) |x| for every x ∈ (−1, 1), where
⎧
⎪
⎪
⎪ 1 if x > 0,
⎪
⎪
⎨
sgn(x) := ⎪
⎪ 0 if x = 0, (1.4.3)
⎪
⎪
⎪
⎩ −1 if x < 0.
Exercise 1.31. Let f : R2 → R be defined by f (x, y) := x|y| for each (x, y) ∈ R2 .

Prove that the weak derivative ∂21 ∂2 f exists, while the weak derivative ∂1 ∂22 f does
not.
Exercise 1.32. Let f : R2 → R be defined by f (x, y) := H(x) − sgn(y) for each
(x, y) ∈ R2 , and let α = (α1 , α2 ) ∈ N20 . Prove that ∂α f exists in the weak sense if and
only if α1 ≥ 1 and α2 ≥ 1.
1.4 Additional Exercises for Chapter 1 15
Exercise 1.33. Let f : R → R be defined by f (x) := sin |x| for every x ∈ R. Does
f exist in the weak sense? How about f ?
Exercise 1.34. Let ω, Ω be open subsets of Rn with ω ⊆ Ω. Suppose f ∈ Lloc 1
(Ω) is
α
such that ∂ f exists in the weak sense in Ω for some α ∈ N0 . Show that the weak
n
derivative ∂α ( f |ω ) exists and equals (∂α f )|ω almost everywhere in ω.
Exercise 1.35. Suppose n ∈ N, n ≥ 2 and a ∈ (0, n). Let f (x) := |x|1a for each
x ∈ Rn \ {0} and note that f ∈ Lloc
1
(Rn ). Prove that ∂ j f , j ∈ {1, . . . , n}, exists in the
weak sense if and only if a < n − 1.
Exercise 1.36. Let Ω be an open subset of Rn and let α, β ∈ N0 . Suppose f belongs
1
to Lloc (Ω) and is such that the weak derivatives ∂α f and ∂β (∂α f ) exist. Prove that
∂ f exists and equals the weak derivative ∂β (∂α f ).
α+β
Exercise 1.37. Let ε ∈ (0, 1) and consider the function f : Rn → R defined by

⎧ −ε
⎪
⎨ |x| if x ∈ R \ {0},
⎪ n
f (x) := ⎪
⎪ ∀ x ∈ Rn .
⎩1 if x = 0,
Prove that ∂ j f exists in the weak sense for each j ∈ {1, . . . , n} if and only if n ≥ 2.
Also compute the weak derivatives ∂ j f , j ∈ {1, . . . , n}, in the case when n ≥ 2.
Exercise 1.38. Assume that a, b ∈ R are such that a < b.
1
(a) Prove that if f ∈ Lloc (a, b) is such that the weak derivative f exists and is
equal to zero almost everywhere on (a, b), then there exists some complex number
c such that f = c almost everywhere on (a, b).

b
Hint: Fix ϕ0 ∈ C0∞ (a, b) with a ϕ0 (t) dt = 1. Then every ϕ ∈ C0∞ (a, b) is of the
b
form ϕ = ϕ0 a ϕ(t) dt + ψ , for some ψ ∈ C0∞ (a, b) .
x that g ∈ Lloc (a, b) and x0 ∈ (a, b). Prove that the function defined
(b) Assume
by f (x) := x g(t) dt, for x ∈ (a, b), belongs to Lloc
1
((a, b)) and has a weak derivative
0
that is equal to g almost everywhere on (a, b).
(c) Let f ∈ Lloc (a, b) be such that the weak derivative f (k) exists for some k ∈ N.
Prove that all the weak derivatives f ( j) exist for each j ∈ N with j < k.
x
Hint: Prove that if g(x) := x h(t) dt where x0 ∈ (a, b) is a fixed point and h := f (k) ,
0
and if ϕ0 is as in the hint to (a), then
b b
f (k−1) = g − g(t)ϕ0 (t) dt + (−1)k−1 f (t)ϕ(k−1)
0 (t) dt.
a a
(d) Let f ∈ Lloc (a, b) such that f (k) = 0 for some k ∈ N. Prove that there exist

k−1
a0 , a1 , . . . , ak−1 ∈ C such that f (x) = a j x j for almost every x ∈ (a, b).
j=0
(e) If Ω is an open set in Rn for n ≥ 2 and f ∈ Lloc 1

(Ω) is such that the weak
derivative ∂ f exists for some α ∈ N0 , does it follow that ∂β f exists in a weak sense
α n
for all β ≤ α?
Exercise 1.39. Let θ ∈ C0∞ (Rn ) and m ∈ N. Prove that the sequence
ϕ j (x) := e− j jm θ( jx), ∀ x ∈ Rn , j ∈ N,
converges in D(Rn ).
Exercise 1.40. Let θ ∈ C0∞ (Rn ), h ∈ Rn \ {0}, and set
ϕ j (x) := θ(x + jh), ∀ x ∈ Rn , j ∈ N.

E(Rn ) D(Rn )
Prove that ϕ j −−−−→ 0. Is it true that ϕ j −−−−→ 0?
j→∞ j→∞
Exercise 1.41. Let θ ∈ C0∞ (Rn ) be not identically zero, and for each j ∈ N define
1
ϕ j (x) := θ( jx), ∀ x ∈ Rn .
j
Prove that the sequence {ϕ j } j∈N does not converge in D(Rn ).
Exercise 1.42. Let θ ∈ C0∞ (Rn ), h ∈ Rn \ {0}. Prove that the sequence
ϕ j (x) := e− j θ( j(x − jh)), ∀ x ∈ Rn , j ∈ N,
converges in D(Rn ) if and only if θ(x) = 0 for all x ∈ Rn .
Exercise 1.43. Consider θ ∈ C0∞ (Rn ) not identically zero, and for each j ∈ N define
1 x
ϕ j (x) := θ , ∀ x ∈ Rn .
j j
Does {ϕ j } j∈N converge in D(Rn )? How about in E(Rn )?
Exercise 1.44. Suppose that {ϕ j } j∈N is a sequence of functions in C0∞ (Ω) with the
D(Rn )
property that lim Rn f (x)ϕ j (x) dx = 0 for every f ∈ Lloc
1
(Ω). Is it true that ϕ j −−−−→
j→∞ j→∞
0?
Chapter 2
The Space D (Ω) of Distributions
Abstract In this chapter the space of distributions is introduced and studied from
the perspective of a topological vector space with various other additional features,
such as the concept of support, multiplication with a smooth function, distributional
derivatives, tensor product, and a partially defined convolution product. Here the
nature of distributions with higher order gradients continuous or bounded is also
discussed.
2.1 The Definition of Distributions
Building on the idea emerging in (1.3.1), we now make the following definition that
is central to all subsequent considerations.
Definition 2.1. u : D(Ω) → C is called a distribution on Ω if u is linear and

continuous.
By design, distributions are simply elements of the dual space of the topological
vector space D(Ω). Given a functional u : D(Ω) → C and a function ϕ ∈ C0∞ (Ω), we
use the traditional notation u, ϕ in place of u(ϕ) (in particular, u, ϕ is a complex
number).
While any linear and continuous functional is sequentially continuous, the con-
verse is not always true. Nonetheless, for linear functionals on D(Ω), continuity
is equivalent with sequential continuity. This remarkable property, itself a conse-
quence of Theorem 14.6, is formally recorded below.
Fact 2.2 Let u : D(Ω) → C be a linear map. Then u is a distribution on Ω if

and only if for every sequence {ϕ j } j∈N contained in C0∞ (Ω) with the property that
D(Ω)
ϕ j −−−−→ ϕ for some function ϕ ∈ C0∞ (Ω), we have lim u, ϕ j = u, ϕ (where the
j→∞ j→∞
latter limit is considered in C).

18 2 The Space D (Ω) of Distributions
Remark 2.3. In general, if X, Y are topological vector spaces and Λ : X → Y is a

linear map, then Λ is sequentially continuous on X if and only if Λ is sequentially
continuous at the zero vector 0 ∈ X. When combined with Fact 2.2, this shows that
a linear map u : D(Ω) → C is a distribution on Ω if and only if lim u, ϕ j = 0 for
j→∞
D(Ω)
every sequence {ϕ j } j∈N ⊂ C0∞ (Ω) with ϕ j −−−−→ 0.
j→∞
Another important characterization of continuity of complex-valued linear func-

tionals u defined on D(Ω) is given by the following proposition.
Proposition 2.4. Let u : D(Ω) → C be a linear map. Then u is a distribution if and

only if for each compact set K ⊂ Ω there exist k ∈ N0 and C ∈ (0, ∞) such that
|u, ϕ| ≤ C sup |∂α ϕ(x)| for all ϕ ∈ C0∞ (Ω) with supp ϕ ⊆ K. (2.1.1)
x∈K
|α|≤k
Proof. Fix u : D(Ω) → C linear and suppose that for each compact set K ⊂ Ω there
exist k ∈ N0 and C ∈ (0, ∞) satisfying (2.1.1). To show that u is a distribution, let
D(Ω)
ϕ j −−−−→ 0. Then, there exists a compact set K ⊆ Ω such that supp ϕ j ⊆ K for all
j→∞
j ∈ N, and ∂α ϕ j −−−−→ 0 uniformly on K for any α ∈ Nn0 . For this compact set K, by
j→∞
our hypotheses, there exist C > 0 and k ∈ N0 such that (2.1.1) holds, and hence,
|u, ϕ j | ≤ C sup |∂α ϕ j (x)| −−−−→ 0, (2.1.2)

x∈K j→∞
|α|≤k
which implies that u, ϕ j −−−−→ 0. From this and Remark 2.3 it follows that u is a
j→∞
distribution in Ω.
To prove the converse implication we reason by contradiction. Suppose that there
exists a compact set K ⊆ Ω such that for every j ∈ N, there exists a function
ϕ j ∈ C0∞ (Ω) with supp ϕ j ⊆ K and
|u, ϕ j | > j sup |∂α ϕ j (x)|. (2.1.3)

x∈K
|α|≤ j
Define ψ j := u,ϕ1
j
ϕ j . Then for each j ∈ N we have ψ j ∈ C ∞ (Ω), supp ψ j ⊆ K, and
u, ψ j = 1. On the other hand, from (2.1.3) we see that
1
sup |∂α ψ j (x)| < ∀ j ∈ N. (2.1.4)
x∈K j
|α|≤ j
Now let α ∈ Nn0 be arbitrary. Then (2.1.4) implies that
1
sup |∂α ψ j | < whenever j ≥ |α|, (2.1.5)
x∈K j
|α|≤ j
2.1 The Definition of Distributions 19
D(Ω)
thus ψ j −−−−→ 0. Since u is a distribution in Ω, the latter implies lim u, ψ j = 0,
j→∞ j→∞
contradicting the fact that u, ψ j = 1 for each j ∈ N. This completes the proof of
the proposition.
Remark 2.5. Recall that for each compact set K ⊂ Ω we denote by DK (Ω) the vector
space of functions in C ∞ (Ω) with support contained in K endowed with the topology
inherited from E(Ω).
A closer look at the topology in DK (Ω) reveals that Proposition 2.4 may be
rephrased as saying that a linear map u : D(Ω) → C is a distribution in Ω if and
only if uD (Ω) is continuous for each compact set K ⊂ Ω. In fact, the topology on
K
D(Ω) is the smallest topology on C0∞ (Ω) with this property.
Definition 2.6. Let u be a distribution in Ω. If the nonnegative integer k intervening

in (2.1.1) may be taken to be independent of K, then u is called a distribution
of finite order. If u is a distribution of finite order, then the order of u
is by definition the smallest k ∈ N0 satisfying condition (2.1.1) for every compact
set K ⊂ Ω.
Here are a few important examples of distributions.
Example 2.7. For each f ∈ Lloc1

(Ω) define the functional u f : D(Ω) → C by

u f (ϕ) := f (x)ϕ(x) dx, ∀ ϕ ∈ C0∞ (Ω). (2.1.6)
Ω
Then clearly u f is linear and, if K is an arbitrary compact set contained in Ω, then

|u f (ϕ)| ≤ sup |ϕ(x)| K | f | dx = C sup |ϕ(x)|,
x∈K x∈K (2.1.7)
∞
for all ϕ ∈ C0 (Ω) with supp ϕ ⊆ K.
Hence, by Proposition 2.4, u f is a distribution in Ω. Moreover, (2.1.7) also shows

that u f is a distribution of order 0.
Remark 2.8. A distribution whose action is defined as in (2.1.6) will be referred to

as a distribution of function type.
To simplify notation, if f ∈ Lloc

1
(Ω) we will often simply use f (in place of u f )
for the distribution of function type defined as in (2.1.6). This is justified by the fact
that the linear map
ι : Lloc
1
(Ω) → {u : D(Ω) → C : u linear and continuous}
(2.1.8)
ι( f ) := u f for each f ∈ Lloc
1
(Ω),

is one-to-one. Indeed, if ι( f ) = 0 for some f ∈ Lloc
1
(Ω), then Ω f ϕ dx = 0 for all
functions ϕ ∈ C0∞ (Ω), which in turn, based on Theorem 1.3, implies that f = 0
almost everywhere in Ω. Since ι is also linear, the desired conclusion follows.
Example 2.9. We have that ln |x| ∈ Lloc

1
(Rn ), thus ln |x| is a distribution in Rn .
To see that indeed ln |x| is locally integrable in Rn , observe first that
1
ln t ≤ max tε , t−ε for all t > 0 and ε > 0. (2.1.9)
ε
This is justified by starting with the elementary inequality ln t ≤ t for all t > 0, then
replacing t by t±ε . In turn, for every R > 0 and ε ∈ (0, 1), estimate (2.1.9) gives

ln |x| dx ≤ ε−1 max |x|ε , |x|−ε dx < ∞. (2.1.10)
B(0,R) B(0,R)
Let us now revisit the functional from (1.3.1).
Example 2.10. For a given f ∈ Lloc 1

(Ω) and multi-index α ∈ Nn0 , consider the func-
tional gα : D(Ω) → C defined by

gα (ϕ) := (−1)|α| f ∂α ϕ dx for each ϕ ∈ C0∞ (Ω). (2.1.11)
Ω
D(Ω)
Clearly this is a linear mapping. Moreover, if ϕ j −−−−→ 0, then there exists K compact
j→∞
subset of Ω such that supp ϕ j ⊆ K for every j ∈ N, hence

|gα (ϕ j )| ≤ | f ||∂α ϕ j | dx
K
≤ ∂α ϕ j L∞ (K) f L1 (K) → 0 as j → ∞. (2.1.12)
By invoking Remark 2.3, it follows that gα is a distribution in Ω.
Next we consider a set of examples of distributions that are not of function type.
As a preamble, observe that f (x) := 1x for x 0, is not locally integrable on R.
Nonetheless, it is possible to associate to this function a certain distribution, not as
in (2.1.6), but in the specific manner described below.
Example 2.11. Consider the mapping P.V. 1x : D(R) → C defined by

1 ϕ(x)
P.V. (ϕ) := lim+ dx, ∀ ϕ ∈ C0∞ (R). (2.1.13)
x ε→0 |x|≥ε x
We claim that P.V. 1x is a distribution of order one in R.

First, we prove that the mapping (2.1.13) is well defined. Let ϕ ∈ C0∞ (R) and
suppose R > 0 is such that supp ϕ ⊂ (−R, R). Fix ε ∈ (0, R) and observe that since 1x
is odd on R \ {0}, we have

ϕ(x) ϕ(x) ϕ(x) − ϕ(0)
dx = dx = dx. (2.1.14)
|x|≥ε x ε≤|x|≤R x ε≤|x|≤R x

In addition, ϕ(x)−ϕ(0) ≤ sup|y|≤R |ϕ (y)| for each x ∈ R \ {0}. Thus, by Lebesgue’s
x
Dominated Convergence Theorem the limit lim+ |x|≥ε ϕ(x) x dx exists and is equal to
ϕ(x)−ϕ(0)
ε→0
1
|x|≤R x dx. This shows that the mapping P.V. x is well defined, and from def-
inition it is clear that P.V. 1x is linear. Furthermore, it is implicit in the argument
above that

P.V. 1 (ϕ) ≤ 2R sup |ϕ (x)|,
∀ ϕ ∈ C0∞ (−R, R) . (2.1.15)

x |x|≤R
In concert with Proposition 2.4 this shows that P.V. 1x is a distribution in R of order
at most one. We are left with showing that P.V. 1x does not have order 0. Consider the

compact K = [0, 1] and for each j ∈ N let ϕ j ∈ C0∞ (0, 1) be such that 0 ≤ ϕ j ≤ 1
1
and ϕ j ≡ 1 on j+2 , 1 − j+21
. Then from the very definition of P.V. 1x and the fact
that ϕ j vanishes near zero,
1 ϕ j (x) 1− 1
P.V. 1 , ϕ j =
j+2 1
dx ≥ dx = ln( j + 1) (2.1.16)
x 0 x 1
j+2
x
for each j ∈ N. Since sup x∈K |ϕ j (x)| ≤ 1 and lim ln( j + 1) = ∞, the inequality in
j→∞
(2.1.16) shows that there is no constant C ∈ (0, ∞) with the property that

P.V. 1 , ϕ ≤ C sup |ϕ(x)| for all ϕ ∈ C ∞ (R) with supp ϕ ⊆ K. (2.1.17)
x x∈K
0
This proves that P.V. 1x does not have order 0.
Remark 2.12. An inspection of the proof in Example 2.11 shows that

1 ϕ(x) − ϕ(0) ϕ(x)
P.V. , ϕ = dx + dx (2.1.18)
x |x|≤1 x |x|>1 x
for each ϕ ∈ C0∞ (R).
Example 2.13. An important distribution is the Dirac distribution δ defined

by
δ(ϕ) := ϕ(0), ∀ ϕ ∈ C0∞ (Rn ). (2.1.19)
It is not difficult to check that δ is a distribution in Rn of order 0. A natural ques-
tion to ask is whether δ is a distribution of function type. To answer this question,
suppose there exists f ∈ Lloc
1
(Rn ) such that

ϕ(0) = δ, ϕ = f ϕ dx for all ϕ ∈ C0∞ (Rn ). (2.1.20)
Rn

This implies that Rn f ϕ dx = 0 for every ϕ ∈ C0∞ (Rn ) with the property that 0
supp ϕ. Hence, by Theorem 1.3 we have f = 0 almost everywhere on Rn \{0}, thus
f = 0 almost everywhere in Rn . Consequently,

ϕ(0) = δ, ϕ = f ϕ dx = 0, ∀ ϕ ∈ C0∞ (Rn ), (2.1.21)
Rn
which is false. This proves that the Dirac distribution is not of function type.
Example 2.14. The Dirac distribution δ is sometimes referred to as having “mass
at zero” since, for each x0 ∈ Rn , we may similarly δ x0 : C0∞ (Rn ) → C by setting
δ x0 (ϕ) := ϕ(x0 ). Then δ x0 is a distribution in Rn (called the Dirac distribution with
mass at x0 ) and the convention we make is to drop the subscript x0 if x0 = 0 ∈ Rn .
Example 2.15. Let μ be either a complex Borel measure on Ω, or a Borel positive

measure on Ω that is locally finite (i.e., satisfies μ(K) < ∞ for every compact K ⊂
Ω). Consider

μ : D(Ω) → C, μ(ϕ) := ϕ dμ, ∀ ϕ ∈ C0∞ (Ω). (2.1.22)
Ω
The mapping in (2.1.22) is well defined, linear, and if K is an arbitrary compact set
in Ω, then
|μ(ϕ)| ≤ |μ|(K) sup |ϕ(x)|, ∀ ϕ ∈ C ∞ (Ω), supp ϕ ⊂ K. (2.1.23)

x∈K
By Proposition 2.4 we have that μ induces a distribution in Ω. The estimate in

(2.1.23) also shows that μ is a distribution of order zero.
Next we discuss the validity of a converse implication to the implication in Ex-
ample 2.15.
Proposition 2.16. Let u be a distribution in Ω of order zero. Then the distribution
u extends uniquely to a linear map Λu : C00 (Ω) → C that is locally bounded, in
the following sense: for each compact set K ⊂ Ω there exists C K ∈ (0, ∞) with the
property that
|Λu (ϕ)| ≤ C K sup |ϕ(x)|, ∀ ϕ ∈ C00 (Ω) with supp ϕ ⊆ K. (2.1.24)

x∈K
In addition, the functional Λu satisfies the following properties.

(i) Let {K j } j∈N be a sequence of compact subsets of Ω satisfying K j ⊂ K̊ j+1 for j ∈ N

∞
and Ω = K j . Then there exists a sequence of complex regular Borel measures
j=1
μ j on K j , j ∈ N, with the following properties:
(a) μ j (E) = μ (E) for every ∈ N, every Borel set E ⊂ K̊ , and every j ≥ ;
(b) for each j ∈ N one has

Λu (ϕ) = ϕ dμ j , ∀ ϕ ∈ C00 (Ω) with supp ϕ ⊂ K j . (2.1.25)
Kj
(ii) There exist two Radon measures μ1 , μ2 , taking Borel sets from Ω into [0, ∞] (i.e.,
measures satisfying the regularity properties (ii)–(iv) in Theorem 14.25), such
that

Re Λu (ϕ) = ϕ dμ1 − ϕ dμ2 , ∀ ϕ ∈ C00 (Ω) real-valued. (2.1.26)
Ω Ω
Furthermore, a similar conclusion is valid for Im Λu .
The following approximation result is useful in the proof of Proposition 2.16.
Lemma 2.17. Let k ∈ N0 and suppose K is a compact subset of Ω. Define the com-
pact set

K0 := x ∈ Ω : dist (x, K) ≤ 12 dist (K, ∂Ω) . (2.1.27)
Then for every ϕ ∈ C0k (Ω) with supp ϕ ⊆ K there exists a sequence {ϕ j } j∈N of func-
tions in C0∞ (Ω) with supp ϕ j ⊆ K0 and
∀ K1 ⊂ Ω compact and ∀ α ∈ Nn0 with |α| ≤ k, we

have lim sup |∂α ϕ j (x) − ∂α ϕ(x)| = 0. (2.1.28)
j→∞ x∈K1
Proof. Recall the sequence of functions {φ1/ j } j∈N defined in (1.2.3)–(1.2.5) corre-
sponding to ε := 1j , j ∈ N. Pick j0 ∈ N such that j10 < 12 dist (K, ∂Ω) and then
set

ϕ j (x) : = ϕ(y)φ1/ j (x − y) dy
Rn

= ϕ(x − y)φ1/ j (y) dy, ∀ x ∈ Ω, j ≥ j0 . (2.1.29)
Rn
In light of (1.2.5) and (2.1.29) we have that
ϕ j ∈ C0∞ (Ω) and supp ϕ j ⊆ K0 , for all j ≥ j0 . (2.1.30)
Pick some j1 ∈ N such that j11 < 14 dist(K, ∂Ω) (note that j1 ≥ j0 ). Then the set
:= K0 + B(0, 1/ j1 ) is a compact subset of Ω.
K
Fix ε > 0 arbitrary. Since ϕ is continuous in Ω, it is uniformly continuous on the
compact K,
hence there exists δ > 0 such that |ϕ(x1 )−ϕ(x2 )| ≤ ε for every x1 , x2 ∈ K
satisfying |x1 − x2 | ≤ δ. Furthermore, choose j2 ∈ N satisfying j12 ≤ δ and j2 ≥ j1 .
At this point, for each x ∈ K0 , and each j ≥ j2 , we may write

|ϕ j (x) − ϕ(x)| = [ϕ(x − y) − ϕ(x)]φ1/ j (y) dy
B(0,1/ j)

= |ϕ(x − y) − ϕ(x)|φ1/ j (y) dy
B(0,1/ j)

≤ε φ1/ j (y) dy = ε. (2.1.31)
B(0,1/ j)
In summary, we have proved that for each ε > 0 there exists j2 ∈ N such that
|ϕ j (x) − ϕ(x)| ≤ ε for every x ∈ K0 . This shows that the sequence defined in (2.1.29)
satisfies
lim sup |ϕ j (x) − ϕ(x)| = 0. (2.1.32)
j→∞ x∈K0
Next, observe that since ϕ is of class C k , from (2.1.29) we also have

∂α ϕ j (x) = (∂α ϕ)(y)φ1/ j (x − y) dy, ∀ x ∈ Ω, (2.1.33)
Rn
for every α ∈ N0 , |α| ≤ k, and all j ≥ j0 . This and an argument similar to that used
for the proof of (2.1.32) imply
lim sup |∂α ϕ j (x) − ∂α ϕ(x)| = 0, ∀ α ∈ N0 , |α| ≤ k. (2.1.34)

j→∞ x∈K0
Now the fact that the sequence {ϕ j } j≥ j0 satisfies (2.1.28) follows from the support
condition (2.1.30) and (2.1.34). This finishes the proof of the lemma.
We are now ready to present the proof of Proposition 2.16.
Proof of Proposition 2.16. Let u be a distribution in Ω of order zero and let K
be a compact set contained in Ω. Fix ϕ ∈ C00 (Ω) such that supp ϕ ⊆ K and apply
Lemma 2.17 with k := 0. Hence, with K0 as in (2.1.27), there exists a sequence
{ϕ j } j∈N of functions in C0∞ (Ω) with supp ϕ j ⊆ K0 satisfying (2.1.28).
The fact that u is a distribution of order zero, implies the existence of some C =
C(K0 ) ∈ (0, ∞) such that (2.1.1) holds with k = 0. The latter combined with (2.1.28)
implies
|u, ϕ j − u, ϕk | ≤ C sup |ϕ j (x) − ϕk (x)| −−−−−→ 0. (2.1.35)
x∈K0 j,k→∞

Hence, the sequence of complex numbers u, ϕ j j∈N is Cauchy, thus convergent in
C, which allows us to define
Λu (ϕ) := lim u, ϕ j . (2.1.36)

j→∞
Proving that this definition is independent of the selection of {ϕ j } j∈N is done by

interlacing sequences. Specifically, if {ϕj } j∈N is another sequence of functions in
C0∞ (Ω), supported in K0 and satisfying (2.1.28), then by considering the sequence
{ψ j } j≥2 defined by ψ2 j+1 := ϕ j and ψ2 j := ϕj for every j ≥ 2, we obtain that

u, ψ j j≥2 is convergent, thus its subsequences u, ϕ j j∈N and u, ϕj j∈N are also
convergent and have the same limit. In turn, the independence of the definition of
Λu (ϕ) on the approximating sequence {ϕ j } j∈N readily implies that Λu : C00 (Ω) → C
is a linear mapping.
In addition, the fact that u is a distribution of order zero implies the existence of
a finite constant C = C(K0 ) > 0 such that
|u, ϕ j | ≤ C sup |ϕ j (x)| for each j ∈ N. (2.1.37)

x∈K0
Taking the limit as j → ∞ in (2.1.37) gives |Λu (ϕ)| ≤ C sup |ϕ(x)| on account
x∈K
of (2.1.36). Moreover, since K0 has an explicit construction in terms of K (recall
(2.1.27)), the constant C above is ultimately dependent on K, thus C = C(K) as
wanted. This proves the local boundedness of Λu in the sense of (2.1.24). To show
that this linear extension Λu of u to C00 (Ω) is unique in the class of linear and lo-
cally bounded mappings, it suffices to prove that if Λ : C00 (Ω) → C is a linear lo-
cally bounded mapping that vanishes on C0∞ (Ω) then Λ is identically zero on C00 (Ω).
To this end, pick an arbitrary ϕ ∈ C0∞ (Ω), set K := supp ϕ and, as before, apply
Lemma 2.17 to obtain a sequence of functions ϕ j ∈ C0∞ (Ω), j ∈ N, supported in the
fixed compact neighborhood K0 of K such that (2.1.28) holds with k = 0. Then
|Λ(ϕ)| = |Λ(ϕ − ϕ j )| ≤ C sup |ϕ(x) − ϕ j (x)| −−−−→ 0, (2.1.38)

x∈K0 j→∞
proving that Λ(ϕ) = 0, as wanted.

Moving on, consider a sequence of compact sets satisfying the hypothesis of (i).
For example the sequence of compacts {K j } j∈N defined by (1.3.11) will do. Based
on what we have proved so far and Riesz’s representation theorem for complex
measures (see Theorem 14.26), it follows that there exists a sequence of complex
regular Borel measures μ j on K j , j ∈ N, such that (2.1.25) holds.
Now fix ∈ N and a Borel set E ⊂ K̊ . If j ≥ , since the measures μ j , μ are
regular, in order to prove that μ j (E) = μ (E) it suffices to show that μ j (F) = μ (F)
for every compact set F ⊆ E. Fix such a compact set and choose a sequence of
compact sets {Fk }k∈N contained in K̊ , such that Fk+1 ⊂ F̊k for each k ∈ N and
∞
k=1 F k = F. Applying Uryshon’s lemma (see Proposition 14.28) we obtain a se-
quence ϕk ∈ C00 (Ω), 0 ≤ ϕk ≤ 1, supp ϕk ⊆ Fk and ϕk ≡ 1 on F, for each k ∈ N.
Then lim ϕk (x) = χF (x) for x ∈ Ω and we may write
k→∞

μ j (F) = χF dμ j = lim ϕk dμ j = lim u, ϕk = lim ϕk dμ
Kj k→∞ Kj k→∞ k→∞ K

= χF dμ = μ (F). (2.1.39)
K
This completes the proof of (i). Finally, the claim in (ii) follows by invoking Riesz’s
representation theorem for locally bounded functionals (cf. Theorem 14.27).
Among other things, Proposition 2.16 is a useful ingredient in the following rep-
resentation theorem for positive distributions.
Theorem 2.18. Let u be a distribution in Ω such that u, ϕ ≥ 0 for every nonnega-
tive function ϕ ∈ C0∞ (Ω). Then there exists a unique positive Borel regular measure
μ on Ω such that
u, ϕ = ϕ dμ, ∀ ϕ ∈ C0∞ (Ω). (2.1.40)
Ω
Proof. First we prove that u has order zero. To do so, let K be a compact set in Ω
and fix ψ ∈ C0∞ (Ω), ψ ≥ 0 and satisfying ψ ≡ 1 on K. Then, if ϕ ∈ C0∞ (Ω) has
supp ϕ ⊆ K and is real valued, it follows that
(sup |ϕ(x)|) ψ ± ϕ ≥ 0, (2.1.41)

x∈K
hence by the positivity of u, we have
(sup |ϕ(x)|)u, ψ ± u, ϕ ≥ 0.

x∈K
Thus,
|u, ϕ| ≤ u, ψ sup x∈K |ϕ(x)|,
(2.1.42)
for all real valued ϕ ∈ C00 (Ω) with supp ϕ ⊆ K.
If now ϕ ∈ C0∞ (Ω) with supp ϕ ⊆ K is complex valued, say ϕ = ϕ1 + iϕ2 and ϕ1 , ϕ2 ,
are real valued, then by using (2.1.42) we may write

|u, ϕ| = |u, ϕ1 |2 + |u, ϕ2 |2 ≤ u, ψ sup |ϕ1 (x)| + sup |ϕ2 (x)|
x∈K x∈K
≤ 2 u, ψ sup |ϕ(x)|. (2.1.43)
x∈K
The fact that the distribution u has order zero now follows from (2.1.43). Having
established this, we may apply Proposition 2.16 to conclude that u may be uniquely
extended to a linear map Λu : C00 (Ω) → C that is locally bounded.
We next propose to show that this linear map is positive. In this respect, we note
that if ϕ ≥ 0 then the functions {Φ j } j∈N used in the proof of Proposition 2.16 (which
are constructed by convolving ϕ with a nonnegative mollifier) may also be taken
to be nonnegative. When combined with (2.1.36) and the fact that u is positive,
this shows that the extension Λu of u to C00 (Ω) as defined in (2.1.36) is a positive
functional. Consequently, Riesz’s representation theorem for positive functionals
(see Theorem 14.25) may be invoked to conclude that there exists a unique positive
Borel regular measure μ on Ω such that

Λu (ϕ) = ϕ dμ, ∀ ϕ ∈ C00 (Ω). (2.1.44)
Ω
Now (2.1.40) follows by specializing (2.1.44) to ϕ ∈ C0∞ (Ω). This finishes the proof
of the theorem.
The first part in the statement of Proposition 2.16 has a natural generalization
corresponding to distributions of any finite order.
Proposition 2.19. Let u be a distribution in Ω of order k ∈ N0 . Then there exists a

linear map Λu : C0k (Ω) → C satisfying Λu C ∞ (Ω) = u and with the property that
0
for each compact set K ⊂ Ω there exists C ∈ (0, ∞) such that

(2.1.45)
|Λu (ϕ)| ≤ C sup x∈K |∂α ϕ(x)|, ∀ ϕ ∈ C0k (Ω) with supp ϕ ⊆ K.
|α|≤k
2.2 The Topological Vector Space D (Ω) 27
Moreover, Λu is unique satisfying these properties.
Proof. Let K be an arbitrary compact subset of Ω and let ϕ ∈ C0k (Ω) be such that
supp ϕ ⊆ K. Apply Lemma 2.17 to obtain a sequence {ϕ j } j∈N ⊂ C0∞ (Ω) of functions
supported in the compact K0 defined in relation to K as in (2.1.27), and satisfying
(2.1.28). Using the fact that u is a distribution of order k and the properties of the
sequence {ϕ j } j∈N we may run an argument similar to that from the first part of the
proof Proposition 2.16 to define a mapping Λu : C0k (Ω) → C as in (2.1.36). That this
mapping is well defined, linear, and satisfies the desired properties is proved much
as is the corresponding result in Proposition 2.16 (with the obvious adjustments due
to the fact that u has finite order k, rather than order zero).
Example 2.20. For each multi-index α ∈ Nn0 the distribution ∂α δ is a distribution of

order |α| in Rn . By Proposition 2.19 (and formula (2.1.36)) this distribution extends
uniquely to the linear map
Λ∂α δ : C0|α| (Rn ) → C defined by

(2.1.46)
Λ∂α δ (ϕ) := (−1)|α| (∂α ϕ)(0) for each ϕ ∈ C0|α| (Rn ),
which satisfies (2.1.45) with k := |α|.
Remark 2.21. Let k ∈ N0 .

(1) If w : C0k (Ω) → C is a linear map, then for every α ∈ Nn0 we may define the map-
ping ∂α w : C0k+|α| (Ω) → C by ∂α w(ϕ) := (−1)|α| w(∂α ϕ) for every ϕ ∈ C0k+|α| (Ω).
(2) If u is a distribution in Ω of order k, then for each α ∈ Nn0 the distribution ∂α u has
finite order k + |α| and the extension Λ∂α u to C0k+|α| (Ω) given by Proposition 2.16
is equal to ∂α Λu on C0k+|α| (Ω), the latter being the mapping defined in part (1) in
relation to Λu in place of w. Indeed, since both Λ∂α u and ∂α Λu are linear func-
tionals on C0k+|α| (Ω) extending the distribution ∂α u and satisfying (2.1.45), the
uniqueness portion of Proposition 2.16 guarantees that they must be equal.
2.2 The Topological Vector Space D (Ω)
The space of distributions in Ω endowed with the natural addition and scalar multi-
plication of linear mappings becomes a vector space over C. Indeed, if u1 , u2 , u are
distributions in Ω and λ ∈ C, we define u1 + u2 : D(Ω) → C and λu : D(Ω) → C
by setting
(u1 + u2 )(ϕ) := u1 , ϕ + u2 , ϕ and (λu)(ϕ) := λu, ϕ,

(2.2.1)
for each test function ϕ ∈ C0∞ (Ω).
It is not difficult to check that u1 + u2 and λu are distributions in Ω.

The topology we consider on the vector space of distributions in Ω is the weak∗-

topology induced by D(Ω) (for details see Section 14.1.0.5), which makes it a topo-
logical vector space over C and we denote this topological vector space by D (Ω).
As a byproduct of this definition we have the following important properties of
D (Ω).
Fact 2.22 D (Ω) is a locally convex topological vector space over C. In addition, a
sequence {u j } j∈N in D (Ω) converges to some u ∈ D (Ω) as j → ∞ in D (Ω) if and
only if u j , ϕ −−−−→ u, ϕ for every ϕ ∈ C0∞ (Ω), in which case we use the notation
j→∞
D (Ω)
u j −−−−→ u.
j→∞
Moreover, the topological space D (Ω) is complete, in the following sense. If the
sequence {u j } j∈N ⊂ D (Ω) is such that lim u j , ϕ exists (in C) for every ϕ ∈ C0∞ (Ω)
j→∞
then the functional u : D(Ω) → C defined by u(ϕ) := lim u j , ϕ for every ϕ ∈
j→∞
C0∞ (Ω) is a distribution in Ω.
Note that from Fact 2.22 it is easy to see that if a sequence {u j } j∈N in D (Ω) is
convergent then its limit is unique. Indeed, if such a sequence would have two limits,
say u, v ∈ D (Rn ), then it would follow that for each ϕ ∈ C0∞ (Rn ), the sequence of
numbers {u j , ϕ} j∈N would converge to both u, ϕ and v, ϕ, thus u, ϕ = v, ϕ.
Hence, u = v.
Remark 2.23. Assume that we are given u ∈ D (Ω) and a sequence uε ∈ D (Ω),
D (Ω)
ε ∈ (0, ∞). We make the convention that uε −−−−→
+
u is understood in the sense
ε→0
that for every sequence of positive numbers {ε j } j∈N satisfying lim ε j = 0 we have
j→∞
D (Ω)
uε j −−−−→ u.
j→∞
Example 2.24. Let φ be as in (1.2.3) and recall the sequence of functions {φ j } j∈N
from (1.3.7). Interpreting each φ j ∈ Lloc
1
(Rn ) as distribution in Rn , for each function
∞
ϕ ∈ C0 (R ) we have
n

φ j , ϕ = φ j (x)ϕ(x) dx = φ(y)ϕ(y/ j) dy, ∀ j ∈ N. (2.2.2)
Rn Rn
Thus, by Lebesgue’s Dominated Convergence Theorem (cf. Theorem 14.15),

φ j , ϕ = φ(y)ϕ(y/ j) dy −−−−→ ϕ(0) φ(y) dy = ϕ(0) = δ, ϕ. (2.2.3)
Rn j→∞ Rn
D (Rn )
This proves that φ j −−−−−→ δ.
j→∞
Exercise 2.25. Prove that if p ∈ [1, ∞] and if { f j } j∈N is a sequence of functions in

D (Ω)
L p (Ω) which converges in L p (Ω) to some f ∈ L p (Ω) then f j −−−−→ f .
j→∞
2.2 The Topological Vector Space D (Ω) 29
Hint: Use Hölder’s inequality.
Exercise 2.26. (a) Assume that f ∈ L1 (Rn ) and for each ε > 0 define the func-
tion fε (x) := ε−n f (x/ε) for each x ∈ Rn . Prove that for each g ∈ L∞ (Rn ) that is
continuous at 0 ∈ Rn we have

fε (x)g(x) dx −−−−→
+
f (x) dx g(0). (2.2.4)
Rn ε→0 Rn
(b) Use part (a) to prove that if f ∈ L1 (Rn ) is given and for every j ∈ N we
define f j (x) := jn f ( jx) for each x ∈ Rn , then each f j belongs to L1 (Rn ) (hence
f j ∈ D (Rn )) and

D (Rn )
f j −−−−−→ c δ where c := f (x) dx. (2.2.5)
j→∞ Rn
Hint:
To justify the claim in part (a), make a change of variables to write the integral
R n f ε (x)g(x) dx as Rn f (y)g(εy) dy, then use Lebesgue’s Dominated Convergence
Theorem.
In the last part of this section we discuss the composition with invertible linear
maps of distributions in Rn . Specifically, let A ∈ Mn×n (R) be such that det A 0.
Then for every f ∈ Lloc 1
(Rn ) one has f ◦ A ∈ Lloc1
(Rn ). By Example 2.7 we have

f, f ◦ A ∈ D (R ). In addition,
n

−1
f ◦ A, ϕ = f (Ax)ϕ(x) dx = | det A| f (y)ϕ(A−1 y) dy
Rn Rn
= | det A|−1 f, ϕ ◦ A−1 , ∀ ϕ ∈ C0∞ (Rn ). (2.2.6)
This and Exercise 1.24 justify extending the operator of composition with linear
maps to D (Rn ) as follows.
Proposition 2.27. Let A ∈ Mn×n (R) be such that det A 0. For each distribution
u ∈ D (Rn ), define the mapping u ◦ A : D(Rn ) → C by setting

u ◦ A (ϕ) := | det A|−1 u, ϕ ◦ A−1 , ∀ ϕ ∈ D(Rn ). (2.2.7)
Then u ◦ A ∈ D (Rn ).
Proof. This is an immediate consequence of (1.3.18).
Exercise 2.28. Let A, B ∈ Mn×n (R) be such that det A 0 and det B 0. Then the
following identities hold in D (Rn ):
(1) (u ◦ A) ◦ B = u ◦ (AB) for every u ∈ D (Rn );
(2) u ◦ (λA) = λu ◦ A for every u ∈ D (Rn ) and every λ ∈ R;
(3) (u + v) ◦ A = u ◦ A + v ◦ A for every u, v ∈ D (Rn ).
2.3 Multiplication of a Distribution with a C∞ Function
The issue we discuss in this section is the definition of the multiplication of a dis-
tribution u ∈ D (Ω) with a smooth function a ∈ C ∞ (Ω). First we consider the case
when u is of function type, i.e., u = u f for some f ∈ Lloc 1
(Ω). In this particular case,
a f ∈ Lloc (Ω) thus it defines a distribution ua f on Ω and
1

ua f , ϕ = (a f )ϕ dx = f (a ϕ) dx = f, a ϕ, ∀ ϕ ∈ C0∞ (Ω). (2.3.1)
Ω Ω
In the general case, this suggests defining au as in the following proposition.
Proposition 2.29. Let u ∈ D (Ω) and a ∈ C ∞ (Ω). Then the mapping
au : D(Ω) → C defined by (au)(ϕ) := u, aϕ, ∀ ϕ ∈ C0∞ (Ω), (2.3.2)
is linear and continuous, hence a distribution on Ω.
Proof. The fact that au is linear is immediate. To show that au is also continuous
D(Ω)
we make use of Remark 2.3. To this end, consider a sequence ϕ j −−−−→ 0. By (2)
j→∞
D(Ω)
in Exercise 1.20 we have a ϕ j −−−−→ 0. Since u is a distribution on Ω, the latter
j→∞
convergence implies lim u, aϕ j = 0. Moreover, from (2.3.2) we have (au)(ϕ j ) =
j→∞
u, aϕ j for each j ∈ N. Hence, lim (au)(ϕ j ) = 0 proving that au is continuous.
j→∞
Exercise 2.30. Let f ∈ Lloc 1

(Ω) and a ∈ C ∞ (Ω). With the notation from (2.1.6),

prove that au f = ua f in D (Ω).
Remark 2.31.
(1) When more information about u ∈ D (Ω) is available, (2.3.2) may continue to
yield a distribution under weaker regularity demands on the function a than
the current assumption that a ∈ C ∞ (Ω). In general, however, the condition
a ∈ C ∞ (Ω) may not be weakened if (2.3.2) is to yield a distribution for arbi-
trary u ∈ D (Ω).
(2) As observed later (see Remark 2.36), one may not define the product of two
arbitrary distributions in a way that ensures associativity.
(3) Based on (2.3.2) and (2.2.1), it follows that if u, u1 , u2 ∈ D (Ω), and if a, a1 ,
a2 ∈ C ∞ (Ω), then a(u1 + u2 ) = au1 + au2 , (a1 + a2 )u = a1 u + a2 u, and a1 (a2 u) =
(a1 a2 )u, where the equalities are considered in D (Ω).
Exercise 2.32. The following properties hold.

E(Ω) D (Ω)
(1) If u ∈ D (Ω) and a j −−−−→ a, then a j u −−−−→ au.
j→∞ j→∞
∞ D (Ω) D (Ω)
(2) If a ∈ C (Ω) and u j −−−−→ u, then au j −−−−→ au.
j→∞ j→∞
2.3 Multiplication of a Distribution with a C ∞ Function 31
(3) For each a ∈ C ∞ (Ω) the mapping D (Ω) u → au ∈ D (Ω) is linear and
continuous.
Hint: Observe that the map in (3) is the transpose (recall the definition from
(14.1.10)) of the linear and continuous map in Exercise 1.21, hence Proposition 14.2
applies.
Example 2.33. Recall the Dirac distribution defined in (2.1.19) and assume that
some function a ∈ C ∞ (Ω) has been given. Then, for every ϕ ∈ C0∞ (Rn ) we may
write
aδ, ϕ = δ, aϕ = (aϕ)(0) = a(0)ϕ(0) = a(0)δ, ϕ.
This shows that
aδ = a(0)δ in D (Rn ) for every a ∈ C ∞ (Ω). (2.3.3)
As a consequence,
xm δ = 0 in D (R), for every m ∈ N. (2.3.4)
Example 2.34. The goal is to solve the equation
xu = 1 in D (R). (2.3.5)
Clearly, this equation does not have a solution u of function type. Recall the distri-
bution defined in Example 2.11. Then for every ϕ ∈ C0∞ (R) we may write

1 1 x ϕ(x)
x P.V. , ϕ = P.V. , xϕ(x) = lim+ dx
x x ε→0 |x|≥ε x

= ϕ(x) dx = 1, ϕ. (2.3.6)
R
Thus,
1
x P.V. =1 in D (R). (2.3.7)
x
Given (2.3.4), it follows that u := P.V. 1x + c δ will also be a solution of (2.3.5) for
any c ∈ C. We will see later (c.f. Remark 2.78) that in fact any solution of (2.3.5) is
of the form P.V. 1x + c δ, where c ∈ C.
Exercise 2.35. Let ψ ∈ C ∞ (R). Determine a solution u ∈ D (R) of the equation

xu = ψ in D (R).
Remark 2.36. Suppose one could define the product of distributions as an associative
operation, in a manner compatible with the multiplication by a smooth function.
Considering then δ, x, and P.V. 1x ∈ D (R), one would then necessarily have
1 1
0 = (δ · x) P.V. = δ x · P.V. =δ·1=δ in D (R), (2.3.8)
x x
thanks to (2.3.4) and (2.3.7), leading to the false conclusion that δ = 0 in D (R).
2.4 Distributional Derivatives
We are now ready to define derivatives of distributions. One of the most basic
attributes of the class of distributions, compared with other classes of locally inte-
grable functions, is that distributions may be differentiated unrestrictedly within this
environment (with the resulting objects being still distributions), and that the oper-
ation of distributional differentiation retains some of the most basic properties as
in the case of ordinary differentiable functions (such as a suitable product formula,
symmetry of mixed derivatives, etc.). In addition, the differentiation of distributions
turns out to be compatible with the pointwise differentiation in the case when the
distribution in question is of function type, given by a sufficiently regular function.
To develop some sort of intuition, we shall start our investigation by looking first
at a distribution of function type, and try to generalize the notion of weak (Sobolev)
derivative from Definition 1.2. As noted earlier, if f ∈ Lloc
1
(Ω), the mapping defined
in (1.3.1) is a distribution on Ω. This suggests making the following definition.
Definition 2.37. If α ∈ Nn0 and u ∈ D (Ω), the distributional derivative

(or the derivative in the sense of distributions) of order α of the distribution u is the
mapping ∂α u : D(Ω) → C defined by
∂α u(ϕ) := (−1)|α| u, ∂α ϕ, ∀ ϕ ∈ C0∞ (Ω). (2.4.1)
Remark 2.38. Note that if u ∈ D (Ω) is of function type, say u = u f for some
f ∈ Lloc
1
(Ω), and if the weak derivative ∂α f exists, i.e., one can find g ∈ Lloc
1
(Ω) such
that (1.2.2) holds, then according to Definition 2.37 we have that the distributional
derivative ∂α u is equal to the distribution ug in D (Ω). In short, ∂α u f = u∂α f in this
case. Thus, Definition 2.37 generalizes Definition 1.2.
That the class of distributions in Ω is stable under taking distributional derivatives

is proved in the next proposition.
Proposition 2.39. For each α ∈ Nn0 and each u ∈ D (Ω) we have ∂α u ∈ D (Ω).
Proof. Fix α ∈ Nn0 and u ∈ D (Ω). That ∂α u : D(Ω) → C is a linear map is

D(Ω)
easy to see. To prove that it is also continuous, let ϕ j −−−−→ 0. Since by item (1) in
j→∞
α D(Ω)
Exercise 1.20 we have ∂ ϕ j −−−−→ 0 and u ∈ D (Ω), we may write
j→∞
∂α u(ϕ j ) = (−1)|α| u, ∂α ϕ j −−−−→ 0. (2.4.2)

j→∞
Remark 2.3 then shows that ∂α u ∈ D (Ω).

2.4 Distributional Derivatives 33
Exercise 2.40. Suppose that m ∈ N and f ∈ C m (Ω). Prove that for any α ∈ Nn0
satisfying |α| ≤ m, the distributional derivative of order α of u f is the distribution of
function type given by the derivative, in the classical sense, of order α of f , that is,
∂α (u f ) = u∂α f in D (Ω).
Proposition 2.41. Let m ∈ N0 and assume that

P(x, ∂) := aα (x)∂α , aα ∈ C ∞ (Ω), α ∈ Nn0 , |α| ≤ m. (2.4.3)
|α|≤m
Also, suppose that u ∈ C m (Ω). Then P(x, ∂)u, computed in D (Ω), coincides as a
distribution with the distribution induced by P(x, ∂)u, computed pointwise in Ω.
Proof. This follows from (2.4.3), Exercise 2.40, and Exercise 2.30.
Remark 2.42. Recall that, given a set E ⊆ Rn , a function f : E → C is called

Lipschitz provided there exists M ∈ [0, ∞) such that
| f (x) − f (y)| ≤ M|x − y| ∀ x, y ∈ E. (2.4.4)
The number C := inf {M for which (2.4.4) holds} is referred to as the Lipschitz
constant of f . We agree to denote by Lip(E) the collection of all Lipschitz func-
tions on E. A classical useful result (due to E. McShane) concerning this space is
that for any set E ⊆ Rn we have

Lip(E) = F|E : F ∈ Lip(Rn ) . (2.4.5)
Indeed given any f ∈ Lip(E) with Lipschitz constant M one may check that the
function F(x) := inf{ f (y) + M|x − y| : y ∈ E} for all x ∈ Rn belongs to Lip(Rn ), has
Lipschitz constant M, and satisfies f = F|E .
We will also prove (see Theorem 2.114) that if Ω ⊆ Rn is an arbitrary open set
and f : Ω → C is Lipschitz then the distributional derivatives ∂k f , k = 1, . . . , n,
belong to L∞ (Ω). Consequently,
f : Ω → R Lipschitz =⇒ ∂k (u f ) = u∂k f in D (Ω) for k = 1, . . . , n. (2.4.6)
Some basic properties of differentiation in the distributional sense are summa-

rized below.
Proposition 2.43. The following properties of distributional differentiation hold.

(1) Any distribution is infinitely differentiable (i.e., D (Ω) is stable under the action
of ∂α for any α ∈ N0 ).
(2) If u ∈ D (Ω) and k, ∈ {1, ..., n} then ∂k ∂ u = ∂ ∂k u in D (Ω).
D (Ω) D (Ω)
(3) If u j −−−−→ u and α ∈ Nn0 , then ∂α u j −−−−→ ∂α u.
j→∞ j→∞
(4) For any u ∈ D (Ω) and any a ∈ C ∞ (Ω) we have ∂ j (au) = (∂ j a)u + a(∂ j u) in
D (Ω).
Proof. The first property follows immediately from the definition of distributional
derivatives. To prove the remaining properties, fix an arbitrary ϕ ∈ C0∞ (Ω). Then,
using (2.4.1) repeatedly and the symmetry of mixed partial derivatives for smooth
functions (Schwarz’s theorem), we have
∂k ∂ u, ϕ = −∂ u, ∂k ϕ = u, ∂ ∂k ϕ = u, ∂k ∂ ϕ

= −∂k u, ∂ ϕ = ∂ ∂k u, ϕ, ∀ k, ∈ {1, . . . , n},
which implies (2). Let now {u j } j∈N , u, and α satisfy the hypotheses in (3). Based on
(2.4.1) and Fact 2.22 we may write
∂α u j , ϕ = (−1)|α| u j , ∂α ϕ −−−−→ (−1)|α| u, ∂α ϕ = ∂α u, ϕ,

j→∞
α α
hence ∂ u j converges to ∂ u in D (Ω) as j → ∞ and the proof of (3) is complete.
Finally, for u ∈ D (Ω) and a ∈ C ∞ (Ω), using (2.4.1) and Leibniz’s product formula
for derivatives of smooth functions we can write
∂ j (au), ϕ = −au, ∂ j ϕ = −u, a(∂ j ϕ) = −u, ∂ j (aϕ) + u, (∂ j a)ϕ

= ∂ j u, aϕ + (∂ j a)u, ϕ = a(∂ j u) + (∂ j a)u, ϕ , (2.4.7)
from which (4) follows.
Example 2.44. Recall the Heaviside function H from (1.2.9). This is a locally inte-
grable function thus it defines a distribution on R that we denote also by H. Then,
the computation in (1.2.12) implies
H , ϕ = −H, ϕ = δ, ϕ, ∀ ϕ ∈ C0∞ (R). (2.4.8)
Hence,
H = δ in D (R). (2.4.9)
Exercise 2.45. Prove that for every function a ∈ C ∞ (Ω) and every α ∈ Nn0 we have
α!
a(∂α δ) = (−1)|β| (∂β a)(0)∂α−β δ in D (Ω). (2.4.10)
β≤α
β!(α − β)!
Hint: Use formula (14.2.6) when computing ∂α (aϕ) for ϕ ∈ C0∞ (Ω).
Exercise 2.46. Prove that for every c ∈ R one has

e−c|x| = −c e−cx H(x) + c ecx H(−x) in D (R). (2.4.11)
Hint: Show e−c|x| = e−cx H(x) + ecx H(−x) in D (R) and then use part (4) in Proposi-
tion 2.43 and (2.4.9).
Next, we look at the issue of existence of antiderivatives for distributions on open
intervals.
2.4 Distributional Derivatives 35
Proposition 2.47. Let I be an open interval in R and suppose u0 ∈ D (I).

(1) The equation u = u0 in D (I) admits at least one solution.
(2) If u1 , u2 ∈ D (I) are such that (u1 ) = u0 in D (I) and (u2 ) = u0 in D (I), then
there exists c ∈ C such that u1 − u2 = c in D (I).
Proof. Suppose I = (a, b), where a ∈ R ∪ {−∞} and b ∈ R ∪ {+∞}, and define the
set A(I) := {ϕ : ϕ ∈ C0∞ (I)}. We claim that

∞
if ϕ ∈ C0 (I), then ϕ ∈ A(I) ⇐⇒ ϕ(x) dx = 0. (2.4.12)
I
The left-to-right implication in (2.4.12) is clear by the fundamental theorem of

∞
calculus.
To prove the converse implication, x suppose ϕ ∈ C0 (I) is such that
I
ϕ(x) dx = 0. Then the function ψ(x) := a ϕ(t) dt, x ∈ I, satisfies ψ ∈ C ∞ (I),
supp ψ ⊆ supp ϕ, and ψ = ϕ on I, thus ϕ ∈ A(I). This finishes the justification of
(2.4.12).
Next, fix ϕ0 ∈ C0∞ (I) with the property that I ϕ0 (x) dx = 1 and consider the map
Θ : D(I) → D(I) defined by
Θ(ϕ) := θϕ for each ϕ ∈ C0∞ (I), where

x
(2.4.13)
θϕ (x) := ϕ(t) − ϕ(y) dy ϕ0 (t) dt, ∀ x ∈ I.
a I

Since the integral of ϕ − I ϕ(x) dx ϕ0 over I is zero, our earlier discussion shows
that Θ is well defined. Since Θ is linear, Theorem 14.6 and Fact 1.15 imply that Θ
is continuous if and only if it is sequentially continuous. The latter property may
be verified from definitions. Finally, from (2.4.13) and the fundamental theorem of
calculus we have
Θ(ϕ ) = ϕ for every ϕ ∈ C0∞ (I). (2.4.14)
Fix an arbitrary distribution u0 on I and define u := −u0 ◦Θ. Thanks to the properties
of Θ, we have that u is a distribution on I. In concert with (2.4.14), the definition of
u implies that u, ϕ = −u0 , ϕ for every ϕ ∈ C0∞ (I), proving that u = u0 in D (I).
This finishes the proof of the statement in (1).
Moving on, suppose u ∈ D (I) is such that u = 0 in D (I). Then if ϕ0 is as earlier
in the proof, for any ϕ ∈ C0∞ (I) we may write

u, ϕ = u, ϕ − ϕ(x) dx ϕ0 + ϕ(x) dx u, ϕ0
I I

= u, (θϕ ) + u, ϕ0 , ϕ = c, ϕ , (2.4.15)
where c := u, ϕ0 ∈ C. Hence, u = c in D(I). By linearity, this readily implies the

statement in (2). The proof of the proposition is now complete.
Proposition 2.48. Let I be an open interval in R, g ∈ C ∞ (I), and f ∈ C k (I) for some
k ∈ N0 . If u ∈ D (I) satisfies u + gu = f in D (I), then u ∈ C k+1 (I).
x
Proof. Fix a ∈ I and define F(x) := e a g0 (t) dt for x ∈ I. Then F ∈ C ∞ (I) and we
may use (4) in Proposition 2.43 and the equation satisfied by u to write (keeping in
mind that f is continuous)
x
Fu − F(t) f (t) dt = F u + Fu − F f = 0 in D (I). (2.4.16)
a
x
By Proposition 2.47, Fu = a F(t) f (t) dt +c in D (I) for some constant c ∈ C. Since
x
a
F(t) f (t) dt ∈ C k+1 (I) and F1 ∈ C ∞ (I), we conclude that
x
1 c
u= F(t) f (t) dt + ∈ C k+1 (I), (2.4.17)
F(x) a F(x)
as desired.
We close this section by presenting a higher degree version of the product for-
mula for differentiation from part (4) in Proposition 2.43.
Proposition 2.49 (Generalized Leibniz Formula). Suppose f ∈ C ∞ (Ω) and let

u ∈ D (Ω). Then for every α ∈ Nn0 one has
α!
∂α ( f u) = (∂β f )(∂α−β u) in D (Ω). (2.4.18)
β≤α
β!(α − β)!
Proof. The first step is to observe that for each j ∈ {1, . . . , n} and each k ∈ N0 we
have k!
∂kj ( f u) = (∂j f )(∂k−
j u) in D (Ω), (2.4.19)
0≤≤k
!(k − )!
which is proved by induction on k making use of part (4) in Proposition 2.43. Hence,
given any α = (α1 , α2 , . . . , αn ) ∈ Nn0 , via repeated applications of (2.4.19) we obtain
α1 !
∂α1 1 ( f u) = (∂β1 f )(∂α1 1 −β1 u) (2.4.20)
0≤β1 ≤α1
β1 !(α1 − β1 )! 1
and
∂α1 1 ∂α2 2 ( f u) (2.4.21)

α1 ! α2 !
= (∂β1 ∂β2 f )(∂α1 1 −β1 ∂α2 2 −β2 u).
0≤β1 ≤α1 0≤β2 ≤α2
β1 !(α1 − β1 )! β2 !(α2 − β2 )! 1 2
By induction, we may then infer

2.5 The Support of a Distribution 37
α1 ! αn !
∂α ( f u) = ··· ··· ×
0≤β1 ≤α1 0≤βn ≤αn
β1 !(α1 − β1 )! βn !(αn − βn )!
× (∂β11 . . . ∂βnn f )(∂α1 1 −β1 . . . ∂αn n −βn u)

α!
= (∂β f )(∂α−β u), (2.4.22)
0≤β≤α
β!(α − β)!
as claimed.
2.5 The Support of a Distribution
In preparation to discussing the notion of support of a distribution, we first define

the restriction of a distribution to an open subset of the Euclidean domain on which
the distribution is considered. Necessarily, such a definition should generalize re-
strictions at the level of locally integrable functions. We start from
the observation
that if f ∈ Lloc
1
(Ω) and ω is a non-empty open subset of Ω, then f ω ∈ Lloc
1
(ω). Thus,

f ω , ϕ = f ϕ dx = f ι(ϕ) dx for each ϕ ∈ C0∞ (ω), (2.5.1)
ω Ω
where ι is the map from (1.3.16).
Proposition 2.50. Let Ω be a non-empty open subset of Rn and suppose ω is a non-

empty open subset of Ω. Also, recall the map ι from (1.3.16). Then for every u ∈
D (Ω), the mapping arising as the restriction of the distribution u to ω, i.e.,

uω : D(ω) → C defined by uω (ϕ) := u, ι(ϕ), ∀ ϕ ∈ C0∞ (ω), (2.5.2)

is linear and continuous. Hence, uω ∈ D (ω).
Proof. It is immediate that the map in (2.5.2) is well defined and linear. To see that
it is also continuous we use Proposition 2.4. Let K be a compact set contained in
ω. Then K ⊂ Ω and, since u ∈ D (Ω), Proposition 2.4 applies and gives k ∈ N0
and C ∈ (0, ∞) such that (2.1.1) holds. In particular, for each ϕ ∈ C0∞ (ω) with
supp ϕ ⊆ K,

u (ϕ) = u, ι(ϕ) ≤ C sup |∂α ϕ(x)|. (2.5.3)
ω
x∈K
|α|≤k

The conclusion
that uω ∈ D (ω) now follows. For an alternative proof of the conti-
nuity of uω one may use Fact 2.2 and Exercise 1.18.
Exercise 2.51.
(1) Prove that the definition of the restriction of a distribution from (2.5.2) general-
izes the usual restriction of functions. More specifically, using the notation in-
troduced in (2.1.6), show that if ω is an open subset of Ω and f ∈ Lloc

1
(Ω), then

u f ω = u f |ω in D (ω).
(2) Prove that the operation of differentiation of a distribution commutes with the
operation of restriction of a distribution to open sets, that is, if ω is an open
subset of Ω, then

∂α uω = (∂α u)ω , ∀ u ∈ D (Ω), ∀ α ∈ Nn0 . (2.5.4)
(3) Prove that the operation of multiplication of a distribution by a smooth function

behaves naturally relative to restriction to open subsets. Specifically, show that if
ω is an open subset of Ω then

(ϕu)ω = ϕω uω , ∀ u ∈ D (Ω), ∀ ϕ ∈ C ∞ (Ω). (2.5.5)
The next proposition shows that a distribution is uniquely determined by its local
behavior.
Proposition 2.52. If u1 , u2 ∈ D (Ω) are such that

for each
x0 ∈ Ω there exists an
open subset ω of Ω with x0 ∈ ω and satisfying u1 ω = u2 ω in D (ω), then u1 = u2 in
D (Ω).
Proof. Observe that this proposition may be viewed as a reconstruction problem,

thus it is meaningful to try to use a partition of unity. Let ϕ ∈ C0∞ (Ω) be arbitrary,
fixed and set K := supp ϕ. The goal is to prove that u1 , ϕ = u2 , ϕ. From hypothe-
x∈ K
ses it follows that for each there exists an open neighborhood ω x ⊂ Ω of x
with the property that u1 ω = u2 ω . Based on the fact that K is compact, the cover
x x
{ω x } x∈K of K may be refined to a finite one, consisting of, say ω1 , . . . , ωN . These are
open subsets of Ω and satisfy

N
K⊂ ωj and u1 ω = u2 ω for j = 1, . . . , N. (2.5.6)
j j
j=1
Consider a partition of unity {ψ j : j = 1, . . . , N} subordinate to the cover {ω j }Nj=1

of K, as given by Theorem 14.37. In particular, for each j = 1, . . . , N, the function
N
ψ j ∈ C0∞ (Ω) satisfies supp ψ j ⊂ ω j . Moreover, ψ j = 1 on K. Consequently, using
j=1
the linearity of distributions and (2.5.6), we obtain

N N
N
u1 , ϕ = u1 , ϕ ψj = u1 , ϕψ j = u2 , ϕψ j
j=1 j=1 j=1
N
= u2 , ϕψ j = u2 , ϕ. (2.5.7)
j=1
Thus, u1 , ϕ = u2 , ϕ and the proof of the proposition is complete.

2.5 The Support of a Distribution 39
Exercise 2.53. Let k ∈ N0 ∪ {∞} and suppose u ∈ D (Ω) is such that for each x ∈ Ω
there exists a number r x > 0 and a function f x ∈ C k (B(x, r x )) such that B(x, r x ) ⊂ Ω
and uB(x,r ) = f x in D (B(x, r x )). Prove that u ∈ C k (Ω).
x
Hint: Use Theorem 14.42 to obtain a partition of unity {ψ j } j∈J subordinate to the

cover {B(x, r x )} x∈Ω of Ω, then show that f := ψ j f j is a function in C k (Ω) satisfying
j∈J
u = f in D (Ω).
Now we are ready to define the notion of support of a distribution. Recall that
if f ∈ C 0 (Ω) then its support is defined to be the closure relative to Ω of the set
{x ∈ Ω : f (x) 0}. However, the value of an arbitrary distribution at a point is not
meaningful. The fact that Ω \ supp f is the largest open set contained in Ω on which
f = 0 suggests the introduction of the following definition.
Definition 2.54. The support of a distribution u ∈ D (Ω) is defined as
supp u (2.5.8)

:= x ∈ Ω : there is no ω open such that x ∈ ω ⊆ Ω and uω = 0 .
Based on (2.5.8), it follows that

Ω \ supp u = x ∈ Ω : ∃ ω open set such that x ∈ ω ⊆ Ω and uω = 0 , (2.5.9)
which is an open set. Hence, supp u is relatively closed in Ω. Moreover, if we apply

Proposition 2.52 to the distributions u and 0 ∈ D (Ω \ supp u) we obtain that

uΩ\supp u = 0, ∀ u ∈ D (Ω). (2.5.10)
In other words, Ω \ supp u is the largest open subset of Ω on which the restriction of
u is zero.
Example 2.55. Recall the Dirac distribution δ from (2.1.19). We claim that supp δ =
∞
Indeed, if ϕ ∈ C0 (R \{0}) it follows that δ, ϕ = ϕ(0) = 0. By Proposition 2.52,
{0}. n
δRn \{0} = 0, thus supp δ ⊆ {0}. To prove the opposite inclusion, consider an arbitrary
open subset ω of Ω such that 0 ∈ ω. Then there exists ϕ ∈ C0∞ (ω) such that ϕ(0) =
1, and hence, δ, ϕ = 1 0, which in turn implies that δω 0. Consequently,
0 ∈ supp δ as desired. Similarly, if x0 ∈ Rn , then supp δ x0 = {x0 }, where δ x0 is as in
Example 2.14.
Example 2.56. If f ∈ C 0 (Ω) then supp u f = supp f , where u f is the distribution
from (2.1.6). Indeed, since f = 0 in Ω \ supp f , we have Ω f (x)ϕ(x) dx = 0 for
every ϕ ∈ C0∞ (Ω \ supp f ), hence supp u f ⊆ supp f . Also, if x ∈ Ω \ supp u f then
there exists an open neighborhood ω of x with ω ⊆ Ω and such that u f ω = 0. Thus,

for every ϕ ∈ C0∞ (ω) one has 0 = u f , ϕ = ω f (x)ϕ(x) dx. Invoking Theorem 1.3
we arrive at the conclusion that f = 0 almost everywhere in ω hence, ultimately,
f = 0 in ω (since f is continuous in ω). Consequently, x supp f and this proves
that supp f ⊆ supp u f .
Exercise 2.57. Let u, v ∈ D (Ω) be such that supp u ∩ supp v = ∅ and u + v = 0 in

D (Ω). Prove that u = 0 and v = 0 in D (Ω).

Hint: Note that vΩ\supp u = (u + v)Ω\supp u = 0 and vΩ\supp v = 0. Combine these with
the fact that (Ω \ supp u) ∪ (Ω \ supp v) = Ω and Proposition 2.52 to deduce that
v = 0.
Exercise 2.58. Prove that
supp(∂α u) ⊆ supp u, ∀ u ∈ D (Ω), ∀ α ∈ Nn0 . (2.5.11)
Hint: Use (2.5.10) and (2.5.4).

We propose to extend the scope of the discussion in Example 2.56 as to make
it applicable to functions that are merely locally integrable (instead of continuous).
This requires defining a suitable notion of support for functions that lack continuity,
and we briefly address this issue first.
Given an arbitrary set E ⊆ Rn and an arbitrary function f : E → C, we define
the support of f as

supp f := x ∈ E : r > 0 such that f = 0 a.e. in B(x, r) ∩ E . (2.5.12)
From this definition one may check without difficulty that

E \ supp f = E ∩ B(x, r x ) (2.5.13)
x∈E\supp f
where for each x ∈ E\supp f the number r x > 0 is such that f = 0 a.e. in B(x, r x )∩E.
Moreover, since Rn has the Lindelöf property, the above union can be refined to a
countable one. Based on these observations, the following basic properties of the
support may be deduced:
supp f is a relatively closed subset of E, (2.5.14)
f = 0 a.e. in E \ supp f, (2.5.15)
supp f ⊆ F if F relatively closed subset of E, f = 0 a.e. on E \ F, (2.5.16)
supp f = supp g if g : E → C is such that f = g a.e. on E. (2.5.17)
In addition, if the set E ⊆ Rn is open and the function f : E → C is continuous, then

supp f may be described as the closure in E of the set {x ∈ E : f (x) 0}, which is
precisely our earlier notion of support in this context.
Exercise 2.59. If f ∈ Lloc

1
(Ω) then supp u f = supp f , where u f is the distribution
from (2.1.6).
Hint: Use (2.5.12), (2.5.9), part (1) in Exercise 2.51, and the fact that the injection
in (2.1.8) is one-to-one.
2.6 Compactly Supported Distributions and the Space E (Ω) 41
2.6 Compactly Supported Distributions and the Space E (Ω)
Next we discuss the issue of extending the action of a distribution u ∈ D (Ω) to

a subclass of C ∞ (Ω) that is possibly larger than C0∞ (Ω). Observe that if f belongs
1
to Lloc (Ω), the expression Ω f ϕ dx is meaningful for functions ϕ ∈ C ∞ (Ω) with
the property that supp f ∩ supp ϕ is a compact subset of Ω. A particular case is
when supp ϕ ∩ supp f = ∅ in which scenario Ω f ϕ dx = 0. This observation is the
motivation behind the following theorem.
Theorem 2.60. Let u ∈ D (Ω) and consider a relatively closed subset F of Ω satis-
fying supp u ⊆ F. Set
MF := {ϕ ∈ C ∞ (Ω) : supp ϕ ∩ F is a compact set in Rn }. (2.6.1)
Then there exists a unique linear map

u : MF → C satisfying the following condi-
tions:
u, ϕ = u, ϕ for every ϕ ∈ C0∞ (Ω), and
(i)
u, ϕ = 0 for every ϕ ∈ C ∞ (Ω) with supp ϕ ∩ F = ∅,
(ii)
where, if ψ ∈ MF then u, ψ denotes u(ψ).
Moreover, extensions of u constructed with respect to different choices of F act
in a compatible fashion. More precisely, if F1 , F2 ⊆ Ω are two relatively closed sets
in Ω with the property that supp u ⊆ F j , j = 1, 2, then u1 , ϕ =
u2 , ϕ for every
ϕ ∈ MF1 ∩ MF2 , where u1 ,
u2 , are the extensions of u constructed as above relative
to the sets F1 and F2 , respectively.
Before presenting the proof of this theorem, a few comments are in order.
Remark 2.61. Retain the context of Theorem 2.60.

(a) One has C0∞ (Ω) ⊆ MF and {ϕ ∈ C ∞ (Ω) : supp ϕ ∩ F = ∅} ⊆ MF .
(b) MF is a vector subspace of C ∞ (Ω), albeit not a topological subspace of E(Ω).
(c) If F = supp u we are in the setting discussed prior to the statement of Theo-
rem 2.60. Also, if F1 ⊆ F2 then MF2 ⊆ MF1 . In particular, the largest MF
corresponds to the case when F = supp u.
(d) If supp u is compact and we take F = supp u then MF = C ∞ (Ω). In such a
scenario, Theorem 2.60 gives an extension of u, originally defined as linear func-
tional on C0∞ (Ω), to a linear functional defined on the larger space C ∞ (Ω). From
a topological point of view, this extension turns out to be a continuous mapping
of E(Ω) into C (as we will see later, in Theorem 2.67).
Proof of Theorem 2.60. Fix a relatively closed subset F of Ω satisfying supp u ⊆ F.

First we prove the uniqueness statement in the first part of the theorem. Suppose
u1 ,
u2 : MF → C satisfy (i) and (ii). Fix ϕ ∈ MF and consider a function ψ ∈ C0∞ (Ω)

such that ψ ≡ 1 in an open neighborhood W of F ∩ supp ϕ. That such a function
ψ exists is guaranteed by Proposition 14.34. Decompose ϕ = ϕ0 + ϕ1 where ϕ0 :=
ψϕ ∈ C0∞ (Ω) and ϕ1 := (1 − ψ)ϕ ∈ C ∞ (Ω).
In general, if A ⊆ Rn and f ∈ C 0 (Rn ), it may be readily verified that f = 0

on A if and only if supp f ⊆ (Ac ). Making use of this observation we obtain that

supp (1 − ψ) ⊆ W c = W̊ c = W c . It follows that supp ϕ1 ⊆ W c ∩ supp ϕ, hence

supp ϕ1 ∩ F = ∅. Thus, by (i) and (ii) written for
u1 and
u2 , we have

u1 , ϕ =
u1 , ϕ0 +
u1 , ϕ1 = u, ϕ0 + 0 =
u2 , ϕ0
=
u2 , ϕ0 +
u2 , ϕ1 =
u2 , ϕ, (2.6.2)
which implies that u1 = u2 .

To prove the existence of an extension satisfying properties (i) and (ii), we make
use of the decomposition of ϕ already employed in the proof of uniqueness. The
apparent problem is that such a decomposition is not unique. However that is not
the case. Suppose ϕ ∈ MF is such that ϕ = ϕ0 + ϕ1 = ϕ0 + ϕ1 for some functions
ϕ0 , ϕ0 ∈ C0∞ (Ω) and ϕ1 , ϕ1 ∈ C ∞ (Ω) satisfying
supp ϕ1 ∩ F = ∅ = supp ϕ1 ∩ F.
Then, ϕ0 − ϕ0 = ϕ1 − ϕ1 , and since supp (ϕ1 − ϕ1 ) ∩ F = ∅, we also have
supp (ϕ0 − ϕ0 ) ∩ F = ∅,
which in turn implies supp (ϕ0 − ϕ0 ) ⊆ Ω \ supp u. The latter condition entails
0 = u, ϕ0 − ϕ0 = u, ϕ0 − u, ϕ0 .
This suggests defining the extension

u : MF −→ C,
u, ϕ := u, ψϕ for each ϕ ∈ MF and
(2.6.3)
each ψ ∈ C0∞ (Ω) with ψ ≡ 1 in a neighborhood of supp ϕ ∩ F.
Clearly u as in (2.6.3) is linear and, based on the previous reasoning, independent

of the choice of ψ, thus well defined. We claim that this extension also satisfies
(i) and (ii). Indeed, if ϕ ∈ C0∞ (Ω), we choose ψ ≡ 1 on supp ϕ. Then necessarily
u, ϕ = u, ϕ, so the extension in (2.6.3) satisfies (i). Also, if ϕ ∈ C ∞ (Ω) is such

that supp ϕ ∩ F = ∅, we may choose ψ ∈ C0∞ (Ω) such that supp ψ ∩ F = ∅ which
forces u, ϕ = u, ψϕ = 0, hence our extension satisfies (ii) as well. This proves
the claim.
We are left with proving the compatibility of extensions. Let F1 , F2 ⊆ Ω be
relatively closed sets in Ω each containing supp u. Denote by u1 and
u2 the linear
extensions of u to MF1 and MF2 , respectively, constructed as above relative to the
sets F1 and F2 . For ϕ ∈ MF1 ∩ MF2 let ψ ∈ C0∞ (Ω) be such that ψ ≡ 1 on an
open neighborhood of the set supp ϕ ∩ F1 and on an open neighborhood of the set
supp ϕ ∩ F2 . Then by (2.6.3), u1 , ϕ = u, ψϕ =
u2 , ϕ. The proof of the theorem
is now complete.
Remark 2.62 In the context of Theorem 2.60 consider u ∈ D (Ω), a ∈ C ∞ (Ω), and
α ∈ Nn0 . Then the extension given in Theorem 2.60 satisfies the following properties:
(1)
au, ϕ =
u, aϕ for every ϕ ∈ MF ;

(2) ∂ u, ϕ = (−1)|α|
α u, ∂α ϕ for every ϕ ∈ MF .
Indeed, since by Theorem 2.60 an extension with properties (i) and (ii) is unique,
the statement in (1) above will follow if one proves that the actions of the linear
functionals considered in the left- and right-hand sides of the equality in (1) coincide
on C0∞ (Ω) and on C ∞ (Ω) functions with supports outside F, which are immediate
from (2.6.3) and properties of distributions. A similar approach works for the proof
of (2).
We introduce the following notation
Dc (Ω) := {u ∈ D (Ω) : supp u is a compact subset of Ω}. (2.6.4)
By applying Theorem 2.60 to u ∈ Dc (Ω) and F := supp u, in which case MF =

C ∞ (Ω), it follows that there exists a linear map u : C ∞ (Ω) → C satisfying (i) and
(ii) in the statement of this theorem. In fact, this extension turns out to be continuous
with respect to the topology E(Ω), an issue that we will address shortly.
The dual of E(Ω) is the space
{v : E(Ω) → C : v linear and continuous}. (2.6.5)
Whenever v : E(Ω) → C is linear and continuous, and whenever ϕ ∈ C ∞ (Ω), we use

the notation v, ϕ in place of v(ϕ). The following is an equivalent characterization
of continuity for linear functionals on E(Ω) (see Section 14.1.0.2 for more details).
Fact 2.63 A linear functional v : E(Ω) → C is continuous (for details see (14.1.22))
if and only if there exist a compact K ⊂ Ω, a number m ∈ N0 , and a constant
C ∈ (0, ∞), such that
|v(ϕ)| ≤ C sup sup |∂α ϕ(x)| , ∀ ϕ ∈ C ∞ (Ω). (2.6.6)

α∈Nn0 , |α|≤m x∈K
In the current setting, functionals on E(Ω) are continuous if and only if they are
sequentially continuous. This can be seen by combining the general result presented
in Theorem 14.1 with Fact 1.8. A direct proof, applicable to the specific case of
linear functionals on E(Ω), is given in the next proposition.
Proposition 2.64. Let v : E(Ω) → C be a linear map. Then v is continuous if and
only if v is sequentially continuous.
Proof. The general fact that any linear and continuous functional on topological
vector spaces is sequentially continuous gives the left-to-right implication. To prove
the converse implication, it suffices to check continuity at zero. This is done reason-
ing by contradiction. Assume that
E(Ω)
v(ϕ j ) −−−−→ 0 whenever ϕ j −−−−→ 0, (2.6.7)
j→∞ j→∞
but that v is not continuous at 0 ∈ E(Ω). Then for each compact subset K of Ω and
every j ∈ N, there exists ϕ j ∈ E(Ω) such that
|v(ϕ j )| > j sup |∂α ϕ j (x)|. (2.6.8)

x∈K
|α|≤ j

∞
Consider now a nested sequence of compact sets {K j } j∈N such that K j = Ω. For
j=1
each j ∈ N, let ϕ j be as given by (2.6.8) corresponding to K := K j and define the
ϕ
function ψ j := v(ϕjj ) which belongs to E(Ω). Then
1
v(ψ j ) = 1 and sup |∂α ψ j (x)| ≤ for every j ∈ N. (2.6.9)
x∈K j , |α|≤ j j
Thus, for each fixed α ∈ Nn0 and every compact subset K of Ω there exists some
j0 ≥ |α| with the property that that K ⊂ K j0 and sup x∈K |∂α ψ j (x)| < 1j for all j ≥ j0 .
E(Ω)
The latter implies ψ j −−−−→ 0 which, in light of (2.6.7), further implies v(ψ j ) −−−−→ 0.
j→∞ j→∞
Since this contradicts the fact that v(ψ j ) = 1 for every j ∈ N, the proof is finished.

The topology we consider on the dual of E(Ω) is the weak∗-topology, and we
denote the resulting topological vector space by E (Ω) (see Section 14.1.0.2 for
more details). A significant byproduct of this set up is singled out next.
Fact 2.65 E (Ω) is a locally convex topological vector space over C, which is not
metrizable, but is complete.
In addition, we have the following important characterization of continuity in E (Ω).
Fact 2.66 A sequence {u j } j∈N ⊂ E (Ω) converges to u ∈ E (Ω) as j → ∞ in E (Ω),

E (Ω)
something we will indicate by writing u j −−−−→ u, if and only if u j , ϕ −−−−→ u, ϕ
j→∞ j→∞
for every ϕ ∈ E(Ω).
We are now ready to state and prove a result that gives a complete characteriza-
tion of the class of functionals that are extensions as in Theorem 2.60 of distributions
u ∈ D (Ω) with compact support.
Theorem 2.67. The spaces Dc (Ω) and E (Ω) are algebraically isomorphic.
Proof. Consider the mapping ι : Dc (Ω) → E (Ω), ι(u) :=

u, where
u is the extension
of u given by Theorem 2.60 corresponding to F := supp u. Then Msupp u = C ∞ (Ω)
and, to conclude that ι is well defined, there remains to show that the functional

u is continuous on E(Ω). With this goal in mind, note that while in general the
function ψ ∈ C0∞ (Ω) used in the construction of u (as in (2.6.3)) depends on ϕ,
given that we are currently assuming that supp u is compact, we may take ψ ≡ 1
on a neighborhood of supp u (originally we only needed ψ ≡ 1 on a neighborhood

of supp u ∩ supp ϕ ⊆ supp u). Let K0 := supp ψ. Then, for each ϕ ∈ C ∞ (Ω) we
have that ϕψ belongs to C0∞ (Ω), satisfies the support condition supp (ϕψ) ⊆ K0 , and
u, ϕ = u, ψϕ. Fix ϕ ∈ C ∞ (Ω). Since u ∈ D (Ω), corresponding to the compact

set K0 there exist k0 ∈ N0 and a finite constant C ≥ 0 such that
u, ϕ| = |u, ψϕ| ≤ C sup |∂α (ψϕ)|.

| (2.6.10)
x∈K0
|α|≤k0
Starting with Leibniz’s formula (14.2.6) applied to ψϕ, we estimate

α!
|∂ (ψϕ)| =
α
∂ ψ∂ ϕ ≤ C sup |∂β ϕ(x)|,
α−β β
(2.6.11)
β!(α − β)! x∈K0
β≤α
|β|≤k0
for some finite constant C = C (α, ψ) > 0. Combining (2.6.10) and (2.6.11), we
obtain
u, ϕ| ≤ C · C sup |(∂β ϕ)(x)|,
| (2.6.12)
x∈K0
|β|≤k0
hence u ∈ E (Ω), proving that ι is well defined.

Moving on, it is clear that ι is linear, hence to conclude that it is injective, it
suffices to show that if ι(u) = 0 for some u ∈ Dc (Ω), then u = 0. Consider u ∈ Dc (Ω)
such that ι(u) = 0. Then by (i) in Theorem 2.60, u = ι(u)C ∞ (Ω) = 0, as desired.
0
Consider now the task of proving that ι is surjective. To get started, pick an
arbitraryv ∈ E (Ω) and set u := vC ∞ (Ω) . Clearly u : D(Ω) → C is linear. Since
0
v ∈ E (Ω), Fact 2.63 ensures the existence of a compact set K ⊂ Ω, nonnegative
integer k, and finite constant C > 0, such that
|v, ϕ| ≤ C sup |∂α ϕ(x)|, ∀ ϕ ∈ C ∞ (Ω). (2.6.13)

x∈K
|α|≤k
Then, for each compact subset A of Ω and ϕ ∈ C0∞ (Ω) with supp ϕ ⊆ A, by regarding
ϕ as being in E(Ω) we may use (2.6.13) to write
|u, ϕ| ≤ C sup |∂α ϕ(x)| = C sup |∂α ϕ(x)| ≤ C sup |∂α ϕ(x)|. (2.6.14)
x∈K x∈K∩A x∈A
|α|≤k |α|≤k |α|≤k
From (2.6.14) we may now conclude (invoking Proposition 2.4) that u ∈ D (Ω).
Next, we claim that supp u ⊆ K. Indeed, if ϕ ∈ C0∞ (Ω) is a test function with
supp ϕ ∩ K = ∅ then from (2.6.13) we obtain |u, ϕ| = 0, thus u = 0 on Ω \ K.
Hence, the claim is proved which, in turn, shows that u ∈ Dc (Ω).
To finish the proof of the surjectivity of ι, it suffices to show that ι(u) = v. Denote
by uK the extension of u given by Theorem 2.60 with F := K. Then reasoning
as in the proof of the fact that ι is well defined, we obtain uK ∈ E (Ω). By (i) in
Theorem 2.60 it follows that
∞
uK C ∞ (Ω) = u. Also, if ϕ ∈ C (Ω) satisfies supp ϕ ∩
0
K = ∅, then by (ii) in Theorem 2.60 we have uK , ϕ = 0, while (2.6.13) implies

v, ϕ = 0. Hence, the uniqueness result in Theorem 2.60 yields uK = v. On the
other hand, since K and supp u are compact, we have MK = C ∞ (Ω) = Msupp u . Now
the last conclusion in Theorem 2.60 gives uK = ι(u). Consequently, ι(u) = v, and the
surjectivity of ι is proved. This finishes the proof of the theorem.
In light of the significance of Dc (Ω), Theorem 2.67 provides a natural algebraic
identification

E (Ω) = u ∈ D (Ω) : supp u is a compact subset of Ω . (2.6.15)
Remark 2.68. The spaces Dc (Ω) and E (Ω) are not topologically isomorphic since
there exist sequences of distributions with compact support that converge in D (Ω)
but not in E (Ω). For example, take the sequence {δ j } j∈N ⊂ D (R) of Dirac distribu-
tions with mass at j ∈ N, that have been defined in Example 2.14. Then it is easy to
check that the sequence {δ j } j converges to 0 in D (R) but not in E (R).
Theorem 2.67 nonetheless proves that the identity mapping is well defined from
E (Ω) into D (Ω). Keeping this in mind and relying on (1.3.15) and Proposition 14.4,
we see that
E (Ω) is continuously embedded into D (Ω). (2.6.16)
This corresponds to the dual version of (1.3.15). In particular, the operation of re-
striction to an open subset ω of Ω is a well-defined linear mapping

E (Ω) u → uω ∈ D (ω). (2.6.17)

Moreover, uω ∈ E (ω) whenever the support of u ∈ E (Ω) is contained in ω.
Proposition 2.69. Let ω and Ω be open subsets of Rn such that ω ⊆ Ω. Then every
u ∈ E (ω) extends to a functional
u ∈ E (Ω) by setting

u : E(Ω) → C,
u, ϕ := u, ψϕ, ∀ ϕ ∈ C ∞ (Ω), (2.6.18)
where ψ ∈ C0∞ (ω) is such that ψ ≡ 1 in a neighborhood of supp u.

Proof. We first claim that the mapping in (2.6.18) is well defined. To see why this
is the case, suppose ψ j ∈ C0∞ (ω), ψ j ≡ 1 in a neighborhood of supp u, for j = 1, 2.
Then for each function ϕ ∈ C ∞ (Ω) we have (ψ1 − ψ2 )ϕ ∈ C0∞ (ω) and in addition
supp u∩supp [(ψ1 −ψ2 )ϕ] = ∅. These further imply that necessarily u, (ψ1 −ψ2 )ϕ =
0. This proves that the definition of
u is independent of the choice of ψ with the given
properties. The functional defined in (2.6.18) is also linear while its continuity is a
consequence of Proposition 2.64 and Exercise 1.22. In addition, if ϕ ∈ C ∞ (ω) is
given and if ψ ∈ C0∞ (ω) is such that ψ ≡ 1 in a neighborhood of supp u, then
supp [(1 − ψ)ϕ] ∩ supp u = ∅. Hence

u, ϕ = u, ψϕ = u, (1 − ψ)ϕ + u, ϕ = u, ϕ, (2.6.19)
proving that
u is an extension of u.
Exercise 2.70. In the context of Proposition 2.69 prove that
u = ∂
∂α αu in D (Ω) (2.6.20)
for every u ∈ E (ω) and every α ∈ Nn0 .

Remark 2.71.
(1) In the sequel, we will often drop· from the notation of the extension (as defined
in the proof of Theorem 2.67 or Proposition 2.69) of a compactly supported
distribution. More precisely, if u ∈ Dc (Ω) we will simply use u for the extension
of u to a functional in E (Ω), as well as for its extension to a functional in E (O),
where O is an open subset of Rn containing Ω.
(2) Whenever necessary, if u ∈ E (Ω), ϕ ∈ C0∞ (Ω), and ψ ∈ C ∞ (Ω), we will use
the notation D u, ϕD for the action of u on ϕ as a functional in D (Ω), and the
notation E u, ψE for the action of u on ψ as a functional in E (Ω).
Proposition 2.72. Let u ∈ D (Ω) and ψ ∈ C0∞ (Ω). Then ψu ∈ E (Ω) and
E ψu, ϕE = D u, ψϕD , ∀ ϕ ∈ E(Ω). (2.6.21)
Proof. Since ψu ∈ D (Ω) and supp ψu ⊆ supp ψ, we have ψu ∈ Dc (Ω), thus ψu ∈
E (Ω) (i.e., ψu extends as an element in E (Ω)). Let φ ∈ C0∞ (Ω) be such that φ ≡ 1
on a neighborhood of supp ψ. Then for every ϕ ∈ E(Ω),
E ψu, ϕE = D ψu, φϕD = D u, ψφϕD
= D u, ψϕD (2.6.22)
proving (2.6.21).
D (Ω)
Exercise 2.73. Let u j −−−−→ u be such that there exists a compact K in Rn that is
j→∞
contained in Ω and with the property that supp u j ⊆ K for every j ∈ N. Prove that
E (Rn ) E (Ω)
supp u ⊆ K and u j −−−−→ u. Consequently, we also have u j −−−−→ u.
j→∞ j→∞
Exercise 2.74. Let k ∈ N0 and assume that cα ∈ C for α ∈ Nn0 with |α| ≤ k. Prove
that
cα ∂α δ = 0 in D (Rn ) ⇐⇒ each cα = 0. (2.6.23)
|α|≤k
Hint: Use (14.2.7).

Exercise 2.75. Prove that if u ∈ D (Rn ) and supp u ⊆ {a} for some a ∈ Rn , then u
has a unique representation of the form

u= cα ∂α δa , (2.6.24)
|α|≤k
for some k ∈ N0 and coefficients cα ∈ C.

Sketch of proof:
(I) Via a translation, reduce matters to the case a = 0.
(II) Use Fact 2.63 to determine k ∈ N0 .
(III) Fix ψ ∈ C0∞ (B(0, 1)) such that ψ ≡ 1 on B(0, 12 ) and for ε > 0 define the
function ψε (x) := ψ( εx ) for every x ∈ Rn . Prove that u = ψε u in D (Rn ).
(IV) For ϕ ∈ C0∞ (Rn ) consider the k-th order Taylor polynomial for ϕ at 0, i.e.,
1
ϕk (x) := ∂β ϕ(0) xβ , ∀ x ∈ Rn . (2.6.25)
|β|≤k
β!
Prove that for α ∈ Nn0 satisfying |α| ≤ k one has
∂α (ϕ − ϕk ) = ∂α ϕ − (∂α ϕ)k−|α| .
(V) Show that for each ϕ ∈ C0∞ (Rn ) there exists some constant c ∈ (0, ∞) such that
|u, (ϕ − ϕk )ψε | ≤ c ε.
(VI) Combine all the above to obtain that
(−1)|α|
u, ϕ = u, xα ∂α δ , ϕ ∀ ϕ ∈ C0∞ (Rn ). (2.6.26)
|α|≤k
α!
(VII) Prove that the representation in (VI) is unique.
Example 2.76. Let m ∈ N. We are interested in solving the equation
xm u = 0 in D (R). (2.6.27)
In this regard, assume that u ∈ D (R) solves (2.6.27) and note that if ϕ belongs to
C0∞ (R \ {0}), then x1m ϕ ∈ C0∞ (R \ {0}). This observation permits us to write
1
u, ϕ = xm u, m ϕ = 0, ∀ ϕ ∈ C0∞ (R \ {0}), (2.6.28)
x
which proves that supp u ⊆ {0}. In particular, u ∈ E (R). Applying Exercise 2.75 we
N
conclude that there exists N ∈ N0 such that u = ck δ (k) in D (R), for some ck ∈ C,
k=0
k = 0, 1, 2, . . . , N. We claim that
c = 0 whenever m ≤ ≤ N. (2.6.29)
To see why this is true, observe that since u ∈ E (R) it makes sense to apply u to
any function in C ∞ (R). In particular, it is meaningful to apply u to any polynomial.
Concerning (2.6.29), if N ≤ m − 1 there is nothing to prove, while in the case when
N ≥ m for each ∈ {m, . . . , N} we may write

N
0 = xm u, x−m = u, x = ck δ (k) , x
k=0

N dk

= (−1)k ck (x ) x=0 = (−1) ! c . (2.6.30)
k=0
dxk
This proves (2.6.29) which, in turn, forces u to have the form

m−1
u= ck δ (k) for some ck ∈ C, k = 0, 1, . . . , m − 1. (2.6.31)
k=0
Conversely, one may readily verify that any distribution u as in (2.6.31) solves
(2.6.27). In conclusion, any solution u of (2.6.27) is as in (2.6.31).
Exercise 2.77. Let m ∈ N, a ∈ R, and u ∈ D (R). Prove u is a solution of the

m−1
equation (x − a)m u = 0 in D (R) if and only if u is of the form u = ck δa(k) for
k=0
some ck ∈ C, k = 0, 1, . . . , m − 1.
Remark 2.78. You have seen in Example 2.34 that P.V. 1x is a solution of the equa-
tion xu = 1 in D (R). Hence, if v ∈ D (R) is another solution of this equation, then
the distribution v − P.V. 1x is a solution of the equation xu = 0 in D (R). By Exam-
ple 2.76, it follows that v − P.V. 1x = c δ, where c ∈ C. Thus, the general solution of
the equation xu = 1 in D (R) is u = P.V. 1x + c δ, for c ∈ C.
Example 2.79. Let N ∈ N and a j ∈ Rn , j ∈ {1, . . . , N}, be a finite family of distinct

points. If u ∈ D (Rn ) is such that supp u ⊆ {a1 , a2 , . . . , aN }, then u has a unique
representation of the form

N
u= cα, j ∂α δa j , k j ∈ N0 , cα, j ∈ C. (2.6.32)
j=1 |α|≤k j
To justify formula (2.6.32), fix a family of pairwise disjoint balls B j := B(a j , r j ),

j ∈ {1, . . . , N}, and for each j select a function ψ j ∈ C0∞ (B j ) satisfying ψ j ≡ 1
in a neighborhood of B(a j , r j /2). Then for each j ∈ {1, . . . , N} we have that ψ j u
is a compactly supported distribution in Rn with supp (ψ j u) ⊆ {a j }. We may now
apply Exercise 2.75 to obtain that there exist k j ∈ N0 and cα, j ∈ C such that ψ j u =

N
cα, j ∂α δa j in D (Rn ). In addition, since ψ j ≡ 1 in a neighborhood of supp u,
|α|≤k j j=1
we have

N
N
N
N
uψ j , ϕ = ψ j u, ϕ = u, ψ j ϕ = u, ϕ ψj
j=1 j=1 j=1 j=1
= u, ϕ for each ϕ ∈ C0∞ (Rn ). (2.6.33)


N
Hence, u = ψ j u in D (Rn ) which, given (2.6.33), proves (2.6.32).
j=1
Example 2.80. Let a, b ∈ R be such that a b. We are interested in solving the

equation
(x − a)(x − b)u = 0 in D (R). (2.6.34)
The first observation is that any solution u of this equation satisfies the support
condition supp u ⊆ {a, b}. Indeed, if we take an arbitrary ϕ ∈ C0∞ (R \ {a, b}), then
∞
(x−a)(x−b) ϕ belongs to the space C 0 (R \ {a, b}) and
1

u, ϕ = (x − a)(x − b)u , 1
(x−a)(x−b) ϕ = 0. (2.6.35)
Hence, we may apply Example 2.79 to conclude that

N1
N2
u= c j δa( j) + d j δb( j) in D (R), (2.6.36)
j=0 j=0
where N1 , N2 ∈ N, {c j }0≤ j≤N1 ⊂ C and {d j }0≤ j≤N2 ⊂ C. Moreover, by dropping terms

with zero coefficients, there is no loss of generality in assuming that
cN1 0 and dN2 0. (2.6.37)
In this scenario, we make the claim that N1 = N2 = 0. To prove this claim, suppose
first that N1 ≥ 1. Then, using (2.6.36) and the hypotheses on u, we obtain

0 = (x − a)(x − b)u, (x − a)N1 −1 (x − b)N2 = u, (x − a)N1 (x − b)N2 +1

N1

N2

= c j δa( j) , (x − a)N1 (x − b)N2 +1 + d j δb( j) , (x − a)N1 (x − b)N2 +1
j=0 j=0

N1 dj
= (−1) j c j j
(x − a)N1 (x − b)N2 +1
j=0
dx x=a

N2 dj
N2 +1
+ (−1) j d j (x − a) N1
(x − b) x=b
j=0
dx j
= (−1)N1 cN1 N1 ! (a − b)N2 +1 . (2.6.38)
Since by assumption a b, from (2.6.38) we obtain cN1 = 0. This contradicts

(2.6.37) and shows that necessarily N1 = 0. Similarly, we obtain that N2 = 0, hence
any solution of (2.6.34) has the form
u = c δa + d δb in D (R), c, d ∈ C. (2.6.39)
2.7 Tensor Product of Distributions 51
Conversely, it is clear that any distribution as in (2.6.39) solves (2.6.34). To sum up,
(2.6.39) describes all solutions of (2.6.34).
2.7 Tensor Product of Distributions
Let m, n ∈ N, U be an open subset of Rm , V be an open subset of Rn , and consider

two complex-valued functions f ∈ Lloc1
(U) and g ∈ Lloc1
(V). Then the tensor
product of the functions f and g is defined as
f ⊗ g : U × V → C, ( f ⊗ g)(x, y) := f (x)g(y) for each (x, y) ∈ U × V. (2.7.1)
In particular, it follows from (2.7.1) that f ⊗ g ∈ Lloc

1
(U × V). When f , g, and f ⊗ g
∞
are regarded as distributions, for each ϕ ∈ C0 (U × V) we obtain that

f ⊗ g, ϕ = f (x)g(y)ϕ(x, y) dx dy = f (x) g(y)ϕ(x, y) dy dx
U×V U V

= g(y) f (x)ϕ(x, y) dx dy (2.7.2)
V U
or, concisely,

f ⊗ g, ϕ = f (x), g(y), ϕ(x, y) = g(y), f (x), ϕ(x, y) . (2.7.3)
If, in addition, the test function ϕ has the form ϕ1 ⊗ ϕ2 , for some ϕ1 ∈ C0∞ (U) and
ϕ2 ∈ C0∞ (V), then (2.7.3) becomes
f ⊗ g, ϕ1 ⊗ ϕ2 = f, ϕ1 g, ϕ2 . (2.7.4)
This suggests a natural way to define tensor products of general distributions granted
the availability of the following density result for D(U × V).
Proposition 2.81. Let m, n ∈ N, U be an open subset of Rm , and V be an open

subset of Rn . Then the set
⎧ ⎫
⎪
⎪
⎪ N ⎪
⎪
⎪
∞ ∞ ⎨ ∞ ∞ ⎬
C0 (U) ⊗ C0 (V) := ⎪ ⎪ ϕ j ⊗ ψ j : ϕ j ∈ C0 (U), ψ j ∈ C0 (V), N ∈ N⎪
⎪ (2.7.5)
⎪
⎩ j=1 ⎪
⎭
is sequentially dense in D(U × V).
Before proceeding with the proof of Proposition 2.81 we state and prove two
lemmas.
Lemma 2.82. Suppose that the sequence { f j } j∈N ⊂ E(Rn ) and f ∈ E(Rn ) are such
that
1
∂α f j − ∂α f L∞ (B(0, j)) < for all α ∈ Nn0 satisfying |α| ≤ j. (2.7.6)
j
E(Rn )
Then f j −−−−→ f .
j→∞
Proof. Suppose { f j } j∈N and f satisfy the current hypotheses. Pick an arbitrary
ε > 0, a multi-index α ∈ Nn0 , and fix a compact subset K of Rn . Then there
exists j0 ∈ N such that K ⊂ B(0, j0 ). If we now fix j∗ ∈ N with the property
that j∗ > max{ 1ε , |α|, j0 }, it follows that for each j ≥ j∗ we have |α| ≤ j∗ ≤ j and
1 1
∂α f j − ∂α f L∞ (K) ≤ ∂α f j − ∂α f L∞ (B(0, j)) < < ∗ < ε. (2.7.7)
j j
Hence, ∂α f j converges uniformly on K to ∂α f . Since α and K are arbitrary, we
E(Rn )
conclude that f j −−−−→ f .
j→∞
Lemma 2.83. For every f ∈ C0∞ (Rn ) there exists a sequence {P j } j∈N of polynomials
E(Rn )
in Rn such that P j −−−−→ f .
j→∞
Proof. For each t > 0 define the function

|x−y|2
ft (x) := (4πt)− 2 e− 4t f (y) dy,
n
∀ x ∈ Rn . (2.7.8)
Rn
The first goal is to prove that

E(Rn )
ft −−−−→
+
f. (2.7.9)
t→0
To this end, consider the function u defined by

⎧
⎪
⎪
⎨ ft (x), x ∈ R , t > 0,
⎪ n
u(x, t) := ⎪
⎪ ∀ (x, t) ∈ Rn × [0, ∞). (2.7.10)
⎪
⎩ f (x), x ∈ Rn , t = 0,
From definition it is clear that u is continuous on Rn × (0, ∞). We claim that, in fact,
u is continuous
√ on Rn × [0, ∞). Indeed, by making use of the change of variables
x − y = 2 tz, we may write

− n2
√
e−|z| f (x − 2 tz) dz,
2
u(x, t) = π ∀ x ∈ Rn , ∀ t > 0. (2.7.11)
Rn
Hence, for each x∗ ∈ Rn , Lebesgue Dominated Convergence Theorem (cf. Theo-

rem 14.15) gives

− n2
e−|z| dz = f (x∗ ) = u(x∗ , 0),
2
lim
x→x
u(x, t) = f (x∗ )π (2.7.12)
∗
t→0+ Rn
proving that u is continuous at points of the form (x∗ , 0).

Being continuous on Rn × [0, ∞), u is uniformly continuous on every compact

subset of Rn ×[0, ∞), thus uniformly continuous on sets of the form K ×[0, 1], where
K ⊂ Rn is compact. Fix such a compact K and fix ε ∈ (0, 1) arbitrary. Then, there
exists δ > 0 such that if (x1 , t1 ), (x2 , t2 ) ∈ K × [0, 1] satisfy |(x1 , t1 ) − (x2 , t2 )| ≤ δ then
we necessarily have |u(x1 , t1 ) − u(x2 , t2 )| ≤ ε. In particular, if x ∈ K and t ∈ (0, δ),
then |u(x, t) − u(x, 0)| ≤ ε, that is ft − f L∞ (K) ≤ ε for all 0 < t < δ. This proves that
lim+ ft (x) = f (x) uniformly on compact sets in Rn .
t→0
The derivatives of ft enjoy the same type of properties as ft . More precisely,
√ by
a direct computation (involving also the change of variables x − y = 2 tz) we see
that for each t > 0 and each α ∈ Nn0 we have

|x−y|2
α − n2
∂ x ft (x) = (4πt) e− 4t (∂α f )(y) dy, ∀ x ∈ Rn . (2.7.13)
Rn
In addition, as before, we obtain that lim+ (∂αx ft ) = ∂α f uniformly on compact sets in

t→0
Rn . This completes the proof of (2.7.9).
Next, recall that the Taylor expansion of the function e s , s ∈ R, about the origin

∞ j
is e s = s
j! with the series converging uniformly on compact subsets of R. In
j=0

N
sj
addition, for each N ∈ N, the remainder RN (s) := e s − j! satisfies |RN (s)| ≤
j=0
N+1
C
eC (N+1)! whenever |s| ≤ C. Fix t > 0 and a compact subset K of Rn . Then there
|x−y|2
exists C > 0 such that 4t ≤ C for every x ∈ K and every y ∈ supp f , so
|x − y|2 N+1
lim RN − ≤ lim eC C = 0. (2.7.14)
N→∞ 4t N→∞ (N + 1)!
Consequently,
∞ j
1 |x − y|2
∂αx ft (x) = (4πt) − n2
− ∂α f (y) dy, (2.7.15)
j=0
j! Rn 4t
for each α ∈ Nn0 and each t > 0, and the series in (2.7.15) converges uniformly for x
in a compact set in Rn . In addition, integrating by parts, we may write

∞ |x − y|2 j
(−1)|α|
∂αx ft (x) = (4πt) − n2
∂αy − f (y) dy, (2.7.16)
j=0
j Rn 4t
where, for each t > 0 fixed, the series in (2.7.16) converges uniformly on compact
sets in Rn . Hence, if for each t > 0 we define the sequence of polynomials
k j
1 |x − y|2
Pt,k (x) := (4πt)− 2
n
− f (y) dy, ∀ x ∈ Rn , ∀ k ∈ N, (2.7.17)
j=0
j! Rn 4t
E(Rn )
then the above proof implies that for each t > 0 we have Pt,k −−−−→ ft .
k→∞
Next, we claim that there exists a sequence of positive numbers {t j } j∈N with the
property that for each j ∈ N we have
∂α ft j − ∂α f L∞ (B(0, j)) < 1

2j for every α ∈ Nn0 with |α| ≤ j. (2.7.18)
To construct a sequence {t j } j∈N satisfying (2.7.18) we proceed by induction. First,

consider the compact set B(0, 1). For each α ∈ Nn0 satisfying |α| ≤ 1, based on
(2.7.9), there exists α1 ∈ N with the property that
'' α '
'∂ ft − ∂α f ''L∞ (B(0,1)) < 12 for all t ∈ 0, 1/α1 . (2.7.19)
(1 )
Define t1 := min : α ∈ N n
, |α| ≤ 1 .
α1 0
Suppose that, for some j ≥ 2, we have already selected t1 , . . . , t j−1 satisfying

(2.7.18). Let α ∈ Nn0 be such that |α| ≤ j. Based on (2.7.9), there exists αj ∈ N with
the property that αj ≥ αj−1 whenever |α| ≤ j − 1, and such that
'' α '
'∂ ft − ∂α f ''L∞ (B(0, j)) < 21j for all t ∈ 0, 1/αj . (2.7.20)
(1 )
Now define t j := min : α ∈ Nn0 , |α| ≤ j . In particular, this choice ensures that
αj
t j ≤ t j−1 . Proceeding by induction it follows that the sequence {t j } j∈N constructed in
this manner satisfies (2.7.18).
Our next claim is that for each t > 0 and each j ∈ N there exists kt, j ∈ N such
that
∂α Pt,kt, j − ∂α ft L∞ (B(0, j)) < 1

2j for every α ∈ Nn0 with |α| ≤ j. (2.7.21)
E(Rn )
To prove this, fix t > 0 and j ∈ N. Since Pt,k −−−−→ ft and B(0, j) is a compact subset
k→∞
of Rn , it follows that for each α ∈ Nn0 satisfying |α| ≤ j there exists kα∗ ∈ N such that
∂α Pt,k − ∂α ft Lα (B(0, j)) < 2j,

1
for k ≥ kα∗ . (2.7.22)
If we now define kt, j := max{kα∗ : α ∈ Nn0 , |α| ≤ j}, then estimate (2.7.21) holds for
this kt, j . This completes the proof the claim.
Here is the endgame in the proof of the lemma. For each j ∈ N, let t j > 0 be
as constructed above so that (2.7.18) holds, for this t j let kt j , j be as defined above so
that (2.7.21) holds, and set P j := Pt j ,kt j , j . Hence, for each j ∈ N and every α ∈ Nn0
satisfying |α| ≤ j we have
∂α P j − ∂α f L∞ (B(0, j)) = ∂α Pt j ,kt j , j − ∂α f L∞ (B(0, j))

≤ 1
2j + 1
2j = 1j . (2.7.23)
E(R )
n
The fact that P j −−−−→ f now follows from (2.7.23) by invoking Lemma 2.82.
j→∞
Before turning to the proof of Proposition 2.81 we introduce some notation. For
m, n ∈ N, if U is an open subset of Rm , V is an open subset of Rn , and A ⊆ U × V,
the projections of A on U and V, respectively, are
πU (A) := {x ∈ U : ∃ y ∈ V such that (x, y) ∈ A},

(2.7.24)
πV (A) := {y ∈ V : ∃ x ∈ U such that (x, y) ∈ A}.
We are ready to present the proof of the density result stated at the beginning of
this section.
Proof of Proposition 2.81. Let ϕ ∈ C0∞ (U × V). By Lemma 2.83, there exists a
E(Rn+m )
sequence of polynomials {P j } j∈N in Rn+m with the property that P j −−−−−−→ ϕ. Set
j→∞
K := supp ϕ, K1 := πU (K), and K2 := πV (K). Then K1 and K2 are compact sets
in Rm and Rn , respectively. Fix a compact set L1 ⊂ U such that K1 ⊂ L˚1 and a
compact set L2 ⊂ V such that K2 ⊂ L˚2 . Then there exists a function ϕ1 ∈ C0∞ (U)
with supp ϕ1 ⊆ L1 , ϕ1 ≡ 1 in a neighborhood of K1 , and a function ϕ2 ∈ C0∞ (V)
satisfying supp ϕ2 ⊆ L2 and ϕ2 ≡ 1 in a neighborhood of K2 . Consequently,
ϕ1 ⊗ ϕ2 ∈ C0∞ (Rn+m ) and supp (ϕ1 ⊗ ϕ2 ) ⊆ L1 × L2 . (2.7.25)
By (2) in Exercise 1.10 it follows that

E(Rn+m )
(ϕ1 ⊗ ϕ2 )P j −−−−−−→ (ϕ1 ⊗ ϕ2 )ϕ. (2.7.26)
j→∞
Hence, since
supp [(ϕ1 ⊗ ϕ2 )P j ] ⊆ L1 × L2 for every j ∈ N
and since (ϕ1 ⊗ ϕ2 )ϕ = ϕ, we obtain
D(U×V)
(ϕ1 ⊗ ϕ2 )P j −−−−−−→ ϕ. (2.7.27)
j→∞
Upon observing that (ϕ1 ⊗ ϕ2 )P j ∈ C0∞ (U) ⊗ C0∞ (V) for every j ∈ N, the desired
conclusion follows.
The next proposition is another important ingredient used to define the tensor
product of two distributions.
Proposition 2.84. Fix m, n ∈ N, let U be an open subset of Rm , and let V be an open
subset of Rn . Then for each distribution u ∈ D (U) the following properties hold.
(a) If for each ϕ ∈ C0∞ (U × V) we define the mapping

ψ : V → C, ψ(y) := u(x), ϕ(x, y) for all y ∈ V, (2.7.28)
then ψ ∈ C0∞ (V).

(b) The mapping D(U × V) ϕ → ψ ∈ D(V), with ψ as defined in (a), is linear and
continuous.
Prior to presenting the proof of this result we make the following remark.
Remark 2.85. In the definition of ψ in part (a) of Proposition 2.84, the use of the
notation u(x) does NOT mean that the distributions u is evaluated at x since the

latter is not meaningful. The notation u(x), ϕ(x, y) should be understood in the
following sense: for each y ∈ V fixed, the distribution u acts on the function ϕ(·, y).
We now turn to the task of presenting the proof of Proposition 2.84.

Proof of Proposition 2.84. Fix ϕ ∈ C0∞ (U × V) and let K := supp ϕ that is a compact
subset of U × V. Also, consider ψ as in (2.7.28) and recall the projections πU , πV
from (2.7.24). Then clearly supp ψ ⊆ πV (K), thus ψ has compact support. Next we
prove that ψ is continuous on V. Let {y j } j∈N be a sequence in V such that lim y j = y0
j→∞
for some y0 ∈ V. Since u ∈ D (U), based on the definition of ψ and Fact 2.2, in order
D(U)
to conclude that lim ψ(y j ) = ψ(y0 ) it suffices to show that ϕ(·, y j ) −−−−→ ϕ(·, y0 ). It
j→∞ j→∞
is clear that for every j ∈ N we have ϕ(·, y j ) ∈ C0∞ (U) and supp ϕ(·, y j ) ⊆ πU (K).
Moreover, since ϕ ∈ C ∞ (U × V) it follows that ∂αx ϕ is continuous on K for every
α ∈ Nm 0 , thus uniformly continuous on K. Consequently,
(∂αx ϕ)(·, y j ) −−−−→ (∂αx ϕ)(·, y0 ) uniformly on πU (K).

j→∞
This completes the proof of the fact that ψ is continuous on V.

To continue, we claim that ψ is of class C 1 on V. Fix y ∈ V and j in {1, ..., n}.
Recall that e j is the unit vector in Rn with the j-th component equal to 1, and let
h ∈ R \ {0}. Since V is open, there exists ε0 > 0 such that if |h| < ε0 then y + he j ∈ V.
Make the standing assumption that |h| < ε0 and set
ϕ(x, y + he j ) − ϕ(x, y) ∂ϕ
Rh (x, y) := − (x, y), ∀ x ∈ U. (2.7.29)
h ∂y j
Then
ψ(y + he j ) − ψ(y) ∂ϕ
− u(x), (x, y) = u(x), Rh (x, y), ∀ x ∈ U. (2.7.30)
h ∂y j
Suppose
lim Rh (·, y) = 0 in D(U). (2.7.31)
h→0
Then limu, Rh (·, y) = 0, which in view of (2.7.30) implies

h→0
∂ϕ
∂ j ψ(y) = u, (·, y) . (2.7.32)
∂y j
∂ϕ
Moreover, since ∂y j
∈ C0∞ (U × V) by reasoning as in the proof of the continuity of
ψ on V, we also obtain that ∂ j ψ is continuous on V. Hence, since j ∈ {1, ..., n} is
arbitrary, to complete the proof of the claim, we are left with showing (2.7.31).
Clearly supp [Rh (·, y)] ⊆ πU (K). Applying Taylor’s formula to ϕ in the variable y
for each fixed x ∈ U we obtain
1
∂ϕ ∂2 ϕ
ϕ(x, y + he j ) = ϕ(x, y) + h (x, y) + h2 (1 − t) 2 (x, y + the j ) dt. (2.7.33)
∂y j 0 ∂y j
Hence, (2.7.29) and (2.7.33) imply

1
∂2 ϕ
Rh (x, y) = h (1 − t) (x, y + the j ) dt. (2.7.34)
0 ∂y2j
Consequently, for every β ∈ Nm

0 , we have
1
∂2 ϕ
∂βx Rh (x, y) =h (1 − t)∂βx (x, y + the j ) dt, ∀ x ∈ U. (2.7.35)
0 ∂y2j
Since the integral in the right-hand side of (2.7.35) is bounded by a constant in-
dependent of h, x and y, it follows that lim ∂βx Rh (·, y) = 0 uniformly on πU (K).
h→0
Combined with the support information on Rh (·, y), this implies (2.7.31) and com-
pletes the proof of the claim that ψ ∈ C 1 (V). By induction, we obtain ψ ∈ C ∞ (V),
completing the proof of the statement in part (a) of the proposition.
The linearity of the mapping in part (b) is immediate since u is a linear mapping.
To show that the mapping in (b) is also continuous, since D(V) is locally convex,
by Theorem 14.6 it suffices to prove that it is sequentially continuous. To this end,
D(U×V)
let ϕ j −−−−−−→ ϕ. In particular, there exists a compact subset K of U × V such that
j→∞
supp ϕ j ⊆ K for all j ∈ N and
∂α ϕ j −−−−→ ∂α ϕ uniformly on K, for every α ∈ N0m+n . (2.7.36)

j→∞
To proceed, for every y ∈ V set ψ(y) := u, ϕ(·, y) and ψ j (y) := u, ϕ j (·, y) for
D(V)
each j ∈ N. The goal is to prove that ψ j −−−−→ ψ. Applying Proposition 2.4 to the
j→∞
distribution u and compact πU (K) yields k ∈ N0 and C > 0 for which (2.1.1) holds
with K replaced by πU (K). Then, for every β ∈ Nn0 , we have

sup ∂β ψ j (y) − ∂β ψ(y) = sup u(x), ∂βy ϕ j (·, y) − ∂βy ϕ(·, y)
y∈πV (K) y∈πV (K)

≤ sup sup ∂γx ∂βy ϕ j (x, y) − ∂γx ∂βy ϕ(x, y)
y∈πV (K) x∈πU (K)
|γ|≤k

= C sup ∂γx ∂βy ϕ j (x, y) − ∂γx ∂βy ϕ(x, y) −−−−→ 0, (2.7.37)
(x,y)∈K j→∞
|γ|≤k
where γ ∈ Nm
0 and the convergence to zero in (2.7.37) is due to (2.7.36) applied for
D(V)
α := (γ, β). Thus, ψ j −−−−→ ψ and the proof of the statement in part (b) is complete.
j→∞

Remark 2.86. Let m, n ∈ N, and consider an open subset U of Rm along with an
open subset V of Rn .
(1) For each ϕ ∈ C ∞ (V × U) define the function
ϕ : U × V → C, ϕ (x, y) := ϕ(y, x), ∀ x ∈ U, ∀ y ∈ V. (2.7.38)
Then ϕ ∈ C ∞ (U × V). In addition, if ϕ has compact support in V × U then

ϕ also has compact support in U × V. Moreover, by invoking Fact 1.15, Theo-
rem 14.6, and Fact 1.16, we obtain that the mapping
D(V × U) ϕ → ϕ ∈ D(U × V) (2.7.39)
is linear and continuous.

(2) Let v ∈ D (V) and for each ϕ ∈ C0∞ (U × V) define the mapping

η : U → C, η(x) := v(y), ϕ(x, y) for all x ∈ U. (2.7.40)
Then η ∈ C0∞ (U) and the mapping D(U × V) ϕ → η ∈ D(U) is linear and
continuous. Indeed, this follows by applying Proposition 2.84 with the current
V, U, v, ϕ , x, y in place of U, V, u, ϕ, y, x.
(3) The fact that the map in (2.7.39) is linear and continuous yields the following
result:
if w ∈ D (U × V), then the map w : D(V × U) → C defined by
w (ϕ) := w, ϕ for each ϕ ∈ D(V × U) is actually a distribution on
V × U, i.e., w ∈ D (V × U).
(2.7.41)
We are now ready to define the tensor product of distributions.
Theorem 2.87. Let m, n ∈ N, U be an open subset of Rm , and V be an open subset
of Rn . Consider u ∈ D (U) and v ∈ D (V). Then the following statements are true.
(i) There exists a unique distribution u ⊗ v ∈ D (U × V), called the tensor
product of u and v, with the property that

u ⊗ v, ϕ1 ⊗ ϕ2 = u, ϕ1 v, ϕ2
(2.7.42)
∀ ϕ1 ∈ C0∞ (U), ∀ ϕ2 ∈ C0∞ (V).
(ii) The action of the distributions u ⊗ v ∈ D (U × V) is given by

u ⊗ v, ϕ = v(y), u(x), ϕ(x, y) (2.7.43)

= u(x), v(y), ϕ(x, y) for each ϕ ∈ C0∞ (U × V).
(iii) The tensor product just defined satisfies u ⊗ v = (v ⊗ u) in D (U × V), where
(v ⊗ u) is the distribution defined in (2.7.41) corresponding to u replaced by
v ⊗ u.
Proof. For each ϕ ∈ C0∞ (U × V) consider the function
ψ(y) := u(x), ϕ(x, y) for y ∈ V.
By Proposition 2.84, we have ψ ∈ C0∞ (V) and the mapping
D(U × V) ϕ → ψ ∈ D(V) is linear and continuous. (2.7.44)
Hence, v, ψ is meaningful and we may define
u ⊗ v : D(U × V) −→ C
(2.7.45)
u ⊗ v, ϕ := v(y), u(x), ϕ(x, y) for every ϕ ∈ C0∞ (U × V).
As defined, the mapping u ⊗ v is the composition of two linear and continuous

mappings, hence it is linear and continuous. Also, if ϕ1 ∈ C0∞ (U) and ϕ2 ∈ C0∞ (V),
then

u ⊗ v, ϕ1 ⊗ ϕ2 = v(y) , u(x), ϕ1 (x)ϕ2 (y)

= v(y) , ϕ2 (y)u(x), ϕ1 (x)

= v(y), ϕ2 (y) u(x), ϕ1 (x) , (2.7.46)
thus the mapping u ⊗ v defined in (2.7.45) satisfies (2.7.42).

To prove the uniqueness statement in part (i), suppose w1 , w2 ∈ D (U × V) are
such that
w j , ϕ1 ⊗ ϕ2 = u, ϕ1 v, ϕ2 , j = 1, 2,
(2.7.47)
for every ϕ1 ∈ C0∞ (U), ϕ2 ∈ C0∞ (V).
Then it follows that w1 , ϕ = w2 , ϕ for every ϕ ∈ C0∞ (U) ⊗ C0∞ (V), which in
concert with Proposition 2.81 and the continuity of w1 and w2 implies w1 = w2 in
D (U × V). This completes the proof of the statement in (i).
The reasoning used to prove the statement in (i), this time relying on (2) in Re-
mark 2.86 in place of Proposition 2.84, also yields that the mapping
w : D(U × V) −→ C
(2.7.48)
w, ϕ := u(x), v(y), ϕ(x, y) for every ϕ ∈ C0∞ (U × V)
is linear and continuous and satisfies

w, ϕ1 ⊗ ϕ2 = v(y), ϕ2 (y) u(x), ϕ1 (x) , (2.7.49)
for every ϕ1 ∈ C0∞ (U) and every ϕ2 ∈ C0∞ (V). The uniqueness result proved in (i)
then gives u ⊗ v = w. Hence, (2.7.43) holds as wanted.
As for the statement in (iii), observe that based on (2.7.45) we have
v ⊗ u : D(V × U) → C and
(2.7.50)
v ⊗ u, ψ = u(x), v(y), ψ(y, x) for every ψ ∈ C0∞ (V × U).
Hence, (v ⊗ u) : D(U × V) → C and
(v ⊗ u) , ϕ = v ⊗ u, ϕ = u(x), v(y), ϕ (y, x)
= u(x), v(y), ϕ(x, y) for every ϕ ∈ C0∞ (U × V). (2.7.51)
In particular, for every ϕ1 ∈ C0∞ (U) and ϕ2 ∈ C0∞ (U) we have
(v ⊗ u) , ϕ1 ⊗ ϕ2 = u, ϕ1 v, ϕ2 = u ⊗ v, ϕ1 ⊗ ϕ2 . (2.7.52)
The uniqueness result from part (i) now implies u ⊗ v = (v ⊗ u) in D (U × V).
Remark 2.88. If u ∈ D (U) and v ∈ Lloc

1
(V), then the statement in part (iii) in Theo-
rem 2.87 becomes

u(x), v(y)ϕ(x, y) dy = v(y) u(x), ϕ(x, y) dy, ∀ ϕ ∈ C0∞ (U × V). (2.7.53)
V V
The interpretation of (2.7.53) is that the distribution u commutes with the integral.
We next establish a number of basic properties for the tensor products of distri-
butions.
Theorem 2.89. Let m, n ∈ N, U be an open subset of Rm , and V be an open subset

of Rn . Assume that u ∈ D (U) and v ∈ D (V). Then the following properties hold.
(a) supp u ⊗ v = supp u × supp v.
(b) ∂αx ∂βy (u ⊗ v) = (∂αx u) ⊗ (∂βy v) for every α ∈ Nm
0 and every β ∈ N0 .
n
(c) ( f ⊗ g) · (u ⊗ v) = ( f u) ⊗ (gv) for every f ∈ C (U) and every g ∈ C ∞ (V).

∞
(d) The mapping D (U) × D (V) (u, v) → u ⊗ v ∈ D (U × V) is bilinear and

separately sequentially continuous.
(e) The tensor product of distributions is associative.
Proof. We start by proving the set theoretic equality from (a). For the right-to-left
inclusion, fix (x0 , y0 ) ∈ supp u × supp v. If C ⊆ U × V is an open neighborhood
of (x0 , y0 ), then there exists an open set A ⊆ U containing x0 and an open set B ⊆
V containing y0 such that A × B ⊂ C. In particular, since x0 ∈ supp u and y0 ∈
supp v, there exist ϕ1 ∈ C0∞ (A) and ϕ2 ∈ C0∞ (B) with the property that u, ϕ1 0
and v, ϕ2 0. If we now set ϕ := ϕ1 ⊗ ϕ2 , then ϕ ∈ C0∞ (C) and u ⊗ v, ϕ =
u, ϕ1 v, ϕ2 0. Hence (x0 , y0 ) ∈ supp (u ⊗ v), finishing the proof of the right-to-
left inclusion in (a).
To prove the opposite inclusion, observe that supp (u ⊗ v) ⊆ supp u × supp v is
equivalent to
(U × V) \ (supp u × supp v) ⊆ (U × V) \ supp (u ⊗ v). (2.7.54)
Write the left-hand side of (2.7.54) as D1 ∪ D2 , where D1 := (U \ supp u) × V and

D2 := U × (V \ supp v). Note that D1 and D2 are open sets in Rm × Rn . Since the
support of a distribution is the smallest relatively closed set outside of which the
distribution vanishes, for (2.7.54) to hold it suffices to show that u ⊗ v, ϕ = 0 for
every ϕ ∈ C0∞ (D1 ∪ D2 ). Fix such a function ϕ, set K := supp ϕ, and consider a
partition of unity subordinate to the covering {D1 , D2 } of K, say
ψ j ∈ C0∞ (D j ), j ∈ {1, 2}, ψ1 + ψ2 = 1 in a neighborhood of K. (2.7.55)
Then ϕψ1 ∈ C0∞ (D1 ), ϕψ2 ∈ C0∞ (D2 ) (with the understanding that ψ1 and ψ2 have
been extended by zero outside their supports), and ϕ = ϕψ1 + ϕψ2 on U × V. Since
πU (D1 ) ∩ supp u = ∅ and πV (D2 ) ∩ supp v = ∅, we may write

u ⊗ v, ϕ = v(y), u(x), ϕ(x, y)ψ1 (x, y)

+ u(x), v(y), ϕ(x, y)ψ2 (x, y) = 0. (2.7.56)
This completes the proof of the equality of sets from part (a).
∞ ∞
To prove the identity in (b), fix α ∈ Nm0 , β ∈ N0 , and let ϕ1 ∈ C 0 (U), ϕ2 ∈ C 0 (V).
n
Then starting with the definition of distributional derivatives and then using (2.7.42)
we may write

∂αx ∂βy (u ⊗ v), ϕ1 ⊗ ϕ2 = (−1)|α|+|β| u ⊗ v, ∂αx ∂βy (ϕ1 ⊗ ϕ2 )
= (−1)|α|+|β| u ⊗ v, ∂αx ϕ1 ⊗ ∂βy ϕ2
= (−1)|α|+|β| u, ∂αx ϕ1 v, ∂βy ϕ2
= ∂αx u, ϕ1 ∂βy u, ϕ2
= (∂αx u) ⊗ (∂βy v), ϕ1 ⊗ ϕ2 . (2.7.57)

By the uniqueness statement in part (i) of Theorem 2.87 we deduce from (2.7.57)
that (∂αx u) ⊗ (∂βy v) = ∂αx ∂βy (u ⊗ v), completing the proof of the identity in (b).
Moving on to the proof of the statement in (c), note that if f ∈ C ∞ (U) and
g ∈ C ∞ (V), then f ⊗ g ∈ C ∞ (U × V). The latter, combined with the definition of
multiplication of a distribution with a smooth function and (2.7.42), permits us to
write

( f ⊗ g) · (u ⊗ v), ϕ1 ⊗ ϕ2 = u ⊗ v, ( f ⊗ g) · (ϕ1 ⊗ ϕ2 ) (2.7.58)

= u ⊗ v, ( f ϕ1 ) ⊗ (gϕ2 ) = u, f ϕ1 v, gϕ2

= f u, ϕ1 gv, ϕ2 = ( f u) ⊗ (gv), ϕ1 ⊗ ϕ2 ,
for every ϕ1 ∈ C0∞ (U), ϕ2 ∈ C0∞ (V). The identity in (c) now follows from (2.7.58)
by once again invoking the uniqueness result from part (i) in Theorem 2.87.
The bilinearity of the mapping (u, v) → u⊗v is a consequence of the definition of
u ⊗ v and (2.2.1). To prove that this mapping is also separately sequentially continu-
D (U)
ous, let u j −−−−→ u and fix v ∈ D (V). If ϕ ∈ C0∞ (U ×V), then v(y), ϕ(x, y) ∈ C0∞ (U)
j→∞
by Proposition 2.84, and we may use Fact 2.22 to write
u j ⊗ v, ϕ = u j (x), v(y), ϕ(x, y) −−−−→ u(x), v(y), ϕ(x, y) = u ⊗ v, ϕ. (2.7.59)
j→∞
D (V) D (U×V)
Similarly, if u ∈ D (U) is fixed and v j −−−−→ v, then u ⊗ v j −−−−−−−→ u ⊗ v.
j→∞ j→∞
Finally, we are left with proving the associativity of the tensor product of dis-
tributions. To this end, let k ∈ N, W be an open subset of Rk , and w ∈ D (W) be
arbitrary. By Theorem 2.87, u ⊗ v ∈ D (U × V), v ⊗ w ∈ D (V × W) and, furthermore,
(u ⊗ v) ⊗ w ∈ D (U × V × W), u ⊗ (v ⊗ w) ∈ D (U × V × W). (2.7.60)
The goal is to prove that
(u ⊗ v) ⊗ w = u ⊗ (v ⊗ w) in D (U × V × W). (2.7.61)
In this regard we first note that for each ϕ ∈ C0∞ (U), ψ ∈ C0∞ (V), and η ∈ C0∞ (W),
we may write

(u ⊗ v) ⊗ w, (ϕ ⊗ ψ) ⊗ η = u, ϕv, ψw, η

= u ⊗ (v ⊗ w), ϕ ⊗ (ϕ ⊗ η) . (2.7.62)
Define C0∞ (U) ⊗ C0∞ (V) ⊗ C0∞ (W) as
(
N )
ϕ j ⊗ ψ j ⊗ η j : ϕ j ∈ C0∞ (U), ψ j ∈ C0∞ (V), η j ∈ C0∞ (W), N ∈ N , (2.7.63)
j=1
and note that this set is sequentially dense in D(U × V × W) (which can be proved
by reasoning as in the proof of Proposition 2.81). Granted this, (2.7.61) is implied
by (2.7.62), completing the proof of the theorem.
Exercise 2.90. Let n, m ∈ N and pick x0 ∈ Rn and y0 ∈ Rm arbitrary. Prove that
δ x0 ⊗ δy0 = δ(x0 ,y0 ) in D (Rn+m ).
We close this section by revisiting the result proved in Proposition 2.84 and es-
tablishing a related version that is going to be useful later on.
Proposition 2.91. Let m, n ∈ N, U be an open subset of Rm , and V be an open
subset of Rn . Assume that u ∈ E (U), ϕ ∈ C ∞ (U × V), and define the function
ψ : V → C, ψ(y) := u(x), ϕ(x, y), ∀ y ∈ V. (2.7.64)
Then ψ ∈ C ∞ (V) and for every α ∈ Nn0 we have

∂α ψ(y) = u(x) , ∂αy ϕ(x, y) , ∀ y ∈ V. (2.7.65)
Proof. Fix some η ∈ C0∞ (U) that satisfies η ≡ 1 in a neighborhood of supp u. Then
for each θ ∈ C0∞ (V) we may write

(θψ)(y) = (ηu)(x), θ(y)ϕ(x, y) = u(x), (η ⊗ θ)(x, y)ϕ(x, y)

= u(x), (η ⊗ θ)ϕ (x, y) , ∀ y ∈ V. (2.7.66)
Given that (η ⊗ θ)ϕ ∈ C0∞ (U × V), Proposition 2.84 applies and gives that the right-
most side of (2.7.66) depends in a C ∞ manner on the variable y ∈ V. Hence, θψ ∈
C ∞ (V) and since θ ∈ C0∞ (V) has been arbitrarily chosen we deduce that ψ ∈ C ∞ (V).
This takes care of the first claim in the statement of the proposition.
Moving on, observe that it suffices to prove (2.7.65) when |α| = 1, since the
general case then follows by iteration. With this in mind, fix some j ∈ {1, . . . , n} and
pick an arbitrary point y∗ ∈ V. Also, select θ ∈ C0∞ (V) such that θ ≡ 1 near y∗ . These
properties of θ permit us to compute
( )
∂y j (η ⊗ θ)ϕ (x, y∗ ) = η(x) (∂ j θ)(y∗ )ϕ(x, y∗ ) + θ(y∗ )(∂y j ϕ)(x, y∗ )
= η(x)(∂y j ϕ)(x, y∗ ), ∀ x ∈ U. (2.7.67)
Making use of (2.7.66), (2.7.32), and (2.7.67), we may then write

∂ j ψ(y∗ ) = ∂ j (θψ)(y∗ ) = u(x), ∂y j (η ⊗ θ)ϕ (x, y∗ )

= u(x), η(x)(∂y j ϕ)(x, y∗ )

= u(x), (∂y j ϕ)(x, y∗ ) . (2.7.68)
This corresponds precisely to formula (2.7.65) written at the point y = y∗ and for
the multi-index α = (0, . . . , 0, 1, 0, . . . , 0) ∈ Nn0 with the nonzero component on the
j-th slot. As remarked earlier, this suffices to finish the proof.
2.8 The Convolution of Distributions in R n
Recall that, as a consequence of Fubini’s Theorem, given any f, g ∈ L1 (Rn ), the

function h : Rn × Rn → C defined by h(x, y) := f (x − y)g(y) for every point
(x, y) ∈ Rn × Rn , is absolutely integrable on Rn × Rn and

|h(x, y)| dx dy = | f (x − y)g(y)| dx dy
Rn ×Rn
R R
n n

= |g(y)| | f (x − y)| dx dy
Rn Rn
= f L1 (Rn ) g L1 (Rn ) . (2.8.1)
Hence, the convolution of f and g defined as

f ∗ g : R → C, ( f ∗ g)(x) :=
n
f (x − y)g(y) dy for each x ∈ Rn , (2.8.2)
Rn
satisfies f ∗ g ∈ L1 (Rn ) (and a natural estimate). We would like to extend this

definition to functions that are not necessarily in L1 (Rn ). Specifically, assume that
f, g ∈ Lloc
1
(Rn ) have the property that

MB(0,r) := (x, y) ∈ supp f × supp g : x + y ∈ B(0, r)
(2.8.3)
is a compact set in Rn × Rn for every r ∈ (0, ∞).
In this scenario, consider the function G : Rn → [0, ∞] defined by

G(x) := | f (x − y)| |g(y)| dy for each x ∈ Rn . (2.8.4)
Rn
Note that for every r ∈ (0, ∞), by making a natural change of variables and using
the fact that f ⊗ g ∈ Lloc
1
(Rn × Rn ), we obtain

G(x) dx = | f (z)||g(y)| dy dz < ∞. (2.8.5)
|x|≤r MB(0,r)
Thus the function G is locally integrable, hence finite almost everywhere in Rn .

f, g ∈ Lloc (R ) satisfy (2.8.3), for almost every x ∈ R
1 n n
In conclusion, whenever
the integral ( f ∗ g)(x) := Rn f (x − y)g(y) dy is absolutely convergent and we have
f ∗ g ∈ Lloc
1
(Rn ). Furthermore, having fixed an arbitrary ϕ ∈ C0∞ (Rn ) we may write

f ∗ g, ϕ = f (x − y)g(y)ϕ(x) dy dx
Rn Rn

= f (z)g(y)ϕ(z + y) dy dz. (2.8.6)
Rn Rn
2.8 The Convolution of Distributions in Rn 65
To proceed, observe that the function ϕΔ defined by
ϕΔ : Rn × Rn → C, ϕΔ (x, y) := ϕ(x + y) for every x, y ∈ Rn , (2.8.7)
satisfies ϕΔ ∈ C ∞ (Rn × Rn ) though, in general, the support of ϕΔ is not compact.

Formally, the last double integral in (2.8.6) has the same expression as f ⊗ g, ϕΔ .
However, under the current assumptions on ϕ, f , and g, it is not clear that this pairing
may be interpreted in the standard distributional sense. Indeed, even though f ⊗ g is
a well-defined distribution in D (Rn × Rn ) (cf. Theorem 2.87), the function ϕΔ does
not belong to C0∞ (Rn × Rn ), as it lacks the compact support property. Nonetheless,
(2.8.3) implies
supp ϕΔ ∩ supp ( f ⊗ g) is a compact set in Rn × Rn . (2.8.8)
Theorem 2.60 applies with F := supp ( f ⊗ g) and allows us to uniquely extend the
action of the distribution f ⊗ g to the set of functions ψ ∈ C ∞ (Rn × Rn ) satisfying
the property that supp ψ ∩ supp ( f ⊗ g) is a compact set in Rn × Rn . Denote this
unique extension by f ⊗ g. Then f ⊗ g, ϕΔ is well defined, and it is meaningful to
set f ∗g, ϕ :=
f ⊗ g, ϕΔ . This discussion justifies making the following definition.
Definition 2.92. Suppose u, v ∈ D (Rn ) are such that
for every compact subset K of Rn the set

(2.8.9)
MK := {(x, y) ∈ supp u × supp v : x + y ∈ K} is compact in Rn × Rn .
Granted this, define the convolution of the distributions u and v as the functional
u ∗ v : D(Rn ) → C whose action on each ϕ ∈ C0∞ (Rn ) is given by
⊗ v, ϕΔ
u ∗ v, ϕ := u (2.8.10)
where ϕΔ (x, y) := ϕ(x + y) for every x, y ∈ Rn , and u

⊗ v is the unique extension of
u ⊗ v obtained by applying Theorem 2.60 with F := supp (u ⊗ v).
Remark 2.93. Retain the context of Definition 2.92.
(1) If ϕ ∈ C0∞ (Rn ) and ψ ∈ C0∞ (Rn × Rn ) is such that ψ ≡ 1 in a neighborhood of
Msupp ϕ then

u ∗ v, ϕ = u ⊗ v, ψϕΔ . (2.8.11)
(2) If (2.8.9) holds for the compacts K j = B(0, j), j ∈ N, then (2.8.9) holds for
arbitrary compact sets K ⊂ Rn .
(3) Condition (2.8.9) is always satisfied if either u or v is compactly supported.
The issue of continuity of the convolution map introduced in Definition 2.92 is
discussed next.
Theorem 2.94. If u, v ∈ D (Rn ) are such that (2.8.9) holds, then u ∗ v belongs to
D (Rn ). In particular, the convolution between two distributions in Rn , one of which
is compactly supported, is always well defined and is a distribution in Rn .
Proof. Let u, v ∈ D (Rn ) be such that (2.8.9) is satisfied. From Theorem 2.60, we
D(Rn )
have that u
⊗ v is linear, hence u∗v is linear as well. Let ϕ j −−−−→ 0. Then there exists
j→∞
a compact subset K of Rn such that supp ϕ j ⊆ K for each j ∈ N, and lim ∂α ϕ j = 0
j→∞
uniformly on K, for every α ∈ Nn0 . In particular, Msupp ϕ j ⊆ MK for every j ∈ N.
Hence, if we fix ψ ∈ C0∞ (Rn × Rn ) such that ψ ≡ 1 in a neighborhood of MK , then
part (1) in Remark 2.93 gives

u ∗ v, ϕ j = u ⊗ v, ψϕΔj , ∀ j ∈ N. (2.8.12)
Moreover, we claim that

D(Rn ×Rn )
ψ ϕΔj −−−−−−−→ 0. (2.8.13)
j→∞
To prove this claim, note that supp (ψ ϕΔj ) ⊆ supp ψ for every j ∈ N and for every
α1 , α2 , β1 , β2 ∈ Nn0 we have

sup (∂αx 1 ∂βy 1 ψ)(x, y)(∂αx 2 ∂βy 2 ϕΔj )(x, y)
(x,y)∈supp ψ
α β
≤ sup ∂ x 1 ∂y 1 ψ(x, y) ∂α2 +β2 ϕ j L∞ (Rn )
(x,y)∈supp ψ
≤ ∂αx 1 ∂βy 1 ψ L∞ (supp ψ) ∂α2 +β2 ϕ j L∞ (K) −−−−→ 0. (2.8.14)

j→∞
Hence, (2.8.13) follows by combining (2.8.14) with Leibniz’s formula (14.2.6).

Since u ⊗ v ∈ D (Rn × Rn ), from (2.8.12) and (2.8.13) we deduce

u ∗ v, ϕ j = u ⊗ v, ψϕΔj −−−−→ 0. (2.8.15)
j→∞
On account of Remark 2.3, this proves that u ∗ v ∈ D (Rn ). Finally, the last state-
ment of the theorem is a consequence of what we proved so far and part (3) in
Remark 2.93.
Remark 2.95. Combined, Theorem 2.94 and the discussion regarding (2.8.6) yield
the following result. If f, g ∈ Lloc 1
(Rn ) are such that for each compact K ⊂ Rn the
set {(x, y) : x ∈ supp f, y ∈ supp g, x + y ∈ K} is compact, then u f ∗ ug = u f ∗g in
D (Rn ). That is, f ∗ g ∈ Lloc
1
(Rn ), the convolution between the distributions u f and
ug (recall (2.1.6)) is well defined, and u f ∗ ug is a distribution of function type that
is equal to the distribution u f ∗g . In particular,
⎫
f ∈ Lloc
1
(Rn ), g ∈ Lcomp
1
(Rn ) ⎪
⎪
⎪
⎪
⎪
⎬
or ⎪
⎪ =⇒ u f ∗ ug = u f ∗g . (2.8.16)
⎪
⎪
n ⎪
f ∈ Lcomp (R ), g ∈ Lloc (R ) ⎭
1 n 1
The main properties of the convolution of distributions, whenever meaningfully

defined, are stated and proved in the next theorem. Recall that for any A, B ⊆ Rn the
set A ± B is defined as {x ± y : x ∈ A, y ∈ B}.
Theorem 2.96. The following statements are true.
(a) If u, v ∈ D (Rn ) are two distributions with the property that (2.8.9) is satisfied,
then supp (u ∗ v) ⊆ supp u + supp v.
(b) If u, v ∈ D (Rn ) are such that (2.8.9) is satisfied, then u ∗ v = v ∗ u.
(c) If u, v, w ∈ D (Rn ) are such that
⎧
⎪
⎪
⎪ for every compact subset K of Rn the set
⎪
⎪
⎪
⎨
⎪
⎪ MK := {(x, y, z) ∈ supp u × supp v × supp w : x + y + z ∈ K} (2.8.17)
⎪
⎪
⎪
⎪
⎩ is compact in Rn × Rn × Rn ,
then (u ∗ v) ∗ w and u ∗ (v ∗ w) are well defined, belong to D (Rn ), and are equal.
(d) Let u ∈ D (Rn ), α ∈ Nn0 . Then (∂α u) ∗ δ = ∂α u = u ∗ ∂α δ. In particular, u ∗ δ = u.
(e) If the distributions u, v ∈ D (Rn ) are such that (2.8.9) is satisfied and α ∈ Nn0 ,
then ∂α (u ∗ v) = (∂α u) ∗ v = u ∗ (∂α v).
Proof. Let u, v ∈ D (Rn ) be such that (2.8.9) holds. Since supp u + supp v is closed,
the inclusion in (a) will follow as soon as we show that

u ∗ vRn \(supp u+supp v) = 0. (2.8.18)
Pick an arbitrary function ϕ ∈ D(Rn \ (supp u + supp v)). Then
supp ϕΔ ∩ supp (u ⊗ v) = ∅, (2.8.19)
hence u ⊗ v, ϕΔ = 0, which implies u ∗ v, ϕ = 0, as wanted.

Next we show the statement in (b). Take u, v ∈ D (Rn ) for which (2.8.9) holds
and let ϕ ∈ C0∞ (Rn ). Choose some ψ ∈ C0∞ (Rn × Rn ) with the property that ψ ≡ 1 in
a neighborhood of
M1 = {(x, y) ∈ supp u × supp v : x + y ∈ supp ϕ}. (2.8.20)
By Remark 2.93 and (iii) in Theorem 2.87 we have

u ∗ v, ϕ = u ⊗ v, ψϕΔ = (v ⊗ u) , ψϕΔ

= v ⊗ u, ψ (ϕΔ ) . (2.8.21)
Definition (2.7.38) in our current setting implies ψ ∈ C0∞ (Rn × Rn ) and ψ ≡ 1 in

a neighborhood of
M2 = {(y, x) ∈ supp v × supp u : x + y ∈ supp ϕ}. (2.8.22)
Also, it is immediate that (ϕΔ ) = ϕΔ . Hence, by Remark 2.93, we see that


v ⊗ u, ψ (ϕΔ ) = v ⊗ u, ψ ϕΔ = v ∗ u, ϕ. (2.8.23)
From (2.8.21) and (2.8.23) it follows that u ∗ v, ϕ = v ∗ u, ϕ, finishing the proof
of the statement in (b).
To prove the statement in (c), suppose u, v, w ∈ D (Rn ) satisfy (2.8.17). Define
the functional u ∗ v ∗ w : D(Rn ) → C by setting

⊗ v ⊗ w, ϕΔ ,
u ∗ v ∗ w, ϕ := u for every ϕ ∈ C0∞ (Rn ), (2.8.24)

where ϕΔ (x, y, z) := ϕ(x + y + z) for each x, y, x ∈ Rn , and where we have denoted
by u
⊗ v ⊗ w the unique extension of u ⊗ v ⊗ w obtained by applying Theorem 2.60
for F := supp (u ⊗ v ⊗ w). The mapping in (2.8.24) is well defined since if ϕ is an

arbitrary test function in C0∞ (Rn ) then ϕΔ ∈ C ∞ (Rn ×Rn ×Rn ) and, based on (2.8.17),
the set
Δ
Msupp ϕ = supp (u ⊗ v ⊗ w) ∩ supp ϕ is compact in Rn × Rn × Rn . (2.8.25)
Reasoning as in the proof of Theorem 2.94, it follows that u∗v∗w belongs to D (Rn )
and

u ∗ v ∗ w, ϕ = u ⊗ v ⊗ w, ψϕΔ for ϕ ∈ C0∞ (Rn ) and for each
(2.8.26)
ψ ∈ C0∞ (Rn × Rn × Rn ) with ψ ≡ 1 in a neighborhood of Msupp

ϕ.
Given the freedom in selecting ψ as in (2.8.26), we choose to take ψ as follows.

Let π j : Rn × Rn × Rn → Rn , j = 1, 2, 3, be the projections defined for each

x, y, z ∈ Rn by π1 (x, y, z) := x, π2 (x, y, z) := y, and π3 (x, y, z) := z. Given a

function ϕ ∈ C0∞ (Rn ), fix
ψ j ∈ C0∞ (Rn ) with ψ j ≡ 1 near π j (Msupp

ϕ ), j = 1, 2, 3, (2.8.27)
then choose
ψ := ψ1 ⊗ ψ2 ⊗ ψ3 ∈ C0∞ (Rn × Rn × Rn ). (2.8.28)
The next two claims are designed to prove that u ∗ (v ∗ w) exists.
Claim 1. For every compact K in Rn the set
NK := {(y, z) ∈ supp v × supp w : y + z ∈ K}
is compact in Rn × Rn .
To see why this is true, start by observing that, for every x0 ∈ supp u, the set K + x0
is compact in Rn and
B := {(x, y, z) ∈ {x0 } × supp v × supp w : x + y + z ∈ K + x0 } (2.8.29)

is closed and contained in MK+x 0
. Since (2.8.17) ensures that MK+x0
is compact in
R × R × R , it follows that B is compact in R × R × R . In addition, the mapping
n n n n n n
θ : Rn × Rn × Rn → Rn × Rn , defined by θ(x, y, z) := (y, z) for each x, y, z ∈ Rn , is

continuous and θ(B) = NK . Therefore, NK must be compact, and Claim 1 is proved.
The latter ensures that v ∗ w exists.
Claim 2. For every K ⊂ Rn compact, the set
PK := {(x, z) ∈ supp u × supp (v ∗ w) : x + z ∈ K}
is compact in Rn × Rn .
By part (a) in the theorem, we have supp (v ∗ w) ⊆ supp v + sup w. Thus
PK ⊆ {(x, z) ∈ supp u × (supp v + supp w) : x + z ∈ K}

= (x, y + t) : x ∈ supp u, y ∈ supp v, t ∈ supp w, x + y + t ∈ K (2.8.30)
and the last set in (2.8.30) is closed in Rn × Rn . If we now set
σ : Rn × Rn × Rn → Rn × Rn ,
(2.8.31)
σ(x, y, t) := (x, y + t) for every x, y, t ∈ Rn ,
then σ is continuous and PK ⊆ σ(MK ). By (2.8.17), we have that MK is compact

in Rn × Rn × Rn , hence σ(MK ) is compact in Rn × Rn . The set PK being closed in
Rn × Rn , we may conclude that PK is compact in Rn × Rn . This proves Claim 2 and,
as a consequence, the fact that u ∗ (v ∗ w) exists.
With an eye toward proving u ∗ (v ∗ w) = u ∗ v ∗ w, we dispense of two more
claims.
Claim 3. Let K be an arbitrary compact subset of Rn and for u, v, w as before
introduce the set A := (supp v + supp w) ∩ (K − supp u). The A is also compact in
Rn .
Rewrite A as
A = {t ∈ supp v + supp w : t = ω − x, for some ω ∈ K, x ∈ supp u}.
Then if t ∈ A, it follows that there exist y ∈ supp v, z ∈ supp w, ω ∈ K and x ∈ supp u

such that t = y+z = ω−x. Hence, (x, y, z) ∈ supp u×supp v×supp w and x+y+z = ω,
which implies that (x, y, z) ∈ MK and A ⊆ ν(MK ), where
ν : Rn × Rn × Rn → Rn ,
(2.8.32)
ν(x1 , x2 , x3 ) := x2 + x3 for every x1 , x2 , x3 ∈ Rn .
We may now conclude that A is compact since the map ν is continuous, MK is
compact and A is closed. This proves Claim 3.
Claim 4. Fix ϕ ∈ C0∞ (Rn ) and set K := supp ϕ. Also, let A be as in Claim 3 cor-
responding to this K, and suppose η ∈ C0∞ (Rn ) is such that η ≡ 1 in a neighborhood
of A. Then, with ψ1 as in (2.8.27), we have
ψ1 ⊗ η = 1 on (supp u × supp (v ∗ w)) ∩ supp ϕΔ (2.8.33)
where, as before, ϕΔ (x, y) = ϕ(x + y) for x, y ∈ Rn .

To prove this claim, since ψ1 ≡ 1 on π1 (MK ) and η ≡ 1 on A, it suffices to show that

supp u × supp (v ∗ w) ∩ supp ϕΔ ⊆ π1 (MK ) × A. (2.8.34)
To justify the latter inclusion, let x ∈ supp u, y ∈ supp v, and z ∈ supp w, be such
that (x, y + z) ∈ supp ϕΔ . Then x + y + z ∈ K which forces x ∈ π1 (MK ) as well as
y + z ∈ K − supp u ⊆ A. Thus, (x, y + z) ∈ π1 (MK ) × A and this completes the proof
of Claim 4.
Consider now an arbitrary function ϕ ∈ C0∞ (Rn ) and set K := supp ϕ. Also,
assume that ψ1 , ψ2 , ψ3 are as in (2.8.27), and let η be as in Claim 4. Making use of
the definition of the convolution and tensor products, and keeping in mind (2.8.33),
we may write

u ∗ (v ∗ w), ϕ = u(x) ⊗ (v ∗ w)(t), ψ1 (x)η(t)ϕ(x + t) (2.8.35)

= u(x), (v ∗ w)(t), ψ1 (x)η(t)ϕ(x + t)

= u(x), v(y) ⊗ w(z), ψ1 (x)η(y + z)ϕ(x + y + z)ψ2 (y)ψ3 (z) .
A few words explaining the origin of the last equality are in order. According to the
definition of the convolution, passing from v ∗ w to v ⊗ w requires that we consider
the set
C := (supp v × supp w) ∩ supp ηΔ ∩ supp[ϕ(x + ·)]Δ . (2.8.36)
Since C is closed and satisfies C ⊂ π2 (MK ) × π3 (MK ), it follows that C is compact.
Now, the fact that ψ2 ⊗ ψ3 ≡ 1 in a neighborhood of C justifies the presence of
ψ2 ⊗ ψ3 in the last term in (2.8.35).
Going further, since
ψ1 (x)ψ2 (y)ψ3 (z) = ψ1 (x)ψ2 (y)ψ3 (z)η(y + z) for (x, y, z) near MK , (2.8.37)
referring to (2.8.26) and (2.8.28) allows us to rewrite (2.8.35) in the form

u ∗ (v ∗ w), ϕ = u(x), v(y) ⊗ w(z), ψ1 (x)ψ2 (y)ψ3 (z)ϕ(x + y + z)

= u ⊗ v ⊗ w, (ψ1 ⊗ ψ2 ⊗ ψ3 )ϕΔ̃ = u ∗ v ∗ w, ϕ. (2.8.38)
Since ϕ ∈ C0∞ (Rn ) was arbitrary, it follows that u ∗ (v ∗ w) = u ∗ v ∗ w. Similarly, it

can be seen that (u ∗ v) ∗ w = u ∗ v ∗ w and this completes the proof of the statement
in (c).
Moving on to (d), fix u ∈ D (Rn ) and α ∈ Nn0 . Since the Dirac distribution
δ has compact support, by Theorem 2.94 it follows that both δ ∗ u and (∂α δ) ∗ u
are well defined and belong to D (Rn ). Let ϕ ∈ C0∞ (Rn ) and consider a function
ψ ∈ C0∞ (Rn × Rn ) with ψ = 1 on a neighborhood of the set ({0} × supp u) ∩ supp ϕΔ .
Starting with Definition 2.92, then using (2.4.1) combined with Proposition 2.39,
then Leibniz’s formula (14.2.6), then (2.1.19), the support condition for ψ and then
(2.4.1) again, we may write
α
∂ δ ∗ u, ϕ = (∂αx δ) ⊗ u(y), ψ(x, y)ϕ(x + y)

= u(y), ∂α δ(x), ψ(x, y)ϕ(x + y)

= (−1)|α| u(y), δ(x), ∂αx (ψ(x, y)ϕ(x + y))

β α−β
= (−1)|α| u(y), δ(x), α!
β!(α−β)! ∂ x ψ(x, y)(∂ x ϕ)(x + y)
β≤α

β α−β
= (−1)|α| u(y) , α!
β!(α−β)! (∂ x ψ)(0, y)(∂ x ϕ)(y)
β≤α
|α|
= (−1) u(y), ψ(0, y)∂α ϕ(y)

= (−1)|α| u, ∂α ϕ = ∂α u, ϕ . (2.8.39)
In particular, if |α| = 0, the above implies δ ∗ u = u. When combined with (b), this
finishes the proof of the statement in (d).
Finally, by making use of the results from (d) and (c) we have
∂α (u ∗ v) = ∂α (δ ∗ (u ∗ v)) = ∂α δ ∗ (u ∗ v) = (∂α δ ∗ u) ∗ v
= (∂α u) ∗ v. (2.8.40)
A similar argument also shows that ∂α (u ∗ v) = u ∗ (∂α v). The proof of the theorem
is now complete.
Exercise 2.97. Prove that for a distribution u ∈ D (Rn ),
u = δ ⇐⇒ u ∗ f = f for each f ∈ C0∞ (Rn ). (2.8.41)
Hint: For the right-to-left implication use f = φ j , where φ j is as in Example 2.24,

and let j → ∞.
Next, we extend the translation map (1.3.17) to distributions.
Proposition 2.98. For each x0 ∈ Rn and each u ∈ D (Rn ) fixed, the translation
mapping D(Rn ) ϕ → u, t−x0 (ϕ) ∈ C is linear and continuous.
Denoting this map by t x0 u thus yields a distribution in Rn that satisfies
t x0 u, ϕ = u, t−x0 (ϕ), ∀ ϕ ∈ C0∞ (Rn ). (2.8.42)

Proof. This follows by observing that the mapping in question is the composition
u ◦ t−x0 where the latter translation operator is consider in the sense of Exercise 1.19.

Exercise 2.99. Fix x0 ∈ Rn and recall the distribution δ x0 from Example 2.14. Prove
that δ x0 = t x0 δ in D (Rn ). Also show that δ x0 ∗ u = t x0 u for every u ∈ D (Rn ). In
particular, if x1 ∈ Rn is arbitrary, then δ x0 ∗ δ x1 = δ x0 +x1 in D (Rn ).
Remark 2.100.
(1) If u, v ∈ E (Rn ) then u ∗ v ∈ E (Rn ). This is an immediate consequence of part (a)
in Theorem 2.96.
(2) There exists a sequence {u j } j∈N of compactly supported distributions in R that
converges to some u ∈ D (R) and such that u j ∗ v does not necessarily converge to
u ∗ v in D (R) for every v ∈ D (R).
To see an example in this regard, consider the sequence of compactly supported
D (R)
distributions {δ j } j∈N that satisfies δ j −−−−→ 0. Then if 1 denotes the distribution on
j→∞
R given by the constant function 1, Exercise 2.99 gives δ j ∗ 1 = 1 for each j, and
the constant sequence {1} j∈N ⊂ D (R) does not converge in D (R) to 0 ∗ 1 = 0. This
shows that sequential continuity for convolution of distributions cannot be expected
in general.
(3) Condition (2.8.17) is necessary for the operation of convolution of distributions
to be associative. To see this, consider the distributions 1, δ , and H on R. Then we
have supp δ = {0}, supp 1 = R, supp H = [0, ∞). If K is a compact set in R, the set

MK = (x, 0, z) : x ∈ R, z ∈ [0, ∞), x + z ∈ K (2.8.43)
is not compact in R×R×R, thus (2.8.17) does not hold. Furthermore, 1∗δ = 1 ∗δ =
0 ∗ δ = 0 so (1 ∗ δ ) ∗ H = 0, while 1 ∗ (δ ∗ H) = 1 ∗ (δ ∗ H ) = 1 ∗ (δ ∗ δ) = 1 ∗ δ = 1
and clearly (1 ∗ δ ) ∗ H 1 ∗ (δ ∗ H) in D (R).
Proposition 2.101. The following statements are true.

D (Rn ) D (Rn )
(1) If u ∈ E (Rn ) and v j −−−−−→ v, then u ∗ v j −−−−−→ u ∗ v.
j→∞ j→∞
D (Rn )
(2) If u j −−−−−→ u and there exists K ⊂ R compact with supp u j ⊆ K for every
n
j→∞
D (Rn )
j ∈ N, then u j ∗ v −−−−−→ u ∗ v for every v ∈ D (Rn ).
j→∞
Proof. To see why (1) is true, fix ϕ ∈ C0∞ (Rn ). Then by definition, for each j ∈ N
we have u ∗ v j , ϕ = u ⊗ v j , ψ j ϕΔ for any smooth compactly supported function
ψ j with ψ j = 1 on a neighborhood of (supp u × supp v j ) ∩ supp ϕΔ . Note that
(supp u × supp v j ) ∩ supp ϕΔ ⊆ supp u × (supp ϕ − supp u) (2.8.44)
and supp u × (supp ϕ − supp u) is compact since both supp u and supp ϕ are compact.
Hence, if we fix ψ ∈ C0∞ (Rn × Rn ) such that ψ = 1 in a neighborhood of the set
supp u × (supp ϕ − supp u), then u ∗ v j , ϕ = u ⊗ v j , ψϕΔ for every j ∈ N. Based

on (d) in Theorem 2.89, we may write
u ∗ v j , ϕ = u ⊗ v j , ψϕΔ −−−−→ u ⊗ v, ψϕΔ = u ∗ v, ϕ, (2.8.45)

j→∞
and the desired conclusion follows.

Assume next the hypotheses in part (2) of the proposition. In particular, these
entail supp u ⊆ K. Let ϕ ∈ C0∞ (Rn ) and note that
(supp u j × supp v) ∩ supp ϕΔ ⊆ K × (supp ϕ − K), ∀ j ∈ N. (2.8.46)
Then if ψ ∈ C0∞ (Rn ×Rn ) is a function with the property that ψ = 1 in a neighborhood
of K × (supp ϕ − K), we have u j ∗ v, ϕ = u j ⊗ v, ψϕΔ for every j ∈ N. Hence,
u j ∗ v, ϕ = u j ⊗ v, ψϕΔ −−−−→ u ⊗ v, ψϕΔ = u ∗ v, ϕ, (2.8.47)

j→∞
where for the convergence in (2.8.47) we used part (d) in Theorem 2.89.
When convolving an arbitrary distribution with a distribution of function type
given by a compactly supported smooth function, the resulting distribution is of
function type. This fact is particularly useful in applications and we prove it next.
Proposition 2.102. If u ∈ D (Rn ) and g ∈ C0∞ (Rn ), then the distribution u ∗ g is of

function type given by the function
f : Rn → C, f (x) := u(y), g(x − y), ∀ x ∈ Rn , (2.8.48)
that satisfies f ∈ C ∞ (Rn ). Moreover, if u is compactly supported then so is f . In

short,
D (Rn ) ∗ C0∞ (Rn ) ⊆ C ∞ (Rn ) and E (Rn ) ∗ C0∞ (Rn ) ⊆ C0∞ (Rn ). (2.8.49)
Proof. Let φ : Rn × Rn → C be defined by φ(x, y) := g(x − y) for each point

(x, y) ∈ Rn × Rn . Then φ ∈ C ∞ (Rn × Rn ) and φ(x, ·) ∈ C0∞ (Rn ) for each x ∈ Rn . This
shows that the function f in (2.8.48) is well defined.
To prove that f is of class C ∞ in Rn , pick an arbitrary point x∗ ∈ Rn and pick a
function ψ ∈ C0∞ (Rn ) with the property that ψ = 1 in a neighborhood of B(x∗ , 1).
Then, for every x ∈ B(x∗ , 1),

f B(x∗ ,1) (x) = u(y), g(x − y) = u(y), ψ(x)g(x − y) . (2.8.50)
Since the function Rn × Rn (x, y) → ψ(x)g(x − y) ∈ C is of class C0∞ , we may

invoke Proposition 2.84 to conclude that f B(x∗ ,1) ∈ C ∞ B(x∗ , 1) . Given that x∗ ∈ Rn
has been arbitrarily chosen, it follows that f ∈ C ∞ (Rn ).
We now turn to the task of showing that the distribution u ∗ g is of function type
and is given by f . To this end, fix ϕ ∈ C0∞ (Rn ) and consider a function
ψ ∈ C0∞ (Rn × Rn ) such that ψ = 1 in a neighborhood of

(2.8.51)
the set (x, y) : x ∈ supp u, y ∈ supp g, x + y ∈ supp ϕ .
Then, starting with Definition 2.92 we write

u ∗ g, ϕ = u ⊗ g, ψϕΔ

= u(x), g(y), ψ(x, y)ϕ(x + y)

= u(x) , g(y)ψ(x, y)ϕ(x + y) dy
Rn

= u(x) , g(y)ϕ(x + y) dy
Rn

= u(x) , g(z − x)ϕ(z) dz
Rn

= u(x), g(z − x)ϕ(z) dz
Rn
= f, ϕ. (2.8.52)
For the third equality in (2.8.52) we have used the fact that g is a distribution of
function type, for the fourth equality we used condition (2.8.51), the fifth equality
is based on a change of variables, the sixth equality follows from (2.7.53), while the
last equality is a consequence of the definition of f .
To complete the proof of the proposition there remains to notice that, by part (a)
in Theorem 2.96, we have supp f ⊆ supp u + supp g. In particular, if u is compactly
supported then so is f .
Exercise 2.103. Prove that if u ∈ E (Rn ) and g ∈ C ∞ (Rn ) then the distribution u ∗ g
is of function type given by the function
f : Rn → C, f (x) := u(y), g(x − y), ∀ x ∈ Rn , (2.8.53)
that satisfies f ∈ C ∞ (Rn ). In short, E (Rn ) ∗ C ∞ (Rn ) ⊆ C ∞ (Rn ).

Hint: Use Proposition 2.91 to show that f ∈ C ∞ (Rn ), then reason as in Proposi-
tion 2.102 to take care of the remaining claims.
Exercise 2.104. Let Ω be a bounded open set in Rn . Suppose u ∈ L1 (Ω) and define

u to be the extension by zero outside Ω of u. Also assume that g ∈ Lloc 1
(Rn ) is given.
u ∈ D (Rn ), it satisfies supp u ⊆ Ω, hence
Prove that u ∗ g is well defined in D (Rn ),
and that the distribution u ∗ g is of function type given by the function

u ∗ g)(x) =
( g(x − y)u(y) dy, ∀ x ∈ Rn . (2.8.54)
Ω
Hint: Apply (2.8.16) with f =

u.
D(Rn )
Exercise 2.105. Prove that if u ∈ D (Rn ) is arbitrary and ϕ j −−−−→ ϕ, then one has
j→∞
E(Rn )
u ∗ ϕ j −−−−→ u ∗ ϕ.
j→∞
Next, we prove that distributions of function type given by smooth, compactly

supported functions are dense in the class of distributions. First we treat the simpler
case when the distributions are considered in Rn . The case when the distributions
are considered on arbitrary open subsets of Rn needs to be handled with a bit more
care.
Theorem 2.106. The set C0∞ (Rn ) is sequentially dense in D (Rn ).
Proof. First we will show that C ∞ (Rn ) is sequentially dense in D (Rn ). Let φ be
as in (1.2.3) and recall the sequence of functions {φ j } j∈N from (1.3.7). In particular,
we have that supp φ j ⊆ B(0, 1) for every j ∈ N. Recall from Example 2.24 that
D (Rn )
φ j −−−−−→ δ.
j→∞
Let u ∈ D (Rn ) be arbitrary and define
u j := u ∗ φ j in D (Rn ), ∀ j ∈ N. (2.8.55)
By Proposition 2.102 we have u j ∈ C ∞ (Rn ) for all j ∈ N. Also, by part (2) in

Proposition 2.101 and part (d) in Theorem 2.96 we obtain
D (Rn )
u j = u ∗ φ j −−−−−→ u ∗ δ = u. (2.8.56)
j→∞
This completes the proof of the fact that C ∞ (Rn ) is sequentially dense in D (Rn ).
Moving on, and keeping the notation introduced so far, fix ψ ∈ C0∞ (Rn ) satisfying
ψ(x) = 1 if |x| < 1 and set
w j (x) := ψ(x/ j)(u ∗ φ j )(x) ∀ x ∈ Rn , ∀ j ∈ N. (2.8.57)
It is immediate that w j ∈ C0∞ (Rn ). Moreover, if ϕ ∈ C0∞ (Rn ) is given, then there
exists j0 ∈ N with the property that supp ϕ ⊂ B(0, j0 ). Therefore, for all j ≥ j0 we
may write
x
w j , ϕ = ψ (u ∗ φ j )(x)ϕ(x) dx = (u ∗ φ j )(x)ϕ(x) dx
Rn j Rn
= u ∗ φ j , ϕ −−−−→ u, ϕ. (2.8.58)

j→∞
D (Rn )
This shows that w j −−−−−→ u. Hence, ultimately, C0∞ (Rn ) is sequentially dense in
j→∞
D (Rn ), finishing the proof of the theorem.
The same type of result as in Theorem 2.106 actually holds in arbitrary open
subsets of the Euclidean ambient.
Theorem 2.107. The set C0∞ (Ω) is sequentially dense in D (Ω).
Proof. Fix u ∈ D (Ω) and recall the sequence of compact sets introduced in (1.3.11).

Then K j = Ω and K j ⊂ K̊ j+1 for every j ∈ N. For each j ≥ 2 consider a function
j∈N
ψ j ∈ C0∞ (Ω), ψ j = 1 on a neighborhood of K j−1 , supp ψ j ⊆ K j , (2.8.59)
and define u j := ψ j u ∈ D (Ω). Since supp u j ⊆ K j , Proposition 2.69 gives that each
u j may be extended to a distribution in E (Rn ), which we continue to denote by u j .
If we now set
1
ε j := dist (K j , ∂K j+1 ) > 0, ∀ j ∈ N, (2.8.60)
4
then the definition of the compacts in (1.3.11) forces εj 0 as j → ∞. Having
fixed some φ ∈ C0∞ (Rn ) satisfying supp φ ⊆ B(0, 1) and Rn φ(x) dx = 1, define
φ j (x) := ε−n
j φ(x/ε j ), ∀ x ∈ Rn , ∀ j ∈ N. (2.8.61)
Thus,
φ j ∈ C0∞ (Rn ), supp φ j ⊆ B(0, ε j ) ⊆ B(0, 1), ∀ j ∈ N, (2.8.62)
and reasoning as in Example 2.24 we see that

D (Rn )
φ j −−−−−→ δ. (2.8.63)
j→∞
Fix j ∈ N with j ≥ 2 and introduce w j := u j ∗ φ j . By combining part (1) in

Remark 2.100 with Proposition 2.102 we obtain that w j ∈ C0∞ (Rn ). In addition, (a)
in Theorem 2.96 and (2.8.60) imply that supp w j ⊆ K j + B(0, ε j ) ⊂ K j+1 , so in fact
w j ∈ C0∞ (Ω).
D (Ω)
To complete the proof of the theorem it suffices to show that w j −−−−→ u. To this
j→∞
end, fix ϕ ∈ C0∞ (Ω) and observe that based on (1.3.11) there exists some j0 ∈ N such
that supp ϕ ⊆ K j0 . Then for j > j0 + 1 we may write

w j , ϕ = u j ∗ φ j , ϕ = (u j ∗ φ j )(x)ϕ(x) dx
Ω

= u j (y), φ j (x − y)ϕ(x) dx = u(y), ψ j (y)φ j (x − y)ϕ(x) dx
Ω Ω

= u(y), ψ j0 +2 (y)φ j (x − y)ϕ(x) dx = u j0 +2 ∗ φ j , ϕ. (2.8.64)
Ω
For the second equality in (2.8.64) we used the fact that u j ∗ φ j ∈ C0∞ (Rn ) for
j ≥ 2, for the third equality we used Proposition 2.102, while for the fourth the fact
that u j = ψ j u for every j ∈ N. These observations also give the last equality in
(2.8.64). As for the penultimate equality in (2.8.64), observe that if j > j0 + 1,
x ∈ supp ϕ ⊆ K j0 and x − y ∈ supp φ j ⊆ B(0, ε j ), then
y ∈ K j0 − B(0, ε j ) ⊆ K j0 − B(0, ε j0 ) ⊂ K j0 +1 , (2.8.65)
thus ψ j (y) = 1 = ψ j0 +2 (y).

If we now combine (2.8.63) with (2.8.62) and part (2) in Proposition 2.101, it
D (Rn )
follows that u j0 +2 ∗ φ j −−−−−→ u j0 +2 ∗ δ = u j0 +2 . Hence, if we return with this in
j→∞
(2.8.64), we may write
lim w j , ϕ = lim u j0 +2 ∗ φ j , ϕ = u j0 +2 , ϕ = ψ j0 +2 u, ϕ
j→∞ j→∞
= u, ϕ, (2.8.66)
since ψ j0 +2 = 1 in a neighborhood of the support of ϕ. This finishes the proof of the

theorem.
Proposition 2.108. Suppose Ω1 , Ω2 are open sets in Rn and let F : Ω1 → Ω2 be a

C ∞ diffeomorphism. For each u ∈ D (Ω2 ), define the mapping u ◦ F : D(Ω1 ) → C
by setting

u ◦ F (ϕ) := u, | det(DF −1 )|(ϕ ◦ F −1 ) , ∀ ϕ ∈ D(Ω1 ), (2.8.67)
where DF −1 denotes the Jacobian matrix of F −1 . Then the following are true.
(1) u ◦ F ∈ D (Ω1 ) and the mapping
D (Ω2 ) u → u ◦ F ∈ D (Ω1 ) is sequentially continuous. (2.8.68)
(2) If u = u f for some f ∈ Lloc 1

(Ω2 ), then u ◦ F = u f ◦F in D (Ω1 ).
(3) If F = (F1 , . . . , Fn ), then the following generalized Chain Rule formula holds

n
∂ j (u ◦ F) = (∂ j Fk )[(∂k u) ◦ F] in D (Ω1 ) (2.8.69)
k=1
for all j ∈ {1, . . . , n} and all u ∈ D (Ω2 ).
Proof. Fix u ∈ D (Ω2 ). The fact that u ◦ F ∈ D (Ω1 ) is an immediate consequence

D (Ω2 )
of Exercise 1.25. Also, if u j −−−−−→ u and ϕ ∈ C0∞ (Ω1 ), since Exercise 1.25 gives
j→∞
| det(DF −1 )|(ϕ◦ F −1 ) ∈ C0∞ (Ω1 ), in view of definition (2.8.67) and Fact 2.22 we have
D (Ω1 )
u j ◦ F −−−−−→ u ◦ F.
j→∞
Next, suppose u = u f for some f ∈ Lloc
1
(Ω2 ). Then for each ϕ ∈ C0∞ (Ω1 ) we may
write

u ◦ F, ϕ = u, | det(DF −1 )|(ϕ ◦ F −1 )

= f (y)(ϕ ◦ F −1 )(y)| det(DF −1 (y))| dy
Ω2

= ( f ◦ F)(x)ϕ(x) dx = u f ◦F , ϕ. (2.8.70)
Ω1
The first equality in (2.8.70) is based on (2.8.67), the second uses the fact that u =
u f , the third uses the change of variables y = F(x), and the last one uses the fact that
f ◦ F ∈ Lloc 1
(Ω1 ). This proves the statement in item (2).
There remains to prove the Chain Rule formula (2.8.69). Suppose first that u is a
distribution of function type given by a function ψ ∈ C0∞ (Ω2 ), that is u = uψ . Then
by item (2) we have u ◦ F = uψ◦F in D (Ω1 ). Hence, invoking Exercise 2.40 and the
Chain Rule for pointwise differentiable functions we have
∂ j (u ◦ F) = u∂ j (ψ◦F) = unk=1 (∂ j Fk )(x)[(∂k ψ)◦F]

n
= u(∂ j Fk )(x)[(∂k ψ)◦F]
k=1

n
= (∂ j Fk )[(∂k u) ◦ F] in D (Ω1 ). (2.8.71)
k=1
This proves (2.8.69) in the case when u = uψ for some ψ ∈ C0∞ (Ω2 ). Recall that by
Theorem 2.107 the set C0∞ (Ω2 ) is sequentially dense in D (Ω2 ). Since the following
operations with distributions are sequentially continuous: differentiation (item (3)
in Proposition 2.43), composition with C ∞ diffeomorphisms (item (1) in the current
proposition), and multiplication with C ∞ functions ((4) in Exercise 2.32), we may
conclude that (2.8.69) holds for arbitrary u ∈ D (Ω2 ).
2.9 Distributions with Higher Order Gradients Continuous or

Bounded
We have seen that if u ∈ C m (Rn ) for some m ∈ N, then its distributional deriva-
tives up to order m are distributions of function type, each given by the correspond-
ing pointwise derivative of u. A more subtle question pertains to the possibility of
deducing regularity results for distributions whose distributional derivatives of a
certain order are of function type, and these functions exhibit a certain amount of
smoothness. In this section we prove two main results in this regard. In the first one
(see Theorem 2.112), we show that if a distribution u ∈ D (Ω) has all distributional
derivatives of order m ∈ N continuous, then in fact u ∈ C m (Ω). In the second main
result (see Theorem 2.114) we prove that a distribution in D (Rn ) is of function type
2.9 Distributions with Higher Order Gradients Continuous or Bounded 79
given by a Lipschitz function if and only if all its first-order distributional derivatives
are bounded functions in Rn .
We start by proving a weaker version of Theorem 2.112.
Proposition 2.109. If u ∈ D (Rn ) and there exists some m ∈ N0 such that for each
α ∈ Nn0 satisfying |α| ≤ m the distributional derivative ∂α u belongs to C 0 (Rn ), then
u ∈ C m (Rn ).
Proof. Recall the sequence of distributions {φ j } j∈N from Example 2.24 and for u
satisfying the hypothesis of the proposition let {u j } j∈N be as in (2.8.55). In particular,
(2.8.56) holds, thus the distributional and classical derivatives of each u j coincide.
Next we make the following claim:
lim ∂α u j = ∂α u uniformly on compact sets in Rn

j→∞
(2.9.1)
for every multi-index α ∈ Nn0 satisfying |α| ≤ m.
To prove this claim, observe that by part (e) in Theorem 2.96, for each j ∈ N we
have ∂α u j = (∂α u) ∗ φ j for every α ∈ Nn0 . Fix α ∈ Nn0 satisfying |α| ≤ m. Since by
the current hypotheses the distributional derivative ∂α u is continuous, by invoking
Proposition 2.102 we may write
∂α u j (x) = ((∂α u) ∗ φ j )(x) = ∂α u(y), φ j (x − y)

α
= (∂ u)(y)φ j (x − y) dy = (∂α u)(x − z)φ j (z) dz
Rn Rn

= (∂α u) x − y/ j φ(y) dy, ∀ x ∈ Rn . (2.9.2)

Rn
Fix a compact set K in Rn . Making use of (2.9.2) and the properties of φ (recall
(1.2.3)) we estimate

α
sup |∂ u j (x) − ∂ u(x)| = sup [(∂ u) x − y/ j − (∂ u)(x)]φ(y) dy

α α
α
x∈K x∈K Rn

≤ sup (∂α u) x − y/ j − (∂α u)(x). (2.9.3)
x∈K
y∈B(0,1)
Since ∂α u is continuous in Rn it follows that ∂α u is uniformly continuous on compact

subsets of Rn , thus
lim sup |∂α u j (x) − ∂α u(x)| (2.9.4)

j→∞ x∈K

≤ lim sup (∂α u) x − z/ j − (∂α u)(x) = 0,
j→∞ x∈K
y∈B(0,1)
completing the proof of the claim.

With (2.9.1) in hand, we may invoke Lemma 2.110 below and proceed by induc-
tion on |α| to conclude that, as desired, u ∈ C m (Rn ).
Lemma 2.110. Suppose the functions {u j } j∈N and u are such that:
(i) u j ∈ C 1 (Ω) for every j ∈ N,
(ii) lim u j = u uniformly on compact subsets of Rn contained in Ω, and
j→∞
(iii) for each k ∈ {1, . . . , n} there exists a function vk ∈ C 0 (Ω) with the property that
lim ∂k u j = vk uniformly on compact subsets of Rn contained in Ω.
j→∞
Then u ∈ C 1 (Ω) and ∂k u = vk for each k ∈ {1, . . . , n}.

Proof. From the start, since uniform convergence on compact sets preserves conti-
nuity, we have that u ∈ C 0 (Ω). Fix x ∈ Ω and k ∈ {1, 2, . . . , n} and let t0 > 0 be such
that x + tek ∈ Ω whenever t ∈ [−t0 , t0 ] where, as before, ek is the unit vector in Rn
whose k-th component is equal to 1. Then, for each t ∈ [−t0 , t0 ], we may write
u(x + tek ) − u(x) = lim [u j (x + tek ) − u j (x)] (2.9.5)

j→∞
t
d
= lim u j (x + s ek ) ds
j→∞ 0 ds
t t
= lim (∂k u j )(x + s ek ) ds = vk (x + s ek ) ds.
j→∞ 0 0
The first equality in (2.9.5) is based on (ii) while the last one is a consequence of
(iii). Next, since vk is continuous on Ω, by the Mean Value Theorem for integrals it
t there exists some st , belonging to the interval with end-points 0 and t,
follows that
such that 0 vk (x + s ek ) ds = tvk (x + st ek ). Hence,
u(x + tek ) − u(x)

lim = lim vk (x + st ek ) = vk (x), (2.9.6)
t→0 t t→0
which proves that (∂k u)(x) exists and is equal to vk (x). Thus, ∂k u = vk ∈ C 0 (Ω) for
every k, which shows that u ∈ C 1 (Ω).
Lemma 2.111. Let u ∈ D (Ω) be such that for each j ∈ {1, . . . , n}, the distributional
derivatives ∂ j u are of function type and belong to C 0 (Ω). Then u ∈ C 0 (Ω).
1
Proof. Since ∇u ∈ [C 0 (Ω)]n the function v(x) := 0 (∇u)(tx) · x dt for x ∈ Ω (where
“·” denotes the dot product of vectors) is well defined and belongs to C 0 (Ω). Given
the current goals, by Exercise 2.53 it suffices to show that for each x0 ∈ Ω there
exists an open set ω ⊂ Ω such that x0 ∈ ω and uω = vω in D (ω).
To this end, fix x0 ∈ Ω and let r ∈ (0, dist(x0 , ∂Ω)). Consequently, we have
B(x0 , r) ⊂ Ω. Without loss of generality in what follows we may assume that x0 = 0
(since translations interact favorably, in a reversible manner, with both hypotheses
and conclusion). Let ϕ ∈ C0∞ (Ω) be such that supp ϕ ⊂ B(0, r) and fix j ∈ {1, . . . , n}.
Then we have
1
∂ j v, ϕ = −v, ∂ j ϕ = − (∇u)(tx) · x ∂ j ϕ(x) dt dx
Ω 0
1
=− (∇u)(tx) · x ∂ j ϕ(x) dx dt
0 Ω
1
= − lim+ (∇u)(tx) · x ∂ j ϕ(x) dx dt
ε→0 ε Ω
1
y
= − lim+ ∇u(y) · (∂ j ϕ)(y/t) dy dt
ε→0 ε Ω tn+1

n 1
yk
= − lim+ ∂k u(y) n+1
(∂ j ϕ)(y/t) dt dy
ε→0 Ω k=1 ε t
n
1 yk
= lim+ u(y), ∂yk n+1
(∂ j ϕ)(y/t) dt . (2.9.7)
ε→0 ε t
k=1
For the fourth equality in (2.9.7) we have used Lebesgue’s Dominated Convergence
Theorem (cf. Theorem 14.15) and for the fifth one the change of variables tx = y
(note that supp ϕ(·/t) ⊂ B(0, r) ⊂ Ω since supp ϕ ∈ B(0, r) and t ∈ (0, 1]). Further-
more, for each ε > 0, differentiating with respect to y and then integrating by parts
in t, gives
n
1 yk
∂yk n+1
(∂ j ϕ)(y/t) dt
k=1 ε t
1 n
1 yk
= (∂ j ϕ)(y/t) + (∂ k ∂ j ϕ)(y/t) dt
ε k=1
tn+1 tn+2
n 1
1 d
= n+1
(∂ j ϕ)(y/t) − n (∂ j ϕ)(y/t) dt
ε t t dt
1
t=1 1
= − n (∂ j ϕ)(y/t) = −∂ j ϕ(y) + n (∂ j ϕ)(y/ε). (2.9.8)
t t=ε ε
By combining (2.9.7) and (2.9.8) we obtain
1
∂ j v, ϕ = −u, ∂ j ϕ + lim+ u(y), n−1 ∂y j ϕ(y/ε)
ε→0 ε
1
= ∂ j u, ϕ + lim+ ∂ j u(y), n−1 ϕ(y/ε) . (2.9.9)
ε→0 ε
Since ∂ j u ∈ C 0 (Ω), the pairing under the limit in the rightmost term in (2.9.9) is
given by an integral in which we make the change of variables x = y/ε to further
compute

1 1
lim+ ∂ j u(y), n−1 ϕ(y/ε) = lim+ n−1 (∂ j u)(y)ϕ(y/ε) dy
ε→0 ε ε→0 ε Ω

= lim+ ε (∂ j u)(εx)ϕ(x) dx = 0 (2.9.10)
ε→0 Ω
given that, by the continuity of ∂ j u at 0 ∈ Ω,

lim+ (∂ j u)(εx)ϕ(x) dx = ∂ j u(0) ϕ(x) dx. (2.9.11)
ε→0 Ω Ω
In summary, from (2.9.9), (2.9.10), (2.9.11), the fact that ϕ is an arbitrary

element
in
D(B(0, r)), and that j is arbitrary in {1, . . . , n}, we conclude that ∇vB(0,r) = ∇uB(0,r)

in D (B(0, r)). By Exercise 2.158 there exists c ∈ C such that uB(0,r) − vB(0,r) = c in

D (B(0, r)). Since v ∈ C 0 (Ω), the latter implies u ∈ C 0 (B(0, r)) as desired. This
B(0,r)
completes the proof of the lemma.
After these preparations, we are ready to state and prove our first main result.
Theorem 2.112. Let u ∈ D (Ω) and suppose that there exists m ∈ N0 such that for
each α ∈ Nn0 satisfying |α| = m the distributional derivative ∂α u is continuous on Ω.
Then u ∈ C m (Ω).
Proof. We prove the theorem by induction on m. For m = 0 there is nothing to
prove. Suppose m = 1. Applying Lemma 2.111, we obtain that u ∈ C 0 (Ω). To prove
that u ∈ C 1 (Ω), it suffices to show that u is of class C 1 in a neighborhood of any
point x0 ∈ Ω. Fix x0 ∈ Ω, pick a number r > 0 with the property that B(x0 , r) ⊂ Ω
and a function ψ ∈ C0∞ (Rn ) such that supp ψ ⊂ B(x0 , r) and ψ ≡ 1 on B(x0 , r/2).
Then by Proposition 2.69 the distribution ψu extends to v := ψ u ∈ E (Rn ). Also,
v ∈ C (R ) since u ∈ C (Ω) and ψ is compactly supported in Ω. Using (2.5.4) and
0 n 0
part (4) in Proposition 2.43 we obtain
∂ j v = ∂
j (ψu) = (∂ j ψ)u + ψ∂ j u ∈ C (R ).
0 n
(2.9.12)
2.109 with m = 1 then gives v ∈ C (R ). In addition, since by

1 n
Applying Proposition

design v B(x ,r/2) = u B(x ,r/2) , we conclude that u B(x ,r/2) ∈ C (B(x0 , r/2)), as wanted.
1
0 0 0
Assume now that the theorem is true for all nonnegative integers up to, and in-
cluding, some m ∈ N. Take u ∈ D (Ω) with the property that ∂α u ∈ C 0 (Ω) for all
α ∈ Nn0 satisfying |α| = m + 1 and fix j ∈ {1, . . . , n}. Then ∂ j u ∈ D (Ω) satisfies
∂α (∂ j u) ∈ C 0 (Ω) for all α ∈ Nn0 with |α| = m. By the induction hypothesis, it follows
that ∂ j u ∈ C m (Ω). In particular, ∂ j u ∈ C 0 (Ω). This being true for all j ∈ {1, . . . , n},
what we already proved for m = 1 implies u ∈ C 1 (Ω). Thus, u is a C 1 function
in Ω whose first-order partial derivatives are of class C m in Ω. Then necessarily
u ∈ C m+1 (Ω) as wanted. The proof of the theorem is finished.
Exercise 2.113. Let u ∈ D (Rn ) be such that for some N ∈ N0 we have ∂α u = 0 in
D (Rn ) for each α ∈ Nn0 with |α| > N. Prove that u is a polynomial of degree less
than or equal to N.
Hint: Use Theorem 2.112 to conclude that u ∈ C ∞ (Rn ) then invoke Taylor’s formula
(14.2.9).
Moving on to the second issue discussed at the beginning of this section we recall
(cf. Remark 2.42) that a function f : Ω → C is called Lipschitz (in Ω) provided there
exists some constant M ∈ [0, ∞) such that
| f (x) − f (y)| ≤ M|x − y|, ∀ x, y ∈ Ω, (2.9.13)
and that the Lipschitz constant of f is the smallest M for which (2.9.13) holds.
Our next task is to prove the following theorem, which provides a distributional
characterization of Lipschitzianity.
Theorem 2.114. For f ∈ D (Rn ) and a number M ∈ [0, ∞) the following two state-
ments are equivalent:
(i) f is given by a Lipschitz function in Rn with Lipschitz constant less than or equal
to M;
(ii) for each k ∈ {1, . . . , n}, the distributional derivative ∂k f belongs to L∞ (Rn ) and
∂k f L∞ (Rn ) ≤ M.
As a consequence of this distributional characterization of Lipschitzianity and
the extension result recorded in (2.4.5) we have that
if Ω ⊆ Rn is an open set and f : Ω → C is a Lipschitz function then
(2.9.14)
∂ j f ∈ L∞ (Ω) for each j = 1, . . . , n.
Proof. Fix a distribution f ∈ D (Rn ) and let φ be as in (1.2.3). Consider the se-
quence of functions {φ j } j∈N from (1.3.7), that satisfies the properties listed in (1.3.8).
Furthermore, set
f j := f ∗ φ j in D (Rn ), ∀ j ∈ N. (2.9.15)
By Proposition 2.102 we have
f j ∈ C ∞ (Rn ) and f j (x) = f, φ j (x − ·), ∀ j ∈ N. (2.9.16)

D (Rn )
Also, since (as proved in Example 2.24) one has φ j −−−−−→ δ, by part (2) in Propo-
j→∞
sition 2.101 and part (d) in Theorem 2.96 one obtains
D (Rn )
f j = f ∗ φ j −−−−−→ f ∗ δ = f. (2.9.17)
j→∞
Next we proceed with the proof of (i) =⇒ (ii). Suppose f is Lipschitz with Lips-
chitz constant ≤ M. In particular, the formula in (2.9.16) becomes

f j (x) = f (y)φ j (x − y) dy = f (x − y)φ j (y) dy,
Rn Rn (2.9.18)
for all x ∈ Rn and all j ∈ N.
We claim that for each j ∈ N one has
| f j (x) − f j (y)| ≤ M|x − y|, ∀ x, y ∈ Rn and ∇ f j L∞ (Rn ) ≤ M. (2.9.19)
Indeed, if j ∈ N is fixed, then from (2.9.18) we obtain

| f j (x) − f j (y)| ≤ | f (x − z) − f (y − z)|φ j (z) dz ≤ M|x − y|, (2.9.20)
Rn
for every x, y ∈ Rn . In turn, since f j is smooth, we have
f j (x + hek ) − f j (x)
(∂k f j )(x) = lim , ∀ x ∈ Rn , ∀ k ∈ {1, . . . , n}. (2.9.21)
h→0 h
In combination with (2.9.20) this implies ∇ f j L∞ (Rn ) ≤ M, completing the proof of
the claims made in (2.9.19).
Next, fix k ∈ {1, . . . , n} and ϕ ∈ C0∞ (Rn ). Then based on (2.9.17) we may write
∂k f, ϕ = − f, ∂k ϕ = − lim f j , ∂k ϕ = lim ∂k f j , ϕ

j→∞ j→∞

= lim (∂k f j )(x)ϕ(x) dx. (2.9.22)
j→∞ Rn
Using the second estimate in (2.9.19) we obtain

(∂ f )(x)ϕ(x) dx ≤ ∇ f ∞ n ϕ 1 n ≤ M ϕ 1 n . (2.9.23)
k j j L (R ) L (R ) L (R )
Rn
Hence, from (2.9.23) and (2.9.22) it follows that

∂k f, ϕ ≤ lim sup (∂k f j )(x)ϕ(x) dx ≤ M ϕ L1 (Rn ) . (2.9.24)

j→∞ Rn
Consequently, the linear assignment
C0∞ (Rn ) ϕ → ∂k f, ϕ ∈ R (2.9.25)
is continuous in the L1 -norm and has norm less than or equal to M. Since C0∞ (Rn )
is dense in L1 (Rn ), the linear functional in (2.9.25) extends to a linear, bounded

functional Λk : L1 (Rn ) → C with norm less or equal to M. Thus, Λk ∈ L1 (Rn ) ∗ =

∞ n
L (R ) has norm less than or equal to M, which implies that there exists a unique
gk ∈ L∞ (Rn ) with gk L∞ (Rn ) ≤ M (2.9.26)
and such that

Λk (h) = gk (x)h(x) dx, ∀ h ∈ L1 (Rn ). (2.9.27)
Rn
Granted (2.9.27) and keeping in mind that Λk is an extension of the linear assignment
in (2.9.25), we arrive at the conclusion that

∂k f, ϕ = gk (x)ϕ(x) dx, ∀ ϕ ∈ C0∞ (Rn ). (2.9.28)
Rn
The identity in (2.9.28) yields ∂k f = gk in D (Rn ), which proves (ii) in view of

(2.9.26).
Conversely, suppose (ii) is true. Fix j ∈ N and note that from (2.9.15) and part
(e) in Theorem 2.96, for every k ∈ {1, . . . , n} one has ∂k f j = (∂k f ) ∗ φ j in D (Rn ).
Thus, using Proposition 2.102, the current assumptions on f , and the properties of
φ j , we have

|(∂k f j )(x)| = |∂k f, φ j (x − ·)| = (∂k f )(y)φ j (x − y) dy
Rn

≤ ∂k f L∞ (Rn ) φ j (x − y) dy ≤ M, ∀ x ∈ Rn , (2.9.29)
Rn
for each k ∈ {1, . . . , n}. Now fix x0 ∈ Rn and consider the sequence of functions
{g j } j∈N given by
g j (x) := f j (x) − f j (x0 ), for x ∈ Rn . (2.9.30)
Then g j ∈ C ∞ (Rn ) and based on the Mean Value Theorem, (2.9.30), and (2.9.29)
we also obtain
|g j (x) − g j (y)| ≤ M|x − y|, ∀ x, y ∈ Rn , and (2.9.31)
|g j (x)| ≤ M|x − x0 |, ∀ x ∈ Rn . (2.9.32)
By a corollary of the classical Arzelá–Ascoli theorem (see Theorem 14.24), there

exists a subsequence {g j }∈N that converges uniformly on any compact subset of Rn
to some function g ∈ C 0 (Rn ). As such, for every ϕ ∈ C0∞ (Rn ) we may write

lim g j , ϕ = lim g j (x)ϕ(x) dx
→∞ →∞ supp ϕ

= g(x)ϕ(x) dx = g, ϕ (2.9.33)
supp ϕ
D (Rn )
which goes to show that g j −−−−−→ g. The latter, (2.9.17), and (2.9.30), give that
→∞
whenever ψ ∈ C0∞ (Rn ) is such that Rn ψ(x) dx = 0 we have

g, ψ = lim g j , ψ = lim g j (x)ψ(x) dx
→∞ →∞ Rn

= lim f j (x)ψ(x) dx = lim f j , ψ = f, ψ. (2.9.34)
→∞ Rn →∞
Thus, we may apply Exercise 2.157 to conclude that there exists some constant
c ∈ C with the property that f = g + c in D (Rn ). This proves that the distribution f
is of function type and is given by the function g + c. Moreover, writing the estimate
in (2.9.31) with j replaced by j and then taking the limit as → ∞ (recall that
{g j } converges pointwise to g) leads to
|g(x) − g(y)| ≤ M|x − y|, ∀ x, y ∈ Rn , (2.9.35)
which in concert with the fact that f (x) − f (y) = g(x) − g(y) proves that f is a
Lipschitz function with Lipschitz constant ≤ M. Now the proof of (ii) =⇒ (i) is
finished.
In connection with Theorem 2.114 we recall Rademacher’s theorem which gives
that Lipschitz functions are pointwise differentiable at almost every point.
Theorem 2.115. Any Lipschitz function f : Rn → R is differentiable almost every-

where.
Exercise 2.116. Assume f : Rn → R is a Lipschitz function. On the one hand,

according to Rademacher’s theorem for each k ∈ {1, . . . , n} the pointwise partial
pw
derivative with respect to the k-th variable, temporarily denoted by ∂k f , exists a.e.
in R . On the other hand, Theorem 2.114 gives that for each k ∈ {1, . . . , n} the partial
n
derivative ∂k f computed in the sense of distributions belongs to L∞ (Rn ). Follow the

outline below to show that the two brands of derivatives mentioned above coincide
a.e. in Rn . As a consequence, the pointwise partial derivatives of order one of f are
bounded.
Step I. Rademacher’s theorem ensures that

pw
∂k f (x) = lim g j (x) for a.e. x ∈ Rn , (2.9.36)

j→∞
f x + j−1 ek − f (x)
where for each j ∈ N we have set g j (x) := for all x ∈ Rn . Since
j−1
pw
each g j is continuous, conclude that ∂k f is measurable in Rn .
| f (x) − f (y)|
Step II. If M := sup ∈ [0, ∞) then for every j ∈ N we have
x,y∈Rn ,
xy |x − y|
|g j (x)| ≤ M for all x ∈ Rn . Use this and Step I to conclude that ∂k f ∈ L∞ (Rn ).
pw
Step III. Fix a test function ϕ ∈ C0∞ (Rn ) and use Lebesgue’s Dominated Conver-
gence Theorem to write

pw
∂k f (x)ϕ(x) dx = lim g j (x)ϕ(x) dx

Rn j→∞ Rn
* +

= lim j f x + j−1 ek ϕ(x) dx − j f (x)ϕ(x) dx

j→∞ Rn Rn
* +

= lim j f (y)ϕ y − j−1 ek dy − j f (y)ϕ(y) dy

j→∞ Rn Rn

ϕ y − j−1 ek − ϕ(y)
= lim f (y) dy
j→∞ Rn j−1

=− f (y)(∂k ϕ)(y) dy. (2.9.37)
Rn
Use this to derive the desired conclusion.
Exercise 2.117. Let Ω be an arbitrary open subset of Rn and fix f ∈ Lip(Ω). Then
for each k ∈ {1, . . . , n} the following properties hold.
(i) The pointwise partial derivative with respect to the k-th variable, denoted by
∂k f , exists a.e. in Ω. Moreover, as a function, ∂k f belongs to L∞ (Ω).
pw pw
(ii) The partial derivative ∂k f computed in the sense of distributions in Ω coincides

pw
with the locally integrable function ∂k f .
Hint: Use (2.4.5) (with E := Ω) and Exercise 2.116.

We close this section by a discussion regarding locally Lipschitz functions.
Specifically, given an open set Ω in Rn we define the set of locally Lipschitz func-
tions in Ω by

Liploc (Ω) := f : Ω → C : f ∈ Lip(K) ∀ K ⊂ Ω compact set , (2.9.38)
There are other useful characterizations of local Lipschitzianity, as discussed

next.
Exercise 2.118. Let Ω be an open subset of Rn and suppose f : Ω → C is a given

function. Then the following are equivalent.
(1) f ∈ Liploc (Ω);

(2) for each x ∈ Ω there exists r x ∈ 0, dist (x, ∂Ω) such that the restriction f B(x,r )

x
belongs to Lip B(x, r x ) .
Hint: The implication (1) ⇒ (2) is immediate, while the reverse implication may be
established with the help of the Lebesgue Number Theorem (cf. Theorem 14.44).
Exercise 2.119. If Ω is an arbitrary open subset of Rn then C 1 (Ω) ⊆ Liploc (Ω).
Hint: Use Exercise 2.118 and the Mean Value Theorem.

Exercise 2.120. Let Ω ⊆ Rn , O ⊆ Rm be two arbitrary open sets, and assume that
Ψ : O → Ω has Lipschitz components. Then if f ∈ Liploc (Ω) it follows that f ◦ Ψ ∈
Liploc (O).
Hint: For any given compact set K ⊆ O it follows that W := Ψ (K) is a compact

subset of Ω, and ( f ◦ Ψ )K = f W ◦ Ψ K ∈ Lip(K).
Proposition 2.121. Let Ω be an open subset in Rn and let f ∈ Liploc (Ω). Then for
pw
each j ∈ {1, . . . , n} the pointwise partial derivative ∂ j f exists at almost every point
∞ pw
in Ω and belongs to Lloc (Ω). Moreover, as a locally integrable function, ∂ j f equals
the distributional derivative ∂ j f ∈ D (Ω). In other words,

∀ ϕ ∈ C0∞ (Ω).
pw
∂ j f ϕ dx = − f ∂ j ϕ dx, (2.9.39)
Ω Ω
Proof. If for each k ∈ N we introduce

Ωk := x ∈ Ω : dist (x, ∂Ω) > 1/k and |x| < k , (2.9.40)
then
each Ωk is an open, relatively compact, subset of Ω,
(2.9.41)
Ωk ⊆ Ωk+1 , and k∈N Ωk = Ω.

Since f ∈ Liploc (Ω), it follows that for each k ∈ N we have f Ω ∈ Lip(Ωk ). As such,
k
Exercise 2.117 ensures that
pw

for each index j ∈ {1, . . . , n} the pointwise derivative ∂ j f Ω exists

pw

a.e. in Ωk , belongs to L∞ (Ωk ), and ∂ j f Ω = ∂ j f Ω in D (Ωk ).

k
(2.9.42)
k k
Granted these, it is then immediate from (2.9.41) that for each j ∈ {1, . . . , n} the
pw
pointwise partial derivative ∂ j f exists at almost every point in Ω and belongs to
∞
Lloc (Ω). There remains to prove (2.9.39). To this end, fix an arbitrary ϕ ∈ C0∞ (Ω)
and pick k ∈ N such that supp ϕ ⊆ Ωk . Then, having fixed some j in {1, . . . , n}, we
may write

f ∂ j ϕ dx = f ∂ j ϕ dx = f Ω , ∂ j ϕΩ
k k
Ω Ωk

pw

= − ∂ j f Ω , ϕΩ = − ∂j f Ω ϕ dx

k k k
Ωk

pw
=− ∂ j f ϕ dx. (2.9.43)
Ω
Above, the first and last equalities are consequences of the support condition on ϕ,
the second equality is the interpretation of the integral as the distributional pairing
over the open set Ωk , the third equality is based on the definition of the distributional
derivative (considered in D (Ωk )), while the fourth equality is a consequence of the
last property in (2.9.42).
Further Notes for Chapter 2. The material in this chapter is at the very core of the theory of
distributions since it provides a versatile calculus for distributions that naturally extends the scope
of the standard calculus for ordinary functions. The definition of distributions used here is essen-
tially that of the French mathematician Laurent Schwartz (1915–2002), cf. [66], though nowadays
there are many books dealing at length with the classical topics discussed here. These include [10],
[17], [21], [20], [22], [23], [27], [28], [33], [34], [65], [71], [74], [76], [77], [78], and the reader is
referred to these sources for other angles of exposition. In particular, in [34], [27], [74], distribu-
tions are defined on smooth manifolds, while in [54] the notion of distributions is adapted to rough
settings.
Exercise 2.122. Consider the mapping u : D(R) → C by setting

∞
1
u(ϕ) := ϕ 2 − ϕ(0) , ∀ ϕ ∈ C0∞ (R). (2.10.1)
j=1
j
Prove that u is well defined. Is u a distribution? If yes, what is the order of u?

Exercise 2.123. Prove that there exists u ∈ D (Ω) for which the following statement
is false:
For each compact K contained in Ω there exist C > 0 and k ∈ N0 such that
|u, ϕ| ≤ C sup |∂α ϕ(x)|, ∀ ϕ ∈ C0∞ (Ω).
x∈K, |α|≤k
(2.10.2)
Exercise 2.124. Prove that |x|N ln |x| ∈ Lloc
1
(Rn ) whenever N is a real number satis-
fying N > −n. Thus, in particular, |x| ln |x| ∈ D (Rn ) when N > −n.
N
Exercise 2.125. Suppose n ≥ 2 and given ξ ∈ S n−1 define f (x) := ln |x · ξ| for each
x ∈ Rn \ {0}. Prove that f ∈ Lloc
1
(Rn ). In particular, ln |x · ξ| ∈ D (Rn ).
Exercise 2.126. Prove that (ln |x|) = P.V. 1x in D (R), where P.V. 1x is the distribu-
tion defined in (2.1.13).
Exercise 2.127. Let f : R → R be defined by f (x) = x ln |x| − x if x ∈ R \ {0}
and f (0) = 0. Prove that f ∈ C 0 (R) and the distributional derivative of f in D (R)
equals ln |x|.
xj
Exercise 2.128. Suppose n ≥ 2. Prove that ∂ j (ln |x|) = |x|2
in D (Rn ) for every
j ∈ {1, . . . , n}.
1
x
Exercise 2.129. Suppose n ≥ 3. Prove that ∂ j |x|n−2
= (2 − n) |x|jn in D (Rn ) for every
j ∈ {1, . . . , n}.
∞
Exercise 2.130. Let θ ∈ C0 (R) be supported in the interval (−1, 1) and such that
R
θ(t) dt = 1. For each j ∈ N, define
j
ψ j (x) := j θ( jx − jt) dt, ∀ x ∈ R.
1/ j
D (R)
Prove that ψ j −−−−→ H, where H is the Heaviside function.
j→∞

∞
Exercise 2.131. Let u : D(R) → C defined by u(ϕ) := ϕ( j) ( j) for each function
j=1
ϕ ∈ C0∞ (R). Prove that u ∈ D (R) and that this distribution does not have finite
order.
Exercise 2.132. For each j ∈ N define
1
f j (x) := 1
for x ∈ Rn \ {0}.
j|x|n− j
D (Rn )
Prove that f j −−−−−→ ωn−1 δ, where ωn−1 is the area of the unit sphere in Rn .
j→∞
Exercise 2.133. For each ε > 0 define

1 ε
fε (x) := for x ∈ R.
π x2 + ε2
D (R)
Prove that fε −−−−→
+
δ.
ε→0
Exercise 2.134. For each ε > 0 define

|x|2
fε (x) := (4πε)− 2 e− 4ε
n
for x ∈ Rn .
D (Rn )
Prove that fε −−−−−→
+
δ.
ε→0
√
Exercise 2.135. Recall that i = −1 ∈ C and, for each ε ∈ (0, ∞), define
1
fε± (x) := for x ∈ R.
x ± iε
Also, recall the distribution from (2.1.13). Prove that
1 1
−→ ∓iπδ + P.V. as ε → 0+ in D (R). (2.10.3)
x ± iε x
This is the so-called Sokhotsky’s formula.
( sin( jx) )
Exercise 2.136. Prove that the sequence πx converges to δ in D (R) as
j∈N
j → ∞.
Exercise 2.137. In each case determine if the given sequence of distributions in
D (R) indexed over j ∈ N converges √ and determine its limit whenever convergent.
Below m ∈ N is fixed and i = −1.
√
j
(a) f j (x) := √ e− jx
2
for x ∈ R;
π
(b) f j (x) ::= jm cos( jx) for x ∈ R;
2 j3 x 2
(c) f j (x) = for x ∈ R;
π(1 + j2 x2 )2
(d) u j := (−1) j δ 1j ;
j
(e) u j := δ 1j − δ− 1j ;
2
1
(f) f j (x) := χ|x|≥ 1j for x ∈ R \ {0};
x
1 sin2 ( jx)
(g) f j (x) := for x ∈ R \ {0};
jπ x2
(h) f j (x) := ⎧jm ei jx for x ∈ R;
⎪
⎨ je if x > 0,
⎪ i jx
(j) f j (x) := ⎪⎪ for x ∈ R.
⎩0 if x ≤ 0,
Exercise 2.138. Let a ∈ R. Compute (H(· − a)) in D (R).
Exercise 2.139. Consider the function
⎧
⎪
⎨ x if x > a,
⎪
f : R −→ R defined by f (x) := ⎪
⎪ ∀ x ∈ R, (2.10.4)
⎩ 0 if x ≤ a,
where a ∈ R is fixed. Compute (u f ) in D (R), where u f is defined as in (2.1.6).

Exercise 2.140. Let f : R → R be defined by f (x) := sin |x| for every x ∈ R.
Compute (u f ) and (u f ) in D (R).
Exercise 2.141. Let I ⊆ R be an open interval, x0 ∈ I, and f ∈ C 1 (I \ {x0 }) be such
that f ∈ Lloc
1
(I) (here f is the pointwise derivative of f in I \ {x0 }). Prove that the
one-sided limits lim+ f (x), lim− f (x) exist and are finite, that f ∈ Lloc
1
(I), and that
x→x0 x→x0

(u f ) = u f + lim+ f (x) − lim− f (x) δ x0 in D (I).
x→x0 x→x0
Remark 2.142. Prove that there exist pointwise differentiable functions on R for
which the distributional derivative does not coincide with the classical derivative.

You may consider the function f defined by f (x) := x2 cos x12 for x 0 and
f (0) := 0, and show that f ∈ C 1 (R \ {0}) and also f is differentiable at the ori-
gin, while f Lloc
1
(R).
Exercise 2.143. Let I ⊆ R be an open interval, x0 ∈ I, and let m ∈ N. Suppose

that f ∈ C ∞ (I \ {x0 }) is such that the pointwise derivatives f , f , . . . , f (m) , belong
1
to Lloc (I). Prove that for every k ∈ {0, , 1, . . . , m − 1} the limits lim+ f (k) (x) and
x→x0
lim− f (k) (x) exist, are finite, and that
x→x0

u(m)
f = u f (m) + lim+
f (x) − lim−
f (x) δ(m−1)
x0
x→x0 x→x0

+ lim+ f (x) − lim− f (x) δ(m−2)
x0 + ···
x→x0 x→x0

+ lim+ f (m−1) (x) − lim− f (m−1) (x) δ x0 in D (I).
x→x0 x→x0
Exercise 2.144. Let I ⊆ R be an open interval and consider a sequence {xk }k∈N
of points in I with no accumulation point in I. Suppose we are given a function
f ∈ C 1 (I \ {xk : k ∈ N}) such that its pointwise derivative f belongs to Lloc
1
(I). Prove
that for each k ∈ N the limits lim± f (x) exist and are finite, f belongs to Lloc 1
(I), and
x→xk

∞
(u f ) = u f + lim+ f (x) − lim− f (x) δ xk in D (I).
x→xk x→xk
k=1
Exercise 2.145. Let f : R → R be the function defined by f (x) := x for each
x ∈ R, where x is the integer part of x. Determine (u f ) .
Exercise 2.146. Let Σ ⊂ Rn be a surface of class C 1 as in Definition 14.45, and

denote byσ the surface measure on Σ. Define the mapping δΣ : C0∞ (Rn ) → C by
δΣ (ϕ) := Σ ϕ(x) dσ(x) for each ϕ ∈ C0∞ (Rn ). Prove that δΣ ∈ D (Rn ), it has order
zero, and supp δΣ = Σ. Also show that if g ∈ L∞ (K ∩ Σ) for each compact set K in
Rn , and if we define

gδΣ (ϕ) := g(x)ϕ(x) dσ(x), ∀ ϕ ∈ C0∞ (Rn ), (2.10.5)

Σ

then gδΣ ∈ D (R ). n
Exercise 2.147. Let Ω ⊂ Rn be a domain of class C 1 (recall Definition 14.59) and

denote by ν = (ν1 , . . . , νn ) its outward unit normal. Denote by δ∂Ω the distribution
defined as in Exercise 2.146 corresponding to Σ := ∂Ω.
Set Ω+ := Ω and Ω− := Rn \ Ω. Suppose f ∈ Lloc 1
(Rn ) has the property that,
for each k ∈ {1, 2 . . . , n}, its distributional derivative ∂k f belongs to Lloc
1
(Rn ). In
addition, assume that the restrictions f± := f |Ω± satisfy f± ∈ C (Ω± ) and that they
1
may be extended to ensure that f± ∈ C 0 (Ω± ). Prove that for each k in {1, 2 . . . , n}
the following equality holds:
∂k u f = u∂k f + s∂Ω ( f )νk δ∂Ω in D (Rn ),

where s∂Ω ( f ) : ∂Ω → C is defined by
s∂Ω ( f )(x) := f− (x) − f+ (x)

= lim f (y) − lim f (y) for every x ∈ ∂Ω. (2.10.6)
Rn \Ωy→x Ωy→x
Exercise 2.148. Let Ω ⊂ Rn be a bounded domain of class C 1 with outward unit nor-
mal ν = (ν1 , . . . , νn ). Prove that ∂k χΩ = −νk δ∂Ω in D (Rn ) for each k ∈ {1, 2 . . . , n}.
Exercise 2.149. Suppose R ∈ (0, ∞) and let u ∈ D (Rn ) be such that
(|x|2 − R2 )u = 0 in D (Rn ). (2.10.7)
Prove that u has compact support. Give an example of a distribution u satisfying

condition (2.10.7).
Exercise 2.150. Let f ∈ C 0 (Ω) be such that u f ∈ E (Ω). Prove that f has compact
support and supp u f = supp f .
Exercise 2.151. Compute the derivatives of order m ∈ N of each distribution on R
given below.
(a) |x| (b) sgn x (c) cos x H (d) sin x H (e) x2 χ[−1,1]
Exercise 2.152. Consider the set A := {(x, y) ∈ R2 : |x − 2| + |y − 1| < 1} ⊂ R2 .
Compute (∂21 − ∂22 )χA in D (R2 ).
Exercise 2.153. Let f : R2 → R be defined by f (x, y) := χ[0,1] (x − y) for x, y ∈ R.
Compute ∂1 (u f ), ∂2 (u f ) in D (R2 ). Prove that ∂21 (u f ) − ∂22 (u f ) = 0 in D (R2 ).
Exercise 2.154. Let ψ ∈ C ∞ (Ω) be such that ψ(x) 0 for every x ∈ Ω. Prove that
for each v ∈ D (Ω) there exists a unique solution u ∈ D (Ω) of the equation ψu = v
in D (Ω).
Exercise 2.155. Let ψ ∈ C ∞ (Ω) and suppose u1 , u2 ∈ D (Ω) are such that u1 u2
and ψu1 = ψu2 in D (Ω). Prove that the set {x ∈ Ω : ψ(x) = 0} is not empty.
Exercise 2.156. Suppose {Ω j } j∈I is an open cover of the open set Ω ⊆ Rn and there

exists
of distributions {u j } j∈I such that u j ∈ D (Ω j ) for each j ∈ I and
a family

u j Ω ∩Ω = uk Ω ∩Ω in D (Ω j ∩ Ωk ) for every j, k ∈ I such that Ω j ∩ Ωk ∅. Prove
j k j k
that there exists a unique distribution u ∈ D (Ω) with the property that u = u in Ωj j
D (Ω j ) for every j ∈ I.
Exercise 2.157.
Let u ∈ D (Rn ) be such that u, ϕ = 0 for every ϕ ∈ C0∞ (Rn )
satisfying Rn ϕ(x) dx = 0. Prove that there exists c ∈ C such that u = c in D (Rn ).
Exercise 2.158. Let Ω ⊆ Rn be open and connected and let u ∈ D (Ω) be such that
∂1 u = ∂2 u = · · · = ∂n u = 0 in D (Ω). Prove that there exists c ∈ C such that u = c
in D (Ω).
Exercise 2.159. Let u ∈ D (Rn ) be such that xn u = 0 in D (Rn ). Prove that there
exists v ∈ D (Rn−1 ) such that u(x , xn ) = v(x ) ⊗ δ(xn ) in D (Rn ).
Exercise 2.160. Let u ∈ D (Rn ) be such that x1 u = · · · = xn u = 0 in D (Rn ).

Determine the expression for u.
Exercise 2.161. Let u ∈ D (Rn ) be such that ∂n u = 0 in D (Rn ). Prove that there
exists v ∈ D (Rn−1 ) such that u(x , xn ) = v(x ) ⊗ 1 in D (Rn ), where 1 denotes the
constant function (equal to 1) in R.
Exercise 2.162. Let v, w ∈ D (R) and define the distribution
u(x1 , x2 ) := 1 ⊗ v(x2 ) + w(x1 ) ⊗ 1 in D (R2 ),
where 1 denotes the constant function (equal to 1) in R. Prove that ∂1 ∂2 u = 0 in

D (R2 ).
Exercise 2.163. Let u(x1 , x2 , x3 ) := H(x1 ) ⊗ δ(x2 ) ⊗ δ(x3 ) in D (R3 ), where H is the
Heaviside function on the real line. Compute ∂1 u, ∂2 u, ∂3 u in D (R3 ).
Exercise 2.164. Consider the sequence

f j (x) := (2π)−n eix·ξ dξ, ∀ x ∈ Rn , ∀ j ∈ N. (2.10.8)
[− j, j]×···×[− j, j]
D (Rn )
Prove that f j −−−−−→ δ.
j→∞
Exercise 2.165. Solve each equation in D (R) for u.

(1) (x − 1)u = δ;
(2) xu = a, where a ∈ C ∞ (R);
(3) xu = v, where v ∈ D (R).
Exercise 2.166. Prove that the given convolutions exist and then compute them.
(a) H ∗ H
(b) H(−x) ∗ H(−x)
(c) x2 H ∗ (sin x H)
(d) χ[0,1] ∗ (xH)
(e) |x|2 ∗ δ∂B(0,r) where r > 0 and δ∂B(0,r) is as defined in Exercise 2.146 corre-
sponding to the surface Σ := ∂B(0, r).
Exercise 2.167. Let a ∈ Rn \ {0}, u j := δ ja , v j := δ− ja , for each j ∈ N. Determine

lim u j , lim v j , lim (u j ∗ v j ), in D (Rn ).
j→∞ j→∞ j→∞
(−1) j
j

Exercise 2.168. For each j ∈ N, consider the functions f j (x) := 2 χ − 1j , 1j (x) and

g j (x) := (−1) , for every x ∈ R. Determine if the given limits exist in D (R).
j
(a) lim f j
j→∞
(b) lim g j
j→∞
(c) lim ( f j ∗ g j )
j→∞
Exercise 2.169. Let u ∈ D (Rn ) and consider the map Λ : D(Rn ) → E(Rn ) given
by Λ(ϕ) := u ∗ ϕ, for every ϕ ∈ C0∞ (Rn ). Prove that Λ is a well-defined, linear, and
continuous map. Also prove that Λ commutes with translations, that is, if x0 ∈ Rn

and ϕ ∈ C0∞ (Rn ), then t x0 Λ(ϕ) = Λ t x0 ϕ , where t x0 is the map from (1.3.17).
Exercise 2.170. Suppose Λ : D(Rn ) → E(Rn ) is a linear, continuous map that com-
mutes with translations (in the sense explained in Exercise 2.169). Prove that there
exists a unique u ∈ D (Rn ) such that Λ(ϕ) = u ∗ ϕ for every ϕ ∈ C0∞ (Rn ).
Exercise 2.171. Let u ∈ E (Rn ) be such that u, xα = 0 for every α ∈ Nn . Prove that
u = 0 in E (Rn ).
Exercise 2.172. Let u : E(R) → C be the functional defined by
⎡ k ⎤
⎢⎢⎢
⎥⎥⎥
u(ϕ) := lim ⎢⎢⎣ ⎢ ϕ j − kϕ(0) − ϕ (0) ln k⎥⎥⎥⎦ ,
1
∀ ϕ ∈ C ∞ (R).
k→∞
j=1
Prove that u ∈ E (R) and determine supp u.

Exercise 2.173. For each j ∈ N consider the function f j : R → R defined by
j E (R)
f j (x) := 2 if |x| ≤ 1
j and f j (x) := 0 if |x| > 1j . Prove that f j −−−−→ δ.
j→∞
Exercise 2.174. For each j ∈ N consider the function f j : R → R defined by

f j (x) := 1j if |x| ≤ j and f j (x) := 0 if |x| > j. Prove that the sequence { f j } j∈N
converges in D (R) but not in E (R).

Exercise 2.175. Let ψ ∈ C0∞ (Rn ) be such that Rn ψ(x) dx = 1 and for each j ∈ N
E (Rn )
define f j : Rn → C by f j (x) := jn ψ( jx) for each x ∈ Rn . Prove that f j −−−−→ δ.
j→∞
Exercise 2.176. Let {x j } j∈N be a sequence of points in R . Prove that {x j } j∈N is con-
n
vergent in Rn if and only if {δ x j } j∈N is convergent in E (Rn ).

Exercise 2.177. Let a ∈ R and k ∈ N0 be given. Prove that
(x + a)δa(k) = 2a δa(k) + kδa(k−1) in D (R), (2.10.9)
(x2 − a2 )δa(k) = −2k a δa(k−1) + k(k − 1)δa(k−2) in D (R), (2.10.10)
with the convention that δa(−m) := 0 ∈ D (R) for each m ∈ N.

Chapter 3
The Schwartz Space and the Fourier
Transform
Abstract This chapter contains material pertaining to the Schwartz space of func-
tions rapidly decaying at infinity and the Fourier transform in such a setting.
3.1 The Schwartz Space of Rapidly Decreasing Functions
Recall that if f ∈ L1 (Rn ) then the Fourier transform of f is the mapping

f : Rn → C defined by

f (ξ) := e−ix·ξ f (x) dx for each ξ ∈ Rn , (3.1.1)
Rn
√
where i := −1 ∈ C. Note that under the current assumptions the integral in (3.1.1)
is absolutely convergent (which means that f is well-defined pointwise in Rn ) and
one has
sup |
f (ξ)| ≤ f L1 (Rn ) and f ∈ C 0 (Rn ). (3.1.2)
ξ∈Rn
where the second condition is seen by applying Lebesgue’s Dominated Convergence

Theorem. Hence, the mapping
F : L1 (Rn ) → {g ∈ C 0 (Rn ) : g is bounded}, F f :=

f, ∀ f ∈ L1 (Rn ), (3.1.3)
called the Fourier transform, is well defined. Besides being continuous, functions
belonging to the image of F also vanish at infinity. This property is proved next.
Proposition 3.1. If f ∈ L1 (Rn ), then lim

f (ξ) = 0.
|ξ|→∞
Proof. First, consider the case when f ∈ C0∞ (Rn ). In this scenario, integrating by
f (ξ) for every ξ ∈ Rn \ {0}, where Δ f := ∂2 f . Hence,
n
parts gives |ξ|2
f (ξ) = −Δ j
j=1

98 3 The Schwartz Space and the Fourier Transform
f (ξ)| Δ f L1 (Rn )

|Δ
|
f (ξ)| ≤ ≤ , ∀ ξ ∈ Rn \ {0}, (3.1.4)
|ξ|2 |ξ|2
from which it is clear that lim

f (ξ) = 0 in this case.
|ξ|→∞
Consider now the case when f is an arbitrary function in L1 (Rn ). Since C0∞ (Rn )
is dense in the latter space, for each ε > 0 fixed there exists g ∈ C0∞ (Rn ) such
that f − gL1 (Rn ) ≤ 2ε . Also, from what we proved so far, there exists R > 0 with
the property that | g(ξ)| ≤ 2ε whenever |ξ| > R. Keeping this in mind and using the
linearity of the Fourier transform as well as the estimate in (3.1.2) we may write
ε
|
f (ξ)| ≤ |( g(ξ)| ≤ f − gL1 (Rn ) + ≤ ε,
f − g)(ξ)| + | if |ξ| > R. (3.1.5)
2
From this, the desired conclusion follows. The proof of the proposition is therefore
complete.
We are very much interested in the possibility of extending the action of the
Fourier transform from functions to distributions, though this is going to be accom-
plished later. For now, we note the following consequence of Fubini’s theorem:

f (ξ)g(ξ) dξ = f (x)
g(x) dx, ∀ f, g ∈ L1 (Rn ). (3.1.6)
Rn Rn
Identity (3.1.6) might suggest defining the Fourier transform of a distribution based
on duality. However, there is a serious impediment in doing so. Specifically, while
for every ϕ ∈ C0∞ (Rn ) we have ϕ ∈ C ∞ (Rn ) (as may be seen directly from (3.1.1))
and ϕ decays at infinity (as proved in Proposition 3.1), we nonetheless have
F (C0∞ (Rn )) C0∞ (Rn ). (3.1.7)
In fact, we claim that
ϕ ∈ C0∞ (Rn ) and

ϕ ∈ C0∞ (Rn ) =⇒ ϕ = 0. (3.1.8)
To see that this is the case, suppose ϕ ∈ C0∞ (Rn ) is such that ϕ has compact support
in Rn , and pick an arbitrary point x∗ = (x1∗ , . . . , xn∗ ) ∈ Rn . Define the function Φ :
C → C by setting
n ∗
Φ(z) := e−izx1 + j=2 x j x j ϕ(x1 , . . . , xn ) dx1 · · · dxn , for z ∈ C. (3.1.9)
Rn
Then Φ is analytic in C and Φ(t) = ϕ(t, x2∗ , . . . , xn∗ ) for every t ∈ R. Given that ϕ has
compact support, this implies that Φ vanishes on R \ [−R, R] if R > 0 is suitably
large. The identity theorem for ordinary analytic functions of one complex variable
then forces Φ = 0 everywhere in C. In particular, ϕ(x∗ ) = Φ(x1∗ ) = 0. Since x∗ ∈ Rn
has been chosen arbitrarily, we conclude that ϕ = 0 in Rn . However, as we will see
in the sequel, the Fourier transform is injective on C0∞ (Rn ), so (3.1.8) follows.
3.1 The Schwartz Space of Rapidly Decreasing Functions 99
To overcome the difficulty highlighted in (3.1.7), we introduce a new (topological

vector) space of functions, that contains C0∞ (Rn ), is invariant under F , and whose
dual is a subspace of D (Rn ). This is the space of Schwartz functions, named after
the French mathematician Laurent–Moı̈se Schwartz (1915–2002) who pioneered the
theory of distributions and first considered this space in connection with the Fourier
transform.
Before presenting the definition of Schwartz functions, we introduce some nota-
tion, motivated by the observation that each time a partial derivative of ϕ is taken,
the exponential introduces i as a multiplicative factor. To adjust for this factor, it
is therefore natural to re-normalize the ordinary partial differentiation operators as
follows:
D j := 1i ∂ j , j = 1, . . . , n, D := (D1 , . . . , Dn ),
(3.1.10)
Dα := Dα1 1 · · · Dαn n , ∀ α = (α1 , . . . , αn ) ∈ Nn0 .
At times, we will also use subscripts to specify the variable with respect to which
the differentiation is taken. For example, Dαx stands for Dα with the additional spec-
ification that the differentiation is taken with respect to the variable x ∈ Rn .
Fix now α, β ∈ Nn0 and observe that for each ϕ ∈ C0∞ (Rn ) integration by parts
implies

β α β −ix·ξ α
ξ Dξ
ϕ(ξ) = ξ e (−x) ϕ(x) dx = (−D x )β (e−ix·ξ ) (−x)α ϕ(x) dx
Rn Rn

= e−ix·ξ Dβx (−x)α ϕ(x) dx. (3.1.11)
Rn

Hence,

ϕ(ξ) ≤
sup ξβ Dαξ |Dβx (xα ϕ(x))| dx < ∞. (3.1.12)
ξ∈Rn Rn
The conclusion from (3.1.12) is that derivatives of any order of

ϕ decrease at ∞
faster than any polynomial. This suggests making the following definition.
Definition 3.2. The Schwartz class of rapidly decreasing functions is defined

as

S(Rn ) := ϕ ∈ C ∞ (Rn ) : sup |xβ ∂α ϕ(x)| < ∞, ∀ α, β ∈ Nn0 . (3.1.13)

x∈Rn
We shall simply say that ϕ is a Schwartz function if ϕ ∈ S(Rn ). Obviously,
C0∞ (Rn ) ⊂ S(Rn ) ⊂ C ∞ (Rn ) (3.1.14)
though both inclusions are strict. An example of a Schwartz function that is not
compactly supported is provided below.
Exercise 3.3. Prove that for each fixed number a ∈ (0, ∞), the function f , defined
by f (x) := e−a|x| for each x ∈ Rn , belongs to S(Rn ) and has the property that
2
supp f = Rn .
Other elementary observations pertaining to the Schwartz class from Defini-

tion 3.2 are recorded below.
Remark 3.4. One has

S(Rn ) = ϕ ∈ C ∞ (Rn ) : sup |xβ Dα ϕ(x)| < ∞, ∀ α, β ∈ Nn0 , (3.1.15)

x∈Rn
and if ϕ ∈ C ∞ (Rn ), then ϕ ∈ S(Rn ) if and only if

sup (1 + |x|)m |∂α ϕ(x)| < ∞, ∀ m ∈ N0 , ∀ α ∈ Nn0 , |α| ≤ m. (3.1.16)
x∈Rn
Indeed, (3.1.15) is immediate from Definition 3.2. Also, the second claim in the
remark readily follows from the observation that for each m ∈ N there exists a
constant C ∈ [1, ∞) with the property that

C −1 |x|m ≤ |xγ | ≤ C|x|m , ∀ x ∈ Rn . (3.1.17)
|γ|=m
In turn, the second inequality in (3.1.17) is seen by noting that for each α ∈ Nn0 with
nonempty support and each x = (x1 , . . . , xn ) ∈ Rn we have (recall (0.0.5), (0.0.8))

|xα | = |x j |α j ≤ |x|α j = |x||α| . (3.1.18)
j∈supp α j∈supp α

To justify the first inequality in (3.1.17), consider the function g(x) := |xγ | for
|γ|=m
x ∈ Rn . Then its restriction to S n−1 attains a nonzero minimum, and the desired
inequality follows by rescaling.
Exercise 3.5. Prove that if f ∈ S(Rn ) then for every α, β ∈ Nn0 and N ∈ N there
exists C = C f,N,α,β ∈ (0, ∞) such that
α β
x ∂ f (x) ≤
C
for every x ∈ Rn . (3.1.19)
(1 + |x|)N
Use this to deduce that S(Rn ) ⊂ L p (Rn ) for every p ∈ [1, ∞].
In particular, S(Rn ) ⊂ L1 (Rn ) which, in concert with (3.1.3), allows us to con-

sider the Fourier transform on S(Rn ). Also, F C0∞ (Rn ) ⊆ S(Rn ) as seen from the
computation in (3.1.12).
Clearly, S(Rn ) is a vector space when endowed with the canonical operations of
addition of functions and multiplication by complex numbers. For a detailed discus-
sion regarding the topology we consider on S(Rn ) see Section 14.1.0.6. We continue
to denote by S(Rn ) the respective topological vector space and we single out here a
few important facts that are useful for our analysis.
Fact 3.6. S(Rn ) is a Frechét space, i.e., S(Rn ) is a locally convex, metrizable, com-
plete, topological vector space over C.
Fact 3.7. A sequence {ϕ j } j∈N ⊂ S(Rn ) converges in S(Rn ) to some ϕ ∈ S(Rn ) pro-
vided
sup xβ ∂α ϕ j (x) − ϕ(x) −−−−→ 0, ∀ α, β ∈ Nn0 , (3.1.20)
x∈Rn j→∞
S(R )
n
in which case we use the notation ϕ j −−−−→ ϕ.
j→∞
S(Rn )
Exercise 3.8. Use (3.1.17) to prove that ϕ j −−−−→ ϕ if and only if
j→∞

sup (1 + |x|)m ∂α [ϕ j (x) − ϕ(x)] −−−−→ 0, ∀ m ∈ N0 . (3.1.21)
x∈Rn j→∞
α∈Nn0 , |α|≤m
It is useful to note that the Schwartz class embeds continuously into Lebesgue
spaces.
Exercise 3.9. Prove that if p ∈ [1, ∞] and a sequence of functions {ϕ j } j∈N in S(Rn )
converges in S(Rn ) to some ϕ ∈ S(Rn ) then {ϕ j } j∈N also converges in L p (Rn ) to ϕ.
Hint: Use Exercise 3.8.

For further reference we also single out the analogue of (2.7.39) for the class of
Schwartz functions.
Remark 3.10. Let m, n ∈ N. Given ϕ ∈ C ∞ (Rn × Rm ) recall the definition of the

function ϕ ∈ C ∞ (Rm × Rn ) from (2.7.38) (presently used with V := Rn and U :=
Rm ). Then a combination of Fact 3.6, Theorem 14.1, Fact 3.7, and (2.7.38), implies
that the mapping
S(Rn × Rm ) ϕ → ϕ ∈ S(Rm × Rn ) (3.1.22)
Definition 3.11. The space of slowly increasing functions in Rn is de-

fined as

L(Rn ) := a ∈ C ∞ (Rn ) : ∀ α ∈ Nn0 ∃ k ∈ N0 such that

sup (1 + |x|)−k |∂α a(x)| < ∞ . (3.1.23)

x∈Rn
Note that an immediate consequence of Definition 3.11 is that L(Rn ) is stable

under differentiation (i.e., if a ∈ L(Rn ) then ∂α a ∈ L(Rn ) for every α ∈ Nn0 ). Also,
S(Rn ) ⊂ L(Rn ), (3.1.24)
though L(Rn ) contains many additional functions that lack decay as, for example,
the class of polynomials (other examples are contained in the two exercises below).
2
Exercise 3.12. Prove that the function f (x) := ei|x| , x ∈ Rn , belongs to L(Rn ).
Exercise 3.13. Prove that for each s ∈ R the function f (x) := (1 + |x|2 ) s , x ∈ Rn ,
belongs to L(Rn ).
Some other basic properties of the Schwartz class are collected in the next theo-
rem.

(a) For each a ∈ L(Rn ), the mapping S(Rn ) ϕ → aϕ ∈ S(Rn ) is well defined,
linear, and continuous.
(b) For every α ∈ Nn0 , the mapping S(Rn ) ϕ → ∂α ϕ ∈ S(Rn ) is well defined, linear,
and continuous.
(c) D(Rn ) → S(Rn ) → E(Rn ) and the embeddings are continuous.
(d) C0∞ (Rn ) is sequentially dense in S(Rn ). Also, the Schwartz class S(Rn ) is sequen-
tially dense in E(Rn ).
(e) If m, n ∈ N and f ∈ S(Rm ), g ∈ S(Rn ), then f ⊗ g ∈ S(Rm × Rn ) and the mapping
S(Rm ) × S(Rn ) ( f, g) → f ⊗ g ∈ S(Rm × Rn ) (3.1.25)
is bilinear and continuous.

(f) If f, g ∈ S(Rn ) then f ∗ g ∈ S(Rn ) and the mapping
S(Rn ) × S(Rn ) ( f, g) → f ∗ g ∈ S(Rn ) (3.1.26)
is bilinear and continuous.
Proof. Clearly, the mappings in (a) and (b) are linear. By Fact 3.6 and Theo-
rem 14.1, their continuity is equivalent with sequential continuity at 0, something
that can be easily checked using Fact 3.7. Moving on to the statement in (c), we
first prove that D(Rn ) embeds continuously into S(Rn ). Consider the mapping
ι : D(Rn ) → S(Rn ) defined by ι(ϕ) := ϕ for each ϕ ∈ C0∞ (Rn ). From (3.1.14)
this is a well-defined and linear mapping. To see that ι is sequentially continuous at
D(Rn )
0 ∈ D(Rn ), consider ϕ j −−−−→ 0. Then there exists a compact set K ⊂ Rn with the
j→∞
property that supp ϕ j ⊆ K for every j ∈ N, and lim sup |∂α ϕ j | = 0 for every α ∈ Nn0 .
j→∞ x∈K
Hence, for any α, β ∈ Nn0 ,

sup xβ ∂α ϕ j (x) = sup xβ ∂α ϕ j (x) ≤ C sup ∂α ϕ j (x) −−−−→ 0, (3.1.27)
x∈Rn x∈K x∈K j→∞
proving that ι is sequentially continuous at the origin. Recalling now Fact 3.6 and
Theorem 14.6, we conclude that ι is continuous.
Our next goal is to show that S(Rn ) embeds continuously in E(Rn ). From (3.1.14)
we have that the identity ι : S(Rn ) → E(Rn ) given by ι( f ) := f , for each f ∈ S(Rn ),
is a well-defined linear map. By Fact 3.6, Fact 1.8, and Theorem 14.1, ι is continuous
S(Rn )
if and only if it is sequentially continuous at zero. However, if f j −−−−→ 0 then for
j→∞
any compact set K ⊂ Rn and any α ∈ Nn0 ,
lim sup |∂α f j (x)| ≤ lim sup |∂α f j (x)| = 0. (3.1.28)

j→∞ x∈K j→∞ x∈Rn
This shows that ι is sequentially continuous at zero, finishing the proof (c).
Next, we prove the statement in (d). Let f ∈ S(Rn ) be arbitrary and, for some
fixed ψ ∈ C0∞ (Rn ) satisfying ψ ≡ 1 in a neighborhood of B(0, 1), define the sequence

of functions f j : Rn → C by setting f j (x) := ψ xj f (x) for every x ∈ Rn and every
j ∈ N. Then f j ∈ C0∞ (Rn ) and f j = f on B(0, j) for each j ∈ N. We claim that
S(Rn )
f j −−−−→ f. (3.1.29)
j→∞
To see that this is the case, if α, β ∈ Nn0 are arbitrary, by making use of Leibniz’s

formula (14.2.6) and the fact that ψ xj = 1 for each x ∈ B(0, j), we may write

β α β α! x
γ α−γ
sup x ∂ f j (x) − f (x) = sup x ∂ f (x)∂ ψ − 1
x∈Rn x∈Rn γ≤α
γ!(α − γ)! j

β α! x
γ α−γ
≤ sup x ∂ f (x)∂ ψ
|x|≥ j γ<α
γ!(α − γ)! j
x
+ sup xβ ∂α f (x) ψ − 1 . (3.1.30)
|x|≥ j j
Since ψ ∈ C0∞ (Rn ), it follows that there exists a finite constant C > 0, depending
only on ψ and α, such that
x C
sup ∂α−γ ψ ≤ , ∀ γ ∈ Nn0 , γ < α, ∀ j ∈ N. (3.1.31)
|x|≥ j j j
Also, since f ∈ S(Rn ), we may invoke (3.1.19) to conclude that there exists some
C = C f,α,β ∈ (0, ∞) such that
x C
sup xβ ∂α f (x) ψ − 1 ≤ 1 + ψL∞ (Rn ) . (3.1.32)
|x|≥ j j j
Combining (3.1.30), (3.1.31), (3.1.32), and keeping in mind that f ∈ S(Rn ), we

obtain
sup |xβ ∂α ( f j (x) − f (x))| (3.1.33)

x∈Rn

β α! C
∂ f (x) + −−−−→ 0.
C γ
≤ sup x
j x∈Rn γ≤α γ!(α − γ)! j j→∞
S(R ) n
This shows that f j −−−−→ f and completes the proof of the fact that C0∞ (Rn ) is
j→∞
sequentially dense in S(Rn ). The sequential continuity of S(Rn ) in E(Rn ) is a con-
sequence of Exercise 1.13 and (3.1.14). This completes the proof of (d).
The claims in part (e) follow using the observation that

(x, y)(α,β) ∂γx ∂μy ( f ⊗ g)(x, y) = xα ∂γ f (x)yβ ∂μ g(y), (3.1.34)
for every (x, y) ∈ Rm × Rn , for every f ∈ S(Rm ), g ∈ S(Rn ), and every α, γ ∈ Nm 0,

β, μ ∈ Nn0 .
Consider now the statement in (f). Since S(Rn ) ⊂ L1 (Rn ) (cf. Exercise 3.5) the
convolution between two functions in S(Rn ) is meaningfully defined. To see that
S(Rn ) ∗ S(Rn ) ⊂ S(Rn ), fix some arbitrary f, g ∈ S(Rn ) and α, β ∈ Nn0 . Then,
making use of the binomial theorem (cf. Theorem 14.9) as well as Exercise 3.5, we
may estimate
β α

sup x ∂ x ( f ∗ g)(x) = sup ((x − y) + y) ∂ x f (x − y)g(y) dy
β α
x∈Rn x∈Rn Rn

β!
≤ sup |(x − y)β−γ (∂α f )(x − y)||yγ g(y)| dy (3.1.35)
x∈Rn γ≤β γ!(β − γ)! Rn

≤ Cα,β sup (1 + |z|)|β| ∂α f (z) (1 + |y|)|β| |g(y)| dy
z∈Rn Rn

≤ Cα,β sup (1 + |z|)|β| ∂α f (z) sup (1 + |y|)|β|+n+1 g(y) < ∞.
z∈Rn y∈Rn
This implies f ∗ g ∈ S(Rn ). The fact that the mapping in (e) is bilinear is immedi-
ate from definitions. As regards its continuity, we may invoke again Theorem 14.1
and Fact 3.6 to reduce matters to proving sequential continuity instead. However,
the latter is apparent from the estimate in (3.1.35). This finishes the proof of the
theorem.
Exercise 3.15. Assume that ψ ∈ S(Rn ) is given and, for each j ∈ N, define the

function ψ j (x) := ψ xj for every x ∈ Rn . Prove that
S(Rn )
ψ j f −−−−→ ψ(0) f for every f ∈ S(Rn ). (3.1.36)
j→∞
Hint: Adapt estimates (3.1.30)–(3.1.31) to the current setting and, in place of

(3.1.32), this time use the Mean Value Theorem for the term ψ(x/ j) − ψ(0) to get a
decay factor of the order 1/ j.
Proposition 3.16. Let m, n ∈ N. Then C0∞ (Rm ) ⊗ C0∞ (Rn ) is sequentially dense in
S(Rm × Rn ).
Proof. Since the topology on S(Rm × Rn ) is metrizable (recall Fact 3.6), there exists
a distance function d : S(Rm × Rn ) × S(Rm × Rn ) → [0, ∞) that induces its topology.
Hence,
S(Rm ×Rn )
f j −−−−−−−→ f if and only if lim d( f j , f ) = 0. (3.1.37)
j→∞ j→∞
Now let f ∈ S(Rm × Rn ). Then by part (d) in Theorem 3.14, there exists a sequence
{ f j } j∈N ⊂ C0∞ (Rm × Rn ) with the property that d( f j , f ) < 1j for every j ∈ N. Fur-
thermore, by Proposition 2.81, for each fixed number j ∈ N, there exists a sequence
D(Rm ×Rn )
{g jk }k∈N ⊂ C0∞ (Rm ) ⊗ C0∞ (Rn ) such that g jk −−−−−−−−→ f j . In particular, by (c) in
k→∞
Theorem 3.14,
S(Rm ×Rn )
g jk −−−−−−−→ f j for each j ∈ N, (3.1.38)
k→∞
thus
lim d(g jk , f j ) = 0 for each j ∈ N. (3.1.39)
k→∞
Condition (3.1.39) implies that for each j ∈ N there exists k j ∈ N with the property
that d(g jk j , f j ) < 1j . Now the sequence {g jk j } j∈N ⊂ C0∞ (Rm ) ⊗ C0∞ (Rn ) satisfies
2
d(g jk j , f ) ≤ d(g jk j , f j ) + d( f j , f ) < for every j ∈ N. (3.1.40)
j
S(Rm ×Rn )
In turn, this forces g jk j −−−−−−−→ f , from which the desired conclusion follows.
j→∞
The analogue of Lemma 1.24 corresponding to the Schwartz class is stated next.
Exercise 3.17. Suppose A ∈ Mn×n (R) is such that det A 0. Prove that the compo-
sition mapping
S(Rn ) ϕ → ϕ ◦ A ∈ S(Rn ) is well defined, linear, and continuous. (3.1.41)
Hint: To prove continuity you may use the linearity of the map in (3.1.41), Fact 3.6,
and Theorem 14.1, to reduce matters to proving sequential continuity at 0.
We conclude this section by proving that L(Rn ) ∗ S(Rn ) ⊆ C ∞ (Rn ).
Proposition
3.18. For every function f ∈ L(Rn ) and every function g in S(Rn ) one
has Rn | f (x − y)||g(y)| dy < ∞ for each x ∈ Rn , and the convolution f ∗ g defined by

( f ∗ g)(x) := f (x − y)g(y) dy for each x ∈ Rn , (3.1.42)
Rn
has the property that f ∗ g ∈ C ∞ (Rn ).
Proof. If f, g are as in the statement, then from (3.1.23) and Exercise 3.5 it follows
that there exists M ∈ N such that for every N ∈ N there exists C ∈ (0, ∞) such that

| f (x − y)||g(y)| dy ≤ C (1 + |x − y|) M (1 + |y|)−N dy (3.1.43)
Rn Rn
for every x ∈ Rn . For each fixed point x ∈ Rn choose now N ∈ N such that N > M+n
and note that this ensures

(1 + |x − y|) M (1 + |y|)−N dy < ∞, (3.1.44)
Rn
proving the first claim in the statement. The fact that f ∗ g ∈ C ∞ (Rn ) is now seen in
a similar fashion given that ∂α f continues to be in L(Rn ) for every α ∈ Nn0 .
Exercise 3.19. Prove that L p (Rn ) ∗ S(Rn ) ⊆ C ∞ (Rn ) for every p ∈ [1, ∞].
Hint: Use the blue print as in the proof of Proposition 3.18, using Hölder’s inequality
in place of estimates for slowly increasing functions, and arrange matters so that all
derivatives fall on the Schwartz function.
We conclude this section with an integration by parts formula that will be useful
shortly.
Lemma 3.20. If f ∈ L(Rn ) and g ∈ S(Rn ), then for every α ∈ Nn0 the following
integration by parts formula holds:

α |α|
(∂ g)(x) f (x) dx = (−1) g(x)(∂α f )(x) dx. (3.1.45)
Rn Rn
Proof. Fix f ∈ L(Rn ) and g ∈ S(Rn ). Since the classes L(Rn ) and S(Rn ) are stable
under differentiation, it suffices to show that for each j ∈ {1, ..., n} we have

(∂ j g)(x) f (x) dx = − g(x)(∂ j f )(x) dx, (3.1.46)
Rn Rn
since (3.1.45) then follows by iterating (3.1.46). To this end, fix some j ∈ {1, ..., n}
along with some arbitrary R ∈ (0, ∞). The classical integration by parts formula in
the bounded, smooth, domain B(0, R) ⊂ Rn then reads (cf. (14.8.4))

(∂ j g)(x) f (x) dx = − g(x)(∂ j f )(x) dx
B(0,R) B(0,R)

+ g(x) f (x)(x j /R) dσ(x). (3.1.47)
∂B(0,R)
From part (a) in Theorem 3.14 we know that f g ∈ S(Rn ). Based on this and Exer-
cise 3.5, it follows that

g(x) f (x)(x j /R) dσ(x) ≤ ωn−1 Rn−1 sup |( f g)(x)|
∂B(0,R) |x|=R
≤ CR−1 −−−−→ 0. (3.1.48)

R→∞
3.2 The Action of the Fourier Transform on the Schwartz Class 107
On the other hand, (∂ j g) f, g∂ j f ∈ S(Rn ) ⊂ L1 (Rn ). As such, taking the limit with
R → ∞ in (3.1.47) yields (3.1.46) on account of Lebesgue’s Dominated Conver-
gence Theorem and (3.1.48).
3.2 The Action of the Fourier Transform on the Schwartz Class
Originally, we have defined the Fourier transform in (3.1.3), as a mapping acting

on functions from L1 (Rn ). Since S(Rn ) is contained in L1 (Rn ), it makes sense to
consider the Fourier transform acting on the Schwartz class. In this section, we study
the main properties of the Fourier transform in this setting. The reader is advised that
we use the symbols F and· interchangeably to denote this Fourier transform.
To state our first major result pertaining to the Fourier transform in this setting,
recall the notation introduced in (3.1.10).

(a) If f ∈ S(Rn ) and α ∈ Nn0 are arbitrary, then D α f (ξ) = ξ α
f (ξ) for every ξ ∈ Rn .
n n
(b) If f ∈ S(R ) and α ∈ N0 are arbitrary, then x f (ξ) = (−D)α
α f (ξ) for every
ξ ∈ Rn .
(c) The Fourier transform, originally introduced in the context of (3.1.3), induces a
mapping F : S(Rn ) → S(Rn ) that is linear and continuous.
(d) If m, n ∈ N, f ∈ S(Rm ) and g ∈ S(Rn ), then f ⊗g= f ⊗ g.
Proof. Fix f ∈ S(Rn ) and α ∈ Nn0 . Then the decay of f (cf. (3.1.19)) ensures that
we may differentiate under the integral sign in (3.1.1) and obtain

Dα
f (ξ) = e−ix·ξ (−x)α f (x) dx
Rn

= (−x) α f (ξ), ∀ ξ ∈ Rn . (3.2.1)
From this, the statement in (b) readily follows. Also, if β ∈ Nn0 is arbitrary, then using
the first identity in (3.2.1), the fact that ξβ e−ix·ξ = (−D x )β (e−ix·ξ ), and the integration
by parts formula from Lemma 3.20, we obtain

β α
ξ D f (ξ) = (−D x )β (e−ix·ξ )(−x)α f (x) dx
Rn

= e−ix·ξ Dβx (−x)α f (x) dx, ∀ ξ ∈ Rn . (3.2.2)
Rn
The formula in (a) follows by specializing (3.2.2) to the case when the multi-index
is α = (0, . . . , 0) ∈ Nn0 . In addition, starting with (3.2.2) we may estimate

sup |ξβ Dα
f (ξ)| ≤ (1 + |x|2 )−n dx ×
ξ∈Rn Rn

× sup (1 + |x|2 )n Dβx xα f (x) < ∞, (3.2.3)
x∈Rn
where the finiteness condition is a consequence of the membership of f to S(Rn ).

Clearly, (3.2.1) also implies that
f ∈ C ∞ (Rn ) which, in combination with (3.2.3),
shows that f ∈ S(Rn ). Hence, the mapping in (c) is well defined. The fact that this
S(Rn )
mapping is linear is immediate from definition. In addition, if f j −−−−→ 0, then based
j→∞
on the first inequality in (3.2.3) we have that, for each m, k ∈ N0 ,
sup |ξβ ∂α
f j (ξ)|
ξ∈Rn
|α|≤m, |β|≤k

≤C sup (1 + |x|2 )n ∂β xα f j (x) → 0 as j → ∞. (3.2.4)
n
x∈R
|α|≤m, |β|≤k
S(Rn )
In view of Exercise 3.8, this proves F f j −−−−→ 0. The latter combined with Fact 3.6
j→∞
and Theorem 14.1 then implies that F is continuous from S(Rn ) into S(Rn ).
At this stage we are left with proving the statement in (d). To this end, fix some
f ∈ S(Rm ) and some g ∈ S(Rn ). Then by (e) in Theorem 3.14, we have f ⊗ g ∈
S(Rm ×Rn ), so F ( f ⊗g) is well defined. Furthermore, by applying Fubini’s theorem,
we may write

f ⊗ g(ξ, η) = e−ix·ξ−iy·η ( f ⊗ g)(x, y) dy dx
Rm Rn

= e−ix·ξ f (x) dx e−iy·η g(y) dy
Rm Rn
= g(η) = (
f (ξ) f ⊗
g)(ξ, η), ∀ ξ ∈ Rm , ∀ η ∈ Rn . (3.2.5)
This finishes the proof of the theorem.
Example 3.22. Suppose λ ∈√ C satisfies Re(λ) > 0 and if λ = reiθ for r > 0 and
−π/2 < θ < π/2, set λ 2 := reiθ/2 . Consider the function f (x) := e−λ|x| for x ∈ Rn .
1 2
Then f ∈ S(Rn ) and

n2 |ξ|2

f (ξ) = π
e− 4λ for each ξ ∈ Rn . (3.2.6)
λ
Proof. Fix λ ∈ C satisfying the given hypotheses. Then Exercise 3.3 ensures that f
is a Schwartz function. Also, f (x) = e−λx1 ⊗ · · · ⊗ e−λxn for each point x = (x1 . . . , xn )
2 2
in Rn . As such, part (d) in Theorem 3.21 shows that it suffices to prove (3.2.6) when
n = 1, in which case f (x) = e−λx for every x ∈ R. Suppose that this is the case
2
and observe that f satisfies f + 2λx f = 0 in R. By taking the Fourier transform

of both sides of this differential equation, and using (a)-(b) in Theorem 3.21, we

arrive at ξ
f + 2λ f = 0. The solution to this latter ordinary differential equation

2 1
is f (ξ) = f (0)e − ξ4λ
for ξ ∈ R. There remains to show that f (0) = λπ 2 . Since by

definition, f (x) dx = e−λx dx, we are left with showing that
2
f (0) = R R

π 12
e−λx dx =
2
λ whenever λ ∈ C has Re(λ) > 0. (3.2.7)

R
1
Corresponding to the case when λ ∈ R+ , the identity R e−λx dx = λπ 2 is a standard
2
exercise in basic calculus. To extend this to complex λ’s observe that the function

π 12
e−zx dx −
2
h(z) := for z ∈ C with Re(z) > 0,
R z
is analytic and equal to zero for every z ∈ R+ . This forces h(z) = 0 for all z in C
1
with Re(z) > 0. Thus, f (0) = λπ 2 , as desired.
Exercise 3.23. Let a ∈ (0, ∞) and b ∈ R be fixed. Show that if x ∈ R then

π 12 (ξ−b)2
F (e−ax +ibx
e−
2
)(ξ) = a
4a for every ξ ∈ R. (3.2.8)
Hint: First prove that F (e−ax +ibx

)(ξ) = (F (e−ax ))(ξ − b) for every ξ ∈ R then use
2 2
Example 3.22.
Exercise 3.24. Prove that if A ∈ Mn×n (R) is such that det A 0, then
ϕ ϕ ◦ A ),
◦ A−1 = | det A| ( ∀ ϕ ∈ S(Rn ), (3.2.9)
where A−1 and A denote, respectively, the inverse and the transpose of the matrix
A.
Next we note a consequence of Theorem 3.21 of basic importance. As motiva-

tion, suppose P(D) = aα Dα is a differential operator of order m ∈ N, with
|α|≤m
constant coefficients aα ∈ C, for every α ∈ Nn0 with |α| ≤ m. Furthermore, assume
that f ∈ S(Rn ) has been given. Then any solution u ∈ S(Rn ) of the differential
equation P(D)u = f in Rn also satisfies P(ξ) u(ξ) = f (ξ) for each ξ ∈ Rn , where we

have set P(ξ) := aα ξα . In particular, if P(ξ) has no zeros, then f (ξ)
u(ξ) = P(ξ) , for
|α|≤m
every ξ ∈ Rn . This gives a formula for the Fourier transform of u. In order to find a
formula for u itself, the natural question that arises is whether we can reconstruct u
from
u. The next theorem provides a positive answer to this question in the class of
Schwartz functions.
Theorem 3.25. The mapping F : S(Rn ) → S(Rn ) is an algebraic and topologic

isomorphism, that is, it is bijective, continuous, and its inverse is also continuous.
In addition, its inverse is the operator F −1 : S(Rn ) → S(Rn ) given by the formula

(F −1 g)(x) = (2π)−n eix·ξ g(ξ) dξ, ∀ x ∈ Rn , ∀ g ∈ S(Rn ). (3.2.10)
Rn
Proof. The proof of the fact that the mapping F −1 : S(Rn ) → S(Rn ) defined as in
(3.2.10) is well defined, linear, and continuous is similar to the proof of part (c) in
Theorem 3.21. There remains to show that F −1 ◦ F = I = F ◦ F −1 on S(Rn ), where
I is the identity operator on S(Rn ). To proceed, observe that the identity F −1 ◦F = I
is equivalent to

(2π) −n
eix·ξ
f (ξ) dξ = f (x), ∀ x ∈ Rn , ∀ f ∈ S(Rn ). (3.2.11)
Rn
As regards (3.2.11), fix a function f ∈ S(Rn ) along with a point x ∈ Rn . Recall (cf.
(3.1.1)) that

f (ξ) = e−iy·ξ f (y) dy, ∀ ξ ∈ Rn . (3.2.12)
Rn
As such, one is tempted to directly replace f (ξ) in (3.2.11) by the right-hand side
of (3.2.12) and then use Fubini’s theorem to reverse the order of integration in the
variables ξ and y. There is, however, a problem in carrying out this approach, since
the function ei(x−y)·ξ f (y), considered jointly in the variable (ξ, y) ∈ Rn × Rn , does
not belong to L1 (Rn × Rn ), hence Fubini’s theorem is not necessarily applicable. To
remedy this problem, we introduce a “convergence factor” in the form of a suitable
family of Schwartz functions ψε , indexed by ε > 0 (to be specified shortly), de-
signed to provide control in the variable ξ thus ensuring the applicability of Fubini’s
theorem.
The idea is to consider Rn eix·ξ ψε (ξ) f (ξ) dξ in place of Rn eix·ξ
f (ξ) dξ and write
(granted that ψε ∈ S(Rn ))

eix·ξ ψε (ξ)
f (ξ) dξ = eix·ξ ψε (ξ) e−iy·ξ f (y) dy dξ
Rn Rn Rn

= e−i(y−x)·ξ ψε (ξ) f (y) dy dξ
Rn ×Rn

−i(y−x)·ξ ε
= e ψ (ξ) dξ f (y) dy
Rn Rn

= ψε (y − x) f (y) dy = ψε (z) f (x + z) dz. (3.2.13)
Rn Rn
Given the goal we have in mind (cf. (3.2.11)), as well as the format of the current
identity, we find it convenient to define ψε by setting
ψε (ξ) := ϕ(ε ξ) for each ξ ∈ Rn and ε > 0, (3.2.14)
where ϕ ∈ S(Rn ) is to be specified momentarily. The rationale behind this choice

is that, as we will see next, the limits as ε → 0+ of the most extreme sides in
(3.2.13) are reasonably easy to compute. This will eventually allow us to deduce
(3.2.11) from (3.2.13) by letting ε → 0+ . Concretely, from (3.2.14) we obtain that
lim+ ψε (ξ) = ϕ(0), while from the definition of the Fourier transform it is immediate
ε→0
that z
ψε (z) = ε−n ϕ = ϕ ε (z) for every z ∈ Rn . (3.2.15)
ε
Keeping this in mind and employing part (a) in Exercise 2.26 we obtain

lim+ ψε (z) f (x + z) dz = lim+
ϕ ε (z) f (x + z) dz
ε→0 Rn ε→0 Rn

=
ϕ(z) dz f (x). (3.2.16)
Rn
Also, on account of (3.2.14) and the fact that

f ∈ S(Rn ) ⊂ L1 (Rn ), Lebesgue’s
Dominated Convergence Theorem gives

lim+ ix·ξ ε
e ψ (ξ) f (ξ) dξ = lim+ eix·ξ ϕ(ε ξ)
f (ξ) dξ
ε→0 Rn ε→0 Rn

= ϕ(0) eix·ξ
f (ξ) dξ. (3.2.17)
Rn
In summary,
(3.2.13), (3.2.16), and (3.2.17), show that whenever ϕ ∈ S(Rn ) is such
that Rn
ϕ(z) dz 0 we have

C eix·ξ
f (ξ) dξ = f (x), ∀ x ∈ Rn , ∀ f ∈ S(Rn ), (3.2.18)
Rn
where the normalization constant C is given by

ϕ(0)
C := . (3.2.19)
Rn

ϕ(z) dz
As such, (3.2.11) will follow as soon as we prove that C = (2π)−n . For this task, we
have the freedom of choosing the function ϕ ∈ S(Rn ) and a candidate that springs to
mind is the Schwartz function from Example 3.22 (say, in the particular case when
λ = 1). Hence, if ϕ(x) := e−|x| for each x ∈ Rn , formula (3.2.6) gives
2
|ξ|2
ϕ(ξ) = π 2 e−
n
4 for each ξ ∈ Rn . (3.2.20)
Consequently,
2 n
|t|2
− |ξ|4
e−
n n n n

ϕ(ξ) dξ = π 2 e dξ = π 2 4 dt = π 2 (4π) 2 = (2π)n . (3.2.21)
Rn Rn R
where the second equality is simply Fubini’s theorem, while the third equality is
provided by (3.2.7) with λ := 1/4. Since in this case we also have ϕ(0) = 1, it
follows that C = (2π)−n , as wanted. This finishes the justification of the identity
F −1 ◦ F = I on S(Rn ). The same approach also works to show F ◦ F −1 = I,
completing the proof of the theorem.
In what follows, for an arbitrary function f : Rn → C we define
f ∨ (x) := f (−x), ∀ x ∈ Rn . (3.2.22)
Exercise 3.26. Prove that the mapping
S(Rn ) f → f ∨ ∈ S(Rn ) (3.2.23)
is well defined, linear, and continuous.
Hint: Use Fact 3.6 and Theorem 14.1.

Recall that z denotes the complex conjugate of z ∈ C.
Exercise 3.27. Let f ∈ S(Rn ). Then the following formulas hold:

(1) f∨ = (f )∨ ;
∨
(2) f = f ;

(3) f = (2π)n f ∨ ;
(4) Rn f (x) dx = f (0) and Rn
f (ξ) dξ = (2π)n f (0).
Proposition 3.28. Let f, g ∈ S(Rn ). Then the following identities hold:

(a) Rn f (x)g(x) dx = Rn
f (ξ)g(ξ) dξ;

(b) Rn f (x)g(x) dx = (2π)−n Rn f (ξ)g(ξ) dξ an identity referred to in the literature
as Parseval’s identity;
(c) f∗g= f ·
g;
(d) f· g = (2π)−n
f ∗
g.
Proof. The identity in (a) follows via a direct computation using Fubini’s theorem.

Also, based on Exercise 3.27, we have g = g∨ = (2π)n g which, when combined
with (a) gives (b). The identity in (c) follows using Fubini’s theorem. Specifically,
for each ξ ∈ Rn we may write

f ∗ g(ξ) = e −ix·ξ
( f ∗ g)(x) dx = e−ix·ξ
f (x − y)g(y) dy dx
Rn Rn Rn

= g(y) e−ix·ξ f (x − y) dx dy
Ru Rn

= e−iy·ξ
g(y) e−iz·ξ f (z) dz dy =
f (ξ)
g(ξ), (3.2.24)
R Rn
as wanted. Next, the identity from (c) combines with Exercise 3.27 to yield

g=
f ∗ g = (2π)2n f ∨ · g∨ = (2π)2n ( f · g)∨ .
f · (3.2.25)
Applying now the Fourier transform to the most extreme sides of (3.2.25) and once
again invoking Exercise 3.27, we obtain
∨
−n

−2n
∨
f · g)∨ = f
(2π) f ∗ g = (2π) f ∗ g = (
· g. (3.2.26)
This completes the proof of the proposition.
Remark 3.29.
(1) It is not difficult to see via a direct computation that we also have

g(x) dx =
f (x)
f (ξ)g(ξ) dξ, ∀ f ∈ L1 (Rn ), ∀ g ∈ S(Rn ). (3.2.27)
Rn Rn
(2) Parseval’s identity written for f = g ∈ S(Rn ) becomes

n
f L2 (Rn ) = (2π) 2 f L2 (Rn ) . (3.2.28)
As a consequence, since C0∞ (Rn ) is dense in L2 (Rn ), the Fourier transform F

may be extended to a linear operator from L2 (Rn ) into itself, and the latter iden-
tity continues to hold for every f ∈ L2 (Rn ). In summary, this extension of F ,
originally considered as in part (c) of Theorem 3.21, satisfies
F : L2 (Rn ) → L2 (Rn ) is linear and continuous and

n (3.2.29)
F f L2 (Rn ) = (2π) 2 f L2 (Rn ) , ∀ f ∈ L2 (Rn ).
Based on this, part (3) in Exercise 3.27, the continuity of the linear mapping
L2 (Rn ) f → f ∨ ∈ L2 (Rn ), and the density of Schwartz functions in L2 (Rn ), we
further deduce that

F F f = (2π)n f ∨ , ∀ f ∈ L2 (Rn ). (3.2.30)
Combined with (3.2.29), this proves that
F : L2 (Rn ) → L2 (Rn ) is a linear, continuous, isomorphism,

(3.2.31)
and F −1 f = (2π)−n F f ∨ = (2π)−n F f ∨ , ∀ f ∈ L2 (Rn ).
We will continue to use the notation f for F f whenever f ∈ L2 (Rn ). The identity
in (3.2.29) is called Plancherel’s identity. The same type of density
argument shows that formula from part (b) of Proposition 3.28 extends to

f (x)g(x) dx = (2π)−n
f (ξ)g(ξ) dξ, ∀ f, g ∈ L2 (Rn ), (3.2.32)
Rn Rn
to which we continue to refer as Parseval’s identity.

(3) An inspection of the computation in (3.2.24) reveals that the identity f ∗g= f ·
g
remains valid if f, g ∈ L (R ).
1 n

Exercise 3.30. Prove that Rn f (x)g(x) dx = Rn
f (ξ)g(ξ) dξ for all f, g ∈ L2 (Rn ).
Hint: Use part (a) in Proposition 3.28, (3.2.29), and the fact that C0∞ (Rn ) is se-
quentially dense in L2 (Rn ) to first prove the desired identity for f ∈ L2 (Rn ) and
g ∈ S(Rn ).
Further Notes for Chapter 3. The basic formalism associated with the Fourier transform goes
back to the French mathematician and physicist Joseph Fourier (1768–1830) in a more or less
precise form. A distinguished attribute of this tool, of fundamental importance in the context of
partial differential equations, is the ability to render the action of a constant coefficient differential
operator simply as ordinary multiplication by its symbol on the Fourier transform side. As the
name suggest, the Schwartz space of rapidly decreasing functions has been formally introduced
by Laurent Schwartz who was the first to recognize its significance in the context of the Fourier
transform. Much of the elegant theory presented here is due to him.
Exercise 3.31. Prove that if f ∈ Lcomp

1
(Rn ) then
f ∈ C ∞ (Rn ).
Exercise 3.32. Prove that if f ∈ L1 (Rn ) is real-valued and odd, then so is

f.
Exercise 3.33. Prove that if f ∈ S(Rn ) then for every α, β ∈ Nn0 one has

lim sup xα ∂β f (x) = 0. (3.3.1)
R→∞ |x|≥R
Exercise 3.34. Let ϕ ∈ C0∞ (Rn ) be such that ϕ 0 and for each j ∈ N set
ϕ j (x) := e− j ϕ(x/ j), ∀ x ∈ Rn .

S(Rn )
Prove that ϕ j −−−−→ 0 but the sequence {ϕ j } j∈N does not converge in D(Rn ).
j→∞
Exercise 3.35. Let ϕ ∈ C0∞ (Rn ) be such that ϕ 0 and for each j ∈ N set
1
ϕ j (x) := ϕ(x/ j), ∀ x ∈ Rn .
j
E(Rn )
Prove that ϕ j −−−−→ 0 but the sequence {ϕ j } j∈N does not converge in S(Rn ).
j→∞
Exercise 3.36. Let θ ∈ C0∞ (R) be such that θ(x) = 1 for |x| ≤ 1, and let ψ in C ∞ (R)
be such that ψ(x) = 0 for x ≤ −1 and ψ(x) = e−x for x ≥ 0. For each j ∈ N then set
1
ϕ j (x) := ψ(x)θ(x/ j), ∀ x ∈ R.
j
Prove that the sequence {ϕ j } j∈N converges in S(R).

Exercise 3.37. Determine which of the following functions belongs to S(Rn ).
(a) e−(x1 +x2 +···+xn )

2 2
(b) (x12 + x22 + · · · + xn2 )n! e−|x|

2
(c) (1 + |x|2 )−2

n
2
sin(e−|x| )
(d) 1+|x|2
2
cos(e−|x| )
(e) (1+|x|2 )n
(f) e−|x| sin(e x1 )

2 2
(g) e−(Ax)·x , where A ∈ Mn×n (R) is symmetric and satisfies (Ax) · x > 0 for all
x ∈ Rn \ {0} (as before, “·” denotes the dot product of vectors in Rn ).
Exercise 3.38. Let A ∈ Mn×n (R) be symmetric and such that (Ax) · x > 0 for every
x ∈ Rn \ {0}. Prove that if we define f (x) := e−(Ax)·x for x ∈ Rn , then
f (ξ) =
n −1 ξ)·ξ
√π
(A
2
det A
e− 4 for every ξ ∈ Rn .
Exercise 3.39. Prove that f : R2 → R defined by f (x1 , x2 ) := e−(x1 +x1 x2 +x2 ) for
2 2
(x1 , x2 ) ∈ R2 belongs to S(R2 ), then compute its Fourier transform.

Exercise 3.40. If P(x) is a polynomial in Rn , compute the Fourier transform of the
function defined by f (x) := P(x)e−|x| for each x ∈ Rn .
2
Exercise 3.41. If a ∈ (0, ∞) and x0 ∈ Rn are fixed, compute the Fourier transform
of the function defined by f (x) := e−a|x| sin(x · x0 ) for each x ∈ Rn .
2
Exercise 3.42. Let ϕ ∈ S(R). Prove that the equation ψ = ϕ has a solution ψ ∈
S(R) if and only if R ϕ(x) dx = 0.
Exercise 3.43. Does the equation ψ = e−x have a solution in S(R)?

2
Exercise 3.44. Fix x0 ∈ Rn . Prove that the translation map t x0 from (1.3.17) extends
linearly and continuously as a map from S(Rn ) into itself. More precisely, show that
the translation map t x0 : S(Rn ) → S(Rn ) defined by t x0 (ϕ) := ϕ(· − x0 ) for every
ϕ ∈ S(Rn ), is linear and continuous.
Also, prove that for every ϕ ∈ S(Rn ) the following identities hold in S(Rn )

F t x0 (ϕ) (ξ) = e−ix0 ·ξ
ϕ(ξ) and ϕ = F eix0 ·x ϕ .
t x0 (3.3.2)
Chapter 4
The Space of Tempered Distributions
Abstract The action of the Fourier transform is extended to the setting of tem-
pered distributions, and several distinguished subclasses of tempered distributions
are introduced and studied, including homogeneous and principal value distribu-
tions. Significant applications to harmonic analysis and partial differential equa-
tions are singled out. For example, a general, higher dimensional jump-formula is
deduced in this chapter for a certain class of tempered distributions, which includes
the classical harmonic Poisson kernel that is later used as the main tool in deriv-
ing information about the boundary behavior of layer potential operators associated
with various partial differential operators and systems. Also, one witnesses here how
singular integral operators of central importance to harmonic analysis, such as the
Riesz transforms, naturally arise as an extension to the space of square-integrable
functions, of the convolution product of tempered distributions of principal value
type with Schwartz functions.
4.1 Definition and Properties of Tempered Distributions
The algebraic dual of S(Rn ) is the vector space

u : S(Rn ) → C : u is linear and continuous . (4.1.1)
Functionals u belonging to this space are called tempered distributions

(a piece of terminology justified a little later). An important equivalent condition
for a linear functional on S(Rn ) to be a tempered distribution is stated next (see
(14.1.33)).
Fact 4.1 A linear functional u : S(Rn ) → C is continuous if and only if there exist
m, k ∈ N0 , and a finite constant C > 0, such that

|u(ϕ)| ≤ C sup sup xβ ∂α ϕ(x) , ∀ ϕ ∈ S(Rn ). (4.1.2)
α,β∈Nn0 , |α|≤m, |β|≤k x∈Rn

118 4 The Space of Tempered Distributions
Exercise 4.2. Prove that a linear functional u : S(Rn ) → C is continuous if and only
if there exist m, k ∈ N0 , and a finite constant C > 0, such that

|u(ϕ)| ≤ C sup sup (1 + |x|) j ∂α ϕ(x) , ∀ ϕ ∈ S(Rn ). (4.1.3)
α∈Nn0 , |α|≤m, 0≤ j≤k x∈Rn
Hint: Use Fact 4.1 and (3.1.17).

From Fact 3.6 and Theorem 14.1 we also know that any f : S(Rn ) → C (linear
or not) is continuous if and only if it is sequentially continuous. As a consequence,
we have the following characterization of tempered distributions.
Proposition 4.3. Let u : S(Rn ) → C be linear. Then u is a tempered distribution if

S(Rn )
and only u(ϕ j ) −−−−→ 0 whenever ϕ j −−−−→ 0.
j→∞ j→∞
As discussed in Example 2.7, to any locally integrable function f in Rn one

can associate an “ordinary” distribution u f ∈ D (Rn ). This being said, it is not
always the case that u f is actually a tempered distribution (for more on this, see
Remark 4.16). This is, however, true if the locally integrable function f becomes
integrable at infinity after being tempered by a polynomial. We elaborate on this
issue in the next example.
Example 4.4. Let f ∈ Lloc

1
(Rn ) be such that there exist some m ∈ [0, ∞) and some
R ∈ (0, ∞) with the property that

|x|−m | f (x)| dx < ∞. (4.1.4)
|x|≥R
We claim that the distribution of function type defined by f is a tempered distribu-

tion, that is, the mapping

u f : S(Rn ) → C, u f (ϕ) := f ϕ dx, ∀ ϕ ∈ S(Rn ), (4.1.5)
Rn
is a well-defined tempered distribution. To see that this is the case, pick N ∈ N0 such
that N ≥ m and, for an arbitrary ϕ ∈ S(Rn ), estimate

| f ϕ| dx ≤ | f (x)ϕ(x)| dx + |x|−m | f (x)||x|m |ϕ(x)| dx
Rn |x|<R |x|≥R

≤ C sup |x|k |ϕ(x)| | f (x)| dx + |x|−m | f (x)| dx
x∈Rn , 0≤k≤N |x|<R |x|≥R

≤ C sup |x| |ϕ(x)| ,
k
(4.1.6)
x∈Rn , 0≤k≤N
where we have used the easily checked fact that, since N ≥ m, we have

sup |x|m |ϕ(x)| ≤ Rm−N sup |x|N |ϕ(x)| . (4.1.7)
|x|≥R x∈Rn
4.1 Definition and Properties of Tempered Distributions 119
Estimate (4.1.6) shows that the functional u f is well defined on S(Rn ). Clearly u f is
linear, and (4.1.6) also implies (based on Fact 4.1) that u f is continuous on S(Rn ).
Hence, u f is a tempered distribution.
In the sequel, we will often use the notation f instead of u f , whenever f is such
that the operator u f as in (4.1.5) is linear and continuous.
Exercise 4.5. Prove that for each a ∈ (−n, ∞) the function |x|a is a tempered distri-
bution in Rn .
Exercise 4.6. Let f : Rn → C be a Lebesgue measurable function with the property
that there exists m ∈ R such that

(1 + |x|)m | f (x)| dx < ∞. (4.1.8)
Rn
Then the mapping u f as in (4.1.5) is well defined and is a tempered distribution.

Hint: Use Example 4.4.
Exercise 4.7. Prove that L(Rn ), the space of slowly increasing functions defined in
(3.1.23), is contained in the space of tempered distributions.
Example 4.8. Let p ∈ [1, ∞] and f ∈ L p (Rn ) be arbitrary. We claim that u f defined
in (4.1.5) is a tempered distribution. To prove this claim, since f is measurable, by
Exercise 4.6, it suffices to show that f satisfies (4.1.8). If p = 1, then (4.1.8) holds
for m = 0, while if p = ∞, (4.1.8) holds for any m < −n. If p ∈ (1, ∞), by applying
Hölder’s inequality, we have

p−1
p
mp
(1 + |x|)m | f (x)| dx ≤ f L p (Rn ) (1 + |x|) p−1 dx <∞
Rn Rn
provided we choose m < − n(p−1)

p . In summary, we proved that
for each p ∈ [1, ∞] the space L p (Rn ) is

(4.1.9)
a subspace of the space of tempered distributions.
As an immediate consequence of Exercise 3.5 and (4.1.9) we therefore obtain
S(Rn ) is a subspace of the space of tempered distributions. (4.1.10)
Example 4.9. From Example 2.9 we know that ln |x| ∈ Lloc 1

(Rn ) and, hence, defines
a distribution in R . We claim that the distribution ln |x| is, in fact, a tempered dis-
n
tribution. Indeed, this follows from Exercise 4.6, since (4.1.8) holds for any m < −n
(seen by using estimate (2.1.9) with 0 < ε < min {n, −m − n}).
The topology we consider on the space of tempered distributions is the weak∗-
topology and we denote this topological vector space by S (Rn ). In particular, from
the general discussion in Section 14.1 we have:
Fact 4.10 S (Rn ) is a locally convex topological space.

Fact 4.11 A sequence {u j } j∈N ⊂ S (Rn ) converges to some u ∈ S (Rn ) in S (Rn )
if and only if u j , ϕ −−−−→ u, ϕ for every ϕ ∈ S(Rn ), in which case we write
j→∞
S (Rn )
u j −−−−−→ u.
j→∞
It is useful to note that, for each p ∈ [1, ∞], the space of p-th power integrable
functions in Rn embeds continuously in the space of tempered distributions.
Exercise 4.12. Prove that if p ∈ [1, ∞] and a sequence of functions { f j } j∈N in L p (Rn )
S (Rn )
converges in L p (Rn ) to some f ∈ L p (Rn ) then f j −−−−−→ f .
j→∞
Hint: Use Hölder’s inequality and (4.1.9).

Exercise 4.13. Assume that φ is as in (1.2.3) and recall the sequence of functions
S (Rn )
{φ j } j∈N from (1.3.7). Prove that φ j −−−−−→ δ.
j→∞
Hint: Reason as in Example 2.24.

Let m, n ∈ N be arbitrary. As a consequence of Remark 3.10 we obtain (compare
to (2.7.41)):
if w ∈ S (Rm × Rn ), then the map w : S(Rn × Rm ) → C defined by

w (ϕ) := w, ϕ for each ϕ ∈ S(Rn × Rm ) is a tempered distribution in
Rn × Rm , i.e., w ∈ S (Rn × Rm ).
(4.1.11)
Theorem 4.14. For each n, m ∈ N the following statements are true:
(a) E (Rn ) → S (Rn ) → D (Rn ), where all the embeddings are injective and con-
tinuous.
(b) For each a ∈ L(Rn ) and each u ∈ S (Rn ), the distribution au ∈ D (Rn ) extends
uniquely to a tempered distribution, which will be denoted by au and its action is
given by au, ϕ = u, aϕ for every ϕ ∈ S(Rn ). Moreover, the mapping
S (Rn ) u → au ∈ S (Rn ) (4.1.12)
is linear, and continuous.

(c) For every α ∈ Nn0 and each u ∈ S (Rn ), the distribution ∂α u ∈ D (Rn ) extends
uniquely to a tempered distribution, that will be denoted by ∂α u and its action is
given by
∂α u, ϕ = (−1)|α| u, ∂α ϕ for every ϕ ∈ S(Rn ). (4.1.13)
Moreover, the mapping
S (Rn ) u → ∂α u ∈ S (Rn ) (4.1.14)

(d) If u ∈ S (Rm ) and v ∈ S (Rn ) then the distribution u ⊗ v originally defined in

D (Rm × Rn ) extends uniquely to a tempered distribution, that will be denoted by
u ⊗ v and its action is given by

u ⊗ v, ϕ = u(x), v(y), ϕ(x, y) = v(y), u(x), ϕ(x, y) ,
(4.1.15)
∀ ϕ ∈ S(Rm × Rn ).
Hence,
u ⊗ v, ϕ1 ⊗ ϕ2 = u, ϕ1 v, ϕ2
(4.1.16)
for each ϕ1 ∈ S(Rm ) and each ϕ2 ∈ S(Rn ).
In addition, u ⊗ v = (v ⊗ u) in S (Rm × Rn ), for every u ∈ S (Rm ) and every
v ∈ S (Rn ). Moreover, the mapping
S (Rm ) × S (Rn ) (u, v) → u ⊗ v ∈ S (Rm × Rn ) (4.1.17)
is bilinear and separately continuous.
Proof. The statement in (a) is a consequence of parts (c) and (d) in Theorem 3.14
and duality (cf. Proposition 14.4). Next, fix a ∈ L(Rn ) and let u ∈ S (Rn ) be arbi-
trary. Then based on part (a) we have u ∈ D (Rn ). Hence, by Proposition 2.29, au
exists and belongs to D (Rn ). We will show that au may be extended uniquely to a
tempered distribution. Define

au : S(Rn ) → C,
au(ϕ) := u, aϕ ∀ ϕ ∈ S(Rn ). (4.1.18)
au is the composition of u ∈ S (Rn ) with the map in part (a) of Theorem 3.14,
Since
of which are linear and continuous, it follows that
both au ∈ S (Rn ). In addition,

au ∞ n = au, so if we invoke (d) in Theorem 3.14, we obtain that au is the unique
C0 (R )
continuous extension of au to S(Rn ). The map in (4.1.12) is also continuous as
seen from Theorem 3.14 and the general fact that the transpose of any linear and
continuous operator between two topological vector spaces is continuous at the level
of dual spaces equipped with weak∗-topologies (cf. Proposition 14.2). Re-denoting

au simply as au now finishes the proof of the statement in (b). The proof of (c) is
similar to the proof of (b).
Turning our attention to (d), let u ∈ S (Rm ) and v ∈ S (Rn ). By part (a) we have
u ∈ D (Rm ) and v ∈ D (Rn ), hence (by Theorem 2.87) u⊗v belongs to D (Rm ×Rn ).
We construct an extension u ⊗ v : S(Rn+m ) → C by setting

u
⊗ v, ϕ := u(x), v(y), ϕ(x, y) , ∀ ϕ ∈ S(Rm × Rn ). (4.1.19)
To see that this mapping is in S (Rm × Rn ), fix ϕ ∈ S(Rm × Rn ). Clearly we have

ϕ(x, ·) ∈ S(Rn ) for each x ∈ Rm . We claim that
ψ(x) := v(y), ϕ(x, y) for each x ∈ Rm =⇒ ψ ∈ S(Rm ). (4.1.20)

Indeed, by reasoning as in the proof of Proposition 2.84, we obtain that ψ belongs

to C ∞ (Rm ). Also, for α, β ∈ Nm β α β α
0 we have x ∂ ψ(x) = v(y), x ∂ x ϕ(x, y) and, since
v ∈ S (Rn ), there exist C > 0, and , k ∈ N0 , such that v satisfies (4.1.2). Thus, for
every x ∈ Rm , we have
β α
x ∂ ψ(x) = v(y), xβ ∂αx ϕ(x, y) ≤ C sup yγ xβ ∂αx ∂δy ϕ(x, y). (4.1.21)
y∈Rn
|γ|≤, |δ|≤k
Therefore, since ϕ ∈ S(Rm × Rn ), estimate (4.1.21) further yields

sup xβ ∂α ψ(x) ≤ C sup yγ xα ∂βx ∂δy ϕ(x, y) < ∞. (4.1.22)
x∈Rm x∈Rm , y∈Rn
|γ|≤, |δ|≤k
This shows that ψ ∈ S(Rm ), finishing the proof of the claim. Consider next the
mapping
S(Rm × Rn ) ϕ → ψ ∈ S(Rm ). (4.1.23)
This is linear by design, as well as continuous (as seen from (4.1.22), Fact 3.6, and
Theorem 14.1). Thus, the composition between u and the map in (4.1.23) gives rise
to a linear and continuous map, which proves that u ⊗ v is a tempered distribution
in Rm × Rn . In addition, (4.1.19) and (ii) in Theorem 2.87 imply

⊗ vC ∞ (Rm ×Rn ) = u ⊗ v,
u (4.1.24)
0
which in combination with (d) in Theorem 3.14, gives that u

⊗ v is the unique con-
tinuous extension of u ⊗ v to S(Rm × Rn ).
Next, we note that the reasoning above also yields that

η(y) := u(x), ϕ(x, y) for each y ∈ Rn =⇒ η ∈ S(Rn ) (4.1.25)
and that the mapping
S(Rm × Rn ) ϕ → η ∈ S(Rn ) is linear and continuous. (4.1.26)
Consequently, the composition between v and the map in (4.1.26) gives rise to a
linear and continuous map w defined by

w(ϕ) := v(y), u(x), ϕ(x, y) for each ϕ ∈ S(Rm × Rn ). (4.1.27)

By (ii) in Theorem 2.87, we have that wC ∞ (Rm ×Rn ) = u ⊗ v. Since we have proved
0
that u
⊗ v is the unique continuous extension of u ⊗ v to S(Rm × Rn ), we must have
w = u ⊗ v. This completes the proof of (4.1.15). Moreover, from (4.1.19) we see
that

u
⊗ v, ϕ1 ⊗ ϕ2 = u, ϕ1 v, ϕ2 , ∀ ϕ1 ∈ S(Rm ), ∀ ϕ2 ∈ S(Rn ). (4.1.28)
Similarly, we define v ⊗ u ∈ S (Rn ×Rm ), the unique extension of the distribution

v ⊗ u ∈ D (R × R ) to a tempered distribution. Recall (4.1.11) and consider the
m n

tempered distribution v ⊗ u in S (Rm × Rn ). Then for each test function ϕ ∈
C0∞ (Rm × Rn ) we may write

⊗ u , ϕ = v
v ⊗ u, ϕ = v ⊗ u, ϕ = (v ⊗ u) , ϕ

= u ⊗ v, ϕ = u ⊗ v, ϕ , (4.1.29)
where for the fourth equality in (4.1.29) we have used item (iii) in Theorem 2.87.
This proves that

u⊗ vC ∞ (Rm ×Rn ) = v⊗ u ∞ m n . (4.1.30)
0 C0 (R ×R )

Thus, using (d) in Theorem 3.14, we conclude u ⊗ v = v ⊗ u .
Clearly, S (Rm ) × S (Rn ) (u, v) → u
⊗ v ∈ S (Rm × Rn ) is bilinear, and our goal
is to show that this is also separately continuous. First we will prove that this map
is continuous in the first variable. For this, we shall rely on the abstract description
of open sets in the weak∗-topology from (14.1.9). Specifically, having fixed v ∈
S (Rn ), pick an arbitrary finite set A ⊂ S(Rm × Rn ) along with some number ε ∈
(0, ∞), and introduce

OA,ε := w ∈ S (Rm × Rn ) : |w, ψ | < ε, ∀ ψ ∈ A . (4.1.31)
If we now define A := v(y), ψ(·, y) : ψ ∈ A, then from what we proved earlier
is a subset of S (Rm ). Also, A
(cf. (4.1.20)) it follows that A is finite since A is finite.
Hence, if we now set
:= u ∈ S (Rm ) : |u, ϕ | < ε, ∀ ϕ ∈ A
O , (4.1.32)
A,ε
using (4.1.19) we have u ⊗ v ∈ OA,ε for every u ∈ O . In light of (14.1.9), this
A,ε
shows that u → u ⊗ v is continuous. On account of formula (4.1.15), a similar proof
also gives that v → u ⊗ v is continuous.
Finally, abbreviating matters by simply writing u ⊗ v in place of u⊗ v, all claims
in part (d) of the statement of the theorem now follow.
Remark 4.15.
(i) In view of (a) in Theorem 4.14 and (d) in Theorem 3.14 we have:
if u, v ∈ S (Rn ) and u = v in D (Rn ), then u = v in S (Rn ). (4.1.33)
(ii) Whenever u ∈ S (Rn ) its support is understood as defined in (2.5.8).

Remark 4.16. The inclusion S (Rn ) → D (Rn ) (which goes to show that any tem-
pered distribution is indeed an ordinary distribution) is actually strict. This follows
from the observation that, in contrast to the case of ordinary distributions,
1
Lloc (Rn ) S (Rn ). (4.1.34)
To justify (4.1.34), take λ > 0 arbitrary, fixed, and consider the function
λ
f : Rn → C, f (x) := e|x| for every x ∈ Rn . (4.1.35)
Clearly f ∈ Lloc 1
(Rn ), hence f ∈ D (Rn ). However, this distribution cannot be
extended to a tempered distribution. To see why this is true, suppose there exists
u ∈ S (Rn ) such that uC ∞ (Rn ) = f in D (Rn ). Since u is a tempered distribution there
0
exist some finite constant C ≥ 0 and numbers k, m ∈ N0 such that
β α
u, ϕ ≤ C sup x ∂ ϕ(x), ∀ ϕ ∈ S(Rn ). (4.1.36)
x∈Rn , |α|≤k, |β|≤m
Now fix a function
ψ ∈ C0∞ (Rn ), ψ ≥ 0, supp ψ ⊆ B(0, 1),

(4.1.37)
ψ(x) dx = 1,
1/2≤|x|≤1
and for each j ∈ N define ψ j (x) := ψ(x/ j) for every x ∈ Rn . Then by (4.1.36), for
each j ∈ N, we have
β α
u, ψ j ≤ C sup x ∂ ψ j (x) ≤ C jm−k , (4.1.38)
x∈B(0, j), |α|≤k, |β|≤m
for some C ∈ [0, ∞) independent of j. On the other hand, (4.1.37) forces the lower
bound
|x| λ

λ
u, ψ j = e ψ j (x) dx ≥ e|x| ψ j (x) dx
|x|≤ j j/2≤|x|≤ j

λ λ
≥ e( j/2) j−n ψ(x) dx = e( j/2) j−n , ∀ j ∈ N. (4.1.39)
1/2≤|x|≤1
Comparing (4.1.38) and (4.1.39) yields a contradiction choosing j large enough.

Hence, f cannot be extended to a tempered distribution.
Remark 4.17. What prevents locally integrable functions of the form (4.1.35) from
belonging to S (Rn ) is the fact that their growth at infinity is not tempered enough
and, in fact, this observation justifies the very name “tempered distribution.”
Exercise 4.18. Let λ ∈ R be such that λ < n − 1, and fix some j in {1, . . . , n}. For
these choices, consider the functions f, g defined, respectively, by f (x) := |x|−λ and
g(x) := −λx j |x|−λ−2 for each x ∈ Rn \ {0}. Prove that f, g ∈ S (Rn ) and that ∂ j f = g
in S (Rn ). In short,

∂ j |x|−λ = −λx j |x|−λ−2 in S (Rn )
(4.1.40)
whenever λ < n − 1 and j ∈ {1, . . . , n}.
Hint: To prove (4.1.40) use (4.1.13), and integration by parts coupled with a limiting
argument to extricate the singularity at the origin.
Given a tempered distribution, we may consider the convolution between its
restriction to C0∞ (Rn ) and any compactly supported distribution. A natural ques-
tion, addressed in the next theorem, is whether the distribution obtained via such a
convolution may be extended to a tempered distribution.
Theorem 4.19. The following statements are true:

(a) If u ∈ E (Rn ) and v ∈ S (Rn ), then the distribution u ∗ v ∈ D (Rn ) extends
uniquely to a tempered distribution, that will be denoted by u ∗ v. Also, the distri-
bution v∗u ∈ D (Rn ) extends uniquely to a tempered distribution and u ∗v = v∗u
in S (Rn ). Moreover, the mapping
E (Rn ) × S (Rn ) (u, v) → u ∗ v ∈ S (Rn ) (4.1.41)
is bilinear, and for every u ∈ E (Rn ), v ∈ S (Rn ), we have
∂α (u ∗ v) = (∂α u) ∗ v = u ∗ (∂α v) in S (Rn ), ∀ α ∈ Nn0 . (4.1.42)

D (Rn )
(b) If u j −−−−−→ u and there exists a compact set K ⊂ Rn with the property that
j→∞
S (Rn )
supp u j ⊂ K for every j ∈ N, then u j ∗ v −−−−−→ u ∗ v for each v ∈ S (Rn ).
j→∞
S (Rn ) S (Rn )
(c) If v j −−−−−→ v, then u ∗ v j −−−−−→ u ∗ v for each u ∈ E (Rn ).
j→∞ j→∞
(d) If u ∈ S (Rn ) then δ ∗ u = u = u ∗ δ for every u ∈ S (Rn ).
(e) Let a ∈ S(Rn ) and u ∈ S (Rn ). Then the mapping
a ∗ u : S(Rn ) → C, (a ∗ u)(ϕ) := u, a∨∗ ϕ ∀ ϕ ∈ S(Rn ), (4.1.43)
is well defined, linear, and continuous, hence a ∗ u belongs to S (Rn ). We also

define u ∗ a := a ∗ u. If a ∈ C0∞ (Rn ), then the map in (4.1.43) is the unique
continuous extension of a ∗ u ∈ D (Rn ) to a tempered distribution. In addition,
the mapping
S(Rn ) × S (Rn ) (a, u) → a ∗ u ∈ S (Rn ) (4.1.44)
is bilinear and separately sequentially continuous. Also, for every a ∈ S(Rn ) and
u ∈ S (Rn ) we have
∂α (a ∗ u) = (∂α a) ∗ u = a ∗ (∂α u) in S (Rn ), ∀ α ∈ Nn0 . (4.1.45)
Proof. Fix u ∈ E (Rn ) and v ∈ S (Rn ). Because of (a) in Theorem 4.14 and The-
orem 2.94, it follows that u ∗ v exists as an element in D (Rn ). We will prove that
u ∗ v extends uniquely to an element in S (Rn ). Fix ψ ∈ C0∞ (Rn ) such that ψ ≡ 1 in
a neighborhood of supp u. Recall the notation introduces in (2.8.7) and let 1 denote
the constant function equal to 1 in Rn . We then claim that the map
S(Rn ) ϕ → (ψ ⊗ 1)ϕΔ ∈ S(Rn × Rn ) is linear and continuous. (4.1.46)
Indeed, if ϕ ∈ S(Rn ), then (ψ ⊗ 1)ϕΔ ∈ C ∞ (Rn × Rn ) and if we pick arbitrary

α, β, γ, δ ∈ Nn0 , then by Leibniz’s formula (cf. Proposition 14.10), the binomial the-
orem (cf. Theorem 14.9), and the compactness of the support of ψ, we may write

sup xα yβ ∂γx ψ(x)∂δy ϕ(x + y)
x∈Rn , y∈Rn
α β γ
= sup x y ∂ x ψ(x)(∂δ ϕ)(x + y)
x∈supp ψ, y∈Rn

≤C sup (y + x − x)β (∂μ ψ)(x)(∂γ−μ+δ ϕ)(x + y)
μ≤γ x∈supp ψ, y∈R
n

≤C sup (y + x)η (∂γ−μ+δ ϕ)(x + y)
μ≤γ η≤β x∈supp ψ, y∈R
n

≤C sup zη ∂γ−μ+δ ϕ(z) < ∞, (4.1.47)
n
μ≤γ η≤β z∈R
where all constants are independent of ϕ. Hence, (ψ ⊗ 1)ϕΔ belongs to S(Rn × Rn )

which proves that the map in (4.1.46) is well defined. Since this map is also linear,
it suffices to check its continuity at zero. This, in turn, follows from Fact 3.6, The-
orem 14.1, and the fact that the final estimate in (4.1.47) implies that the map in
(4.1.46) is sequentially continuous. If we now define
u∗ v : S(Rn ) −→ C,
(4.1.48)
u
∗ v (ϕ) := u(x) ⊗ v(y), ψ(x)ϕ(x + y) , ∀ ϕ ∈ S(Rn ),
then (d) in Theorem 4.14 combined with (4.1.46) gives that u ∗ v is a tempered
distribution in Rn × Rn . In addition, this definition is independent of the choice of
ψ selected as above. Indeed, if ψ1 , ψ2 ∈ C0∞ (Rn ) are such that each equals 1 in some
neighborhood of supp u then, for each ϕ ∈ S(Rn ),

u(x) ⊗ v(y) , ψ1 (x) − ψ2 (x) ϕ(x + y) = 0. (4.1.49)
To see that the map in (4.1.48) is an extension of u ∗ v ∈ D (Rn ), pick some

ϕ ∈ C0∞ (Rn ) along with a function η ∈ C0∞ (Rn ) with the property that η ≡ 1 in a
neighborhood of supp ϕ − supp u. Then the function Ψ (x, y) := ψ(x)η(y) for each
x, y ∈ Rn , belongs to C0∞ (Rn × Rn ) and is equal to 1 in a neighborhood of (supp u ×
supp v) ∩ supp ϕΔ . Upon recalling (2.8.11), this permits us to write

u(x) ⊗ v(y), ψ(x)ϕ(x + y) = u(x) ⊗ v(y), Ψ (x, y)ϕ(x + y)
= u ∗ v, ϕ . (4.1.50)
This shows that the mapping in (4.1.48) is an extension of u ∗ v ∈ D (Rn ) which,

together with (d) in Theorem 3.14, implies that u ∗ v is the unique extension of
u ∗ v ∈ D (Rn ) to a tempered distribution.
A similar construction realizes v ∗ u as a tempered distribution. Moreover, since
u ∗ v = v ∗ u in D (Rn ), from (4.1.33) we deduce that u ∗ v = v∗ u in S (Rn ). It is
also clear from (4.1.48) that the mapping
E (Rn ) × S (Rn ) (u, v) → u

∗ v ∈ S (Rn ) (4.1.51)
is bilinear. After dropping the tilde, all claims in statement (a) now follow, with
the exception of (4.1.42). Regarding the latter, we first note that the equalities in
(4.1.42) hold when interpreted in D (Rn ), thanks to part (e) in Theorem 2.96. Given
that all distributions involved are tempered, we may invoke (4.1.33) to conclude that
the named equalities also hold in S (Rn ).
D (Rn )
Moving on, let u j −−−−−→ u be such that there exists a compact set K ⊂ Rn with
j→∞
the property that supp u j ⊆ K for each j ∈ N. Then, by Exercise 2.73, we have
E (Rn )
supp u ⊆ K and u j −−−−→ u. The latter combined with (a) in Theorem 4.14 implies
j→∞
S (Rn )
u j −−−−−→ u. Fix now an arbitrary v ∈ S (Rn ). From the current part (a) it follows
j→∞
that u ∗ v ∈ S (Rn ) and u j ∗ v ∈ S (Rn ), for every j ∈ N. If ϕ ∈ S(Rn ) then definition
(4.1.48) implies that

u j ∗ v, ϕ = u j (x) ⊗ v(y), ψ(x)ϕ(x + y) , ∀ j ∈ N, (4.1.52)
where ψ ∈ C0∞ (Rn ) is a fixed function chosen so that ψ ≡ 1 in a neighborhood of K.

The proof of (4.1.46) shows that (ψ ⊗ 1)ϕΔ ∈ S(Rn × Rn ). Since the map in (4.1.17)
is separately continuous, we may write

u j ∗ v, ϕ = u j ⊗ v, (ψ ⊗ 1)ϕΔ −−−−→ u ⊗ v, (ψ ⊗ 1)ϕΔ = u ∗ v, ϕ . (4.1.53)
j→∞
This proves the statement in (b). The proof of the statement in (c) is similar. Also
the statement in (d) is an immediate consequence of part (d) in Theorem 2.96 and
(4.1.33).
Turning our attention to the statement in (e), we first note that, as noted in Ex-
ercise 3.26, the mapping S(Rn ) a → a∨ ∈ S(Rn ) is well defined, linear, and
continuous. Based on this and (f) in Theorem 3.14 we may then conclude that the
map in (4.1.43) is well defined, linear, and continuous, as a composition of linear
and continuous maps. Assume next that a ∈ C0∞ (Rn ). Then by Proposition 2.102,
for each u ∈ D (Rn ) we have a ∗ u ∈ C ∞ (Rn ) and (a ∗ u)(x) = u(y), a(x − y) for
every x ∈ Rn , hence

a ∗ u, ϕ = u(y), a(x − y) ϕ(x) dx (4.1.54)
Rn

= u(y), a(x − y)ϕ(x) dx = u, a∨ ∗ ϕ , ∀ ϕ ∈ C0∞ (Rn ).
Rn
This shows that the map in (4.1.43) is an extension of a ∗ u ∈ D (Rn ) to a tem-

pered distribution. The uniqueness of such an extension is then seen from (d) in
Theorem 3.14.
S(Rn )
Regarding the map in (4.1.44), its bilinearity is immediate. If a j −−−−→ a then
j→∞
S(Rn ) S(Rn )
a∨j −−−−→ ∨
a , so by (f) in Theorem 3.14, a∨j ∨
∗ ϕ −−−−→ a ∗ ϕ for every ϕ in S(Rn ),
j→∞ j→∞
hence
u, a∨j ∗ ϕ −−−−→ u, a∨ ∗ ϕ for each u ∈ S (Rn ),
j→∞
proving that the map in (4.1.44) is sequentially continuous in the first variable.
S (Rn )
Moreover, if u j −−−−−→ u and a ∈ S(Rn ), then
j→∞
a ∗ u j , ϕ = u j , a∨ ∗ ϕ −−−−→ u, a∨ ∗ ϕ = a ∗ u, ϕ (4.1.55)
j→∞
S (Rn )
for each ϕ ∈ S(Rn ). Thus, a ∗ u j −−−−−→ a ∗ u. This finishes the proof of the fact
j→∞
that the mapping (4.1.44) is bilinear and separately sequentially continuous. Finally,
(4.1.45) is a consequence of (4.1.42), part (d) in Theorem 3.14, the separate sequen-
tial continuity of (4.1.44), and part (c) in Theorem 4.14. This completes the proof
of the theorem.
In the last part of this section we present a density result.
Proposition 4.20. The space C0∞ (Rn ) is sequentially dense in S (Rn ). In particular,
the Schwartz class S(Rn ) is sequentially dense in S (Rn ).
Proof. Pick an arbitrary u ∈ S (Rn ). Fix a function ψ ∈ C0∞ (Rn ) such that ψ ≡ 1 on
B(0, 1) and, for each j ∈ N, define ψ j (x) := ψ(x/ j) for every x ∈ Rn . We claim that
S (Rn )
ψ j u −−−−−→ u. (4.1.56)
j→∞
Indeed, if ϕ ∈ S(Rn ) is arbitrary, then by reasoning as in the proof of (3.1.29) we

S(Rn )
deduce that ψ j ϕ −−−−→ ϕ. Hence,
j→∞
ψ j u, ϕ = u, ψ j ϕ −−−−→ u, ϕ , (4.1.57)
j→∞
from which (4.1.56) follows. In light of (4.1.56) it suffices to show that any tempered
distribution with compact support is the limit in S (Rn ) of a sequence from C0∞ (Rn ).
To this end, suppose that u ∈ S (Rn ) is compactly supported. Also, choose a function
4.2 The Fourier Transform Acting on Tempered Distributions 129
φ as in (1.2.3) and recall the sequence {φ j } j∈N ⊂ C0∞ (Rn ) from (1.3.7). Exercise 4.13
S (Rn )
then gives φ j −−−−−→ δ which, in concert with parts (d)–(e) in Theorem 4.19, implies
j→∞
S (Rn )
that φ j ∗ u −−−−−→ u. At this stage there remains to observe that φ j ∗ u ∈ C0∞ (Rn ) for
j→∞
every j ∈ N, thanks to (2.8.49).
4.2 The Fourier Transform Acting on Tempered Distributions
The goal here is to extend the action of the Fourier transform, considered earlier on
the Schwartz class of rapidly decreasing functions, to the space of tempered distri-
butions. To get some understanding of how this can be done in a natural fashion, we
begin by noting that if f ∈ L1 (Rn ) is arbitrary then (3.2.27) gives

f (ξ)ϕ(ξ) dξ = ϕ(x) dx,
f (x) ∀ ϕ ∈ S(Rn ). (4.2.1)
Rn Rn
Since f ∈ L1 (Rn ) and

f ∈ L∞ (Rn ), based on (4.1.9) we have f, f ∈ S (Rn ). Hence,
(4.2.1) may be rewritten as
f , ϕ = f,
ϕ , for every ϕ ∈ S(R ). This suggests the
n
definition made below for the Fourier transform of a tempered distribution.
Proposition 4.21. Let u ∈ S (Rn ). Then the mapping

u : S(Rn ) → C,
u(ϕ) := u,
ϕ , ∀ ϕ ∈ S(Rn ), (4.2.2)
u ∈ S (Rn ).
is well defined, linear, and continuous. Hence,
Proof. This is an immediate consequence of the fact that u ∈ S (Rn ), part (c) in
Theorem 3.21, and the identity
u = u ◦ F , where F is the Fourier transform on
S(Rn ).
Example 4.22. Since δ ∈ S (Rn ) we may write

δ, ϕ = δ,
ϕ =
ϕ(0) = ϕ(x) dx = 1, ϕ , ∀ ϕ ∈ S(Rn ), (4.2.3)
Rn
thus

δ=1 in S (Rn ). (4.2.4)
Also, if x0 ∈ R , then for each ϕ ∈ S(R ) we may write
n n

δ x0 , ϕ = δ x0 ,
ϕ =
ϕ(x0 ) = e−ix0 ·x ϕ(x) dx = e−ix0 ·x , ϕ , (4.2.5)
Rn
proving
δ
x0 = e
−ix0 ·x
in S (Rn ). (4.2.6)
Remark 4.23. The map S (Rn ) u → u ∈ S (Rn ) (that is further analyzed in part
(a) of Theorem 4.26) is an extension of the map in (3.1.3). Indeed, if f ∈ L1 (Rn )
then the map u f from (4.1.5) is a tempered distribution and we may write

uf , ϕ = u f ,
ϕ = ϕ(x) dx =
f (x) f (ξ)ϕ(ξ) dξ
Rn Rn
=
f , ϕ ∀ ϕ ∈ S(Rn ), (4.2.7)
where
f is as in (3.1.1) and for the third equality we used (3.2.27). This shows that
uf = u f in S (Rn ), ∀ f ∈ L1 (Rn ). (4.2.8)
In particular, (4.2.8) also holds if f ∈ S(Rn ) (since S(Rn ) ⊂ L1 (Rn ) by Exercise 3.5).
Example 4.24. Suppose a ∈ (0, ∞) and consider the function f (x) := x2 +a

1
2 for every
x ∈ R. Since f ∈ L (R), by (4.1.9) we may regard f as a tempered distribution. The

1
goal is to compute f in S (R). From (4.2.8) we know that this is the same as the
Fourier transform of f as a function in L1 (R), that is

e−ixξ
f (ξ) = dx, ∀ ξ ∈ R. (4.2.9)
R x +a
2 2
To compute the integral in (4.2.9) we use the calculus of residues applied to the
function
e−izξ
g(z) := 2 for z ∈ C \ {ia, −ia}. (4.2.10)
z + a2
In this regard, denote the residue of g at z by Resg (z). We separate the computation
in two cases.
Case 1. Assume ξ ∈ (−∞, 0). For each R ∈ (a, ∞) consider the contour Γ :=
Γ1 ∪ Γ2 , where

Γ1 := t : −R ≤ t ≤ R , (4.2.11)

Γ2 := Reiθ : 0 ≤ θ ≤ π . (4.2.12)
The function g has one pole z = ia enclosed by Γ and the residue theorem gives

g(z) dz + g(z) dz = 2πi Resg (ia). (4.2.13)
Γ1 Γ2
Let us analyze each term in (4.2.13). First,
e−izξ eaξ e−a|ξ|

Resg (ia) = z=ia = = . (4.2.14)
z + ia 2ia 2ia
−itξ
Second, given that for each fixed ξ ∈ R the function R t → te2 +a2 ∈ C is absolutely
integrable, Lebesgue’s Dominated Convergence Theorem gives

R
e−itξ e−itξ
g(z) dz = dt −−−−→ dt =
f (ξ). (4.2.15)
Γ1 −R t + a2
2 R→∞ R t2+ a2
Third,
π ξR sin θ−iξR cos θ
e C
g(z) dz = iRe dθ ≤ −−−−→ 0,
iθ
(4.2.16)
Γ2 0 R e +a
2 i2θ 2 R R→∞
since eξR sin θ ≤ 1 for θ ∈ [0, π] (recall that we are assuming ξ < 0). Hence, letting
R → ∞ in (4.2.13) and making use of (4.2.14), (4.2.15), and (4.2.16), we arrive at
e−a|ξ| π −a|ξ|
f (ξ) = 2πi = e whenever ξ ∈ (0, ∞). (4.2.17)
2ia a
Case 2. Assume ξ ∈ (0, ∞). This time we consider a contour that encloses the
pole z = −ia. Specifically, we keep Γ1 as before (with R ∈ (a, ∞)) and set

Γ3 := Reiθ : π ≤ θ ≤ 2π . (4.2.18)
Applying the residue theorem for the contour Γ1 ∪ Γ3 we then obtain

e−izξ e−a|ξ|
− g(z) dz + g(z) dz = 2πi Resg (−ia) = z=−ia = − . (4.2.19)
Γ1 Γ3 z − ia 2ia
ξR sin θ
A computation similar to that in (4.2.16)
(note that e ≤ 1 for each ξ > 0 and

each θ ∈ [π, 2π]) gives lim Γ g(z) dz = 0, which when used in (4.2.19) yields
R→∞ 2

f (ξ) = πa e−a|ξ| for ξ ∈ (0, ∞).
Case 3. Assume ξ = 0. It is immediate from (4.2.9) that

1 π
f (0) = dx = . (4.2.20)
R x +a
2 2 a
In summary, we have proved

1 π
F (ξ) = e−a|ξ| for every ξ ∈ R, as well as in S (R). (4.2.21)
x2 + a2 a
In the next example we compute the Fourier transform of a tempered distribution
induced by a slowly growing function that is not absolutely integrable in Rn .
Example 4.25. Let a ∈ (0, ∞) and consider the function f (x) := e−ia|x| for each
2
x ∈ Rn . It is not difficult to check that f ∈ L(Rn ), thus by Exercise 4.7 the func-
tion f may be regarded as a tempered distribution (identifying it with u f ∈ S (Rn )
associated with f as in (4.1.5)). We claim that
π n2
e
|ξ|2
−ia|x|2 = e i 4a in S (Rn ). (4.2.22)
ia
To prove (4.2.22), fix ϕ ∈ S(Rn ) and starting with the definition of the Fourier
transform on S (Rn ) write

e
−ia|x|2 , ϕ = e−ia|x| , e−ia|ξ|
2 2
ϕ = ϕ(ξ) dξ
Rn

2
e−(ia+ε)|ξ| ϕ , e−(ia+ε)|ξ|
2
= lim+ ϕ(ξ) dξ = lim+
ε→0 Rn ε→0
2
|ξ|2
π 2 − 4(ia+ε)
n
= lim+ ϕ , F e−(ia+ε)|x| = lim+ ϕ , ia+ε e
ε→0 ε→0
π n2 |ξ|2
= lim+ e− 4(ia+ε) ϕ(ξ) dξ
ε→0 ia + ε Rn
π n2 |ξ|2
= e− 4ia ϕ(ξ) dξ
ia Rn

π 2 |ξ|2
n

= e i 4a , ϕ . (4.2.23)
ia
Above, the first equality is based on (4.2.2), while the second equality is a conse-
quence of the way in which the slowly growing function e−ia|x| is regarded as a tem-
2
pered distribution. The third and second-to-last equalities are based on Lebesgue’s
Dominated Convergence Theorem. In the fourth equality we interpret the Schwartz
ϕ as being in S (Rn ) and rely on the fact that e−(ia+ε)|x| ∈ S(Rn ) for each
2
function
ε > 0 (as noted in Example 3.22). The fifth equality once uses Remark 4.23 and
once again (4.2.2). The sixth equality is a consequence of formula (3.2.6), while the
rest is routine.

(a) The mapping F : S (Rn ) → S (Rn ) defined by F (u) := u, for every u in S (Rn ),
is bijective, continuous, and its inverse is also continuous.
(b) D α u = ξα
u for all α ∈ Nn0 and all u ∈ S (Rn ).

(c) x u = (−D)α
α u for all α ∈ Nn0 and all u ∈ S (Rn ).
(d) If u ∈ S (R ) and v ∈ S (Rn ), then u
m
⊗v = u ⊗v.
Proof. Recall from Theorem 3.25 that the map F : S(Rn ) → S(Rn ) is linear, con-
tinuous, bijective, and its inverse is also continuous. Since the transpose of this map
(in the sense of (14.1.10)) is precisely the Fourier transform in the context of part
(a) of the current statement, from Propositions 14.2–14.3 it follows that the mapping
F : S (Rn ) → S (Rn ) is also well defined, linear, continuous, bijective, and has a
continuous inverse.
Consider next u ∈ S (Rn ) and α ∈ Nn0 . Then for every ϕ ∈ S(Rn ), using part (c)
in Theorem 4.14, part (a) in Theorem 3.21, (4.2.2), and part (b) in Theorem 4.14,
we may write

Dα u, ϕ = Dα u, ϕ = u, ξ
ϕ = (−1)|α| u, Dα αϕ
u, ξα ϕ = ξα
= u, ϕ , (4.2.24)
and

xα u, ϕ = xα u,
ϕ = u, xα
ϕ = u, Dαϕ
u, Dα ϕ = (−D)α
= u, ϕ . (4.2.25)
This proves the claims in (b)–(c).

We are left with the proof of the statement in (d). Fix u ∈ S (Rm ) and some
v ∈ S (Rn ). Then based on statement (d) in Theorem 4.14 we have that u ⊗ v belongs
to S (Rm × Rn ). Hence, starting with (4.2.2), then using (d) in Theorem 3.21 and
(4.1.16), we may write
⊗ v, ϕ ⊗ ψ = u ⊗ v, ϕ
u ϕ⊗
⊗ ψ = u ⊗ v, ψ (4.2.26)
ϕ v,
= u, ψ =
u, ϕ
v, ψ
= u ⊗ v, ϕ ⊗ ψ , ∀ ϕ ∈ S(Rm ), ∀ ψ ∈ S(Rn ).

Consequently, u⊗ vC ∞ (Rm )⊗C ∞ (Rn ) = vC ∞ (Rm )⊗C ∞ (Rn ) , which in combination with
u ⊗
0 0 0 0
Proposition 3.16 proves the statement in (d).
Exercise 4.27. Recall (4.1.5) and (4.1.9). Prove that uf = u f in S (Rn ) for each
f ∈ L2 (Rn ) and each f ∈ L1 (Rn ).
Hint: Use (3.2.27) and Exercise 3.30.
Lemma 4.28 (Riemann–Lebesgue Lemma). If f ∈ L1 (Rn ), then the tempered

distribution u f satisfies uf ∈ C 0 (Rn ).
Proof This is a consequence of Exercise 4.27 and (3.1.3).
Proposition 4.29. Prove that for every f, g ∈ L2 (Rn ) one has f

∗g= g in S (Rn ).
f ·
Proof. Let f, g ∈ L2 (Rn ) be arbitrary. Using Young’s inequality (cf. Theorem 14.17)
we see that f ∗ g ∈ L∞ (Rn ), hence f ∗ g ∈ S (Rn ) (recall Example 4.8). Also,
from (3.2.29) we have g ∈ L2 (Rn ), which further implies that
f, f ·
g belongs to

L (R ) ⊆ S (R ). This shows that the equality f ∗ g = f ·
1 n n g is meaningful in S (Rn ).
The proof is now based on a density argument. First, since C0∞ (Rn ) is dense in
L (Rn ), from (3.1.14) and Exercise 3.9 it follows that S(Rn ) is dense in L2 (Rn ).
2
Consequently there exist sequences { f j } j∈N and {g j } j∈N in S(Rn ) such that
L2 (Rn ) L2 (Rn )
f j −−−−−→ f and g j −−−−−→ g. (4.2.27)
j→∞ j→∞
Invoking (3.2.29) we obtain

2 n
L (R ) L2 (Rn )

f j −−−−−→ f and gj −−−−−→
g. (4.2.28)
j→∞ j→∞
Then, for each j ∈ N, using Hölder’s inequality we write

f j · gj − g L1 (Rn ) ≤
f · g ) L1 (Rn ) + (
gj −
f j ( fj − g L1 (Rn )
f )
≤ gj −
f j L2 (Rn ) g L2 (Rn )
g L2 (Rn )
+ fj −
f L2 (Rn ) , (4.2.29)
2 n
L (R )
which when combined with (4.2.28) implies
f j · gj −−−−−→ f ·
g. The latter and
j→∞
Exercise 4.12 further yield

S (R )
n
f j · gj −−−−−→
f ·
g. (4.2.30)
j→∞
On the other hand, from (4.2.27) and Young’s inequality (cf. (14.2.16) applied with
L∞ (Rn ) S (Rn )
p = p = 2) we have f j ∗ g j −−−−−→ f ∗ g, hence f j ∗ g j −−−−−→ f ∗ g according to
j→∞ j→∞
Exercise 4.12. Since the Fourier transform is continuous on S (Rn ), it follows that

S (R )n
f j ∗ g j −−−−−→ f
∗ g. (4.2.31)
j→∞
The desired conclusion follows from (4.2.30), (4.2.31), and the formula fj ∗ gj =
f j · gj in S(Rn ) (see part (c) of Proposition 3.28).

Exercise 4.30. Let θ ∈ S(Rn ) be such that Rn θ(x) dx = 1. For each ε > 0 set
θε (x) := ε−n θ(x/ε) for every x ∈ Rn . Also, let f ∈ L2 (Rn ) and for each ε > 0 define
the function fε := f ∗ θε , which by Young’s inequality belongs to L2 (Rn ). Prove that
L2 (Rn )
fε −−−−−→ f .
j→∞
Hint: Use Plancherel’s identity (cf. the identity in (3.2.29)), Proposition 4.29, and
the fact that for each ε > 0 we have θε ∈ L∞ (Rn ) and θε (ξ) =
θ(εξ) for every ξ ∈ Rn .
Example 4.31. Let a ∈ (0, ∞). We are interested in computing the Fourier transform

2 , viewed as a distribution in S (R). In this vein, con-
x
of the bounded function x2 +a
2 for x ∈ R, and recall from (4.2.21) that
1
sider the auxiliary function f (x) := x2 +a
π
f (ξ) = a e−a|ξ|
in S (R). This formula, part (c) in Theorem 4.26, and (2.4.11), allow
us to write
x d d π −a|ξ|
F (ξ) = F x f (ξ) = i f (ξ) = i e
x2 + a2 dξ dξ a
πi
= − a e−aξ H(ξ) + a eaξ H(−ξ)
a
= −πi(sgn ξ)e−a|ξ| in S (R). (4.2.32)
Proposition 4.32. For each u ∈ S (Rn ) define the mapping
u∨ : S(Rn ) → C, u∨ (ϕ) := u, ϕ∨ , ∀ ϕ ∈ S(Rn ). (4.2.33)
Then u∨ ∈ S (Rn ) and

u = (2π)n u∨ .
(4.2.34)
Proof. Fix u ∈ S (Rn ). Then (4.2.33) is simply the composition of u with the
mapping from Exercise 3.26. Since both are linear and continuous, it follows that
u∨ ∈ S (Rn ). Formula (4.2.34) then follows by combining (4.2.2) with the identity
in part (3) of Exercise 3.27.
Exercise 4.33. Show that

1 = (2π)n δ in S (Rn ). (4.2.35)
Exercise 4.34. Prove that for any tempered distribution u the following equivalence
holds:
u∨ = −u in S (Rn ) ⇐⇒ ( u )∨ = −u in S (Rn ). (4.2.36)
One suggestive way of summarizing (4.2.36) is to say that a tempered distribution
is odd if and only if its Fourier transform is odd.
(a) If a ∈ S(Rn ) and u ∈ S (Rn ), then u
∗a =au in S (Rn ), where
a is viewed as an
element from S(R ).n
(b) If u ∈ E (Rn ) then the tempered distribution u is of function type given by the

formula u(ξ) = u(x), e−ix·ξ for every ξ ∈ Rn , and
u ∈ L(Rn ).

(c) If u ∈ E (R ), v ∈ S (R ) then u
n n
∗v =
uv, where u is viewed as an element in
L(Rn ).
Proof. By (d) in Theorem 4.19 we have u ∗ a ∈ S (Rn ), hence u ∗ a exists and
belongs to S (Rn ). Also, since a ∈ S(Rn ), by (3.1.24) and (b) in Theorem 4.14 it
follows that
au ∈ S (Rn ). Then, we may write

u
∗ a, ϕ = u ∗ a, ϕ = (2π)−n u,
ϕ = u, a∨ ∗
a∗ϕ

= u,
aϕ = aϕ =
u, au, ϕ , ∀ ϕ ∈ S(Rn ). (4.2.37)
For the second equality in (4.2.37) we used (4.1.43), for the third we used part (3)
in Exercise 3.27, while for the fourth we used (d) in Proposition 3.28. This proves
the statement in (a).
Moving on to the proof of (b), fix some u ∈ E (Rn ) and introduce the function

f (ξ) := u(x), e−ix·ξ for every ξ ∈ Rn . From Proposition 2.91 it follows that f ∈
C ∞ (Rn ) and

∂α f (ξ) = u(x) , ∂αξ [e−ix·ξ ] , ∀ ξ ∈ Rn , for every α ∈ Nn0 . (4.2.38)
In addition, since u ∈ E (Rn ), there exist a compact subset K of Rn , along with a

constant C ∈ (0, ∞), and a number k ∈ N0 , such that u satisfies (2.6.6). Combining
all these facts, for each α ∈ Nn0 , we may estimate
α
∂ f (ξ) = u(x), ∂αξ e−ix·ξ ≤ C sup ∂βx ∂αξ e−ix·ξ
x∈K, |β|≤k

≤ C sup xα ξβ ≤ C(1 + |ξ|)k . (4.2.39)
x∈K, |β|≤k
From (4.2.39) and the fact that f is smooth we deduce that f ∈ L(Rn ). Hence, if we
now recall Exercise 4.7, it follows that f ∈ S (Rn ).
We are left with proving that u = f as tempered distributions. To this end, fix
θ ∈ C0∞ (Rn ) such that θ ≡ 1 in a neighborhood of supp u. Then, for every ϕ ∈ C0∞ (Rn )
one has

u, ϕ = u,
ϕ = u(ξ) , θ(ξ) ϕ(ξ) = u(ξ), e−ix·ξ θ(ξ)ϕ(x) dx . (4.2.40)
Rn
At this point recall Remark 2.88 (note that the function e−ix·ξ θ(ξ)ϕ(x) belongs to
C0∞ (Rn × Rn ) and one may take v = 1 in (2.7.53)) which allows one to rewrite the
last term in (4.2.40) and conclude that

u, ϕ = u(ξ), e−ix·ξ ϕ(x) dx = f, ϕ , ∀ ϕ ∈ C0∞ (Rn ). (4.2.41)
Rn
Hence the tempered distributions u and f coincide on C0∞ (Rn ). By (4.1.33), we

therefore have u = f in S (R ), completing the proof of the statement in (b).
n
Regarding the formula in part (c), while informally this is similar to the formula
proved in part (a), the computation in (4.2.37) through which the latter has been
deduced no longer works in the current case, as various ingredients used to justify it
break down (given that u is now only known to belong to L(Rn ) and not necessarily
to S(R )). This being said, we may employ what has been established in part (a)
n
together with a limiting argument to get the desired result.

Specifically, assume that u ∈ E (Rn ) and v ∈ S (Rn ). Then u ∗ v is a tempered
distribution in Rn (recall item (a) in Theorem 4.19) hence, u ∗ v is well defined in
S (Rn ). Also, from (b) we have u ∈ L(Rn ), thus uv is meaningful and belongs to
S (Rn ). To proceed, recall the sequence {φ j } j∈N from Example 1.11. In particular,
D (Rn )
(1.3.8) holds and φ j −−−−−→ δ (see Example 2.24). Thus, the statement from part (b)
j→∞
S (Rn )
in Theorem 4.19 applies and gives that u ∗ φ j −−−−−→ u ∗ δ = u. Moreover, since
j→∞
u ∈ E (Rn ), it follows that u ∗ φ j ∈ E (Rn ) (by statement (a) in Theorem 2.96) and
supp (u ∗ φ j ) ⊆ supp u + B(0, 1) for every j ∈ N (by part (1) in Remark 2.100).
Hence, one may apply (b) in Theorem 4.19 to further conclude that the convergence
S (Rn )
(u ∗ φ j ) ∗ v −−−−−→ u ∗ v holds. Given (a) in Theorem 4.26, the latter implies (recall
j→∞
Fact 4.11)

lim F (u ∗ φ j ) ∗ v , ϕ = u
∗ v, ϕ , ∀ ϕ ∈ S(Rn ). (4.2.42)
j→∞
Note that (2.8.49) gives u ∗ φ j ∈ C0∞ (Rn ) for every j ∈ N. Hence, based on what we
have proved already in part (a), we obtain (keeping in mind that u ∈ L(Rn ))

F (u ∗ φ j ) ∗ v = u v = φj
∗ φj uv= u
φ(·/ j) v in S (Rn ), ∀ j ∈ N. (4.2.43)
In concert, (4.2.42) and (4.2.43) yield that for each ϕ ∈ S(Rn )

u
∗ v, ϕ = lim
uv, φ(·/ j) ϕ =
uv, φ(0) ϕ = u
v, ϕ (4.2.44)
j→∞
S(Rn )
φ(·/ j) ϕ −−−−→
since φ(0) ϕ by Exercise 3.15, and
φ(0) = 1. This proves the state-
j→∞
ment in (c) and finishes the proof of the theorem.
Example 4.36. If a ∈ (0, ∞) then χ[−a,a] , the characteristic function of the interval
[−a, a], belongs to E (R) and by statement (b) in Theorem 4.35 we have
⎧ sin(aξ)
a ⎪
⎪
−ixξ ⎨ 2 ξ for ξ ∈ R \ {0},
⎪
χ
(ξ) = e dx = ⎪
⎪ (4.2.45)
[−a,a]
−a ⎪
⎩ 2a for ξ = 0.
Exercise 4.37. Use Exercise 2.75 and statement (b) in Theorem 4.35 to prove that
if u ∈ S (Rn ) is such that supp
u ⊆ {a} for some a ∈ Rn , then there exist k ∈ N0 and
constants cα ∈ C, for α ∈ N0 with |α| ≤ k, such that
n

u= cα xα eix·a in S (Rn ). (4.2.46)
|α|≤k
In particular, if a = 0 then u is a polynomial in Rn .
Example 4.38. Let a ∈ R and consider the function f (x) := sin(ax) for x ∈ R.
Then f ∈ C ∞ (R) ∩ L∞ (R), hence f ∈ S (R). Suppose a 0. We shall compute
the Fourier transform of f in S (R) by making use of a technique relying on the
ordinary differential equation f satisfies. More precisely, since f + a2 f = 0 in R,
the same equation holds in S (R), thus by (a) and (b) in Theorem 4.26 we have
(ξ2 − a2 )
f = 0 in S (R). By Example 2.80 this implies
f = C1 δa + C2 δ−a in S (R)
for some C1 , C2 ∈ C. Applying again the Fourier transform to the last equation,
using (4.2.34) and part (b) from Theorem 4.35, we have

= C1 δa + C2 δ
2π sin(−ax) = sin(ax) −a
= C1 δa (ξ), e−ixξ + C2 δ−a (ξ), e−ixξ

= C1 e−iax + C2 eiax
= (C1 + C2 ) cos(ax) + i(C1 − C2 ) sin(−ax). (4.2.47)
The resulting identity in (4.2.47) forces C1 = −iπ and C2 = iπ. Plugging these
constants back in the expression for
f yields
= −iπδa + iπδ−a
sin(ax) in S (R). (4.2.48)
It is immediate that (4.2.48) also holds if a = 0.
Example 4.39. Let a, b ∈ R. Then the function g(x) := sin(ax) sin(b x) for x ∈ R
satisfies g ∈ L∞ (R), thus g ∈ S (R) (cf. (4.1.9)). Applying the Fourier transform to
the identity in (4.2.48) and using (4.2.34) we obtain
i i
sin(ax) = δa − δ−a in S (R). (4.2.49)
2 2
Also, making use of (4.2.49), part (c) in Theorem 4.35, and then Exercise 2.99, we
may write
i i i i
sin(ax) sin(bx) = δa − δ
−a δb − δ
−b
2 2 2 2
1
= − F δa+b − δa−b − δb−a + δ−a−b in S (R). (4.2.50)
4
Hence, another application of the Fourier transform gives (relying on (4.2.34))
π
F sin(ax) sin(bx) = − δa+b − δa−b − δb−a + δ−a−b in S (R). (4.2.51)
2
2
Example 4.40. Let a ∈ (0, ∞) and consider the function f (x) := e ia|x| for x ∈ Rn .
Then f ∈ L∞ (Rn ) thus f ∈ S (Rn ) by (4.1.9). The goal is to compute the Fourier
transform of f in S (Rn ). The starting point is the observation that
2 2 2
f (x) = e iax1 ⊗ e iax2 ⊗ · · · ⊗ e iaxn ∀ x = (x1 , . . . , xn ) ∈ R. (4.2.52)
Invoking part (d) in Theorem 4.26 then reduces matters to computing the Fourier
transform of f in the case n = 1.
2
Assume that n = 1, in which case f (x) = e iax for x ∈ R. Then f satisfies the
differential equation: f −2ia x f = 0 in S (R). Taking the Fourier transform in S (R)
and using the formulas from (b)–(c) in Theorem 4.26 we obtain
i
(
f ) + ξ f = 0 in S (R). (4.2.53)
2a
The format of (4.2.53) suggests that we consider the ordinary differential equation
y + 2ai ξy = 0 in R, in the unknown y = y(ξ). One particular solution of this o.d.e. is
ξ2
y(ξ) = e−i 4a . Note that both y and 1/y belong to L(R). In particular, it makes sense
to consider the tempered distribution u := (1/y) f , whose derivative is
u = (1/y) f ) = −(y /y2 )

f + (1/y)( f + (1/y)(
f )
i i
= ξ /y
f − ξ(1/y)
f = 0 in S (R). (4.2.54)
2a 2a
From this and part (2) in Proposition 2.47 we then deduce that there exists some
constant c ∈ C such that u = c in S (R) which, given the significance of u and y,
ξ2
forces
f = c e−i 4a in S (R) for some c ∈ C. We are therefore left with determining
the constant c. This may be done by choosing a suitable Schwartz function and
ξ2
then computing the action of
√ f on it. Take ϕ(ξ) := e− 4a for ξ ∈ R. Since
ϕ(x) =
4aπe−ax for x ∈ R (recall Example 3.22), we may write
2
√
ξ2 ξ2
e−i 4a − 4a dξ = e iax −ax dx.
2 2
c f , ϕ = f,
ϕ = 4aπ (4.2.55)
R R
The two integrals in (4.2.55) may be computed by applying formula (3.2.7) with
4a and λ := a(1 − i), respectively. After some routine algebra (i.e., computing
λ := 1+i
these$integrals, replacing their values in (4.2.55), then solving for c), we find that
π
c = πa e i 4 . In summary, this analysis proves that
n nπ |ξ|
e
2
ia|x|2 (ξ) = π 2 e i 4 e−i 4a in S (Rn ). (4.2.56)
a
Partial Fourier Transforms

In the last part of this section we define partial Fourier transforms. To set the stage,
fix m, n ∈ N. We shall denote by x, ξ generic variables in Rn , and by y, η generic
variables in Rm . The partial Fourier transform with respect to the variable x of a
function ϕ ∈ S(Rn+m ), denoted by ϕ x or F x ϕ, is defined by

ϕ x (ξ, y) := e−ix·ξ ϕ(x, y) dx, ∀ ξ ∈ Rn , ∀ y ∈ Rm . (4.2.57)
Rn
Reasoning in a similar manner as in the proof of Theorem 3.25, it follows that
F x : S(Rn+m ) → S(Rn+m ) is bijective,

(4.2.58)
continuous, with continuous inverse
and its inverse is given by


(F x−1 ψ)(x, η) := (2π)−n eix·ξ ψ(ξ, η) dξ,
Rn (4.2.59)
for all (x, η) ∈ R n+m
and all ψ ∈ S(R n+m
).
Furthermore, analogously to Proposition 4.21, the partial Fourier transform F x

extends to S (Rn+m ) as a continuous map by setting
F x u, ϕ := u, F x ϕ , ∀ u ∈ S (Rn+m ), ∀ ϕ ∈ S(Rn+m ), (4.2.60)
and this extension is an isomorphism from S (Rn+m ) into itself, with continuous
inverse denoted by F x−1 . Moreover, the action of F x enjoys properties analogous to
those established for the “full” Fourier transform in Theorem 3.21, Exercise 3.27,
Theorem 4.26, and Proposition 4.32.
Exercise 4.41. Let · denote the full Fourier transform in Rn+m . Prove that for each
function ϕ ∈ S(Rn × Rm ) and each (ξ, η) ∈ Rn × Rm we have

F x Fy ϕ(x, y) (ξ, η) = Fy F x ϕ(x, y) (ξ, η) =
ϕ(ξ, η). (4.2.61)
Also, show that
u in S (Rn+m ),
F x Fy u = Fy F x u = ∀ u ∈ S (Rn × Rm ). (4.2.62)
Exercise 4.42. Prove that

F x δ(x) ⊗ δ(y) = 1(ξ) ⊗ δ(y) in S (Rn+m ). (4.2.63)
4.3 Homogeneous Distributions
Let A ∈ Mn×n (R) be such that det A 0. Then for every f ∈ L1 (Rn ) one has
f ◦ A ∈ L1 (Rn ), thus f, f ◦ A ∈ S (Rn ) by (4.1.9). Moreover,

f ◦ A, ϕ = f (Ax)ϕ(x) dx = | det A|−1 f (y)ϕ(A−1 y) dy
Rn Rn
= | det A|−1 f, ϕ ◦ A−1 , ∀ ϕ ∈ S(Rn ). (4.3.1)
The resulting identity in (4.3.1) and Exercise 3.17 justifies the following extension
(compare with Proposition 2.27).
Proposition 4.43. Let A ∈ Mn×n (R) be such that det A 0. For each u in S (Rn ),
define the mapping u ◦ A : S(Rn ) → C by setting

u ◦ A (ϕ) := | det A|−1 u, ϕ ◦ A−1 , ∀ ϕ ∈ S(Rn ). (4.3.2)
Then u ◦ A ∈ S (Rn ).
4.3 Homogeneous Distributions 141
Proof. This is an immediate consequence of (3.1.41).

The identities in Exercise 2.28 have natural analogues in the current setting.
Exercise 4.44. Let A, B ∈ Mn×n (R) be such that det A 0 and det B 0. Then the
following identities hold in S (Rn ):
(1) (u ◦ A) ◦ B = u ◦ (AB) for every u ∈ S (Rn );
(2) u ◦ (λA) = λu ◦ A for every u ∈ S (Rn ) and every λ ∈ R;
(3) (u + v) ◦ A = u ◦ A + v ◦ A for every u, v ∈ S (Rn ).
In the next proposition we study how the Fourier transform interacts with the
operator of composition by an invertible matrix. Recall that the transpose of a matrix
A is denoted by A .
Proposition 4.45. Assume that A ∈ Mn×n (R) is such that det A 0. Then for each
u ∈ S (Rn ),

u u ◦ A −1 .
◦ A = | det A|−1 (4.3.3)
Proof. For each ϕ ∈ S(Rn ), based on (4.2.2), (4.3.2), and (3.2.9), we may write

u ϕ = | det A|−1 u,
◦ A, ϕ = u ◦ A, ϕ ◦ A−1

= u, ϕ
◦ A = u, ϕ ◦ A
−1
= | det A|−1
u◦ A ,ϕ . (4.3.4)
This proves (4.3.3).

The mappings in (3.1.41) and (4.3.2) corresponding to A := tIn×n , for some num-
ber t ∈ (0, ∞), are called dilations and will be denoted by τt . More precisely,
for each t ∈ (0, ∞) we have
τt : S(Rn ) → S(Rn ), (τt ϕ)(x) := ϕ(tx), ∀ ϕ ∈ S(Rn ), ∀ x ∈ Rn , (4.3.5)
and
τt : S (Rn ) → S (Rn ),
(4.3.6)
τt u, ϕ := t−n u, τ 1t ϕ , ∀ u ∈ S (Rn ), ∀ ϕ ∈ S(Rn ).
Exercise 4.46. Prove that for each t ∈ (0, ∞) the following are true:
F (τt ϕ) = t−n τ 1t F (ϕ) in S(Rn ), ∀ ϕ ∈ S(Rn ), (4.3.7)
F (τt u) = t−n τ 1t F (u) in S (Rn ), ∀ u ∈ S (Rn ). (4.3.8)
Hint: Use (3.2.9) with A = 1

t In×n to prove (4.3.7), then use (4.3.7) and (4.3.6) to
prove (4.3.8).
To proceed, we make a couple of definitions.

Definition 4.47. A linear transformation A ∈ Mn×n (R) is called orthogonal pro-

vided A is invertible and A−1 = A .
Some of the most basic attributes of an orthogonal matrix A are
(A )−1 = A, |det A| = 1, |Ax| = |x| for every x ∈ Rn . (4.3.9)
Definition 4.48. A tempered distribution u ∈ S (Rn ) is called invariant under

orthogonal transformations provided u = u ◦ A in S (Rn ) for every
orthogonal matrix A ∈ Mn×n (R).
Proposition 4.49. Let u ∈ S (Rn ). Then u is invariant under orthogonal transfor-

mations if and only if
u is invariant under orthogonal transformations.
Proof. This is a direct consequence of (4.3.3) and the fact that any orthogonal matrix
A satisfies (4.3.9).
Next we take a look at homogeneous functions to gain some insight into how this
notion may be defined in the setting of distributions.
Definition 4.50. (1) A nonempty open set O in Rn is called a cone-like region

if tx ∈ O whenever x ∈ O and t ∈ (0, ∞).
(2) Given a cone-like region O ⊆ Rn , call a function f : O → C positive
homogeneous of degree k ∈ R if f (tx) = tk f (x) for every t > 0 and every
x ∈ O.
Exercise 4.51. Prove that if O ⊆ Rn is a cone-like region, N ∈ N, and f ∈ C N (O) is

positive homogeneous of degree k ∈ R on O, then ∂α f is positive homogeneous of
degree k − |α| on O for every α ∈ Nn0 with |α| ≤ N.
Exercise 4.52. Prove that if f ∈ C 0 (Rn \ {0}) is positive homogeneous of degree

1 − n on Rn \ {0}, and g ∈ C 0 (S n−1 ), then

g(x/R) f (x) dσ(x) = g(x) f (x) dσ(x), ∀ R ∈ (0, ∞). (4.3.10)
∂B(0,R) S n−1
Exercise 4.53. Prove that if f ∈ C 0 (Rn \ {0}) is positive homogeneous of degree

k ∈ R, then | f (x)| ≤ f L∞ (S n−1 ) |x|k for every x ∈ Rn \ {0}.

Hint: Write f (x) = f |x|x |x|k for every x ∈ Rn \ {0}.
Exercise 4.54. Show that if f ∈ C 0 (Rn \ {0}) is positive homogeneous of degree

k ∈ R with k > −n, then f ∈ S (Rn ).
Hint: Make use of Exercise 4.53 and the result discussed in Example 4.4.
After this preamble, we are ready to extend the notion of positive homogeneity
to tempered distributions.
Definition 4.55. A distribution u ∈ S (Rn ) is called positive homogeneous

of degree k ∈ R provided τt u = tk u in S (Rn ) for every t > 0.
Exercise 4.56. Prove that δ ∈ S (Rn ) is positive homogeneous of degree −n.
Exercise 4.57. Let f ∈ Lloc

1
(Rn ) be such that the integrability condition (4.1.4) is
satisfied for some m ∈ [0, ∞) and some R ∈ (0, ∞), and let k ∈ R. Show that the
tempered distribution u f is positive homogeneous of degree k, if and only if f is
positive homogeneous of degree k.
Exercise 4.58. Prove that if u ∈ S (Rn ) is positive homogeneous of degree k ∈ R

then for every α ∈ Nn0 the tempered distribution ∂α u is positive homogeneous of
degree k − |α|. Deduce from this that for every α ∈ Nn0 the tempered distribution ∂α δ
is positive homogeneous of degree −n − |α|.
Proposition 4.59. Let k ∈ R. If u ∈ S (Rn ) is positive homogeneous of degree k,

then
u is positive homogeneous of degree −n − k.
Proof. Let u ∈ S (Rn ) be positive homogeneous of degree k, and fix t > 0. Then
(4.3.8) and the assumption on u give

u = t−n F τ 1t u = t−n F t−k u = t−n−k
τt u in S (Rn ), (4.3.11)
hence
u is positive homogeneous of degree −n − k.

Proposition 4.60. If u ∈ S (Rn ), uRn \{0} ∈ C ∞ (Rn \{0}), and uRn \{0} is positive homo-

geneous of degree k, for some k ∈ R, then uRn \{0} ∈ C ∞ (Rn \ {0}).
Proof. Fix u satisfying the hypotheses of the proposition. By (c) in Proposition 4.26,
for each α ∈ Nn0 one has Dαξ
u = (−x) α u in S (Rn ). Also, it is not difficult to check
α
that (−x) u ∈ S (R ) continues to satisfy the hypotheses of the proposition with
n
k
replaced by k + |α|. Hence, the desired conclusion follows once we prove that uRn \{0}
is continuous on Rn \ {0}.
To this end, assume first that k < −n and fix ψ ∈ C0∞ (Rn ) such that ψ ≡ 1
on B(0, 1). Use this to decompose u = ψu + (1 − ψ)u. Since ψu ∈ E (Rn ) part
(b) in Theorem 4.35 gives ψ u ∈ C ∞ (Rn ). Furthermore, (1 − ψ)u vanishes near the

origin while outside supp ψ becomes u(x) = u |x| |x|x = |x|k u |x|x . Given the current
assumption on k, this behavior implies (1 − ψ)u ∈ L1 (Rn ), hence (1 − ψ)u ∈ C 0 (Rn )
by Lemma 4.28. To summarize, this analysis shows that

u ∈ C 0 (Rn ) whenever k + n < 0. (4.3.12)
To deal with the case k + n ≥ 0, pick some multi-index α ∈ Nn0 arbitrary and define
vα := Dα u ∈ S (Rn ). Since uRn \{0} ∈ C ∞ (Rn \ {0}) differentiating u(tx) = tk u(x)
yields
t|α| (Dα u)(tx) = tk (Dα u)(x) for x ∈ Rn \ {0} and t > 0. (4.3.13)

Given that vα Rn \{0} ∈ C ∞ (R \ {0}), the latter translates into
vα (tx) = tk−|α| vα (x) for x ∈ Rn \ {0} and t > 0. (4.3.14)

Hence, vα Rn \{0} is homogeneous of degree k − |α|. Based on what we proved earlier
(cf. (4.3.12)), it follows that vα ∈ C 0 (Rn ) whenever k − |α| < −n. In terms of the
original distribution u this amounts to saying that
ξα
u ∈ C 0 (Rn ), ∀ α ∈ Nn0 with |α| > k + n. (4.3.15)
The endgame in the proof is then as follows. Given an arbitrary k ∈ R, pick a

natural number N with the property that 2N > k + n. Writing (cf. (14.2.3)) for each
ξ ∈ Rn
N!
|ξ|2N = ξ2α , (4.3.16)
|α|=N
α!
we obtain N!
|ξ|2N
u= u in S (Rn ).
ξ2α (4.3.17)
|α|=N
α!
Collectively, (4.3.17), (4.3.15), and the assumption on N imply
|ξ|2N
u ∈ C 0 (Rn ). (4.3.18)

Since |ξ|12N Rn \{0} ∈ C ∞ (Rn \ {0}), the membership in (4.3.18) further implies that

uRn \{0} ∈ C 0 (Rn \ {0}). This completes the proof of the proposition.

An inspection of the proof of Proposition 4.60 shows that several other useful
versions could be derived, two of which are recorded below.

Exercise 4.61. If u ∈ S (Rn ), uRn \{0} ∈ C ∞ (Rn \ {0}) and uRn \{0} is positive homo-
geneous of degree k, for some k ∈ R satisfying k < −n, then u ∈ C k0 (Rn ), where
k0 := max{ j ∈ N0 , j + k < −n}.

Exercise 4.62. Assume that u ∈ S (Rn ), uRn \{0} ∈ C N (Rn \ {0}) where N ∈ N is

even, and uRn \{0} is positive homogeneous of degree k, for some k ∈ R satisfying

k < N − n. Then uRn \{0} ∈ C m (Rn \ {0}) for every m ∈ N0 satisfying m < N − n − k.
Exercise 4.63. Consider a function b ∈ C 0 (Rn \ {0}) that is positive homogeneous

of degree k, for some k ∈ R satisfying k > −n. Prove that the following are true.
(1) The mapping ub : S(Rn ) → C acting on each ϕ ∈ S(Rn ) according to

ub (ϕ) := b(x)ϕ(x) dx, (4.3.19)
Rn
is a well-defined (via an absolutely convergent integral) tempered distribution in

Rn , i.e., ub ∈ S (Rn ).
(2) Let N ∈ N be even and such that N ≥ 4. Suppose b ∈ C N (Rn \ {0}) is positive
homogeneous of degree k, for some k ∈ R satisfying 1 − n < k < N − n − 1. Fix
an index j ∈ {1, . . . , n} and consider the tempered distribution u∂ j b defined as in
(4.3.19) with ∂ j b in place of b. Then, when restricted to Rn \ {0}, the distributions
u b are given by functions of class C m , for each integer 0 ≤ m < N − n −
∂ j b and u
k − 1, and satisfy the pointwise equality

∂ j b Rn \{0} (ξ) = iξ j u
u b Rn \{0} (ξ), ∀ ξ ∈ Rn \ {0}. (4.3.20)
Sketch of proof of (2):

Step I. Consider the function a j (ξ) := iξ j for each ξ ∈ Rn which belongs to L(Rn ).
The goal here is to show that
u b in S (Rn ).
∂ jb = a ju (4.3.21)
One way to see this is to use Remark 2.38 to obtain that u∂ j b = ∂ j ub in S (Rn ) and
then apply item (2) in Theorem 4.26.
A more direct proof of (4.3.21) is as follows. Pick ϕ ∈ S(Rn ) and use integration
by parts, Lebesgue’s Dominated Convergence Theorem, and the fact that for each
x ∈ Rn we have ∂ j ϕ(x) = − a j ϕ(x) to write

u
∂ jb , ϕ = u∂ jb ,
ϕ = ϕ(x) dx
(∂ j b)(x)
Rn

= lim+ lim ϕ(x) dx
(∂ j b)(x)
ε→0 R→∞ ε<|x|<R

xj
= lim+ lim − b(x)∂ j
ϕ(x) dx + ϕ(x)
b(x) dσ(x)
ε→0 R→∞ ε<|x|<R |x|=R R

xj
− ϕ(x)
b(x) dσ(x)
|x|=ε ε

= a j ϕ(x) dx = ub , a
b(x) jϕ = ub , a j ϕ
Rn

= a j ub , ϕ . (4.3.22)
In justifying the fifth equality in (4.3.22) we also need to observe that the limits as
R → ∞ and as ε → 0+ of the integrals over |x| = R and over |x| = ε, respectively,
are equal to zero. Indeed, by Exercise 4.53 we have |b(x)| ≤ b L∞ (S n−1 ) |x|k for every
x ∈ Rn \ {0}. Hence, for each R > 0, using also Remark 3.4, we can estimate

xj
ϕ(x) dσ(x)
b(x) (4.3.23)
|x|=R R
Rk+n−1
≤ b L∞ (S n−1 ) sup (1 + |x|)k+n |
ϕ(x)| ωn−1 −−−−→ 0.
x∈Rn (1 + R)k+n R→∞
Similarly,

xj
ϕ(x) dσ(x)
b(x)
|x|=ε ε
≤ b L∞ (S n−1 )
ϕ L∞ (Rn ) ωn−1 εk+n−1 −−−−→
+
0, (4.3.24)
ε→0
where for the limit we used the current assumption that k > 1 − n.
Step II. By Exercise 4.51 we have that ∂ j b is positive homogeneous of degree k − 1
in Rn \{0}. Now apply Exercise 4.62 (with k −1 in place of k and N −2 in place of N;
note that the smoothness N in Exercise 4.62 should be an even natural number) and

use the assumption k < N − n − 1 to conclude that u ∂ j b Rn \{0} ∈ C (R \ {0}) for each
m n

integer 0 ≤ m < N − n − k − 1. By Exercise 4.62 we also have ub Rn \{0} ∈ C m (Rn \ {0})
for each integer 0 ≤ m < N − n − k. Now invoke (4.3.21) to finish the proof of
(4.3.20).
Next, we take on the task of computing the Fourier transform of certain homoge-
neous tempered distributions that will be particularly important later in applications.
Recall the gamma function Γ from (14.5.1).
−λ
Let λ ∈ (0, n) and set fλ (x) := |x| , for each x ∈ R \ {0}. Then
n
Proposition 4.64.
n ∞
fλ ∈ S (R ), fλ Rn \{0} ∈ C (R \ {0}), and
n

Γ n−λ

fλ (ξ) = 2n−λ π
n
2 |ξ|λ−n
2
for every ξ ∈ Rn \ {0}. (4.3.25)
λ
Γ 2
Proof. Fix λ ∈ (0, n). Exercise 4.5 then shows that fλ ∈ S (Rn ). Clearly, |x|−λ is
invariant under orthogonal transformations and is positive homogeneous of degree
−λ. Hence, by Proposition 4.49 and Proposition 4.59, it follows that fλ is invariant
under orthogonal transformations and positive homogeneous of degree −n + λ. In
addition, fλ Rn \{0} ∈ C ∞ (Rn \ {0}) by Proposition 4.60.
Fix ξ ∈ Rn \ {0} and choose an orthogonal matrix A ∈ Mn×n (R) with the property
that Aξ = (0, 0, . . . , 0, |ξ|) (such a matrix may be obtained by completing the vector
vn := |ξ|ξ to an orthonormal basis {v1 , . . . , vn } in Rn and then taking A to be the matrix
mapping each v j into e j for j = 1, . . . , n). Then
fλ (ξ) =
fλ (Aξ) =
fλ (0, . . . , 0, |ξ| ) = cλ |ξ|λ−n , (4.3.26)
where cλ := fλ (0, . . . , 0, 1) ∈ C. As such, we are left with determining the value

|x|2
of cλ . We do so by apply fλ to the particular Schwartz function ϕ(x) := e− 2 for
|ξ|2
ϕ(ξ) = (2π) 2 e− 2 for every ξ ∈ Rn .
n
x ∈ Rn . From Example 3.22 we know that
Based on this formula and (4.3.26), the identity
fλ , ϕ = fλ ,
ϕ may be rewritten
as
|ξ|2 |x|2
|ξ|λ−n e− 2 dξ = (2π) 2 |x|−λ e− 2 dx.
n
cλ (4.3.27)
Rn Rn
The two integrals in (4.3.27) may be computed simultaneously, by adopting a

slightly more general point of view, as follows. For any p > −n use polar coor-
dinates (cf. (14.9.9)) and a natural change of variables to write (for the definition
and properties of the gamma function Γ see (14.5.1) and the subsequent comments)
2
∞
ρ2
p − |ξ|2
|ξ| e dξ = ωn−1 ρ p+n−1 e− 2 dρ
Rn 0
∞
p+n−2 p+n−2
= ωn−1 2 2 t 2 e−t dt
0
p+n−2 p + n
=2 2 ωn−1 Γ . (4.3.28)
2
When used with p := λ − n and p := −λ, formula (4.3.28) allows us to rewrite
(4.3.27) as
λ−n+n−2 λ n −λ+n−2 −λ + n
cλ 2 2 ωn−1 Γ = (2π) 2 2 2 ωn−1 Γ . (4.3.29)
2 2
n Γ ( n−λ )
This gives cλ = 2n−λ π 2 Γ( λ2 ) , finishing the proof of (4.3.25).
2
A remarkable consequence of Proposition 4.64 is singled out below.

Corollary 4.65. Assume that n ∈ N, n ≥ 2, and fix λ ∈ [0, n − 1). Then for each
j ∈ {1, . . . , n}, we have

n Γ
n−λ
xj ξj
n−λ in S (Rn ).
2
F = −i 2 π
n−λ−1 2
(4.3.30)
Γ 2 + 1 |ξ|
|x|λ+2 λ
In particular, corresponding to the case when λ = n − 2, formula (4.3.30) becomes

xj ξj
F = −i ωn−1 2 in S (Rn ). (4.3.31)
|x|n |ξ|
Proof. Fix an integer n ≥ 2, and suppose first that λ ∈ (0, n − 1). In this regime, both
(4.3.25) and (4.1.40) hold. In concert with part (b) in Theorem 4.26, these give

xj
−λ F = F ∂ j f λ = iξ j
fλ (ξ)
|x|λ+2

Γ n−λ
ξ j |ξ|λ−n in S (Rn ).
n−λ n2 2
= i2 π (4.3.32)
λ
Γ 2
Hence, whenever λ ∈ (0, n − 1),

xj
F = Cλ ξ j |ξ|λ−n in S (Rn ), (4.3.33)
|x|λ+2
where we have set


n Γ n−λ
2 nΓ n−λ
2
Cλ := −i 2 n−λ−1
π 2 = −i 2n−λ−1
π
2 (4.3.34)
λ λ λ
2Γ 2 Γ 2 +1
and the last equality follows from (14.5.2). This proves formula (4.3.30) in the case
when λ ∈ (0, n − 1). The case when λ = 0 then follows from what we have just
proved, by passing to limit λ → 0+ in (4.3.30) and observing that all quantities
involved depend continuously on λ, in an appropriate sense. Finally, (4.3.31) is a
direct consequence of (4.3.30) and (14.5.6).
4.4 Principal Value Tempered Distributions
Recall the distribution P.V. 1x ∈ D (R) from Example 2.11. As seen from Exer-
cise 4.107, we have P.V. 1x ∈ S (R). The issue we address in this section is the gen-
eralization of this distribution to higher dimensions. The key features of the function
Θ(x) := 1x , x ∈ R \ {0}, that allowed us to define P.V. 1x as a tempered distribution on
the real line are as follows: first, Θ ∈ C 0 (R\{0}), second, Θ is positive homogeneous
of degree −1, and third, Θ(1) + Θ(−1) = 0.
Moving from one dimension to Rn , this suggests considering the class of func-
tions satisfying
Θ ∈ C 0 (Rn \ {0}), positive homogeneous of degree −n,

(4.4.1)
S n−1
Θ dσ = 0.
It is worth noting that a function Θ as above typically fails to be in Lloc

1
(Rn ) (though
obviously Θ ∈ Lloc (R \ {0})). As such, associating a distribution with Θ necessarily
1 n
1
has to be a more elaborate process than the one identifying functions from Lloc (Rn )
with distributions in R . This has already been the case when defining P.V. x and
n 1
here we model the same type of definition in higher dimensions.

Specifically, given Θ as in (4.4.1) consider the linear mapping
P.V. Θ : S(Rn ) → C,

(4.4.2)
P.V. Θ (ϕ) := lim+ Θ(x)ϕ(x) dx, ∀ ϕ ∈ S(Rn ).
ε→0 |x|≥ε
That this definition does the job is proved next.
Proposition 4.66. Let Θ be a function satisfying (4.4.1). Then the map P.V. Θ con-
sidered in (4.4.2) is well defined and is a tempered distribution in Rn . In addition,

P.V. Θ Rn \{0} = ΘRn \{0} in D (Rn \ {0}).
Before proceeding with the proof of Proposition 4.66 we recall a definition and
introduce a class of functions that will be used in the proof.
4.4 Principal Value Tempered Distributions 149
Definition 4.67. A function ψ : Rn → C is called radial if ψ(x) depends only on

|x| for every x ∈ Rn , that is, ψ(x) = f (|x|) for all x ∈ Rn , where f : R → C.
Next, consider the class of functions

⎧
⎪
⎪
⎪ψ : Rn → C, ψ ∈ C 1 (Rn ),
⎪
⎪
⎨
⎪
⎪ ψ is radial and ψ(0) = 1, (4.4.3)
⎪
⎪
⎪
⎩∃ εo ∈ (0, ∞) such that ψ decays like |x|−εo at ∞,
|x|2
(for example ψ(x) = e− 2 , x ∈ Rn , satisfies (4.4.3)) and set

Q := ψ : ψ satisfies (4.4.3) . (4.4.4)
Now we are ready to return to Proposition 4.66.

Proof of Proposition 4.66. Fix an arbitrary ψ satisfying (4.4.3). Then, making use of
formula (14.9.9) and the properties of Θ and ψ, for each ε ∈ (0, ∞) we have
Θ( |x|x )
Θ(x)ψ(x) dx = ψ(x) dx
|x|≥ε |x|≥ε |x|n
∞
ψ(ρ)
= Θ(ω) dσ(ω) dρ = 0. (4.4.5)
ε ρ S n−1
Hence, Θ[ϕ − ϕ(0)ψ] ∈ L1 (Rn ) for every ϕ ∈ S(Rn ), which when combined with
(4.4.5) and Lebesgue Dominated Convergence Theorem 14.15 yields

P.V. Θ (ϕ) = Θ(x)[ϕ(x) − ϕ(0)ψ(x)] dx, ∀ ϕ ∈ S(Rn ). (4.4.6)
Rn
Note that because of (4.4.5), the right-hand side in (4.4.6) is independent of the
choice of ψ. Estimating the right-hand side of (4.4.6) (using Exercise 4.53, the decay
at infinity of functions from Q and the Schwartz class, and the Mean Value Theorem
near the origin) shows that there exists a constant C ∈ (0, ∞) independent of ϕ with
the property that
α β
P.V. Θ (ϕ) ≤ C sup x ∂ ϕ(x), ∀ ϕ ∈ S(Rn ). (4.4.7)
x∈Rn , |α|≤1, |β|≤1
Since from (4.4.6) we see that P.V. Θ is linear, in light of Fact 4.1 estimate (4.4.7)
implies P.V. Θ ∈ S (Rn ) as wanted. The fact that the restriction in the distributional
sense of P.V. Θ to Rn \ {0} is equal to the restriction of the function Θ to Rn \ {0} is
immediate from definitions.
Remark 4.68. (1) As already alluded to, if n = 1 and we take Θ(x) := 1

x for each
x ∈ R \ {0}, then P.V.Θ = P.V. 1x .
(2) Suppose Θ is as in (4.4.1). Since identity (4.4.6) holds for any ψ ∈ Q, we may
select ψ ∈ Q that also satisfies ψ ≡ 1 on B(0, 1) and observe that for this choice of ψ
we have

Θ(x)[ϕ(x) − ϕ(0)ψ(x)] dx (4.4.8)
Rn

= Θ(x) ϕ(x) − ϕ(0) dx + Θ(x)ϕ(x) dx, ∀ ϕ ∈ S(Rn ).
|x|≤1 |x|>1
Hence,

P.V. Θ, ϕ = Θ(x) ϕ(x) − ϕ(0) dx
|x|≤1

+ Θ(x)ϕ(x) dx, ∀ ϕ ∈ S(Rn ). (4.4.9)
|x|>1
x
Example 4.69. If j ∈ {1, . . . , n}, the function Θ defined by Θ(x) := |x|n+1 j
for each
x
x ∈ Rn \ {0} satisfies (4.4.1). By Proposition 4.66 we have P.V. |x|n+1 belongs toj
S (Rn ) and part (2) in Remark 4.68 gives that for every ϕ ∈ S(Rn )

xj x j ϕ(x)
P.V. n+1 , ϕ = lim+ dx (4.4.10)
|x| ε→0 |x|≥ε |x|
n+1

x j (ϕ(x) − ϕ(0)) x j ϕ(x)
= dx + dx.
|x|≤1 |x|n+1
|x|>1 |x|
n+1
The next proposition elaborates on the manner in which principal value tempered
distributions convolve with Schwartz functions.
Proposition 4.70. Let Θ be a function satisfying the conditions in (4.4.1). Then for

each ϕ ∈ S(Rn ) one has that P.V. Θ ∗ ϕ ∈ S (Rn ) ∩ C ∞ (Rn ) and

P.V. Θ ∗ ϕ (x) = lim+ Θ(x − y)ϕ(y) dy, ∀ x ∈ Rn . (4.4.11)
ε→0 |x−y|≥ε
Proof. Fix an arbitrary ϕ ∈ S(Rn ) and note that since P.V. Θ ∈ S (Rn ) (c.f. Proposi-

tion 4.66), part (e) in Theorem 4.19 implies P.V. Θ ∗ ϕ ∈ S (Rn ). Let ψ ∈ C0∞ (Rn )
be such that ψ ≡ 1 near the origin. Then 1 − ψ ∈ L(Rn ) and it makes sense to

consider (1 − ψ) P.V. Θ in S (Rn ) (cf. part (b) in Theorem 4.14). Hence, we may
decompose P.V. Θ = u + v where

u := ψ P.V. Θ ∈ E (Rn ) and v := (1 − ψ) P.V. Θ ∈ S (Rn ). (4.4.12)
The last part in Proposition 4.66 also permits us to identify v = (1 − ψ)Θ in Lloc
1
(Rn ).
By Exercise 4.53 we therefore have v ∈ L (R ) for every p ∈ (1, ∞) which, in
p n
combination with Exercise 3.19, allows us to conclude that v ∗ ϕ belongs to C ∞ (Rn )

4.4 Principal Value Tempered Distributions 151
and
(v ∗ ϕ)(x) = (1 − ψ(y))Θ(y)ϕ(x − y) dy, ∀ x ∈ Rn . (4.4.13)
Rn
Since the above integral is absolutely convergent, by Lebesgue’s Dominated Con-
vergence Theorem permits we further express this as

(v ∗ ϕ)(x) = lim+ (1 − ψ(y))Θ(y)ϕ(x − y) dy, ∀ x ∈ Rn . (4.4.14)
ε→0 |y|≥ε
Thanks to Exercise 2.103 we also have u ∗ ϕ ∈ C ∞ (Rn ) and
(u ∗ ϕ)(x) = ψ P.V. Θ, ϕ(x − ·) for each x ∈ Rn . (4.4.15)
On the other hand, the definition of the principal value gives that for each x ∈ Rn

ψ P.V. Θ, ϕ(x − ·) = P.V. Θ, ψ(·)ϕ(x − ·)

= lim+ Θ(y)ψ(y)ϕ(x − y) dy. (4.4.16)
ε→0 |y|≥ε
Collectively, these arguments show that, for each x ∈ Rn ,

P.V. Θ ∗ ϕ (x) = (u ∗ ϕ)(x) + (v ∗ ϕ)(x)

= lim+ Θ(y)ψ(y)ϕ(x − y) dy
ε→0 |y|≥ε

+ lim+ (1 − ψ(y))Θ(y)ϕ(x − y) dy
ε→0 |y|≥ε

= lim+ Θ(x − y)ϕ(y) dy, (4.4.17)
ε→0 |x−y|≥ε
proving (4.4.11).
The next example discusses a basic class of principal value tempered distribu-
tions arising naturally in applications.
Example 4.71. Let Φ ∈ C 1 (Rn \ {0}) be positive homogeneous of degree 1 − n.

Then for each j ∈ {1, . . . , n} it follows that ∂ j Φ satisfies the conditions in (4.4.1).
Consequently, P.V.(∂ j Φ) is a well-defined tempered distribution.
To see why this is true fix j ∈ {1, . . . , n} and note that ∂ j Φ ∈ C 0 (Rn \ {0}) and
∂ j Φ is positive homogeneous of degree −n (cf. Exercise 4.51). Moreover, using
Exercise 4.52, then integrating by parts based on (14.8.4), and then using (14.9.5),
we obtain

xj
0= Φ(x) dσ(x) − Φ(x)x j dσ(x) = ∂ j Φ(x) dx
|x|=2 2 |x|=1 1<|x|<2

2 2
dρ
= (∂ j Φ)(ρω) ρ n−1
dσ(ω) dρ = (∂ j Φ)(ω) dσ(ω)
1 S n−1 1 ρ S n−1

= (ln 2) (∂ j Φ)(ω) dσ(ω). (4.4.18)
S n−1

This shows that S n−1
∂ j Φ dσ = 0, hence ∂ j Φ satisfies all conditions in (4.4.1).
Principal value tempered distributions often arise when differentiating certain

types of functions exhibiting a point singularity. Specifically, we have the following
theorem.
Theorem 4.72. Let Φ ∈ C 1 (Rn \ {0}) be a function that is positive homogeneous of

degree 1 − n. Then for each j ∈ {1, . . . , n}, the distributional derivative ∂ j Φ satisfies

∂ jΦ = Φ(ω)ω j dσ(ω) δ + P.V.(∂ j Φ) in S (Rn ). (4.4.19)
S n−1
Proof. From the properties of Φ, Exercise 4.53, Exercise 4.54, and Exercise 4.6
it follows that Φ ∈ S (Rn ). Fix j ∈ {1, . . . , n}. By Example 4.71 we have that
P.V.(∂ j Φ) ∈ S (Rn ). Hence, invoking (4.1.33), to conclude (4.4.19) it suffices to
prove that the equality in (4.4.19) holds in D (Rn ). To this end, fix some ϕ ∈ C0∞ (Rn )
and using Lebesgue Dominated Convergence Theorem 14.15 and integration by
parts (based on (14.8.4) write

∂ j Φ, ϕ = −Φ, ∂ j ϕ = − Φ(x)∂ j ϕ(x) dx
Rn

= − lim+ Φ(x)∂ j ϕ(x) dx
ε→0 |x|≥ε

xj
= lim+ ∂ j Φ(x)ϕ(x) dx + lim+ Φ(x)ϕ(x) dσ(x)
ε→0 |x|≥ε ε→0 |x|=ε ε

xj
= P.V.(∂ j Φ), ϕ + lim+ Φ(x)ϕ(x) dσ(x). (4.4.20)
ε→0 |x|=ε ε
Also, for each ε ∈ (0, ∞) we have

4.5 The Fourier Transform of Principal Value Distributions 153

xj
Φ(x)ϕ(x) dσ(x)
|x|=ε ε

xj xj
= Φ(x) ϕ(x) − ϕ(0) dσ(x) + ϕ(0) Φ(x) dσ(x)
|x|=ε ε |x|=ε ε

xj
= Φ(x) ϕ(x) − ϕ(0) dσ(x) + ϕ(0) ω j Φ(ω) dσ(ω), (4.4.21)
|x|=ε ε S n−1
where for the last equality in (4.4.21) we used Exercise 4.52. In addition, using the
fact that ϕ ∈ C0∞ (Rn ) and Exercise 4.53 we may estimate

xj
Φ(x) ϕ(x) − ϕ(0) dσ(x)
|x|=ε ε

1
≤ ε ∇ϕ L∞ (Rn ) Φ L∞ (S n−1 ) dσ(x)
|x|=ε |x|
n−1
= ∇ϕ L∞ (Rn ) Φ L∞ (S n−1 ) ωn−1 ε −−−−→

+
0. (4.4.22)
ε→0
Combining (4.4.20), (4.4.21), and (4.4.22), we conclude

∂ j Φ, ϕ = P.V.(∂ j Φ), ϕ + ω j Φ(ω) dσ(ω) δ, ϕ . (4.4.23)
S n−1

This yields that ∂ j Φ = P.V.(∂ j Φ) + S n−1
ω j Φ(ω) dσ(ω) δ in D (Rn ) and completes
the proof of the theorem.
4.5 The Fourier Transform of Principal Value Distributions
From Proposition 4.66 we know that whenever Θ is a function satisfying the condi-
tions in (4.4.1), the principal value distribution P.V. Θ belongs to S (Rn ). As such,
its Fourier transform makes sense as a tempered distribution. This being said, in
many applications (cf. the discussion in Remark 4.95), it is of basic importance to
actually identify this distribution. The general aim of this section is to do just that,
though as a warm-up, we deal with the following particular (yet relevant) case.
Proposition 4.73. For each k ∈ {1, . . . , n} consider the function

xk
Φk (x) := , ∀ x ∈ Rn \ {0}. (4.5.1)
|x|n
Then for each j ∈ {1, . . . , n} the function ∂ j Φk satisfies the conditions in (4.4.1) and
ξ j ξk ωn−1
F P.V. (∂ j Φk ) = ωn−1 2 − δ jk in S (Rn ). (4.5.2)
|ξ| n
Proof. Fix j, k ∈ {1, . . . , n}. From Example 4.71 it is clear that the function ∂ j Φk is
as in (4.4.1). Moreover, Theorem 4.72 gives that

∂ j Φk = ωk ω j dσ(ω) δ + P.V.(∂ j Φk ) in S (Rn ). (4.5.3)
S n−1
Taking into account (14.9.45), and applying the Fourier transform to both sides, this
yields
ωn−1
F P.V. (∂ j Φk ) = F (∂ j Φk ) − δ jk in S (Rn ). (4.5.4)
n
On the other hand, since by part (b) in Theorem 4.26 and Corollary 4.65 we have
ξ j ξk
F (∂ j Φk ) = i ξ j F (Φk ) = ωn−1 in S (Rn ), (4.5.5)
|ξ|2
formula (4.5.2) follows from (4.5.4)–(4.5.5).
The next theorem shows that the Fourier transform of principal value distribu-
tions P.V. Θ is given by bounded functions. As we shall see in Section 4.9, such a
result plays a key role in establishing the L2 boundedness of singular integral oper-
ators.
Theorem 4.74. Let Θ be a function satisfying the conditions in (4.4.1). Then the
function given by the formula

mΘ (ξ) := − Θ(ω) log(i(ξ · ω)) dσ(ω) for ξ ∈ Rn \ {0}, (4.5.6)
S n−1
is well-defined, positive homogeneous of degree zero, satisfies

mΘ (ξ) dσ(ξ) = 0, (4.5.7)
S n−1
%% %%
%mΘ %L∞ (Rn ) ≤ Cn Θ L∞ (S n−1 ) , (4.5.8)
where Cn ∈ (0, ∞) is defined as

Cn :=
πωn−1
+ ln ξ · ω dσ(ω), (4.5.9)
2 |ξ|
S n−1
and
∨
mΘ = mΘ∨ . (4.5.10)
Moreover, the Fourier transform of the tempered distribution P.V. Θ is of function
type and

F P.V. Θ = mΘ in S (Rn ), (4.5.11)
and
mΘ = mΘ if Θ is an even function. (4.5.12)

In addition,
if Θ ∈ C k (Rn \ {0}) for some k ∈ N0 ∪ {∞},
(4.5.13)
then m ∈ C k (Rn \ {0}).
Θ Rn \{0}
Proof. First, we show that the integral in (4.5.6) is absolutely convergent for each
vector ξ ∈ Rn \ {0}. To see this, fix an arbitrary ξ ∈ Rn \ {0} and observe that for each
ω ∈ S n−1 we have
π
log(i(ξ · ω)) = ln |ξ · ω| + i sgn (ξ · ω)
2
π
= ln |ξ| + ln |ξ|ξ · ω + i sgn (ξ · ω). (4.5.14)
2
ξ
Applying Proposition 14.65 with f (t) := ln |t|, t ∈ R, and v := |ξ| , yields
√
ln ξ · ω dσ(ω) = 2ωn−2
1
|ξ|
ln s( 1 − s2 )n−3 ds < ∞. (4.5.15)
S n−1 0
As a by-product, we note that this implies that the constant Cn from (4.5.9) is finite.
Next, from (4.5.15) and (4.5.9) we obtain

Θ(ω) ln ξ · ω + i π sgn (ξ · ω) dσ(ω)
|ξ| 2
S n−1

πωn−1 ξ
≤ Θ L∞ (S n−1 ) + ln |ξ| · ω dσ(ω)
2 S n−1
= Cn Θ L∞ (S n−1 ) < ∞. (4.5.16)
From (4.5.14) and (4.5.16) it is now clear that the integral in (4.5.6) is absolutely
convergent for each fixed ξ ∈ Rn \{0}. This proves that mΘ is well defined in Rn \{0}.
Going further, since S n−1 Θ dσ = 0, from (4.5.6) and (4.5.14) we see that, for
each ξ ∈ Rn \ {0},
π
mΘ (ξ) = − Θ(ω) ln |ξ|ξ · ω + i sgn (ξ · ω) dσ(ω). (4.5.17)
S n−1 2
Having justified this, we see that mΘ is positive homogeneous of degree zero and that
(4.5.8) follows based on (4.5.16). Also, (4.5.10) is obtained directly from (4.5.6) by
changing variables ω → −ω.
Next, we turn to (4.5.7). Observe that if ω ∈ S n−1 is arbitrary, then there exists
some unitary transformation R in Rn such that Rω = e1 . This combined with
(14.9.11) allows us to write

log(i(ξ · ω)) dσ(ξ) = log(i(R ξ) · ω) dσ(ξ)
S n−1 S n−1

= log(i(ξ · e1 ) dσ(ξ) = C (4.5.18)
S n−1
Hence, (4.5.6), Fubini’s Theorem, (4.5.18), and the last condition in (4.4.1), further
imply

mΘ (ξ) dσ(ξ) = − θ(ω) log(i(ξ · ω)) dσ(ξ) dσ(ω)
S n−1 S n−1 S n−1

= −C θ(ω) dσ(ω) = 0. (4.5.19)
S n−1
This proves (4.5.7).

To show (4.5.11), fix an arbitrary function ϕ ∈ S(Rn ). Using (4.4.2), Lebesgue’s
Dominated Convergence Theorem, Fubini’s Theorem, and (14.9.7) write

F P.V. Θ , ϕ = P.V. Θ, ϕ = lim+ Θ(x)
ϕ(x) dx (4.5.20)
ε→0 |x|≥ε

= lim+ Θ(x)
ϕ(x) dx
ε→0 ε≤|x|≤R
R→∞

= lim+ ϕ(ξ) Θ(x) e−ix·ξ dx dξ
ε→0 Rn ε≤|x|≤R
R→∞
R
dρ
= lim+ ϕ(ξ) Θ(ω) e−i(ω·ξ)ρ dσ(ω) dξ
ε→0 Rn S n−1 ε ρ
R→∞
R
−i(ω·ξ)ρ dρ
= lim+ ϕ(ξ) Θ(ω) e − cos ρ dσ(ω) dξ,
ε→0 Rn S n−1 ε ρ
R→∞
R
where the last equality uses S n−1 Θ(ω) ε cosρ ρ dρ dσ(ω) = 0 for each ε, R > 0
(itself a consequence of the fact that Θ has mean value zero over S n−1 ). At this
stage, we wish to invoke Lebesgue’s Dominated Convergence Theorem in order to
absorb the limit inside the integral. To see that this theorem is applicable in the
current context, we first note that for each ξ ∈ Rn \ {0} and each ω ∈ S n−1 such that
ξ · ω 0, formulas (4.11.5) and (4.11.6) give
R
dρ
lim+ e−i(ω·ξ)ρ) − cos ρ
ε→0 ε ρ
R→∞
R
R
cos (ω · ξ)ρ − cos ρ sin (ω · ξ)ρ
= lim+ dρ − i dr
ε→0 ε ρ ε ρ
R→∞
π
= − ln |ω · ξ| − i sgn (ω · ξ) = − log i(ω · ξ) . (4.5.21)
2
This takes care of the pointwise convergence aspect of Lebesgue’s theorem. To ver-
ify the uniform domination aspect, based on (4.11.7)–(4.11.8) we first estimate

R
dρ
|ϕ(ξ)| |Θ(ω)| sup e−i(ω·ξ)ρ − cos ρ dσ(ω) dξ
Rn S n−1 0<ε<R ε ρ

≤ |ϕ(ξ)| |Θ(ω)| 2 ln |ω · ξ| + 4 dσ(ω) dξ
Rn S n−1
≤ Θ L∞ (S n−1 ) × (4.5.22)

× 4ωn−1 ϕ L1 (Rn ) + 2 |ϕ(ξ)| ln |ω · ξ| dσ(ω) dξ ,
Rn S n−1
and note that, further,

|ϕ(ξ)| ln |ω · ξ| dσ(ω) dξ ≤ |ϕ(ξ)| ln ξ · ω dσ(ω) dξ
|ξ|
Rn S n−1 Rn S n−1

+ ωn−1 |ϕ(ξ)| ln |ξ| dξ. (4.5.23)
Rn
From (4.5.22)–(4.5.23), the fact that ϕ ∈ S(Rn ), and (4.5.15), we may therefore
conclude that
R
dρ
|ϕ(ξ)| |Θ(ω)| sup e−i(ω·ξ)ρ − cos ρ dσ(ω) dξ < ∞. (4.5.24)
Rn S n−1 0<ε<R ε ρ
Having established (4.5.21) and (4.5.24) we may now use Lebesgue’s Dominated
Convergence Theorem in the context of (4.5.20) to obtain

F P.V. Θ , ϕ = ϕ(ξ) mΘ (ξ) dξ. (4.5.25)
Rn
From this, the fact that mΘ ∈ L∞ (Rn ), and keeping in mind that ϕ ∈ S(Rn ) was
arbitrary, (4.5.11) follows.
In order to show (4.5.12), note that if Θ is assumed to be even, then

Θ(ω) sgn (ξ · ω) dσ(ω) = Θ(ω) dσ(ω) − Θ(ω) dσ(ω)
S n−1
ω∈S n−1 ω∈S n−1
ω·ξ>0 ω·ξ<0

= Θ(ω) dσ(ω) − Θ(−ω) dσ(ω)
ω∈S n−1 ω∈S n−1
ω·ξ>0 (−ω)·ξ<0

= Θ(ω) − Θ(−ω) dσ(ω) = 0. (4.5.26)
ω∈S n−1
ω·ξ>0
Consequently,
π
mΘ (ξ) = − Θ(ω) ln |ξ|ξ · ω − i sgn (ξ · ω) dσ(ω)
S n−1 2
ξ π
=− Θ(ω) ln |ξ| · ω + i sgn (ξ · ω) dσ(ω)
S n−1 2

+ iπ Θ(ω) sgn (ξ · ω) dσ(ω)
S n−1
= mΘ (ξ), (4.5.27)
proving (4.5.12).
There remains to prove (4.5.13). To this end, suppose Θ ∈ C k (Rn \ {0}) for some
k ∈ N0 ∪ {∞}. Fix ∈ {1, ..., n} and let e be the unit vector in Rn with one on the
-th component. Introduce the open set

O := ξ = (ξ1 , . . . , ξn ) ∈ Rn : ξ > 0 . (4.5.28)
For each ξ = (ξ1 , . . . , ξn ) ∈ O define the linear map R,ξ : Rn → Rn by
x · ξ + x |ξ| x (|ξ| + 2ξ ) − x · ξ
R,ξ (x) := x − e + ξ, ∀ x ∈ Rn . (4.5.29)
|ξ| + ξ |ξ|(|ξ| + ξ )
ξ
By Exercise 4.130 (presently used with ζ := e and η := |ξ| ) we have that this is a
unitary transformation,
ξ · R,ξ (x) = |ξ|x , ∀ ξ ∈ O , ∀ x ∈ Rn , (4.5.30)
and the joint application
O × Rn (ξ, x) → R,ξ (x) ∈ Rn is of class C ∞ . (4.5.31)
Starting with (4.5.17), then using the invariance under unitary transformations of
the operation of integration over S n−1 (cf. (14.9.11)) and (4.5.30), for each ξ ∈ O
we may then write
π
mΘ (ξ) = − Θ(ω) ln |ξ|ξ · ω + i sgn (ξ · ω) dσ(ω)
S n−1 2
π

=− Θ R,ξ (ω) ln |ξ|ξ · R,ξ (ω) + i sgn ξ · R,ξ (ω) dσ(ω)
S n−1 2

π
=− Θ R,ξ (ω) ln |ω | + i sgn (ω ) dσ(ω). (4.5.32)
S n−1 2
In turn, (4.5.32), (4.5.31), and the current assumption Θ ∈ C k (Rn \ {0}) imply

mΘ O ∈ C k (O ). (4.5.33)

If we also set

O− := x = (x1 , . . . , xn ) ∈ Rn : x < 0 (4.5.34)
and for each given ξ ∈ O− define the linear map R−,ξ : R → R by
n n
x · ξ − x |ξ| x (|ξ| − 2ξ ) + x · ξ
R−,ξ (x) := x + e − ξ, ∀ x ∈ Rn , (4.5.35)
|ξ| − ξ |ξ|(|ξ| − ξ )
running the reasoning that yielded (4.5.33) this time with O replaced by O− and
R,ξ replaced by R−,ξ (now invoking Exercise 4.130 with ζ := e and η := − |ξ|ξ ), and
having identity (4.5.30) replaced by ξ · R−,ξ (x) = −|ξ|x for all ξ ∈ O− and all x ∈ Rn ,
ultimately implies
mΘ O− ∈ C k (O− ). (4.5.36)

&
n
Upon observing that Rn \{0} = (O ∪O− ), from (4.5.33) and (4.5.36), we conclude
=1
that m ∈ C k (Rn \ {0}), as wanted. This finishes the proof of the theorem.
Θ Rn \{0}
Exercise 4.75. Complete the following outline aimed at extending the convolution
product to the class of principal value distributions in Rn . Assume that Θ1 , Θ2 are
two given functions as in (4.4.1).
Step 1. Pick an arbitrary function ψ ∈ C0∞ (Rn ) with the property that ψ ≡ 1 near
the origin, and show that the following convolutions are meaningfully defined in
S (Rn ):

u00 := ψ P.V. Θ1 ∗ ψ P.V. Θ2

u01 := ψ P.V. Θ1 ∗ (1 − ψ) P.V. Θ2

u10 := (1 − ψ) P.V. Θ1 ∗ ψ P.V. Θ2

u11 := (1 − ψ) P.V. Θ1 ∗ (1 − ψ) P.V. Θ2 . (4.5.37)
For u00 , u01 , and u10 , use Proposition 4.66 and part (a) of Theorem 4.19. Show that
u11 = f1 ∗ f2 where f j := (1 − ψ)Θ j , j = 1, 2, are functions belonging to L2 (Rn ) (here
the behavior of Θ1 , Θ2 at infinity is relevant). Use Young’s inequality to conclude

that u11 is meaningfully defined in L∞ (Rn ).
Step 2. With the same cutoff function ψ as in Step 1, define

P.V. Θ1 ∗ P.V. Θ2 := u00 + u01 + u10 + u11 in S (Rn ), (4.5.38)
and use this definition to show that

F P.V. Θ1 ∗ P.V. Θ2 = mΘ1 mΘ2 in S (Rn ), (4.5.39)
where mΘ1 , mΘ2 are associated with Θ1 , Θ2 as in (4.5.6). To do so, compute first ujk
for j, k ∈ {0, 1}, using part (c) in Theorem 4.35, Proposition 4.29, and Theorem 4.74.
Step 3. Use (4.5.39) to show that the definition in (4.5.38) is independent of the
cutoff function ψ chosen at the beginning.
As an application, show that
1 1
P.V. ∗ P.V. = −π2 δ in S (R). (4.5.40)
x x
4.6 Tempered Distributions Associated with |x|−n
Let us consider the effect of dropping the cancelation condition in (4.4.1). Con-
Ψ ∈ C (R \ {0}) that is positive homogeneous of degree
0 n
cretely, given a function
−n, if one sets C := S n−1 Ψ dσ, then Ψ (x) = Ψ0 (x) + C|x|−n for every x in Rn \ {0},
where Ψ0 satisfies (4.4.1).
From Section 4.4 we know that one may associate to Ψ0 the tempered distribution
P.V. Ψ0 . As such, associating a tempered distribution to the original function Ψ
hinges on how to meaningfully associate a tempered distribution to |x|1n . In fact, we
shall associate to |x|1n a family of tempered distributions in the manner described
below.
Recall the class of functions Q from (4.4.4), fix ψ ∈ Q, and define

1
wψ (ϕ) := [ϕ(x) − ϕ(0)ψ(x)] dx, ∀ ϕ ∈ S(Rn ). (4.6.1)
R n |x|n
This gives rise to a well-defined, linear mapping wψ : S(Rn ) → C. In addition,

much as in the proof of (4.4.7), there exists a constant C ∈ [0, ∞) with the property
that α β
wψ (ϕ) ≤ C sup x ∂ ϕ(x), ∀ ϕ ∈ S(Rn ). (4.6.2)
x∈Rn , |α|≤1, |β|≤1
In light of Fact 4.1 this shows

that wψ is a tempered distribution. An inspection of
(4.6.1) also reveals that wψ Rn \{0} = |x|−n in D (Rn \ {0}). This is the sense in which
4.6 Tempered Distributions Associated with |x|−n 161
we shall say that we have associated to |x|−n the family of tempered distributions
{wψ }ψ∈Q .
Our next goal is to determine a formula for the Fourier transform of the tem-
pered distribution in (4.6.1). As a preamble, note that the class Q is invariant under
dilations (recall (4.3.5)); that is, for every t ∈ (0, ∞) we have
τt (ψ) ∈ Q, ∀ t ∈ (0, ∞), ∀ ψ ∈ Q. (4.6.3)
Theorem 4.76. For each ψ ∈ Q, the Fourier transform of the tempered distribution
wψ defined in (4.6.1) is of function type and
ψ (ξ) = −ωn−1 ln |ξ| + Cψ ,

w ∀ ξ ∈ Rn \ {0}, (4.6.4)
for some constant Cψ ∈ C.

Proof. Fix ψ ∈ Q. As noted earlier, wψ Rn \{0} = |x|1n , hence Proposition 4.60 applies
and yields
ψ Rn \{0} ∈ C ∞ (Rn \ {0}).
w (4.6.5)
To establish (4.6.4) we proceed by proving a series of claims.
Claim 1. w ψ is invariant under orthogonal transformations. let A ∈ Mn×n (R) be
an orthogonal matrix. Then

1
wψ ◦ A, ϕ = wψ , ϕ ◦ A−1 = ϕ(A−1 x) − ϕ(0)ψ(A−1 x) dx
Rn |x|
n
= wψ , ϕ , ∀ ϕ ∈ S(Rn ). (4.6.6)
This shows that wψ is invariant under orthogonal transformations which, together

ψ is also invariant under orthogonal transfor-
with Proposition 4.49, implies that w
mations.
Claim 2. τt (wψ ) = t−n wτt ψ in S (Rn ). To see why this is true, for any t ∈ (0, ∞)
and any ϕ ∈ S(Rn ) write

1 ' x (
τt wψ , ϕ = t−n wψ , τ 1t ϕ = t−n ϕ − ϕ(0)ψ(x) dx
Rn |x|
n t

1
= n |y|n
ϕ(y) − ϕ(0)ψ(ty) dy = t−n wτt ψ , ϕ , (4.6.7)
R n t
from which the desired identity follows.
Claim 3. For every t ∈ (0, ∞) we have

ψ(x) − ψ(tx)
τt wψ − t−n wψ = t−n dx δ in S (Rn ). (4.6.8)
Rn |x|n
To prove the current claim, consider ψ1 , ψ2 ∈ Q and write


ϕ(0)
wψ1 − wψ2 , ϕ = [ψ2 (x) − ψ1 (x)] dx
Rn |x|
n
)
*
ψ2 (x) − ψ1 (x)
= dx δ , ϕ ∀ ϕ ∈ S(Rn ). (4.6.9)
Rn |x|n
As a result,

1
wψ1 − wψ2 = [ψ2 (x) − ψ1 (x)] dx δ in S (Rn ). (4.6.10)
Rn |x|n
Combining now the identity from Claim 2 and (4.6.10) yields
τt wψ − t−n wψ = t−n [wτt ψ − wψ ]

−n ψ(x) − ψ(tx)
= t dx δ in S (Rn ) (4.6.11)
Rn |x|n
for every t ∈ (0, ∞). This proves Claim 3.

Claim 4. The following identity is true:

ψ(x) − ψ(tx)
dx = ωn−1 ln t ∀ t ∈ (0, ∞). (4.6.12)
R n |x|n
To set the stage for the proof of Claim 4 observe that there exists some C 1 func-
tion η : [0, ∞) → R decaying at ∞, satisfying η(0) = 1, and ψ(x) = η(|x|) for every
x ∈ Rn . In particular, (∇ψ)(x) = η (|x|) |x|x for every x ∈ Rn \ {0}. Focusing attention
on (4.6.12), note that the intervening integral is absolutely convergent. In turn, this
allows us to write for each fixed t ∈ (0, ∞)

ψ(x) − ψ(tx) 1 1 d
dx = lim ψ(sx) ds dx
Rn |x|n R→∞ |x|<R |x|n t ds

1 1
= lim η (s|x|) ds dx
R→∞ |x|<R |x|n−1 t
R 1
= ωn−1 lim η (sρ) ds dρ (4.6.13)
R→∞ 0 t
1 R
1 d
= ωn−1 lim η(sρ) dρ ds
R→∞ t 0 s dρ
1
1
= ωn−1 lim η(sR) − η(0) ds = ωn−1 ln t,
R→∞ t s
as wanted.
Claim 5. The following identity holds:
ψ = w
τt w ψ − ωn−1 ln t in Rn \ {0}, ∀ t ∈ (0, ∞). (4.6.14)
To prove (4.6.14), we first combine (4.6.8) and (4.6.12) and obtain
τt wψ = t−n wψ + (ωn−1 t−n ln t)δ in S (Rn ), ∀ t ∈ (0, ∞). (4.6.15)
Hence, for each t ∈ (0, ∞) and each ϕ ∈ S(Rn ), we have

ψ , ϕ = t−n w
τt w ϕ = t−n τ 1t wψ ,
ψ , τ 1 ϕ = wψ , τt ϕ
t

= t−n tn wψ − (ωn−1 tn ln t)δ ,
ϕ = wψ − (ωn−1 ln t)δ ,
ϕ

= w ψ − ωn−1 ln t , ϕ . (4.6.16)
The first and third equality in (4.6.16) is based on the definition of τt acting on
S (Rn ), the second uses (4.3.7), while the fourth
makes use of (4.6.15). Now (4.6.14)
follows from (4.6.16) and the fact that w ψ Rn \{0} ∈ C ∞ (Rn \ {0}).
We are ready to complete the proof of Theorem 4.76. First, Claim 1 ensures that
ψ Rn \{0} is constant on S n−1 , thus
w

ψ S n−1
Cψ := w (4.6.17)
is a well-defined complex number. Hence, given any ξ ∈ Rn \ {0}, taking t := |ξ|1 in

(4.6.14) yields

ξ
ψ
Cψ = w = τ |ξ|1 w ψ (ξ) + ωn−1 ln |ξ|,
ψ (ξ) = w (4.6.18)
|ξ|
from which (4.6.4) follows.
Remark 4.77. It is possible to enlarge the scope of our earlier considerations by

relaxing the hypotheses on the function ψ. Specifically, in place of (4.4.3) assume
now that ψ : Rn → C is such that
ψ is radial, of class C 1 near origin, satisfies ψ(0) = 1,

(4.6.19)
and ∃ εo , C ∈ (0, ∞) such that |ψ(x)| ≤ C(1 + |x|)−εo .
Such a function ψ still induces a tempered distribution as in (4.6.1), and we claim

that formula (4.6.4) continues to be valid in this more general case as well.
To justify this claim, observe that the proof of Theorem 4.76 goes through for
ψ as in (4.6.19), albeit (4.6.12) requires more care, since we are now dropping the
global C 1 requirement on η. However, this may be remedied by working with ηε in
place of η, where ηε is defined via mollification as follows. If {φε } is the sequence
from (1.2.3)–(1.2.5), then

ηε (x) := η(y)φε (x − y) dy, ∀ x ∈ Rn . (4.6.20)
Rn
Now observe that if ψε (x) := ηε (|x|) then ψε /ψε (0) ∈ Q, hence

ψε (x) − ψε (tx)
dx = ωn−1 ψε (0) ln t ∀ t ∈ (0, ∞), (4.6.21)
R n |x|n
and finally passing to limit ε → 0+ in (4.6.21).

Remark 4.77 applies, in particular, to the function ψ := χB(0,1) . This gives that
ϕ(x)

ϕ(x) − ϕ(0)
wχB(0,1) , ϕ = dx + dx, ∀ ϕ ∈ S(Rn ), (4.6.22)
|x|>1 |x| |x|n
n
|x|≤1

is a tempered distribution, F wχB(0,1) ∈ S (Rn ) is of function type, and there exists a
constant c ∈ C such that

F wχB(0,1) (ξ) = −ωn−1 ln |ξ| + c, ∀ ξ ∈ Rn \ {0}. (4.6.23)
In the last part of this section we compute the constant c from (4.6.23) in the case
n = 1. Before doing so we recall that Euler’s constant γ is defined by

k
1
γ := lim − ln k . (4.6.24)
k→∞
j=1
j
The number γ plays an important role in analysis, as it appears prominently in a

number of basic formulas. Two such identities which are relevant for us here are as
follows:
∞ ∞ √
π
e−x ln x dx = −γ, e−x ln x dx = −
2
(γ + 2 ln 2). (4.6.25)
0 0 4
Example 4.78. We claim that

F wχ(−1,1) (ξ) = −2 ln |ξ| − 2γ in S (R), (4.6.26)
where wχ(−1,1) is defined in (4.6.22). To see why this is true, note that (4.6.23) written
for n = 1 (in which case ωn−1 = 2) becomes

F wχ(−1,1) (x) = −2 ln |x| + c in S (R). (4.6.27)
Consequently, to obtain (4.6.26) it remains to prove that c = −2γ when n = 1.

The idea for determining the value of c is to apply F wχ(−1,1) to the Schwartz
function e−x . First, from (4.6.23) we have
2

F wχ(−1,1) , e−x = −2 ln |x| + c, e−x .
2 2
(4.6.28)
Second, we compute the term in the left-hand side of (4.6.28). Specifically, based
on the definition of the Fourier transform of a tempered distribution, (3.2.6), and
(4.6.22) we may write
2 √
F wχ(−1,1) , e−x = wχ(−1,1) , F (e−x ) = wχ(−1,1) , πe−x /4
2 2

√ e−x /4 √ e−x /4
2 2
−1
= π dx + π dx
|x|>1 |x| |x|≤1 |x|
∞
√ e−x √ e−x − 1
2 1/2 2
=2 π dx + 2 π dx
1/2 x 0 x
=: I + II. (4.6.29)
Going further, integrating by parts then changing variables gives

∞
√ √
I = 2 πe−1/4 ln 2 + 4 π (ln x)e−x x dx
2
1/2
∞
√ √
= 2 πe−1/4 ln 2 + π √ (ln x)e−x dx. (4.6.30)
1/ 2
Regarding II, for each ε > 0 an integration by parts and then a change of variables
yield

e−x − 1
1/2 2
dx (4.6.31)
ε x
1/2
−1/4 −ε2
(ln x)e−x x dx
2
= −(e − 1) ln 2 − (e − 1) ln ε + 2
ε
√
1/ 2
−1/4 −ε2 1
= −(e − 1) ln 2 − (e − 1) ln ε + √
(ln x)e−x dx.
2 ε
Observe that since

1 ln ln y − ln(y − 1)
lim+ (e−ε − 1) ln ε =
2
lim = 0, (4.6.32)
ε→0 2 y→∞ y
by taking the limit as ε → 0+ in (4.6.31) we may conclude that

√
√ −1/4
√ 1/ 2
II = −2 π(e − 1) ln 2 + π (ln x)e−x dx. (4.6.33)
0
Hence, from (4.6.30), (4.6.33), and the first identity in (4.6.25), it follows that
∞
√ √ √ √
I + II = 2 π ln 2 + π (ln x)e−x dx = 2 π ln 2 − πγ. (4.6.34)
0
This takes care of the term in the left-hand side of (4.6.26). To compute the term
in the right-hand side of (4.6.26), write

−2 ln |ξ| + c, e−x = −2 (ln |x|)e−x dx + c e−x dx
2 2 2
R R
∞ √
(ln x)e−x dx + c π
2
= −4
0
√ √
= π(γ + 2 ln 2) + c π. (4.6.35)
For the last equality in (4.6.35) we used the second identity in (4.6.25). A combina-
tion of (4.6.28), (4.6.29), (4.6.34), and (4.6.35), then implies c = −2γ, as desired.
4.7 A General Jump-Formula in the Class of Tempered

Distributions
The aim in this section is to prove a very useful formula expressing the limits of cer-
tain sequences of tempered distributions {Φε }ε0 as the index parameter ε ∈ R \ {0}
approaches the origin from either side. The trademark feature (which also justifies
the name) of this formula is the presence of a jump-term of the form ±Cδ (where
the sign is correlated to sgn ε) in addition to a suitable principal value tempered dis-
tribution. A conceptually simple example of this phenomenon has been presented
in Exercise 2.135 which our theorem contains as a simple special case (see Re-
mark 4.82).
With the notational convention that for points x ∈ Rn we write x = (x , t), where

x = (x1 , . . . , xn−1 ) ∈ Rn−1 and t ∈ R, our main result in this regard reads as follows.
Theorem 4.79. If Φ ∈ C 4 (Rn \{0}) is odd and positive homogeneous of degree 1−n,
then
i
lim± Φ(x , ε) = ± Φ(0 , 1) δ(x ) + P.V. Φ(x , 0) in S (Rn−1 ). (4.7.1)
ε→0 2
A few comments before presenting the proof of this result are in order. First,
above we have employed the earlier convention of writing u(x ) for a distribution u
in Rn−1 simply to stress that the test functions to which u is applied are considered
in the variable x ∈ Rn−1 .
Second, for each fixed ε ∈ R \ {0}, applying Exercise 4.53 yields

|Φ(x , ε)| ≤ Φ L∞ (S n−1 ) |x |2 + ε2 −(n−1)/2 for each x ∈ Rn−1 . (4.7.2)
Having observed this, the discussion in Example 4.4 then shows that Φ(·, ε) belongs
to S (Rn−1 ).
Third, it is worth recalling an earlier convention to the effect that given a family
of tempered distributions uε , indexed by ε ∈ I, I = (a, b) ⊆ R open interval, we say
4.7 A General Jump-Formula in the Class of Tempered Distributions 167
that uε → u ∈ D (Rn ) in S (Rn ) as I ε → a+ provided uε j → u in S (Rn ) for every

+ ,
sequence ε j ⊆ I such that ε j → a+ as j → ∞. In particular, this interpretation
j∈N
is in effect for (4.7.1).
Fourth, from Exercise 4.54 we know that Φ ∈ S (Rn ). In addition, Exercise 4.62
applied to Φ shows that Φ n ∈ C 1 (Rn \{0}). In particular, Φ(0
, 1) is meaningfully
R \{0}
defined.
Fifth, Φ(·, 0) (viewed as a function in Rn−1 \{0 }) is continuousand positive homo-
geneous of degree −(n−1) in Rn−1 \{0 } and, being odd, satisfies S n−2 Φ(·, 0) dσ = 0.
As such, the conditions in (4.4.1) are satisfied (with n − 1 in place of n); hence,
Proposition 4.66 ensures that P.V. Φ(·, 0) is a well-defined tempered distribution in
Rn−1 .
Proof of Theorem 4.79. Assume Φ ∈ C 4 (Rn \ {0}) is odd and positive homogeneous
of degree 1 − n, and let ϕ ∈ S(Rn−1 ). Then, for any fixed ε > 0, write

lim+ Φ(x , t)ϕ(x ) dx = lim+ Φ(x , t)ϕ(x ) dx
t→0 Rn−1 t→0
|x |>ε

+ lim+ Φ(x , t)[ϕ(x ) − ϕ(0 )] dx
t→0
|x |<ε

+ ϕ(0 ) lim+ Φ(x , t) dx
t→0
|x |<ε
=: Iε + IIε + IIIε . (4.7.3)
Note that, making the change of variable y := x /t, we obtain

IIIε = ϕ(0 ) lim+ Φ(y , 1) dy = ϕ(0 ) lim Φ(y , 1) dy , (4.7.4)
t→0 r→∞
|y |<ε/t |y |<r
that is independent of ε. Also, since for every x ∈ Rn−1 and t ∈ R \ {0},
Φ|L∞ (S n−1 ) Φ|L∞ (S n−1 )

|Φ(x , t)| ≤ ≤ , (4.7.5)
|(x , t)|n−1 |x |n−1
|ϕ(x ) − ϕ(0 )| ≤ ∇ϕ|L∞ (Rn−1 ) |x |, (4.7.6)
it follows that |IIε | ≤ Cε, hence
lim IIε = 0. (4.7.7)

ε→0+
Finally, it is clear from Lebesgue’s Dominated Convergence Theorem (cf. Theo-

rem 14.15) that
Iε = Φ(x , 0)ϕ(x ) dx . (4.7.8)
|x |>ε
By passing to the limit as ε → 0+ we therefore conclude that for any function

ϕ ∈ S(Rn−1 )

lim+ Φ(x , t)ϕ(x ) dx = lim+ Φ(x , 0)ϕ(x ) dx
t→0 Rn−1 ε→0
|x |>ε

+ ϕ(0 ) lim Φ(x , 1) dx . (4.7.9)
r→∞
|x |<r
To proceed, make the change of variable x → −x in each integral of (4.7.9)

and use the fact that Φ is odd. After re-denoting t by −t and ϕ∨ by ϕ this yields the
identity

lim− Φ(x , t)ϕ(x ) dx = lim+ Φ(x , 0)ϕ(x ) dx
t→0 Rn−1 ε→0
|x |>ε

+ ϕ(0 ) lim Φ(x , −1) dx , (4.7.10)
r→∞
|x |<r
for any ϕ ∈ S(Rn−1 ). Collectively, (4.7.9) and (4.7.10) may then be written as

lim± Φ(x , t)ϕ(x ) dx (4.7.11)
t→0 Rn−1

= a± ϕ(0 ) + lim+ Φ(x , 0)ϕ(x ) dx , ∀ ϕ ∈ S(Rn−1 ),
ε→0
|x |>ε
where
a± := lim Φ(x , ±1) dx . (4.7.12)
r→∞
|x |<r
Regarding (4.7.12), we note that the limits

lim Φ(x , ±1) dx exist in C. (4.7.13)
r→∞ |x |<r
To see that this is the case, first observe that by Exercise 4.53,

|Φ(x , t)| dx < ∞, ∀ t 0, ∀ r > 0. (4.7.14)
|x |<r
Consider the case of the choice of the sign “plus” in (4.7.13) (the case of the sing
“minus” is treated analogously). For this choice of sign we write for each fixed r > 0

1 1
Φ(x , 1) dx = Φ(x , 1) dx + Φ(x , 1) dx
|x |<r 2 |x |<r 2 |x |<r

1 1
= Φ(x , 1) dx − Φ(x , −1) dx
2
|x |<r 2 |x |<r

1
= 2 Φ(x , 1) − Φ(x , −1) dx , (4.7.15)
|x |<r
where we have used the fact that Φ is odd. Next, the Mean Value Theorem and the
fact that ∇Φ is positive homogeneous of degree −n allow us to estimate
2 ∇Φ|L∞ (S n−1 )
|Φ(x , 1) − Φ(x , −1)| ≤ for |x | large, (4.7.16)
|x |n
which implies that

|Φ(x , 1) dx − Φ(x , −1)| dx < ∞. (4.7.17)
Rn−1
Consequently, by Lebesgue’s Dominated Convergence Theorem

1
lim Φ(x , 1) dx =
2 Φ(x , 1) dx − Φ(x , −1) dx ∈ C, (4.7.18)
r→∞ |x |<r Rn−1
proving (4.7.13).
There remains to identify the actual values of a± and we organize the remainder
of the proof as a series of claims, starting with:
Claim 1: If u0 ∈ E (Rn ) and u1 ∈ L1 (Rn ), then
u := u0 + u1 has the property that

u ∈ C 0 (Rn ). (4.7.19)
Proof of Claim 1. First note that

u =
u0 + u0 ∈ C ∞ (Rn ) by part (b) in
u1 . Since
Theorem 4.35 and u1 ∈ C (R ) by Lemma 4.28, it follows that
0 n
u ∈ C 0 (Rn ).
Claim 2: Assume that ξn 0 and that ξ ∈ Rn−1 is such that |ξ | ≤ C for some
C ∈ (0, ∞). Also, suppose that Θ ∈ C 1 (Rn \ {0}) satisfies
Θ(λx) = λ−1 Θ(x), ∀ x ∈ Rn \ {0}, ∀ λ ∈ R \ {0}. (4.7.20)
Then there exists C ∈ (0, ∞) independent of ξn such that

C
|Θ(ξ , −ξn ) + Θ(0 , ξn )| ≤ . (4.7.21)
|ξn |2
Proof of Claim 2. From (4.7.20) it follows that (∇Θ)(λx) = λ−2 (∇Θ)(x) for every
x ∈ Rn \ {0} and every λ ∈ R \ {0}. Based on (4.7.20) and this observation, we may
estimate
1
|Θ(ξ , −ξn ) + Θ(0 , ξn )| = |Θ(−ξ /ξn , 1) − Θ(0 , 1)|
|ξn |

1 ξ
≤ · · sup |(∇Θ)(−tξ /ξn , 1)|
|ξn | ξn t∈[0,1]

1 ξ ∇Θ|L∞ (S n−1 )
≤ · · sup
|ξn | ξn t∈[0,1] |(−tξ /ξn , 1)|2
C
≤ · ∇Θ|L∞ (S n−1 ) , (4.7.22)
|ξn |2
where the last line uses the assumption that |ξ | ≤ C . Since ∇Θ is continuous,
hence bounded, on S n−1 , the desired result follows by taking the constant to be
C := C ∇Θ|L∞ (S n−1 ) < ∞.
Claim 3: If ϕ ∈ S(Rn−1 ) is such that
ϕ has compact support, then

, ξn )
F(ξn ) := −(2π)1−n Φ(ξ , −ξn ) + Φ(0 ϕ(ξ ) dξ . (4.7.23)
Rn−1
is well defined and integrable on R \ [−1, 1].

Proof of Claim 3. From Proposition 4.59 it follows that Φ is a positive homogeneous

tempered distribution of degree −n − (1 − n) = −1. This readily implies that Φ n ,
R \{0}
viewed as a continuous function in Rn \ {0}, is also positive homogeneous of degree
−1. Moreover, from (4.2.36) we deduce is an odd function in Rn \ {0}. As a
that Φ

consequence, the function Θ := Φ n ∈ C (Rn \ {0}) satisfies (4.7.20).
1
R \{0}
Hence, Claim 2 applies to this Θ and, granted the compact support condition on

ϕ, it follows that there exists C ∈ (0, ∞) such that
C
|F(ξn )| ≤ , ∀ ξn 0, (4.7.24)
|ξn |2
from which the desired conclusion follows.
Claim 4: Assume ϕ ∈ S(Rn−1 ) is such that ϕ has compact support. Then the function

1 , 1)ϕ(0 ),
f (t) := Φ(x , t)ϕ(x ) dx + (sgn t)Φ(0 (4.7.25)
Rn−1 2i
originally defined for t ∈ R \ {0}, has a continuous extension to all of R.

Proof of Claim 4. Decompose f = f1 + f2 where f1 (t) := Rn−1 Φ(x , t)ϕ(x ) dx and
, 1)ϕ(0 ) for each t ∈ R \ {0}. From (13.4.18) and (3.2.10) we
f2 (t) := 2i1 (sgn t)Φ(0
know that

1
n ) = −2i P.V.
sgn(ξ , and ϕ(0 ) = (2π)1−n ϕ(ξ ) dξ .
(4.7.26)
ξn R n−1

Also, Φ is odd and positive homogeneous of degree −1 in Rn \{0}, and since P.V. 1
ξn
restricted to R \ {0} is simply ξ1n , it follows that

1 , ξn )
P.V. Φ(0 , 1) = Φ(0 on R \ {0}.
ξn
Keeping these in mind we conclude that

f2 (ξn ) = −(2π) 1−n , ξn )
Φ(0 ϕ(ξ ) dξ for ξn ∈ R \ {0}. (4.7.27)
Rn−1
As far as f1 is concerned, observe first that f1 ∈ S (R). Indeed, for every ψ ∈

S(R) we have f1 , ψ = Φ, ϕ ⊗ ψ . Since ϕ ⊗ ψ ∈ S(Rn ) and, as already noted,
Φ ∈ S (Rn ), the desired conclusion follows. The next order of business is to compute
the Fourier transform of f1 . With this goal in mind, pick an arbitrary ψ ∈ S(R) and
write (keeping in mind Exercise 3.27)

f1 , ψ = f1 ,
ψ = Φ, ϕ ⊗ ψ
∨
= (2π)1−n Φ, F ( ϕ∨ ⊗ ψ) = (2π)1−n Φ,
ϕ ⊗ψ

= (2π)1−n Φ(ξ , ξn ),
ϕ(−ξ ) , ψ(ξn ) (4.7.28)
which proves that

f1 (ξn ) = (2π)1−n , ξn )
Φ(ξ ϕ(−ξ ) dξ for ξn ∈ R \ {0}. (4.7.29)
Rn−1
By combining (4.7.27) with (4.7.29) we arrive at the conclusion that, for each ξn ∈
R \ {0},

f (ξn ) = (2π)1−n , ξn )
Φ(ξ ϕ(−ξ ) dξ − (2π)1−n , ξn )
Φ(0 ϕ(ξ ) dξ
Rn−1 Rn−1

= −(2π)1−n , −ξn )
Φ(ξ ϕ(ξ ) dξ − (2π)1−n , ξn )
Φ(0 ϕ(ξ ) dξ
Rn−1 Rn−1
= F(ξn ), (4.7.30)
where the second equality uses the fact that Φ is odd in Rn \ {0}. Hence,
f = F on
R \ {0}, where F is defined in Claim 3. Now select θ ∈ C0∞ (R) with θ ≡ 1 on [−1, 1]
and write

f = (1 − θ)
f + θf. (4.7.31)
Since (1 − θ) f = (1 − θ)F ∈ L1 (R) by Claim 3, and θ
f ∈ E (R), we may conclude
from Claim 1 that the Fourier transform of f belongs to C 0 (R) hence, ultimately,
0
that f itself belongs to C (R). This completes the proof of Claim 4.
Claim 5: Assume that ϕ ∈ S(Rn−1 ) is such that
ϕ has compact support. Then

lim Φ(x , t)ϕ(x ) dx − lim− Φ(x , t)ϕ(x ) dx
t→0+ Rn−1 t→0 Rn−1
, 1)ϕ(0 ).
= i Φ(0 (4.7.32)
Proof of Claim 5. This follows from Claim 4 by writing for each t ∈ R \ {0}

1 , 1)ϕ(0 ) + f (t)
Φ(x , t)ϕ(x ) dx = − (sgn t)Φ(0 (4.7.33)
R n−1 2i
with f continuous on R.
Claim 6: For a± originally defined in (4.7.12) one has
i
a± = ± Φ(0 , 1). (4.7.34)
2
Proof of Claim 6. Let ψ ∈ C0∞ (Rn−1 ) be such that

ψ(x ) dx = 1, (4.7.35)
Rn−1
and set ϕ := ϕ = (2π)n−1 ψ∨ has compact support, and

ψ. Then ϕ ∈ S(Rn−1 ),

ϕ(0 ) =
ψ(0 ) = ψ(x ) dx = 1. (4.7.36)
Rn−1
From (4.7.11) and Claim 5, one obtains

, 1) = a+ − a− .
i Φ(0 (4.7.37)
However, since Φ is odd, from the definition of a± in (4.7.12) we see that a− = −a+ .
This forces the equalities in (4.7.34) and finishes the proof of Claim 6.
At this stage, we note that (4.7.1) is a consequence of (4.7.11) and (4.7.34). The
proof of Theorem 4.79 is therefore complete.
The proof just completed offers a bit more and, below, we bring to the forefront
one such by-product.
Proposition 4.80. Assume that Φ ∈ C 4 (Rn \ {0}) is odd and positive homogeneous
of degree 1 − n. Also, let ξ ∈ S n−1 be arbitrary and set
Hξ := {x ∈ Rn : x · ξ = 0}. (4.7.38)
Then
= −2i lim
Φ(ξ) Φ(x + ξ) dσ(x), (4.7.39)
r→∞
x∈Hξ , |x|<r
where σ denotes the surface measure on Hξ (viewed as surface in Rn ).

Proof. We start by observing that (4.7.12) and (4.7.34) imply that

Ψ (en ) = −2i lim Ψ (x , 0) + en dx
r→∞
x ∈Rn−1 , |x |<r
(4.7.40)
for each function Ψ of class C 4 , odd, and positive
homogeneous of degree 1 − n, in Rn \ {0}.
To proceed, fix a vector ξ ∈ S n−1 and denote by Hξ the hyperplane in Rn orthogonal

to ξ. Consider then an orthogonal matrix A ∈ Mn×n (R) satisfying
AHξ = Rn−1 × {0} and Aξ = en . (4.7.41)
For example, if v1 , ..., vn−1 is an orthonormal basis in Hξ and we set vn := ξ, then

the conditions Av j = e j for j ∈ {1, . . . , n} define an orthogonal matrix A such that
the conditions in (4.7.41) hold. If we now introduce Ψ : Rn \ {0} → C by setting
Ψ := Φ ◦ A−1 then Ψ ∈ C 4 (Rn \ {0}) is odd, positive homogeneous of degree 1 − n,
and Ψ=Φ ◦ A−1 (as seen from (4.3.3)). Using these observations and (4.7.40), we
may write

=Ψ (Aξ) = Ψ (en ) = −2i lim
Φ(ξ) Ψ (x , 0) + en dx . (4.7.42)
r→∞
x ∈Rn−1 , |x |<r
Given a number r ∈ (0, ∞), consider now the surface Σ := {x ∈ Hξ : |x| < r} and
note that if O := {x ∈ Rn−1 : |x | < r} then
P : O → Σ, P(x ) := A−1 (x , 0) for each x ∈ O, (4.7.43)
is a global C ∞ parametrization of Σ, satisfying |∂1 P × · · · × ∂n−1 P| = 1 at each point

in the set O. Consequently, (14.6.6) gives

Φ(x + ξ) dσ(x) = Φ(P(x ) + ξ) dx
x∈Hξ , |x|<r O

= Φ A−1 (x , 0) + A−1 en dx
O

= Ψ (x , 0) + en dx . (4.7.44)
x ∈Rn−1 , |x |<r
At this stage, (4.7.39) is clear from (4.7.42) and (4.7.44).

The power of Theorem 4.79 is most apparent by considering the consequence
discussed in Corollary 4.81 below, which sheds light on the boundary behavior
lim+ F ± (x , t) of functions of the following type:
t→0

F ± (x , t) := Φ(x − y , ± t)ϕ(y ) dy , ∀ (x , t) ∈ Rn+ , (4.7.45)
Rn−1
where Φ is as in the statement of Theorem 4.79 and ϕ is a Schwartz function on

Rn−1 (which is canonically identified with ∂Rn+ ). Functions of the form (4.7.45)
play a prominent role in partial differential equations and harmonic analysis as they
arise naturally in the treatment of boundary value problems via boundary integral
methods. We shall return to this topic in Section 11.6.
Corollary 4.81. Let the function Φ ∈ C 4 (Rn \{0}) be odd and positive homogeneous
of degree 1 − n, and assume that ϕ ∈ S(Rn−1 ). Then for every x ∈ Rn−1 one has

i
lim± Φ(x − y , t)ϕ(y ) dy = ± Φ(0 , 1)ϕ(x )
t→0 Rn−1 2

+ lim+ Φ(x − y , 0)ϕ(y ) dy . (4.7.46)
ε→0
y ∈Rn−1
|x −y |>ε
Proof. Given any ϕ ∈ S(Rn−1 ) and any x ∈ Rn−1 , write

lim± Φ(x − y , t)ϕ(y ) dy = lim± Φ(z , t)ϕ(x − z ) dz
t→0 Rn−1 t→0 Rn−1

= lim± Φ(·, t), ϕ(x − ·) (4.7.47)
t→0
i
= ± Φ(0 , 1)ϕ(x ) + P.V. Φ(·, 0), ϕ(x − ·)
2
then notice that

P.V. Φ(·, 0), ϕ(x − ·) = lim+ Φ(z , 0)ϕ(x − z ) dz
ε→0
z ∈Rn−1
|z |>ε

= lim+ Φ(x − y , 0)ϕ(y ) dy . (4.7.48)
ε→0

y ∈R n−1
|x −y |>ε
Now (4.7.46) follows from (4.7.47)–(4.7.48).
Remark 4.82. Consider Φ : R2 \ {(0, 0)} → C given by Φ(x, y) := x+iy 1

for all
(x, y) ∈ R \ {(0, 0)}. Then Φ is odd and homogeneous of degree −1. Moreover,
2
since under the canonical identification R2 ≡ C the function Φ becomes Φ(z) = 1z

for z ∈ C \ {0}, it follows that Φ(ξ) = 2πi · 1ξ for all ξ ∈ C \ {0} (for details, see
1) = 2π · 1 = −2π. Consequently,
Proposition 7.44). In particular, this yields Φ(0, i i
(4.7.1) becomes in this case
1 1
lim = ∓iπ δ + P.V. in S (R), (4.7.49)
ε→0+ x ± iε x
which is in agreement with Sokhotsky’s formula from (2.10.3).
Theorem 4.79 also suggests a natural procedure for computing the Fourier trans-
form of certain principal value distributions. While the latter topic has been treated
in Section 4.5, where the general formula (4.5.11) has been established, such a pro-
cedure remains of interest since the integral in (4.5.6) may not always be readily
computed from scratch. Specifically, we have the following result.
Corollary 4.83. Assume that the function Φ ∈ C 4 (Rn \ {0}) is odd and positive
homogeneous of degree 1 − n. Then
i
F P.V. Φ( ·, 0) = ∓ Φ(0 , 1) + lim± F Φ( ·, ε) in S (Rn−1 ), (4.7.50)
2 ε→0
where F denotes the Fourier transform in Rn−1 .

Proof. Formula (4.7.50) is a direct consequence of Theorem 4.79, the continuity of

F on S (Rn−1 ), and the fact that F δ(x ) = 1.
Here is an example implementing this procedure in a case that is going to be
useful later on, when discussing the Riesz transforms in Rn .
Proposition 4.84. For each j ∈ {1, . . . , n} we have (with F denoting the Fourier
transform in Rn )
xj iωn ξ j
F P.V. n+1 = − in S (Rn ). (4.7.51)
|x| 2 |ξ|
Proof. Fix some j ∈ {1, . . . , n} and consider the function defined by
xj
Φ(x, t) := n+1 , ∀ (x, t) ∈ Rn+1 \ {0}. (4.7.52)
|x|2 + t2 2
Note that the function Φ is C ∞ , odd, and positive homogeneous of degree −n in

Rn+1 \ {0}. Moreover, if hat denotes the Fourier transform in Rn+1 , from Corol-
lary 4.65 we have
ξj
η) = −iωn
Φ(ξ, in S (Rn+1 ), (4.7.53)
|ξ|2 + η2
1) = 0. Keeping this in mind, formula (4.7.50) (used

from which we see that Φ(0,
with n + 1 in place on n) then yields
xj
F P.V. n+1 = lim+ F x Φ(x, t) in S (Rn ), (4.7.54)
|x| t→0
where F x denotes the Fourier transform in the variable x in Rn . Next, recall the dis-
cussion at the end of Section 4.2 regarding partial Fourier transforms and compute
1
F x Φ(x, t) (ξ) = Fη−1 Φ(ξ, η) (t) = −iωn ξ j Fη−1 (t)
|ξ|2 + η2
1
= −iωn (2π)−1 ξ j Fη (t)
|ξ|2 + η2
iωn ξ j −|ξ||t|
=− e in S (Rn ), (4.7.55)
2 |ξ|
where the last equality makes uses of (4.2.21). Since by Lebesgue’s Dominated
Convergence Theorem

ξ j −|ξ||t| ξj
lim e = in S (Rn ), (4.7.56)
t→0 |ξ| |ξ|
formula (4.7.51) now follows from (4.7.54)–(4.7.56).
Remark 4.85. Alternatively, one may prove (4.7.51) by combining (4.3.25) and
(4.4.19). Indeed, if j ∈ {1, . . . , n} is fixed, identity (4.4.19) written for the function
Φ(x) := 1−n1
|x|−(n−1) , x ∈ Rn \ {0}, becomes
1 xj
∂ j |x|−(n−1) = P.V. n+1 in S (Rn ). (4.7.57)
1−n |x|
Hence, taking the Fourier transform of (4.7.57), then using (b) in Theorem 4.26,
then (4.3.25) with λ := n − 1, and formulas (14.5.2) and (14.5.2), we obtain
xj 1 i
F P.V. n+1 = F ∂ j |x|−(n−1) = ξ j F |x|−(n−1)
|x| 1−n 1−n
n

i 2π 2 Γ 2 ξ j
1
iωn ξ j
=− =− in S (Rn ). (4.7.58)
n − 1 Γ n−1 |ξ| 2 |ξ|
2
Corollary 4.83 may also be combined with Theorem 4.74 to obtain the following
result pertaining to certain limits of Fourier transforms of tempered distributions.
Corollary 4.86. Suppose that Φ ∈ C 4 (Rn \ {0}) is odd and positive homogeneous of
degree 1 − n. Then in S (Rn−1 ),

i
lim F Φ( ·, ε) (ξ ) = ± Φ(0 , 1) − Φ(θ, 0) log(i(ξ · θ)) dσ(θ) (4.7.59)
ε→0± 2 S n−2
where F denotes the Fourier transform in Rn−1 and S n−2 denotes the unit sphere
centered at the origin in Rn−1 .
Proof. This follows from Corollary 4.83, (4.5.11), and (4.5.6).

4.8 The Harmonic Poisson Kernel 177
4.8 The Harmonic Poisson Kernel
The goal of this section is to introduce and discuss the harmonic Poisson kernel.
Definition 4.87. Define the harmonic Poisson kernel P : Rn \ {0} → R by
setting
2 xn
P(x , xn ) := , ∀ x = (x , xn ) ∈ Rn \ {0}. (4.8.1)
ωn−1 |x|n
Furthermore, for each x ∈ Rn−1 set p(x ) := P(x , 1), i.e., consider
2 1
p(x ) := n , ∀ x ∈ Rn−1 .
ωn−1 1 + |x |2 2
(4.8.2)
In our next result we discuss the boundary behavior of a mapping P defined

below, taking Schwartz functions from Rn−1 into functions defined in Rn± , which in
partial differential equations is referred to as the harmonic double layer potential
operator.
Proposition 4.88. Given any ϕ ∈ S(Rn−1 ) and any x = (x , xn ) ∈ Rn with xn 0,
define (where P denotes the harmonic Poisson kernel)

(Pϕ)(x) := P(x − y , xn )ϕ(y ) dy
Rn−1

2 xn
= n2 ϕ(y ) dy . (4.8.3)
ωn−1 Rn−1 |x − y |2 + xn2
Then for each x ∈ Rn−1 one has
lim (Pϕ)(x , xn ) = ± ϕ(x ). (4.8.4)

xn →0±
Proof. The function P : Rn \ {0} → R from (4.8.1) is C ∞ , odd, and positive homo-
geneous of degree 1 − n. Moreover, Corollary 4.65 gives
= −2i ξn
P(ξ) in S (Rn ). (4.8.5)
|ξ|2
, 1) = −2i. Also, since P(x , 0) = 0 we have P.V. P(x , 0) = 0. At

Consequently, P(0
this point, Corollary 4.81 applies (with P playing the role of Φ) and yields (4.8.4).

Remark 4.89. When specializing Theorem 4.79 to the case Φ := P, with P defined
as in (4.8.1), we obtain (making use of (4.8.5))
lim P(x , xn ) = ± δ(x ) in S (Rn−1 ). (4.8.6)

xn →0±
This may be regarded as a higher dimensional generalization of the result presented

in Exercise 2.133, to which (4.8.6) reduces in the case when n = 2 and xn = ε > 0.
Among other things, the following proposition sheds light on the normalization
of the harmonic Poisson kernel introduced in (4.8.1).
Proposition 4.90. The function p defined in (4.8.2) satisfies the following proper-
ties:
- q n−1
(1) One has p ∈ L (R ) and
1≤q≤∞

p(x ) dx = 1. (4.8.7)
Rn−1
(2) For each t > 0 set
pt (x ) := t1−n p(x /t) for x ∈ Rn−1 . (4.8.8)

-
Then for each t ∈ (0, ∞) we have pt ∈ Lq (Rn−1 ) and its Fourier transform
1≤q≤∞
is

pt (ξ ) = e−t|ξ |
in S (Rn−1 ). (4.8.9)
(3) The family {pt }t>0 has the semigroup property, that is,
pt1 ∗ pt2 = pt1 +t2 ∀ t1 , t2 ∈ (0, ∞). (4.8.10)
Proof. That p ∈ Lq (Rn−1 ) for any q ∈ [1, ∞] is immediate from its expression. Fix
t ∈ (0, ∞) and let pt be as in part (2). Applying Exercise 2.26 (with n − 1 in place of
n, p in place of f , and t in place of 1/ j), we obtain

D (R )
n−1

pt (x ) −−−−−−+
→ c δ(x ) where c := p(x ) dx . (4.8.11)
t→0 Rn−1
On the other hand, as seen from (4.8.2) and (4.8.1),
pt (x ) = P(x , t) for each t > 0 and x ∈ Rn−1 , (4.8.12)
so (4.8.6) gives
pt (x ) → δ(x ) in S (Rn ) as t → 0+ . (4.8.13)
Now (4.8.7) follows by combining (4.8.13) with (4.8.11).

Moving on to the proof of part (2), observe first that pt ∈ Lq (Rn−1 ) for each
t > 0, if q ∈ [1, ∞], thus pt ∈ S (Rn−1 ) (as a consequence of (4.1.9)). In particular,
its Fourier transform in S (Rn−1 ) is meaningfully defined and equal to its Fourier
transform as a function in L1 (Rn−1 ) (by Remark (4.23)) and belongs to C 0 (Rn−1 ) (by
(3.1.3)).
In what follows we will be using partial Fourier transforms (cf. the discussion
at the end of Section 4.2) and the notation ξ ∈ Rn−1 and η ∈ R. The idea we will

pursue is to compute F x pt (x ) (ξ ) by making use of (4.8.12) and the fact that we
4.8 The Harmonic Poisson Kernel 179
, η), the Fourier transform of P in S (Rn ) (cf. (4.8.5)). With this in mind,
know P(ξ
for each ξ ∈ Rn−1 \ {0} and t ∈ R we write
η
F x P(x , t) (ξ ) = Fη−1 P(ξ , η) (t) = Fη−1 − 2i 2 (t)
|ξ | + η 2
η i
= 2i(2π)−1 Fη (t) = (−πi)(sgn t)e−|t| |ξ |
|ξ |2 +η 2 π

= (sgn t)e−|t| |ξ | . (4.8.14)
For the second equality in (4.8.14) we have used (4.8.5), for the third the fact that
Fη−1 g = (2π)−1 Fη g∨ for every g ∈ S (Rn ), while for the fourth we have used (4.2.32)
(applied with a := |ξ |). In particular, (4.8.14) implies that

F x pt (x ) (ξ ) = e−t|ξ | for ξ ∈ Rn−1 \ {0} and t > 0. (4.8.15)
Now (4.8.9) follows from (4.8.15) and the fact that pt ∈ C 0 (Rn−1 ) for every t > 0.
Regarding (4.8.10), fix t1 , t2 ∈ (0, ∞) and by using the property of the Fourier trans-
form singled out in part (3) of Remark 3.29 and (4.8.10) we obtain

F x pt1 ∗ pt2 (ξ ) = F x pt1 (ξ )F x pt2 (ξ )

= e−t1 |ξ | e−t2 |ξ | = e−(t1 +t2 )|ξ |

= F x pt1 +t2 ∀ ξ ∈ Rn−1 , ∀ t > 0. (4.8.16)
Now (4.8.10) follows from (4.8.16) by recalling that F x is an isomorphism on

S (Rn−1 ).
We wish to note that the harmonic Poisson kernel presented above plays a basic
role in the treatment of boundary value problems for the Laplacian Δ := ∂21 + · · · +
∂2n in the upper-half space. More specifically, for any given function ϕ ∈ S(Rn−1 )
(called the boundary datum), the Dirichlet problem
⎧
⎪
⎪
⎪ u ∈ C ∞ (Rn+ ),
⎪
⎪
⎪
⎨ Δu = 0 in Rn+ ,
⎪
⎪
⎪ (4.8.17)
⎪
⎪ ver
⎪
⎩ u n = ϕ on R ≡ ∂R+ ,
n−1 n
∂R+
has as a solution the function
u := Pϕ in Rn+ , (4.8.18)
i.e., (cf. (4.8.3))

2 xn
u(x) = n ϕ(y ) dy , x = (x , xn ) ∈ Rn+ . (4.8.19)
ωn−1 Rn−1
|x − y |2 + xn2 2
ver
In the last line of (4.8.17), the symbol u n stands for the “vertical limit” of u to the
∂R +
boundary of the upper-half space, understood at each point x ∈ Rn−1 as
ver
u n (x ) := lim+ u(x , xn ). (4.8.20)
∂R+ xn →0
To see why u as in (4.8.19) is a solution of (4.8.17) note that (4.8.1) implies u ∈

C ∞ (Rn+ ), that u has the limit to the boundary equal to ϕ based on (4.8.4), while the
fact that Δu = 0 in Rn+ follows from (4.8.19) by checking directly that, for each
n
y ∈ Rn−1 fixed, Δ xn |x − y |2 + xn2 − 2 = 0 for all (x , xn ) ∈ Rn+ .
Definition 4.91. The conjugate harmonic Poisson kernels are the map-
pings Q j : Rn \ {0} → R, j ∈ {1, . . . , n − 1}, defined by
2 xj
Q j (x , xn ) := , ∀ x = (x , xn ) ∈ Rn \ {0}, j ∈ {1, . . . , n − 1}. (4.8.21)
ωn−1 |x|n
Furthermore, for each x ∈ Rn−1 set q j (x ) := Q j (x , 1), that is,
2 xj
q j (x ) := , ∀ x ∈ Rn−1 , j ∈ {1, . . . , n − 1}.
ωn−1 1 + |x |2 2
n (4.8.22)
Proposition 4.92. The functions defined in (4.8.2) satisfy the following properties:
(1) For each j ∈ {1, . . . , n − 1} and each p ∈ (1, ∞) one has q j ∈ L p (Rn−1 ).
(2) For each t > 0 set
(q j )t (x ) := t1−n q j (x /t) for x ∈ Rn−1 , j ∈ {1, . . . , n − 1}. (4.8.23)
Then for each j ∈ {1, . . . , n − 1}, each t ∈ (0, ∞), and each p ∈ (1, ∞), we have
(q j )t ∈ L p (Rn−1 ) and its Fourier transform is
ξ j −t|ξ |

(q
j )t (ξ ) = −i e in S (Rn−1 ). (4.8.24)
|ξ |
(3) The following identity holds:
d n−1
pt (x ) = − ∂ j (q j )t (x ) ∀ t ∈ (0, ∞), ∀ x ∈ Rn−1 , (4.8.25)
dt j=1
where pt is as in (4.8.23).
Proof. The claim in (1) is an immediate consequence of (4.8.22). To prove (2), fix
j ∈ {1, . . . , n − 1}. Using (4.8.23) we obtain
2 xj
(q j )t (x ) = = Q j (x , t) (4.8.26)
ωn−1 (t2
n
+ |x |2 ) 2
4.9 Singular Integral Operators 181
for each x ∈ Rn−1 and each t ∈ (0, ∞).

As in the proof of Proposition 4.90, we will be using partial Fourier transforms
(cf. the discussion at the end of Section 4.2) and the notation ξ ∈ Rn−1 and η ∈ R.

The idea we will pursue is to compute F x (q j )t (x ) (ξ ) by making use of (4.8.26)
and the fact that, by (4.3.31), the Fourier transform of Q j is
ξj

Q j (ξ , η) = −2i in S (Rn ). (4.8.27)
|ξ |2+ η2
Hence, for each t ∈ (0, ∞) we may write
ξj
F x (q j )t (x ) (ξ ) = Fη−1 Q j (ξ , η) (t) = Fη−1 − 2i (t)
|ξ |2+η 2
1 iξ j π −t |ξ |
= −2i(2π)−1 ξ j Fη (t) = − e
|ξ |2 + η2 π |ξ |
iξ j −t |ξ |
=− e in S (Rn−1 ). (4.8.28)
|ξ |
For the third equality in (4.8.28) we used the fact that Fη−1 g = (2π)−1 Fη g∨ for every
g ∈ S (Rn ), while for the fourth we used (4.2.21) (applied with a := |ξ |). This
completes the proof of (2). Finally, (4.8.25) follows by a direct computation based
on the Chain Rule.
4.9 Singular Integral Operators
In Example 4.69 we have already encountered the principal value tempered distri-
xj
butions P.V. |x|n+1 , for j ∈ {1, . . . , n}. The operators R j , j ∈ {1, ..., n}, defined by
convolving with these distributions, i.e.,
xj
R j ϕ := P.V. n+1 ∗ ϕ, ∀ ϕ ∈ S(Rn ), (4.9.1)
|x|
are called the Riesz transforms in Rn . In the particular case when n = 1 the
corresponding operator, i.e.,
1
Hϕ := P.V. ∗ ϕ, ∀ ϕ ∈ S(R), (4.9.2)
x
is called the Hilbert transform. These operators play a fundamental role in
harmonic analysis and here the goal is to study a larger class of operators containing
the aforementioned examples. We begin by introducing this class.
The format of Proposition 4.70 suggests making the following definition.
Definition 4.93. For eachfunction Θ ∈ C 0 (Rn \ {0}) that is positive homogeneous of
degree −n and such that S n−1 Θ(ω) dσ(ω) = 0, define the singular integral
operator

(T Θ ϕ)(x) := lim+ Θ(x − y)ϕ(y) dy, ∀ x ∈ Rn , ∀ ϕ ∈ S(Rn ). (4.9.3)
ε→0 |x−y|≥ε
Proposition 4.70 ensures that the above definition is meaningful. Moreover, for
the class of singular integral operators just defined, the following result holds (fur-
ther properties are deduced in Theorem 4.100).
Proposition 4.94. Let Θ be a function satisfying the conditions in (4.4.1) and con-
sider the singular integral operator T Θ associated with Θ as in (4.9.3). Then
T Θ : S(Rn ) −→ S (Rn ) is linear and sequentially continuous. (4.9.4)
Moreover, for each ϕ ∈ S(Rn ) we have

T Θ ϕ ∈ C ∞ (Rn ) and T Θ ϕ = P.V. Θ ∗ ϕ in S (Rn ), (4.9.5)
as well as

T
Θϕ =
ϕ F P.V. Θ =
ϕ mΘ in S (Rn ), (4.9.6)
where mΘ is as in (4.5.6).
Proof. All claims are consequences of Definition 4.93, Proposition 4.70, part (e) in
Theorem 4.19, part (a) in Theorem 4.35, and (4.5.11).
Remark 4.95. In the harmonic analysis parlance, a mapping
T : S(Rn ) → S (Rn ) (4.9.7)
with the property that there exists m ∈ S (Rn ) such that Tϕ = ϕ m in S (Rn ) for
every ϕ ∈ S(R ) is called a multiplier (and the tempered distribution m is
n
referred to as the symbol of the multiplier). Note that any multiplier T

is necessarily a linear and sequentially bounded mapping from S(Rn ) into S (Rn ).
Also, with this piece of terminology, for any function Θ as in (4.4.1) and mΘ as in
(4.5.6) we have

the operator T Θ is a multiplier with symbol F P.V. Θ = mΘ . (4.9.8)
To set the stage for the subsequent discussion, we recall that given a linear and
bounded map (also referred to as a bounded operator) T : L2 (Rn ) → L2 (Rn ), its
adjoint is the unique map T ∗ : L2 (Rn ) → L2 (Rn ) with the property that

(T f )(x)g(x) dx = f (x)(T ∗ g)(x) dx, ∀ f, g ∈ L2 (Rn ). (4.9.9)
Rn Rn
Exercise 4.96. Let T : L2 (Rn ) → L2 (Rn ) be a linear and bounded operator which is
a multiplier with a bounded symbol, i.e., there exists m ∈ L∞ (Rn ) with the property
that Tϕ = ϕ m in S (Rn ) (hence also in L2 (Rn )) for every ϕ ∈ S(Rn ). Prove that its
∗
adjoint T is also a multiplier with symbol m.
Hint: Use (4.9.9) and Parseval’s identity (3.2.32).

Some of the properties of the symbol F P.V. Θ are discussed in Theorem 4.74.
An explicit computation of this Fourier transform has been done in the case cor-
xj
responding to Θ(x) := |x|n+1 for some j ∈ {1, . . . , n} (see (4.7.51)). The fact that
such functions appear in the definition of the Riesz transforms in Rn (cf. (4.9.1)),
warrants revisiting these operators.
Theorem 4.97. Consider the Riesz transforms, originally introduced as in (4.9.1).

Then for each j ∈ {1, . . . , n} the following properties hold:
(a) The operator
R j : S(Rn ) → S (Rn ) is well defined,
(4.9.10)
linear, and sequentially continuous.
(b) For each ϕ ∈ S(Rn ) we have R j ϕ ∈ C ∞ (Rn ) and R j may be expressed as the
singular integral operator

xj − yj
(R j ϕ)(x) = lim+ ϕ(y) dy, ∀ x ∈ Rn . (4.9.11)
ε→0 |x−y|≥ε |x − y|n+1
(c) The j-th Riesz transform is a multiplier with symbol given by the function
ξ
m j (ξ) := − iω2n |ξ|j ∈ L∞ (Rn ). That is, for each ϕ ∈ S(Rn ),
iωn ξ j
R j ϕ(ξ) = −
ϕ(ξ) in S (Rn ). (4.9.12)
2 |ξ|
(d) For each ϕ ∈ S(Rn ), we have that R j ϕ originally viewed as a tempered distribu-
tion belongs to the subspace L2 (Rn ) of S (Rn ), and
R j ϕ L2 (Rn ) ≤ (ωn /2) ϕ L2 (Rn ) . (4.9.13)
(e) The j-th Riesz transform R j , originally considered as in (4.9.10) extends, by den-
sity to a linear and bounded operator
R j : L2 (Rn ) −→ L2 (Rn ) (4.9.14)
and the operator R j is skew-symmetric (i.e., R∗j = −R j ) in this context. In addi-

tion,
iωn ξ j
R j f (ξ) = − f (ξ) a.e. in Rn , ∀ f ∈ L2 (Rn ). (4.9.15)
2 |ξ|
(f) For every k ∈ {1, . . . , n}, we have
R j Rk = Rk R j as operators on L2 (Rn ), (4.9.16)
and

n
ωn 2
R2k = − 2 I as operators on L2 (Rn ), (4.9.17)
k=1
where I denotes the identity operator on L2 (Rn ).

Proof. The claims in (a)–(b) are immediate consequence of (4.9.1)–(4.9.2) and
Proposition 4.70, upon recalling the discussion in Example 4.69. Also, formula
(4.9.12) follows from (4.9.1), (4.9.6), and Proposition 4.84.
Turning our attention to (d), fix an index j ∈ {1, . . . , n} along with two arbi-
trary functions ϕ, ψ ∈ S(Rn ). Based on Exercise 3.27, (4.2.2), (4.9.12), Cauchy–
Schwarz’s inequality, and (3.2.28), we may then write

ωn ξ j
R j ϕ, ψ = (2π)−n R ψ∨ = (2π)−n
j ϕ, ψ∨
ϕ,

2 |ξ|

ωn ξj
= (2π)−n ϕ(ξ)
ψ(−ξ) dξ
2 Rn |ξ|

ωn
≤ (2π)−n ϕ(ξ)
ψ(−ξ) dξ
2 Rn
ωn
≤ (2π)−n ϕ L2 (Rn )
ψ L2 (Rn )
2
ωn
= 2 ϕ L (R ) ψ L (R ) .
2 n 2 n (4.9.18)
This computation may be summarized by saying that, for each ϕ ∈ S(Rn ) fixed, the
linear functional
Λϕ : S(Rn ) → C, Λϕ (ψ) := R j ϕ, ψ , ∀ ψ ∈ S(Rn ), (4.9.19)

ω
satisfies Λϕ (ψ) ≤ 2n ϕ L2 (Rn ) ψ L2 (Rn ) for every ψ ∈ S(Rn ). Given that S(Rn ) is a
dense subspace of L2 (Rn ), it follows that the mapping Λϕ from (4.9.19) has a unique
extension Λ ϕ : L2 (Rn ) → C satisfying

Λϕ (g) ≤ ωn ϕ L2 (Rn ) g L2 (Rn ) for every g ∈ L2 (Rn ). (4.9.20)
2
According to Riesz’s representation theorem for such functionals, there exists a

unique fϕ ∈ L2 (Rn ) satisfying the following two properties:

ϕ (g) =
Λ fϕ (x)g(x) dx for every g ∈ L2 (Rn ), (4.9.21)
Rn
and
ωn
fϕ L2 (Rn ) ≤ 2 ϕ L (R ) .
2 n (4.9.22)

Since L (R ) ⊆ S (R ) (cf. (4.1.9)), it follows that fϕ may be regarded as a tempered
2 n n
distribution. At this stage we make the claim that R j ϕ = fϕ as tempered distributions.

Indeed, for every ψ ∈ S(Rn ) we have

ϕ (ψ) =
R j ϕ, ψ = Λϕ (ψ) = Λ fϕ (x)ψ(x) dx = fϕ , ψ , (4.9.23)
Rn
from which the claim follows. With this in hand, (4.9.13) now follows from (4.9.22),
finishing the proof of part (d). In turn, estimate (4.9.13) and a standard density argu-
ment give that R j extends to a linear and bounded operator in the context of (4.9.14).
The next item in part (e) is showing that R j in (4.9.14) is skew-symmetric. To
this end, first note that, given any ϕ, ψ ∈ S(Rn ), by reasoning much as in (4.9.18)
we obtain
∨ ωn ξ j ∨
R j ϕ, ψ = (2π)−n R j ϕ, ψ = −i(2π)−n ϕ,
ψ
2 |ξ|

ωn ξj dξ
= −i(2π)−n
ϕ(ξ)ψ(−ξ)
2 Rn |ξ|

ωn ξj
= i(2π)−n ψ(ξ)
ϕ(−ξ) dξ
2 Rn |ξ|
ωn ξ j ∨ ∨
= i(2π)−n ϕ = −(2π)−n R
ψ, j ψ,
ϕ
2 |ξ|
= −R j ψ, ϕ = −R j ψ, ϕ , (4.9.24)
where the last step uses the fact that, as seen from (4.9.11), R j ψ = R j ψ. Since
all functions involved are (by part (d)) in L2 (Rn ), the distributional pairings may be
interpreted as pairings in L2 (Rn ). With this in mind, the equality of the most extreme
sides of (4.9.24) reads

(R j ϕ)(x)ψ(x) dx = − ϕ(x)(R j ψ)(x) dx. (4.9.25)
Rn Rn
By density and part (d), we deduce from (4.9.25) that

(R j f )(x)g(x) dx = − f (x)(R j g)(x) dx, ∀ f, g ∈ L2 (Rn ). (4.9.26)
Rn Rn
In light of (4.9.9), this shows that R∗j = −R j as mappings on L2 (Rn ), as wanted. To

complete the treatment of part (e), there remains to observe that, collectively, the
boundedness of R j and of the Fourier transform on L2 (Rn ), (4.9.12), and the density
of the Schwartz class in L2 (Rn ), readily yield (4.9.15) via a limiting argument.
In turn, given k ∈ {1, . . . , n}, for each f ∈ L2 (Rn ) formula (4.9.15) allows us to
compute
iωn ξ j ωn 2 ξ j ξk
F R j (Rk f ) = − Rk f = −
f in L2 (Rn ). (4.9.27)
2 |ξ| 2 |ξ|2
2 ξ ξ
A similar computation also gives that F Rk (R j f ) = − ω2n |ξ|j 2k f in L2 (Rn ). From
these, (4.9.16) follows upon recalling (cf. (3.2.31)) that the Fourier transform is an
isomorphism of L2 (Rn ).
Finally, as regards (4.9.17), for any function f ∈ L2 (Rn ), formula (4.9.27) (with
j = k) permits us to write

n n
ωn 2 n
ξk2
F R2k f = F Rk (Rk f ) = −
f
k=1 k=1
2 k=1
|ξ| 2
ωn 2
=−
f in L2 (Rn ), (4.9.28)
2
.
n ξk2
where the last equality uses the fact that |ξ|2
= 1 a.e. in Rn . Now identity (4.9.17)
k=1
follows from (4.9.28) and (3.2.31).
Corollary 4.98. Let j ∈ {1, . . . , n − 1} and t ∈ (0, ∞), and recall the functions pt and
(q j )t from (4.8.8) and (4.8.23), respectively. Then
2
(q j )t = R j pt in L2 (Rn−1 ), (4.9.29)
ωn−1
where R j in (4.9.29) is the j-th Riesz transform from Theorem 4.97 corresponding
to n replaced by n − 1.
Proof. By (3.2.31), it suffices to check identity (4.9.29) on the Fourier transform
side. That the latter holds is an immediate consequence of part (2) in Proposi-
tion 4.92, (4.9.15), and part (2) in Proposition 4.90.
In the one dimensional setting, Theorem 4.97 yields the following type of infor-
mation about the Hilbert transform.
Corollary 4.99. The Hilbert transform defined in (4.9.2) satisfies the following
properties:
(a) The operator
H : S(R) → S (R) is well defined,
(4.9.30)
(b) For each ϕ ∈ S(R) we have Hϕ ∈ C ∞ (R) and H may be expressed as the singular
integral operator

1
(Hϕ)(x) = lim+ ϕ(y) dy, ∀ x ∈ R. (4.9.31)
ε→0 |x−y|≥ε x − y
(c) The Hilbert transform is a multiplier with symbol m(ξ) := −iπ (sgn ξ) belonging
to L∞ (R). In other words, for each ϕ ∈ S(R),

Hϕ(ξ) ϕ(ξ) in S (R).
= −iπ (sgn ξ) (4.9.32)
(d) For each ϕ ∈ S(R), we have that Hϕ originally viewed as a tempered distribution
belongs to the subspace L2 (R) of S (R), and
Hϕ L2 (Rn ) ≤ π ϕ L2 (Rn ) . (4.9.33)
(e) The Hilbert transform H, originally considered as in (4.9.30) extends, by density

to a linear and bounded operator
H : L2 (R) −→ L2 (R) (4.9.34)
and the operator H is skew-symmetric (i.e., H ∗ = −H) in this context. In addition,

f (ξ) = −iπ (sgn ξ)
H f (ξ) a.e. in R, ∀ f ∈ L2 (R). (4.9.35)
(f) In the context of (4.9.34),
H 2 = −π2 I in L2 (R), (4.9.36)
where I denotes the identity operator on L2 (R).

Proof. This corresponds to Theorem 4.97 in the case when n = 1.
The Riesz transforms are prototypes of for the more general class of operators
introduced in Definition 4.93 and most of the properties deduced for the Riesz trans-
forms in Theorem 4.97 have natural counterparts in this more general setting.
Theorem 4.100. Assume that Θ is a function satisfying the conditions in (4.4.1) and
let T Θ be the singular integral operator associated with Θ as in (4.9.3). Then the
following statements are true.
(a) The operator
T Θ : S(Rn ) → S (Rn ) is well defined,
(4.9.37)
(b) For each ϕ ∈ S(Rn ) we have T Θ ϕ ∈ C ∞ (Rn ) and the singular integral operator
T Θ may be expressed as the convolution

T Θ ϕ = P.V. Θ ∗ ϕ ∀ ϕ ∈ S(Rn ). (4.9.38)
(c) The function mΘ , associated with Θ as in (4.5.6), is the symbol of the multiplier
T Θ . This means that for each ϕ ∈ S(Rn ),
T ϕ(ξ) in S (Rn ).
Θ ϕ(ξ) = mΘ (ξ) (4.9.39)
(d) For each ϕ ∈ S(Rn ), we have that T Θ ϕ originally viewed as a tempered distribu-
tion belongs to the subspace L2 (Rn ) of S (Rn ), and
T Θ ϕ L2 (Rn ) ≤ Cn Θ L∞ (S n−1 ) ϕ L2 (Rn ) , (4.9.40)
where the constant Cn ∈ (0, ∞) is as in (4.5.9).

(e) The operator T Θ , originally considered as in (4.9.37) extends, by density to a

linear and bounded operator
T Θ : L2 (Rn ) −→ L2 (Rn ). (4.9.41)
Moreover, in this context,
T
Θ f (ξ) = mΘ (ξ) f (ξ) a.e. in Rn , ∀ f ∈ L2 (Rn ), (4.9.42)
and the operator T Θ satisfies

∗
TΘ = T Θ∨ in L2 (Rn ). (4.9.43)
In particular, if Θ is odd and real-valued, then the operator T Θ is skew-symmetric

(i.e., T Θ ∗ = −T Θ ), while if Θ is even and real-valued then the operator T Θ is

self-adjoint (i.e., T Θ ∗ = T Θ ).
Proof. Parts part (a)–(b) are contained in Proposition 4.94, while part (c) follows
from (4.9.6) and Theorem 4.74. Part (d) is justified by reasoning as in the proof of
part (d) in Theorem 4.97, keeping in mind (4.5.8). As for part (e), in a first stage we
note that, for any ϕ, ψ ∈ S(Rn ), by reasoning much as in (4.9.24) while relying on
(4.9.39) and (4.5.10), we obtain

T Θ ϕ, ψ = (2π)−n T ∨ = (2π)−n m ∨
Θ ϕ, ψ Θ ϕ, ψ

= (2π) −n
mΘ (ξ) dξ
ϕ(ξ)ψ(−ξ)
Rn

= (2π)−n mΘ (−ξ) ψ(ξ)
ϕ(−ξ) dξ
Rn

= (2π)−n mΘ∨ ϕ∨ = (2π)−n T
ψ, ϕ∨
Θ∨ ψ,
= T Θ∨ ψ, ϕ = T Θ∨ ψ, ϕ , (4.9.44)
where the last step uses the fact that, as seen from (4.9.3), T Θ∨ ψ = T Θ∨ ψ. Passing
from (4.9.44) to (4.9.43) may now be done by arguing as in the proof of part (e) of
Theorem 4.97 (a pattern of reasoning which also gives all remaining claims in the
current case).
We conclude this section with a discussion of the following result of basic impor-
tance from Calderón–Zygmund theory (originally proved in [7]; for a more timely
presentation see, e.g., [26], [68], [69]).
Theorem 4.101. Let Θ be a function satisfying (4.4.1) and, in analogy with (4.9.3),
given some p ∈ (1, ∞) set

(T Θ f )(x) := lim+ Θ(x − y) f (y) dy, x ∈ Rn , f ∈ L p (Rn ). (4.9.45)
ε→0 |x−y|≥ε
4.10 Derivatives of Volume Potentials 189
Then for every f ∈ L p (Rn ) we have that (T Θ f )(x) exists for almost every x ∈ Rn and
T Θ f L p (Rn ) ≤ CΘ f L p (Rn ) , (4.9.46)
where CΘ is a finite positive constant independent of f .
Compared with Theorem 4.100, a basic achievement of Calderón and Zygmund

is allowing functions from L p (Rn ) in lieu of S(Rn ). While the estimate in (4.9.46)
for p = 2 is essentially contained in part (d) of Theorem 4.100, the fact that for each
f ∈ L2 (Rn ) the limit defining (T Θ f )(x) as in (4.9.45) exists for almost every x ∈ Rn
is a bit more subtle. To elaborate on this issue, recall that we have already proved
that the limit in question exists at every point when f is a Schwartz function (cf.
Proposition 4.70). Starting from this, one can pass to arbitrary functions in L2 (Rn )
granted the availability of additional tools from harmonic analysis, including the
boundedness in L2 (Rn ) of the so-called maximal operator

T ∗ f (x) := sup Θ(x − y) f (y) dy for each x ∈ Rn . (4.9.47)
ε>0 |x−y|≥ε
Having dealt with (4.9.46) in the case p = 2, its proof for p ∈ (1, ∞) proceeds as
follows. In a first step, a suitable version of (4.9.46) is established for p = 1 which,
when combined with the case p = 2 already treated yields (via a technique called
interpolation) (4.9.46) for p ∈ (1, 2). In a second step, one uses duality to handle the
case p ∈ (2, ∞). There are two important factors at play here: that the dual of L p (Rn )

with p ∈ (1, 2) is L p (Rn ) with p = p−1 p
∈ (2, ∞), and the formula for the adjoint of
T Θ we deduced in (4.9.43).
In particular, Theorem 4.101 gives that, given p ∈ (1, ∞), the j-th Riesz transform
R j , j ∈ {1, . . . , n}, originally considered as in (4.9.10) extends by density so that
R j : L p (Rn ) −→ L p (Rn ) is linear and bounded, (4.9.48)
and for each f ∈ L p (Rn ) we have

xj − yj
(R j f )(x) = lim+ f (y) dy for a.e. x ∈ Rn . (4.9.49)
|x−y|≥ε |x − y|
ε→0 n+1
Of course, the same type of result is true for the Hilbert transform on the real line.
4.10 Derivatives of Volume Potentials
One basic integral operator in analysis is the so-called Newtonian potential,

given by

−1 1
(NΩ f )(x) := f (y) dy, ∀ x ∈ Rn , (4.10.1)
(n − 2)ωn−1 Ω |x − y|n−2
where Ω is an open set in Rn , n ≥ 3, and f ∈ L∞ (Ω) with bounded support. In this

regard, an important issue is that of computing ∂ j ∂k NΩ f in the sense of distributions
in Rn , where j, k ∈ {1, . . . , n}. First, we will show (cf. Theorem 4.105) that NΩ f ∈
C 1 (Rn ) and we have

1 xk − yk
∂k (NΩ f )(x) = f (y) dy, ∀ x ∈ Rn . (4.10.2)
ωn−1 Ω |x − y|n

Hence, ∂k (NΩ f )(x) = Ω Φ(x − y) f (y) dy where Φ(z) := ωn−1 1 zk
|z|n for each point
z ∈ R \ {0}. Note that for x fixed, Φ(x − ·) is locally integrable, though this is
n

no longer the case for (∂ j Φ)(x − ·). This makes the job of computing ∂ j ∂k (NΩ f )
considerably more subtle. There is a good reason for this since, as it turns out, the
latter distributional derivative involves (as we shall see momentarily) singular inte-
gral operators. We note that the function Φ considered above is C ∞ and positive
homogeneous of degree 1 − n in Rn \ {0}. These are going to be key features in our
subsequent analysis.
Our most general results encompassing the discussion about the Newtonian
potential introduced above are contained in Theorem 4.104 and Theorem 4.105.
We begin by first considering the case when Ω = Rn and f is a Schwartz function.
Proposition 4.102. Let Φ ∈ C 1 (Rn \ {0}) be a function that is positive homogeneous

of degree 1 − n and let f ∈ S(Rn ). Then Rn Φ(x − y) f (y) dy is differentiable in Rn
and for each j ∈ {1, . . . , n} and at every x ∈ Rn we have

∂x j Φ(x − y) f (y) dy = Φ(ω)ω j dσ(ω) f (x) (4.10.3)
Rn S n−1

+ lim+ (∂ j Φ)(x − y) f (y) dy.
ε→0 |y−x|≥ε
Proof. The key ingredient in the proof of (4.10.3) is formula (4.4.19). To see how
(4.4.19) applies, first note that by using Exercise 4.53 we have for each R > 0,

Φ(x − y) f (y) dy (4.10.4)
Rn

| f (y)|
≤ Φ L∞ (S n−1 ) dy
Rn |x − y|n−1

dy
≤ Φ L∞ (S n−1 ) f L∞ (Rn )
|y−x|≤R |x − y|n−1

dy
+ Φ L∞ (S n−1 ) (1 + |y|)2 f L∞ (Rn ) < ∞,
|y−x|>R (1 + |y|2 )|x − y|n−1

thus Rn
Φ(x − y) f (y) dy is well defined. Second, since

Φ(x − y) f (y) dy = Φ(y) f (x − y) dy ∀ x ∈ Rn , (4.10.5)
Rn Rn

we have that Rn
Φ(x − y) f (y) dy is differentiable and

∂x j Φ(x − y) f (y) dy = Φ(y)(∂ j f )(x − y) dy ∀ x ∈ Rn . (4.10.6)
Rn Rn
Third, for each x ∈ Rn , with t x as in (2.8.42) (and recalling (3.2.22)), we may write

Φ(x − y)ϕ(y) dy = t x (Φ∨ ), ϕ , ∀ ϕ ∈ S(Rn ). (4.10.7)
Rn
Now fix j ∈ {1, . . . , n}. Then for x ∈ Rn we have

∂x j Φ(x − y) f (y) dy (4.10.8)
Rn
= t x (Φ∨ ), ∂ j f = Φ∨ , t−x (∂ j f ) = Φ∨ , ∂ j [t−x f ] = −∂ j [Φ∨ ], t−x f

=− ω j Φ∨ (ω) dσ(ω) δ, t−x f − P.V. ∂ j (Φ∨ ) , t−x f
S n−1

= f (x) ω j Φ(ω) dσ(ω) + lim+ (∂ j Φ)(−y) f (x + y) dy
S n−1 ε→0 |y|≥ε

= f (x) ω j Φ(ω) dσ(ω) + lim+ (∂ j Φ)(x − y) f (y) dy.
S n−1 ε→0 |y|≥ε
The first equality in (4.10.8) uses (4.10.6) and (4.10.7), the fifth uses (4.4.19), the
sixth the fact that ∂ j (Φ∨ ) = −(∂ j Φ)∨ , and the last one a suitable change of variables.
This completes the proof of the corollary.
Next, we present a version of Proposition 4.102 when the function f is lacking
any type of differentiability properties. Here we make use of the basic Calderón–
Zygmund result recorded in Theorem 4.101.

degree 1 − n and let f ∈ L p (Rn ) for some p ∈ (1, n). Then Φ ∗ f ∈ Lloc
1
(Rn ) and for
each j ∈ {1, . . . , n} we have T ∂ j Φ f ∈ L p (Rn ) and

∂ j (Φ ∗ f ) = Φ(ω)ω j dσ(ω) f + T ∂ j Φ f in D (Rn ), (4.10.9)
S n−1
where T ∂ j Φ is the operator from (4.9.46) with ψ replaced by ∂ j Φ.
Proof. Fix p ∈ (1, n), f ∈ L p (Rn ), and R ∈ (0, ∞). Then we write

| f (y)| 1
dy dx = | f (y)| dx dy
B(0,R) Rn |x − y|n−1 |y|≤2R B(0,R) |x − y|n−1

1
+ | f (y)| dx dy
|y|>2R B(0,R) |x − y|n−1
=: I + II. (4.10.10)
If y ∈ B(0, 2R) and x ∈ B(0, R), then |x − y| ≤ 3R, thus

1
I≤ | f (y)| dx dy
|y|≤2R |x−y|≤3R |x − y|n−1
1 1
≤ dz |B(0, 2R)|1− p f L p (B(0,2R) < ∞, (4.10.11)
|z|≤3R |z|
n−1
where the second inequality in (4.10.11) uses Hölder’s inequality. Also, whenever
x ∈ B(0, R) and y ∈ Rn \ B(0, 2R) we have
|y|
|y| ≤ |y − x| + |x| ≤ |y − x| + R ≤ |y − x| + , (4.10.12)
2
which implies |y − x| ≥ |y|/2. Using this, II is estimated as

1
II ≤ 2n−1 | f (y)| dx dy
|y|>2R |x|≤R |y| n−1

| f (y)|
≤ 2 |B(0, R)|
n−1
dy
|y|>2R |y|
n−1
dy 1/p
≤ 2n−1 |B(0, R)| f L p (Rn ) < ∞, (4.10.13)
|y|>2R |y|
(n−1)p
where p := p−1 p
is the Hölder conjugate exponent for p. In the third inequality in
(4.10.13) the assumption p ∈ (1, n) has been used to ensure that (n − 1)p > n.
A combination of (4.10.10), (4.10.11), and (4.10.13) yields the following con-
clusion: for every R ∈ (0, ∞) there exists some finite positive constant C, depending
on R, n, and p, such that

| f (y)|
dy dx ≤ C f L p (Rn ) , ∀ f ∈ L p (Rn ). (4.10.14)
B(0,R) Rn |x − y|
n−1
In turn, (4.10.14) entails several useful conclusions. Recall that the assumptions on
Φ imply |Φ(x)| ≤ Φ L∞ (S n−1 ) |x|1−n for each x ∈ Rn \ {0} (cf. Exercise 4.53). The first
conclusion is that for each f ∈ L p (Rn ) and each R ∈ (0, ∞),

|Φ(x − y)|| f (y)| dy < ∞ for a.e. x ∈ B(0, R). (4.10.15)
Rn
This shows that (Φ ∗ f )(x) is well defined for a.e. x ∈ Rn . Second, from what we
have just proved and (4.10.14) we may conclude that Φ ∗ f ∈ Lloc 1
(Rn ).
∞
Next, recall that C0 (R ) is dense in L (R ) for each p ∈ (1, ∞). As such there
n p n
exists a sequence { fk }k∈N of function in C0∞ (Rn ) with the property that lim fk = f in
k→∞
L p (Rn ). Based on (4.10.14) we may conclude that the sequence {Φ∗ fk }k∈N converges
in L p (Rn ) to Φ ∗ f . Therefore, by Exercise 2.25,
D (Rn )
Φ ∗ fk −−−−−→ Φ ∗ f. (4.10.16)
k→∞
Moving on, fix j ∈ {1, . . . , n} and note that, by Exercise 4.51, ∂ j Φ is positive
homogeneous of degree −n. The function Φ ∗ fk belongs to C ∞ (Rn ) and Proposi-
tion 4.102 applies and implies that

∂ j Φ ∗ fk = Φ(ω)ω j dσ(ω) fk + T ∂ j Φ fk , ∀ k ∈ N, (4.10.17)
S n−1
pointwise in Rn . In particular, from (4.10.17) we infer that T ∂ j Φ fk ∈ C ∞ (Rn ) for all

k ∈ N, and that the equality in (4.10.17) also holds in D (Rn ). Since Theorem 4.101
L p (Rn )
gives that T ∂ j Φ fk −−−−−→ T ∂ j Φ f , based on Exercise 2.25 it follows that
k→∞
D (Rn ) D (Rn )
T ∂ j Φ fk −−−−−→ T ∂ j Φ f and fk −−−−−→ f. (4.10.18)
k→∞ k→∞
Since an immediate consequence of (4.10.16) is that

D (Rn )
∂ j (Φ ∗ fk ) −−−−−→ ∂ j (Φ ∗ f ), (4.10.19)
k→∞
combining this, (4.10.18), and the fact that (4.10.17) holds in D (Rn ), we obtain
(4.10.9).
We are prepared to state and prove our most general results regarding distribu-
tional derivatives of volume potentials. In particular, the next two theorems contain
solutions to the questions formulated at the beginning of this section.

degree 1−n. Consider
a measurable set Ω ⊆ Rn and assume that f ∈ L p (Ω) for some
p ∈ (1, n). Then Ω Φ(x − y) f (y) dy is absolutely convergent for a.e. point x ∈ Rn
1
and belongs to Lloc (Rn ) as a function of the variable x. Moreover, if f denotes the
extension of f by zero to Rn , then for each j ∈ {1, . . . , n} we have

∂x j Φ(x − y) f (y) dy = Φ(ω)ω j dσ(ω) f (x)
Ω S n−1

+ lim+ (∂ j Φ)(x − y) f (y) dy (4.10.20)
ε→0 y∈Ω\B(x,ε)
where the derivative in the left-hand side is taken in D (Rn ) and the equality is also
understood in D (Rn ).
In particular, formula (4.10.20) holds for any function f belonging to some
Lq (Ω), with q ∈ (1, ∞), that vanishes outside of a measurable subset of Ω of finite
measure.
Proof. The main claim in the statement follows by applying Theorem 4.103 to the
function ⎧
⎪
⎨ f in Ω,
⎪

f := ⎪ (4.10.21)
⎪
⎩ 0 in Rn \ Ω,
upon observing that f ∈ L p (Rn ). Moreover, if f is as in the last claim in the statement
then Hölder’s inequality may be invoked to show that f belongs to L p (Ω) for some
p ∈ (1, n).
Recall that if a ∈ R, the integer part of a is denoted by a and is by definition the
largest integer that is less than or equal to a. To state the next theorem we introduce
⎧
⎪
⎨ a if a Z,
⎪
a := ⎪
⎪ (4.10.22)
⎩ a − 1 if a ∈ Z.
Hence, a is the largest integer strictly less than a (thus, in particular, a < a for
every a ∈ R).
Theorem 4.105. Let Φ ∈ C ∞ (Rn \ {0}) be a function that is positive homogeneous

of degree m ∈ R where m > −n. Define the generalized volume potential
∞
associated with Φ by setting for each f ∈ Lcomp (Rn )

(ΠΦ f )(x) := Φ(x − y) f (y) dy, ∀ x ∈ Rn . (4.10.23)
Rn
∞
Then for each f ∈ Lcomp (Rn ) one has ΠΦ f ∈ C m+n (Rn ) and
∂α (ΠΦ f ) = Π∂α Φ f pointwise in Rn ,

(4.10.24)
for each α ∈ Nn0 with |α| ≤ m + n .
∞
Moreover, if α ∈ Nn0 is such that |α| = m + n, then for each f ∈ Lcomp (Rn ), the

distributional derivative ∂α ΠΦ f is of function type and satisfies

∂α ΠΦ f (x) = (∂β Φ)(ω)ω j dσ(ω) f (x) (4.10.25)
S n−1

+ lim+ (∂α Φ)(x − y) f (y) dy in D (Rn ),
ε→0 Rn \B(x,ε)
for any β ∈ Nn0 and j ∈ {1, . . . , n} such that β + e j = α.

∞
Proof. Fix f ∈ Lcomp (Rn ) and let K := supp f which is a compact set in Rn . By
Exercise 4.51 and Exercise 4.53,
for every α ∈ Nn0 we have that ∂α Φ ∈ C ∞ (Rn \ {0}),

∂α Φ is positive homogeneous of degree m − |α|, and (4.10.26)

∂α Φ(x − y) ≤ ∂α Φ L∞ (S n−1 ) |x − y|m−|α| ∀ x, y ∈ Rn , x y.
Fix now α ∈ Nn0 such that |α| ≤ m + n . Then m − |α| ≥ m + n > −n, hence
(4.10.26) further yields

|(∂α Φ)(x − y) f (y)| dy ≤ C f L∞ (Rn ) |x − y|m−|α| dy < ∞. (4.10.27)
Rn K
This proves that Π∂α Φ f is well defined.

Next, we focus on proving
Π∂α Φ f ∈ C 0 (Rn ). (4.10.28)
To see this, fix x0 ∈ Rn and pick an arbitrary sequence {xk }k∈N of points in Rn
satisfying lim xk = x0 . Consider the following functions defined a.e. in Rn :
k→∞
vk := (∂α Φ)(xk − ·) f, ∀ k ∈ N, and v := (∂α Φ)(x0 − ·) f. (4.10.29)
To conclude that Π∂α Φ f is continuous at x0 , it suffices to show that

lim vk (y) dy = v(y) dy. (4.10.30)
k→∞ K K
The strategy for proving (4.10.30) is to apply Vitali’s theorem (cf. Theorem 14.29)
with X := K and μ being the restriction to K of the Lebesgue measure in Rn . Since
K is compact we have μ(X) < ∞ and from (4.10.27) we know that vk ∈ L1 (X, μ) for
all k ∈ N. Clearly, |v(x)| < ∞ for μ-a.e. x ∈ K. Also, lim vk (y) = v(y) for μ-almost
k→∞
every y ∈ K. Hence, in order to obtain (4.10.30), the only hypothesis left to verify in
Vitali’s theorem is that the sequence {vk }k∈N is uniformly integrable in (X, μ). With
this goal in mind, let ε > 0 be fixed and consider a μ-measurable set A ⊂ K such
that μ(A) is sufficiently small, to a degree to be specified later. Then for every k ∈ N,
based on (4.10.26), we have

vk (x) dμ(x) ≤ ∂α Φ L∞ (S n−1 ) f L∞ (Rn ) |xk − y|m−|α| dy
A
A
=C |x| m−|α|
dx, (4.10.31)
A−xk
where C := ∂α Φ L∞ (S n−1 ) f L∞ (Rn ) . Note that μ(A − xk ) = μ(A) for each k ∈ N. Also,
since the sequence {xk }k∈N and the set K are bounded,
∃ R ∈ (0, ∞) such that A − xk ⊂ K − xk ⊂ B(0, R) ∀ k ∈ N. (4.10.32)
Given that m − |α| > −n, we have |x|m−|α| ∈ L1 (B(0, R)) and we may invoke Proposi-
tion 14.30 to conclude that there exists δ > 0 such that for every Lebesgue measur-
able set E ⊂ B(0, R) with Lebesgue measure less than δ we have E |x|m−|α| dx < ε/C.
At this point return with the latter estimate in (4.10.31) to conclude that

if μ(A) < δ then vk (x) dμ(x) < ε ∀ k ∈ N. (4.10.33)
A
This proves that the sequence {vk }k∈N is uniformly integrable in (X, μ) and finishes
the proof of the fact that Π∂α Φ f is continuous at x0 . Since x0 was arbitrary in Rn the
membership in (4.10.28) follows.
Our next goal is to show (4.10.24). First note that based on what we proved so
far we have ΠΦ f, Π∂α Φ f ∈ Lloc1
(Rn ), thus they define distributions in Rn . We claim
that the distribution Π∂α Φ f is equal to the distributional derivative ∂α [ΠΦ f ]. To see
this, fix ϕ ∈ C0∞ (Rn ) and using the definition of distributional derivatives and the
definition of ΠΦ f write

∂α ΠΦ f, ϕ = (−1)|α| ΠΦ f, ∂α ϕ

= (−1)|α| Φ(x − y) f (y) dy ∂α ϕ(x) dx
Rn Rn

= (−1)|α| f (y) Φ(x − y)∂α ϕ(x) dx dy. (4.10.34)
Rn Rn
Based on (4.10.26), the assumptions on ϕ, and the fact that m − |α| > −n, we may
use Lebesgue’s Dominated Convergence Theorem and formula (4.10.26) repeatedly
to further write

α
Φ(x − y)∂ ϕ(x) dx = lim+ Φ(x − y)∂α ϕ(x) dx (4.10.35)
Rn ε→0 Rn \B(y,ε)
/
|α|
= lim+ (−1) (∂α Φ)(x − y)ϕ(x) dx
ε→0 Rn \B(y,ε)
⎫
⎪
⎪
j β xj − yj γ ⎪
⎬
+ cβγ (∂ Φ)(x − y) ∂ ϕ(x) dσ(x)⎪
⎪ ,
∂B(y,ε) ε ⎪
⎭
α=β+γ+e j
j
where cβγ are suitable constants independent of ε. Note that, for each β, γ, and j
such that α = β + γ + e j , in light of (4.10.26) we have

β xj − yj γ
(∂ Φ) (x − y) ∂ ϕ(x) dσ(x)
∂B(y,ε) ε

≤ |(∂β Φ)(x − y)| |∂γ ϕ(x)| dσ(x)
∂B(y,ε)

≤ ∂γ ϕ L∞ (Rn ) ∂β Φ L∞ (S n−1 ) |x − y|m−|β| dσ(x)
∂B(y,ε)
= Cεm−|β|+n−1 −−−−→
+
0, (4.10.36)
ε→0
since m − |β| + n − 1 ≥ m + n − |α| > 0. Returning with (4.10.36) to (4.10.35), and

applying one more time Lebesgue’s Dominated Convergence Theorem, it follows
that
Φ(x − y)∂α ϕ(x) dx = (−1)|α| (∂α Φ)(x − y)ϕ(x) dx. (4.10.37)
Rn Rn
This, when used in (4.10.34) further yields

α
∂ ΠΦ f, ϕ = f (y) (∂α Φ)(x − y)ϕ(x) dx dy
Rn Rn

= (∂α Φ)(x − y) f (y) dy ϕ(x) dx
Rn Rn

= Π∂α Φ f, ϕ . (4.10.38)
Since ϕ ∈ C0∞ (Rn ) is arbitrary, from (4.10.38) we conclude
∂α ΠΦ f = Π∂α Φ f in D (Rn ), ∀ α ∈ Nn0 with |α| ≤ m + n . (4.10.39)
Upon observing that (4.10.28) and (4.10.39) hold for every α ∈ Nn0 with the property
that |α| = m + n , we may invoke Theorem 2.112 to infer that ΠΦ f belongs to
C m+n (Rn ) and that (4.10.24) is valid.
We are left with proving the very last statement in the theorem. To this end,
suppose there exists α ∈ Nn0 satisfying |α| = m + n. In particular, we have m ∈ Z
and |α| ≥ 1. Hence, there exists j ∈ {1, . . . , n} with the property that α j ≥ 1. Set
β := (α1 , . . . , α j−1 , α j − 1, α j+1 , . . . , αn ), so that
α = β + e j and |β| = m + n − 1 = m + n . (4.10.40)
Based on what we proved earlier, we have ∂β [ΠΦ f ] = Π∂β Φ f pointwise in Rn . Also,

∂β Φ is of class C ∞ and positive homogeneous of degree 1 − n in Rn \ {0}. Thus, we
may apply Theorem 4.104 with Φ replaced by ∂β Φ and Ω replaced by K and obtain

∂α ΠΦ f (x) = ∂ j ∂β [ΠΦ f ] (x) = ∂ j Π∂β Φ f (x)

= ∂x j (∂β Φ)(x − y) f (y) dy
K

= (∂β Φ)(ω)ω j dσ(ω) f (x)
S n−1

+ lim+ (∂α Φ)(x − y) f (y) dy, (4.10.41)
ε→0 Rn \B(x,ε)
where the derivative in the left-hand side of (4.10.41) is taken in D (Rn ) and the
equality is understood in D (Rn ). This proves (4.10.25) and finishes the proof of the
theorem.
Exercise 4.106. In the context of Theorem 4.105 prove directly, without relying on
∞
Theorem (2.112), that whenever f ∈ Lcomp (Rn ) one has ΠΦ f ∈ C m+n (Rn ).
Hint: Show that for each x ∈ Rn and each j ∈ {1, . . . , n} fixed, one has

lim (ΠΦ f )(x + he j ) − (ΠΦ f )(x) = (Π∂ j Φ f )(x) (4.10.42)
h→0
The proof of (4.10.42) may be done by using Vitali’s theorem (cf. Theorem 14.29) in
a manner analogous to the proof of (4.10.28). Once (4.10.42) is established, iterate
to allow higher order partial derivatives.
Further Notes for Chapter 4. The significance of the class of tempered distributions stems from
the fact that this class is stable under the action of the Fourier transform. The topics discussed
in Sections 4.1–4.6 are classical and a variety of expositions is present in the literature, though
they differ in terms of length and depth, and the current presentation is no exception. For example,
while the convolution product of distributions is often confined to the case in which one of the
distributions in question is compactly supported, i.e., E (Rn ) ∗ D (Rn ), in Theorem 4.19 we have
seen that S(Rn ) ∗ S (Rn ) continues to be meaningfully defined in S (Rn ). For us, extending the
action of the convolution product in this manner is motivated by the discussion in Section 4.9,
indicating how singular integral operators may be interpreted as multipliers.
The main result in Section 4.7, Theorem 4.79, appears to be new at least in the formulation
and the degree of generality in which it has been presented. This result may be regarded as a far-
reaching generalization of Sokhotsky’s formula (2.10.3); cf. Remark 4.82 for details. Theorem 4.79
has a number of remarkable consequences, and we use it to offer a new perspective on the treatment
of the classical harmonic Poisson kernel in Section 4.8. Later on, in Section 11.6, Theorem 4.79
resurfaces as the key ingredient in the study of boundary behavior of layer potential operators in
the upper-half space.
The treatment of the singular integral operators from in Section 4.9 highlights the interplay
between distribution theory, harmonic analysis, and partial differential equations. Specifically, first
of a singular integral operator of the form T Θ on a Schwartz function ϕ is interpreted as
the action
P.V. Θ ∗ϕ (which, as pointed out earlier, is a well-defined object in S(Rn )∗S (Rn ) ⊂ S (Rn )). Sec-

ond, the Fourier analysis of tempered distributions is invoked to conclude that T Θϕ = ϕ F P.V. Θ
which shifts the focus of understanding the properties of thesingular integral operator T Θ to clarify-
ing the nature of the tempered distribution mΘ := F P.V. Θ . Third, it turns out that the distribution
mΘ is of function type and, in fact, a suitable pointwise formula may be deduced for it. In partic-
ular, it is apparent from this formula that mΘ ∈ L∞ (Rn ). Having established this, in a fourth step
we may return to the original focus of our investigation, namely the singular integral operator T Θ ,
and use the fact that T Θ ϕ = mΘ ϕ, along with Plancherel’s formula and the boundedness of mΘ to
eventually conclude (via a density argument) that T Θ extends to a linear and bounded operator on
L2 (Rn ). In turn, singular integral operators naturally intervene in the derivatives of volume poten-
tials discussed in Section 4.10, and the boundedness result just derived ultimately becomes the key
tool in obtaining estimates for the solution of the Poisson problem, treated later.
The discussion in the previous paragraph also sheds light on some of the main aims of the
theory of singular integral operators (SIO), as a subbranch of harmonic analysis. For example,
one would like to extend the action of SIO to L p (Rn ) for any p ∈ (1, ∞), rather than just L2 (Rn ).
Also, it is desirable to consider SIO which are not necessarily of convolution type, and/or in a
setting in which Rn is replaced by a more general type of ambient (e.g., some type of surface in
Rn+1 ). The execution of this program, essentially originating in the work of A. P. Calderón, A.
Zygmund, and S. G. Mikhlin in the 1950s, stretches all the way to the present day. The reader is
referred to the exposition in the monographs [8], [9], [19], [26], [52], [68], [69], where refer-
ences to specific research articles may be found. Here, we only wish to mention that, in turn, such
progress in harmonic analysis has led to significant advances in those areas of partial differential
equations where SIO play an important role. In this regard, the reader is referred to the discussion
in [31], [39], [52], [57].
Exercise 4.107. Prove that P.V. 1x ∈ S (R).
Exercise 4.108. Prove that |x|N ln |x| ∈ S (Rn ) if N is a real number satisfying
N > −n.
Exercise 4.109. Prove that (ln |x|) = P.V. 1x in S (R).
Exercise 4.110. Let a ∈ (0, ∞) and recall the Heaviside function H from (1.2.9).
Prove that the function eax H(x), x ∈ R, does not belong to S (R). Do any of the
functions e−ax H(−x), eax H(−x), or e−ax H(x), defined for x ∈ R, belong to S (R)?
Exercise 4.111. In each case determine if the given sequence of tempered distribu-
tions indexed over j ∈ N converges in S (R). Whenever convergent, determine its
limit.
(a) f j (x) = x
x2 + j−2
, x ∈ R;
(b) f j (x) = 1
j · 1
x2 + j−2
, x ∈ R;
1 sin jx
(c) f j (x) = π x , x ∈ R \ {0};
2
(d) f j (x) = e j δ j ;
Exercise 4.112. For each j ∈ N let f j (x) := jn θ( jx) for each x ∈ Rn , where the
function θ ∈ S(Rn ) is fixed and satisfies Rn θ(x) dx = 1. Does the sequence { f j } j∈N
converge in S (Rn )? If yes, determine its limit.
2
Exercise 4.113. For each j ∈ N consider the function f j (x) := e x χ[− j, j] (x) defined
for x ∈ R. Prove that the sequence { f j } j∈N converges in D (R) but not in S (R).
Exercise 4.114. For each j ∈ N let f j (x) := χ[ j−1, j] (x) for x ∈ R. Prove that the
sequence { f j } j∈N does not converge in L p (R) for any p ≥ 1 but it converges to zero
in S (R).
Exercise 4.115. For each j ∈ N consider the tempered distribution u j ∈ S (R)
D (R)
defined by u j := χ[− j, j] e x sin(e x ). Prove that u j −−−−→ e x sin(e x ). Does the sequence
j→∞
{u j } j∈N converge in S (R)? If yes, what is the limit?
Exercise 4.116. Let m ∈ N. Prove that any solution u of the equation xm u = 0
in D (R) satisfies u ∈ S (R). Use the Fourier transform to show that the general
.
m−1
solution to this equation is u = ck δ(k) in S (R), where ck ∈ C, for every k =
k=0
0, 1, . . . , m − 1.
Exercise 4.117. Prove that for any f ∈ S (R) there exists u ∈ S (R) such that xu = f
in S (R).
Exercise 4.118. Does the equation e−|x| u = 1 in S (Rn ) have a solution?
2
Exercise 4.119. Prove that the Heaviside function H belongs to S (R) and compute
its Fourier transform in S (R).
Exercise 4.120. Compute the Fourier transform of P.V. 1x in S (R).

Exercise 4.121. Compute the Fourier transform in S (R) of each of the following
tempered distribution (all are considered in S (R) and recall the definition of the sgn
function from (1.4.3):
(a) sgn x;
(b) |x|k for k ∈ N;
sin(ax)
(c) x for a ∈ R \ {0};
sin(ax) sin(bx)
(d) x for a, b ∈ R \ {0};
(e) sin(x2 );
(f) ln |x|.
Exercise 4.122. Suppose a, b ∈ R are such that a > 0. Consider f (x) := x2 −(b+ia)
1
2

for every x ∈ R. Note that f ∈ L (R) hence (4.1.9) implies f ∈ S (R). Compute the
1
Fourier transform in S (R) of f .

Exercise 4.123. Let n = 3 and R ∈ (0, ∞). Compute the Fourier transform of the
tempered distribution δ∂B(0,R) in S (R3 ), where δ∂B(0,R) the distribution defined as in
Exercise 2.146 corresponding to Σ := ∂B(0, R).
Exercise 4.124. Let k ∈ N0 and suppose f ∈ L1 (Rn ) has the property that xα f
belongs to L1 (Rn ) for every α ∈ Nn0 satisfying |α| ≤ k. Prove that
f , the Fourier
n
transform of f in S (R ) satisfies f ∈ C (R ).
k n
Exercise 4.125. Show that χ[−1,1] ∈ S (R) and compute χ

[−1,1] in S (R). Do we have
χ
[−1,1] ∈ L (R)?
1
Exercise 4.126. Fix x0 ∈ Rn and set
t x0 : S (Rn ) → S (Rn )
(4.11.1)
t x0 u, ϕ) := u, t−x0 (ϕ) , ∀ ϕ ∈ S(Rn ),
where t−x0 (ϕ) is understood as in Exercise 3.44. Prove that the map in (4.11.1) is well
defined, linear, and continuous and is an extension of the map from Exercise 3.44 in
the sense that, if f ∈ S(Rn ), then t x0 u f = utx0 ( f ) in S (Rn ). Also show that

F t x0 u = e−ix0 ·(·)
u in S (Rn ) for every u ∈ S (Rn ). (4.11.2)
Exercise 4.127. Let a ∈ R and consider the function g(x) := cos(ax) for every
x ∈ R. Prove that g ∈ S (R) and compute
g in S (R).
Exercise 4.128. Let m ∈ N and suppose P is a polynomial in Rn of degree 2m

that has no real roots. Prove that P1 ∈ S (Rn ), and that if 2m > n then F P1 ∈
C 2m−n−1 (Rn ).
Exercise 4.129. Let P be a polynomial in Rn and suppose k ∈ N. Prove that if P is

homogeneous of degree −k then P ≡ 0.
Exercise 4.130. Let ζ, η ∈ Rn , n ≥ 2, be two unit vectors such that ζ · η −1, and
consider the linear mapping R : Rn → Rn define by
ξ · (η + ζ) ξ · [(1 + 2η · ζ)ζ − η]
Rξ := ξ − ζ+ η, ∀ ξ ∈ Rn . (4.11.3)
1+η·ζ 1+η·ζ
Show that this is an orthogonal transformation in Rn that satisfies
Rζ = η and R η = ζ. (4.11.4)
Exercise 4.131. Prove that for every c ∈ R \ {0} the following formulas hold:
R
cos(cρ) − cos ρ
lim+ dρ = − ln |c|, (4.11.5)
ε→0 ε ρ
R→∞
R
sin(cρ) π
lim+ dρ = sgn c, (4.11.6)
ε→0 ε ρ 2
R→∞
R
cos(cρ) − cos ρ
sup dρ ≤ 2 ln |c|, (4.11.7)
0<ε<R ε ρ
R
sin(cρ)
sup dρ ≤ 4. (4.11.8)
0<ε<R ε ρ
Chapter 5
The Concept of Fundamental Solution
Abstract The first explicit encounter with the notion of fundamental solution takes
place in this chapter. We consider constant coefficient linear differential operators
and discuss the existence of a fundamental solution for such operators based on the
classical Malgrange–Ehrenpreis theorem.
5.1 Constant Coefficient Linear Differential Operators
Recall (3.1.10). To fix notation, for m ∈ N0 , let

P(D) := aα Dα, aα ∈ C, ∀ α ∈ Nn0 , |α| ≤ m. (5.1.1)
|α|≤m

Also, whenever P(D) is as above, define P(−D) := (−1)|α| aα Dα , and set
|α|≤m

P(ξ) := aα ξα , ∀ ξ ∈ Rn . (5.1.2)
|α|≤m
Typically, an object P(D) as in (5.1.1) is called a constant coefficient linear differen-

tial operator, while P(ξ) from (5.1.2) is referred to as its total (or full) symbol.
Moreover,
P(D) is said to have order m provided
(5.1.3)
there exists α∗ ∈ Nn0 with |α∗ | = m and aα∗ 0.
Whenever P(D) is a constant coefficient linear differential operator of order m we
define its principal symbol to be

Pm (ξ) := aα ξα , for ξ ∈ Rn . (5.1.4)
|α|=m

204 5 The Concept of Fundamental Solution
Remark 5.1. Obviously, any operator of the form

Q(∂) := bα ∂α , (5.1.5)
|α|≤m
may be expressed as P(D) for the choice of coefficients aα := i|α| bα for each α ∈ Nn0
with |α| ≤ m. Hence, we may move freely back and forth between the writings
of a differential operator in (5.1.1) and (5.1.5). While we shall naturally define

in this case Q(ξ) := bα ξα , the reader is advised that the full symbol of Q(∂)
|α|≤m
|α| α
from (5.1.5) is the expression i bα ξ , for ξ ∈ Rn , while its principal symbol is
|α|≤m
α
|α|=m i bα ξ , for ξ ∈ R .
m n
The terminology and formalism described above will play a significant role in
our subsequent work. In the last part of this section, we prove a regularity result
involving constant coefficient linear differential operators with nonvanishing total
symbol outside the origin, that is further used to establish a rather general Liouville
type theorem for this class of operators.
Proposition 5.2. Suppose P(D) is a constant coefficient linear differential operator

in Rn with the property that
P(ξ) 0 for every ξ ∈ Rn \ {0}. (5.1.6)
Then if u ∈ S (Rn ) is such that P(D)u = 0 in S (Rn ) it follows that u is a polynomial

in Rn and P(D)u = 0 pointwise in Rn .
Proof. If u ∈ S (Rn ) is such that P(D)u = 0 in S (Rn ), after taking the Fourier
transform, we obtain P(ξ)u = 0 in S (Rn ). Granted (5.1.6), the latter identity implies
suppu ⊆ {0}. By Exercise 4.37, it follows that u is a polynomial in Rn . Since u ∈
C (Rn ), the condition P(D)u = 0 in S (Rn ) further yields P(D)u = 0 pointwise in
∞
Rn .
In preparation for the general Liouville type theorem advertised earlier, we prove
the following result relating the growth of a polynomial to its degree.
Lemma 5.3. Let P(x) be a polynomial in Rn with the property that there exist N ∈
N0 and C, R ∈ [0, ∞) such that
|P(x)| ≤ C|x|N whenever |x| ≥ R. (5.1.7)
Then P(x) has degree at most N. In particular, the only bounded polynomials in Rn
are constants.
Proof. We begin by noting that the case n = 1 is easily handled. In the multidimen-

sional setting, assume that P(x) = aα xα is a polynomial in Rn satisfying (5.1.7)
|α|≤m
for some N ∈ N0 and C, R ∈ [0, ∞). Suppose m ≥ N + 1, since otherwise there
5.2 A First Look at Fundamental Solutions 205
is nothing to prove. Fix an arbitrary ω ∈ S n−1 and consider the one-dimensional

polynomial pω (t) := P(tω) for t ∈ R, i.e.,
m

pω (t) = aα ωα t j , t ∈ R. (5.1.8)
j=0 |α|= j
Given that |pω (t)| = |P(tω)| ≤ C|tω|N = Ct N whenever |t| ≥ R, the one-dimensional
result applies and gives that the coefficient of t j in (5.1.8) vanishes if j ≥ N + 1.
Thus,
N

P(tω) = pω (t) = aα ωα t j = aα (tω)α , ∀ t ∈ R. (5.1.9)
j=0 |α|= j |α|≤N

Given that ω ∈ S n−1 was arbitrary, it follows that P(x) = aα xα for every x ∈ Rn .
|α|≤N

All ingredients are now in place for dealing with the following result.
Theorem 5.4 (A general Liouville type theorem). Assume P(D) is a constant co-
efficient linear differential operator in Rn such that (5.1.6) holds. Also, suppose
u ∈ Lloc
1
(Rn ) satisfies P(D)u = 0 in S (Rn ) and has the property that there exist
N ∈ N0 and C, R ∈ [0, ∞) such that
|u(x)| ≤ C|x|N whenever |x| ≥ R. (5.1.10)
Then u is a polynomial in Rn of degree at most N.

In particular, if u ∈ L∞ (Rn ) satisfies P(D)u = 0 in S (Rn ) then u is necessarily a
constant.
Proof. Since (5.1.10) implies that the locally integrable function u belongs to
S (Rn ) (cf. Example 4.4), Proposition 5.2 implies that u is a polynomial in Rn .
Moreover, Lemma 5.3 gives that the degree of this polynomial is at most N.
5.2 A First Look at Fundamental Solutions
Definition 5.5. Let P(D) be a constant coefficient linear differential operator in Rn .

Then E ∈ D (Rn ) is called a fundamental solution for P(D) provided
P(D)E = δ in D (Rn ). (5.2.1)
A simple, useful observation is that E ∈ D (Rn ) is a fundamental solution for

P(D) if and only if

E, P(−D)ϕ = ϕ(0), ∀ ϕ ∈ C0∞ (Rn ). (5.2.2)
Indeed, this is immediate from the fact that

P(D)E, ϕ =
E, P(−D)ϕ for each
function ϕ ∈ C0∞ (Rn ).
Remark 5.6. Suppose P(D) is a constant coefficient linear differential operator in

Rn . Then the equation
P(D)u = f (5.2.3)
is called the Poisson equation for the operator P(D). If the operator
P(D) has a fundamental solution E ∈ D (Rn ), then for any f ∈ D (Rn ) for which
E ∗ f exists as an element of D (Rn ) (for example, this is the case if f ∈ E (Rn )) the
Poisson equation (5.2.3) is solvable in D (Rn ). Indeed, if we set u := E ∗ f , then
P(D)u = (P(D)E) ∗ f = δ ∗ f = f in D (Rn ). (5.2.4)
This is one of the many reasons for which fundamental solutions are important when
solving partial differential equations.
Remark 5.7. Let P(D) and Q(D) be two constant coefficient linear differential oper-
ators in Rn and let E ∈ D (Rn ) be a fundamental solution for the operator Q(D)P(D)
in D (Rn ). Then P(D)E is a fundamental solution for Q(D) in D (Rn ). Moreover,
P(D)E ∈ S (Rn ) if E ∈ S (Rn ).
In general, there are two types of questions we are going to be concerned with in
this regard, namely whether fundamental solutions exist, and if so whether they can
all be cataloged. The proposition below shows that for any differential operator with
a nonvanishing total symbol on Rn \ {0} a fundamental solution which is a tempered
distribution is unique up to polynomials.
Proposition 5.8. Suppose P(D) is a constant coefficient linear differential operator

in Rn with the property that P(ξ) 0 for every ξ ∈ Rn \ {0}.
If P(D) has a fundamental solution E that is a tempered distribution, then any
other fundamental solution of P(D) belonging to S (Rn ) differs from E by a polyno-
mial Q satisfying P(D)Q = 0 pointwise in Rn .
Proof. Suppose u ∈ S (Rn ) is such that P(D)u = δ in S (Rn ). Then P(D)(u − E) = 0

in S (Rn ) and the desired conclusion is provided by Proposition 5.2.
Consider now the task of finding a formula for a fundamental solution. In a first
stage, suppose P(D) is as in (5.1.1) and make the additional assumption that
there exists c ∈ (0, ∞) such that |P(ξ)| ≥ c for every ξ ∈ Rn . (5.2.5)

1
In such a scenario, P(ξ) ∈ L(Rn ), hence P(ξ)
1
∈ S (Rn ) by Exercise 4.7. Thus, if

some E ∈ S (R ) happens to be a fundamental solution of P(D) (we will see later
n
in Theorem 5.14 that such an E always exists), then P(D)E = δ in S (Rn ). After
= 1 in S (Rn ) (recall part (b) in
applying the Fourier transform, this implies P(ξ)E
Theorem 4.26 and Example 4.22). In particular, since P(ξ) P(ξ)1
= 1 in S (Rn ), we
obtain
5.2 A First Look at Fundamental Solutions 207

− 1 = 0 in
P(ξ) E(ξ) S (Rn ). (5.2.6)
P(ξ)
Multiplying by 1 =
∈ L(Rn ) then gives E(ξ) 1
hence, further,
P(ξ) P(ξ)
∨
, ϕ = (2π)−n E,

E, ϕ = (2π)−n E ϕ∨

ϕ(−ξ)
= (2π)−n dξ, ∀ ϕ ∈ S(Rn ). (5.2.7)
Rn P(ξ)
Formula (5.2.7), derived under the assumption that (5.2.5) holds, is the departure
point for the construction of a fundamental solution for a constant coefficient linear
differential operator P(D) in the more general case when (5.2.5) is no longer as-
sumed. Of course, we would have to assume that P(D) 0 (since otherwise (5.2.1)
clearly does not have any solution).
The following comments are designed to shed some light into how the issues
created by the zeros of P(ξ) (in the context of (5.2.7)) may be circumvented when
condition (5.2.5) is dropped. The first observation is that if ϕ ∈ C0∞ (Rn ) we may
extend its Fourier transform ϕ to Cn by setting
n

ϕ(ζ) := e−i j=1 x j ζ j ϕ(x) dx for every ζ = (ζ1 , . . . , ζn ) ∈ Cn . (5.2.8)
Rn
Note that the compactsupport condition on ϕ is needed since for each ζ ∈ Cn \ Rn

−i nj=1 x j ζ j
fixed the function |e | grows rapidly as |x| → ∞.
Second, if (5.2.5) holds then P(ξ) 0 for each ξ ∈ Rn . Consequently, for each
ξ ∈ Rn there exists a small number r(ξ) > 0 such that P(z) 0 whenever z ∈ Cn
satisfies |z − ξ| < 2r(ξ). Hence, for every η ∈ S n−1 the mapping

ϕ(−ξ − τη)
{τ ∈ C : |τ| < 2r(ξ)} τ → ∈C (5.2.9)
P(ξ + τη)
is well defined and analytic. Using Cauchy’s formula (and keeping in mind the con-
vention (5.2.8)) we may therefore rewrite (5.2.7) in the form

ϕ(−ξ)

E, ϕ = (2π)−n dξ (5.2.10)
Rn P(ξ)

1 ϕ(−ξ − τη) dτ

= (2π)−n · dξ, ∀ ϕ ∈ C0∞ (Rn ).
R n 2πi |τ|=r(ξ) P(ξ + τη) τ
The advantage of the expression in the last line of (5.2.10), compared to the
expression in the last line of (5.2.7), is the added flexibility in the choice of the
set over which P1 is integrated. This discussion suggests that we try to construct a
fundamental solution for P(D) in a manner similar to the expression in the last line
of (5.2.10), with the integration taking place on a set avoiding the zeros of P, even
in the case when (5.2.5) is dropped.
We shall implement this strategy in Section 5.3. For the time being, we discuss a
preliminary result that will play a basic role in this endeavor. To state, it recalls the
notion of principal symbol from (5.1.4).
Lemma 5.9. Let m ∈ N and let P(D) be a differential operator of order m as in

(5.1.1). In addition, suppose that Pm (η) 0 for some η ∈ S n−1 . Then there exist
ε > 0, r0 , r1 , . . . , rm > 0 and closed sets F0 , F1 , . . . , Fm in Rn such that

m
(a) Rn = F j and
j=0
(b) if ξ ∈ F j , τ ∈ C, |τ| = r j , then |P(ξ + τη)| ≥ ε, for each j = 0, 1, . . . , m.
−m
Proof. Choose r j := j+1 m , for j ∈ {0, 1, . . . , m} and set ε := (2m) |Pm (η)| > 0. Also,
for each j ∈ {1, . . . , m} set

F j := ξ ∈ Rn : |P(ξ + τη)| ≥ ε, ∀ τ ∈ C, |τ| = r j . (5.2.11)
These sets are closed in Rn and clearly satisfy (b). There remains to show that (a)
holds.
Fix ξ ∈ Rn , and consider the mapping C τ → P(ξ + τη) ∈ C. This is a
polynomial in τ of order m with the coefficient of τm equal to Pm (η). Let τ j (ξ),
j = 1, . . . , m, be the zeros of this polynomial. Then

m

P(ξ + τη) = Pm (η) τ − τ j (ξ) , ∀ τ ∈ C. (5.2.12)
j=1
Consider
1
M j := τ ∈ C : dist τ, ∂B(0, r j ) < , for j ∈ {0, 1, . . . , m}. (5.2.13)
2m
Given that r j − r j−1 = 1/m for j = 1, . . . , m, the sets in the collection {M j }mj=0 are
pairwise disjoint. Consequently, among these m+1 sets there is at least one that does
not contain any of the numbers τ j (ξ), j = 1, . . . , m (a simple form of the pigeon hole
principle). Hence, there exists k ∈ {0, 1, . . . , m} such that
1
dist τ j (ξ), ∂B(0, rk ) ≥ , ∀ j ∈ {1, . . . , m}. (5.2.14)
2m
Then, |τ j (ξ) − τ| ≥ 1
2m , for every |τ| = rk and every j = 1, . . . , m, which when used
in (5.2.12) implies
m
|Pm (η)|
|P(ξ + τη)| = Pm (η) τ − τ j (ξ) ≥ = ε, if |τ| = rk . (5.2.15)
j=1
(2m)m
This shows that ξ ∈ Fk , thus (a) is satisfied for the sets defined in (5.2.11).
5.3 The Malgrange–Ehrenpreis Theorem 209
5.3 The Malgrange–Ehrenpreis Theorem
The theorem referred to in the title reads as follows.

Theorem 5.10 (Malgrange–Ehrenpreis). Let P(D) be a linear constant coefficient
differential operator in Rn which is not identically zero. Then there exists some
E ∈ D (Rn ) which is a fundamental solution for P(D).
Proof. Let m ∈ N0 and assume that P(D) is as in (5.1.1) of order m. If m = 0 then
P(D) = a ∈ C \ {0} and the distribution E := 1a δ is a fundamental solution for P(D).
Suppose now that m ≥ 1. Since the principal symbol Pm does not vanish identically
in Rn \ {0} and since
ξ
Pm (ξ) = |ξ|m Pm for all ξ ∈ Rn \ {0}, (5.3.1)
|ξ|
we conclude that there exists η ∈ S n−1 such that Pm (η) 0. Fix such an η and retain
the notation introduced in the proof of Lemma 5.9. Then the functions
⎧
⎪
⎪
⎨1, ξ ∈ F j ,
⎪
χ j : R → R, χ j (ξ) := ⎪
n
⎪ j ∈ {0, 1, . . . , m}, (5.3.2)
⎪
⎩0, ξ ∈ Rn \ F j ,
and χj
f j : Rn → R, f j := , j ∈ {0, 1, . . . , m}, (5.3.3)

m
χk
k=0

m
are well defined, measurable (recall that the F j ’s are closed), and f j = 1 on Rn .
j=0
In addition, 0 ≤ f j ≤ 1, and f j = 0 on Rn \ F j for each j ∈ {0, 1, . . . , m}.
Define now the linear functional E : D(Rn ) → C by setting, for each function
ϕ ∈ C0∞ (Rn ),
m
1
−n ϕ(−ξ − τη) dτ
E(ϕ) := (2π) f j (ξ) · dξ. (5.3.4)
j=0 Rn 2πi |τ|=r j P(ξ + τη) τ
In order to prove that E is well defined, fix an arbitrary compact set K ⊂ Rn and
suppose ϕ ∈ C0∞ (Rn ) is such that supp ϕ ⊆ K. First, observe that for each j =
0, 1, . . . , m, if ξ ∈ F j = supp f j , then |P(ξ + τη)| ≥ ε for every τ ∈ C, |τ| = r j , which
makes the integral in (5.3.4) over |τ| = r j finite. Second, we claim that the integrals
over Rn in (5.3.4) are convergent. Indeed, since ϕ is compactly supported and

ϕ(−ξ − τη) = eix·ξ+iτx·η ϕ(x) dx, (5.3.5)
Rn
it follows that
ϕ(−ξ − τη) is analytic in τ. Then, if we fix j ∈ {0, 1, . . . , m}, and take
N ∈ N0 , using (5.3.5) and integration by parts, for each |τ| = r j we may write

n N
(1 + |ξ|2 )N
ϕ(−ξ − τη) = 1− ∂2x eix·ξ eiτx·η ϕ(x) dx
Rn =1

n N
= eix·ξ −1 − ∂2x eiτx·η ϕ(x) dx, (5.3.6)
Rn =1
for each ξ ∈ Rn . An inspection of the last integral in (5.3.6) then yields

ϕ(−ξ − τη)| ≤ C j (1 + |ξ|2 )−N sup ∂αx ϕ(x)
| if |τ| = r j , (5.3.7)
x∈K
|α|≤2N
for some constant C j = C j (K, r j , η) ∈ (0, ∞). In turn, we may use (5.3.7) in (5.3.4)
to obtain
m
1 2πr j
|E(ϕ)| ≤ (2π) −n
Cj f j (ξ) · (1 + |ξ|2 )−N sup ∂α ϕ(x) dξ
j=0 R n 2π εr j x∈K
|α|≤2N
α
≤ C sup ∂ ϕ(x) (1 + |ξ|2 )−N dξ
x∈K Rn
|α|≤2N

= C sup ∂α ϕ(x) (5.3.8)
x∈K
|α|≤2N
for some C ∈ (0, ∞) independent of ϕ, by choosing N > n2 . This proves that E is

well defined.
Clearly E is linear which, together with (5.3.8) and Proposition 2.4, implies that
E ∈ D (Rn ). To finish the proof of the theorem we are left with showing that

P(D)E = δ in D (Rn ), i.e., that (5.2.2) holds. Since F P(−D)ϕ = P∨ ϕ, recalling
(5.3.4) we may write
m ⎡ ⎤
−n
⎢⎢⎢ 1 ϕ(−ξ − τη) ⎥⎥⎥

E, P(−D)ϕ = (2π) f j (ξ) ⎣ ⎢ dτ⎥⎦ dξ
j=0 R
n 2πi |τ|=r j τ
m

−n
= (2π) ϕ(−ξ) dξ
f j (ξ)
j=0 Rn

−n
= (2π)
ϕ(ξ) dξ
Rn
= ϕ(0). (5.3.9)
For the second equality in (5.3.9), we used

# the fact that
ϕ(−ξ − τη) is analytic in τ,

ϕ(−ξ−τη)
hence by Cauchy’s formula we have 2πi 1
|τ|=r j τ dτ =ϕ(−ξ). The third equality

m
relies on f j = 1 on Rn , while the last one follows from part (4) in Exercise 3.27.
j=0
On account of (5.3.9) and (5.2.2) we may therefore conclude that E is a fundamental

solution for P(D).
One of the consequences of Theorem 5.10 is the fact that the Poisson problem for
nonzero constant coefficient linear differential operators is locally solvable. Specif-
ically, we have the following result.
Corollary 5.11. Let P(D) be a constant coefficient linear differential operator in Rn
which is not identically zero. Suppose Ω is a non-empty, open subset of Rn and that
f ∈ D (Ω) is given. Then for every non-empty, open subset ω of Ω with the property
that ω is a compact subset of Ω, there exists u ∈ D (Ω) such that P(D)u = f in
D (ω).
Proof. Let E ∈ D (Rn ) be a fundamental solution of P(D), which is known to exist
by Theorem 5.10. Also, fix ψ ∈ C0∞ (Ω) such that ψ ≡ 1 on ω. Then ψ f ∈ E (Ω)
hence, by Proposition 2.69, it has a unique extension ψ $f ) ⊆
$f ∈ E (Rn ) and supp (ψ

supp ψ. Now let v := E ∗ (ψ $f ) ∈ D (R ). By Proposition 2.50 we have v ∈ D (Ω)
n
Ω
and we claim that u := v is a solution of P(D)u = f in D (ω) (in the sense that
Ω
P(D) uω = f ω in D (ω)). Indeed,

P(D) uω = P(D)u ω = P(D)(vΩ ) ω = [P(D)(E ∗ ψ
$f )]
ω

$ $
= P(D)E ∗ ψ f = δ∗ψ f ω
ω

$f ] = [ψ f ] = f
= [ψ in D (ω). (5.3.10)
ω ω ω
For the first and third equality in (5.3.10) we used (2.5.4), for the fourth equality we
used (e) in Theorem 2.96, while for the sixth equality we used (d) in Theorem 2.96.

The next example gives an approach for computing fundamental solutions for
linear constant coefficient operators on the real line.
d m
m−1 d j
Example 5.12. Let m ∈ N and let P := dx + aj dx be a differential operator
j=0
of order m in R with constant coefficients. Suppose that v is the solution to
⎧
⎪
⎪ v ∈ C ∞ (R),
⎪
⎪
⎪
⎪
⎪
⎪
⎨ Pv = 0 in R,
⎪
⎪
⎪
⎪
(5.3.11)
⎪
⎪
⎪ v( j) (0) = 0, 0 ≤ j ≤ m − 2,
⎪
⎪
⎪
⎩ (m−1)
v (0) = 1,
with the understanding that the third condition is void if m = 1. Recall the Heaviside
function H from (1.2.9) and define the distributions
u+ := vH and u− := −vH ∨ in D (R). (5.3.12)

We claim that u+ and u− are fundamental solutions for P in D (R). To see that this
j 1 d j
d m m−1
is the case, express P as P(D) = im 1i dx + i a j i dx , in line with (5.1.1), and
j=0
note that this forces
1 d m
m−1
1 d j
P(−D) = (−i)m + (−i) j a j
i dx j=0
i dx
d m
m−1
d j
= (−1)m + (−1) j a j . (5.3.13)
dx j=0
dx
As such, for every ϕ ∈ C0∞ (R) we have

∞

P(D)u+ , ϕ = u+ , P(−D)ϕ = v(x) P(−D)ϕ (x) dx
0
∞
= v(m−1) (0)ϕ(0) + (Pv)(x)ϕ(x) dx =
δ, ϕ, (5.3.14)
0
where the third equality uses (5.3.11) and repeated integrations by parts. This proves
that Pu+ = δ in D (R). Similarly, one obtains that Pu− = δ in D (R). We note that
the two fundamental solutions just described satisfy supp u+ ⊆ [0, ∞) and supp u− ⊆
(−∞, 0].
Example 5.13. Let c ∈ C and consider the differential operator P := dx d

+ c in R.
−cx
It is easy to see that v(x) = e , x ∈ R is the solution (in the classical sense) of
the initial value problem Pv = 0 in R, v(0) = 1. Hence, by Example 5.12, the
distributions u+ := vH and u− := −vH ∨ in D (R) are fundamental solutions for P.
Note that instead of using Example 5.12, one may show that Pu+ = δ (and similarly
that Pu− = δ) in D (R) by applying (4) in Proposition 2.43 and the fact that H = δ
in D (R) to write
P(vH) = (vH) + cvH = v H + vH + cvH = vδ = δ in D (R). (5.3.15)
We close this section by proving a strengthened version of the Malgrange–

Ehrenpreis theorem that addresses the issue of existence of fundamental solutions
that are tempered distributions. Such a result is important in the study of partial
differential equations as it opens the door for using the Fourier transform.
Theorem 5.14. Let P(D) be a nonzero constant coefficient linear differential oper-
ator in Rn . Then there exists E ∈ S (Rn ) which is a fundamental solution for P(D).
In particular, the equality P(D)E = δ holds in S (Rn ).
We prove this theorem by relying on a deep result due to L. Hörmander and S.

Lojasiewicz. For a proof of Theorem 5.15 the interested reader is referred to [32]
and [44].
Theorem 5.15 (Hörmander-Lojasiewicz). Let P be a polynomial in Rn that does

not vanish identically and consider the pointwise multiplication mapping
MP : S(Rn ) → S(Rn ), MP (ϕ) := Pϕ, ∀ ϕ ∈ S(Rn ). (5.3.16)
Then MP is linear, continuous, injective, with closed range SP (Rn ) in S(Rn ). More-
over, the mapping MP : S(Rn ) → SP (Rn ) has a linear and continuous inverse
T : SP (Rn ) → S(Rn ).
Granted this, we can now turn to the proof of the theorem stated earlier.
Proof of Theorem 5.14. Let P(D) be as in the statement and consider the polyno-
mial P(ξ) associated to it as in (5.1.2). Denote by SP (Rn ) the range of the map-
ping MP from (5.3.16) corresponding to this polynomial and by T the inverse of
MP : S(Rn ) → SP (Rn ). Then Theorem 5.15 combined with Proposition 14.2 give
that the transpose of T is linear and continuous as a mapping
T t : S (Rn ) → SP (Rn ) (5.3.17)
where SP (Rn ) is the dual of SP (Rn ) endowed with the weak-∗ topology.
Next, assume that some f ∈ S (Rn ) has been fixed and define u := T t f . Then
u ∈ SP (Rn ) and, making use of (14.1.3), (14.1.7), and Theorem 14.5, it is possible
to extend u to an element in S (Rn ), which we continue to denote by u. We claim that
Pu = f in S (Rn ), where Pu is the tempered distribution obtained by multiplying u
with the Schwartz function P(ξ). To see this, fix ϕ ∈ S(Rn ) and write

Pu, ϕ =
u, Pϕ = T t f, MP ϕ =
f, T MP ϕ =
f, ϕ, (5.3.18)
proving the claim. In particular, the above procedure may be implemented for the
distribution f := 1 ∈ S (Rn ). Corresponding to this choice, we obtain u ∈ S (Rn )
with the property that Pu = 1 in S (Rn ). Define E := F −1 u. Then E belongs to
S (Rn ) and
= P(ξ)E
P(D)E = Pu = 1 = δ in S (Rn ). (5.3.19)
Taking the Fourier transform we see that, as desired, E is a tempered distribution
fundamental solution for P(D).
Further Notes for Chapter 5. The existence of fundamental solutions for arbitrary nonzero, con-
stant coefficient, linear, partial differential operators were first established in this degree of gener-
ality by L. Ehrenpreis and B. Malgrange (cf. [13], [45]), via proofs relying on the Hahn–Banach
theorem. This basic result served as a first compelling piece of evidence of the impact of the theory
of distributions in the field of partial differential equations. The general discussion initiated here is
going to be further augmented by considering in later chapters concrete examples of fundamental
solutions associated with basic operators arising in mathematics, physics, and engineering.
Exercise 5.16. Let m ∈ N. Prove that

u ∈ S (R) : u(m) = δ in S (R) (5.4.1)

m−1
= 1
(m−1)! xm−1 H + ck xk : ck ∈ C, k = 1, . . . , m − 1 .
k=0
Exercise 5.17. Let n, m ∈ N and k1 , k2 ∈ N0 . Consider two scalar operators

P1 (D x ) := aα Dαx , aα ∈ C, ∀ α ∈ Nn0 , |α| ≤ k1 , (5.4.2)
|α|≤k1

P2 (Dy ) := bβ Dβy , bβ ∈ C, ∀ β ∈ Nm
0 , |β| ≤ k2 , (5.4.3)
|β|≤k2
and assume that E1 ∈ D (Rn ) and E2 ∈ D (Rm ) are fundamental solutions for P1
in Rn and P2 in Rm , respectively. Prove that E1 ⊗ E2 is a fundamental solution for

P1 (D x ) ⊗ P2 (Dy ) := aα bβ Dαx Dβy in Rn × Rm .
|α|≤k1 |β|≤k2
Chapter 6
Hypoelliptic Operators
Abstract The concept of hypoelliptic operator is introduced and studied. A classical

result, due to L. Schwartz, is proved here to the effect that a necessary and sufficient
condition for a linear, constant coefficient differential operator to be hypoelliptic in
the entire ambient space is that the named operator possesses a fundamental solu-
tion with singular support consisting of the origin alone. In this chapter an integral
representation formula and interior estimates for a subclass of hypoelliptic operators
are proved as well.
6.1 Definition and Properties
As usual, assume that Ω ⊆ Rn is an open set, and fix m ∈ N. From Theorem 2.112
we know that if u ∈ D (Ω) is such that for each α ∈ Nn0 satisfying |α| = m the
distributional derivative ∂α u is continuous on Ω, then u ∈ C m (Ω). To re-phrase
this result in a manner which lends itself more naturally to generalizations, define

Pm u := ∂α u |α|=m for each u ∈ D (Ω). In this notation, the earlier result reads
u ∈ D (Ω) and Pm u ∈ C 0 (Ω) =⇒ u ∈ C m (Ω). (6.1.1)
We are interested in proving results in the same spirit with the role of mapping Pm
played by (certain) linear differential operators.
Recall the notation from (3.1.10) and consider the linear differential operator

P(x, D) = aα (x)Dα , aα ∈ C ∞ (Ω), for all α ∈ Nn0 with |α| ≤ m. (6.1.2)
|α|≤m
By Proposition 2.41, if f ∈ C 0 (Ω) and u ∈ C m (Ω) is a solution of P(x, D)u = f in

D (Ω), then u is a solution of P(x, D)u = f in Ω in the classical sense. However,
this reasoning requires knowing in advance that u ∈ C m (Ω). It is desirable to have a
more general statement to the effect that if u ∈ D (Ω) is a solution of the equation

216 6 Hypoelliptic Operators
P(x, D)u = f in D (Ω) and f has certain regularity properties, then the distribution
u also belongs to an appropriate smoothness class (e.g., if f ∈ C ∞ (Ω) then u ∈
C ∞ (Ω)). While such a phenomenon is not to be expected in general, there exist
classes of operators P(x, D) for which there are a priori regularity results for the
solution u based on the regularity of the datum f . Such a class of operators is singled
out in this chapter.
Definition 6.1. An operator P(x, D) as in (6.1.2) is called hypoelliptic in Ω if
for all open subsets ω of Ω the following property holds: whenever u ∈ D (ω) is
such that P(x, D)u ∈ C ∞ (ω) then u ∈ C ∞ (ω).
Remark 6.2. It is easy to see from Definition 6.1 that if an operator is hypoelliptic
in Ω, then it is also hypoelliptic in any other open subset of Ω.
Definition 6.3. Let u ∈ D (Ω) be arbitrary. Then the singular support of u,
denoted by sing supp u, is defined by

sing supp u := x ∈ Ω : there is no open set ω

such that x ∈ ω ⊆ Ω and uω ∈ C ∞ (ω) . (6.1.3)
Example 6.4. Not surprisingly,
sing supp δ = {0}. (6.1.4)
To see why this is true, note that clearly sing supp δ ⊂ {0} and sing supp δ ∅ since,
by Example 2.13, the distribution δ is not of function type.
Remark 6.5.(1) Since Ω \ sing supp u is an open set, it follows that sing supp u is
relatively closed in Ω.
(2) By Exercise 2.53, we have that uΩ\sing supp u ∈ C ∞ (Ω \ sing supp u).
(3) The singular support of a distribution u ∈ D (Ω) is the smallest relatively closed
set in Ω with the property that u restricted to its complement in Ω is of class C ∞ .
(4) If u ∈ D (Ω) is such that its singular support is a compact subset of Ω, then u
is the sum between a distribution of function type and a compactly supported
distribution in Ω. Indeed, if we denote by f the function obtained by extending
the restriction uΩ\sing supp u by zero to Ω, then f ∈ Lloc
1
(Ω) and supp(u − u f ) =

sing supp u, hence u − u f ∈ E (Ω).
(5) If the operator P(x, D) is as in (6.1.2), then
sing supp [P(x, D)u] ⊆ sing supp u, ∀ u ∈ D (Ω). (6.1.5)
Indeed, if x0 ∈ Ω \ sing supp u, then there exists an open neighborhood ω of x0

contained in Ω and such that uω ∈ C ∞ (ω). Hence,

[P(x, D)u]ω = P(x, D)[uω ] ∈ C ∞ (ω), (6.1.6)
which implies that x0 ∈ Ω \ sing supp [P(x, D)u].

6.1 Definition and Properties 217
The quality of being a hypoelliptic operator turns out to be equivalent to (6.1.5)

being valid with equality.
Proposition 6.6. Let P(x, D) be as in (6.1.2). Then P(x, D) is hypoelliptic in Ω if
and only if
sing supp [P(x, D)u] = sing supp u, ∀ u ∈ D (Ω). (6.1.7)
Proof. Suppose P(x, D) is hypoelliptic in Ω. The left-to-right inclusion in (6.1.7) has

been proved in part (5) of Remark 6.5. To prove the opposite inclusion, fix a point
x0 ∈ Ω \ sing supp
[P(x, D)u]. Then there exists an open subset ω of Ω with x0 ∈ ω
and [P(x, D)u]ω ∈ C ∞ (ω). Since P(x, D) is hypoelliptic in Ω it is also hypoelliptic
in ω and, as such, uω ∈ C ∞ (ω). This proves that x0 ∈ Ω \ sing supp u, as wanted.
Now suppose (6.1.7) holds and let ω be an open subset of Ω. Fix some u ∈ D (ω)
such that P(x, D)u ∈ C ∞ (ω). The desired conclusion will follow as soon as we show
that u ∈ C ∞ (ω). By Exercise 2.53, it suffices to prove that
for each point x0 ∈ ω there
exists an open neighborhood ω0 ⊆ ω of x0 such that uω belongs to C ∞ (ω0 ). To this
0
end, fix x0 ∈ ω along with some open neighborhood ω0 of x0 whose closure is a
compact subset of ω. Furthermore, pick a function ψ in C0∞ (ω) with the property
that ψ ≡ 1 in a neighborhood of ω0 , and note that necessarily ψu ∈ E (ω). By
Proposition 2.69, every v ∈ E (ω) has a unique extension to E (Ω) that we will
denote by Ext(v). Since the operation of taking the extension Ext commutes with
applying P(x, D) (recall Exercise 2.70) we may write

P(x, D)[Ext(ψu)] = Ext P(x, D)(ψu) = Ext ψP(x, D)u + g

= Ext ψP(x, D)u + Ext(g) in D (Ω), (6.1.8)
where g := P(x, D)(ψu) − ψP(x,D)u ∈ E (ω). On the other hand, from part (4) in
Proposition 2.43 it follows that gω = 0, thus
0
(sing supp g) ∩ ω0 = ∅. (6.1.9)
In concert, (6.1.9), the fact that ψ(P(x, D)u) ∈ C0∞ (ω) (recall that, by assumption,
P(x, D)u ∈ C ∞ (ω)), and (6.1.8), imply that

sing supp P(x, D)[Ext(ψu)] ∩ ω0 = ∅. (6.1.10)
Invoking the hypothesis (6.1.7) then yields sing supp [Ext(ψu)] ∩ ω0 = ∅ which,
in turn, forces [Ext(ψu)]ω ∈ C ∞ (ω0 ). The desired conclusion now follows upon
0
observing that [Ext(ψu)]ω = (ψu)ω = uω .
0 0 0
Remark 6.7. A noteworthy consequence of Proposition 6.6 is as follows: any two

fundamental solutions of a hypoelliptic operator in Rn differ by a C ∞ function. More
specifically, if an operator P(x, D) as in (6.1.2) is hypoelliptic in Rn , then E1 − E2 ∈
C ∞ (Rn ) for every E1 , E2 ∈ D (Rn ) satisfying P(x, D)E1 = δ and P(x, D)E2 = δ in
D (Rn ).
In particular, if an operator P(x, D) as in (6.1.2) is hypoelliptic in Rn and has a

fundamental solution E ∈ D (Rn ) with the property that sing supp E = {0} then the
singular support of any other fundamental solution of P(x, D) is also {0}.
6.2 Hypoelliptic Operators with Constant Coefficients
In this subsection we further analyze hypoelliptic operators with constant coeffi-

cients. First we present a result due to L. Schwartz regarding a necessary and suf-
ficient condition for a linear, constant coefficient operator to be hypoelliptic in the
entire space.
Theorem 6.8. Let P(D) be a constant coefficient linear differential operator in Rn .

Then P(D) is hypoelliptic in Rn if and only if there exists E ∈ D (Rn ) fundamental
solution of P(D) in D (Rn ) satisfying sing supp E = {0}.
Proof. Suppose P(D) is hypoelliptic. Then necessarily P(D) is not identically zero
(since the zero operator sends the Dirac distribution to the C ∞ function zero).
Granted this, Theorem 5.10 applies and gives that the operator P(D) has a fun-
damental solution E ∈ D (Rn ). In concert with the hypoellipticity of P(D), Propo-
sition 6.6, and (6.1.4), this allows us to write
sing supp E = sing supp [P(D)E] = sing supp δ = {0}, (6.2.1)
as wanted.
Consider next the converse implication. Suppose that there exists E ∈ D (Rn )
with P(D)E = δ and sing supp E = {0}. Fix u ∈ D (Rn ) arbitrary. Combining (5) in
Remark 6.5 with Proposition 6.6, in order to conclude that P(D) is hypoelliptic, it
suffices to prove that
sing supp u ⊆ sing supp [P(D)u]. (6.2.2)
To this end, fix a point x0 ∈ Rn \ sing supp [P(D)u]. Then, there exists an open
neighborhood ω of x0 such that [P(D)u]ω ∈ C ∞ (ω). The fact that x0 does not belong
to sing suppu will follow if we are able to show that there exists r ∈ (0, ∞) such that
uB(x ,r) ∈ C ∞ (B(x0 , r)).
0
Pick r1 ∈ (0, ∞) with the property that B(x0 , r1 ) ⊂ ω. Also, fix r0 ∈ (0, r1 )
∞
select a function ψ ∈ C0 (ω) such that ψ ≡ 1 on B(x0 , r1 ). Set v := ψu. Then
and

vB(x ,r ) = uB(x ,r ) and v ∈ E (ω). By virtue of Proposition 2.69 we may extend v to
0 1 0 1
a distribution in E (Rn ) (and we continue to denote this extension by v).
Define g := P(D)v − ψ[P(D)u]. Since [P(D)u]ω ∈ C ∞ (ω) and supp ψ ⊂ ω, we
deduce that ψ[P(D)u] ∈ C0∞ (ω). In particular, ψ[P(D)u] ∈ C0∞ (Rn ). Consequently,
g ∈ E (Rn ). In addition, combining
(4) in Proposition 2.43 with the fact that ψ ≡ 1
on B(x0 , r1 ) shows that gB(x ,r ) = 0. Thus,
0 1
6.2 Hypoelliptic Operators with Constant Coefficients 219
sing supp g ∩ B(x0 , r1 ) = ∅. (6.2.3)
On the other hand, invoking (d) and (e) in Theorem 2.96 we may write
v = δ ∗ v = (P(D)E) ∗ v = E ∗ (P(D)v)
= E ∗ g + E ∗ (ψP(D)u) in D (Rn ). (6.2.4)
We claim that
sing supp(E ∗ g) ∩ B(x0 , r0 ) = ∅. (6.2.5)
∞
is true, fix ε ∈ (0, r1 − r0 ) and φ ∈ C0 (B(0, ε)) with the property
To see why (6.2.5)
ε
that φ ≡ 1 on B 0, 2 . Then E ∗ g = [(1 − φ)E] ∗ g + (φE) ∗ g. Hence, recalling (a)
in Theorem 2.96, we may write
supp [(φE) ∗ g] ⊆ supp (φE) + supp g ⊆ B(0, ε) + [Rn \ B(x0 , r1 )]

⊆ Rn \ B(x0 , r0 ). (6.2.6)
This shows that

sing supp[(φE) ∗ g] ∩ B(x0 , r0 ) = ∅. (6.2.7)
In addition, since by assumption sing supp E = {0}, it follows that (1 − φ)E belongs
to C ∞ (Rn ) and invoking Exercise 2.103, we obtain that [(1 − φ)E] ∗ g is of class C ∞
in Rn . The latter combined with (6.2.7) now yields (6.2.5).
After using (6.2.5) back in (6.2.4) and observing that E ∗ (ψP(D)u) belongs to
C ∞ (Rn ) (itself a consequence of Proposition 2.102) we conclude that vB(x ,r ) is of
0 0
∞
C in B(x0 , r0 ). Upon recalling that ψ ≡ 1 on B(x0 , r0 ), we necessarily have
class
uB(x ,r ) ∈ C ∞ (B(x0 , r0 )). The proof of the theorem is now complete.
0 0
Exercise 6.9. Show that if P(D) is a constant coefficient linear differential operator
which is hypoelliptic in Rn , then any fundamental solution E ∈ D (Rn ) of P(D) has
the property that sing supp E = {0}.
Definition 6.10. Let P(D) be a constant coefficient linear differential operator in Rn .
Then F ∈ D (Rn ) is called a parametrix of P(D) if there exists w ∈ C ∞ (Rn ) with
the property that P(D)F = δ + w in D (Rn ).
It is natural to think of the concept of parametrix as a relaxation of the notion of
fundamental solution since, clearly, any fundamental solution is a parametrix. We
have seen that Theorem 6.8 is an important tool in determining if a given operator
P(D) is hypoelliptic. One apparent drawback of this particular result is the fact that
constructing a fundamental solution for a given operator P(D) may be a delicate
task. It is therefore desirable to have a more flexible criterion for deciding whether
a certain operator is hypoelliptic and, remarkably, Theorem 6.11 shows that one can
use a parametrix in place of a fundamental solution to the same effect.
Theorem 6.11. Let P(D) be a constant coefficient linear differential operator in Rn .
Then P(D) is hypoelliptic in Rn if and only if there exists F ∈ D (Rn ) which is a
parametrix of P(D) and satisfies sing supp F = {0}.
Proof. In one direction, if P(D) is hypoelliptic in Rn , then by Theorem 6.8 there

exists a fundamental solution E of P(D) with sing supp E = {0}. The desired con-
clusion follows by taking F := E. Conversely, suppose there exists a parametrix
F ∈ D (Rn ) of P(D) satisfying sing supp F = {0}. Let w ∈ C ∞ (Rn ) be such
that P(D)F − δ = w in D (Rn ). To show that P(D) is hypoelliptic in Rn , we
adjust the argument used in the proof of Theorem 6.8. Specifically, starting with
u ∈ D (Rn ) arbitrary, as before, matters are reduced to proving that for each point
x0 in Rn \ sing supp [P(D)u] there exists r ∈ (0, ∞) such that uB(x ,r) ∈ C ∞ (B(x0 , r)).
0
Retaining the notation and context from the proof of Theorem 6.8, the reasoning
there applies and allows us to write (in place of (6.2.4))
v = δ ∗ v = (P(D)F − w) ∗ v = F ∗ (P(D)v) − w ∗ v
= F ∗ g + F ∗ (ψP(D)u) − w ∗ v in D (Rn ). (6.2.8)
Since sing supp F = {0}, the same argument as in the proof of (6.2.5) shows that
sing supp(F ∗ g) ∩ B(x0 , r0 ) = ∅, (6.2.9)
while since ψ[P(D)u] ∈ C0∞ (Rn ), Proposition 2.102 gives that
F ∗ (ψP(D)u) ∈ C ∞ (Rn ). (6.2.10)
The new observation is that from v ∈ E (Rn ), w ∈ C ∞ (Rn ), and Exercise 2.103, we
have
w ∗ v ∈ C ∞ (Rn ). (6.2.11)

Combining (6.2.8)–(6.2.11) it follows that vB(x ,r ) ∈ C ∞ (B(x0 , r0 )). The latter forces

uB(x ,r ) ∈ C ∞ (B(x0 , r0 )) (recall that ψ ≡ 1 on B(x0 , r0 )). The proof of the theorem is
0 0
0 0
now complete.
In the class of constant coefficient linear differential operators, Theorem 6.8 and
Theorem 6.11 provide useful characterizations of hypoellipticity. In order to offer a
wider perspective on this matter, we state, without proof, another such characteriza-
tion1 due to L. Hörmander.
Theorem 6.12. Let P(D) be a nonzero constant coefficient linear differential oper-
ator in Rn . Then P(D) is hypoelliptic if and only if there exist constants C, R, c ∈
(0, ∞) with the property that
β
∂ P(ξ) −c|β|
≤ C|ξ| , ∀ ξ ∈ Rn with |ξ| ≥ R, (6.2.12)
P(ξ)
for every β ∈ Nn0 .
1
This characterization is not going to play a significant role for us here. For a proof, the interested
reader is referred to [35, Theorem 11.1.3, p. 62].
An important subclass of linear, constant coefficient operators is that of elliptic

operators. Assume that an operator P(D) of order m ∈ N0 as in (5.1.1) has been

given and recall that its principal symbol is Pm (ξ) = aα ξα , ξ ∈ Rn .

|α|=m
Definition 6.13. Let P(D) be a constant coefficient linear differential operator of

order m ∈ N0 in Rn . Then P(D) is called elliptic if its principal symbol vanishes
only at zero, i.e., if Pm (ξ) 0 for every ξ ∈ Rn \ {0}.
Remark 6.14. An example of a constant coefficient elliptic operator playing a sig-

nificant role in partial differential equations is the poly-harmonic operator of
order 2m, where m ∈ N, defined as

n
Δm := ∂2j1 · · · ∂2jm in Rn . (6.2.13)
j1 ,..., jm =1

n 2 m
Note that Δm is simply j=1 ∂ j ; hence, its principal symbol is (−1) |ξ|
m 2m
and it
does not vanish for any ξ ∈ R \ {0}. The operators obtained for m = 1 and m = 2
n
are called, respectively, the Laplacian and the bi-Laplacian operator. By the
above discussion they are also elliptic.
The relevance of the class of elliptic operators is apparent from the following
important consequence of Theorem 6.11.
Theorem 6.15. Linear, constant coefficient elliptic operators are hypoelliptic in Rn .
Proof. Let P(D) be a constant coefficient linear elliptic operator of order m ∈ N0 .

By Theorem 6.11, to conclude that P(D) is hypoelliptic, matters reduce to showing
that P(D) has a parametrix F with sing supp F = {0}.
Starting from the ellipticity condition Pm (ξ) 0 for every ξ ∈ Rn \ {0}, we note
that if C0 := inf |Pm (ξ)|, then C0 > 0. Keeping in mind that Pm (ξ) is a homogeneous
|ξ|=1
polynomial of degree m we may then estimate

ξ m ξ
|Pm (ξ)| = Pm |ξ| = |ξ| Pm ≥ C0 |ξ| ,
m
∀ ξ ∈ Rn \ {0}. (6.2.14)
|ξ| |ξ|
In addition, there exists C1 ∈ (0, ∞) such that
|P(ξ) − Pm (ξ)| ≤ C1 (1 + |ξ|)m−1 ∀ ξ ∈ Rn . (6.2.15)
Hence, from (6.2.14) and (6.2.15) it follows that for every ξ ∈ Rn \ {0} we have
|P(ξ)| ≥ |Pm (ξ)| − |P(ξ) − Pm (ξ)| ≥ C0 |ξ|m − C1 (1 + |ξ|)m−1 . (6.2.16)
If |ξ| is sufficiently large, then C1 (1 + |ξ|)m−1 ≤ 2 |ξ| ,

C0 m
which implies that there exists
R ∈ (0, ∞) such that
C0 m
|P(ξ)| ≥ |ξ| whenever |ξ| ≥ R. (6.2.17)
2
Hence, if we take ψ ∈ C0∞ (Rn ) satisfying ψ ≡ 1 on B(0, R), then
1 − ψ(ξ)
∈ C ∞ (Rn ) ∩ L∞ (Rn ). (6.2.18)
P(ξ)
Recalling (4.1.9) and (a) in Theorem 4.26, it follows that there exists some tempered
distribution F ∈ S (Rn ) with the property that
= 1 − ψ(ξ)
F in S (Rn ). (6.2.19)
P(ξ)
= 1 − ψ(ξ) in S (Rn ). Making use of (b)

Since P(ξ) ∈ L(Rn ), we further have P(ξ)F
in Theorem 4.26 and (4.2.4), then taking the Fourier transform of the latter identity
and invoking (4.2.34), we arrive at
P(D)F − δ = −(2π)−n
ψ∨ in S (Rn ). (6.2.20)
Since
ψ ∈ S(Rn ), identity (6.2.20) shows that F is a parametrix for P(D).
There remains to prove sing supp F = {0}. The inclusion {0} ⊆ sing supp F is
immediate from (6.2.20) given (6.1.4) and (6.1.5). The opposite inclusion will fol-
low once we show
F Rn \{0} ∈ C ∞ (Rn \ {0}). (6.2.21)
To this end, we claim that if β, α ∈ Nn0 then

F Dβ xα F ∈ L1 (Rn ) whenever |β| − |α| − m < −n. (6.2.22)
Assume (6.2.22) for now. Then, recalling (3.1.3), it follows that
Dβ (xα F) ∈ C 0 (Rn ) for β, α ∈ Nn0 , |β| − |α| − m < −n. (6.2.23)
Next, fix k ∈ N0 and choose N ∈ N0 with the property that 2N > n + k − m. Since

N! 2α
|x|2N = α! x for every x ∈ Rn (cf. (14.2.3)) from (6.2.23) we may conclude
|α|=N
that Dβ (|x|2N F)
∈ C (R ) for every β ∈ N0 such that |β| = k. Hence, |x| F ∈ C (R ).
0 n n 2N k n
1 ∞
Because |x|2N Rn \{0} ∈ C (R \ {0}), the latter further gives F Rn \{0} ∈ C (R \ {0}). The
n k n
membership in (6.2.21) now follows since k ∈ N0 is arbitrary.

Returning to the proof of (6.2.22), fix β, α ∈ Nn0 with |β| − |α| − m < −n. In light
of (b)–(c) in Theorem 4.26, it suffices to show ξβ Dα F ∈ L1 (Rn ). Note that based on
(6.2.19) and (2.4.18) we may write
α! 1
=
Dα F Dγ Dα−γ 1 − ψ(ξ) in S (Rn ). (6.2.24)
γ≤α
γ!(α − γ)! P(ξ)
We now claim that for each μ, ν ∈ Nn0 and with R as in (6.2.17), there exists a
finite constant C > 0 independent of ξ such that

ξμ Dν 1 ≤ C|ξ||μ|−|ν|−m for |ξ| > R. (6.2.25)
P(ξ)

Indeed, since by induction one can see that Dν P(ξ) 1 Q(ξ)
= P(ξ)|ν|+1 for some polynomial
Q of degree at most (m − 1)|ν|, by making also use of (6.2.17), for |ξ| ≥ R we may
write

ξμ Dν 1 = |ξ||μ| |Q(ξ)| ≤ C|ξ||μ| |ξ|
(m−1)|ν|
= C|ξ||μ|−|ν|−m , (6.2.26)
P(ξ) |P(ξ)||ν|+1 |ξ|(|ν|+1)m
where C ∈ (0, ∞) is independent of ξ. This proves (6.2.25).
At this point, we combine (6.2.24), (6.2.25), (6.2.17), and the properties of ψ to
estimate

≤
ξβ Dα F α! ξβ Dγ 1 Dα−γ 1 − ψ(ξ)
γ!(α − γ)!
γ<α
P(ξ)
1

+ ξβ Dα 1 − ψ(ξ)
P(ξ)

≤ Cα,γ |ξ||β|−|γ|−m χsupp ψ\B(0,R)
γ<α
+ C|ξ||β|−|α|−m χRn \B(0,R) , ∀ ξ ∈ Rn . (6.2.27)
Recalling that |β| − |α| − m < −n, from (6.2.27) we obtain that ξβ Dα F belongs to
L1 (Rn ), as desired. This completes the proof of (6.2.22), and with it the proof of the
theorem.
Exercise 6.16. Use Theorem 6.12 to give an alternative proof of Theorem 6.15, i.e.,
that linear, constant coefficient elliptic operators are hypoelliptic in Rn .
Hint: Show that C1 may be chosen sufficiently large so that, in addition to (6.2.15),
one also has |∂β P(ξ)| ≤ C1 (1 + |ξ|)m−|β| . Then use the latter and (6.2.17) to show that
(6.2.12) is verified by taking C := 2CC0 (1 + R ) and c := 1.
1 1 m
Proposition 6.17. The poly-harmonic operator is hypoelliptic. In particular, the

operators Δ and Δ2 are hypoelliptic.
Proof. This is a consequence of Theorem 6.15 and Remark 6.14.
Corollary 6.18. Let P(D) be a linear, constant coefficient elliptic operator in Rn

and let Ω be an open subset of Rn . If u ∈ D (Ω) is such that for some open subset ω
of Ω we have P(D)uω ∈ C ∞ (ω), then uω ∈ C ∞ (ω).
Proof. From Theorem 6.15 and Remark 6.2 we know that the restriction of P(D) to
ω is a hypoelliptic operator. Since P(D)uω ∈ D (ω) has an empty singular support,

Proposition
6.6 gives that the singular support of uω is also empty. Consequently,
uω ∈ C ∞ (ω).
An example of a linear, constant coefficient, differential operator that is not
hypoelliptic is the operator P = ∂21 − ∂22 used in (1.1.1) to describe the equation
governing the displacement of a vibrating string in R2 . Indeed, as we shall see later
(cf. (9.1.28)), the distribution u f associated with the locally integrable function
f (x1 , x2 ) := H(x2 − |x1 |), ∀ (x1 , x2 ) ∈ R2 , (6.2.28)
(where H stands for the Heaviside function) satisfies (∂21 − ∂22 )u f = 2δ in D (R2 )
and sing supp u f = {(x1 , x2 ) ∈ R2 : x2 = |x1 |}. Hence,
sing supp u f sing supp Pu f (6.2.29)
which, in light of Proposition 6.6, shows that P is not hypoelliptic.
6.3 Integral Representation Formulas and Interior Estimates
Consider the constant coefficient, linear, differential operator

P(∂) = aα ∂α , aα ∈ C, for all α ∈ Nn0 with |α| ≤ m. (6.3.1)
|α|≤m
Theorem 6.19 (A general integral representation formula). Assume that the con-
stant coefficient, linear, differential operator P(∂) is as in (6.3.1) and is hypoelliptic
in Rn . Let E ∈ D (Rn ) be a fundamental solution for P(∂) as provided by Theo-
rem 6.8 (hence, in particular, E ∈ C ∞ (Rn \ {0})).
Let Ω be an open subset of Rn and suppose u ∈ D (Ω) satisfies P(∂)u = 0 in

D (Ω). Then u ∈ C ∞ (Ω) and for each x0 ∈ Ω, each r ∈ 0, dist (x0 , ∂Ω) , and each

function ψ ∈ C0∞ B(x0 , r) such that ψ ≡ 1 near B(x0 , r/2), we have
α!
u(x) = − (−1)|α|+|γ| aα × (6.3.2)
|α|≤m γ<α
(α − γ)!γ!

× (∂γ E)(x − y)(∂α−γ ψ)(y)u(y) dy,
B(x0 ,r)\B(x0 ,r/2)
for each x ∈ B(x0 , r/2). In particular, for every μ ∈ Nn0 ,

6.3 Integral Representation Formulas and Interior Estimates 225
α!
(∂μ u)(x) = − (−1)|α|+|γ| aα × (6.3.3)
|α|≤m γ<α
(α − γ)!γ!

× (∂γ+μ E)(x − y)(∂α−γ ψ)(y)u(y) dy,
B(x0 ,r)\B(x0 ,r/2)
for each x ∈ B(x0 , r/2).
Proof. The fact that u ∈ C ∞ (Ω) is a consequence of the hypoellipticity of the oper-
ator P(∂) (cf. Definition 6.1 and Remark 6.2). As regards (6.3.2), pick an arbitrary

x0 ∈ Ω, fix r ∈ 0, dist (x0 , ∂Ω) , and let ψ ∈ C0∞ B(x0 , r) be such that ψ ≡ 1 near
B(x0 , r/2). Then ψu ∈ C0∞ (Rn ) and since P(∂)E = δ in D (Rn ) we may write, for
each point x ∈ B(x0 , r/2),

u(x) = (ψu)(x) = δ ∗ (ψu) (x)

= (P(∂)E) ∗ (ψu) (x) = E ∗ (P(∂)(ψu)) (x)
α!
= aα E ∗ (∂β ψ∂α−β u) (x) (6.3.4)
|α|≤m 0<β≤α
β!(α − β)!
where we have used (6.3.1), part (e) of Theorem 2.96, and that P(∂)u = 0 in Ω. Note
that for each β > 0 the function ∂β ψ is compactly supported in B(x0 , r) \ B(x0 , r/2).
Keeping this in mind, we may then integrate by parts in order to further write the
last expression in (6.3.4) as
α!
aα E ∗ (∂β ψ∂α−β u) (x) (6.3.5)
|α|≤m 0<β≤α
β!(α − β)!

α!
= aα E(x − y)∂β ψ(y)∂α−β u(y) dy
|α|≤m 0<β≤α
β!(α − β)! B(x0 ,r)\B(x0 ,r/2)

(−1)|α|+|β| α!
= aα ∂α−β
y E(x − y)∂β ψ(y) u(y) dy
|α|≤m 0<β≤α
β!(α − β)! B(x0 ,r)\B(x0 ,r/2)
(−1)|α|+|β|+|γ| α! (α − β)!
= aα ×
|α|≤m 0<β≤α γ≤α−β
β!(α − β)! γ!(α − β − γ)!

× (∂γ E)(x − y)∂α−γ ψ(y) u(y) dy.
B(x0 ,r)\B(x0 ,r/2)
To proceed, observe that whenever γ < α formula (14.2.5) gives

(α − γ)! (α − γ)!
(−1)|β| = (−1)|β| −1
0<β≤α−γ
β!(α − γ − β)! β≤α−γ
β!(α − γ − β)!
= 0 − 1 = −1. (6.3.6)
Making use of (6.3.6) back in (6.3.5) yields

α!
aα E ∗ (∂β ψ∂α−β u) (x) (6.3.7)
|α|≤m 0<β≤α
β!(α − β)!
(−1)|α|+|γ| α!
=− aα ×
|α|≤m γ<α
γ!(α − γ)!

× (∂γ E)(x − y)∂α−γ ψ(y) u(y) dy,
B(x0 ,r)\B(x0 ,r/2)
and (6.3.2) follows from (6.3.4) and (6.3.7). Finally, (6.3.3) is obtained by differen-
tiating (6.3.2).
To state our next result, recall that the operator P(∂) from (6.3.1) is said to be
homogeneous (of degree m) whenever aα = 0 for all multi-indices α with |α| < m.
Theorem 6.20 (Interior estimates). Let P(∂) be a constant coefficient, linear, dif-
ferential operator, of order m ∈ N0 , which is hypoelliptic in Rn . Also suppose
E ∈ D (Rn ) is a fundamental solution for P(∂) as provided by Theorem 6.8 (thus, in
particular, E ∈ C ∞ (Rn \ {0})).
Then for every μ ∈ Nn0 there exists a constant Cμ ∈ (0, ∞) (which also depends
on the coefficients of P(∂) and n) with the following significance. If Ω be an open
subset of Rn and u ∈ D (Ω) satisfies P(∂)u = 0 in D (Ω), then u belongs to C ∞ (Ω)

and for each x0 ∈ Ω and each r ∈ 0, dist (x0 , ∂Ω) we have

sup |(∂μ u)(x)| ≤ Cμ max r|γ|−|α| · sup |(∂γ+μ E)(z)| ×
|α|≤m
x∈B(x0 ,r/2) γ<α r/4<|z|<r

× |u(y)| dy. (6.3.8)
B(x0 ,r)\B(x0 ,r/2)
In the particular case when P(∂) is homogeneous (of degree m) and one also
assumes that there exists k ∈ N such that the fundamental solution E satisfies
|∂β E(x)| ≤ Cβ |x|−n+m−|β| ,

(6.3.9)
∀ x ∈ Rn \ {0}, ∀ β ∈ Nn0 with |β| ≥ k,
then whenever |μ| ≥ k we have

μ Cμ
|(∂ u)(x0 )| ≤ |μ| − |u(y)| dy. (6.3.10)
r B(x0 ,r)

Proof. Pick a function φ ∈ C0∞ B(0, 1) such that φ ≡ 1 near B(0, 3/4), and note that
if we set

ψ(x) := φ (x − x0 )/r , ∀ x ∈ Rn , (6.3.11)
∞
then ψ ∈ C0 B(x0 , r) and ψ ≡ 1 near B(x0 , 3r/4). In particular, for every γ ∈ Nn0
with |γ| > 0 there exists cγ ∈ (0, ∞) independent of r such that
supp (∂γ ψ) ⊆ B(x0 , r) \ B(x0 , 3r/4) and ∂γ ψL∞ (Rn ) ≤ cγ r−|γ| . (6.3.12)
For such a choice of ψ, estimate (6.3.8) then follows in a straightforward manner

from the integral representation formula (6.3.3). In turn, (6.3.8) readily implies
(6.3.10) in the case when P(∂) is homogeneous (of degree m), E satisfies (6.3.9)
for some k ∈ N, and |μ| ≥ k.
The procedure described abstractly in the following lemma is useful for deriving
higher order interior estimates, with precise control of the constants involved, for
various classes of functions.
Lemma 6.21. Let A ⊂ C ∞ (Ω) be a family of functions satisfying the following two
conditions:
(1) ∂α u ∈ A for every u ∈ A and every α ∈ Nn0 ;
(2) there exists C ∈ (0, ∞) such that for every u ∈ A we have
C
|∂ j u(x)| ≤ max |u(y)|, ∀ j ∈ {1, . . . , n}, (6.3.13)
r y∈B(x,r)

whenever x ∈ Ω and r ∈ 0, dist (x, ∂Ω) .

Then given any u ∈ A, for every x ∈ Ω, every r ∈ 0, dist (x, ∂Ω) , every k ∈ N,
and every λ ∈ (0, 1), we have (with C as in (6.3.13))
C k (1−λ)−k ek−1 k!
max |∂α u(y)| ≤ rk
max |u(y)|,
y∈B(x,λr) y∈B(x,r) (6.3.14)
for every multi-index α ∈ Nn0 with |α| = k.
Proof. In a first stage, we propose to prove by induction over k that, given any

u ∈ A, for every x ∈ Ω, every r ∈ 0, dist (x, ∂Ω) , and every k ∈ N we have (with C
as in (6.3.13))
C k ek−1 k!
|∂α u(x)| ≤ max |u(y)|, ∀ α ∈ Nn0 with |α| = k. (6.3.15)
rk y∈B(x,r)
Note that if k = 1, this is contained in (6.3.13). Suppose now that (6.3.15) holds for

some k ∈ N and every u ∈ A, every x ∈ Ω, and every r ∈ 0, dist (x, ∂Ω) , and pick
an arbitrary α ∈ Nn0 with |α| = k + 1. Then there exist j ∈ {1, . . . , n} and β ∈ Nn0 for
which ∂α = ∂ j ∂β . Fix such j and β and note that, in particular, |β| = k. Next, take

x ∈ Ω and r ∈ 0, dist (x, ∂Ω) arbitrary, and pick some ε ∈ (0, 1) to be specified
shortly. Since ∂β u ∈ A we may use (6.3.13) with u replaced by ∂β u and r replaced
by (1 − ε)r to obtain
C
|∂α u(x)| ≤ max |∂β u(y)|. (6.3.16)
(1 − ε)r y∈B(x,(1−ε)r)

Note that if y ∈ B x, (1 − ε)r then B(y, εr) ⊂ B(x, r) ⊂ Ω. Thus, for every point

y ∈ B x, (1 − ε)r , we may use the induction hypothesis to estimate
C k ek−1 k! C k ek−1 k!
|∂β u(y)| ≤ max |u(z)| ≤ max |u(z)|. (6.3.17)
k
(εr) z∈B(y,εr) εk rk z∈B(x,r)
Combined, (6.3.16) and (6.3.17) yield
C k+1 ek−1 k!
|∂α u(x)| ≤ max |u(z)|. (6.3.18)
(1 − ε)εk rk+1 z∈B(x,r)
Set now ε := k
k+1 ∈ (0, 1) in (6.3.18) which, given that ε−k < e, further implies
C k+1 ek (k + 1)!
|∂α u(x)| ≤ max |u(z)|. (6.3.19)
rk+1 z∈B(x,r)
Hence, (6.3.15) holds with k + 1 in place of k, as desired.

As far as (6.3.14) is concerned, pick x0 ∈ Ω, r ∈ 0, dist (x0 , ∂Ω) , λ ∈ (0, 1), as
well as k ∈ N. Next, select an arbitrary point x ∈ B(x0 , λr) and note that this forces

(1 − λ)r ∈ 0, dist (x, ∂Ω) . In concert with the easily seen fact that B x, (1 − λ)r ⊆
B(x0 , r), this permits us to invoke (6.3.15) in order to estimate
C k ek−1 k!
|∂α u(x)| ≤ max |u(y)|
(1 − λ)r k y∈B(x,(1−λ)r)
C k (1 − λ)−k ek−1 k!
≤ max |u(y)|, (6.3.20)
rk y∈B(x0 ,r)
whenever α ∈ Nn0 satisfies |α| = k. Taking the supremum over x ∈ B(x0 , λr) then
yields (6.3.14) (written with x0 in place of x).
Definition 6.22. A function u ∈ C ∞ (Ω) is called real-analytic in Ω provided

for every x ∈ Ω there exists r x ∈ (0, ∞) with the property that B(x, r x ) ⊆ Ω and the
Taylor series for u at x converges uniformly to u on B(x, r x ), i.e.,
1
u(y) = (y − x)α (∂α u)(x) uniformly for y ∈ B(x, r x ). (6.3.21)
α∈Nn
α!
0
The following observation justifies the name “real-analytic” used for the class of
functions introduced above.
Remark 6.23. In the context of Definition 6.22, it is clear that, for each x ∈ Ω,
1

u(z) := (z − x)α (∂α u)(x) for z ∈ Cn with |z − x| < r x , (6.3.22)
n α!
α∈N 0
is a holomorphic function in {z ∈ Cn : |z − x| < r x } which locally extends the

original real-analytic function u. Such local extensions may be constructed near
each point x ∈ Ω and Lemma 7.67 ensures that any two local extensions coincide
on their common domain. The conclusion is that the real-analytic function u has a
well-defined extension to a holomorphic function defined in a neighborhood of Ω in
Cn .
Our next lemma gives a sufficient condition (which is in the nature of best possi-
ble) ensuring real-analyticity.
Lemma 6.24. Suppose u ∈ C ∞ (Ω) is a function with the property that for each x ∈ Ω

there exist r = r(x) ∈ 0, dist (x, ∂Ω) , M = M(x) ∈ (0, ∞), and C = C(x) ∈ (0, ∞),
such that for every k ∈ N the following estimate holds:
max |∂α u(y)| ≤ MC k k! ∀ α ∈ Nn0 with |α| = k. (6.3.23)

y∈B(x,r)
Then u is real-analytic in Ω.

Proof. Fix x ∈ Ω and let r ∈ 0, dist (x, ∂Ω) be such that (6.3.23) holds for every
k ∈ N. Write Taylor’s formula (14.2.10) for u at x to obtain that for each N ∈ N and
each y ∈ B(x, r/2) there exists θ ∈ (0, 1) such that
1
u(y) = (y − x)α (∂α u)(x) + RN,u (y), (6.3.24)
|α|≤N−1
α!
where 1
RN,u (y) := (y − x)α (∂α u)(x + θ(y − x)). (6.3.25)
|α|=N
α!
Using (6.3.23), for each α ∈ Nn0 with |α| = N and y ∈ B(x, r/2), we may estimate
α
(∂ u)(x + θ(y − x)) ≤ max |(∂α u)(z)| ≤ MC N N! (6.3.26)
z∈B(x,r)

since B x + θ(y − x), r ⊂ B(x, r). Together, (6.3.25) and (6.3.26) imply that, for each
y ∈ B(x, r/2),
N!
|RN,u (y)| ≤ MC N |y − x|N = M (nC|y − x|)N . (6.3.27)
|α|=N
α!
For the equality in (6.3.27) we used formula (14.2.3) for x1 = · · · = xn = 1. Hence,

1 r
if say, |y − x| < min 2nC , 2 , then lim RN,u (y) = 0 uniformly with respect to y.
N→∞
Consequently, the Taylor series for u converges uniformly to u in a neighborhood of
x. Since x is arbitrary in Ω it follows that u is real-analytic in Ω.
Theorem 6.25 (Unique continuation). Suppose Ω ⊆ Rn is an open and connected

set. Then any real-analytic function u in Ω with the property that there exists x0 ∈ Ω
such that ∂α u(x0 ) = 0 for all α ∈ Nn0 (which is the case if, e.g., u is zero in a
neighborhood of x0 ) vanishes identically in Ω.
Proof. Suppose u satisfies the hypotheses of the theorem and define the set

U := x ∈ Ω : ∂α u(x) = 0 for all α ∈ Nn0 . (6.3.28)
Since x0 ∈ U we have U ∅. Also, U is relatively closed in Ω given that it is

the intersection of the relatively closed sets (∂α u)−1 {0} , α ∈ Nn0 . In addition, if
x ∈ U then, on the one hand, the Taylor series of u at x is identically zero, while
on the other hand, this series converges to u in an open neighborhood of x. Thus, u
is identically zero in that neighborhood of x, which ultimately proves that U is also
open. Recalling that Ω is connected, it follows that U = Ω as desired.
Theorem 6.25 highlights the much more restrictive nature of real-analyticity
compared to indefinite differentiability, since obviously there are plenty of C ∞ func-
tions which vanish in a neighborhood of a point without being identically zero.
Theorem 6.26. Let P(∂) be a constant coefficient, linear, differential operator in

Rn , which is homogeneous of degree m ∈ N0 . Assume that P(∂) has a fundamental
solution E ∈ D (Rn ) with the property that E ∈ C ∞ (Rn \ {0}) and
|∂β E(x)| ≤ Cβ |x|−n+m−|β| ,

(6.3.29)
∀ x ∈ Rn \ {0}, ∀ β ∈ Nn0 with |β| > 0.
Finally, suppose that Ω ⊆ Rn is open and u ∈ D (Ω) satisfies P(∂)u = 0 in D (Ω).

Then u ∈ C ∞ (Ω) and there exists a constant C ∈ (0, ∞) such that if λ ∈ (0, 1) we
have
C |α| (1 − λ)−|α| |α|!
max |∂α u(y)| ≤ max |u(y)|, ∀ α ∈ Nn0 , (6.3.30)
y∈B(x,λr) r|α| y∈B(x,r)

whenever x ∈ Ω and r ∈ 0, dist (x, ∂Ω) . In particular, u is real-analytic in Ω.
Proof. The fact that P(∂) has a fundamental solution E ∈ D (Rn ) with the property
that E ∈ C ∞ (Rn \ {0}) implies sing supp E = {0}; hence, P(∂) is hypoelliptic in
Rn by Theorem 6.8. Granted the properties enjoyed by P(∂) and condition (6.3.29),
estimate (6.3.30) follows from (6.3.9) (used with |μ| = 1) and Lemma 6.21 (in which
we take A := {u ∈ C ∞ (Ω) : P(∂)u = 0}). Finally, the last claim in the statement of
the theorem is a consequence of (6.3.30) and Lemma 6.24.
Further Notes for Chapter 6. For more extensive discussion of the notion of hypoellipticity the
interested reader is referred to the excellent presentation in [35]. Further information about real-
analytic functions may be found in, e.g., [40].
Exercise 6.27. Let k ∈ (0, ∞). Prove that

d2 u
u ∈ S (R) : + ku = δ in S (R) (6.4.1)
dx2
i i
= − eikt H + e−ikt H + c1 eikt + c2 e−ikt : c1 , c2 ∈ C .
2k 2k
Exercise 6.28. Let a > 0 be fixed.
d2
2 − a in R.
2
(a) Compute a fundamental solution for the operator dx
(b) Use the result from part (a) to compute the Fourier transform in S (R) of the
2 for each x ∈ R
1
tempered distribution associated with the function f (x) := x2 +a
(compare with Example 4.24).
Exercise 6.29. Prove that sing supp u ⊆ supp u for every u ∈ D (Ω).
Exercise 6.30. Give an example of a distribution u for which the inclusion from
Exercise 6.29 is strict. Give an example of a distribution u for which sing supp u =
supp u.
Exercise 6.31. Let x0 ∈ Rn . Prove that sing supp δαx0 = {x0 } for every α ∈ Nn0 .
Exercise 6.32. Let a ∈ C0∞ (Ω) and u ∈ D (Ω). Prove that
sing supp (au) ⊆ sing supp u.
Exercise 6.33. Determine sing supp (P.V. 1x ), where P.V. 1x is the distribution defined
in (2.1.13).
Exercise 6.34. Let m ∈ N and let P be a polynomial in Rn of degree 2m with no

real roots. Recall from Exercise 4.128 that P1 ∈ S (Rn ). Prove that in this scenario

sing supp F P1 is nonempty.
Exercise 6.35. Recall P.V. 1x ∈ D (R) from (2.1.13) and define the distribution u ∈
D (R2 ) by u := P.V. 1x ⊗ δ(y). Determine sing supp u.
Chapter 7
The Laplacian and Related Operators
Abstract Starting from first principles, all fundamental solutions (that are tempered
distributions) for scalar elliptic operators are identified in this chapter. While the
natural starting point is the Laplacian, this study encompasses a variety of related
operators, such as the bi-Laplacian, the poly-harmonic operator, the Helmholtz op-
erator and its iterations, the Cauchy–Riemann operator, the Dirac operator, the per-
turbed Dirac operator and its iterations, as well as general second-order constant
coefficient strongly elliptic operators. Having accomplished this task then makes
it possible to prove the well-posedness of the Poisson problem (equipped with a
boundary condition at infinity), and derive qualitative/quantitative properties for the
solution. Along the way, Cauchy-like integral operators are also introduced and their
connections with Hardy spaces are brought to light in the setting of both complex
and Clifford analyses.
7.1 Fundamental Solutions for the Laplace Operator
One of the most important operators in partial differential equations is the Laplace
n
operator1 Δ (also called the Laplacian) in Rn which is defined as Δ := ∂2j .
j=1
Functions f satisfying Δ f = 0 pointwise in Rn are called harmonic in Rn . The
Laplace operator arises in many applications such as in the modeling of heat conduc-
tion, electrical conduction, and chemical concentration, to name a few. The focus for
us will be on finding all fundamental solutions for Δ that are tempered distributions.
Part of the motivation is that, as explained in Remark 5.6, this greatly facilitates the
study of the Poisson equation Δu = f in D (Rn ), a task we take up in Section 7.2.
The goal in this section is to determine all fundamental solutions of Δ that are
also in S (Rn ). This is done by employing the properties of the Fourier transform
we have proved so far. To get started, fix some E ∈ S (Rn ) that is fundamental
1
Named after the French mathematician and astronomer Pierre Simon de Laplace (1749–1827).
234 7 The Laplacian and Related Operators
solution for Δ. Note that the existence of such a tempered distribution is guaranteed
by Theorem 5.14. Since ΔE = δ in D (Rn ) and δ ∈ S (Rn ), by (4.1.33) it follows
that
ΔE = δ in S (Rn ). (7.1.1)
2
n
Applying F to the equation in (7.1.1) (and using that −Δ = D x j , part (b) in
j=1
Theorem 4.26, and (4.2.3)) we obtain
= 1
−|ξ|2 E in S (Rn ). (7.1.2)
To proceed with determining E, we discuss separately the cases n ≥ 3, n = 2, and

n = 1.
Case n ≥ 3. From Exercise 4.5 we have that |ξ|12 ∈ S (Rn ). In addition, we have
|ξ|2 ∈ L(Rn ), thus |ξ|2 · |ξ|12 ∈ S (Rn ) (recall (b) in Theorem 4.14), and it is not
1
difficult to check that |ξ|2 · |ξ|12 = 1 in S (Rn ). Hence, |ξ|2 E + |ξ|2 = 0 in S (Rn ).
1
Thus, supp E + |ξ|2 ⊆ {0} and by Exercise 4.37, it follows that
+ 1 = P(ξ)
E in S (Rn ), (7.1.3)
|ξ|2
where P is a polynomial in Rn satisfying |ξ|2

P = 0 in S (Rn ). The latter implies that
∞
ΔP = 0 in S (R ). Since P ∈ C (R ), we may conclude that ΔP = 0 pointwise
n n
in Rn , hence P is a harmonic polynomial in Rn . To compute E (which is equal to

apply F −1 to the identity in (7.1.3) and then recall Proposition 4.64 with
F −1 (E))
λ = 2 to write
1
E(x) = −F −1 2 (x) + P(x)
|ξ|

Γ n−2
−2 − n2 2
= −2 π |x|2−n + P(x)
Γ(1)
1 1
=− · n−2 + P(x), ∀ x ∈ Rn \ {0}, (7.1.4)
(n − 2)ωn−1 |x|

where the last equality in (7.1.4) is also based on the fact that Γ n2 is equal to

2 − 1 Γ 2 − 1 (recall that n ≥ 3), that Γ(1) = 1, and on (14.5.6). Moreover,
n n

1 1
Δ − · =δ in S (Rn ), (7.1.5)
(n − 2)ωn−1 |x|n−2
since (7.1.5) is equivalent (via the Fourier transform, as a consequence of (a) and

(b) in Theorem 4.26, as well as (4.2.3)) with |ξ|2 F (n−2)ω 1
n−1
· |x|
1
n−2 = 1 in S (Rn ), or
equivalently, with |ξ|2 · |ξ|12 = 1 in S (Rn ), which we know to be true.
7.1 Fundamental Solutions for the Laplace Operator 235
Case n = 2. Since |ξ|12 Lloc

1
(R2 ), we cannot proceed as in the case n ≥ 3. Instead,
we rely on Theorem 4.76. Retaining the notation from Theorem 4.76, fix ψ ∈ Q and
observe that
|ξ|2 wψ = 1 in S (R2 ). (7.1.6)
To see why this is true, first note that since |ξ|2 ∈ L(R2 ), by (b) in Theorem 4.14
it follows that |ξ|2 wψ ∈ S (R2 ) while by (a) in Theorem 3.14 we have that |ξ|2 ϕ
belongs to S(R2 ) for every ϕ ∈ S(R2 ). Hence, we may write

|ξ|2 ϕ(ξ) − (|ξ|2 ϕ(ξ)) ψ(ξ)
ξ=0
|ξ| wψ , ϕ = wψ , |ξ| ϕ =
2 2
dξ
R n |ξ| 2

= ϕ(ξ) dξ = 1, ϕ , ∀ ϕ ∈ S(R2 ), (7.1.7)

R2
finishing the proof of (7.1.6).

+wψ ) = 0 in S (R2 ). Reasoning
A combination of (7.1.2) and (7.1.6) yields |ξ|2 (E
as in the case n ≥ 3, the latter implies
= −wψ + P
E in S (Rn ), (7.1.8)
for some harmonic polynomial P. After applying the Fourier transform to the equal-
ity in (7.1.8) and recalling (4.6.4), we arrive at
1
E(x) = ln |x| + P(x), ∀ x ∈ R2 \ {0}. (7.1.9)
2π
In addition, by making use of the Fourier transform (recall (a) and (c) in Theo-
rem 4.26, (4.2.3), and Proposition 4.32), equation (7.1.6) is equivalent with
wψ ) = (2π)2 δ
−Δ( in S (R2 ). (7.1.10)
Combining (7.1.10) with (4.6.4), we arrive at

Δ 2π
1
ln |x| = δ in S (R2 ). (7.1.11)
Case n = 1. In this case we use Exercise 5.16 to conclude that E = xH + c1 x + c0 in

S (R), for some c0 , c1 ∈ C. Hence, E is of function type and E(x) = xH + c1 x + c0 ,
for every x ∈ R.
Remark 7.1. From Example 5.12 we know that −xH ∨ is another fundamental so-
lution for Δ in D (R). This is in agreement with what we proved above since
H + H ∨ = 1 in S (R), thus −xH ∨ = xH + x in S (R).
In summary, we have proved the following result.
Theorem 7.2. The function E ∈ Lloc

1
(Rn ) defined as
⎧ −1
⎪
⎪
⎪
1
⎪
⎪ if x ∈ Rn \ {0}, n ≥ 3,
⎪
⎪
⎪ (n − 2)ωn−1 |x|n−2
⎪
⎨
E(x) := ⎪
⎪ (7.1.12)
2π ln |x| if x ∈ R2 \ {0}, n = 2,
1
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎩ xH(x) if x ∈ R, n = 1,

belongs to S (R ) and is a fundamental solution2 for the Laplace operator Δ in Rn .
n
Moreover,

u ∈ S (Rn ) : Δu = δ in S (Rn (7.1.13)

= E + P : P harmonic polynomial in Rn .
Remark 7.3. Note that, as a consequence of Proposition 5.2 and Example 4.4, for a
harmonic function in Rn , the respective function is a polynomial if and only if it is
a tempered distribution.
The collection of harmonic functions is strictly larger than the collection of
harmonic polynomial. For example, the function u(x, y) := e x cos y defined for
(x, y) ∈ R2 , is harmonic in R2 without being a polynomial. In particular, we obtain
that u S (R2 ).
Moving on, we recall that if Ω ⊂ Rn is a domain of class C 1 (see Definition 14.59)

with outward unit normal ν = (ν1 , . . . , νn ) and u is a function of class C 1 in an open
neighborhood of ∂Ω, then the normal derivative of u on ∂Ω, denoted by ∂u ∂ν , is the
directional derivative of the function u along the unit vector ν, i.e.,
∂u n
(y) := ν j (y)∂ j u(y), ∀ y ∈ ∂Ω. (7.1.14)
∂ν j=1
Remark 7.4. Via a direct computation, one may check that E as in (7.1.12) defines
a fundamental solution for the Laplacian. We outline the computation for n ≥ 3 and
leave the cases n = 2 and n = 1 as an exercise.
First, observe that E ∈ C ∞ (Rn \ {0}) and ΔE = 0 in Rn \ {0}. Next, pick a function
ϕ ∈ C0∞ (Rn ) such that supp ϕ ⊂ B(0, R), for some R ∈ (0, ∞). Then, starting with
the definition of distributional derivatives and then using Lebesgue’s Dominated
Convergence Theorem, we write
2
In the case n = 3, the expression for E was used in 1789 by Pierre Simon de Laplace to show that
for f smooth, compactly supported, Δ(E ∗ f ) = 0 outside the support of f (cf. [41]).
7.1 Fundamental Solutions for the Laplace Operator 237

ΔE, ϕ = E, Δϕ = lim+ E(x)Δϕ(x) dx

ε→0 ε≤|x|≤R

∂ϕ
= lim+ ΔE(x)ϕ(x) dx + E dσ
ε→0 ε≤|x|≤R ∂B(0,R) ∂ν

∂ϕ ∂E ∂E
− E dσ − ϕ dσ + ϕ dσ
∂B(0,ε) ∂ν ∂B(0,R) ∂ν ∂B(0,ε) ∂ν

∂E ∂ϕ
= lim+ ϕ−E dσ = ϕ(0), (7.1.15)
ε→0 ∂B(0,ε) ∂ν ∂ν
where for each integral considered over ∂B(0, r), for some r ∈ (0, ∞), ν denotes the
outward unit normal to B(0, r). For the third equality in (7.1.15) we used (14.8.5)
while for the fourth equality in (7.1.15) we used the fact that the support of ϕ is
contained inside B(0, R). In addition, to see why the last equality in (7.1.15) holds,
use the fact that ∇E(x) = ωn−1
1
· |x|xn for every x ∈ Rn \ {0} to write

∂E 1
(x) ϕ(x) dσ(x) = ϕ(x)dσ(x) (7.1.16)
∂B(0,ε) ∂ν ωn−1 εn−1 ∂B(0,ε)

1
= [ϕ(x) − ϕ(0)]dσ(x) + ϕ(0)
ωn−1 εn−1 ∂B(0,ε)

1
= [ϕ(εy) − ϕ(0)]dσ(y) + ϕ(0) −−−−→ ϕ(0)
ωn−1 S n−1 ε→0+
where the convergence follows by invoking Lebesgue’s Dominated Convergence

Theorem. Also,

∂ϕ ∇ϕL∞ 1
E(x) (x) dσ(x) ≤ dσ(x)
∂B(0,ε) ∂ν (n − 2)ωn−1 ∂B(0,ε) |x|n−2
= ε C(n, ϕ) −−−−→
+
0. (7.1.17)
ε→0
This computation shows that if n ≥ 3, the Lloc

1
(Rn ) function E from (7.1.12) satisfies
∞
ΔE, ϕ = δ, ϕ for every ϕ ∈ C0 (R ), thus E is a fundamental solution for the
n
Laplacian.
We conclude this section by presenting a result relating the composition of Riesz

transforms to singular integral operators whose kernels are two derivatives on the
fundamental solution for the Laplacian.
Proposition 7.5. Let E be the fundamental solution for Δ from (7.1.12) and recall
the Riesz transforms in Rn (cf. Theorem 4.97). Then for every function ϕ ∈ S(Rn )
and each j, k ∈ {1, . . . , n} we have
ωn 2

R j Rk ϕ (x) = − lim (∂ j ∂k E)(x − y)ϕ(y) dy
2 ε→0+ |x−y|>ε
ωn 2 δ jk
− ϕ(x) in S (Rn ), (7.1.18)
2 n
with the left-hand side initially understood in L2 (Rn ) (cf. Theorem 4.97) and the
right-hand side considered as a tempered distribution (cf. Proposition 4.70).
Moreover, if T ∂ j ∂k E is the operator as in part (e) of Theorem 4.100 (for the choice
Θ := ∂ j ∂k E), then for every f ∈ L2 (Rn ) we have
ωn 2 ωn 2 δ jk
R j Rk f = − T ∂ j ∂k E f − f in L2 (Rn ). (7.1.19)
2 2 n
Proof. Fix j, k ∈ {1, . . . , n} along with some ϕ ∈ S(Rn ). Since the Fourier transform
is an isomorphism of S (Rn ), it suffices to show that (7.1.18) holds on the Fourier
transform side. With this in mind, note that on the one hand,
ωn ξ j ωn 2 ξ j ξk
F R j Rk ϕ = −i F Rk ϕ = − ϕ in S (Rn ),
(7.1.20)
2 |ξ| 2 |ξ|2
by a twofold application of (4.9.15). On the other hand, since the function ∂ j ∂k E
is as in (4.4.1) (here Example 4.71 is also used), we may invoke Proposition 4.70,
(4.9.6), and (4.5.2), to conclude that the Fourier transform of the right-hand side of
(7.1.18) is
ωn 2 δ jk
−
ϕ F P.V. (∂ j ∂k E) +
ϕ
2 n
ωn 2 1 δ jk
=−
ϕ F P.V. (∂ j Φk ) +
ϕ
2 ωn−1 n
ωn 2 ξ j ξk δ jk δ jk
=− −
ϕ+
ϕ
2 |ξ|2 n n
ωn 2 ξ j ξk
=− ϕ in S (Rn ).
(7.1.21)
2 |ξ|2
That (7.1.18) holds now follows from (7.1.20) and (7.1.21).
Finally, (7.1.19) is a consequence of what we have proved so far, part (e) in
Theorem 4.97, part (e) in Theorem 4.100, and the fact that S(Rn ) is dense in L2 (Rn ).

Making use of the full force of Theorem 4.101 we obtain the following L p ver-
sion, 1 < p < ∞, of Proposition 7.5.
Proposition 7.6. Let E be the fundamental solution for Δ as in (7.1.12) and recall
the Riesz transforms in Rn (cf. Theorem 4.97). Also, fix p ∈ (1, ∞). Then for every
j, k ∈ {1, . . . , n} and f ∈ L p (Rn ) we have
7.2 The Poisson Equation and Layer Potential Representation Formulas 239
ωn 2

R j Rk f (x) = − lim (∂ j ∂k E)(x − y) f (y) dy
2 ε→0+ |x−y|>ε
ωn 2 δ jk
− f (x) for a.e. x ∈ Rn , (7.1.22)
2 n
with the Riesz transforms are understood as bounded operators on L p (Rn ) (cf.
(4.9.48)–(4.9.49)) while the singular integral operator in the right-hand side is also
considered as a bounded mapping on L p (Rn ) (cf. Theorem 4.101).
In particular, if T ∂ j ∂k E is interpreted in the sense of Theorem 4.101 (for the choice
Θ := ∂ j ∂k E), then for every f ∈ L p (Rn ) we have
ωn 2 ωn 2 δ jk
R j Rk f = − T ∂ j ∂k E f − f in L p (Rn ). (7.1.23)
2 2 n
Proof. All claims are consequences of Proposition 7.5, Theorem 4.101, and the
density of S(Rn ) in L p (Rn ).
7.2 The Poisson Equation and Layer Potential Representation

Formulas
In this section we use the fundamental solutions for the Laplacian determined in
the previous section in a number of applications. The first application concerns the
Poisson equation3 for the Laplacian in Rn :
Δu = f in Rn . (7.2.1)
In (7.2.1) the unknown is u and f is a given function. As a consequence of Re-

mark 5.6, this equation is always solvable in D (Rn ) if f ∈ C0∞ (Rn ). We will see that
in fact the solution obtained via the convolution of such an f with the fundamental
solution for Δ from (7.1.12) is smooth and that we can solve the Poisson equation
in D (Rn ) for a less restrictive class of functions f . The second application is an
integral representation formula, involving layer potentials, for functions of class C 2
on a neighborhood of the closure of a domain of class C 1 .
Proposition 7.7. Let E be the fundamental solution for Δ from (7.1.12). Suppose
f ∈ C0∞ (Rn ). Then the function

u(x) := E(x − y) f (y) dy, ∀ x ∈ Rn , (7.2.2)

Rn
3
In 1813, the French mathematician, geometer, and physicist Siméon Denis Poisson (1781–1840)
proved that Δ(E ∗ f ) = f in dimension n = 3 (cf. [61]), where E(x) = − 4π|x|
1
, x ∈ R3 \ {0}, and
∞
f ∈ C0 (R ).
3
satisfies u ∈ C ∞ (Rn ) and is a classical solution of the Poisson equation (7.2.1) for
the Laplacian in Rn .
Proof. Fix f ∈ C0∞ (Rn ) and let E be as in (7.1.12). Define u := E ∗ f in D (Rn ).
Then Remark 5.6 implies that u is a solution of the equation Δu = f in D (Rn ). In
addition, by Proposition 2.102 and the fact that E ∈ Lloc1
(Rn ), we have u ∈ C ∞ (Rn )
and

u(x) = E(x − y), f (y) = E(x − y) f (y) dy, ∀ x ∈ Rn , (7.2.3)

Rn
or even more explicitly (based on the expressions in (7.1.12)), that

⎧
⎪
⎪ −1 f (y)
⎪
⎪
⎪ dy, if n ≥ 3,
⎪
⎪
⎪ (n − 2)ωn−1 Rn |x − y|n−2
⎪
⎪
⎪
⎪
⎨ 1
u(x) = ⎪
⎪ ln |x − y| f (y) dy, if n = 2, ∀ x ∈ Rn . (7.2.4)
⎪
⎪
⎪ 2π R
⎪
2
⎪
⎪
⎪
⎪
⎪
⎪
⎩ (x − y)H(x − y) f (y) dy, if n = 1,
R
Furthermore, based on Proposition 2.41, we may conclude that Δu = f pointwise in

Rn . This proves that u is a solution of (7.7) as desired.
An inspection of the right-hand side of (7.2.2) reveals that this expression con-
tinues to be meaningful under weaker assumptions on f . This observation may be
used to prove that (7.7) is solvable in the class of distributions for f belonging to a
larger class of functions than C0∞ (Rn ).
Proposition 7.8. Let E be the fundamental solution for Δ from (7.1.12). Suppose Ω
is an open set in Rn , n ≥ 2, and f ∈ L∞ (Ω) vanishes outside a bounded measurable
subset of Ω. Then the function

u(x) := E(x − y) f (y) dy, ∀ x ∈ Ω, (7.2.5)

Ω
is a distributional solution of the Poisson equation for the Laplacian in Ω, that is,
Δu = f in D (Ω). In addition, u ∈ C 1 (Ω) and for each j ∈ {1, . . . , n},

∂ j u(x) = (∂ j E)(x − y) f (y) dy, ∀ x ∈ Ω. (7.2.6)

Ω
Proof. Suppose first that n ≥ 3. Note that E ∈ C ∞ (Rn ) and is positive homogeneous
of degree 2 − n. Also the extension of f by zero outside Ω, which we continue to
∞
denote by f , satisfies f ∈ Lcomp (Rn ). From Theorem 4.105 it follows that u ∈ C 1 (Ω)
and (7.2.6) holds for each j ∈ {1, . . . , n}.
If n = 2, a reasoning based on Vitali’s theorem, similar to the one used in the
proof of Theorem 4.105, yields (∂ j E) ∗ f ∈ C 0 (Ω) for each j ∈ {1, . . . , n}. Also, the
computation in (4.10.35) may be adapted to give that the distributional derivative

∂ j u is equal to (∂ j E) ∗ f for each j ∈ {1, . . . , n}. Hence, Theorem 2.112 applies and
yields u ∈ C 1 (Ω), as desired.
To finish the proof of the proposition, there remains to show that Δu = f in
D (Ω). Since f ∈ Lloc 1
(Ω) we have f ∈D (Ω). Thus, for each ϕ∈C0∞ (Ω) we may write

Δu, ϕ = u, Δϕ = u(x)Δϕ(x) dx = E(x − y) f (y) dy Δϕ(x) dx
Ω Ω Ω

= f (y) E(x − y)Δϕ(x) dx dy
Ω Ω

= f (y) E(x)Δϕ(x + y) dx dy
Ω Ω−y

= f (y) E, Δϕ(· + y) dy = f (y) ΔE, ϕ(· + y) dy
Ω Ω

= f (y)δ, ϕ(· + y) dy = f (y)ϕ(y) dy = f, ϕ , (7.2.7)

Ω Ω
where for each y ∈ Ω we set Ω − y := {x − y : x ∈ Ω}. Note that for each y ∈ Ω fixed
one has ϕ(· + y) ∈ C0∞ (Ω − y), 0 ∈ Ω − y, and E ∈ D (Ω − y), thus the sixth equality
in (7.2.7) is justified. The eighth equality in (7.2.7) is a consequence of the fact that
E is a fundamental solution for the Laplacian and that 0 ∈ Ω − y. This completes the
proof of the proposition.
Exercise 7.9. Assume the hypothesis of Proposition 7.8 and recall the fundamen-
tal solution E for the Laplacian from (7.1.12). Using Vitali’s Theorem 14.29,
show that u(x) = Ω E(· − y) f (y) dy is differentiable in Ω (without making use of
Theorem 2.112). As a by-product, show that, for every j ∈ {1, . . . , n},

∂ j u(x) = ∂ j E(x − y) f (y) dy, ∀ x ∈ Ω. (7.2.8)

Ω
We next establish mean value formulas for harmonic functions. A clar-

ification of the terminology employed is in order. Traditionally, the name harmonic
function in an open set Ω ⊆ Rn has been used for functions u ∈ C 2 (Ω) with the
n
property that ∂2j u = 0 in a pointwise sense, everywhere in Ω. Such a function is
j=1
called a classical solution of the equation Δu = 0 in Ω.
Theorem 7.10. For an open set Ω ⊆ Rn the following are equivalent:

(i) u is harmonic in Ω;
(ii) u ∈ D (Ω) and Δu = 0 in D (Ω);
(iii) u ∈ C ∞ (Ω) and Δu = 0 in a pointwise sense in Ω.
Proof. This is a direct consequence of Theorem 6.8 and Theorem 7.2 (or, alterna-
tively, of Remark 6.14 and Corollary 6.18).
Recall the symbol for integral averages from (15).
Theorem 7.11. Let Ω be an open set in Rn and u be a harmonic function in Ω. Then

for every x ∈ Ω and every r ∈ 0, dist(x, ∂Ω) we have

u(x) = − u(y) dy and u(x) = − u(y) dσ(y). (7.2.9)

B(x,r) ∂B(x,r)
Proof. From Theorem 7.10 we know that u ∈ C ∞ (Ω) and Δu = 0 in a pointwise

sense in Ω. Fix x ∈ Ω and define the function

φ : 0, dist(x, ∂Ω) → R,

(7.2.10)
φ(r) := − u(y) dσ(y), ∀ r ∈ 0, dist(x, ∂Ω) .
∂B(x,r)

A change of variables gives that φ(r) = ωn−1
1
S n−1
u(x+rω) dσ(ω). Taking the deriva-
tive of φ and then using the integration by parts formula from Theorem 14.60 give
that

1
φ (r) = (∇u)(x + rω) · ω dσ(ω) (7.2.11)
ωn−1 S n−1

r r
= (Δu)(x + ry) dy = − Δu(z) dz.
ωn−1 B(0,1) n B(x,r)
Since Δu = 0 in Ω we have that φ (r) = 0, thus φ(r) = C for some constant C. We

claim that lim+ φ(r) = u(x). Then this claim implies C = u(x) and the first formula
r→0
in (7.2.9) follows. To see the claim, fix ε > 0 and by the continuity of u at x find

δ ∈ 0, dist (x, ∂Ω) such that |u(y) − u(x)| < ε if |y − x| < δ. Consequently,

r ∈ (0, δ) =⇒ φ(r) − u(x) = − [u(y) − u(x)] dy ≤ ε, (7.2.12)
B(x,r)
as desired.
With the help of (14.5.6) and (14.9.5), the first formula in (7.2.9) may be rewritten
as

r
ωn−1 rn
u(x) = u(y) dσ(y) = u(ω) dσ(ω) dρ (7.2.13)
n B(x,r) 0 ∂B(x, ρ)

for r ∈ 0, dist (x, ∂Ω) . Differentiating the left- and right-most terms in (7.2.13) with
respect to r and then dividing by ωn−1 rn−1 gives the second formula in (7.2.9).
Interior estimates for harmonic functions, as in (6.3.30), may be obtained as a
particular case of Theorem 6.26 by observing that the fundamental solution for the
Laplacian from (7.1.12) satisfies (6.3.29) (with m = 2). Below we take a slightly
more direct route which also yields explicit constants.
Theorem 7.12 (Interior estimates for the Laplacian). Suppose u is harmonic in

Ω. Then for each x ∈ Ω, each r ∈ 0, dist (x, ∂Ω) , and each k ∈ N, we have
(2n)k ek−1 k!
max |∂α u(y)| ≤ max |u(y)|, ∀ α ∈ Nn0 with |α| = k. (7.2.14)
y∈B(x,r/2) rk y∈B(x,r)
Proof. Let j ∈ {1, . . . , n} be arbitrary and note that, by Theorem 7.10, u ∈ C ∞ (Ω)
and hence, ∂ j u is harmonic in Ω. Thus, by (7.2.9) and the integration by parts for-

mula (14.8.4), for each x ∈ Ω and each r ∈ 0, dist (x, ∂Ω) , we may write
n

|∂ j u(x)| = ∂ j u(y) dy
ωn−1 rn B(x,r)
n
yj − xj
= u(y) dσ(y)
ωn−1 rn ∂B(x,r) r
n
≤ max |u(y)|. (7.2.15)
r y∈B(x,r)
With this in hand, Lemma 6.21 applies (with A the class of harmonic functions in
Ω and C := n) and yields (7.2.14).
Theorem 7.13. Any harmonic function in Ω is real analytic in Ω.
Proof. This is an immediate consequence of Theorem 7.12 and Lemma 6.24 (or,
alternatively, Theorem 6.26).
Corollary 7.14. Suppose u is harmonic in an open, connected set Ω ⊆ Rn with the

property that there exists x0 ∈ Ω such that ∂α u(x0 ) = 0 for all α ∈ Nn0 . Then u
vanishes identically in Ω.
As a consequence, if a harmonic function defined in an open connected set Ω ⊆
Rn vanishes on an open subset of Ω, then the respective function vanishes identically
in Ω.
Proof. This is an immediate consequence of Theorem 6.25 and Theorem 7.13.
Theorem 7.15 (Liouville’s Theorem for the Laplacian). If u is a bounded har-

monic function in Rn then there exists c ∈ C such that u = c in Rn .
Proof. This may be justified in several ways. For example, it suffices to note that this
is a particular case of Theorem 5.4. Another proof may be given based on interior
estimates. Specifically, let j ∈ {1, . . . , n} and using (7.2.14), for each x ∈ Rn , we
may write
n
lim |∂ j u(x)| ≤ lim uL∞ (Rn ) = 0. (7.2.16)
r→∞ r→∞ r
Hence, ∇u = 0 proving that u is locally constant in Rn . Since Rn is connected, the
desired conclusion follows.
All ingredients are now in place for proving the following basic well-posedness
result for the Poisson problem for the Laplacian in Rn .
∞
Theorem 7.16. Assume n ≥ 3. Then for each f ∈ Lcomp (Rn ) and each c ∈ C the
Poisson problem for the Laplacian in R ,
n
⎧
⎪
⎪
⎪ u ∈ C 0 (Rn ),
⎪
⎪
⎪
⎪
⎨ Δu = f in D (Rn ),
⎪
⎪
⎪
(7.2.17)
⎪
⎪
⎪
⎪
⎩ |x|→∞
lim u(x) = c,
has a unique solution. Moreover, the solution u satisfies the following additional
properties.
(1) The function u belongs to C 1 (Rn ) and has the integral representation formula

u(x) = c + E(x − y) f (y) dy, x ∈ Rn , (7.2.18)

Rn
where E is the fundamental solution for the Laplacian in Rn from (7.1.12).

(2) If in fact f ∈ C0∞ (Rn ) then actually u ∈ C ∞ (Rn ).
(3) For every j, k ∈ {1, . . . , n}, we have
2 2
∂ j ∂k u = − R j Rk f in D (Rn ), (7.2.19)
ωn
where R j and Rk are ( j-th and k-th) Riesz transforms in Rn .
(4) For every p ∈ (1, ∞), the solution u of (7.2.17) satisfies ∂ j ∂k u ∈ L p (Rn ) for each
j, k ∈ {1, ..., n}, where the derivatives are taken in D (Rn ), and there exists a
constant C = C(p, n) ∈ (0, ∞) with the property that

n
∂ j ∂k uL p (Rn ) ≤ C f L p (Rn ) . (7.2.20)
j,k=1
Proof. The fact that the function u defined as in (7.2.18) is of class C 1 (Rn ) and
solves Δu = f in D (Rn ) has been established in Proposition 7.8. To proceed, let
R ∈ (0, ∞) be such that supp f ⊂ B(0, R), and note that if |x| ≥ 2R, then for every
y ∈ B(0, R) we have |x − y| ≥ |x| − |y| ≥ R ≥ |y|. Hence, using (7.1.12) (recall that we
are assuming n ≥ 3) and Lebesgue’s Dominated Convergence Theorem we obtain

dy
E(x − y) f (y) dy ≤ C f L∞ (Rn ) →0 (7.2.21)
B(0,R) |x − y|
n−2
Rn
as |x| → ∞. It is then clear from (7.2.21) that the function u from (7.2.18) also sat-
isfies the limit condition in (7.2.17). That (7.2.18) is the unique solution of (7.2.17)
follows from linearity and Theorem 7.15. Next, that u ∈ C ∞ (Rn ) if f ∈ C0∞ (Rn ), is a
consequence of Proposition 7.7 (or, alternatively, Corollary 6.18 and the ellipticity
of the Laplacian). This proves (1)–(2). As regards (3), start by fixing j, k in {1, . . . , n}.
Then from (7.2.6), Theorem 4.103, (14.9.45), and Proposition 7.6, we deduce that

(∂ j ∂k u)(x) = ∂ j ∂k u(x) = ∂ j (∂k E)(x − y) f (y) dy
Rn
1

= ωk ω j dσ(ω) f (x)
ωn−1 S n−1

+ lim+ (∂ j ∂k E)(x − y) f (y) dy

ε→0 |x−y|>ε

δ jk
= f (x) + lim+ (∂ j ∂k E)(x − y) f (y) dy
n ε→0 |x−y|>ε
2 2
=− R j Rk f (x) in D (Rn ). (7.2.22)
ωn
Finally, the claim in (4) follows from (3) and the boundedness of the Riesz trans-
forms on L p (Rn ) (cf. (4.9.48)).
In the last part of this section, we prove an integral representation formula in-
volving layer potentials (a piece of terminology we elaborate on a little later) for
arbitrary (as opposed to harmonic) functions u ∈ C 2 (Ω ), where Ω is a bounded do-
main of class C 1 in Rn (as in Definition 14.59). To state this recall the definition of
the normal derivative from (7.1.14).
Proposition 7.17. Suppose n ≥ 2 and let Ω ⊂ Rn be a bounded domain of class C 1

and let ν denote its outward unit normal. Also let E be the fundamental solution for
the Laplacian from (7.1.12). If u ∈ C 2 (Ω), then

∂u
Δu(y)E(x − y) dy − E(x − y) (y) dσ(y) (7.2.23)
Ω ∂Ω ∂ν

⎧
⎪
⎨ u(x), x ∈ Ω,
⎪
− [ν(y) · (∇E)(x − y)]u(y) dσ(y) = ⎪ ⎪
∂Ω ⎩ 0, x ∈ Rn \ Ω.
Several strategies may be employed to prove Proposition 7.17, and here we

choose an approach that highlights the role of distribution theory. Another approach
(based on isolating the singularity in the fundamental solution and a limiting argu-
ment), more akin to the proof of (7.1.15), is going to be used in the proof of Theo-
rem 7.71 (presented later) where a result of similar flavor to (7.2.23) is established
for more general differential operators than the Laplacian.
Proof of Proposition 7.17. In this proof, for a function v defined in Ω, we let v
denote the extension of v with zero outside Ω. Fix some function u ∈ C 2 (Ω )
and observe that
u ∈ Lloc
1
(Rn ) ⊂ D (Rn ). In addition, for each ϕ ∈ C0∞ (Rn ) one has

Δ
u, ϕ =
u, Δϕ = uΔϕ dx
Ω

∂ϕ ∂u
= ϕΔu dx + u −ϕ dσ, (7.2.24)
Ω ∂Ω ∂ν ∂ν
where the last equality in (7.2.24) is based on (14.8.5).

∂
For a ∈ C 0 (∂Ω) define the mappings aδ∂Ω , ∂ν (aδ∂Ω ) : C0∞ (Rn ) → C by setting

aδ∂Ω , ϕ := a(x)ϕ(x) dσ(x), ∀ ϕ ∈ C0∞ (Rn ), (7.2.25)

∂Ω
and
∂
∂ϕ
(aδ∂Ω ), ϕ := − a(x) (x) dσ(x), ∀ ϕ ∈ C0∞ (Rn ). (7.2.26)
∂ν ∂Ω ∂ν
By Exercise 2.146 corresponding to Σ := ∂Ω we have aδ∂Ω ∈ D (Rn ). By a similar
∂
reasoning, one also has ∂ν (aδ∂Ω ) ∈ D (Rn ). In addition, it is easy to see that the
supports of the distributions in (7.2.25) and (7.2.26) are contained in ∂Ω, hence
∂
aδ∂Ω , ∂ν (aδ∂Ω ) ∈ E (Rn ). In light of these definitions, (7.2.24) is equivalent with
− ∂u ∂
Δ
u = Δu δ∂Ω − (uδ∂Ω ) in D (Rn ). (7.2.27)
∂ν ∂ν
Note that the supports of the distributions Δ are contained in the compact
u and Δu
set Ω. Hence, all distributions in (7.2.27) are compactly supported and their convo-
lutions with E are well defined. Furthermore, since
u∗E =
Δ u ∗ ΔE = u in D (Rn ),
u∗δ =
after convolving the left- and right-hand sides of (7.2.27) with E we arrive at

∗ E − ∂u δ∂Ω ∗ E − ∂ (uδ∂Ω ) ∗ E
u = Δu in D (Rn ). (7.2.28)
∂ν ∂ν
Moreover, since Rn \ ∂Ω is open, it follows that the equality in (7.2.28) also holds
in D (Rn \ ∂Ω). The goal is to show that all distributions in (7.2.28) when restricted
to Rn \ ∂Ω are of function type and to determine the respective functions that define
them.
uRn \∂Ω is of function type follows from the defini-
The fact that the distribution

tion of
u. Also, it is immediate that uRn \∂Ω is given by the function equal to u in Ω
and equal to zero in Rn \ Ω, which is precisely the function in the right-hand side of
(7.2.23).
E ∈ L1 (Rn ), with Δu
Since Δu, being compactly supported in Ω, by Exer-
loc
cise 2.104 we have that Δu ∗ E is of function type, determined by the function
∗ E ∈ L1 (Rn ) defined as
Δu loc

∗ E)(x) =
(Δu E(x − y)Δu(y) dy, ∀ x ∈ Rn . (7.2.29)
Ω
To proceed with identifying the restrictions to Rn \ ∂Ω of the other two convolu-

tions in the right-hand side of (7.2.28), let ϕ ∈ C0∞ (Rn \ ∂Ω) and set K := supp ϕ.
Then, the set
MK := {(x, y) ∈ Rn × ∂Ω : x + y ∈ K}
is compact in R2n and if we take ψ ∈ C0∞ (R2n ) such that ψ ≡ 1 in a neighborhood of
MK , we have
∂u ∂u
δ∂Ω ∗ E, ϕ = δ∂Ω (y), E(x), ψ(x, y)ϕ(x + y) (7.2.30)
∂ν ∂ν
∂u

= δ∂Ω (y), E(x)ψ(x, y)ϕ(x + y) dx .
∂ν Rn
Note that via a change of variables we have

E(x)ψ(x, y)ϕ(x + y) dx = E(z − y)ψ(z − y, y)ϕ(z) dz, ∀ y ∈ Rn . (7.2.31)

Rn Rn
If we now consider the function

h(y) := E(z − y)ψ(z − y, y)ϕ(z) dz, ∀ y ∈ Rn , (7.2.32)

Rn
then h is well defined. Moreover, since derivatives of order 2 or higher of E are

not integrable near the origin, it follows that h does not belong to C ∞ (Rn ), hence
(7.2.31) may not be used to rewrite the last term in (7.2.30). To fix this drawback,
we impose an additional restriction on ψ. Specifically, since ∂Ω ∩ K = ∅, there
exists ε > 0 sufficiently small such that (∂Ω + B(0, ε)) ∩ (K + B(0, ε)) = ∅. Then, if
U, the neighborhood of MK where ψ ≡ 1, is such that whenever (x, y) ∈ U we have
y ∈ ∂Ω + B(0, ε/2), we may further require that ψ(·, y) = 0 for y ∈ K + B(0, ε/2).
Under this requirement, if z ∈ K = supp ϕ and y ∈ Rn is such that |z − y| < ε/2,
then ψ(z − y, y) = 0. This ensures that derivatives in y of any order of the function
E(z − y)ψ(z − y, y)ϕ(z) are integrable in z over Rn , thus h ∈ C ∞ (Rn ). Now (7.2.31)
may be used in the last term in (7.2.30) to obtain
∂u ∂u

δ∂Ω ∗ E, ϕ = δ∂Ω (y), E(z − y)ψ(z − y, y)ϕ(z) dz . (7.2.33)
∂ν ∂ν Rn
∂u
Since ∂ν ∈ C 0 (∂Ω), by (7.2.25) one has
∂u

δ∂Ω (y), E(x − y)ψ(x − y, y)ϕ(x) dx
∂ν Rn

∂u
= (y) ϕ(x)E(x − y)ψ(x − y, y) dx dσ(y)
∂Ω ∂ν Rn

∂u
= (y)E(x − y) dσ(y) ϕ(x) dx. (7.2.34)
Rn ∂Ω ∂ν

Combining (7.2.33) and (7.2.34) it follows that ∂u
∂ν δ∂Ω ∗ E Rn \∂Ω is of function
type and
∂u
∂u
δ∂Ω ∗ E (x) = (y)E(x − y) dσ(y), ∀ x ∈ Rn \ ∂Ω. (7.2.35)
∂ν ∂Ω ∂ν
Similarly, with ϕ and ψ as previously specified, since u ∈ C 0 (∂Ω), definition (7.2.26)

applies and we obtain
∂ ∂

(uδ∂Ω ) ∗ E, ϕ = (uδ∂Ω )(y), E(x)ψ(x, y)ϕ(x + y) dx
∂ν ∂ν Rn
∂

= (uδ∂Ω )(y), E(x − y)ψ(x − y, y)ϕ(x) dx
∂ν Rn

∂
=− u(y) ϕ(x) ν(y) · (E(x − y)) dx dσ(y)
∂Ω Rn ∂y

= ϕ(x) u(y)[ν(y) · (∇E)(x − y)] dσ(y) dx. (7.2.36)

Rn ∂Ω

Hence ∂ν∂
(uδ∂Ω ) ∗ E n is of function type and
R \∂Ω
∂
(uδ∂Ω ) ∗ E (x) (7.2.37)
∂ν

= u(y)[ν(y) · (∇E)(x − y)] dσ(y), ∀ x ∈ Rn \ ∂Ω.

∂Ω

Combining (7.2.28), the earlier comments about uRn \∂Ω , (7.2.29), (7.2.35), and
(7.2.37), we may conclude that

∂u

u(x) = E(x − y)Δu(y) dy − E(x − y) (y) dσ(y)
Ω ∂Ω ∂ν

− [ν(y) · (∇E)(x − y)]u(y) dσ(y), ∀ x ∈ Rn \ ∂Ω, (7.2.38)

∂Ω
proving (7.2.23).
7.3 Fundamental Solutions for the Bi-Laplacian 249
As the observant reader has perhaps noticed, the solid integral appearing in
(7.2.23) is simply the volume (or Newtonian) potential defined in (4.10.1) acting
on f := Δu. Starting from this observation, formula (7.2.23) also suggests the con-
sideration of two other types of integral operators associated with a given bounded
domain Ω ⊂ Rn . Specifically, for each given complex-valued function ϕ ∈ C 0 (∂Ω)
set

Dϕ (x) := − ν(y) · (∇E)(x − y) ϕ(y) dσ(y)
∂Ω

1 ν(y) · (y − x)
= ϕ(y) dσ(y),
x ∈ Rn \ ∂Ω, (7.2.39)
ωn−1 ∂Ω |x − y|n

and S ϕ (x) := ∂Ω E(x − y)ϕ(y) dσ(y), for x ∈ Rn \ ∂Ω. Thus,
⎧ −1
⎪
⎪ (n−2)ωn−1 ∂Ω |x−y|n−2 ϕ(y) dσ(y) if n ≥ 3,
⎪ 1
⎨
S ϕ (x) = ⎪
⎪ (7.2.40)
⎪
⎩ 1
2π ∂Ω ln |y − x|ϕ(y) dσ(y) if n = 2.
The operators D and S are called, respectively, the double and single layer poten-
tials for the Laplacian. In this notation, (7.2.23) reads
⎧
∂u ⎪
⎨ u(x), x ∈ Ω,
⎪
NΩ (Δu) (x) − S (x) + D u∂Ω (x) = ⎪ ⎪ (7.2.41)
∂ν ⎩ 0, x ∈ Rn \ Ω,
for every u ∈ C 2 (Ω ). In particular, this formula shows that we may recover any
function u ∈ C 2 (Ω ) knowing Δu in Ω as well as ∂u∂ν and u on ∂Ω.
The above layer potential operators play a crucial role in the treatment of bound-
ary value problems. In this vein, we invite the reader to check that the version of the
double layer operator (7.2.39) corresponding to the case when Ω = Rn+ is precisely
twice the operator P from (4.8.3) whose relevance in the context of boundary value
problems for the Laplacian has been highlighted in (4.8.17)–(4.8.18).
7.3 Fundamental Solutions for the Bi-Laplacian

n
The bi-Laplacian in Rn is the operator Δ2 = ΔΔ = ∂2j 2 , sometimes also
j=1
referred to as the biharmonic operator. Functions u of class C 4 in an open
set Ω of Rn satisfying Δ2 u = 0 pointwise in Ω are called biharmonic in Ω.
Theorem 5.14 guarantees the existence of E ∈ S (Rn ) such that Δ2 E = δ in

S (Rn ). The goal in this section is to determine all such fundamental solution for
Δ2 . In the case when n ≥ 4, these may be computed by following the approach
from Section 7.1 (corresponding to n ≥ 2 there). Such a line of reasoning will then
require treating the cases n = 1, 2, 3 via another method. In what follows, we will
proceed differently by employing the fact that we have a complete description of
all fundamental solutions in S (Rn ) for the Laplacian and that Δ2 = ΔΔ. This latter
approach, will take care of the cases n ≥ 2, leaving n = 1 to be treated separately.
Fix E ∈ S (Rn ) satisfying Δ2 E = δ in D (Rn ). By (4.1.33) we have Δ2 E = δ
in S (Rn ). Keeping in mind that Δ2 = ΔΔ, it is natural to consider the following
equation:
ΔE = EΔ in S (Rn ), (7.3.1)
where EΔ is the fundamental solution for the Laplacian from (7.1.12). Note that
any E as in (7.3.1) is a fundamental solution for Δ2 . We proceed by analyzing three
cases.
Case n ≥ 3. Under the current assumptions, EΔ (x) = − (n−2)ω 1

n−1
|x|2−n for each x ∈
R \ {0}. The key observation is that the following identity holds
n

Δ |x|m = m(m + n − 2)|x|m−2 , pointwise in Rn \ {0}, ∀ m ∈ R. (7.3.2)
This suggests investigating the validity of a similar identity when derivatives are
taken in the distributional sense. As seen in the next lemma, whose proof we post-
pone for later, a version of identity (7.3.2) holds in S (Rn ) for a suitable range of
exponents that depends on the dimension n.
Lemma 7.18. Let N be a real number such that N > 2 − n. Then

Δ |x|N = N(N + n − 2)|x|N−2 in S (Rn ). (7.3.3)
In view of Lemma 7.18 and the definition of EΔ , it is justified to look for a solution
to (7.3.1) of the form E(x) = cn |x|4−n , x ∈ Rn \ {0}, where cn is a constant to be
determined. This candidate is in S (Rn ) (recall Exercise 4.5) and satisfies (7.3.1) if
2cn (4 − n) = − (n−2)ω
1
n−1
. Hence, we obtain cn = 2(n−2)(n−4)ω
1
n−1
if n 4.
To handle the case n = 4, we use another result that we state next (for a proof see
the last part of this section).

Lemma 7.19. Let n ≥ 3. Then Δ ln |x| = (n − 2)|x|−2 in S (Rn ).
Lemma 7.19 suggests to take when n = 4 the candidate E(x) = c ln |x| for each
x ∈ Rn \ {0}, where c is a constant to be determined. From Example 4.9 we know
that E ∈ S (Rn ). Also, Lemma 7.19 and the expression of EΔ corresponding to
n = 4 imply that E(x) = c ln |x| satisfies (7.3.1) if 2c = − 2ω1 3 . Since ω3 = 2π2 (recall
(14.5.6) and (14.5.3)), the latter implies c = − 8π1 2 .
In summary, we proved that
⎧
⎪
⎪
⎪
1
⎪
⎪ |x|4−n for n ≥ 3, n 4,
⎨ 2(n − 2)(n − 4)ωn−1
E(x) = ⎪
⎪ x ∈ Rn \ {0},
⎪
⎪
⎪ (7.3.4)
⎩ − 1 ln |x| for n = 4,
8π2
satisfies E ∈ S (Rn ) and Δ2 E = δ in S (Rn ), n ≥ 3.

The treatment of the case n = 2 is different from the above considerations since,
in such a setting, EΔ (x) = 2π1
ln |x| for x ∈ R2 \ {0}. Given the format of EΔ , for some
insight on how to choose a candidate E we start with the readily verified identity

Δ |x|m ln |x| = |x|m−2 m(m + n − 2) ln |x| + 2m + n − 2
(7.3.5)
pointwise in Rn \ {0}, ∀ m ∈ R.
The function |x|a belongs to L(Rn ) only when a ∈ N0 and is even, in which case
|x|a ln |x| ∈ S (Rn ) (by Example 4.9 and (b) in Theorem 4.14). Hence, these restric-
tions should be taken into account when considering the analogue of identity (7.3.5)
in the distributional sense in S (Rn ). The latter is stated next and its proof, which is
very much in the spirit of the proofs for Lemmas 7.18 and 7.19 (here Exercise 4.108
is also relevant), is left as an exercise.
Lemma 7.20. Let N be a real number such that N > 2 − n. Then

Δ |x|N ln |x| = |x|N−2 N(N + n − 2) ln |x| + 2N + n − 2 in S (Rn ). (7.3.6)
We are ready to proceed in earnest with considering:

Case n = 2. In view of Lemma 7.20, it is natural to start with a candidate of the form
E(x) := c|x|2 ln |x|, x ∈ R2 \ {0}, with c a constant yet to be determined. Lemma 7.20
applied for N = 2 = n implies ΔE = 4c ln |x| + 4c in S (R2 ), hence (7.3.1) is satisfied
for this E provided 4c = 2π 1
, or equivalently, if c = 8π
1
. The bottom line is that, for

this value of c, we have E ∈ S (R ) and ΔE = EΔ + 2π
2 1
in S (R2 ). Hence, since

Δc = 0 in S (R ), we conclude that
2
E(x) = 8π |x|
1 2
ln |x| for x ∈ R2 \ {0},
(7.3.7)
satisfies E ∈ S (R2 ) and Δ2 E = δ in S (R2 ).
Finally, we are left with:
Case n = 1. In this situation we use Exercise 5.16 to conclude that any fundamental
solution for Δ2 that is a tempered distribution has the form 16 x3 H + P for some
polynomial P in R of degree less than or equal to 3.
The main result emerging from this analysis is summarized next.
Theorem 7.21. The function E ∈ Lloc

1
(Rn ) defined as
⎧
⎪
⎪
⎪
1
⎪
⎪ |x|4−n if x ∈ Rn \ {0}, n ≥ 3, n 4,
⎪
⎪
⎪ 2(n − 2)(n − 4)ω n−1
⎪
⎪
⎪
⎪
⎪
⎨ − 8π1 2 ln |x| if x ∈ R4 \ {0}, n = 4,
E(x) := ⎪
⎪ (7.3.8)
⎪
⎪
⎪
⎪
⎪ 8π |x| ln |x| if x ∈ R \ {0}, n = 2,
1 2 2
⎪
⎪
⎪
⎪
⎪
⎪
⎩ 1 x3 H if x ∈ R, n = 1,
6
belongs to S (Rn ) and is a fundamental solution for the operator Δ2 in Rn . Moreover,

u ∈ S (Rn ) : Δ2 u = δ in S (Rn ) (7.3.9)

= E + P : P biharmonic polynomial in Rn .
Proof. It is clear from (7.3.8) that E ∈ Lloc

1
(Rn ) ∩ S (Rn ). The fact that Δ2 E = δ in

S (R ) has been checked in (7.3.4), (7.3.7), and Exercise 5.16. Finally, (7.3.9) is a
n
consequence of Proposition 5.8.

We now turn to the task of proving Lemmas 7.18 and 7.19.
Proof of Lemma 7.18. Fix N > 2 − n. By Exercise 4.5 we have that both |x|N−2 and
|x|N belong to S (Rn ). Thus, invoking (4.1.33), there remains to show that
N
Δ |x| , ϕ = N(N + n − 2)|x|N−2 , ϕ , ∀ ϕ ∈ C0∞ (Rn ). (7.3.10)
To this end, fix ϕ ∈ C0∞ (Rn ) and let R ∈ (0, ∞) be such that supp ϕ ⊂ B(0, R).
Then, starting with the definition of distributional derivatives, then using the support
condition for ϕ and Lebesgue’s Dominated Convergence Theorem, and then (14.8.5)
(keeping in mind that supp ϕ ⊂ B(0, R)), we have

N
Δ |x| , ϕ = |x| Δϕ(x) dx = lim+
N
|x|N Δϕ(x) dx
Rn ε→0 ε<|x|<R
∂ϕ
= lim+ ϕ(x)Δ|x|N dx − |x|N (x) dσ(x)
ε→0 ε<|x|<R ∂B(0,ε) ∂ν

∂|x|N
+ ϕ(x) dσ(x) , (7.3.11)
∂B(0,ε) ∂ν
where ν(x) = εx , for x ∈ ∂B(0, ε). By (7.3.2) and Lebesgue’s Dominated Conver-
gence Theorem 14.15 it follows that

lim+ ϕ(x)Δ|x|N dx = N(N + n − 2) |x|N−2 ϕ(x) dx

ε→0 ε<|x|<R Rn

= N(N + n − 2) |x|N−2 , ϕ . (7.3.12)
∂|x|N
Since ∇(|x|N ) = N|x|N−2 x for every x 0, we have ∂ν = NεN−1 on ∂B(0, ε). Hence,

N ∂ϕ ∂|x|N
− |x| (x) dσ(x) + ϕ(x) dσ(x) (7.3.13)
∂B(0,ε) ∂ν ∂B(0,ε) ∂ν
≤ ωn−1 ∇ϕL∞ (Rn ) εN+n−1 + Nωn−1 ϕL∞ (Rn ) εN+n−2 −−−−→

+
0
ε→0
given that N > 2 − n. Now (7.3.10) follows from (7.3.11)–(7.3.13). The proof of the
lemma is complete.
Proof of Lemma 7.19. From Exercise 4.5 and the assumption n ≥ 3 it follows that
|x|−2 ∈ S (Rn ), while from Example 4.9 we have ln |x| ∈ S (Rn ). Hence, by (4.1.33)
matters reduce to showing

Δ ln |x| , ϕ = (n − 2) |x|−2 , ϕ , ∀ ϕ ∈ C0∞ (Rn ). (7.3.14)
Fix ϕ ∈ C0∞ (Rn ) and let R ∈ (0, ∞) be such that supp ϕ ⊂ B(0, R). Using in the
current setting a reasoning similar to that applied when deriving (7.3.11) one obtains

Δ ln |x| , ϕ = ln |x|Δϕ(x) dx = lim+ ln |x|Δϕ(x) dx
Rn ε→0 ε<|x|<R
∂ϕ
= lim+ ϕ(x)Δ(ln |x|) dx − ln |x| (x) dσ(x)
ε→0 ε<|x|<R ∂B(0,ε) ∂ν

∂ ln |x|
+ ϕ(x) dσ(x) . (7.3.15)
∂B(0,ε) ∂ν
It is easy to see that Δ(ln |x|) = (n − 2)|x|−2 pointwise in Rn \ {0}, thus invoking
Lebesgue’s Dominated Convergence Theorem 14.15 we obtain

lim+ ϕ(x)Δ(ln |x|) dx = (n − 2) |x|−2 ϕ(x) dx

ε→0 ε<|x|<R Rn
= (n − 2)|x|−2 , ϕ . (7.3.16)
Also, since ∇(ln |x|) = |x|x2 pointwise in Rn \ {0}, we have ∂ ln∂ν|x| = ε−1 on ∂B(0, ε).
Hence,

∂ϕ ∂ ln |x|
− ln |x| (x) dσ(x) + ϕ(x) dσ(x) (7.3.17)
∂B(0,ε) ∂ν ∂B(0,ε) ∂ν
≤ ωn−1 ∇ϕL∞ (Rn ) εn−1 | ln ε| + ωn−1 ϕL∞ (Rn ) εn−2 −−−−→

+
0,
ε→0
where for the convergence in (7.3.17) we used the fact that n ≥ 3. To finish the proof
of the lemma we combine (7.3.15)–(7.3.17).
Exercise 7.22. Let E be the fundamental solution for the bi-Laplacian operator from
(7.3.8). Prove that for every α ∈ Nn0 with |α| = 3 the function ∂α E is of class C ∞ and
positive homogeneous of degree 1 − n in Rn \ {0}.
7.4 The Poisson Equation for the Bi-Laplacian
Analogously to Theorem 7.10 we have the following regularity result for the bi-
Laplacian.
Theorem 7.23. For an open set Ω ⊆ Rn the following are equivalent:

(i) u is biharmonic in Ω;
(ii) u ∈ D (Ω) and Δ2 u = 0 in D (Ω);
(iii) u ∈ C ∞ (Ω) and Δ2 u = 0 in a pointwise sense in Ω.
Proof. This is a direct consequence of Theorem 6.8 and Theorem 7.21 (or, alterna-
tively, of Remark 6.14 and Corollary 6.18).
We begin by establishing mean value formulas for biharmonic functions. Recall
the notation from (0.0.15).
Theorem 7.24. Let Ω be an open set in Rn and let u be a biharmonic function in Ω.

Then for every x ∈ Ω and every r ∈ 0, dist(x, ∂Ω) the following formulas hold:

r2
u(x) = − u(y) dσ(y) − (Δu)(x), (7.4.1)
∂B(x,r) 2n

r2
u(x) = − u(y) dy − (Δu)(x). (7.4.2)
B(x,r) 2(n + 2)
Proof. Fix x ∈ Ω and recall the function φ defined in (7.2.10). Then (7.2.11) holds.
Since Δ(Δu) = 0 in Ω we have that Δu is harmonic in Ω and we can apply (7.2.9)
to obtain that φ (r) = nr Δu(x). Consequently, φ(r) = 2n
r2
Δu(x) + C for some constant
C. To determine C we use the fact that lim+ φ(r) = u(x) (for the latter see (7.2.15)).
r→0
Hence formula (7.4.1) follows.
Next, write (7.4.1) as

ωn−1 rn+1
ωn−1 r n−1
u(x) = u(y) dσ(y) − Δu(x). (7.4.3)
∂B(x,r) n

Integrating (7.4.3) with respect to r for r ∈ (0, R) and R ∈ 0, dist (x, ∂Ω) , and then
dividing by ωn−1 Rn /n gives (7.4.2) with r replaced by R.
Theorem 7.25 (Liouville’s Theorem for Δ2 ). Any bounded biharmonic function in

Rn is constant.
Proof. One way to justify this result is by observing that it is a particular case of
Theorem 5.4. Another proof based on interior estimates goes as follows. Let u be a
bounded biharmonic function in Rn . Formula (7.4.2) in the current setting gives that

2(n + 2) 2(n + 2)
Δu(x) = − u(y) dy − u(x) (7.4.4)
r2 B(x,r) r2
7.4 The Poisson Equation for the Bi-Laplacian 255
for every x ∈ Rn and every r ∈ (0, ∞). Since

− u(y) dy ≤ uL∞ (Rn ) , (7.4.5)
B(x,r)
if we let r → ∞ in (7.4.4), we see that Δu(x) = 0, thus u is harmonic. By Liouville’s

Theorem 7.15 for harmonic functions it follows that there exists c ∈ R such that
u(x) = c for all x ∈ Rn .
We shall next show that biharmonic functions satisfy interior estimates and are
real analytic.
Theorem 7.26. Suppose u ∈ D (Ω) satisfies Δ2 u = 0 in D (Ω). Then u is real ana-
lytic in Ω and there exists a dimensional constant C ∈ (0, ∞) such that
C |α| |α|!
max |∂α u(y)| ≤ max |u(y)|, ∀ α ∈ Nn0 , (7.4.6)
y∈B(x,r/2) r|α| y∈B(x,r)

for each x ∈ Ω and each r ∈ 0, dist (x, ∂Ω) .
Proof. In the case when n = 1 or n ≥ 3, all claims are direct consequences of
Theorem 6.26 and Theorem 7.21. To treat the case n = 2 we shall introduce a
“dummy” variable. Specifically, in this setting consider the open set Ω := Ω × R in
R and define the function
3
u(x1 , x2 , x3 ) := u(x1 , x2 ) for (x1 , x2 , x3 ) ∈ Ω. Observe that
As such, the higher dimensional theory applies and yields that
u is biharmonic in Ω.

u satisfies, for some universal constant C ∈ (0, ∞),
C |α| |α|!
max |∂α
u(y)| ≤ max |u(y)|, ∀ α ∈ N30 , (7.4.7)
for each x ∈ Ω and each r ∈ 0, dist (x, ∂Ω)

. Applying (7.4.7) for points of the form
and for multi-indices α = (α1 , α2 , 0) then yields (7.4.6) for the
x = (x1 , x2 , 0) ∈ Ω
original function u. In turn, this and Lemma 6.24 imply that u is real analytic in Ω.

We are now prepared to discuss the well-posedness of the Poisson problem for
the bi-Laplacian in Rn .
Theorem 7.27. Assume n ∈ N satisfies n ≥ 3 and n 4. Then for each function
∞
f ∈ Lcomp (Rn ) and each c ∈ C the Poisson problem for the bi-Laplacian in Rn ,
⎧
⎪
⎪
⎪ u ∈ C 0 (Rn ),
⎪
⎪
⎪
⎪
⎨ Δ2 u = f in D (Rn ),
⎪
⎪
⎪
(7.4.8)
⎪
⎪
⎪
⎪
⎩ lim u(x) = c,
|x|→∞
properties.
(1) The function u has the integral representation formula

u(x) = c + E(x − y) f (y) dy, x ∈ Rn , (7.4.9)

Rn
where E is the fundamental solution for the bi-Laplacian in Rn from (7.3.8).

Moreover, u belongs to C 3 (Rn ) and for each α ∈ Nn0 with 0 < |α| ≤ 3 we have

∂α u(x) = (∂α E)(x − y) f (y) dy, x ∈ Rn . (7.4.10)

Rn
(2) If in fact f ∈ C0∞ (Rn ) then u ∈ C ∞ (Rn ).

(3) For every α ∈ Nn0 with |α| = 4 there exists a constant cα ∈ C such that
∂α u = cα f + T ∂α E f in D (Rn ), (7.4.11)
where T ∂α E is the singular integral operator associated with Θ := ∂α E as in

Definition 4.93.
(4) For every p ∈ (1, ∞), the solution u of (7.4.8) satisfies ∂α u ∈ L p (Rn ) for each
α ∈ Nn0 with |α| = 4, where the derivatives are taken in D (Rn ). Moreover, there
exists a constant C = C(p, n) ∈ (0, ∞) with the property that

n
∂α uL p (Rn ) ≤ C f L p (Rn ) . (7.4.12)
|α|=4
Proof. That the function u defined as in (7.4.9) is of class C 3 (Rn ) and formula
(7.4.10) holds for each α ∈ Nn0 with |α| ≤ 3 can be established much as in the
proof of Proposition 7.8 keeping in mind that ∂α E ∈ Lloc 1
(Rn for each α ∈ Nn0 with

|α| ≤ 3. Also, this function solves Δ u = f in D (R ) by reasoning in a similar
2 n
fashion to the computation in (7.2.7). Given the format of E from (7.3.8) under the
current assumptions on n, and the conditions on f , it is clear that the function u from
(7.4.9) satisfies the limit condition in (7.4.8) (see (7.2.21) in the case of the Laplace
operator). The fact that (7.4.9) is the unique solution of (7.4.8) is a consequence of
Theorem 7.25. Moving on, that u ∈ C ∞ (Rn ) if f belongs to C0∞ (Rn ) follows from
Corollary 6.18 and the ellipticity of Δ2 . The above arguments cover the claims in
parts (1)–(2).
Turning to part (3), start by fixing α ∈ Nn0 with |α| = 4. Then there exists some
β ∈ Nn0 with |β| = 3 and j ∈ {1, . . . , n} such that ∂α = ∂ j ∂β . If we define Φ := ∂β E,
it follows that Φ is C ∞ and positive homogeneous of degree 1 − n in Rn \ {0} (cf.
Exercise 7.22). Consequently, the function Θ := ∂α E is as in (4.4.1), thanks to
the discussion in Example 4.71. Then from what we have proved in part (1) and
Theorem 4.103 we deduce that the following formula holds in D (Rn ):
7.5 Fundamental Solutions for the Poly-harmonic Operator 257

∂α u(x) = ∂ j ∂β u(x) = ∂ j (∂β E)(x − y) f (y) dy (7.4.13)
Rn

β
= (∂ E)(ω)ω j dσ(ω) f (x) + lim+ (∂α E)(x − y) f (y) dy.
S n−1 ε→0 |x−y|>ε

Choosing cα := S n−1 (∂β E)(ω)ω j dσ(ω) the above formula may be written as in
(7.4.11), and this finishes the proof of part (3). Finally, the claim in (4) follows
from (3) and the boundedness of the singular integral operators T ∂α E on L p (Rn ) (cf.
Theorem 4.101).
7.5 Fundamental Solutions for the Poly-harmonic Operator
Let m ∈ N and consider the poly-harmonic operator

⎛ n ⎞
⎜⎜⎜ ⎟⎟⎟m m!
Δm := ⎜⎜⎜⎝ ∂2j ⎟⎟⎟⎠ = ∂2γ , in Rn . (7.5.1)
j=1 |γ|=m
γ!
The goal in this section is to identify all fundamental solutions for this operator in
Rn that are tempered distributions. The case m = 1 has been treated in Theorem 7.2
and the case m = 2 is discussed in Theorem 7.21. We remark that the formulas
from Lemma 7.18, Lemma 7.19, and Lemma 7.20, suggest that we should look for
a fundamental solution for Δm that is a constant multiple of |x|2m−n or a constant
multiple of |x|2m−n ln |x|. In precise terms, we will prove the following result (recall
that H denotes the Heaviside function).
Theorem 7.28. Let m, n ∈ N and consider the function

⎧
⎪
⎪ (−1)m Γ(n/2 − m) 2m−n
⎪
⎪
⎪ |x| if n > 2m or n is odd and n 1,
⎪
⎪
⎪ πn/2 4m (m − 1)!
⎪
⎪
⎪
⎪
⎪
⎪
⎨ (−1)n/2+1 |x|2m−n
Fm,n (x) := ⎪
⎪
⎪ ln |x| if n is even and ≤ 2m, (7.5.2)
⎪
⎪
⎪ 2π 4 (m − n/2)!(m − 1)!
n/2 m−1
⎪
⎪
⎪
⎪
⎪
⎪ 1
⎪
⎪
⎩ x2m−1 H if n = 1,
(2m − 1)!
defined for x ∈ Rn \ {0} if n ≥ 2 and for x ∈ R if n = 1. Then Fm,n belongs to
1
Lloc (Rn ) ∩ S (Rn ) and is a fundamental solution for Δm in Rn . Moreover,

u ∈ S (Rn ) : Δm u = δ in S (Rn ) (7.5.3)

= Fm,n + P : P poly-harmonic polynomial in Rn .
Proof. The case n = 1 is immediate from Exercise 5.16. Assume in what follows
that n ≥ 2. Since n − 2m < n, we clearly have |x|2m−n ∈ Lloc 1
(Rn ) and furthermore, by

Exercise 4.5, that |x|2m−n
∈ S (R ). By Exercise 2.124 and Exercise 4.108 we also
n
have |x|2m−n ln |x| ∈ Lloc

1
(Rn ) ∩ S (Rn ). This proves that Fm,n ∈ Lloc1
(Rn ) ∩ S (Rn ).
Once we show that Fm,n is a fundamental solution for Δ in R , Proposition 5.8
m n
readily yields (7.5.3).

There remains to prove Δm Fm,n = δ in S (Rn ). If m = 1, then it is easy to see that
F1,n is the same as the distribution in (7.1.12) (corresponding to n ≥ 2). Thus, by
Theorem 7.2 we have
ΔF1,n = δ in S (Rn ). (7.5.4)
We proceed by breaking up the remaining part of our analysis in a few cases.
The Case n > 2m or n is Odd.
Because of (7.5.4) we may assume m ≥ 2. Applying Lemma 7.18 and (14.5.2) we
obtain
(−1)m Γ(n/2 − m)
ΔFm,n = Δ(|x|2m−n )
πn/2 4m (m − 1)!
(−1)m Γ(n/2 − m)
= (2m − n)(2m − 2)|x|2m−n−2
πn/2 4m (m − 1)!
(−1)m−1 Γ(n/2 − m)(n/2 − m)(m − 1) 2(m−1)−n
= |x|
πn/2 4m−1 (m − 1)!
= Fm−1,n in S (Rn ) and pointwise in Rn \ {0}. (7.5.5)
Hence, inductively we have Δm−1 Fm,n = F1,n in S (Rn ), which when combined with
(7.5.4) yields the desired conclusion in the current case.
The Case n = 2m.
In this scenario n is even, and the goal is to prove
in S (Rn ).
n
Δ 2 F 2n ,n = δ (7.5.6)
Note that (7.5.6) holds for n = 2 as seen from (7.5.4). Consequently, we may assume
that n ≥ 4. Inductively, using Lemma 7.18, it is not difficult to show that

k −2
(−1)k 4k k! n2 − 2 ! −2−2k
Δ |x| = n |x| in S (Rn )
2 −k−2 ! (7.5.7)
n
if n is even, n ≥ 4 and k ∈ N0 is such that k < − 1.
2
In particular, (7.5.7) specialized to k = n
2 − 2 yields
n n 2
Δ 2 −2 |x|−2 = (−1) 2 4 2 in S (Rn ).
n n
− 2 ! |x|2−n (7.5.8)
2
Starting with the expression from (7.5.2) we obtain
(−1) 2 +1
n
−1
Δ 2 −1 F n2 ,n =
n n
n n
−1 n Δ 2 ln |x|
2π 2 4 2 2 − 1 !
(−1) 2 +1
n
n
−2 −2
= n n
−1 n (n − 2)Δ 2 |x|
2π 4 2 2 − 1 !
2
(−1) 2 +1 n n 2
n
n
= n n
−1 (n − 2)(−1) 2 4 2 − 2 ! |x|2−n
2π 4 2 2 − 1 !
2
n 2
−1
= |x|2−n in S (Rn ). (7.5.9)
(n − 2)ωn−1
The second equality in (7.5.9) is based on Lemma 7.19, the third equality uses
(7.5.8), while the last equality follows by straightforward calculations and (14.5.6).
In summary, we proved that if n is even and n ≥ 4, then Δ 2 −1 F 2n ,n = EΔ in S (Rn ),
n
where EΔ is the fundamental solution for the Laplacian from (7.1.12). The latter
combined with Theorem 7.2 now yields (7.5.6) for n ≥ 4, and finishes the proof of
(7.5.6).
The Case n ≤ 2m and n is Even.
Fix an even number n ∈ N. We will prove that if m ≥ n
2 then
Δm Fm,n = δ in S (Rn ) (7.5.10)
by induction on m. The case m = n2 has been treated in the previous case. Suppose
(7.5.10) holds for some m ≥ n2 . Then clearly 2(m + 1) > n and invoking Lemma 7.20
we obtain
ΔFm+1,n = Fm,n + Cm+1,n (4m − n + 2)|x|2m−n in S (Rn ), (7.5.11)
where Cm+1,n is the coefficient of |x|2(m+1)−n ln |x| in the expression for Fm+1,n . Note
that since n is even the expression |x|2m−n is a polynomial of degree 2m − n. Hence,
Δm [|x|2m−n ] = 0 given that Δm is a homogeneous differential operator of degree
2m > 2m − n. Combined with (7.5.11) and the induction hypothesis on Fm,n , this
gives
Δm+1 Fm+1,n = Δm Fm,n + Cm+1,n (4m − n + 2)Δm [|x|2m−n ] = δ (7.5.12)
in S (Rn ). By induction, it follows that (7.5.10) holds in the current setting. This
finishes the proof of the theorem.
We remark that Theorem 7.28 specialized to m = 2 gives another proof of the
fact that when n ≥ 2 the expression from (7.3.8) is a fundamental solution for Δ2 in
Rn .
Proposition 7.29. Let m, n ∈ N be such that n ≥ 2 and let Fm,n be the fundamental
solution for Δm in Rn as defined in (7.5.2). Then if N ∈ N0 is such that 0 ≤ N < m
the following identity holds

ΔN Fm,n = Fm−N,n + Pm,N in S (Rn ) and pointwise in Rn \ {0}, (7.5.13)
for some homogeneous polynomial Pm,N in Rn of degree at most max {2m−n−2N, 0}

which is identically zero if n > 2m or n is odd, or if n is even such that n ≤ 2m and
N > m − n2 (in short, Pm,N ≡ 0 if either n is odd or 2N + n > 2m).
Proof. We start by observing that using an inductive reasoning and Lemma 7.20 we
obtain that for each N ∈ N0

ΔN |x| M ln |x| = C (1)
M,N |x|
M−2N
ln |x| + C (2)
M,N |x|
M−2N
in S (Rn )
(7.5.14)
if M is a real number such that M − 2N > −n,
for some real constants C (1) (2)

M,N , C M,N , depending on n, M, and N.
Now suppose n is even and n ≤ 2m. Specializing (7.5.14) to the case M := 2m−n

and taking N ∈ 0, 1, . . . , m − n2 , we conclude
ΔN Fm,n = c(1) (2)

m,N F m−N,n + cm,N |x|
2m−n−2N
in S (Rn ), (7.5.15)
for some real constants c(1) (2)

m,N , cm,N , depending only on n, m, and N. For each number
n
N ∈ 0, 1, . . . , m − 2 set
Pm,N (x) := c(2)

m,N |x|
2m−n−2N
, x ∈ Rn .
Note that since n is even and n ≤ 2m, each Pm,N is a homogeneous polynomial

of degree at most 2m − n − 2N. Also, for each N ∈ 0, 1, . . . , m − n2 the operator
Δm−N is a homogeneous differential operator of degree 2m − 2N > 2m − n − 2N,
thus Δm−N Pm,N = 0 both in S (Rn ) and pointwise in Rn . The latter, (7.5.15), and
Theorem 7.28 imply
m−N
δ = Δm−N ΔN Fm,n = c(1)
m,N Δ Fm−N,n + Δm−N Pm,N

= c(1)
m,N δ in S (R )
n
(7.5.16)
n
for each N ∈ 0, 1, . . . , m − n2 . Thus, c(1)
m,N = 0 for each N ∈ 0, 1, . . . , m − 2 .
Consequently, (7.5.15) becomes
n
ΔN Fm,n = Fm−N,n + Pm,N in S (Rn ), ∀ N ∈ 0, 1, . . . , m − . (7.5.17)
2
That the above equality also holds pointwise in Rn is immediate, hence we proved
(7.5.13) for n even such that n ≤ 2m and 0 ≤ N ≤ m − n2 . Note that if n = 2 then

m − n2 = m − 1 so we actually have proved (7.5.13) for all N ∈ 0, 1, . . . , m − 1 .
To finish with the proof of (7.5.13) when n is even and satisfies n ≤ 2m, there

remains the case N ∈ m − n2 + 1, . . . , m − 1 under the assumption that n ≥ 4. For
starters, observe that if we write (7.5.17) for N = m − n2 , then Pm,N is just a constant
and using the expression for F 2n ,n we have
in S (Rn ),
n (2) (2)
Δm− 2 Fm,n = F 2n ,n + Cm,m− n = cn ln |x| + c
m,m− n
(7.5.18)
2 2
for some real constant cn . Therefore, (7.5.18), Lemma 7.19, and Lemma 7.18 (which
we can apply N − m + n2 − 1 times since N < m) allow us to conclude that for each

N ∈ m − n2 + 1, . . . , m − 1 there exists some constant cm,N such that
n
ΔN Fm,n = ΔN−m+ 2 Δm− 2 Fm,n = cm,N |x|−2N+2m−2n in S (Rn ).
n
(7.5.19)
Upon observing that the condition N > m − n2 is equivalent with n > 2(m − N), we
may conclude that cm,N |x|−2N+2m−2n = c(1) (1)
m,N F m−N,n for some constant cm,N . Hence, for

each N ∈ m − 2 + 1, . . . , m − 1 we have
n
ΔN Fm,n = c(1)
m,N F m−N,n in S (Rn ). (7.5.20)
Applying Δm−N to (7.5.20) and invoking Theorem 7.28 we obtain c(1)

m,N = 0 for each

N ∈ m − n2 + 1, . . . , m − 1 , so that
n
ΔN Fm,n = Fm−N,n in S (Rn ), ∀ N ∈ m − + 1, . . . , m − 1 . (7.5.21)
2
This proves (7.5.13) when n is even, satisfies n ≤ 2m, and m − n2 < N < m if we take
Pm,N := 0.
Suppose next that n > 2m or n is odd. Then as seen in the proof of Theorem 7.28,
identity (7.5.12) holds in S (Rn ) and pointwise in Rn \ {0}. Iterating this identity we
arrive at

ΔN Fm,n ] = Fm−N,n in S (Rn ) and pointwise in Rn \ {0}. (7.5.22)
Consequently, formula (7.5.13) holds in the current setting if we take Pm,N := 0 for
all N ∈ {0, . . . , m − 1}. This completes the proof of the proposition.
We next address the issue of interior estimates and real analyticity for poly-
harmonic functions.
Theorem 7.30. Whenever u ∈ D (Ω) satisfies Δm u = 0 in D (Ω), it follows that u is

real analytic in Ω and there exists a constant C = C(n, m) ∈ (0, ∞) with the property
that
C |α| |α|!
max |∂α u(y)| ≤ max |u(y)|, ∀ α ∈ Nn0 , (7.5.23)

for each x ∈ Ω and each r ∈ 0, dist (x, ∂Ω) .
Proof. In the case when either n is odd, or n ≥ 2m, all claims are consequences of
Theorem 6.26 and Theorem 7.28. The remaining cases may be reduced to the ones
just treated by introducing “dummy” variables, as in the proof of Theorem 7.26.
In the last part of this section we prove an integral representation formula for a
fundamental solution for the poly-harmonic operator in Rn involving a sufficiently
large power of the Laplacian (i.e., Δm with m positive integer bigger than or equal
to n/2). The difference between this fundamental solution and that in (7.5.2) turns
out to be a homogeneous polynomial that is a null solution of the respective poly-
harmonic operator. In what follows log denotes the principal branch of the complex
logarithm that is defined for points z ∈ C \ (−∞, 0].
Lemma 7.31. Let n ∈ N, n ≥ 2, and suppose q ∈ N0 is such that n + q is an even

number. Then the function

x · ξ
1
Eq (x) := − (x · ξ)q log dσ(ξ), ∀ x ∈ Rn \ {0}, (7.5.24)
(2πi)n q! S n−1 i
n+q
is a fundamental solution for the poly-harmonic operator Δ 2 in Rn .
Moreover, Eq ∈ S (Rn ) and
Eq (x) = F n+q
2 ,n
(x) + C(n, q)|x|q , ∀ x ∈ Rn \ {0}, (7.5.25)
n+q
where F n+q
2 ,n
is the fundamental solution for Δ 2 in Rn given in (7.5.2) and
⎧
1 √
⎪
⎪
⎪ 2ωn−2
⎪
⎨ (2πi)n q! 0 s (ln s)( 1 − s ) ds if n is even,
⎪ − q 2 n−3
C(n, q) := ⎪
⎪ (7.5.26)
⎪
⎪
⎪
⎩0 if n is odd.
Proof. Fix n and q as in the hypotheses of the lemma. Fix x ∈ Rn \ {0} and use the

fact that log x·ξ
i = ln |x · ξ| − i π2 sgn (x · ξ) for every ξ ∈ S n−1 to write

x · ξ
(x · ξ)q log dσ(ξ)
S n−1 i

= (x · ξ)q ln |x · ξ| − i π2 dσ(ξ)
ξ∈S n−1 , x·ξ>0

π
+ (−1)q (x · ξ)q ln |x · ξ| + i dσ(ξ)
ξ∈S n−1 , x·ξ>0 2

1 π
= |x · ξ|q ln |x · ξ| − i dσ(ξ)
2 S n−1 2
q

(−1)
+ |x · ξ|q ln |x · ξ| + i π2 dσ(ξ)
2 S n−1
q
1 + (−1)
= |x · ξ|q ln |x · ξ| dσ(ξ)
2 S n−1

((−1)q − 1)πi
+ |x · ξ|q dσ(ξ) (7.5.27)
4 S n−1
To compute the integrals in the right-most side of (7.5.27) we use Proposition 14.65.
First, applying (14.9.15) with f (t) := |t|q for t ∈ R \ {0} and v := x, we obtain

1 √
|x · ξ| dσ(ξ) = 2ωn−2 |x|
q q
sq ( 1 − s2 )n−3 ds (7.5.28)
S n−1 0

π
2 q n−2
= 2ωn−2 |x|q sin θ cos θ dθ
0
q+1 n−1 n−1
Γ 2 Γ 2 2π 2 Γ q+1
= ωn−2 |x| =
q
q+n |x| .
2 q
Γ q+n
2 Γ 2
For the third equality in (7.5.28) we used (14.5.10), while the last one is based on
(14.5.6). Second, formula (14.9.15) with f (t) := |t|q ln |t| for t ∈ R and v := x, in
concert with (7.5.28) further yields

√
1

|x · ξ|q ln |x · ξ| dσ(ξ) = 2ωn−2 |x|q sq ln |x| + ln s ( 1 − s2 )n−3 ds
S n−1 0
n−1 q+1
2π 2 Γ 2
= q+n |x|q ln |x| + c(n, q)|x|q , (7.5.29)
Γ 2
where
1 √
c(n, q) := 2ωn−2 sq (ln s)( 1 − s2 )n−3 ds < ∞. (7.5.30)
0
Hence, a combination of (7.5.27), (7.5.28), (7.5.29), and (7.5.24) implies
⎧
⎪
⎪
⎪ (−1) 2 +1 Γ q+1
n
c(n, q)
⎪
⎪
⎪
2
|x|q ln |x| − |x|q if n is even,
⎪
⎪
⎪ n−1 π
n+1
2 q! Γ
q+n (2πi) n q!
⎪
⎨ 2 2
Eq (x) = ⎪
⎪ (7.5.31)
⎪
⎪
⎪
n−1 q+1
⎪
⎪ (−1) 2 Γ 2
⎪
⎪
⎪ q+n |x|
q
if n is odd,
⎩ 2n π n−1
2 q! Γ
2
for every x ∈ Rn \ {0}. That this expression belongs to S (Rn ) follows from Exer-
cise 4.108.
Let F n+q
2 ,n
be as in (7.5.2). The goal is to prove that
c(n, q) q
Eq − F n+q
2 ,n
= |x| pointwise in Rn \ {0}. (7.5.32)
(2πi)n q!
Once such an identity is established, Theorem 7.28 may be used to conclude that Eq
n+q
is a fundamental solution for Δ 2 in Rn (note that if n is even, thenn+qq is even and
|x| is a homogeneous polynomial of order q that is annihilated by Δ 2 ). Also, it is
q
c(n,q)
easy to see that C(n, q) = (2πi)n q! .
With an eye toward proving (7.5.32), suppose first that n is even. Choosing m :=
n+q
2 in (7.5.2), we have n ≤ 2m, hence
(−1) 2 +1
n
F n+q ,n (x) = |x| ln |x|,

q
∀ x ∈ Rn \ {0}. (7.5.33)
π 2 2n+q−1 q2 ! n+q
n
2
2 − 1 !
Applying (14.5.4) with q/2 in place of n, and using (14.5.3) (recall that n+q is even)
we have
q + 1 q! π1/2 q + n q + n
Γ = and Γ = − 1 !. (7.5.34)
2 2q (q/2)! 2 2
Using now (7.5.34) in (7.5.33) and the expression for Eq , it follows that (7.5.32)
holds.
If next we assume that n is odd, then

n+q
(−1) Γ − q2
2
F n+q
2 ,n
(x) = n n+q n+q |x| ln |x|,
q
∀ x ∈ Rn \ {0}. (7.5.35)
π2 2 2 −1 !
Applying (14.5.5) and (14.5.3) with (q + 1)/2 in place of n we see that

q+1
q + 1 q! π1/2 q (−1)(q+1)/2 2q+1 2 !π1/2
Γ = q and Γ − = (7.5.36)
2 2 (q/2)! 2 (q + 1)!
which in concert with the second equality in (7.5.34), the expression in (7.5.35), and
the formula for Eq , implies Eq (x) = F n+q
2 ,n
(x) in Rn \ {0}. The proof of the lemma is
now complete.
7.6 Fundamental Solutions for the Helmholtz Operator 265
7.6 Fundamental Solutions for the Helmholtz Operator
For the purpose of the discussion in this section let k ∈ (0, ∞) be arbitrary, fixed. The
Helmholtz operator in Rn , n ∈ N, is the operator Δ + k2 . The goal for us is to deter-
mine a fundamental solution for the Helmholtz operator which also enjoys a specific
decay condition at infinity. As it happens, in dimensions n ≥ 2, this fundamental so-
lution is related to the Hankel function of the first kind which is a null solution for
the Bessel differential operator. As such, we first introduce the latter operator.
The ordinary Bessel differential operator with parameter λ ∈ R acting on a

complex-valued function v ∈ C 2 (0, ∞) is defined as
(Bλ v)(r) := r2 v (r) + rv (r) + (r2 − λ2 )v(r), r > 0. (7.6.1)
To reveal the connection between the Bessel differential operator and the Helmholtz
operator Δ+k2 in Rn , assume that a smooth radial complex-valued function u defined

in Rn \ {0}, has been given. Then there exists w ∈ C ∞ (0, ∞) such that
u(x) = w(|x|) for each x ∈ Rn \ {0}. (7.6.2)
Via differentiation, formula (7.6.2) implies that for each j ∈ {1, . . . , n} and each
x ∈ Rn \ {0}, we have
xj
(∂ j u)(x) = w (|x|) · |x| and
x2j x2j
(7.6.3)
(∂2j u)(x) = w (|x|) · |x|2
+ w (|x|) · 1
|x| − w (|x|) · |x|3
,
so that
n−1
[(Δ + k2 )u](x) = w (|x|) + w (|x|) · + k2 w(|x|) (7.6.4)
|x|
for each x ∈ Rn \ {0}. To bring to light the Bessel differential operator define
v(r) := r(n−2)/2 w(r/k) for each r ∈ (0, ∞). (7.6.5)
Taking derivatives we obtain
v (r) = w (r/k),
n−4 n−2
n−2
2 r 2 w(r/k) + 1k r 2
(7.6.6)
v (r) = w (r/k) + w (r/k),
(n−2)(n−4) n−6 n−4 n−2
4 r 2 w(r/k) + n−2
k r 2
1
k2
r 2
for each r ∈ (0, ∞). This may be then used to check that

1 n+2 k
B(n−2)/2 v(r) = 2 r 2 w (r/k) + (n − 1)w (r/k) + k2 w(r/k) (7.6.7)
k r
for each r ∈ (0, ∞). From (7.6.7) and (7.6.4) it follows that

(Δ + k2 )u (x) = (k|x|)−(n+2)/2 k2 (B(n−2)/2 v)(k|x|) (7.6.8)
for each x ∈ Rn \ {0}. Having obtained (7.6.8), it is natural to consider the Hankel
function of the first kind with index λ ∈ R, denoted by Hλ(1) (·), which is a null
solution of Bλ in the sense that
Bλ Hλ(1) = 0 on (0, ∞), for each λ ∈ R. (7.6.9)
(The definition of the Hankel function of the first kind and some of the properties
this function enjoys are reviewed in Appendix 14.10.) In particular, corresponding to
(1)
choosing v := H(n−2)/2 , which via (7.6.5) and (7.6.2) and r = k|x|, yields a function
u(x) of the form (k|x|)−(n−2)/2 H(n−2)/2
(1)
(k|x|), the identities (7.6.9) and (7.6.8) imply
that
if n ≥ 2 and u(x) := |x|−(n−2)/2 H(n−2)/2(1)
(k|x|) for
∞
each x ∈ R \ {0}, then u ∈ C (Rn \ {0}) and u
n (7.6.10)
satisfies the equation (Δ + k2 )u = 0 in Rn \ {0}.
In fact, as a function defined a.e. in Rn , it turns out that u is locally integrable and a
fundamental solution for the Helmholtz operator in Rn , n ≥ 2, may be obtained by
suitably normalizing u from (7.6.10).
Specifically, consider the function
⎧
⎪
⎪
⎪ H (1) (k|x|)
⎪
⎪
⎪ c k (n−2)/2 (n−2)/2
if x ∈ Rn \ {0}, n ≥ 2,
⎪
⎨ n |x|(n−2)/2
Φk (x) := ⎪
⎪ (7.6.11)
⎪
⎪
⎪
⎪
⎪
⎩− e
i ik|x|
if x ∈ R, n = 1,
2k
where
1
cn := . (7.6.12)
4i(2π)(n−2)/2
Corresponding' to the case n = 3, the Hankel function takes a simple form, namely
(1)
H1/2 (r) = −i πr2 eir for r > 0 (see, e.g., [74, (6.37), p. 231]), hence formula (7.6.11)
becomes
eik|x|
Φk (x) = − , ∀ x ∈ R3 \{0}. (7.6.13)
4π|x|
To better understand the behavior of Φk near the origin let us introduce, for each
λ ∈ R, the function Ψλ defined at each r ∈ (0, ∞) by
⎧ π (1)
⎪
⎪
⎪ H0 (r)(ln(r))−1 if λ = 0,
⎪
⎪
⎪ 2i
⎪
⎪
⎪
⎪
⎪
⎨ iπ H (1) (r) rλ
Ψλ (r) := ⎪ if λ ∈ (0, ∞), (7.6.14)
⎪
⎪
⎪ 2λ Γ(λ) λ
⎪
⎪
⎪
⎪
⎪
⎪ iπ
⎪
⎩ −iπλ −λ H (1) (r) r−λ if λ ∈ (−∞, 0).
e 2 Γ(−λ) λ
By Lemma 14.72 we have
lim Ψλ (r) = 1, ∀ λ ∈ R. (7.6.15)

r→0+
For further reference we also note that item (3) in Lemma 14.71 combined with the
Chain Rule implies
(1) (1) (1)
Hλ (k|x|) λ
Hλ (k|x|) λ
Hλ+1 (k|x|) x
∇ =k ∇ = −k k
|x|λ (k|x|)λ (k|x|)λ |x|
(1)
Hλ+1 (k|x|)
= −k λ+1
x (7.6.16)
|x|
for each λ ∈ R and every x ∈ Rn \ {0}.
Clearly, Ψλ is closely related to the function Φk and to the fundamental solution
for the Laplacian in Rn for n ≥ 2. The latter, which we now denote by EΔ , is given
by (when n ≥ 2)
⎧ −1
⎪
⎪
⎪
1
if x ∈ Rn \ {0}, n ≥ 3,
⎪
⎨ (n − 2)ωn−1 |x|n−2
EΔ (x) = ⎪
⎪ (7.6.17)
⎪
⎪
⎩ 1 ln |x|
2π if x ∈ R2 \ {0}, n = 2,
See (7.1.12). Here is the result linking Φk to EΔ .
Proposition 7.32. Let k ∈ (0, ∞) and let n ∈ N be such that n ≥ 2. Then for each
x ∈ Rn \ {0} one has
⎧
⎪
⎨ Ψ(n−2)/2 (k|x|) EΔ (x)
⎪
⎪ if n ≥ 3,
Φk (x) = ⎪
⎪ (7.6.18)
⎪
⎩ Ψ0 (k|x|) 2π
1
ln k + EΔ (x) if n = 2.
In addition, if n ≥ 2 then for each x ∈ Rn \ {0} one has

1 x
∇Φk (x) = Ψn/2 (k|x|) n = Ψn/2 (k|x|)∇EΔ (x). (7.6.19)
ωn−1 |x|
Finally, if n ≥ 2 and j, ∈ {1, . . . , n}, then
n x j x
(∂ ∂ j Φk )(x) = − Ψ(n+2)/2 (k|x|) n+3
k ωn−1 |x|
1 δ j
+ Ψn/2 (k|x|) , (7.6.20)
ωn−1 |x|n
for each x ∈ Rn \ {0}.
Proof. Assume for now that n ≥ 3. Then, starting with the expression in (7.6.11), at
each point x ∈ Rn \ {0} we may write
n−2 (1)
cn 2(n−2)/2 Γ H(n−2)/2 (k|x|)
Φk (x) = 2
· n−2 (k|x|)(n−2)/2 |x|2−n . (7.6.21)
iπ (n−2)/2
2 Γ 2
iπ
n−2
Upon recalling (7.6.12), the fact that Γ n2 = n−2 2 Γ 2 (cf. (14.5.2)), and the
expression for ωn−1 from (14.5.6), a direct computation gives

cn 2(n−2)/2 Γ n−2 1
2
=− . (7.6.22)
iπ (n − 2)ωn−1
Formula (7.6.18) for n ≥ 3 now follows by combining (7.6.21), (7.6.22), the defini-
tion in (7.6.14) with λ = (n − 2)/2 > 0 and r = k|x|, and (7.6.17).
The same circle of ideas gives (7.6.18) in the two-dimensional case. Specifically,
if we assume n = 2, then starting with (7.6.11) and (7.6.12), then using (7.6.14) with
λ = 0 and r = k|x|, and then invoking (7.6.17) (for n = 2), we may write
1 (1) 1
Φk (x) = H (k|x|) = Ψ0 (k|x|) ln(k|x|)
4i 0 2π
1
= Ψ0 (k|x|) ln k + EΔ (x) , ∀ x ∈ R2 \ {0}, (7.6.23)
2π
as wanted.
Next, we prove the statement regarding ∇Φk . Suppose n ≥ 2. Using formula
(7.6.16) with λ = (n − 2)/2 we obtain
(1)
H(n−2)/2 (k|x|)
∇Φk (x) = cn k(n−2)/2 ∇
|x|(n−2)/2
(1)
Hn/2 (k|x|)
= −cn kn/2 x, ∀ x ∈ Rn \ {0}. (7.6.24)
|x|n/2
Furthermore, for each x ∈ Rn \ {0} we have
(1) n (1)
Hn/2 (k|x|) cn 2n/2 Γ Hn/2 (k|x|)
−cn kn/2 =− 2
· n (k|x|)n/2 |x|−n
|x|n/2 iπ n/2
2 Γ 2
iπ
1
= Ψn/2 (k|x|) |x|−n (7.6.25)
ωn−1
where in the last equality we have used (7.6.12), (7.6.14) with λ = n/2, and the
expression for ωn−1 from (14.5.6). In concert, (7.6.24) and (7.6.25) give the first
equality in (7.6.19). The second equality in (7.6.19) is obtained by a direct compu-
tation which makes use of (7.6.17).
Moving on to the proof of (7.6.20), fix j, ∈ {1, . . . , n} and x ∈ Rn \ {0}. Then
from (7.6.24) we know that
(1)
Hn/2 (k|x|)
(∂ j Φk )(x) = −cn kn x j. (7.6.26)
(k|x|)n/2
Apply ∂ . Using the product rule, the Chain Rule, and item (3) in Lemma 14.71
(with λ = n/2) we further obtain
(1)
H(n+2)/2 (k|x|) x H (1) (k|x|)
n n/2
(∂ ∂ j Φk )(x) = cn k n
k x j − cn k δ j
(k|x|)(n+2)/2 |x| (k|x|)n/2
cn (1) x j x
= H(n+2)/2 (k|x|)(k|x|)(n+2)/2 n+3
k |x|
1 δ j
+ Ψn/2 (k|x|) (7.6.27)
ωn−1 |x|−n
where the second equality in (7.6.27) uses (7.6.25). Moreover, since
n+2
cn 2(n+2)/2 Γ 2 2(n+2)/2 Γ n+2
· = 2
k iπ k 4i 2(n−2)/2 π(n−2)/2 iπ
n n
Γ 2 n
= − 2 n/2 =− , (7.6.28)
kπ k ω n−1
by recalling (7.6.14) (with λ = (n + 2)/2) we may write

cn (1) n
H (k|x|)(k|x|)(n+2)/2 = − Ψ(n+2)/2 (k|x|). (7.6.29)
k (n+2)/2 k ωn−1
At this stage, (7.6.20) follows by combining (7.6.27) and (7.6.29). This finishes the
proof of Proposition 7.32.
In the next theorem we show that the function Φk is a fundamental solution for
the Helmholtz operator Δ + k2 .
Theorem 7.33. Suppose n ∈ N and fix k ∈ (0, ∞). Then the function Φk defined in
(7.6.11)–(7.6.12) satisfies Φk ∈ Lloc
1
(Rn ) ∩ S (Rn ) and is a fundamental solution for
the Helmholtz operator Δ + k2 in Rn . Moreover,

u ∈ S (Rn ) : (Δ + k2 )u = δ in S (Rn (7.6.30)

= Φk + f : f ∈ C ∞ (Rn ), (Δ + k2 ) f = 0 in Rn .
Corresponding to the case n = 1,

u ∈ S (R) : u + ku = δ in S (R) (7.6.31)
( i )
= − eik|x| + c1 eikx + c2 e−ikx : c1 , c2 ∈ C .
2k
Proof. Consider first the case n ≥ 2. The fact that Φk ∈ Lloc

1
(Rn ) for n ≥ 2 is a
consequence of (7.6.18) in Proposition 7.32, formula (7.6.15), and the membership
EΔ ∈ Lloc
1
(Rn ). Also, item (9) in Lemma 14.71 implies
*
2 i(k|x|−(n−1)π/4) −(n−1)/2
Φk (x) = cn k (n−3)/2
e |x| + O(|x|−(n+1)/2 )
π (7.6.32)
as |x| → ∞,
which further guarantees the existence of some constant C ∈ (0, ∞) such that
|Φk (x)| ≤ C|x|−(n−1)/2 for every x ∈ Rn \ B(0, 1). Hence, Example 4.4 applies (condi-
tion (4.1.4) is satisfied for any m > (n + 1)/2) and gives that Φk ∈ S (Rn ).
Next we take up the task of proving that Φk in (7.6.11) is a fundamental solution
for Δ + k2 in Rn , n ≥ 2. By (7.6.10) and (7.6.11) we have
Φk ∈ C ∞ (Rn \ {0}) and (Δ + k2 )Φk = 0 in Rn \ {0}. (7.6.33)
Now let ϕ ∈ C0∞ (Rn ) be arbitrary. To conclude that Φk is a fundamental solution for
Δ + k2 , since Φk is locally integrable, it suffices to establish that

Φk (x) (Δ + k2 )ϕ (x) dx = ϕ(0), ∀ ϕ ∈ C0∞ (Rn ). (7.6.34)
Rn
Applying Lebesgue’s Dominated Convergence Theorem, then using integrations by

parts twice (cf. Theorem 14.60), and then (7.6.33), we may write

Φk (x) (Δ + k2 )ϕ (x) dx = lim+ Φk (x) (Δ + k2 )ϕ (x) dx
Rn ε→0 Rn \B(0,ε)
+
,

= lim+ (Δ + k )Φk (x)ϕ(x) dx + Iε + IIε
2
ε→0 Rn \B(0,ε)
= lim+ Iε + lim+ IIε , (7.6.35)

ε→0 ε→0
where we have set (recall the definition of the normal derivative from (7.1.14))

∂ϕ
Iε := − Φk (x) (x) dσ(x), (7.6.36)
∂B(0,ε) ∂ν

∂Φk
IIε := (x)ϕ(x) dσ(x). (7.6.37)
∂B(0,ε) ∂ν
Above, ν denotes the outward unit normal to B(0, ε), that is, ν(x) = εx for each
x ∈ ∂B(0, ε), and σ denotes the surface measure on ∂B(0, ε).
We impose for now the restriction that n ≥ 3. Note that for each ε ∈ (0, 1),
formula (7.6.18) permits us to estimate

∂ϕ
|Iε | ≤ |Φk (x)| (x) dσ(x)
∂B(0,ε) ∂ν

≤ ∇ϕL∞ (B(0,1)) Ψ(n−2)/2 (kε) EΔ (x) dσ(x)
∂B(0,ε)
ε
= ∇ϕL∞ (B(0,1)) Ψ(n−2)/2 (kε) . (7.6.38)
n−2
Taking the limit as ε → 0+ and invoking (7.6.15), we obtain
lim Iε = 0. (7.6.39)
ε→0+
We claim that (7.6.39) is also valid if n = 2. Indeed, the estimate for Iε from (7.6.38)
now becomes

1
|Iε | ≤ ∇ϕL∞ (B(0,1)) |Ψ0 (kε)| ln k + EΔ (x) dσ(x)
∂B(0,ε) 2π
= ∇ϕL∞ (B(0,1)) |Ψ0 (kε)|| ln(kε)|ε. (7.6.40)
Taking the limit as ε → 0+ and invoking (7.6.15) yields (7.6.39). To summarize, we

have proved (7.6.39) for n ≥ 2.
To handle IIε use (7.6.19) to write, for each x ∈ ∂B(0, ε),
∂Φk x 1
(x) = ∇Φk (x) · = Ψn/2 (kε)ε−(n−1) . (7.6.41)
∂ν |x| ωn−1
When used back in (7.6.37), this allows us to recast IIε as

1
IIε = Ψn/2 (kε) · ϕ(x) dσ(x). (7.6.42)
ωn−1 εn−1 ∂B(0,ε)

Since ∂B(0,ε) 1 dσ = ωn−1 εn−1 , we may write

1
ϕ(x) dσ(x)
ωn−1 εn−1 ∂B(0,ε)

1
= (ϕ(x) − ϕ(0)) dσ(x) + ϕ(0). (7.6.43)
ωn−1 εn−1 ∂B(0,ε)
By the Mean Value Theorem we have |ϕ(x) − ϕ(0)| ≤ ε∇ϕL∞ (Rn ) for each x in
∂B(0, ε), thus the integral in the right-hand side of (7.6.43) converges to zero as
ε → 0+ . This proves that

1
lim+ ϕ(x) dσ(x) = ϕ(0). (7.6.44)
ε→0 ωn−1 εn−1 ∂B(0,ε)
In concert, (7.6.42), (7.6.44), and (7.6.15) (applied with λ = n/2) imply

lim IIε = ϕ(0). (7.6.45)

ε→0+
At this point formula (7.6.34) follows from (7.6.35), (7.6.39), and (7.6.45). This
completes the proof of the fact that Φk is a fundamental solution for Δ + k2 in Rn
when n ≥ 2.
To treat the case when n = 1, observe first that
i ik|x| i i
Φk (x) = − e = − eikx H(x) − e−ikx H ∨ (x), (7.6.46)
2k 2k 2k
for every x ∈ R \ {0}. Clearly Φk ∈ Lloc
1
(R) and e±ikx ∈ L(R). Also, Exercise 4.119

guarantees that H ∈ S (R). Granted these properties, item (b) in Theorem 4.14
applies and yields Φk ∈ S (R). In addition,
i −ikx i i
Φk (x) + e = − eikx H(x) + e−ikx H(x), (7.6.47)
2k 2k 2k
for every x ∈ R \ {0}. With (7.6.47) in hand, (7.6.31) becomes a consequence of
Exercise 6.27.
There remains to check (7.6.30). The right-to-left inclusion is clear from what
we have proved already. Also, if u ∈ S (Rn ) is a fundamental solution for Δ + k2
in Rn , then (Δ + k2 )(u − Φk ) = 0 in D (Rn ). Observe that the operator Δ + k2 is
elliptic (cf. Definition 6.13), hence hypoelliptic by Theorem 6.15. Consequently, by
Remark 6.7, we have f := u − Φk ∈ C ∞ (Rn ), which in turn implies (Δ + k2 ) f = 0
pointwise in Rn . With this the left-to-right inclusion in (7.6.30) also follows.
Next, we concern ourselves with the behavior of Φk at infinity. In this regard, we
shall make use of the asymptotic results for Hankel functions from §14.10 in order
to establish the following proposition.
Proposition 7.34. Fix k ∈ (0, ∞) along with n ∈ N, n ≥ 2, and set
k(n−3)/2 2 1/2 −iπ(n−1)/4

bn,k := e . (7.6.48)
4i(2π)(n−2)/2 π
Also, throughout abbreviate
x

x := for each x ∈ Rn \ {0}. (7.6.49)
|x|
Then for each multi-index α ∈ Nn0 one has
eik|x| e−iky,x
(∂α Φk )(x − y) = bn,k x )α + O |x|−(n+1)/2
(ik (7.6.50)
|x|(n−1)/2

x )α Φk (x − y) + O |x|−(n+1)/2
= (ik as |x| → ∞,
uniformly for y in compact subsets of Rn , and

eik|y| e−ikx,y
(∂α Φk )(x − y) = bn,k y )α + O |y|−(n+1)/2
(−ik (7.6.51)
|y|(n−1)/2

y )α Φk (x − y) + O |y|−(n+1)/2
= (−ik as |y| → ∞,
uniformly for x in compact subsets of Rn .

In particular, for each multi-index α ∈ Nn0 one has

(∂α Φk )(x − y) = O |x|−(n−1)/2 as |x| → ∞,
(7.6.52)
uniformly for y in compact subsets of Rn ,
and

(∂α Φk )(x − y) = O |y|−(n−1)/2 as |y| → ∞,
(7.6.53)
Proof. Fix α ∈ Nn0 arbitrary. As a first step we shall prove that
eik|x|
(∂α Φk )(x) = bn,k x )α + O |x|−(n+1)/2
(ik as |x| → ∞. (7.6.54)
|x|(n−1)/2
To get started, combine (7.6.12) and item (9) in Lemma 14.71 to write
-
k(n−2)/2 −(n−2)/2 2 .1/2 i(|x|−(n−2)π/4−π/4) −3/2
Φk (x) = |x| e + O |x|
4i(2π)(n−2)/2 πk|x|
k(n−3)/2 2 1/2 −iπ(n−1)/4 eik|x|

= e + O |x|−(n+1)/2
4i(2π)(n−2)/2 π |x|(n−1)/2
eik|x|
= bn,k + O |x|−(n+1)/2 as |x| → ∞, (7.6.55)
|x|(n−1)/2
which proves (7.6.54) for α = (0, . . . , 0).
Next, from (7.6.12) and Leibniz’ formula we have
(1)
(∂α Φk )(x) = cn k(n−2)/2 ∂α H(n−2)/2 (k|x|) |x|−(n−2)/2 (7.6.56)
α! β (1)
+ cn k(n−2)/2 ∂ H(n−2)/2 (k|x|) ∂γ |x|−(n−2)/2 .
β+γ=α, |γ|>0
β!γ!
By invoking (14.10.17), the first term in the right-hand side of (7.6.56) becomes
α (1)
cn k(n−2)/2 ∂ H(n−2)/2 (k|x|) |x|−(n−2)/2
(1)
= cn k(n−2)/2 H(n−2)/2−|α| x )α + O |x|−3/2 |x|−(n−2)/2
(k|x|)(k
(1)
H(n−2)/2−|α| x )α
(k|x|)(k
= cn k(n−2)/2 + O |x|−(n+1)/2
|x|(n−2)/2

2 1/2 i(k|x|−(n−2)π/4+|α|π/2−π/4)
πk|x| e + O |x|−3/2 (k
x )α
= cn k (n−2)/2
|x|(n−2)/2

+ O |x|−(n+1)/2
eik|x|
= bn,k x )α + O |x|−(n+1)/2
(ik as |x| → ∞, (7.6.57)
|x|(n−1)/2
where for the third equality in (7.6.57) we used item (9) in Lemma 14.71. Re-
garding the second term in the right-hand side of (7.6.56), it is immediate that

∂γ |x|−(n−2)/2 = O |x|−(n−2)/2−|γ| as |x| → ∞ for γ ∈ Nn0 , while from (14.10.10) it

follows that ∂β H(n−2)/2
(1)
(k|x|) = O(|x|−1/2 ) as |x| → ∞. As such, we may conclude

that the sum in (7.6.56) is O |x|−(n+1)/2 as |x| → ∞. This, together with (7.6.57),
finishes the proof of (7.6.54).
With an eye on (7.6.50), let K be a compact subset of Rn and assume for now
that y ∈ K and that |x| is sufficiently large. In this case |x − y| is proportional to |x|,

hence O |x − y|−(n+1)/2 = O |x|−(n+1)/2 as |x| → ∞, uniformly for y ∈ K. Moreover,
the Mean Value Theorem gives

x−y x 1
= +O , as |x| → ∞, (7.6.58)
|x − y| |x| |x|
uniformly for y ∈ K. Thus, on the one hand,

α α
x−y x 1
= +O as |x| → ∞, (7.6.59)
|x − y| |x| |x|
uniformly for y ∈ K. On the other hand, as a consequence of (7.6.54) we obtain
eik|x−y| α
(∂α Φk )(x − y) = bn,k ik(x − y)
|x − y|(n−1)/2

+ O |x − y|−(n+1)/2 as |x − y| → ∞. (7.6.60)
Next we claim that

eik|x−y| eik|x| e−iky,x
− = O |x|−(n+1)/2
|x − y|(n−1)/2 |x| (n−1)/2
(7.6.61)
uniformly for y ∈ K, as |x| → ∞.
Indeed
eik|x−y| eik|x| e−iky,x
− (7.6.62)
|x − y|(n−1)/2 |x|(n−1)/2
+ ,
1 1
=e ik|x−y|
−
|x − y|(n−1)/2 |x|(n−1)/2
eik|x−y|
+ 1 − e−ik|x−y|+ik|x|−iky,x := I + II.
|x|(n−1)/2
Based on the Mean Value Theorem we see that

I = O |x|−(n+1)/2 as |x| → ∞, uniformly for y ∈ K. (7.6.63)
Consequently, (7.6.61) will follow once we show that

1 − e−ik|x−y|+ik|x|−iky, x = O |x|−1
(7.6.64)
as |x| → ∞, uniformly for y ∈ K.
With this goal in mind, write (for x large)

|x − y| = x − y, x − y 1/2 = |x|2 − 2x, y + |y|2 1/2
1/2

x, y |y|2
= |x| 1 − 2 + 2 . (7.6.65)
|x| |x|
Recalling that the Taylor series expansion of the function t → (1 + t)1/2 around zero
is (1 + t)1/2 = 1 + 2t + O(t2 ), from (7.6.65) we further deduce that
⎛ 2 ⎞
1 |y|2
x, y ⎜⎜⎜ |y|2
x , y ⎟⎟⎟⎟
|x − y| = |x| 1 + −2
+ O ⎜⎜⎝ 2 − 2 ⎟
2 |x| 2 |x| |x| |x| ⎠

x, y −2
= |x| 1 − + O |x| (7.6.66)
|x|

x, y + O |x|−1 as |x| → ∞, uniformly for y ∈ K.
= |x| −

Now (7.6.64) follows from (7.6.66) and the fact that 1 − eia ≤ 2|a| for every a ∈ R.
With this, the proof of (7.6.61) is finished.
A combination of (7.6.60), (7.6.59), and (7.6.61) yields
eik|x−y| α
(∂α Φk )(x − y) = bn,k ik(x − y) + O |x − y|−(n+1)/2
|x − y|(n−1)/2
ik|x| −iky,x
e e −(n+1)/2 α 1
= bn,k + O |x| x +O
ik
|x|(n−1)/2 |x|

+ O |x|−(n+1)/2
eik|x| e−iky,x α
= bn,k x + O |x|−(n+1)/2
ik
|x|(n−1)/2
as |x| → ∞, uniformly for y ∈ K, (7.6.67)
and proves the first equality in (7.6.50). In particular, corresponding to |α| = 0, we

have
eik|x| e−iky,x
Φk (x − y) = bn,k + O |x|−(n+1)/2
|x| (n−1)/2
as |x| → ∞, uniformly for y ∈ K. (7.6.68)
Making use of (7.6.68) back in (7.6.67) implies

(∂α Φk )(x − y) = Φk (x − y) + O |x|−(n+1)/2 (ik x )α + O |x|−(n+1)/2

x )α Φk (x − y) + O |x|−(n+1)/2
= (ik
as |x| → ∞, uniformly for y ∈ K. (7.6.69)
This finishes the proof of (7.6.50).

Finally, upon noting that Φk is even, it follows that
(∂α Φk )(x − y) = (−1)|α| (∂α Φk )(y − x), x y, (7.6.70)
so (7.6.51) becomes a consequence of (7.6.70) and (7.6.50).

Exercise 7.35. Prove that Φk satisfies Sommerfeld’s radiation condition
n
xj
iku(x) − (∂ j u)(x) = o |x|−(n−1)/2 as |x| → ∞. (7.6.71)
j=1
|x|
Exercise 7.36. Consider k ∈ (0, ∞) along with some n ∈ N, n ≥ 2. Show that for
every multi-index α ∈ Nn0 one has

x , ∇(∂α Φk ) (x − y) − ik(∂α Φk )(x − y)

= O |x|−(n+1)/2 as |x| → ∞, (7.6.72)
uniformly for y in compact subsets of Rn . As a consequence, for each y ∈ Rn fixed,

conclude that
(∂α Φk )(· − y) satisfies Sommerfeld’s radiation condition (7.6.71). (7.6.73)
Exercise 7.37. Fix k ∈ (0, ∞) together with n ∈ N, n ≥ 2.

(1) Show that for any two multi-indices α, β ∈ Nn0 one has

(∂α+β Φk )(x − y) = (ik
x )α (∂β Φk )(x − y) + O |x|−(n+1)/2
(7.6.74)
as |x| → ∞, uniformly for y in compact subsets of Rn ,
and

(∂α+β Φk )(x − y) = (−ik
y )α (∂β Φk )(x − y) + O |y|−(n+1)/2
(7.6.75)
as |y| → ∞, uniformly for x in compact subsets of Rn .
(2) Use the results in part (1) to show that for any multi-index α ∈ Nn0 and any
indexes j, ∈ {1, . . . , n},

x j (∂ ∂α Φk )(x − y) −
x (∂ j ∂α Φk )(x − y) = O |x|−(n+1)/2
(7.6.76)
as |x| → ∞, uniformly for y in compact subsets of Rn ,
and

y j (∂ ∂α Φk )(x − y) −
y (∂ j ∂α Φk )(x − y) = O |y|−(n+1)/2
(7.6.77)
as |y| → ∞, uniformly for x in compact subsets of Rn .
We close this section by isolating a result which is useful in the computation of

fundamental solutions for the perturbed Dirac operator and its iterations (later on,
in §7.10–§7.11).
Proposition 7.38. Let λ ∈ (−∞, n/2) and consider the function
Hλ(1) (k|x|)
Fλ (x) := , ∀ x ∈ Rn \ {0}. (7.6.78)
|x|λ
1
Then Fλ belongs to Lloc (Rn ), hence it defines a distribution of function type. Also,
if λ < (n − 1)/2 then for each j ∈ {1, . . . , n} it follows that Fλ+1 (x) x j belongs to
1
Lloc (Rn ) and
∂ j Fλ = −kFλ+1 (x) x j in D (Rn ). (7.6.79)
Proof. First work under the assumption that λ ∈ (−∞, n/2). Recall the function Ψλ
from (7.6.14). Then for each x ∈ Rn \ {0} we have
⎧ 2i
⎪
⎪
⎪
⎪
⎪ Ψ0 (k|x|) ln(k|x|) if λ = 0,
⎪
⎪
⎪ π
⎪
⎪
⎪
⎪
⎪
⎨ 2λ Γ(λ) 1
Fλ (x) = ⎪
⎪ Ψλ (k|x|) 2λ if λ ∈ (0, n/2), (7.6.80)
⎪
⎪
⎪ iπ k λ |x|
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪ e−iπλ 2−λ Γ(−λ) kλ
⎩ Ψλ (k|x|) if λ ∈ (−∞, 0),
iπ
In combination with (7.6.15) this implies that Fλ ∈ Lloc
1
(Rn ).
Next, suppose λ ∈ (−∞, (n − 1)/2) and fix j ∈ {1, . . . , n}. From (7.6.16) we know
that when differentiating pointwise we have
(∂ j Fλ )(x) = −kFλ+1 (x) x j ∀ x ∈ Rn \ {0}. (7.6.81)
Pick an arbitrary ϕ ∈ C0∞ (Rn ). Since Fλ ∈ Lloc 1

(Rn ) and ϕ has compact support we
may use Lebesgue’s Dominated Convergence Theorem, integration by parts, and
(7.6.81), to write

∂ j Fλ , ϕ = −Fλ , ∂ j ϕ = − Fλ (x)∂ j ϕ(x) dx

Rn

= − lim+ Fλ (x)∂ j ϕ(x) dx (7.6.82)

ε→0 |x|≥ε

xj
= lim+ − k Fλ+1 (x) x j ϕ(x) dx − Fλ (x) ϕ(x) dσ(x) .
ε→0 |x|≥ε |x|=ε ε
The reasoning that has produced (7.6.80) also gives that for each x ∈ Rn \ {0} we
have
⎧
⎪
⎪
⎪ C Ψ0 (k|x|) ln(k|x|) x j if λ = −1,
⎪
⎪
⎪
⎪
⎪
⎨ xj
Fλ+1 (x)x j = ⎪
⎪ C Ψλ+1 (k|x|) |x|2(λ+1) if λ ∈ (−1, (n − 1)/2), (7.6.83)
⎪
⎪
⎪
⎪
⎪
⎪
⎩ C Ψλ+1 (k|x|) x j if λ ∈ (−∞, −1),
where C denotes in each case some constant which may depend on n, λ, and k,
but not on x. Bearing in mind that we are currently assuming λ < (n − 1)/2, by
1
combining (7.6.83) and (7.6.15) we see that Fλ+1 (x)x j belongs to Lloc (Rn ). Also,
Lebesgue’s Dominated Convergence Theorem applies and gives

−k lim+ Fλ+1 (x) x j ϕ(x) dx = −k Fλ+1 (x) x j ϕ(x) dx

ε→0 |x|≥ε Rn
= −kFλ+1 (x)x j , ϕ . (7.6.84)

7.7 Fundamental Solutions for the Iterated Helmholtz Operator 279
Moreover, (7.6.80) used in the range λ < (n − 1)/2 gives

⎧
⎪
⎪ O(ln(kε)) if λ = 0,
⎪
⎪⎪
⎪
⎨
sup Fλ (x) = ⎪
⎪ O(ε−2λ ) if λ ∈ (0, (n − 1)/2), as ε → 0+ . (7.6.85)
|x|=ε ⎪
⎪
⎪
⎪
⎩ O(1) if λ ∈ (−∞, 0),
This, together with the fact that ϕ is bounded, implies

xj
lim+ Fλ (x) ϕ(x) dσ(x) = 0. (7.6.86)
ε→0 |x|=ε ε
From (7.6.82), (7.6.84), and (7.6.86), we conclude
∂ j Fλ , ϕ = −kFλ+1 (x)x j , ϕ (7.6.87)
which, in view of the arbitrariness of ϕ ∈ C0∞ (Rn ), proves (7.6.79).
7.7 Fundamental Solutions for the Iterated Helmholtz Operator
Suppose n ∈ N is such that n ≥ 2. The goal in this section is to determine a funda-

mental solution for the iterated Helmholtz operator in Rn , that is, for the operator
(Δ + k2 )N in Rn , where k ∈ (0, ∞) and N ∈ N. The case N = 1 has been treated in the
previous section; cf. Theorem 7.33. To get started, recall the function Φk defined in
(7.6.11). Bearing in mind that the Hankel function of the first kind is of class C ∞ in
(0, ∞) (cf. Lemma 14.71), this implies that
(0, ∞) × (Rn \ {0}) (k, x) → Φk (x) ∈ C

(7.7.1)
is a mapping of class C ∞ .
d
This allows us to apply dk to both sides of the identity in (7.6.33) and compute
(recall that the Laplacian Δ is taken in the variable x)
d d d
0= (Δ + k2 )Φk = ΔΦk + k2 Φk
dk dk dk
d d
=Δ Φk + 2kΦk + k2 Φk
dk dk
d
= (Δ + k2 ) Φk + 2kΦk pointwise in Rn \ {0}. (7.7.2)
dk
Hence,
1 d
(Δ + k2 ) − · Φk = Φk pointwise in Rn \ {0}. (7.7.3)
2k dk
In particular, if we set
1 d
Φ(2)
k := − · Φk for each (k, x) ∈ (0, ∞) × (Rn \ {0}), (7.7.4)
2k dk
then (7.7.3) becomes
(Δ + k2 )Φ(2)
k = Φk pointwise in Rn \ {0}. (7.7.5)
In light of (7.6.33), formula (7.7.5) also implies that for each k ∈ (0, ∞) there
holds (Δ + k2 )2 Φ(2)
k = (Δ + k )Φk = 0 pointwise in R \ {0}. Given that from The-
2 n

orem 7.33 we know that (Δ + k )Φk = δ in D (R ) it is natural to ask whether the
2 n

identity (Δ + k2 )2 Φ(2)
k = δ holds in D (R ). The good news is that the answer to
n
this question is yes, and the approach employed to arrive at a formula for Φ(2) k is
(N)
resourceful enough to provide a good candidate Φk for a fundamental solution for
the iterated Helmholtz operator (Δ + k2 )N with N ∈ N arbitrary. This is done in
the proposition below via a recurrence relation. Before stating the aforementioned

proposition, we wish to note that if u ∈ C ∞ (0, ∞) × (Rn \ {0}) is arbitrary, then for
k ∈ (0, ∞) and N ∈ N the following differentiation formula holds
d d
(Δ + k2 )N u = (Δ + k2 )N u + 2Nk(Δ + k2 )N−1 u (7.7.6)
dk dk
pointwise in Rn \ {0}. This may be easily justified via an induction over N.
Proposition 7.39. Let n ∈ N, n ≥ 2, and take k ∈ (0, ∞). Recall the function Φk from
(7.6.11)–(7.6.12). Consider the sequence of functions {Φ(N)
k }N∈N recurrently defined
by setting
Φ(1)
k (x) := Φk (x), ∀ x ∈ Rn \ {0}, (7.7.7)
and, for each N ∈ N,
−1 d (N)
Φ(N+1)
k (x) := · Φ (x), ∀ x ∈ Rn \ {0}. (7.7.8)
2Nk dk k
Then these functions satisfy the following properties:
(1) The assignment (0, ∞) × (Rn \ {0}) (k, x) → Φ(N)
k (x) ∈ C is a function of class
C ∞ , for each N ∈ N.
(2) For each N ∈ N and k ∈ (0, ∞) one has (Δ+k2 )N Φ(N)
k (x) = 0 for every x ∈ R \{0}.
n
(N+1) (N)
(3) For each N ∈ N and k ∈ (0, ∞) one has (Δ + k )Φk2
(x) = Φk (x) for every
x ∈ Rn \ {0}.
(4) Having fixed k ∈ (0, ∞) and N ∈ N, the following formula holds for each point
x ∈ Rn \ {0}:
(1)
(−1)N−1 cn k(n−2N)/2 H(n−2N)/2 (k|x|)
Φ(N)
k (x) = , (7.7.9)
2N−1 (N − 1)! |x|(n−2N)/2
where cn is as in (7.6.12).
(5) Recall the function Ψλ from (7.6.14). Then, for each x ∈ Rn \ {0},
⎧
⎪
⎪ C1 (n, N) Ψ(n−2N)/2 (k|x|) |x|−(n−2N) , if n > 2N,
⎪
⎪
⎪
⎪
⎨
Φ(N)
k (x) = ⎪
⎪ C2 (n, N) Ψ0 (k|x|) ln(k|x|), if n = 2N, (7.7.10)
⎪
⎪
⎪
⎪
⎩ C (n, N, k) Ψ
3 (n−2N)/2 (k|x|), if n < 2N,
where

(−1)N 2−2N Γ n−2N
C1 (n, N) := 2
, (7.7.11)
(N − 1)! πn/2
(−1)N−1
C2 (n, N) := , (7.7.12)
2(n+2N−2)/2 (N − 1)! πn/2
2N−n
(−1)N kn−2N e−iπ(n−2N)/2 Γ
C3 (n, N, k) := 2
. (7.7.13)
(N − 1)! 2n πn/2
Consequently,
⎧ −(n−2N)
⎪
⎪ O |x| , if n > 2N,
⎪
⎪
⎪
(N) ⎪
⎨
Φk (x) = ⎪
⎪ O ln(k|x|) , if n = 2N, as x → 0. (7.7.14)
⎪
⎪
⎪
⎪
⎩ O(1), if n < 2N,
(6) For each k ∈ (0, ∞) and each N ∈ N one has Φ(N)k ∈ Lloc (R ).
1 n
(7) For each k ∈ (0, ∞) and each N ∈ N one has

√
(N) (−1)N−1 cn k(n−2N−1)/2 2 ei(k|x|−(2n−4N+1)π/4)
Φk (x) = √ ·
2N−1 (N − 1)! π |x|(n−2N+1)/2
+ O(|x|−(n−2N+3)/2 ) as |x| → ∞. (7.7.15)

(8) For each k ∈ (0, ∞) and each N ∈ N one has Φ(N) k ∈ S (R ).
n
(9) For each k ∈ (0, ∞) and each N ∈ N one has

⎧ x
⎪
⎪
⎪ C4 (n, N) Ψ(n−2N+2)/2 (k|x|) · n−2N+2 , if n > 2N − 2,
⎪
⎪
⎪ |x|
⎪
⎪
⎨
∇Φ(N) (x) = ⎪
⎪ C5 (n, N) Ψ0 (k|x|) ln(k|x|) x, if n = 2N − 2, (7.7.16)
k ⎪
⎪
⎪
⎪
⎪
⎪
⎩ C (n, N, k) Ψ (k|x|) x, if n < 2N − 2,
6 (n−2N+2)/2
where
n−2N+2
(−1)N+1 Γ
C4 (n, N) := 2
, (7.7.17)
(N − 1)! 22N−1 πn/2
(−1)N
C5 (n, N) := = −C2 (n, N), (7.7.18)
2(n+2N−2)/2 (N − 1)! πn/2

(−1)N kn−2N+2 e−iπ(n−2N)/2 Γ 2N−2−n
C6 (n, N, k) := 2
. (7.7.19)
(N − 1)! 2n+1 πn/2
In particular,
⎧ −(n−2N+1)
⎪
⎪
⎪ O |x| , if n > 2N − 2,
⎪
⎪
⎪
⎨
∇Φ(N)
k (x) = ⎪
⎪ O ln(k|x|)|x| , if n = 2N − 2, as x → 0. (7.7.20)
⎪
⎪
⎪
⎪
⎩ O(|x|), if n < 2N − 2,
Proof. We shall first prove by induction on N ∈ N the properties claimed in items

(1)–(3) hold. That this is the case when N = 1 is seen from (7.7.1), (7.6.33), and
(7.7.3). Assume next that the properties claimed in items (1)–(3) hold for some
N ∈ N. Then the fact that
(0, ∞) × (Rn \ {0}) (k, x) → Φ(N+1)

k (x) ∈ C (7.7.21)
is a function of class C ∞ is a consequence of (7.7.8) and the induction hypothesis.

Also, applying (Δ + k2 )N to the identity in item (3) yields
(Δ + k2 )N+1 Φ(N+1)
k = (Δ + k2 )N Φ(N)
k = 0 in R \ {0},
n
(7.7.22)
where the last equality uses item (2) in the current proposition. Finally, to prove the
version of item (3) written for N + 1 in place of N, start by applying dk d
to both
(N) (N+1)
sides of the formula Φk = (Δ + k )Φk
2
and, reasoning much as in (7.7.2) (while
bearing (7.7.8) in mind), obtain
d (N) d d (N+1) d 2 (N+1)

Φk = (Δ + k2 )Φ(N+1)
k = ΔΦk + k Φk
dk dk dk dk
d (N+1) d
=Δ Φk + 2kΦ(N+1)
k + k2 Φ(N+1)
dk dk k
d (N+1)
= (Δ + k2 ) Φ + 2kΦ(N+1)
dk k k
= −2(N + 1)k(Δ + k2 )Φ(N+2)

k + 2kΦ(N+1)
k (7.7.23)
pointwise in Rn \ {0}. Together, (7.7.23) and (7.7.8) give
−2NkΦ(N+1)
k = −2(N + 1)k(Δ + k2 )Φ(N+2)
k + 2kΦ(N+1)
k (7.7.24)
pointwise in Rn \ {0}, hence (Δ + k2 )Φ(N+2)

k = Φ(N+1)
k pointwise in Rn \ {0}. This
concludes the proof (by induction on N) of the statements in items (1)–(3) of the
proposition.
To prove the statement in item (4) fix k ∈ (0, ∞). We will show by induction
on N ∈ N that formula (7.7.9) holds. If N = 1, formula (7.7.9) is an immediate
consequence of (7.6.11) since Φ(1)
k = Φk and we are assuming n ≥ 2. Suppose
(7.7.9) holds for some N ∈ N. Then, (7.7.8), the induction hypothesis, item (2)
in Lemma 14.71 (applied with λ = (n − 2N)/2), and the Chain Rule permit us to
compute
−1 d (N)
Φ(N+1)
k (x) = · Φ (x)
2Nk dk k
(−1)N cn d (1)

= · (k|x|)(n−2N)/2 H(n−2N)/2 (k|x|)
2N N! k |x|n−2N dk
(−1)N cn (1)
= · (k|x|)(n−2N)/2 H(n−2N−2)/2 (k|x|) |x|
2N N! k |x|n−2N
(1)
(−1)N cn k(n−2N−2)/2 H(n−2N−2)/2 (k|x|)
= (7.7.25)
2N N! |x|(n−2N−2)/2
for all x ∈ Rn \ {0}. Hence, (7.7.9) holds for N + 1 in place of N. This finishes the
proof of the statement in item (4).
Moving on to item (5), let k ∈ (0, ∞), N ∈ N be arbitrary and recall the function
Ψλ from (7.6.14). We separate our discussion in three cases: n > 2N, n = 2N, and
n < 2N.
Assume first that n > 2N. Then based on (7.7.9) and (7.6.14) corresponding to
the case λ = (n − 2N)/2 > 0 we may write

(−1)N−1 cn 2(n−2N)/2 Γ n−2N 1
Φ(N) (x) = 2
· Ψ(n−2N)/2 (k|x|) · n−2N (7.7.26)
k
2N−1 (N − 1)! iπ |x|
for every x ∈ Rn \ {0}. Using (7.6.12), the notation from (7.7.11), and some elemen-
tary algebra, this gives the equality in (7.7.10) if n > 2N. Moreover, (7.7.26) and
−(n−2N)
(7.6.15) imply that Φ(N)
k (x) = O |x| as x → 0 whenever n > 2N.
Next, suppose n = 2N. Then formulas (7.7.9) and (7.6.14) applied with λ = 0
give
(−1)N−1 cn 2i
Φ(N)
k (x) = · Ψ0 (k|x|) ln(k|x|) (7.7.27)
2N−1 (N − 1)! π
for every x ∈ Rn \ {0}. Making use of (7.6.12) and some algebra yields the equality
in (7.7.10) corresponding to n = 2N. In addition, (7.7.27) and (7.6.15) ensure that

we have Φ(N)
k (x) = O ln(k|x|) as x → 0 in the current case.
Finally, if n < 2N, from (7.7.9) and (7.6.14) (used with λ = (n − 2N)/2 < 0) we
obtain
2N−n
(−1)N−1 cn kn−2N e−iπ(n−2N)/2 2(2N−n)/2 Γ
Φ(N)
k (x) = 2
· Ψ(n−2N)/2 (k|x|) (7.7.28)
2N−1 (N − 1)! iπ
for every x ∈ Rn \ {0}. From this and (7.6.12) the equality in (7.7.10) correspond-
ing to n < 2N follows. Invoking (7.6.15) in concert with (7.7.28) we also see that
Φ(N)
k (x) = O(1) as x → 0 if n < 2N. This completes the proof of the statement in
item (5).
The statement in item (6) follows immediately from (7.7.14) since the functions
in the right-hand side of (7.7.14) are locally integrable in Rn . The behavior of Φ(N) k
at infinity claimed in item (7) is a direct consequence of (7.7.9) and item (9) in
Lemma 14.71.
As regards item (8), in order to show that for each k ∈ (0, ∞) and each N ∈ N the
function Φ(N)
k defines a tempered distribution in R , we note that (7.7.15) implies the
n
−(n−2N+1)/2
existence of some C ∈ (0, ∞) with the property that |Φ(N) k (x)| ≤ C|x| for
(N)
every x ∈ R \ B(0, 1). From this, the membership Φk ∈ Lloc (R ), and Example 4.4
n 1 n
(note that condition (4.1.4) holds for any m > (n + 2N − 1)/2) the desired conclusion
follows.
Turning to the proof of the statements in item (9), consider first (7.7.16). As an
application of (7.6.16) (with λ = (n − 2N)/2) we have
⎡ (1) ⎤
⎢⎢⎢ H(n−2N)/2 (k|x|) ⎥⎥⎥ H (1) (k|x|)
∇ ⎢⎢⎣ ⎢ ⎥⎥⎥ = −k (n−2N+2)/2 x (7.7.29)
|x|(n−2N)/2 ⎦ |x|(n−2N+2)/2
for every x ∈ Rn \ {0}. Together, (7.7.29) and (7.7.9) then give

(1)
(−1)N cn k(n−2N+2)/2 H(n−2N+2)/2 (k|x|)
∇Φ(N)
k (x) = x, (7.7.30)
2N−1 (N − 1)! |x|(n−2N+2)/2
for every x ∈ Rn \ {0}. This, the definition of the function Ψλ from (7.6.14), and
with (7.6.12) then yield (7.7.16). Finally, (7.7.20) is an immediate consequence of
(7.7.16). This finishes the proof of Proposition 7.39.
We now take up the task of establishing that Φ(N)
k is a fundamental solution for
the iterated Helmholtz operator (Δ + k2 )N .
Theorem 7.40. Let n ∈ N, n ≥ 2, and k ∈ (0, ∞). For N ∈ N recall the function
Φ(N)
k ∈ Lloc (R ) from (7.7.9), which is explicitly given by
1 n
(1)
(−1)N i k(n−2N)/2 H(n−2N)/2 (k|x|)
Φ(N)
k (x) = (n+2N)/2 · (7.7.31)
2 (N − 1)!π(n−2)/2 |x|(n−2N)/2
for every x ∈ Rn \ {0}. Then, the distribution of function type defined by Φ(N)
k , which
by Proposition 7.39 is known to belong to S (Rn ), is a fundamental solution for the
iterated Helmholtz operator (Δ + k2 )N in Rn .
Proof. We shall reason by induction over N. The case N = 1 has been dealt with
in Theorem 7.33 (recall that Φ(1)
k = Φk ). Suppose next that for some positive inte-
ger N the distribution Φ(N) k is known to be a fundamental solution for the oper-
ator (Δ + k ) in R . In order to show that Φ(N+1)
2 N n
k is a fundamental solution for
the operator (Δ + k2 )N+1 in Rn , take ϕ ∈ C0∞ (Rn ). The idea is to reason as in the
proof of Theorem 7.33 and employ Proposition 7.39. Starting with the fact that the
distribution Φ(N+1)
k is of function type (cf. (6) in Proposition 7.39), then applying
Lebesgue’s Dominated Convergence Theorem, then integrating by parts twice (cf.
Theorem 14.60), then invoking item (3) in Proposition 7.39, then item (6) in Propo-
sition 7.39 and Lebesgue’s Dominated Convergence Theorem, we may write

(Δ + k2 )N+1 Φ(N+1)
k , ϕ = Φ(N+1)
k (x) (Δ + k2 )N+1 ϕ (x) dx
Rn

= lim+ Φ(N+1)
k (x) (Δ + k2 )N+1 ϕ (x) dx
ε→0 Rn \B(0,ε)
+
,

= lim+ (Δ + k2 )Φ(N+1)
k (x) (Δ + k2 )N ϕ (x) dx + Iε + IIε
ε→0 Rn \B(0,ε)

= ΦkN (x) (Δ + k2 )N ϕ (x) dx + lim+ Iε + lim+ IIε , (7.7.32)
Rn ε→0 ε→0
where we have set (recall the definition of the normal derivative from (7.1.14))

∂
Iε := − Φ(N+1) (x) (Δ + k2 )N ϕ (x) dσ(x), (7.7.33)
∂B(0,ε)
k ∂ν

∂Φ(N+1)
IIε := k
(x) (Δ + k2 )N ϕ (x) dσ(x). (7.7.34)
∂B(0,ε) ∂ν
Above, ν(x) = εx for each x ∈ ∂B(0, ε), and σ denotes the surface measure on
∂B(0, ε). By (6) in Proposition 7.39 and the induction hypothesis we have

ΦkN (x) (Δ + k2 )N ϕ (x) dx = ΦkN , (Δ + k2 )N ϕ
Rn

= (Δ + k2 )N ΦkN , ϕ = δ, ϕ = ϕ(0). (7.7.35)
Next, let ε ∈ (0, 1) be arbitrary and focus on estimating |Iε |. From (7.7.33) we see
that

∂
|Iε | ≤ (N+1)
(x) (Δ + k ) ϕ (x) dσ(x)
Φk
2 N
∂B(0,ε) ∂ν
55 α 55

≤ max 5∂ ϕ5L∞ (B(0,1)) Φ(N+1) (x) dσ(x), (7.7.36)
α∈N0 , |α|≤2N+1
n k
∂B(0,ε)
Bringing in (7.7.14) (applied with N replaced by N +1), and denoting by C constants

that may depend on ϕ, N, n, k, but are independent of ε, we may further estimate
⎧ 2N+1
⎪
⎪ Cε , if n > 2N + 2,
⎪
⎪
⎪
⎪
⎨
|Iε | ≤ ⎪
⎪ C| ln(kε)|εn−1 , if n = 2N + 2, (7.7.37)
⎪
⎪
⎪
⎪
⎩ n−1
Cε , if n < 2N + 2.
Recalling that we are assuming n ≥ 2, it is immediate that the terms in the right-hand
side of (7.7.37) vanish as ε → 0+ . Consequently,
lim Iε = 0. (7.7.38)
ε→0+
Moving on, for each ε ∈ (0, 1) starting with (7.7.34) we estimate

∂
|IIε | ≤ Φ(N+1) (x)(Δ + k2 )N ϕ (x) dσ(x),
∂B(0,ε) ∂ν
k
55 α 55
∂
≤ max 5 ∂ ϕ 5 ∞
Φ(N+1) (x) dσ(x). (7.7.39)
∂B(0,ε) ∂ν
α∈Nn0 , |α|≤2N L (B(0,1)) k
On account of (7.7.20) (applied with N replaced by N + 1),

⎧ 2N
⎪
⎪
⎪ Cε , if n > 2N,
⎪
⎪
⎪
⎪
⎨
|IIε | ≤ ⎪
⎪ C| ln(kε)|εn , if n = 2N, (7.7.40)
⎪
⎪
⎪
⎪
⎪
⎩ Cεn , if n < 2N,
hence,
lim IIε = 0. (7.7.41)

ε→0+
Combining (7.7.32), (7.7.35), (7.7.38), and (7.7.41), we arrive at

(Δ + k2 )N+1 Φ(N+1)
k , ϕ = ϕ(0). (7.7.42)
Since ϕ ∈ C0∞ (Rn ) is arbitrary, this implies (Δ+k2 )N+1 Φ(N+1)

k = δ in D (Rn ), finishing
the proof of the theorem.
It turns out that the pointwise identity from item (3) of Proposition 7.39 is a man-
ifestation of a more general formula, valid in the sense of distributions, described
below. This may then be used to give an alternative proof of Theorem 7.40 (see
Exercise 7.42).
Proposition 7.41. Let n ∈ N, n ≥ 2, and k ∈ (0, ∞). For N ∈ N recall the function
Φ(N)
k from (7.7.9). Then for each N ∈ N one has
(Δ + k2 )Φ(N+1)
k = Φ(N)
k in D (Rn ). (7.7.43)
Proof. Fix N ∈ N and recall the family of functions Fλ , indexed by λ, as in (7.6.78).

In terms of this, Φ(N)
k may be expressed as
Φ(N)
k (x) = aN,k F (n−2N)/2 (x), ∀ x ∈ Rn \ {0}, (7.7.44)
where we have abbreviated

(−1)N i k(n−2N)/2
aN,k := . (7.7.45)
2(n+2N)/2 (N − 1)!π(n−2)/2
Then, for each j ∈ {1, . . . , n}, formula (7.6.79) in Proposition 7.38 (applied with
λ := (n − 2N − 2)/2 < (n − 1)/2) implies
∂ j Φ(N+1)
k = −aN+1,k kF(n−2N)/2 (x)x j in D (Rn ). (7.7.46)
A second application of (7.6.79) (this time with λ := (n − 2N)/2 < (n − 1)/2) yields
that for each j ∈ {1, . . . , n} we have
∂2j Φ(N+1)
k = aN+1,k k2 F(n−2N+2)/2 (x)x2j − aN+1,k kF(n−2N)/2 (7.7.47)
in D (Rn ). Consequently, (7.7.47) and the definition of Fλ from (7.6.78) give
(Δ + k2 )Φ(N+1)
k = aN+1,k k2 F(n−2N+2)/2 (x)|x|2 − naN+1,k kF(n−2N)/2
+ k2 aN+1,k F(n−2N−2)/2
(1)
= k2 aN+1,k |x|−(n−2N−2)/2 H(n−2N+2)/2 (1)
(k|x|) + H(n−2N−2)/2 (k|x|)
− n aN+1,k k|x|−(n−2N)/2 H(n−2N)/2

(1)
(k|x|) (7.7.48)
in D (Rn ). By (6) in Lemma 14.71 we have
(1) (1) n − 2N (1)

H(n−2N+2)/2 (k|x|) + H(n−2N−2)/2 (k|x|) = H(n−2N)/2 (k|x|) (7.7.49)
k|x|
for each x ∈ Rn \ {0}, hence also in D (Rn ). In concert, (7.7.48) and (7.7.49) imply
(Δ + k2 )Φ(N+1)
k = −2NaN+1,k k|x|−(n−2N)/2 H(n−2N)/2
(1)
(k|x|) (7.7.50)
in D (Rn ). The desired conclusion now follows from (7.7.50) and (7.7.9) upon
observing that −2NaN+1,k k = aN,k .
Exercise 7.42. Use Proposition 7.41 and induction over N to give another proof of
Theorem 7.40, i.e., that Φ(N)
k as in (7.7.9) is a fundamental solution for the iterated
Helmholtz operator (Δ + k2 )N in Rn .
7.8 Fundamental Solutions for the Cauchy–Riemann Operator
In this section we determine all fundamental solutions for the Cauchy-Riemann

operator ∂z∂ := 12 (∂1 +i∂2 ) and its conjugate, ∂z∂ := 12 (∂1 −i∂2 ) that belong to S (R2 ). It
is immediate that ∂z∂ ∂z∂ = 14 Δ when acting on distributions from D (R2 ). Given this,
Remark 5.7 is particularly useful. Specifically, since by Theorem 7.2 we have that
E(x) = 2π1
ln |x|, x ∈ R2 \ {0}, is a fundamental solution for Δ in R2 and E ∈ S (R2 ),
Remark 5.7 implies that
∂ 2 ∂
ln |x| ∈ S (R2 ) is a fundamental solution for in S (R2 ), (7.8.1)
∂z π ∂z
∂ 2 ∂
ln |x| ∈ S (R2 ) is a fundamental solution for in S (R2 ). (7.8.2)
∂z π ∂z

We proceed with computing ∂z∂ π2 ln |x| in the distributional sense in S (R2 ). By (c)

in Theorem 4.14 it suffices to compute the distributional derivative ∂z∂ π2 ln |x| in
D (R2 ). We will show that this distributional
derivative
is equal to the distribution
given by the function obtained by taking ∂z∂ π2 ln |x| in the classical sense for x 0.
To this end, fix a function ϕ ∈ C0∞ (R2 ). Since ln |x| ∈ Lloc
1
(R2 ) (recall Example 2.9)
we may write
∂ 2 2 ∂ϕ
ln |x| , ϕ = − ln |x|,
∂z π π ∂z

1
=− ln |x|[(∂1 + i∂2 )ϕ(x)] dx. (7.8.3)
π R2
Let R > 0 be such that supp ϕ ⊂ B(0, R). By Lebesgue’s Dominated Convergence
Theorem 14.15 we further have

ln |x| [(∂1 + i∂2 )ϕ(x)] dx = lim+ ln |x| [(∂1 + i∂2 )ϕ(x)] dx. (7.8.4)
R2 ε→0 ε≤|x|≤R
Fix ε ∈ (0, R/2) and with x = (x1 , x2 ) use (14.8.4) to write

ln |x| (∂1 + i∂2 )ϕ(x) dx (7.8.5)

ε≤|x|≤R

ln ε
=− ϕ(x)(∂1 + i∂2 ) ln |x| dx − (x1 + ix2 )ϕ(x) dσ(x),
|x|≥ε ε |x|=ε
where we have used the fact that the outward unit normal to B(0, ε) at a point x on
∂B(0, ε) is 1ε (x1 , x2 ) and that ϕ vanishes on ∂B(0, R). Using polar coordinates, we
further write
7.8 Fundamental Solutions for the Cauchy–Riemann Operator 289
ln ε

(x1 + ix2 )ϕ(x) dσ(x)
ε |x|=ε

2π

= ε ln ε (cos θ + i sin θ)ϕ(ε cos θ, ε sin θ) dθ
0
≤ Cε| ln ε| −−−−→
+
0. (7.8.6)
ε→0
Since (∂1 + i∂2 )[ln |x|] ∈ Lloc

1
(R2 ) and ϕ is compactly supported, by Lebesgue’s
Dominated Convergence Theorem we have

lim+ ϕ(x)(∂1 + i∂2 ) ln |x| dx = ϕ(x)(∂1 + i∂2 ) ln |x| dx. (7.8.7)

ε→0 |x|≥ε R2
By combining (7.8.3)–(7.8.7) we conclude that

∂ 2 1 1 1
ln |x| = (∂1 + i∂2 )[ln |x|] = · in S (R2 ). (7.8.8)
∂z π π π x1 − ix2
Similarly,

∂ 2 1 1 1
ln |x| = (∂1 − i∂2 )[ln |x|] = · in S (R2 ). (7.8.9)
∂z π π π x1 + ix2
With this at hand, by invoking Proposition 5.8, we have a complete description of

all fundamental solutions for ∂z∂ and ∂z∂ that are also tempered distributions.
In summary, we proved the following theorem.
Theorem 7.43. Consider the functions

1 1 1 1
E(x1 , x2 ) := · , F(x1 , x2 ) := · , (7.8.10)
π x1 − ix2 π x1 + ix2
defined for x = (x1 , x2 ) ∈ R2 \ {0}. Then E ∈ S (R2 ) and is a fundamental solution

for the operator ∂z∂ . Also F ∈ S (R2 ) and is a fundamental solution for the operator
∂
∂z .
Moreover,

u ∈ S (R2 ) : ∂z∂ u = δ in S (R2 ) (7.8.11)
∂
= E + P : P polynomial in R2 satisfying P = 0 in R2
∂z
and

u ∈ S (R2 ) : ∂z∂ u = δ in S (R2 ) (7.8.12)
∂
= F + P : P polynomial in R2 satisfying P = 0 in R2 .
∂z
Having identified the fundamental solutions in R2 for the Cauchy–Riemann oper-

ator ∂ = 12 (∂ x +i∂y ), we shall now compute the Fourier transform of 1/z. Throughout,
we shall repeatedly identify R2 with C.
Proposition 7.44. Let u(z) := 1
z for all z ∈ C \ {0}. Then u ∈ S (R2 ) and
2π

u(ξ) = in S (R2 ). (7.8.13)
iξ
Proof. That u ∈ S (R2 ) follows by observing that u is locally integrable near the
−1
|u(z)| decays like |z| for |z| > 1. To prove (7.8.13), first note that,
origin, while
since ∂ πz = δ in S (R ) by Theorem 7.43, it follows that ∂u = π δ in S (R2 ).
1 2
Hence
i 2π
π = ∂u(ξ) = ξ u(ξ) =⇒ ξ
u(ξ) = . (7.8.14)
2 i

Thus, ξ 2π u(ξ) = 0 in S (R2 ) which when combined with Example 2.76 implies
iξ −
2π
−
u(ξ) = c δ, for some c ∈ C. (7.8.15)
iξ
Thus,
2π
u −
u = cδ in S (R2 ). (7.8.16)
i
Taking another Fourier transform yields
2π
u + (2π)2 u = c in S (R2 ),
(7.8.17)
i
given that

u = −(2π)2 u (keeping in mind that u is odd). A linear combination
of (7.8.16) and (7.8.17) which eliminates u then leads us to the conclusion that
(−2iπδ + 1) c = 0 in S (R2 ). In turn, this forces c = 0, concluding the proof of the
proposition.
Exercise 7.45. Consider u(z) := 1/z for all z ∈ C \ {0}. Show that u ∈ S (R2 ) and
2π

u(ξ) = in S (R2 ). (7.8.18)
iξ
Hint: Use Proposition 7.44 and Exercise 3.27.
Proposition 7.46. Let Ω be an open set in C that contains the origin. Suppose u ∈
C 0 (Ω) is such that ∂u 1 1 ∂u
∂z ∈ Lloc (Ω) and z ∂z ∈ Lloc (Ω). Then z ∈ Lloc (Ω) and
1 u 1
∂ u 1 ∂u
= πu(0)δ + in D (Ω). (7.8.19)
∂z z z ∂z
Proof. Pick a function θ ∈ C ∞ (C) with the property that θ = 0 on B(0, 1) and θ ≡ 1
on C \ B(0, 2). For each ε ∈ (0, 1) define the function θε : C → C by setting
7.8 Fundamental Solutions for the Cauchy–Riemann Operator 291
θε (z) := θ(z/ε) for each z ∈ C. Then there exists some constant C ∈ (0, ∞) such that
θε ∈ C ∞ (C), supp (∇θε ) ⊆ B(0, 2ε) \ B(0, ε), (7.8.20)

lim θε (z) = 1 and |θε (z)| ≤ C ∀ z ∈ C. (7.8.21)
ε→0+
Next, fix ϕ ∈ C0∞ (Ω) and, with L2 denoting the Lebesgue measure in R2 , write
∂ u u ∂ϕ
u ∂ϕ
,ϕ = − , =− dL2
∂z z z ∂z Ω z ∂z

u ∂ϕ
= − lim+ θε dL2 , (7.8.22)
ε→0 Ω z ∂z
where for the last equality in (7.8.22) we used (7.8.21) and Lebesgue’s Dominated
Convergence Theorem. Note that
1 ∂ 1
ϕ θε ∈ C0∞ (Ω) and =0 in C \ {0}. (7.8.23)
z ∂z z
Therefore,
∂ 1 1 ∂ϕ 1 ∂θε
ϕ θε = θ ε + ϕ in Ω. (7.8.24)
∂z z z ∂z z ∂z
Combining (7.8.22) and (7.8.24) we obtain
∂ u

∂ 1 u ∂θε
, ϕ = − lim+ u ϕ θε dL + lim+
2
ϕ dL2
∂z z ε→0 Ω ∂z z ε→0 Ω z ∂z
∂u 1
u ∂θε
= lim+ , ϕ θε + lim+ ϕ dL2 =: I + II. (7.8.25)
ε→0 ∂z z ε→0 Ω z ∂z
Using the hypotheses on u, (7.8.21), and Lebesgue’s Dominated Convergence The-
orem, we obtain

1 ∂u
∂u 1 ∂u 1
I = lim+ ϕ θε dL =
2
ϕ dL2 = ,ϕ . (7.8.26)
ε→0 Ω ∂z z Ω ∂z z z ∂z
To compute II, first write

u ∂θε u − u(0) 1 ∂θ 1 ∂θε

ϕ dL2 = ϕ (·/ε) dL2 + u(0) ϕ dL2
Ω z ∂z Ω z ε ∂z Ω z ∂z
=: III + IV. (7.8.27)
Using the support condition from (7.8.20), the continuity of u and the properties of
ϕ, term III may be estimated by

|III| ≤ CϕL∞ (Ω) ∇θL∞ (Rn ) sup |u − u(0)| −−−−→
+
0. (7.8.28)
B(0,2ε) ε→0
As for IV, we have

1 ∂θε 1 ∂ϕ
lim IV = u(0) lim+ ϕ dL2 = −u(0) lim+ θε dL2
ε→0+ ε→0 Ω z ∂z ε→0 Ω z ∂z

1 ∂ϕ ∂ 1
1 ∂ϕ
= −u(0) dL2 = −u(0) , = u(0) ,ϕ
Ω z ∂z z ∂z ∂z z
= πu(0)δ, ϕ . (7.8.29)
For the second equality in (7.8.29), for each ε ∈ (0, 1) fixed, we used integration
by parts (cf. (14.8.4)) on D \ B(0, ε/2), where D is a bounded open subset of Ω
with the property that supp ϕ ⊂ D. At this step is also useful to recall (7.8.23). For
the third equality in (7.8.29), we applied Lebesgue’s Dominated Convergence The-
orem, while the fifth is based on Theorem 7.43. Now (7.8.19) follows by combining
(7.8.25)–(7.8.29), since ϕ is arbitrary in C0∞ (Ω).
Exercise 7.47. Let Ω be an open set in C and let z0 ∈ Ω. Suppose u ∈ C 0 (Ω) is such
that ∂u 1 1 ∂u
∂z ∈ Lloc (Ω) and z−z0 ∂z ∈ Lloc (Ω). Then z−z0 ∈ Lloc (Ω) and
1 u 1
∂ u 1 ∂u
= πu(z0 )δz0 + in D (Ω). (7.8.30)
∂z z − z0 z − z0 ∂z
Proposition 7.48. If ϕ ∈ S(R) is a complex-valued function, define the Cauchy

operator by

1 ϕ(x)
(C ϕ)(z) := dx, ∀ z ∈ C \ R. (7.8.31)
2πi R x − z
Then the following Plemelj jump-formula holds at every x ∈ R:

1 1 ϕ(y)
lim (C ϕ)(x + iy) = ± ϕ(x) + lim+ dy. (7.8.32)
y→0± 2 ε→0 2πi y−x
y∈R
| x−y|>ε
Proof. Apply Corollary 4.81 to the function Φ : R2 \ {(0, 0)} → C defined by

−1
Φ(x, y) := for all (x, y) ∈ R2 \ {(0, 0)} . (7.8.33)
2πi(x + iy)
Note that Φ is C ∞ , odd, and homogeneous of degree −1 and, under the canonical
−1
identification R2 ≡ C, takes the form Φ(z) = 2πiz for z ∈ C \ {0}. Proposition 7.44

then gives Φ(ξ) 1) = −i.
= ξ for all ξ ∈ C \ {0} which, in particular, yields Φ(0,
1
Having established this, (7.8.32) follows directly from (4.7.46).

Remark 7.49. Upon recalling formula (4.9.31) for the Hilbert transform H on the
real line, we may recast the version of Plemelj jump-formula (7.8.32) corresponding
to considering the Cauchy operator in the upper half-plane in the form
7.9 Fundamental Solutions for the Dirac Operator 293
ver 1 i
C ϕ 2 = ϕ + Hϕ in R, ∀ ϕ ∈ S(R), (7.8.34)
∂R+ 2 2π
where the “vertical limit” of C ϕ to the boundary of the upper half-plane is under-
stood as in (4.8.20). In turn, formula (7.8.34) suggests the consideration of the oper-
ator (with I denoting the identity)
1 i
P := I+ H. (7.8.35)
2 2π
From Corollary 4.99 it follows that P is a well-defined, linear, and bounded oper-
ator on L2 (R). Using the fact that H 2 = −π2 I and H ∗ = −H on L2 (R) (again, see
Corollary 4.99), we may then compute
1 i 2 1 i i 2
P2 = I+ H = I+ H+ H2
2 2π 4 2π 2π
1 i 1 1 i
= I+ H+ I= I+ H
4 2π 4 2 2π
= P, (7.8.36)
and
1 i ∗ 1 i
P∗ = I− H = I+ H = P. (7.8.37)
2 2π 2 2π
Any linear and bounded operator on L2 (R) satisfying these two properties (i.e., P 2 =
P and P∗ = P) is called a projection. Then one may readily verify that I − P =
2 I − 2π H is also a projection and if we introduce (what are commonly referred to
1 i
as Hardy spaces on the real line)

( 1 i )
H±2 (R) := I± H f : f ∈ L2 (R) , (7.8.38)
2 2π
then any complex-valued function f ∈ L2 (R) may be uniquely decomposed as f =
f+ + f− with f± ∈ H±2 (R) and, moreover, any two functions f± ∈ H±2 (R) are
orthogonal, in the sense that

f+ (x) f− (x) dx = 0. (7.8.39)

R
7.9 Fundamental Solutions for the Dirac Operator
In a nutshell, Dirac operators are first-order differential operators factoring the

Laplacian. When n = 1, Δ = d2 /dx2 , hence if we set D := i(d/dx), then D2 = −Δ.
We seek a higher dimensional generalization of the latter factorization formula. The
natural context in which such a generalization may be carried out is the Clifford
algebra setting.
The Clifford algebra with n generators Cn is the associative algebra with unit

Cn , , +, 1 freely generated over R by the family {e j }1≤ j≤n , of the standard or-
thonormal base in Rn , now called imaginary units, subject to the following axioms:
e j ek + ek e j = −2δ jk , ∀ j, k ∈ {1, ..., n}. (7.9.1)
Hence,
e j ek = −ek e j if 1 ≤ j k ≤ n,
(7.9.2)
and e j e j = −1 for j ∈ {1, ..., n}.
The first condition above indicates that Cn is noncommutative if n > 1, while the
second condition justifies calling {e j }1≤ j≤n imaginary units. Elements in the Clifford
algebra Cn can be uniquely written in the form

n

a= aI eI (7.9.3)
l=0 |I|=l
with aI ∈ C, where eI stands for the product ei1 ei2 · · · eil if I = (i1 , i2 , . . . , il )
with 1 ≤ i1 < i2 < · · · < il ≤ n, e∅ := 1 ∈ R (that plays the role of the multiplicative

unit in Cn ) and indicates that the sum is performed over strictly increasingly
|I|=l
ordered indexes I with l components (selected from the set {1, ..., n}). In the writing
(7.9.3) we shall refer to the numbers aI ∈ C as the scalar components of a.
Exercise 7.50. Given any a, b ∈ Cn with scalar components aI , bI ∈ C define

n

n

(a, b) := aI bI , a := aI eI , (7.9.4)
l=0 |I|=l l=0 |I|=l
and abbreviate Ma b := a b. Prove that
Ma b = Ma b and (Ma b , c) = (b , Ma c) for every a, b, c ∈ Cn . (7.9.5)
Clifford algebra-valued functions defined in an open set Ω ⊆ Rn may be defined

naturally. Specifically, any function f : Ω → Cn is an object of the form

n

f = fI eI , (7.9.6)
l=0 |I|=l
where each component fI is a complex-valued function defined in Ω. Given any

k ∈ N0 ∪ {∞}, we shall denote by C k (Ω, Cn ) the collection of all Clifford algebra-
valued functions f whose scalar components fI are of class C k in Ω. In a similar
manner we may define C0∞ (Ω, Cn ), Lloc
1
(Ω, Cn ), L p (Ω, Cn ), etc.
In fact, we may also consider Clifford algebra-valued distributions in an open set
Ω ⊆ Rn . Specifically, write u ∈ D (Ω, Cn ) provided

n

u= uI eI with uI ∈ D (Ω) for each I. (7.9.7)
l=0 |I|=l
In particular we have that the Dirac distribution δ = δe∅ , and the action of a Clifford
algebra-valued distribution u ∈ D (Ω, Cm ) as in (7.9.7) on a test function ϕ ∈ C0∞ (Ω)
is naturally defined by
m

u(ϕ) := uI , ϕ eI . (7.9.8)
=0 |I|=
Much of the theory originally developed for scalar-valued distributions extends in a

natural fashion to the current setting. For example, if u ∈ D (Ω, Cn ) is as in (7.9.7),
we may define
n
α
∂α u := ∂ uI eI , ∀ α ∈ Nn0 , (7.9.9)
l=0 |I|=l
and

n

n

f u := f J uI e J eI , (7.9.10)
=0 |J|= k=0 |I|=k

n

for any f = f J e J ∈ C ∞ (Ω, Cn ).
k=0 |J|=k
We are ready to introduce the Dirac operator D associated with Cn . Specifically,
given a Clifford algebra-valued distribution u ∈ D (Ω, Cn ) as in (7.9.7), we define
Du ∈ D (Ω, Cn ) by setting

n
n

Du := ∂ j uI e j eI . (7.9.11)
j=1 l=0 |I|=l
In other words,

n
D := Me j ∂ j , (7.9.12)
j=1
where Me j denotes the operator of Clifford algebra multiplication by e j from the

left.
Proposition 7.51. Let Ω ⊆ Rn be an open set. Then the Dirac operator D satisfies
D2 = −Δ in D (Ω, Cn ), (7.9.13)
where Δ is the Laplacian.

n

Proof. Pick an arbitrary u ∈ D (Ω, Cn ), say u = uI eI with uI ∈ D (Ω) for
l=0 |I|=l
each I. Then using (7.9.11) twice yields

n
n

D2 u = D(Du) = ∂k ∂ j uI ek e j eI . (7.9.14)
j,k=1 l=0 |I|=l
Observe that, on the one hand,

n

∂k ∂ j uI ek e j eI = 0, (7.9.15)
1≤ jk≤n l=0 |I|=l
since for each I, we have ∂k ∂ j uI = ∂ j ∂k uI and ek e j eI = −e j ek eI whenever

j k by the first formula in (7.9.2). On the other hand, corresponding to the case
when j = k

n
n

n
n

∂2k uI ek ek eI = − ∂2k uI eI = −Δu, (7.9.16)
k=1 l=0 |I|=l k=1 l=0 |I|=l
since ek ek = −1 by the second formula in (7.9.2).
Exercise 7.52. Consider the embedding

n
Rn → Cn , Rn x = (x j )1≤ j≤n ≡ x j e j ∈ Cn , (7.9.17)
j=1
which identifies vectors from Rn with elements in the Clifford algebra Cn . With this
identification in mind, show that
x x = −|x|2 for any x ∈ Rn , (7.9.18)

x y + y x = −2x · y for any x, y ∈ Rn . (7.9.19)
In light of the embedding described in (7.9.17) we may regard the assignment

Rn \ {0} x → |x|xn as a Clifford algebra-valued function.
Proposition 7.51 combined with Remark 5.7 yields the following result.
Theorem 7.53. The Clifford algebra-valued function

1 x
E(x) := − ∈ Lloc
1
(Rn , Cn ) ∩ S (Rn , Cn ) (7.9.20)
ωn−1 |x|n
is a fundamental solution for the Dirac operator D in Rn ,, i.e.,
DE = δ in S (Rn , Cn ). (7.9.21)
Moreover, any u ∈ S (Rn , Cn ) satisfying Du = δ in S (Rn , Cn ) is of the form E + P

where E is as in (7.9.20) and P is a Clifford algebra-valued function whose compo-
nents are polynomials and such that DP = 0 in Rn .
Proof. Let EΔ be the fundamental solution for Δ as described in (7.1.12) for n ≥ 2.

From (7.9.13) and Remark 5.7, we may infer that −DEΔ computed in D (Rn ) is a
fundamental solution for the Dirac operator D in Rn . From Exercise 2.128 (when
n = 2) and Exercise 2.129 (when n ≥ 3) we obtain

n n
1 xj
−DEΔ = − ∂ j (EΔ ) e j = − · n ej
j=1 j=1
ω n−1 |x|
1 x
=− · , in D (Rn ). (7.9.22)
ωn−1 |x|n
This proves that the distribution in (7.9.20) is indeed a fundamental solution for the
Dirac operator D in Rn . To justify the last claim in the statement of the theorem,
let F ∈ S (Rn , Cn ) be an arbitrary fundamental solution for the Dirac operator D in
Rn . Then the tempered distribution P := F − E satisfies DP = 0 in S (Rn , Cn ) and,
on the Fourier transform side, we have ξ P = 0 in S (Rn , Cn ). Multiplying (in the
Clifford algebra sense) this equality with the Clifford algebra-valued function with
polynomial growth ξ then yields −|ξ|2 P = 0 in S (Rn , Cn ) (cf. (7.9.18)). In turn, the
latter implies supp P ⊆ {0}, hence the components of P are polynomials in Rn (cf.
Exercise 4.37).
There are other versions of the Dirac operator D from (7.9.12) that are more in
line with the classical Cauchy–Riemann operator ∂z∂ := 12 (∂1 + i∂2 ). Specifically, in
Rn+1 set x = (x0 , x1 , x2 , . . . , xn ) ∈ Rn+1 and consider

n
D± := ∂0 ± D = ∂0 ± Me j ∂ j , (7.9.23)
j=1
Note that in the case when n = 1 the Dirac operator D− corresponds to a constant
multiple of the Cauchy–Riemann operator ∂z∂ . A reasoning similar to that used above
for D also yields fundamental solutions for D± . We leave this as an exercise for the
interested reader.
Exercise 7.54. Let Ω ⊆ Rn+1 be an open set and let x = (x0 , x1 , x2 , . . . , xn ) ∈ Rn+1 .
Then the Dirac operators D± satisfy
D+ D− = D− D+ = Δn+1 in D (Ω, Cn ), (7.9.24)
where Δn+1 is the Laplacian in Rn+1 .

Moreover, use a reasoning similar to the one used in the proof of Theorem 7.53
to show that the functions

n
x0 + x je j
+ 1 j=1
E (x) : = ∈ Lloc
1
(Rn+1 , Cn ) → D (Rn+1 , Cn ) (7.9.25)
ωn |x|n+1

n
x0 − x je j
1 j=1
E − (x) : = ∈ Lloc
1
(Rn+1 , Cn ) → D (Rn+1 , Cn ) (7.9.26)
ωn |x|n+1
are fundamental solution for the Dirac operators D− and D+ , respectively, in Rn+1 .
We next introduce the Cauchy-Clifford operator and discuss its jump-

formulas in the upper and lower half-spaces.
Theorem 7.55. For each Cn -valued function ϕ ∈ S(Rn−1 ) define the Cauchy–
Clifford operator

n−1
1 j=1 (x j − y j )e j + xn en
(C ϕ)(x) := − n en ϕ(y ) dy , (7.9.27)
ωn−1 Rn−1 x − (y , 0)
for each x = (x1 , . . . , xn ) ∈ Rn with xn 0. Then for each x ∈ Rn−1 one has
lim (C ϕ)(x , xn ) (7.9.28)

xn →0±

n−1
1 1 j=1 (y j − x j )e j
= ± ϕ(x ) + lim+ en ϕ(y ) dy .
2 ε→0 ωn−1 |x − y |n
y ∈Rn−1
|x −y |>ε
Proof. Consider the Clifford algebra-value function Φ : Rn \ {0} → Cn given by

n
j=1 x j e j
Φ(x) := − en for each x = (x1 , . . . , xn ) ∈ Rn \ {0}. (7.9.29)
ωn−1 |x|n
Then Φ is C ∞ , odd, and positive homogeneous of degree 1 − n in Rn \ {0}. As such,
Corollary 4.81 may be applied (to each component of Φ). In this regard note from
Corollary 4.65 that
⎡ n ⎤ ⎡ n ⎤
1 ⎢⎢⎢ x j ⎥⎥⎥ ⎢⎢⎢ ξ j ⎥⎥⎥
=−
Φ(ξ) ⎢⎢⎢ F (ξ) e j ⎥⎥⎥⎦ en = i ⎢⎢⎢⎣ e j ⎥⎥⎥ en (7.9.30)
ωn−1 ⎣ j=1
|x|n |ξ|2 ⎦
j=1
in S (Rn ). In particular, since en en = −1, we obtain

, 1) = i en en = −i.
Φ(0 (7.9.31)
Given that, as seen from (7.9.27) and (7.9.29), we have

(C ϕ)(x) = Φ(x − y , xn ) ϕ(y ) dy , (7.9.32)

Rn−1
for each x = (x1 , . . . , xn ) ∈ Rn with xn 0, the jump-formulas for the Cauchy–

Clifford operator in (7.9.28) follow from (4.7.46) and (7.9.31).
Paralleling the discussion in Remark 7.49, in the higher dimensional setting we
have the following connection between the Cauchy–Clifford operator and Riesz
transforms.
Remark 7.56. Let R j , j ∈ {1, . . . , n−1}, be the Riesz transforms in Rn−1 (i.e., singular
integral operators defined as in (4.9.11) with n − 1 in place of n). Also, recall the
definition of the “vertical limit” of a function defined in Rn+ to the boundary of

the upper half-space from (4.8.20). Then we may express the version of the jump-
formula (7.9.28) corresponding to considering the Cauchy–Clifford operator in the
upper half-space as
ver 1 1
n−1
C ϕ n = ϕ − e j en (R j ϕ) in Rn−1 (7.9.33)
∂R+ 2 ωn−1 j=1
for each Cn -valued function ϕ ∈ S(Rn−1 ), where the Riesz transforms act on ϕ
componentwise.
The format of the jump-formula displayed in (7.9.33) suggests considering the

operator acting on Cn -valued functions according to
1
n−1
1
P := I− e j en R j
2 ωn−1 j=1
1 1
n−1
= I+ en ej Rj , (7.9.34)
2 ωn−1 j=1
where I stands for the identity operator. In the second line of (7.9.34) the change in
sign is due to the formula (cf. (7.9.2))
e j en = −en e j for each j ∈ {1, . . . , n − 1}. (7.9.35)
Theorem 4.97 then gives that P is a well-defined, linear, and bounded operator on
L2 (Rn−1 , Cn ). Moreover, since each R j has a real-valued kernel, its action commutes
with multiplication by elements from Cn (i.e., for every a ∈ Cn we have R j Ma =
Ma R j , in the notation from Exercise 7.50). Keeping this in mind and relying on
(7.9.35) and the fact that e2n = −1 (cf. (7.9.2)), we may then write
1 1
n−1 2
P2 = I+ en ej Rj
2 ωn−1 j=1
1 1 n−1 1 2 2 n−1
= I+ en ej Rj + ej Rj . (7.9.36)
4 ωn−1 j=1
ωn−1 j=1
Furthermore, using the fact that, as proved in Theorem 4.97, the Riesz transforms
2
n−1 2
commute with one another and satisfy R j = − ωn−1
2 I on L2 (Rn−1 ), we may
j=1
expand

n−1 2
n−1
ej Rj = e j ek R j Rk
j=1 j,k=1

n−1
= e2j R2j + e j ek R j Rk
j=1 1≤ jk≤n−1

n−1
ωn−1 2
=− R2j = I, (7.9.37)
j=1
2
where the source of the cancelation taking place in the third equality above is the
observation that e j ek R j Rk = −ek e j Rk R j whenever 1 ≤ j k ≤ n − 1.
Combining (7.9.36)–(7.9.37) and recalling (7.9.34) then yields
P 2 = P on L2 (Rn−1 , Cn ). (7.9.38)
Note that this is in agreement with the result obtained in (7.8.36) in the case of
the two-dimensional setting. Let us also consider the higher dimensional analogue
of (7.8.37). In this regard, we first observe that based on Exercise 7.50 for any
f, g ∈ L2 (Rn−1 , Cn ) we may write

n−1
Men Me j (R j f )(x ) , g(x ) dx
Rn−1 j=1

n−1
= (R j f )(x ) , Me j Men g(x ) dx
Rn−1 j=1

n−1
=− f (x ) , Me j Men (R j g)(x ) dx
Rn−1 j=1

n−1
= f (x ) , Men Me j (R j g)(x ) dx . (7.9.39)
Rn−1 j=1
From this and (7.9.34) it follows that for every f, g ∈ L2 (Rn−1 , Cn ) we have

(P f )(x ), g(x ) dx = f (x ), (Pg)(x ) dx , (7.9.40)
Rn−1 Rn−1
a condition that we shall interpret simply as
P∗ = P on L2 (Rn−1 , Cn ). (7.9.41)
In summary, the above analysis shows that the operator P defined as in (7.9.34)
is a projection on L2 (Rn−1 , Cn ). Starting from this result, a corresponding higher
7.10 Fundamental Solutions for the Perturbed Dirac Operator 301
dimensional Hardy space theory may be developed in the Clifford algebra setting as
well.
7.10 Fundamental Solutions for the Perturbed Dirac Operator
In this section we will work within the framework of Cn+1 , the Clifford algebra
with n + 1 imaginary units {e j }1≤ j≤n+1 . This is the Clifford algebra introduced in
Section 7.9 with n replaced by n + 1. We continue to denote by Ω an arbitrary open
set in Rn and will be working with Cn+1 -valued functions defined in Ω, which are
functions f : Ω → Cn+1 of the form

n+1

f = fI eI , (7.10.1)
l=0 |I|=l
whose components fI ’s are complex-valued functions defined in Ω. As before, we

will also use the notation C m (Ω, Cn+1 ), Lloc
1
(Ω, Cn+1 ), L p (Ω, Cn+1 ), etc., for the col-
lection of Cn+1 -valued functions defined in Ω with scalar components in C m (Ω),
1
Lloc (Ω), L p (Ω), etc., respectively (here m ∈ N ∪ {∞}). Similarly, D (Ω, Cn+1 ) de-
notes Cn+1 -valued distributions in Ω. Hence, any u ∈ D (Ω, Cn+1 ) is of the form

n+1

u= uI eI with uI ∈ D (Ω) for each I, (7.10.2)
l=0 |I|=l
and the action of the Dirac operator D from (7.9.12) on u becomes

n
n+1

Du := ∂ j uI e j eI . (7.10.3)
j=1 l=0 |I|=l
Next, fix k ∈ C arbitrary. Then the perturbed Dirac operator is denoted

by Dk and is defined as
Dk := D + ken+1 . (7.10.4)
Hence, if u ∈ D (Ω, Cn+1 ) is as in (7.10.2), then the action of Dk on u is given as

n
n+1

n+1

Dk u = ∂ j uI e j eI + kuI en+1 eI . (7.10.5)
j=1 l=0 |I|=l l=0 |I|=l
It is easy to see by an inspection of the proof of Proposition 7.51 that
D2 = −Δ in D (Ω, Cn+1 ), (7.10.6)
and that
D2k = −(Δ + k2 ) in D (Ω, Cn+1 ), (7.10.7)
n
where Δ := j=1 ∂2j is the Laplacian in Rn . Moreover, since Cn → Cn+1 , the
embedding from (7.9.17) now becomes

n
Rn → Cn+1 , Rn x = (x j )1≤ j≤n ≡ x j e j ∈ Cn+1 , (7.10.8)
j=1
the identities in (7.9.18) and (7.9.19) continue to hold in Cn+1 , and the assignment
Rn \ {0} x → |x|xm is a Cn+1 -valued function for each m ∈ R.
Identity (7.10.7), Remark 5.7, and Theorem 7.33 are the main ingredients in the
proof of the following result.
Theorem 7.57. Let k ∈ (0, ∞) and suppose n ∈ N, n ≥ 2. Then the Clifford algebra-
valued function
cn kn/2 (1) x (1)

Ek (x) := Hn/2 (k|x|) − H(n−2)/2 (k|x|) en+1 , (7.10.9)
|x|(n−2)/2 |x|
for all x ∈ Rn \ {0}, belongs to Lloc

1
(Rn , Cn+1 ) ∩ S (Rn , Cn+1 ) and satisfies
Dk Ek = δ in S (Rn , Cn+1 ), (7.10.10)
thus is a fundamental solution for the perturbed Dirac operator Dk in Rn .

Similarly, the Clifford algebra-valued function
cn kn/2 (1) x (1)

E−k (x) := Hn/2 (k|x|) + H(n−2)/2 (k|x|) en+1 , (7.10.11)
|x|(n−2)/2 |x|
for all x ∈ Rn \ {0}, belongs to Lloc

1
(Rn , Cn+1 ) ∩ S (Rn , Cn+1 ) and is a fundamental
solution for the perturbed Dirac operator D−k in Rn .
Proof. Recall Fλ from (7.6.78). Then Φk from (7.6.11)–(7.6.12) may be written in

terms of Fλ corresponding to λ := (n − 2)/2 as
Φk (x) = cn k(n−2)/2 F(n−2)/2 (x), ∀ x ∈ Rn \ {0}. (7.10.12)
Also, the function Ek from (7.10.9) satisfies
Ek (x) = cn kn/2 Fn/2 (x) x − cn kn/2 F(n−2)/2 (x) en+1 , ∀ x ∈ Rn \ {0}. (7.10.13)
Proposition 7.38 (applied with λ = (n−2)/2) ensures that Ek ∈ Lloc

1
(Rn ). In addition,
(7.10.4) and Proposition 7.38 imply
7.10 Fundamental Solutions for the Perturbed Dirac Operator 303

n
−Dk Φk = − ∂ j Φk e j − kΦk en+1
j=1

n
= −cn k(n−2)/2 ∂ j F(n−2)/2 e j − cn kn/2 F(n−2)/2 en+1
j=1

n
= cn kn/2 Fn/2 x j e j − cn kn/2 F(n−2)/2 en+1
j=1
= Ek in D (Rn , Cn+1 ). (7.10.14)
This, (7.10.7), and Theorem 7.33, imply that Dk Ek = δ in D (Rn , Cn+1 ). Hence, Ek
is a fundamental solution for the perturbed Dirac operator Dk in Rn . Also, item (9)
in Lemma 14.71 implies that
'
Ek (x) + cn k(n−3)/2 π2 ei(k|x|−(n−1)π/4) |x|−(n−1)/2 en+1
(7.10.15)
= O(|x|−(n+1)/2 ) as |x| → ∞.
Consequently, there exists some constant C ∈ (0, ∞) such that the coefficients of the
Clifford algebra-valued function Ek (x) are bounded by C|x|−(n−1)/2 for every point
x ∈ Rn \ B(0, 1). Hence, Example 4.4 applies (condition (4.1.4) is satisfied for any
m > (n + 1)/2) and gives that Ek belongs to S (Rn , Cn+1 ).
1
The same reasoning also gives that E−k belongs to Lloc (Rn , Cn+1 ) ∩ S (Rn , Cn+1 )

and that −D−k Φk = E−k in D (R , Cn+1 ), which ultimately, in concert with (7.10.7)
n
and Theorem 7.33, implies that E−k is a fundamental solution for D−k in Rn .
Recall that the complex number bn,k has been defined in (7.6.48).
Exercise 7.58. Pick k ∈ (0, ∞) and n ∈ N, n ≥ 2. Prove that

x + ken+1 + O |x|−(n+1)/2
(Dk Φk )(x − y) = Φk (x − y) ik
eik|x| e−iky,x
= bn,k x + en+1 + O |x|−(n+1)/2
k i
|x|(n−1)/2
as |x| → ∞, (7.10.16)
uniformly for y in compact subsets of Rn , and

(Dk Φk )(x − y) = Φk (x − y) − iky + ken+1 + O |y|−(n+1)/2
eik|y| e−ikx,y
= bn,k y + en+1 + O |y|−(n+1)/2
k − i
|y| (n−1)/2
as |y| → ∞, (7.10.17)

7.11 Fundamental Solutions for the Iterated Perturbed Dirac

Operator
Suppose n ∈ N is such that n ≥ 2 and let k ∈ (0, ∞). In this section we focus on
determining a fundamental solution for the iterated Dirac operator DkN in Rn , where
N ∈ N. In Section 7.10 we have treated the case N = 1; cf. Theorem 7.57. As seen
in (7.10.7), the identity D2k = −(Δ + k2 ) holds in D (Ω, Cn+1 ). Since in Section 7.7
we have determined a fundamental solution Φ(N) k for the iterated Helmholtz operator
(Δ + k2 )N (cf. Theorem 7.40), we now can obtain a fundamental solution for DkN as
follows.
Case I: N is even. Since DkN = (−1)N/2 (Δ + k2 )N/2 in D (Rn , Cn+1 ) and Φ(N/2)k
is a fundamental solution for (Δ + k2 )N/2 in Rn , it follows that (−1)N/2 Φ(N/2)
k is a
fundamental solution for DkN in Rn .
Case II: N is odd. We write
(−1)(N+1)/2 (Δ + k2 )(N+1)/2 = DkN Dk in D (Ω, Cn+1 ), (7.11.1)
which in light of Remark 5.7 gives that (−1)(N+1)/2 Dk Φ((N+1)/2) k , computed in the
sense of distributions, is a fundamental solution for DkN in Rn . To determine an
explicit expression for this fundamental solution, based on (7.7.9), (7.10.4), and an
application of Proposition 7.38 with λ = (n − N − 1)/2 < (n − 1)/2, we obtain
⎡ (1) ⎤
(−1)(N−1)/2 cn k(n−N−1)/2 ⎢⎢⎢ H(n−N−1)/2 (k|x|) ⎥⎥⎥
Dk Φk ((N+1)/2)
(x) = Dk ⎢⎢⎣⎢ ⎥⎥⎥
2(N−1)/2 N−12 !
|x|(n−N−1)/2 ⎦
(−1)(N−1)/2 cn k(n−N−1)/2 H(n−N−1)/2 (k|x|)

n (1)
= ∂ j ej
2(N−1)/2 N−12 ! j=1
|x|(n−N−1)/2
(1)
(−1)(N−1)/2 cn k(n−N+1)/2 H(n−N−1)/2 (k|x|)
+ · en+1
2(N−1)/2 N−12 !
|x|(n−N−1)/2
(1)
(−1)(N+1)/2 cn k(n−N+1)/2 H(n−N+1)/2 (k|x|)
= · xj ej
2(N−1)/2 N−12 !
|x|(n−N+1)/2
(1)
(−1)(N−1)/2 cn k(n−N+1)/2 H(n−N−1)/2 (k|x|)
+ · en+1
2(N−1)/2 N−12 !
|x|(n−N−1)/2
(−1)(N+1)/2 cn k(n−N+1)/2
= (n−N−1)/2 ×
2(N−1)/2 N−1
2 !|x|
(1) x (1)

× H(n−N+1)/2 (k|x|) − H(n−N−1)/2 (k|x|) en+1 (7.11.2)
|x|
7.12 Fundamental Solutions for General Second-Order Operators 305
in D (Rn , Cn+1 ), where cn is as in (7.6.12). In addition, Proposition 7.38 also gives

that the resulting expression in the variable x ∈ Rn \ {0} defines a function in
1
Lloc (Rn , Cn+1 ) which is a Clifford algebra-valued tempered distribution in Rn .
The above analysis combined with the definition of cn from (7.6.12) justifies the
following result.
Theorem 7.59. Let n ∈ N, n ≥ 2, and let k ∈ (0, ∞). Recall the functions Φ(N) k ,
N ∈ N, from Theorem 7.40 which belong to Lloc 1
(Rn ) ∩ S (Rn ). For each N ∈ N
define, with derivatives taken in the sense of distributions in Rn ,
⎧
⎪
⎪ N/2 (N/2)
(N) ⎨(−1) Φk
⎪ if N even,
Θk := ⎪ ⎪ (7.11.3)
⎪
⎩(−1)(N+1)/2 Dk Φ((N+1)/2) if N odd.
k

Then Θ(N)k belongs to Lloc (R , Cn+1 )∩S (R , Cn+1 ) and is a fundamental solution
1 n n
for the iterated perturbed Dirac operator Dk in Rn . Moreover,

N
ik(n−N)/2
Θ(N)
k (x) = H (1) (k|x|),
2(n+N)/2 2 ! π(n−2)/2 |x|(n−N)/2 (n−N)/2
N−2
(7.11.4)
for all x ∈ R \ {0}, if N is even,
n
and
−ik(n−N+1)/2
Θ(N)
k (x) = (n−2)/2 (n−N−1)/2 ×
2 !π
2(n+N+1)/2 N−1 |x|
(1) x (1)
(7.11.5)
× H(n−N+1)/2 (k|x|) − H(n−N−1)/2 (k|x|) en+1
|x|
for all x ∈ Rn \ {0}, if N is odd.
7.12 Fundamental Solutions for General Second-Order

Operators
Consider a constant, complex coefficient, homogeneous, second-order differential

n
operator L = a jk ∂ j ∂k in Rn . In a first stage, our goal is to find necessary and
j,k=1
sufficient conditions, that can be expressed without reference to the theory of dis-
tributions, guaranteeing that a function E ∈ Lloc1
(Rn ) is a fundamental solution for
L. Several necessary conditions readily present themselves. First, it is clear that LE
is the zero distribution in Rn \ {0}. In addition, if L is elliptic then this necessarily
implies that E ∈ C ∞ (Rn \ {0}). In the absence of ellipticity as a hypothesis for L,
we may wish to assume that E is reasonably regular, say E ∈ C 2 (Rn \ {0}). Second,
if the fundamental solution E is a priori known to be a tempered distribution then
= 1 in S (Rn ). If L is elliptic and n ≥ 3, this forces E = E0 + P where
−L(ξ)E(ξ)
−1 −1
E0 := −F L(ξ) is homogeneous of degree 2 − n and P is a polynomial that is
annihilated by L. Hence, working with E0 in place of E, there is no loss of generality
in assuming that E is homogeneous of degree 2 − n. The case n = 2 may also be
included in this discussion by demanding that ∇E is homogeneous of degree 1 − n.
In summary, it is reasonable to restrict our search for a fundamental solution for
L in the class of functions satisfying E ∈ C 2 (Rn \ {0}) ∩ Lloc 1
(Rn ) with the prop-
erty that ∇E is positive homogeneous of degree 1 − n in R \ {0}. However, these
n
conditions do not rule out such trivial candidates as the zero distribution. In Theo-
rem 7.60 we identify the key nondegeneracy property (7.12.2) guaranteeing that E
is in fact a fundamental solution. Theorem 7.60 is then later used to find an explicit
formula for such a fundamental solution, under a strong ellipticity assumption on L
(cf. Theorem 7.68).
Theorem 7.60. Assume that n ≥ 2 and consider

n
L= a jk ∂ j ∂k , a jk ∈ C. (7.12.1)
j,k=1
Then for a function E ∈ C 2 (Rn \ {0}) ∩ Lloc

1
(Rn ) with the property that ∇E is positive
homogeneous of degree 1 − n in R \ {0}, the following statements are equivalent:
n
1
(1) When viewed in Lloc (Rn ), the function E is a fundamental solution for L in Rn ;
(2) One has LE = 0 pointwise in Rn \ {0} and
n

a jk ω j ∂k E(ω) dσ(ω) = 1. (7.12.2)
j,k=1 S n−1
Remark 7.61. (i) In the partial differential equation parlance, the integrand in

n
(7.12.2), i.e., the expression a jk ω j ∂k E(ω), is referred to as the conormal deriva-
j,k=1
tive of E on S n−1 .
(ii) One remarkable aspect of Theorem 7.60 is that the description of a funda-
mental solution from part (2) is purely in terms of ordinary calculus (i.e., without
any reference to the theory of distributions).
Proof of Theorem 7.60. Let E ∈ C 2 (Rn \ {0}) ∩ Lloc 1

(Rn ) with the property that ∇E
is positive homogeneous of degree 1 − n in R \ {0}. Exercise 4.53 then implies
n
that ∇E ∈ Lloc
1
(Rn ). Fix an arbitrary f ∈ C0∞ (Rn ). Then making use of (4.4.19) and
Proposition 4.70 (applied to each ∂k E) we obtain

n

(LE) ∗ f = a jk ∂ j (∂k E) ∗ f
j,k=1

n

n
= a jk ∂k E(ω)ω j dσ(ω) (δ ∗ f ) + a jk P.V.(∂ j ∂k E) ∗ f
j,k=1 S n−1 j,k=1

n
= a jk ∂k E(ω)ω j dσ(ω) f (7.12.3)
S n−1 j,k=1

n
+ lim+ a jk (∂ j ∂k E)(· − y) f (y) dy in D (Rn ),
ε→0 |y−· |>ε j,k=1
where we have also used the fact that δ ∗ f = f . Having proved this, we now turn in
earnest to the proof of the equivalence in the statement of the theorem.
First, assume that E is a fundamental solution for L in Rn . Then LE = δ in

D (Rn ) implies that L E Rn \{0} = 0 in D (Rn ). Since by assumption E Rn \{0} belongs

to C 2 (Rn \ {0}), we arrive at the conclusion that LE = 0 pointwise in Rn \ {0}.

Explicitly,
n
a jk ∂ j ∂k E(x) = 0, ∀ x ∈ Rn \ {0}. (7.12.4)
j,k=1
This proves the first claim in (2). Next, for each function f ∈ C0∞ (Rn ) we may write
f = δ ∗ f = (LE) ∗ f in D (Rn ) which, in light of (7.12.3) and (7.12.4), forces

n
f = a jk ∂k E(ω)ω j dσ(ω) f. (7.12.5)
S n−1 j,k=1
Since f ∈ C0∞ (Rn ) was arbitrary, (7.12.2) follows. This finishes the proof of (1) ⇒
(2).
Conversely, suppose that E ∈ C 2 (Rn \ {0}) ∩ Lloc
1
(Rn ) with the property that ∇E
is positive homogeneous of degree 1 − n in R \ {0}, such that LE = 0 pointwise
n
in Rn \ {0}, and (7.12.2) holds. Then for every f ∈ C0∞ (Rn ) formula (7.12.3) simply
reduces to (LE)∗ f = f . Now Exercise 2.97 may be invoked to conclude that LE = δ
in D (Rn ), as wanted.
Next, we turn to the task of finding all fundamental solutions that are tempered
distributions for general homogeneous, second-order, constant coefficient operators
that are strongly elliptic. We begin by defining this stronger (than originally intro-
duced in Definition 6.13) notion of ellipticity.

Let A = a jk 1≤ j,k≤n ∈ Mn×n (C) and associated to such a matrix A, consider the
operator
n
LA := LA (∂) := a jk ∂ j ∂k . (7.12.6)
j,k=1
This is a homogeneous, second-order, constant coefficient operator for which the

ellipticity condition reads

n
LA (ξ) := a jk ξ j ξk 0, ∀ ξ = (ξ1 , . . . , ξn ) ∈ Rn \ {0}. (7.12.7)
j,k=1
As a trivial consequence of the Malgrange–Ehrenpreis theorem (cf. Theorem 5.10),

any elliptic operator has a fundamental solution. The goal is to obtain explicit formu-
las for such fundamental solutions for a subclass of homogeneous, elliptic, second-
order, constant coefficient operators satisfying a stronger condition than (7.12.7).
Definition 7.62. Call an operator LA as in (7.12.6) strongly elliptic, if there

exists a constant C ∈ (0, ∞) such that

n
Re a jk ξ j ξk ≥ C|ξ|2 , ∀ ξ = (ξ1 , . . . , ξn ) ∈ Rn . (7.12.8)
j,k=1

By extension, call a matrix A = a jk 1≤ j,k≤n ∈ Mn×n (C) strongly elliptic
provided there exists some C ∈ (0, ∞) with the property that (7.12.8) holds.
Remark 7.63.
(1) It is obvious that any operator LA as in (7.12.6) that is strongly elliptic is
elliptic.
(2) Up to changing L to −L, any elliptic, homogeneous, second- order, constant
coefficient differential operator L with real coefficients is strongly elliptic. To see
why this is the case let A ∈ Mn×n (R) and suppose that LA is elliptic. Consider

n
the function f : S n−1 → R defined by f (ξ) := a jk ξ j ξk for ξ ∈ S n−1 . Then f
j,k=1
is continuous and since LA is elliptic the number 0 is not in the image of f . The
unit sphere S n−1 being compact and connected, it is mapped by f in a compact,
connected, subset of R, not containing 0. This forces the image of f to be a compact
interval that does not contain 0. Hence, there exists c ∈ (0, ∞) with the property that
either f (ξ) ≥ c for every ξ ∈ S n−1 or − f (ξ) ≥ c for every ξ ∈ S n−1 . This implies that
either f (ξ/|ξ|) ≥ c for every ξ ∈ Rn \ {0} or − f (ξ/|ξ|) ≥ c for every ξ ∈ Rn \ {0}, or
equivalently, that either LA or −LA is strongly elliptic.

10
(3) Consider the operator L = ∂1 + i∂2 in R . If we take A :=
2 2 2
then L = LA
0 i
is a homogeneous, second-order, constant complex coefficient differential operator,
and L(ξ) = ξ12 + iξ22 , ξ = (ξ1 , ξ2 ) ∈ R2 . Clearly L(ξ) 0 if ξ 0, so L is elliptic.
However, L is not strongly elliptic since Re [L(ξ)] = ξ12 which cannot be bounded
from below by a constant multiple of |ξ|2 since the latter blows up if |ξ2 | → ∞.

Fix A = a jk 1≤ j,k≤n ∈ Mn×n (C) and consider the operator LA as in (7.12.6).
Due to the symmetry of mixed partial derivatives in the sense of distributions, it is
immediate that
A + A
LA = LAsym , where A sym := . (7.12.9)
2
As such, any fundamental solution for LAsym is also a fundamental solution for LA .
Also, since (A sym ξ) · ξ = (Aξ) · ξ for every ξ ∈ Rn , we have that
LA is strongly elliptic if and only if LAsym is strongly elliptic. (7.12.10)
Consequently,
when computing the fundamental solution for LA

we may assume without loss of generality that A (7.12.11)
is symmetric, i.e., A = A .
For further reference we summarize a few basic properties of symmetric matrices

(throughout, the symbol “·” denotes the real inner product of vectors with complex
components).
A ∈ Mn×n (R), A = A =⇒ (Aζ) · ζ ∈ R, ∀ ζ ∈ Cn , (7.12.12)
A ∈ Mn×n (C), A = A =⇒ Re A = (Re A) and Im A = (Im A) , (7.12.13)

A ∈ Mn×n (C), A = A =⇒ Re (Aζ) · ζ = (Re A)ζ · ζ, ∀ ζ ∈ Cn . (7.12.14)
It is easy to see that (7.12.12)–(7.12.13) hold, while (7.12.14) follows from (7.12.12)–
(7.12.13) after writing A = Re A + i Im A.
Also, recall that a matrix A ∈ Mn×n (C) is said to be positive definite
provided
(Aζ) · ζ is real and strictly positive for each ζ ∈ Cn \ {0}. (7.12.15)

It is easy to see that any positive-definite matrix A ∈ Mn×n (C) satisfies A = A,
and there exists c ∈ (0, ∞) such that
(Aζ) · ζ ≥ c|ζ|2 , ∀ ζ ∈ Cn . (7.12.16)
Remark 7.64. Fix A ∈ Mn×n (C) that is symmetric and satisfies (7.12.8). Then, for
each ζ ∈ Cn we have (with C ∈ (0, ∞) as in (7.12.8))

Re [(Aζ) · ζ ] = Re A(Re ζ + i Im ζ) · (Re ζ − i Im ζ)

= Re (ARe ζ) · Re ζ + (A Im ζ) · Im ζ
≥ C|Re ζ|2 + C|Im ζ|2 = C|ζ|2 . (7.12.17)
The second equality in (7.12.17) uses the fact that A is symmetric, while (7.12.8)
is used for the inequality in (7.12.17). Thus, combining (7.12.17) with the Cauchy–
Schwarz inequality, we obtain
|Aζ| ≥ C|ζ| for every ζ ∈ Cn , (7.12.18)

proving that the linear map A : Cn → Cn is injective, thus invertible. In particular,

detA 0. Also, thanks to (7.12.14) and (7.12.17) we have that Re A is a positive-
definite matrix. More precisely, with C as in (7.12.8), we have

(Re A)ζ · ζ ≥ C|ζ|2 ∀ ζ ∈ Cn . (7.12.19)
From Remark 7.64 and definitions we see that, given any A ∈ Mn×n (C), the
following implications hold:
Re A positive definite =⇒ A strongly elliptic, (7.12.20)
A strongly elliptic ⇐⇒ A sym strongly elliptic, (7.12.21)
A strongly elliptic and symmetric =⇒ Re A positive definite. (7.12.22)
Remark 7.65. Assume that A ∈ Mn×n (C) is symmetric and satisfies (7.12.8). From
Remark 7.64 it follows that A is invertible. Moreover, if we define

A := sup |Aζ| : ζ ∈ Cn , |ζ| = 1 , (7.12.23)
then (7.12.18) ensures that A > 0. We claim that

C
Re (A−1 ζ) · ζ ≥ |ζ|2 ∀ ζ ∈ Cn , (7.12.24)
A2
where C is as in (7.12.8). To justify this, first note that
|ζ|2 = |AA−1 ζ|2 ≤ A2 |A−1 ζ|2 for each ζ ∈ Cn . (7.12.25)
In turn, (7.12.25) and (7.12.17) permit us to estimate

Re [(A−1 ζ) · ζ ] =Re [ (A−1 ζ) · ζ] = Re (A−1 ζ) · (AA−1 ζ) ≥ C|A−1 ζ|2
C
≥ |ζ|2 ∀ ζ ∈ Cn . (7.12.26)
A2
This proves (7.12.24). In particular, (7.12.24) yields
C
A−1 |ξ|2 ≥ |(A−1 ξ) · ξ| ≥ |ξ|2 ∀ ξ ∈ Rn . (7.12.27)
A2
Remark 7.66. Consider the set

M := A ∈ Mn×n (C) : A = A , Re A is positive definite . (7.12.28)
Since the n×n symmetric matrices A = (a jk )1≤ j,k≤n with complex entries are uniquely
determined by the elements a jk with 1 ≤ j ≤ k ≤ n, we may naturally identify M
with an open convex subset of Cn(n+1)/2 . Throughout, this identification is implicitly
assumed. Moreover, every A ∈ M satisfies det A 0, since if Aζ = 0 for some
ζ ∈ Cn , then (7.12.14) and the fact that Re A is positive definite force ζ = 0. The
fact that M is convex, implies that there is a unique analytic branch of the mapping
1 1 1
M A → (detA) 2 ∈ C such that (detA) 2 > 0 when A is real. Thus (detA) 2 is
unambiguously defined for A ∈ M.
To proceed with the discussion regarding determining a fundamental solution for

a strongly elliptic operator LA we first analyze the case when A has real entries.
The Case When A is Real, Symmetric and Satisfies (7.12.8).
Since A is real, symmetric, and positive definite, A is diagonalizable and has the
form A = U −1 DU for some orthogonal matrix U ∈ Mn×n (R) and some diagonal
n × n matrix D whose entries on the main diagonal are strictly positive real numbers.
1
Hence, D 2 is meaningfully defined as the n × n diagonal matrix with the entries on
the diagonal being equal to the square roots of the entries on the main diagonal in
D. In addition, A√2 := U −1 D 2 U is well-defined, symmetric, invertible, A 2 A 2 = A
1 1 1 1
1
and det(A 2 ) = detA. The idea now is to apply Exercise 7.77 with A1 := A and
B := A 2 . In this vein, observe that A2 = A− 2 A(A− 2 ) = A− 2 AA− 2 = I. Thus,
1 1 1 1 1
LA2 = Δ and formula (7.14.3) becomes
(LA u) ◦ A 2 = Δ(u ◦ A 2 ) for every u ∈ D (Rn ).

1 1
(7.12.29)
Our goal is to find E A ∈ D (Rn ) such that LA E A = δ in D (Rn ). The latter equality
is equivalent with (LA E A ) ◦ A 2 = δ ◦ A 2 in D (Rn ). Using (2.2.7) and the definition
1 1
of δ we see that
1 1
δ, ϕ ◦ A− 2 = √
1 1
δ ◦ A 2 , ϕ = √ ϕ(0)
detA detA
1
= √ δ, ϕ , ∀ ϕ ∈ C0∞ (Rn ). (7.12.30)
detA
Hence, in light of (7.12.29) and (7.12.30), it suffices to find E A ∈ D (Rn ) such that
1
in D (Rn ).
1
Δ(E A ◦ A 2 ) = √ δ (7.12.31)
detA
In particular, if we now choose
1 1
E A (y) := √ EΔ ◦ A− 2 (y) for y ∈ Rn \ {0}, (7.12.32)
detA
where EΔ is the fundamental solution for the Laplacian from (7.1.12), then E A sat-
isfies (7.12.31), hence is a fundamental solution for LA in Rn . The additional prop-
erties of EΔ , (7.12.27), and Exercise 4.6, imply that E A ∈ Lloc 1
(Rn ) and E A is a
tempered distribution in R . Keeping in mind that
n
1 1
|A− 2 x|2 = A− 2 x · A− 2 x = A−1 x · x,
1
∀ x ∈ Rn , (7.12.33)
formulas (7.1.12) and (7.12.32) give

⎧ −1
⎪
⎪
⎪
1
⎪
⎪ √ · if n ≥ 3,
⎨ (n − 2)ωn−1 detA [(A−1 x) · x] n−2
E A (x) = ⎪
⎪
2
(7.12.34)
⎪
⎩ √1 ln (A−1 x) · x
⎪
⎪ if n = 2,
4π detA
for every x ∈ Rn \ {0}. This completes the discussion for determining a fundamental
solution for LA in the case when A is real, symmetric, and satisfies (7.12.8).
In preparation to dealing with the case of matrices with complex entries, we state
and prove the following useful complex analysis result.
Lemma 7.67. Let N ∈ N and assume O is an open and convex subset of CN with
the property that O ∩ RN ∅ (where RN is canonically embedded into CN ). Also,
suppose f, g : O → C are two functions which are separately holomorphic (i.e., in
each scalar complex component in CN ) such that

f O∩RN = gO∩RN . (7.12.35)
Then f = g in O.
Proof. Fix an arbitrary point (x1 , . . . , xN ) ∈ O ∩ RN and consider

O1 := z1 ∈ C : (z1 , x2 , . . . , xN ) ∈ O . (7.12.36)
Then O1 is an open convex subset of C, which contains x1 , hence O1 ∩ R ∅.

Define the functions f1 , g1 : O1 → C by
f1 (z) := f (z, x2 , . . . , xN ) and g1 (z) := g(z, x2 , . . . , xN ), ∀ z ∈ O1 . (7.12.37)
Then f1 and g1 are holomorphic functions in O1 which coincide on O1 ∩ R. Since

the latter contains an accumulation point in the convex (hence connected) set O1 , it
follows that f1 = g1 on O1 by the coincidence theorem for holomorphic functions of
one complex variable. Since (x1 , . . . , xN ) ∈ O ∩ RN was arbitrary, we may conclude
that
f O∩(C×RN−1 ) = gO∩(C×RN−1 ) . (7.12.38)
Next, fix (z1 , x2 , . . . , xN ) ∈ O ∩ (C × RN−1 ) and define

O2 := z2 ∈ C : (z1 , z2 , x3 , . . . , xN ) ∈ O . (7.12.39)
Once again, O2 is an open convex subset of C, which contains x2 , hence O2 ∩R ∅.

If we now define the functions f2 , g2 : O2 → C by
f2 (z) := f (z1 , z, x3 , . . . , xN ) and g2 (z) := g(z1 , z, x3 , . . . , xN )

(7.12.40)
for each point z belonging to the set O2 ,
it follows that f2 , g2 are holomorphic in O2 which, by (7.12.38), coincide on

O2 ∩ R. Given that the latter set contains an accumulation point in the convex
set O1 , we deduce that f2 = g2 on O2 by once again invoking the coincidence
theorem for holomorphic functions of one complex variable. Upon recalling that
(z1 , x2 , . . . , xN ) ∈ O ∩ (C × RN−1 ) was arbitrary, we conclude that

f O∩(C2 ×RN−2 ) = gO∩(C2 ×RN−2 ) . (7.12.41)
Continuing this process inductively, we arrive at the conclusion that f = g in O.

After this preamble, we are ready to consider the general case.
The Case When A Has Complex Entries, is Symmetric and Satisfies (7.12.8).
As observed in Remark 7.64, under the current assumptions, A continues to be
invertible. Also, (7.12.27) holds. In addition, under the current assumptions we have
1
that A ∈ M and (detA) 2 is unambiguously defined (see in Remark 7.66).
These comments show that the function E A from (7.12.34) continues to be well
defined under the current assumption on A if ln is replaced by the principal branch
of the complex log (defined for points z ∈ C \ (−∞, 0] so that za = ea log z for each
a ∈ R). In addition, E A continues to belong to Lloc 1
(Rn ) and E A is of class C ∞ in
R \ {0}. Furthermore, from (7.12.27) and Exercise 4.6 it follows that the function
n
E A continues to be a tempered distribution in Rn . The goal is to prove that this

expression is a fundamental solution for LA in the current case.
First, observe that since A−1 is symmetric, for each j, k ∈ {1, . . . , n} we have

∂k A−1 x · x = 2(A−1 x)k and ∂ j A−1 x k = (A−1 )k j . (7.12.42)
Hence, for every x ∈ Rn \ {0}, differentiating pointwise we obtain
2−n
n
−n
LA A−1 x · x 2 = a jk ∂ j (2 − n) A−1 x · x 2 (A−1 x)k
j,k=1

n
−n−2
= −n(2 − n) a jk A−1 x · x 2 (A−1 x) j (A−1 x)k
j,k=1

n
−n
+ (2 − n) a jk A−1 x · x 2 (A−1 )k j = 0. (7.12.43)
j,k=1

Similarly, LA log (A−1 x) · x = 0 for x ∈ R2 \ {0}. Thus, we may conclude that
LA E A (x) = 0 ∀ x ∈ Rn \ {0}. (7.12.44)
Second, by making use of (7.12.42) and the expression for E A we obtain
1 A−1 x
∇E A (x) = √ · −1 n ∀ x ∈ Rn \ {0} (7.12.45)
ωn−1 detA [(A x) · x] 2
which, in particular, shows that ∇E A is positive homogeneous of degree 1 − n in

Rn \ {0}. Furthermore, for each x ∈ S n−1 we have

n

n a jk x j (A−1 x)k
1 j,k=1
a jk x j ∂k E A (x) = √ · n
j,k=1 ωn−1 detA [(A−1 x) · x] 2
1 (A x) · (A−1 x)
= √ · n
ωn−1 detA [(A−1 x) · x] 2
1 1
= √ · −1 n , (7.12.46)
ωn−1 detA [(A x) · x] 2
where the last equality uses the fact that |x| = 1.
Invoking Theorem 7.60 we conclude that E A is a fundamental solution for LA in
Rn if and only if
√ dσ(x)
ωn−1 detA = −1 n . (7.12.47)
S n−1 [(A x) · x] 2
The fact that we already know that E A is a fundamental solution for LA in Rn in

the case when A ∈ Mn×n (R) satisfies A = A and condition (7.12.8), implies that
formula (7.12.47) holds for this class of matrices.
We make the claim that in fact (7.12.47) actually holds for the larger class of
matrices A ∈ Mn×n (C) satisfying A = A and condition (7.12.8). To see why this is
true, recall the open subset M of Cn(n+1)/2 from (7.12.28) and consider the functions

f, g : M → C defined for every A = a jk 1≤ j,k≤n ∈ M by
√
f (a jk )1≤ j≤k≤n := ωn−1 detA, (7.12.48)

dσ(x)
g (a jk )1≤ j≤k≤n := n . (7.12.49)
n−1 −1
[(A x) · x] 2
S
Then f and g are analytic (as functions of several complex variables) on M, which
is an open convex set in Cn(n+1)/2 . If A ∈ M has real entries, then A satisfies (7.12.8),
and (7.12.47) holds for such A, hence f = g on M∩Mn×n (R). Invoking Lemma 7.67
we may therefore conclude that f = g on M. Thus, (7.12.47) holds for every A ∈
Mn×n (C) satisfying A = A and condition (7.12.8).
Finally, we note that thanks to Proposition 5.8 and the current strong ellipticity
assumption, any other fundamental solution of LA belonging to S (Rn ) differs from
E A by a polynomial that LA annihilates.
In summary, the above analysis proves the following result.

Theorem 7.68. Suppose A = a jk 1≤ j,k≤n ∈ Mn×n (C) and consider the operator LA
associated to A as in (7.12.6). If LA is strongly elliptic, then the function defined by
⎧
⎪
⎪
⎪
1 1
⎪
⎪ − 6 · if n ≥ 3,
⎪
⎪
⎨ (n − 2)ωn−1 detA sym [((A sym ) x) · x] 2
n−2
−1
⎪
E A (x) := ⎪
⎪ (7.12.50)
⎪
⎪
⎪ 1
⎪
⎪ log ((A sym )−1 x) · x if n = 2,
⎩ 4π 6detA
⎪
sym
for x ∈ Rn \ {0} belongs to Lloc

1
(Rn ) ∩ S (Rn ) ∩ C ∞ (Rn \ {0}) and is a fundamental
6
solution for LA in R . Above, A sym := 12 (A + A ), detA sym is defined as in Re-
n
mark 7.66, and log denotes the principal branch of the complex logarithm (defined
for complex numbers z ∈ C \ (−∞, 0] so that za = ea log z for each a ∈ R). Moreover,

u ∈ S (Rn ) : LA u = δ in S (Rn ) (7.12.51)
n
= E A + P : P polynomial such that LA P = 0 in R .
We conclude this section with a couple of related exercises about fundamental

solutions for second-order, constant coefficient, differential operators.

n
Exercise 7.69. Let n ≥ 2 and consider a differential operator L = a jk ∂ j ∂k with
j,k=1
complex coefficients. Assume that E ∈ C 2 (Rn \ {0}) ∩ Lloc
1
(Rn ) is a function with the
property that ∇E is positive homogeneous of degree 1 − n in Rn \ {0}. In addition,
suppose that
n
λ := a jk ω j ∂k E(ω) dσ(ω) 0. (7.12.52)

j,k=1 S n−1
Prove that λ−1 E is a fundamental solution for L in Rn . Use this result to find the
proper normalization for the standard fundamental solution for the Laplacian in Rn ,
starting with E(x) := |x|2−n when n ≥ 3, and with E(x) := ln |x| for n = 2.
Exercise 7.70. Let n ≥ 2 and suppose that E ∈ C 2 (Rn \ {0}) ∩ Lloc 1

(Rn ) has the
property that ∇E is odd and positive homogeneous of degree 1 − n in Rn \ {0}.
In addition, assume that the function E is a fundamental solution for the complex

n
constant coefficient differential operator L = a jk ∂ j ∂k in Rn . Prove that for every
j,k=1
ξ ∈ S n−1 one has

n
1
a jk ω j ∂k E(ω) dσ(ω) = . (7.12.53)
j,k=1
2
ω∈S n−1
ω·ξ>0

n
Hint: Let f (ξ) := a jk ω j ∂k E(ω) dσ(ω) for each ξ ∈ S n−1 . Show that f is
j,k=1
ω∈S n−1
ω·ξ>0
even and make use of Theorem 7.60.
7.13 Layer Potential Representation Formulas Revisited
The goal here is to derive a layer potential representation formula generalizing the
identity from Proposition 7.17 for the Laplacian. We begin by describing the setting
in which we intend to work.

Given a matrix A = a jk 1≤ j,k≤n ∈ Mn×n (C), we associate the homogeneous
second-order differential operator

n
LA = a jk ∂ j ∂k in Rn , (7.13.1)
j,k=1
and for every unit vector ν = (ν1 , . . . , νn ) and any complex-valued function u of class
C 1 , define the conormal derivative of u associated with the matrix A (along ν) as

n
∂νA u := ν j a jk ∂k u. (7.13.2)
j,k=1
Theorem 7.71. Suppose n ≥ 2 and let Ω ⊂ Rn be a bounded domain of class C 1 ,

with outward unit normal ν and surface measure σ. In addition, assume that the

matrix A = a jk 1≤ j,k≤n ∈ Mn×n (C) is such that the operator LA associated with A as
in (7.13.1) is strongly elliptic, and recall the fundamental solution E A for LA defined
in (7.12.50).
Then for every complex-valued function u ∈ C 2 (Ω ) one has
(LA u)(y)E A (x − y) dy − E A (x − y)(∂νA u)(y) dσ(y)

Ω ∂Ω

⎧

⎪
⎨ u(x), x ∈ Ω,
⎪
− u(y)(∂νA E)(x − y) dσ(y) = ⎪
⎪ (7.13.3)
∂Ω ⎩ 0, x ∈ Rn \ Ω,
where is the conormal derivative associated with the matrix A (along ν).
Proof. When x ∈ Rn \ Ω, it clear from (7.12.50) that E A (x − ·) ∈ C ∞ (Ω ). Also, since

E A is a fundamental solution for LA in Rn we have that LA [E A (x − ·)] = 0 in Rn \ {x},
hence (LA E A )(x − ·) = 0 in Ω. Based on these and repeated use of (14.8.4) we may
then write
7.13 Layer Potential Representation Formulas Revisited 317

n

(LA u)(y)E A (x − y) dy = a jk (∂ j ∂k u)(y)E A (x − y) dy
Ω j,k=1 Ω

= a jk (∂k u)(y)(∂ j E A )(x − y) dy
j,k=1 Ω

+ a jk ν j (y)(∂k u)(y)E A (x − y) dσ(y)
j,k=1 ∂Ω

= u(y)(LA E A )(x − y) dy + u(y)(∂νA E A )(x − y) dσ(y)
Ω ∂Ω

+ E A (x − y)(∂νA u)(y) dσ(y). (7.13.4)

∂Ω
Upon recalling that (LA E A )(x − ·) = 0 in Ω, the last solid integral drops out and the
resulting identity is in agreement with (7.13.3).
Consider now the case when x ∈ Ω. Since Ω is open, there exists r > 0 such that
B(x, r) ⊆ Ω. For each ε ∈ (0, r) define Ωε := Ω \ B(x, ε) which is a bounded domain
of class C 1 . Since E A (x − ·) ∈ C ∞ (Ωε ) and (LA E A )(x − ·) = 0 in Ωε , the same type
of reasoning as above gives (keeping in mind that ∂Ωε = ∂Ω ∪ ∂B(x, ε))

(LA u)(y)E A (x − y) dy = u(y)(∂νA E A )(x − y) dσ(y)
Ωε ∂Ω

− u(y)(∂νA E A )(x − y) dσ(y) + E A (x − y)(∂νA u)(y) dσ(y)
∂B(x,ε) ∂Ω

− E A (x − y)(∂νA u)(y) dσ(y) =: I + II + III + IV. (7.13.5)

∂B(x,ε)

As seen from (7.12.50), we have |IV| ≤ C A ∇uL∞ (Ω) ε max 1, ln |ε| , from which
we deduce that lim+ IV = 0. Next, split II = II + u(x)II where
ε→0

II := − [u(y) − u(x)](∂νA E A )(x − y) dσ(y), (7.13.6)
∂B(x,ε)
and observe that


II := − (∂νA E A )(x − y) dσ(y)
∂B(x,ε)

n
yk − xk
=− a jk (∂ j E A )(x − y) dσ(y)
∂B(x,ε) j,k=1 ε

n
= a jk ωk (∂ j E A )(ω) dσ(ω)
S n−1 j,k=1

n
= (A ) jk ω j (∂k E A )(ω) dσ(ω) = 1. (7.13.7)
S n−1 j,k=1
Above, the first equality defines II , the second equality uses the definition of the
conormal derivative and the outward unit normal to the ball, the third equality is
based on the change of variables ω = (x − y)/ε and the fact that ∇E A is positive
homogeneous of degree 1 − n. Finally, in the fourth equality we have interchanged
j and k in the summation and used the identities E A = E A , LA = LA , while the last
equality is due to (7.12.2).
In addition, |II | ≤ C A ∇uL∞ (Ω) ε, hence lim+ II = 0, and
ε→0

lim+ (LA u)(y)E A (x − y) dy = (LA u)(y)E A (x − y) dy

ε→0 Ωε Ω
by Lebesgue’s Dominated Convergence Theorem (recall that E A (x−·) ∈ L1 (Ω) since

Ω is bounded). Collectively, the results deduced in the above analysis yield (7.13.3)
in the case when x ∈ Ω, finishing the proof of the theorem.
Further Notes for Chapter 7. As evidenced by the treatment of the Poisson problem for the
Laplacian and bi-Laplacian (from Section 7.2 and Section 7.4, respectively), fundamental solutions
play a key role both for establishing integral representation formulas and for deriving estimates for
the solution. This type of application to partial differential equations amply substantiate the utility
of the tools from distribution theory and harmonic analysis derived in Section 4.10 (dealing with
derivatives of volume potentials) and Section 4.9 (dealing with singular integral operators). The
aforementioned Poisson problems serve as a prototype for other types of boundary value problems
formulated for other differential operators and with the entire space Rn replaced by an open set
Ω ⊂ Rn . In the latter scenario one specifies boundary conditions on ∂Ω in place of ∞ as in the case
of Rn (note that ∞ plays the role of the topological boundary of Rn regarded as an open subset of
its compactification Rn ∪ {∞}).
In Section 7.9 the Dirac operator has been considered in the natural setting of Clifford algebras.
For more information pertaining to this topic, the interested reader is referred to the monographs
[6], [25], [58]. The last two references also contain a discussion of Hardy spaces in the context of
Clifford algebras (a topic touched upon in Section 7.9). A classical reference to Hardy spaces in
the ordinary context of C (which appeared at the end of Section 7.8) is the book [30].
Exercise 7.72. Prove that there exists a unique E ∈ S (Rn ) such that ΔE − E = δ in
S (Rn ).
Exercise 7.73. Does there exist E ∈ L1 (Rn ) such that ΔE = δ in D (Rn )?
Exercise 7.74. Prove that for every u ∈ D (Ω) we have
div(∇u) = Δu in D (Ω). (7.14.1)
Exercise 7.75. Prove that if f ∈ C ∞ (Ω) and u ∈ D (Ω), then
Δ( f u) = (Δ f )u + 2(∇ f ) · (∇u) + f Δu in D (Ω). (7.14.2)
Exercise 7.76. Let Ω be an open set in R2 and suppose K is a compact set with
its restriction to Ω \ K is holomorphic,
K ⊂ Ω. Prove that if u ∈ C 2 (Ω) is such that
then Δu is absolutely integrable on Ω and Ω Δu dx = 0.
Exercise 7.77. Suppose A1 ∈ Mn×n (C) and B ∈ Mn×n (R) are given and B is invert-
ible. Define the matrix A2 ∈ Mn×n (C) by A2 := B−1 A1 (B−1 ) . Recall (7.12.6). Prove
that
(LA1 u) ◦ B = LA2 (u ◦ B) for every u ∈ D (Rn ). (7.14.3)
Exercise 7.78. Suppose n ≥ 2 and denote by EΔ the fundamental solution for the
Laplacian operator Δ given in (7.1.12). Without making use of Corollary 4.65, prove
that for each j ∈ {1, . . . , n} one has
ξj
F (∂ j EΔ ) = −i in S (Rn ). (7.14.4)
|ξ|2
In turn, use (7.14.4) to show that

xj ξj
F = −iωn−1 2 in S (Rn ), (7.14.5)
|x|n |ξ|
and
−1 ξj i xj
F = · in S (Rn ). (7.14.6)
|ξ|2 ωn−1 |x|n
Exercise 7.79. Suppose n ≥ 3 and denote by EΔ2 the fundamental solution for the
bi-Laplacian operator Δ2 given in (7.3.8). Prove that for each j, k ∈ {1, . . . , n} one
has
ξ j ξk
F (∂ j ∂k EΔ2 ) = − 4 in S (Rn ). (7.14.7)
|ξ|
Consequently,

ξ j ξk 1 δ jk 1 x j xk
F −1 = · n−2 − · n in S (Rn ), (7.14.8)
|ξ| 4 2(n − 2)ωn−1 |x| 2ωn−1 |x|
and
x j xk δ jk ξ j ξk
F = ωn−1 2 − 2ωn−1 4 in S (Rn ). (7.14.9)
|x| n |ξ| |ξ|
Exercise 7.80. Let P(D) be a nonzero linear constant coefficient operator of order
m ∈ N0 . Prove that P(D) is elliptic if and only if there exist C, R ∈ (0, ∞) such that
|P(ξ)| ≥ C|ξ|m for every ξ ∈ Rn \ B(0, R).
Exercise 7.81. Give a second proof to Theorem 7.68 without making any appeal to
Theorem 7.60.
Chapter 8
The Heat Operator and Related Versions
Abstract This chapter has a twofold aim: determine all fundamental solutions that
are tempered distributions for the heat operator and related versions (including the
Schrödinger operator), then use this as a tool in obtaining the solution of the gener-
alized Cauchy problem for the heat operator.
8.1 Fundamental Solutions for the Heat Operator
Throughout this chapter we use the notation (x, t) := (x1 , . . . , xn , t) ∈ Rn+1 . The
n
heat operator1 is then defined as L := ∂t − Δ x = ∂t − ∂2x j . The starting point
j=1
in determining all fundamental solutions for the heat operator L that are tempered
distributions is Theorem 5.14 which guarantees that such fundamental solutions do
exist. As we have seen in the case of the Laplace operator, the Fourier transform
is an important tool in determining explicit expressions for fundamental solutions
that are tempered distributions. We will continue to make use of this tool in the case
of the heat operator with the adjustment that, this time, we work with the partial
Fourier transform F x discussed at the end of Section 4.2.
Let E ∈ S (Rn+1 ) be a fundamental solution for L. Thus, in view of Exercise 2.90
and (4.1.33), we have
∂t E − Δ x E = δ(x) ⊗ δ(t) in S (Rn+1 ). (8.1.1)
Applying F x to (8.1.1), denoting F x (E) by Ex , and using Exercise 4.42, it follows
that
x + |ξ|2 E
∂t E x = 1(ξ) ⊗ δ(t) in S (Rn+1 ). (8.1.2)
In particular, for each fixed ξ ∈ Rn , we have
1
First considered in 1809 for n = 1 by Laplace (cf. [42]) and then for higher dimensions by
Poisson (cf. [62] for n = 2).
322 8 The Heat Operator and Related Versions

∂t + |ξ|2 E x = δ(t) in D (R). (8.1.3)
Applying Example 5.13, we obtain that E x (ξ, t) := H(t)e−|ξ|2 t is a solution of (8.1.3),

where as before, H denotes the Heaviside function from (1.2.9). Consequently, we
x (ξ, t) ∈ S (Rn+1 ). Also, if ϕ ∈ C ∞ (Rn+1 ) then, integration by parts yields
have E 0
x + |ξ|2 E
∂t E x , ϕ = −H(t)e−|ξ|2 t , ∂t ϕ(ξ, t) + H(t)e−|ξ|2 t , |ξ|2 ϕ(ξ, t)
∞
e−|ξ| t ∂t ϕ(ξ, t) dt dξ
2
=−
Rn 0
∞
|ξ|2 e−|ξ| t ϕ(ξ, t) dξ dt
2
+
Rn
0
= ϕ(ξ, 0) dξ = 1(ξ) ⊗ δ(t), ϕ, (8.1.4)

Rn
hence E x +|ξ|2 E
x verifies ∂t E x = 1(ξ)⊗δ(t) in D (Rn+1 ). Invoking (4.1.33), it follows

that E x (ξ, t) = H(t)e −|ξ|2 t
verifies (8.1.2). We remark here that while the distribution
−H(−t)e−|ξ| t satisfies (8.1.3), based on our earlier discussion pertaining to the nature
2
of the function in (4.1.35), it does not belong to S (Rn+1 ), thus we cannot apply F x−1
to it.
Starting from the identity F x (F x E(x, t)) = (2π)n E(−x, t) (which is easy to check),
we may write
x (ξ))(x) = (2π)−n H(t)Fξ (e−t|ξ|2 )(x)

E(−x, t) = (2π)−n Fξ (E
π n2 |x|2 |x|2
= (2π)−n H(t) e− 4t = H(t)(4πt)− 2 e− 4t in S (Rn+1 ),
n
(8.1.5)
t
where for the third equality in (8.1.5) we used Remark 4.23 and (3.2.6). Hence, we
may conclude that the tempered distribution from (8.1.5) is a fundamental solution
for the heat operator.
Note that, with the notation from (3.1.10), we have

n
∂t − Δ x = iDn+1 + D2j , (8.1.6)
j=1
hence the heat operator satisfies the hypothesis of Proposition 5.8. Consequently,
if u ∈ S (Rn+1 ) is an arbitrary fundamental solution for the heat operator in Rn+1 ,
then u − E = P(x, t) in S (Rn+1 ), for some polynomial P(x, t) in Rn+1 satisfying
(∂t − Δ x )P(x, t) = 0 in Rn+1 .
As a final remark, we claim that E as in (8.1.5) satisfies E ∈ Lloc
1
(Rn+1 ). Indeed,
if K is a compact subset of R × [−R, R] for some R ∈ (0, ∞), then
n
8.1 Fundamental Solutions for the Heat Operator 323
R
|x|2
(4πt)− 2 e− 4t dx dt
n
0≤ E(x, t) dx dt ≤
K 0 Rn
R
= π− 2 e−|y| dy dt = R < ∞.
n 2
(8.1.7)
0 Rn
In summary, we proved the following theorem.

Theorem 8.1. The function defined as
|x|2
E(x, t) := H(t)(4πt)− 2 e− 4t ,
n
∀ (x, t) ∈ Rn+1 , (8.1.8)
belongs to S (Rn+1 ) ∩ Lloc

1
(Rn+1 ) ∩ C ∞ (Rn+1 \ {0}) and is a fundamental solution for
the heat operator ∂t − Δ x in Rn+1 . Moreover,

u ∈ S (Rn+1 ) : (∂t − Δ x )u = δ(x) ⊗ δ(t) in S (Rn+1 ) (8.1.9)

= E + P : P polynomial in R satisfying (∂t − Δ x )P(x, t) = 0 .
n+1
Corollary 8.2. The heat operator ∂t − Δ x is hypoelliptic in Rn+1 .

Proof. This is a consequence of Theorem 6.8 and Theorem 8.1.
Exercise 8.3. Let E be the function defined in (8.1.9) and for each (x, t) ∈ R set n+1
F(x, t) := −E(x, −t). Prove that F is a fundamental solution for the operator ∂t + Δ
in Rn+1 .
|x|2
Remark 8.4. Given the expression E(x, t) = H(t)(4πt)− 2 e− 4t for (x, t) ∈ Rn+1 , then
n
one may check via a direct computation that this is a fundamental solution for the
heat operator L = ∂t − Δ x . First, the computation in (8.1.7) gives that E ∈ Lloc
1
(Rn+1 ),

which in turn implies E ∈ D (R ). n+1
Second, if ϕ ∈ C0∞ (Rn+1 ) is arbitrary, then using integration by parts we may

write
LE, ϕ = −E, ∂t ϕ + Δ x ϕ
∞
= − lim+ E(x, t)(∂t ϕ(x, t) + Δ x ϕ(x, t)) dx dt
ε→0 ε Rn
∞
= lim+ E(x, ε)ϕ(x, ε) dx + LE(x, t)ϕ(x, t) dx dt
ε→0 Rn ε Rn

|x|2
(4πε)− 2 e− 4ε ϕ(x, ε) dx
n
= lim+
ε→0 Rn

√
π− 2 e−|y| ϕ(2 εy, ε) dy
n 2
= lim+
ε→0 Rn
= ϕ(0) = δ, ϕ. (8.1.10)
For the fourth equality in (8.1.10) we have used the fact that LE = 0 pointwise in
Rn × (0, ∞), for the fifth a suitable change of variables, while for the sixth equality
we applied Lebesgue’s Dominated Convergence Theorem. This proves that LE = δ
in D (Rn+1 ), thus E is a fundamental solution for L.
In closing we record a Liouville type theorem for the operator ∂t − Δ, which is a

particular case of Theorem 5.4.
Theorem 8.5 (Liouville’s Theorem for the heat operator). Any bounded function
in Rn+1 that is a null solution of the heat operator is constant.
8.2 The Generalized Cauchy Problem for the Heat Operator

Let F ∈ C 0 Rn × [0, ∞) , f ∈ C 0 (Rn ) and suppose u is a solution of the Cauchy
problem for the heat operator:
⎧
⎪
⎪
⎪ u ∈ C 0 Rn × [0, ∞) ,
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪ 0 n
⎨∂t u, ∂ x j u ∈ C R × (0, ∞) , j = 1, . . . , n,
2
⎪
⎪
⎪
⎪
(8.2.1)
⎪
⎪
⎪ ∂ u − Δ u = F in Rn
× (0, ∞),
⎪
⎪
⎪
t x
⎪
⎪
⎪
⎩u(·, 0) = f in Rn .
Denote by u and F the extensions by zero of u and F to the entire space Rn+1 . Then,
∞
if ϕ ∈ C0 (R ), integrating by parts and using Lebesgue’s Dominated Convergence
n+1
Theorem we obtain
∞

u, ϕ = −
(∂t − Δ x ) u, ∂t u + Δ x ϕ = − lim+ u(∂t ϕ + Δ x ϕ) dx dt
ε→0 ε Rn
∞
= lim+ (∂t u − Δ x u)ϕ dx dt + u(x, ε)ϕ(x, ε) dx
ε→0 ε Rn Rn
∞
= F(x, t)ϕ(x, t) dx dt + f (x)ϕ(x, 0) dx
0 Rn Rn

= F, ϕ + f (x) ⊗ δ(t), ϕ(x, t) . (8.2.2)
This proves that (∂t − Δ x ) + f (x) ⊗ δ(t) in D (Rn+1 ) and suggests the definition
u=F
made below. As a preamble, we introduce the notation

Rn+1
+ := (x, t) ∈ R
n+1
:t≥0 (8.2.3)
and the space

D+ (Rn+1 ) := {u ∈ D (Rn+1 ) : supp u ⊆ Rn+1
+ }. (8.2.4)
8.2 The Generalized Cauchy Problem for the Heat Operator 325
Definition 8.6. Let F0 ∈ D+ (Rn+1 ) and f ∈ D (Rn ) be given. Then a distribution
u ∈ D+ (Rn+1 ) is called a solution of the generalized Cauchy problem for
the heat operator for the data F0 and f , if u verifies
(∂t − Δ x )u = F0 + f (x) ⊗ δ(t) in D (Rn+1 ). (8.2.5)
The issue of solvability of equation (8.2.5) fits into the framework presented in
Remark 5.6. More precisely, let E be the fundamental solution for the heat operator
as given in (8.1.8). Then, if F0 ∈ D+ (Rn+1 ) and f ∈ D (Rn ) are such that
u := E ∗ [F0 + f ⊗ δ] exists in D (Rn+1 ) (8.2.6)
the distribution u satisfies (8.2.5). For this u to be a solution of the generalized

Cauchy problem for the heat operator, we would also need supp u ⊆ Rn+1 + . Since it
is not difficult to check that supp E ⊆ R+ and supp ( f ⊗ δ) ⊆ R+ , and since by
n+1 n+1
assumption supp F0 ⊆ Rn+1 + , it follows that whenever condition (8.2.6) is verified,

the distribution u defined in (8.2.6) also satisfies supp u ⊆ Rn+1
+ (by (a) in Theo-
rem 2.96). While the convolution of two arbitrary distributions in D+ (Rn ) does not
always exist (you might want to check that an exception is the case n = 1), under the
additional assumptions f ∈ E (Rn ) and F0 ∈ E (Rn+1 ) condition (8.2.6) is verified.
The above discussion is the reason why we analyze in detail the following setting.
Retain the notation introduced at the beginning of the section, and assume that
∈ C0∞ (Rn+1 )
F and f ∈ C0∞ (Rn ). (8.2.7)
Then
+ E ∗ f (x) ⊗ δ(t) exists, belongs to D+ (Rn+1 ),
u := E ∗ F (8.2.8)
and is a solution of (8.2.5). We proceed by rewriting the expression for u in a more

explicit form. First, by Proposition 2.102 and the fact that E ∈ Lloc
1
(Rn+1 ), we have
E∗F ∈ C ∞ (Rn+1 ) and
t) = E(y, τ), F(x

(E ∗ F)(x, − y, t − τ) = E(x − y, t − τ), F(y, τ)
∞
n (x−y)2
= H(t − τ) 4π(t − τ) − 2 e− 4(t−τ) F(y, τ) dy dτ
0 Rn
t
= E(x − y, t − τ)F(y, τ) dy dτ. (8.2.9)
0 Rn
To compute E ∗ ( f ⊗ δ), fix an arbitrary compact set K ⊂ Rn+1 , consider a function

ϕ ∈ C0∞ (Rn+1 ) such that supp ϕ ⊆ K, and pick some ψ ∈ C0∞ (R2n+2 ) with the property
that ψ ≡ 1 in a neighborhood of the set

(x, t, y, 0) ∈ Rn × R × Rn × R : y ∈ supp f and (x + y, t) ∈ K . (8.2.10)
Relying on the definition of convolution we have

E ∗ ( f ⊗ δ), ϕ = E(x, t), f (y) ⊗ δ(τ), ψ(x, t, y, τ)ϕ(x + y, t + τ)

= E(x, t), f (y)ϕ(x + y, t) dy
Rn

= E(x, t) f (y)ϕ(x + y, t) dx dy dt
R Rn Rn

= E(z − y, t) f (y)ϕ(z, t) dz dy dt. (8.2.11)
R Rn Rn

Hence, E ∗ f (x) ⊗ δ(t) is given by the function

Rn × R (x, t) → E(x − y, t) f (y) dy ∈ C, (8.2.12)
Rn
whose restriction to Rn × (0, ∞) is of class C ∞ .

In summary, this analysis proves the following result.

Proposition 8.7. Let f ∈ C0∞ (Rn ) and assume that F ∈ C 0 Rn × [0, ∞) is such that
its extension F by zero to Rn+1 satisfies F
∈ C ∞ (Rn+1 ). Also let E be the fundamental
0
solution for the heat operator ∂t − Δ x as given in (8.1.8).
Then the generalized Cauchy problem for the heat operator for the data F and
f has a solution u ∈ D+ (Rn+1 ) that is of function type, whose restriction to the set
Rn × (0, ∞) is of class C ∞ , and has the expression
t
u(x, t) = E(x − y, t) f (y) dy + τ) dy dτ
E(x − y, t − τ)F(y, (8.2.13)
Rn 0 Rn
for every x ∈ Rn and t ∈ (0, ∞).
Note that the integrals in (8.2.13) are meaningfully defined under weaker assump-
tions on F and f . In fact, starting with u as in (8.2.13) one may prove that this is a
solution to a version of (8.2.1) (corresponding to finite time, i.e., t ∈ (0, T ), for some
T > 0) under suitable yet less stringent conditions on F and f .
8.3 Fundamental Solutions for General Second-Order Parabolic

Operators
We continue to work in Rn+1 and use the notation (x, t) := (x1 , . . . , xn , t) ∈ Rn+1 . We
look at a more general parabolic operator than the heat operator ∂t − Δ x discussed

so far in this chapter. Specifically, if A = a jk 1≤ j,k≤n ∈ Mn×n (C), associate to such a
matrix A the parabolic operator
8.3 Fundamental Solutions for General Second-Order Parabolic Operators 327

n
LA := LA (∂) := ∂t − a jk ∂ j ∂k , (8.3.1)
j,k=1
where ∂ j = ∂ x j , j = 1, . . . , n. The goal is to obtain explicit formulas for all tem-

pered distributions in Rn+1 that are fundamental solutions for LA in Rn+1 under the
additional assumption that there exists a constant C ∈ (0, ∞) such that the matrix A
satisfies the strict positiveness condition

n
Re a jk ξ j ξk ≥ C|ξ|2 , ∀ ξ = (ξ1 , . . . , ξn ) ∈ Rn . (8.3.2)
j,k=1
The approach is an adaptation to the parabolic setting of the ideas used in Sec-
tion 7.12 for the derivation of (7.12.50). In a first stage, we note that, via the same
reasoning as in Section 7.12, when computing the fundamental solution for LA we
may assume without loss of generality that A is symmetric, i.e., A = A . Also, we
treat first the case when A has real entries.
The Case When A is Real, Symmetric and Satisfies (8.3.2).
1
Since A is real, symmetric and positive definite,
√ A 2 is well defined, real, symmet-
1 1 1
ric, invertible, A 2 A 2 = A, and det(A 2 ) = detA. Consider next the matrix B in
M(n+1)×(n+1) (R) whose entries on the positions ( j, k), for j, k ∈ {1, . . . , n}, coincide
1
with the entries of A 2 , has 1 on the entry (n + 1, n + 1) and
zeros on the rest of
1
the entries. Hence, using matrix block notation, B = A 0 . Then B

2
is symmetric,
0 1
−1
√
−1 = A 2 0 and det B
invertible, B = detA. A direct computation based on the
0 1
Chain Rule gives that
−1 ) = (−∂t − Δ x )ϕ ◦ B
n
− ∂t − a jk ∂ j ∂k (ϕ ◦ B −1
j,k=1 (8.3.3)
pointwise in Rn+1 for every ϕ ∈ C0∞ (Rn+1 ).
This and (2.2.7) then allow us to write, for each u ∈ D (Rn+1 ),

ϕ = 1 −1
(LA u) ◦ B, LA u, ϕ ◦ B

|det B|
1 n

= u, − ∂t − −1
a jk ∂ j ∂k ϕ ◦ B

|det B| j,k=1
1 −1
= u, (−∂t − Δ x )ϕ ◦ B

|det B|

= u ◦ B, (−∂t − Δ x )ϕ
ϕ,
= (∂t − Δ x )(u ◦ B), ∀ ϕ ∈ C0∞ (Rn+1 ). (8.3.4)
In addition, with δ denoting the Dirac distribution in Rn+1 , we also have
ϕ = 1 −1 = √ 1 ϕ(0)
δ ◦ B, δ, ϕ ◦ B

det B detA
1
= √ δ, ϕ , ∀ ϕ ∈ C0∞ (Rn+1 ). (8.3.5)
detA
Our goal is to find E A ∈ D (Rn+1 ) such that LA E A = δ in D (Rn+1 ). The latter

= δ◦B
equality is equivalent with (LA E A ) ◦ B in D (Rn+1 ), which furthermore, in
view of (8.3.4)–(8.3.5), becomes
= √ 1 δ in D (Rn+1 ).
(∂t − Δ x )(E A ◦ B) (8.3.6)
detA
If now E∂t −Δ denotes the fundamental solution for the heat operator from (8.1.8), the
ensure that the function
properties of E∂t −Δ , (8.3.6), and the properties of B,
1
E∂t −Δ (A− 2 x, t) for each (x, t) ∈ Rn+1 ,
1
E A (x, t) := √ (8.3.7)
detA
is a fundamental solution for LA in Rn+1 . Keeping in mind that for every x ∈ Rn we

1 1
have |A− 2 x|2 = A− 2 x · A− 2 x = A−1 x · x, the function E A may be re-written as
1
1 (A−1 x)·x
H(t)(4πt)− 2 e−
n
E A (x, t) = √ 4t for each (x, t) ∈ Rn+1 . (8.3.8)
detA
Moreover, from (8.3.7), Theorem 8.1, and Proposition 4.43, it follows that E A be-
longs to S (Rn+1 ) ∩ Lloc
1
(Rn+1 ) ∩ C ∞ (Rn+1 \ {0}).
The Case When A Has Complex Entries, is Symmetric and Satisfies (8.3.2).
As observed in Remark 7.64, under the current assumptions, A continues to be
1
invertible. Also, (7.12.27) holds. In addition, under the current assumptions (detA) 2
is unambiguously defined (see in Remark 7.66).
These comments show that the function E A from (8.3.8) continues to be well-
defined under the current assumption on A if ln is replaced by the principal branch
of the complex log (defined for points z ∈ C \ (−∞, 0] so that za = ea log z for
each a ∈ R). In addition, E A continues to belong to Lloc 1
(Rn+1 ) (this can be seen
by a computation similar to that in (8.1.7), keeping in mind (7.12.24)), and E A ∈
C ∞ (Rn+1 \{0}). Moreover, from (7.12.42) and Exercise 4.6 it follows that E A belongs
to S (Rn+1 ).
The goal is to prove that this expression is a fundamental solution for LA in the
current case. Making use of (7.12.42) for every (x, t) ∈ Rn+1 with t 0 differentiat-
ing pointwise we obtain

n (A−1 x)·x (A−1 x)·x (A
−1
x) · x n
a jk ∂ j ∂k e− 4t = e− 4t − , (8.3.9)
j,k=1
4t2 2t
8.3 Fundamental Solutions for General Second-Order Parabolic Operators 329
while
(A−1 x)·x (A−1 x)·x (A
−1
x) · x n
∂t t−n/2 e− 4t = t−n/2 e− 4t − . (8.3.10)
4t2 2t
From (8.3.9), (8.3.10), and the expression for E A we may conclude that
LA E A (x, t) = 0 ∀ x ∈ Rn , ∀ t ∈ R \ {0}. (8.3.11)
In addition, for each ϕ ∈ C0∞ (Rn+1 ) we may compute

n
LA E A , ϕ = − E A , ∂t ϕ + a jk ∂ j ∂k ϕ
j,k=1
⎡ ⎤
⎢⎢⎢ ∞
n ⎥⎥⎥
= − lim+ ⎢⎢⎢⎣ E A (x, t) ∂t ϕ(x, t) + a jk ∂ j ∂k ϕ(x, t) dx dt⎥⎥⎦⎥
ε→0 ε R n
j,k=1
∞
= lim+ E A (x, ε)ϕ(x, ε) dx + (LA E A )(x, t)ϕ(x, t) dx dt
ε→0 Rn ε Rn

1 (A−1 x)·x
(4πε)− 2 e−
n
= lim+ √ 4ε ϕ(x, ε) dx
ε→0 detA Rn

π− 2 √
n
−1
= lim+ √ e−(A y)·y
ϕ(2 εy, ε) dy
ε→0 detA Rn
− n2
π −1
= ϕ(0) √ e−(A y)·y
dy. (8.3.12)
detA Rn
For the fourth equality

√ in (8.3.12) we have used (8.3.11), for the fifth the change of
variables x = 2 εy, while for the sixth equality we applied Lebesgue’s Dominated
Convergence Theorem. From (8.3.12) we then conclude that E A is a fundamental
solution for LA in Rn+1 if and only if

−1 n √
e−(A y)·y dy = π 2 detA. (8.3.13)
Rn
The fact that we already know that E A is a fundamental solution for LA in Rn+1 in
the case when A ∈ Mn×n (R) satisfies A = A and condition (8.3.2), implies that
formula (8.3.13) holds for this class of matrices. By using the same circle of ideas
as the ones employed in proving (7.12.47) (based on Lemma 7.67), we conclude
that (8.3.13) holds for the larger class of matrices A ∈ Mn×n (C) satisfying A = A
and condition (8.3.2). Hence, E A is indeed a fundamental solution for LA under the
current assumptions on A.
Next, we claim that the hypotheses of Proposition 5.8 are satisfied in the case
when P(D) := LA . To justify this, note that if ξ = (ξ1 , . . . , ξn+1 ) ∈ Rn+1 is such that
P(ξ) = 0 then

n
iξn+1 + a jk ξ j ξk = 0. (8.3.14)
j,k=1
Taking reals parts, condition (8.3.2) implies ξ1 = · · · = ξn = 0 which, in combi-

nation with (8.3.14), also forces ξn+1 = 0. Hence, ξ = 0 as wanted. Applying now
Proposition 5.8 gives that if u ∈ S (Rn+1 ) is an arbitrary fundamental solution for
LA in Rn+1 , then u = E A + P in S (Rn+1 ), for some polynomial P in Rn+1 satisfying
LA P = 0 pointwise in Rn+1 .
In summary, the above analysis proves the following result.

Theorem 8.8. Suppose A = a jk 1≤ j,k≤n ∈ Mn×n (C) satisfies (8.3.2) and consider
the operator LA associated to A as in (8.3.1). Then the function defined by
(A−1
sym x) · x
1 −
− n2
E A (x) := ' H(t)(4πt) e 4t for all (x, t) ∈ Rn+1 , (8.3.15)
detA sym
belongs to S (Rn+1 ) ∩ Lloc

1
(Rn+1 ) ∩ C ∞ (Rn+1 \ {0}) and is a fundamental solution
'
for LA in Rn+1 . Above, A sym := 12 (A + A ), detA sym is defined as in Remark 7.64,
and log denotes the principal branch of the complex logarithm (defined for points
z ∈ C \ (−∞, 0] so that za = ea log z for each a ∈ R).
Moreover,

u ∈ S (Rn+1 ) : LA u = δ in S (Rn+1 ) (8.3.16)

= E A + P : P polynomial in Rn+1 satisfying LA P = 0 .
8.4 Fundamental Solution for the Schrödinger Operator
Let x ∈ Rn and t ∈ R. The operator 1i ∂t − Δ x is called the (time dependent)

Schrödinger operator in Rn+1 (with zero potential). In this section we deter-
mine a fundamental solution for this operator. By Theorem 5.14, there exists some
E ∈ S (Rn+1 ) such that
1
∂t − Δ x E = δ(x) ⊗ δ(t) in S (Rn+1 ). (8.4.1)
i
Fix such a distribution E and take the partial Fourier transform F x of (8.4.1) (recall
the discussion on partial Fourier transform at the end of Section 4.2) to obtain
1 x = 1(ξ) ⊗ δ(t) in S (Rn+1 ).
∂t E x + |ξ|2 E (8.4.2)
i
Fix ξ ∈ Rn and consider the equation
8.4 Fundamental Solution for the Schrödinger Operator 331
1
∂t u + |ξ|2 u = δ in D (R). (8.4.3)
i
Using Example 5.12, we obtain that iH(t)e−i|ξ| t and −iH(−t)e−i|ξ| t are solution of
2 2
this equation. This suggests considering F := iH(t)e−i|ξ| t that belongs to S (Rn+1 )

2
and satisfies (based on a computation similar to that from (8.1.4))

1
∂t F + |ξ|2 F = 1(ξ) ⊗ δ(t) in S (Rn+1 ). (8.4.4)
i
Then F x−1 (F) ∈ S (Rn+1 ) and
1
∂t − Δ x F x−1 (F) = δ(x) ⊗ δ(t) in S (Rn+1 ). (8.4.5)
i
To compute F x−1 (F), pick ϕ ∈ S(Rn+1 ) and, based on the properties of F x as well as
the expression for F, write

F x−1 (F), ϕ = (2π)−n F x F ∨ , ϕ = (2π)−n F ∨ , F x ϕ

−n
H(t)e−i|ξ| t (F x ϕ)(ξ, t) dξ dt
2
= (2π) i
Rn+1
∞
−n −i|ξ|2 t
= (2π) i H(t) e (F x ϕ)(ξ, t) dξ dt. (8.4.6)
0 Rn
Observe that, for each t > 0, we have

e−i|ξ| t (F x ϕ)(ξ, t) dξ = F x (e−it|x| ), ϕ(·, t)
2 2
Rn
π n2 |ξ|2
= e− 4it ϕ(ξ, t) dξ, (8.4.7)
it Rn
where the last equality is a consequence of Example 4.25 (used with a = t). Hence,
for every ϕ ∈ S(Rn+1 ) we have
−1
∞ π n2 −|ξ|2

F x (F), ϕ = (2π)−n i H(t) e 4it ϕ(ξ, t) dξ dt
0 it Rn
∞
|ξ|2
H(t)(4πt)− 2 e− 4it ϕ(ξ, t) dξ dt.
n n
= i1− 2 (8.4.8)
0 Rn
|x|2
We remark here that H(t)(4πt)− 2 e− 4it ∈ Lloc
n
1
(Rn+1 ) only if n = 1. In particular,
F x−1 (F) is of function type only if n = 1. In general, this distribution belongs to
S (Rn+1 ) and its action on ϕ ∈ S(Rn+1 ) is given as in (8.4.8).
In summary we proved the following result.
Theorem 8.9. The distribution E ∈ S (Rn+1 ) defined by

∞
|ξ|2
H(t)(4πt)− 2 e− 4it ϕ(ξ, t) dξ dt
n n
E, ϕ := i1− 2 (8.4.9)
0 Rn
for each ϕ ∈ S(Rn+1 ), is a fundamental solution for the Schrödinger operator 1i ∂t −

Δ x in Rn+1 . In particular, if n = 1 then E is of function type and is given by the
1
Lloc (R2 )
|x|2
E(x, t) = H(t)(4πt)− 2 e− 4it for each x ∈ R and t ∈ R.
1
(8.4.10)
Further Notes for Chapter 8. The heat equation is one example of what is commonly referred to
as linear evolution equations. Originally derived in physics from Fourier’s law and conservation of
energy (see, e.g., [77] for details), the heat equation has come to play a role of fundamental impor-
tance in mathematics and applied sciences. In mathematics, the heat operator is the prototype
for a larger class, called parabolic partial differential operators, that includes the operators studied
in Section 8.3. The Schrödinger operator is named after the Austrian physicist Erwin Schrödinger
who first introduced it in 1926. It plays a fundamental role in quantum mechanics, where it de-
scribes how the quantum state of certain physical systems changes in time.
Chapter 9
The Wave Operator
Abstract Here all fundamental solutions that are tempered distributions for the wave
operator are determined and then used as a tool in the solution of the generalized
Cauchy problem for this operator.
9.1 Fundamental Solution for the Wave Operator
The operator := ∂2t − Δ x , x ∈ Rn , t ∈ R, is called the wave operator in

Rn+1 . The wave operator arises from modeling vibrations in a string, membrane, or
elastic solid. The goal is to determine fundamental solutions for this operator that
are tempered distributions. By Theorem 5.14, we know that the wave operator
admits a fundamental solution E ∈ S (Rn+1 ). Hence, by (4.1.33) and Exercise 2.90,
we have
∂2t E − Δ x E = δ(x) ⊗ δ(t) in S (Rn+1 ). (9.1.1)
In this section we determine an explicit expression for two such fundamental solu-
tions.
Fix E ∈ S (Rn+1 ) that satisfies (9.1.1) and apply the partial Fourier transform F x
to this equation (recall the discussion about partial Fourier transforms from the last
part of Section 4.2) to obtain
x + |ξ|2 E
∂2t E x = 1(ξ) ⊗ δ(t) in S (Rn+1 ), (9.1.2)
x := F x (E) ∈ S (Rn+1 ). For ξ ∈ Rn \ {0} fixed consider the initial value

where E
problem (in the variable t)
⎧ 2
⎪
⎪
⎨ dt2 v + |ξ| v = 0 in R,
d 2
⎪
⎪
⎪
⎪
(9.1.3)
⎩(∂t v)(0) = 1, v(0) = 0,

334 9 The Wave Operator
sin(t|ξ|)
which admits the solution v(t) = |ξ| for t ∈ R. By Example 5.12, it follows
that vH and −vH are fundamental solutions for the operator dtd 2 + |ξ|2 in D (R). In
∨ 2
addition, vH and −vH ∨ belong to S (R) (based on (b) in Theorem 4.14 and Exer-
cise 4.119), thus
d2 d2
+ |ξ|2 (vH) = δ and + |ξ|2 (−vH ∨ ) = δ in S (R). (9.1.4)
dt2 dt2
Moreover, there exists c ∈ (0, ∞) such that

H(±t) sin(t|y|) ≤ c|t| for (y, t) ∈ B(0, 1) \ {0} ⊂ Rn+1 . (9.1.5)
|y|
In particular, if we define the functions
sin(t|y|)
f± (y, t) := H(±t) for (y, t) ∈ Rn+1 with t 0, (9.1.6)
|y|
then (9.1.5) implies f± ∈ Lloc

1
(Rn+1 ). Furthermore,
(1 + |y|2 + t2 )−n−2 f± ∈ L1 (Rn+1 ), (9.1.7)
thus by Example 4.4, we obtain H(±t) sin(t|y|)

|y| ∈ S (Rn+1 ). Based on all these facts, if
we introduce
sin(t|ξ|) sin(t|ξ|)
F + := H(t) , F − := −H(−t) , (9.1.8)
|ξ| |ξ|
then
F + , F − ∈ S (Rn+1 ) and (9.1.9)
∂2t F ± + |ξ|2 F ± = 1(ξ) ⊗ δ(t) in S (Rn+1 ). (9.1.10)
In particular, F x−1 (F + ) and F x−1 (F − ) are meaningfully defined in S (Rn+1 ). Thus, if

we set

sin(t|ξ|)
E+ := F x−1 H(t) , (9.1.11)
|ξ|

−1 sin(t|ξ|)
E− := F x −H(−t) , (9.1.12)
|ξ|
we have
E± ∈ S (Rn+1 ), (∂2t − Δ x )E± = δ in S (Rn+1 ), (9.1.13)
supp E+ ⊆ Rn × [0, ∞), supp E− ⊆ Rn × (−∞, 0]. (9.1.14)

9.1 Fundamental Solution for the Wave Operator 335
The next task is to find explicit expressions for F x−1 (F ± ). To this end, fix a func-
tion ϕ ∈ S(Rn+1 ) and, with the operation ·∨ considered only in the variable x, write

E+ , ϕ = (2π)−n (F x F x (E+ ))∨ , ϕ = (2π)−n F x (E+ ), (F x ϕ)∨

sin t|ξ|
= (2π)−n H(t)
ϕ x (−ξ, t) dξ dt. (9.1.15)
Rn+1 |ξ|

Note that if one replaces ϕ x (−ξ, t) above with Rn eix·ξ ϕ(x, t) dx then the order of
integration in the resulting iterated integral may not be switched since |ξ|1 is not in-
tegrable at infinity, thus Fubini’s theorem does not apply. This is why we should
proceed with more care and, based on Lebesgue’s Dominated Convergence Theo-
rem, we introduce a convergence factor which enables us to eventually make the use
of Fubini’s theorem. More precisely, we have

−n

E+ , ϕ = (2π) F + (ξ, t)
ϕ x (−ξ, t) dξ dt
Rn+1

−n
= lim+ (2π) F + (ξ, t)e−ε|ξ|
ϕ x (−ξ, t) dξ dt
ε→0 Rn+1

−n
= lim+ (2π) ϕ(x, t) eix·ξ−ε|ξ| F + (ξ, t) dξ dx dt
ε→0 Rn+1 Rn
= lim+
Eε , ϕ, (9.1.16)
ε→0
where

sin(t|ξ|)
Eε (x, t) := (2π)−n H(t) eix·ξ−ε|ξ| dξ, ∀ x ∈ Rn , ∀ t ∈ R. (9.1.17)
Rn |ξ|
To compute the last limit in (9.1.16), we separate our analysis into three cases: n = 1,
n = 2p + 1 with p ≥ 1, and n = 2p with p ≥ 1.
9.1.1 The Case n = 1
Fix ε > 0 and x ∈ R and define the function

∞
sin(t|ξ|)
fε (t) := eixξ−ε|ξ| dξ for t ∈ R. (9.1.18)
−∞ |ξ|
Then
∞ ∞
eit|ξ| + e−it|ξ|
∂t fε (t) = eixξ−ε|ξ| cos(t|ξ|) dξ = eixξ−ε|ξ| dξ
−∞ −∞ 2

1 ∞ ∞
= eξ(ix+it−ε) dξ + eξ(ix−it−ε) dξ
2 0 0
1 0 0
+ eξ(ix−it+ε) dξ + eξ(ix+it+ε) dξ
2 −∞ −∞

1 1 1 1 1
= − − + +
2 ix + it − ε ix − it − ε ix − it + ε ix + it + ε
ε ε
= + , ∀ t ∈ R. (9.1.19)
(x + t) + ε
2 2 (x − t)2 + ε2
Consequently, (9.1.19) and the fact that fε (0) = 0 imply
x + t x − t
fε (t) = arctan − arctan for t ∈ R. (9.1.20)
ε ε
Making use of (9.1.20) back in (9.1.17) (written for n = 1) then gives
1 x + t x − t
Eε (x, t) = H(t) arctan − arctan , ∀ x, t ∈ R. (9.1.21)
2π ε ε
Hence, for ϕ ∈ S(R2 ) fixed, we may write
lim
Eε (x, t), ϕ(x, t) (9.1.22)
ε→0+
∞ x + t x − t
1
= lim+ arctan − arctan ϕ(x, t) dt dx.
ε→0 2π R 0 ε ε
To continue with our calculation, we further decompose
∞ x + t
arctan ϕ(x, t) dx dt
0 R ε
∞ −t x + t ∞ ∞ x + t
= arctan ϕ(x, t) dx dt + arctan ϕ(x, t) dx dt
0 −∞ ε 0 −t ε
=: Iε + IIε . (9.1.23)
By Lebesgue’s Dominated Convergence Theorem,

π ∞ −t
lim Iε = − ϕ(x, t) dx dt (9.1.24)
ε→0+ 2 0 −∞
and ∞ ∞
π
lim+ IIε = ϕ(x, t) dx dt. (9.1.25)
ε→0 2 0 −t
Similarly,
∞ x − t
lim arctan ϕ(x, t) dx dt
ε→0+ 0 R ε
∞ ∞ ∞
π t
π
=− ϕ(x, t) dx dt + ϕ(x, t) dx dt. (9.1.26)
2 0 −∞ 2 0 t
Combining (9.1.22)–(9.1.26) permits us to write
lim
Eε (x, t), ϕ(x, t) (9.1.27)
ε→0+
∞ −t ∞
1
= − ϕ(x, t) dx + ϕ(x, t) dx dt
4 0 −∞ −t
∞ t ∞
1
+ ϕ(x, t) dx − ϕ(x, t) dx dt
4 0 −∞ t
∞ t
1 1
= ϕ(x, t) dx dt = H(t − |x|)ϕ(x, t) dx dt.
2 0 −t 2 R2
From (9.1.27) and (9.1.16) we then conclude that1 E+ (x, t) = H(t−|x|)

2 for each point
(x, t) ∈ R2 . In summary, we proved that
⎧
⎪
⎪
⎪ H(t − |x|)
⎪ E+ (x, t) =
⎨ , ∀ (x, t) ∈ R2 , is a tempered distribution
⎪
⎪
⎪
2 (9.1.28)
⎪
⎩ and satisfies (∂2 − ∂2 )E+ = δ in S (R2 ).
t x
Remark 9.1.
(1) The reasoning used to obtain E+ also yields that

⎧
⎪
⎪
⎪ H(−t − |x|)
⎪
⎨ E− (x, t) := , ∀ (x, t) ∈ R2 , is a tempered distribution
⎪
⎪
⎪
2 (9.1.29)
⎪
⎩ and satisfies (∂2 − ∂2 )E− = δ in S (R2 ).
t x
(2) An inspection of (9.1.28) and (9.1.29) reveals that
supp E+ = {(x, t) ∈ R2 : |x| ≤ t}

(9.1.30)
and supp E− = {(x, t) ∈ R2 : |x| ≤ −t}.
1
this expression for a fundamental solution for the wave operator when n = 1 was first used by
Jean d’Alembert in 1747 in connection with a vibrating string
9.1.2 The Case n = 2p + 1, p ≥ 1
It is immediate from (9.1.17) that Eε (x, t) is invariant under orthogonal transforma-

tions in the variable x. Also, if y ∈ Rn \ {0}, we may set v1 := |y|1 y, then complete
to an orthonormal basis {v1 , v2 , . . . , vn } in Rn , and consider the orthogonal transfor-
mation A : Rn → Rn satisfying A(v j ) = e j , for j ∈ {1, . . . , n}. Then A(y) = |y|e1
which when combined with the invariance of Eε (x, t) under orthogonal transforma-
tions in the variable x, implies that Eε (y, t) = Eε (|y|, 0, . . . , 0, t) for every y in Rn and
t ∈ R. As such, in what follows we may assume x = (|x|, 0, . . . , 0). Fix such an x
and set r := |x|. With t ∈ R fixed, using polar coordinates (see (14.9.1), (14.9.2), and
(14.9.3)), rewrite the expression from (9.1.17) as
∞ 2π
sin tρ n−1
Eε (x, t) = (2π)−n H(t) eiρr cos θ1 −ερ ρ ×
0 (0,π)n−2 0 ρ
× (sin θ1 )n−2 · · · sin(θn−2 ) dθn−1 · · · dθ1 dρ

∞ π
= (2π)−n H(t) ρn−2 sin(tρ) eiρr cos θ1 −ερ (sin θ1 )n−2 ×
0 0
2π
× (sin θ2 ) n−3
. . . sin(θn−2 ) dθn−1 · · · dθ2 dθ1 dρ
(0,π)n−3 0
∞ π
= (2π)−n ωn−2 H(t) eiρr cos θ−ερ ρn−2 (sin θ)n−2 sin(tρ) dθ dρ (9.1.31)
0 0
where ωn−2 denotes the surface area of the unit ball in Rn−1 (see also (14.6.6) for
more details on why the expression inside the right brackets in (9.1.31) is equal to
ωn−2 ).
To proceed with the computation of the integrals in the rightmost term in
(9.1.31), recall that n = 2p + 1 and, for ρ > 0 fixed, set
π
I p := ρ2p−1 eiρr cos θ (sin θ)2p−1 dθ, ∀ p ∈ N. (9.1.32)
0
We claim that
2p
∂r I p for all p ≥ 1.
I p+1 = − (9.1.33)
r
In order to prove (9.1.33) note that, for each p ≥ 1, integration by parts yields
π
I p+1 = ρ2p+1 eiρr cos θ (sin θ)2p+1 dθ
0
π
−1
= ρ2p+1 ∂θ (eiρr cos θ )(sin θ)2p dθ
iρr 0
π
i θ=π
= ρ2p eiρr cos θ (sin θ)2p θ=0 −2p eiρr cos θ (sin θ)2p−1 cos θ dθ
r 0
π
2pi 2p
=− ρ eiρr cos θ (sin θ)2p−1 cos θ dθ
r 0
π
2pi 2p 1 2p
=− ρ ∂r eiρr cos θ (sin θ)2p−1 dθ = − (∂r I p ), (9.1.34)
r iρ 0 r
as wanted. By induction, from the recurrence relation in (9.1.33) it follows that
1 (p−1)
I p = (−2) p−1 (p − 1)! ∂r I1 for each p ∈ N. (9.1.35)
r
As for I1 , we have
π
−1 iρr cos θ θ=π 2 sin(ρr)
I1 = ρ eiρr cos θ sin θ dθ = e θ=0 = . (9.1.36)
0 ir r
Recalling (9.1.31) and the fact that n = 2p + 1 for some p ∈ N, formulas (9.1.35)
and (9.1.36) yield
∞
Eε (x, t) = (2π)−n ωn−2 H(t) I p e−ερ sin(tρ) dρ (9.1.37)
0
−n
= (2π) ωn−2 H(t)(−2) p−1 (p − 1)! 2 ×
∞ 1 p−1 sin(ρr)
× e−ερ sin(tρ) ∂r dρ.
0 r r
Furthermore, using (14.5.6) we have
n−1
2π 2 2π p 2π p
ωn−2 = = = (9.1.38)
Γ 2n−1 Γ(p) (p − 1)!
which further simplifies the expression in (9.1.37) to
Eε (x, t) (9.1.39)
1 p−1 1 ∞
= 2(−2π)−p−1 H(t) ∂r e−ερ sin(tρ) sin(ρr) dρ .
r r 0
Our next claim is that

∞
1 ε ε
e−ερ sin(tρ) sin(ρr) dρ = − . (9.1.40)
0 2 ε2 + (t − r)2 ε2 + (t + r)2
Indeed,
∞
e−ερ sin(tρ) sin(ρr) dρ (9.1.41)
0
∞
1
= e−ερ [cos(t − r)ρ − cos(t + r)ρ] dρ
2 0
∞
1
= Re [e−ερ+i(t−r)ρ − e−ερ+i(t+r)ρ ] dρ
2 0

1 1 1
= Re − +
2 −ε + i(t − r) −ε + i(t + r)

1 ε + i(t − r) −ε − i(t + r)
= Re 2 +
2 ε + (t − r)2 ε2 + (t + r)2

1 ε ε
= − ,
2 ε2 + (t − r)2 ε2 + (t + r)2
proving (9.1.40). The identity resulting from (9.1.41) further simplifies the expres-
sion in (9.1.39) as
Eε (x, t) (9.1.42)
1 (p−1) 1
ε ε

= (−2π)−p−1 H(t) ∂r −
r r ε2 + (t − r)2 ε2 + (t + r)2
for every x ∈ Rn and every t ∈ R, where r = |x|.

Recall from (9.1.16) that in order to determine E+ we further need to compute
lim+ Eε (x, t) in S (Rn+1 ). With this goal in mind, fix ϕ ∈ S(Rn+1 ) and use (9.1.42)
ε→0
in concert with 14.9.8 to write

Eε , ϕ (9.1.43)
∞ ∞
1 p−1 ε

1 1
= (−2π)−p−1 ∂r − 2 ×
0 0 ∂B(0,r) r r ε + (t − r)
2 2 ε + (t + r)2
× ϕ(ω, t) dσ(ω) dr dt
∞

−p−1
∞
1 p−1 1 1 1
= (−2π) ε ∂r − ×
0 0 r r ε2 + (t − r)2 ε2 + (t + r)2

× ϕ(ω, t) dσ(ω) dr dt.
∂B(0,r)
A natural question to ask is whether, if p ≥ 2, we may integrate by parts (p − 1)

times in the r variable in the rightmost expression in (9.1.43). Observe that, at least
formally,
∞ ∞
1 d d 1
f (r) g(r) dr = − f (r) g(r) dr (9.1.44)
0 r dr 0 dr r
if
f (r)g(r) f (r)g(r)
= 0 = lim+
lim . (9.1.45)
r r→∞ r→0 r
In the setting of the last expression in (9.1.43), for each ε, t > 0 fixed, and assuming
p ≥ 2, if we set, for each r > 0,
1 p−2 1
1 1

f (r) := ∂r − , (9.1.46)
r r ε2 + (t − r)2 ε2 + (t + r)2

g(r) := ϕ(ω, t) dσ(ω) = rn−1 ϕ(rω, t) dσ(ω), (9.1.47)
∂B(0,r) S n−1
then these functions satisfy (9.1.45) and (9.1.44). Proceeding by induction (with
p ≥ 2), we apply (p − 1) times formula (9.1.44), pick up (−1) p−1 in the process (the
last factor 1r bundled up with the derivative) and write
∞ ∞

ε ε

Eε , ϕ = (2π)−p−1 − × (9.1.48)
0 0 ε2 + (t − r)2 ε2 + (t + r)2
1 p−1 1
× ∂r ϕ(ω, t) dσ(ω) dr dt.
r r ∂B(0,r)
Note that (9.1.48) is also valid if p = 1 without any need of integration by parts.
We are left with taking the limit as ε → 0+ in (9.1.48) a task we complete by
using Lemma 9.3 (which is stated and proved at the end of this subsection). Specif-
ically, we apply Lemma 9.3 with

p−1
1 1
h(r) := ∂r ϕ(ω, t) dσ(ω) . (9.1.49)
r r ∂B(0,r)
Note that the second equality in (9.1.47) and the fact that ϕ ∈ S(Rn+1 ) guarantee
that h in (9.1.49) satisfies the hypothesis of Lemma 9.3. These facts combined with
Lebesgue’s Dominated Convergence Theorem yield
lim
Eε , ϕ (9.1.50)
ε→0+

∞
1 p−1 1
= (2π)−p−1 π ∂r ϕ(ω, t) dσ(ω) dt.
0 r r ∂B(0,r) r=t
In summary, we proved the following result:
If n = 2p + 1, for some p ∈ N, then E+ ∈ S (Rn+1 ) defined by

∞
1 p−1 1

E+ , ϕ = (2π)−p−1 π ∂r ϕ(ω, t) dσ(ω) dt (9.1.51)
0 r r ∂B(0,r) r=t
for ϕ ∈ S(Rn+1 ) is a fundamental solution for the wave operator in Rn+1 .
The reasoning used to obtain (9.1.51) also yields an expression for E− . More
precisely, similar to (9.1.16), we arrive at

E− , ϕ = lim+
Eε , ϕ for each ϕ ∈ S(Rn+1 ), (9.1.52)
ε→0
where, this time,

∞

0
ε ε

Eε , ϕ = −(2π)−p−1 − ×
−∞ 0 ε2 + (t − r)2 ε2 + (t + r)2
1 p−1 1
× ∂r ϕ(ω, t) dσ(ω) dr dt
r r ∂B(0,r)
∞ ∞

−p−1 ε ε
= (2π) − × (9.1.53)
0 0 ε2 + (t − r)2 ε2 + (t + r)2
1 p−1 1
× ∂r ϕ(ω, −t) dσ(ω) dr dt.
r r ∂B(0,r)
Invoking Lemma 9.3 we obtain:


If n = 2p + 1, for some p ∈ N, then E− ∈ S (R n+1
), defined by

−p−1
∞
1 p−1 1

E− , ϕ = (2π) π ∂r ϕ(ω, −t) dσ(ω) dt, (9.1.54)
0 r r ∂B(0,r) r=t
for ϕ ∈ S(Rn+1 ), is a fundamental solution for the wave operator in Rn+1 .
Remark 9.2.(1) The distributions E+ and E− from (9.1.51) and (9.1.54), respec-
tively, satisfy
supp E+ = {(x, t) ∈ R2p+1 × [0, ∞) : |x| = t}, (9.1.55)
supp E+ = {(x, t) ∈ R2p+1 × (−∞, 0] : |x| = −t}. (9.1.56)
(2) In the case when n = 3 (thus, for p = 1), formulas (9.1.51) and (9.1.54) become
∞
1 1

E+ , ϕ = ϕ(ω, t) dσ(ω) dt
4π 0 t ∂B(0,t)
H(t)
= δ∂B(0,t) , ϕ , (9.1.57)
4πt
and
∞
1 1

E− , ϕ = ϕ(ω, −t) dσ(ω) dt
4π 0 t ∂B(0,t)
H(−t)
= − δ∂B(0,−t) , ϕ , (9.1.58)
4πt
for every ϕ ∈ S(R4 ) where, for each R ∈ (0, ∞), the symbol δ∂B(0,R) stands for
the distribution defined as in Exercise 2.146 with Σ := ∂B(0, R).
(3) If n = 2p, p ∈ N, the approach used to obtain (9.1.51) works up to the point
where the general formula for I p was obtained. More precisely, with ρ > 0 fixed,
if we define π
Jn := ρn
eiρr cos θ (sin θ)n dθ, ∀ n ≥ 2, (9.1.59)
0
then the recurrence formula Jn = − n−1 r ∂r Jn−2

πis valid for all n ≥ 2 (observe
that I p = J2p−1 ). However, the integral J0 = 0 eiρr cos θ dθ cannot be computed
explicitly (as opposed to the computation of I1 ), hence the recurrence formula
for Jn may not be used inductively to obtain an explicit expression for Jn . This
is why, in order to obtain explicit expressions for E± when n is even we resort
to a proof different than the one used when n is odd.
Lemma 9.3. If h : [0, ∞) → R is continuous and bounded, then for each t in (0, ∞)
we have
∞

ε ε
lim+ − h(r) dr = πh(t). (9.1.60)
ε→0 0 ε2 + (t − r)2 ε2 + (t + r)2
Proof. If t ≥ 0 is fixed, then via suitable changes of variables we obtain
∞

ε ε
lim − h(r) dr
ε→0+ 0 ε2 + (t − r)2 ε2 + (t + r)2
⎡ ∞ ∞ ⎤
⎢⎢ h(t + ελ) h(−t + ελ) ⎥⎥⎥
= lim+ ⎢⎢⎣ dλ − dλ ⎥⎦
ε→0 − εt 1 + λ2 t
ε
1 + λ2
∞
h(t)
= dλ = πh(t), (9.1.61)
−∞ 1 + λ2
where for the second to the last equality in (9.1.61) we applied Lebesgue’s Domi-
nated Convergence Theorem.
9.1.3 The Method of Descent
In order to treat the case when n is even, we use a procedure called the method of
descent. The ultimate goal is to use the method of descent to deduce from a funda-
mental solution for the wave operator in dimension n + 1 with n even, a fundamental
solution for the wave operator in dimension n. To set the stage we make the follow-
ing definition.
Definition 9.4. A sequence {ψ j } j∈N of functions in C0∞ (R) is said to converge in

a dominated fashion to 1 if the following two conditions are satisfied:
(i) for every compact subset K of R there exists j0 = j0 (K) ∈ N such that ψ j (x) = 1
for all x ∈ K if j ≥ j0 ;
∞
(ii) for every q ∈ N0 , the sequence {ψ(q)
j } j∈N is uniformly bounded on R.
An example of a sequence {ψ j } j∈N converging in a dominated fashion to 1 is given

by
x
ψ j (x) := ψ , ∀ x ∈ R, ∀ j ∈ N, (9.1.62)
j
where
ψ ∈ C0∞ (R) satisfies ψ(x) ≡ 1 whenever |x| ≤ 1. (9.1.63)
In what follows we will use the notation
x = (x , xn ) ∈ Rn , x ∈ Rn−1 , xn ∈ R,
(9.1.64)
∂ = (∂ , ∂n ), ∂ = (∂1 , . . . , ∂n−1 ).
Definition 9.5. Call a distribution u ∈ D (Rn ) integrable with respect to

xn if for any ϕ ∈ C0∞ (Rn−1 ) and any sequence {ψ j } j∈N ⊂ C0∞ (R) converging in a

dominated fashion to 1, the sequence
u, ϕ ⊗ ψ j j∈N is convergent and its limit does
not depend on the selection of the sequence {ψ j } j∈N .
Suppose u ∈ D (Rn ) is integrable with respect to xn and {ψ j } j∈N ⊂ C0∞ (R) is a

sequence converging in a dominated fashion to 1. For each j ∈ N define the linear
mapping
u j : D(Rn−1 ) → C, u j (ϕ) :=
u, ϕ ⊗ ψ j , ∀ ϕ ∈ C0∞ (Rn−1 ). (9.1.65)
Then u j ∈ D (Rn−1 ) for each j ∈ N and lim

u j − uk , ϕ = 0 for every test func-
j,k→∞
tion ϕ ∈ C0∞ (Rn−1 ). Thus, the sequence {u j } j∈N is Cauchy in D (Rn−1 ) (see Sec-
tion 14.1.0.5). Since D (Rn−1 ) is sequentially complete (recall Fact 2.22), it follows
that
there exists some u0 ∈ D (Rn−1 ) with the property that
D (Rn−1 ) (9.1.66)
u j −−−−−−→ u0 and
u0 , ϕ = lim
u j , ϕ ∀ ϕ ∈ C0∞ (Rn−1 ).
j→∞ j→∞
Moreover, u0 is independent of the choice of the sequence {ψ j } j∈N converging in a

dominated fashion to 1 (prove this as an exercise). The distribution u0 willbe called
∞
the integral of u with respect to xn and will be denoted by −∞ u dxn .
The reason for using this terminology and notation for u0 is evident from the next
proposition. We denote by Ln−1 the Lebesgue measure in Rn−1 .
Proposition 9.6. If f ∈ Lloc 1

(Rn ) is a function with the property that

| f (x , xn )| dxn < ∞ for Ln−1 -almost every x ∈ Rn−1
R (9.1.67)
and f ∈ L (K × R)
1
for every compact set K ⊂ R n−1
,
then the distribution u f ∈ D (Rn ) determined by f (recall (2.1.6)) is integrable with
∞
respect to xn . In addition, the distribution u0 := −∞ u f dxn is of function type and is
given by the function

g(x ) := f (x , xn ) dxn defined for Ln−1 -almost every x ∈ Rn−1 . (9.1.68)
R
Proof. Since f ∈ Lloc

1
(Rn ) we have that u f ∈ D (Rn ) as in (2.1.6) is well defined. Fix
∞
ϕ ∈ C0 (R ). Since f is absolutely integrable on supp ϕ × R, whenever {ψ j } j∈N ⊂
n−1
C0∞ (R) is a sequence converging in a dominated fashion to 1, we may apply Fubini’s

theorem and then Lebesgue’s Dominated Convergence Theorem to write

lim u f , ϕ ⊗ ψ j = lim ϕ(x ) f (x , xn )ψ j (xn ) dxn dx
j→∞ j→∞ Rn−1 R

= ϕ(x ) f (x , xn ) dxn dx , (9.1.69)
Rn−1 R
and the desired conclusion follows.

The proposition that is the engine in the method of descent is proved next.
Proposition 9.7. Let m ∈ N0 and let P(∂) = P(∂ , ∂n ) be a constant coefficient,
linear operator of order m in Rn . Define the differential operator P0 (∂ ) := P(∂ , 0)
in Rn−1 . If f ∈ D (Rn−1 ) and u ∈ D (Rn ) are such that
P(∂)u = f (x ) ⊗ δ(xn )

in D (Rn ) (9.1.70)
∞
and u is integrable with respect to xn , then u0 := −∞ u dxn is a solution of the
equation P0 (∂ )u0 = f in D (Rn−1 ).
Proof. Fix a sequence {ψ j } j∈N ⊂ C0∞ (R) that converges in a dominated fashion to 1
and let ϕ ∈ C0∞ (Rn−1 ). Using the definition of u0 we may write

P0 (∂ )u0 , ϕ (9.1.71)

= u0 , P0 (−∂ )ϕ = lim u, (P0 (−∂ )ϕ) ⊗ ψ j
j→∞

= lim u, P(−∂)(ϕ ⊗ ψ j ) + u, (P0 (−∂ )ϕ) ⊗ ψ j − P(−∂)(ϕ ⊗ ψ j ) .
j→∞
We claim that

lim u , (P0 (−∂ )ϕ) ⊗ ψ j − P(−∂)(ϕ ⊗ ψ j ) = 0. (9.1.72)
j→∞
Assume the claim for now. Then returning to (9.1.71) we have

P0 (∂ )u0 , ϕ = lim u, P(−∂)(ϕ ⊗ ψ j ) = lim P(∂)u , ϕ ⊗ ψ j
j→∞ j→∞
!
= lim f ⊗ δ, ϕ ⊗ ψ j = lim f, ϕ ψ j (0) =
f, ϕ, (9.1.73)
j→∞ j→∞
where for the last equality in (9.1.73) we used property (i) in Definition 9.4 with
K = {0}. Hence, the desired conclusion follows.
To prove (9.1.72), observe that
" # $" #
m
P0 (−∂ )ϕ ⊗ ψ j − P(−∂)(ϕ ⊗ ψ j ) = Pq (∂ )ϕ ⊗ ψ(q)
j
q=1
where Pq is a differential operator of order ≤ m − q. Then for each q ∈ {1, . . . m}, the

sequence ψ j + ψ(q)
j j∈N also converges in a dominated fashion to 1, which combined
with the fact that u is integrable with respect to xn , further yields

" #
lim u, Pq (∂ )ϕ ⊗ ψ(q)
j
j→
" # " #
= lim u, Pq (∂ )ϕ ⊗ (ψ j + ψ(q) j ) − lim u, Pq (∂ )ϕ ⊗ ψ j
j→∞ j→∞

= u0 , Pq (∂ )ϕ − u0 , Pq (∂ )ϕ = 0, (9.1.74)
proving (9.1.72). The proof of the proposition is now complete.
9.1.4 The Case n = 2p, p ≥ 1
We are now ready to proceed with determining a fundamental solution for the
wave operator in the case when n = 2p, p ≥ 1. The main idea is to use Propo-
% 2
2p+1
sition 9.7 corresponding to P := ∂2t − ∂ j being the wave operator in Rn+2 ,
j=1
f := δ(x1 , . . . , x2p ) ⊗ δ(t), and u equal to E2p+1 , the fundamental solution from
%
2p
(9.1.51). Note that under these assumptions we have P0 := ∂2t − ∂2j which is the
j=1
wave operator in Rn+1 . Thus, if E2p+1
∞ is integrable with respect to x2p+1 , then by
Proposition 9.7 it follows that u := −∞ E2p+1 dx2p+1 satisfies
$
2p

∂2t − ∂2j u = δ(x1 , . . . , x2p ) ⊗ δ(t) in D (R2p+1 ). (9.1.75)
j=1
Therefore, u is a fundamental solution for the wave operator corresponding to n =

2p.
Let us first show that the distribution E2p+1 given by the formula in (9.1.51) is
integrable with respect to x2p+1 . Fix an arbitrary function ϕ ∈ C0∞ (R2p+1 ) and let
{ψ j } j∈N ⊂ C0∞ (R) be a sequence that converges in a dominated fashion to 1. Then,
using the notation x = (x , x2p+1 ) ∈ R2p × R, we have

lim E2p+1 , ϕ ⊗ ψ j
j→∞
∞
−(p+1) −p 1 p−1 1
= lim 2 π ∂r ϕ(x , t)ψ j (x2p+1 ) dσ(x) dt
j→∞ 0 r r r=t
x∈R2p+1 , |x|=r
∞
1 p−1 1
= 2−(p+1) π−p ∂r ϕ(x , t) dσ(x) dt, (9.1.76)
0 r r r=t
x∈R2p+1 , |x|=r
where for the last equality in (9.1.76) we used Lebesgue’s Dominated Conver-
gence Theorem (here we note that the second equality in (9.1.47) and the prop-
erties satisfied by {ψ j } j∈N play an important role). With (9.1.76) in hand, we may
conclude ∞that E2p+1 is integrable with respect to x2p+1 . Consequently, if one sets
E2p := −∞ E2p+1 dx2p+1 , then E2p ∈ D (R2p+1 ) and
∞
−(p+1) −p 1 p−1 1

E2p , ϕ = 2 π ∂r ϕ(x , t) dσ(x) dt (9.1.77)
0 r r r=t
x∈R2p+1 , |(x ,x2p+1 )|=r
for every ϕ ∈ C0∞ (R2p+1 ). In addition, by Proposition 9.7, it follows that E2p is a
fundamental solution for the wave operator in Rn+1 .
Note that the function ϕ appearing under the second integral in (9.1.77) does
not depend on the variable x2p+1 . Hence, it is natural to proceed further in order
to rewrite (9.1.77) in a form that does not involve the variable x2p+1 . Fix r > 0
and denote by B(0, r) the ball in Rn of radius r and centered at 0 ∈ Rn . Define the
mappings P± : B(0, r) → R2p+1 by setting
&
P± (x ) := (x1 , x2 , . . . , x2p , ± r2 − |x |2 )
(9.1.78)
for each x = (x1 , . . . , x2p ) ∈ B(0, r).
Then P+ and P− are parametrizations of the (open) upper and lower, respectively,
hemispheres of the surface ball in R2p+1 of radius r and centered at 0. Hence, (keep-
ing in mind Definition 14.47 and Definition 14.48) we may write

ϕ(x , t) dσ(x) (9.1.79)
x∈R2p+1 , |(x ,x2p+1 )|=r

= ϕ(x , t) |∂1 P+ × · · · × ∂2p P+ | dx
B(0,r)

+ ϕ(x , t) |∂1 P− × · · · × ∂2p P− | dx .
B(0,r)
A direct computation based on (9.1.78) and 14.6.4 further yields

1 0 . . . 0 ∓ √ 2 1 2
x
r −|x |
0 1 . . . 0 ∓ √ x2
r2 −|x |2

∂1 P± × · · · × ∂2p P± = . . . . . . . . . . . . . . . . . . . . . . =: det A± , (9.1.80)

0 0 . . . 1 ∓ √ x2p
r2 −|x |2

e1 e2 . . . e2p e2p+1
where e j is the unit vector in R2p+1 with 1 on the j-th position, for each j in
{1, . . . , 2p + 1}. Hence, the components of the vector ∂1 P± × · · · × ∂2p P± are
(−1)k+1 det A±k , k ∈ {1, . . . , 2p, 2p + 1}, (9.1.81)
where A±k is the 2p × 2p matrix obtained from A± by deleting column k and row
2p + 1. It is easy to see from (9.1.80) that
∓xk
det A±2p+1 = 1, det A±k = (−1)k & , ∀ k ∈ {1, . . . , 2p}. (9.1.82)
r2 − |x |2
Consequently,
r
|∂1 P± × · · · × ∂2p P± | = & . (9.1.83)
r − |x |2
2
Formula (9.1.83) combined with (9.1.79) and (9.1.77) gives

∞ ⎡⎢ ⎤
⎢⎢⎢ 1 p−1 1 r ⎥
⎥

E2p , ϕ = (2π) −p
⎢⎣ ∂r
ϕ(x , r) & dx ⎥⎥⎦⎥ dt
0 r r
x ∈B(0,r) 2
r − |x |2
r=t
⎡ ⎤
∞ ⎢⎢⎢ 1 p−1 ϕ(x , t) ⎥
⎥
= (2π) −p ⎢⎢⎣ ∂r & dx ⎥⎥⎥⎦ dt. (9.1.84)
0 r B(0,r) r − |x |
2 2
r=t
In summary, we proved:
If n = 2p, for some p ∈ N, then E+ ∈ S (Rn+1 ), defined by
⎡ ⎤
∞ ⎢⎢⎢ 1 p−1 ϕ(x, t) ⎥⎥

E+ , ϕ = (2π) −p ⎢⎢⎣ ∂r & dx ⎥⎥⎥⎦ dt, (9.1.85)
0 r x∈Rn , |x|<r r2 − |x|2 r=t
The reasoning used to obtain E+ also applies if one starts with E2p+1 being the
distribution from (9.1.54). Under this scenario the conclusion is that
If n = 2p, for some p ∈ N, then E− ∈ S (Rn+1 ), defined by

∞ ⎡⎢ ⎤
⎢⎢⎢ 1 p−1 ϕ(x, −t) ⎥⎥
⎥⎥⎥

E− , ϕ = (2π)−p ⎢⎣ ∂ r & dx ⎦ dt, (9.1.86)
0 r x∈R , |x|<r
n
r − |x|
2 2
r=t
Remark 9.8.(1) If E+ and E− are as in (9.1.85) and (9.1.86), respectively, then
supp E+ = {(x, t) ∈ R2p × [0, ∞) : |x| ≤ t}, and (9.1.87)
supp E− = {(x, t) ∈ R2p × (−∞, 0] : |x| ≤ −t}. (9.1.88)

(2) If p = 1, then for each ϕ ∈ S(R3 ) the expression2 in (9.1.85) becomes

∞ ' (
1 ϕ(x, t) H(t − |x|)

E+ , ϕ = & dx dt = & ,ϕ .
2π 0 |x|<t t2 − |x|2 2π t2 − |x|2
Hence, if n = 2 then E+ is of function type and
H(t − |x|)
E+ (x, t) = & for every x ∈ R2 and every t ∈ R.
2π t2 − |x|2
Similarly, when n = 2 it follows that E− is of function type and
H(−t − |x|)
E− (x, t) = & for every x ∈ R2 and every t ∈ R.
2π t2 − |x|2
9.1.5 Summary for Arbitrary n
Here we combine the results obtained in Section 9.1.1, Section 9.1.2, and Sec-
tion 9.1.4 regarding fundamental solutions for the heat operator. These results have
been summarized in (9.1.28), Remark 9.1, (9.1.51), (9.1.54), Remark 9.2, (9.1.85),
(9.1.86), and Remark 9.8.
Theorem 9.9. Consider the wave operator = ∂2t − Δ x in Rn+1 , where x ∈ Rn and
t ∈ R. Then the following are true.
(1) Suppose n = 1. Then the Lloc
1
(R2 ) functions
H(t − |x|)
E+ (x, t) := , ∀ (x, t) ∈ R2 , (9.1.89)
2
H(−t − |x|)
E− (x, t) := , ∀ (x, t) ∈ R2 , (9.1.90)
2
satisfy E+ , E− ∈ S (R2 ) and are fundamental solutions for the wave operator in
R2 . Moreover,
supp E+ = {(x, t) ∈ R2 : |x| ≤ t}, (9.1.91)
supp E− = {(x, t) ∈ R2 : |x| ≤ −t}. (9.1.92)
(2) Suppose n = 2p + 1, for some p ∈ N. Then the distributions E+ ∈ S (Rn+1 ) and

E− ∈ S (Rn+1 ) defined by
2
this expression was first found by Vito Volterra (cf. [79])

∞
1 p−1 1

E+ , ϕ := (2π)−p−1 π ∂r ϕ(ω, t) dσ(ω) dt, (9.1.93)
0 r r ∂B(0,r) r=t

∞
1 p−1 1

E− , ϕ := (2π)−p−1 π ∂r ϕ(ω, −t) dσ(ω) dt, (9.1.94)
0 r r ∂B(0,r) r=t
for every ϕ ∈ S(Rn+1 ), are fundamental solutions for the wave operator in Rn+1 .
Moreover,
supp E+ = {(x, t) ∈ R2p+1 × [0, ∞) : |x| = t}, (9.1.95)
supp E+ = {(x, t) ∈ R2p+1 × (−∞, 0] : |x| = −t}. (9.1.96)
Corresponding to the case n = 3 (thus, for p = 1), formulas (9.1.93) and (9.1.94)
become
∞ H(t)
1 1

E+ , ϕ = ϕ(ω, t) dσ(ω) dt = δ∂B(0,t) , ϕ , (9.1.97)
4π 0 t ∂B(0,t) 4πt
∞ H(−t)
1 1

E− , ϕ = ϕ(ω, −t) dσ(ω) dt = − δ∂B(0,−t) , ϕ , (9.1.98)
4π 0 t ∂B(0,t) 4πt
for every ϕ ∈ S(R4 ).

(3) Suppose n = 2p, for some p ∈ N. Then the distributions E+ ∈ S (Rn+1 ) and
E− ∈ S (Rn+1 ) defined by
∞ ⎡⎢ ⎤
⎢⎢⎢ 1 p−1 ϕ(x, t) ⎥⎥

E+ , ϕ := (2π) −p
⎢⎣ ∂r & dx ⎥⎥⎥⎦ dt, (9.1.99)
0 r x∈Rn , |x|<r r2 − |x|2 r=t
⎡ ⎤
∞ ⎢⎢⎢ 1 p−1 ϕ(x, −t) ⎥⎥

E− , ϕ := (2π) −p ⎢⎢⎣ ∂r & dx ⎥⎥⎥⎦ dt, (9.1.100)
0 r x∈Rn , |x|<r r2 − |x|2 r=t
for every ϕ ∈ S(Rn+1 ), are fundamental solutions for the wave operator in Rn+1 .
Moreover,
supp E+ = {(x, t) ∈ R2p × [0, ∞) : |x| ≤ t}, and (9.1.101)
supp E− = {(x, t) ∈ R2p × (−∞, 0] : |x| ≤ −t}. (9.1.102)
Corresponding to the case n = 2, thus p = 1, the distributions E+ and E− defined

in (9.1.99) and (9.1.100) are of function type and given by the functions
H(t − |x|)
E+ (x, t) = & , ∀ x ∈ R2 , ∀ t ∈ R, (9.1.103)
2π t2 − |x|2
H(−t − |x|)
E− (x, t) = & , ∀ x ∈ R2 , ∀ t ∈ R. (9.1.104)
2π t2 − |x|2
9.2 The Generalized Cauchy Problem for the Wave Operator
In this section we discuss the generalized Cauchy problem for the wave operator.
Let F ∈ C 0 (Rn+1 ), f, g ∈ C 0 (Rn ) and, as before, use the notation x ∈ Rn , t ∈ R.
Suppose u ∈ C 2 (Rn+1 ) solves in the classical sense the Cauchy problem
⎧ 2
⎪
⎨ (∂t − Δ x )u = F in R ,
⎪
n+1
⎪
⎪ (9.2.1)
⎩ u = f, ∂t u = g in Rn .
t=0 t=0
Define ) ) t) := H(t)F(x, t) for every x ∈ Rn and every

u(x, t) := H(t)u(x, t) and F(x,
t ∈ R. Then ) )
u, F ∈ Lloc (R ) and for each ϕ ∈ C0∞ (Rn+1 ) we have
1 n+1
∞
2
u, ϕ = )
(∂t − Δ x )) u, (∂2t − Δ x )ϕ = lim+ u (∂2t − Δ x )ϕ dx dt
ε→0 ε Rn
∞
= lim+ (∂2t − Δ x )u ϕ dx dt (9.2.2)
ε→0 ε Rn

+ lim+ ∂t u(x, ε)ϕ(x, ε) − u(x, ε)∂t ϕ(x, ε) dx
ε→0 Rn
)
= F, ϕ + f (x) ⊗ δ (t), ϕ(x, t) + g(x) ⊗ δ(x), ϕ(x, t) .
This suggests the following definition (recall (8.2.4)).

) ∈ D+ (Rn+1 ) and f, g ∈ D(Rn ) are given. Then a distri-
Definition 9.10. Suppose F

bution )
u ∈ D+ (R ) is called a solution of the generalized Cauchy problem for the
n+1
) f , and g, if
wave operator with data F,
(∂2t − Δ x )) ) + f (x) ⊗ δ (t) + g(x) ⊗ δ(t)

u=F in D (Rn+1 ). (9.2.3)
) ∈ D+ (Rn+1 ) and f, g ∈ D(Rn ) be given. Then the generalized
Theorem 9.11. Let F
Cauchy problem (9.2.3) has the unique solution
) ) + E ∗ ( f ⊗ δ ) + E ∗ (g ⊗ δ),
u=E∗F (9.2.4)
where E is the fundamental solution for the wave operator in Rn+1 as specified in
(9.1.89) if n = 1, in (9.1.93) if n ≥ 3 is odd, and in (9.1.99) if n is even.
Proof. Let E be as in the statement of the theorem. Then, by (9.1.91), by (9.1.95),

and by (9.1.101), for any G ∈ D+ (Rn+1 ) we have that, whenever K is a compact
subset of Rn+1 , the set
" #
MK := (x, t), (y, s) : (x, t) ∈ supp E,

(y, s) ∈ supp G, (x + y, t + s) ∈ K (9.2.5)
is compact in Rn+1 . Hence, by Theorem 2.94, the convolution E ∗ G exists. This

proves that )u as in (9.2.4) is a well-defined element of D (Rn+1 ). Moreover, since
)
F, f ⊗ δ , g ⊗ δ, E ∈ D+ (Rn+1 ), by (a) in Theorem 2.96, we may conclude that )

u
belongs to D+ (Rn+1 ). In addition, by Remark 5.6, we obtain that ) u as in (9.2.4) is a
solution of the generalized Cauchy problem (9.2.3).
u ∈ D+ (Rn+1 ) is such that (∂2t − Δ x ))
To prove uniqueness, observe that if ) u = 0 in

D (R ), then
n+1
)
u =)
u∗δ =)
u ∗ ((∂2t − Δ x )E) = ((∂2t − Δ x ))
u) ∗ E = 0. (9.2.6)
This completes the proof of the theorem.

Further Notes for Chapter 9. The wave operator := ∂2t − Δ x ,
where x ∈ R and t ∈ R, was orig-
n
inally discovered by the French mathematician and physicist Jean le Rond d’Alembert in 1747 in
the case n = 1 in his studies of vibration of strings. For this reason, is also called the d’Alembert
operator or, simply, the d’Alembertian. Like the heat operator discussed in Chapter 8, the wave
operator is another basic example of a partial differential operator governing a linear evolution
equation (though, unlike the heat operator, the wave operator belongs to a class of operators called
hyperbolic operators).
Exercise 9.12. Use the Method of Descent to compute a fundamental solution for
%
n
the Laplace operator Δ = ∂2j in Rn , n ≥ 3, by starting from a fundamental solution
j=1
%
n
for the heat operator L = ∂t − ∂2j in Rn+1 .
j=1
Chapter 10
The Lamé and Stokes Operators
Abstract The material here is centered around two basic systems: the Lamé operator
arising in the theory of elasticity, and the Stokes operator arising in hydrodynamics.
Among other things, all of their fundamental solutions that are tempered distribu-
tions are identified, and the well-posedness of the Poisson problem for the Lamé
system is established.
Throughout this chapter, it is assumed that n ∈ N satisfies n ≥ 2.
10.1 General Remarks About Vector and Matrix Distributions
The material developed up to this point may be regarded as a theory for scalar dis-
tributions. Nonetheless, practical considerations dictate the necessity of considering
vectors/matrices whose components/entries are themselves distributions. It is there-
fore natural to refer to such objects as vector and matrix distributions. A significant
portion of the theory of scalar distributions then readily extends to this more general
setting. The philosophy in the vector/matrix case is that we perform the same type
of analysis as in the scalar case, at the level of individual components, while at the
same time obeying the natural algebraic rules that are now in effect (e.g., keeping
in mind the algebraic mechanism according to which two matrices are multiplies,
etc.).
To offer some examples, fix an open set Ω ⊆ Rn and consider a matrix distribution

U = uk 1≤k≤N ∈ MN×K D (Ω) , (10.1.1)
1≤≤K
i.e., U is an N × K matrix whose entries are from D (Ω). Naturally, equality of

elements in MN×K D (Ω) is understood entry by entry in D (Ω). We agree to define
the support of U as

K N
supp U := supp uk . (10.1.2)
k=1 =1

356 10 The Lamé and Stokes Operators
Note that supp U is the smallest relatively closed subset of Ω outside of which all
entries of U vanish. Similarly, we define the singular support of U as

K
N
sing supp U := sing supp uk . (10.1.3)
k=1 =1
Hence, sing supp U is the smallest relatively closed subset of Ω outside of which all
entries of U are C ∞ .

Next, if A = a jk 1≤ j≤M ∈ M M×N C ∞ (Ω) , then we define (with U as in (10.1.1))
1≤k≤N

N
AU := a jk uk 1≤ j≤M ∈ M M×K D (Ω) . (10.1.4)
1≤≤K
k=1
Some trivial, yet useful properties include

if U, V ∈ MN×K D (Ω) , A ∈ M J×M C ∞ (Ω) , B ∈ M M×N C ∞ (Ω) ,
then A(BU) = (AB)U and
U = V =⇒ AU = AV.
(10.1.5)
Given an operator of the form

L= Aα ∂α , Aα = aαjk 1≤ j≤M ∈ M M×N C ∞ (Ω) , (10.1.6)
1≤k≤N
|α|≤m
referred to as an M × N system (of differential operators) of order m ∈ N, the natural

way in which L acts on U as in (10.1.1) is to regard LU as the matrix distribution

from M M×K D (Ω) whose entries are given by

N
(LU) j := aαjk ∂α uk for all 1 ≤ j ≤ M, 1 ≤ ≤ K. (10.1.7)
|α|≤m k=1
A matrix distribution with a single column is referred to a vector distribution and

simply denoted by u = (uk )1≤k≤N with components uk ∈ D (Ω), 1 ≤ k ≤ N. The

collection of all such vector distributions is denoted by D (Ω) N .
Regarding the convolution product for matrix distributions, if

U = u jk 1≤k≤N ∈ M M×N D (Rn ) and
1≤≤K
(10.1.8)
V = vk 1≤k≤N ∈ MN×K E (Rn )
1≤≤K
(hence, V is an N × K matrix whose entries are from E (Rn )), we define U ∗ V as

being the M × K matrix whose entries are the distributions from D (Rn ) given by
10.1 General Remarks About Vector and Matrix Distributions 357

N
(U ∗ V) j := u jk ∗ vk , 1 ≤ j ≤ M, 1 ≤ ≤ K. (10.1.9)
k=1
Formula (10.1.9) is also taken as definition for U ∗ V in the case when

U = u jk 1≤k≤N ∈ M M×N E (Rn ) (10.1.10)
1≤≤K
and

V = vk 1≤k≤N ∈ MN×K D (Rn ) . (10.1.11)
1≤≤K
The reader is advised that, as opposed to the scalar case, the commutativity property
for the convolution product is lost in the setting of (say, square) matrix distributions.
For each natural number M we agree to denote by δI M×M the matrix distribution

from M M×M D (Rn ) with entries δ jk δ, for all j, k ∈ {1, . . . , M}, where δ jk = 1 if
j = k and 0 otherwise, is the Kronecker symbol. Then, one can check using (10.1.9)
and part (d) in Theorem (2.96) that

(δI M×M ) ∗ U = U, ∀ U ∈ M M×N D (Rn ) , (10.1.12)

V ∗ (δI M×M ) = V, ∀ V ∈ MN×M D (Rn ) . (10.1.13)
Two important first-order systems of differential operators are the gradient ∇

and the divergence div. Specifically, the gradient ∇ := (∂1 , . . . , ∂n ) acts on a scalar

distribution according to ∇u := (∂1 u, . . . , ∂n u) ∈ D (Ω) n for every u ∈ D (Ω),
while the divergence operator div acts on a vector distribution

v := (v1 , . . . , vn ) ∈ D (Ω) n (10.1.14)

n
according to divv := ∂ jv j.
j=1
To give an example of a specific result from the theory of scalar distributions
and scalar differential operators that naturally carries over to the vector/matrix case,
we note here that if L is an J × M system with constant coefficients, then for every

matrix distributions U ∈ M M×N D (Rn ) and V ∈ MN×K E (Rn ) then

L(U ∗ V) = (LU) ∗ V = U ∗ (LV) in M J×K D (Rn ) . (10.1.15)
Of course, similar equalities hold when

U ∈ M M×N E (Rn ) and V ∈ MN×K D (Rn ) . (10.1.16)
Definition 10.1. Given an M × N system of order m ∈ N,

L = L(∂) = Aα ∂α , Aα = aαjk 1≤ j≤M ∈ M M×N C ∞ (Rn ) , (10.1.17)
|α|≤m 1≤k≤N
define

L(ζ) := ζ α Aα ∈ M M×N C ∞ (Rn ) , ∀ ζ ∈ Cn . (10.1.18)
|α|≤m

Also, call a matrix distribution E ∈ MN×M D (Rn ) a fundamental solution for L in
Rn provided

LE = δI M×M in M M×M D (Rn ) . (10.1.19)
We record next the analogue of Proposition 5.2 and Proposition 5.8 in the current
setting. The reader is advised to recall Remark 5.1.
Proposition 10.2. Suppose L is a constant coefficient M × M system of order m ∈ N,

L = L(∂) := Aα ∂α , Aα = aαjk 1≤ j,k≤M ∈ M M×M (C), (10.1.20)
|α|≤m
with the property that
det (L(iξ)) 0 for every ξ ∈ Rn \ {0}. (10.1.21)
Then the following hold.

(1) If U ∈ M M×K S (Rn ) is such that LU = 0 in M M×K S (Rn ) , then the entries of
U are polynomials in R and LU = 0 ∈ M M×K (C) pointwise in Rn .
n

(2) If L has a fundamental solution E ∈ M M×M S (Rn ) , then any other fundamen-
n
tal solution of L belonging to M M×M S (R ) differs from E by an M × M matrix
P whose entries are polynomials in Rn satisfying
LP = 0 ∈ M M×M (C) pointwise in Rn . (10.1.22)

Proof. If U ∈ M M×K S (Rn ) is such that LU = 0 in M M×K S (Rn ) , taking the

= 0 in M M×K S (Rn ) . In particular,
Fourier transform, we obtain L(iξ)U
= 0 in M M×K D (Rn \ {0}).

L(iξ)U (10.1.23)
Since (10.1.21) holds, the matrix L(iξ) is invertible for every ξ ∈ Rn \ {0}, hence

L(iξ) −1 ∈ M M×M C ∞ (Rn \ {0}) . Based on this, (10.1.23), and (10.1.5), we may
write
= L(iξ)−1 L(iξ)U
U = L(iξ)−1 0 = 0 in M M×K D (Rn \ {0}). (10.1.24)
Hence, U = 0 in M M×K D (Rn \ {0}) which implies supp U ⊆ {0}. Exercise 4.37
(applied to each entry of U) then gives that each component of U is a polynomial in
Rn . In addition, LU = 0 pointwise in Rn . This proves (1).

Suppose U ∈ M M×M S (Rn ) is such that LU = δI M×M in M M×M S (Rn ) . Then
n
L(U − E) = 0 in M M×M S (R ) and the desired conclusion follows by applying
part (1).
We also have a the general Liouville type theorem for system with total symbol
invertible outside the origin that we prove next.
10.1 General Remarks About Vector and Matrix Distributions 359
Theorem 10.3 (A general Liouville type theorem for systems). Assume L is an

M × M system of order m ∈ N of the form (10.1.20) which satisfies (10.1.21). Also,
1
suppose u = (u1 , . . . , u M ) ∈ Lloc (Rn ) M satisfies Lu = 0 in S (Rn ) M and has the
property that there exist N ∈ N0 and C, R ∈ [0, ∞) such that
|u j (x)| ≤ C|x|N whenever |x| ≥ R and j ∈ {1, . . . , M}. (10.1.25)
Then u j is a polynomial in Rn of degree at most N for each j ∈ {1, . . . , M}.

In particular, if u ∈ L∞ (Rn ) M satisfies Lu = 0 in S (Rn ) M then the compo-
nents of u are necessarily constants.
Proof. Since (5.1.10) implies that the locally integrable function u j belongs to
S (Rn ) (cf. Example 4.4) for each j ∈ {1, . . . , M}, Proposition 10.2 implies that all
the entries of u are polynomials in Rn . Moreover, Lemma 5.3 gives that the degree
of these polynomials are at most N.
We conclude this section by noting a couple of useful results.
Proposition 10.4. Assume that

L= Aα ∂α , Aα = aαjk 1≤ j,k≤M ∈ M M×M C) , (10.1.26)
|α|≤m
is an M × M system of order m ∈ N with constant complex coefficients and suppose

that E ∈ M M×M D (Rn ) is a fundamental solution for the system L in Rn . Then for
n
any U ∈ M M×M E (R ) one has

L(E ∗ U) = U in M M×M D (Rn ) . (10.1.27)
Proof. This follows from (10.1.15), (10.1.19), and (10.1.12).

Our last result in this section is the following counterpart of Theorem 7.60 at the
level of systems.
Theorem 10.5. Assume that n ≥ 2, M ∈ N, and consider the M × M system

n
L= A jk ∂ j ∂k , A jk ∈ M M×M (C). (10.1.28)
j,k=1
Then for an M × M matrix-valued function E whose components are contained

in C 2 (Rn \ {0}) ∩ Lloc
1
(Rn ) and have the property that their gradients are positive
homogeneous of degree 1 − n in Rn \ {0}, the following statements are equivalent:
1
(1) When viewed in Lloc (Rn ), the matrix-valued function E is a fundamental solution
for the system L in Rn ;
(2) One has LE = 0 pointwise in Rn \ {0} and
n

ω j A jk ∂k E(ω) dσ(ω) = I M×M . (10.1.29)
j,k=1 S n−1
Proof. This is established much as in the scalar case, following the argument in the
proof of Theorem 7.60, keeping in mind the matrix algebra formalism.
10.2 Fundamental Solutions and Regularity for General Systems
Definition 10.6. Let K be a field, and let (R, +, 0) be a vector space over K equipped
with an additional binary operation from R × R to R, called product (or multiplica-
tion). Then R is said to be a commutative associative algebra over K if
the following identities hold for any three elements a, b, c ∈ R, and any two scalars
x, y ∈ K:
commutativity : ab = ba, (10.2.1)
associativity : (ab)c = a(bc), (10.2.2)
distributivity : (a + b)c = ac + bc, (10.2.3)
compatibility with scalars : (xa)(yb) = (xy)(ab). (10.2.4)
Finally, an element e ∈ R is said to be multiplicative unit provided ea =

ae = a for each a ∈ R.
Let R be a commutative associative algebra over a field K, with multiplicative

unit e and additive neutral element 0. We make the convention that a − b := a + (−b)
if a, b ∈ R and −b is the inverse of b with respect to the addition operation in the
algebra R. Also, for j ∈ N, we define (−1) j a simply a when j is even, and as −a
when j is odd.
Given M ∈ N, we shall let M M×M (R) stand for the collection of M × M matrices

with entries from R. Then, if A = a jk 1≤ j,k≤M ∈ M M×M (R) and c ∈ R, one defines
in the usual fashion the operations

A := ak j 1≤ j,k≤M ∈ M M×M (R), (10.2.5)

c A := c a jk 1≤ j,k≤M ∈ M M×M (R), (10.2.6)

det A := (−1)sgn σ a1σ(1) a2σ(2) · · · a Mσ(M) ∈ R, (10.2.7)
σ∈P M
where P M is the collection of all permutations of the set {1, 2, . . . , M}, and the matrix
of cofactors
adj (A) := (−1) j+k det A jk , (10.2.8)
1≤ j,k≤M
where, for any given j0 , k0 ∈ {1, . . . , M}, the minor A j0 k0 is defined by

A j0 k0 := a jk 1≤ j,k≤M, j j0 , kk0 ∈ M(M−1)×(M−1) (R). (10.2.9)
10.2 Fundamental Solutions and Regularity for General Systems 361

Furthermore, given another matrix B = b jk 1≤ j,k≤M ∈ M M×M (R), we set

M
A ± B := a jk ± b jk 1≤ j,k≤M and A · B := a jr brk . (10.2.10)
1≤ j,k≤M
r=1
A number of basic properties satisfied by these operations are collected in the

next proposition.
Proposition 10.7. Let R be a commutative associative algebra over a field K with

multiplicative unit e and additive neutral element 0. Also, let M ∈ N be arbitrary.
Then the following statements are true:
(i) (A · B) = B · A for all A, B ∈ M M×M (R);
⎛ ⎞
⎜⎜⎜⎜ e 0 ... 0 ⎟⎟⎟⎟
⎜⎜ 0 e ... 0 ⎟⎟⎟
(ii) the identity matrix I M×M := ⎜⎜⎜⎜ ⎟ ∈ M M×M (R) has the property
⎜⎜⎝ . . . . . . . . . . . . ⎟⎟⎟⎟⎠
0 0 ... e
that I M×M · A = A · I M×M = A for each A ∈ M M×M (R);
(iii) det (A ) = det A for each A ∈ M M×M (R);
(iv) det (c A) = c M det A for each A ∈ M M×M (R) and each c ∈ K;
(v) given any A ∈ M M×M (R), for each j, k ∈ {1, . . . , M} we have

M
M
det A = a j (−1) j+ det (A j ) = ak (−1)+k det (Ak ); (10.2.11)
=1 =1
(vi) det A = 0 whenever A ∈ M M×M (R) has either two identical columns, or two
identical rows;

(vii) adj (A ) · A = A · adj (A ) = det A I M×M for each A ∈ M M×M (R);
(viii) det (A · B) = (det A) (det B) for all A, B ∈ M M×M (R);
(ix) adj (A · B) = adj (A) · adj (B) for all A, B ∈ M M×M (R);
(x) M M×M (R) is a (in general, noncommutative) ring, with multiplicative unit I M×M ,
and A ∈ M M×M (R) has a multiplicative inverse in M M×M (R) if and only if det (A)
has a multiplicative inverse in R.
Proof. All properties are established much as in the standard case R ≡ C. Here we
only wish to mention that (vii) is a direct consequence of (v)-(vi) (complete proofs
may be found in, e.g., [36]).
An example that is relevant for our future considerations pertaining to the Lamé
system is as follows. Suppose R is a commutative associative algebra over a field K
and that n ∈ N, λ, μ ∈ K, with μ 0. Also, fix a ∈ Rn (i.e., a = (a1 , . . . , an ) with
a j ∈ R for each j ∈ {1, . . . , n}), and consider A ∈ Mn×n (R) given by

n
A := μ a j a j In×n + (λ + μ)a ⊗ a, (10.2.12)
j=1

where a ⊗ a ∈ Mn×n (R) is defined as a ⊗ a := a j ak 1≤ j,k≤n . Then a direct calculation
(compare with Exercise 10.13) shows that

n n
det A = μn−1 (λ + 2μ) a ja j , (10.2.13)
j=1
and
adj (A) (10.2.14)

n n−2
n
= μn−2 a ja j (λ + 2μ) a j a j In×n − (λ + μ)a ⊗ a .
j=1 j=1
Our main interest in the algebraic framework developed so far in this section lies
with the particular case when, for some fixed n ∈ N,

R := aα ∂α : aα ∈ C, m ∈ N0 , (10.2.15)
α∈Nn0 , |α|≤m
with the natural operations of addition and multiplication, and with the convention
that ∂(0,...,0) = 1. Then clearly R is a commutative associative algebra over C with
multiplicative unit 1 and, for each M ∈ N, the set M M×M (R) consists of all M × M
systems of constant, complex coefficient differential operators. Thus, if M ∈ N,
m ∈ N0 , are fixed an M × M system of constant, complex coefficient differential
operators of degree m has the form L(∂) ∈ M M×M (R) with
jk
L(∂) := aα ∂α . (10.2.16)
1≤ j,k≤M
|α|≤m
We shall call L(∂) a homogeneous (linear) system (of order m) if aαjk = 0 whenever
|α| < m and j, k ∈ {1, . . . , M}.
According to (10.2.5), the algebraic transpose of L(∂) from (10.2.16) in the space
M M×M (R) is given by kj
L(∂) = aα ∂α . (10.2.17)
1≤ j,k≤M
|α|≤m
Next, for each L(∂) as in (10.2.16) set
DL (∂) := det(L(∂)) (10.2.18)
and notice that DL (∂) is a scalar, constant (complex) coefficient, differential operator
of order ≤ Mm.
Consequently, statement (vii) in Proposition 10.7, (10.2.17), and (10.2.18) readily
give that

L(∂) · adj L(∂) = DL (∂)I M×M , ∀ L(∂) ∈ M M×M (R). (10.2.19)
10.2 Fundamental Solutions and Regularity for General Systems 363
Let us also observe that, at the level of symbols,
DL (ξ) = det(L(ξ)) for each ξ ∈ Rn (10.2.20)
and that
DL (∂) is a homogeneous scalar operator
(10.2.21)
whenever L(∂) is a homogeneous system.
In particular, from (10.2.20) and (10.2.21) we deduce that
⎫
if L(∂) is a homogeneous system with ⎪⎬
⎪ =⇒ DL (∂) is elliptic. (10.2.22)
det(L(ξ)) 0 for each ξ ∈ R \ {0} ⎭
n
Our goal is to use (10.2.19) as a link between systems and scalar partial differen-
tial operators. We shall prove two results based on this scheme, the first of which is
a procedure to reduce the task of finding a fundamental solution for a given system
to the case of scalar operators. Specifically, we have the following proposition.
Theorem 10.8. Let L(∂) be an M × M system with constant coefficients as in

(10.2.16), and let DL (∂) be the scalar differential operator associated with L as in
(10.2.18). Then if E ∈ D (Rn ) is a fundamental solution for DL (∂) in Rn , it follows
that

E := adj L(∂) EI M×M ∈ M M×M (D (Rn )) (10.2.23)
is a fundamental solution for the system L(∂) in Rn . Moreover,
E ∈ S (Rn ) =⇒ E ∈ M M×M (S (Rn )). (10.2.24)
In particular, if DL (∂) is not identically zero, then L(∂) has a fundamental solution
that is a tempered distribution.
Proof. Let E be a fundamental solution for DL (∂) and define E as in (10.2.23). That
E is a fundamental solution for L(∂) follows from (10.2.19). Also it is clear that
E ∈ S (Rn ) forces E ∈ M M×M (S (Rn )). As for the last claim in the statement of the
theorem, note that if DL (∂) is not identically zero, then Theorem 5.14 ensures the
existence of a fundamental solution E ∈ S (Rn ) for DL (∂) which, in turn, yields a
solution for L(∂) via the recipe (10.2.23).
Next we record the following consequence of (10.2.19) and (10.2.22). As a
preamble, recall the notion of singular support from (10.1.3).
Theorem 10.9. Assume that L = L(∂) is an M × M homogeneous system in Rn , with

constant complex coefficients, such that

det L(ξ) 0, ∀ ξ ∈ Rn \ {0}. (10.2.25)

Then, for each open set Ω ⊆ Rn and for each u ∈ D (Ω) M one has
sing supp u = sing supp(Lu). (10.2.26)


In particular, if u ∈ D (Ω) M is such that Lu ∈ C ∞ (Ω) M then u ∈ C ∞ (Ω) M .
Proof. Let L be a system as in the statement of the theorem, and define the op-
erator DL (∂) := det[L(∂)]. Then (10.2.22) ensures that DL (∂) is an elliptic, scalar
differential operator, with constant coefficients. Next, fix a vector-valued distri-
M
u = (u1 , . . . , u M ) ∈ D (Ω) and let ω ⊆ Ω be an open set such that
bution
(Lu)ω ∈ C ∞ (ω). Then, for each j ∈ {1, . . . , M} there holds

DL (∂)u j ω = adj L(∂) L(∂)u = adj L(∂) L(∂)uω . (10.2.27)
j ω j
From (10.2.27) it follows that DL (∂)u j ∈ C ∞ (ω) for each j ∈ {1, . . . , M}. This and
Corollary 6.18 imply that u j ω ∈ C ∞ (ω) for each j ∈ {1, . . . , M}. Consequently
ω ⊆ Ω \ sing supp u, from which the left-to-right inclusion in (10.2.26) immediately
follows. The opposite inclusion is readily seen from definitions.
Corollary 10.10. Let L = L(∂) be an M × M homogeneous differential system in

Rn , with constant complex coefficients, with the property that det [L(ξ)] 0 for all
ξ ∈ Rn \ {0}. Then L has a fundamental solution E ∈ M M×M (S (Rn )) which also
satisfies sing supp E = {0}.
Proof. This is an immediate consequence of Theorem 10.8 and Theorem 10.9.
Exercise 10.11. Suppose M ∈ N and consider an M × M homogeneous second-

n
order system with constant complex coefficients L(∂) = arsjk ∂r ∂ s in Rn .
r,s=1 1≤ j,k≤M
Assume that there exists c ∈ (0, ∞) such that this system satisfies the Legendre–
Hadamard ellipticity condition

Re L(ξ)η · η ≥ c|ξ|2 |η|2 , ∀ ξ ∈ Rn , ∀ η ∈ C M . (10.2.28)
Then, |det [L(ξ)]| ≥ c M |ξ|2M for every ξ ∈ Rn . In particular, det [L(ξ)] 0 for every
ξ ∈ Rn \ {0}.
Sketch of proof: Show that
A ∈ M M×M (C), |Aη| ≥ c|η|, ∀ η ∈ C M =⇒ |detA| ≥ c M . (10.2.29)
This may be proved by considering the self-adjoint matrix B := A∗ A (where the

matrix A∗ := (A ) is the adjoint of A) that satisfies (Bη) · η = |Aη|2 ≥ c2 |η|2 for
every η ∈ C M . Use the latter and the fact that B is diagonalizable to conclude that
|detA|2 = detB ∈ [c2M , ∞). To finish the proof of the exercise, apply (10.2.29) with
A := L(ξ), ξ ∈ Rn .
10.3 Fundamental Solutions for the Lamé Operator
The Lamé operator L in Rn is a differential operator that acts on vector distributions.

Specifically, if u ∈ [D (Rn )]n , i.e., u = (u1 , ..., un ) where uk ∈ D (Rn ), k = 1, . . . , n,
10.3 Fundamental Solutions for the Lamé Operator 365
then the action of the Lamé operator on u is defined by
Lu := μΔu + (λ + μ)∇div u

n
= μΔu j + (λ + μ) ∂ j ∂ u ∈ D (Rn ) n , (10.3.1)
1≤ j≤n
=1
where the constants λ, μ ∈ C (typically called Lamé moduli) are assumed to satisfy
μ 0 and 2μ + λ 0. (10.3.2)
To study in greater detail the structure of the Lamé operator, we need to discuss
some useful algebraic formalism.
Definition 10.12. Given, a = (a1 , ..., an ) ∈ Cn , b = (b1 , ..., bn ) ∈ Cn , define a ⊗ b to

be the matrix

a ⊗ b := a j bk 1≤ j,k≤n ∈ Mn×n (C). (10.3.3)
Exercise 10.13. Prove that:

(1) (a ⊗ b) = b ⊗ a for all a, b ∈ Cn ;
(2) Tr (a ⊗ b) = a · b for all a, b ∈ Cn ;
(3) (a ⊗ b)c = (b · c)a for all a, b, c ∈ Cn ;
(4) det (In×n + a ⊗ b) = 1 + a · b for all a, b ∈ Cn ;
(5) (a ⊗ b) (c ⊗ d) = (b · c) a ⊗ d for all a, b, c, d ∈ Cn ;
(6) for every a ∈ Rn and every numbers μ, λ ∈ C the matrix μIn×n +λ a⊗a is invertible
if and only if μ 0 and μ −λ|a|2 ; moreover, whenever μ 0 and μ −λ|a|2 ,
1 λ
μIn×n + λ a ⊗ a −1 = In×n − a⊗a . (10.3.4)
μ μ + λ|a|2
Hint for (4): Fix a = (a1 , ..., an ) ∈ Cn and b = (b1 , ..., bn ) ∈ Cn . By continuity, it
suffices to prove the formula in (4) when a j 0 and b j 0 for every j ∈ {1, ..., n}.
Assuming that this is the case, we may write

1 + a1 b1 a1 b2 · · · a1 bn
a b 1 + a2 b2 · · · a2 bn
det (In×n + a ⊗ b) = det 2 1
· · · ··· · · · · · ·
an b1
an b2 · · · 1 + an bn
1
⎛ n ⎞ a1 + b1 b2 ··· bn
⎜⎜⎜ ⎟⎟⎟ b bn
a2 + b2 ···
1
= ⎜⎜⎝⎜ a j ⎟⎟⎠⎟ det 1
· · · ··· ··· · · ·
j=1 b ···
an + bn
1
1 b2
1
⎛ n ⎞⎛ n ⎞ a1 b1 + 1 1 ··· 1
⎜⎜⎜ ⎟⎟⎟ ⎜⎜⎜ ⎟⎟⎟ 1 1
a2 b2 + 1 ···
1
= ⎜⎜⎝⎜ a j ⎟⎟⎠⎟ ⎜⎜⎝⎜ b j ⎟⎟⎟⎠ det
· · · ··· ··· · · ·
j=1 j=1 1 ···
an bn + 1
1
1
⎛ n ⎞⎛ n ⎞⎛ n ⎞⎛ ⎞
⎜⎜⎜ ⎟⎟⎟ ⎜⎜⎜ ⎟⎟⎟ ⎜⎜⎜ 1 ⎟⎟⎟ ⎜⎜⎜
n ⎟⎟⎟
= ⎜⎜⎝⎜ a j ⎟⎟⎠⎟ ⎜⎜⎝⎜ b j ⎟⎟⎟⎠ ⎜⎜⎜⎝ ⎟⎟⎟ ⎜⎜⎜1 +
⎠ ⎝ a j b j ⎟⎟⎟⎠
j=1 j=1 j=1
a jb j j=1

n
=1+ a j b j = 1 + a · b. (10.3.5)
j=1
The format of the Lamé system (10.3.1) suggests the following definition and
result, shedding light on the conditions imposed on the Lamé moduli in (10.3.2).
Proposition 10.14. Given any Lamé moduli λ, μ ∈ C, define the characteristic

matrix for the Lamé system (10.3.1) as
L(ξ) := μ|ξ|2 In×n + (λ + μ)ξ ⊗ ξ, ∀ ξ ∈ Rn . (10.3.6)
Then the following statements are equivalent:

(1) L(ξ) is invertible for every ξ ∈ Rn \ {0};
(2) L(ξ) is invertible for some ξ ∈ Rn \ {0};
(3) μ 0 and λ + 2μ 0.
Moreover, if μ 0 and λ + 2μ 0, then for each ξ ∈ Rn \ {0} one has
−1 1 λ+μ ξ ξ
L(ξ) = In×n − ⊗ . (10.3.7)
μ|ξ|2 λ + 2μ |ξ| |ξ|
Proof. This is a direct consequence of Exercise 10.13.

Now we are ready to tackle the issue of finding all fundamental solutions for
the Lamé operator (10.3.1) when the Lamé moduli satisfy (10.3.2). In this con-
text, according to Definition 10.1, a fundamental solution for the Lamé operator
is a matrix distribution E ∈ Mn×n (D (Rn )) with the property that LE = δIn×n in

M M×M D (Rn ) . It follows that the columns Ek , k = 1, . . . , n, of E satisfy the equa-
tions

LEk = δ ek in D (Rn ) n for k = 1, . . . , n, (10.3.8)
where ek denotes the unit vector in Rn with 1 on the k-th entry, for each k ∈ {1, . . . , n}.
Our goal is to determine, under the standing assumption n ≥ 3, all the funda-
mental solutions for the Lamé operator with entries in S (Rn ). We do so by relying
on the tools developed for scalar operators. To get started, suppose that there ex-

ists some matrix E = E jk 1≤ j,k≤n ∈ Mn×n (S (Rn )) whose columns satisfy (10.3.8).
Since δ ∈ S (Rn ) and (4.1.33) is true, using (10.3.1), the latter is equivalent with

n
μΔE jk + (λ + μ) ∂ j ∂ Ek = δ jk δ in S (Rn ), j, k ∈ {1, . . . , n}. (10.3.9)
=1
Applying the Fourier transform to each equation in (10.3.9), we further write

n
−μ|ξ|2 Ejk (ξ) − (λ + μ)ξ j ξ E
k (ξ) = δ jk in S (R ),
n
(10.3.10)
=1
for each j, k ∈ {1, . . . , n}. For k arbitrary, fixed, multiply (10.3.10) with ξ j and
then sum up the resulting identities over j ∈ {1, . . . , n} to obtain that, for each
k ∈ {1, . . . , n},

n
n
n
−μ|ξ|2 ξ j Ejk (ξ) − (λ + μ) ξ2j ξ E
k (ξ) = ξk in S (R ),
n
(10.3.11)
j=1 j=1 =1
or equivalently, that

n
−(λ + 2μ)|ξ|2 ξ j Ejk (ξ) = ξk in S (Rn ), k ∈ {1, . . . , n}. (10.3.12)
j=1
If we now multiply (10.3.10) by (λ + 2μ)|ξ|2 and make use of (10.3.12), we may

conclude that, for each j, k ∈ {1, . . . , n},
−μ(λ + 2μ)|ξ|4 Ejk (ξ) = (λ + 2μ)|ξ|2 δ jk − (λ + μ)ξ j ξk in S (Rn ). (10.3.13)
To proceed fix j, k ∈ {1, . . . , n} and note that since n ≥ 3, by Exercise 4.5 we have
ξξ
1
|ξ|2
∈ S (Rn ). Also, |ξ|j 4k ∈ Lloc
1
(Rn ) and in view of Example 4.4 one may infer that
ξ j ξk ξ j ξk
|ξ|4
∈ S (Rn ). In addition, |ξ|4 ∈ L(Rn ), thus |ξ|4 · |ξ|4
, |ξ|4 · |ξ|12 ∈ S (Rn ) (recall (b) in
ξξ
Theorem 4.14), and it is not difficult to check that |ξ| · |ξ|j 4k = ξ j ξk and |ξ|4 · |ξ|12 = |ξ|2
4
in S (Rn ). These conclusions combined with (10.3.13) imply (recall also (10.3.2))

4 δ jk 1 (λ + μ) ξ j ξk
μ(λ + 2μ)|ξ| E jk (ξ) + · − · =0 (10.3.14)
μ |ξ|2 μ(λ + 2μ) |ξ|4
in S (Rn ). Thus, by Proposition 5.2 applied with P(D) := Δ2 , it follows that

δ jk 1 (λ + μ) ξ j ξk
Ejk (ξ) + · 2− · jk (ξ)
=P in S (Rn ), (10.3.15)
μ |ξ| μ(λ + 2μ) |ξ|4
where P jk is a polynomial in Rn satisfying Δ2 P jk = 0 pointwise in Rn . To continue

with the computation of E jk we apply the inverse Fourier transform to (10.3.15) and
use Proposition 4.64 with λ = 2 as well as (7.14.8) to write
! !
δ jk −1 1 (λ + μ) −1 ξ j ξk
E jk = − F + F + P jk
μ |ξ|2 μ(λ + 2μ) |ξ|4
1 δ jk
=− · n−2 + P jk (x)
μ(n − 2)ωn−1 |x|

(λ + μ) 1 δ jk 1 x j xk
+ · − ·
μ(λ + 2μ) 2(n − 2)ωn−1 |x|n−2 2ωn−1 |x|n

(λ + μ) 1 1 δ jk
= · − · n−2
μ(λ + 2μ) 2(n − 2)ωn−1 μ(n − 2)ωn−1 |x|
(λ + μ) x j xk
− · + P jk (x)
2ωn−1 μ(λ + 2μ) |x|n

−1 3μ + λ δ jk (μ + λ)x j xk
= + + P jk (x) (10.3.16)
2μ(2μ + λ)ωn−1 n − 2 |x|n−2 |x|n
in S (Rn ) (hence, in particular, for all x ∈ Rn \ {0}).

Next, we claim that the matrix F = F jk 1≤ j,k≤n ∈ Mn×n S (Rn ) with entries
defined by

−1 3μ + λ δ jk (μ + λ)x j xk
F jk := + , (10.3.17)
2μ(2μ + λ)ωn−1 n − 2 |x|n−2 |x|n
for j, k ∈ {1, . . . , n}, is a fundamental solution for the Lamé operator. Note that, based
on the properties of the Fourier transform, this claim is equivalent with having the
entries of F satisfy

n
−μ|ξ|2 Fjk (ξ) − (λ + μ)ξ j ξ F
k (ξ) = δ jk in S (R )
n
(10.3.18)
=1
for each j, k ∈ {1, . . . , n}. To check (10.3.18) we use (10.3.17), Proposition 4.64
(with λ = n − 2) and (7.14.9) to first write

−1 (3μ + λ)δ jk 1 x j xk
Fjk (ξ) = F + (μ + λ)F
2μ(2μ + λ)ωn−1 n−2 |x|n−2 |x|n
−(3μ + λ)δ jk (n − 2)ωn−1
= ·
2μ(2μ + λ)ωn−1 (n − 2) |ξ|2

μ+λ δ jk ξ j ξk
− ωn−1 2 − 2ωn−1 4
2μ(2μ + λ)ωn−1 |ξ| |ξ|
δ jk 1 (λ + μ) ξ j ξk
=− · + · in S (Rn ). (10.3.19)
μ |ξ|2 μ(λ + 2μ) |ξ|4
Next, we use (10.3.19) to rewrite the term in the left-hand side of (10.3.18) as

n
− μ|ξ|2 Fjk (ξ) − (λ + μ)ξ j ξ F
k (ξ)
=1
n
δ jk (λ + μ)ξ j ξk δk (λ + μ)ξ ξk
= μ|ξ|2
− + (λ + μ)ξ j ξ −
μ|ξ|2 μ(λ + 2μ)|ξ|4 =1
μ|ξ|2 μ(λ + 2μ)|ξ|4

(λ + μ)ξ j ξk ξk (λ + μ)ξk
= δ jk − + (λ + μ)ξ j −
(λ + 2μ)|ξ|2 μ|ξ|2 μ(λ + 2μ)|ξ|2
= δ jk in S (Rn ), (10.3.20)
proving that the matrix F is a fundamental solution for the Lamé operator.
The main result emerging from this analysis is summarized next.
Theorem 10.15. Assume n ≥ 3 and let L be the Lamé operator from (10.3.1) such
that the constants the λ, μ ∈ C satisfy (10.3.2). Define the matrix E = E jk 1≤ j,k≤n
1
with entries given by the Lloc (Rn ) functions

−1 3μ + λ δ jk (μ + λ)x j xk
E jk (x) := + (10.3.21)
2μ(2μ + λ)ωn−1 n − 2 |x|n−2 |x|n

for each x ∈ Rn \ {0} and j, k ∈ {1, . . . , n}. Then E belongs to Mn×n S (Rn ) and is
a fundamental solution for the Lamé operator in Rn .

In addition, any fundamental solution U ∈ Mn×n S (Rn ) of the Lamé operator

in Rn is of the form U = E + P, for some matrix P := P jk 1≤ j,k≤n whose entries
are polynomials in Rn and whose columns, Pk , k = 1, . . . , n, satisfy the pointwise
equations LPk = (0, . . . , 0) ∈ Cn in Rn for k = 1, . . . , n.
Proof. Since the entries of E as given in (10.3.21) are the same as the expressions

from (10.3.17), the earlier analysis shows that E belongs to Mn×n S (Rn ) and is
a fundamental solution for the Lamé operator in Rn . To justify the claim in the
last paragraph of the statement of the theorem we shall invoke Proposition 10.2.
Concretely, since condition (10.3.2) and Proposition 10.14 imply det (L(ξ)) 0
for ξ ∈ Rn \ {0}, it follows that det (L(iξ)) = − det (L(ξ)) 0 for ξ ∈ Rn \ {0}.
This shows that (10.1.21) is satisfied, hence Proposition 10.2 applies and yields the
desired conclusion.

Note that, as Lemma 10.17 shows, if P = P jk 1≤ j,k≤n is as in Theorem 10.15,
then Δ2 P jk = 0 pointwise in Rn for every j, k ∈ {1, . . . , n}.
One may check without using of the Fourier transform that the matrix from The-
orem 10.15 is a fundamental solution for the Lamé operator (much as in the spirit
of Remark 7.4).
Exercise 10.16. Follow the outline below to check, without the use of the Fourier

transform, that the matrix E = E jk 1≤ j,k≤n ∈ Mn×n (S (Rn )), with entries of function
type defined by the functions from (10.3.21), is a fundamental solution for the Lamé
operator in Rn , n ≥ 3.

Step 1. Show that μΔE jk + (λ + μ) n=1 ∂ j ∂ Ek = 0 pointwise in Rn \ {0} for each
j, k ∈ {1, . . . , n}.
Step 2. Show that the desired conclusion (i.e., that the given E is a fundamental
solution for the Lamé operator in Rn , n ≥ 3) is equivalent with the condition that
⎡ ⎤
⎢⎢⎢ n ⎥⎥
lim ⎢⎣μ⎢ E jk (x)Δϕ(x) dx + (λ + μ) Ek (x)∂ j ∂ ϕ(x) dx⎥⎥⎥⎦
ε→0+ |x|≥ε |x|≥ε =1
= ϕ(0)δ jk (10.3.22)
for every j, k ∈ {1, . . . , n} and every ϕ ∈ C0∞ (Rn ).

Step 3. Fix j, k ∈ {1, . . . , n}, ϕ ∈ C0∞ (Rn ), and let R ∈ (0, ∞) be such that supp ϕ ⊆
B(0, R), so that one may replace the domain of integration for the integrals in the left-
hand side of (10.3.22) with {x ∈ Rn : ε < |x| < R}. Use this domain of integration,
(14.8.5), (14.8.4), and the result from Step 1, to prove that (10.3.22) is equivalent
with

∂ϕ ∂E jk
lim+ −μ E jk (x) (x) dσ(x) + μ ϕ(x) (x) dσ(x)
ε→0 ∂B(0,ε) ∂ν ∂B(0,ε) ∂ν

n
− (λ + μ) Ek (x)∂ ϕ(x) ν j (x) dσ(x) (10.3.23)
∂B(0,ε) =1
⎤

n ⎥⎥
+(λ + μ) ν (x)∂ j Ek (x) ϕ(x) dσ(x)⎥⎥⎥⎦ = ϕ(0)δ jk ,
∂B(0,ε) =1
where ν(x) = x
ε for each x ∈ ∂B(0, ε).
Step 4. Prove that there exists a constant C ∈ (0, ∞) independent of ε such that each
of the quantities:
( (
E (x) ∂ϕ ∂E jk
(x)[ϕ(x) − ϕ(0)] dσ(x) ,
∂B(0,ε) jk ∂ν (x) dσ(x) , ∂B(0,ε) ∂ν
(

n
E k (x)∂ ϕ(x) ν (x) dσ(x) , and (10.3.24)
∂B(0,ε) =1 j
(

n
ν (x)∂ E k (x) [ϕ(x) − ϕ(0)] dσ(x) ,
∂B(0,ε) =1 j
is bounded by C∇ϕL∞ (Rn ) ε, thus convergent to zero as ε → 0+ .

Step 5. Combine Steps 2-4 to reduce matters to proving that
n
xs
lim+ μ (∂ s E jk )(x) dσ(x) (10.3.25)
ε→0
s=1
ε
∂B(0,ε)
n
x
+ (λ + μ) (∂ j Ek )(x) dσ(x) = δ jk .
=1
ε
∂B(0,ε)
Step 6. Prove that for every x ∈ ∂B(0, ε) we have
n
xs (3μ + λ)δ jk (μ + λ)(2 − n)x j xk
(∂ s E jk )(x) = − (10.3.26)
s=1
ε 2μ(2μ + λ) ω n−1 εn−1 2μ(2μ + λ) ωn−1 εn+1
and
n
x (3μ + λ) − (μ + λ)(1 − n)
(∂ j Ek )(x) = x j xk
=1
ε 2μ(2μ + λ) ωn−1 εn+1
(λ + μ)δ jk
− · (10.3.27)
2μ(2μ + λ) ωn−1 εn−1
Step 7. Using the fact that (cf. (14.9.45))

εn+1 ωn−1
x j xk dσ(x) = δ jk , (10.3.28)
∂B(0,ε) n
integrate the expressions in (10.3.26)–(10.3.27) to conclude that

n
xs
(∂ s E jk )(x) dσ(x) (10.3.29)
∂B(0,ε) s=1 ε
(μ + λ)(2 − n) δ jk
= 3μ + λ − (10.3.30)
n 2μ(2μ + λ)
and
n
x δ jk
(∂ j Ek )(x) dσ(x) = , (10.3.31)
=1
ε n(2μ + λ)
∂B(x,ε)
then finish the proof of (10.3.25).
10.4 Mean Value Formulas and Interior Estimates for the Lamé
Operator
The goal here is to prove that solutions of the Lamé system satisfy certain mean
value formulas similar in spirit to those for harmonic functions. We start by estab-
lishing a few properties of elastic vector fields.
Lemma 10.17. Let λ, μ ∈ C be such that μ 0 and λ + 2μ 0. Assume that the

vector distribution u = (u1 , u2 , . . . , un ) ∈ D (Ω) n satisfies the Lamé system

μΔu + (λ + μ)∇div u = 0 in D (Ω) n . (10.4.1)

Then u ∈ C ∞ (Ω) n and the following statements are true.
(i) The function div u is harmonic in Ω.
(ii) The function u j is biharmonic in Ω for each j = 1, . . . , n.
(iii) The function ∂ j (div u) is harmonic in Ω for each j = 1, . . . , n.

Proof. The fact that u ∈ C ∞ (Ω) n is a consequence of Theorem 10.9, Proposi-
tion 10.14, and our assumptions on λ, μ. To show (i), we apply div to (10.4.1). Since
div∇ = Δ and divΔ = Δ div we obtain (λ + 2μ)Δdiv u = 0 in Ω, thus (i) follows
since λ + 2μ 0. Now if we return to (10.4.1) and apply Δ to it, the second term
will be zero because of the harmonicity of div u. Consequently, μΔ2 u = 0 in Ω
which proves (ii) since μ 0. Finally, applying Δ to (10.4.1) and using (ii) yields
(λ + μ)Δ(∇div u) = 0, and (iii) follows from this in the case λ + μ 0. If the latter
condition fails, then from (10.4.1) we have that u is harmonic, so (iii) also holds in
this case.
As an auxiliary step in the direction of Theorem 10.19 that contains the mean
value formulas alluded to earlier, we prove the following useful result.
Proposition 10.18. Let λ, μ ∈ C be such that μ 0 and λ + 2μ 0. Assume u =
(u1 , u2 , . . . , un ) ∈ [D (Ω)]n satisfies the Lamé system (10.4.1). Then the components
of u belong to C ∞ (Ω) and the following formulas hold
(μ − λ)r2
u j (x) + ∂ j (div u)(x) (10.4.2)
2μ(n + 2)

n
= 2− (y j − x j )(y − x) · u(y) dσ(y),
r ∂B(x,r)
10.4 Mean Value Formulas and Interior Estimates for the Lamé Operator 373
and
[(n + 3)μ + (n + 1)λ]r2
u j (x)− ∂ j (div u)(x) (10.4.3)
2μ(n + 2)(n − 1)

n
=− − (y j − x j )(y − x) · u(y) − r2 u j (y) dσ(y),
(n − 1)r2 ∂B(x,r)

for every x ∈ Ω, every r ∈ 0, dist(x, ∂Ω) , and every j ∈ {1, . . . , n}.
Proof. The fact that u ∈ [C ∞ (Ω)]n follows from Lemma 10.17. To proceed, fix

some x ∈ Ω, r ∈ 0, dist(x, ∂Ω) , and j ∈ {1, . . . , n}. By (iii) in Lemma 10.17, we
have that ∂ j (div u)(x) is harmonic in Ω. Thus, we may apply the mean value formula
on solid balls for harmonic function (c.f. the first formula in (7.2.9)), followed by
an application of the integration by parts formula (14.8.4), to write

n
∂ j (div u)(x) = ∂ j (div u)(y) dy
ωn−1 rn B(x,r)

n yj − xj
= divu(y) dσ(y)
ωn−1 r ∂B(x,r) r
n

n
= div (y j − x j )u(y) dσ(y)
ωn−1 rn+1 ∂B(x,r)

n
− u j (y) dσ(y) =: I + II. (10.4.4)
Since u j is biharmonic (recall (ii) in Lemma 10.17), we may write (7.4.1) for u j ,
then use the latter formula to simplify II, and then replace Δu j by − λ+μ
μ ∂ j (div u)
(recall that u satisfies the Lamé system) to obtain
n 1 n λ+μ
II = − u j (x) − Δu j (x) = − 2 u j (x) + ∂ j (div u)(x). (10.4.5)
r2 2 r 2μ
Combining (10.4.4) and (10.4.5) we see that
μ − λ n+1
r ∂ j (div u)(x) + nrn−1 u j (x) (10.4.6)
2μ

n
= div (y j − x j )u(y) dσ(y).
ωn−1 ∂B(x,r)

Fix R ∈ 0, dist(x, ∂Ω) . Integrating (10.4.6) with respect to r for r ∈ (0, R), applying
(14.9.5) and then (14.8.4), we arrive at
μ−λ
Rn u j (x) + Rn+2 ∂ j (div u)(x) (10.4.7)
2μ(n + 2)

n
= div (y j − x j )u(y) dy
ωn−1 B(x,R)

n y−x
= (y j − x j ) · u(y) dσ(y)
ωn−1 ∂B(x,R) R
which, in turn, yields (10.4.2) (with R in place of r) after dividing by Rn .

To prove (10.4.3) we start with the term in the right-hand side of formula (10.4.2)
in which we add and subtract r2 u j (y) under the integral sign and then split it in two
integrals. One integral, call it I1 , corresponds to the expression
(y j − x j )(y − x) · u(y) − r2 u j (y). (10.4.8)
The other integral, call it I2 , corresponds to u j (y). For I2 we recall that u j is bihar-
monic and use (7.4.1) after which we replace Δu j (x) by − λ+μ μ ∂ j (div u)(x). Hence,

n
(y j − x j )(y − x) · u(y) dσ(y) (10.4.9)

n
= (y j − x j )(y − x) · u(y) − r2 u j (y) dσ(y)
r2 λ + μ
+ nu j (x) − · ∂ j (div u)(x).
2 μ
Finally, (10.4.3) follows by adding (10.4.2) and (10.4.9).
As mentioned before, Proposition 10.18 is an important ingredient in proving the
following mean value formulas for the Lamé system.
Theorem 10.19. Let λ, μ ∈ C be such that μ 0, λ + 2μ 0, and (n + 1)μ + λ 0.
Assume u ∈ [D (Ω)]n satisfies the Lamé system (10.4.1). Then u ∈ [C ∞ (Ω)]n , and

for every x ∈ Ω and every r ∈ 0, dist(x, ∂Ω) the following formulas hold:

n(λ + μ)(n + 2)
u(x) = − (y − x) · u(y) (y − x) dσ(y)
2[(n + 1)μ + λ]r2 ∂B(x,r)

n(μ − λ)
+ − u(y) dσ(y) (10.4.10)
2[(n + 1)μ + λ] ∂B(x,r)
and

n(λ + μ)(n + 2) y−x y−x
u(x) = − · u(y) dy (10.4.11)
2[(n + 1)μ + λ] B(x,r) |y − x| |y − x|

n(μ − λ)
+ − u(y) dy.
2[(n + 1)μ + λ] B(x,r)
Proof. Once again, that u ∈ [C ∞ (Ω)]n is contained in Lemma 10.17. Formula

(10.4.10) follows by taking a suitable linear combination of (10.4.2) and (10.4.3)
so that ∂ j (div u) cancels (here (n + 1)μ + λ 0 is used). To prove (10.4.11), mul-
tiply (10.4.10) by ωn−1 rn−1 , then move r12 from the front of the first integral in the
1
right-hand side of the equality obtained inside of that integral as |y−x| 2 , then integrate
with respect to r ∈ (0, R), where R ∈ (0, dist(x, ∂Ω)) and use (14.9.5). Finally divide
the very last expression by ωn−1 Rn /n which is precisely the volume of B(x, R). This
gives (10.4.11) written with R in place of r.
We shall now discuss how the mean value formulas from Theorem 10.19 can
be used to obtain interior estimates for solutions of the Lamé system. Two such
versions are proved in Theorem 10.20 and Theorem 10.22 (cf. also Exercise 10.21).
Theorem 10.20 (L1 -Interior estimates for the Lamé operator). Let λ, μ ∈ C be
such that μ 0, λ + 2μ 0, and (n + 1)μ + λ 0. Assume u ∈ [D (Ω)]n is
a vector distribution satisfying the Lamé system (10.4.1). Then u ∈ [C ∞ (Ω)]n and
there exists C ∈ (0, ∞) depending only on n, λ, and μ, such that for every x ∈ Ω,

every r ∈ 0, dist(x, ∂Ω) and each k ∈ {1, . . . , n}, one has

C
|∂k u(x)| ≤ − |u(y)| dy. (10.4.12)
r B(x,r)
Proof. The fact that u ∈ [C ∞ (Ω)]n has been established in Lemma 10.17. Fix k ∈
{1, . . . , n} and observe that since u = (u1 , . . . , un ) satisfies the Lamé system in Ω,

then so does ∂k u. Fix x∗ ∈ Ω arbitrary and select R ∈ 0, dist(x∗ , ∂Ω) . Writing
formula (10.4.11) for u replaced by ∂k u and r replaced by R/2 shows that for each
j ∈ {1, . . . , n},
y − x∗ y j − x∗j
∗ c1 2n n
∂k u j (x ) = ∗
· ∂k u(y) dy (10.4.13)
ωn−1 Rn B(x∗ , R/2) |y − x | |y − x∗ |

c2 2n n
+ ∂k u j (y) dy,
ωn−1 Rn B(x∗ , R/2)
where
n(λ + μ)(n + 2) n(μ − λ)
c1 := and c2 := . (10.4.14)
2[(n + 1)μ + λ] 2[(n + 1)μ + λ]
Integrating by parts (using (14.8.4)) in both integrals in (10.4.13) yields
n (y j − x∗j )(y − x∗ )
∗ c1 2n n
∂k u j (x ) = − ∂yk u (y) dy (10.4.15)
ωn−1 Rn |y − x∗ |2
B(x∗ , R/2) =1

c1 2n n yk − xk∗ n (y − x∗ )(y − x∗ )
j j
+ u (y) dσ(y)
ωn−1 Rn |y − x∗ | =1 |y − x∗ |2
∂B(x∗ , R/2)

c2 2n n yk − xk∗
+ u j (y) dσ(y).
ωn−1 Rn |y − x∗ |
∂B(x∗ , R/2)
Thus,

|u(y)|
|∂k u(x∗ )| ≤ CR−n dy + CR−n |u(y)| dσ(y), (10.4.16)
|y − x∗ |
B(x∗ , R/2) ∂B(x∗ , R/2)
where C stands for a finite positive constant depending only on n, λ, and μ. Before
continuing let us note a useful consequence of (10.4.11). Specifically, for every

x ∈ Ω and every r ∈ 0, dist(x, ∂Ω) one has

|u(x)| ≤ C − |u(z)| dz. (10.4.17)
B(x,r)
Taking y ∈ B(x∗ , R/2) forces B(y, R/2) ⊂ B(x∗ , R) which, when used in estimate
(10.4.17) written for x replaced by y and r by R/2, gives

|u(y)| ≤ C − |u(z)| dz ≤ C − |u(z)| dz, ∀ y ∈ B(x∗ , R/2). (10.4.18)
B(y,R/2) B(x∗ , R)
This, in turn, allows us to estimate the integrals in (10.4.16) as follows. For the
boundary integral, (10.4.18) yields

|u(y)| dσ(y) ≤ C − |u(z)| dz Rn−1
B(x∗ , R)
∂B(x∗ , R/2)

C
= |u(z)| dz, (10.4.19)
R B(x∗ , R)
while for the solid integral

|u(y)| dy
∗
dy ≤ C − |u(z)| dz ∗
|y − x | B(x∗ , R) B(x∗ , R/2) |y − x |
B(x∗ , R/2)

C
≤ |u(z)| dz, (10.4.20)
R B(x∗ , R)
( dy
( R
∗| = C ρn−2 dρ = CRn−1 . Now (10.4.12) (with r replaced by R)

2
since B(x∗ , R/2) |y−x 0
follows from (10.4.16), (10.4.19), and (10.4.20).
Exercise 10.21. Under the same background assumptions as in Theorem 10.20,

prove that for every multi-index α ∈ Nn0 there exists a constant Cα ∈ (0, ∞), that
depends only on n, α, λ, μ, and with the property that for every x ∈ Ω and every

r ∈ 0, dist(x, ∂Ω)

α Cα
|∂ u(x)| ≤ |α| − |u(y)| dy. (10.4.21)
r B(x,r)
Hint: Use induction on |α|, (10.4.12) written for r/2, and the fact that ∂α u continues
to be a solution of the Lamé system (10.4.1)
Theorem 10.22 (L∞ -Interior estimates for the Lamé operator). Let λ, μ ∈ C be
such that μ 0, λ + 2μ 0, and (n + 1)μ + λ 0 and suppose u ∈ [D (Ω)]n satisfies
the Lamé system (10.4.1). Then u ∈ [C ∞ (Ω)]n and, with C as in Theorem 10.20, for

each x ∈ Ω, each r ∈ 0, dist(x, ∂Ω) , and each k ∈ N, we have
C k ek−1 k!
|∂α u(x)| ≤ max |u(y)|, ∀ α ∈ Nn0 with |α| = k. (10.4.22)
rk y∈B(x,r)
Proof. The fact that u ∈ [C ∞ (Ω)]n follows from Lemma 10.17. In particular, this
implies that ∂α u satisfies (10.4.1) for every α ∈ Nn0 . The case k = 1 is an immediate
consequence of (10.4.12) since, clearly,

− |u(y)| dy ≤ max |u(y)|. (10.4.23)
B(x,r) y∈B(x,r)
Having established this, the desired conclusion follows by invoking Lemma 6.21
(with A the class of null solutions for the Lamé system in Ω).
Recall Definition 6.22.
Theorem 10.23. Suppose λ, μ ∈ C are such that
μ 0, λ + 2μ 0, and (n + 1)μ + λ 0. (10.4.24)
Then any null solution of the Lamé system (10.4.1) has components that are real-
analytic in Ω.
Proof. This is an immediate consequence of Theorem 10.22 and Lemma 6.24.
Theorem 10.24. Suppose λ, μ ∈ C satisfy the conditions in (10.4.24) and Ω ⊆ Rn is

open and connected. Assume u is a null solution of the Lamé system (10.4.1) with
the property that ∂α u(x0 ) = 0 for some x0 ∈ Ω and for all α ∈ Nn0 . Then u = 0 in Ω.
Proof. This follows from Theorem 10.23 and Theorem 6.25.

Next we record the analogue of the classical Liouville’s theorem for the Lapla-
cian (cf. Theorem 7.15) in the case of the Lamé system.
Theorem 10.25 (Liouville’s Theorem for the Lamé system). Let λ, μ ∈ C be such
that μ 0, λ + 2μ 0 and suppose u ∈ [L∞ (Rn )]n satisfies the Lamé system

μΔu + (λ + μ)∇div u = 0 in D (Rn ) n . (10.4.25)
Then there exists a constant vector c ∈ Cn such that u(x) = c for all x ∈ Rn .
Proof. This is a particular case of Theorem 10.3 since based on the current assump-
tions on λ and μ, Proposition 10.14 ensures that condition (10.1.21) is satisfied.
Exercise 10.26. Assuming that λ, μ ∈ C satisfy the conditions in (10.4.24), give
an alternative proof of Theorem 10.25 by relying on the interior estimates from
(10.4.12).
10.5 The Poisson Equation for the Lamé Operator
Let L be the operator from (10.3.1) with coefficients μ, λ ∈ C satisfying (10.3.2).

Then the Poisson equation for the Lamé operator in an open subset Ω of Rn reads
Lu = f, (10.5.1)
where the vector f is given and the vector u is the unknown. If u is a priori known
to be of class C 2 , then the equality in (10.5.1) is considered in the pointwise sense,
everywhere in Rn . Such a solution is called classical. Often, one starts with u simply

a vector distribution in which scenario (10.5.1) is interpreted in D (Ω) n . In this
case, we shall refer to u as a distributional solution. In this vein, it is worth pointing
out that if the datum f is of class C ∞ in Ω then any distributional solution of (10.5.1)
is also of class C ∞ in Ω, as seen from Theorem 10.9, Proposition 10.14, and (10.3.2).
A key ingredient in solving the Poisson equation (10.5.1) is going to be the fun-
damental solution for the Lamé system derived in (10.3.21).
Proposition 10.27. Let L be the Lamé operator from (10.3.1) such that (10.3.2) is

satisfied. Assume n ≥ 2 and let E = E jk 1≤ j,k≤n be the fundamental solution for L in
R with entries as in (10.3.21) for n ≥ 3 and as in (10.7.2) for n = 2. Suppose that
n
Ω is an open set in Rn and that f ∈ [L∞ (Ω)]n vanishes outside a bounded subset of
Ω. Then
u(x) = E(x − y)f(y) dy, ∀ x ∈ Ω, (10.5.2)
Ω
is a distributional solution of the Poisson equation for the Lamé system in Ω, i.e.,
Lu = f in [D (Ω)]n . In addition, u ∈ [C 1 (Ω)]n .
Proof. This is established by arguing along the lines of the proof of Proposition 7.8.

10.5 The Poisson Equation for the Lamé Operator 379
The main result in this section is the following well-posedness result for the
Poisson problem for the Lamé operator in Rn .
Theorem 10.28. Assume n ≥ 3, and let L be the Lamé operator from (10.3.1) with
λ, μ ∈ C satisfying μ 0 and λ + 2μ 0. Also, suppose a vector-valued function
∞
f ∈ Lcomp (Rn ) n and c ∈ Cn are given. Then the Poisson problem for the Lamé
operator in Rn , ⎧
⎪
⎪
⎪ u ∈ [C 0 (Rn )]n ,
⎪
⎪
⎨ Lu = f in D (Rn ) n ,
⎪
⎪
⎪
⎪
(10.5.3)
⎪
⎪
⎪
⎩ lim u(x) = c,
|x|→∞
properties.
(1) The function u is of class C 1 in Rn and admits the integral representation formula

u(x) = c + E(x − y)f(y) dy, ∀ x ∈ Rn , (10.5.4)
Rn

where E = E jk 1≤ j,k≤n is the fundamental solution for L in Rn with entries as in
(10.3.21). Moreover, for each j ∈ {1, . . . , n} we have

∂ j u(x) = (∂ j E)(x − y)f(y) dy, ∀ x ∈ Rn . (10.5.5)
Rn

(2) If in fact f ∈ C0∞ (Rn ) n then u ∈ C ∞ (Rn ) n .
(3) For every j, k ∈ {1, . . . , n} there exists a matrix C jk ∈ Mn×n (C) such that

∂ j ∂k u = C jk f + T ∂ j ∂k E f in D (Rn ) n , (10.5.6)
where T ∂ j ∂k E is the singular integral operator associated with the matrix-valued

function Θ := ∂ j ∂k E (cf. Definition 4.93).
(4) For every integrability exponent p ∈ (1, ∞), the solution u of (10.5.3) satisfies

∂ j ∂k u ∈ L p (Rn ) n for each j, k ∈ {1, . . . , n}, where the derivatives are taken
n n
in D (R ) . Moreover, there exists a constant C = C(p, n) ∈ (0, ∞) with the
property that
n
∂ j ∂k u[L p (Rn )]n ≤ Cf[L p (Rn )]n . (10.5.7)
j,k=1
Proof. From Proposition 10.27 we have that u defined as in (10.5.4) is of class C 1

in Rn and satisfies Lu = f in D (Rn ) n . In addition, by reasoning as in the proof
of Proposition 7.8 we also see that formula (10.5.5) holds for each j ∈ {1, . . . , n}.
Furthermore, the same type of estimate as in (7.2.21) proves that the function u from
(10.5.4) also satisfies the limit condition in (10.5.3). This concludes the treatment of
the existence. As for uniqueness, suppose u is a solution of (10.5.3) for f = 0 ∈ Cn .
Given that u satisfies (10.5.3), it follows that u is bounded in Rn . As such, Liou-
ville’s theorem (cf. Theorem 10.25) applies and gives that u = 0. This proves that u
defined as in (10.5.4) is the unique solution of (10.5.3). Next, the regularity result in
part (2) may be seen either directly from (10.5.4), or by relying on Theorem 10.9,
Proposition 10.14, and the assumptions on the Lamé moduli λ, μ. This concludes
the proof of the claims made in parts (1)–(2).
Consider now the claim made in part (3). Fix j, k ∈ {1, . . . , n} arbitrary. Then,
as seen from (10.3.21), the function Φ := ∂k E is C ∞ and positive homogeneous
of degree 1 − n in Rn \ {0}. In turn, this implies that the matrix-valued function
Θ := ∂ j ∂k E has entries satisfying the conditions in (4.4.1) (here the discussion
in Example (4.71) is relevant). From what we have proved in part (1) and Theo-

rem 4.103 we then conclude that, in the sense of D (Rn ) n ,

∂ j ∂k u(x) = ∂ j ∂k u(x) = ∂ j (∂k E)(x − y)f(y) dy (10.5.8)
Rn

= (∂k E)(ω)ω j dσ(ω) f(x)
S n−1

+ lim+ (∂ j ∂k E)(x − y)f(y) dy.
ε→0 |x−y|>ε
(
Upon taking C jk := S n−1 (∂k E)(ω)ω j dσ(ω) ∈ Mn×n (C), formula (10.5.6) follows,
finishing the proof of part (3). Lastly, the claim in (4) is a consequence of (3) and

the boundedness of the singular integral operators T ∂ j ∂k E on L p (Rn ) n (as seen by
applying Theorem 4.101 componentwise).
10.6 Fundamental Solutions for the Stokes Operator

Let u = (u1 , . . . , un ) ∈ D (Rn ) n and p ∈ D (Rn ). Then the Stokes operator LS
acting on (u, p) = (u1 , . . . , un , p) is defined by

n
LS (u, p) := Δu1 − ∂1 p, . . . , Δun − ∂n p , ∂ s u s ∈ D (Rn ) n+1 . (10.6.1)
s=1
In practice, u and p are referred to as the velocity field and the pressure function,
respectively.
A fundamental solution for the Stokes operator is given by a pair (E, p), where

E = E jk 1≤ j,k≤n ∈ Mn×n (D (Rn )), p = (p1 , . . . , pn ) ∈ D (Rn ) n , satisfy the follow-

ing conditions. If for each k ∈ {1, . . . , n} we set Γk := E1k , . . . , Enk , pk and en+1 k
denotes the unit vector in Rn+1 with 1 on the k-th entry, then

LS Γk = δ en+1
k in D (Rn ) n+1 for k = 1, . . . , n. (10.6.2)
We propose to determine all the fundamental solutions for the Stokes operator
with entries in S (Rn ) in the case when n ≥ 3, a condition assumed in this section
10.6 Fundamental Solutions for the Stokes Operator 381
unless otherwise specified. Suppose (E, p) is a fundamental solution for LS with

the property that E = E jk 1≤ j,k≤n ∈ Mn×n (S (Rn )) and p = (p1 , . . . , pn ) belongs to
n n
S (R ) . Then (10.6.2) implies
ΔE jk − ∂ j pk = δ jk δ in S (Rn ), ∀ j, k ∈ {1, . . . , n}, (10.6.3)

n
∂ s E sk = 0 in S (Rn ), ∀ k ∈ {1, . . . , n}. (10.6.4)
s=1
Apply the Fourier transform to each of the equalities in (10.6.3) and (10.6.4) to
obtain
− |ξ|2 Ejk − iξ j pk = δ jk in S (Rn ), ∀ j, k ∈ {1, . . . , n}, (10.6.5)

n
ξ s Esk = 0 in S (Rn ), ∀ k ∈ {1, . . . , n}. (10.6.6)
s=1
Fix k ∈ {1, . . . , n} and, for each j = 1, . . . , n, multiply the identity in (10.6.5) corre-
sponding to this j with ξ j , then sum up over j and use (10.6.6) to arrive at
|ξ|2 pk = iξk in S (Rn ). (10.6.7)
Reasoning as in the derivation of (10.3.15) from (10.3.13), the last identity implies
ξk
pk = i + rk (ξ) in S (Rn ), (10.6.8)
|ξ|2
for some harmonic polynomial rk in Rn . An application of the inverse Fourier trans-
form to (10.6.8) combined with (7.14.6) then yields
ξk 1 xk
pk = iF −1 + rk = − · + rk in S (Rn ). (10.6.9)
|ξ|2 ωn−1 |x|n
In particular, if we take pk = − ωn−1

1
· |x|xkn and use it to substitute for pk back in (10.6.5),
we arrive at the condition
ξ j ξk
|ξ|2 Ejk = − δ jk in S (Rn ), ∀ j ∈ {1, . . . , n}. (10.6.10)
|ξ|2
Consequently, since n ≥ 3, by reasoning as in the derivation of (10.3.15) from
(10.3.13), identity (10.6.10) implies
ξ j ξk δ jk
Ejk = − 2 + Rjk in S (Rn ), ∀ j ∈ {1, . . . , n}, (10.6.11)
|ξ|4 |ξ|
for some polynomials R jk in Rn satisfying ΔR jk = 0 in Rn , j = 1, . . . , n. Taking
the Fourier transform in (10.3.13), then using (7.14.8) and Proposition 4.64 with
λ = n − 2, we obtain
ξ j ξk 1
E jk = F −1 − F −1 δ jk + R jk
|ξ|4 |ξ|2
1 δ jk 1 x j xk
=− − + R jk in S (Rn ). (10.6.12)
2(n − 2)ωn−1 |x|n−2 2ωn−1 |x|n
We are now ready to state our main result regarding fundamental solutions for
the Stokes operator that are tempered distributions.
Theorem 10.29. Let n ≥ 3 and let LS be the Stokes operator from (10.6.1). Consider
1
the following functions in Lloc (Rn ):
1 δ jk 1 x j xk
E jk (x) := − − , ∀ j, k ∈ {1, . . . , n}, (10.6.13)
2(n − 2)ωn−1 |x|n−2 2ωn−1 |x|n
1 xk
pk (x) := − , ∀ k ∈ {1, . . . , n}, (10.6.14)
ωn−1 |x|n

defined for x ∈ Rn \ {0}. Then if we set E := E jk 1≤ j,k≤n and p := (p1 , . . . , pn ),
n n n
we have E ∈ Mn×n S (R ) , p ∈ S (R ) , and the pair (E, p) is a fundamental
solution for the Stokes operator in Rn .
Moreover, any fundamental solution (U, q) for the Stokes operator with the prop-

erty that U ∈ Mn×n S (Rn ) , q ∈ S (Rn ) n is of the form U = E + P, q = p + r,

for some matrix P := P jk 1≤ j,k≤n whose entries are polynomials in Rn satisfying
Δ2 P jk = 0 pointwise in Rn for j, k = 1, . . . , n, and some vector r := (r1 , . . . , rn )
whose entries are polynomials in Rn satisfying Δrk = 0 pointwise in Rn for
k = 1, . . . , n.

Proof. From (10.6.12) and (10.6.9) we have E ∈ Mn×n S (Rn ) , p ∈ S (Rn ) n , and
ξ j ξk δ jk ξk
Ejk = − 2, pk = i , in S (Rn ), ∀ j, k ∈ {1, . . . , n}. (10.6.15)
|ξ|4 |ξ| |ξ|2
Based on the properties of the Fourier transform, we have that (E, p) is a fundamen-
tal solution for the Stokes operator if and only if its components satisfy (10.6.5)–
(10.6.6). By making use of (10.6.15) we may write
ξ j ξk ξ j ξk
−|ξ|2 Ejk − iξ j pk = − + δ jk + 2 = δ jk in S (Rn ), ∀ j, k ∈ {1, . . . , n}
|ξ|2 |ξ|
(10.6.16)
and

n
n ξ2 ξ
j k

n
δ jk ξ j
ξ j Ejk = − = 0 in S (Rn ), (10.6.17)
j=1 j=1
|ξ|4 j=1
|ξ|2
proving that (E, p) is a fundamental solution for the Stokes operator.

Suppose (U, q) is another fundamental solution for the Stokes operator such that

U = U jk 1≤ j,k≤n ∈ Mn×n S (Rn ) and q = (q1 , . . . , qn ) ∈ S (Rn ) n . Then for each
k ∈ {1, . . . , n} we have qk = pk +rk for some harmonic polynomial rk in Rn (recall the
10.6 Fundamental Solutions for the Stokes Operator 383
computation that lead to (10.6.9)). From this fact and the equations corresponding
to (10.6.3) written for the components of (E, p) and (U, q), we further obtain that
Δ(U jk − E jk ) = ∂ j rk in S (Rn ) for each j, k ∈ {1, . . . , n}. (10.6.18)
Fix j, k ∈ {1, . . . , n} arbitrary. Then by applying Δ to both sides of the equation in

(10.6.18) it follows that Δ2 (U jk − E jk ) = 0 in S (Rn ). Consequently, after taking the
Fourier transform of the latter identity we arrive at |ξ|4 (U ,
jk − E jk ) = 0 in S (R ).
n
In turn, this implies U jk − E jk = R jk for some polynomials R jk in R satisfying

n
Δ2 R jk = 0 pointwise in Rn . The proof of the theorem is now complete.

Exercise 10.30. Follow the outline below to check, without the use of the Fourier

transform, that (E, p) with entries E = E jk 1≤ j,k≤n and p = (p1 , . . . , pn ) of function
type defined by the functions from (10.6.13)–(10.6.14) is a fundamental solution for
the Stokes operator in Rn , n ≥ 3.

n
Step 1. Show that ΔE jk − ∂ j pk = 0 and ∂ Ek = 0 pointwise in Rn \ {0} for every
=1
j, k ∈ {1, . . . , n}.
Step 2. Show that the desired conclusion (i.e., that the given (E, p) is a fundamental
solution for the Stokes operator in Rn , n ≥ 3) is equivalent with the conditions that

lim+ E jk (x)Δϕ(x) dx + pk (x)∂ j ϕ(x) dx = ϕ(0)δ jk (10.6.19)
ε→0 |x|≥ε |x|≥ε
and ⎡ ⎤
⎢⎢
n ⎥⎥
lim+ ⎢⎢⎢⎣ Ek (x)∂ ϕ(x) dx⎥⎥⎥⎦ = 0 (10.6.20)
ε→0 |x|≥ε =1
for every j, k ∈ {1, . . . , n} and every ϕ ∈ C0∞ (Rn ).

Step 3. Fix j, k ∈ {1, . . . , n}, ϕ ∈ C0∞ (Rn ), and let R ∈ (0, ∞) be such that
supp ϕ ⊆ B(0, R), so that one may replace the domain of integration for the inte-
grals in (10.6.19) and (10.6.20) with {x ∈ Rn : ε < |x| < R}. Use this domain of
integration, (14.8.5), (14.8.4), and the result from Step 1, to prove that (10.6.19) and
(10.6.20) are equivalent with

∂ϕ ∂E jk
lim − E jk (x) (x) dσ(x) + ϕ(x) (x) dσ(x)
ε→0+ ∂B(0,ε) ∂ν ∂B(0,ε) ∂ν

− pk (x)ϕ(x)ν j (x) dσ(x) = ϕ(0)δ jk (10.6.21)
∂B(0,ε)
and ⎡ ⎤
⎢⎢
n ⎥⎥
lim+ ⎢⎢⎢⎣ Ek (x)ϕ(x)νl (x) dσ(x)⎥⎥⎥⎦ = 0, (10.6.22)
ε→0 ∂B(0,ε) =1
where ν(x) = x
ε for each x ∈ ∂B(0, ε).
Step 4. Prove that there exists a constant C ∈ (0, ∞) independent of ε such that each
of the quantities:

∂ϕ
E jk (x) (x) dσ(x) ,
∂B(0,ε) ∂ν

∂E jk
(x)[ϕ(x) − ϕ(0)] dσ(x) , (10.6.23)
∂B(0,ε) ∂ν

pk (x)[ϕ(x) − ϕ(0)]ν j (x) dσ(x) ,
∂B(0,ε)
is bounded by C∇ϕL∞ (Rn ) ε, thus convergent to zero as ε → 0+ , and such that

n
Ek (x)ϕ(x)νl (x) dx ≤ CϕL∞ (Rn ) ε −−−−→ 0. (10.6.24)
∂B(0,ε) =1 ε→0+
Step 5. Combine Steps 2, 3, and 4 to reduce matters to proving that
n
xs

xj
lim+ (∂ s E jk )(x) dσ(x) − pk (x) dσ(x) = δ jk . (10.6.25)
ε→0
s=1
ε ε
∂B(0,ε) ∂B(0,ε)
Step 6. Prove that for every x ∈ ∂B(0, ε) we have

n
xs xj δ jk nx j xk
(∂ s E jk )(x) − pk (x) = + . (10.6.26)
s=1
ε ε 2 ωn−1 ε n−1 2 ωn−1 εn+1
Step 7. Integrate the expression in (10.6.26) and use (10.3.28) to finish the proof of
(10.6.25).
Further Notes for Chapter 10. The discussion in Chapters 5–6 about the existence and nature
of fundamental solutions has been limited to the case of scalar differential operators, and in Sec-
tion 10.2 these scalar results have been extended to generic constant coefficient systems (of arbi-
trary order) via an approach that appears to be new. Moreover, this approach offers further options
for finding an explicit form for a fundamental solution for a given system, such as the Lamé sys-
tem. While we shall pursue this idea later, in Section 11.1, we felt it is natural and beneficial to
first deal with the issue of computing a fundamental solution for the Lamé and Stokes systems via
Fourier analysis (as done in Section 10.3 and Section 10.6, respectively).
The inclusion of a section on mean value formulas for the Lamé operator is justified by the
fact that such formulas directly yield interior estimates, without resorting to quantitative elliptic
regularity, which is typically formulated in the language of Sobolev spaces. In turn, these interior
estimates play a pivotal role in establishing uniqueness for the corresponding Poisson problem. A
more systematic treatment of the issue of mean value formulas for the Lamé operator may be found
in [56].
Exercise 10.31. Fix m, N, M ∈ N and assume

L = L(∂) = Aα ∂α , Aα = aαjk 1≤ j≤M ∈ M M×N (C), (10.7.1)
1≤k≤N
|α|≤m
is a constant, complex coefficient, homogeneous system of order m. Make the as-

sumption that L(ξ) ∈ M M×N (C) is a surjective linear map from CN → C M for each
ξ ∈ Rn \ {0}. Follow the outline below to construct a fundamental solution for L.
(1) Consider L := LL∗ . Show that L is a homogeneous constant coefficient M × M

system of order 2m and L(ξ) = L(ξ)L(ξ)∗ in M M×N (C) for each ξ ∈ Rn .
(2) Show that L(ξ) is injective for each ξ ∈ Rn \ {0} and use this to conclude that
det[L(ξ)] 0 for each ξ ∈ Rn \ {0}.
(3) Corollary 10.10 guarantees the existence of a matrix-valued fundamental solution
E ∈ M M×M (S (Rn )) for L. Show that E := L∗ E is a fundamental solution for L.
Exercise 10.32. Let LS be the Stokes operator from (10.6.1). Also, suppose p in

D (Ω) and u = u1 , . . . , un ) ∈ D (Ω) n satisfy LS (u, p) = 0 in D (Ω) n+1 . Under
∞
these conditions prove that p, u j ∈ C (Ω) for j ∈ {1, . . . , n}, and Δp = 0 pointwise
in Ω while Δ2 u j = 0 pointwise in Ω for j ∈ {1, . . . , n}.
Exercise 10.33. For each x ∈ R2 \ {0} and j, k ∈ {1, 2} consider the functions

1 (μ + λ)x j xk
E jk (x) := (3μ + λ)δ jk ln |x| − . (10.7.2)
4πμ(2μ + λ) |x|2

Prove that E := E j,k is a fundamental solution for the Lamé operator in R2 .
1≤ j,k≤2
Exercise 10.34. Consider the functions

1 1 x j xk 1 xk
E jk (x) := δ jk ln |x| − · , ·
pk (x) := − , (10.7.3)
4π 4π |x|2 2π |x|2

defined for x ∈ R2 \ {0} and j, k ∈ {1, 2}. Prove that if E := E j,k and p =
1≤ j,k≤2
(p1 , p2 ), then (E, p) is a fundamental solution for the Stokes operator in R . 2
Exercise 10.35. Assume n ≥ 3, let L be the Lamé operator from (10.3.1) such that

(10.3.2) is satisfied and consider f = ( f1 , . . . , fn ) ∈ C0∞ (Rn ) n . Prove that for each
A ∈ Mn×n (R) and each b ∈ Rn , there exists a unique solution for the problem
⎧
⎪
⎪
⎪ u = (u1 , . . . , un ) ∈ [C ∞ (Rn )]n ,
⎪
⎪
⎪
⎨ Lu = f pointwise in Rn ,
⎪
⎪
⎪
(10.7.4)
⎪
⎪
⎪
⎩ |x|→∞
lim |u(x) − Ax − B| = 0.
Exercise 10.36. Assume n ≥ 3, let L be the Lamé operator from (10.3.1) such that
(10.3.2) is satisfied. Prove that if m ∈ N0 and u is a solution of the problem
⎧
⎪
⎪
⎪ u ∈ [C ∞ (Rn )]n ,
⎪
⎪
⎨
⎪
⎪ Lu = 0 pointwise in Rn , (10.7.5)
⎪
⎪
⎪
⎩ |u(x)| ≤ C(|x| + 1),
m
for some C ∈ (0, ∞), then the components of u are polynomials of degree ≤ m.
Chapter 11
More on Fundamental Solutions
for Systems
Abstract The issue of identifying fundamental solutions for homogeneous constant

coefficient systems of arbitrary order is a central topic here. As particular cases of
the approach is developed, fundamental solutions that are tempered distributions for
the Lamé and Stokes operators are derived. The fact that integral representation for-
mulas and interior estimates hold for null-solutions of homogeneous systems with
nonvanishing full symbol is also proved. As a consequence, null-solutions are real-
analytic and shown to satisfy reverse Hölder estimates. Finally, layer potentials as-
sociated with arbitrary constant coefficient second- order systems in the upper-half
space, and the relevance of these operators vis-a-vis to the solvability of boundary
value problems for such systems in this setting, are discussed.
11.1 Computing a Fundamental Solution for the Lamé Operator
In this section, we use Theorem 10.8 in order to find a fundamental solution for the
Lamé operator of elastostatics in Rn , with n ≥ 2. This is possible since in the case
when L(∂) is the Lamé operator we have that det(L(∂)) is a constant factor of Δn .
Concretely, fix n ∈ N, n ≥ 2, and consider the algebra R as in (10.2.15). Fix
constants λ, μ ∈ C satisfying μ 0, λ + 2μ 0, and set
L(∂) := μΔ + (λ + μ)∇div ∈ Mn×n (R), (11.1.1)
(recall that ∇ = (∂1 , . . . , ∂n ) and the operator div was defined in Exercise 7.74). This
is precisely the Lamé operator from (10.3.1) and L(∂) has the form (10.2.12), with
a := ∇. Thus, for this choice of a, formula (10.2.13) gives
P(∂) := detL(∂) = μn−1 (λ + 2μ)Δn , (11.1.2)
while (10.2.14) implies

adj [L(∂)] = μn−2 (λ + 2μ)ΔIn×n − (λ + μ)∇div Δn−2 . (11.1.3)

388 11 More on Fundamental Solutions for Systems
Going further, for each m ∈ N, denoted by EΔm , the fundamental solution for the poly-
harmonic differential operator Δm in Rn from (7.5.2) and apply Proposition 7.29 to
conclude that

Δn−2 EΔn = EΔ2 + c(n) in S (Rn ) and pointwise in Rn \ {0}, (11.1.4)
where c(n) = 0 if n 4 and c(4) ∈ R. A direct computation starting with (7.5.2)

corresponding to n = 4 = m and using Lemma 7.20 yields c(4) = − 24π 5
2 . Thus, a
combination of (11.1.4) and (7.3.8) yields

⎧
⎪
⎪
⎪ |x|4−n
⎪
⎪
⎪ if n ≥ 3, n 4,
⎪
⎪
⎪ 2(n − 2)(n − 4)ωn−1
⎪
⎪
⎪
⎪
⎪
⎨ 1
Δn−2 EΔn (x) = ⎪ 5
⎪
⎪ − 2 · ln |x| − if n = 4, (11.1.5)
⎪
⎪
⎪ 8π 24π2
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎩
1
· |x|2 ln |x| if n = 2,
8π
both for each x ∈ Rn \ {0} as well as in S (Rn ). Also, by applying one more Δ to the
identity in (11.1.4) and recalling (7.1.12), we have
⎧
⎪
⎪
⎪
1
⎪
⎪ − · |x|2−n if n ≥ 3,
⎪
⎨ (n − 2)ω n−1
Δn−1 EΔn (x) = EΔ (x) = ⎪
⎪ (11.1.6)
⎪
⎪
⎪ 1
⎪
⎩ · ln|x| if n = 2,
2π
for every x ∈ Rn \ {0} and in S (Rn ).
Hence, based on (10.2.23) and (11.1.2), we obtain that a fundamental (matrix)
solution EL for the operator L(∂) as in (11.1.1) is given by

EL (x) = adj [L(∂)] Eμn−1 (λ+2μ)Δn (x)In×n , x ∈ Rn \ {0}, (11.1.7)
where Eμn−1 (λ+2μ)Δn is a fundamental solution for P(∂) := μn−1 (λ + 2μ)Δn . Using
(11.1.3) and (11.1.4) in the right-most expression in (11.1.7) further gives that, for
each x ∈ Rn \ {0},
μn−2
EL (x) = (λ + 2μ)ΔIn×n − (λ + μ)∇div Δn−2 EΔn (x)In×n
μn−1 (λ
+ 2μ)

1 λ+μ
= δ jk EΔ (x) − (∂ j ∂k EΔ2 )(x) . (11.1.8)
μ μ(λ + 2μ) 1≤ j,k≤n
A quick inspection of (7.3.8) and the last equality in (11.1.6) shows that both for
each x = (x1 , . . . , xn ) ∈ Rn \ {0} as well as in S (Rn ), for each j, k in {1, . . . , n} we
have
11.2 Computing a Fundamental Solution for the Stokes Operator 389
⎧
⎪
⎪ δ jk 1 x j xk
⎪
⎪
⎪ EΔ (x) + · n if n ≥ 3,
⎪
⎪
⎨ 2 2ω n−1 |x|
∂ j ∂k EΔ2 (x) = ⎪
⎪ (11.1.9)
⎪
⎪
⎪ δ jk 1 x j xk δ jk
⎪
⎪
⎩ EΔ (x) + · n + , if n = 2.
2 2ωn−1 |x| 8
δ
Note that, for the purpose of computing E L , the additive constant 8jk from (11.1.9)
may be dropped. Thus, based on (11.1.8) and (11.1.9) (with δ jk /8 dropped), it fol-
lows that
3μ + λ
EL (x) = EΔ (x)In×n (11.1.10)
2μ(2μ + λ)
λ+μ x x
− · ⊗ , ∀ x ∈ Rn \ {0},
2μ(2μ + λ)ωn−1 |x|n/2 |x|n/2
is a tempered distribution that is a fundamental solution for the operator Lamé oper-
ator (11.1.1) (with μ 0, λ + 2μ 0) in Rn . Note that this formula coincides with
the one from (10.2.21) if n ≥ 3 and with the one from (10.7.2) if n = 2.
11.2 Computing a Fundamental Solution for the Stokes Operator
In this section, the goal is to compute a fundamental solution for the Stokes operator
in Rn starting from the explicit expression for the fundamental solution for the oper-
ator Lamé operator EL from (11.1.10). To this end, fix n ∈ N, n ≥ 2 and λ, μ ∈ C
satisfying μ 0, λ + 2μ 0, and let x = (x1 , . . . , xn ) for x ∈ Rn .
For each k ∈ {1, . . . , n}, denote by EkL the k-th column of the fundamental solution
E from (11.1.10). Then a straightforward calculation gives that for each x ∈ Rn \{0}
L
and in S (Rn ), there holds

1 xk
div [EkL (x)] = · , ∀ k ∈ {1, . . . , n}. (11.2.11)
(λ + 2μ)ωn−1 |x|n
Define
1 xk
pk (x) := − · , ∀ x ∈ Rn \ {0}, ∀ k ∈ {1, . . . , n}. (11.2.12)
ωn−1 |x|n
Then, from (11.2.11) it follows that for each k ∈ {1, . . . , n} we have
lim div [EkL ] = 0 pointwise in Rn \ {0} and in S (Rn ), (11.2.13)

λ→∞
and

lim −(λ + μ)div[EkL ] = pk pointwise in Rn \ {0} and in S (Rn ). (11.2.14)
λ→∞
In addition, from (11.1.10) we obtain

1 S
lim EL = E pointwise in Rn \ {0} and in S (Rn ), (11.2.15)
λ→∞ μ
where, we have set
δ jk 1 x j xk
ES (x) := · EΔ (x) − · n , ∀ x ∈ Rn \ {0}, (11.2.16)
2 2ωn−1 |x| 1≤ j,k≤n
with EΔ as in (11.1.6). Consequently, a combination of (11.2.14) and (11.2.16) (in
view of the definition (11.1.1) of the Lamé operator), implies

n
ΔE Sjk − ∂ j pk = lim μE Ljk + (λ + μ) ∂ j ∂ s E sk
L
λ→∞
s=1

= δ jk δ in S (R ), ∀ j, k ∈ {1, . . . , n}.
n
(11.2.17)
Also, (11.2.15) and (11.2.13), with the convention that ESk stands for the k-th column
in the matrix ES from (11.2.16), give

div [ESk ] = lim div [EkL ] = 0 in S (Rn ), ∀ k ∈ {1, . . . , n}. (11.2.18)
λ→∞
A quick inspection of (11.2.16) and (11.2.12), shows that the entries in the matrix
ES and the components of the vector p := (p1 , . . . , pn ) belong to Lloc
1
(Rn ), while
(11.2.17) guarantees that
⎧ S n
⎪
⎨ ΔE − ∇p = δ0 In×n , in Mn×n S (R ) ,
⎪
⎪
⎪
⎩ div ES = 0, in [S (Rn )]n .
(11.2.19)
Recalling now the definition of the Stokes operator from (10.6.1), we may conclude
that
(ES , p) is a fundamental solution for the Stokes operator in Rn . (11.2.20)
Note that the expressions we obtained for (ES , p) as given in (11.2.16) and (11.2.12)
are the same as the ones from (10.6.13)–(10.6.14) if n ≥ 3 and as the one from
(10.7.3) if n = 2.
11.3 Fundamental Solutions for Higher Order Systems
In this section, we discuss an approach for computing all fundamental solutions that
are tempered distributions for a certain subclass of systems of the form 10.2.16. To
specify this class, let us before n ∈ N denote the Euclidean dimension, fix M, m ∈ N,
and consider M× M systems of homogeneous differential operators of order 2m with
11.3 Fundamental Solutions for Higher Order Systems 391
complex constant coefficients that have the form

jk
L = L(∂) := aα ∂α =: L jk (∂) 1≤ j,k≤M (11.3.1)
1≤ j,k≤M
|α|=2m
where aαjk ∈ C, j, k ∈ {1, ..., M}. Define the characteristic matrix of L to be

jk
L(ξ) := aα ξα , ∀ ξ ∈ Rn . (11.3.2)
1≤ j,k≤M
|α|=2m
The standing assumption under which we will identify all fundamental solutions for
L that are tempered distributions is

det L(ξ) 0, ∀ ξ ∈ Rn \ {0}. (11.3.3)
Several types of operators discussed earlier fall under the scope of these specifi-
cations, including the class of strictly elliptic, homogeneous, second order, and con-
stant coefficient operators from Section 7.12, the polyharmonic operator from Sec-
tion 7.5, and the Lamé operator from Section 10.3. Also recall that, as seen in Corol-
lary 10.10, assumption (11.3.3) guarantees that L has a fundamental solution that is
a tempered distribution.
For any operator L as in (11.3.1), its transpose and complex conjugate are, re-
spectively, given by

L := aαjk ∂α , L= aαjk ∂α , (11.3.4)
1≤k, j≤M 1≤ j,k≤M
|α|=2m |α|=2m
while the formal adjoint of L is defined as

L∗ := L . (11.3.5)
The main result in this section is Theorem 11.1 below, describing the nature and
properties of all fundamental solutions for a higher order system L as in (11.3.1)–
(11.3.3) that are tempered distributions. Before presenting this basic result, we first
describe a strategy that points to a natural candidate for a fundamental solution
for L.
Suppose that Q is a scalar constant coefficient differential operator (of an aux-
iliary nature) which, from other sources of information, is known to have a funda-
mental solution in Rn of the form

E Q (x) = F(x · ξ) dσ(ξ), (11.3.6)
S n−1
where F is a sufficiently regular scalar-valued function on the real line. Granted this,
we then proceed to define

E L (x) := Q G(x · ξ) L(ξ) −1 dσ(ξ) , (11.3.7)
S n−1
where G is a scalar-valued function on the real line, chosen in such a way that
p(t) := G(2m) (t) − F(t) is a polynomial in t ∈ R of degree < order Q. (11.3.8)
Then
P(x) := p(x · ξ) dσ(ξ), x ∈ Rn , (11.3.9)
S n−1
is a polynomial in Rn of degree < order Q. Keeping this in mind and observing that

LQ = QL and L G(x · ξ) = G(2m) (x · ξ)L(ξ), (11.3.10)
we may compute

L E L (x) = Q G(2m) (x · ξ)L(ξ) L(ξ) −1 dσ(ξ)
S n−1

=Q F(x · ξ) dσ(ξ) + P(x) I M×M
S n−1
= Q[E Q (x)]I M×M + (QP)(x)I M×M = δI M×M + 0 I M×M
= δI M×M , (11.3.11)
which shows that E L is a fundamental solution for the system L in Rn .

To implement the procedure just described, it is natural to take the polyharmonic
operator ΔN to play the role of the auxiliary scalar differential operator Q. This is
not just a matter of convenience, but from Lemma 7.31 we already know that if
N ∈ N satisfies N ≥ n/2 then there exists a fundamental solution for this operator
that has the form requested in (11.3.6). Specifically, in this case we have (compare
with (7.5.24))
1
F(t) := − t2N−n log t/i . (11.3.12)
(2πi)n (2N − n)!
To find a function G on the real line that satisfies (11.3.8), we may try

G(t) := a t A log t/i , (11.3.13)
where a ∈ R, A ∈ N with A ≥ 2m, and note there exists a constant C A,a,m such that
d 2m A!
G(t) = a t A−2m log t/i + C A,a,m t A−2m , (11.3.14)
dt (A − 2m)!
which means that (11.3.8) is going to hold if we take
1
A := 2N − n + 2m and a := − , (11.3.15)
(2πi)n (2N − n + 2m)!
for a polynomial p on R of degree 2N − n, which is strictly less than the order of

Q = ΔN .
To bring matters more in line with notation employed in the proof of Theo-
rem 11.1, pick now a number q ∈ N0 with the same parity as n, and choose
N := (n + q)/2 (which satisfies N ∈ N and N ≥ n/2). For this choice of N,
n+q
the operator Q becomes Δ 2 , (11.3.16)
and if A, a are as in (11.3.15) the function G from (11.3.13) takes the form
1
G(t) = − t2m+q log t/i . (11.3.17)
(2πi)n (2m + q)!
For these specifications, the function E L from (11.3.7) then becomes precisely the
function
x · ξ
−Δ(n+q)/2
E(x) := x
(x · ξ)2m+q log L(ξ) −1 dσ(ξ). (11.3.18)
(2πi) (2m + q)! S n−1
n i
Having explained the genesis of this formula, we now proceed to rigorously show
that not only is this candidate for a fundamental solution natural but it also does job.
Theorem 11.1. Fix n, m, M ∈ N with n ≥ 2, and assume L is an M × M system in

Rn of order 2m as in (11.3.1)–(11.3.3). Then the M × M matrix E defined at each
x ∈ Rn \ {0} by

Δ(n−1)/2
E(x) := x
(x · ξ)2m−1 sgn (x · ξ) L(ξ) −1 dσ(ξ) (11.3.19)
4(2πi)n−1 (2m − 1)! S n−1
if n is odd, and

−Δn/2
E(x) := x
(x · ξ)2m ln |x · ξ| L(ξ) −1 dσ(ξ) (11.3.20)
(2πi)n (2m)! S n−1
if n is even, satisfies the following properties.

(1) One has

E ∈ M M×M S (Rn ) ∩ M M×M C ∞ (Rn \ {0})
1
∩ M M×M Lloc (Rn ) (11.3.21)
and
E(−x) = E(x) for all x ∈ Rn \ {0}. (11.3.22)
Moreover, the entries in E are real-analytic functions in R \ {0}. n
(2) If I M×M is the M × M identity matrix, then for each y ∈ Rn one has

L x E(x − y) = δy (x) I M×M in M M×M S (Rn ) , (11.3.23)
where the superscript x denotes the fact that the operator L in (11.3.23) is applied
to each column of E in the variable x.
(3) Define the M × M matrix-valued function

−1
P(x) := (x · ξ)2m−n L(ξ) −1 dσ(ξ), ∀ x ∈ Rn . (11.3.24)
(2πi)n (2m − n)! S n−1
Then, the entries of P are identically zero when either n is odd or n > 2m, and are
homogeneous polynomials of degree 2m − n when n ≤ 2m. Moreover, there exists

a function Φ ∈ M M×M C ∞ (Rn \ {0}) that is positive homogeneous of degree
2m − n such that

E(x) = Φ(x) + ln |x| P(x), ∀ x ∈ Rn \ {0}. (11.3.25)
As a consequence,
if either n is odd or n > 2m then E is positive homoge-
(11.3.26)
neous of degree 2m − n in Rn \ {0}.
(4) For each β ∈ Nn0 with |β| ≥ 2m − 1, the restriction to Rn \ {0} of the matrix
distribution ∂β E is of class C ∞ and positive homogeneous of degree 2m − n − |β|.
(5) For each β ∈ Nn0 there exists Cβ ∈ (0, ∞) such that the estimate
⎧
⎪
⎪
⎪
Cβ
⎪
⎪
⎪ if either n is odd, or n > 2m, or |β| > 2m − n,
⎨ |x|
⎪ n−2m+|β|
β
|∂ E(x)| ≤ ⎪
⎪
⎪
⎪
⎪ Cβ (1 + | ln |x||)
⎪
⎪
⎩ if 0 ≤ |β| ≤ 2m − n,
|x|n−2m+|β|
(11.3.27)
holds for each x ∈ Rn \ {0}.
(6) When restricted to Rn \ {0}, the entries of E (with “hat” denoting the Fourier
transform) are C ∞ functions and, moreover,

E(ξ) = (−1)m L(ξ) −1 for each ξ ∈ Rn \ {0}. (11.3.28)
In addition,

if n > 2m then
E = (−1)m L(·) −1
(11.3.29)
as tempered distributions in Rn ,
which also implies

if n > 2m then E = (−1)m (2π)−n L(·) −1
(11.3.30)
as tempered distributions in Rn .
A more general version of (11.3.28) is as follows: given any γ ∈ Nn0 , the tempered
distribution ∂
γ E is of class C ∞ when restricted to Rn \ {0} and, regarded as such,
satisfies
γ −1
∂
γ E(ξ) = (−1)m i|γ| ξ L(ξ) for every ξ ∈ Rn \ {0}. (11.3.31)
(7) Writing EL in place of E to emphasize the dependence on L, the fundamental

solution EL with entries as in (11.3.19)–(11.3.20) satisfies
∗
EL = E L , EL = EL , EL = E L∗ ,
(11.3.32)
and EλL = λ−1 EL ∀ λ ∈ C \ {0}.
As a consequence of this and (11.3.23), for each y ∈ Rn one also has

(L x ) E (x − y) = δy (x) I M×M in M M×M S (Rn ) . (11.3.33)

(8) Any fundamental solution U ∈ M M×M S (Rn ) of the system L in Rn is of the
form U = E + Q where E is as in (11.3.19)–(11.3.20) and Q is an M × M matrix
whose entries are polynomials in Rn and whose columns, Qk , k ∈ {1, . . . , M},
satisfy the pointwise equations LQk = 0 ∈ C M in Rn for each k ∈ {1, . . . , M}.
Remark 11.2.(1) Let m ∈ N and a ∈ R \ {0}. Since a = |a|sgn a, it follows that

a2m−1 sgn a = |a|2m−1 (sgn a)2m = |a|2m−1 . This shows that E in (11.3.19) corre-
sponding to n ∈ N, n odd, may be written as

Δ(n−1)/2
E(x) = x
|x · ξ|2m−1 L(ξ) −1 dσ(ξ) (11.3.34)
4(2πi)n−1 (2m − 1)! S n−1

(2) For further reference, we single out the expressions of E from Theorem 11.1
corresponding to the case m = 1. Specifically, if n, M ∈ N with n ≥ 2, and L
is the M × M system of order 2 with complex constant coefficients in Rn of the
form
L = L(∂) := arsjk ∂r ∂ s , (11.3.35)
1≤ j,k≤M
r,s∈{1,...,n}
where arsjk ∈ C, j, k ∈ {1, ..., M}, r, s ∈ {1, . . . , n}, and L satisfies (11.3.3), then
for each x ∈ Rn \ {0}, the M × M matrix E from (11.3.19)–(11.3.20) has the
expression

Δ(n−1)/2
E(x) = x n−1 |x · ξ| L(ξ) −1 dσ(ξ) (11.3.36)
4(2πi) S n−1
if n is odd, and

−Δn/2
E(x) = x
(x · ξ)2 ln |x · ξ| L(ξ) −1 dσ(ξ) (11.3.37)
2(2πi)n S n−1
if n is even.
Moreover, by making use of the observation that for all x, ξ ∈ Rn \ {0} we have

Δ x (x · ξ)2 ln (x · ξ)2 = 8|ξ|2 + 4|ξ|2 ln |x · ξ|, (11.3.38)
one may absorb Δ x inside the integral in formula (11.3.37) and obtain that if n is
even then

Δ(n−2)/2
E(x) = cn − x
n
ln |x · ξ| L(ξ) −1 dσ(ξ) (11.3.39)
(2πi) S n−1
for each x ∈ Rn \ {0}, where

⎧
⎪
⎪
⎪ 0, if n > 2,
⎪
⎨
cn := ⎪
⎪ 1 (11.3.40)
⎪
⎪
⎩ 2 L(ξ) −1 dσ(ξ), if n = 2.
2π S 1
Proof of Theorem 11.1. To facilitate the subsequent discussion, we agree to denote

by P jk (ξ) 1≤ j,k≤M the inverse of the characteristic matrix L(ξ), i.e.,

P jk (ξ) := L(ξ) −1 , ∀ ξ ∈ Rn \ {0}. (11.3.41)
1≤ j,k≤M

Then the entries E jk 1≤ j,k≤M of the matrix E are given at each x ∈ Rn \ {0} by

Δ(n−1)/2
x
E jk (x) = (x · ξ)2m−1 sgn (x · ξ)P jk (ξ) dσ(ξ) (11.3.42)
4(2πi)n−1 (2m − 1)! S n−1
if n is odd, and

−Δn/2
x
E jk (x) = (x · ξ)2m ln |x · ξ|P jk (ξ) dσ(ξ) (11.3.43)
(2πi)n (2m)! S n−1
if n is even. The proof of the theorem is completed in eight steps.

Step I. We claim that if we set q := 0 if n is even andq := 1if nis odd, then
x · ξ
−Δ(n+q)/2
x
E jk (x) = (x · ξ)2m+q log P jk (ξ) dσ(ξ) (11.3.44)
(2πi)n (2m + q)! S n−1 i
for x ∈ Rn \ {0} and j, k ∈ {1, . . . , M}, where log denotes the principal branch of the
complex logarithm defined for points z ∈ C \ {x : x ∈ R, x ≤ 0}.
Suppose first that n is even, thus q = 0, and start from (11.3.43), then use the

formula log x·ξ i = ln |x · ξ| − πi2 sgn (x · ξ) for the term under the integral sign and
the fact that the integral over S n−1 of the function (x · ξ)2m sgn (x · ξ)P jk (ξ) (which is
odd in ξ) is zero, to obtain that (11.3.44) holds.
Moving on to the case when n is odd, consider the function
⎧ 2m+q
⎪
⎪
⎨t log (t/i) if t ∈ R \ {0},
F(t) := ⎪⎪ (11.3.45)
⎩0 if t = 0.
It is not difficult to see that
F ∈ C ∞ (R \ {0}) ∩ C 2m+q−1 (R), F (k) (0) = 0 if k = 1, . . . , 2m + q − 1, (11.3.46)
and that for each k ∈ {1, . . . , 2m} there exists a constant Cm,q,k such that
d k (2m + q)! 2m+q−k t
F(t) = t log + Cm,q,k t2m+q−k (11.3.47)
dt (2m + q − k)! i
in R \ {0}. In particular, by the Chain Rule, if β ∈ Nn0 is such that |β| ≤ 2m + q − 1,
then

∂βx [F(x · ξ)] = F (|β|) (x · ξ)ξβ , ∀ x ∈ Rn , ∀ ξ ∈ S n−1 . (11.3.48)
Based on (11.3.48) and (11.3.47) for each x ∈ Rn \ {0} we may write
x · ξ
Δx (x · ξ)2m+1 log P jk (ξ) dσ(ξ)
S n−1 i

= Δx F(x · ξ)P jk (ξ) dσ(ξ) = F (x · ξ)P jk (ξ) dσ(ξ)
S n−1 S n−1
x · ξ
= 2m(2m + 1) (x · ξ)2m−1 log P jk (ξ) dσ(ξ)
S n−1 i

+ Cm,q (x · ξ)2m−1 P jk (ξ) dσ(ξ)
S n−1

πi
= 2m(2m + 1) (x · ξ)2m−1 ln |x · ξ| − 2 sgn(x · ξ) P jk (ξ) dσ(ξ)
S n−1

= −πim(2m + 1) (x · ξ)2m−1 sgn(x · ξ)P jk (ξ) dσ(ξ). (11.3.49)
S n−1
Hence, if one starts with the expression in the right-hand side of (11.3.44), then one
transfers Δ under the integral sign using from (11.3.49), one arrives at the expression
of E jk from (11.3.42). This completes the proof of Step I.
Step II. Proof of the fact that the entries of E are C ∞ and even in Rn \ {0}.
That the functions in (11.3.42) and (11.3.43) are even is immediate from their
respective expressions. To show that the components of E belong to the space
C ∞ (Rn \ {0}), for every ∈ {1, ..., n} let e be the unit vector in Rn with one on
the -th component, and consider the open set

O := Rn \ {λ e : λ ≤ 0} = x = (x1 , . . . , xn ) ∈ Rn : x > 0 . (11.3.50)
Fix an arbitrary index ∈ {1, . . . , n} and for each x = (x1 , . . . , xn ) ∈ O define the
linear map R,x : Rn → Rn by
ξ · x + ξ |x| ξ (|x| + 2x ) − ξ · x

R,x (ξ) := ξ − e + x, ∀ ξ ∈ Rn . (11.3.51)
|x| + x |x|(|x| + x )
By Exercise 4.130 (presently used with ζ := e and η := x

|x| ) we have that this is a
unitary transformation and
x · R,x (ξ) = |x|ξ , ∀ x ∈ O , ∀ ξ ∈ Rn . (11.3.52)
Also,
R,λ x = R,x whenever x ∈ O and λ > 0, (11.3.53)
and the joint application
O × Rn (x, ξ) → R,x (ξ) ∈ Rn is of class C ∞ . (11.3.54)
Fix j, k ∈ {1, . . . , M}. Using the invariance under unitary transformations of the
operation of integration over S n−1 , for each x ∈ O we may then write
x · ξ
(x · ξ)2m+q · log P jk (ξ) dσ(ξ)
S n−1 i
x · R,x (ξ)
= (x · R,x (ξ))2m+q · log P jk (R,x (ξ)) dσ(ξ)
S n−1 i
|x|ξ
= (|x|ξ )2m+q · log P jk (R,x (ξ)) dσ(ξ)
S n−1 i
ξ
= |x|2m+q ξ2m+q ln |x| + log P jk (R,x (ξ)) dσ(ξ). (11.3.55)
S n−1 i
From this representation, (11.3.54), and (11.3.44), it is clear that E jk ∈ C ∞ (O ).

Next, for each ∈ {1, ..., n} define the open set

O− := Rn \ {λ e : λ ≥ 0} = x = (x1 , . . . , xn ) ∈ Rn : x < 0 (11.3.56)
and for each given x ∈ O− define the linear map R−,x : Rn → Rn by
ξ · x − ξ |x| ξ (|x| − 2x ) + ξ · x

R−,x (ξ) := ξ + e − x, ∀ ξ ∈ Rn . (11.3.57)
|x| − x |x|(|x| − x )
Running the reasoning above with O replaced by O− and R,x replaced by R−,x (this
time invoking Exercise 4.130 with ζ := e and η := − |x|x , and with identity (11.3.52)
replaced by x · R−,x (ξ) = −|x|ξ for all x ∈ O− and all ξ ∈ Rn ) ultimately implies
E jk ∈ C ∞ (O− ) for each j, k ∈ {1, . . . , M} and each ∈ {1, ..., n}.

n
To complete Step II, there remains to observe that Rn \ {0} = (O ∪ O− ).
=1
Step III. Proof of part (3) in the statement of the theorem.

To facilitate the discussion, fix j, k ∈ {1, . . . , M}, introduce

Q jk (x) := (x · ξ)2m+q P jk (ξ) dσ(ξ), ∀ x ∈ Rn , (11.3.58)
S n−1
and define Ψ jk : Rn \ {0} → C by setting

x · ξ
Ψ jk (x) := (x · ξ)2m+q · log P jk (ξ) dσ(ξ) − ln |x| Q jk (x) (11.3.59)
S n−1 i
for each Rn \ {0}. Observe that Q jk is a polynomial of degree 2m + q that vanishes
when n is odd (since in that case the integrand in (11.3.58) is odd). Also, from
our earlier analysis in (11.3.55), we know that the integral in the right-hand side of
(11.3.59) depends in a C ∞ fashion on the variable x ∈ Rn \ {0}. These comments and
(11.3.59) imply that Ψ jk is of class C ∞ in Rn \ {0}.
Our next goal is to prove that
Ψ jk is positive homogeneous of degree 2m + q in Rn \ {0}. (11.3.60)
To this end, first note that for each ∈ {1, . . . , n} we may write

Q jk (x) = |x|2m+q ξ2m+q P jk (R,x (ξ)) dσ(ξ), ∀ x ∈ O , (11.3.61)
S n−1
by (11.3.58) and (11.3.52). Consequently, from (11.3.61), (11.3.55), and (11.3.59),

we deduce that for each ∈ {1, . . . , n},
ξ
Ψ jk (x) := |x|2m+q ξ2m+q log P jk (R,x (ξ)) dσ(ξ), (11.3.62)
S n−1 i
for all x ∈ O . In turn, (11.3.62) and (11.3.53) readily show that Ψ jk is positive
homogeneous of degree 2m + q when restricted to the cone-like region O . Since
n
Rn \ {0} = O , the claim in (11.3.60) follows.
=1
Next, an induction argument shows that for each ξ ∈ Rn \ {0} and k, N ∈ N
satisfying N ≥ 2k, the following formulas hold:
N!
Δkx (x · ξ)N = (x · ξ)N−2k |ξ|2k (11.3.63)
(N − 2k)!
and

Δkx (x · ξ)N ln |x| = (ln |x|)Δkx (x · ξ)N (11.3.64)

k
+ c(r, k, N, n)|x|−2r (x · ξ)N−2k+2r .
r=1
We now observe that if P jk is the ( j, k)-entry in the matrix P from (11.3.24), then

−1 (n+q)/2
P jk (x) = Δ x (x · ξ)2m+q P jk (ξ) dσ(ξ), (11.3.65)
(2πi)n (2m + q)! S n−1
as seen from identity (11.3.63) with N := 2m + q and k := (n + q)/2. It is also

immediate from (11.3.65) that P jk is identically zero when either n is odd (due
to parity considerations) or n > 2m (due to degree considerations). In addition,
formula (11.3.65) shows that P jk is a homogeneous polynomial of degree 2m − n
when n ≤ 2m. Finally, we note that combining (11.3.58), (11.3.65), and (11.3.64)
used with N := 2m + q and k := (n + q)/2, yields
−1
Δ(n+q)/2
x (ln |x|)Q jk (x) = (ln |x|)P jk (x) (11.3.66)
(2π i)n (2m + q)!

(n+q)/2
+ Cr |x|−2r (x · ξ)2m−n+2r P jk (ξ) dσ(ξ), ∀ x ∈ Rn \ {0},
r=1 S n−1
for some constants Cr depending only on r, n, q, and m. It is easy to see that the sum
in the right-hand side of (11.3.66) gives rise to a function that belongs to C ∞ (Rn \{0})
and is positive homogeneous of degree 2m − n. At this point, if we define
−1
Φ jk (x) := Δ(n+q)/2
x Ψ jk (x) (11.3.67)
(2πi)n (2m + q)!

(n+q)/2
+ Cr |x|−2r (x · ξ)2m−n+2r P jk (ξ) dσ(ξ)
r=1 S n−1
for each x ∈ Rn \ {0}, then (11.3.25) follows from (11.3.44), (11.3.59), (11.3.66),
and (11.3.67). Moreover, from (11.3.60) and (11.3.67), it is clear that Φ jk is positive
homogeneous of degree 2m−n, while the regularity of Ψ jk established earlier entails
Φ jk ∈ C ∞ (Rn \ {0}).
1
Step IV. Proof of the fact that E ∈ M M×M (S (Rn )) ∩ M M×M Lloc (Rn \ {0}) .
Fix j, k ∈ {1, . . . , M} and recall (11.3.25). By what we proved in Step III, Φ jk is
positive homogeneous of degree 2m − n in Rn \ {0}, and since m ≥ 1 we may invoke
Exercise 4.53 to conclude that Φ jk ∈ Lloc 1
(Rn ). Now the estimate from Exercise 4.53

and Example 4.4 give Φ jk ∈ S (R ). In addition, by Exercise 4.18, it follows that
n
(ln |x|)P jk ∈ S (Rn ) ∩ Lloc

1
(Rn \ {0}). In summary, we conclude that E jk ∈ S (Rn ) ∩
Lloc (R \ {0}).
1 n
Step V. Proof of (11.3.23).

Componentwise, (11.3.23) reads as follows: for each j, k ∈ {1, . . . , M},
⎧
⎪⎪ if j k,
M
x ⎨0
L jr Erk (x − y) = ⎪
⎪ in S (Rn ), (11.3.68)
⎩ δy (x) if j = k,
r=1
where the superscript x denotes the fact that the operator L jr (defined as in (11.3.1))
is applied in the variable x. To justify this, fix j, k, r ∈ {1, . . . , M}. By (11.3.47) and
(11.3.48) we have
(2m + q)! x·ξ
L xjr F(x · ξ) = (x · ξ)q log + Cm,q (x · ξ)q L jr (ξ) (11.3.69)
q! i
for every x ∈ Rn \ {0} and every ξ ∈ S n−1 such that x · ξ 0. To continue, fix
ϕ ∈ C0∞ (Rn ) and write

L xjr F(x · ξ)Prk (ξ) dσ(ξ) , ϕ(x)
S n−1

= F(x · ξ)Prk (ξ) dσ(ξ), L jr ϕ(x)
S n−1

= F(x · ξ)Prk (ξ) dσ(ξ)L jr ϕ(x) dx
Rn S n−1

= Prk (ξ) F(x · ξ)L jr ϕ(x) dx dσ(ξ)
S n−1 x∈Rn , x·ξ>0

+ rk
P (ξ) F(x · ξ)L jr ϕ(x) dx dσ(ξ). (11.3.70)
S n−1 x∈Rn , x·ξ<0
At this point, we integrate by parts repeatedly with respect to x (according to formula

(14.8.4)) in the inner most integrals in (11.3.70), until we transfer all the derivatives
from ϕ onto F(x · ξ). This is justified by (11.3.46), the fact that

∂αx F(x · ξ) ∈ Lloc
1
(Rn ) whenever α ∈ Nn0 has |α| ≤ 2m (11.3.71)
(cf. (11.3.48) and Exercise 2.125), and the fact that ϕ has compact support. Note that
in the process, the terms corresponding to boundary integrals (i.e., integrals over the
set {x ∈ Rn : x · ξ = 0} ∩ supp ϕ) are zero thanks to the formulas in (11.3.46). Hence,
summing up over r the resulting identity in (11.3.70), then using (11.3.69) and the

M
fact that L jr (ξ)Prk (ξ) = δ jk for every ξ ∈ S n−1 , we arrive at
r=1

M
L xjr F(x · ξ)Prk (ξ) dσ(ξ) , ϕ(x) (11.3.72)
r=1 S n−1
M

= Prk (ξ) [L xjr F(x · ξ)]ϕ(x) dx dσ(ξ)
r=1 S n−1 x∈Rn
(2m + q)!
x·ξ
= (x · ξ)q log + Cm,q (x · ξ)q δ jk dσ(ξ)ϕ(x) dx.
Rn S n−1 q! i
Consequently, since ϕ ∈ C0∞ (Rn ) is arbitrary, a combination of (11.3.72), (11.3.45),
and (11.3.44), yields

M
L jr Erk = Δ(n+q)/2 Eq + Pq δ jk in D (Rn ), (11.3.73)
r=1
where x · ξ
−1
Eq (x) := (x · ξ)q · log dσ(ξ), (11.3.74)
(2πi)n q! S n−1 i
for each x ∈ Rn \ {0}, and

−Cm,q
Pq (x) := (x · ξ)q dσ(ξ), ∀ x ∈ Rn . (11.3.75)
(2πi)n (2m + q)! S n−1
Given our choice of q, from Lemma 7.31 we have that Eq is a fundamental solution
for Δ(n+q)/2 , so Δ(n+q)/2 Eq = δ in D (Rn ). Also, Pq is a homogeneous polynomial of
degree q in Rn , thus Δ(n+q)/2 Pq = 0 pointwise and in D (Rn ). Therefore, (11.3.73)
becomes

M
L jr Erk = δ jk δ in D (Rn ), ∀ j, k ∈ {1, . . . , M}. (11.3.76)
r=1
The statement in part (2) of the theorem follows from (11.3.76), (4.1.33), and the
result from Step IV.
Step VI. Proof of claims in parts (4)–(7) in the statement of the theorem.
The claim in part (4) follows from part (1), (11.3.25), the fact that each P jk is a
homogeneous polynomial of degree at most 2m − n, and the observation that when
computing ∂β E jk with |β| ≥ 2m − 1 at least one derivative falls on ln. The estimates
in (11.3.27) are a direct consequence of (11.3.25). This takes care of parts (4) and
(5) in the statement of the theorem.
Moving on to the proof of part (6), fix j, k ∈ {1, . . . , M} and recall (11.3.25). From
the proof in Step IV, we have that Φ jk and (ln |x|)P jk are tempered distributions in Rn .
Since Φ jk ∈ C ∞ (Rn \ {0}) and is positive homogeneous of degree 2m − n in Rn \ {0},
Proposition 4.60 implies that its Fourier transform coincides with a C ∞ function on
Rn \{0}. To analyze the effect of taking the Fourier transform of (ln |x|)P jk pick some
θ ∈ C0∞ (Rn ) such that supp θ ⊂ B(0, 2) and θ ≡ 1 on B(0, 1) and write
(ln |x|)P jk = (1 − θ)(ln |x|)P jk + θ(ln |x|)P jk in Rn \ {0}. (11.3.77)
The two terms in the right-hand side of (11.3.77) continue to belong to S (Rn ).
From Example 2.9 and the fact that θ is compactly supported, we obtain that the
function θ(ln |x|)P jk belongs to L1 (Rn ) and has compact support. Consequently,

F θ(ln |x|)P jk ∈ C ∞ (Rn ) (recall Exercise 3.31). Regarding (1 − θ)(ln |x|)P jk , note
that this function is of class C ∞ in Rn . Also, for every β ∈ Nn0 , the function

xβ ∂α (1 − θ)(ln |x|)P jk belongs to L1 (Rn ) provided α ∈ Nn0 and |α| is sufficiently
large. Since the Fourier transform of any L1 function is continuous, this readily im-
plies that for any r ∈ N there exists α ∈ Nn0 such that

F ∂α [(1 − θ)(ln |x|)P jk ] = (iξ)α F (1 − θ)(ln |x|)P jk ∈ C r (Rn ). (11.3.78)

Thus, we necessarily have F (1 − θ)(ln |x|)P jk ∈ C ∞ (Rn \ {0}).
The reasoning above shows that the Fourier transform of (the matrix-valued tem-
pered distribution) E when restricted to Rn \ {0} is a function of class C ∞ . Taking
the Fourier transforms of both sides of (11.3.76) gives

M
(−1)m L jr (ξ)E
rk (ξ) = δ jk in S (R ), for all j, k ∈ {1, . . . , M}.
n
(11.3.79)
r=1
Restricting (11.3.79) to Rn \{0} then readily implies (11.3.28). Also, for each γ ∈ Nn0
we have that ∂γ E is a tempered distribution and item (b) in Theorem 4.26 implies
∂
γ E = (−1)m i|γ| ξ γ
E in S (Rn ). The latter combined with (11.3.28) then yields
(11.3.31).
Consider next the task of justifying (11.3.29). The assumption that n > 2m guar-

antees that the matrix-valued function L(·) −1 has entries that are locally integrable
in Rn and satisfy (4.1.4). As remarked in Example 4.4, the distribution of function

type defined by L(·) −1 is a well-defined tempered distribution in Rn . Moreover,
Exercise 4.57 implies that this tempered distribution is positive homogeneous of
degree −2m. To proceed, define

u :=
E − (−1)m L(·) −1 . (11.3.80)

From the above discussion, it follows that u ∈ M M×M S (Rn ) . In addition, since

from (11.3.26) and Proposition 4.59 we know that
E ∈ M M×M S (Rn ) is positive
homogeneous of degree −2m, we conclude that u is also positive homogeneous of

degree −2m. Observe next that L(ξ) ∈ M M×M L(Rn ) and

L(ξ)u = L(ξ)
E − (−1)m L(ξ) L(·) −1
− (−1)m I M×M = (−1)m δI

= (−1)m LE M×M − (−1) I M×M
m

= (−1)m I M×M − (−1)m I M×M = 0 in M M×M S (Rn ) ; (11.3.81)
cf. item (b) in Theorem 4.26 and (4.2.4). As a consequence of (11.3.81) and (11.3.3),
we deduce that supp u ⊆ {0}. Granted this, we may invoke Exercise 2.75 to conclude

that each entry in u is of the form |α|≤N cα ∂α δ. Hence, on the one hand, u has
polynomial entries (again, see item (b) in Theorem 4.26 and (4.2.4)). On the other

hand, Proposition 4.59 ensures that u ∈ M M×M S (Rn ) is positive homogeneous of
degree −n + 2m. Since the current assumption forces −n + 2m < 0, this implies u=0
hence, ultimately, u = 0. Now (11.3.29) follows from (11.3.80). In turn, (11.3.30) is
a consequence of (11.3.29), (4.2.34), and (11.3.22). This finishes the proof of part
(6) in the statement of the theorem.
Finally, the identities in (11.3.32) can be seen directly from (11.3.42)–(11.3.43).
Step VII. Proof of the claim in part (8) in the statement of the theorem.

Let U ∈ M M×M S (Rn ) be an arbitrary fundamental solution of the system L in Rn

and set Q := U − E. Then LQ = 0 in M M×M S (Rn ) and, on the Fourier transform
= 0 in M M×M S (Rn ). In light of (11.3.3), this implies
side, Q satisfies L(ξ)Q
supp Q ⊆ {0}, hence Q is an M × M matrix whose entries are polynomials in Rn by
Exercise 4.37 (applied to each entry).
Step VIII. Proof of the fact that the entries of E are real-analytic in Rn \ {0}.
This is a direct consequence of the fact that (cf. (11.3.23))

L E = δ I M×M in M M×M S (Rn ) , (11.3.82)
where the operator L is applied to each column of E, which implies that L E = 0 in

M M×M D (Rn \ {0}) , and Theorem 11.7, established in the next section.
The proof of Theorem 11.1 is now complete.
Theorem 11.1 describes all fundamental solutions, which are tempered distri-
butions, for any homogeneous constant coefficient system with an invertible char-
acteristic matrix, and elaborates on the properties of such fundamental solutions.
In specific cases, it is possible to use formulas (11.3.19)–(11.3.20) to find an ex-
plicit expression for a specific fundamental solution. The case of the polyharmonic
operator is discussed in Exercise 11.24. Here we study in detail the case of the three-
dimensional Lamé operator (cf. also Exercise 11.20 and Exercise 11.21). We remark
that the argument in the proof of Proposition 11.3 is different from the one used to
derive the fundamental solution for the Lamé operator in Section 11.1.
Proposition 11.3. Let λ, μ ∈ C be such that μ 0, λ + 2μ 0. A fundamental

solution for the Lamé operator (10.3.1) in R3 is
1 λ + 3μ 1 1 λ+μ x⊗x
E(x) = − I3×3 − , x ∈ R3 \ {0}. (11.3.83)
8π μ(λ + 2μ) |x| 8π μ(λ + 2μ) |x|3
Proof. We start by recalling formula (11.3.42) that gives an expression for the fun-
damental solution of a homogeneous differential operator for n odd. In our case,
L = (L jk )1≤ j,k≤3 , L jk = μδ jk Δ + (λ + μ)∂ j ∂k ,

(11.3.84)
and L(ξ) = μI3×3 + (λ + μ)ξ ⊗ ξ for each ξ ∈ S 2 .
Observe that (x · ξ) sgn (x · ξ) = |x · ξ|. As such, for any x ∈ R3 \ {0} we may compute

1
E(x) = − Δx |x · ξ|(L(ξ))−1 dσ(ξ) (11.3.85)
16π2 S2
1
1 λ+μ
=− Δx |x · ξ| I3×3 − ξ ⊗ ξ dσ(ξ)
16π2 S2 μ μ(λ + 2μ)
1 ω1 λ + μ ω1 x ⊗ x
=− Δx |x|I3×3 − |x|I3×3 +
16π2 μ μ(λ + 2μ) 4 |x|
1 4λ + 8μ − λ − μ λ + μ x ⊗ x
=− Δx |x|I3×3 −
16π 2μ(λ + 2μ) 2μ(λ + 2μ) |x|
1 3λ + 7μ 2 λ+μ 2 4x ⊗ x
=− I3×3 − I3×3 −
16π 2μ(λ + 2μ) |x| 2μ(λ + 2μ) |x| |x|3
1 λ + 3μ 1 1 λ+μ x⊗x
=− I3×3 − .
8π μ(λ + 2μ) |x| 8π μ(λ + 2μ) |x|3
For the second equality in (11.3.85), we have used Proposition 10.14, for the third
we have used Proposition 14.67 and Proposition 14.68, while for the fifth we have
used (7.3.2) and the readily verified fact that if N is a nonzero integer then
x ⊗ x 2 N(N − n − 2)
Δx = In×n + x ⊗ x, (11.3.86)
|x|N |x|N |x|N+2
Exercise 11.4. Fix n ∈ N with n ≥ 2 along with M ∈ N and consider a second order
M × M system
n
L= aαβ
jk ∂ j ∂k , aαβ
jk ∈ C. (11.3.87)
1≤α,β≤M
j,k=1
Also, suppose E = (Eαβ )1≤α,β≤M : Rn \ {0} → M M×M (C) is a matrix-valued func-

tion whose entries belong to C 2 (Rn \ {0}) ∩ Lloc1
(Rn ) and have gradients positive
homogeneous of degree 1 − n in R \ {0}.
n
Prove that the following statements are equivalent:

1
(1) When viewed in M M×M Lloc (Rn ) , the matrix-valued function E is a fundamental
solution for L in Rn ;
(2) With the system L acting on the columns of E one has LE = 0 ∈ M M×M (C)
pointwise in Rn \ {0} and for each α, γ ∈ {1, . . . , M} there holds
M
n
aαβ
jk ω j ∂k E βγ (ω) dσ(ω) = δαγ . (11.3.88)
j,k=1 β=1 S n−1
Hint: Adapt the proof of Theorem 7.60 (this time taking the test function to be vector

valued, i.e., f ∈ C0∞ (Rn ) M ).
Exercise 11.5. Suppose n ∈ N satisfies n ≥ 2, and fix some M ∈ N. Consider a

second order M × M system L as in (11.3.87) with the property that (11.3.3) holds,
and suppose E = (Eαβ )1≤α,β≤M is the fundamental solution associated with L as in
Theorem 11.1 (with m := 1). Finally, fix two indexes α, γ ∈ {1, . . . , M}. Prove that
M
n
aαβ
jk ω j (∂k E βγ )(ω) dσ(ω) = δαγ
j,k=1 β=1 S n−1
M
n
= aβα
k j ω j (∂k E γβ )(ω) dσ(ω), (11.3.89)
j,k=1 β=1 S n−1
and that, given any vector ξ ∈ S n−1 ,

n
M
1
aαβ
jk ω j (∂k E βγ )(ω) dσ(ω) = δαγ
j,k=1 β=1
2
ω∈S n−1
ω,ξ>0

n
M
= aβα
k j ω j (∂k E γβ )(ω) dσ(ω). (11.3.90)
j,k=1 β=1
ω∈S n−1
ω,ξ>0
Hint: For the first equality in (11.3.89) use Exercise 11.4 and Theorem 11.1. Then,
the last equality in (11.3.89) follows from the first (written for L in place of L,
bearing in mind (11.3.32)). Formulas (11.3.90) are consequences of (11.3.89) and
the fact that the integrands are even (cf. Theorem 11.1).
11.4 Interior Estimates and Real-Analyticity for Null-Solutions

of Systems
The aim in this section is to explore the extent to which results such as integral
representation formula and interior estimates, proved earlier in §6.3 in the scalar
context, continue to hold for vector-valued functions which are null-solutions of a
certain class of systems of differential operators.
Proposition 11.6. Fix n, m, M ∈ N with n ≥ 2, and assume that L is an M × M

system in Rn of order 2m of the form

L= Aα ∂α , Aα = aαjk 1≤ j,k≤M ∈ M M×M (C), (11.4.1)
|α|=2m
11.4 Interior Estimates and Real-Analyticity for Null-Solutions of Systems 407
with the property that det [L(ξ)] 0 for each ξ ∈ Rn \ {0}. In addition, suppose

Ω ⊆ Rn is an open set and u = (u1 , . . . , u M ) ∈ D (Ω) M is such that Lu = 0 in
M
D (Ω) .

Then u ∈ C ∞ (Ω) M and for each x0 ∈ Ω, each r ∈ 0, dist (x0 , ∂Ω) , and each

function ψ ∈ C0∞ B(x0 , r) such that ψ ≡ 1 near B(x0 , r/2), we have

M
α!
u (x) = − (−1)|γ| aαjk ×
j,k=1 |α|=2m γ<α
γ!(α − γ)!

× (∂γ E j )(x − y)∂α−γ ψ(y) uk (y) dy (11.4.2)
B(x0 ,r)\B(x0 ,r/2)
for each ∈ {1, . . . , M} and each x ∈ B(x0 , r/2), where

E = E jk 1≤ j,k≤M ∈ M M×M S (Rn ) ∩ M M×M C ∞ (Rn \ {0}) (11.4.3)
is the fundamental matrix for L (the transpose of L) as given by Theorem 11.1.

In particular, for every μ ∈ Nn0 ,

M
α!
∂μ u (x) = − (−1)|γ| aαjk ×
j,k=1 |α|=2m γ<α
γ!(α − γ)!

× (∂γ+μ E j )(x − y)∂α−γ ψ(y) uk (y) dy (11.4.4)
B(x0 ,r)\B(x0 ,r/2)
for each ∈ {1, . . . , M} and each x ∈ B(x0 , r/2). Also, if either n is odd, or n > 2m,
or if |μ| > 2m − n, we have

Cμ
|(∂μ u)(x0 )| ≤ |μ| − |u(y)| dy, (11.4.5)
r B(x0 ,r)
where Cμ ∈ (0, ∞) is independent of u, x0 , r, and Ω.

Proof. Let u = (u )1≤≤M and E = E jk 1≤ j,k≤M be as specified in the statement of
the proposition. In particular,

E ∈ M M×M S (Rn ) ∩ M M×M C ∞ (Rn \ {0}) (11.4.6)

by Theorem 11.1. Also, pick x0 ∈ Ω, a number r ∈ 0, dist (x0 , ∂Ω) , and fix some

function ψ ∈ C0∞ B(x0 , r) such that ψ ≡ 1 near B(x0 , r/2). Then, from Theorem 10.9
∞ M
it follows that u ∈ C (Ω) , hence also ψu ∈ C0∞ (Ω) M . Granted these, for each
x ∈ B(x0 , r/2) we may then write (keeping in mind that E is a fundamental solution
for L , the transposed of L),

u(x) = (ψu)(x) = (ψu )(x) 1≤≤M

M

= δk δ , (ψuk )(x − ·)
1≤≤M
k=1

M

= (L E)k , (ψuk )(x − ·) . (11.4.7)
1≤≤M
k=1
Note that
L = (−1)|α| Aα ∂α = Aα ∂α , (11.4.8)
|α|=2m |α|=2m
where Aα is the transposed of the matrix Aα , for each α. Thus, (11.4.7) implies
M
M
u(x) = (Aα )k j ∂α E j , (ψuk )(x − ·) (11.4.9)
1≤≤M
k=1 |α|=2m j=1
M
M

= aαjk E j , ∂α (ψuk )(x − ·)
1≤≤M
k=1 |α|=2m j=1

M
M
= E j , aαjk (∂α (ψuk ))(x − ·)
1≤≤M
j=1 |α|=2m k=1

M
M α!
= E j (x − ·), aαjk ∂β ψ∂α−β uk
j=1 |α|=2m k=1 0<β≤α
β!(α − β)! 1≤≤M

M
M
+ E j (x − ·), ψ aαjk ∂α uk
1≤≤M
j=1 |α|=2m k=1

M
α!
= aαjk E j (x − ·), ∂β ψ ∂α−β uk
j,k=1 |α|=2m 0<β≤α
β!(α − β)! 1≤≤M
since for each j ∈ {1, . . . , M},

M
ψ aαjk ∂α uk = ψ(Lu) j = 0 in Rn . (11.4.10)
|α|=2m k=1
Next, for each , j, k ∈ {1, . . . , M} and each α, β ∈ Nn0 satisfying |α| = 2m and
0 < β ≤ α, we observe that
11.4 Interior Estimates and Real-Analyticity for Null-Solutions of Systems 409

E j (x − ·), ∂β ψ ∂α−β uk (11.4.11)

= E j (x − y)∂β ψ(y)∂α−β uk (y) dy
B(x0 ,r)\B(x0 ,r/2)

= (−1)|β| ∂α−β
y E j (x − y)∂β ψ(y) uk (y) dy
B(x0 ,r)\B(x0 ,r/2)
(α − β)!
= (−1)|β|+|γ| ×
γ≤α−β
γ!(α − β − γ)!

× (∂γ E j )(x − y)∂α−γ ψ(y) uk (y) dy,
B(x0 ,r)\B(x0 ,r/2)
and that, whenever γ < α,

(α − γ)! (α − γ)!
(−1)|β| = (−1)|β| −1
0<β≤α−γ
β!(α − β − γ)! β≤α−γ
β!(α − β − γ)!
= 0 − 1 = −1. (11.4.12)
At this stage, (11.4.2) follows from (11.4.9), (11.4.11), and (11.4.12). In turn,
(11.4.2) readily implies (11.4.4). Finally, (11.4.5) is a consequence of (11.4.4), the
assumptions on n and μ, and (11.3.27), by choosing ψ as in (6.3.11)–(6.3.12).
An L∞ -version of the interior estimate (11.4.5), valid in all space dimensions,
is established in the next theorem. In particular, this version allows us to prove the
real-analyticity of null-solutions of the systems considered here.
Theorem 11.7. Let n, m, M ∈ N and suppose L is an M × M constant (complex)
coefficient system in Rn , homogeneous of order 2m, and with the property that
det [L(ξ)] 0 for each ξ ∈ Rn \ {0}. Assume also that Ω ⊆ Rn is an open set

and u ∈ D (Ω) M is such that Lu = 0 in D (Ω) M .
∞ M
Then u ∈ C (Ω) and there exists a constant C ∈ (0, ∞) such that
C |α| (1 − λ)−|α| |α|!

max |∂α u(y)| ≤ max |u(y)|, ∀ α ∈ Nn0 , (11.4.13)
y∈B(x,λr) r|α| y∈B(x,r)

for each x ∈ Ω, each r ∈ 0, dist (x, ∂Ω) , and each λ ∈ (0, 1). In particular, each
component of u is real-analytic in Ω.

Proof. Theorem 10.9 gives that u ∈ C ∞ (Ω) M . As far as (11.4.13) is concerned,
consider first the case when either n is odd, or n > 2m. In this scenario, (11.4.5)
gives that there exists some C ∈ (0, ∞), independent of u and Ω, such that

C
|∂ j u(x)| ≤ − |u(y)| dy, ∀ j ∈ {1, . . . , n}, (11.4.14)
r B(x,r)

whenever x ∈ Ω and r ∈ 0, dist (x, ∂Ω) . A quick inspection reveals that the proof
of Lemma 6.21 carries through in the case when the functions involved are vector-
valued. When applied with

A := u ∈ C ∞ (Ω) M : Lu = 0 in Ω , (11.4.15)
this lemma gives (thanks to (11.4.14)) that for every point x ∈ Ω, every radius

r ∈ 0, dist (x, ∂Ω) , every k ∈ N, and every λ ∈ (0, 1), we have (with C as in
(11.4.14))
C k (1 − λ)−k ek−1 k!
max |∂α u(y)| ≤ max |u(y)|,
y∈B(x,λr) rk y∈B(x,r) (11.4.16)
for every multi-index α ∈ Nn0 with |α| = k.
This proves (11.4.13) under the current assumptions on n.

Fix now an arbitrary n ∈ N and pick some k ∈ N. With L as in (11.4.1) consider
the constant coefficient, homogeneous, M × M system in Rn , of order 4m, given by

L (∂) := L∗ L = A∗α Aβ ∂α+β , (11.4.17)
|α|=|β|=2m
and note that for each ξ ∈ Rn we have

∗ β
L (ξ) = ξα+β A∗α Aβ = ξα Aα ξ Aβ
|α|=|β|=2m |α|=|β|=2m
∗
= L(ξ) L(ξ). (11.4.18)
In particular,
L (ξ)η · η = |L(ξ)η|2 , ∀ ξ ∈ Rn , ∀ η ∈ CM . (11.4.19)
Finally, define the constant coefficient, homogeneous, M × M system in Rn+k , of

order 4m,
(∂, ∂n+1 , . . . , ∂n+k ) := L (∂) + ∂4m + · · · + ∂4m I M×M .
L (11.4.20)
n+1 n+k
We claim that

det L (ξ, ξn+1 , . . . , ξn+k ) 0, ∀ (ξ, ξn+1 , . . . , ξn+k ) ∈ Rn+k \ {0}. (11.4.21)
To see that this is the case, fix some (ξ, ξn+1 , . . . , ξn+k ) ∈ Rn+k \ {0} and note
that it suffices to show that the M × M complex matrix L (ξ, ξn+1 , . . . , ξn+k ) acts
injectively on C M . With this in mind, pick some η ∈ C M with the property that
L(ξ, ξn+1 , . . . , ξn+k )η = 0 ∈ C M . Then
(ξ, ξn+1 , . . . , ξn+k )η · η = |L(ξ)η|2 + ξ4m + · · · + ξ4m |η|2

0=L (11.4.22)
n+1 n+k
11.5 Reverse Hölder Estimates for Null-Solutions of Systems 411
which forces
4m 4m 2
L(ξ)η = 0 and ξn+1 + · · · + ξn+k |η| = 0. (11.4.23)
If ξ ∈ Rn \ {0} then the first condition in (11.4.23) implies η = 0, given the assump-
tions on L. If ξ = 0 then necessarily there exists j ∈ {1, . . . , k} such that ξn+ j 0, and
the second condition in (11.4.23) now implies η = 0. Thus, η = 0 in all alternatives
which finishes the proof of (11.4.21).

To proceed, assume u ∈ C ∞ (Ω) M satisfies Lu = 0 in Ω, and define
u(x, xn+1 , . . . , xn+k ) := u(x) for each

(11.4.24)
(x, xn+1 , . . . , xn+k ) ∈ Ω := Ω × Rk .

Then Ω is an open subset of Rn+k and u ∈ C ∞ (Ω) M . Moreover, as is apparent from
(11.4.17), (11.4.20), and (11.4.24), we have
(∂, ∂n+1 , . . . , ∂n+k )u = 0
L in Ω. (11.4.25)
Assume now that n ∈ N is an arbitrary given number and pick k ∈ N such that

n + k is either odd or n + k > 4m. Then from the first part in the proof, applied to L
and u, we know that there exists a constant C ∈ (0, ∞) such that
C |α| (1 − λ)−|α| |α|!

max |∂α u| ≤ max |u|, ∀ α ∈ Nn0 , (11.4.26)
B( x,λr) r|α| B( x,r)

for each x ∈ Ω, each r ∈ 0, dist ( x, ∂Ω) , and each λ ∈ (0, 1), where the balls
appearing in (11.4.26) are considered in Rn+k . Now, given x ∈ Ω and some radius

r ∈ 0, dist (x, ∂Ω) , specializing (11.4.26) to the case when x := (x, 0, . . . , 0) yields
(11.4.13) for the current n. Thus, (11.4.13) holds for any n. Finally, the last claim in
the statement of the theorem is a consequence of (11.4.13) and Lemma 6.24.
11.5 Reverse Hölder Estimates for Null-Solutions of Systems
The principal result in this section is the version of interior estimates for null-
solutions of certain higher order systems stated in Theorem 11.12. In particular,
this contains the fact that such null-solutions satisfy reverse Hölder estimates (a
topic worthy of investigation in its own right). We begin by making the following
definition.
Definition 11.8. A continuous (complex) vector-valued function u defined in Ω is
said to be p-subaveraging for some p ∈ (0, ∞) if there exists a finite constant
C > 0 with the property that

1p
|u(x)| ≤ C − |u(y)| dy
p
(11.5.1)
B(x,r)

for every x ∈ Ω and every r ∈ 0, dist (x, ∂Ω) .
The class of p-subaveraging functions exhibits a number of self-improvement

properties, the first of which is discussed below.
Lemma 11.9. Assume that u is a continuous (complex) vector-valued function de-

fined in Ω, and suppose that 0 < p < ∞. Then u is p-subaveraging if and only if
there exists a finite constant C > 0 such that

1p
−n/p
sup |u(z)| ≤ C(1 − λ) − |u(y)| dy
p
(11.5.2)
z∈B(x,λr) B(x,r)

for every x ∈ Ω, every r ∈ 0, dist (x, ∂Ω) , and every λ ∈ (0, 1).
Proof. The fact that (11.5.2) implies (11.5.1) is obvious. Conversely, suppose that

u is p-subaveraging. Pick x ∈ Ω, r ∈ 0, dist (x, ∂Ω) , and z ∈ B(x, λr). Then, if
R := (1 − λ)r, it follows that z ∈ Ω and 0 < R < dist (z, ∂Ω). Furthermore, B(z, R) ⊆
B(x, r). Consequently, with C as in (11.5.1),

1p
1p
(1 − λ)−n
|u(z)| ≤ C − |u(y)| dy = C
p
|u(y)| dy
p
B(z,R) |B(z, r)| B(z,R)

1p
−n/p
≤ C(1 − λ) − |u(y)| dy ,
p
(11.5.3)
B(x,r)
which readily implies (11.5.2) by taking the supremum over z ∈ B(x, λr).
The second self-improvement within the class of p-subaveraging functions is the
fact that the value of the parameter p is immaterial.
Lemma 11.10. If there exists p0 > 0 such that u is p0 -subaveraging function, then
u is p-subaveraging for every p ∈ (0, ∞).
In light of Lemma 11.10, it is unequivocal to refer to a function u as simply being

subaveraging if it is p-subaveraging for some p ∈ (0, ∞). The optimal constant
which can be used in (11.5.1) is referred to as the p-subaveraging constant of the
function u.
Proof of Lemma 11.10. The proof is based on ideas used in the work of G. Hardy
and J. Littlewood [29] (cf. also [16, Lemma 2, pp. 172–173]). The case p > p0
can be handled directly utilizing Hölder’s inequality with q = pp0 > 1. Henceforth,
we shall focus on the case when p < p0 . Replacing u by a suitable power of |u|,
there is no loss of generality in assuming that, in fact, p0 = 1 and p < 1. We
may !also assume (by rescaling and making a translation) that B(0, 1) ⊆ Ω, x = 0,
and B(0,1) |u(y)| p dy = 1. The goal is then to prove the estimate |u(0)| ≤ C with a
finite constant C > 0 independent of u. Continuing our series of reductions, we may
assume that |u(0)| > 1. Next, for each r ∈ (0, 1] and q ∈ (0, ∞], introduce
⎧ ! 1p
⎪
⎪
⎪
⎪
⎨ B(0,r) |u(y)| p
dy if q < ∞,
mq (r) := ⎪
⎪ (11.5.4)
⎪
⎪
⎩ sup
B(0,r) |u(y)| if q = ∞.
Observe that for each r ∈ (0, 1),

m1 (r) ≤ m p (r) p m∞ (r) 1−p ≤ m∞ (r) 1−p , ∀ p ∈ (0, 1), (11.5.5)
where the last inequality holds by virtue of the trivial estimate m p (r) ≤ m p (1), valid
for every r ∈ (0, 1), and the assumption m p (1) = 1. On the other hand, for every

x ∈ Ω and every r ∈ 0, dist (x, ∂Ω)

C
|u(x)| ≤ n |u(y)| dy, (11.5.6)
r B(x,r)
and, consequently,

C
|u(z)| ≤ |u(y)| dy
(r − ρ)n B(z,r−ρ)

C
≤ |u(y)| dy (11.5.7)
(r − ρ)n B(0,r)
whenever |z| = ρ ∈ (0, r). Then for any z∗ ∈ B(0, ρ) such that |z∗ | = ρ∗ < ρ we obtain

C
|u(z∗ )| ≤ |u(y)| dy
(r − ρ∗ )n B(0,r)

C
≤ |u(y)| dy, (11.5.8)
(r − ρ)n B(0,r)
which, in concert with (11.5.5), yields the estimate

C
m∞ (ρ) ≤ m∞ (r)1−p . (11.5.9)
(r − ρ)n
To continue, set ρ := ra for some a ∈ (1, ∞) to be specified momentarily. Then
(11.5.9) entails
1 1
a dr 1 dr
ln m∞ (r ) ≤C+n ln
1/2 r 1/2 (r − r a) r
1
dr
+ (1 − p) ln m∞ (r) , (11.5.10)
1/2 r
and for the first integral above the change of variables t := ra gives
1 1
dr 1 dt
ln m∞ (ra ) = ln m∞ (t) . (11.5.11)
1/2 r a (1/2)a t
Since our assumption |u(0)| > 1 implies m∞ (t) ≥ 1, the right-hand side of (11.5.11)
is bounded from below by
1
1 dr
ln m∞ (r) . (11.5.12)
a 1/2 r
Therefore, (11.5.10)–(11.5.12) imply

1 1
dr
1
1 dr
−1+ p ln m∞ (r) ≤C +C ln
a 1/2 r 1/2 (r − ra ) r
≤ C < ∞. (11.5.13)
Choose now a > 1 such that 1

a − 1 + p > 0. Then (11.5.13) forces
1
ln m∞ (r) dr ≤ C, (11.5.14)
1/2
and hence, ln m∞ (1/2) ≤ C for some finite constant C > 0 independent of initial
function u. In concert with the inequality |u(0)| ≤ m∞ (1/2), this finishes the proof
of the lemma.
There are certain connections between the subaveraging property and reverse
Hölder estimates, brought to light by our next two results.
Lemma 11.11. Let u be a subaveraging function. Then for every p, q ∈ (0, ∞) and
λ ∈ (0, 1) the following reverse Hölder estimate holds
1q 1p
− |u(y)| dy ≤ C −
q
|u(y)| p dy , (11.5.15)
B(x,λr) B(x,r)
for x ∈ Ω and 0 < r < dist (x, ∂Ω), where C > 0 is a finite constant depending only
on p, q, λ, n and the p-subaveraging constant of u.
Proof. If x ∈ Ω and 0 < r < dist (x, ∂Ω), we write
1q
− |u(y)|q dy ≤ sup |u(z)|
B(x,λr) z∈B(x,λr)
1p
≤C − |u(y)| p dy , (11.5.16)
B(x,r)
by Lemma 11.9.
The main result in this section is the combination of interior estimates and reverse
Hölder estimates contained in the following theorem.
Theorem 11.12. Let n, m, M ∈ N and suppose L is an M × M constant (complex)

coefficient system in Rn , homogeneous of order 2m, and with the property that
det [L(ξ)] 0 for each ξ ∈ Rn \ {0}. Assume also that Ω ⊆ Rn is an open set

and u ∈ D (Ω) M is such that Lu = 0 in D (Ω) M .

Then u ∈ C ∞ (Ω) M and u is subaveraging. Moreover, there exists a constant
C = C(L, n) ∈ (0, ∞) with the property that given p ∈ (0, ∞) there exists some
c = c(L, n, p) ∈ (0, ∞) satisfying

C |α| |α|! 1/p
max |∂α u(y)| ≤ c (1 − λ)−|α|−n/p |α| − |u(y)| p dy (11.5.17)
y∈B(x,λr) r B(x,r)
whenever x ∈ Ω, 0 < r < dist (x, ∂Ω), λ ∈ (0, 1), and α ∈ Nn0 . In particular, the
components of u are real-analytic in Ω.

Proof. As in the past, Theorem 10.9 gives that u ∈ C ∞ (Ω) M . By working with
the system L defined as in (11.4.20) and the function u defined as in (11.4.24), the
same type of reasoning as in the proof of Theorem 11.7 shows that estimate (11.4.5)
actually holds without any restrictions on the dimension n. Keeping this in mind,
the version of this estimate corresponding to μ = (0, . . . , 0) may be interpreted as
saying that u is 1-subaveraging. Hence, by Lemma 11.10 and the ensuing remark,
it follows that u is subaveraging. Based on this and (11.4.13), for each x ∈ Ω, each

r ∈ 0, dist (x, ∂Ω) , each θ, η ∈ (0, 1) and each α ∈ Nn0 we may write
C |α| (1 − θ)−|α| |α|!

max |∂α u(y)| ≤ max |u(y)| (11.5.18)
y∈B(x,θηr) (ηr)|α| y∈B(x,ηr)

C |α| (1 − θ)−|α| |α|! 1/p
≤ c(1 − η)−n/p |α|
− |u(y)| p dy ,
(ηr) B(x,r)
where c = c(L, n, p) ∈ (0, ∞). Next, given any λ ∈ (0, 1), specialize (11.5.18) to
the case when θ := 2λ/(λ + 1) and η := (λ + 1)/2. Note that θ, η ∈ (0, 1), θη = λ,
1 − η = (1 − λ)/2, and 1 − θ = (1 − λ)/(1 + λ). Based on these, we may transform
(11.5.18) into
max |∂α u(y)| (11.5.19)

y∈B(x,λr)
|α| 1/p
−|α|−n/p (2C) |α|!
≤ c2 n/p
(1 − λ) − |u(y)| p dy ,
r|α| B(x,r)
from which (11.5.17) follows after adjusting notation. Also, the last claim in the
statement of the theorem is a consequence of (11.5.17) and Lemma 6.24.
11.6 Layer Potentials and Jump Relations for Systems
The goal of this section is to introduce and study certain types of integral opera-
tors, typically called layer potentials, that are particularly useful in the treatment of
boundary value problems for systems. This is also an excellent opportunity to il-
lustrate how a good understanding of the nature of fundamental solutions coupled
with a versatile command of distribution theory yield powerful tools in the study of
partial differential equations.
To set the stage, we introduce some notation and make some background as-
sumptions. Throughout this section we let

n

L= A jk ∂ j ∂k , A jk = aαβ
jk 1≤α,β≤M ∈ M M×M (C), (11.6.1)
j,k=1
be a second order, homogeneous, complex constant coefficient, M × M system,

where M ∈ N, with the property that its characteristic matrix, L(ξ), defined for
each vector ξ = (ξ1 , . . . , ξn ) ∈ Rn by the formula

n
n
L(ξ) := ξ j ξk A jk = aαβ
jk ξ j ξk ∈ M M×M (C), (11.6.2)
1≤α,β≤M
j,k=1 j,k=1
satisfies the ellipticity condition

det L(ξ) 0, ∀ ξ ∈ Rn \ {0}. (11.6.3)
In particular, since L(e j ) = A j j for each j ∈ {1, . . . , n}, the ellipticity condition
(11.6.3) entails
A j j ∈ M M×M (C) is an invertible matrix for each j ∈ {1, . . . , n}. (11.6.4)
Thanks to (11.6.3), Theorem 11.1 is applicable (with m = 1) and we denote by

E = Eαβ 1≤α,β≤M ∈ M M×M S (Rn ) (11.6.5)
the matrix-valued fundamental solution of L in Rn whose entries are constructed

according to the recipe devised there. Thus, among other things, for each α, β ∈
{1, . . . , M} we have
"
Eαβ ""Rn \{0} is even and belongs to C ∞ (Rn \ {0}), (11.6.6)
∇Eαβ is positive homogeneous of degree 1 − n. (11.6.7)
First, by specializing Theorem 4.79 to the case when Φ is a first-order partial

derivative of E, we obtain the following basic result.
Theorem 11.13. Assume that the system L is as in (11.6.1)–(11.6.3) and suppose
that E is the fundamental solution for L in Rn constructed in Theorem 11.1. Then
11.6 Layer Potentials and Jump Relations for Systems 417
for each j ∈ {1, . . . , n} one has

lim (∂ j E)(x , ε) = ± 12 δ jn δ(x ) Ann −1 + P.V. (∂ j E)(x , 0) (11.6.8)
ε→0±

in M M×M S (Rn−1 ) .
Proof. Fix j ∈ {1, . . . , n}. Then Φ := ∂ j E is of class C ∞ , odd, and positive homo-
geneous of degree 1 − n in Rn \ {0} (cf. (11.6.6)–(11.6.7)). Moreover, by virtue of
(11.3.28) used with m = 1, we have
= ∂ −1
Φ(ξ) j E(ξ) = i ξ j E(ξ) = −i ξ j L(ξ) in M M×M S (Rn−1 ) . (11.6.9)
This further implies that
Φ(0 n ) = −i δ jn L(en )−1 = −i δ jn Ann −1 .

, 1) = Φ(e (11.6.10)
With this in hand, (11.6.8) follows with the help of (4.7.1).

In the next result, we introduce the so-called single layer potential operator asso-
ciated with the system L and study the boundary behavior of its first-order deriva-
tives. Here and elsewhere, the integral of a vector-valued function is applied to each
individual component.
Theorem 11.14. Suppose that the system L is as in (11.6.1)–(11.6.3) and assume
that E is the fundamental solution for L in Rn constructed in Theorem 11.1. Given
any C M -valued function ϕ ∈ S(Rn−1 ), define the C M -valued function

(S ϕ)(x) := E(x − y , t)ϕ(y ) dy (11.6.11)
Rn−1
for each x = (x , t) ∈ Rn with t 0.

Then for any C M -valued function ϕ ∈ S(Rn−1 ),
S ϕ ∈ C ∞ (Rn± ), L(S ϕ) = 0 pointwise in Rn± , (11.6.12)
and for each j ∈ {1, . . . , n} and each x ∈ Rn−1 one has

lim± ∂ j (S ϕ)(x , t) = ± 1
2 δ jn Ann −1 ϕ(x )
t→0

+ lim+ (∂ j E)(x − y , 0)ϕ(y ) dy . (11.6.13)
ε→0
y ∈Rn−1
|x −y |>ε
Proof. Let ϕ ∈ S(Rn−1 ) be an arbitrary C M -valued function. That S ϕ is of class C ∞

in Rn \ {xn = 0} is clear from (11.6.11), (11.6.6), and estimates for the derivatives of
E (see (11.3.27)). In addition, for each x = (x , xn ) ∈ Rn with xn 0,

L(S ϕ)(x) = (LE)(x − y , xn )ϕ(y ) dy = 0 (11.6.14)
Rn−1
since LE = δ I M×M in D (Rn ) which (upon recalling that E ∈ C ∞ (Rn \ {0})) forces
LE = 0 pointwise in Rn± . (11.6.15)
Finally, (11.6.13) follows from Corollary 4.81 applied to the same function Φ used
in the proof of Theorem 11.13.
Remark 11.15. The integral operator S from (11.6.11) is called the single layer po-
tential (or single layer operator) associated to L in the upper and lower half-spaces in
Rn , while (11.6.13) may be naturally regarded as the jump formula for the gradient
of the single layer potential in this setting.
Our next result brings to the forefront another basic integral operator, typically
referred to as the double layer potential.
Theorem 11.16. Let the system L be as in (11.6.1)–(11.6.3) and assume that E is

the fundamental solution for L in Rn constructed in Theorem 11.1. Given any C M -
valued function ϕ ∈ S(Rn−1 ), define for each x = (x , t) ∈ Rn with t 0

n
(Dϕ)(x) := (∂k E)(x − y , t)Akn ϕ(y ) dy . (11.6.16)
Rn−1 k=1
Then, for any C M -valued function ϕ ∈ S(Rn−1 ),
Dϕ ∈ C ∞ (Rn± ), L(Dϕ) = 0 pointwise in Rn± , (11.6.17)
and for every x ∈ Rn−1 one has
lim (Dϕ)(x , t) = ± 1
2 ϕ(x )
t→0±

n
+ lim+ (∂k E)(x − y , 0)Akn ϕ(y ) dy . (11.6.18)
ε→0
k=1
y ∈Rn−1
|x −y |>ε
Proof. Fix some C M -valued function ϕ ∈ S(Rn−1 ). The fact that Dϕ is of class C ∞
in Rn \ {xn = 0} is seen from (11.6.16), (11.6.6), and estimates for the derivatives of
E (cf. (11.3.27)). Moreover, for each x = (x , xn ) ∈ Rn with xn 0,

n

L(Dϕ)(x) = ∂ xk (LE)(x − y , xn ) Akn ϕ(y ) dy = 0 (11.6.19)
Rn−1 k=1
where the last equality uses (11.6.15).

At this stage, there remains to prove (11.6.18). In this regard, if x ∈ Rn−1 has
been fixed, making use of Corollary 11.14 we may write

n
lim± (Dϕ)(x , t) = lim (∂k E)(x − y , t)Akn ϕ(y ) dy
t→0 t→0± Rn−1
k=1

n
= lim ∂k (S Akn ϕ)(x , t)
t→0±
k=1
1 −1
n
=± δkn Ann Akn ϕ(x ) (11.6.20)
2 k=1
n
+ lim+ (∂k E)(x − y , 0)Akn ϕ(y ) dy
ε→0
k=1
y ∈Rn−1
|x −y |>ε

n

= ± 12 ϕ(x ) + lim+ (∂k E)(x − y , 0)Akn ϕ(y ) dy ,
ε→0
k=1
y ∈Rn−1
|x −y |>ε
as wanted.
Remark 11.17. The mapping D from (11.6.16) is called the double layer potential
(or double layer operator) associated to the system L in Rn± and (11.6.18) may be
naturally regarded as the jump formula for the double layer potential in this setting.
When dealing with the so-called Neumann boundary value problem for the sys-
tem L (discussed later in (11.6.34)), as boundary condition one prescribes a certain
n
combination of first-order derivatives of the unknown function, namely − Ank ∂k ,
k=1
amounting to what is called the conormal derivative associated with the system
L. Our next result elaborates on the nature of boundary behavior of this conormal
derivative of the single layer potential S introduced earlier in (11.6.11).
Corollary 11.18. Assume that the system L is as in (11.6.1)–(11.6.3) and let E be

the fundamental solution for L in Rn constructed in Theorem 11.1. Then for every
C M -valued function ϕ ∈ S(Rn−1 ),

n
lim± Ank ∂k (S ϕ)(x , t) = ± 12 ϕ(x ) (11.6.21)
t→0
k=1

n
+ lim+ Ank (∂k E)(x − y , 0)ϕ(y ) dy .
ε→0
k=1
y ∈Rn−1
|x −y |>ε
Proof. For each C M -valued function ϕ ∈ S(Rn−1 ), formula (11.6.21) follows using
Corollary 11.14 by writing

n
n
lim± Ank ∂k (S ϕ)(x , t) = Ank lim± ∂k (S ϕ)(x , t)
t→0 t→0
k=1 k=1
1
n

=± Ank δkn Ann −1 ϕ(x ) (11.6.22)
2 k=1
n
+ lim+ Ank (∂k E)(x − y , 0)ϕ(y ) dy
ε→0
k=1
y ∈Rn−1
|x −y |>ε

n

= ± 12 ϕ(x ) + lim+ Ank (∂k E)(x − y , 0)ϕ(y ) dy ,
ε→0
k=1
y ∈Rn−1
|x −y |>ε
at every point x ∈ Rn−1 .
Exercise 11.19. State and prove the analogue of Theorem 4.100 in the scenario
when Θ : Rn \ {0} → M M×M (C), for some M ∈ N, is a matrix-valued function
whose scalar entries satisfy the conditions in (4.4.1). In particular, pay attention to
the fact that the format of (4.9.43) now becomes
∗ M
TΘ = T Θ ∨ in L2 (Rn ) (11.6.23)
where, as before, Θ denotes the transposed of the matrix Θ.

Hint: Prove and use the fact that mΘ = mΘ .

The fact that for any ϕ ∈ S(Rn−1 ) M we have L(Dϕ) = 0 in Rn± (cf. (11.6.17))
means that we may regard the double layer potential operator D defined in (11.6.16)
as a mechanism for generating an abundance of solutions of the partial differential
equation Lu = 0 in Rn± . In addition, the jump formula (11.6.18) fully clarifies the
boundary behavior of the function
u := Dϕ. (11.6.24)
This is particularly relevant in the context of the Dirichlet problem for the M × M
system L (assumed to be as in (11.6.1)–(11.6.3)), whose formulation in, say, the
upper-half space reads (compare with (4.8.17) in the case L := Δ)
⎧
⎪
⎪
⎪ u ∈ C ∞ (Rn+ ) M ,
⎪
⎪
⎪
⎪
⎨ Lu = 0 in Rn+ ,
⎪
⎪
⎪
(11.6.25)
⎪
⎪ "
"ver
⎪
⎪
⎩ u"" n = ψ on Rn−1 ≡ ∂Rn+ .
∂R+

Above, ψ ∈ S(Rn−1 ) M is a given function" (called the boundary datum) and, much
"ver
as in the case of (4.8.17), the symbol u"" n stands for the “vertical limit” of u
∂R+
to the boundary of the upper-half space, understood at each point x ∈ Rn−1 as

lim+ u(x , xn ).
xn →0
Indeed, focusing the search for a solution of (11.6.25) in the class of functions

u defined as in (11.6.24) (with ϕ ∈ S(Rn−1 ) M yet to be determined) has the dis-
tinct advantage that the first two conditions in (11.6.25) are automatically satisfied
(thanks to (11.6.17)), irrespective of the choice of ϕ. Keeping in mind (11.6.18),
matters have been therefore reduced to solving the boundary integral equation

n
1
2 ϕ(x ) + lim+ (∂k E)(x − y , 0)Akn ϕ(y ) dy = ψ(x ), (11.6.26)
ε→0
k=1
y ∈Rn−1
|x −y |>ε
with x ∈ Rn−1 , so named since Rn−1 may be regarded as the boundary of Rn+ . To
streamline notation, introduce the singular integral operator

n

(Kϕ)(x ) := lim+ (∂k E)(x − y , 0)Akn ϕ(y ) dy , (11.6.27)
ε→0
k=1
y ∈Rn−1
|x −y |>ε
for each x ∈ Rn−1 , typically referred to as the principal value (or boundary version)
of the double layer associated with the system L in the upper-half space. Employing
(11.6.27), we may write in place of (11.6.26)
1
2 I M×M +K ϕ=ψ in Rn−1 . (11.6.28)
Reducing the entire boundary value problem (11.6.25) to solving the boundary inte-
gral equation (11.6.28) is perhaps the most distinguished feature of the technology
described thus far for dealing with (11.6.25), called the method of boundary layer
potentials (in standard partial differential equations parlance).
Regarding the boundary integral equation (11.6.28), involving the singular inte-
gral operator K introduced in (11.6.27), we wish to note that by taking the Fourier
transform (in Rn−1 ), then invoking Theorem 4.74 (whose applicability in the present
setting is justified by virtue of the properties of E from (11.6.6)–(11.6.7)), we arrive
at
1
ϕ=
2 I M×M + aL ψ in S (Rn−1 ), (11.6.29)
where the M × M matrix-valued function aL is defined by the formula
n

aL (ξ ) := − (∂k E)(ω , 0) log(i(ξ · ω )) dσ(ω ) Akn (11.6.30)
k=1 S n−2
for each ξ ∈ Rn−1 \ {0 }. Thus, in the case when the matrix 12 I M×M + aL is invertible,
at least at the formal level we may express the solution of (11.6.29) in the form

ϕ = F −1 12 I M×M + aL −1
ψ, (11.6.31)
then ultimately conclude from (11.6.24) and (11.6.31) that

u = D F −1 12 I M×M + aL −1
ψ (11.6.32)
solves the Dirichlet problem (11.6.25) for the system L in Rn+ with ψ as boundary
datum.
In certain concrete cases of practical importance, the “symbol” function aL from
(11.6.30) is simple enough in order for us to make sense of the expression appearing
in the right-hand side of (11.6.31). For example, it is visible from (11.6.30) that
aL = 0 in Rn−1 \ {0 } whenever

n
(∂k E)(x , 0)Akn = 0 for all x ∈ Rn−1 \ {0 }. (11.6.33)
k=1
This is indeed the case when L = Δ, the Laplacian in Rn . To see that this happens,
note that in this scalar (M = 1) case, A jk = δ jk for each j, k ∈ {1, . . . , n}, and if EΔ
is the fundamental solution for the Laplacian from (7.1.12) then (∂n EΔ )(x , xn ) =

ωn−1 |x|n for each x = (x , xn ) ∈ R \ {0}, hence (∂n E Δ )(x , 0) = 0 for each
1 xn n
x ∈ Rn−1 \ {0 }. The bottom line is that the symbol function aΔ vanishes iden-
tically. In light of (11.6.31), this shows that for the Laplacian we must chose the
(originally unspecified) function ϕ to be equal to the (given) boundary datum ψ.
For this choice, the general formula (11.6.32) then reduces precisely to the classical
expression (4.8.19, as the reader may verify without difficulty. −1
−1 1
However, in general, the assignment ψ → F 2 I M×M + aL ψ can be of an
intricate nature, leading to operators that are well beyond the class of singular inte-
gral operators introduced in Definition 4.93. This leads to the consideration of more
exotic classes of operators, such as pseudodifferential operators, Fourier integral op-
erators, etc., which are outside of the scope of the present monograph. This being
said, the material developed here serves both as preparation and motivation for the
reader interested in further pursuing such matters.
Similar considerations apply in the case of the Neumann problem for a system L
as in (11.6.1)–(11.6.3), whose formulation reads
⎧
⎪
⎪
⎪ u ∈ C ∞ (Rn+ ) M ,
⎪
⎪
⎪
⎪
⎪
⎨ Lu = 0 in Rn+ ,
⎪
⎪
⎪
(11.6.34)
⎪
⎪ ""ver
⎪
⎪
n
⎩ − Ank ∂k u ""∂Rn = ψ on R ≡ ∂R+ ,
⎪ n−1 n
k=1 +

where ψ ∈ S(Rn−1 ) M is the boundary datum. Granted the results proved in Theo-
rem 11.14 and Corollary 11.18, this time it is natural to seek a solution in the form
(compare with (11.6.24))
u := S ϕ in Rn+ , (11.6.35)
n−1 M
where the function ϕ ∈ S(R ) is subject to the boundary integral equation
(implied by (11.6.21))

n
− 21 ϕ(x ) − lim+ Ank (∂k E)(x − y , 0)ϕ(y ) dy = ψ(x ) (11.6.36)
ε→0
k=1
y ∈Rn−1
|x −y |>ε
for each x ∈ Rn−1 . Once again, Theorem 4.74 may be used in order to rewrite
(11.6.36) on the Fourier transform side as

ϕ=
− 12 I M×M + aL ψ in S (Rn−1 ), (11.6.37)
where the M × M matrix-valued function aL is now defined by the formula

n

aL (ξ ) := Ank (∂k E)(ω , 0) log(i(ξ · ω )) dσ(ω ) (11.6.38)
k=1 S n−2
for each ξ ∈ Rn−1 \ {0 }. As regards the symbol function aL for the Neumann
problem, one can see from (11.6.38), (11.6.30), and (11.3.32) that

a L = − a L (11.6.39)
which shows that aL has, up to transpositions, the same nature as the symbol func-
tion for the Dirichlet problem. Hence, the same type of considerations as in the
latter case apply. For example, corresponding to the case when L = Δ we have
aΔ = 0 which means that a solution to the Neumann problem for the Laplacian in
the upper-half space, i.e.,
⎧
⎪
⎪
⎪ u ∈ C ∞ (Rn+ ),
⎪
⎪
⎪
⎪
⎨ Δu = 0 in Rn+ ,
⎪
⎪
⎪ "
(11.6.40)
⎪
⎪
⎪ "ver
⎩ − ∂n u "" n = ψ on Rn−1 ≡ ∂Rn+ ,
⎪
∂R+
where ψ ∈ S(Rn−1 ) is the boundary datum, is given by

u(x) := EΔ (x − y , xn )ψ(y ) dy for x = (x , xn ) ∈ Rn+ , (11.6.41)
Rn−1
with EΔ denoting the fundamental solution for the Laplacian (cf. (7.1.12)).
Further Notes for Chapter 11. As already mentioned, the approach developed in Section 10.2
contains a constructive procedure for reducing the task of finding a fundamental solution for the
Lamé operator to the scalar case. This scheme has been implemented in Section 11.1 based on
the knowledge of a fundamental solution for the polyharmonic operator from Section 7.5. Sub-
sequently, in Section 11.2 a fundamental solution for the Stokes operator is computed indirectly
by rigorously making sense of the informal observation that this operator is a limiting case of the
Lamé operator (taking μ = 1 and sending λ → ∞).
From the considerations in Section 10.2 it was also already known that the higher order
homogeneous elliptic constant coefficient linear systems considered in Section 11.3 posses funda-
mental solutions that are tempered distributions with singular support at the origin. The new issue
addressed in Theorem 11.1 is to find an explicit formula and to study other properties of such
fundamental solutions. This theorem refines and further builds on the results proved in [34], [38],
[57], [59], and [67], in various degrees of generality. Fundamental solutions for variable coefficient
elliptic operators on manifolds have been studied in [55].
Exercise 11.20. Use Theorem 11.1 in a similar manner as in the proof of Propo-
sition 11.3 in order to derive a formula for a fundamental solution for the Lamé
operator in Rn when n ≥ 3 is arbitrary and odd.
Exercise 11.21. Similarly to Exercise 11.20, derive a formula for a fundamental
solution for the Lamé operator in Rn when n is even.
Exercise 11.22. Let L be a homogeneous differential operator in Rn , n ≥ 2, of order
2m, m ∈ N, with complex constant coefficients as in (11.3.1) that satisfies (11.3.3),
and recall (11.3.41). Also, let q ∈ N0 be such that n + q is even. Consider the matrix

E := E jk 1≤ j,k≤M with entries given by
x · ξ
−1
E jk (x) = (x · ξ)2m+q log P jk (ξ) dσ(ξ) (11.7.42)
(2πi)n (2m + q)! S n−1 i
for x ∈ Rn \ {0} and j, k ∈ {1, . . . , M}, where log denotes the principal branch of
the complex logarithm defined for points z ∈ C \ {x : x ∈ R, x ≤ 0}. Prove that

E ∈ M M×M (S Rn ) is a fundamental solution for the system Δ(n+q)/2
x L in Rn .
Exercise 11.23. Let L1 , L2 be two homogeneous M × M systems of differential oper-
ators in Rn , n ≥ 2, of orders 2m1 and 2m2 , respectively, where m1 , m2 ∈ N, with
complex constant coefficients, satisfying det L1 (ξ) 0 and det L2 (ξ) 0 for every
ξ ∈ Rn \ {0}. Prove that, with the notational convention employed in part (7) of
Theorem 11.1, one has
L2 EL1 L2 = EL1 + P in S (Rn ), (11.7.43)
where P is zero if either n is odd or 2m1 < n, and P is an M × M matrix whose

entries are homogeneous polynomials of degree 2m1 − n if 2m1 ≥ n.
Exercise 11.24. Recall the notational convention employed in part (7) of Theo-
rem 11.1 as well as Fm,n from (7.5.2). Prove that for every m ∈ N one has
EΔm = Fm,n + P in S (Rn ), (11.7.44)
where P is zero if either n is odd or 2m < n, and P is a homogeneous polynomial of

degree 2m − n if 2m ≥ n.
Chapter 12
Sobolev Spaces
Abstract While Lebesgue spaces play a most basic role in analysis, it is highly desir-
able to consider a scale of spaces which contains provisions for quantifying smooth-
ness (measured in a suitable sense). This is the key feature of the so-called Sobolev
spaces, introduced and studied at some length in this chapter in a completely self-
contained manner. The starting point is the treatment of global L2 -based Sobolev
spaces of arbitrary smoothness in the entire Euclidean space, using the Fourier trans-
form as the main tool. We then proceed to define Sobolev spaces in arbitrary open
sets, both via restriction (which permits the consideration of arbitrary amounts of
smoothness) and in an intrinsic fashion (for integer amounts of smoothness, de-
manding that distributional derivatives up to a certain order are square-integrable in
the respective open set). When the underlying set is a bounded Lipschitz domain,
both these brands of Sobolev spaces (defined intrinsically and via restriction) co-
incide for an integer amount of smoothness. A key role in the proof of this result
is played by Calderón’s extension operator, mapping functions originally defined
in the said Lipschitz domain to the entire Euclidean ambient with preservation of
Sobolev class. Finally, the fractional Sobolev space of order 1/2 is defined on the
boundary of a Lipschitz domain as the space of square-integrable functions satis-
fying a finiteness condition involving a suitable Gagliardo–Slobodeckij semi-norm.
This is then linked to Sobolev spaces in Lipschitz domains via trace and extension
results.
12.1 Global Sobolev Spaces H s (R n), s ∈ R n
Throughout this section we will use the notation
ξ := (1 + |ξ|2 )1/2 , ∀ ξ ∈ Rn . (12.1.1)
Then, for s ∈ R, the Sobolev space H s (Rn ) is defined by

426 12 Sobolev Spaces

H s (Rn ) := u ∈ S (Rn ) : ξ s
u ∈ L2 (Rn ) (12.1.2)
where u is the Fourier transform of u (cf. Proposition 4.21). Note that for each distri-
bution u ∈ S (Rn ) we have u ∈ S (Rn ) and ξ s
u is understood as the multiplication
between the slowly increasing function ξ s (cf. Exercise 3.13) and the tempered
distribution u (see (b) in Theorem 4.14). Hence, we are demanding that the tem-
pered distribution ξ su is given by a square-integrable function in Rn (cf. (4.1.9) in
this regard). In particular, this implies that
u ∈ Lloc
1
(Rn ).
The space H s (Rn ) introduced in (12.1.2) should be thought of as the collection of
all tempered distributions which, in a suitable sense, have “derivatives up to order s”
in L2 (Rn ). This heuristic principle is substantiated by Theorem 12.24 which shows
this is indeed the case when s is a natural number. In particular, it is natural to think
of the parameter s as being a smoothness index.
We begin by systematically studying the basic properties of the scale of global
Sobolev spaces just introduced. For starters, it is easy to check that H s (Rn ) is a
complex vector space. When endowed with an appropriate inner product, H s (Rn )
becomes a Hilbert space. This issue is discussed in the next theorem which also
contains an important density result.
Theorem 12.1. Let s ∈ R. Then the following are true.
(1) Define

(u, v)H s (Rn ) := ξ2s
u(ξ)
v(ξ) dξ, ∀ u, v ∈ H s (Rn ). (12.1.3)
Rn
Then this is an inner product on H s (Rn ) with respect to which H s (Rn ) is a Hilbert
space, with norm

uL2 (Rn ) ,
uH s (Rn ) := (u, u)H s (Rn ) = ξ s ∀ u ∈ H s (Rn ). (12.1.4)
(2) The following topological inclusions hold: S(Rn ) → H s (Rn ) → S (Rn ).

(3) The space C0∞ (Rn ) is dense in H s (Rn ).
Remark 12.2. From items (1)–(2) in Theorem 12.1 and part (b) of Theorem 14.1
it follows that whenever {u j } j∈N is a sequence that converges both in H s1 (Rn ) and
in H s2 (Rn ), for some real numbers s1 , s2 , then necessarily the limits coincide (as
tempered distributions).
Before presenting the proof of Theorem 12.1 we isolate a number of technical
results in the next three lemmas.
Lemma 12.3. The following are true.
(1) If s1 , s2 ∈ R then ξ s1 +s2 = ξ s1 ξ s2 for every ξ ∈ Rn .
(2) If s ∈ R then ξ + η s ≤ 2|s|/2 ξ s η|s| for every ξ, η ∈ Rn .
Proof. The identity in item (1) is immediate. To prove the inequality in item (2),
note that for each ξ, η ∈ Rn we have 1 + |ξ + η|2 ≤ 2(1 + |ξ|2 )(1 + |η|2 ). If s ≥ 0 the
12.1 Global Sobolev Spaces H s (Rn ), s ∈ Rn 427
latter inequality raised to power s/2 gives the inequality in (2). Also, if s < 0, from
what we just proved (written for −s > 0) it follows that ξ + η−s ≤ 2−s/2 ξ−s η−s
for every ξ, η ∈ Rn . In particular, replacing ξ by ξ + η and η by −η, we obtain
ξ−s ≤ 2−s/2 ξ + η−s η−s for every ξ, η ∈ Rn . The desired conclusion follows since
−s = |s|.
Lemma 12.4. Let s ∈ R. Define the space

L2s (Rn ) := v : Rn → C : ξ s v ∈ L2 (Rn ) , (12.1.5)
and consider
vL2s (Rn ) := ξ s vL2 (Rn ) , ∀ v ∈ L2s (Rn ). (12.1.6)
Then the following are true.
(1) The expression in (12.1.6) defines a norm on L2s (Rn ), which is actually induced
by the inner product

(u, v)L2s (Rn ) := ξ2s u(ξ)v(ξ) dξ, ∀ u, v ∈ L2s (Rn ) (12.1.7)
Rn
with respect to which L2s (Rn ) becomes a Hilbert space.

(2) The space of Schwartz functions S(Rn ) is continuously imbedded in L2s (Rn ).
(3) The identification of functions from L2s (Rn ) with tempered distributions via inte-
gration against Schwartz functions induces a continuous embedding of the space
L2s (Rn ) into S (Rn ).
(4) Let ϕ ∈ S(Rn ), v ∈ L2s (Rn ), and define the function

f (ξ) := ϕ(ξ − η)v(η) dη, ∀ ξ ∈ Rn . (12.1.8)
Rn
Then f is well-defined (via an absolutely convergent integral for each ξ ∈ Rn ),

and may be alternatively expressed as

f (ξ) = ϕ(η)v(ξ − η) dη, ∀ ξ ∈ Rn . (12.1.9)
Rn
Moreover, f belongs to the space of slowly increasing functions L(Rn ) (in par-
ticular, f induces a tempered distribution via integration against Schwartz func-
tions), and for each multi-index α ∈ Nn0 the following formula holds

α
(∂ f )(ξ) = (∂α ϕ)(ξ − η)v(η) dη, ∀ ξ ∈ Rn . (12.1.10)
Rn
(5) Fix ϕ ∈ S(Rn ) and v ∈ L2s (Rn ). Regarding v as a tempered distribution (cf. item
(3) above) define ϕ ∗ v ∈ S (Rn ) in the sense of item (e) in Theorem 4.19. Then,
with f as in item (4) above, the following is true
ϕ ∗ v = f in S (Rn ). (12.1.11)
Proof. That (12.1.6) is a norm on L2s (Rn ) is easily seen from the properties of the
norm on L2 (Rn ). Also, it is immediate that the norm (12.1.6) is induced by the inner
product (12.1.7). To prove that L2s (Rn ) is complete with respect to the norm (12.1.6),
consider the mapping
L2s (Rn ) v → ξ s v ∈ L2 (Rn ). (12.1.12)
Clearly, this mapping is well-defined, injective, and an isometry. We claim that it
is also surjective. Indeed, if w ∈ L2 (Rn ) then the function v := ξ−s w belongs to
L2s (Rn ) and satisfies ξ s v = w. Hence, the mapping in (12.1.12) is a bijective isom-
etry. Consequently, since L2 (Rn ) is complete, we have that L2s (Rn ) is also complete.
As regards the statement in item (2), let ϕ ∈ S(Rn ). Choose a positive number m
satisfying m > n/2 + s and estimate
1/2

ξ2s |ϕ(ξ)|2 dξ ≤ sup (1 + |ξ|)m |ϕ(ξ)| × (12.1.13)
Rn ξ∈Rn
1/2
× ξ2(s−m) dξ < ∞,
Rn
where for the last inequality in (12.1.13) we have used Remark 3.4. This shows that
S(Rn ) ⊆ L2s (Rn ). In addition, the estimate in (12.1.13) and Exercise 3.8 also give
that the latter embedding is sequentially continuous, thus continuous (recall that
S(Rn ) is metrizable; cf. Fact 3.6). This completes the proof of the statement in item
(2).
Moving on to item (3), let v ∈ L2s (Rn ) and ϕ ∈ S(Rn ). By item (2) (used with −s
in place of s) we know that ϕ ∈ L−s 2
(Rn ). Based on Hölder’s inequality, (12.1.13)
(with −s in place of s, and m > n/2 − s), and Remark 3.4 we may estimate

|v(ξ)ϕ(ξ)| dξ ≤ vL2s (Rn ) ϕL−s
2 (Rn )
Rn

≤ CvL2s (Rn ) sup (1 + |ξ|)m |ϕ(ξ)| < ∞, (12.1.14)
ξ∈Rn
for a finite constant C = C(s, n) > 0 independent of ϕ. This proves that the measur-
able function vϕ is absolutely integrable in Rn and that

≤ Cv 2

L s (Rn ) sup (1 + |ξ|) |ϕ(ξ)| .
m
v(ξ)ϕ(ξ) dξ (12.1.15)
Rn ξ∈Rn
By invoking Exercise 4.2 we conclude that v induces a tempered distribution de-

fined via integration against Schwartz functions. Finally, the assignment taking any
function v ∈ L2s (Rn ) into the tempered distribution

S(Rn ) ϕ → v(ξ)ϕ(ξ) dξ ∈ C (12.1.16)
Rn
is linear and injective (thanks to Theorem 1.3). In addition, estimate (12.1.15) and
Fact 4.11 imply that the respective embedding is sequentially continuous, thus con-
tinuous, as wanted. This shows that we have the embedding claimed in item (3).
To prove the statement in item (4), fix ϕ ∈ S(Rn ) and v ∈ L2s (Rn ). Also, pick
ξ ∈ Rn . Based on the Cauchy–Schwarz inequality, item (2) in Lemma 12.3, and the
current item (2) we may estimate
1/2
|ϕ(ξ − η)v(η)| dη ≤ vL2s (Rn ) η−2s |ϕ(ξ − η)|2 dη
Rn Rn
≤ 2|s| ξ−2s vL2s (Rn ) ϕL|s|2 (Rn ) < ∞. (12.1.17)

Thus, the integral Rn ϕ(ξ − η)v(η) dη is absolutely convergent for each ξ ∈ Rn .
Hence, f is well defined and a simple change of variables shows that (12.1.9) holds.
In addition, (12.1.17) gives that, if C := 2|s| vL2s (Rn ) ϕL|s|2 (Rn ) ∈ [0, ∞), then
| f (ξ)| ≤ Cξ−2s , ∀ ξ ∈ Rn . (12.1.18)
To conclude that f belongs to L(Rn ) there remains to prove that f is smooth and its
derivatives satisfy estimates similar in spirit to (12.1.18). To this end, based on what
we have proved so far and induction over |α| ∈ N, it suffices to show that for each
j ∈ {1, . . . , n} the following formula holds

(∂ j f )(ξ) = (∂ j ϕ)(ξ − η)v(η) dη, ∀ ξ ∈ Rn . (12.1.19)
Rn
With the goal of proving (12.1.19), fix j ∈ {1, . . . , n} and ξ ∈ Rn . From (12.1.17)
(used with ∂ j ϕ in place of ϕ) we know that the integral in (12.1.19) is absolutely
convergent. Next, for each h ∈ [−1, 0) ∪ (0, 1] define
ϕ(ξ + he j − η) − ϕ(ξ − η)
Fh (η) := , ∀ η ∈ Rn . (12.1.20)
h
Since ϕ ∈ C ∞ (Rn ), we have that
lim Fh (η) = (∂ j ϕ)(ξ − η) for every η ∈ Rn , (12.1.21)

h→0
while the Mean Value Theorem gives

|Fh (η)| ≤ sup (∂ j ϕ)(ξ + te j − η) for each η ∈ Rn . (12.1.22)
|t|≤1
Note that if we set c(ξ) := |ξ| + 2 then
1 + |η| ≤ c(ξ)(1 + |ξ + te j − η|), ∀ t ∈ [−1, 1], ∀ η ∈ Rn . (12.1.23)
This, combined with Remark 3.4 applied for m > n/2 + |s|, gives that there exists
some constant C ∈ (0, ∞) depending only on ϕ, m, and ξ, such that

sup (∂ j ϕ)(ξ + te j − η) ≤
C
, ∀ η ∈ Rn . (12.1.24)
|t|≤1 (1 + |η|)m
In concert, (12.1.22) and (12.1.24) imply
η−s (1 + |η|)|s|
sup |Fh (η)||v(η)| ≤ C η s
|v(η)| ≤ C η s |v(η)|
|h|≤1 (1 + |η|)m (1 + |η|)m
1
≤C η s |v(η)|, ∀ η ∈ Rn . (12.1.25)
(1 + |η|)m−|s|
Recalling that η s |v(η)| ∈ L2 (Rn ) and m − |s| > n/2, from (12.1.25) and an ap-
plication of Hölder’s inequality we obtain sup|h|≤1 |Fh ||v| ≤ G for some function
G ∈ L1 (Rn ). The latter, (12.1.21), and Lebesgue’s Dominated Convergence Theo-
rem allow us to write

f (ξ + he j ) − f (ξ)
lim = lim Fh (η)v(η) dη
h→0 h h→0 Rn

= (∂ j ϕ)(ξ − η)v(η) dη. (12.1.26)
Rn
This completes the proof of (12.1.19), finishing the treatment of item (4).
On to item (5), given any ψ ∈ S(Rn ) we may use (4.1.43), item (3) in the current
lemma, the definition of the convolution at the level of functions, Fubini’s theorem,
and (12.1.8) to write

ϕ ∗ v, ψ = v, ϕ∨∗ ψ = v(ξ) ϕ∨∗ ψ (ξ) dξ
Rn

= v(ξ) ϕ(η − ξ)ψ(η) dη dξ
Rn Rn

= ψ(η) ϕ(η − ξ)v(ξ) dξ dη
Rn Rn

= ψ(η) f (η) dη. (12.1.27)
Rn
To justify the applicability of Fubini’s theorem in the fourth equality, note that
Hölder’s inequality, item (2) in Lemma 12.3 which gives
ξ−2s ≤ 2|s| η − ξ−2s η2|s| for ξ, η ∈ Rn , (12.1.28)
item (2) in the current lemma, and Remark 3.4, imply


|v(ξ)||ϕ(η − ξ)||ψ(η)| dξ dη
Rn Rn
1/2
≤ v L2s (Rn ) |ψ(η)| ξ−2s |ϕ(η − ξ)|2 dξ dη
Rn Rn
1/2
≤ 2|s| vL2s (Rn ) |ψ(η)|η2|s| η − ξ−2s |ϕ(η − ξ)|2 dξ dη
Rn Rn
≤ 2|s| vL2s (Rn ) ϕL−s

2 (Rn ) ×

dη
× sup (1 + |η|)2|s|+n+1 |ψ(η)| < ∞. (12.1.29)
η∈Rn Rn (1 + |η|)n+1

Thus Rn Rn v(ξ)ϕ(η − ξ)ψ(η) dξ dη is absolutely convergent, so we may indeed
reverse the order of integration. Having established (12.1.27), we conclude that
(12.1.11) holds. This finishes the proof of the lemma.
Lemma 12.5. If u ∈ H s (Rn ), with s ∈ R arbitrary, then

u, ϕ = (2π)−n ϕ(−ξ) dξ, ∀ ϕ ∈ S(Rn ).
u(ξ) (12.1.30)
Rn
Consequently, for each s ∈ R the following estimate holds:

u, ϕ ≤ (2π)−n
u L2 (Rn )
ϕ L2 (Rn )
s −s
(12.1.31)
∀ u ∈ H (R ), ∀ ϕ ∈ S(R ).
s n n
Proof. For ϕ ∈ S(Rn ) write

∨ ∨ ∨
u, ϕ = (2π)−n u , ϕ = (2π)−n
u, ϕ = (2π)−n u, ξ s ξ−s ϕ

= (2π)−n ξ s ∨
u, ξ−s ϕ

= (2π)−n u(ξ)ξ−s
ξ s ϕ(−ξ) dξ
Rn

= (2π)−n ϕ(−ξ) dξ.
u(ξ) (12.1.32)
Rn
The first and second equality in (12.1.32) use Proposition 4.32 and Proposition 4.21.
The fourth equality is based on the definition of the multiplication of the tempered
∨
u with the slowly increasing function ξ s (keeping in mind that ξ−s ϕ
distribution
belongs to S(R ) by (a) in Theorem 3.14 and (c) in Theorem 3.21). Also, the fifth
n
equality uses the fact that ξ s

u ∈ L2 (Rn ) and Example 4.8. This proves (12.1.30).
Finally, the estimate in (12.1.31) becomes a consequence of (12.1.30), item (2)
in Lemma 12.4, and Hölder’s inequality.
After this preamble we are ready to proceed with the proof of Theorem 12.1.
Proof of Theorem 12.1. The fact that (12.1.3) is an inner product on H s (Rn ) follows
from the properties of the Fourier transform, those of the inner product in L2 (Rn ),
and (12.1.31).
We will show that H s (Rn ) is a Hilbert space by proving that it is isometrically
isomorphic to the Hilbert space L2s (Rn ). The starting point is the observation that

H s (Rn ) = u ∈ S (Rn ) :
u ∈ L2s (Rn ) . (12.1.33)
Then a combination of (12.1.33), item (3) in Lemma 12.4, and part (a) of Theo-
rem 4.26, give that (the restriction of) the Fourier transform F at the level of tem-
pered distributions is a well-defined and injective map from H s (Rn ) into L2s (Rn ).
This map is also surjective since for v ∈ L2s (Rn ), if we take u := F −1 v (where
F −1 is the inverse of the map in part (a) of Theorem 4.26) then u ∈ S (Rn ) and

u = v ∈ L2s (Rn ), thus u ∈ H s (Rn ). In addition, (12.1.4) and (12.1.6) imply that the
Fourier transform is an isometry from H s (Rn ) into L2s (Rn ). In summary,
F : H s (Rn ) → L2s (Rn ) is a bijective isometry. (12.1.34)

Having established that L2s (Rn ) is a Hilbert space (cf. item (1) in Lemma 12.4), from
(12.1.34) we deduce that H s (Rn ) is also a Hilbert space. This proves the statement
in item (1).
The statement in item (2) is a consequence of items (2)–(3) in Lemma 12.4 and
(12.1.34). We are left with proving the statement in item (3). In a first stage, we
observe that as a consequence of the fact that C0∞ (Rn ) is dense in L2 (Rn ), we have
C0∞ (Rn ) ⊆ L2s (Rn ) densely. (12.1.35)
Specifically, if v ∈ L2s (Rn ), then ξ s v ∈ L2 (Rn ), so there exists a sequence {ϕ j } j∈N
in C0∞ (Rn ) that converges to ξ s v in L2 (Rn ). Taking ψ j := ξ−s ϕ j for each j ∈ N,
it follows that {ψ j } j∈N is a sequence of functions in C0∞ (Rn ) which converges to v in
L2s (Rn ).
If we now take u ∈ H s (Rn ), then there exists a sequence {ψ j } j∈N ⊆ C0∞ (Rn )
which converges to u in L2s (Rn ). Hence, by (12.1.34) and Remark 4.23, we have that
{F ψ j } j∈N is a sequence of Schwartz functions which converges to F −1 (
−1
u) = u
in H s (Rn ). This shows that S(Rn ) is dense in H s (Rn ). To finish the proof of the
statement in item (3) we now invoke the first inclusion in item (2) of the current
theorem and the density result from part (d) of Theorem 3.14.
Example 12.6. Corresponding to s = 0, from (12.1.2) and Remark 3.29 we have

H 0 (Rn ) = u ∈ S (Rn ) :
u ∈ L2 (Rn )

= u ∈ S (Rn ) : u ∈ L2 (Rn ) = L2 (Rn ) (12.1.36)
and uH 0 (Rn ) = (2π)n/2 uL2 (Rn ) . Hence, H 0 (Rn ) = L2 (Rn ) with equivalent norms.
The Sobolev spaces H s (Rn ) are stable under multiplication with Schwartz func-
tions and are nested with respect to s. These and other useful properties are proved
next. Let us also note that, as seen from Exercise 12.54, the scale {H s (Rn )} s∈R con-
sists of distinct spaces.
Theorem 12.7. The following are true.
(1) Let s ∈ R and ϕ ∈ S(Rn ). Then for every u ∈ H s (Rn ) one has ϕu ∈ H s (Rn ) and
ϕuH s (Rn ) ≤ C(ϕ, s)uH s (Rn ) , (12.1.37)
for some constant C(ϕ, s) ∈ (0, ∞) independent of u.

(2) If s1 , s2 ∈ R are two numbers satisfying s1 ≤ s2 then H s2 (Rn ) ⊆ H s1 (Rn ) and
uH s1 (Rn ) ≤ uH s2 (Rn ) for every u ∈ H s2 (Rn ). Consequently, the continuous
embedding H s2 (Rn ) → H s1 (Rn ) holds. In particular, the sequence of spaces
{H s (Rn )} s∈R is nested with respect to the smoothness index s.
(3) If P(D) is a constant coefficient operator of order m ∈ N0 (recall (5.1.1)) then
P(D) : H s (Rn ) → H s−m (Rn ) is well-defined and continuous for each s ∈ R.
(4) For each compactly supported distribution u in Rn there exists s ∈ (−∞, 0) with
the property that u belongs to the Sobolev space H s (Rn ). Consequently, we have

E (Rn ) ⊆ s<0 H s (Rn ).
Proof. Let ϕ ∈ S(Rn ) and u ∈ H s (Rn ). Then ϕ ∈ S(Rn ) and u ∈ S (Rn ) and

ϕ ∗u ∈ S (R ) (see (e) in Theorem 4.19). Hence, by item (a) in Theorem 4.35 and
n
(4.2.34) we have ϕ u=

∗ ϕ
u = (2π)2n (ϕu)∨ in S (Rn ). Taking the Fourier transform

of the latter identity (combined with (4.2.34)) yields ϕ ∗ u in S (Rn ).
u = (2π)n ϕ
This, item (5) in Lemma 12.4, item (2) in Lemma 12.3, and the Cauchy–Schwarz
inequality give
s
u(ξ) = (2π)−n ξ s (
ξ ϕ u)(ξ)
ϕ ∗

≤ (2π)−n 2|s|/2 η|s| |
ϕ(η)| ξ − η s |
u(ξ − η)| dη
Rn
1/2
≤ (2π)−n 2|s|/2 η|s| |
ϕ(η)| dη ×
Rn
1/2
× η|s| |
ϕ(η)| ξ − η2s |
u(ξ − η)|2 dη (12.1.38)
Rn
for each ξ ∈ Rn . Squaring the left-most terms in (12.1.38), then integrating them
with respect to ξ over Rn , and finally using Fubini’s theorem and a simple change
of variables further yields
2
ξ s ϕ −2n
u(ξ) dξ ≤ (2π) 2 |s|
η|s| |
ϕ(η)| dη ×
Rn Rn

× η|s| |
ϕ(η)| ξ − η2s |
u(ξ − η)|2 dη dξ
Rn Rn
2 s 2
= (2π)−2n 2|s| η|s| |
ϕ(η)| dη u(z) dz.
z (12.1.39)
Rn Rn
This now implies (12.1.37) with

|s|
C(ϕ, s) := (2π) 2 −n |s|/2 η ϕ(η) dη. (12.1.40)
Rn
Next, let s1 , s2 ∈ R be such that s1 ≤ s2 . Then ξ s1 ≤ ξ s2 for every ξ ∈ Rn .

Hence, if u ∈ H s2 (Rn ) then ξ s2
u ∈ L2 (Rn ) which further forces

u = ξ s1 −s2 ξ s2
ξ s1 u ∈ L∞ (Rn ) · L2 (Rn ) ⊆ L2 (Rn ). (12.1.41)
This proves that u ∈ H s1 (Rn ) and also gives that uH s1 (Rn ) ≤ uH s2 (Rn ) . Ultimately,
H s2 (Rn ) ⊆ H s1 (Rn ) and the embedding H s2 (Rn ) → H s1 (Rn ) is continuous.

Moving on, fix m ∈ N0 and suppose P(D) = aα Dα is a constant (complex)
|α|≤m
coefficient differential operator of order m. Then there exists C ∈ (0, ∞), depending
only on P, such that
|P(ξ)| ≤ Cξm ∀ ξ ∈ Rn . (12.1.42)
Indeed, making use of (3.1.18) we may estimate
|α|
|P(ξ)| ≤ max |aα | |ξ| ≤ max |aα | (1 + |ξ|2 )|α|/2
|α|≤m |α|≤m
|α|≤m |α|≤m
≤ C(1 + |ξ|2 )m/2 , ∀ ξ ∈ Rn , (12.1.43)
for some finite constant C = C(P, n, m) > 0, and (12.1.42) follows. Going further, if

s ∈ R and u ∈ H s (Rn ), then P(D)u ∈ S (Rn ) and P(D)u(ξ) u(ξ) in S (Rn ) (cf.
= P(ξ)
part (b) in Theorem 4.26). As such, we may compute

ξ s−m P(D)u(ξ) = ξ s−m P(ξ)
u(ξ) (12.1.44)
P(ξ)
= u(ξ) ∈ L∞ (Rn ) · L2 (Rn ) ⊆ L2 (Rn ).
ξ s
ξm
This implies
P(D)u ∈ H s−m (Rn ) (12.1.45)
and

P(D)uH s−m (Rn ) = ξ s−m P(D)u(ξ)
2 n
L (R )
(12.1.46)
|P(ξ)|
≤ sup ξ s
uL2 (Rn ) = CuH s (Rn )
ξ∈R n ξ m
for some constant C = C(P) ∈ (0, ∞) independent of u. This establishes the state-
ment in item (3) of the current theorem.
Finally, if u ∈ E (Rn ), then
u ∈ L(Rn ) (cf. item (b) in Theorem 4.35) thus
u is

a smooth function and there exists k ∈ N0 such that supξ∈Rn (1 + |ξ|)−k |
u(ξ)| < ∞
(recall Definition 3.11). Consequently, there exists a constant C ∈ (0, ∞) with the
u(ξ)| ≤ Cξk for all ξ ∈ Rn . In turn, this implies
property that |

ξ |
2s
u(ξ)| dξ ≤ C
2
ξ2s+2k dξ < ∞ (12.1.47)
Rn Rn
whenever 2s + 2k < −n. Thus, u ∈ H s (Rn ) for s < −k − n/2.

In the next theorem we show that under some additional assumptions the embed-
ding from item (2) in Theorem 12.7 is compact.
Theorem 12.8. Let s1 , s2 ∈ R satisfy s1 < s2 and suppose K is a compact set in

Rn . Then the embedding H s2 (Rn ) ∩ EK (Rn ) → H s1 (Rn ) is compact, where EK (Rn )
denotes the collection of distributions in Rn with support contained in K.
We remark that, as seen from Exercise 12.52, the aforementioned embedding is

no longer compact if s1 = s2 .
Proof of Theorem 12.8. Let {u j } j∈N be a sequence in H s2 (Rn ) ∩ EK (Rn ) such that
there exists C0 ∈ (0, ∞) with
u j H s2 (Rn ) ≤ C0 , ∀ j ∈ N. (12.1.48)
The goal is to prove that this sequence contains a subsequence that is convergent in
H s1 (Rn ). To proceed, pick ϕ ∈ C0∞ (Rn ) satisfying ϕ ≡ 1 in a neighborhood of K. Fix
j ∈ N. Then u j = ϕu j in S (Rn ) and, reasoning as in the first part of the proof of
Theorem 12.7, we obtain

uj = ϕ u j = (2π)−n
ϕ∗
uj in S (Rn ). (12.1.49)
This and item (5) in Lemma 12.4 imply that the action of u j on Schwartz functions
is given by integration against the function

f j (ξ) := (2π)−n
ϕ(ξ − η)
u j (η) dη, ∀ ξ ∈ Rn . (12.1.50)
Rn
Fix now α ∈ Nn0 and invoke (4) in Lemma 12.4 together with (b) in Theorem 3.21
to compute

(∂α f j )(ξ) = (2π)−n (∂α
ϕ)(ξ − η)
u j (η) dη
Rn

= (2π)−n (−i)|α| ζ
α ϕ(ξ − η)
u j (η) dη, ∀ ξ ∈ Rn . (12.1.51)
Rn
This identity, item (2) in Lemma 12.3, and the Cauchy–Schwarz inequality, allow
us to further estimate
s α
ξ 2 (∂ f j )(ξ) ≤ (2π)−n 2|s2 |/2 ξ − η|s2 | |ζ
α ϕ(ξ − η)| η s2 |
u j (η)| dη
Rn
≤ (2π)−n 2|s2 |/2 ζ α ϕH |s2 | (Rn ) u j H s2 (Rn )

≤ Cζ α ϕH |s2 | (Rn ) , ∀ ξ ∈ Rn , (12.1.52)
where C := (2π)−n 2|s2 |/2C0 . This goes to show that the sequence {∂α f j } j∈N is uni-
formly bounded on compact sets in Rn . In particular, the sequence { f j } j∈N is equicon-
tinuous at each point in Rn . Hence, Arzelà–Ascoli’s theorem applies (see Theo-
rem 14.24) and proves the existence of a subsequence { f jk }k∈N that converges uni-
formly on compact subsets of Rn to a continuous function. Upon recalling that
fj =
u j for each j ∈ N, this uniform convergence on compacts implies
for each R > 0 and δ > 0 there exists N ∈ N such that

(12.1.53)
u jk (ξ) −
| u j (ξ)| < δ for all ξ ∈ B(0, R) if k, ≥ N.
The claim we make at this stage is that
the sequence {u jk }k∈N is Cauchy in H s1 (Rn ). (12.1.54)
To prove (12.1.54), fix a threshold ε > 0 and observe that since we have s1 − s2 < 0
there exists Rε > 0 such that ξ s1 −s2 < ε/(4C0 ) if |ξ| ≥ Rε . In particular,
ε
ξ s1 = ξ s1 −s2 ξ s2 < ξ s2 , ∀ ξ ∈ Rn \ B(0, Rε ). (12.1.55)
4C0
Then, using (12.1.55) and (12.1.48), for any k, ∈ N we may write
1/2
u jk − u j H s1 (Rn ) = u jk (ξ) −
ξ2s1 | u j (ξ)|2 dξ
Rn
1/2
≤ u jk (ξ) −
ξ2s1 | u j (ξ)|2 dξ
B(0,Rε )

ε 1/2
+ u jk (ξ) −
ξ2s2 | u j (ξ)|2 dξ
4C0 Rn \B(0,Rε )
1/2
≤ u jk (ξ) −
ξ2s1 | u j (ξ)|2 dξ
B(0,Rε )
ε
+ u j − u j H s2 (Rn )
4C0 k
1/2 ε
≤ u jk (ξ) −
ξ2s1 | u j (ξ)|2 dξ + . (12.1.56)
B(0,Rε ) 2
By applying now (12.1.53) with δ := 2ε ξ s1 −1L2 (B(0,Rε ))

and R := Rε , we conclude
that there exists N ∈ N such that the first term in the right-most side of (12.1.56) is
≤ ε/2 if k, ≥ N. When used in concert with (12.1.56), this shows that
u jk − u j H s1 (Rn ) < ε for k, ≥ N. (12.1.57)
The proof of the claim in (12.1.54) is therefore finished. Now we may invoke the
fact that H s1 (Rn ) is complete (cf. item (1) in Theorem 12.1) in order to conclude
that the subsequence {u jk }k∈N of {u j } j∈N is convergent in H s1 (Rn ).
While for arbitrary s ∈ R one cannot talk about pointwise differentiability for
distributions belonging to the Sobolev space H s (Rn ), if s is sufficiently large these
distributions are actually given by differentiable functions. This fact is made precise
in the next theorem.
Theorem 12.9. If s ∈ R and k ∈ N0 are such that k < s − n/2 then H s (Rn ) is
contained in C k (Rn ).
Proof. Suppose s, k satisfy the hypotheses of the theorem and let α ∈ Nn0 such that
|α| ≤ k. Then s − |α| > n/2, thus ξα ξ−s ∈ L2 (Rn ). Let now u ∈ H s (Rn ). Then

ξα
u(ξ) = ξα ξ−s ξ s
u(ξ) ∈ L2 (Rn ) · L2 (Rn ) ⊆ L1 (Rn ). (12.1.58)
By combining (4.2.34), (4.2.33), (4.2.2), and (b) in Theorem 4.26, we obtain

∨
∨
∂α u, ϕ = (2π)−n ∂
α u , ϕ = (2π)−n ∂ α u, ϕ
∨ ,
= (2π)−n i|α| ξα
u, ϕ ∀ ϕ ∈ S(Rn ). (12.1.59)
From (12.1.58) and (4.1.9) (see also Example 4.8), we know that the last pairing in
(12.1.59) is given by an integral, thus

∂α u, ϕ = (2π)−n i|α| ξα
u(ξ)ϕ∨ (ξ) dξ
Rn

= (2π)−n i|α| ξα
u(ξ)eix·ξ ϕ(x) dx dξ, ∀ ϕ ∈ S(Rn ). (12.1.60)
Rn Rn
Combined, (12.1.58) and the fact that ϕ ∈ L1 (Rn ) (cf. Exercise 3.5) ensure that the
double integral in (12.1.60) is absolutely convergent and we may apply Fubini’s
theorem to further write

∂α u, ϕ = (2π)−n i|α| ϕ(x) ξαu(ξ)eix·ξ dξ dx
Rn Rn

= (2π)−n i|α| ϕ(x)F ξα
u(ξ) (−x) dx
Rn

u(ξ) ∨ , ϕ ,
= (2π)−n i|α| F ξα ∀ ϕ ∈ S(Rn ). (12.1.61)
Hence,

u(ξ) ∨ in S (Rn ).
∂α u = (2π)−n i|α| F ξα (12.1.62)
In turn, from (12.1.62), (12.1.58), and (3.1.3) we see that the distributional derivative
∂α u belongs to C 0 (Rn ). Since this conclusion holds for all α ∈ Nn0 with |α| ≤ k, we
may now invoke Proposition 2.109 to obtain that u ∈ C k (Rn ).
A corollary of Theorem 12.9 worth singling out is the fact that

H s (Rn ) ⊂ C ∞ (Rn ). (12.1.63)
s∈R
As seen from Exercise 12.55 this is a strict inclusion.
12.2 Restriction Sobolev Spaces H s (Ω), s ∈ R
Let Ω be an open subset of Rn and let s ∈ R. Then the Sobolev space of order s over
Ω, denoted by H s (Ω), is defined via restriction, as

H s (Ω) := u ∈ D (Ω) : there exists some U ∈ H s (Rn ) (12.2.1)

with the property that U = u in D (Ω) .
Ω
It is immediate that H s (Ω) is a vector space over C. This may be endowed with a
natural norm, as explained in the next lemma.
Lemma 12.10. For each open subset Ω of Rn and each s ∈ R the mapping
H s (Ω) u −→ uH s (Ω) := inf UH s (Rn ) (12.2.2)

U∈H s (Rn ), U|Ω =u
defines a norm on H s (Ω). Moreover, when regarding H s (Ω) equipped with this norm,
the inclusion H s (Ω) → D (Ω) is continuous. (12.2.3)
Proof. The homogeneity, subadditivity, and positivity are consequences of the fact
that · H s (Rn ) is a norm on H s (Rn ). There remains to prove that if u ∈ H s (Ω) is
such that uH s (Ω) = 0 then u = 0 in D (Ω). Consider such a u. Then by (12.2.2)
there exists a sequence {U j } j∈N contained in H s (Rn ) such that U j Ω = u in D (Ω) for
each j ∈ N and lim j→∞ U j H s (Rn ) = 0. For each ϕ ∈ C0∞ (Ω) we have ϕ ∈ C0∞ (Rn )
(considering ϕ extended by zero outside Ω) and

ϕ| ≤ U j H s (Rn )
|u, ϕ| = |U j , ϕL2 (Rn ) , ∀ j ∈ N.
(12.2.4)
−s
The equality in (12.2.4) uses Proposition 2.50, while the inequality is based on
(12.1.31). Passing to the limit with j → ∞ in (12.2.4) we arrive at u, ϕ = 0.
Since ϕ is arbitrary in C0∞ (Ω), we conclude that u = 0 in D (Ω), as wanted.
Finally, the claim in (12.2.3) may be justified based on (12.2.1), (12.2.2), item
(2) in Theorem 12.1, and part (b) of Theorem 14.1.
12.2 Restriction Sobolev Spaces H s (Ω), s ∈ R 439
Example 12.11. Corresponding to s = 0 we have H 0 (Ω) = L2 (Ω) as vector spaces,

and
uH 0 (Ω) = (2π)n/2 uL2 (Ω) , ∀ u ∈ H 0 (Ω). (12.2.5)
Specifically, if u ∈ H (Ω)
0
This is a consequence of Example 12.6 and Remark 3.29.
then there exists U ∈ H 0 (Rn ) = L2 (Rn ) satisfying U Ω = u, hence u belongs to
L2 (Ω). For the opposite inclusion, if u ∈ L2 (Ω) and we denote by u the extension
by zero of u to Rn , then u ∈ L2 (Rn ) = H 0 (Rn ), uΩ = u, thus u ∈ H 0 (Ω). In
addition, for every u ∈ L2 (Ω) we have uH 0 (Rn ) = (2π)n/2
uL2 (Rn ) = (2π)n/2 uL2 (Ω) ,
2 n
and any function U ∈ L (R ) with the property that U Ω = u necessarily satisfies
UL2 (Rn ) ≥ U|Ω L2 (Ω) = uL2 (Ω) . Hence, (12.2.5) holds.
Remark 12.12. As regards the scale of restriction Sobolev spaces defined in this
section, we wish to remark that if Ω ⊆ Rn is an open set and s ∈ R, then the
restriction operator RΩ : D (Rn ) → D (Ω), defined by RΩ (U) := U Ω for every
U ∈ D (Rn ), induces a well-defined, linear, continuous, and surjective map
RΩ : H s (Rn ) → H s (Ω). (12.2.6)
That (12.2.6) is well-defined, linear, and surjective follows from the definition of
H s (Ω), while its continuity is a consequence of (12.2.2).
Remark 12.13. As a consequence of (12.2.3), if s1 , s2 ∈ R and {u j } j∈N is a sequence
that converges both in H s1 (Ω) and in H s2 (Ω), then the limits coincide (as distribu-
tions in Ω).
Theorem 12.14. Let Ω be an open subset of Rn and let s ∈ R. Then the following
are true.
(1) The space H s (Ω) endowed with the norm · H s (Ω) is complete, hence Banach.

(2) The set C0∞ (Ω) := ϕΩ : ϕ ∈ C0∞ (Rn ) is dense in H s (Ω) with respect to the norm
· H s (Ω) .

Proof. Let A := {v ∈ H s (Rn ) : vΩ = 0 in D (Ω)}. Then A is a closed subspace
s n
of the Banach space H (R ), · H s (Rn ) . Hence the quotient space H s (Rn )/A is a
Banach space when equipped with the natural quotient norm

[U]H s (Rn )/A := inf U − vH s (Rn ) , ∀ [U] ∈ H s (Rn )/A, (12.2.7)
v∈A
where [U] denotes the class of U ∈ H s (Rn ) in the space H s (Rn )/A. Define the
mapping
H s (Ω) u −→ [Uu ] ∈ H s (Rn )/A, (12.2.8)

where for u ∈ H s (Ω) we denote by Uu an element in H s (Rn ) with
Uu Ω = u in

D (Ω). This map is well defined since if U, V ∈ H (R ) satisfy U Ω = u = V Ω in
s n
D (Ω) then U −V ∈ A, so [U] = [V]. The map (12.2.8) is clearly linear and injective.
To see that it is also surjective, note that if [U] ∈ H s (Rn )/A and V ∈ H s (Rn ) satisfies

U − V ∈ A, then by setting u := U Ω = V Ω we have u ∈ H s (Ω) and u is mapped
by (12.2.8) into [U]. In addition, for each u ∈ H s (Ω), and each Uu ∈ H s (Rn ) with
Uu Ω = u in D (Ω), we have

U ∈ H s (Rn ) : U Ω = u = {Uu − v : v ∈ A}. (12.2.9)
In concert with (12.2.7) and (12.2.2), this ultimately implies that the map (12.2.8) is

an isometry. In summary, the space H s (Ω), · H s (Ω) is isomorphically isometric to
a Banach space, hence it is Banach.
the density statement in (2), let u ∈ H (Ω) and let U ∈ H (R )
s s n
Moving on to
be such that U Ω = u in D (Ω). By (3) in Theorem 12.1, there exists a sequence
{ϕ j } j∈N ⊂ C0∞ (Rn ) that converges to U in H s (Rn ). If we set ψ j := ϕ j Ω for each
j ∈ N, then {ψ j } j∈N ⊂ C0∞ (Ω) and we claim that lim j→∞ ψ j = u in H (Ω). Indeed,
s
for each j ∈ N we have U − ϕ j ∈ H (R ) and (U − ϕ j )Ω = u − ψ j in D (Ω). Hence,

s n
u − ψ j H s (Ω) ≤ U − ϕ j H s (Rn ) , ∀ j ∈ N. (12.2.10)
The desired conclusion follows by passing to the limit in (12.2.10) with j → ∞.

The analogues of the results in Theorem 12.7 and Theorem 12.9 for spaces H s (Ω)
are discussed next.
Theorem 12.15. The following are true.

(1) Let s ∈ R and ϕ ∈ C0∞ (Ω). Then for every u ∈ H s (Ω) we have ϕu ∈ H s (Ω) and
ϕuH s (Ω) ≤ CuH s (Ω) , (12.2.11)
for some constant C ∈ (0, ∞) independent of u.

(2) If s1 , s2 ∈ R satisfying s1 ≤ s2 then H s2 (Ω) ⊆ H s1 (Ω) and
uH s1 (Ω) ≤ uH s2 (Ω) for every u ∈ H s2 (Ω). (12.2.12)
As such, one has the continuous embedding H s2 (Ω) → H s1 (Ω). Consequently,

the sequence of spaces {H s (Ω)} s∈R is nested with respect to s.
(3) If s1 , s2 ∈ R are such that s1 < s2 and Ω is bounded then the embedding
H s2 (Ω) → H s1 (Ω) is in fact compact.

(4) If m ∈ N0 and P(D) = aα Dα with aα ∈ C0∞ (Ω) for all α ∈ Nn0 , |α| ≤ m, then
|α|≤m
P(D) : H s (Ω) → H s−m (Ω) is a well-defined and continuous operator for each
s ∈ R.
(5) If s ∈ R and k ∈ N0 are such that k < s − n/2 then H s (Ω) ⊂ C k (Ω).
∞ ∞
H (Ω) and ϕ ∈ C0 (Ω) then there exist U ∈ H (R ) and Φ ∈ C0 (R )
Proof. If u ∈ s s n n

such that U Ω = u in D (Ω) and ΦΩ = ϕ in Ω. By (1) in Theorem 12.7 we have
ΦU ∈ H s (Rn ) and
ΦUH s (Rn ) ≤ C(Φ, s)UH s (Rn ) . (12.2.13)
12.2 Restriction Sobolev Spaces H s (Ω), s ∈ R 441

Also, (2.5.5) ensures that (ΦU)Ω = ϕu in D (Ω), hence ϕu ∈ H s (Ω). In addition, by
making use of (12.2.13) we have
ϕuH s (Ω) ≤ ΦUH s (Rn ) ≤ CUH s (Rn ) . (12.2.14)

Taking now the infimum over all U ∈ H s (Rn ) with U Ω = u in D (Ω), from (12.2.14)
we obtain
ϕuH s (Ω) ≤ CuH s (Ω) . (12.2.15)
This proves the statement in item (1).
Next, let s1 , s2 ∈ R be such that s1 ≤ s2 and consider u ∈ H (Ω). Then there
s2
exists U ∈ H s2 (Rn ) satisfying U Ω = u in D (Ω) and (2) in Theorem 12.7 implies

U ∈ H s1 (Rn ) and UH s1 (Rn ) ≤ UH s2 (Rn ) . Hence, u ∈ H s1 (Ω) and since we have
uH s1 (Ω) ≤ UH s1 (Rn ) it follows that uH s1 (Ω) ≤ UH s2 (Rn ) . Taking the infimum
over all such U’s we obtain uH s1 (Ω) ≤ uH s2 (Ω) , proving the statement in item (2).
In order to prove the statement in item (3) make the additional assumption that Ω
is bounded. Let {u j } j∈N be a bounded sequence in H s2 (Ω), say u j H s2 (Ω) ≤ C0 , for
some C0 ∈ (0, ∞). Then for each j ∈ N there exists U j ∈ H s2 (Rn ) with the property
that U j Ω = u j in D (Ω) and such that U j H s2 (Rn ) ≤ 2C0 (recall the definition from
(12.2.2)). Pick now ϕ ∈ C0∞ (Rn ) with the property that ϕ ≡ 1 in a neighborhood of
Ω. Then (1) in Theorem 12.7 implies ϕU j ∈ H s2 (Rn ) and the existence of a con-
stant C(ϕ, s2 ) ∈ (0, ∞) such that ϕU j H s2 (Rn ) ≤ 2C0C(ϕ, s2 ) for all j ∈ N. Thus,
the sequence {ϕU j } j∈N is bounded in H s2 (Rn ). Being also a sequence of distribu-
tions supported in supp ϕ, we may invoke Theorem 12.8 to extract a subsequence

{ϕU jk }k∈N which converges in H s1 (Rn ). Since for each k ∈ N we have ϕU jk Ω = u jk
in D (Ω), Remark 12.12 implies that the sequence {u jk }k∈N is convergent in H s1 (Ω)
and finishes the proof of the statement in item (3).
Moving on, observe that having proved the statements in items (1) and (2), to
show that the statement in item (4) is true it suffices to prove that ∂ j is a continuous
operator from H s (Ω) into H s−1 (Ω). To this end, start with u ∈ H s (Ω) and pick some
U ∈ H s (Rn ) such that U Ω = u in D (Ω). We may then apply (3) in Theorem 12.7
to obtain that ∂ j U ∈ H s−1 (Rn ) and
∂ j UH s−1 (Rn ) ≤ UH s (Rn ) . (12.2.16)
(The fact that the norm inequality in (12.2.16) holds as stated is seen from (12.1.46).)
In addition, since distributional derivatives commute with restrictions to open sets
(cf. item (2) in Exercise 2.51), we have

(∂ j U)Ω = ∂ j u in D (Ω). (12.2.17)
In concert, (12.2.2), (12.2.17), and (12.2.16) imply

∀ U ∈ H s (Rn ), U Ω = u.
∂ j uH s−1 (Ω) ≤ UH s (Rn ) , (12.2.18)

Taking now the infimum over all U ∈ H s (Rn ) such that U Ω = u in D (Ω), we obtain
∂ j uH s−1 (Ω) ≤ uH s (Ω) . (12.2.19)
This proves that ∂ j is a continuous operator from H s (Ω) into H s−1 (Ω) and completes
the proof of the statement in item (4).
Finally, the statement in item (5) is an immediate consequence of (12.2.1) and
Theorem 12.9.
12.3 Intrinsic Sobolev Spaces H m(Ω), m ∈ N
Recall the restriction Sobolev spaces considered in §12.2. At least for (positive) inte-
ger amounts of smoothness there is yet another venue, which is intrinsic in nature,
of introducing a natural brand of Sobolev spaces on any given open set. Specifically,
let Ω be an open set in Rn and fix some m ∈ N0 . We then define the intrinsic Sobolev
space H m (Ω) by setting

H m (Ω) := u ∈ L2 (Ω) : ∂α u ∈ L2 (Ω), ∀ α ∈ Nn0 , |α| ≤ m , (12.3.1)
where the partial derivatives are considered in the sense of distributions. As is appar-
ent from the above definition,
H 0 (Ω) = L2 (Ω). (12.3.2)
Also, H m (Ω) is a vector space over C, and it is easy to check that

(u, v)H m (Ω) := ∂α u ∂α v dx, ∀ u, v ∈ H m (Ω), (12.3.3)
α∈Nn0 , |α|≤m Ω
is an inner product on H m (Ω). The norm induced on H m (Ω) by this inner product is

uH m (Ω) := (u, u)H m (Ω)
1/2
= ∂α u2L2 (Ω) , ∀ u ∈ H m (Ω). (12.3.4)
α∈Nn0 , |α|≤m
In particular,
· H 0 (Ω) = · L2 (Ω) , (12.3.5)
and for each u ∈ H m (Ω) we have
∂α uL2 (Ω) ≤ uH m (Ω) , ∀ α ∈ Nn0 , |α| ≤ m. (12.3.6)
Theorem 12.16. For each open subset Ω of Rn and each m ∈ N0 , the space H m (Ω)
endowed with the inner product (12.3.3) is a Hilbert space.
Proof. In order to prove that H m (Ω) is complete with respect to the norm (12.3.4),
consider a Cauchy sequence {u j } j∈N ⊆ H m (Ω). In view of (12.3.6), this implies that
12.3 Intrinsic Sobolev Spaces H m (Ω), m ∈ N 443
for each α ∈ Nn0 with |α| ≤ m the sequence {∂α u j } j∈N is Cauchy in L2 (Ω). Since
L2 (Ω) is complete, for each α ∈ Nn0 with |α| ≤ m there exists uα ∈ L2 (Ω) such that
lim ∂α u j = uα in L2 (Ω). (12.3.7)

j→∞
Abbreviate u := u(0,...,0) . Then, in order to conclude that lim j→∞ u j = u in H m (Ω),

there remains to show that for each α ∈ Nn0 with |α| ≤ m the weak derivative ∂α u
exists and equals uα . Fix such an α and pick ϕ ∈ C0∞ (Ω). We may then compute

∂α u, ϕ = (−1)α u, ∂α ϕ = (−1)α u ∂α ϕ dx
Ω

= lim (−1)α u j ∂α ϕ dx = lim (∂α u j )ϕ dx
j→∞ Ω j→∞ Ω

= uα ϕ dx = uα , ϕ. (12.3.8)
Ω
Above, the first equality is based on the definition of the weak derivative ∂α , the
second equality uses the fact that u ∈ L2 (Ω) ⊆ Lloc 1
(Ω), the third equality is justified
by observing that

u ∂α ϕ dx − u j ∂α ϕ dx ≤ |u − u j ||∂α ϕ| dx (12.3.9)
Ω Ω Ω
≤ u − u j L2 (Ω) ∂α ϕL2 (Ω) → 0 as j → ∞,
(thanks to (12.3.7)), the fourth equality is integration by parts (recall that ϕ has
compact support in Ω), and the fifth equality has a justification similar to (12.3.9).
Hence, ∂α u = uα , and the proof of the theorem is finished.
Remark 12.17. Let Ω1 ⊆ Ω2 be open sets in Rn and let m ∈ N. Then the restriction
operator RΩ2 : D (Ω1 ) → D (Ω2 ) defined by RΩ2 (w) := wΩ in D (Ω2 ) for every
2
w ∈ D (Ω1 ), induces a well-defined, linear, and continuous map
RΩ2 : H m (Ω1 ) → H m (Ω2 ). (12.3.10)
Indeed this is a consequence of (12.3.1), (12.3.4), the fact that distributional deriva-
tives commute with restrictions to open sets, and the continuity of the restriction to
Ω2 as an operator from L2 (Ω1 ) into L2 (Ω2 ).
Exercise 12.18. If Ω is an open set in Rn and m ∈ N0 then the following are true.
(1) For each ϕ ∈ C ∞ (Rn ) with ∂α ϕ bounded for all α ∈ Nn0 with |α| ≤ m (hence, in
particular, for each ϕ ∈ C0∞ (Ω)) the mapping
H m (Ω) u −→ ϕu ∈ H m (Ω) (12.3.11)
is well-defined, linear, and continuous.


(2) If Ω is also bounded then C m Ω ⊂ H m (Ω).
Hint: To prove (1) use the Generalized Leibniz Formula from Proposition 2.49 while
for the claim in (2) recall Exercise 2.40.
Example 12.19. Suppose Ω is an open set in Rn and u ∈ H m (Ω) for some m ∈ N0 .
If ψ ∈ C0∞ (Ω) then ψ u, which is the extension by zero of ψu outside Ω belongs
m n
to H (R ) and ψuH m (Rn ) ≤ CuH m (Ω) for some C ∈ (0, ∞) depending on ψ and
independent of u.
Indeed, if ϕ ∈ C0∞ (Rn ) and ξ ∈ C0∞ (Ω) satisfies ξ ≡ 1 near supp ψ, then for each
α ∈ Nn0 we have that ∂α ϕ − ∂α (ϕξ) vanishes in a neighborhood of the support of ψ u,
thus

(−1)|α| u ∂α ϕ dx = (−1)|α|
ψ u ∂α (ξϕ) dx
ψ
Rn Rn

= (−1)|α| ψu ∂α (ξϕ) dx
Ω

= ∂α (ψu) ξϕ dx = ∂α (ψu) ϕ dx. (12.3.12)
Ω Ω
u) = ∂
This implies that ∂α (ψ α (ψu) in D (Rn ) which further entails

u2 m n =
ψ ∂ α 2
( ψ u) = ∂α (ψu)2L2 (Ω)
H (R ) 2 n
L (R )
|α|≤m |α|≤m
= ψu2H m (Ω) ≤ Cu2H m (Ω) (12.3.13)
for some C ∈ (0, ∞) dependent on ψ and independent of u (for the inequality in

(12.3.13) we used item (1) in Exercise 12.18).
Example 12.20. Suppose Ω0 , Ω1 , and C are open sets in Rn such that Ω0 ∩ C =
Ω1 ∩ C. Fix m ∈ N and suppose u ∈ H m (Ω0 ) is such that u = 0 a.e. in Ω0 \ K for
some compact set K ⊂ C. If we define v : Ω1 → C by
⎧
⎪
⎨ u in Ω1 ∩ C,
⎪
v := ⎪
⎪ (12.3.14)
⎩ 0 in Ω1 \ C,
then v ∈ H m (Ω1 ), for each α ∈ Nn0 with |α| ≤ m we have

⎧ α
⎪
⎨ ∂ u in Ω1 ∩ C,
⎪
α
∂ v=⎪ ⎪ (12.3.15)
⎩ 0 in Ω1 \ C,
and vH m (Ω1 ) = uH m (Ω2 ) .

To see why v has the aforementioned properties fix some ξ ∈ C0∞ (C) with the
property that ξ ≡ 1 near K. Then for each ϕ ∈ C0∞ (Ω1 ) and each α ∈ Nn0 we may
write

(−1)|α| v ∂α ϕ dx = (−1)|α| u ∂α ϕ dx = (−1)|α| u ∂α (ϕξ) dx
Ω1 Ω1 ∩C Ω0 ∩C

= (−1)|α| u ∂α (ϕξ) dx = (∂α u)ϕξ dx
Ω0 Ω0

α
= (∂ u)ϕ dx = (∂α u)ϕ dx. (12.3.16)
Ω0 ∩C Ω1 ∩C
The first equality in (12.3.16) is due to (12.3.14), the second uses the fact that Ω0 ∩
C = Ω1 ∩C and the observation that ∂α ϕ−∂α (ϕξ) is identically zero near K, thus has
support disjoint from the support of u. The third equality uses the support condition
on u. The fourth equality uses the fact that u ∈ H m (Ω0 ) and ϕξ is a smooth function
compactly supported in Ω1 ∩ C = Ω0 ∩ C, hence compactly supported in Ω0 . The
fifth equality is based on the inclusions supp(∂α u) ⊆ supp u ⊆ K ∩C and the fact that
ξ ≡ 1 near K, while the last equality is a consequence of the fact that Ω0 ∩C = Ω1 ∩C.
The resulting equality in (12.3.16) proves (12.3.15). Also, from (12.3.15) we see that

v2H m (Ω1 ) = ∂α u2L2 (Ω1 ∩C)
α∈Nn0 , |α|≤m

= ∂α u2L2 (Ω0 ∩C) = u2H m (Ω0 ) , (12.3.17)
α∈Nn0 , |α|≤m
hence the rest of the claims about v follow.
Moving on, we elaborate on the manner on which the approximation by convo-

lution with a smooth mollifier behaves on the scale of intrinsic Sobolev spaces in
the Euclidean ambient.
Lemma 12.21. Let φ be a function as in (1.2.3) and for each ε > 0 define φε as in
(1.2.4). Also, let u ∈ H m (Rn ) for some m ∈ N and set uε := u ∗ φε for every ε > 0.
H m (Rn )
Then uε ∈ C ∞ (Rn ) ∩ H m (Rn ) for each ε > 0 and uε −−−−−−
→ u.
+ ε→0
Proof. That uε ∈ C ∞ (Rn ) for each ε > 0 is a consequence of Proposition 2.102.

Next, pick a multi-index α ∈ Nn0 such that |α| ≤ m. By part (e) in Theorem 2.96
we have ∂α uε = (∂α u) ∗ φε for each ε > 0. Since by assumption ∂α u ∈ L2 (Rn ),
we may invoke Lemma 14.19 to conclude that ∂α uε ∈ L2 (Rn ) for each ε > 0, and
L2 (Rn )
that ∂α uε −−−−−→
+
∂α u. Granted these, the desired conclusion follows from (12.3.1)–
ε→0
(12.3.4).
Our focus next is proving a couple of density results for intrinsic Sobolev spaces
defined in arbitrary open sets. Here is the first such result.
Theorem 12.22. Let Ω be an open set in Rn and suppose m ∈ N. Then the set
C ∞ (Ω) ∩ H m (Ω) is dense in H m (Ω).
Proof. To get started fix u ∈ H m (Ω) along with some threshold ε∗ > 0. The desired
conclusion then will follow once we show that
there exists a function w ∈ C ∞ (Ω) ∩ H m (Ω)
(12.3.18)
with the property that u − wH m (Ω) < ε∗ .
To this end, for each k ∈ N, define the open set

Ok := x ∈ Ω : k+11
< dist (x, ∂Ω) < k−1
1
and |x| < k . (12.3.19)
This definition ensures that
Ok is open, bounded, Ok ⊂ Ω, for each k ∈ N,

(12.3.20)
and k∈N Ok = Ω.
By Theorem 14.42, there exists a partition of unity subordinate to the family (Ok )k∈N .
Specifically, there exists an at most countable collection (ϕ j ) j∈J of C ∞ functions
ϕ j : Ω → R, neither of which is identically zero, and satisfying the following
properties:
(a) For every j ∈ J one has 0 ≤ ϕ j ≤ 1 in Ω and there exists k ∈ I such that ϕ j is
compactly supported in Ok ;

(b) The family of sets {x ∈ Ω : ϕ j (x) 0} j∈J is locally finite in Ω;

(c) ϕ j (x) = 1 for every x ∈ Ω.
j∈J
Next, for each k ∈ N define
"
k−1
Jk := j ∈ J : supp ϕ j ⊆ Ok \ Oi . (12.3.21)
i=1
We claim that
Jk is finite for each k ∈ N. (12.3.22)
Suppose the claim in (12.3.22) is false. Then there exists k0 ∈ N and an infinite
sequence { ji }i∈N ⊆ Jk0 . Upon recalling that neither function in the partition of unity
is identically zero, for each i ∈ N we may pick xi ∈ Ω such that ϕ ji (xi ) 0. By
definition (12.3.21), it follows that the sequence {xi }i∈N is contained in the compact
set Ok0 , thus it contains a subsequence {xi }∈N convergent to some point x∗ ∈ Ok0 .
In particular,
for each r ∈ (0, ∞) there exists r ∈ N such that

(12.3.23)
xi ∈ B(x∗ , r) ∩ x ∈ Ω : ϕ ji (x) 0} for all ≥ r .
However, by property (b) above, there exists some r0 > 0 such that B(x∗ , r0 ) inter-

sects only finitely many of the sets x ∈ Ω : ϕ ji (x) 0}, ∈ N. The latter is in
contradiction with (12.3.23). This shows that the claim in (12.3.22) must be true.
Having proved (12.3.22), we may define the functions

ψk := ϕ j, for each k ∈ N. (12.3.24)
j∈Jk

Based on (12.3.22), the observation that k∈N Jk = J (where the union is disjoint),
and the properties of the partition of unity recalled earlier we conclude that

ψk ∈ C0∞ (Ok ), ∀ k ∈ N, and ψk = ϕ j = 1 in Ω. (12.3.25)
k∈N j∈J
Next, fix k ∈ N and proceed as follows. First, by Example 12.19, we have

#
ψ k u ∈ H (R ), where tilde denotes extension by zero outside Ω. Second, invok-
m n
ing Lemma 12.21 (with u := ψ # #

k u), we construct the functions (ψ
∞
k u)ε ∈ C (R ), for
n
each ε > 0, with the property that

H m (Rn )
#
(ψk u)ε −
−−−−−
→ψ#k u. (12.3.26)
+ ε→0
#
In addition, since ψk u is compactly supported in Ok , we also have that
#
(ψ k u)ε has compact supported contained in Ok + B(0, ε), (12.3.27)
which is a consequence of the fact that the support of (ψ #k u)ε is contained in

#
supp (ψk u) + supp φε , where φε is as in Lemma 12.21. Third, we define the bounded
open set

Uk := x ∈ Ω : 2(k+1)
1
< dist (x, ∂Ω) < k−1
2
and |x| < 2k . (12.3.28)
This is a slight enlargement of Ok . Invoking (12.3.26) and (12.3.27) we may pick

some εk > 0 sufficiently small to ensure that
ε∗
#
supp (ψ k u)εk ⊆ Uk and #
(ψ #
k u)εk − ψ k uH m (Rn ) < k . (12.3.29)
2
Now we are ready to finish the proof of the theorem. Specifically, set

∞
w := #
(ψ k u)εk Ω in Ω. (12.3.30)
k=1
Note that as a consequence of the definition in (12.3.28) and the support property in
(12.3.29), the sum in (12.3.30) is locally finite, which further implies that w is well
defined and belongs to C ∞ (Ω). In fact, if we consider the open subsets of Ω given
by
Ω j := x ∈ Ω : dist (x, ∂Ω) > 1j , for each j ∈ N, (12.3.31)
then, for each j ∈ N, it follows (from the support property in (12.3.29) and the
definition in (12.3.28)) that

2j
whenever x ∈ Ω j we have w(x) = #
(ψ k u)εk (x). (12.3.32)
k=1
Also, from (12.3.25) and (12.3.19), we may conclude that, for each j ∈ N,

j
2j
whenever x ∈ Ω j we have u(x) = #
(ψk u)(x) = #
(ψ k u)(x) (12.3.33)
k=1 k=1
(note that the terms in the last sum corresponding to k ∈ { j + 1, . . . , 2 j} are zero).
In concert, (12.3.32), (12.3.33), Minkowski’s inequality (cf. Theorem 14.22), and
(12.3.29) imply
1/2
∂α w − ∂α u2L2 (Ω j )
|α|≤m

2j
α 2 1/2
#
∂ (ψ α #
≤ k u)εk − ∂ (ψ k u)L2 (Ω )
j
|α|≤m k=1

2j
2 1/2
#
∂α (ψ α #
≤ k u)εk − ∂ (ψ k u)L2 (Ω )
j
k=1 |α|≤m

2j

≤ #
(ψ #
k u)εk − ψ k uH m (Rn )
k=1

2j
ε∗
≤ ≤ ε∗ , for all j ∈ N. (12.3.34)
k=1
2k
Observing that Ω j Ω as j → ∞ and relying on Lebesgue’s Monotone Conver-

gence Theorem, from (12.3.34) we deduce that
1/2
∂α w − ∂α u2L2 (Ω) ≤ ε∗ . (12.3.35)
|α|≤m
In particular,
1/2
∂α w2L2 (Ω)
|α|≤m
1/2 1/2
≤ ∂α w − ∂α u2L2 (Ω) + ∂α u2L2 (Ω)
|α|≤m |α|≤m
≤ ε∗ + uH m (Ω) < ∞, (12.3.36)

which goes to show that w ∈ H m (Ω). Thus, ultimately, w ∈ C ∞ (Ω) ∩ H m (Ω) and we
may recast (12.3.35) simply as w − uH m (Ω) ≤ ε∗ . This establishes (12.3.18) which
completes the proof of the theorem.
Here is our second density result for intrinsic Sobolev spaces defined in open
sets, advertised earlier.
Theorem 12.23. Suppose Ω ⊆ Rn is an open set and fix m ∈ N. Then the set
{u ∈ H m (Ω): u vanishes outside of a bounded subset of Ω} is dense in the intrinsic
Sobolev space H m (Ω).
Proof. Pick some θ ∈ C0∞ (Rn ) with the property that θ ≡ 1 on B(0, 1) and, for each
R > 0, define θR (x) := θ(x/R), x ∈ Rn . Having fixed a function u ∈ H m (Ω), we
may invoke part (1) of Exercise 12.18 to conclude that for each R > 0 the function
uR := θR u belongs to H m (Ω). In addition, it is clear that uR vanishes outside of a
bounded subset of Ω. As such, the proof is complete as soon as we establish that
H m (Ω)
uR −−−−−→ u. (12.3.37)
R→∞
To justify this, observe that for each α ∈ Nn0 with |α| ≤ m the Generalized Leibniz
Formula from Proposition 2.49 gives
α! β γ
∂α uR = θR ∂α u + ∂ θR ∂ u in D (Ω). (12.3.38)
β+γ=α
β!γ!
|β|>0
Lebesgue’s Dominated Convergence Theorem (cf. Theorem 14.16 for p = 2) en-

L2 (Ω) L2 (Ω)
sures that θR ∂α u −−−−→ ∂α u and that ∂β θR ∂γ u −−−−→ 0 for each β, γ ∈ Nn0 with
R→∞ R→∞
0 < |β| ≤ m. On account of these observations, (12.3.37) follows from (12.3.38).
We next take up the task of determining how the space H m (Ω) relates to H m (Ω)
for m ∈ N0 . First we consider the case Ω = Rn .
Theorem 12.24. For each m ∈ N0 one has H m (Rn ) = H m (Rn ) as vector spaces,
with equivalent norms.
Proof. If u ∈ H m (Rn ), for each multi-index α ∈ Nn0 with |α| ≤ m we may write
(keeping in mind (b) in Theorem 4.26)
ξα
∂
α u(ξ) = ξ α
u(ξ) = ξm
u(ξ)
ξ m
∈ L∞ (Rn ) · L2 (Rn ) ⊆ L2 (Rn ). (12.3.39)
Hence, for each multi-index α ∈ Nn0 with |α| ≤ m we have ∂ α u ∈ L2 (Rn ) and
∂
α u 2 n ≤ Cu m n for some constant C ∈ (0, ∞) independent of u. Via
L (R ) H (R )
Plancherel’s theorem (cf. part (2) in Remark 3.29), this ultimately shows that ∂α u
belongs to L2 (Rn ) and ∂α uL2 (Rn ) ≤ CuH m (Rn ) for each α ∈ Nn0 with |α| ≤ m. Thus,
u ∈ H m (Rn ) and uH m (Rn ) ≤ CuH m (Rn ) . This proves that we have a continuous
embedding H m (Rn ) → H m (Rn ).
To deal with the converse embedding, we first note that (3.1.7) implies that there
exist constants C1 , C2 ∈ (0, ∞) depending only on n and m such that

C1 ξ2m ≤ |ξ2α | ≤ C2 ξ2m , ∀ ξ ∈ Rn , (12.3.40)
|α|≤m
(recall that we adopted the convention that ξ(0,...,0) := 1). If u ∈ H m (Rn ) then
∂α u ∈ L2 (Rn ) for all α ∈ Nn0 with |α| ≤ m. Hence, invoking (3.2.31) and (b) in
Theorem 4.26 we have i|α| ξαu = ∂ α u ∈ L2 (Rn ) for all α ∈ Nn with |α| ≤ m. The
0
latter combined with the first inequality in (12.3.40) yields ξmu ∈ L2 (Rn ), thus
u ∈ H (R ), and the estimate
m n
uH m (Rn ) ≤ (2π)n/2C1−1/2 uH m (Rn ) (12.3.41)
follows after one more use of (3.2.29).

In general, for an open set Ω in R and m ∈ N0 , one cannot expect to have
n
H m (Ω) = H m (Ω). Nonetheless, the left-to-right inclusion always holds.
Proposition 12.25. If Ω is an open set in Rn and m ∈ N0 then H m (Ω) embeds con-

tinuously into H m (Ω).

Proof. Let u ∈ H m (Ω) and pick U ∈ H m (Rn ) with U Ω = u in D (Ω). By The-
orem 12.24 we have U ∈ H m (Rn ). Given that taking restrictions of distributions
to open sets commutes with taking distributional derivatives, the latter implies that
u ∈ H m (Ω) and
uH m (Ω) ≤ UH m (Rn ) . (12.3.42)
Since by Theorem 12.24 there exists a constant C ∈ (0, ∞) independent of U, such
that UH m (Rn ) ≤ CUH m (Rn ) , we may conclude
uH m (Ω) ≤ CUH m (Rn ) . (12.3.43)
(12.2.2), if we take in (12.3.43) the infimum over all U ∈ H (R ) such

m n
Recalling

that U Ω = u in D (Ω), we arrive at uH m (Ω) ≤ CuH (Ω) , as wanted.
m
There are open sets Ω in Rn for which the embedding from Proposition 12.25 is
strict. Here is an example.
Example 12.26. Let Ω := {(x, y) ∈ Rn : 0 < x < 1, 0 < y < x4 }. Note that if λ < 5/2
then
1 x4 1
1
x−λ L2 (Ω) = x−2λ dy dx = x4−2λ dx = . (12.3.44)
0 0 0 5 − 2λ
Hence, if we define u(x, y) := x−1/4 for all (x, y) ∈ Ω then we have u ∈ C ∞ (Ω),
∂1 u(x, y) = − 41 x−5/4 , ∂21 u(x, y) = 16
5 −9/4
x , and ∂2 u(x, y) = ∂22 u(x, y) = 0, for all
(x, y) ∈ Ω. Invoking (12.3.44) it follows that ∂α u ∈ L2 (Ω) for all α ∈ N20 , |α| ≤ 2.
This implies u ∈ H 2 (Ω). On the other hand, if we assume that u ∈ H 2 (Ω), then
by (5) in Theorem 12.15 there exists a function U ∈ C 0 (Ω) such that U Ω = u. The
latter is not possible since u cannot be extended continuously near the origin. Hence,
necessarily u H 2 (Ω).
Example 12.26 points to the fact that the equality between the space H m (Ω)
and H m (Ω) requires additional assumptions on the open set Ω. In Theorem 12.30
we will show that this equality holds in the case of bounded Lipschitz domains
(recall Definition 14.49). Among other things, the proof of Theorem 12.30 relies
on a density result. In part (2) of Theorem 12.14 we have established an important

density result for H s (Ω), · H s (Ω) with s ∈ R and Ω an open set in Rn . A natu-
ral question is whether a similar result is available for the intrinsic Sobolev spaces
m
H (Ω), · H m (Ω) with m ∈ N. The answer turns out to be more delicate, and the
smoothness of the domain Ω plays an important role as is seen in the next theorem.
Theorem 12.27. Fix m ∈ N and let Ω be either an upper-graph Lipschitz domain,
or a bounded Lipschitz domain in Rn . Then the set C0∞ (Ω) is dense in H m (Ω) with
respect to the norm · H m (Ω) .
Proof. We divide the proof into two steps.

Step I. The case of an upper-graph Lipschitz domain. Fix an upper-graph Lipschitz
domain Ω ⊆ Rn and a function uo ∈ H m (Ω). Also, pick ε > 0. According to The-
orem 12.23 there exists some u ∈ H m (Ω) which vanishes a.e. outside of a bounded
subset of Ω and such that
uo − uH m (Ω) ≤ ε/2. (12.3.45)
Suppose we are able to show that there exists a sequence
H m (Ω)
{ϕk }k∈N ⊂ C0∞ (Rn ) satisfying ϕk Ω −−−−−→ u. (12.3.46)
k→∞

Then if k ∈ N is sufficiently large we have ϕk Ω − uH m (Ω) ≤ ε/2 which, in concert

with (12.3.45) proves that ϕ − u
k Ω ≤ ε. Since ϕ ∈ C ∞ (Ω) and ε > 0 is
o H m (Ω) k Ω 0
arbitrary, we conclude that C0∞ (Ω) is dense in H m (Ω) in this case.
Back to the issue of constructing a sequence as in (12.3.46), first recall the nota-
tion for a circular cone from (14.7.23). Lemma 14.55 guarantees that there exists an
angle θ ∈ (0, π/2) (which is determined by the Lipschitz constant of the function
whose upper-graph is Ω) such that
Γθ (x, −en ) ⊆ Rn \ Ω for each x ∈ ∂Ω. (12.3.47)
To simplify notation, abbreviate Γ := Γθ (0, −en ).

Next, pick a function Θ ∈ C0∞ (Γ) with the property that Rn Θ dx = 1 and for
each ε ∈ (0, ∞) define Θε (x) := ε−n Θ(x/ε), for all x ∈ Rn . It follows that for each
ε ∈ (0, ∞) we have

Θε ∈ C0∞ (Γ), supp Θε ⊆ Γ, and Θε dx = 1, (12.3.48)
Rn
Also, denote by tilde the extension by zero outside Ω to Rn , and pick some α ∈ Nn0
with |α| ≤ m. Since u belongs to H m (Ω) and vanishes a.e. outside of a bounded
subset of Ω, it follows that the distribution ∂#
α u belongs to L2 (Rn ) and has compact
support. Moreover, u is a compactly supported distribution in Rn and so is ∂α u.

Hence, if we define ωα := ∂α u − ∂#
α u, then ω ∈ E (Rn ) and, in fact,
α
supp ωα ⊆ ∂Ω. (12.3.49)
Now fix ε ∈ (0, ∞). Proposition 2.102 ensures that u ∗ Θε ∈ C0∞ (Rn ), hence

uε := u ∗ Θε Ω ∈ C0∞ (Ω). (12.3.50)
Given that differentiation of distributions commutes with restrictions to open sets,

part (e) in Theorem 2.96 permits us to write
α
u ∗ Θε Ω = ∂#
∂α uε = ∂α u ∗ Θε Ω + ωα ∗ Θε Ω (12.3.51)
in D (Ω). In light of (12.3.49), the support inclusion in (12.3.48), item (a) in Theo-
rem 2.96, and (12.3.47), we obtain that
supp (ωα ∗ Θε ) ⊆ ∂Ω + Γ ⊆ Rn \ Ω, (12.3.52)

which implies ωα ∗ Θε Ω = 0 in D (Ω). In concert with (12.3.51) this ultimately
yields
α
∂α uε = ∂# u ∗ Θε Ω for all ε > 0. (12.3.53)
2 n
In addition, from Exercise 4.30 we know that ∂# ∂#
L (R )
αu ∗ Θ − α u which, in turn,
ε −−−−→
+ ε→0
implies
# L2 (Ω) α
∂α u ∗ Θε Ω −−−−→
+
∂ u (12.3.54)
ε→0
since the operation of restriction to Ω is continuous from L2 (Rn ) into L2 (Ω). From
(12.3.53) and (12.3.54) it follows that
L2 (Ω)
∂α uε −−−−→
+
∂α u, for each α ∈ Nn0 , |α| ≤ m. (12.3.55)
ε→0
Finally, if we now define ϕk := u ∗ Θ 1k for every k ∈ N, then from (12.3.55) and

(12.3.50) we conclude that {ϕk }k∈N is as in (12.3.46). This finishes the proof of Step I.
Step II. The case of a bounded Lipschitz domain. Suppose Ω is a bounded Lipschitz
domain in Rn and fix u ∈ H m (Ω). By Remark 14.54 there exist N ∈ N, points
x1∗ , . . . , x∗N ∈ ∂Ω, open cylinders {C x∗j }Nj=1 satisfying (14.7.19), upper-graph Lipschitz
domains {Ω j }Nj=1 such that
Ω j ∩ C x∗j = Ω ∩ C x∗j , ∀ j ∈ {1, . . . , N}, (12.3.56)

N
a set O ⊂ Ω with O ⊂ Ω and such that O ∪ C x∗j is an open cover of Ω, and a
j=1
partition of unity {ψ j }Nj=0 subordinate to this cover (cf. (14.7.22)).
Corresponding to j = 0, let ψ # 0 u denote the extension of ψ0 u by zero outside
Ω to Rn . By Example 12.19 we have ψ #0 u ∈ H (R ) which when combined with
m n
Theorem 12.24 gives ψ #0 u ∈ H (R ). Thus, we may apply (3) in Theorem 12.1 to

m n
find a sequence
H m (Rn )
{ϕ0k }k∈N ⊂ C0∞ (Rn ) such that #
ϕ0k −−−−−→ ψ0 u. (12.3.57)
k→∞
By Remark 12.17 (applied with Ω1 := Rn and Ω2 := Ω) it follows that RΩ (ϕ0k )

# 0
converges to RΩ ψ 0 u in H (Ω) as k → ∞. Upon observing that RΩ (ϕk ) = ϕk Ω for
m 0

each k ∈ N, and that RΩ ψ #
0 u = ψ0 u, the latter becomes
H m (Ω)
ϕ0k Ω −−−−−→ ψ0 u. (12.3.58)
k→∞
Next, fix j ∈ {1, . . . , N}. By Exercise 12.18 we have ψ j u ∈ H m (Ω). If we now

define the function ⎧
⎪
⎨ ψ j u in Ω j ∩ C x∗j ,
⎪
u j := ⎪
⎪ (12.3.59)
⎩0 in Ω j \ C x∗ , j
then Example 12.20, applied with Ω0 := Ω, Ω1 := Ω j , u := ψ j u, and v := u j (whose

applicability is ensured by (12.3.59), (12.3.56), and the properties of ψ j u) gives
u j ∈ H m (Ω j ), u j H m (Ω j ) = ψ j uH m (Ω) , and

⎧ α
⎪
⎨ ∂ (ψ j u) in Ω j ∩ C x∗j ,
⎪ (12.3.60)
α
∂ uj = ⎪ ⎪ ∀ α ∈ Nn0 , |α| ≤ m.
⎩ 0 in Ω j \ C x∗ ,
j
At this point we may apply Step I corresponding to the upper-graph Lipschitz do-
main Ω j and the function u j ∈ H m (Ω j ) to obtain
a sequence {ϕkj }k∈N of functions

in C0∞ (Rn ) with
the property that lim ϕkj Ω = u j in H m (Ω j ).
(12.3.61)
k→∞ j
Also, let η j ∈ C0∞ (C x∗j ) satisfying η j ≡ 1 near supp ψ j . In particular, η j ≡ 1 near the
support of u j , hence η j u j = u j in D (Ω j ). Moreover, if we set
Φkj := η j ϕkj for each k ∈ N, (12.3.62)

then
Φkj ∈ C0∞ (C x∗j ) for each k ∈ N. (12.3.63)
A combination of (12.3.62), (12.3.61), and part (1) in Exercise 12.18 gives

lim Φkj Ω = η j u j = u j in H m (Ω j ). (12.3.64)
k→∞ j
In addition, for each α ∈ Nn0 with |α| ≤ m, we may write

α j
∂ Φk Ω − ∂α (ψ j u)L2 (Ω) = ∂α Φkj Ω∩C ∗ − ∂α (ψ j u)L2 (Ω∩C ∗ )
x x
j j

= ∂α Φkj Ω ∩C ∗ − ∂α u j L2 (Ω ∩C ∗ )
j x j x
j j

= ∂α Φkj Ω − ∂α u j L2 (Ω ) . (12.3.65)
j j
The first equality in (12.3.65) follows from (12.3.63) and the fact that the support
of ψ j is contained in C x∗j , the second equality is due to (12.3.56) and (12.3.60),
while the last equality uses (12.3.63) and (12.3.60). Recalling the definition of the
H m -norm (cf. (12.3.4)), from (12.3.65) and (12.3.64) we ultimately conclude that

lim Φkj Ω = ψ j u in H m (Ω), for each j ∈ {1, . . . , N}. (12.3.66)
k→∞
At this stage, for each k ∈ N define

N
Φk := ϕ0k + Φkj . (12.3.67)
j=1
Then (12.3.57) and (12.3.63) guarantee that Φk ∈ C0∞ (Rn ) for each k ∈ N. Also,
based on (12.3.67), (12.3.58), (12.3.66), and the last line in (14.7.22) we obtain

N
lim Φk Ω = ψ ju = u in H m (Ω). (12.3.68)
k→∞
j=0
This finishes the proof of Theorem 12.27.

Extending functions from Sobolev spaces defined intrinsically on open sets to
the entire Euclidean space with preservation of Sobolev class is a delicate issue.
As is apparent from Example 12.26, such an extension is not possible for arbitrary
bounded open sets. We will prove that an extension exists in the case of bounded
Lipschitz domains. This will be done via the so-called Calderón extension operator.
To set the stage, in the next lemma we prove a useful integral representation formula
for smooth functions.
Lemma 12.28. Let Γ be an infinite open upright circular cone with vertex at the
origin and aperture θ ∈ (0, π), that is Γ := {x ∈ Rn : cos(θ/2) |x| < x · en }. Fix m ∈ N
and suppose
φ ∈ C ∞ (S n−1 ) is supported in Γ ∩ S n−1 (12.3.69)
and satisfies
(−1)m
φ(ω) dω = . (12.3.70)
S n−1 (m − 1)!
Then for every Θ ∈ C0∞ (Rn ) there holds
m! y yα
Θ(0) = φ (∂α Θ)(y) dy. (12.3.71)
|α|=m Γ
α! |y| |y|n
Proof. To get started, fix ω ∈ S n−1 and apply formula (14.2.8) to f (t) := Θ(tω),
t ∈ R, to obtain
∞
(−1)m dm
Θ(0) = tm−1 [Θ(tω)] dt. (12.3.72)
(m − 1)! 0 dtm
We claim that
dm m!
[Θ(tω)] = ωα (∂α Θ)(tω), t ∈ R. (12.3.73)
dt m
|α|=m
α!
This is proved by induction over m. If m = 1, formula (12.3.73) is just the Chain

Rule. Suppose (12.3.73) holds for some m ∈ N. Then taking one more derivative in
t and using the Chain Rule from (12.3.73) we obtain
dm+1 n
m! α+e j α+e j
[Θ(tω)] = ω (∂ Θ)(tω), t ∈ R. (12.3.74)
dtm+1 j=1 |α|=m
α!
Now fix j ∈ {1, . . . , n} and invoke Lemma 14.7 in the following setting:
D := {α ∈ Nn0 : |α| = m}, R := {β ∈ Nn0 : |β| = m + 1},
F (α) := α + e j , for each α ∈ D, G := C, (12.3.75)

G(β) := m!
(β−e j )! ωβ (∂β Θ)(tω), for each β ∈ R,

In particular, if suppβ := i ∈ {1, . . . , n} : βi 0 for each β = (β1 , . . . , βn ) ∈ Nn0 , we
have
⎧
⎪
⎨ 0 if j suppβ,
⎪
−1
#F ({β}) = ⎪
⎪ ∀ β ∈ R, (12.3.76)
⎩ 1 if j ∈ suppβ,
and (14.2.1) (applied for each j) in the current yields

n
m! α+e j α+e j
ω (∂ Θ)(tω)
j=1 |α|=m
α!

n m!
= 1suppβ ( j) ωβ (∂β Θ)(tω)
j=1 |β|=m+1
(β − e j )!
1
= m! ωβ (∂β Θ)(tω) (12.3.77)
|β|=m+1 j∈suppβ
(β − e j )!
for each t ∈ R. In addition, since

1 βj 1 m+1
= = βj = , (12.3.78)
j∈suppβ
(β − e j )! j∈suppβ
β! β! j∈suppβ
β!
formula (12.3.77) further gives, for each t ∈ R,

n
m! α+e j α+e j (m + 1)
ω (∂ Θ)(tω) = ωβ (∂β Θ)(tω). (12.3.79)
j=1 |α|=m
α! |β|=m+1
β!
This and (12.3.74) completes the proof by induction of (12.3.73).

A combination of (12.3.72) and (12.3.73) then gives

(−1)m m! ∞ m−1 α α
Θ(0) = t ω (∂ Θ)(tω) dt. (12.3.80)
(m − 1)! |α|=m α! 0
Multiplying (12.3.80) by (m−1)!

(−1)m φ(ω), then integrating in ω ∈ S
n−1
and using
(12.3.70), permits us to write
m! ∞
Θ(0) = tm−1 ωα φ(ω)(∂α Θ)(tω) dt dω
|α|=m
α! S n−1 0
m! y α 1
= |y|m−1 φ(y/|y|)(∂α Θ)(y) n−1 dy
|α|=m
α! Γ |y| |y|
m!
yα
= φ(y/|y|)(∂α Θ)(y) dy. (12.3.81)
|α|=m
α! Γ |y|n
For the second equality in (12.3.81) we changed variables y = tω and observed that
if ω ∈ supp φ ⊆ Γ ∩ S n−1 then y ∈ Γ. The proof of (12.3.71) is now finished.
The construction of Calderón’s extension operator makes the object of the next
theorem.
Theorem 12.29 (Calderón’s Extension Operator). Let Ω be a bounded Lipschitz

domain in Rn and let m ∈ N. Then there exists a linear and bounded operator
Em : H m (Ω) → H m (Rn ) = H m (Rn ) (12.3.82)

Em u Ω = u for every u ∈ H m (Ω). (12.3.83)
Proof. We split the proof of the theorem in a number of steps.

Step I. The Extension operator for upper-graph Lipschitz domains and functions
that vanish outside of a fixed bounded set.
Suppose Ω is the upper-graph of a Lipschitz function ϕ : Rn−1 → R, with Lipschitz

constant M, i.e., Ω = (y , yn ) ∈ Rn−1 × R : yn > ϕ(y ) . In this setting, the goal is to
show that
for each compact set K ⊂ Rn and each η ∈ C0∞ (Rn )
such that η ≡ 1 near K, there exists a linear and
bounded operator Em : H m (Ω) → H m (Rn ) satisfy- (12.3.84)
ing supp (Em v) ⊆ supp η and (Em v)Ω = η2 v a.e. in Ω,
for every v ∈ H m (Ω).
To this end, fix a compact set K and a function η as in (12.3.84). Also, pick an

angle θ ∈ 0, 2 arctan ( M1 ) and apply Lemma 14.55 to conclude that (14.7.24) holds.
To lighten notation, abbreviate Γ := Γθ (0, en ) (recall (14.7.23)). Then (14.7.24) im-
plies
x + Γ ⊆ Ω for every x ∈ Ω. (12.3.85)
Let φ be as in Lemma 12.28 associated with this cone Γ. For each α ∈ Nn0 with
|α| = m define the function
m! x xα
φα (x) := (−1)m φ− , ∀ x ∈ Rn \ {0}. (12.3.86)
α! |x| |x|n
From (12.3.69) and (12.3.86) we see that
supp φα ⊂ −Γ, φα ∈ C ∞ (Rn \ {0}),

(12.3.87)
φα (λx) = λm−n φα (x), ∀ x ∈ Rn \ {0}, ∀ λ ∈ (0, ∞).
The properties listed in (12.3.87) allow us to define the generalized volume potential
∞
associated with φα (cf. (4.10.23)). This acts on each f ∈ Lcomp (Rn ) according to

Πφα f (x) := φα (x − y) f (y) dy, ∀ x ∈ Rn . (12.3.88)
Rn
∞
From Theorem 4.105 we know that Πφα f ∈ C m−1 (Rn ) for each f ∈ Lcomp (Rn ) (note
that φα is positive homogeneous of degree m − n ∈ Z) and
∂γ [Πφα f ] = Π∂γ φα f pointwise in Rn

(12.3.89)
for every γ ∈ Nn0 with |γ| ≤ m − 1.
Moreover, Theorem 4.105 gives that if γ = (γ1 , . . . , γn ) ∈ Nn0 has |γ| = m then for
∞

each f ∈ Lcomp (Rn ) the distributional derivative ∂γ Πφα f is of function type and for
each j ∈ {1, . . . , n} such that γ j 0 we have

γ
∂ [Πφα f ](x) = (∂γ−e j φα )(ω)ω j dσ(ω) f (x) (12.3.90)
S n−1

+ lim+ (∂γ φα )(x − y) f (y) dy for a.e. x ∈ Rn ,
ε→0 |y−x|≥ε
where the derivative in the left-hand side is taken in D (Rn ).

Next, consider the operator Em : C0∞ (Ω) → C acting on each w ∈ C0∞ (Ω) accord-
ing to
(Em w)(x) := η(x) Πφα ∂
α (ηw) (x), ∀ x ∈ Rn , (12.3.91)
|α|=m
where tilde denotes extension by zero from Ω to Rn . Note that since w ∈ C0∞ (Ω) then
∂
α (ηw) ∈ L∞ (Rn ). In light of the fact that the operator in (12.3.88) is well defined,
comp
this ensures that the operator in (12.3.91) is also well defined. We make two claims
pertaining to the nature of Em .

Claim 1. (E w) = η2 w pointwise in Ω, for each w ∈ C ∞ (Ω).
m Ω 0
Claim 2. There exists a constant C ∈ (0, ∞) such that
Em wH m (Rn ) ≤ CwH m (Ω) , ∀ w ∈ C0∞ (Ω). (12.3.92)
Let us assume Claim 1 and Claim 2 for the moment and see how they may be used
to finish the construction of the desired extension operator.
The operator Em is obviously linear which, when combined with the estimate in

(12.3.92), implies that Em : C0∞ (Ω) , · H m (Ω) → H m (Rn ) is linear and bounded.
Since C0∞ (Ω) is dense in H m (Ω) (cf. Theorem 12.27) it follows that Em extends
continuously to H m (Ω), thus
there exists Em : H m (Ω) → H m (Rn ) linear and

bounded and such that Em C ∞ (Ω) = Em . (12.3.93)
0
This extension by density means that
Em v = lim Em w j for each given v ∈ H m (Ω) and each

j→∞
H m (Ω) (12.3.94)
sequence w j ∈ C0∞ (Ω), j ∈ N, such that w j −−−−−→ v.
j→∞
In particular, since supp (Em w) ⊆ supp η for every w ∈ C0∞ (Ω) (as seen from an
inspection of formula (12.3.91)), we also have that
supp (Em v) ⊆ supp η for all v ∈ H m (Ω). (12.3.95)
Hence, to show that the operator Em from (12.3.93) satisfies all the conditions listed
in (12.3.84), there remains to prove (modulo the justifications of the two earlier
claims) that (Em v)Ω = η2 v, a.e. in Ω, for each v ∈ H m (Ω). Fix such a function v and
apply Theorem 12.27 to obtain a sequence {w j } j∈N of functions in C0∞ (Ω) such that
H m (Ω) L2 (Ω)
w j −−−−−→ v. In particular, w j −−−−→ v and, by passing to a subsequence (which we
j→∞ j→∞
still denote by {w j } j∈N ) and multiplying by η2 , we have
η2 w j → η2 v pointwise a.e. in Ω as j → ∞. (12.3.96)
In addition, Em v ∈ H m (Rn ) and the continuity of Em further yields

H m (Rn )
Em w j = Em w j −−−−−→ Em v. (12.3.97)
j→∞
L2 (Rn )
In particular, (12.3.97) implies Em w j −−−−−→ Em v and, passing to a subsequence
j→∞
(which is a subsequence of the subsequence in (12.3.96)) we obtain (again, keeping
the same notation for this sub-subsequence)
Em w j → Em v pointwise a.e. in Rn as j → ∞. (12.3.98)

By restricting to Ω and invoking Claim 1, to the effect that η2 w j = Em w j Ω , the
convergence in (12.3.98) further gives

η2 w j → (Em v)Ω pointwise a.e. in Ω as j → ∞. (12.3.99)
In concert, (12.3.96) and (12.3.99) imply

(Em v)Ω = η2 v a.e. in Ω. (12.3.100)
In summary, we have established (12.3.84), modulo the proofs of Claim 1 and

Claim 2 which we shall deal with in the next two steps.
Step II. Proof of Claim 1 from Step I.
Let w ∈ C0∞ (Ω) be given. Pick W ∈ C0∞ (Rn ) such that W Ω = w, and fix x ∈ Ω.
Starting with (12.3.91) and recalling (12.3.88) we may write

(Em w)(x) = η(x) φα (x − y)∂α (ηW)(y) dy
|α|=m Ω

= η(x) φα (z)∂α (ηW)(x − z) dz, (12.3.101)
|α|=m −Γ
where for the second equality in (12.3.101) we made the change of variables y = x−z
and used the support condition supp φα ⊂ −Γ (cf. (12.3.87)) and the fact that if y ∈ Ω
then z ∈ (x − Ω) ∩ (−Γ) = −Γ (due to (12.3.85)). Next, we bring in formula (12.3.86)

and make the change of variables z = −y under the integral to further obtain
m! y yα
(Em w)(x) = η(x) φ ∂α (ηW)(x + y) dy. (12.3.102)
|α|=m
α! Γ |y| |y|n
This last expression in the right-hand side of (12.3.102) allows us to apply (12.3.71)
in Lemma 12.28 with Θ(y) := ∂α (ηW)(x + y) for y ∈ Rn (recall that x ∈ Ω is fixed)
and conclude that
(Em w)(x) = η(x)(ηW)(x) = η2 (x)w(x), (12.3.103)
as desired.
Step III. Proof of Claim 2 from Step I.
Fix again w ∈ C0∞ (Ω) and pick some β ∈ Nn0 with |β| ≤ m. Then (12.3.91) and the
Generalized Leibniz Formula from Proposition 2.49 give
β!
∂β (Em w) = (∂β−γ η)∂γ Πφα ∂
α (ηw) (12.3.104)
|α|=m γ≤β
γ!(β − γ)!
in D (Rn ). To estimate the L2 (Rn ) norm of the terms in the right-hand side we dis-
tinguish two cases.
First, suppose γ ∈ Nn0 is such that γ ≤ β and |γ| ≤ m − 1. Observe that since η is
compactly supported, the set

Kη := x − y : x, y ∈ supp η is compact in Rn . (12.3.105)
From (12.3.87) it follows that the function ∂γ φα is smooth outside the origin and
is positive homogeneous of degree m − n − |γ| ≥ 1 − n (recall Exercise 4.51) so it
1
belongs to Lloc (Rn ). The latter and (12.3.105) then ensure (∂γ φα )1Kη ∈ L1 (Rn ), so
for each x ∈ supp η we may write

γ
Π∂γ φα ∂α (ηw) (x) = (∂ φα )1Kη (x − y)∂
α (ηw)(y) dy
Rn
α
= (∂γ φα )1Kη ∗ ∂
(ηw) (x). (12.3.106)
Since ∂α (ηw) ∈ L2 (Rn ) we may invoke Young’s Inequality and Example 12.19 to
estimate

α (ηw)
(∂β−γ η) Π∂γ φ ∂ L2 (Rn )
α

≤ ∂β−γ ηL∞ (Rn ) (∂γ φα )1Kη L1 (Rn ) ∂
α (ηw)
L2 (Rn )

≤ C ∂β−γ ηL∞ (Rn ) (∂γ φα )1Kη L1 (Rn ) wH m (Ω)
≤ CwH m (Ω) , (12.3.107)
for some C = C(Ω, η, φ, α, γ) ∈ (0, ∞) independent of w.

Second, suppose γ ∈ Nn0 is such that γ ≤ β and |γ| = m (a scenario in which we
necessarily have |β| = m and, in fact, γ = β). If γ = (γ1 , . . . , γn ) and j ∈ {1, . . . , n} is
such that γ j 0, then according to (12.3.90) we have
α (ηw)] = C(γ, φ )∂

∂γ [Πφα ∂ α
α (ηw) + T γ ∂
∂ φα
α (ηw) (12.3.108)
where the constant C(γ, φα ) is defined as

C(γ, φα ) := (∂γ−e j φα )(ω)ω j dσ(ω) (12.3.109)
S n−1
and, for a.e. x ∈ Rn ,

T ∂γ φα ∂
α (ηw)(x) := lim
+
(∂γ φα )(x − y)∂
α (ηw)(y) dy. (12.3.110)
ε→0 |y−x|≥ε
From (12.3.87), the current assumptions on γ, Exercise 4.51, and Example 4.71 (ap-
plied to ∂γ−e j φα ), it follows that the function ∂γ φα satisfies the conditions in (4.4.1).
Consequently, part (e) in Theorem 4.100 ensures that the operator T ∂γ φα is bounded
from L2 (Rn ) into L2 (Rn ), thus

α (ηw)
T ∂γ φα ∂ L2 (Rn ) ≤ C ∂
α (ηw)
L2 (Rn ) . (12.3.111)
In concert, (12.3.108), (12.3.111), and Example 12.19 give

α (ηw)
η Π∂γ φ ∂ α (ηw)
α L2 (Rn ) ≤ C ηL∞ (Rn ) ∂ L2 (Rn )
≤ CwH m (Ω) , (12.3.112)
for some C = C(Ω, η, φ, α, γ) ∈ (0, ∞) independent of w.

At this stage, combining (12.3.104), (12.3.107), (12.3.112) proves that there ex-
ists a constant C ∈ (0, ∞) with the property that
∂β (Em w)L2 (Rn ) ≤ CwH m (Ω) , (12.3.113)
for all w ∈ C0∞ (Ω) and β ∈ Nn0 with |β| ≤ m. The desired estimate in Claim 2 now
follows by invoking Theorem 12.24 and (12.3.4).
Step IV. The extension operator for a bounded Lipschitz domain.

Work in the case when Ω is a bounded Lipschitz domain. By Remark 14.54 there
exist a natural number N ∈ N, points x1∗ , . . . , x∗N ∈ ∂Ω, upper-graph Lipschitz do-
mains {Ω j }Nj=1 , open cylinders {C x∗j }Nj=1 satisfying (14.7.19), a set O ⊂ Ω with O ⊂ Ω
N
and such that O ∪ C x∗j is an open cover of Ω, and a partition of unity {ψ j }Nj=0
j=1
subordinate to this cover (cf. (14.7.22)). In addition, for each j ∈ {1, . . . , N}, consider

η j ∈ C0∞ C x∗j , η j ≡ 1 near supp ψ j . (12.3.114)
To proceed, select u ∈ H m (Ω) and fix j ∈ {1, . . . , n}. Making use of Exer-
cise 12.18 we obtain
ψ j u ∈ H m (Ω) and ψ j uH m (Ω) ≤ CuH m (Ω) , (12.3.115)
for some constant C = C(Ω) ∈ (0, ∞) (given that the partition of unity ultimately
depends only on Ω). Since by (14.7.19) we have
Ω j ∩ C x∗j = Ω ∩ C x∗j , (12.3.116)
we may define the function

⎧
⎪
⎨ ψ j u in Ω j ∩ C x∗j ,
⎪
v j := ⎪
⎪ (12.3.117)
⎩0 in Ω j \ C x∗j .
Granted (12.3.115) and (12.3.117), Example 12.20 may be invoked to conclude that
there exists some C = C(Ω) ∈ (0, ∞) such that
v j belongs to H m (Ω j ), vanishes outside a bounded subset
(12.3.118)
of Ω j , and obeys the estimate v j H m (Ω j ) ≤ CuH m (Ω) .
Next, let Emj denote the extension operator from Step I associated with the upper-
graph Lipschitz domain Ω j , the compact set K := supp ψ j , and the function η := η j .
In particular, we have

Emj v j ∈ H m (Rn ), Emj v j H m (Rn ) ≤ Cv j H m (Ω j ) ,
supp (Emj v j ) ⊆ supp η j ⊂ C x∗j , and (12.3.119)

(Emj v j )Ω = η2j v j a.e. in Ω j .
j
Taking into account (12.3.114) and (12.3.116), the last condition in (12.3.119) fur-
ther yields

Emj v j )Ω∩ C ∗ = Emj v j )Ω ∩ C ∗ = η2j v j Ω ∩ C ∗
x j x j x
j j j

= η2j ψ j u Ω ∩ C ∗ = ψ j u Ω ∩ C ∗
j x j x
j j

= ψ j u Ω∩ C ∗ . (12.3.120)
x
j
In concert, (12.3.120), the support condition in (12.3.119), and the fact that ψ j is
supported in C x∗j imply

Emj v j )Ω = Emj
v j )Ω∩ C ∗ = ψ j
u Ω∩ C ∗ = ψ j u, (12.3.121)
x x
j j
where tilde denotes the extension by zero to Ω. Also, by combining the norm esti-
mate in (12.3.119) with the norm estimate in (12.3.118) we obtain
j
Em v j H m (Rn ) ≤ CuH m (Ω) . (12.3.122)
At this point, we are ready to define an extension of u to H m (Rn ). Specifically,

set

N
Em u := #
Emj v j + ψ0 u, (12.3.123)
j=1
where ψ#0 u denotes the extension of ψ0 u by zero outside Ω to R . From Exam-

n
ple 12.19 we know that
#
ψ m n #
0 u ∈ H (R ) and ψ 0 uH m (Rn ) ≤ CuH m (Ω) , (12.3.124)
for some C = C(ψ0 ) = C(Ω) ∈ (0, ∞). From (12.3.123), (12.3.122), and (12.3.124),
it follows that

Em u ∈ H m (Rn ) and Em uH m (Rn ) ≤ CuH m (Ω) , (12.3.125)
for some finite constant C = C(Ω) independent of u. Moreover, (12.3.123),

N
#
(12.3.121), the fact that ψ0 uΩ = ψ0 u, and that j=0 ψ j = 1 in Ω, imply

N
j
N
(Em u)Ω = #
Em v j )Ω + ψ0 uΩ = ψ j u = u. (12.3.126)
j=1 j=0
Since u ∈ H m (Ω) is arbitrary, in light of (12.3.125) and (12.3.126), we have that the
operator defined in (12.3.123) is linear and bounded from H m (Ω) into H m (Rn ) and
is an extension operator, in the sense that (Em u)Ω = u for every u ∈ H m (Ω). This
finishes the proof of Theorem 12.29.
An important consequence of the existence of Calderón’s extension operator as
proved in the last theorem is the fact that for a bounded Lipschitz domain Ω the
spaces H m (Ω) and H m (Ω) are equal.
Theorem 12.30. Let Ω be a bounded Lipschitz domain in Rn and fix m ∈ N. Then

H m (Ω) = H m (Ω) as vector spaces, with equivalent norms. In particular,
1/2
uH m (Ω) ≈ ∂α u2L2 (Ω) , ∀ u ∈ H m (Ω). (12.3.127)
α∈Nn0 , |α|≤m
Proof. Since the inclusion H m (Ω) ⊆ H m (Ω) and the corresponding norm inequality
hold for arbitrary open sets Ω (cf. Proposition 12.25) there remains to prove the
opposite inclusion and the naturally accompanying norm inequality. To see this,

pick an arbitrary u ∈ H m (Ω) and recall from Theorem 12.29 that u = Em u Ω ,
where Em : H m (Ω) → H m (Rn ) is Calderón’s extension operator. Since the latter is
bounded, the desired conclusion follows.
Collectively, Theorem 12.30 and Theorem 12.15 yield the following result.
Corollary 12.31. Let Ω be a bounded Lipschitz domain in Rn . Then the following

statements are true.
(1) For every m ∈ N the embedding H m (Ω) → H m−1 (Ω) is compact.
(2) If m ∈ N and k ∈ N0 are such that k < m − n/2 then H m (Ω) ⊂ C k (Ω).
We close this section by proving a Chain Rule formula for functions in in-
trinsic Sobolev spaces of order one. To set the stage, the reader is reminded that
a function Ψ : E → F is said to be bi-Lipschitz (where E, F ⊆ Rn ) provided there
exists c ∈ (0, 1) such that c|x − y| ≤ |Ψ (x) − Ψ (y)| ≤ c−1 |x − y| for each x, y ∈ E.
Theorem 12.32. Suppose Ψ : Rn → Rn is a bijective and bi-Lipschitz function. Let

O be an open set in Rn and consider U := Ψ (O). Then U is an open subset of
Rn , for every u ∈ H 1 (U) we have u ◦ Ψ ∈ H 1 (O), and for each j ∈ {1, . . . , n} the
following Chain Rule formula holds

n

∂ j (u ◦ Ψ ) = (∂k u) ◦ Ψ · ∂ j Ψk a.e. in O. (12.3.128)
k=1
Moreover, there exists a constant C = C(n, Ψ ) ∈ (0, ∞) such that
u ◦ Ψ H 1 (O) ≤ CuH 1 (U) , for all u ∈ H 1 (U). (12.3.129)
Proof. Since Ψ is bijective and bi-Lipschitz, it follows that both Ψ and its inverse
Ψ −1 are Lipschitz functions (in particular, continuous). The fact that U is the pre-
image of the open set O under the continuous function Ψ −1 then implies that U is
itself open.
To proceed, we first observe that for any measurable function w : U → C, the
change of variables y = Ψ (x) (cf. Theorem 14.58) yields

w(Ψ (x))2 dx =
|w(y)|2 det (DΨ −1 )(y) dy
O U

≤C |w(y)|2 dy, (12.3.130)
U
where the inequality in (12.3.130) is based on the fact that pointwise partial deriva-
tives of Lipschitz functions are bounded (cf. Theorem 2.115 and Exercise 2.116). In
particular, from (12.3.130) we obtain the implication
w ∈ L2 (U) =⇒ w ◦ Ψ ∈ L2 (O) and w ◦ Ψ L2 (O) ≤ CwL2 (U) , (12.3.131)
for some finite constant C = C(n, Ψ ) > 0 independent of w.

Now fix u ∈ H 1 (U). Since u ∈ L2 (U) and ∂k u ∈ L2 (U) for k ∈ {1, . . . , n},
implication (12.3.131) gives that u ◦ Ψ ∈ L2 (O) and (∂k u) ◦ Ψ ∈ L2 (O) for all
integers k ∈ {1, . . . , n}. Recalling that all the entries in DΨ are uniformly bounded,
to complete the proof of the theorem there remains to show

n

∂ j (u ◦ Ψ ) = (∂k u) ◦ Ψ · ∂ j Ψk in D (O). (12.3.132)
k=1
To this end, apply Theorem 12.22 (with m := 1) to obtain a sequence of functions

{um }m∈N ⊆ C ∞ (U) ∩ H 1 (U) with the property that
H 1 (U)
um −−−−−→ u. (12.3.133)
m→∞
Next, pick j ∈ {1, . . . , n}. Since the Lipschitz function Ψ : O → U is differentiable

a.e. on O, for every m ∈ N we may rely on the classical Chain Rule to write that

n

∂ j (um ◦ Ψ ) = (∂k um ) ◦ Ψ · ∂ j Ψk pointwise a.e. in O. (12.3.134)
k=1
Pick k ∈ {1, . . . , n} and write the inequality resulting when applying (12.3.130) to
L2 (U)
the function w := ∂k um − ∂k u. Since ∂k um −−−−→ ∂k u (a consequence of (12.3.133))
m→∞
we may further conclude that
L2 (O)
(∂k um ) ◦ Ψ −−−−→ (∂k u) ◦ Ψ, ∀ k ∈ {1, . . . , n}. (12.3.135)
m→∞
A similar reasoning (working with um in place of ∂k um ) yields
L2 (O)
um ◦ Ψ −−−−→ u ◦ Ψ. (12.3.136)
m→∞
At this point we take ϕ ∈ C0∞ (O) and write


∂ j (u ◦ Ψ ), ϕ = −u ◦ Ψ, ∂ j ϕ = − (u ◦ Ψ )(x)(∂ j ϕ)(x) dx
O

= − lim (um ◦ Ψ )(x)(∂ j ϕ)(x) dx
m→∞ O

= lim ∂ j (um ◦ Ψ )(x)ϕ(x) dx
m→∞ O
n

= lim (∂k um ) ◦ Ψ (x)(∂ j Ψ )(x)ϕ(x) dx
m→∞ O
k=1
n

= (∂k u) ◦ Ψ (x)(∂ j Ψ )(x)ϕ(x) dx
k=1 O

n

= (∂k u) ◦ Ψ · ∂ j Ψk , ϕ . (12.3.137)
k=1
The second and last equality in (12.3.137) are consequences of the fact that the dis-

tributions u ◦ Ψ and (∂k u) ◦ Ψ · ∂ j Ψk , for each k ∈ {1, . . . , n}, are given by locally
integrable functions in O. The third equality in (12.3.137) is based on (12.3.136),
Cauchy–Schwarz’s inequality, and the (obvious) membership of ∂ j ϕ to L2 (O). The
fourth equality in (12.3.137) uses the integration by parts formula (2.9.39) (bearing
in mind that um ◦ Ψ is a locally Lipschitz function in O; cf. Exercise 2.120 and Exer-
cise 2.119). The fifth equality in (12.3.137) is a consequence of (12.3.134), while
for the sixth equality we used (12.3.135), Hölder’s inequality, and the membership
ϕ ∈ L2 (O). This proves (12.3.132) and finishes the proof of the theorem.
Corollary 12.33. Let Ω be an upper-graph Lipschitz domain in Rn and let Φ be the

bijective bi-Lipschitz flattening map associated with Ω as in Remark 14.52. Then
the following are true.
(i) For a measurable function u in Ω we have
u ∈ L2 (Ω) ⇐⇒ u ◦ Φ ∈ L2 (Rn+ ) (12.3.138)
and
u ∈ H 1 (Ω) ⇐⇒ u ◦ Φ ∈ H 1 (Rn+ ), (12.3.139)
and the equivalences hold with naturally accompanying estimates.
(ii) For a measurable function v in Rn+ we have
v ∈ L2 (Rn+ ) ⇐⇒ v ◦ Φ−1 ∈ L2 (Ω) (12.3.140)
and
v ∈ H 1 (Rn+ ) ⇐⇒ v ◦ Φ−1 ∈ H 1 (Ω), (12.3.141)
Proof. All equivalences are consequences of Theorem 12.32. For example, for the
left-to-right implications in (i) apply Theorem 12.32 with u := u, Ψ := Φ, O := Rn+
and U := Ω.
In the last portion of this section we present yet another extension result, which is
a companion to Theorem 12.29. While the latter theorem dealt with bounded Lips-
chitz domains and intrinsic Sobolev spaces of any order (via the Calderón extension
operator), the new result treats Sobolev spaces of order one in upper-graph Lipschitz
domains (via flattening and extension by reflection).
Theorem 12.34. Let Ω be an upper-graph Lipschitz domain in Rn . Then there exists

a linear and bounded operator
EΩ : H 1 (Ω) → H 1 (Rn ) (12.3.142)

EΩ u Ω = u for every u ∈ H 1 (Ω). (12.3.143)
Proof. The proof is organized into two steps.

Step I. The case of the upper-half space. Fix u ∈ C0∞ (Rn+ ) and define the function
Eu ∈ C 0 (Rn ) by setting
⎧
⎪
⎪
⎨ u(x) if x ∈ Rn+ ,
(Eu)(x) := ⎪
⎪ (12.3.144)
⎩ u(x , −x ) if x = (x , x ) ∈ Rn ,
n n −
for each x ∈ Rn . Note that

√
EuL2 (Rn ) ≤ 2uL2 (Rn+ ) . (12.3.145)
For each j ∈ {1, . . . , n} and each x ∈ Rn let us also introduce

⎧
⎪
⎪
⎨ (∂ j u)(x) if x ∈ Rn+ ,
f j (x) := ⎪⎪ (12.3.146)
⎩ (∂ u)(x , −x ) if x = (x , x ) ∈ Rn ,
j n n −
if 1 ≤ j ≤ n − 1 and, corresponding to j = n,
⎧
⎪
⎪
⎨ (∂n u)(x) if x ∈ Rn+ ,
fn (x) := ⎪
⎪ (12.3.147)
⎩ −(∂ u)(x , −x ) if x = (x , x ) ∈ Rn .
n n n −
It is then clear that f j ∈ L2 (Rn ) for each j ∈ {1, . . . , n} and there exists a purely
dimensional constant Cn ∈ (0, ∞) such that

n
f j L2 (Rn ) ≤ Cn ∇uL2 (Rn+ ) . (12.3.148)
j=1
Fix ϕ ∈ C0∞ (Rn ) and assume j ∈ {1, . . . , n−1}. Then based on (12.3.144), (12.3.146),
and straightforward integrations by parts we may compute

(Eu)(x)(∂ j ϕ)(x) dx = u(x)(∂ j ϕ)(x) dx + u(x , −xn )(∂ j ϕ)(x) dx
Rn Rn+ Rn−

=− (∂ j u)(x)ϕ(x) dx − (∂ j u)(x , −xn )ϕ(x) dx
Rn+ Rn−

=− f j (x)ϕ(x) dx. (12.3.149)
Rn
Also, corresponding to j = n,

(Eu)(x)(∂n ϕ)(x) dx = u(x)(∂n ϕ)(x) dx + u(x , −xn )(∂n ϕ)(x) dx
Rn Rn+ Rn−

=− (∂n u)(x)ϕ(x) dx − u(x , 0)ϕ(x , 0) dx
Rn+ Rn−1

+ (∂n u)(x , −xn )ϕ(x) dx + u(x , 0)ϕ(x , 0) dx
Rn− Rn−1

=− fn (x)ϕ(x) dx. (12.3.150)
Rn
All together, (12.3.149)–(12.3.150) prove that
∂ j (Eu) = f j in D (Rn ) for each j ∈ {1, . . . , n}. (12.3.151)
In particular, from (12.3.145), (12.3.151), and (12.3.148) we conclude that
for each u ∈ C0∞ (Rn+ ) we have Eu ∈ H 1 (Rn )

(12.3.152)
and EuH 1 (Rn ) ≤ Cn uH 1 (Rn+ ) .
Hence the assignment C0∞ (Rn+ ) u → Eu ∈ H 1 (Rn ) is linear and bounded. In

concert with the density result from Theorem 12.27 (used for Ω := Rn+ ), this allows
us to conclude that there exists a (unique) linear and bounded mapping
ERn+ : H 1 (Rn+ ) → H 1 (Rn ) (12.3.153)
ERn+ u = Eu for every u ∈ C0∞ (Rn+ ). (12.3.154)
In particular, if RRn+ : H 1 (Rn ) → H 1 (Rn+ ) denotes the operator of restriction to Rn+

which is linear and bounded in this context (cf. Remark 12.17), from (12.3.154) and
12.4 The Space H 1/2 (∂Ω) on Boundaries of Lipschitz Domains 469
(12.3.144) we conclude that

RRn+ ◦ ERn+ u = u for every u ∈ C0∞ (Rn+ ). (12.3.155)
Thus, RRn+ ◦ ERn+ and I, the identity operator, are two continuous operators from
H 1 (Rn+ ) into itself which agree on C0∞ (Rn+ ), a dense subset of the latter space (cf.
Theorem 12.27). As such, the said operators agree on the entire space H 1 (Rn+ ), i.e.,

ERn+ u Rn = u for every u ∈ H 1 (Rn+ ). (12.3.156)
+
This concludes the treatment of the case when the underlying domain is the upper-
half space.
Step II. The case of an upper-graph Lipschitz domain. Assume Ω ⊆ Rn is an upper-
graph Lipschitz domain and bring in the bijective bi-Lipschitz map Φ associated
with Ω as in Remark 14.52. In this context, we define
EΩ : H 1 (Ω) → H 1 (Rn )

(12.3.157)
EΩ u := ERn+ (u ◦ Φ) ◦ Φ−1 for each u ∈ H 1 (Ω).
From Corollary 12.33 and Step I we see that this is a well-defined, linear, and
bounded operator. Moreover, for x ∈ Ω we have Φ−1 (x) ∈ Rn+ hence given any
u ∈ H 1 (Ω) we may write

EΩ u (x) = ERn+ (u ◦ Φ) Φ−1 (x) = (u ◦ Φ) Φ−1 (x) = u(x), (12.3.158)
where the first equality is the definition in (12.3.157), and the second equality is
a consequence of (12.3.156) (bearing in mind that u ◦ Φ ∈ H 1 (Rn+ ); cf. Corol-
lary 12.33). This establishes (12.3.143) and finishes the proof of Theorem 12.34.

Here is a companion result to Theorem 12.30.
Corollary 12.35. Suppose Ω is an upper-graph Lipschitz domain in Rn . Then
H 1 (Ω) = H 1 (Ω) as vector spaces, with equivalent norms.
Proof. This is an immediate consequence of Proposition 12.25, Theorem 12.34,
Theorem 12.24, and definitions.
12.4 The Space H1/2 (∂Ω) on Boundaries of Lipschitz Domains
In this section we define the Sobolev space of order s = 1/2 on the boundary of
a set Ω ⊂ Rn which is either an upper-graph Lipschitz domain, or a bounded Lip-
schitz domain. To set the stage for the subsequent discussion, we first revisit the
space H 1/2 (Rn ) and define an equivalent norm on it which may be adapted to more
general geometries. Recall from (12.1.2) that u ∈ H 1/2 (Rn ) provided u ∈ S (Rn ) and
ξ1/2
u ∈ L2 (Rn ). Since
1 1/2
√ (1 + |ξ|) ≤ 1 + |ξ|2 ≤ 1 + |ξ| for all ξ ∈ Rn , (12.4.1)
2
it follows that for any u ∈ S (Rn ) we have

ξ
u ∈ L (R ) ⇐⇒ u ∈ L (R ) and
1/2 2 n 2 n
|ξ||
u(ξ)|2 dξ < ∞, (12.4.2)
Rn
plus naturally accompanying estimates.

A useful fact about the integral in (12.4.2) is proved in the next lemma. In what
follows, we employ the notation ω = (ω1 , . . . , ωn ) for ω ∈ S n−1 .
Lemma 12.36. Let u ∈ L2 (Rn ). Then

|u(x) − u(y)|2
c0 |ξ||
u(ξ)|2 dξ = dx dy (12.4.3)
Rn Rn Rn |x − y|n+1
∞ 4 sin2 (tω1 /2)
where c0 := (2π)−n 0 S n−1 t2
dσ(ω) dt ∈ (0, ∞).
Proof. Starting with the substitution x = y + h, then reversing the order of inte-
gration and applying Plancherel’s identity (cf. formula (3.2.29)), using the fact that

F u(· + h) − u (ξ) = (eih·ξ − 1)u(ξ) for each ξ ∈ Rn , and again reversing the order of
integration, we obtain

|u(x) − u(y)|2
dx dy
Rn Rn |x − y|
n+1
F u(· + h) − u2
−n L2 (Rn )
= (2π) dh
R n |h|n+1

−n |eih·ξ − 1|2
= (2π) |
u(ξ)|2
dh dξ. (12.4.4)
Rn Rn |h|n+1
Making use of polar coordinates the inner integral becomes
∞
|eih·ξ − 1|2 1
dh = |eiρω·ξ − 1|2 dσ(ω) dρ
R n |h|n+1
0 ρ2
S n−1
∞
1 ξ
= |ξ| 2
|eitω· |ξ| − 1|2 dσ(ω) dt
0 t S n−1
∞
1
= |ξ| 2
|eitω1 − 1|2 dσ(ω) dt, (12.4.5)
0 t S n−1
where in the second equality we made the change of variables ρ = t/|ξ|, and the last
ξ
equality is based on (14.9.11) applied with f (ω) := |eitω· |ξ| − 1|2 for each ω ∈ S n−1 ,
and R being the unitary transformation mapping |ξ|ξ into the vector (1, 0, . . . , 0). Note
that |eitω1 − 1|2 = 4 sin2 (tω1 /2) which, together with the fact that | sin(x)| ≤ |x| for
every x ∈ R, implies
∞
1
|eitω1 − 1|2 dσ(ω) dt
0 t2 S n−1
1 ∞
1 1
≤ t2 ω21 dσ(ω) dt + 4 dσ(ω) dt < ∞. (12.4.6)
0 t2 S n−1 1 t2 S n−1
Now (12.4.3) follows from (12.4.4)–(12.4.6).

An important consequence of this discussion is the following description of the
vector space H 1/2 (Rn ).
Proposition 12.37. The vector space H 1/2 (Rn ) is equal to the collection of all func-
tions u ∈ L2 (Rn ) with the property that

|u(x) − u(y)|2
dx dy < ∞. (12.4.7)
Rn Rn |x − y|n+1
In addition, if for each u ∈ H 1/2 (Rn ) one defines

|u(x) − u(y)|2 1/2
|u| 12 ,Rn := uL2 (Rn ) + dx dy , (12.4.8)
Rn Rn |x − y|n+1
then | · | 12 ,Rn is a norm on H 1/2 (Rn ) which is equivalent with the norm · H 1/2 (Rn ) .
Proof. Let us check that | · | 12 ,Rn is indeed a norm on the vector space H 1/2 (Rn ).
Start by considering the ambient set X := Rn × Rn endowed with the measure
μ(x, y) := |x − y|−n−1 Ln (x) ⊗ Ln (y) for (x, y) ∈ X, (12.4.9)
where Ln is the Lebesgue measure in Rn . For u, v ∈ H 1/2 (Rn ), define the functions
F(x, y) := u(x) − u(y) and G(x, y) := v(x) − v(y), for every (x, y) ∈ X. Using the
properties of a generic L2 -norm we may then write
|u + v| 12 ,Rn = u + vL2 (Rn ) + F + GL2 (X,μ)
≤ uL2 (Rn ) + vL2 (Rn ) + FL2 (X,μ) + GL2 (X,μ)

= |u| 12 ,Rn + |v| 12 ,Rn (12.4.10)
hence |·| 12 ,Rn satisfies the triangle inequality. The homogeneity is immediate, while
the nondegeneracy of | · | 12 ,Rn is inherited from that of · L2 (Rn ) .
Moving on, the first statement in the proposition is a consequence of (12.1.2),
(12.4.2), and (12.4.3). These and the definition of · H 1/2 (Rn ) also imply the equiva-
lence of the latter norm with | · | 12 ,Rn .
An immediate consequence of Proposition 12.37 and part (3) in Theorem 12.1 is
the next density result.

Corollary 12.38. The space C0∞ (Rn ) is dense in H 1/2 (Rn ) , | · | 12 ,Rn .
Proposition 12.37 (used with n replaced by n − 1) motivates the following defini-

tion for Sobolev spaces of order 1/2 on Lipschitz surfaces in Rn .
Definition 12.39. Let Ω be either an upper-graph Lipschitz domain in Rn , or a
bounded Lipschitz domain in Rn . Then the Sobolev space H 1/2 (∂Ω) is defined as
|u(x) − u(y)|2 1/2
H (∂Ω) := u ∈ L (∂Ω) :
1/2 2
dσ(x) dσ(y) < ∞ .
∂Ω ∂Ω |x − y|n
(12.4.11)
Proposition 12.40. Let Ω be an upper-graph Lipschitz domain in Rn , or a bounded
Lipschitz domain in Rn . For each function u ∈ H 1/2 (∂Ω) set
uH 1/2 (∂Ω) := uL2 (∂Ω) (12.4.12)

|u(x) − u(y)|2 1/2
+ dσ(x) dσ(y) .
∂Ω ∂Ω |x − y|n

Then (12.4.12) is a norm on H 1/2 (∂Ω) and H 1/2 (∂Ω), · H 1/2 (∂Ω) is a Banach
space.
Proof. The proof of the fact that (12.4.12) is a norm on H 1/2 (∂Ω) is similar to the
proof used in Proposition 12.37, this time considering the set X := ∂Ω×∂Ω endowed
with the measure μ(x, y) := |x − y|−n σ(x) ⊗ σ(y), for (x, y) ∈ X.
To show that H 1/2 (∂Ω) is complete with respect to the norm · H 1/2 (∂Ω) , let
{u j } j∈N ⊆ H 1/2 (∂Ω) be a Cauchy sequence. Recalling (12.4.12) it follows that
{u j } j∈N is Cauchy in L2 (∂Ω) and {F j } j∈N is Cauchy in L2 (X, μ), where we have set
F j (x, y) := u j (x) − u j (y) for each point (x, y) ∈ X. In particular, there exists some
u ∈ L2 (∂Ω) such that {u j } j∈N converges to u both pointwise σ-a.e. on ∂Ω and in
L2 (∂Ω) as j → ∞, as well as some G ∈ L2 (X, μ) such that {F j } j∈N converges to G
both pointwise μ-a.e. on X and in L2 (X, μ) as j → ∞. The pointwise convergence
of these sequences forces G(x, y) to be equal to F(x, y) := u(x) − u(y) for σ-a.e.
x, y ∈ ∂Ω. Consequently, u ∈ H 1/2 (∂Ω) and
lim u j − uH 1/2 (∂Ω)

j→∞
= lim u j − uL2 (∂Ω) + lim F j − FL2 (X,μ) = 0. (12.4.13)

j→∞ j→∞

This proves that H 1/2 (∂Ω), · H 1/2 (∂Ω) is indeed a Banach space.
The next lemma ensures that the Sobolev space in Definition 12.39 is rather rich.
For a set E ⊂ Rn we denote by Lipcomp (E) the collection of functions in Lip(E)
which are compactly supported.
Lemma 12.41. Let Ω be either an upper-graph Lipschitz domain in Rn , or a
bounded Lipschitz domain in Rn . Then
Lipcomp (∂Ω) ⊆ H 1/2 (∂Ω). (12.4.14)

Proof. Since Lipcomp (∂Ω) ⊆ L2 (∂Ω), in view of (12.4.11) there remains to prove
that

|u(x) − u(y)|2
dσ(x) dσ(y) < ∞ for each u ∈ Lipcomp (∂Ω). (12.4.15)
∂Ω ∂Ω |x − y|n
Pick u ∈ Lipcomp (∂Ω). Then we may estimate

|u(x) − u(y)|2
1|x−y|>1 dσ(x) dσ(y)
∂Ω ∂Ω |x − y|n
1|x−y|>1
≤4 |u(y)|2
dσ(x) dσ(y)
∂Ω |x − y|
n
∂Ω
≤ Cu2L2 (∂Ω) < ∞, (12.4.16)
where the second inequality is due to Lemma 14.57 (applied with α := 1 and r := 1).
Fix a reference point x0 ∈ ∂Ω and pick some radius R > 0 large enough so that
supp u ⊆ B(x0 , R). Then for every choice of y ∈ ∂Ω \ B(x0 , R + 2) and x ∈ ∂Ω
with |x − y| ≤ 1 we necessarily have u(x) = 0 = u(y). Similarly, u(x) = 0 = u(y)
whenever x ∈ ∂Ω \ B(x0 , R + 2) and y ∈ ∂Ω satisfy |x − y| ≤ 1. Hence, if M denotes
the Lipschitz constant of u, i.e., M := sup x,y∈∂Ω, xy |u(x)−u(y)|
|x−y| , we may write for some
C = C(Ω) ∈ (0, ∞),

|u(x) − u(y)|2
1|x−y|≤1 dσ(x) dσ(y)
∂Ω ∂Ω |x − y|n

|u(x) − u(y)|2
= 1|x−y|≤1 dσ(x) dσ(y)
∂Ω∩B(x0 ,R+2) ∂Ω∩B(x0 ,R+2) |x − y|n

1|x−y|≤1
≤ M2 dσ(x) dσ(y)
∂Ω |x − y|
n−2
∂Ω∩B(x0 ,R+2)

≤ CM 2 σ ∂Ω ∩ B(x0 , R + 2) < ∞. (12.4.17)
Above, the inner-most integral in the penultimate line has been estimated using
(14.7.33) with α := 1 and r := 1, and the last inequality uses (14.7.34). Now
(12.4.15) follows from (12.4.16) and (12.4.17).
Exercise 12.42. Let Ω be an upper-graph Lipschitz domain in Rn and let Φ be

the bijective bi-Lipschitz flattening map associated with Ω as in Remark 14.52.
Throughout, identify ∂Rn+ canonically with Rn−1 . Prove that the following are true.
(i) For a measurable function u on ∂Ω we have
u ∈ L2 (∂Ω) ⇐⇒ u ◦ Φ ∈ L2 (∂Rn+ ) (12.4.18)
and
u ∈ H 1/2 (∂Ω) ⇐⇒ u ◦ Φ ∈ H 1/2 (∂Rn+ ), (12.4.19)

(ii) For a measurable function v on ∂Rn+ we have
v ∈ L2 (∂Rn+ ) ⇐⇒ v ◦ Φ−1 ∈ L2 (∂Ω) (12.4.20)
and
v ∈ H 1/2 (∂Rn+ ) ⇐⇒ v ◦ Φ−1 ∈ H 1/2 (∂Ω), (12.4.21)
Hint: If Ω ⊆ Rn is the upper-graph of a Lipschitz function ϕ : Rn−1 → R with

Lipschitz constant M, then for all x , y ∈ Rn−1 we have
√
1 ≤ 1 + |∇ϕ(x )|2 ≤ 1 + M 2 and
2 (12.4.22)
|x − y |2 ≤ (x , ϕ(x )) − (y , ϕ(y )) ≤ (1 + M 2 )|x − y |2 .
Use (12.4.12), (14.7.31), and (12.4.22) to prove the equivalences in (i). The state-
ments in (ii) are implied by (i) used with u := v ◦ Φ−1 .
Lemma 12.43. Let Ω be either a bounded Lipschitz domain, or an upper-graph

Lipschitz domain in Rn . Also, suppose Ω1 is either a bounded Lipschitz domain, or
an upper-graph Lipschitz domain in Rn and assume that there exist x0 ∈ ∂Ω ∩ ∂Ω1
and an open neighborhood C of x0 with the property that ∂Ω ∩ C = ∂Ω1 ∩ C. Let K
be a compact set in Rn such that K ⊂ C. For u ∈ H 1/2 (∂Ω) which vanishes outside
∂Ω \ K define the function
⎧
⎪
⎨ u on ∂Ω1 ∩ C = ∂Ω ∩ C,
⎪
v := ⎪
⎪ (12.4.23)
⎩ 0 on ∂Ω1 \ C.
Then v ∈ H 1/2 (∂Ω1 ) and there exist two constants C1 , C2 ∈ (0, ∞) depending only
Ω, Ω1 , C, and K, such that
C1 vH 1/2 (∂Ω1 ) ≤ uH 1/2 (∂Ω) ≤ C2 vH 1/2 (∂Ω1 ) . (12.4.24)
Proof. It is immediate that
vL2 (∂Ω1 ) = uL2 (∂Ω1 ∩C) = uL2 (∂Ω∩C) = uL2 (∂Ω) . (12.4.25)

Next, pick ε ∈ 0, dist (K , ∂C) and break up the double integral in the H 1/2 (∂Ω1 )
norm of v as follows (with σ1 denoting the surface measure on ∂Ω1 )

|v(x) − v(y)|2
dσ1 (x) dσ1 (y) =: I + II, (12.4.26)
∂Ω1 ∂Ω1 |x − y|n
where

|v(x) − v(y)|2
I := 1|x−y|>ε (y) dσ1 (x) dσ1 (y), (12.4.27)
∂Ω1 ∂Ω1 |x − y|n

|v(x) − v(y)|2
II := 1|x−y|≤ε (y) dσ1 (x) dσ1 (y). (12.4.28)
∂Ω1 ∂Ω1 |x − y|n
To estimate I write

|v(x)|2 + |v(y)|2
I≤2 1|x−y|>ε dσ1 (x) dσ1 (y)
∂Ω1 ∂Ω1 |x − y|n

1|x−y|>ε
=4 |v(x)|2 dσ1 (y) dσ1 (x)
∂Ω1 |x − y|
n
∂Ω1
≤ Cε−1 v2L2 (∂Ω1 ) = Cε−1 u2L2 (∂Ω) , (12.4.29)
where for the first equality in (12.4.29) we used Fubini’s theorem, the next inequality
is due to (14.7.39) in Lemma 14.57 (with α := 1 and r := ε), and the last equality
uses (12.4.25).
In order to estimate II observe that if x ∈ ∂Ω1 \ C, y ∈ ∂Ω1 , and |x − y| ≤ ε,
then the choice of ε and the fact that supp v ⊆ K imply v(x) = 0 = v(y). Similarly,
v(x) = 0 = v(y) whenever x ∈ ∂Ω1 , y ∈ ∂Ω1 \ C, and |x − y| ≤ ε. Consequently,

|v(x) − v(y)|2
II = 1|x−y|≤ε dσ1 (x) dσ1 (y)
∂Ω1 ∩C ∂Ω1 ∩C |x − y|n

|u(x) − u(y)|2
= 1|x−y|≤ε dσ(x) dσ(y)
∂Ω∩C ∂Ω∩C |x − y|n

|u(x) − u(y)|2
≤ dσ(x) dσ(y). (12.4.30)
∂Ω ∂Ω |x − y|n
In concert, (12.4.25)–(12.4.30) and (12.4.12) yield the first inequality in (12.4.24).
The second inequality in (12.4.24) is proved similarly.
Theorem 12.44. Let Ω be either an upper-graph Lipschitz domain in Rn , or a

bounded Lipschitz domain in Rn . Then the following are true.
(1) If ϕ ∈ Lipcomp (∂Ω) then ϕu ∈ H 1/2 (∂Ω) for every u ∈ H 1/2 (∂Ω) and
ϕuH 1/2 (∂Ω) ≤ C(ϕ, Ω)uH 1/2 (∂Ω) , (12.4.31)
for some constant C(ϕ, Ω) ∈ (0, ∞) independent of u.

(2) The set Lipcomp (∂Ω) is dense in H 1/2 (∂Ω).
Proof. Fix ϕ ∈ Lipcomp (∂Ω) and u ∈ H 1/2 (∂Ω). Then
ϕuL2 (∂Ω) ≤ ϕL∞ (∂Ω) uL2 (∂Ω) . (12.4.32)
Also, adding and subtracting ϕ(x)u(y) we may write


|(ϕu)(x) − (ϕu)(y)|2
dσ(x) dσ(y) (12.4.33)
∂Ω ∂Ω |x − y|n

|ϕ(x)|2 |u(x) − u(y)|2
≤2 dσ(x) dσ(y)
∂Ω ∂Ω |x − y|n

|ϕ(x) − ϕ(y)|2
+2 |u(y)|2 dσ(x) dσ(y) =: I + II.
∂Ω ∂Ω |x − y|n
It is immediate that

|u(x) − u(y)|2
I ≤ 2ϕL∞ (∂Ω) dσ(x) dσ(y). (12.4.34)
∂Ω ∂Ω |x − y|n
Moreover, if M denotes the Lipschitz constant of ϕ, for each y ∈ ∂Ω we may apply
Lemma 14.56 and Lemma 14.57 with α := 1 and r := 1 to estimate

|ϕ(x) − ϕ(y)|2
dσ(x) dσ(y) (12.4.35)
∂Ω |x − y|n

1|x−y|≤1 1|x−y|>1
≤ M2 dσ(x) + 4ϕ L ∞ (∂Ω) dσ(x) ≤ C,
∂Ω |x − y| ∂Ω |x − y|
n−2 n
for some constant C = C(ϕ, Ω) ∈ (0, ∞) independent of y. In turn, (12.4.35) may be

used to estimate II as

II ≤ C |u(y)|2 dσ(x). (12.4.36)
∂Ω
Now (12.4.31) follows by combining (12.4.12), (12.4.32), (12.4.33), (12.4.34), and

(12.4.36).
Moving on to the proof of the density result, consider first the case when Ω is
an upper-graph Lipschitz domain and let Φ be the function associated with Ω as in
Remark 14.52. We pick some u ∈ H 1/2 (∂Ω) and make use of (12.4.19) to observe
that u ◦ Φ ∈ H 1/2 (∂Rn+ ) ≡ H 1/2 (Rn−1 ). Apply Corollary 12.38 to obtain a sequence
{gk }k∈N ⊆ C0∞ (Rn−1 ) ⊆ Lipcomp (Rn−1 ) (12.4.37)
H 1/2 (Rn−1 )
with the property that gk −−−−−−−→ u ◦ Φ. Since Φ is a bijective bi-Lipschitz func-
k→∞
tion we have that gk ◦ Φ−1 ∈ Lipcomp (∂Ω) for each k ∈ N. Also, (12.4.21) and its
H 1/2 (∂Ω)
accompanying estimate further imply gk ◦ Φ−1 −−−−−−→ u. This proves the density
k→∞
result stated in (2) when Ω is an upper-graph Lipschitz domain.
Next, consider the case when Ω is a bounded Lipschitz domain. Let {C x∗j }Nj=1 ,
{Ω j }Nj=1 , and {ψ j }Nj=1 be the families of cylinders, of upper-graph Lipschitz domains,
and the partition of unity associated with Ω as in Remark 14.54. Also, for each j
in {1, . . . , N} fix ξ j ∈ C0∞ (C x∗j ) with the property that ξ j ≡ 1 near supp ψ j . Then, if
12.5 Traces and Extensions 477
N
u ∈ H 1/2 (∂Ω) we have u = j=1 ψ j u. Theorem 12.44 implies ψ j u ∈ H 1/2 (∂Ω) for all
j ∈ {1, . . . , N}. Fix j ∈ {1, . . . , N} and define the function
⎧
⎪
⎨ ψ j u on ∂Ω j ∩ C x j = ∂Ω ∩ C x j ,
⎪
⎪ ∗ ∗
u j := ⎪⎪ (12.4.38)
⎪
⎩0 on ∂Ω j \ C x∗ . j
By Lemma 12.43 applied with Ω := Ω, Ω1 := Ω j , u := ψ j u, and v := u j we have

u j ∈ H 1/2 (∂Ω j ). Since Ω j is an upper-graph Lipschitz domain, based on what we
proved earlier, there exists a sequence {gkj }k∈N ⊆ Lipcomp (∂Ω j ) with the property that
H 1/2 (∂Ω j )
gkj −−−−−−−→ u j . In addition, by relying on item (1) in Theorem 12.44, the fact that
k→∞
supp u j ⊆ (supp ψ j ) ∩ ∂Ω j , and that ξ j ≡ 1 near supp ψ j , we obtain
H 1/2 (∂Ω j )
ξ j gkj −−−−−−−→ ξ j u j = u j . (12.4.39)
k→∞
If we now define
⎧
⎪ j
⎨ ξ j gk on ∂Ω ∩ C x∗j = ∂Ω j ∩ C x∗j ,
⎪
Ψkj := ⎪
⎪ ∀ k ∈ N, (12.4.40)
⎩0 on ∂Ω \ C x∗j ,
then {Ψkj }k∈N ⊆ Lipcomp (∂Ω) and we may invoke Lemma 12.43 to conclude that
H 1/2 (∂Ω)
Ψkj −−−−−−→ ψ j u. Since the latter holds for each j ∈ {1, . . . , N} and u = Nj=1 ψ j u, the
k→∞
desired conclusion follows.
12.5 Traces and Extensions
In the first part of this section we address the issue of existence of traces of functions
in the Sobolev space H 1 (Ω) considered in a bounded Lipschitz domain Ω on the
boundary of the respective domain. The reader is reminded that in such a setting we
have H 1 (Ω) = H 1 (Ω), with equivalent norms (cf. Theorem 12.30).
Theorem 12.45. Let Ω be a bounded Lipschitz domain in Rn . Then the following
are true.
(1) The vector space V(Ω) := C 0 (Ω) ∩ Liploc (Ω) ∩ H 1 (Ω) is a dense subspace of
H 1 (Ω).
(2) For each u ∈ V(Ω) its restriction to the boundary u∂Ω ∈ C 0 (∂Ω) also belongs to
H 1/2 (∂Ω) and there exists C = C(Ω) ∈ (0, ∞) such that

u∂Ω H 1/2 (∂Ω) ≤ CuH 1 (Ω) . (12.5.1)

(3) The map V(Ω) u → u∂Ω ∈ C 0 (∂Ω) ∩ H 1/2 (∂Ω) has a unique extension to a
linear and bounded operator
Tr : H 1 (Ω) −→ H 1/2 (∂Ω). (12.5.2)
Proof. The density statement in (1) follows by observing that C0∞ (Ω) ⊆ V(Ω) and
recalling that the set C0∞ (Ω) is dense in H 1 (Ω) (cf. Theorem 12.14). Also, the state-
ment in (3) is an immediate consequence of items (1)-(2). Hence, we are left with
showing (12.5.1) whose proof we divide in three steps.
Step I. The case when Ω = Rn+ . In this scenario fix R > 0 and assume that

u ∈ C 0 Rn+ ∩ Lip loc (Rn+ ) ∩ H 1 (Rn+ )
(12.5.3)
and supp u ⊆ [−R, R]n−1 × [0, R].
The goal is to show that

u∂Rn belongs to the space H 1/2 (∂Rn+ ) ≡ H 1/2 (Rn−1 )
+
(12.5.4)
and u∂Rn H 1/2 (∂Rn ) ≤ CuH 1 (Rn+ ) ,
+ +
for some constant C = C(R, n) ∈ (0, ∞) independent of u. To get started, introduce

the notation x = (x , 0) for points x ∈ ∂Rn+ . In view of (12.4.12) the aim is to estimate
1/2
|u(x , 0)|2 dx (12.5.5)
Rn−1

|u(x , 0) − u(y , 0)|2 1/2
+ dy dx
Rn−1 Rn−1 |x − y |n
by a fixed multiple of uH 1 (Rn+ ) .
Using the assumptions on u, for each x ∈ [−R, R]n−1 , we may estimate
R 2
|u(x , 0)|2 = |u(x , 0) − u(x , R)|2 ≤ |(∂n u)(x , t)| dt
0
R
≤R |∇u(x , t)|2 dt, (12.5.6)
0
where for the first inequality we used Lemma 14.20 with f (t) := u(x , t), t ∈ [0, R],
and for the second inequality we used Hölder. Consequently,

|u(x , 0)|2 dx = |u(x , 0)|2 dx
Rn−1 |x |≤R
R
≤R |∇u(x , t)|2 dt dx
|x |≤R 0

=R |(∇u)(x)|2 dx. (12.5.7)
Rn+
To treat the double integral in (12.5.5), apply the triangle inequality to write

|u(x , 0) − u(y , 0)|2

dy dx ≤ 3 I1 + I2 + I3 ), (12.5.8)
Rn−1 Rn−1 |x − y | n
where

|u(x , 0) − u(x , |x − y |)|2
I1 := dy dx , (12.5.9)
Rn−1 Rn−1 |x − y |n

|u(x , |x − y |) − u(y , |x − y |)|2
I2 := dy dx , (12.5.10)
Rn−1 Rn−1 |x − y |n

|u(y , |x − y |) − u(y , 0)|2
I3 := dy dx . (12.5.11)
Rn−1 Rn−1 |x − y |n
We proceed to estimate the integrals in (12.5.9)–(12.5.11), starting with integral I1 .
For each x , y ∈ Rn−1 we apply Lemma 14.20 to the real-valued function defined by
f (t) := u(x , t|x − y |) for t ∈ [0, 1] and write

1
1
2
I1 ≤ |x − y | (∂n u) x , (1 − t)|x − y | dt dy dx
Rn−1 Rn−1 |x − y |n 0

1 1
2
≤
|(∂n u) x , t|x − y | | dt dy dx . (12.5.12)
Rn−1 Rn−1 |x − y |n−2
0
Next, for each fixed x ∈ Rn−1 , we use polar coordinates to write y = x + ρ ω,

where ω ∈ S n−2 and ρ ∈ (0, ∞). Then (12.5.12) further yields

ρ n−2 1 ∞ 2
I1 ≤ |(∂n u)(x , ρ t)| dt dρ dω dx
Rn−1 S n−2 0 ρ
n−2
0
∞ 1 2
= ωn−2 |(∂n u)(x , ρ t)| dt dρ dx
Rn−1 0 0
∞ ρ 2
= ωn−2 ρ−2 |(∂n u)(x , τ)| dτ dρ dx (12.5.13)
Rn−1 0 0
where ωn−2 denotes the area of the unit sphere in Rn−1 and the last equality in
(12.5.13) is based on the change of variables τ := ρ t. Moreover, for each x ∈ Rn−1
we apply Hardy’s inequality (cf. (14.2.31) in Theorem 14.21, presently used with
p := 2, r := 1, and f := |(∂n u)(x , ·)|) to obtain
∞ ρ 2 ∞
−2
ρ |(∂n u)(x , τ)| dτ dρ ≤ 4 |(∂n u)(x , τ)|2 dτ. (12.5.14)
0 0 0
Combining (12.5.13), (12.5.14), and Fubini’s theorem we arrive at


I1 ≤ 4 ωn−2 |(∂n u)(x)|2 dx. (12.5.15)
Rn+
By interchanging the roles of x and y , a similar argument yields

I3 ≤ 4 ωn−2 |(∂n u)(x)|2 dx. (12.5.16)
Rn+
Next, we estimate I2 . By the integral version of the Mean Value Theorem in Rn−1
and the Cauchy–Schwarz inequality we may write
1
1
I2 = (x , |x − y |) − (y , |x − y |) ·
Rn−1 Rn−1 |x − y |n 0
2
· (∇u) t(x , |x − y |) + (1 − t)(y , |x − y |) dt dy dx

1
= ×
Rn−1 Rn−1 |x − y |n
1 2
× (x − y , 0) · (∇u) tx + (1 − t)y , |x − y | dt dy dx
0
$ 1 1/2
1
≤ ×
Rn−1 Rn−1 0 |x − y |n−2
%2
× |(∇n−1 u) tx + (1 − t)y , |x − y | | dt dy dx , (12.5.17)
where ∇n−1 u denotes the first n − 1 components of ∇u. To further estimate the last
expression in (12.5.17) we invoke the generalized Minkowski inequality (cf. Theo-
rem 14.22) and obtain
$ 1
1
I2 ≤ − y |n−2
×
0 R n−1 R n−1 |x
1/2 %2
× |(∇n−1 u) y + t(x − y ), |x − y | |2 dx dy dt . (12.5.18)
Next, we make a series of changes of variables. First, for each fixed y ∈ Rn−1 we
set z := x − y and then we use Fubini’s theorem to write
$ 1
1 1/2 %2
I2 ≤ |(∇n−1 u)(y + tz , |z |)|2 dy dz dt . (12.5.19)
0 Rn−1 Rn−1 |z |n−2
Second, for each t ∈ [0, 1] and each z ∈ Rn−1 fixed we let ξ := y + tz and the
estimate in (12.5.19) becomes
$ 1
1 1/2 %2
I2 ≤ |(∇n−1 u)(ξ , |z |)|2 dξ dz dt
0 Rn−1 Rn−1 |z |n−2

1
= |(∇n−1 u)(ξ , |z |)|2 dξ dz . (12.5.20)
Rn−1 Rn−1 |z |n−2
Third, we pass to polar coordinates in z by setting z := ρ ω with ρ ∈ (0, ∞) and
ω ∈ S n−2 , which allows us to write the last double integral in (12.5.20) as

1
|n−2
|(∇n−1 u)(ξ , |z |)|2 dξ dz
R n−1 R n−1 |z
∞
ρn−2
= |(∇n−1 u)(ξ , ρ)|2 dξ dω dρ
S n−2 Rn−1 ρ
n−2
0
∞
= ωn−2 |(∇n−1 u)(ξ , ρ)|2 dξ dρ
0 Rn−1

= ωn−2 |(∇n−1 u)(x)| p dx, (12.5.21)
Rn+
where the last equality uses Fubini’s theorem. From (12.5.21) and (12.5.20) it fol-
lows that
I2 ≤ ωn−2 |(∇n−1 u)(x)|2 dx. (12.5.22)
Rn+
In concert, (12.5.8), (12.5.22), (12.5.16), and (12.5.15), then yield

|u(x , 0) − u(y , 0)|2
dy dx (12.5.23)
Rn−1 Rn−1 |x − y |n

≤ 27 ωn−2 |(∇u)(x)|2 dx.
Rn+
Finally, by combining (12.5.7), (12.5.23), and (12.3.127), it follows that if u is as in

(12.5.3) then u satisfies (12.5.4). This finishes the proof of Step I.
Step II. The upper-graph Lipschitz domain case. The goal is to prove that
if Ω ⊆ Rn is an upper-graph Lipschitz domain, K is some com-

pact subset of Rn , and u ∈ C 0 (Ω) ∩ Liploc (Ω) ∩ H 1 (Ω) vanishes
outside Ω\ (12.5.24)
K, there exists a finite constant C = C(Ω, K) > 0
then
such that u∂Ω H 1/2 (∂Ω) ≤ CuH 1 (Ω) .
To this end, let ϕ : Rn−1 → R be the Lipschitz function with the property that

Ω = (y , yn ) ∈ Rn−1 × R : yn > ϕ(y ) (12.5.25)
and denote by M its Lipschitz constant. In particular, we have

∂Ω = (y , ϕ(y )) : y ∈ Rn−1 . (12.5.26)
Recall the function Φ from Remark 14.52. Having fixed some compact set K ⊂ Rn ,
pick u ∈ C 0 (Ω) ∩ Liploc (Ω) ∩ H 1 (Ω) which vanishes identically on Ω \ K, and define

the function w(x , t) := u x , ϕ(x ) + t for each x ∈ Rn−1 and each t ∈ [0, ∞). Since
w = u ◦ Φ, with Φ as in (14.7.6), we have that
w ∈ C 0 (Rn+ ) ∩ Liploc (Rn+ ) ∩ H 1 (Rn+ ) and has bounded support. (12.5.27)
Also, from the chain rule formula (12.3.128) proved in Theorem 12.32 and the
fact that DΦ and DΦ−1 are bounded, we see that there exists a constant C ∈ (0, ∞)
depending only on Ω such that

(∇w)(x , t) ≤ C (∇u)(Φ(x , t)) for a.e. (x , t) ∈ Rn+ . (12.5.28)
After this preamble we are ready to proceed with the estimate in (12.5.24). Based
on (14.7.30) we obtain
2 &
|u(x)|2 dσ(x) = u(x , ϕ(x )) 1 + |∇ϕ(x )|2 dx
∂Ω Rn−1
√ 2
≤ 1 + M2 w(x , 0) dx (12.5.29)
Rn−1
√ 2
≤ CK 1 + M2 (∇w)(x) dx
Rn+
2
≤C (∇u)(Φ(x)) dx ≤ C |(∇u)(y)|2 dy.
Rn+ Ω
For the first inequality in (12.5.29) we used the fact that ϕ is Lipschitz and the
definition of w, for the second one we applied (12.5.7) (bearing in mind (12.5.27)),
for the third inequality we invoked (12.5.28), while for the last inequality we have
used Corollary 12.33.
A similar circle of ideas also gives

|u(x) − u(y)|2
dσ(x) dσ(y) (12.5.30)
∂Ω ∂Ω |x − y|n
2
u(x , ϕ(x )) − u(y , ϕ(y ))
= n ×
Rn−1 R
n−1 (x − y , ϕ(x ) − ϕ(y ))
& &
× 1 + |∇ϕ(x )|2 1 + |∇ϕ(y )|2 dx dy
2
w(x , 0) − w(y , 0)
≤ (1 + M ) 2
dx dy
Rn−1 Rn−1 |x − y |n
2
≤ (1 + M 2 ) (∇w)(x) dx ≤ C (∇u)(Φ(x))2 dx
Rn+ Rn+

≤C |(∇u)(y)|2 dy.
Ω
The equality in (12.5.30) comes from (14.7.30). The first inequality in (12.5.30) uses
the fact that ϕ is Lipschitz and the definition of w, the second inequality is based on
(12.5.23) (whose applicability is ensured by (12.5.27)), the third inequality uses
(12.5.28), while the last inequality is justified by Corollary 12.33.
Finally, combining
(12.5.29), (12.5.30),
and (12.4.12) at this stage we may con-
clude that u∂Ω ∈ H 1/2 (∂Ω) and u∂Ω H 1/2 (∂Ω) ≤ CuH 1 (Ω) , finishing the proof of
(12.5.24).
Step III. Localization. Assume Ω is a bounded Lipschitz domain and recall that in
such a setting we have H 1 (Ω) = H 1 (Ω), with equivalent norms (cf. Theorem 12.30).
Remark 14.54 ensures the existence of a family of cylinders {C x∗j }Nj=1 satisfying
(14.7.19), a family of upper-graph Lipschitz domains {Ω j }Nj=1 , and a partition of
unity {ψ j }Nj=0 (cf. (14.7.22)). In view of (14.7.19) and Lemma 14.53, we have
Ω j ∩ C x∗j = Ω ∩ C x∗j , (12.5.31)
Ω j ∩ C x∗j = Ω ∩ C x∗j , (12.5.32)
∂Ω j ∩ C x∗j = ∂Ω ∩ C x∗j , (12.5.33)
for each j ∈ {0, 1, . . . , N}. Pick u ∈ V(Ω). Item (1) in Theorem 12.15 then implies
that ψ j u ∈ H 1 (Ω) for each j ∈ {0, 1, . . . , N}.
Next, fix j ∈ {1, . . . , N} and define the function
⎧
⎪
⎨ ψ j u in Ω j ∩ C x∗j = Ω ∩ C x∗j ,
⎪
u j := ⎪
⎪ (12.5.34)
⎩0 in Ω j \ C x∗j .
From (12.5.34), the properties of u, and Example 12.20 (also mindful of Theo-
rem 12.30 and the generalized Leibniz formula (2.4.18)) it follows that
u j ∈ C 0 (Ω j ) ∩ Liploc (Ω j ) ∩ H 1 (Ω j ), it vanishes outside a

bounded subset of Ω j , and has u j H 1 (Ω j ) ≤ CuH 1 (Ω) for (12.5.35)
some constant C ∈ (0, ∞) depending only on Ω.
Granted these properties, we may invoke (12.5.24) to conclude that

u j ∂Ω belongs to H 1/2 (∂Ω j ) and
j (12.5.36)
u ≤ Cu 1
j ∂Ω H 1/2 (∂Ω ) ≤ Cu 1 .
j H (Ω j ) H (Ω)
j j
Let us also observe that (12.5.34) implies that

⎧
⎪
⎨ u j ∂Ω j on ∂Ω ∩ C x∗j = ∂Ω j ∩ C x∗j ,
⎪
(ψ j u)∂Ω = ⎪
⎪ (12.5.37)
⎩0 on ∂Ω \ C x∗j .
With (12.5.36)–(12.5.37)
in
hand, Lemma 12.43 applies (with Ω := Ω j , Ω1 := Ω,
u := u j ∂Ω , and v := (ψ j u)∂Ω ) and gives that (ψ j u)∂Ω ∈ H 1/2 (∂Ω) and
j

(ψ j u)∂Ω H 1/2 (∂Ω) ≤ C u j ∂Ω H 1/2 (∂Ω ) ≤ CuH 1 (Ω) , (12.5.38)
j j
for some constant C ∈ (0, ∞) depending only on Ω. Upon recalling that we may

express u∂Ω = (ψ j u)∂Ω on ∂Ω, and employing (12.5.38), we obtain
N
j=1

u∂Ω ∈ H 1/2 (∂Ω) and u∂Ω H 1/2 (∂Ω) ≤ CuH 1 (Ω) (12.5.39)
for some constant C ∈ (0, ∞) depending only on Ω. This finishes the proof of
Theorem 12.45.
It turns out that the trace operator from Theorem 12.45 is surjective in the context
of (12.5.2). Remarkably, the said trace operator has an inverse from the right, which
is the extension mapping described in the theorem below.
Theorem 12.46. Let Ω ⊆ Rn be a bounded Lipschitz domain. Then there exists a
mapping
Ex : H 1/2 (∂Ω) −→ H 1 (Ω) (12.5.40)
that is linear, bounded, and satisfies

Tr Ex( f ) = f, ∀ f ∈ H 1/2 (∂Ω). (12.5.41)
where Tr is the trace operator from (12.5.2).

Proof. We divide the proof in a number of steps.
Step I. The case when Ω = Rn+ . The goal in this step is to prove that there exists a
linear operator ExRn+ : Lipcomp (Rn−1 ) → C 0 (Rn+ ) ∩ C ∞ (Rn+ ) satisfying

ExRn+ f ∂Rn = f for each f ∈ Lipcomp (Rn−1 ), (12.5.42)
+
and with the property that there exists a finite positive constant Cn such that

| f (y ) − f (z )|2 1/2
∇(ExRn+ f )L2 (Rn ) ≤ Cn dy dz (12.5.43)
+
Rn−1 Rn−1 |y − z |n
for every f ∈ Lipcomp (Rn−1 ), and for each compact set K ⊂ Rn there exists a finite
positive constant C K such that

ExRn+ f L2 (Rn ∩K) ≤ C K f L2 (Rn−1 ) (12.5.44)
+
for every f ∈ Lipcomp (Rn−1 ).
To this end, we begin by picking a function
η ∈ C ∞ (Rn ), 0 ≤ η ≤ 1 on Rn ,
(12.5.45)
supp η ⊆ B(0, 4), η ≡ 1 on B(0, 2).
Next, define the kernel

k : Rn+ × Rn+ −→ R (12.5.46)
by setting
x − y $ x − (z , 0) %−1
k(x, y) := η η dz ,
xn Rn−1 xn (12.5.47)
for each x = (x1 , . . . , xn ) ∈ Rn+ and eachy ∈ Rn+ .
Note that k is well-defined. Indeed,

x − (z , 0) x − z

η dz = η , 1 dz
Rn−1 xn Rn−1 xn

= xnn−1 η(w , 1) dw = co xnn−1 > 0, (12.5.48)
Rn−1

where co := Rn−1 η(w , 1) dw is a real nonzero constant (as seen from (12.5.45)).
In particular, we have
x − y
k(x, y) = co η xn1−n ,
xn (12.5.49)
for each x = (x , xn ) ∈ Rn−1 × (0, ∞) and eachy ∈ Rn+ .
Clearly k is a nonnegative function and, by design,

k x, (y , 0) dy = 1, ∀ x ∈ Rn+ . (12.5.50)
Rn−1
Moreover, (12.5.49) and the regularity of η, imply k ∈ C ∞ (Rn+ × Rn+ ).

In what follows we will need estimates for derivatives in the first variable of k
(see (12.5.57) below), a task we take up next. To get started, fix α ∈ Nn0 and use
Leibniz’s rule to compute (starting from (12.5.49))
α! β x − y γ 1−n
∂αx k(x, y) = co ∂x η ∂ x [xn ]. (12.5.51)
β+γ=α
β!γ! xn
Furthermore,
⎧
⎪
⎪ |γ|−1
'
⎪
⎪
γ 1−n ⎨ (1 − n − j) xn1−n−|γ| , if γ = (0, ..., 0, γn ),
∂ x xn =⎪
⎪ (12.5.52)
⎪
⎪
⎩ 0,
j=0
otherwise,
while by the Chain Rule we have

$ x − y %
∂βx η (12.5.53)
xn
β,δ
x − y P2|β| −|β| (x1 − y1 , ..., xn − yn , xn )
= (∂δ η) .
|δ|≤|β|
xn xn2|β|
Above, generally speaking, Pβ,δ r (t1 , ..., tn , tn+1 ) is a homogeneous polynomial of

degree r in the variables t1 , ..., tn+1 , that is,

Pβ,δ
r (t) = aβ,δ
γ t ,
γ
t = (t1 , ..., tn , tn+1 ) ∈ Rn+1 , (12.5.54)
|γ|=r
where the aβ,δ

γ ’s are real-coefficients. Formula (12.5.53) may be justified by starting
from the observation that, for each j ∈ {1, ..., n} and for each differentiable function
F, there holds
$ x − y % n x − y δ jk xn − (xk − yk )δ jn
∂x j F = (∂k F) , (12.5.55)
xn k=1
xn xn2
and then using induction on the length of the multi-index β ∈ Nn0 . In particular, from
(12.5.53) we see that for each x = (x , xn ) ∈ Rn+ and each y ∈ Rn+ we have
x−y
∈ supp ∂δ η =⇒ |x − y| ≤ 4xn
xn

=⇒ Pβ,δ
|β|
|β|
2 −|β|
(x1 − y 1 , ..., xn − yn , xn ) ≤ Cn,β,δ xn2 −|β|
(
β x − y ) −|β|
=⇒ ∂ x η ≤ C xn χ|x−y|<4xn . (12.5.56)
xn
Collectively, (12.5.51), (12.5.52), and (12.5.56) imply that the function k satisfies

(∂αx k)(x, y) ≤ Cn,α xn1−n−|α| χ|x−y|<4xn ,
(12.5.57)
∀ x = (x , xn ) ∈ Rn+ , ∀ y ∈ Rn+ , ∀ α ∈ Nn0 .
We are now ready to define the extension operator

(ExRn+ f )(x) := k x, (y , 0) f (y ) dy ,
Rn−1 (12.5.58)
for each f ∈ Lipcomp (R n−1
) and all x ∈ Rn+ .
Note that (12.5.57) (applied with α = (0, . . . , 0)) ensures that ExRn+ f is well-defined
whenever f ∈ Lipcomp (Rn−1 ). Also, thanks to (12.5.57), we have that ExRn+ f inherits
the regularity of k, i.e., ExRn+ f ∈ C ∞ (Rn+ ).
We claim that there exists some Cn ∈ (0, ∞) such that (12.5.43) holds for each
f ∈ Lipcomp (Rn−1 ). To justify this claim, fix f ∈ Lipcomp (Rn−1 ). Also fix x ∈ Rn+ and
z ∈ Rn−1 . Then (12.5.50) gives

f (z ) = k x, (y , 0) f (z ) dy , (12.5.59)
Rn−1
which further implies

0= ∂ x j k x, (y , 0) f (z ) dy , ∀ j ∈ {1, . . . , n}. (12.5.60)
Rn−1
Combined, (12.5.60) and (12.5.58) imply

[∇(ExRn+ f )](x) ≤ (∇ x k)(x, (y , 0)) | f (y ) − f (z )| dy . (12.5.61)
Rn−1
In turn, from (12.5.61), (12.5.57), and Hölder’s inequality we obtain

2 2
−n
[∇(ExRn+ f )](x) ≤ C xn | f (y ) − f (z )| dy
|x−(y ,0)|<4xn

≤ Cxn−2n · xnn−1 | f (y ) − f (z )|2 dy . (12.5.62)
|x−(y ,0)|<4xn
At this stage, average the most extreme sides of (12.5.62) in z ∈ Rn−1 such that
|x − (z , 0)| < 4xn in order to obtain
2
[∇(ExRn+ f )](x) (12.5.63)

≤ Cxn−2n | f (y ) − f (z )|2 dy dz
|x−(z ,0)|<4xn |x−(y ,0)|<4xn
for each x = (x , xn ) ∈ Rn+ . Consequently, Fubini’s theorem allows us to write

[∇(Ex n f )](x)2 dx
R+
Rn+
* +
≤C | f (y ) − f (z )|2 xn−2n dx dy dz . (12.5.64)
Rn−1 Rn−1
|x−(z ,0)|<4xn
|x−(y ,0)|<4xn
Observe
√ that on the domain
√ of integration of the√ inner-most integral |x − z | <
15 xn and |x − y | < 15 xn , hence |y − z | < 2 15 xn by the triangle inequality.

Bearing this in mind and using Fubini’s theorem, we may write

∞
−2n
xn dx ≤ √ √ 1 dx xn−2n dxn
|y −z |/(2 15) |x −z |< 15 xn
|x−(z ,0)|<4xn
|x−(y ,0)|<4xn
∞
=C √ xn−1−n dxn = C|y − z |−n , (12.5.65)
|y −z |/(2 15)
where C = C(n) > 0 is a finite constant given that −1 − n < −1. At this stage,
(12.5.43) follows from (12.5.64) and (12.5.65).
Moving on, we claim that for each compact set K ⊂ Rn there exists a constant
C K ∈ (0, ∞) with the property that the estimate recorded in (12.5.44) holds for
each function f ∈ Lipcomp (Rn−1 ). To see why this is the case, given such a compact
K choose some radius R ∈ (0, ∞) with the property that K ⊂ B(0, R). Also, fix
f ∈ Lipcomp (Rn−1 ) and let x ∈ Rn+ . Then formula (12.5.58), estimate (12.5.57), and
Hölder’s inequality give
2
|(ExRn+ f )(x)| ≤
2
|k x, (y , 0) || f (y )| dy
Rn−1
2
≤C xn1−n | f (y )| dy
|x−(y ,0)|<4xn

≤ Cxn2−2n xnn−1 | f (y )|2 dy
|x−(y ,0)|<4xn

= Cxn1−n | f (y )|2 dy (12.5.66)
|x−(y ,0)|<4xn
for each x = (x , xn ) ∈ Rn+ . Hence,


|(ExRn+ f )(x)|2 dx ≤ |(ExRn+ f )(x)|2 dx
Rn+ ∩K Rn+ ∩B(0,R)

≤C xn1−n | f (y )|2 dy dx
Rn+ ∩B(0,R) |x−(y ,0)|<4xn
$ R %
≤C | f (y )|2 √ 1 dx xn1−n dxn dy
Rn−1 0 |x −y |< 15 xn

≤ CR | f (y )|2 dy , (12.5.67)
Rn−1
where for the second inequality we applied Fubini’s theorem. This proves (12.5.44).
To complete the proof of the goal for Step I we are left with the task of establish-
ing that for each f ∈ Lipcomp (Rn−1 )
ExRn+ f extends continuously to Rn+ and

(12.5.68)
(ExRn+ f )∂Rn (x ) = f (x ), ∀ x ∈ Rn−1 ≡ ∂Rn+ .
+
To this end, fix f ∈ Lipcomp (Rn−1 ) along with some x∗ ∈ Rn−1 . Also, let ε > 0 be
fixed. Since f is continuous at x∗ , there exists δ > 0 such that
y ∈ Rn−1 , |y − x∗ | < δ =⇒ | f (y ) − f (x∗ )| < ε. (12.5.69)
Then for each x = (x , xn ) ∈ Rn+ we may estimate

(ExRn+ f )(x) − f (x∗ ) =

k x, (y , 0) f (y ) − f (x∗ ) dy
R
n−1
≤ k x, (y , 0) | f (y ) − f (x )| dy

∗
Rn−1

≤ Cn − √ | f (y ) − f (x∗ )| dy , (12.5.70)
|x −y |< 15 xn
where the equality uses (12.5.50), while for the last inequality we have used

(12.5.57) and the fact that the set √ {y ∈ R
n−1
: |x − (y , 0)| < 4xn } is contained

in the set {y ∈ R : |x − y | < 15 xn }. Thus, (12.5.69) and (12.5.70) yield
n−1
√
(ExRn+ f )(x) − f (x∗ ) ≤ ε if |x − x∗ | < δ/2 and xn < δ/(2 15), (12.5.71)
and the claims in (12.5.68) readily follow from this. In particular, (12.5.71) implies

ExRn+ f ∂Rn = f . This completes the proof of the stated goal for Step I.
+
Step II. The case when Ω is an upper-graph Lipschitz domain. The claim we make
is that there exists a linear map ExΩ : Lipcomp (∂Ω) → C 0 (Ω) ∩ Liploc (Ω) satisfying

ExΩ f ∂Ω = f for each f ∈ Lipcomp (∂Ω), (12.5.72)
and with the property that there exists a finite positive constant Cn such that
| f (y) − f (z)|2 1/2
∇(ExΩ f )L2 (Ω) ≤ Cn dσ(y) dσ(z) (12.5.73)
∂Ω ∂Ω |y − z|n
for every f ∈ Lipcomp (∂Ω), and for each compact set K ⊂ Rn , there exists a finite
positive constant C K such that

ExΩ f L2 (Ω∩K) ≤ C K f L2 (∂Ω) (12.5.74)
for every f ∈ Lipcomp (∂Ω).
In order to prove this claim suppose Ω is the upper-graph of some Lipschitz

function ϕ : Rn−1 → R and recall the bijective bi-Lipschitz flattening map Φ from
Remark 14.52. Pick f ∈ Lipcomp (∂Ω). Identifying Rn−1 × {0} canonically with Rn−1 ,
it follows that the function f ◦ Φ : Rn−1 × {0} → R belongs to Lipcomp (Rn−1 ). Then,
according to Step I, the function ExRn+ ( f ◦ Φ) belongs to C 0 (Rn+ ) ∩ C ∞ (Rn+ ), satisfies
the trace identity

ExRn+ ( f ◦ Φ) ∂Rn = f ◦ Φ on Rn−1 , (12.5.75)
+
the estimate

∇(ExRn+ ( f ◦ Φ))L2 (Rn ) (12.5.76)
+

|( f ◦ Φ)(y ) − ( f ◦ Φ)(z )|2 1/2
≤ Cn dy dz
Rn−1 Rn−1 |y − z |n
and, for each compact K ⊂ Rn , the estimate

ExRn+ ( f ◦ Φ)L2 (Rn ∩K) ≤ C K f ◦ ΦL2 (Rn−1 ) , (12.5.77)
+
where Cn , C K ∈ (0, ∞) are independent of f . Hence, if we further compose this

extension of f ◦ Φ to Rn+ with Φ−1 , we obtain the function ExRn+ ( f ◦ Φ) ◦ Φ−1
which belongs to C 0 (Ω) ∩ Liploc (Ω). This suggests defining the mapping
ExΩ : Lipcomp (∂Ω) → C 0 (Ω) ∩ Liploc (Ω)

(12.5.78)
ExΩ f := ExRn+ ( f ◦ Φ) ◦ Φ−1 , ∀ f ∈ Lipcomp (∂Ω).
To see why this map satisfies (12.5.72)–(12.5.74), fix some f ∈ Lipcomp (∂Ω). If
x ∈ Rn−1 , then (x , ϕ(x )) ∈ ∂Ω, we have Φ−1 (x , ϕ(x ) = (x , 0), and we may write

(ExΩ f )(x , ϕ(x )) = ExRn+ ( f ◦ Φ) Φ−1 (x , ϕ(x ))

= ExRn+ ( f ◦ Φ) (x , 0)
= ( f ◦ Φ)(x , 0) = f (x , ϕ(x )), (12.5.79)
where the third equality in (12.5.79) is just (12.5.75). This proves (12.5.72).
Moving on, for each index j ∈ {1, . . . , n} apply the standard Chain Rule (recall
that ExRn+ ( f ◦ Φ) is smooth in Rn+ and Φ is bijective and bi-Lipschitz, hence the
components of Φ and Φ−1 are differentiable a.e. by Rademacher’s theorem) to obtain
that

n

∂ j ExRn+ ( f ◦ Φ) ◦ Φ−1 = ∂k ExRn+ ( f ◦ Φ) ◦ Φ−1 · ∂ j Φ−1
k (12.5.80)
k=1
at a.e. point in Ω. On account of this, (12.5.78), and Proposition 2.121 (guaran-

teeing that the distributional derivatives of any locally Lipschitz function agree
with the pointwise partial derivatives) we therefore conclude that, for each index
j ∈ {1, . . . , n},

n

∂ j ExΩ f = ∂k ExRn+ ( f ◦ Φ) ◦ Φ−1 · ∂ j Φ−1
k in D (Ω). (12.5.81)
k=1
Hence, the distributional partial derivatives of ExΩ f in Ω are of function type and

∇ ExΩ f L2 (Ω) ≤ C ∇ ExRn+ ( f ◦ Φ) ◦ Φ−1 L2 (Ω)

≤ C ∇ ExRn+ ( f ◦ Φ) L2 (Rn )
+

| f (y , ϕ(y )) − f (z , ϕ(z ))|2 1/2

≤C dy dz
Rn−1 Rn−1 |y − z |n

| f (y) − f (z)|2 1/2
≤C dσ(y) dσ(z) , (12.5.82)
∂Ω ∂Ω |y − z|n
for some constant C ∈ (0, ∞) depending only on n and Φ. The first inequality in
(12.5.82) is a consequence of (12.5.80) and the fact that DΦ−1 has bounded com-
ponents. The second inequality is due to (12.3.138) and the corresponding norm
estimate, the third inequality follows from (12.5.76), while the last inequality is
based on (12.4.22) and (14.7.31). This shows that (12.5.73) is also satisfied.
In order to check (12.5.74), let K be a fixed compact subset of Rn . Then, since
−1
Φ is continuous, K := Φ−1 (K) is a compact set. Thus, we may write

ExΩ f L2 (Ω∩K) = ExRn+ ( f ◦ Φ) ◦ Φ−1 L2 (Ω∩K)

≤ C(Ω)ExRn+ ( f ◦ Φ)L2 (Rn ∩K)
+
Ω) f ◦ ΦL2 (Rn−1 ) ≤ C(K, Ω) f L2 (∂Ω) ,

≤ C(K, (12.5.83)
where the first inequality uses (12.3.140) (for v := [ExRn+ ( f ◦ Φ)]1K ), the second
in place of K), while the third inequality is
inequality is (12.5.77) (applied with K
obtained by applying (12.3.138) and the accompanying norm estimate. This shows
that (12.5.74) holds for the choice of the extension (12.5.78). This finishes the proof
of the claim in Step II.
Step III. The case when Ω is a bounded Lipschitz domain and the extension operator
acts on Lipschitz functions defined on ∂Ω. The goal here is to show that there exists
→ C (Ω)∩Lip
0 1
Ex : Lip(∂Ω) loc (Ω)∩H (Ω) linear map such
that (Ex f )∂Ω = f and Ex f H 1 (Ω) ≤ C f H 1/2 (∂Ω) for every
(12.5.84)
f ∈ Lip(∂Ω), where C is a constant depending only on the
set Ω and is independent of f .
Recall that in the present setting we have (cf. Theorem 12.30)
H 1 (Ω) = H 1 (Ω), with equivalent norms. (12.5.85)
Start by applying Remark 14.54 to obtain a family of cylinders {C x∗j }Nj=1 satisfying
(14.7.19), a family of upper-graph Lipschitz domains {Ω j }Nj=1 , and a partition of
unity {ψ j }Nj=0 as in (14.7.22). In view of (14.7.19) and Lemma 14.53, we have
Ω j ∩ C x∗j = Ω ∩ C x∗j , (12.5.86)
∂Ω j ∩ C x∗j = ∂Ω ∩ C x∗j , (12.5.87)
for each j ∈ {0, 1, . . . , N}. In addition, consider functions ξ j ∈ C0∞ (C x∗j ) with ξ j ≡ 1
near supp ψ j , for every j ∈ {0, 1, . . . , N}.

Fix f ∈ Lip(∂Ω). Then we may decompose f = Nj=1 ψ j f with each term satis-
fying
ψ j f ∈ Lip(∂Ω) and supp (ψ j f ) ⊆ ∂Ω ∩ C x∗j = ∂Ω j ∩ C x∗j . (12.5.88)
Pick j ∈ {1, . . . , N} and define

⎧
⎪
⎨ ψ j f in ∂Ω j ∩ C x∗j ,
⎪
f j := ⎪
⎪ (12.5.89)
⎩0 in ∂Ω j \ C x∗j .
Then f j ∈ Lipcomp (∂Ω j ). As such, we may apply Step II to construct an extension of

f j . Specifically, if ExΩ j denotes the extension operator from Step II associated with
the upper-graph Lipschitz domain Ω j , then what we have proved in Step II implies
ExΩ j f j ∈ C 0 (Ω j ) ∩ Liploc (Ω j ), hence

ξ j ExΩ j f j ) ∈ C 0 (Ω j ) ∩ Liploc (Ω j )

(12.5.90)
and supp ξ j ExΩ j f j ) ⊆ C x∗j .
Also, by Step II (cf. (12.5.72)–(12.5.73) with Ω replaced by Ω j and f replaced by

f j ) we have the trace property

ExΩ j f j ∂Ω = f j , (12.5.91)
j
and the estimate (with σ j denoting the surface measure on ∂Ω j )

| f j (y) − f j (z)|2 1/2
∇(ExΩ j f j )L2 (Ω ) ≤ C dσ j (y) dσ j (z)
j
∂Ω j ∂Ω j |y − z|n
≤ C f j H 1/2 (∂Ω j )
≤ Cψ j f H 1/2 (∂Ω) ≤ C f H 1/2 (∂Ω) , (12.5.92)
for some constant C = C(n, Ω) > 0. The first inequality in (12.5.92) is a rewriting of
(12.5.73) in the current setting, the second inequality is immediate from the defini-
tion of the H 1/2 -norm, the third is a consequence of Lemma 12.43, and the last one
follows from item (1) in Theorem 12.44.
Going further, use estimate (12.5.74) from Step II with Ω := Ω j , f := f j , and
K := supp ξ j to write

ξ j ExΩ j f j L2 (Ω ) = ξ j ExΩ j f j L2 (Ω ∩K)
j j

≤ ξ j L∞ (Rn ) ExΩ j f j L2 (Ω ∩K)
j
≤ C f j L2 (∂Ω j ) = Cψ j f L2 (∂Ω j ∩Cx∗ )

j
≤ C f L2 (∂Ω∩Cx∗ ) ≤ C f L2 (∂Ω) , (12.5.93)

j
for some finite constant C = C(n, Ω) > 0.

Let us now set
⎧
⎪
⎨ ξ j ExΩ j f j ) in Ω ∩ C x∗j = Ω j ∩ C x∗j ,
⎪
u j := ⎪
⎪ (12.5.94)
⎩0 in Ω \ C x∗j ,
and note that from (12.5.90) and (12.5.93) we have
u j ∈ C 0 (Ω) ∩ Liploc (Ω) and u j L2 (Ω) ≤ C f L2 (∂Ω) . (12.5.95)
Also, by the Chain Rule, (12.5.93) (applied with ξ j replaced by ∇ξ j ) and (12.5.92),
we obtain

∇u j L2 (Ω) = ∇ ξ j (ExΩ j f j ) L2 (Ω )
j

≤ (∇ξ j )(ExΩ j f j )L2 (Ω ) + ξ j ∇(ExΩ j f j )L2 (Ω )
j j

≤ C f H 1/2 (∂Ω) + ξ j L∞ (Rn ) ∇(ExΩ j f j )L2 (Ω )
j
≤ C f H 1/2 (∂Ω) , (12.5.96)

for some finite constant C = C(n, Ω) > 0. From (12.5.95), (12.5.96), and (12.5.85)
we conclude that there exists some C = C(n, Ω) ∈ (0, ∞) such that
u j ∈ H 1 (Ω) and u j H 1 (Ω) ≤ C f H 1/2 (∂Ω) . (12.5.97)
We are now ready to define the extension operator by setting

N
Ex f := u j. (12.5.98)
j=1
From (12.5.95), (12.5.97), and (12.5.98) it follows that
Ex f ∈ C 0 (Ω) ∩ Liploc (Ω) ∩ H 1 (Ω) and

(12.5.99)
Ex f H 1 (Ω) ≤ C f H 1/2 (∂Ω) .
Moreover, for each j ∈ {1, . . . , N} we have

u j ∂Ω∩C ∗ = ξ j ExΩ j f j ) ∂Ω ∩C ∗
x j x
j j

= ξ j ∂Ω ∩C ∗ ExΩ j f j ) ∂Ω ∩C ∗
j x j x
j j

= ξ j ∂Ω ∩C ∗ f j ∂Ω ∩C ∗
j x j x
j j

= ξ j ∂Ω ∩C ∗ (ψ j f )∂Ω ∩C ∗
j x j x
j j

= (ψ j f )∂Ω∩C ∗ , (12.5.100)
x
j
taking into account (12.5.94), (12.5.91), (12.5.88)–(12.5.89), and the fact that ξ j is
identically one on the support of ψ j . Hence, if tilde denotes the extension by zero
outside support (of functions defined on subsets of ∂Ω) to the entire set ∂Ω, from
(12.5.98) and (12.5.100) we obtain

N
N
Ex f ∂Ω = u j ∂Ω = u
j ∂Ω∩C ∗
x
j
j=1 j=1

N
N
= f )∂Ω∩C ∗ =
(ψ j ψ j f = f. (12.5.101)
x
j
j=1 j=1
Thus, the operator Ex defined in (12.5.98) satisfies all the properties listed in
(12.5.84). The proof of Step III is therefore complete.
Step IV. The proof of the existence of an extension operator as in the statement of
the theorem. Having proved the existence of the extension operator Ex as in Step III
(cf. (12.5.84)), since for bounded Lipschitz domains Ω in Rn we know that Lip(∂Ω)
is a dense subspace of H 1/2 (∂Ω) (cf. item (2) in Theorem 12.44) and H 1 (Ω) is a
Banach space (cf. Theorem 12.14) it follows that the operator Ex from (12.5.84) has
a unique extension to a linear and bounded operator
: H 1/2 (∂Ω) → H 1 (Ω).
Ex (12.5.102)
In addition, if f ∈ Lip(∂Ω) then

f = Ex f ∈ C 0 (Ω) ∩ Liploc (Ω) ∩ H 1 (Ω) = V(Ω),
Ex (12.5.103)
(where the last equality is a definition; cf. item (1) in Theorem 12.45), hence if Tr
denotes the trace operator from Theorem 12.45 we have

Tr ◦ Ex f = TrEx f = Ex f = f,
f = TrEx (12.5.104)
∂Ω
where for the third equality in (12.5.104) we used the fact that Tr acts as the point-
wise restriction to ∂Ω for functions in the space V(Ω). This shows that the oper-
acting from H 1/2 (∂Ω) into H 1/2 (∂Ω) is well-defined, linear, bounded,
ator Tr ◦ Ex
and coincides with I, the identity operator, on Lip(∂Ω). By density, it follows that
= I on H 1/2 (∂Ω). The proof of Theorem 12.46 is now finished.
Tr ◦ Ex
Further Notes for Chapter 12. Sobolev spaces constitute a classical topic at the confluence of
harmonic analysis, functional analysis, and partial differential equations and there is a vast litera-
ture dedicated to this subject. In this regard, the interested reader may consult [2], [43]. See also
[15], [47], [48], [49], for more specialized, in-depth treatments.

,
Exercise 12.47. From Theorem 12.1 we know that S(Rn ) ⊂ s>0 H s (Rn ). Prove
'
that this is a strict inclusion by considering the function u(x) := nj=1 f (x j ) for all
x = (x1 , . . . , xn ) ∈ Rn , where
⎧
⎪
⎪
⎪ sin ξ
⎪
⎨ , if ξ 0,
f (ξ) := ⎪ ξ for all ξ ∈ R. (12.6.1)
⎪
⎪
⎪
⎩ 1, if ξ = 0,

Exercise 12.48. By Theorem 12.7 we know that E (Rn ) ⊂ s<0 H s (Rn ). Prove that
this is a strict inclusion.
Exercise 12.49. Show that for every ε > 0 we have δ ∈ H −n/2−ε (Rn ) while, at the
same time, δ H −n/2 (Rn ).
Exercise 12.50. Suppose u ∈ H −n/2 (Rn )∩E (Rn ) and supp u ⊂ {a1 , . . . , aN } for some
distinct points a1 , . . . , aN ∈ Rn . Show that u ≡ 0.
Exercise 12.51. Let t ∈ R and ϕ ∈ S(Rn ) and define

−n
Bt ϕ(x) := (2π) eix·ξ ξ−t
ϕ(ξ) dξ, ∀ x ∈ Rn . (12.6.2)
Rn
Prove the following statements.

(1) Bt ϕ ∈ S(Rn ) for each t ∈ R and ϕ ∈ S(Rn ).
(2) Bt ϕH s+t (Rn ) = ϕH s (Rn ) for each s, t ∈ R and ϕ ∈ S(Rn ).
(3) Bt ◦ Br = Bt+r for all t, r ∈ R.
(4) B0 ϕ = ϕ for all ϕ ∈ S(Rn ).
(5) For each s, t ∈ R the mapping Bt extends uniquely as a linear and continuous
map from H s (Rn ) into H s+t (Rn ).
Exercise 12.52. Consider a nonzero function θ ∈ C0∞ (R) with supp θ ⊆ [−1, 1] and
define
sin( jx)
u j (x) := θ(x) , ∀ x ∈ R, ∀ j ∈ N. (12.6.3)
j
Prove that the sequence {u j } j∈N is bounded in H 1 (R), it is convergent in H s (R) pro-
vided s < 1, but it has no convergent subsequence in H 1 (R).
Exercise 12.53. Consider a function θ ∈ C0∞ (R) with supp θ ⊆ [0, 1] and define
$ cos( jx) %
u j (x) := θ(x) 1 + (−1) j + , ∀ x ∈ R, ∀ j ∈ N. (12.6.4)
j
Prove that the sequence {u j } j∈N has a convergent subsequence in H 1/2 (R).
Exercise 12.54. Let s1 , s2 ∈ R be such that s1 < s2 . Prove that the embedding
H s2 (Rn ) → H s1 (Rn ) is strict.
,
Exercise 12.55. Prove that the inclusion s∈R H s (Rn ) ⊂ C ∞ (Rn ) (cf. (12.1.63)) is
strict.

Exercise 12.56. Prove that |x| ∈ H 1 (−1, 1) and |x| H 2 (−1, 1) .
Exercise 12.57. Let Ω be a bounded open set in Rn . Prove that Lip(Ω) ⊂ H 1 (Ω).

Exercise 12.58. Show that the membership |x|−λ ∈ H 1 B(0, 1) , when considered in
Rn , holds for all λ < 1 if and only if n ≥ 4.
Chapter 13
Solutions to Selected Exercises
Abstract In this chapter, solutions to various problems listed at the end of each of
the Chapters 1–12 are provided.
13.1 Solutions to Exercises from Section 1.4
Exercise 1.27 a = 0; see Examples 1.4 and 1.5.

Exercise 1.28 No; see Examples 1.4 and 1.5.
Exercise 1.29 Let ϕ ∈ C0∞ (R2 ). Based on integration by parts and the fact that ϕ is
compactly supported we have

∂1 ∂2 f, ϕ = f (x, y)(∂1 ∂2 ϕ)(x, y) dx dy
R2
∞ ∞
= ∂2 (∂1 ϕ)(x, y) dy dx + ∂1 (∂2 ϕ)(x, y) dx dy = 0. (13.1.1)
0 R 0 R
This shows that ∂1 ∂2 f = 0 in the weak sense. Similarly, the other weak derivatives
are ∂2 (∂1 ∂2 f ) = 0 and ∂21 ∂2 f = 0. Suppose now that ∂1 f exists in the weak sense,
i.e., there exists g ∈ Lloc1
(R2 ) such that

f (x, y)(∂1 ϕ)(x, y) dx dy = − g(x, y)ϕ(x, y) dx dy, (13.1.2)
R2 R2

for all ϕ ∈ C0∞ (R2 ). Since R2 f (x, y)(∂1 ϕ)(x, y) dx dy = − R ϕ(0, y) dy (as seen using
the definition of f and integration by parts), (13.1.2) becomes

ϕ(0, y) dy = g(x, y)ϕ(x, y) dx dy, ∀ ϕ ∈ C0∞ (R2 ). (13.1.3)
R R2

498 13 Solutions to Selected Exercises
Taking now ϕ in (13.1.3) to be supported in {(x, y) ∈ R2 : x 0} and recalling

Theorem 1.3, it follows that g = 0 almost everywhere on {(x, y) ∈ R2 : x 0}.
Thus,
g = 0 almost everywhere in R2 . The latter, when used in (13.1.3), implies
R
ϕ(0, y) dy = 0 for every ϕ ∈ C0∞ (R2 ). However this leads to a contradiction by
taking ϕ(x, y) = ϕ1 (x)ϕ2 (y) for (x, y) ∈ R2 , where ϕ1 , ϕ2 ∈ C0∞ (R) are such that
ϕ1 (0) = 1 and R ϕ2 (y) dy = 1.

Exercise 1.30 Let ϕ ∈ C0∞ (−1, 1) be arbitrary. Then Lebesgue’s Dominated Con-
vergence Theorem and integration by parts give

1

−ε √ 1 √
− f (x)ϕ (x) dx = lim+ −x ϕ (x) dx − lim+ x ϕ (x) dx
−1 ε→0 −1 ε→0 ε

√ −ε
1
= lim+ −x ϕ(x)|−ε
−1 + √ ϕ(x) dx
ε→0 −1 2 −x

√ 1
1
− lim+ xϕ(x)|ε−1 − √ ϕ(x) dx
ε→0 ε 2 x
1
1
=
√ ϕ(x) dx, (13.1.4)
−1 2 |x|

where we have used the fact that 2 √1|x| ∈ L1 (−1, 1) . Hence the weak derivative of
order one of f is the function given by g(x) := √1
2 |x|
for x ∈ (−1, 1) \ {0}.
Exercise 1.31 Consider a function ϕ ∈ C0∞ (R2 ) and pick a number R > 0 such that
supp ϕ ⊂ (−R, R) × (−R, R). Then

− x|y|(∂21 ∂2 ϕ)(x, y) dx dy = − |y| x(∂21 ∂2 ϕ)(x, y) dx dy = 0, (13.1.5)
R2 R R
where for the last equality is obtained integrating by parts twice with respect to x,
keeping in mind that ϕ has compact support. This proves that ∂21 ∂2 f = 0 in the weak
sense. On the other hand,

− x|y|(∂1 ∂22 ϕ)(x, y) dx dy = |y|(∂22 ϕ)(x, y) dx dy
R2 R2
0 R
= − y(∂22 ϕ)(x, y) dy + y(∂22 ϕ)(x, y) dy dx
R −R 0

=− sgn(y)(∂2 ϕ)(x, y) dy dx = 2 ϕ(x, 0) dx (13.1.6)
R2 R
after integrating by parts, first with respect to x, then twice with respect to y. Now
1
reasoning as in the proof of Exercise 1.29 one can show that if g belongs to Lloc (R2 )
is such that
13.1 Solutions to Exercises from Section 1.4 499

2 ϕ(x, 0) dx = g(x, y)ϕ(x, y) dx dy, ∀ ϕ ∈ C0∞ (R2 ),
R R2
leads to a contradiction. Hence, ∂1 ∂22 f does not exist in the weak sense.
Exercise 1.32 Reasoning as in the proofs of Exercise 1.29 and Exercise 1.31, one
can show that ∂k1 f and ∂k2 f do not exist in the weak sense for any k ∈ N, and that
∂α f = 0 in the weak sense whenever α = (α1 , α2 ) ∈ N20 with α1 0 and α2 0.
Exercise 1.33 Let ϕ ∈ C ∞ (R) with supp ϕ ⊂ (−R, R) for some R ∈ (0, ∞). Then,
0 R
− f (x)ϕ (x) dx = − sin(−x)ϕ (x) dx − sin(x)ϕ (x) dx
R −R 0
0 R
=− cos(x)ϕ(x) dx + cos(x)ϕ(x) dx
−R 0

=− sgn(x) cos(x)ϕ(x) dx. (13.1.7)
R
Since sgn(x) cos(x) ∈ Lloc

1
(R), we deduce from (13.1.7) that f exists in the weak

sense and that f (x) = sgn(x) cos(x) for x ∈ R. Also,
0 R

f (x)ϕ (x) dx = cos(x)ϕ (x) dx − cos(x)ϕ (x) dx
R −R 0

= 2ϕ(0) − sgn(x) sin(x)ϕ(x) dx. (13.1.8)
R
If the weak derivative f were to exist we could find g ∈ Lloc

1
(R) such that

g(x)ϕ(x) dx = 2ϕ(0) − sgn(x) sin(x)ϕ(x) dx, ∀ ϕ ∈ C0∞ (R). (13.1.9)
R R
In particular,(13.1.9) would also hold for all ϕ ∈ C0∞ (R) supported

in (0, ∞), which
would give g(0,∞) = − sin(x). Similarly, we would conclude that g(−∞,0) = sin(x) and,
hence, returning to (13.1.9), we would obtain ϕ(0) = 0 for all ϕ ∈ C0∞ (R). The latter
is false, thus the weak derivative f does not exist.
Exercise 1.35 Suppose ∂ j f , j = 1, . . . , n, exist in the weak sense. Fix j and note
that by Exercise 1.34 we have that the weak derivative (∂ j f )Rn \{0} equals the weak

derivative ∂ f
j . The function f
Rn \{0}
= 1 is smooth in Rn \ {0} so the latter
Rn \{0} |x|a
x
weak derivative equals the classical pointwise derivative which is −a |x|a+2 j
for each
x ∈ R \ {0}. Since the Lebesgue measure of the set {0} is zero, we obtain that the
n
xj
weak derivative ∂ j f (x) = −a |x|a+2 for almost every x ∈ Rn . However, by assumption,
this weak derivative is in Lloc (R ) and since j is arbitrary, we obtain that |x|aa+1 =
1 n
1/2
n
j=1 (∂ j f )
2
∈ Lloc
1
(Rn ). This necessarily implies a < n − 1.
x
Conversely, suppose a < n − 1 and consider the functions g j (x) := (−a) |x|a+2 j
defined for almost every x ∈ Rn and each j ∈ {1, . . . , n}. Then g j ∈ Lloc
1
(Rn ) for each
j ∈ {1, . . . , n}. To complete the proof it suffices to show that

1 xj
(∂ j ϕ)(x) dx = a ϕ(x) dx (13.1.10)
Rn |x| Rn |x|
a a+2
for every ϕ ∈ C0∞ (Rn ) and each j ∈ {1, . . . , n}. For ϕ and j arbitrary, fixed, using
Lebesgue’s Dominated Convergence Theorem and integration by parts (cf. formula
(14.8.4)) we write

1 1
(∂ j ϕ)(x) dx = lim (∂ j ϕ)(x) dx
R n |x|a ε→0 +
R \B(0,ε)
n |x|a

xj x j ϕ(x)
= lim+ a ϕ(x) dx − lim dσ(x)
Rn \B(0,ε) |x| |x|=ε |x|
ε→0 a+2 ε→0 + a+1

xj
=a ϕ(x) dx. (13.1.11)
Rn |x|
a+2
For the last equality in (13.1.11) we have used the fact that
x j ϕ(x)
dσ(x) ≤ ϕL∞ (Rn ωn−1 εn−1−a → 0 as ε → 0+ (13.1.12)
|x|=ε |x|
a+1 n
since n − 1 − a > 0. This proves (13.1.10) as wanted.

Exercise 1.36 Let ϕ ∈ C0∞ (Ω) be arbitrary. Then ∂α+β ϕ = ∂α (∂β ϕ) and since the
weak derivative ∂α f exists we have

f ∂α+β ϕ dx = (−1)|α| (∂α f )(∂β ϕ) dx. (13.1.13)
Ω Ω
Also, since the weak derivative ∂β (∂α f ) exists we have

α β |β|
(∂ f )(∂ ϕ) dx = (−1) ∂β (∂α f )ϕ dx. (13.1.14)
Ω Ω
The desired conclusion follows from (13.1.13), (13.1.14), and the definition of weak
derivative.
Exercise 1.37 Note that f ∈ Lloc 1
(Rn ) for every n ∈ N. Suppose n ≥ 2, and fix
some index j ∈ {1, . . . , n} along with a function ϕ ∈ C0∞ (Rn ). If R > 0 is such
that supp ϕ ⊂ B(0, R), then using Lebesgue’s Dominated Convergence Theorem and
(14.8.4) we obtain

− f (x)∂ j ϕ(x) dx = − lim+ f (x)∂ j ϕ(x) dx
Rn r→0 r<|x|<R
xj
= lim+ ε+1
ϕ(x) dσ(x)
r→0 ∂B(0,r) r

xj
−ε ϕ(x) dx
r<|x|<R |x|ε+2

xj
= −ε ϕ(x) dx, (13.1.15)
Rn |x|ε+2
since, given that under the current assumptions n − 1 − ε > 0, we have

xj
ϕ(x) dσ(x) ≤ ωn−1 ϕL∞ (Rn ) rn−1−ε → 0 as r → 0+ . (13.1.16)
∂B(0,r) r ε+1
This proves that if n ≥ 2 then the weak derivative ∂ j f exists and is equal to the
xj
function −ε |x|ε+2 ∈ Lloc
1
(Rn ).
Assume nextthat n = 1 and suppose that there exists some g ∈ Lloc (R) with the
1
property that − R f ϕ dx = R gϕ dx for every ϕ ∈ C0∞ (R). Then, with R > 0 such
that supp ϕ ⊂ (−R, R), using Lebesgue’s Dominated Convergence Theorem we may
write
−r R
ϕ (x) ϕ (x)
f ϕ dx = lim+ ε
dx + dx (13.1.17)
R r→0 −R (−x) r xε
−r R
ϕ(−r) ϕ(x) ϕ(r) ϕ(x)
= lim+ −ε dx − ε + ε dx .
r→0 rε −R (−x)
ε+1 r r x
ε+1
In particular, identity (13.1.17) holds if supp ϕ ⊂ (0, ∞), in which case (13.1.17)
implies g(x) = εx−ε−1 for x > 0, and if supp ϕ ⊂ (−∞, 0), when we obtain g(x) =
−ε(−x)−ε−1 for x < 0 (recall Theorem 1.3). However, the function g thus obtained
1
does not belong to Lloc (R). Consequently, f does not have a weak derivative of order
one in the case when n = 1.
1
Exercise 1.38 (a) Let f ∈ Lloc (a, b) be such that the weak derivative f exists and
equals 0. That is,
b

f (x)ϕ (x) dx = 0, ∀ ϕ ∈ C0∞ (a, b) . (13.1.18)
a
b
Select ϕ0 ∈ C0∞ (a, b) such that a ϕ0 (s) ds = 1. Then, each ϕ ∈ C0∞ (a, b) may be
written as
b
ϕ(t) = ϕ0 (t) ϕ(s) ds + ψ (t), ∀ t ∈ (a, b), (13.1.19)
a
where
x b
ψ(x) := ϕ(t) − ϕ0 (t) ϕ(s) ds dt, ∀ x ∈ (a, b). (13.1.20)
a a

An inspection of (13.1.20) reveals that ψ ∈ C ∞ (a, b) , and if [a0 , b0 ] ⊆ (a, b) is an
interval such that supp ϕ, supp ϕ0 ⊆ [a0 , b0 ], then we also have supp ψ ⊆ [a0 , b0 ] (as
seen by analyzing the cases x ∈ (a, a0 ) and x ∈ (b0 , b)). Now define the constant
b
c := a f (x)ϕ0 (x) dx and use (13.1.18) to obtain

b b

f (x)ϕ(x) dx = c ϕ(s) ds, ∀ ϕ ∈ C0∞ (a, b) . (13.1.21)
a a
Hence, by invoking Theorem 1.3 we obtain f = c almost everywhere in (a, b).
x
1
(b) Let g ∈ Lloc (a, b) , x0 ∈ (a, b), and set f (x) := x g(t) dt for every x ∈ (a, b).
0
By Lebesgue’s Dominated Convergence Theorem, it follows that f is continuous

on (a, b), hence f ∈ Lloc
1
(a, b) . There remains to prove that g is the weak derivative
of f .
Parenthetically, we note that under the current assumptions, one may not expect f
to necessarily be pointwise differentiable in (a, b). As an example, take the function
f : R → R defined by f (x) := x for x ≥ 0 and f (x) := 0 for x < 0. Then f is not
differentiable at zero but its weak derivative f exists and is equal to χ(0,∞) ∈ Lloc
1
(R).
Also, observe that the fundamental theorem of calculus does not apply in this case
1
since while g is in Lloc (R) it is not continuous.
∞
Let ϕ ∈ C0 (a, b) and c ∈ (a, b) such that supp ϕ ⊂ (c, b). Consider two cases.
Case 1. Assume c > x0 . Then supp ϕ ⊂ (x0 , b) and using the expression for f and
Fubini’s theorem we may write
b b x b b
f (x)ϕ (x) dx = g(t) dt ϕ (x) dx = ϕ (x) dx g(t) dt
a x0 x0 x0 t
b b
=− ϕ(t)g(t) dt = − g(t)ϕ(t) dt. (13.1.22)
x0 a
Case 2. Assume c < x0 . Again, using the expression for f and Fubini’s theorem
it follows that
b x0 b
f (x)ϕ (x) dx = f (x)ϕ (x) dx + f (x)ϕ (x) dx
a a x0
x0 x b x
= g(t) dt ϕ (x) dx + g(t) dt ϕ (x) dx
a x0 x0 x0
x0 t b b
=− g(t) ϕ (x) dx dt + g(t) ϕ (x) dx dt
a a x0 t
b
=− g(t)ϕ(t) dt. (13.1.23)
a
From (13.1.22) and (13.1.23) the desired conclusion follows.

1
(c) Denote by h ∈ Lloc (a, b) the weak derivative f (k) and fix an arbitrary point
x
x0 ∈ (a, b). If one sets g(x) := x h(s) ds for x ∈ (a, b), then by Lebesgue’s Dom-
0
inated Convergence Theorem, it follows that g is continuous. Thus, in particular,
b
1
g ∈ Lloc (a, b) . Fix ϕ0 ∈ C0∞ (a, b) with a ϕ0 (t) dt = 1. Using induction and what
we have proved in part (b), the desired conclusion will follow if we show that
b b
f (k−1) = g − g(t)ϕ0 (t) dt + (−1)k−1 f (t)ϕ(k−1)
0 (t) dt a.e. on (a, b). (13.1.24)
a a

Let ϕ ∈ C0∞ (a, b) and write it as in (13.1.19)–(13.1.20). Then
b b
b
b
g(x)ϕ(x) dx = g(x)ϕ0 (x) dx ϕ(t) dt + g(x)ψ (x) dx
a a a a
b
b
b
= g(x)ϕ0 (x) dx h(x)ψ(x) dx ϕ(t) dt −
a a a
b
b
b
= g(x)ϕ0 (x) dx ϕ(t) dt + (−1)k−1 f (x)ψ(k) (x) dx
a a a
b
b
b
= g(x)ϕ0 (x) dx ϕ(t) dt + (−1)k−1 f (x)ϕ(k−1) (x) dx
a a a
b
b
+ (−1) k
f (x)ϕ0(k−1) (x) dx ϕ(t) dt . (13.1.25)
a a
Now (13.1.24) follows from (13.1.25) and Theorem 1.3.

(d) Using what we have proved in part (c), we obtain that the weak derivative f ( j)
exists for each j ∈ N satisfying j ≤ k. From what we proved in part (a) we know
that f (k−1) = ak−1 ∈ C, while from part (c) we know that f (k−2) = ak−1 x − ak−1 x0 − c.
The proof may be now completed by induction.
(e) No; see Exercise 1.29.
Exercise 1.39 Let R > 0 be such that supp θ ⊆ B(0, R). Then ϕ j ∈ C0∞ (Rn ) and
supp ϕ j ⊆ B(0, R) for each j ∈ N. Also, given any α ∈ Nn0 we have
∂α ϕ j = e− j jm+|α| (∂α θ)( j·) for every j ∈ N. (13.1.26)
Hence,
sup |∂α ϕ j (x)| ≤ e− j jm+|α| θL∞ (Rn ) −−−−→ 0.
j→∞
x∈B(0,R)
D(Rn )
This shows that ϕ j −−−−→ 0.
j→∞
Exercise 1.40 Clearly, for every j ∈ N, ϕ j ∈ C0∞ (Rn ) and supp ϕ j = supp θ − jh.
Hence, for every R > 0 there exists j0 ∈ N such that
(supp θ − jh) ∩ B(0, R) = ∅ for j ≥ j0 . (13.1.27)
This shows that there is no compact set K ⊂ Rn such that supp ϕ j ⊆ K for all j ∈ N,
which implies that {ϕ j } j∈N does not converge in D(Rn ). In addition, if α ∈ Nn0 , K is
a compact set in Rn , and R > 0 is such that K ⊆ B(0, R), then sup x∈K |∂α ϕ j (x)| =
sup x∈K |(∂α θ)(x + jh)| = 0 for all j ≥ j0 , where j0 ∈ N is such that (supp θ − jh) ∩
E(Rn )
B(0, R) = ∅. This proves that ϕ j −−−−→ 0.
j→∞
E(Rn )
Exercise 1.41 Suppose there exists ϕ ∈ C0∞ (Rn ) such that ϕ j −−−−→ ϕ. Then nec-
j→∞
essarily ϕ j −−−−→ ϕ pointwise (explain why). Since lim ϕ j (x) = 0 for every point
j→∞ j→∞
x ∈ Rn , this forces ϕ = 0. However, if α ∈ N0 is such that |α| = 1, then
we have ∂α ϕ j = (∂α θ)( j·) for every j ∈ N, hence for each compact K we have
sup x∈K |∂α ϕ j | = ∂α θL∞ (K) which does not converge to zero as j → ∞, leading to a
contradiction.
Exercise 1.42 One implication is easy. If θ 0, then supp ϕ j = jh + 1j supp θ for
each j ∈ N, and since h 0, there is no compact K ⊂ Rn such that supp ϕ j ⊆ K for
all j ∈ N. This shows that {ϕ j } j∈N does not converge in D(Rn ).
Exercise 1.43 {ϕ j } j∈N does not converges in D(Rn ) since supp ϕ j = j supp θ for
every j ∈ N, thus there is no compact K ⊂ Rn such that supp ϕ j ⊆ K for all j ∈ N.
On the other hand, if α ∈ Nn0 then |∂α ϕ j (x)| ≤ j−1−|α| ∂α θL∞ (Rn ) for every x ∈ Rn .
E(Rn )
Thus, ∂α ϕ j converges to zero uniformly on Rn , proving that ϕ j −−−−→ 0.
j→∞
Exercise 1.44 Let x0 ∈ Ω and R > 0 be such that B(x0 , R) ⊂ Ω. Consider a nonzero
function ϕ ∈ C0∞ (R) with the property that supp ϕ ⊂ B(0, R) and define the sequence

ϕ j (x) := 1j ϕ j(x − x0 ) , x ∈ Ω, j ∈ N. Show that the sequence {ϕ j } j∈N satisfies the
hypotheses in the problem but it does not converge in D(Rn ).
Exercise 2.122 Let K be a compact set in R and fix an arbitrary ϕ ∈ C0∞ (R) with

supp ϕ ⊆ K. By the Mean Value Theorem, |ϕ j12 − ϕ(0)| ≤ j12 ϕ L∞ (K) . This shows
that the series in (2.10.1) is absolutely convergent, hence u is well defined. Clearly,
∞
u is linear and since |u(ϕ)| ≤ j−2 sup |ϕ (x)| it follows that u ∈ D (R), it has
j=1 x∈K
finite order, and its order is at most 1. To see that u is not of order zero, consider the
sequence of functions {ϕk }k∈N satisfying, for each k ∈ N, the conditions

ϕk ∈ C0∞ (0, 2) , 0 ≤ ϕk ≤ 1k ,
(13.2.1)
ϕk (x) = 0 if x ≤ (k+1)
1
2 , ϕk (x) = k if x ∈ k2 , 1 .
1 1

k 1
Then u, ϕk = ϕk j2
= 1. If we assume that u has order zero, then there exists a
j=1
finite positive number C such that |u, ϕ| ≤ C sup x∈[0,2] |ϕ(x)| for every ϕ ∈ C0∞ (R)
with supp ϕ ⊆ [0, 2]. This implies 1 = |u, ϕk | ≤ Ck for every k ∈ N, which leads to
a contradiction. Hence the order of u is 1.
Exercise 2.123 Take u to be the distribution given by the constant function 1. Then
this u does not satisfy (2.10.2). Indeed,
for any compact K ⊂ Ω we can choose
ϕ ∈ C0∞ (Ω) with supp ϕ ⊆ Ω\ K and Ω ϕ dx 0 which would lead to a contradiction
if (2.10.2) were true for u = 1.
Exercise 2.124 Use a reasoning similar in spirit to the one from Example 2.9.

Exercise 2.125 Estimate B(0,R) | f (x)| dx by working in polar coordinates (cf. (14.9.7))
and using Proposition 14.65, as well as (2.1.9).
Exercise 2.126 Let ϕ ∈ C0∞ (R) with supp ϕ ⊂ (−R, R). Then using integration by
parts we have

(ln |x|) , ϕ = − ln |x|, ϕ

=− ϕ (x) ln |x| dx = − lim+ ϕ (x) ln |x| dx
R ε→0 ε<|x|<R
ϕ(x) −ε R

= lim+ dx − ϕ(x) ln |x| − ϕ(x) ln |x|
ε→0 ε<|x|<R x −R ε
1
= P.V. , ϕ + lim+ ϕ(ε) − ϕ(−ε) ln ε
x ε→0
1
= P.V. , ϕ . (13.2.2)
x
For the last equality in (13.2.2) note that

ϕ(ε) − ϕ(−ε) ln ε ≤ 2ϕ L∞ (R) ε| ln ε| −−−−→ 0. (13.2.3)
ε→0+
Exercise 2.127 It is immediate that f is continuous on R \ {0}, while its continuity

at x = 0 follows from the fact that
lim x ln |x| = 0. (13.2.4)

x→0
Fix ϕ ∈ C0∞ (R). Starting with the definition of distributional derivative, then apply-
ing Lebesgue’s Dominated Convergence Theorem (in concert with the properties of
ϕ), then integrating by parts and using (13.2.4), we obtain

f , ϕ = − f, ϕ
−ε ∞

= − lim+ (x ln(−x) − x)ϕ (x) dx + (x ln x − x)ϕ (x) dx
ε→0 −∞ ε
−ε ∞
= lim+ ln(−x)ϕ(x) dx + (ln x)ϕ(x) dx
ε→0 −∞ ε

= ln |x|, ϕ . (13.2.5)
The last equality in (13.2.5) uses the fact that ln |x| ∈ Lloc
1
(R).
Alternatively, using (4) in Proposition 2.43, Exercise 2.126, the fact that x = 1
in D (R), and (2.3.7), we have
1
(x ln |x| − x) = ln |x| − x P.V. − 1 = ln |x| in D (R). (13.2.6)

x
Exercise 2.128 Fix j ∈ {1, . . . , n} and ϕ ∈ C0∞ (Rn ). Then

∂ j (ln |x|), ϕ = − ln |x|, ∂ j ϕ = − (ln |x|)∂ j ϕ(x) dx
Rn

= − lim+ (ln |x|)∂ j ϕ(x) dx
ε→0 |x|≥ε

xj xj
= lim+ ϕ(x) dx + lim+ (ln ε) ϕ(x) dσ(x)
ε→0 |x|≥ε |x|2 ε→0 |x|=ε ε

xj
= ϕ(x) dx + lim+ εn−1 (ln ε) ω j ϕ(εω) dσ(ω) (13.2.7)
Rn |x|2 ε→0 S n−1
For the third and last equality in (13.2.7) we used Lebesgue’s Dominated Con-
x
vergence Theorem (note that |x|j2 , ln |x| ∈ Lloc
1
(Rn ) and ϕ is compactly supported).
The fourth equality uses the integration by parts formula (14.8.4). Also, for the last
equality, in the integral on ∂B(0, ε) we made the change of variables x = εω with
ω ∈ S n−1 . One more application of Lebesgue’s Dominated Convergence Theorem
yields

lim ω j ϕ(εω) dσ(ω) = ϕ(0) ω j dσ(ω) = 0, (13.2.8)
ε→0+ S n−1 S n−1
where the last equality is due to the fact that the integral of an odd function over
the unit sphere is zero. Moreover, (2.1.9) implies lim+ εn−1 (ln ε) = 0 (recall that we
ε→0
are assuming n ≥ 2). Returning with all these comments to (13.2.7) we arrive at the
conclusion that
xj
∂ j (ln |x|), ϕ = , ϕ for every ϕ ∈ C0∞ (Rn ).
|x|2
xj
Hence, if n ≥ 2 then ∂ j (ln |x|) = |x|2
in D (Rn ).
Exercise 2.129 Fix j ∈ {1, . . . , n} and ϕ ∈ C0∞ (Rn ). Then
1
1
1
∂ j n−2 , ϕ = − n−2 , ∂ j ϕ = − ∂ ϕ(x) dx
n−2 j
|x| |x| R n |x|

1
= − lim+ ∂ ϕ(x) dx
n−2 j
ε→0 |x|≥ε |x|

(2 − n)x j xj
= lim+ ϕ(x) dx + lim ϕ(x) dσ(x)
ε→0 |x|≥ε |x| n ε→0 +
|x|=ε ε
n−1

xj
= (2 − n) ϕ(x) dx + lim+ ω j ϕ(εω) dσ(ω) (13.2.9)
Rn |x|
n ε→0 S n−1
For the third and last equality in (13.2.7) we used Lebesgue’s Dominated Con-
x
vergence Theorem (note that |x|jn , |x|1n−2 ∈ Lloc
1
(Rn ) and ϕ is compactly supported).
The fourth equality uses the integration by parts formula (14.8.4). Also, for the last
equality, in the integral on ∂B(0, ε) we made the change of variables x = εω with
x ∈ S n−1 . One more application of Lebesgue’s Dominated Convergence Theorem
yields

lim+ ω j ϕ(εω) dσ(ω) = ϕ(0) ω j dσ(ω) = 0, (13.2.10)
ε→0 S n−1 S n−1
where the last equality is due to the fact that the integral of an odd function over the
unit sphere is zero. Returning with all these comments to (13.2.9) we arrive at the
conclusion that
1 xj
∂ j n−2 , ϕ = (2 − n) n , ϕ for every ϕ ∈ C0∞ (Rn ).
|x| |x|
x
Hence, ∂ j |x|1n−2 = (2 − n) |x|jn in D (Rn ).
Exercise 2.130 Fix j ∈ N. Using the change of variables y = jx − jt the expression
for ψ j becomes
jx−1
ψ j (x) = θ(y) dy for x ∈ R.
jx− j2
1
Hence ψ j (x) = 0 if x ≤ 0, while for x > 0 we have ψ j (x) = −1 θ(y) dy = 1 if
j ≥ j0 , where j0 ∈ N is such that j0 x − j20 ≤ −1 and j0 x − 1 ≥ 1. This shows that
1
ψ j converges pointwise to H as j → ∞. In addition, |ψ j (x)| ≤ −1 |θ(y)| dy < ∞ for
every j ∈ N. Hence, by applying Lebesgue’s Dominated Convergence Theorem,

lim ψ j (x)ϕ(x) dx = H(x)ϕ(x) dx for each ϕ ∈ C0∞ (R).
j→∞ R R
D (R)
This proves that ψ j −−−−→ H as wanted.
j→∞
Exercise 2.131 If ϕ ∈ C0∞ (R) then the support condition for ϕ guarantees that

∞
only finitely many terms in the sum ϕ( j) ( j) are nonzero, hence u is well defined.
j=1
Clearly u is linear. If K is a compact in R and k ∈ N is such that K ⊆ [−k, k], then

k
|u(ϕ)| ≤ ϕ( j) ( j) ≤ k sup ϕ( j) (x)
j=1 x∈K, j≤k
for ϕ ∈ C0∞ (R) with supp ϕ ⊆ K. This proves that u ∈ D (R).

Suppose the distribution u has finite order. Then there exists a nonnegative integer
k0 with the property that for each compact set K ⊂ R there is a finite constant C K ≥ 0
such that |u, ϕ| ≤ C K sup |ϕ( j) (x)| for every ϕ ∈ C0∞ (R) with supp ϕ ⊆ K. In
x∈K, j≤k0
particular, from this and the definition of u it follows that there exists C ∈ [0, ∞) such

that whenever ϕ ∈ C0∞ (R) satisfies the support condition supp ϕ ⊆ k0 + 12 , k0 + 32
we have
|ϕ(k0 +1) (k0 + 1)| = |u, ϕ| ≤ C sup |ϕ( ) (x)|. (13.2.11)
x∈ k0 + 12 ,k0 + 13 , ≤k0

Now consider θ ∈ C0∞ (−1/2, 1/2) satisfying θ(0) = 1 and construct the
sequence of smooth functions ϕ j (x) := θ( jx − j(k0 + 1)) for x ∈ R and j ∈ N.
Then, for each j ∈ N, we have
1 3
supp ϕ j ⊆ k0 + , k0 + ,
2 2
( )
ϕ( )
j = j θ ( j · − j(k0 + 1)), ∀ ∈ N0 ,
and ϕ(kj 0 +1) (k0 + 1) = jk0 +1 .
Combining all these facts with (13.2.11) it follows that for each j ∈ N,
jk0 +1 ≤ C sup j |θ( ) (x)|

1 1
x∈ − 2 , 2 , ≤k0

≤ C max θ( ) L∞ ([−1/2,1/2]) : ≤ k0 jk0 . (13.2.12)
Choosing now j sufficiently large in (13.2.12) leads to a contradiction. Hence u does

not have finite order.
Exercise 2.132 Note that for each j ∈ N we have f j ∈ Lloc 1
(Rn ). Pick some ϕ ∈
∞
C0 (R ) and suppose R > 0 is such that supp ϕ ⊂ B(0, R). Then
n

1 ϕ(x) − ϕ(0) ϕ(0) 1
f j , ϕ = dx + dx. (13.2.13)
j B(0,R) |x|n− 1j j 1
B(0,R) |x|n− j
Using the Mean Value Theorem and then (14.9.6) we may further write

1 ϕ(x) − ϕ(0) ∇ϕL∞ (Rn ) 1

dx ≤ dx
j B(0,R) |x| j
n− 1
j B(0,R) |x|n−1− 1
j

ωn−1 ∇ϕL∞ (Rn ) R j +1
1
ωn−1 ∇ϕL∞ (Rn ) R
1
= ρ dρ =
j −−−−→ 0. (13.2.14)
j 0 j+1 j→∞
One more use of (14.9.6) implies

1 1 ωn−1 R
ρ j −1 dρ
1
dx =
j n− 1j j
B(0,R) |x| 0
1
= ωn−1 R j −−−−→ ωn−1 . (13.2.15)
j→∞
By combining (13.2.13)–(13.2.15) we obtain
lim f j , ϕ = ωn−1 ϕ(0) = ωn−1 δ, ϕ. (13.2.16)

j→∞
The desired conclusion now follows.

Exercise 2.133 Note that fε ∈ Lloc1
(R) for every ε > 0, hence fε ∈ D (R). Also,
lim+ fε (x) = 0 for every x ∈ R \ {0}. For a given function ϕ ∈ C0∞ (R) let R > 0 be
ε→0
such that supp ϕ ⊂ (−R, R) and for ε > 0 write

1 R
xε ϕ(x) − ϕ(0) ϕ(0)ε R
dx
fε , ϕ = · dx +
π −R x +ε
2 2 x π −R x2 + ε2
=: I + II. (13.2.17)

Then x2xε
+ε2
≤ 1/2 for each x ∈ R if ε > 0, so we may apply Lebesgue’s Dominated
Convergence Theorem to conclude that lim+ I = 0. Also, integrating the second
ε→0
2ϕ(0)
term and then taking the limit yields lim+ II = lim+
π ε→0 arctan(R/ε) = ϕ(0). In
ε→0
conclusion, lim+ fε , ϕ = ϕ(0) = δ, ϕ as desired.
ε→0
∞
√ 2.134 Let ϕ ∈ C0 (R ) and ε > 0. Using first the change of variables
n
Exercise
x = 2 εy and then Lebesgue’s Dominated Convergence Theorem, we have

|x|2
fε , ϕ = (4πε)− 2 e− 4ε ϕ(x) dx
n
(13.2.18)
Rn

√
= π− 2 e−|y| ϕ(2 εy) dy −−−−→ π− 2 e−|y| ϕ(0) dy = ϕ(0).
n 2 n 2
+ ε→0
Rn Rn
Exercise 2.135 Note that for every ε > 0 we have | fε± | ≤ 1ε , hence these are functions
1
in Lloc (R) ⊂ D (R). Let ϕ ∈ C0∞ (R) and let R > 0 be such that supp ϕ ⊂ (−R, R).
Then
R R
± x ∓ iε x ∓ iε
fε (x), ϕ = ϕ(0) 2 + ε2
dx + 2 + ε2
[ϕ(x) − ϕ(0)] dx =: I + II. (13.2.19)
−R x −R x
x
Since x2 +ε2
is odd, we further obtain
I = ∓2iϕ(0) arctan (R/ε) −−−−→

+
∓iπϕ(0) = ∓iπδ, ϕ. (13.2.20)
ε→0
As for II, since lim+ x∓iε

= 1
for every x 0 and
ε→0 x +ε
2 2 x

x ∓ iε ϕ(x) − ϕ(0) ≤ |ϕ(x) − ϕ(0)| ∈ L1 (−R, R),
x2 + ε2 |x|
we may apply Lebesgue’s Dominated Convergence Theorem to obtain

R
ϕ(x) − ϕ(0) 1
lim+ II = dx = P.V. , ϕ . (13.2.21)
ε→0 −R x x
Exercise 2.136 You may use the following outline:

(a) Show that for f ∈ L1 (R) one has lim R f (x) sin( jx) dx = 0 by reducing to the
j→∞
case f ∈ C0∞ (R) based on density arguments.
(b) For ϕ ∈ C0∞ (R) and R > 0, write the expression π1 sinx jx , ϕ(x) as the sum of two
integrals, one over the region {x ∈ R : |x| ≤ R}, the other
over {x ∈ R : |x| > R}, and
use (a) to obtain the desired conclusion. Recall that R sinx x dx = π.
Exercise 2.137 Let ϕ ∈ C0∞ (R). √
(a) In the expression for f j , ϕ use the change of variables x = y/ j and then
D (R)
Lebesgue’s Dominated Convergence Theorem to conclude f j −−−−→ δ.
j→∞
(b) Integrate by parts m + 1 times to conclude that

| f j , ϕ| = jm cos( jx)ϕ(x) dx ≤ j−1 |ϕ(m+1) (x)| dx −−−−→ 0,
R R j→∞
D (R)
hence f j −−−−→ 0.
j→∞
D (R)
(c) Use a reasoning similar to one in the proof of (a) to conclude that f j −−−−→ δ,
j→∞
this time via the change of variables x = y/ j.
(d) Not convergent since if the function ϕ ∈ C0∞ (R) is such that ϕ(0) 0, then
u j , ϕ = (−1) j ϕ(1/ j) and the sequence {(−1) j ϕ(1/ j)} j∈N is not convergent.
(e) Use the Mean Value Theorem to obtain that u j , ϕ −−−−→ ϕ (0) = −δ , ϕ.
j→∞
D (R)
(f) f j −−−−→ P.V. 1
x.
j→∞
D (R)
(g) Use a reasoning similar to the one in the proof of (a) to conclude that f j −−−−→
j→∞
2
δ, this time via the change of variables x = y/ j and recalling that R (siny2y) dy = π.
(h) Integrate by parts m + 1 times and then use Lebesgue’s Dominated Conver-
D (R)
gence Theorem to conclude that f j −−−−→ 0.
j→∞
(j) Integrate by parts twice to obtain
∞
u j , ϕ = iϕ(0) + i ei jx ϕ (x) dx
0
∞
1 1
= iϕ(0) − ϕ (0) − ei jx ϕ (x) dx −−−−→ iϕ(0) = iδ, ϕ. (13.2.22)
j j 0 j→∞
Exercise 2.138 (H(· − a)) = δa in D (R).

Exercise 2.139 (u f ) = aδa + H(· − a) in D (R).
Exercise 2.140 Let ϕ ∈ C0∞ (R). Then integration by parts yields
0 ∞
− f (x)ϕ (x) dx = sin(x)ϕ (x) dx − sin(x)ϕ (x) dx
R −∞ 0
0 ∞
=− cos(x)ϕ(x) dx + cos(x)ϕ(x) dx, (13.2.23)
−∞ 0
hence (u f ) = cos(x)H(x) − cos(x)H(−x) in D (R). Similarly,

0 ∞
f (x)ϕ (x) dx = 2ϕ(0) + sin(x)ϕ(x) dx − sin(x)ϕ(x) dx, (13.2.24)
R −∞ 0
hence (u f ) = 2δ − sin(x)H(x) + sin(x)H(−x) in D (R).

Exercise 2.141 Let a, b ∈ I be such that a < x0 < b. Since for every x ∈ [a, x0 ) we
x
have f (x) = f (a) + a f (t) dt, Lebesgue’s Dominated Convergence Theorem gives
x0
that lim− f (x) exists and equals f (a) + a f (t) dt. A similar argument proves that
x→x0
b
lim+ f (x) = f (b) − x f (t) dt. Note that f ∈ L1 [a, b] . Suppose now that ϕ ∈ C0∞ (I)
x→x0 0
satisfies supp ϕ ⊂ [c, d] for some c < x0 < d. Then for ε > 0 small enough we have
d
(u f ) , ϕ = −u f , ϕ = − f (t)ϕ (t) dt
c
x0 −ε x0 +ε

= − f (x0 − ε)ϕ(x0 − ε) + f (t)ϕ(t) dt − f (t)ϕ (t) dt
c x0 −ε
d
+ f (x0 + ε)ϕ(x0 + ε) + f (t)ϕ(t) dt. (13.2.25)
x0 +ε
x0 +ε
Send ε → 0+ in (13.2.25) and observe that lim+ x0 −ε
f (t)ϕ (t) dt = 0 by Lebesgue’s
ε→0
Dominated Convergence Theorem. The case when x0 is not in the interior of the
support of ϕ is simpler, and can be handled via a direct integration by parts.
Exercise 2.143 Use Exercise 2.141 and induction.
Exercise 2.144 Use Exercise 2.141 and the fact that since {xk }k∈N has no accumula-
tion point in I, for each R > 0 only finitely many terms of the sequence {xk }k∈N are
contained in (−R, R).
Exercise 2.145 Use Exercise 2.144.
Exercise 2.146 Clearly δΣ is well defined and linear. Also, for each compact set
K ⊂ Rn and every ϕ ∈ C0∞ (Rn ) satisfying supp ϕ ⊆ K we have
|δΣ (ϕ)| ≤ σ(Σ ∩ K) sup |ϕ(x)|. (13.2.26)

x∈K
This shows that δΣ ∈ D (Rn ) and has order zero. Also, if supp ϕ ∩ Σ = ∅ then
δΣ , ϕ = 0, thus supp δΣ ⊆ Σ.
To prove that Σ ⊆ supp δΣ , note that if x∗ ∈ Σ, then there exists a neighbor-
hood U(x∗ ) of x∗ and a local parametrization P of class C 1 near x∗ as in (14.6.2)–
(14.6.3). In particular, if u0 ∈ O is such that P(u0 ) = x∗ , then the vectors ∂1 P(u0 ), ...,
∂n−1 P(u0 ), are linearly independent. Upon recalling the cross product from (14.6.4),
this ensures that
c0 := ∂1 P(u0 ) × · · · × ∂n−1 P(u0 ) 0. (13.2.27)
Since P is of class C 1 , it follows that ∂1 P(u) × · · · × ∂n−1 P(u) ≥ c0 /2 for every u
belonging to some small open neighborhood O ⊆ O of u0 in Rn−1 . Using the fact
that P : O → P(O) is a homeomorphism (see Proposition 14.46), it follows that
there exists some open set V(x∗ ) in Rn contained in U(x∗ ) and containing x∗ with
= V(x∗ ) ∩ Σ. By further shrinking O
the property that P(O) if necessary, there is no
∗
loss of generality in assuming that V(x ) is bounded.
Now consider 0 < r1 < r2 < ∞ such that V(x∗ ) ⊆ B(x∗ , r1 ) ⊆ B(x∗ , r2 ). Pick a

function ϕ ∈ C0∞ B(x∗ , r2 ) with ϕ ≥ 0, ϕ ≡ 1 in a neighborhood of B(x∗ , r1 ) (see
Proposition 14.34). Then

δΣ , ϕ = ϕ(x) dσ(x) ≥ ϕ(x) dσ(x) (13.2.28)
Σ Σ∩V(x∗ )

= > 0.
∂1 P(u) × · · · × ∂n−1 P(u) du ≥ c0 |O| (13.2.29)

O
In a similar manner, for each function g satisfying g ∈ L∞ (K ∩Σ) for any compact
set K ⊆ Rn , we have gδΣ ∈ D (Rn ) and for each compact K one has
|(gδΣ )(ϕ)| ≤ gL∞ (K) σ(Σ ∩ K) sup |ϕ(x)|, ∀ ϕ ∈ C0∞ (Rn ), supp ϕ ⊆ K.

x∈K
Exercise 2.147 Use integration by parts (see Theorem 14.60) and Exercise 2.146.
Exercise 2.148 Use the definition of distributional derivative, integration by parts
(cf. Theorem 14.60), and Exercise 2.146.
Exercise 2.149 Let ϕ ∈ C0∞ (Rn ) be such that supp ϕ∩∂B(0, R) = ∅. In this scenario,
we have |x|2ϕ−R2 ∈ C0∞ (Rn ), hence we may write
ϕ
u, ϕ = (|x|2 − R2 )u, = 0.
|x|2 −R2
This proves that supp u ⊆ ∂B(0, R), thus u is compactly supported. Examples of
distributions satisfying the given equation include δ∂B(0,R) and δ x0 , for any x0 ∈ Rn
with |x0 | = R.
Exercise 2.150 Use Example 2.56.
Exercise 2.151 The derivative of order m is equal to:
(a) sgn(x) if m = 1 and 2δ(m−2) if m ≥ 2;
(b) 2δ(m−1) ;

n
n
(c) cos x H + [δ(4 j−1) − δ(4 j−3) ] if m = 4 j, j ∈ N0 ; − sin x H + δ + [δ(4 j) − δ(4 j−2) ]
j=1 j=1

n
if m = 4 j + 1, j ∈ N0 ; − cos x H + δ + [δ(4 j+1) − δ(4 j−1) ] if m = 4 j, j ∈ N0 ;
j=1

n
sin x H + [δ(4 j+2) − δ(4 j) ] if m = 4 j, j ∈ N0 ; the convention is that a sum is void if
j=0
the upper bound is lower than the lower bound in the summation sign;
(d) (sin x H) = cos x H and use (c);
(e) −δ1 + δ−1 + 2xχ[−1,1] if m = 1, −(δ1 ) + (δ−1 ) − 2δ1 − 2δ−1 + 2χ[−1,1] if m = 2,
and δ(m−1)
−1 − δ(m−1)
1 − 2δ(m−2)
1 − 2δ(m−2)
−1 − 2δ(m−3)
1 + 2δ(m−3)
−1 if m ≥ 3.
Exercise 2.152 For ϕ ∈ C0∞ (R2 ) use the change of variables u = x + y, v = x − y and
u−v
2 , 2
the reasoning in (1.1.6)–(1.1.8) with ψ(u, v) := ϕ u+v for u ∈ [2, 4], v ∈ [0, 2],
to write

(∂21 − ∂22 )χA , ϕ = [(∂21 ϕ)(x, y) − (∂22 ϕ)(x, y)] dx dy
A
4 2
=2 ∂u ∂v ψ(u, v) dv du
2 0
= 2[ψ(4, 2) − ψ(2, 2) − ψ(4, 0) + ψ(2, 0)]

= 2[ϕ(3, 1) − ϕ(2, 0) − ϕ(2, 2) + ϕ(1, 1)]

= 2[δ(3,1) − δ(2,0) − δ(2,2) + δ(1,1) ], ϕ . (13.2.30)
Exercise 2.153 Fix ϕ ∈ C0∞ (R2 ). Then, using the change of variables x = t + y we
obtain

∂1 (u f ), ϕ = − χ[0,1] (x − y)∂1 ϕ(x, y) dx dy
R2

=− χ[0,1] (t)(∂1 ϕ)(t + y, y) dt dy
R2

= [ϕ(y, y) − ϕ(1 + y, y)] dy. (13.2.31)
R
Similarly,

∂2 (u f ), ϕ = [ϕ(x, x − 1) − ϕ(x, x)] dx. (13.2.32)
R
Combining (13.2.31)–(13.2.32) it follows that ∂1 (u f ) − ∂2 (u f ) = 0 in D (R2 ) and, in

turn, that ∂21 (u f ) − ∂22 (u f ) = (∂1 + ∂2 ) ∂1 (u f ) − ∂2 (u f ) = 0 in D (R2 ).
Exercise 2.156 The uniqueness statement is a consequence of Proposition 2.52. Let
K be a compact set in Rn such that K ⊂ Ω. Refine {Ω j } j∈I to a finite subcover
{Ω k }k=1
N
, with k ∈ I for k = 1, . . . , N, of K. Consider a partition of unity {ϕk }1≤k≤N
subordinate to the cover {Ω k }k=1N
of K, as given by Theorem 14.37. Hence, ϕk ∈
N
C0∞ (Ω) with supp ϕk ⊂ Ω k for each k, and ϕk = 1 in a neighborhood of K. Next,
k=1
N
for each function ϕ ∈ C0∞ (Ω) with supp ϕ ⊆ K define uK (ϕ) := k=1 u k , ϕk ϕ. Show
N
that k=1 u k , ϕk ϕ is independent of the cover of K chosen and of the partition of
unity, thus uK : DK (Ω) → C is well defined. The map uK is clearly linear. Now set
u : D(Ω) → C to be the map given by u(ϕ) := uK (ϕ) for each ϕ ∈ C0∞ (Ω) such that
supp ϕ ⊆ K. Show that this map is well defined, satisfies u ∈ D (Ω), and uΩ = u j
j
in D (Ω j ) for every j ∈ I.

Exercise 2.157 Fix ϕ0 ∈ C0∞ (Rn ) with the property that Rn ϕ0 (x) dx = 1. Next, let

ϕ ∈ C0∞ (Rn ) be arbitrary and set λ := Rn ϕ(x) dx. Hence, if ψ := ϕ − λϕ0 , then

ψ ∈ C0∞ (Rn ) and Rn ψ(x) dx = 0, so
0 = u, ψ = u, ϕ − λu, ϕ0 . (13.2.33)

Consequently, u, ϕ = λu, ϕ0 = c, ϕ, where c := u, ϕ0 ∈ C.
Exercise 2.158 You may proceed by completing the following steps. Let ϕ ∈
C0∞ (Rn ).

Step I. Prove that Rn ϕ(x) dx = 0 if and only if there exist ϕ1 , . . . , ϕn ∈ C0∞ (Rn )
n
such that ϕ = ∂ j ϕ j . Do so by induction over n. Corresponding to n = 1 show
j=1
x
that if a, b ∈ R satisfy a < b and for ϕ ∈ C0∞ (a, b) one defines φ1 (x) := a ϕ(t) dt,
b
x ∈ (a, b), then φ1 ∈ C0∞ (a, b) if and only if a ϕ(x) dx = 0.
Step II. Show that the statement from Step I continues to hold if Rn is replaced
by (a1 , b1 ) × · · · × (an , bn ), where a j , b j ∈ R, a j < b j for each j = 1, . . . , n.
Step III. Fix a j , b j ∈ R, a j < b j for j = 1, . . . , n and consider the n-dimensional

rectangle Q := (a1 , b1 ) × · · · × (an , bn ). Let ϕ0 ∈ C0∞ (Q) be such that Q ϕ0 dx =

1. Then, if ϕ ∈ C0∞ (Q) is arbitrary, the function ψ := ϕ − Q ϕ dx ϕ0 belongs to

C0∞ (Q) and satisfies Q ψ dx = 0. As such, Step II applies and shows that there exist
n
ϕ1 , . . . , ϕn ∈ C0∞ (Q) such that ϕ = Q ϕ dx ϕ0 + ∂ j ϕ j . In turn, this permits us to
j=1
write
n

u, ϕ = u, ϕ0 ϕ dx − ∂ j u, ϕ j = u, ϕ0 , ϕ , (13.2.34)
Q j=1

which shows if cQ := u, ϕ0 ∈ C, then uQ = cQ in D (Q).
Step IV. Since Ω is connected and open, it is path connected. Now combine this
with the fact that u is locally constant (as proved in Step III) to finish the proof.
Exercise 2.159 By Proposition 2.81, it suffices to prove that there exists v ∈
D (Rn−1 ) such that u, ϕ ⊗ ψ = v, ϕδ, ψ for every ϕ ∈ C0∞ (Rn−1 ) and every
ψ ∈ C0∞ (R). Fix ϕ ∈ C0∞ (Rn−1 ), ψ ∈ C0∞ (R), and consider some ψ0 ∈ C0∞ (R) with the
property that ψ0 (0) = 1. Then there exists h ∈ C0∞ (R) satisfying ψ(xn )−ψ(0)ψ0 (xn ) =
xn h(xn ) for every xn ∈ R. This and the fact that xn u = 0 allows us to write
u, ϕψ = u, ϕ(ψ − ψ(0)ψ0 ) + u, ϕψ0 ψ(0)
= xn u, ϕh + u, ϕψ0 ψ(0) = u, ϕψ0 δ, ψ. (13.2.35)
Now define v : D(Rn−1 ) → C by v(ϕ) := u, ϕψ0 for ϕ ∈ C0∞ (Rn−1 ), and show that
v ∈ D(Rn−1 ). By (13.2.35) this v does the job.
Exercise 2.160 Fix ψ ∈ C0∞ (R) with the property that ψ(0) = 1. Use Exercise 2.159
and induction to show that u = c δ(x1 ) ⊗ · · · ⊗ δ(xn ), where c := u, ψ ⊗ · · · ⊗ ψ ∈ C.

Exercise 2.161 Fix ψ ∈ C0∞ (R) with the property that R ψ(s) ds = 1. Given any
function ϕ ∈ C0∞ (Rn ), at each point x = (x , xn ) ∈ Rn−1 × R we may write

ϕ(x) = ϕ(x) − ψ(xn ) ϕ(x , s) ds + ψ(xn ) ϕ(x , s) ds
R R
xn

= ∂ xn ϕ(x , t) − ψ(t) ϕ(x , s) ds dt + ψ(xn ) ϕ(x , s) ds.
−∞ R R

Since ∂n u = 0 in D (R ), this yields n

u, ϕ = u , ψ(xn ) ϕ(x , s) ds . (13.2.36)
R
In particular, if ϕ = ϕ1 ⊗ ϕ2 for some ϕ1 ∈ C0∞ (Rn−1 ) and some ϕ2 ∈ C0∞ (R), then

u, ϕ = u , ϕ1 ⊗ ψ ϕ2 (s) ds = u, ϕ1 ⊗ ψ1, ϕ2 . (13.2.37)
R
Define v : C0∞ (Rn−1 ) → C by v(θ) := u, θ ⊗ ψ for every θ ∈ C0∞ (Rn−1 ). Then
v ∈ D (Rn−1 ) and u(x , xn ) = v(x ) ⊗ 1 when restricted to C0∞ (Rn−1 ) ⊗ C0∞ (R). The
desired conclusion follows by recalling Proposition 2.81.
Exercise 2.164 Fix j ∈ N and note that for each x = (x1 , . . . , xn ) ∈ Rn we may write
j j
1 ix1 ξ1 1
f j (x) = e dξ1 ⊗ · · · ⊗ eixn ξn dξn . (13.2.38)
2π −j 2π −j
Also, for each j ∈ N and each k ∈ {1, . . . , n}, the fundamental theorem of calculus
gives
j j
ixk ξk sin( jxk )
e dξk = cos(xk ξk ) dξk = 2 , assuming xk 0. (13.2.39)
−j −j xk
Now use (13.2.38), (13.2.39), Exercise 2.136, and part (d) in Theorem 2.89 to finish
the proof.
Exercise 2.165 Note that u = −δ is a solution for the equation in (1). Hence, if u is
any other solution of the equation (x − 1)u = δ, then setting v := u + δ it follows
that (x − 1)v = 0 in D (R). Next use this and the reasoning from Example 2.76 to
conclude that the general solution for the equation in (1) is −δ + c δ1 , with c ∈ C.
Fix ψ ∈ C0∞ (R) with the property that ψ(0) = 1. Show that any solution u of the
equation in (2) satisfies

ϕ(x) − ϕ(0)ψ(x)
u, ϕ = a(x) dx + u, ψδ, ϕ , ∀ ϕ ∈ C0∞ (R) (13.2.40)
R x
and use this to obtain that the general solution of the equation in (2) is va +c δ, c ∈ C,
where va is the distribution given by

ϕ(x) − ϕ(0)ψ(x)
va , ϕ := a(x) dx, ∀ ϕ ∈ C0∞ (R).
R x
Similarly, any solution u of the equation in (3) satisfies
ϕ − ϕ(0)ψ
u, ϕ = v, + u, ψδ, ϕ , ∀ ϕ ∈ C0∞ (R), (13.2.41)
x
so the general solution for (3) is w + c δ, where c ∈ C and w is the distribution given
by
ϕ − ϕ(0)ψ
w, ϕ := v, , ∀ ϕ ∈ C0∞ (R).
x
Exercise 2.166 (a) Since H ∈ Lloc

1
(R) and

MR := (x, y) ∈ [0, ∞) × [0, ∞) : |x + y| ≤ R
is a compact set in R2 for each R > 0, by Remark 2.93 and Theorem 2.94, it follows
that H ∗ H is well defined and belongs to D (R). Fix a compact set K in R and let
R > 0 be such that K ⊂ (−R, R). Pick now ϕ ∈ C0∞ (R) with supp ϕ ⊆ K, and suppose
that ψ ∈ C0∞ (R2 ) satisfies ψ ≡ 1 on the set
MK := {(x, y) ∈ [0, ∞) × [0, ∞) : x + y ∈ K}.
Then

H ∗ H, ϕ = H(x)H(y)ϕ(x + y)ψ(x, y) dy dx
R R
∞ ∞
= ϕ(x + y)ψ(x, y) dy dx (13.2.42)
0 0
Note that

(x, y) ∈ [0, ∞) × [0, ∞) : |x + y| ≤ R (13.2.43)

= (x, y) : 0 ≤ x ≤ R, 0 ≤ y ≤ R − x ,
hence
R R−x R R−x
H ∗ H, ϕ = ϕ(x + y)ψ(x, y) dy dx = ϕ(x + y) dy dx
0 0 0 0
R R R z R
= ϕ(z) dz dx = ϕ(z) dx dz = zϕ(z) dz
0 x 0 0 0
= xH, ϕ. (13.2.44)

Alternatively, one may use Remark 2.95, to observe that H ∗ H in the distributional
sense is the distribution given by the function obtained by taking the convolution,
∞ in
the sense of (2.8.2), of the function H with itself. Hence, (H ∗ H)(x) = 0 χ[0,∞) (x −
y) dy = xH(x) for every x ∈ R.
2 2
(b) −xH(−x) (c) (x2 − 2 + 2 cos x)H(x) (d) x2 H(x) − (x−1)
2 H(x − 1)
(e) Exercise 2.146 tells us that δ∂B(0,r) is compactly supported, so the given con-
volution is well defined. Also, by Exercise 2.103, |x|2 ∗ δ∂B(0,r) ∈ C ∞ (Rn ) and equals

δ∂B(0,r) , |x − y|2 = |x − y|2 dσ(y) = rn−1 |x − rω|2 dσ(ω)
∂B(0,r) S n−1

= rn−1 [r2 + |x|2 − 2rx · ω] dσ(ω)
S n−1
= (rn+1 + |x|2 rn−1 )ωn−1 , ∀ x ∈ Rn . (13.2.45)
For the second equality in (13.2.45) we used the change of variables y = rω, ω ∈
S n−1 , while for the last equality we used the fact that since x · ω as a function in ω
is odd, its integral over S n−1 is zero.
D (Rn ) D (Rn )
Exercise 2.167 u j −−−−−→ 0, v j −−−−−→ 0. Given that u j , v j ∈ E (Rn ), we deduce that
j→∞ j→∞
D (Rn )
u j ∗ v j ∈ E (Rn ) for each j ∈ N. Also, u j ∗ v j −−−−−→ δ.
j→∞
Exercise 2.168 The limits in (a) and (b) do not exist. Since for each j ∈ N we
have f j ∈ E (R) and g j ∈ C ∞ (R), Exercise 2.103 may be used to conclude that
D (Rn )
f j ∗ g j ∈ C ∞ (R) and that f j ∗ g j = 1 for every j. Thus, f j ∗ g j −−−−−→ 1.
j→∞
Exercise 2.169 Note that Λ is well defined based on Proposition 2.102. You may
want to use Theorem 14.6 to prove the continuity of Λ.
Exercise 2.170 For f : Rn → C set f ∨ (x) := f (−x), x ∈ Rn . If u ∈ D (Rn ) is
such that u ∗ ϕ = 0 for every ϕ ∈ C0∞ (Rn ), then 0 = (u ∗ ϕ∨ )(0) = u, ϕ for every
ϕ ∈ C0∞ (Rn ), thus u = 0. This proves the uniqueness part in the statement. As for
existence, given Λ as specified, define u0 : D(Rn ) → C by u0 (ϕ) := δ, Λ(ϕ∨ ) for
ϕ ∈ C0∞ (Rn ). Being a composition of linear and continuous maps, u0 is linear and
continuous, thus u0 ∈ D (Rn ). Also, if ϕ ∈ C0∞ (Rn ) is fixed, we have

(u0 ∗ ϕ)(x) = u0 , ϕ(x − ·) = u0 , (ϕ∨ )(· − x) = δ, Λ(ϕ(· + x)

= δ, (Λϕ)(· + x) = Λ(ϕ)(x), ∀ x ∈ Rn . (13.2.46)
For the first equality in (13.2.46) we used Proposition 2.102, the third equality is
based on the definition of u0 , while the forth equality uses the fact that Λ is com-
mutes with translations.
Exercise 2.171 From hypotheses we obtain u, P = 0 for every polynomial P in

Rn . Now use Lemma 2.83 to conclude that u, ϕ = 0 for every ϕ ∈ C0∞ (Rn ). Let
ψ ∈ C0∞ (Rn ) be such that ψ ≡ 1 in a neighborhood of supp u. Then for every function
ϕ ∈ C ∞ (Rn ) we have 0 = u, ψϕ = u, ϕ since the support condition on u implies
u, (1 − ψ)ϕ = 0.
Exercise 2.172 Fix ϕ ∈ C ∞ (R) and write

k

ϕ 1j − kϕ(0) − ϕ (0) ln k
j=1
⎡ k ⎤

k
1 ⎢⎢⎢ ⎥⎥⎥
= 1
ϕ j − ϕ(0) − j ϕ (0) + ϕ (0) ⎢⎢⎣ ⎢ 1
− ln k ⎥⎥⎥ . (13.2.47)
j ⎦
j=1 j=1
Since
ϕ 1 − ϕ(0) − 1 ϕ (0) ≤ 1
ϕ L∞ ([0,1]) ∀ j ∈ N,
j j j2
taking the limit as k → ∞ in (13.2.47) (also recall Euler’s constant γ from (4.6.24))
we obtain

∞

u(ϕ) = ϕ 1j − ϕ(0) − 1j ϕ (0) + γ ϕ (0).
j=1

∞
Now apply Fact 2.63 with K := [0, 1], m := 2 and C := 1
j2
+ γ to conclude
! j=1
u ∈ E (R). Also, show that supp u = {0} ∪ 1j : j ∈ N .
Exercise 2.173 Note that f j ∈ Lcomp 1

(R) for each j ∈ N, hence { f j } j∈N ⊂ E (R). If
∞
ϕ ∈ C (R) then we may write

1/ j j
j 1/ j
f j , ϕ − δ, ϕ = ϕ(x) − ϕ(0) dx ≤ |ϕ(x) − ϕ(0)| dx
−1/ j 2 2 −1/ j
1
≤ ϕ L∞ ([−1,1]) → 0 as j → ∞. (13.2.48)
j
Exercise 2.174 Since f j ∈ Lcomp 1

(R) for each j ∈ N, we have { f j } j∈N ⊂ E (R). Let
∞
ϕ ∈ C0 (R) and suppose R ∈ (0, ∞) is such that supp ϕ ⊂ (−R, R). Then for j ∈ N
with j ≥ R we have
j
ϕ(x) R ϕ(x) 2RϕL∞ (R)
| f j , ϕ| = dx = dx ≤ (13.2.49)
−j j −R j j
D (Rn )
which proves that f j −−−−−→ 0. Suppose next that there exists a distribution u ∈
j→∞
E (Rn )
E (R ) such that f j −−−−→ u. In particular, we have lim f j , ϕ = u, ϕ for every
n
j→∞ j→∞
ϕ ∈ C0∞ (R). Together with what we have proved before, this forces u = 0. However,
f j , 1 = 2 for every j ∈ N, which contradicts the fact that u = 0. Thus, { f j } j∈N does
not converge in E (Rn ).
Exercise 2.175 Since f j ∈ C0∞ (R) for each j ∈ N, we have { f j } j∈N ⊂ E (R). Let
ϕ ∈ C ∞ (Rn ). Then using the change of variables jx = y we may write

| f j , ϕ − δ, ϕ| = jn ψ( jx)ϕ(x) dx − ϕ(0) ψ(x) dx
Rn Rn

= ψ(y)ϕ(y/ j) dy − ϕ(0) ψ(x) dx
Rn Rn

≤ |ϕ(y/ j) − ϕ(0)||ψ(y)| dy → 0 as j → ∞, (13.2.50)
supp ψ
where the convergence is based on Lebesgue’s Dominated Convergence Theorem.

Exercise 2.176 For each k ∈ {1, . . . , n} consider the sequence {δ x j , ϕk } j∈N where
the function ϕk is defined by ϕk (x) := xk for each x = (x1 . . . , xn ) ∈ Rn .
Exercise 3.33 Let f ∈ S(Rn ) and α, β ∈ Nn0 . Let N := |α| + 1. Then Exercise 3.5
implies the existence of a constant C ∈ (0, ∞) such that ∂β f (x) ≤ C(1 + |x|)−N for
all x ∈ Rn . Hence,
C|xα |
sup xα ∂β f (x) ≤ sup −N
≤ CR−1 . (13.3.1)
|x|≥R |x|≥R (1 + |x|)
The desired conclusion follows after taking the limit as R → ∞.

Exercise 3.34 Let R > 0 be such that supp ϕ ⊂ B(0, R). Then ϕ j ∈ C0∞ (Rn ) and
supp ϕ j ⊂ B(0, jR) for each j ∈ N. Since there is no compact K ⊂ Rn such that
supp ϕ j ⊂ K for all j ∈ N, the sequence {ϕ j } j∈N does not converge in D(Rn ). Also,
for each α ∈ Nn0 we have ∂α ϕ j (x) = j1|α| e− j (∂α ϕ)(x/ j). Hence if we also take β ∈ Nn0
arbitrary, then
sup |xβ ∂α ϕ j (x)| ≤ e− j j−|α| sup |xβ (∂α ϕ)(x/ j)|

x∈Rn x∈B(0, jR)
≤ e− j j−|α| sup |xβ |∂α ϕL∞ (B(0,R))

x∈B(0, jR)
≤ e− j j|β|−|α| R|β| ∂α ϕL∞ (B(0,R)) −−−−→ 0, (13.3.2)

j→∞
which implies that {ϕ j } j∈N converges to zero in S(Rn ).

Exercise 3.35 For each α ∈ Nn0 we have ∂α ϕ j (x) = j|α|+1

1
(∂α ϕ)(x/ j), for each x ∈ Rn .
Consequently, for every compact subset K of R we may write
n
sup |∂α ϕ j (x)| ≤ j−|α|−1 sup |(∂α ϕ)(x/ j)|

x∈K x∈K
−|α|−1
≤ j ∂α ϕL∞ (Rn ) −−−−→ 0, (13.3.3)
j→∞
E(Rn )
which proves that ϕ j −−−−→ 0. Moreover, if
j→∞
O := {x = (x1 , . . . , xn ) ∈ Rn : x j 0 for j = 1, . . . , n}
we claim that there exists x∗ ∈ O with the property that ϕ(x∗ ) 0. Indeed, if this
were not to be the case, we would have ϕ = 0 in O, hence ϕ = 0 in Rn since ϕ is
continuous and O is dense in Rn . Having fixed such a point x∗ we then proceed to
estimate
sup |xβ ϕ j (x)| = j−1 sup |xβ ϕ(x/ j)| = j|β|−1 sup |xβ ϕ(x)|
x∈Rn x∈Rn x∈Rn
≥ j|β|−1 |x∗β | |ϕ(x∗ )| −−−−→ ∞, (13.3.4)

j→∞
for β ∈ Nn0 satisfying |β| > 1. Thus, {ϕ j } j∈N does not converge in S(Rn ).
Exercise 3.36 Clearly {ϕ j } j∈N ⊂ C ∞ (R). If m ∈ N then for each j ∈ N0 we have
1
m
m! 1
ϕ(m)
j (x) = ψ(k) (x)θ(m−k) (x/ j) m−k , ∀ x ∈ R. (13.3.5)
j k=0 k!(m − k)! j
Hence, if ∈ N0 , then using the properties of θ and ψ we may write
1
m
m! x
sup |x ϕ(m)
j (x)| = sup ψ(k) (x)θ(m−k) (x/ j) m−k
x∈R j x∈R k=0 k!(m − k)! j
1 1
m
m! x
≤ sup |ψ(m) (x)| + sup e−x θ(m−k) (x/ j) m−k
j |x|≤1 j x>1 k=0 k!(m − k)! j
C C C
≤ + sup |e−x x | ≤ −−−−→ 0. (13.3.6)
j j x>1 j j→∞
S(Rn )
In conclusion, ϕ j −−−−→ 0.
j→∞
Exercise 3.37 (a) Not in S(Rn ) since it is not bounded since lim e−x1 = ∞.
x1 →−∞
(b) Since e−|x| ∈ S(Rn ) and |x|2n! ∈ L(Rn ), by (a) in Theorem 3.14, their product
2
is in S(Rn ).
(c) (1 + |x|2 )2 is a polynomial function and if (1 + |x|2 )−2 ∈ S(Rn ) then their
n n
product, which is equal to 1, would belong to S(Rn ), which is not true.

(d) Show first that sin(e−|x| ) ∈ S(Rn ). Then, given that 1+|x|
2
2 ∈ L(R ), it follows
1 n
2
sin(e−|x| )
that 1+|x|2
∈ S(Rn ).
−|x|2
(e) Not in S(Rn ) since lim (1 + |x|2 )n+1 cos(e
(1+|x|2 )n
)
= ∞.
|x|→∞
(f) Set ϕ(x) := e−|x| sin(e x1 ), x ∈ Rn . If ϕ were to belong to S(Rn ) then ∂1 ϕ and
2 2
x1 ∂ϕ would be bounded. However, for every x = (x1 , ..., xn ) ∈ Rn ,
∂1 ϕ(x) = 2e−x2 −x3 −···−xn x1 [cos(e x1 ) − e−x1 sin(e x1 )].

2 2 2 2 2 2
(13.3.7)
2
In particular, (∂1 ϕ)(x1 , 0, ..., 0) + 2x1 ϕ(x1 , 0, ..., 0) = 2x1 cos(e x1 ) which is not
bounded.
(g) Since A is positive definite, there exists a real, symmetric, positive definite
n×n matrix B such that B2 = A. Hence, ϕ(x) = e−|Bx| for every x ∈ Rn , which means
2
that ϕ is the composition between the function e−|x| ∈ S(Rn ) (recall Exercise 3.3)
2
and the linear transformation B that maps S(Rn ) into itself. Recalling Exercise 3.17
we may conclude that ϕ ∈ S(Rn ).
Exercise 3.38 From part (g) in Exercise 3.37 we have f ∈ S(Rn ). Pick some B ∈
Mn×n (R) with the property that B2 = A, so that f (x) = e−|Bx| for every x ∈ Rn . Now
2
use Exercise 3.24 and Example 3.22.

1 1/2
Exercise 3.39 Let A := . Then f (x) = e−(Ax)·x for every x ∈ R2 and
1/2 1
Exercise 3.38 applies and yields
" 2π
f (ξ) = √ e−(ξ1 −ξ1 ξ2 +ξ2 )/3
2 2
for ξ = (ξ1 , ξ2 ) ∈ R2 . (13.3.8)
3
Exercise 3.40 Use (b) in Theorem 3.21 and Example 3.22.

Exercise 3.41 Since sin(x · x0 ) = 2i1 eix·x0 − e−ix·x0 matters reduced to the case
n = 1 after which we may apply Exercise 3.23. Hence, if x = (x1 , . . . , xn ), x0 =
(x01 , . . . , x0n ), and ξ = (ξ1 , . . . , ξn ), we may write
#
n
# −ax2j −ix j x0 j
n
F e−a|x| sin(x · x0 ) (ξ) = F e−ax j +ix j x0 j −
2 2
1
2i F e
j=1 j=1
#
n #
n
1 π2
n (ξ j −x0 j )2 (ξ j +x0 j )2
= 2i a e− 4a − e− 4a
j=1 j=1

1 π2
n |ξ−x0 |2 |ξ+x0 |2
= 2i a e− 4a − e− 4a . (13.3.9)
Exercise 3.42 Fix ϕ ∈ S(R). Suppose ψ ∈ S(R) is such that ψ = ϕ. Then

R2
ϕ(x) dx = lim ψ (x) dx = lim [ψ(R2 ) − ψ(R1 )] = 0.
R R1 ,R2 →∞ R1 R1 ,R2 →∞
x
Conversely, if R ϕ(x) dx = 0, then the function ψ(x) := 0 ϕ(t) dt, for x ∈ R,
belongs to S(R) and ψ = ϕ.
Exercise 4.107 Using the change of variables y = −x on the region corresponding

to x < 0, we have
∞
ϕ(x) ϕ(x) − ϕ(−x)
dx = dx, ∀ ϕ ∈ S(Rn ). (13.4.1)
|x|≥ε x ε x
Moreover, for each ε > 0 and each ϕ ∈ S(R) we may write

1 ∞
ϕ(x) |ϕ(x) − ϕ(−x)| |ϕ(x) − ϕ(−x)|x
dx ≤ dx + dx
|x|≥ε x ε x 1 x2
∞ dx

≤ sup |ϕ (x)| + 2 sup |xϕ(x)| (13.4.2)
x∈R 1 x2 x∈R
so that P.V. 1x is well defined and

1 dx
P.V. , ϕ ≤ 1 + sup |xk ϕ( j) (x)| (13.4.3)

|x|>1 |x|
x 2
x∈R, k=0,1, j=0,1
for all ϕ ∈ S(R) Hence, (4.1.2) holds for m = k = 1. Since P.V. 1x is also linear, we
obtain P.V. 1x ∈ S (Rn ).
Exercise 4.108 Use Exercise 2.124, (2.1.9), and Exercise 4.6.
Exercise 4.109 Use Exercise 2.126 and (4.1.33).
Exercise 4.110 Use a reasoning similar to the proof of the fact that the function in
(4.1.35) is not a tempered distribution to conclude that eax H(x) and e−ax H(−x) do
not belong to S (R) while e−ax H(x), eax H(−x) ∈ S (R).
Exercise 4.111 Let ϕ ∈ S(R) and j ∈ N.
(a) We may write

x x x2 ϕ(x) − ϕ(0)
−2
,ϕ = ϕ(x) dx + · dx
x2 +j |x|>1 x2 + j−2 |x|≤1 x2 + j−2 x

x
+ ϕ(0) dx. (13.4.4)
|x|≤1 x2 + j−2
The last integral in (13.4.4) is equal to zero (the function integrated is odd) while
for the other two integrals in (13.4.4) apply Lebesgue’s Dominated Convergence
Theorem to obtain that the first integral converges to |x|>1 ϕ(x) x dx and the second
ϕ(x)−ϕ(0) S (R)
integral converges to |x|≤1 x dx. In conclusion, x2 + j−2 −−−−→ P.V. 1x .
x
j→∞
(b) Making the change of variables x = y/ j and then applying Lebesgue’s Dom-
inated Convergence Theorem write
1 ϕ(x)

ϕ(y/ j)
−2
, ϕ = −2
dx = dy −−−−→ πϕ(0), (13.4.5)
j(x + j )
2
R j(x + j )
2
R y +1
2 j→∞
S (R)
hence 1
−−−−→
j(x2 + j−2 ) j→∞
πδ.

sin( jx) S (R)
(c) See Exercise 2.136. πx − −−−→
j→∞
δ.
D (R)
(d) Prove first that e j δ j −−−−→ 0. If {e j δ j } j∈N would converge in S (R) it would
2 2
j→∞
x2
have to converge to 0. However, for the test function e− 2 ∈ S(R) one has
2 x2 j2
e j δ j , e− 2 = e 2 −−−−→ ∞.
j→∞
Exercise 4.112 Use the change of variables jx = y and Lebesgue’s Dominated

S (R)
Convergence Theorem to show that jn θ( jx) −−−−→ δ.
j→∞
Exercise 4.113 Use Lebesgue’s Dominated Convergence Theorem to prove that

D (R)
f j −−−−→ e x . On the other hand, the sequence { f j } j∈N does not convergence in S (R)
2
j→∞
since
/2
j
f j , e−x /2 /2
2 2 2
= ex dx −−−−→ ex dx = ∞.
−j j→∞ R
Exercise 4.114 If there were some f ∈ L p (R) such that lim f j − f L p (R) = 0, then
j→∞
lim f j (x) = f (x) for almost every x ∈ R. Since lim f j (x) = 0 for every x ∈ R, we
j→∞ j→∞
have that f = 0 almost everywhere on R. This leads to a contradiction given that we
j
would have 0 = lim f j Lp p (R) = j−1 1 dx = 1. Hence, the sequence { f j } j∈N does not
j→∞
converge in L p (R). If ϕ ∈ S(R) then
j
dx C
| f j , ϕ| ≤ sup |x2 ϕ(x)| ≤ −−−−→ 0, (13.4.6)
x∈R j−1 x2 j( j − 1) j→∞
S (R)
which shows that f j −−−−→ 0.
j→∞
Exercise 4.115 Using Lebesgue’s Dominated Convergence Theorem it is not diffi-

cult to check that u j , ϕ −−−−→ e x sin(e x ), ϕ for every ϕ ∈ C0∞ (R).
j→∞
If ϕ ∈ S(R), integration by parts yields
j j j
e x sin(e x )ϕ(x) dx = − cos(e x )ϕ(x)− j + cos(e x )ϕ (x) dx. (13.4.7)
−j −j
Since ϕ ∈ S(R) one has cos(e x )ϕ ∈ L1 (R) and

cos(e x )ϕ(x) j = cos(e ) xϕ(x) j ≤ C −−−−→ 0.
x
−j x −j j j→∞

Thus, lim u j , ϕ = R cos(e x )ϕ (x) dx and if u : S(R) → C is such that
j→∞

u, ϕ := cos(e x )ϕ (x) dx for every ϕ ∈ S(R), (13.4.8)
R
then u is a well defined, linear mapping and

dx
|u, ϕ| ≤ sup |(1 + x2 )ϕ (x)| for every ϕ ∈ S(R). (13.4.9)
R 1+x
2
x∈R
S (R)
Hence u ∈ S (R) and u j −−−−→ u.
j→∞
Exercise 4.117 Fix f ∈ S (R) and let ψ ∈ S(R) be such that ψ(0) = 1. Then, if u is
a solution of the equation xu = f in S (R), for each ϕ ∈ S(R) we have
ϕ − ϕ(0)ψ
u, ϕ = u, x + u, ϕ(0)ψ
x
ϕ − ϕ(0)ψ
= f, + u, ψδ, ϕ. (13.4.10)
x
Set ⎧ ϕ(x)−ϕ(0)ψ(x)
⎪
⎪
⎨ if x ∈ R \ {0},
x
g(x) := ⎪
⎪ (13.4.11)
⎩ ϕ (0) − ϕ(0)ψ (0) if x = 0.
1

Note that g ∈ S(R), since g(x) = ϕ (tx) − ϕ(0)ψ (tx) dt for x ∈ R. Thus, if we
0
define
⎧ ϕ(x)−ϕ(0)ψ(x)
⎪
⎪
⎨ if |x| > 1,
x

ϕ(x) := ⎪
⎪ (13.4.12)
⎩ g(x) if |x| ≤ 1,
then
ϕ ∈ S(R) and (
xϕ = ϕ. There remains to observe that the mapping
w f : S(R) → C, w f , ϕ := f,
ϕ ∀ ϕ ∈ S(R), (13.4.13)
ϕ is as in (13.4.12), satisfies w f ∈ S (R) in order to conclude that u = w f +cδ,

where
some c ∈ R.
Exercise 4.118 Suppose u ∈ S (Rn ) is a solution of the equation e−|x| u = 1 in
2
S (Rn ). Then for ϕ ∈ C0∞ (Rn ) we have

2 2 2
u, ϕ = u, e−|x| e|x| ϕ = 1, e|x| ϕ = e|x| , ϕ ,
2
(13.4.14)

which shows that uC ∞ (Rn ) = e|x| . Since C0∞ (Rn ) is sequentially dense in S(Rn ), this
2
would imply that u = e|x| . However, as proved following Remark 4.16, e|x| does not
2 2
belong to S (Rn ).
Exercise 4.119 Since H is Lebesgue measurable and (1 + x2 )−1 H ∈ L1 (R), by
Exercise 4.6 it follows that H defines a tempered distribution. We have seen that
H ∈ D (R) and H = δ in D (R) (recall Example 2.44). Since δ ∈ S (R), by
(4.1.33) it follows that H = δ in S (R). Taking the Fourier transform

of the last
" = 1 in S (R). On the other hand, ξ P.V. 1 = 1 in S (R)
equation we arrive at iξ H ξ
"
by Exercise 4.107 and (2.3.6). Consequently, ξ iH − P.V. 1 = 0 in S (R). Now ξ
" − P.V. 1 = cδ, for some c ∈ C.
Example 2.76 may be used to conclude that iH ξ
Hence
" = −iP.V. 1 − cδ
H in S (R). (13.4.15)
ξ
" to the function ϕ(x) := e−x ∈ S(R). First, by Example 3.22
To determine c apply H
2
write
∞
" −x2 ) √ √
= H, e−x2 = H, πe−x /4 = π e−x /4 dx
2 2
H, e
0
√
∞
π
e−x /4
dx = e
2
= −x2 /4 (0) = π. (13.4.16)
2
−∞
2
e−x
Second, from (13.4.15) and the fact that x is an odd function obtain that

e−x
2
1
− i P.V. − cδ, e−x = −i lim+
2
dx − c = −c. (13.4.17)
ξ ε→0 |x|>ε x
" = −i P.V. 1 + πδ.

Combining (13.4.15), (13.4.16), and (13.4.17) arrive at H ξ
Exercise 4.120 You may use Exercise 4.119 to show that

1
F P.V. (ξ) = 2πiH(ξ) − iπ = iπ sgn (ξ) in S (R).

x
Exercise 4.121 (a) sgn = H − H ∨ , hence using Exercise 4.119 one may write
* =H
sgn *∨ = πδ − i P.V. 1 − πδ − i P.V. 1 = −2i P.V. 1
"− H (13.4.18)
ξ ξ ξ
in S (R). Alternatively, you may use Exercise 4.120 and take another Fourier trans-
form.
*k = x*
(b) If k is even, then |x|k = xk , so that |x| δ = (−D)k"
k" "
δ = 2πik δ(k) in S (R). If k is
odd, then |x| = x sgn x, thus
k k
*k = x 1
|x| gn x = −2ik+1 P.V. (k)
k sgn x = (−D)k s in S (R). (13.4.19)
ξ
(c) You may use Example 4.36. Alternatively, take the Fourier transform of the
identity x · sin(ax)
x = sin(ax) in S (R), then use (c) in Theorem 4.26 and (4.2.48) to
conclude that
sin(ax)
F = πδ−a − πδa in S (R).

x
Now use Exercise 2.138 and Proposition 2.47 to obtain
sin(ax)
F = π sgn(a)χ[−|a|,|a|] + c in S (R), (13.4.20)

x
and then show that c = 0 by applying the Fourier transform to the last identity and
recalling (4.2.45).
(d) Take the Fourier transform of the identity
sin(ax) sin(bx)
x· = sin(ax) sin(bx) in S (R),
x
then use (c) in Theorem 4.26 and (4.2.48) to conclude that

sin(ax) sin(bx)
iπ
F = [δa+b − δa−b − δb−a + δ−a−b ] in S (R).
x 2
Now use Exercise 2.138 and Proposition 2.47 to obtain

sin(ax) sin(bx)
F
x
iπ
= [H(x − a − b) − H(x − a + b) − H(x − b + a) + H(x + a + b)] + c0
2
iπ
= sgn(b) χ[−a−|b|,−a+|b|] − χ[a−|b|,a+|b|] + c0 (13.4.21)
2
in S (R), and then show that c0 = 0 by applying the Fourier transform to the last
identity. 2
(e) Using the identity sin(x2 ) = 2i1 eix − e−ix , (4.2.22) and (4.2.56), we may write
2
1 +* ,

sin(x 2) = eix2 − e)
−ix2
2i
√ + ,
π i π −i ξ2 π ξ2
= e 4 e 4 − e−i 4 e i 4
2i
√ + ,
π i π−ξ2 π−ξ2
= e 4 − e−i 4 . (13.4.22)
2i
(f) Use (4.6.26), (4.2.3), and Proposition 4.32 to show
)
ln |x| = −2πγδ − πwχ(−1,1) in S (R).
Exercise 4.122 The computation of the Fourier transform in Example 4.24 naturally
adapts to the current setting and yields that
1
πieib|ξ| −a|ξ|
F (ξ) = e , ∀ ξ ∈ R. (13.4.23)
x2 − (b + ia)2 b + ia
Exercise 4.123 From Exercise 2.146 we know that δ∂B(0,R) ∈ E (R3 ), thus by (b) in
Theorem 4.35 it follows that

−ix·ξ
δ∂B(0,R) (ξ) = δ∂B(0,R) , e = e−ix·ξ dσ(x), ∀ ξ ∈ R3 . (13.4.24)
∂B(0,R)
Check that δ∂B(0,R) is invariant under orthogonal transformations, and conclude that
δ
∂B(0,R) is invariant under orthogonal transformations. Fix ξ ∈ R \ {0} and show that
3
there exists an orthogonal transformation

A ∈ M3×3 (R) such that Aξ = (|ξ|, 0, 0),

and furthermore that δ∂B(0,R) (ξ) = ∂B(0,R) e−ix1 |ξ| dσ(x). Now use spherical coordi-
nates to compute the latter integral and conclude that δ ∂B(0,R) (ξ) = 4πR
sin(R|ξ|)
|ξ| . Treat
separately the case ξ = (0, 0, 0).
Exercise 4.124 Use Lemma 4.28.
Exercise 4.125 Since χ[−1,1] ∈ L1 (R), from (4.1.9) it follows that χ[−1,1] ∈ S (R).

Also, by Exercise 4.27, we have that χ [−1,1] in S (R) is the tempered distribution
given by the function
1
2 sin ξ
e−ixξ dx = , ξ ∈ R \ {0}.
−1 ξ
sin ξ
In particular, since ξ L1 (R) we have χ
[−1,1] L (R).
1
g = πδa + πδ−a in S (R).

Exercise 4.127 See Example 4.38 and deduce that "

Exercise 4.129 If P(x) = |α|≤m aα xα , then the condition P(tx) = t−k P(x) in Rn for

each t > 0 implies |α|≤m t|α| aα xα = t−k |α|≤m aα xα in Rn for each t > 0. Hence, for
each α ∈ Nn0 we have t|α| aα = t−k aα for every t > 0, or equivalently that t|α|+k aα = aα
for every t > 0. Now take t → 0+ in the last equality to obtain aα = 0.
Exercise 4.130 The first identity in (4.11.4) is easy to verify (keeping in mind that
|η| = |ζ| = 1). Check, via a direct calculation, that
w · [ζ − (1 + 2η · ζ)η] w · (ζ + η)
R w := w − ζ− η, ∀ w ∈ Rn . (13.4.25)
1+η·ζ 1+η·ζ
To show that R R = In×n , let ξ ∈ Rn be arbitrary and set a := ζ · η ∈ R \ {−1}. From

(13.4.25) and (4.11.3) we have
(Rξ) · [ζ − (1 + 2a)η] (Rξ) · (ζ + η)

R (Rξ) = Rξ − ζ− η
1+a 1+a
ξ · (η + ζ) ξ · [(1 + 2a)ζ − η]
=ξ− ζ+ η
1+a 1+a
1 ξ · (η + ζ)
− ξ · [ζ − (1 + 2a)η] − 1 − (1 + 2a)a ζ
1+a 1+a
ξ · [(1 + 2a)ζ − η]
− a − (1 + 2a) ζ
(1 + a)2
1 ξ · (η + ζ)
− ξ · (η + ζ) − (1 + a) η
1+a 1+a
ξ · [(1 + 2a)ζ − η]
− (1 + a)η = ξ. (13.4.26)
(1 + a)2
Hence, R R = In×n as wanted. The latter identity now implies the second identity in
(4.11.4) and the fact that RR = In×n .
Exercise 4.131 Fix c ∈ R \ {0} and use the fact that cosine is an odd function, the
fundamental theorem of calculus, and Fubini’s theorem to write

R
cos(cρ) − cos ρ cos(|c|ρ) − cos ρ
R
dρ = dρ
ε ρ ε ρ
R |c|
=− sin(rρ) dr dρ
ε 1
|c| R
=− sin(rρ) dρ dr
1 ε
cos(Rr)
|c|
cos(εr)
= − dr. (13.4.27)
1 r r
In particular, estimate (4.11.7) readily follows from this formula. Going further, we
note that an integration by parts gives
|c| sin(Rr) r=|c| 1 |c| sin(Rr)
cos(Rr)
dr = + dr. (13.4.28)
1 r Rr r=1 R 1 r2
By combining (13.4.27) with (13.4.28) we arrive at
|c|
R
cos(cρ) − cos ρ sin(|c|R) sin R 1 sin(Rr)
dρ = − + dr
ε ρ |c|R R R 1 r2
|c|
cos(εr)
− dr. (13.4.29)
1 r
Passing to limit R → ∞ and ε → 0+ in (13.4.29) yields (4.11.5) after observing

(e.g., by Lebesgue’s Dominated Convergence Theorem) that
|c| |c|
cos(εr) dr
lim+ dr = = ln |c|. (13.4.30)
ε→0 1 r 1 r
Next,
R Rc
sin(cρ) sin t
lim+ dρ = lim+ dt
ε→0 ε ρ ε→0 εc t
R→∞ R→∞

R
sin t π
= sgn c lim+ dt = sgn c, (13.4.31)
ε→0 ε t 2
R→∞
by a suitable change of variables and the well-known fact (based on a residue calcu-
R
lation) that lim ε→0+ ε sint t dt = π2 . This proves (4.11.6). Finally, to justify (4.11.8)
R→∞
use the fact that whenever 0 < a < b < ∞ an integration by parts gives
b cos t t=b b cos t
sin t −
dt = − dt. (13.4.32)
a t t t=a a t2

d m
Exercise 5.16 The ordinary differential equation dx v = 0, with initial conditions
v(0) = v (0) = · · · = v(m−2) (0) = 0, and v(m−1) (0) = 1, has the unique solution
v(x) = (m−1)!
1
xm−1 for x ∈ R. Hence, by Example 5.12 the function u := (m−1)! 1
xm−1 H
d m
is a fundamental solution for dx in R. In addition, by Exercise 4.119 we have
H ∈ S (R) which, when combined with the fact that (m−1)!1
xm−1 ∈ L(R), implies (as
a consequence of (b) in Theorem 4.14) that
1
xm−1 H ∈ S (R).
(m − 1)!
d m
If E ∈ S (R) is an arbitrary fundamental solution for dx , then by Proposition 5.8
d m
we have that E − u = P for some polynomial in R satisfying dx P = 0 in R. This
forces P to be a polynomial of degree less than or equal to m − 1.
Exercise 5.17 First prove that
P1 (D x ) ⊗ P2 (Dy ), ϕ1 (x) ⊗ ϕ2 (y) = δ x ⊗ δy , ϕ1 (x) ⊗ ϕ2 (y)
for every ϕ1 ∈ C0∞ (Rn ) and every ϕ2 ∈ C0∞ (Rm ), and then use Proposition 2.81.
Exercise 6.27 The ordinary differential equation v ∈ C ∞ (R), v + k2 v = 0 in R, with

initial conditions v(0) = 0 and v (0) = 1, has the unique solution
i ikt i
v(x) = − e + e−ikt for all x ∈ R. (13.6.1)
2k 2k
Hence, by Example 5.12 the function u := vH = − 2ki eikt H + 2ki e−ikt H is a fundamen-
d 2
tal solution for dx + k2 in R. In addition, by Exercise 4.119 we have H ∈ S (R)
which, when combined with the fact that v ∈ L(R), implies (as a consequence of (b)
in Theorem 4.14) that u ∈ S (R). If E ∈ S (R) is an arbitrary fundamental solution
d 2
for dx + k2 in R, then
d
2
+ k2 (E − u) = 0 in S (R).
dx
d 2
Since dx + k2 is elliptic (cf. Definition 6.13), it is hypoelliptic (cf. Theorem 6.15)
in R, hence E − u ∈ C ∞ (R). Thus, E − u is a solution in the classical sense of the
ordinary differential equation w + k2 w = 0 in R. The general solution of the latter
equation is c1 eikt + c2 e−ikt for any c1 , c2 ∈ C. Consequently, E = u + c1 eikt + c2 e−ikt
for c1 , c2 ∈ C.
Exercise 6.28 (a) The general solution for the ordinary differential equation v −
a2 v = 0 in R is v(x) := c1 eax + c2 e−ax for x ∈ R, where c1 , c2 ∈ C. Hence, if we
further impose the conditions v(0) = 0 and v (0) = 1 then c1 = 2a 1
= −c2 , and by
Example 5.12, we have that u := 2a e H − 2a e H satisfies u − a2 u = δ in D (R).
1 ax 1 −ax
d2 2
(b) By Theorem 5.14 there exists E ∈ S (R) such that dx 2 − a E = δ in S (R).
Fix such an E and apply the Fourier transform to the last equation. Then −(ξ2 +
" = 1 in S (R). Since 2 1 2 ∈ L(R) we may multiply the last equality by 2 1 2 to
a2 )E ξ +a ξ +a
conclude that −E " = 2 1 2 = f , and, furthermore, that "
f = −2πE ∨ . Therefore we can
ξ +a
determine " f as soon as we find a fundamental solution E ∈ S (R) for the operator
d2
dx2
−a .2
d2
2 − a is a hypoelliptic operator in R. As a consequence of
2
By Theorem 6.15, dx
Remark 6.7 we have (with u as in (a)) that u − E ∈ C ∞ (R) and is a classical solution
of the ordinary differential equation v − a2 v = 0 in R. Thus,
E = u + c1 eax + c2 e−ax
1 −ax
= 1 ax
2a e H − 2a e H + c1 eax H + c1 eax H ∨ + c2 e−ax H + c2 e−ax H ∨

= 1
2a + c1 eax H + − 2a
1
+ c2 e−ax H + c1 eax H ∨ + c2 e−ax H ∨ . (13.6.2)
The condition E ∈ S (R) implies (in view of Exercise 4.110) that 2a

1
+ c1 = 0 and
c2 = 0, which when used in (13.6.2) give E = − 2a e . Consequently, "
1 −a|x|
f (x) =
π −a|x|
a e for x ∈ R.
Exercise 6.30 Recall Exercise 6.4. Consider u := 1 + δ in D (Rn ). It follows that
sing supp u = {0} and supp u = Rn , thus sing supp u ⊂ supp u.

Exercise 6.31 Since ∂α δ = 0 we have sing supp ∂α δ ⊆ {x }. To see that the
x0 Rn \{x } x0 0
0
latter inclusion is in fact equality, use Example 2.13 if α = (0, . . . , 0) and if |α| > 0
use the fact that the order of the distribution ∂α δ x0 equals |α| while any distribution
of function type is of order zero.
Exercise
6.32 If x ∈ Ω is such that there exists an open set ω ⊆ Ω with x ∈ ω and
uω ∈ C ∞ (ω), then (au)ω ∈ C ∞ (ω), which gives
Ω \ sing supp u ⊆ Ω \ sing supp (au).

Exercise 6.33 Since P.V. 1x n = 1x and 1x belongs to C ∞ (Rn \ {0}), it follows

R \{0}
that sing supp P.V. 1x ⊆ {0}. Now using Example 2.11 we have that the distribution
P.V. 1x is not of function type.
Exercise 6.34 First prove that P(x) · 1

P(x) = 1 in S (Rn ). Then take the Fourier
transform to arrive at
1
P(−D)F (ξ) = (2π)n δ in S (Rn ). (13.6.3)

P
Use (13.6.3), (6.1.5), and Example 6.4, to show that it is not possible to have
1
sing supp F = ∅.
P
Exercise 6.35 sing supp u = R × {0}. To prove this you may want to use the fact that
1
u, ϕ = P.V. , ϕ(x, 0) (13.6.4)
x

ϕ(x, 0) ϕ(x, 0) − ϕ(0, 0)
= dx + dx
|x|≥1 x |x|≤1 x
for every ϕ ∈ C0∞ (R2 ).
Exercise 7.72 If the equation has two solutions E1 , E2 ∈ S (Rn ), set E := E1 − E2 .

Then ΔE − E = 0 in S (Rn ) implies (after an application of the Fourier transform)
" = 0 in S (Rn ). Since 1 2 ∈ L(Rn ) (recall Exercise 3.13), by part (a)
−(|ξ|2 + 1)E 1+|ξ|
2 ϕ ∈ S(R ) for every ϕ ∈ S(R ). Thus,
1 n n
in Theorem 3.14 we have 1+|ξ|
1
" ϕ = (1 + |ξ|2 )E,
E, " ϕ = 0, ∀ ϕ ∈ S(Rn ).
1 + |ξ|2
This proves that E " = 0 in S (Rn ), thus after applying the Fourier transform we
obtain E = 0 in S (Rn ). In turn, the latter implies E1 = E2 in S (Rn ) and the proof
of the uniqueness statement is complete. Regarding the existence statement, suppose
E ∈ S (Rn ) satisfies ΔE − E = δ in S (Rn ). This equation, via the Fourier transform,
is equivalent with −(|ξ|2 + 1)E " = 1 in S (Rn ). Since 1 2 ∈ S (Rn ) (use Exercise
1+|ξ|

4.7) and (|ξ|2 + 1) 1+|ξ| 1
= 1 in S (Rn
), it follows that the tempered distribution

−1
E := −F 1
1+|ξ|2
is a solution of ΔE − E = δ in S (R ), which must be unique
n
based on the earlier reasoning.

Exercise 7.73 Suppose there exists E ∈ L1 (Rn ) with ΔE = δ. Then −|ξ|2 E " = 1. We
know that E "
" ∈ L1 ⊂ C (R ) (recall (3.1.3)), hence [|ξ| E]
0 n 2 " = 0. This leads to
ξ=(0,...,0)
a contradiction.
Exercise 7.76 Since ∂z∂ ∂z∂ = 14 Δ and u is holomorphic on Ω \ K, it follows that
supp (Δu) ⊆ K. It follows that Δu ∈ C00 (Ω), hence Δu is absolutely integrable on Ω.
Take ϕ ∈ C0∞ (Ω) satisfying ϕ ≡ 1 in a neighborhood of K. Then integration by parts,
∂ ∂
the fact that ∂z u = 0 pointwise on Ω \ K, and that ∂z ϕ = 0 near K, imply

∂ ∂
Δu dx = ϕΔu dx = −4 ϕ u dx = 0. (13.7.1)
Ω Ω Ω ∂z ∂z
Exercise 7.77 Let ϕ ∈ C0∞ (Rn ). A direct computation based on the Chain Rule gives
that LA1 (ϕ ◦ B−1 ) = (LA2 ϕ) ◦ B−1 pointwise in Rn . This and (2.2.7) then allow us to
write
1 1
(LA1 u) ◦ B, ϕ = LA1 u, ϕ ◦ B−1 = u, LA1 (ϕ ◦ B−1 )
|detB| |detB|
1
= u, (LA2 ϕ) ◦ B−1 ) = u ◦ B, LA2 ϕ
|detB|
= LA2 (u ◦ B), ϕ, ∀ u ∈ D (Rn ). (13.7.2)
Exercise 7.78 Since EΔ ∈ S (Rn ) is a fundamental solution for Δ, we have ΔEΔ = δ

in S (Rn ). Fix j ∈ {1, . . . , n}. Then, Δ(∂ j EΔ ) = ∂ j δ in S (Rn ). Take the Fourier
transform of the last equation to arrive at
|ξ|2 F (∂ j EΔ ) = −iξ j in S (Rn ). (13.7.3)

ξj
Since n ≥ 2, by Example 4.4 we have |ξ|2
∈ S (Rn ). Also, |ξ|2 ∈ L(Rn ), thus
ξj
|ξ|2 · |ξ|2
∈ S (Rn ) (recall (b) in Theorem 4.14), and it is not difficult to check that
ξj
|ξ|2 · |ξ|2
= ξ j in S (Rn ). These facts combined with (13.7.3) imply

ξj
|ξ| F (∂ j EΔ ) + i 2 = 0 in S (Rn ).
2
(13.7.4)
|ξ|
ξ

Thus, supp F (∂ j EΔ ) + i |ξ|j2 ⊆ {0} and by Exercise 4.37, it follows that
ξj
F (∂ j EΔ ) = −i + P"j (ξ) in S (Rn ), (13.7.5)
|ξ|2
where P j is a polynomial in Rn . Now a direct computation gives that if n ≥ 2, then
1 xj
∂ j EΔ = · in S (Rn ). (13.7.6)
ωn−1 |x|n
Hence, ∂ j EΔ is positive homogeneous of degree 1 − n, which in turn when combined
with Proposition 4.59 implies that F (∂ j EΔ ) is positive homogeneous of degree −1.
Thus, the term in the right-hand side of (13.7.5) is positive homogeneous of degree
ξ
−1. Since |ξ|j2 is positive homogeneous of degree −1, we conclude that P"j (ξ) is
positive homogeneous of degree −1, and furthermore, by Proposition 4.59 and Ex-
ercise 4.57, that the polynomial P j is positive homogeneous of degree 1 − n ≤ −1.
Now invoking Exercise 4.129 we obtain P j ≡ 0. The latter when used in (13.7.5)
proves (7.14.4). Identities (7.14.5) and (7.14.6) follow from (7.14.4) and (13.7.6).
Exercise 7.79 Since EΔ2 ∈ S (Rn ) is a fundamental solution for Δ2 , we have Δ2 EΔ2 =
δ in S (Rn ). Fix j, k ∈ {1, . . . , n}. Then, Δ2 (∂ j ∂k EΔ2 ) = ∂ j ∂k δ in S (Rn ). Take the
Fourier transform of the last equation to arrive at
|ξ|4 F (∂ j ∂k EΔ2 ) = −ξ j ξk in S (Rn ). (13.7.7)
Under the current assumption on n, by Exercise 4.5 we have 1

|ξ|2
∈ S (Rn ). Also,
ξ j ξk ξ j ξk
|ξ|4
∈ Lloc
1
(Rn ) and in view of Example 4.4 one may infer that |ξ|4
∈ S (Rn ). In
ξ j ξk
addition, |ξ|4 ∈ L(Rn ), thus |ξ|4 · |ξ|4
, |ξ|4 · |ξ|12 ∈ S (Rn ) (recall (b) in Theorem 4.14),
ξ j ξk
and |ξ|4 · |ξ|4
= ξ j ξk , |ξ|4 · 1
|ξ|2
= |ξ| in S (Rn ). These facts combined with (13.7.7)
2
imply
ξ j ξk
|ξ|4 F (∂ j ∂k EΔ2 ) + 4 = 0 in S (Rn ). (13.7.8)
|ξ|
ξξ

Thus, supp F (∂ j ∂k EΔ2 ) + |ξ|j 4k ⊆ {0} and by Exercise 4.37, it follows that
ξ j ξk
F (∂ j ∂k EΔ2 ) = − + R*jk (ξ) in S (Rn ), (13.7.9)
|ξ|4
where R jk is a polynomial in Rn .
It is easy to check that
1 δ jk 1 x j xk
∂ j ∂ k E Δ2 = − · n−2 + · n in S (Rn ). (13.7.10)
2(n − 2)ωn−1 |x| 2ωn−1 |x|
Hence, ∂ j ∂k EΔ2 is positive homogeneous of degree 2 − n which, in turn, when com-
bined with Proposition 4.59 implies that F (∂ j ∂k EΔ2 ) is positive homogeneous of
degree −2. Thus, the term in the right-hand side of (13.7.9) is positive homoge-
ξξ
neous of degree −2. Since |ξ|j 4k is positive homogeneous of degree −2, we may con-
clude that R*jk (ξ) is positive homogeneous of degree −2 and furthermore (by Propo-
sition 4.59 and Exercise 4.57) that the polynomial R jk is positive homogeneous of
degree 2 − n ≤ −1. Now invoking Exercise 4.129 we obtain R jk ≡ 0. The latter
when used in (13.7.9) proves (7.14.7). Identity (7.14.8) follows from (7.14.7) and
(13.7.10). As for identity (7.14.9), apply the Fourier transform to (7.14.8) and then
use Proposition 4.64 with λ = n − 2.
Exercise 7.80 If P(D) is elliptic then first show that there exists C ∈ (0, ∞) such that
Pm (ξ) ≥ C for ξ ∈ S n−1 , thus conclude Pm (ξ) ≥ C|ξ|m for every ξ ∈ Rn \ {0}. Use the
latter and the fact that

|P(ξ)| ≥ |Pm (ξ)| − aα ξα

|α|≤m−1
to obtain the desired conclusion. Conversely, suppose that there exist finite, positive
numbers C, R such that |P(ξ)| ≥ C|ξ|m for every ξ ∈ Rn \ B(0, R) and that Pm (ξ∗ ) = 0
for some ξ∗ ∈ Rn \ {0}. Then for every λ > R/|ξ∗ | we have
m−1
0 < Cλm |ξ∗ |m ≤ |P(λξ∗ )| = aα λ|α| ξ∗α ≤ c jλ j
|α|<m j=1
and obtain a contradiction by dividing with λm and then letting λ → ∞.

Exercise 7.81 Fix ϕ ∈ C0∞ (Rn ). Using the fact that ϕ is compactly supported,
Lebesgue’s Dominated Convergence Theorem, formula (14.8.4) twice, and the fact
that A is symmetric, we may write

E A (x)LA ϕ(x) dx = lim+ E A (x)LA ϕ(x) dx
Rn ε→0 Rn \B(0,ε)

1
= lim+ (LA E A )(x)ϕ(x) dx − E A (x)(A∇ϕ(x)) · x dσ(x)
ε→0 R \B(0,ε)
n ε ∂B(0,ε)

1
+ [ϕ(x) − ϕ(0)] A∇E A (x) · x dσ(x)
ε ∂B(0,ε)

ϕ(0)
+ A∇E A (x) · x dσ(x) . (13.7.11)
ε ∂B(0,ε)
For the second equality in (13.7.11) we also used the fact that the outward unit
normal to ∂B(0, ε) is ν(x) = 1ε x, for x ∈ ∂B(0, ε).
Analyze each of the integrals in the rightmost side of (13.7.11). First, a direct
computation shows that LA E A = 0 pointwise in Rn \{0}. Second, (7.12.45), (7.12.27),
and the Mean Value Theorem for ϕ, imply

1
[ϕ(x) − ϕ(0)] A∇E A (x) · x dσ(x) ≤ Cε −−−−→ 0, (13.7.12)
ε ∂B(0,ε) ε→0+
where the constant C in (13.7.12) depends on A, ∇ϕL∞ (Rn ) and n but not on ε.

Third, thanks to (7.12.27) we have
1
E A (x)(A∇ϕ(x)) · x dσ(x)
ε ∂B(0,ε)
⎧
⎪
⎨ Cε if n ≥ 3
≤⎪ ⎩ C(| ln ε| + 1)ε if n = 2 −ε→0
−−−→
+
0. (13.7.13)
Fourth, (7.12.45) and a natural change of variables allow us to write


1 ε dσ(x)
A∇E A (x) · x dσ(x) = √
ε
n
∂B(0,ε) ωn−1 detA ∂B(0,ε) [(A−1 x)· x] 2

1 dσ(x)
= √ n . (13.7.14)
ωn−1 detA S n−1 [(A−1 x) · x] 2
Combining (13.7.11), (7.12.44), (13.7.12), (13.7.13), and (13.7.14), we conclude

that, under the current assumptions on A, for every ϕ ∈ C0∞ (Rn ) we have

ϕ(0) dσ(x)
E A (x)LA ϕ(x) dx = √ −1 x) · x] n2
. (13.7.15)
R n ωn−1 detA S n−1 [(A
In particular, (13.7.15) holds if A has real entries, is symmetric and satisfies

(7.12.8). In this latter situation, we already proved that E A is a fundamental solution
for LA , thus the left-hand side in (13.7.15) is equal to E A , LA ϕ = LA E A , ϕ =
δ, ϕ = ϕ(0), which means that (7.12.47) holds in the case when A has real entries,
is symmetric and satisfies (7.12.8). As in the proof of Theorem 7.68, the identity
(7.12.47) may be extended to the class of complex symmetric matrices satisfying
condition (7.12.8).
Based on what we have just proved and (13.7.15) we deduce that

E A (x)LA ϕ(x) dx = ϕ(0) (13.7.16)
Rn
holds for every ϕ ∈ C0∞ (Rn ) and every matrix A ∈ Mn×n (Cn ) satisfying A = A
as well as condition (7.12.8). The latter, combined with the fact that E A ∈ Lloc
1
(Rn )
further implies that E A is a fundamental solution for LA whenever A ∈ Mn×n (C) is
symmetric and satisfies (7.12.8).
Exercise 9.12 Let E be the fundamental solution for the operator ∂t − Δ x in Rn+1 as
in (8.1.8). Then E ∈ Lloc 1
(Rn+1 ) and we claim that for each x ∈ Rn \ {0}, the function
E(x, t) is absolutely integrable with respect
to t ∈ R. Indeed,

using

the change of
variables τ = |x|4t , (14.5.1), the fact that Γ n2 = n2 − 1 Γ n2 − 1 , and (14.5.6), we
2
obtain
∞ ∞
1 Γ(n/2 − 1)
e−τ τ 2 −2 dτ = − n/2 n−2
n
E(x, t) dt = − n/2 n−2
−∞ 4π |x| 0 4π |x|
1 1
=− · if x ∈ Rn \ {0}. (13.8.1)
(n − 2)ωn−1 |x|n−2
Hence, we may apply Proposition 9.6 and Proposition 9.7, to conclude that the dis-
tribution
∞
1 1
− E(x, t) dt = · n−2 , x ∈ Rn \ {0},
−∞ (n − 2)ωn−1 |x|
is a fundamental solution for Δ in Rn , when n ≥ 3.

Exercise 10.32 We have Δu + ∇p = 0 in D (Ω) n and div u = 0 in D (Ω). Apply
div to the first equation and conclude that Δp = 0 in D (Ω). Next apply Δ to the first
equation. Finally, use Theorem 7.10 and Theorem 7.23.
Exercise 10.33 Follow the outline from Exercise 10.16 with the adjustment that the
integrals in (10.3.24) this time will be bounded by C∇ϕL∞ (Rn ) ε(1 + | ln ε|).
Exercise 10.34 Follow the outline from Exercise 10.30 with the adjustment that the
integrals in (10.6.23) this time will be bounded by C∇ϕL∞ (Rn ) ε(1 + | ln ε|).
Exercise 10.35 Note that u is a solution for (10.7.4) if and only if u − Ax − b is a
solution for (10.5.3).
Exercise 10.36 Show that Theorem 10.3 applies.
Exercise 11.22 Follow the outline of the proof of Theorem 11.1.

Exercise 11.23 Compute L2 EL1 L2 using the analogue of (11.3.44) written for the
operator L := L1 L2 , and make use of a suitable version of (11.3.69). Deduce that

P(x) = cm1 ,m2 ,n,q Δ(n+q)/2 (x · ξ)2m1 +q L2 (ξ) −1 dσ(ξ) (13.10.1)
S n−1
for every x ∈ Rn , from which all the desired properties of P follow.

Exercise 11.24 Start with (11.3.44) written for L = Δm and use (7.5.24), (7.5.25),
and (7.5.13), in order to deduce that (11.7.3) holds with
(n+q)/2 q
P(x) = Pm+ n+q n+q (x) + C(n, q)Δ
2 , 2
|x| , x ∈ Rn , (13.10.2)
where C(n, q) is as in (7.5.26).


-
Exercise 12.47 Example 4.36 implies " u(ξ) = πn nj=1 χ[−1,1] (ξ j ) for all points
ξ = (ξ1 , . . . , ξn ) ∈ Rn . Use this to show that u ∈ H s (Rn ) for every s ∈ R. In
particular, by (12.1.63), we have u ∈ C ∞ (Rn ). In addition, Exercise 3.33 applied
with α = (1, . . . , 1)and β = (0, . . . , 0), gives
that if we assume u ∈ S(R ) then
n

necessarily limR→∞ sup|x|≥R x1 · · · xn · u(x) = 0. However, for each R > 0 we
have sup|x|≥R |x1 · · · xn u(x)| = 1 so the latter limits cannot be zero. This proves that
u S(Rn ).
Exercise 12.48 Consider the function u(x) := e−|x| for x ∈ Rn and use Exercise 3.3
2
and item (2) in Theorem 12.1.

Exercise 12.49 Since " δ = 1, we have δ ∈ H s (Rn ) if and only if ξ s ∈ L2 (Rn ). The
latter membership is true if and only if s < −n/2.
Exercise 12.50 Let r > 0 be such that the sets B(a j , r), j ∈ {1, . . . , N}, are pair-

wise disjoint and for each j ∈ {1, . . . , N} pick a function ψ ∈ C0∞ B(a j , r) with
N
ψ ≡ 1 on B(a j , r/2). Then u = j=1 ψ j u and it suffices to show that ψ j u ≡ 0 for
each j ∈ {1, . . . , N}. Fix an arbitrary j ∈ {1, . . . , N} and note that ψ j u ∈ H −n/2 (Rn )
(cf. item (1) in Theorem 12.7) and supp(ψ j u) ⊆ {a j }. By Exercise 2.75, we

have ψ j u = |α|≤k j cαj ∂α δa j , for some k j ∈ N0 and coefficients cαj ∈ C. Taking
the Fourier transform, applying item (2) in Theorem 4.26 and (4.2.6), we obtain
j |α| α −ia j ·ξ
)
ψ j u(ξ) = |α|≤k j cα i ξ e for all ξ ∈ Rn . The membership ψ j u ∈ H −n/2 (Rn )
implies 2
cαj i|α| ξα (1 + |ξ|2 )−n/2 ∈ L1 (Rn ). (13.11.1)
|α|≤k j

However, (13.11.1) holds if and only if |α|≤k j cαj i|α| ξα ≡ 0, thus cαj = 0 for all
|α| ≤ k j . This proves ψ j u ≡ 0.
Exercise 12.51 Pick ϕ ∈ S(Rn ). Then "
ϕ ∈ S(Rn ). By Exercise 3.13 we know that
ξ ∈ L(R ), hence item (a) in Theorem 3.14 implies ξ−t"
−t n
ϕ ∈ S(Rn ). The expres-
sion for Bt may now be written as

Bt ϕ(x) = F −1 ξ−t"
ϕ (x), ∀ x ∈ Rn . (13.11.2)
The claim in (1) now follows by invoking Theorem 3.25. In addition, identity
(13.11.2) further yields

F Bt ϕ (ξ) = ξ−t"
ϕ(ξ), ∀ ξ ∈ Rn . (13.11.3)
This and (12.1.4) then give

Bt ϕ2H s+t (Rn ) = ξ2s+2t ξ−2t |"
ϕ(ξ)|2 dξ = ϕ2H s (Rn ) , (13.11.4)
Rn
proving (2).
To show the identity in (3), let t, r ∈ R and write (making use of (13.11.3))

−n
(Bt ◦ Br )(ϕ)(x) = (2π) eix·ξ ξ−t F Br ϕ (ξ) dξ
Rn

= (2π)−n eix·ξ ξ−t ξ−r"
ϕ(ξ) dξ
Rn

= Bt+r ϕ (x), ∀ x ∈ Rn . (13.11.5)
Also, the identity in (4) is immediate from (13.11.2). Finally, since Bt is linear, given
the norm estimate in (13.11.4), and the fact that S(Rn ) is dense in H s (Rn ) (cf. (3)
in Theorem 12.1), the extension of Bt by continuity to a linear and continuous map
from H s (Rn ) into H s+t (Rn ) follows.
Exercise 12.52 For each j ∈ N we have u j ∈ C0∞ (R), hence u j ∈ H s (R) for all s ∈ R.
Also, since
sin( jx)
uj (x) = θ (x) + θ(x) cos( jx), ∀ x ∈ R, ∀ j ∈ N, (13.11.6)
j
we have

u j 2H 1 (R) = (1 + ξ2 )|"
u j (ξ)|2 dξ = u j (ξ)|2 + |u"j (ξ)|2 dξ
|"
R R

= 2π |u j (x)|2 + |uj (x)|2 dx
R

≤ 2π 3|θ(x)|2 + 2|θ (x)|2 dx ≤ 6πθ2H 1 (R) < ∞. (13.11.7)
R
This proves that the sequence {u j } j∈N is bounded in H 1 (R).

Observe that |u j | ≤ |θ|/ j for all j ∈ N, hence
H 0 (R)
u j −−−−→ 0. (13.11.8)
j→∞
Fix now s < 1. Then any subsequence {u jk }k∈N is bounded in H s (R) (since it is
bounded in H 1 (R)). By Theorem 12.8 we know that, with K := [−1, 1], the embed-
ding H 1 (R) ∩ EK (R) → H s (R) is compact for all s < 1. Thus, any subsequence
{u jk }k∈N contains a subsequence convergent in H s (R), which by Remark 12.2 and
(13.11.8) should in fact converge to zero. This ultimately implies that {u j } j∈N con-
verges to zero in H s (R).
Assume now that there exists a subsequence {u jk }k∈N which is convergent in
H 1 (R). Then, by Remark 12.2 and (13.11.8) the respective limit must be zero.
Recalling the computation in the first two equalities in (13.11.7), the latter and
(13.11.8) imply limk→∞ R |ujk (x)|2 dx = 0, and furthermore (in view of (13.11.6)),
that

0 = lim |θ(x) cos( jk x)|2 dx
k→∞ R

1 1
= θ(x) dx + lim
2
θ(x)2 cos(2 jk x) dx. (13.11.9)
2 R k→∞ 2 R

Note that R θ(x)2 cos(2 jk x) dx is equal to the real part of F (θ2 )(−2 jk ) and, accord-

ing to Proposition 3.1, it converges to zero as k → ∞. Hence, R θ(x)2 dx = 0 which
forces θ ≡ 0. The latter is in contradiction with the properties of θ. Ultimately, this
proves that {u j } j∈N has no convergent subsequence.
Exercise 12.53 If θ ≡ 0 the conclusion is obvious. If θ is nonzero, consider the
subsequence {u2k+1 }k∈N and see Exercise 12.52.
Exercise 12.54 By Exercise 12.49 we have δ ∈ H s (Rn ) for all s < −n/2 and
δ H −n/2 (Rn ). Applying Exercise 12.51, it follows that u := Bn/2+s2 δ belongs to
H s+n/2+s2 (Rn ) for all s < −n/2. Since s + n/2 + s2 < s2 , we have that u ∈ H s (Rn ) for
all s < s2 , hence u ∈ H s1 (Rn ). Also, u H s2 (Rn ). Otherwise, by Exercise 12.51 we
would have δ = B−n/2−s2 u ∈ H −n/2−s2 +s2 (Rn ) = H −n/2 (Rn ), a contradiction.
Exercise 12.55 Consider u(x) = 1, for all x ∈ Rn .

Exercise 12.56 We have |x| ∈ L2 (−1, 1) and by (a) in Exercise 2.151, we have

|x| = sgn x ∈ L2 (−1, 1) . Hence, |x| ∈ H 1 (−1, 1) . Also, by Exercise 2.151 we

have |x| = 2δ L2 (−1, 1) , thus |x| H 2 (−1, 1) .
Exercise 12.57 Pick a function ψ ∈ C0∞ (Rn ) with the property that ψ ≡ 1 near Ω.
Given any f ∈ Lip(Ω), use (2.4.5) to find F ∈ Lip(Rn ) such that f = F|Ω . Use the
distributional characterization of Lipschitzianity from Theorem 2.114 to show that
u := ψF belongs to H 1 (Rn ). Finally, observe that f = u|Ω , hence f ∈ H 1 (Ω).
Chapter 14
Appendix
Abstract The appendix contains a summary of topological and functional analysis

results in reference to the description of the topology and equivalent characteriza-
tions of convergence in spaces of test functions and in spaces of distributions. In
addition, a variety of foundational results from calculus, measure theory, and spe-
cial functions originating outside the scope of this book are included here.
14.1 Summary of Topological and Functional Analytic Results
A topology on a set X is a family τ of subsets of X satisfying the following

properties:
(1) X, ∅ ∈ τ;

(2) if {Aα }α∈I is a family of sets contained in τ, then α∈I Aα ∈ τ;
(3) if A1 , A2 ∈ τ, then A1 ∩ A2 ∈ τ.
If X and τ are as above, the pair (X, τ) is called a topological space and the
elements of τ are called open sets. If x is an element of X (sometimes referred to
as point), then any open set containing x is called an open neighborhood of x.
Any set that contains an open neighborhood of x is called a neighborhood of x. A
sequence {x j } j∈N of elements of a topological space (X, τ) are said to converge to
x ∈ X, and we write lim x j = x, if every neighborhood of x contains all but a finite
j→∞
number of the elements of the sequence.
A family B of subsets of X is a base for a topology τ on X if B is a
subfamily of τ and for each x ∈ X, and each neighborhood U of x, there is V ∈ B
such that x ∈ V ⊂ U. We say that the base B generates the topology τ. An equivalent
characterization of bases that is useful in applications is as follows. A base for a
topological space (X, τ) is a collection B of open sets in τ such that every open set
in τ can be written as a union of elements of B.
It is important to note that not any family of subsets of a given set is a base for
some topology on the set. Given a set X and a family B of subsets of X, this family
544 14 Appendix
B is a base for some topology on X if and only if

⎧
⎪
⎨ B∈B B = X,
⎪
⎪
⎪
⎩ ∀ B1 , B2 ∈ B and ∀ x ∈ B1 ∩ B2 , ∃ B ∈ B such that x ∈ B ⊆ B1 ∩ B2 .
(14.1.1)
If τ1 and τ2 are two topologies on a set X, such that every member of τ2 is also
a member of τ1 , then we say that τ1 is finer (or, larger) than τ2 , and that τ2 is
coarser (or, smaller) than τ1 . If (X, τ) is a topological space and E ⊆ X, then
the topology induced by τ on E is the topology for which all open sets are
intersections of open sets in τ with E.
Let (X, τ) and (Y, τ
) be two topological spaces. The topology on X × Y with
base the collection of products of open sets in X and open sets in Y (which satisfies
(14.1.1)) is called the product topology. A function f : X → Y is called
continuous at x ∈ X if the inverse image under f of every open neighborhood
of f (x) is an open neighborhood of x. It is called continuous on X if it is continuous
at every x ∈ X. In particular, it is easy to see that if f : X → Y is continuous then
lim f (x j ) = f (x) for every convergent sequence {x j } j∈N of X converging to x ∈ X,
j→∞
i.e., f is sequentially continuous. While in general the converse is false,
in the case when X is in fact a metric space (more on this shortly), if f : X → Y is
sequential continuous, then f is continuous (see Theorem 14.1 below).
A topological space (X, τ) is said to be separated, or Hausdorff, if for
any two distinct elements x and y of X, there exist U, a neighborhood of x and V,
a neighborhood of y, such that U ∩ V = ∅. In a Hausdorff space the limit of a
convergent sequence is unique.
A metric space is a set X equipped with a distance function (also called
metric). This is a function d : X × X → [0, ∞) satisfying the following properties:
(i) if x, y ∈ X, then d(x, y) = 0 if and only if x = y (nondegeneracy);
(ii) d(x, y) = d(y, x) for every x, y ∈ X (symmetry);
(iii) d(x, y) ≤ d(x, z) + d(z, y) for every x, y, z ∈ X (triangle inequality).
Let d be a metric on X. The open balls B(x, r) := {y ∈ X : d(x, y) < r}, x ∈ X,
r ∈ (0, ∞), are then the base of a Hausdorff topology on X (since it satisfies (14.1.1)),
denoted τd . For a sequence {x j } j∈N of points in X and x ∈ X one has lim x j = x in
j→∞
the topology τd , if and only if the sequence of numbers {d(x j , x)} j∈N converges to 0
as j → ∞. A sequence {x j } j∈N of points in (X, τd ) is called Cauchy if d(x j , xk ) → 0
as j, k → ∞. A metric space is called complete if every Cauchy sequence is
convergent.
A topological space X is called metrizable if there exists a distance func-
tion for which the open balls form a base for the topology (i.e., the topology on X
coincides with τd ). For a proof of the theorem below see [65, Theorem, p. 395].
Theorem 14.1. Let X and Y be two Hausdorff topological spaces.
(a) If Λ : X → Y is continuous, then Λ is sequentially continuous.
(b) If Λ : X → Y is sequentially continuous and X is metrizable, then Λ is continuous.
14.1 Summary of Topological and Functional Analytic Results 545
Moving on, recall that a vector space X over C becomes a topological

vector space if equipped with a topology that is compatible with its vector
structure, that is, the operation of vector addition and of multiplication by a com-
plex number are continuous maps X × X → X and C × X → X, respectively (where
X × X and C × X are endowed each with the corresponding product topology). Ob-
serve that in order to specify a topology on a vector space, it suffices to give
the system of neighborhoods of the zero element 0 ∈ X, since the system of neigh-
borhoods of any other element of X is obtained from this via translation. In fact, it
suffices to give a base of neighborhoods of 0. This is a family B of neighborhoods
of 0 such that every neighborhood of 0 contains a member of B. A set E ⊆ X is then
open if and only if, for every x ∈ E, there exists U ∈ B such that x + U ⊆ E, where
x + U := {x + y : y ∈ U}.
A topological vector space is called locally convex if it has a base of neigh-
borhoods of 0 consisting of sets that are convex, balanced, and absorbing. A set U
is called convex provided tx + (1 − t)y ∈ U whenever x, y ∈ U and t ∈ [0, 1].
A set U is called balanced if cx ∈ U whenever x ∈ U, c ∈ C and |c| ≤ 1. A
set U is called absorbing provided for each x ∈ X there exists t > 0 such that
x ∈ tU := {ty : y ∈ U}.
A seminorm on a vector space X is a function p : X → R satisfying the
properties
(1) p(cx) = |c|p(x), for every c ∈ C and every x ∈ X (positively homogeneous);
(2) p(x + y) ≤ p(x) + p(y), for every x, y ∈ X (sub-additive).
In particular, a seminorm p on X satisfies p(0) = 0 and p(x) ≥ 0 for every x ∈ X.
A family P of seminorms on X is called separating if, for each x ∈ X, x 0,
there exists p ∈ P such that p(x) 0. Given a separating family of seminorms P on
X, let B be the collection of sets of the form

x : p(x) < ε ∀ p ∈ P0 , P0 ⊆ P, P0 finite, ε > 0. (14.1.2)
Then B is a base of neighborhoods of 0 of a locally convex vector space topology

τP on X called the topology generated by the family of seminorms
P. In this context, it may be readily verified that if Y is a linear subspace of X and
P|Y := {p|Y : p ∈ P}, then
the topology induced on Y by τP coincides with τP|Y . (14.1.3)
Conversely, if X is a locally convex topological vector space, for each U convex,

balanced, and absorbing neighborhood of 0, the mapping pU : X → R defined by
pU (x) := inf{t > 0 : t−1 x ∈ U}, x ∈ X (called the Minkowski functional associated
with U) is a seminorm. It is then not hard to see that the topology on X is generated
by this family of seminorms (in the manner described above).
Let P = {p j } j∈N be a countable family of seminorms that is also separating (thus,
if x ∈ X and p j (x) = 0 for all j ∈ N, then x = 0). The topology generated by this
family is metrizable. Indeed, the function d : X × X → R defined by
546 14 Appendix

∞
p j (x − y)
d(x, y) := 2− j , for each x, y ∈ X, (14.1.4)
j=1
1 + p j (x − y)
is a distance on X and the topology τd induced by the metric d coincides with the
topology generated by P. In the converse direction, it can be shown that the topology
of a locally convex space that is metrizable, endowed with a translation invariant
metric, can be generated by a countable family of seminorms.
A locally convex topological vector space that is metrizable and complete is
called a Fréchet space. Thus, if a family P = {p j } j∈N of seminorms generates
the topology of a Fréchet space X, then whenever p j (xk − xl ) → 0 as k, l → ∞ for
every j ∈ N, there exists x ∈ X such that p j (xk − x) → 0 as k → ∞ for every j ∈ N.
Let X be a vector space, {X j } j∈J be a family of vector subspaces of X such that

X= X j and if j1 , j2 ∈ J then there exists j3 ∈ J with X j1 ⊂ X j3 and X j2 ⊂ X j3 .
j∈J
Also assume that there exist topologies τ j on X j such that if X j1 ⊂ X j2 then the
topology τ j1 is finer than the topology τ j2 j1 induced on X j1 by X j2 . Let
W := {W ⊂ X : W balanced, convex and such that W ∩ X j

is a neighborhood of 0 in X j , ∀ j ∈ N}. (14.1.5)
Then W is a base of neighborhoods of 0 in a locally convex topology τ on X. Call

this topology the inductive limit topology on X.
If in addition whenever X j1 ⊂ X j2 we also have τ j2 j1 = τ j1 , we call the inductive
limit strict. If the topology τ on X is the strict inductive limit of the topologies of an
increasing sequence {Xn }n∈N , then the topology induced on Xn by the topology τ on
X coincides with the initial topology τn on Xn , for every n ∈ N.
In general, if X is a topological vector space, its dual space is the collection of
all linear mappings f : X → C (also referred to as functionals) that are continuous
with respect to the topology on X. In the case when X is a locally convex topological
vector space and P is a family of seminorms generating the topology of X, then
a seminorm q on X is continuous if and only if there exist N ∈ N, seminorms
p1 , ..., pN ∈ P, and a constant C ∈ (0, ∞), such that
|q(x)| ≤ C max {p1 (x), . . . , pn (x)}, ∀ x ∈ X. (14.1.6)
This fact then gives a criterion for continuity for functionals on X, since if f : X →
C is a linear mapping, then q(x) := | f (x)|, for x ∈ X, is a seminorm on X. In addition,
if the family of seminorms P has the property that for any p1 , p2 ∈ P there exists
p3 ∈ P with the property that max{p1 (x), p2 (x)} ≤ p3 (x) for every x ∈ X, then a
linear functional f : X → C is continuous if and only if there exist a seminorm
p ∈ P and a constant C ∈ (0, ∞), such that
| f (x)| ≤ C p(x), ∀ x ∈ X. (14.1.7)

Given a nonempty set X and a family F of mappings f : X → C, denote by τF

the collection of all unions of finite intersections of sets f −1 (V), for f ∈ F and V
open set in C. Then τF is a topology on X, and is the weakest topology on X that
makes every f ∈ F continuous. We will refer to it as the F -topology on X.
Let X be a vector space, F be a separating vector space of linear functionals on
X, and τF be the F -topology on X. Then (X, τF ) is a locally convex topological
space and its dual is F . In particular, a sequence {x j } j in X satisfies x j → 0 in τF as
j → ∞ if and only if f (x j ) → 0 as j → ∞ for every f ∈ F .
If X is a topological vector space and X
denotes its dual space, then every x ∈ X
induces a linear functional F x on X
defined by F x (Λ) := Λ(x) for every Λ ∈ X
and if we set F := {F x : x ∈ X}, then F separates points in X

. In particular,
the F -topology on X
, called the weak∗-topology on X
, is locally convex and
every linear functional on X
that is continuous with respect to the weak∗-topology
is precisely of the form F x , for some x ∈ X. Moreover, a sequence {Λ j } j ⊂ X
converges to some Λ ∈ X
, in the weak∗-topology on X
, if and only if Λ j (x) → Λ(x)
as j → ∞ for every x ∈ X.
An inspection of the definition of the weak∗-topology yields a description of the
open sets in this topology. More precisely, if X is a topological vector space and for
A ⊆ X and ε ∈ (0, ∞) we set

OA,ε := f ∈ X
: | f (x)| < ε, ∀ x ∈ A , (14.1.8)
then the following equivalence holds:
O ⊆ X
is a weak∗-open neighborhood of 0 ∈ X
⇐⇒ there exist a set I,

A j ⊆ X finite and ε j > 0 for each j ∈ I, such that O = OA j ,ε j .
j∈I
(14.1.9)
The transpose of any linear and continuous operator between two topological
vector spaces is always continuous at the level of dual spaces equipped with weak∗-
topologies.
Proposition 14.2. Assume X and Y are two given topological vector spaces, and
denote by X
, Y
their duals, each endowed with the corresponding weak∗-topology.
Also, suppose T : X → Y is a linear and continuous operator, and define its trans-
pose T t as the mapping
T t : Y
→ X
, T t (y
) := y
◦ T for each y
∈ Y
. (14.1.10)
Then T t is well-defined, linear, and continuous.

Proof. For each y
∈ Y
it follows that y
◦ T is a composition of two linear and
continuous mappings. Hence, y
◦ T ∈ X
which proves that T t : Y
→ X
is well-
defined. It is also clear from (14.1.10) that T t is linear. There remains to prove that
T t is continuous. By linearity it suffices to check that T t is continuous at 0. With this
goal in mind, fix an arbitrary finite subset A of X along with an arbitrary number
ε ∈ (0, ∞), and define
548 14 Appendix
OA,ε := {x
∈ X
: |x
(x)| < ε, ∀ x ∈ A}. (14.1.11)

:= {T x : x ∈ A} and set
Furthermore, introduce A

:= {y
∈ Y
: |y
(y)| < ε, ∀ y ∈ A}.
O
(14.1.12)
A,ε
Then T t (O
) ⊆ OA,ε . Invoking the description from (14.1.9) we may conclude that

A,ε
T t is continuous at 0, and this finishes the proof of the proposition.
Proposition 14.3. Suppose that X, Y, Z are topological vector spaces, and denote
by X
, Y
, Z
their duals, each endowed with the corresponding weak∗-topology. In
addition, assume that T : X → Y and R : Y → Z are two linear and continuous
operators. Then
(R ◦ T )t = T t ◦ Rt . (14.1.13)
In particular, if T : X → Y is a linear, continuous, bijective map, with continuous
inverse T −1 : Y → X, then T t : Y
→ X
is also bijective and has a continuous
inverse (T t )−1 : X
→ Y
that satisfies (T t )−1 = (T −1 )t .
Proof. Formula (14.1.13) is immediate from definitions, while the claims in the
last part of the statement are direct consequences of (14.1.13) and the fact that the
transpose of the identity is also the identity.
We also state and prove an embedding result at the level of dual spaces endowed
with the weak∗-topology.
Proposition 14.4. Suppose X and Y are topological vector spaces such that X ⊆ Y
densely and the inclusion map ι : X → Y, ι(x) := x for each x ∈ X, is continuous.
Then Y
endowed with the weak∗-topology embeds continuously in the space X
endowed with the weak∗-topology, in the sense that the mapping
ιt : Y
−→ X
(14.1.14)
is well-defined, linear, injective, and continuous. Under the identification of Y

with
ιt (Y
) ⊆ X
, we therefore have
Y
→ X
. (14.1.15)
Proof. The fact that ιt is well-defined, linear, and continuous follows directly from
Proposition 14.2 and assumptions. Assume now that y
∈ Y
is such that t
ι (y ) = 0.

Then from the fact that y ◦ ι : X → C is zero we deduce that y X = 0. Since
y
: Y → C is continuous and X is dense in Y, we necessarily have that y
vanishes
on Y, forcing y
= 0 in Y
. Keeping in mind that ιt is linear, this implies that ιt is
injective.
Theorem 14.5 (Hahn–Banach Theorem). Let X be a vector space (over complex

numbers) and suppose p : X → [0, ∞) is a seminorm on X. Also assume that Y is a
linear subspace of X and that φ : Y → C is a linear functional which is dominated
by p on Y, i.e.,
|φ(y)| ≤ p(y), ∀ y ∈ Y. (14.1.16)

Then there exists a linear functional Φ : X → C satisfying
Φ(y) = φ(y) ∀ y ∈ Y, |Φ(x)| ≤ p(x) ∀ x ∈ X. (14.1.17)
In the next installment we shall specialize the above considerations to various

specific settings used in the book. The reader is reminded that Ω ⊆ Rn denotes a
fixed, nonempty, arbitrary open set.
14.1.0.1 The Topological Vector Space E(Ω)
By τ we denote the topology on C ∞ (Ω) generated by the following family of semi-

norms:
pK,m : C ∞ (Ω) → R, pK,m (ϕ) := sup |∂α ϕ(x)|, ∀ ϕ ∈ C ∞ (Ω), (14.1.18)

x∈K
α∈Nn0 , |α|≤m
where K ⊂ Ω is a compact set and m ∈ N0 . Consequently, a sequence ϕ j ∈ C ∞ (Ω),

j ∈ N, converges in τ to a function ϕ ∈ C ∞ (Ω) as j → ∞, if and only if for any
compact set K ⊂ Ω and any m ∈ N0 one has

lim sup sup ∂α (ϕ j − ϕ)(x) = 0. (14.1.19)
j→∞ α∈Nn , |α|≤m x∈K
0
We will use the notation

E(Ω) = C ∞ (Ω), τ . (14.1.20)
The space E(Ω) is locally convex and metrizable since its topology is defined by the
family of countable seminorms {pKm ,m }m∈N0 where

∞
Km ⊂ K̊m+1 ⊂ Ω, Km is compact for each m ∈ N0 and Ω = Km . (14.1.21)
m=0
In addition, τ is independent of the family {Km }m∈N0 with the above properties and
E(Ω) is complete. Thus, E(Ω) is a Frechét space.
14.1.0.2 The Topological Vector Space E (Ω)
Based on the discussion on dual spaces for locally convex topological vector spaces
(see (14.1.7) and the remarks preceding it), it follows that the dual space of E(Ω) is
the collection of all linear functionals u : E(Ω) → C for which there exist m ∈ N, a
compact set K ⊂ Rn with K ⊂ Ω, and a constant C > 0 such that
|u(ϕ)| ≤ C sup sup |∂α (ϕ)(x)| , ∀ ϕ ∈ C ∞ (Ω). (14.1.22)

550 14 Appendix
This dual space will be endowed with the weak∗-topology induced by E(Ω) and
we denote this topological space by E
(Ω). Hence, if for each ϕ ∈ C ∞ (Ω) we con-
sider the evaluation mapping Fϕ taking any functional u from the dual of E(Ω) into
the number Fϕ (u) := u(ϕ) ∈ C, then the family F := {Fϕ : ϕ ∈ E(Ω)} separates
points in the dual of E(Ω), and the weak∗-topology on this dual is the F -topology
on it. In particular, if {u j } j∈N is a sequence in E
(Ω) and u ∈ E
(Ω), then
u j → u in E
(Ω) as j → ∞ ⇐⇒
(14.1.23)
u j (ϕ) → u(ϕ) as j → ∞, ∀ ϕ ∈ C ∞ (Ω).
Moreover, a sequence {u j } j∈N in E

(Ω) is Cauchy provided lim (u j − uk )(ϕ) = 0 for
j,k→∞
every ϕ ∈ C ∞ (Ω) and the weak∗-topology on the dual of E(Ω) is locally convex and
complete.
14.1.0.3 The Topological Vector Space D K (Ω)
Let K ⊆ Ω be a compact set in Rn . Denote by DK (Ω) the topological vector space of

functions { f ∈ C ∞ (Ω) : supp f ⊆ K} with the topology induced by τ, the topology
in E(Ω). Then the topology on DK (Ω) is generated by the family of seminorms
⎧
⎪
⎪
⎪ pm : DK (Ω) → R,
⎪
⎪
⎨
⎪
⎪
⎪ pm (ϕ) := sup |∂α ϕ(x)|, ∀ ϕ ∈ C ∞ (Ω), supp ϕ ⊆ K, (14.1.24)
⎪
⎪
⎩ x∈K
α∈Nn , |α|≤m
0
where m ∈ N0 . Hence, DK (Ω) is a Fréchet space. In addition, a linear map u :

DK (Ω) → C is continuous if and only if there exist m ∈ N and a constant C > 0,
both depending on K, such that
|u(ϕ)| ≤ C sup sup |∂α (ϕ)(x)| , ∀ ϕ ∈ C ∞ (Ω), supp ϕ ⊆ K. (14.1.25)

14.1.0.4 The Topological Vector Space D(Ω)
The topological vector space on C0∞ (Ω) endowed with the inductive limit topology
of the Frechét spaces DK (Ω) will be denoted by D(Ω). In this setting, we have
ϕj → ϕ in D(Ω) as j → ∞
⎧
⎪
⎪
⎨∃ K ⊆ Ω compact set such that ϕ j ∈ DK (Ω), ∀ j, and
⎪
⇐⇒ ⎪⎪ (14.1.26)
⎪
⎩ϕ j →
− ϕ in DK (Ω) as j → ∞.
The topology D(Ω) is locally convex and complete but not metrizable (thus not
Fréchet) and is the strict inductive limit of the topologies {DK j (Ω)} j∈N , where the
family {K j } j∈N is as in (14.1.21).
We also record an important result that is proved in [65, Theorem 6.6, p. 155].
Theorem 14.6. Let X be a locally convex topological vector space and suppose the
map Λ : D(Ω) → X is linear. Then Λ is continuous if and only if for every sequence
D(Ω)
{ϕ j } j∈N in C0∞ (Ω) satisfying ϕ j −−−−→ 0 we have lim Λ(ϕ j ) = 0 in X.
j→∞ j→∞
14.1.0.5 The Topological Vector Space D (Ω)
The dual space of D(Ω) endowed with the weak∗-topology is denoted by D

(Ω).
Hence, if {u j } j∈N is a sequence in D
(Ω) and u ∈ D
(Ω), then
u j → u in D
(Ω) as j → ∞ ⇐⇒
(14.1.27)
u j (ϕ) → u(ϕ) as j → ∞, ∀ ϕ ∈ C0∞ (Ω).
In addition, a sequence {u j } j∈N in D

(Ω) is called Cauchy provided
lim (u j − uk )(ϕ) = 0 for every ϕ ∈ C0∞ (Ω). (14.1.28)

j,k→∞
The weak∗-topology on the dual of D(Ω) is locally convex and complete and an
inspection of this topology reveals that it coincides with the topology defined by the
family of seminorms
D
(Ω) u → max |u(ϕ j )|, m ∈ N, ϕ1 , . . . , ϕm ∈ D(Ω). (14.1.29)
1≤ j≤m
14.1.0.6 The Topological Vector Space S(R n)
The Schwartz class of rapidly decreasing functions is the vector space

S(Rn ) := ϕ ∈ C ∞ (Rn ) : ∀ α, β ∈ Nn0 , sup |xβ ∂α ϕ(x)| < ∞ , (14.1.30)
x∈Rn
endowed with the topology generated by the family of seminorms {pk,m }k,m∈N0 de-
fined by
pk,m : S(Rn ) → R,
pk,m (ϕ) := sup |xβ ∂α ϕ(x)|, ∀ ϕ ∈ S(Rn ). (14.1.31)
n
x∈R
|α|≤m, |β|≤k
Hence, the topology generated by the family of seminorms {pk,m }k,m∈N0 on S(Rn ) is
locally convex, metrizable, and since it is also complete, the space S(Rn ) is Frechét.
Moreover, a sequence ϕ j ∈ S(Rn ), j ∈ N, converges in S(Rn ) to a function ϕ ∈
552 14 Appendix
S(Rn ) as j → ∞, if and only if for every m, k ∈ N0 one has

lim sup xβ ∂α (ϕ j − ϕ)(x) = 0. (14.1.32)
j→∞ x∈Rn
|α|≤m, |β|≤k
14.1.0.7 The Topological Vector Space S (R n)
By the discussion about dual spaces for locally convex topological vector spaces,
it follows that the dual space of S(Rn ) is the collection of all linear functions u :
S(Rn ) → C for which there exist m, k ∈ N0 , and a finite constant C > 0, such that

|u(ϕ)| ≤ C sup sup xβ ∂α ϕ(x) , ∀ ϕ ∈ S(Rn ). (14.1.33)
α,β∈Nn0 , |α|≤m, |β|≤k x∈Rn
We endow the dual of S(Rn ) with the weak∗-topology and denote the resulting
locally convex topological vector space by S
(Rn ). Hence, if {u j } j∈N is a sequence
in S
(Rn ) and u ∈ S
(Rn ), then
u j → u in S
(Rn ) as j → ∞ ⇐⇒
(14.1.34)
u j (ϕ) → u(ϕ) as j → ∞, ∀ ϕ ∈ S(Rn ).
Also, a sequence {u j } j∈N in S

(Rn ) is called Cauchy provided
lim (u j − uk )(ϕ) = 0 for every ϕ ∈ S(Rn ). (14.1.35)

j,k→∞
14.2 Basic Results from Calculus, Measure Theory, and

Topology
The next lemma contains a discrete change of variable formula (for sums). To state
it, given a set E we denote by #E its cardinality.
Lemma 14.7. Let F : D → R be a function between finite sets. If (G, +) is an

Abelian group, then for every function G : R → G one has

G(F (a)) = #F −1 ({b}) G(b), (14.2.1)
a∈D b∈R
where F −1 ({b}) := {a ∈ D : F (a) = b} for each b ∈ R.

As corollary, under the additional assumption that F : D → R is a bijection,

G(F (a)) = G(b). (14.2.2)
a∈D b∈R
Proposition 14.8 (Multinomial Theorem). If x = (x1 , . . . , xn ) ∈ Rn and N ∈ N are

arbitrary, then
14.2 Basic Results from Calculus, Measure Theory, and Topology 553

n N N!
xj = xα . (14.2.3)
j=1 |α|=N
α!
Theorem 14.9 (Binomial Theorem). For any x, y ∈ Cn and any γ ∈ Nn0 we have
γ! α β
(x + y)γ = x y , (14.2.4)
α+β=γ
α!β!
(with the convention that z0 := 1 for each z ∈ C).
In the particular case when x = (1, . . . , 1) ∈ Cn and y = (−1, . . . , −1) ∈ Cn ,

formula (14.2.4) yields
γ!
0= (−1)|β| , ∀ γ ∈ Nn0 with |γ| > 0. (14.2.5)
α+β=γ
α!β!
Proposition 14.10 (Leibniz’s Formula). Suppose that U ⊆ Rn is an open set,

N ∈ N, and f, g : U → C are two functions of class C N in U. Then
α!
∂α ( f g) = (∂β f )(∂α−β g) in U, (14.2.6)
β≤α
β!(α − β)!
for every multi-index α ∈ Nn0 of length ≤ N.
It is useful to note that for each α, β ∈ Nn0 and x ∈ Rn ,

⎧ α!
⎪
⎪
⎪ α−β
if β ≤ α,
⎨ (α − β)! x
∂β (xα ) = ⎪
⎪ (14.2.7)
⎪
⎩0 otherwise.
Theorem 14.11 (Taylor’s Formula in dimension one). For every function f be-
longing to C0∞ (R) and every m ∈ N the following formula holds:
∞
(−1)m dm
f (0) = tm−1 m [ f (t)] dt. (14.2.8)
(m − 1)! 0 dt
Theorem 14.12 (Taylor’s Formula). Assume U ⊆ Rn is an open convex set, and

that N ∈ N. Also, suppose that f : U → C is a function of class C N+1 on U. Then
for every x, y ∈ U one has
1
f (x) = (x − y)α (∂α f )(y) (14.2.9)
|α|≤N
α!
N+1 1
+ (1 − t)N (x − y)α (∂α f )(tx + (1 − t)y) dt.
|α|=N+1
α! 0
In particular, for each x, y ∈ U there exists θ ∈ (0, 1) with the property that
554 14 Appendix
1
f (x) = (x − y)α (∂α f )(y) (14.2.10)
|α|≤N
α!
1
+ (x − y)α (∂α f )(θx + (1 − θ)y).
|α|=N+1
α!
Theorem 14.13 (Rademacher’s Theorem). If f : Rn → R is a Lipschitz function

with Lipschitz constant less than or equal to M, for some M ∈ (0, ∞), then f is
differentiable almost everywhere and ∂k f L∞ (Rn ) ≤ M for each k = 1, . . . , n.
Theorem 14.14 (Lebesgue’s Differentiation Theorem). If f ∈ Lloc 1

(Rn ), then

1
lim | f (y) − f (x)| dy = 0 for almost every x ∈ Rn . (14.2.11)
ε→0+ |B(x, ε)| B(x,ε)
In particular,

1
lim+ f (y) dy = f (x) for almost every x ∈ Rn . (14.2.12)
ε→0 |B(x, ε)| B(x,ε)
Theorem 14.15 (Lebesgue’s Dominated Convergence Theorem). Let (X, μ) be a

positive measure space and assume that g ∈ L1 (X, μ) is a nonnegative function. If
{ f j } j∈N is a sequence of μ-measurable, complex-valued functions on X, such that
| f j (x)| ≤ g(x) for μ-almost every x ∈ X and f (x) := lim f j (x) exists (in C) for
j→∞
μ-almost every x ∈ X, then f ∈ L1 (X, μ) and lim X | f j − f | dμ = 0. In particular,
j→∞
lim X f j dμ = X f dμ.
j→∞
The same circle of ideas that are used to prove Theorem 14.15 also yields a
corresponding L p -version which is stated next.
Theorem 14.16 (L p version of Lebesgue’s Dominated Convergence Theorem).

Let (X, μ) be a positive measure space and assume that g ∈ L p (X, μ) for some p ∈
[1, ∞) is a nonnegative function. If { f j } j∈N is a sequence of μ-measurable, complex-
valued functions on X, such that | f j (x)| ≤ g(x) for μ-almost every x ∈ X and f (x) :=
lim f j (x) exists (in C) for μ-almost every x ∈ X, then f ∈ L p (X, μ) and lim X | f j −
j→∞ j→∞
f | p dμ = 0. In particular, lim X | f j | p dμ = X | f | p dμ.
j→∞
One can prove Theorem 14.16 by applying Fatou’s lemma to the sequence of
functions {2 p−1 g p − | f j − f | p } j∈N .
Theorem 14.17 (Young’s Inequality). Assume that 1 ≤ p, q, r ≤ ∞ are such that

p + q = r + 1. Then for every f ∈ L (R ), g ∈ L (R ) it follows that
1 1 1 p n q n

| f (x − y)||g(y)| dy < ∞ for a.e. x ∈ Rn , (14.2.13)
Rn
the function defined at a.e. x ∈ Rn by

( f ∗ g)(x) := f (x − y)g(y) dy (14.2.14)
Rn
belongs to Lr (Rn ), and f ∗ gLr (Rn ) ≤ f L p (Rn ) gLq (Rn ) .
We single out two corollaries of Theorem 14.17 obtained by taking particular

values for q and r. Specifically, if we take q = 1 = r, then Theorem 14.17 implies
L p (Rn ) ∗ L1 (Rn ) → L1 (Rn ) for all p ∈ [1, ∞], (14.2.15)

p
and if q = p−1 , r = ∞, we obtain

1 1
L p (Rn ) ∗ L p (Rn ) → L∞ (Rn ) if p, p
∈ [1, ∞], + = 1. (14.2.16)
p p
In turn, (14.2.15) readily yields the following result.

Lemma 14.18. Let θ ∈ L1 (Rn ) and for each ε > 0 define θε (x) := ε−n θ(x/ε) for
x ∈ Rn . Then for each p ∈ [1, ∞] and each f ∈ L p (Rn ), the functions fε := f ∗ θε ,
ε > 0, satisfy fε ∈ L p (Rn ) for each ε > 0 and
sup fε L p (Rn ) ≤ θL1 (Rn ) f L p (Rn ) . (14.2.17)

ε>0
Lemma 14.18 is an ingredient in the proof of the following basic approximation

result.

Lemma 14.19. Let θ ∈ Lcomp 1
(Rn ) be such that Rn θ dx = 1. For each ε > 0 set
θε (x) := ε−n θ(x/ε) for x ∈ Rn . Consider f ∈ L p (Rn ) with p ∈ [1, ∞) and define the
sequence of functions fε := f ∗ θε , ε > 0. Then fε ∈ L p (Rn ) for each ε > 0 and
fε → f in L p (Rn ) as ε → 0+ .
Proof. First we show that the conclusion of the lemma holds if f ∈ C00 (Rn ). To this
end pick R > 0 with supp θ ⊆ B(0, R). Then
supp θε ⊆ B(0, εR) ⊆ B(0, R) (14.2.18)
for every ε ∈ (0, 1) and if we set K := supp f + B(0, R) then K is compact and
supp f ⊆ K, supp fε ⊆ K for all ε ∈ (0, 1). (14.2.19)
Let δ > 0 be arbitrary. From the uniform continuity of f it follows that there exists
η > 0 such that if x, y ∈ Rn are such that |x − y| < η then | f (x) − f (y)| < δ. Hence, if
0 < ε < Rη and we use the fact that Rn θε dx = 1, for each x ∈ Rn we may write
556 14 Appendix

| fε (x) − f (x)| ≤ | f (y) − f (x)||θε (x − y)| dy
B(x,εR)
≤ δθL1 (Rn ) . (14.2.20)
This proves that fε → f in L∞ (Rn ) as ε → 0+ . The latter combined with (14.2.19)

readily implies fε → f in L p (Rn ) as ε → 0+ , as wanted.
Suppose now that f ∈ L p (Rn ) for some p ∈ [1, ∞) and let δ > 0 be arbitrary.
Since C00 (Rn ) is dense in L p (Rn ) we can find g ∈ C00 (Rn ) with the property that
f − gL p (Rn ) < δ/3. Moreover, based on what we proved earlier, there exists ε0 > 0
such that g − gε L p (Rn ) < δ/3 if ε ∈ (0, ε0 ). Hence, if ε ∈ (0, ε0 ) we may estimate
f − fε L p (Rn ) ≤ f − gL p (Rn ) + g − gε L p (Rn ) + (g − f )ε L p (Rn )

δ δ
< + + g − f L p (Rn ) < δ, (14.2.21)
3 3
where for the second inequality in (14.2.21) we have also used (14.2.17). This fin-
ishes the proof of the lemma.
Lemma 14.20. Let a, b ∈ R satisfying a < b. Then for every real-valued function

f ∈ C 0 [a, b] ∩ Lip loc (a, b) one has
b
| f (b) − f (a)| ≤ | f
(t)| dt. (14.2.22)
a

Proof. Assume first that (14.2.22) holds for functions in Lip [a, b] and pick some

f ∈ C 0 [a, b] ∩Lip loc (a, b) . Then for each ε ∈ 0, (b−a)/4 the function f belongs

to Lip [a + ε, b − ε] and, by the current assumption,
b−ε
| f (b − ε) − f (a + ε)| ≤ | f
(t)| dt. (14.2.23)
a+ε
That f satisfies estimate (14.2.22) is now seen by passing to the limit in (14.2.23)
as ε → 0+ and using the continuity of f combined with the Lebesgue Monotone
Convergence Theorem.

We are left with proving that (14.2.22) holds for each function f ∈ Lip [a, b] .
Fix such an f and observe that without loss of generality we may assume that ac-
tually f ∈ Lipcomp (R). Indeed, a classical result (due to E.J. McShane [51]) states
that
any real-valued Lipschitz function defined on a subset of the

Euclidean space may be extended to a Lipschitz function on the (14.2.24)
entire Euclidean space with preservation of Lipschitz constant.
Multiplying the extension of f provided by (14.2.24) with a smooth compactly sup-

ported function that is identically one near [a, b] then yields an extension of f in
the space Lipcomp (R). In particular, we have that f
∈ Lcomp
∞
(Rn ). Let θ ∈ C0∞ (R)

be even, positive, supported in [−1, 1] with R θ dx = 1 (cf. Lemma 14.32) and
set θε (x) := ε−1 θ(x/ε), x ∈ R for each ε > 0. Then for each ε > 0 the function
fε := f ∗ θε belongs to C0∞ (R), satisfies ( fε )
:= f
∗ θε and

( fε ) (t) ≤ | f (s)|θε (t − s) ds, ∀ t ∈ R. (14.2.25)

R
Consequently
b
b
( fε ) (t) dt ≤ | f
(s)|θε (t − s) ds dt (14.2.26)
a a R
b
= | f
(s)| θε (t − s) dt ds
R a

= | f
(s)| 1[a,b] ∗ θε (s) ds.
R
Lemma 14.19 applied with n := 1, f := 1[a,b] , and p := 1 implies
1[a,b] ∗ θε → 1[a,b] in L1 (R) as ε → 0+ . (14.2.27)
This and the fact that f

∈ Lcomp
∞
(Rn ) allow us to take the limit in (14.2.26) as
+
ε → 0 to obtain
b b
lim+

( fε ) (t) dt ≤ | f
(s)| ds. (14.2.28)
ε→0 a a
b
In addition, the Fundamental Theorem of Calculus implies fε (b)− fε (a) = a ( fε )
(t) dt
for each ε > 0 hence
b

fε (b) − fε (a) ≤ ( fε ) (t) dt, (14.2.29)
a
for each ε > 0. Since fε converges to f uniformly on compact subsets of R as
ε → 0+ (this may be seen as in the proof of (2.1.32)) taking the limit in (14.2.29) as
ε → 0+ gives
b
| f (b) − f (a)| ≤ lim ( f )
(t) dt. (14.2.30)
ε
ε→0+ a
Combining (14.2.30) and (14.2.28) it is immediate that f satisfies (14.2.22) as
wanted.
Theorem 14.21 (Hardy’s Inequality). Suppose p ∈ [1, ∞), r ∈ (0, ∞), and con-
sider a measurable function f : [0, ∞] −→ [0, ∞]. Then
∞ ρ p p p ∞
ρ−1−r f (θ) dθ dρ ≤ f (θ) p θ p−r−1 dθ. (14.2.31)
0 0 r 0
See e.g., [68, p. 272, A.4].

558 14 Appendix
Theorem 14.22 (Minkowski’s Inequality). Suppose (X, M, μ) and (Y, N, ν) are σ-

finite measure spaces, and let f be an (M ⊗ N)-measurable function on X × Y. If
f ≥ 0 and p ∈ [1, ∞), then
p 1/p 1/p
f (x, y) dν(y) dμ(x) ≤ f (x, y) p dμ(x) dν(y). (14.2.32)
X Y Y X
For a proof of Minkowski’s Inequality see [18, 6.19, p. 194].
Definition 14.23. A rigid transformation, or isometry, of the Euclidean

space Rn is any distance preserving mapping of Rn , i.e., any function T : Rn → Rn
satisfying
|T (x) − T (y)| = |x − y|, ∀ x, y ∈ Rn . (14.2.33)
A rigid transformation of Rn is any distance preserving mapping of Rn , i.e., any

function T : Rn → Rn satisfying |T x − T y| = |x − y| for every x, y ∈ Rn . The
rigid transformations of the Euclidean space Rn are precisely those obtained by
composing a translation with a mapping in Rn given by an orthogonal matrix. In
other words, a mapping T : Rn → Rn is a rigid transformation of Rn if and only if
there exist x0 ∈ Rn and an orthogonal matrix A : Rn → Rn with the property that
T (x) = x0 + Ax, ∀ x ∈ Rn . (14.2.34)
For a proof of the following version of the Arzelà–Ascoli theorem see [63, Corol-
lary 34, p. 179].
Theorem 14.24 (Arzelà–Ascoli’s Theorem). Let F be an equicontinuous family
of real-valued functions on a sepa0rable space X. Then each sequence { f j } j∈N in F
which is bounded at each point has a subsequence { f jk }k∈N that converges pointwise
to a continuous function, the converges being uniform on each compact subset of X.
Theorem 14.25 (Riesz’s Representation Theorem for Positive Functionals). Let

X be a locally compact Hausdorff topological space and Λ a positive linear func-
tional on the space of continuous, compactly supported functions on X (denoted by
C00 (X)). Then there exists a unique σ-algebra M on X, which contains all Borel sets
on X, and a unique measure μ : M → [0, ∞] that represents Λ, that is, the following
hold:

(i) Λ f = X f dμ for every continuous, compactly supported function f on X;
(ii) μ(K) < ∞ for every compact K ⊂ X;
(iii) for every E ∈ M we have μ(E) = inf{μ(V) : E ⊂ V, V open};
(iv) μ(E) = sup {μ(K) : K ⊂ E, K compact} for every open set E and every E ∈ M
with μ(E) < ∞;
(v) if E ∈ M, A ⊂ E, and μ(E) = 0, then μ(A) = 0.
Theorem 14.26 (Riesz’s Representation Theorem for Complex Functionals).

Let X be a locally compact Hausdorff topological space and consider the space of
continuous functions on X vanishing at infinity, i.e.,
Coo (X) := { f ∈ C 0 (X) : ∀ ε >0, ∃ compact K ⊂ X such that
| f (x)| < ε for x ∈ X \ K}. (14.2.35)
Then Coo (X) is the closure in the uniform norm of C00 (X) and for every bounded
linear functional Λ : Coo (X) → C there exists a unique regular complex Borel
measure μ on X such that Λ f = X f dμ for every f ∈ Coo (X) and Λ = |μ|(X).
Theorem 14.27 (Riesz’s Representation Theorem for Locally Bounded Func-
tionals). Let X be a locally compact Hausdorff topological space and assume that
Λ : C00 (X) → R is a linear functional that is locally bounded, in sense that for each
compact set K ⊂ X there exists a constant C K ∈ (0, ∞) such that
|Λ f | ≤ C K sup | f (x)|, ∀ f ∈ C00 (X) with supp f ⊆ K. (14.2.36)

x∈K
Then there exist two measures μ1 , μ2 , taking Borel sets from X into [0, ∞], and sat-
isfying properties (ii)-(iv) in Theorem 14.25, such that

Λf = f dμ1 − f dμ2 for every f ∈ C00 (X). (14.2.37)
X X
The reader is warned that since both μ1 and μ2 are allowed to take the value ∞,
their difference μ1 − μ2 is not well-defined in general. This being said, μ1 − μ2 is a
well-defined finite signed measure on each compact subset of X.
Proposition 14.28 (Urysohn’s Lemma). If X is a locally compact Hausdorff space
and K ⊂ U ⊂ X are such that K is compact and U is open, then there exists a
function f ∈ C00 (U) that satisfies f = 1 on K and 0 ≤ f ≤ 1.
Theorem 14.29 (Vitali’s Convergence Theorem). Let (X, μ) be a positive measure
space with μ(X) < ∞. Suppose { fk }k∈N is a sequence of functions in L1 (X, μ) and
that f is a function on X (all complex-valued) satisfying:
(i) fk (x) → f (x) for μ-almost every x ∈ X as k → ∞;
(ii) | f | < ∞ μ-almost everywhere in X;
(iii) { fk }k∈N is uniformly integrable, in the sense every ε > 0 there exists
that for
δ > 0 such that for every k ∈ N we have E fk dμ < ε whenever E ⊆ X is a
μ-measurable set with μ(E) < δ.
Then f ∈ L1 (X, μ) and

lim f − f dμ = 0. (14.2.38)
k
k→∞ X

In particular, lim fk dμ = f dμ.
k→∞ X X
560 14 Appendix
See, e.g., [64, p. 133].

Proposition 14.30. Let (X, μ) be a positive measure space and suppose f ∈ L1 (X, μ).
Then for every ε > 0 there exists
δ > 0 such that for every μ-measurable set A ⊆ X
satisfying μ(A) < δ we have A | f | dμ < ε.
Proof. Consider the measure λ := | f |μ on X. Then λ is absolutely continuous with
respect to μ and the (ε, δ) characterization of absolute continuity of measures (see,
e.g., [64, Theorem 6.11, p. 124]) yields the desired conclusion.
14.3 Custom-Designing Smooth Cut-off Functions
Lemma 14.31. Let f : R → R be the function defined by

⎧ −1/x
⎪
⎪
⎨e , if x > 0,
f (x) := ⎪
⎪ ∀ x ∈ R. (14.3.1)
⎩ 0, if x ≤ 0,
Then f is of class C ∞ on R.
Proof. Denote by C the collection of functions g : R → R for which there exists a

polynomial P such that
⎧ −1/x
⎪
⎪
⎨e P(1/x), if x > 0,
g(x) := ⎪
⎪ ∀ x ∈ R. (14.3.2)
⎩ 0, if x ≤ 0,
Recall that if h : R → R is a continuous function that is differentiable on R \ {0}

and for which there exists L ∈ R such that lim− h
(x) = L = lim+ h
(x), then h is also
x→0 x→0
differentiable at the origin and h
(0) = L. An immediate consequence of this fact is
that any g ∈ C is differentiable and g
∈ C. In turn, this readily gives that any g ∈ C
is of class C ∞ on R. Since f defined in (14.3.1) clearly belongs to C, it follows that
f is of class C ∞ on R.
Lemma 14.32. The function φ : Rn → R defined by

⎧
⎪
⎪
⎨ Ce
1
|x|2 −1 if x ∈ B(0, 1),
φ(x) := ⎪
⎪ (14.3.3)
⎩0 if x ∈ Rn \ B(0, 1),
1 −1
where C := ωn−1 0 e1/(ρ −1) ρn−1 dρ ∈ (0, ∞), satisfies the following properties:
2
φ ∈ C ∞ (Rn ), φ ≥ 0, φ is even,
(14.3.4)
supp φ ⊆ B(0, 1), and Rn
φ(x) dx = 1.
Proof. That φ ≥ 0 and supp φ ⊆ B(0, 1) is immediate from its definition. Also, since
φ(x) = C f (1 − |x|2 ) for x ∈ Rn where f is as in (14.3.1), invoking Lemma 14.31
14.3 Custom-Designing Smooth Cut-off Functions 561

it follows that φ ∈ C ∞ (Rn ). Finally, the condition Rn φ(x) dx = 1 follows upon
1
observing that based on (14.9.9) we have B(0,1) e |x|2 −1 dx = 1/C.
Proposition 14.33. Let F0 , F1 ⊂ Rn be two nonempty sets with the property that
dist(F0 , F1 ) > 0. Then there exists a function ψ : Rn → R with the following
properties:
ψ ∈ C ∞ (Rn ), 0 ≤ ψ ≤ 1, ψ = 0 on F0 , ψ ≡ 1 on F1 , and
(14.3.5)
∀ α ∈ Nn0 ∃ Cα ∈ (0, ∞) such that |∂α ψ(x)| ≤ Cα
dist (F0 ,F1 )|α|
∀ x ∈ Rn .
1 := x ∈ Rn : dist (x, F1 ) ≤ r/4. Also,

Proof. Let r := dist(F0 , F1 ) > 0 and set F

with φ as in Lemma 14.32, define the function θ(x) := 4r n φ(4x/r) for x ∈ Rn . Then

∞
θ ∈ C (R ), θ ≥ 0, supp θ ⊆ B(0, r/4), and
n
θ(x) dx = 1. (14.3.6)
Rn
We claim that the function ψ : Rn → R defined by ψ := χF1 ∗ θ has the desired

properties. To see why this is true note that since

ψ(x) = θ(x − y) dy ∀ x ∈ Rn , (14.3.7)
1
F
from the properties of θ it is immediate that ψ ∈ C ∞ (Rn ) and 0 ≤ ψ ≤ 1. Further-

more, for α ∈ Nn0 and x ∈ Rn we may write

α 4 |α| 4 n α
|∂ ψ(x)| ≤ r r
(∂ φ) 4(x − y)/r dy
1
F
α
4 |α|
∂ φ(y) dy = Cα r−|α| =
Cα
≤ , (14.3.8)
r
Rn dist (F0 , F1 )|α|

where Cα := 4|α| Rn ∂α φ(y) dy is a positive, finite number independent of r.
We are left with checking the fact that ψ = j on F j , j = 0, 1. First, if x ∈ F0 , then
|x − y| ≥ 3r/4 for every y ∈ F 1 , hence θ(x − y) = 0 for every y ∈ F 1 , which when
combined with (14.3.7) implies ψ(x) = 0. Second, if x ∈ F1 , then
B(x, r/4) ⊆ F1 .
Since the support of θ(x − ·) ⊂ B(x, r/4) the latter implies ψ(x) = Rn θ(x − y) dy = 1.
The proof of the proposition is complete.
A trivial yet useful consequence of the above result is as follows.
Proposition 14.34. If U ⊆ Rn is open and K ⊂ U is compact, then there exists a

function ψ : Rn → R that is of class C ∞ , satisfies 0 ≤ ψ(x) ≤ 1 for every x ∈ Rn and
ψ(x) = 1 for every x ∈ K, and which has compact support, contained in U.
Proof. Since K is a compact set contained in U we may define r := dist (K, Rn \ U)

which is a positive number. Now Proposition 14.33 applied with F1 := K and F0 :=
{x ∈ Rn : dist (x, K) > r/2} yields the desired function ψ.
562 14 Appendix
14.4 Partition of Unity
Lemma 14.35. If C ⊂ Rn is compact, and U ⊆ Rn is an open set such that C ⊂ U,

then there exists a compact set D ⊆ Rn such that C ⊂ D̊ ⊂ D ⊂ U.
Proof. Let V = U c ∩ B(0, R), where R > 0 is large enough so that U ⊂ B(0, R). Then
V is compact and disjoint from C so r := dist(V, C) = inf x − y > 0. Hence, if we
x∈C
y∈V
set
D := B(x, r/4), (14.4.1)
x∈C
then D is compact and C ⊂ D̊ ⊂ D ⊂ U.

Lemma 14.36. Suppose that K ⊆ Rn is a compact set and that O j is a finite
1≤ j≤k
open cover of K. Then there are compact sets D j ⊂ O j , 1 ≤ j ≤ k, with the property
that
k
K⊂ D̊ j . (14.4.2)
j=1
k
Proof. Set C1 := K \ j=2 O j ⊆ O1 . Since C1 is compact, Lemma 14.35 shows
that there exists a compact set D1 with the property that C1 ⊂ D̊1 ⊂ D1 ⊂ O1 .
Next, proceeding inductively, suppose that m ∈ N is such that 1 ≤ m < k and
that m compact sets D1 , . . . , Dm have been constructed with the property that K ⊂
m k
j=1 D̊ j ∪ j=m+1 O j . Introduce
⎛m ⎞
⎜⎜⎜
k ⎟⎟⎟
Cm+1 ⎜
:= K \ ⎜⎝⎜ D̊ j ∪ O j ⎟⎟⎠⎟ . (14.4.3)
j=1 j=m+2
Clearly, Cm+1 is a compact subset of Om+1 . By once again invoking Lemma 14.35,
there exists a compact set Dm+1 ⊂ Om+1 such that Cm+1 ⊂ D̊m+1 . After k iterations,
this procedure yields a family of sets C1 , ..., Ck which have all the desired properties.

Theorem 14.37 (Partition of Unity for Compact Sets). Let K ⊂ Rn be a compact

set, and let {O j }1≤ j≤N be a finite open cover of K. Then there exists a finite collection
of C ∞ functions ϕ j : Rn → R, 1 ≤ j ≤ N, satisfying the following properties:
(i) For every 1 ≤ j ≤ N, the set supp (ϕ j ) is compact and contained in O j ;
(ii) For every 1 ≤ j ≤ N, one has 0 ≤ ϕ j ≤ 1;
"N
(iii) ϕ j (x) = 1 for every x ∈ K.
j=1
The family {ϕ j : 1 ≤ j ≤ N} is called a partition of unity subordinate

to the cover {O j }1≤ j≤N of K.

Proof. Let O j be any finite open cover for K. From Lemma 14.36 we know
1≤ j≤N

N
that there exist compact sets D j ⊆ O j , 1 ≤ j ≤ N, such that K ⊂ D̊ j . By
j=1
14.4 Partition of Unity 563
Proposition 14.34, for each 1 ≤ j ≤ N, choose a C ∞ function η j : Rn → [0, ∞) that

"
N
is positive on D j and has compact support in O j . It follows that η j (x) > 0 for all
j=1

N
x∈ D̊ j , so we can define for each j ∈ {1, ..., N} the function
j=1

N
η j (x)
N
ψj : D̊ j → R, ψ j (x) := , ∀x ∈ D̊ j . (14.4.4)
"
N
i=1 ηk (x) i=1
k=1

N
By Lemma 14.36, there exists a compact set U ⊆ Rn with K ⊆ Ů ⊆ U ⊆ D̊ j .
j=1
We apply Proposition 14.34 to obtain a C ∞ function f : Rn → [0, 1] that satisfies
f (x) = 1 for x ∈ K and having compact support contained in Ů. Then for each
j ∈ {1, ..., N} we define the function ϕ j := f ψ j acting from Rn into R. It is not hard
to see that each ϕ j is C ∞ in Rn , has compact support, contained in O j , 0 ≤ ϕ j ≤ 1,
"
and ϕ j (x) = 1 for all x ∈ K.
1≤ j≤N
Definition 14.38. (i) A family (F j ) j∈I of subsets of Rn is said to be locally

finite in E ⊆ Rn provided every x ∈ E has a neighborhood O ⊆ Rn with
the property that the set { j ∈ I : F j ∩ O ∅} is finite.
(ii) Given a collection of functions f j : Ω → R, j ∈ I, defined in some fixed
"
subset Ω of Rn , the sum f j is called locally finite in E ⊆ Rn provided the
j∈I
family of sets {x ∈ Ω : f j (x) 0}, indexed by j ∈ I, is locally finite in E.
Exercise 14.39. Show that a family (F j ) j∈I of subsets of Rn is locally finite in the
open set E ⊆ Rn if and only if for every compact K ⊆ E the set { j ∈ I : F j ∩ K ∅}
is finite.
Exercise 14.40. Show that if the family (F j ) j∈I of subsets of Rn is locally finite in
E ⊆ Rn , then (F j ) j∈I is also locally finite in E.
Exercise 14.41. Show that if the family (F j ) j∈I of closed subsets of Rn is locally

finite in E ⊆ Rn , then (E ∩ F j ) is a relatively closed subset of E.
j∈I
Proof. Let x ∈ E be such that x∗ F j for every j ∈ I. Then there exists r > 0 with
∗
the property that B(x∗ , r) ∩ F j = ∅ for every j ∈ I \ I∗ , where I∗ is a finite subset of I.

Hence, by eventually further decreasing r, it can be assumed that B(x∗ , r) ∩ F j = ∅
for every I∗ as well. Thus, B(x∗ , r) ∩ E is a relative neighborhood of x∗ in E that

is disjoint from F j . This proves that E \ F j is relatively open in E; hence,
j∈I j∈I
(E ∩ F j ) is a relatively closed subset of E.
j∈I
Theorem 14.42 (Partition of Unity for Arbitrary Open Covers). Let (Ok )k∈I be

an arbitrary family of open sets in Rn and set Ω := Ok . Then there exists an at
k∈I
564 14 Appendix
most countable collection (ϕ j ) j∈J of C ∞ nonzero functions ϕ j : Ω → R satisfying

the following properties:
(i) For every j ∈ J there exists k ∈ I such that ϕ j is compactly supported in Ok ;
(ii) For every j ∈ J, one has 0 ≤ ϕ j ≤ 1 in Ω;
(iii) The family of sets {x ∈ Ω : ϕ j (x) 0}, indexed by j ∈ J, is locally finite in Ω;
"
(iv) ϕ j (x) = 1 for every x ∈ Ω.
j∈J
The family (ϕ j ) j∈J is called a partition of unity subordinate to the

family (Ok )k∈I .
Proof. Start by defining
Ω j := {x ∈ Ω : x ≤ j and dist (x, ∂Ω) ≥ 1j }, j ∈ N, (14.4.5)

∞
Then Ω = Ω j and
j=1
Ω j ⊆ Rn is compact, Ω j ⊂ Ω̊ j+1 for every j ∈ N. (14.4.6)
Proceed now to define the compact sets
K2 := Ω2 , K j := Ω j \ Ω̊ j−1 for every j ≥ 3. (14.4.7)
As such, from (14.4.7) and (14.4.6) we have

K j = Ω j ∩ Ω̊ j−1 c ⊆ Ω̊ j+1 ∩ (Ω j−2 )c for every j ≥ 3. (14.4.8)
Finally, we define the following families of open sets:

O2 := Ok ∩ Ω̊3 : k ∈ I , O j := Ok ∩ Ω j+1 \ Ω̊ j−2 : k ∈ I ∀ j ≥ 3. (14.4.9)
Making use of (14.4.6), for every j ≥ 3 we have that

Ω j+1 \ Ω j−2 ◦ = Ω j+1 ∩ (Ω j−2 )c ◦ = Ω̊ j+1 ∩ (Ω j−2 )c ◦

= Ω̊ j+1 ∩ Ω j−2 c = Ω̊ j+1 ∩ (Ω j−2 )c . (14.4.10)
Hence, (14.4.7), (14.4.9), and (14.4.10) imply that O j is an open cover for K j for
every j ≥ 3. Also, from the definitions of K2 and O2 and (14.4.6), we obtain that
O2 is an open cover for K2 . Since the K j ’s are compact, these open covers can be
refined to finite subcovers in each case. As such, for each j = 2, 3, . . . , we can apply
Theorem 14.37 to obtain a finite partition of unity {ϕ : ϕ ∈ Φ j } for K j subordinate to
O j . Also note that due to (14.4.10) and (14.4.9), we necessarily have that, for each
j ∈ {2, 3, . . . }, Ω j ∩ O = ∅ for every O ∈ Ok , for every k ∈ N satisfying k ≥ j + 2.
This ensures that the family {supp ϕ : ϕ ∈ Φ j , j ≥ 2} is locally finite in Ω, so we
can define
s(x) := ϕ(x), for every x ∈ Ω. (14.4.11)
j≥2 ϕ∈Φ j
14.4 Partition of Unity 565
Given that differentiability is a local property, it follows that s is of class C ∞ in

Ω. Moreover, note that s(x) > 0 for every x ∈ Ω, since 0 ≤ ϕ ≤ 1 for all ϕ ∈ Φ j ,
"
j = 2, 3, . . . , and if x ∈ Ω j , for some j ∈ {2, 3, . . . }, then ϕ(x) = 1. Consequently,
ϕ∈Φ j
1/s is also a C ∞ function in Ω. It is then clear that the collection Φ := {ϕ/s : ϕ ∈
Φ j , j = 2, 3, . . . } is a partition of unity subordinate to the family of open sets (Ok )k∈I
(in the sense described in the statement of the theorem).
Theorem 14.43 (Partition of Unity with Preservation of Indexes). Let (Ok )k∈I be

an arbitrary family of open sets in Rn and set Ω := Ok . Then there exists a
k∈I
collection (ψk )k∈I of C ∞ functions ψk : Ω → R satisfying the following properties:
(i) For every k ∈ I the function ψk vanishes outside of a relatively closed subset of
Ok ;
(ii) For every k ∈ I, one has 0 ≤ ψk ≤ 1 in Ω;
(iii) The family of sets {x ∈ Ω : ψk (x) 0}, indexed by k ∈ I, is locally finite in Ω;
"
(iv) ψk (x) = 1 for every x ∈ Ω.
k∈I
Proof. Let (ϕ j ) j∈J be a partition of unity subordinate to the family (Ok )k∈I , and
denote by f : J → I a function with the property that, for every j ∈ J, the function
ϕ j is compactly supported in O f ( j) . That this exists is guaranteed by Theorem 14.42.
For every k ∈ I then define

ψk (x) := ϕ j (x), ∀ x ∈ Ω. (14.4.12)
j∈ f −1 ({k})
Note that the sum is locally finite in Ω, hence ψk : Ω → R is a well-defined non-

negative function, of class C ∞ in Ω, for every k ∈ I. In addition, the result from
Exercise 14.41 shows that, for every k ∈ I, ϕk vanishes outside a relatively closed
subset of Ok . Furthermore, since ( f −1 ({k}))k∈I is a partition of J into mutually dis-
joint subsets, we may compute

ψk (x) = ϕ j (x) = ϕ j (x) = 1 ∀ x ∈ Ω. (14.4.13)
k∈I k∈I j∈ f −1 ({k}) j∈J
Incidentally, this also shows that, necessarily, 0 ≤ ψk (x) ≤ 1 for every k ∈ I and
x ∈ Ω. Finally, the fact that the family of sets {x ∈ Ω : ψk (x) 0}, k ∈ I, is locally
finite in Ω is inherited from the corresponding property of the ϕ j ’s.
Theorem 14.44 (Lebesgue Number Theorem). If (X, d) is a compact metric space

and {C j }Nj=1 is an open cover of X, then there exists a number r∗ > 0 such that
every subset of X having diameter less than r∗ is contained in some member of the
respective cover.
566 14 Appendix
14.5 The Gamma and Beta Functions
The Gamma function is defined as

∞
Γ(z) := tz−1 e−t dt, for z ∈ C, Re z > 0. (14.5.1)
0
√
It is easy to check that Γ(1) = 1, Γ(1/2) = π and via integration by parts that
Γ(z + 1) = zΓ(z) for z ∈ C, Re z > 0. By analytic continuation, the function Γ(z) is
extended to a meromorphic function defined for all complex numbers z except for
z = −n, n ∈ N0 , where the extended function has simple poles with residue (−1)n /n!
and this extension satisfies
Γ(z + 1) = z Γ(z) for z ∈ C \ {−n : n ∈ N0 }. (14.5.2)
By induction it follows that for every n ∈ N we have
Γ(n) = (n − 1)!, (14.5.3)

1 1 · 3 · 5 · · · (2n − 1) √ (2n)! √
Γ +n = π = 2n π, (14.5.4)
2 2n 2 n!
1 (−1)n 2n √ (−1)n 22n n! √
Γ −n = π= π. (14.5.5)
2 1 · 3 · 5 · · · (2n − 1) (2n)!
The volume of the unit ball B(0, 1) in Rn , which we denote by |B(0, 1)|, and the
surface area of the unit sphere in Rn , denoted here by ωn−1 , have the following
formulas:
n n
π2 2π 2
|B(0, 1)| = , ωn−1 = n|B(0, 1)| = . (14.5.6)
Γ n
2 +1 Γ n2
Next, consider the so-called beta function

1
B(z, w) := tz−1 (1 − t)w−1 dt, Re z, Re w > 0. (14.5.7)
0
Clearly B(z, w) = B(w, z). Making the change of variables t = u/(u + 1) for each
u ∈ (0, ∞), it follows that
∞ 1 z+w
B(z, w) = uw−1 du (14.5.8)
0 u+1
whenever Re z, Re w > 0.
The basic identity relating the Gamma and Beta functions reads
Γ(z)Γ(w)
B(z, w) = , Re z, Re w > 0. (14.5.9)
Γ(z + w)
14.6 Surfaces in Rn and Surface Integrals 567
∞
This is easily proved starting with (14.5.8), writing Γ(z + w) = 0 tz+w−1 e−t dt and
expressing B(z, w)Γ(z + w) as a double integral, then making the change of variables
s := t/(u + 1). A useful consequence of identity (14.5.9) is the following formula:
π/2
1 a+1 b+1
(sin θ)a (cos θ)b dθ = B( 2 , 2 )
0 2
1 Γ( 2 )Γ( 2 )
a+1 b+1
= if a, b > −1. (14.5.10)
2 Γ( a+b+2
2 )
Indeed, making the change of variables u := (sin θ)2 , the integral in the leftmost side
1
of (14.5.10) becomes 12 0 u(a−1)/2 (1 − u)(b−1)/2 du.
For further reference, let us also note here that
π
1 + (−1)b
(sin θ)a (cos θ)b dθ = · B( a+1
2 , 2 )
b+1
0 2
1 + (−1)b Γ( 2 )Γ( 2 )
a+1 b+1
= · , (14.5.11)
2 Γ( a+b+2
2 )
whenever a, b > −1. This is proved by splitting the domain of integration into
(0, π/2) ∪ (π/2, π), making a change of variables θ → θ − π/2 in the second integral,
and invoking (14.5.10).
14.6 Surfaces in R n and Surface Integrals
Definition 14.45. Given n ≥ 2 and k ∈ N ∪ {∞}, a C k surface (or, surface of class

C k ) in Rn is a subset Σ of Rn with the property that for every x∗ ∈ Σ there exists an
open neighborhood U(x∗ ) such that
Σ ∩ U(x∗ ) = P(O) (14.6.1)
where O is an open subset of Rn−1 and
P : O −→ Rn is an injective function of class C k satisfying

(14.6.2)
rank [DP(u)] = n − 1, for all u = (u1 , . . . , un−1 ) ∈ O,
where DP is the Jacobian matrix of P, i.e.,

# $
D(P1 , . . . , Pn )
DP(u) = (u), u ∈ O. (14.6.3)
D(u1 , . . . , un−1 )
The function P in (14.6.2) is called a local parametrization of class C k

near x∗ and Σ ∩ U(x∗ ) a parametrizable patch.
568 14 Appendix
In the case when (14.6.1) holds when we formally take r = +∞, i.e., in the case
when Σ = P(O), we call P a global parametrization of the surface Σ.
Proposition 14.46. If O is an open subset of Rn−1 and P : O → Rn satisfies

(14.6.2)–(14.6.3) for k = 1, then P : O → P(O) is a homeomorphism.
Definition 14.47. Assume n ≥ 3. If v1 = (v11 , ..., v1n ), . . . , vn−1 = (vn−1 1 , ..., vn−1 n )
are n − 1 vectors in Rn , their cross product is defined as
⎛ ⎞
⎜⎜⎜ v11 v12 . . . v1n ⎟⎟⎟
⎜⎜⎜ v . . . v2n ⎟⎟⎟
⎜⎜⎜ 21 v22 ⎟⎟⎟
⎜ ⎟⎟⎟
v1 × v2 × · · · × vn−1 := det ⎜⎜⎜⎜ ... ..
.
.
. . . .. ⎟⎟⎟ , (14.6.4)
⎜⎜⎜ ⎟
⎜⎜⎜ vn−1 1 vn−1 2
⎝ . . . vn−1 n ⎟⎟⎟⎟⎠
e1 e2 . . . en
where the determinant is understood as computed by formally expanding it with
respect to the last row, the result being a vector in Rn . More precisely,
v1 × . . . × vn−1 (14.6.5)
⎛ ⎞
⎜⎜⎜ v11 . . . v j−1 v j+1 . . . v1n ⎟⎟⎟
n
⎜⎜⎜ . . . .. ⎟⎟⎟
:= (−1) det ⎜⎜⎜ .. . . . ..
j+1 .. ... . ⎟⎟⎟ e j .
⎜⎝ ⎟⎠
j=1
vn−1 1 . . . vn−1 j−1 vn−1 j+1 . . . vn−1 n
Definition 14.48. Let Σ ⊂ Rn , n ≥ 2, be a C 1 surface and assume that P : O → Rn ,

with O open subset of Rn−1 , is a local parametrization of Σ of class C 1 near some
point X ∗ ∈ Σ. Also, suppose that f : Σ → R is a continuous function on Σ that
vanishes outside of a compact subset of P(O). We then define

f (x) dσ(x) (14.6.6)
Σ

:= ( f ◦ P)(u) |∂1 P(u) × . . . × ∂n−1 P(u)| du1 . . . dun−1 if n ≥ 3,
O
and
f (x) dσ(x) := ( f ◦ P)(u)|P
(u)| du if n = 2. (14.6.7)
Σ O
In (14.6.6), dσ stands for the surface measure (or, surface area element),
whereas in (14.6.7), dσ stands for the arc-length measure.
14.7 Lipschitz Domains
Definition 14.49. Call Ω ⊆ Rn a bounded Lipschitz domain if Ω is a nonempty,

proper, bounded open subset of Rn for which there exist r, c > 0 with the following
significance. For each point x∗ ∈ ∂Ω there exist an (n−1)-dimensional plane H ⊆ Rn
14.7 Lipschitz Domains 569
passing through the point x∗ and a choice ν of the unit normal to H such that if one
defines the open cylinder

C x∗ := C(x∗ , H, ν, r, c) := x
+ tν : x
∈ H, |x
− x∗ | < r, |t| < c (14.7.1)
then
C x∗ ∩ Ω = C x∗ ∩ {x
+ tν : x
∈ H and t > ϕ(x
)}, (14.7.2)
for some Lipschitz function ϕ : H → R satisfying
ϕ(x∗ ) = 0 and |ϕ(x

)| < c if |x
− x∗ | ≤ r. (14.7.3)
Definition 14.50. Call Ω ⊆ Rn an upper-graph Lipschitz domain if there exists a

Lipschitz function ϕ : Rn−1 → R with the property that

Ω = (y
, yn ) ∈ Rn−1 × R : yn > ϕ(y
) . (14.7.4)
Remark 14.51. We note that if Ω ⊆ Rn is an upper-graph Lipschitz domain and ϕ

is a Lipschitz function so that (14.7.4) holds, the outward unit normal to Ω is the
vector valued function ν = (ν j )1≤ j≤n with ν j : ∂Ω → R, j ∈ {1, . . . , n}, and where
∂ j ϕ(x
)
ν j x
, ϕ(x
) := % , for 1 ≤ j ≤ n − 1,
1 + |∇ϕ(x
)|2
(14.7.5)
−1
and νn (x
) := % , for a.e. x
∈ Rn−1 .
1 + |∇ϕ(x
)|2
Remark 14.52. A very useful feature of upper-graph Lipschitz domains is that they
are homeomorphic, via bijective bi-Lipschitz maps, to the upper-half space. To see
this, suppose Ω is the upper-graph of a Lipschitz function ϕ : Rn−1 → R, i.e., that
(14.7.4) holds. Introduce the function

Φ : Rn → Rn , Φ(x
, t) := x
, ϕ(x
) + t ,
(14.7.6)
for each x
∈ Rn−1 and each t ∈ R.
Clearly Φ is a Lipschitz function and Φ is bijective with inverse Φ−1 : Rn → Rn

given by Φ−1 (x
, t) := x
, t − ϕ(x
) for each x
∈ Rn−1 and each t ∈ R. In particular,
Φ is a bijective bi-Lipschitz map. Also, Rademacher’s Theorem (cf. Theorem 2.115)
ensures that Φ is differentiable (in a classical sense) at almost every point. A direct
computation shows that the Jacobian matrix of Φ is
⎛ ⎞
⎜⎜⎜ 1 0 . . . 0 0 ⎟⎟⎟
⎜⎜⎜ 0 1 . . . 0 0 ⎟⎟⎟
⎜⎜⎜ ⎟⎟
DΦ = ⎜⎜⎜⎜ ...
⎜ .. .. .. ⎟⎟⎟⎟ a.e. in Rn .
⎜⎜⎜ . . . . . . ⎟⎟⎟ (14.7.7)
⎜⎜⎜ 0 0 . . . 1 0 ⎟⎟⎟⎟⎟
⎝ ⎠
∂1 ϕ ∂2 ϕ . . . ∂n−1 ϕ 1
570 14 Appendix
It is apparent from this that det(DΦ) = det(DΦ−1 ) = 1 a.e. in Rn . Also, Exer-

cise 2.116 guarantees that
the entries in DΦ and DΦ−1 belong to L∞ (Rn ). (14.7.8)
In addition, based on (14.7.4) it is immediate that the following restrictions are bi-
Lipschitz bijections:
Φ : Rn+ → Ω, (14.7.9)
Φ : Rn+ → Ω, (14.7.10)
Φ : ∂Rn+ → ∂Ω, (14.7.11)
A few useful properties of bounded Lipschitz domains are collected in the lemma
below. For a proof see [4, Proposition 2.8].
Lemma 14.53. Assume that Ω is a bounded Lipschitz domain in Rn . Let x∗ ∈ ∂Ω
and let C x∗ and ϕ be associated with x∗ as in Definition 14.49. Then, in addition to
(14.7.2), one also has
C x∗ ∩ ∂Ω = C x∗ ∩ {x
+ tν : x
∈ H, t = ϕ(x
)}, (14.7.12)
C x∗ \ Ω = C x∗ ∩ {x
+ tν : x
∈ H, t < ϕ(x
)}. (14.7.13)
Furthermore,
C x∗ ∩ Ω = C x∗ ∩ {x
+ tν : x
∈ H, t ≥ ϕ(x
)}, (14.7.14)
C x∗ ∩ Ω̊ = C x∗ ∩ {x
+ tν : x
∈ H, t > ϕ(x
)}, (14.7.15)
and, consequently,
E ∩ ∂Ω = E ∩ ∂( Ω ), ∀ E ⊆ C x∗ . (14.7.16)
In the proof of a number of results stated for bounded Lipschitz domains one first
reduces matters to an upper-graph Lipschitz domain. This is done via a localization
procedure which we describe next.
Remark 14.54. Let Ω be a bounded Lipschitz domain in Rn . Then from Defini-
tion 14.49 it follows that for each x∗ ∈ ∂Ω there exist an (n − 1)-dimensional plane
H ⊆ Rn passing through x∗ , a choice ν ∈ S n−1 of the unit normal to H, an open
cylinder C x∗ , and a Lipschitz function ϕ : H → R with ϕ(x∗ ) = 0 and such that
C x∗ ∩ Ω = C x∗ ∩ {x
+ tν : x
∈ H and t > ϕ(x
)}. (14.7.17)

Tautologically, ∂Ω ⊆ C x∗ . Since Ω is assumed to be bounded, it follows that
x∗ ∈∂Ω
∂Ω is a compact set. As such, there exist N ∈ N and x1∗ , . . . , x∗N ∈ ∂Ω with the

N
property that ∂Ω ⊆ C x∗j . For each j = 1, . . . , N, denote by H j , ν j , ϕ j the respective
j=1
hyperplane, unit normal, and Lipschitz function associated with x∗j . In particular, if
for each j ∈ {1, . . . , N} we set

Ω j := x
+ tν j : x
∈ H j and t > ϕ j (x
) , (14.7.18)
then Ω j is an upper-graph Lipschitz domain in Rn (when described in a suitable

system of coordinates corresponding to H j being ∂Rn+ ) and
C x∗j ∩ Ω = C x∗j ∩ Ω j , (14.7.19)
C x∗j ∩ ∂Ω = C x∗j ∩ ∂Ω j , (14.7.20)
for each j ∈ {1, . . . , N}. Also, we complete the collection {C x∗j }Nj=1 of cylinder to an
open cover for Ω by picking an open set O ⊂ Ω with the property that O ⊂ Ω and

N
Ω⊆O∪ C x∗j . (14.7.21)
j=1
Then, we invoke Theorem 14.37 to further obtain a partition of unity subordinate to

this open cover of Ω, i.e., a collection {ψ j }Nj=0 of functions satisfying:
ψ j ∈ C0∞ (Rn ), 0 ≤ ψ j ≤ 1, ∀ j ∈ {0, 1, . . . , N},

supp (ψ0 ) ⊂ O, supp (ψ j ) ⊂ C x∗j , ∀ j ∈ {1, . . . , N},
(14.7.22)
"
N
and ψ j ≡ 1 near Ω.
j=0
This machinery is used, for example, in the proof of Theorem 12.27, Theorem 12.29,
Theorem 12.44, Theorem 12.45, and Theorem 12.46.
To state the next lemma we introduce the (open, infinite) one-component circular
cone in Rn with vertex x∗ ∈ Rn , axis h ∈ S n−1 , and (full) aperture θ ∈ (0, π), denoted
by Γθ (x∗ , h), as the set
Γθ (x∗ , h) := {x ∈ Rn : cos(θ/2) |x − x∗ | < (x − x∗ ) · h}. (14.7.23)
Lemma 14.55. Assume Ω ⊆ Rn is an upper-graph Lipschitz domain in Rn and let

ϕ : Rn−1 → R be the Lipschitz function such that (14.7.4) holds. Denote by M the
&
Lipschitz constant of ϕ and fix θ ∈ 0, 2 arctan ( M1 ) . Then
Γθ (x, en ) ⊆ Ω and Γθ (x, −en ) ⊆ Rn \ Ω for each x ∈ ∂Ω. (14.7.24)
Proof. As far as the first inclusion in (14.7.24) is concerned, it suffices to show that
572 14 Appendix
if x
, y
∈ Rn−1 , s ∈ R are such that
(14.7.25)
(y
, s) ∈ Γθ (x
, ϕ(x
)), en then s > ϕ(y
).

Fix x
, y
, s such that (y
, s) ∈ Γθ (x
, ϕ(x
)), en . Then based on (14.7.23) we have
'
cos(θ/2) |y
− x
|2 + (s − ϕ(x
))2 + ϕ(x
) < s. (14.7.26)
To conclude that s > ϕ(y

) it suffices to show
cos (θ/2)(|y
− x
|2 + (s − ϕ(x
))2 ) 2 + ϕ(x
) ≥ ϕ(y
).
1
(14.7.27)
This inequality is trivially true if y

= x
. Suppose x
y
. Then (14.7.27) is equiv-
alent with
(
(s − ϕ(x
))2 ϕ(x
) − ϕ(y
)
cos(θ/2) 1 + + ≥ 0. (14.7.28)
|y
− x
|2 |y
− x
|
|s−ϕ(x
)|2
Here is how (14.7.26) implies (14.7.28). Start by setting A := |y
−x
|2
. Then
(14.7.26) becomes cos ( 2θ )(1 + A) 2 < A 2 . Since
1 1
θ 0, squaring both sides this
further implies A > cot2 ( 2θ ), and furthermore that
( )
θ (s − ϕ(x
))2 θ θ
cos 1+

> cos 1 + cot2

2 |y − x |2 2 2
θ
= cot > M, (14.7.29)
2
where for the last inequality in (14.7.29) we have used our assumption on θ. Now
(14.7.28) follows from (14.7.29) upon recalling that ϕ is a Lipschitz function with
Lipschitz constant M.
Next we discuss integration on boundaries of Lipschitz domains. In the case
when Ω ⊆ Rn is an upper-graph Lipschitz domain and ϕ : Rn−1 → R is the Lipschitz
function with the property that (14.7.4) holds, for each σ-measurable nonnegative
function f : ∂Ω → R we define its integral over ∂Ω as

%
f dσ := f x
, ϕ(x
) 1 + |∇ϕ(x
)| dx
. (14.7.30)
∂Ω Rn−1
The space L1 (∂Ω) is then defined as the collection (of equivalence classes, under
the identification of a.e. coincidence) of σ-measurable functions f : ∂Ω → R with
the property that ∂Ω | f | dσ < ∞. For each f ∈ L1 (∂Ω) we may then define ∂Ω f dσ

as ∂Ω f+ dσ − ∂Ω f− dσ where f± := max{± f, 0} ∈ L1 (∂Ω). This ultimately implies
that

%
f dσ = f x
, ϕ(x
) 1 + |∇ϕ(x
)| dx
for all f ∈ L1 (∂Ω). (14.7.31)
∂Ω Rn−1
In the case when Ω ⊆ Rn is a bounded Lipschitz domain the integral over ∂Ω of

a σ-measurable nonnegative function f : ∂Ω → R is defined as
N

f dσ := ψ j f dσ j , (14.7.32)
∂Ω j=1 ∂Ω j
where {Ω j }Nj=1 are upper-graph Lipschitz domains and {ψ j }Nj=0 is a partition of unity
associated with Ω as in Remark 14.54. We note that the value of the sum in (14.7.32)
is independent of the latter choice of families of upper-graph domains and parti-
tion of unity. As before, this definition leads to a naturally defined Lebesgue space
L1 (∂Ω), within which the integration on ∂Ω is meaningful.
Lemma 14.56. Let α > 0. Suppose Ω is either an upper-graph Lipschitz domain

in Rn , or a bounded Lipschitz domain in Rn . Then there exists some constant C =
C(Ω, α) ∈ (0, ∞) such that for each x ∈ ∂Ω and each r > 0 we have

1|x−y|<r
dσ(y) ≤ Crα . (14.7.33)
∂Ω |x − y|n−1−α
In particular, corresponding to the special case when α = n − 1, estimate (14.7.33)

yields

σ ∂Ω ∩ B(x, r) ≤ Crn−1 , ∀ x ∈ ∂Ω, ∀ r ∈ (0, ∞). (14.7.34)
Proof. Suppose first that Ω is the upper-graph of a Lipschitz function ϕ : Rn−1 → R

with Lipschitz constant M. Pick arbitrary x ∈ ∂Ω and r > 0. Then, by (14.7.31) we
have

1|x−y|<r
dσ(y)
∂Ω |x − y|n−1−α
'
1|(x
−y
,ϕ(x
)−ϕ(y
))|<r
2
= n−1−α 1 + |∇ϕ(y )| dy
R n−1
x − y , ϕ(x ) − ϕ(y )

√
dy
≤ 1+M 2

n−1−α
dy
y
∈Rn−1 , |y
−x
|<r |x − y |
√ r n−2
ρ ωn−2 √
= ωn−2 1 + M 2 dρ = 1 + M 2 rα . (14.7.35)
0 ρ
n−1−α α
In the second to the last equality in (14.7.35) we used polar coordinates to write
y
= x
+ ρ ω, for ω ∈ S n−2 and ρ ∈ [0, r]. This proves (14.7.33). Also, (14.7.34) in
the case of an upper-graph Lipschitz domain now follows from (14.7.33) by taking
α = n − 1.
574 14 Appendix
Next, assume Ω is a bounded Lipschitz domain. Let the integer N ∈ N, the

points x1∗ , . . . , x∗N ∈ ∂Ω, the upper-graph Lipschitz domains {Ω j }Nj=1 , and the cylin-
ders {C x∗j }Nj=1 be as in Remark 14.54. These cylinders have finite diameters, hence
R := max1≤ j≤N diam(C x∗j ) is finite. In particular, ∂Ω ∩ C x∗j ⊆ ∂Ω j ∩ B(x∗j , R), for each
j ∈ {1, . . . , N}, and we may use (14.7.34) to obtain

N

N
σ(∂Ω) ≤ σ ∂Ω ∩ C x∗j ≤ C(Ω j )Rn−1 < ∞. (14.7.36)
j=1 j=1
Next, apply the Lebesgue Number Theorem 14.44 to the metric space ∂Ω en-
dowed with the Euclidean distance and cover {C x∗j }Nj=1 to find r∗ > 0 such that for
every x ∈ ∂Ω the ball B(x, r∗ ) is contained in one of the cylinders in the respec-
tive cover. Pick an arbitrary x ∈ ∂Ω and r > 0. Let j ∈ {1, . . . , N} be such that
B(x, r∗ ) ⊆ C x∗j . Invoking (14.7.20) we have ∂Ω ∩ B(x, r∗ ) = ∂Ω j ∩ B(x, r∗ ). We have
two cases. First, if r ≤ r∗ then ∂Ω∩B(x, r) = ∂Ω j ∩B(x, r). Hence, with dσ j denoting
the surface measure on ∂Ω j , we obtain

1|x−y|<r 1|x−y|<r
dσ(y) = dσ j (y)
∂Ω |x − y| ∂Ω j |x − y|
n−1−α n−1−α
≤ Crα , ∀ r ∈ (0, r∗ ], (14.7.37)
where for the inequality in (14.7.37) we used (14.7.33) for the upper-graph Lipschitz
domain Ω j . Second, if r∗ ≤ r, then we may use (14.7.37) with r = r∗ and (14.7.36)
to write

1|x−y|<r 1|x−y|<r∗ 1r∗ <|x−y|<r
dσ(y) = dσ(y) + dσ(y)
∂Ω |x − y| ∂Ω |x − y| ∂Ω |x − y|
n−1−α n−1−α n−1−α

≤ Cr∗α + r∗−n+1+α σ B(x, r) ∩ ∂Ω

≤ r∗α C + r∗−n+1 σ(∂Ω) = Cr∗α ≤ Crα , (14.7.38)
for some C ∈ (0, ∞) depending on Ω (including the cover {C x∗j }Nj=1 ) and α. A combi-
nation of (14.7.37) and (14.7.38) yields (14.7.33) in the case when Ω is a bounded
Lipschitz domain.
Lemma 14.57. Let α > 0. Assume that Ω ⊆ Rn is either an upper-graph Lipschitz

domain, or a bounded Lipschitz domain. Then for each x ∈ ∂Ω and each r > 0 there
exists a constant C ∈ (0, ∞) depending on Ω, n, and α such that

1|x−y|>r
dσ(y) ≤ Cr−α . (14.7.39)
∂Ω |x − y|n−1+α
Proof. Given x ∈ ∂Ω and r > 0 arbitrary we have

14.8 Integration by Parts and Green’s Formula 575
∞

1|x−y|>r 12 j+1 r≥|x−y|>2 j r
dσ(y) = dσ(y)
∂Ω |x − y|n−1+α j=0 ∂Ω |x − y|n−1+α

∞

≤ (2 j r)−n+1−α σ B(x, 2 j+1 r) ∩ ∂Ω)
j=0

∞
≤ (2 j r)−n+1−αC(2 j+1 r)n−1
j=0

∞
= Cr−α 2− jα = Cr−α , (14.7.40)
j=0
where for the last inequality we have used (14.7.34).

We record here a change of variable formula via bijective bi-Lipschitz functions.
Theorem 14.58. Suppose Ψ : Rn → Rn is a bijective and bi-Lipschitz function. Let

O and U be open sets in Rn such that U = Ψ (O). Then for every function u ∈ L1 (U)
one has

u(x)det D(Ψ −1 )(x) dx = u(Ψ (y)) dy. (14.7.41)
U O
Proof. This is a particular case of [15, Theorem 2, p. 99] applied with n := m,

f := Ψ −1 , and g :=
u, the extension by zero of u outside U.
14.8 Integration by Parts and Green’s Formula
Definition 14.59. We say that a nonempty open set Ω ⊆ Rn , where n ≥ 2, is a C k

domain (or, a domain of class C k ), for some k ∈ N0 ∪ {∞}, provided the following
holds. For every point x∗ ∈ ∂Ω there exist R > 0, an open interval I ⊂ R with 0 ∈ I,
a rigid transformation T : Rn → Rn with T (x∗ ) = 0, along with a function φ of class
C k that maps B(0, R) ⊆ Rn−1 into I with the property that φ(0) = 0 and, if C denotes
the (open) cylinder B(0, R) × I ⊆ Rn−1 × R = Rn , then
C ∩ T (Ω) = {x = (x
, xn ) ∈ C : xn > φ(x
)}, (14.8.1)
C ∩ ∂T (Ω) = {x = (x
, xn ) ∈ C : xn = φ(x
)}, (14.8.2)
C ∩ (T (Ω))c = {x = (x
, xn ) ∈ C : xn < φ(x
)}. (14.8.3)
If Ω ⊆ Rn is a C k domain for some k ∈ N0 ∪ {∞}, then it may be easily verified

that ∂Ω is a C k surface. The fact that boundaries of bounded domains of class C k ,
k ∈ N ∪ {∞}, may be locally flattened via diffeomorphisms of class C k is stated next.
576 14 Appendix
Theorem 14.60 (Integration by Parts Formula). Suppose Ω ⊂ Rn is a domain

of class C 1 and ν = (ν1 , . . . , νn ) denotes its outward unit normal. Let k ∈ {1, . . . , n}
and assume f, g ∈ C 1 (Ω) ∩ C 0 (Ω ) are such that ∂k f , ∂k g ∈ L1 (Ω) and there exists a
compact set K ⊂ Rn with the property that f = 0 on Ω \ K. Then,

(∂k f )g dx = − f (∂k g) dx + f gνk dσ, (14.8.4)
Ω Ω ∂Ω
where σ is the surface measure on ∂Ω.

For the sense in which the last integral in (14.8.4) should be understood see Defini-
tion 14.48.
An immediate corollary of Theorem 14.60 is Green’s formula that is stated next
(recall 7.1.14).
Theorem 14.61 (Green’s formula). Suppose Ω ⊂ Rn is a bounded domain of class
C 1 and ν denotes its outward unit normal. If f, g ∈ C 2 (Ω), then
# $
∂g ∂f
f Δg dx = gΔ f dx + f −g dσ. (14.8.5)
Ω Ω ∂Ω ∂ν ∂ν
14.9 Polar Coordinates and Integrals on Spheres
Assume that n ≥ 3 and R > 0 are fixed. For ρ ∈ (0, R), θ j ∈ (0, π), j ∈ {1, . . . , n − 2},
and θn−1 ∈ (0, 2π), set
x1 := ρ cos θ1 ,
x2 := ρ sin θ1 cos θ2 ,
x3 := ρ sin θ1 sin θ2 cos θ3 , (14.9.1)

..
.
xn−1 := ρ sin θ1 sin θ2 . . . sin θn−2 cos θn−1 ,
xn := ρ sin θ1 sin θ2 . . . sin θn−2 sin θn−1 .
The variables θ1 , . . . , θn−1 , ρ are called polar coordinates.

Definition 14.62. Assume that x∗ ∈ Rn , n ≥ 3, is a fixed point, and R > 0 is ar-
bitrary. The standard parametrization of the ball B(x∗ , R) is defined
as
P : (0, π)n−2 × (0, 2π) × (0, R) −→ Rn ,
(14.9.2)
P(θ1 , θ2 , . . . , θn−1 , ρ) := x∗ + (x1 , x2 , . . . , xn ),
where x1 , ..., xn are as in (14.9.1).
14.9 Polar Coordinates and Integrals on Spheres 577
The function P in (14.9.2) is injective, of class C ∞ , takes values in B(x∗ , R), its
image differs from B(x∗ , R) by a subset of measure zero and
det (DP)(θ1 , θ2 , . . . , θn−1 , ρ) = ρn−1 (sin θ1 )n−2 (sin θ2 )n−3 . . . (sin θn−2 ), (14.9.3)
at every point in its domain, where DP denotes the Jacobian of P. Using this stan-
dard parametrization for the unit sphere in Rn , we see that
π π 2π
ωn−1 = ... (sin ϕ1 )n−2 (sin ϕ2 )n−3 · · · (sin ϕn−3 )2 ×
0 0 0
× (sin ϕn−2 ) dϕn−1 dϕn−2 . . . dϕ1 . (14.9.4)
This parametrization of the sphere B(x∗ , R) may also be used to prove the following
theorem.
Theorem 14.63 (Spherical Fubini and Polar Coordinates). Let f ∈ Lloc 1
(Rn ),
n ≥ 2. Then for each x∗ ∈ R and each R > 0 the following formulas hold:
n
R
f dx = f dσ dρ, (14.9.5)
B(x∗ ,R) 0 ∂B(x∗ ,ρ)
R
f dx = f (x∗ + ρω)ρn−1 dσ(ω) dρ (14.9.6)
B(x∗ ,R) 0 S n−1
R
= f (ρω)ρn−1 dσ(ω) dρ. (14.9.7)
0 ∂B(x∗ ,1)
Moreover, if f ∈ L1 (Rn ), then

∞
f dx = f dσ dρ, (14.9.8)
Rn 0 ∂B(0,ρ)
∞
f dx = f (ρω)ρn−1 dσ(ω) dρ. (14.9.9)
Rn 0 S n−1
Note that if P : O → S n−1 is a parametrization of the unit sphere in Rn and if R

is a unitary transformation in Rn , then R ◦ P : O → S n−1 is also a parametrization
of the unit sphere in Rn . Indeed, this function is injective, has C 1 components, its
image is S n−1 (up to a negligible set) and

Rank D(R ◦ P) = dim Im [R(DP)]

= dim Im (DP) = Rank (DP) = n − 1. (14.9.10)
Hence,
578 14 Appendix

f ◦ R dσ = f (R ◦ P)|∂u1 P × · · · × ∂un−1 P| du1 . . . dun−1
S n−1 O

= f (R ◦ P)|∂u1 (R ◦ P) × · · · × ∂un−1 (R ◦ P)| du1 . . . dun−1
O

= f dσ. (14.9.11)
S n−1
The same type of reasoning also yields the following result.

Proposition 14.64. For each x = (x1 , ..., xn ) ∈ Rn , and each j, k ∈ {1, . . . , n} with
j ≤ k, define
R j (x) := (x1 , ..., x j−1 , −x j , x j+1 , ..., xn ),

(14.9.12)
R jk (x) := (x1 , ..., x j−1 , xk , x j+1 , ..., xk−1 , x j , xk+1 , ..., xn ).
Then

f ◦ R j dσ = f dσ, 1 ≤ j ≤ n, (14.9.13)
S n−1 S n−1

f ◦ R jk dσ = f dσ, 1 ≤ j ≤ k ≤ n. (14.9.14)
S n−1 S n−1
Proposition 14.65. Let v ∈ Rn \ {0}, n ≥ 2, be fixed. Then for any measurable and
nonnegative function f defined on the real line, there holds
1 √
f (v · θ) dσ(θ) = ωn−2 f (s|v|)( 1 − s2 )n−3 ds. (14.9.15)
S n−1 −1
Proof. Since integrals over the unit sphere are invariant under orthogonal transfor-
mations, we may assume that v/|v| = e1 and, hence, using polar coordinates and
(14.9.3), we have

f (v · θ) dσ(θ) = f (|v|θ1 ) dσ(θ)
S n−1 S n−1
2π π π *
n−2
= ... f (|v| cos ϕ1 ) (sin ϕ j )n−1− j dϕ1 ... dϕn−2 dϕn−1
0 0 0 j=1
π
= ωn−2 f (|v| cos ϕ1 )(sin ϕ1 )n−2 dϕ1 . (14.9.16)
0
Making the change of variables s := cos ϕ1 in the last integral above shows that this
matches the right-hand side of (14.9.15).
Remark 14.66. Of course (14.9.15) remains valid for measurable and non-positive
functions. In general, if f is merely measurable and real-valued, then one can write
(14.9.15) for f+ and f− and subtract these identities in order to obtain (14.9.15) for
f , as long as one does not run into ∞ − ∞.
Proposition 14.67. Let f ∈ C 0 (R) be positive homogeneous of degree m ∈ R and

fix η = (η1 , . . . , ηn ) ∈ Rn . Then if n ∈ N with n ≥ 2, for j, k ∈ {1, . . . , n} one has

f (η · ξ)ξ j ξk dσ(ξ) = α|η|m δ jk + β|η|m−2 η j ηk (14.9.17)
S n−1
where
1
α= f (ξ1 )(1 − ξ12 ) dσ(ξ),
n − 1 S n−1
(14.9.18)
1
β= f (ξ1 )(nξ12 − 1) dσ(ξ).
n − 1 S n−1
Proof. For j, k ∈ {1, . . . , n} set

q jk (η) := f (η · ξ)ξ j ξk dσ(ξ), ∀ η ∈ Rn , (14.9.19)
S n−1
and define the quadratic form

n
Q(ζ, η) := q jk (η)ζ j ζk , ∀ ζ, η ∈ Rn . (14.9.20)
j,k=1

Observe that we can write Q(ζ, η) = S n−1 f (η · ξ)(ζ · ξ)2 dσ(ξ). By the invariance
under rotations of integrals over S n−1 (see (14.9.11)), we have that for any rotation
R in Rn

Q(ζ, η) = f (η · R ξ)(ζ · R ξ)2 dσ(ξ) = Q(Rζ, Rη), ζ, η ∈ Rn . (14.9.21)
S n−1
A direct computation also gives that
Q(λ1 ζ, λ2 η) = λ21 λm
2 Q(ζ, η) for all λ1 , λ2 > 0, ζ, η ∈ Rn . (14.9.22)
Next we claim that
Q(ζ, η) = α + β(η · ζ)2 , ∀ ζ, η ∈ S n−1 . (14.9.23)
To show that (14.9.23) holds, we first observe that it suffices to prove (14.9.23) when
η = e1 . Indeed, if we assume that (14.9.23) is true for η = e1 , then for an arbitrary η
let R be the rotation such that Rη = e1 . Then if we also take into account (14.9.20),
we have
Q(ζ, η) = Q(Rζ, Rη) = Q(Rζ, e1 ) = α + β(Rζ · e1 )2 (14.9.24)
= α + β(ζ · R e1 )2 = α + β(η · ζ)2 .

580 14 Appendix
Hence, (14.9.23) will follow if we prove that

f (ξ1 )(ζ · ξ)2 dσ(ξ) = α + β ζ12 , ∀ ζ ∈ S n−1 . (14.9.25)
S n−1

To see the later let A jk := S n−1 f (ξ1 )ξ j ξk dσ(ξ) for j, k ∈ {1, . . . , n}. Assume j
k. Then either j 1 or k 1. If, say, j 1 we use (14.9.13) to conclude that
A jk = 0 in this case. A similar reasoning applies in the case k 1. Clearly, we have
A11 = S n−1 f (ξ1 )ξ12 dσ(ξ). Corresponding to j = k 1, we first observe that A22 =
· · · = Ann since A j j is independent of j ∈ {2, . . . , n} due to (14.9.14). Moreover,
"n
A j j = S n−1 f (ξ1 ) dσ(ξ), which in turn implies that
j=1

1
Ajj = f (ξ1 ) dσ(ξ) − A11 = α for each j ∈ {2, . . . , n}. (14.9.26)
n − 1 S n−1
Combining all these we have that for ζ ∈ S n−1 ,

n
f (ξ1 )(ζ · ξ)2 dσ(ξ) = ζ j ζk A jk
S n−1 j,k=1

n
= ζ12 f (ξ1 )ξ12 dσ(ξ) + α ζ 2j
S n−1 j=2
= α + βζ12 . (14.9.27)
This concludes the proof of (14.9.25) which, in turn, implies (14.9.23). Now if ζ,
η ∈ Rn \ {0}, we make use of (14.9.22) and (14.9.23) to write
ζ η
Q(ζ, η) = |ζ|2 |η|m Q , = α|ζ|2 |η|m + β|η|m−2 (ζ · η)2 (14.9.28)
|ζ| |η|

n

= α|η|m δ jk + β|η|m−2 η j ηk ζ j ζk
j,k=1
which proves (14.9.17).
Proposition 14.68. Consider f (t) := |t| for t ∈ R, and let α, β be as in (14.9.17) for
this choice of f . Then
2ωn−2
α=β= 2 , (14.9.29)
n −1
where ωn−2 denotes the surface measure of the unit ball in Rn−1 .
Proof. Using the standard parametrization of S n−1 (see (14.9.1) with R = 1) we

have
π π π 2π
1
α= ... | cos ϕ1 |(1 − (cos ϕ1 )2 )(sin ϕ1 )n−2 (sin ϕ2 )n−3 ×
n−1 0 0 0 0
× · · · × (sin ϕn−2 ) dϕn−1 dϕn−2 . . . dϕ1
π
ωn−2
= | cos ϕ1 |(sin ϕ1 )n dϕ1 . (14.9.30)
n−1 0
π π
The change of variables θ = π − y yields − π cos θ(sin θ)n dθ = 0
2
cos y(sin y)n dy.
2
Using this back in (14.9.30) then gives
π
2 ωn−2 (sin θ)n+1 2 2 ωn−2
π
2 ωn−2 2
α= cos θ (sin θ) dθ = n
= 2 . (14.9.31)
n−1 0 n−1 n + 1 0 n −1
Similar arguments in computing β give that

π
ωn−2
β= | cos θ|[n(cos θ)2 − 1](sin θ)n−2 dθ (14.9.32)
n−1 0
π
n ωn−2 π
= ωn−2 | cos θ|(sin θ)n−2 dθ − | cos θ|(sin θ)n dθ
0 n−1 0
π π
2 ωn−2 2 2
= (n − 1) cos θ (sin θ) dθ − n
n−2
cos θ (sin θ)n dθ
n−1 0 0
2 ωn−2 π (sin θ)n+1 π2 2 ωn−2

= (sin θ)n−1 − n = 2 .
2
n−1 0 n+1 0 n −1
This finishes the proof.
Recall from (0.0.8) that zα = zα1 1 zα2 2 · · · zαn n whenever z = (z1 , ..., zn ) ∈ Rn and
α = (α1 , α2 , . . . , αn ) ∈ Nn0 . Let us also introduce

2Nn0 := (2α1 , 2α2 , . . . , 2αn ) : α = (α1 , α2 , . . . , αn ) ∈ Nn0 . (14.9.33)
The next proposition deals with the issue of integrating arbitrary monomials on the
unit sphere centered at the origin.
Proposition 14.69. For each multi-index α = (α1 , . . . , αn ) ∈ Nn0 ,

⎧
⎪
⎪
⎪ 0 if α 2Nn0 ,
⎪
⎪
⎪
⎪
⎨ |α|
zα dσ(z) = ⎪
⎪
n−1
α! 2π 2 Γ( 2 )
1+|α| (14.9.34)
⎪
⎪ 2 !
S n−1 ⎪
⎪
⎪ · if α ∈ 2Nn0 ,
⎩ |α|! α Γ( |α|+n
2 ! 2 )
where Γ is the Gamma function introduced in (14.5.1).
Proof. Fix an arbitrary k ∈ N and set

582 14 Appendix

qα := zα dσ(z), ∀ α ∈ Nn0 with |α| = k. (14.9.35)
S n−1
Also, with “dot”’ standing for the standard inner product in Rn , introduce
k!
Qk (x) := qα xα , ∀ x = (x1 , x2 , . . . , xn ) ∈ Rn . (14.9.36)
α∈Nn
α!
0
|α|=k
Then (14.2.3) implies that

Qk (x) = (z · x)k dσ(z), ∀ x ∈ Rn . (14.9.37)
S n−1
Let us also observe here that, if x ∈ S n−1 is arbitrary but fixed and if R is a rotation
about the origin in Rn such that R−1 x = en := (0, ..., 0, 1) ∈ Rn , then by (14.9.36)
and the rotation invariance of integrals on S n−1 (cf. (14.9.11)), we have

Qk (x) = (Rz · x)k dσ(z) = Qk (en ). (14.9.38)
S n−1
By the homogeneity of Qk , (14.9.38) implies that Qk (x) = |x|k Qk (en ) for all x ∈ Rn
and, hence,
k!
qα xα = |x|k Qk (en ) for all x ∈ Rn . (14.9.39)
|α|=k
α!
We now consider two cases, starting with:

Case I: k is odd. In this scenario, the mapping S n−1 z → zkn ∈ R is odd. In
particular, Qk (x) = S n−1 zkn dσ(z) = 0. This, in turn, along with (14.9.39) then force
" k! α
α! qα x = 0 for every x ∈ R . From (14.2.7) it is easy to deduce that for each
n
|α|=k
β ∈ Nn0 we have
⎧
⎪
⎪
⎪ if α γ,
⎨0
∂γx [xα ] = ⎪
⎪ (14.9.40)
x=0 ⎪
⎩α! if α = γ.

γ "

We may therefore conclude that qγ = ∂ x 1
|α|=k α! qα xα = 0 for each γ ∈ Nn0 with
|γ| = k, in agreement with (14.9.34).
"
Case II: k is even. Suppose k = 2m for some m ∈ N. Then, |x|k = m! 2β
β! x and
|β|=m
(14.9.39) becomes
m! k!
x2β Qk (en ) = qα xα for all x ∈ Rn . (14.9.41)
|β|=m
β! |α|=k
α!
Fix γ ∈ Nn0 such that |γ| = k and observe that

14.10 Hankel Functions 583
⎧ m!
⎪
⎪
⎪ if γ = 2β, some β ∈ Nn0 , |β| = m,
m! 2β ⎪
⎨ β! (2β)!Qk (en )
∂γx x Qk (en ) = ⎪
⎪
β! x=0
⎪
⎪
⎩0 otherwise.
(14.9.42)
This, (14.9.41), and (14.9.40) then imply that
⎧
⎪
⎪
⎪ 0 if γ 2Nn0 ,
⎪
⎪
⎨ |γ|
⎪
qγ = ⎪
⎪ 2 ! γ! (14.9.43)
⎪
⎪
⎪ γ · Q|γ| (en ) if γ ∈ 2Nn0 .
⎪
⎩ |γ|!
! 2
We are now left with computing Q2m (en ) when m ∈ N. Using spherical coordinates,
a direct computation gives that

Q2m (en ) = (z · en )2m dσ(z)
S n−1
2π π π *
n−2
= ... (cos θ1 )2m (sin θ j )n−1− j dθ1 . . . dθn−2 dθn−1
0 0 0 j=1
π n−1
2π 2Γ( 21 + m)
= ωn−2 (cos θ)2m (sin θ)n−2 dθ = , (14.9.44)
0 Γ(m + n2 )
by (14.9.4) and (14.5.6) (considered with n − 1 in place of n), and (14.5.11). This
once again agrees with (14.9.34), and the proof of Proposition 14.69 is finished.
A simple useful consequence of the general formula (14.9.34) is the fact that

ωn−1
z j zk dσ(z) = δ jk whenever 1 ≤ j, k ≤ n. (14.9.45)
S n−1 n
14.10 Hankel Functions
Let Hλ(1) (·) denote the Hankel function of the first kind with index λ ∈ R. Its defini-
tion and some of its basic properties are reviewed next. For more on this topic see
e.g., [1], [46], [60], [74].
Definition 14.70. Fix λ ∈ R. The regular Bessel function Jλ is defined by

∞
z λ+2 j (−1) j
Jλ (z) = , z ∈ C, (14.10.1)
j=0
2 Γ(λ + j + 1) j!
where Γ is the gamma function from (14.5.1). Also, by Nλ we denote the irregular
Bessel function given by
584 14 Appendix
Jλ (z) cos(πλ) − J−λ (z)

Nλ (z) := , z ∈ C, (14.10.2)
sin(πλ)
for λ N0 and, corresponding to the case when λ ∈ N0 ,
2 z 1 (λ − j − 1)! z 2 j−λ
λ−1
Nλ (z) := Jλ (z) log −
π 2 π j=0 j! 2

1 z 2 j+λ (−1) j Γ
( j + 1) Γ
( j + λ + 1)
∞
− + , (14.10.3)
π j=0 2 j!( j + λ)! Γ( j + 1) Γ( j + λ + 1)
for z ∈ C.
The Hankel function of the first kind with index λ is then defined by
Hλ(1) (·) := Jλ (·) + i Nλ (·). (14.10.4)
The following lemma summarizes some well-known standard properties of Han-

kel functions of the first kind (cf. [1], [60]).
Lemma 14.71. Let λ ∈ R. Then the Hankel function of the first kind Hλ(1) is of class
C ∞ in (0, ∞) and the following properties hold for each r > 0:
(1)
(1) H−λ (r) = eiπλ Hλ(1) (r)
d λ (1)
(2) r Hλ (r) = rλ Hλ−1 (1)
(r)
dr

d −λ (1)
(3) r Hλ (r) = −r−λ Hλ+1 (1)
(r)
dr
d λ
(4) Hλ(1) (r) = Hλ−1 (1)
(r) − Hλ(1) (r)
dr r
d (1) λ
(5) Hλ (r) = −Hλ+1 (r) + Hλ(1) (r)
(1)
dr r
2λ (1) (1) (1)
(6) Hλ (r) = Hλ−1 (r) + Hλ+1 (r)
r
d
1 (1)
(7) Hλ(1) (r) = Hλ−1 (r) − Hλ+1 (1)
(r)
dr 2 # $
d N (1) 1
N
N (1)
(8) for each N ∈ N we have Hλ (r) = N (−1) j H (r)
dr 2 j=0 j λ−N+2 j
2 1/2
(9) Hλ(1) (r) = ei(r−λπ/2−π/4) + O(r−3/2 ) as r → ∞.
πr
The nature of the singularity of Hankel functions of the first kind at the origin is
studied next.
Lemma 14.72. The following formulas are valid:
H0(1) (r)
lim+ =1 (14.10.5)
r→0 2i
ln(r)
π
and
Hλ(1) (r)
lim = 1, ∀ λ ∈ (0, ∞). (14.10.6)
r→0+ 2λ
Γ(λ) r−λ
iπ
Moreover,
Hλ(1) (r)
lim+ = 1, ∀ λ ∈ (−∞, 0). (14.10.7)
r→0 e−iπλ 2−λ
Γ(−λ) rλ
iπ
Proof. The limits in (14.10.5) and (14.10.6) may be found in [60, 10.7.2 and 10.7.7].
If λ ∈ (−∞, 0), then item (1) in Lemma 14.71 and (14.10.6) imply (14.10.7).
We continue by collecting useful asymptotic expansions at infinity for Hankel
functions of the first kind in the next lemma.
Lemma 14.73. Let λ ∈ R, N ∈ N, and suppose r > 0. Then the following asymptotic
expansions of Hankel functions of the first kind and their derivatives hold:
Hλ(1) (r) = O(r−1/2 ) as r → ∞, (14.10.8)
d (1) (1)
H (r) = Hλ−1 (r) + O(r−3/2 ) as r → ∞, (14.10.9)
dr λ
d N (1)
Hλ (r) = O(r−1/2 ) as r → ∞, (14.10.10)
dr
d N (1) (1)
Hλ (r) = Hλ−N (r) + O(r−3/2 ) as r → ∞. (14.10.11)
dr
Proof. Property (14.10.8) follows directly from item (9) in Lemma 14.71, while
(14.10.8) combined with (4) in Lemma 14.71 yields (14.10.9). Also, (14.10.8)
together with (8) in Lemma 14.71 gives (14.10.10). We are left with proving
(14.10.11). First, we claim that for each N ∈ N,
d N (1) N
1
λ (1)
Hλ (r) = C N, j N− j Hλ− j (r) (14.10.12)
dr j=0
r
λ
where C N, j ∈ C are constants depending only on N, j, and λ, defined as follows.
Corresponding to N = 1 we take
λ λ
C1,0 := −λ and C1,1 := 1, (14.10.13)
then for each N ∈ N we recursively define

586 14 Appendix
⎧
⎪
⎪
⎪ 1 if j = N + 1,
⎪
⎪
⎪
λ ⎨ C λ (2 j − N − λ) + C λ
C N+1, j ⎪
:= ⎪
⎪ N, j N, j−1 if 1 ≤ j ≤ N, (14.10.14)
⎪
⎪
⎪
⎩ C λ (−N − λ) if j = 0.
N,0
We shall now prove that formula (14.10.12) holds for the choice of coefficients as
in (14.10.13)–(14.10.14) via an induction argument over N. That the corresponding
statement for N = 1 is true is seen directly from (4) in Lemma 14.71 and (14.10.13).
Suppose next that (14.10.12) holds for some N ∈ N. By differentiating (14.10.12)
one more time and using (4) in Lemma 14.71 we arrive at
d N+1 (1) N
λ j−N (1)
Hλ (r) = C N, j N+1− j Hλ− j (r)
dr j=0
r

N
1 λ − j (1)
λ (1)
+ C N, j Hλ− j−1 (r) − Hλ− j (r)
j=0
r N− j r

N
2 j − N − λ (1) N+1
1
λ λ (1)
= C N, j N+1− j
Hλ− j (r) + C N, j−1 N+1− j Hλ− j (r)
j=0
r j=1
r

N+1
1
λ
= C N+1, H (1) (r), (14.10.15)
j=0
j
r N+1− j λ− j
where the last step uses the recurrence formula (14.10.14). This completes the proof
of (14.10.12).
Moving on, observe that formula (14.10.12) may be written as
d N (1)
N−1
1
(1) λ (1)
Hλ (r) = Hλ−N (r) + C N, j N− j Hλ− j (r). (14.10.16)
dr j=0
r
When used in concert with (14.10.8), this now readily yields (14.10.11).
A combination of Lemma 14.73 and the Chain Rule yields asymptotic expan-
sions for derivatives with respect to x of Hλ(1) (k|x|). These are collected in the next
proposition. The reader is reminded that for each x ∈ Rn \ {0} we abbreviate
+
x := x/|x|.
Proposition 14.74. Let λ ∈ R, k ∈ (0, ∞), and fix a multi-index β ∈ Nn0 with |β| > 0.
Then the following asymptotic expansions hold:
β, &
∂ Hλ(1) (k|x|) = Hλ−|β|
(1)
x )β + O |x|−3/2
(k|x|)(k+ as |x| → ∞, (14.10.17)
## $1/2 $
β, & 2
∂ Hλ(1) (k|x|) = e i(k|x|−λπ/2−π/4)
x )β
(ik+
πk|x|

+ O |x|−3/2 as |x| → ∞. (14.10.18)
Proof. Fix a multi-index β = (β1 , . . . , βn ) ∈ Nn0 of positive length. For starters ob-
serve that repeated applications of the Chain Rule give that, for x ∈ Rn \ {0},
, & d |β| (1)
∂β Hλ(1) (k|x|) = k|β| Hλ (k|x|)×
dr
× (∂1 (|x|))β1 · (∂2 (|x|))β2 · · · (∂n (|x|))βn

|β|−1 d (1)
+ k Hλ (k|x|)×
=1
dr

× Cα1 ,...,α ∂α1 (|x|) · · · ∂α (|x|), (14.10.19)
α1 +···+α =β
with the convention that the sum over is void if |β| = 1. Above, Cα1 ,...,α are
constants depending only on the multi-indices α1 , . . . , α ∈ Nn0 . Since for each
x
j ∈ {1, . . . , n} we have ∂ j (|x|) = |x|j = +
x j , we may write (14.10.19) as
, & d |β| (1)
∂β Hλ(1) (k|x|) = x )β
Hλ (k|x|) (k+ (14.10.20)
dr

|β|−1
d (1)
+ k Hλ (k|x|)×
=1
dr

× Cα1 ,...,α ∂α1 (|x|) · · · ∂α (|x|),
α1 +···+α =β
again, with the convention that the sum over is void in the case when |β| = 1.
Invoking (14.10.11), we further transform
d |β| (1) (1)
x )β = Hλ−|β|
Hλ (k|x|) (k+ (k|x|) + O(|x|−3/2 ) (k+
x )β
dr
(1)
= Hλ−|β| x )β + O(|x|−3/2 )
(k|x|)(k+
as |x| → ∞. (14.10.21)
588 14 Appendix
On the other hand, note that ∂γ (|x|) = O(|x|1−|γ| ) as |x| → ∞, for any γ ∈ Nn0 . On
account of this observation we then conclude that, if |β| > 1, then for each index
∈ {1, . . . , |β| − 1} we have

Cα1 ,...,α ∂α1 (|x|) · · · ∂α (|x|)
α1 +···+α =β
−|β|
= O(|x| ) = O(|x|−1 ) as |x| → ∞. (14.10.22)
Now (14.10.17) follows by combining (14.10.20), (14.10.21), (14.10.22), and

(14.10.10). Finally, formula (14.10.18) is a direct consequence of (14.10.17) and
item (9) in Lemma 14.71.
14.11 Tables of Fourier Transforms
Fourier Transforms of Schwartz Functions
f F(f) Location
n2 |ξ|2
e−λ|x| , λ ∈ C, Re(λ) > 0 π
e− 4λ
2
λ Example 3.22
π 12 (ξ−b)2
e−ax +ibx
e−
2
, a
4a Exercise 3.23
if a > 0 and b ∈ R are fixed
n
π2 (A−1 ξ)·ξ
e−(Ax)·x , A ∈ Mn×n (R), A = A , √ e− 4 Exercise 3.38
det A
(Ax) · x > 0 for x ∈ Rn \ {0}
2π
e−(x1 +x1 x2 +x2 ) for (x1 , x2 ) ∈ R2 √ e−(ξ1 −ξ1 ξ2 +ξ2 )/3
2 2 2 2
Exercise 3.39
3
|ξ−x0 |2 |ξ+x0 |2
1 π 2
n
e−a|x| sin(x · x0 ), for x ∈ Rn , e− 4a − e− 4a
2
2i a Exercise 3.41
if a > 0 and x0 ∈ Rn are fixed
14.11 Tables of Fourier Transforms 589
Fourier Transforms of Tempered Distributions in R

Below x, ξ ∈ R.
u F (u) Location
π −a|ξ|
1
x2 +a2
, a ∈ (0, ∞) given ae Example 4.24
πieib|ξ|
1
x2 −(b+ia)2
, a ∈ (0, ∞) given b+ia e−a|ξ| Exercise 4.122
x
x2 +a2
, a ∈ (0, ∞) given −πi(sgn ξ)e−a|ξ| Example 4.31
⎧
⎪
⎪
⎪
⎪
⎪
⎪ sin(aξ)
⎨ 2 ξ for ξ ∈ R \ {0}
⎪
χ[−a,a] , a ∈ (0, ∞) given ⎪
⎪ Example 4.36
⎪
⎪
⎪
⎪
⎪
⎩ 2a for ξ = 0
sin(b x), b ∈ R fixed −iπδb + iπδ−b Example 4.38
cos(b x), b ∈ R πδb + πδ−b Exercise 4.127

sin(b x) sin(c x), − π2 δb+c − δb−c − δc−b + δ−b−c Example 4.39
b, c ∈ R fixed
H −iP.V. 1ξ − c δ Exercise 4.119
P.V. 1x iπ sgn (ξ) Exercise 4.120
sgn x −2i P.V. 1ξ Exercise 4.121
|x|k , k ∈ N even 2πik δ(k) Exercise 4.121

|x|k , k ∈ N odd −2ik+1 P.V. 1ξ (k) Exercise 4.121
sin(b x)
x , b ∈ R fixed π sgn(b)χ[−|b|,|b|] Exercise 4.121
sin(b x) sin(c x) , &
x , iπ
2 sgn(c) χ[−b−|c|,−b+|c|] − χ[b−|c|,b+|c|] Exercise 4.121
b, c ∈ R fixed
√ - π−ξ2 π−ξ2
.
π
sin(x2 ) 2i ei 4 − e−i 4 Exercise 4.121
ln |x| −2πγδ − πwχ(−1,1) Exercise 4.121
wχ(−1,1) from (4.6.22), γ from (4.6.24)

590 14 Appendix
Fourier Transforms of Tempered Distributions

Below x, ξ ∈ Rn and x
, ξ
∈ Rn−1 .
u F (u) Location
δ 1 Example 4.22
1 (2π)n δ Exercise 4.33

n2 |ξ|2
π
e−ia|x| , a ∈ (0, ∞) given e−i 4 e i 4a
2 nπ
a Example 4.25
π n |ξ|2
e i 4 e−i 4a
2 nπ
e ia|x| , a ∈ (0, ∞) given a
2
Example 4.40
Γ ( n−λ
2 )
|x|−λ , λ ∈ (0, n) fixed |ξ|λ−n
n
2n−λ π 2 Γ ( λ2 )
Proposition 4.64
xj ξj
|x|n , n≥2 −i ωn−1 |ξ|2
Corollary 4.65
xj n Γ ( n−λ ) ξj
|x|λ+2
, n ≥ 2, −i 2n−λ−1 π 2 Γ( λ +1
2
) |ξ|n−λ
Corollary 4.65
2
λ ∈ [0, n − 1), j ∈ {1, . . . , n}

x j xk δ ξξ
n ≥ 3, j, k ∈ {1, . . . , n}
|x|n , ωn−1 |ξ|jk2 − 2ωn−1 |ξ|j 4k Exercise 7.79
, & ξξ ωn−1
P.V. ∂ j |x|xkn , j, k ∈ {1, . . . , n} ωn−1 |ξ|j 2k − n δ jk Proposition 4.73

P.V. Θ, with Θ as in (4.4.1) − S n−1
Θ(ω) log(i(ξ · ω)) dσ(ω) Theorem 4.74
xj ξj
P.V. |x|n+1 , j ∈ {1, . . . , n} − iω2n |ξ| Proposition 4.84

2 t
ωn−1 (t2 +|x
|2 ) n2 e−t|ξ | Proposition 4.90
with t > 0 fixed

xj ξ
2
ωn−1 (t2 +|x
|2 ) n2 −i |ξ
j | e−t|ξ | Proposition 4.92
t > 0 fixed, j ∈ {1, . . . , n − 1}

References
1. M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions, Dover, New York,

1972.
2. R. A. Adams and J. J. F. Fournier, Sobolev Spaces, Pure and Applied Mathematics, 140, 2nd
edition, Academic Press, 2003.
3. D. R. Adams and L. I. Hedberg, Function Spaces and Potential Theory, Grundlehren der
Mathematischen Wissenschaften, Vol. 314, Springer, Berlin, 1996.
4. R. Alvarado, D. Brigham, V. Maz’ya, M. Mitrea, and E. Ziadé, On the regularity of domains
satisfying a uniform hour-glass condition and a sharp version of the Hopf-Oleinik Boundary
Point Principle, Journal of Mathematical Sciences, 176 (2011), no. 3, 281–360.
5. S. Axler, P. Bourdon, and W. Ramey, Harmonic Function Theory, 2nd edition, Graduate Texts
in Mathematics, 137, Springer, New York, 2001.
6. F. Brackx, R. Delanghe, and F. Sommen, Clifford Analysis, Research Notes in Mathematics,
76, Pitman, Advanced Publishing Program, Boston, MA, 1982.
7. A. P. Calderón and A. Zygmund, On the existence of certain singular integrals, Acta Math.,
88 (1952), 85–139.
8. M. Christ, Lectures on Singular Integral Operators, CBMS Regional Conference Series in
Mathematics, No. 77, AMS, 1990.
9. G. David and S. Semmes, Singular Integrals and Rectifiable Sets in Rn : Beyond Lipschitz
Graphs, Astérisque, No. 193, 1991.
10. J. J. Duistermaat and J. A. C. Kolk, Distributions. Theory and applications, Translated from
the Dutch by J. P. van Braam Houckgeest, Cornerstones, Birkhäuser Boston, Inc., Boston,
MA, 2010.
11. J. Duoandikoetxea. Fourier Analysis, Graduate Studies in Mathematics, AMS, 2000.
12. R. E. Edwards, Functional Analysis. Theory and Applications, Dover Publications, Inc., New
York, 1995.
13. L. Ehrenpreis, Solution of some problems of division. I. Division by a polynomial of derivation,
Amer. J. Math. 76, (1954).
14. L. Evans, Partial Differential Equations, 2nd edition, Graduate Studies in Mathematics, 19,
American Mathematical Society, Providence, RI, 2010.
15. L. C. Evans and R. F. Gariepy, Measure Theory and Fine Properties of Functions, Studies in
Advanced Mathematics, CRC Press, Boca Raton, FL, 1992.
16. C. Fefferman and E. M. Stein, H p spaces of several variables, Acta Math., 129 (1972), no.
3-4, 137–193.
17. G. Friedlander and M. Joshi, Introduction to the Theory of Distributions, Cambridge Univer-
sity Press, 2nd edition, 1998.
18. G. B. Folland, Real Analysis, Modern Techniques and Their Applications, 2-nd edition, John
Wiley and Sons, Inc., New York, 1999.

D. Mitrea, Distributions, Partial Differential Equations, and Harmonic Analysis,
Universitext, https://doi.org/10.1007/978-3-030-03296-8
592 References
19. J. Garcia-Cuerva and J. Rubio de Francia, Weighted Norm Inequalities and Related Topics,
North Holland, Amsterdam, 1985.
20. I. M. Gel’fand and G. E. Šilov, Generalized Functions, Vol. 1: Properties and Operations,
Academic Press, New York and London, 1964.
21. I. M. Gel’fand and N. Y. Vilenkin, Generalized Functions, Vol. 4: Applications of Harmonic
Analysis, Academic Press, New York and London, 1964.
22. I. M. Gel’fand, M. I. Graev, and N. Y. Vilenkin, Generalized Functions, Vol. 5: Integral Geom-
etry and Representation theory, Academic Press, 1966.
23. I. M. Gel’fand and G. E. Šilov, Generalized Functions, Vol. 2: Spaces of Fundamental and
Generalized Functions, Academic Press, New York and London, 1968.
24. D. Gilbarg and N.S. Trudinger, Elliptic Partial Differential Equations of Second Order,
Springer, 2001.
25. J. Gilbert and M. A. M. Murray, Clifford Algebras and Dirac Operators in Harmonic Analysis,
Cambridge Studies in Advanced Mathematics, 26, Cambridge University Press, Cambridge,
1991.
26. L. Grafakos, Classical and modern Fourier analysis, Pearson Education, Inc., Upper Saddle
River, NJ, 2004.
27. G. Grubb, Distributions and Operators, Springer, 2009.
28. M. A. Al-Gwaiz, Theory of Distributions, Pure and Applied Mathematics, CRC Press, 1992.
29. G. Hardy and J. Littlewood, Some properties of conjugate functions, J. Reine Angew. Math.,
167 (1932), 405–423.
30. K. Hoffman, Banach Spaces of Analytic Functions, Dover Publications, 2007.
31. S. Hofmann, M. Mitrea and M. Taylor, Singular integrals and elliptic boundary problems
on regular Semmes-Kenig-Toro domains, International Math. Research Notices, 2010 (2010),
2567–2865.
32. L. Hörmander, On the division of distributions by polynomials, Ark. Mat., 3 (1958), 555–568.
33. L. Hörmander, Linear Partial Differential Operators, Die Grundlehren der Mathematischen
Wissenschaften in Einzeldarstellungen, Vol. 116, Springer, 1969.
34. L. Hörmander, The Analysis of Linear Partial Differential Operators I, Distribution Theory
and Fourier Analysis, Springer-Verlag, 2003.
35. L. Hörmander, The Analysis of Linear Partial Differential Operators II, Differential Operators
with Constant Coefficients, Springer, 2003.
36. R. Howard, Rings, determinants, and the Smith normal form, preprint (2013).
http://www.math.sc.edu/ howard/Classes/700c/notes2.pdf
37. G.C. Hsiao and W.L. Wendland, Boundary integral equations, Applied Mathematical Sci-
ences, 164, Springer, Berlin, 2008.
38. F. John, Plane Waves and Spherical Means Applied to Partial Differential Equations, Inter-
science Publishers, New York-London, 1955.
39. C. E. Kenig, Harmonic Analysis Techniques for Second Order Elliptic Boundary Value Prob-
lems, CBMS Regional Conference Series in Mathematics, 83, American Mathematical Soci-
ety, Providence, RI, 1994.
40. S. G. Krantz and H. R. Parks, A Primer of Real Analytic Functions, 2-nd edition, Birkhäuser,
Boston, 2002.
41. P. S. Laplace, Mémoire sur la théorie de l’anneau de Saturne, Mém. Acad. Roy. Sci. Paris
(1787/1789), 201–234.
42. P. S. Laplace, J. École Polytéch. cah., 15 (1809), p. 240 (quoted in Enc. Math. Wiss. Band II,
1. Teil, 2. Hälfte, p. 1198).
43. G. Leoni, A First Course in Sobolev Spaces, Vol. 105, Graduate Studies in Mathematics, Amer-
ican Mathematical Soc., 2009.
44. S. Lojasiewicz, Division d’une distribution par une fonction analytique de variables réelles,
C. R. Acad. Sci. Paris, 246 (1958), 683–686.
45. B. Malgrange, Existence et approximation des solutions des équations aux dérivées partielles
et des équations de convolution, (French) Ann. Inst. Fourier, Grenoble 6 (1955–1956), p. 271–
355.
References 593
46. E. Marmolejo-Olea, D. Mitrea, I. Mitrea, and M. Mitrea, Radiation Conditions and Inte-
gral Representations for Clifford Algebra-Valued Null-Solutions of the Helmholtz Operator,
Journal of Mathematical Sciences, Vol. 231, no. 3 (2018), 367–472; https://doi.org/10.1007/
s10958-018-3826-9.
47. V. G. Maz’ya, Sobolev Spaces, Springer, Berlin-New York, 1985.
48. V. G. Maz’ya, Sobolev spaces with applications to elliptic partial differential equations, Sec-
ond, revised and augmented edition, Grundlehren der Mathematischen Wissenschaften [Fun-
damental Principles of Mathematical Sciences], 342, Springer, Heidelberg, 2011.
49. V. G. Maz’ya and T. Shaposhnikova, Theory of Sobolev multipliers. With applications to differ-
ential and integral operators, Grundlehren der Mathematischen Wissenschaften [Fundamental
Principles of Mathematical Sciences], 337, Springer, Berlin, 2009.
50. W. McLean, Strongly Elliptic Systems and Boundary Integral Equations, Cambridge Univer-
sity Press, Cambridge, 2000.
51. E.J. McShane, Extension of range of functions, Bull. Amer. Math. Soc., 40 (1934), 837–842.
52. Y. Meyer, Ondelettes et Opérateurs II. Opérateurs de Calderón-Zygmund, Actualités
Mathématiques, Hermann, Paris, 1990.
53. Y. Meyer and R. R. Coifman, Ondelettes et Opérateurs III. Opérateurs Multilinéaires, Actu-
alités Mathématiques, Hermann, Paris, 1991.
54. D. Mitrea, I. Mitrea, M. Mitrea, and S. Monniaux, Groupoid Metrization Theory With Appli-
cations to Analysis on Quasi-Metric Spaces and Functional Analysis, Birkhäuser, Springer
New York, Heidelberg, Dordrecht, London, 2013.
55. D. Mitrea, M. Mitrea and M. Taylor, Layer potentials, the Hodge Laplacian and global bound-
ary problems in nonsmooth Riemannian manifolds, Memoirs of American Mathematical Soci-
ety, Vol. 150, No. 713, 2001.
56. D. Mitrea and H. Rosenblatt, A General Converse Theorem for Mean Value Theorems in Lin-
ear Elasticity, Mathematical Methods for the Applied Sciences, Vol. 29, No. 12 (2006), 1349–
1361.
57. I. Mitrea and M. Mitrea, Multi-Layer Potentials and Boundary Problems for Higher-Order
Elliptic Systems in Lipschitz Domains, Lecture Notes in Mathematics 2063, Springer, 2012.
58. M. Mitrea, Clifford Wavelets, Singular Integrals, and Hardy Spaces, Lecture Notes in Mathe-
matics, 1575, Springer, Berlin, 1994.
59. C. B. Morrey, Second order elliptic systems of differential equations. Contributions to the
theory of partial differential equations, Ann. Math. Studies, 33 (1954), 101–159.
60. F. Olver, NIST Digital Library of Mathematical Functions, Chapter 10, http://dlmf.nist.gov/10,
[edited by] Frank W. J. Olver and L. C. Maximon, New York: Cambridge University Press:
NIST, 2010.
61. S. D. Poisson, Remarques sur une équation qui se présente dans la théorie de l’attraction des
sphéroides, Bulletin de la Société Philomathique de Paris, 3 (1813), 388–392.
62. S. D. Poisson, Sur l’intégrale de léquation relative aux vibrations des surfaces élastiques et
au mouvement des ondes, Bulletin de la Société Philomathique de Paris, (1818), 125–128.
63. H. L. Royden, Real Analysis, Macmillian Publishing Co., Inc., New York, 1968.
64. W. Rudin, Real and Complex Analysis, 2nd edition, International Series in Pure and Applied
Mathematics, McGraw-Hill, Inc., 1986.
65. W. Rudin, Functional Analysis, 2nd edition, International Series in Pure and Applied Mathe-
matics, McGraw-Hill, Inc., 1991.
66. L. Schwartz, Théorie des Distributions, I, II, Hermann, Paris, 1950-51.
67. Z. Shapiro, On elliptical systems of partial differential equations, C. R. (Doklady) Acad. Sci.
URSS (N. S.), 46 (1945).
68. E.M. Stein, Singular Integrals and Differentiability Properties of Functions, Princeton Math-
ematical Series, 30, Princeton University Press, Princeton, N.J., 1970.
69. E.M. Stein, Harmonic Analysis: Real-Variable Methods, Orthogonality, and Oscillatory Inte-
grals, Princeton Mathematical Series, 43, Monographs in Harmonic Analysis, III, Princeton
University Press, Princeton, NJ, 1993.
594 References
70. E. M. Stein and G. Weiss, Introduction to Fourier Analysis on Euclidean Spaces, Princeton
University Press, Princeton, New Jersey, 1971.
71. R. Strichartz, A guide to Distribution Theory and Fourier Transforms, World Scientific Pub-
lishing Co., Inc., River Edge, NJ, 2003.
72. H. Tanabe, Functional Analytic Methods for Partial Differential Equations, Monographs and
Textbooks in Pure and Applied Mathematics, 204, Marcel Dekker, Inc., New York, 1997.
73. M. E. Taylor, Pseudodifferential Operators, Princeton Mathematical Series, 34, Princeton Uni-
versity Press, Princeton, N.J., 1981.
74. M. E. Taylor, Partial Differential Equations. I. Basic Theory, Applied Mathematical Sciences,
115, Springer, New York, 1996.
75. M. E. Taylor, Tools for PDE. Pseudodifferential Operators, Paradifferential Operators, and
Layer Potentials, Mathematical Surveys and Monographs, 81, American Mathematical Soci-
ety, Providence, RI, 2000.
76. F. Tréves, Topological Vector Spaces, Distributions and Kernels, Dover Publications, 2006.
77. V. S. Vladimirov, Equations of Mathematical Physics, Marcel Dekker Inc., 1971.
78. V. S. Vladimirov, Methods of the Theory of Generalized Functions, Analytical Methods and
Special Functions, 6, Taylor & Francis, London, 2002.
79. V. Voltera, Sur les vibrations des corps élastiques isotropes, Acta Math., 18 (1894), 161–232.
80. W. Wendland, Integral Equation Methods for Boundary Value Problems, Springer, 2002.
81. J. T. Wloka, B. Rowley, and B. Lawruk, Boundary Value Problems for Elliptic Systems, Cam-
bridge University Press, 1995.
82. W. P. Ziemer, Weakly Differentiable Functions: Sobolev Spaces and Functions of Bounded
Variation, Graduate Text in Mathematics, 120, Springer, New York, 1989.
Subject Index
absorbing set, 545 heat, 321

Arzelà–Ascoli’s Theorem, 85, 558 higher order systems, 391
balanced set, 545 Lamé, 364
base for a topology, 543 Laplacian, 221, 233, 236
Bessel function, 583 poly-harmonic, 221, 257
Bessel operator, 265 scalar second order, 305
Beta function, 566 Schrödinger, 330
bi-harmonic, 254 second order parabolic, 326
bi-Lipschitz function, 464 Stokes, 380
Binomial Theorem, 553 wave, 352
Calderón’s extension operator, 456 dilations, 141
Cauchy problem Dirac distribution δ, 21, 22, 34, 39, 129
heat operator, 324 Dirichlet problem, 179, 420
vibrating infinite string, 1 distance, 544
wave operator, 352 distribution
Cauchy–Clifford operator, 298 compactly supported, 43, 44
Cauchy operator, 292 definition of, 17
Cauchy sequence, 544 of finite order, 19
Chain Rule, 77, 464 of function type, 19
Clifford algebra, 293 tempered, 117
conjugate Poisson kernels, 180 distributional derivative, 32
conormal derivative, 316 elliptic operators, 221
continuous, 544 Euler’s constant, 164
convergence in Fourier transform, 97, 107, 113, 129, 132,
D(Ω), 10 135, 154, 175
E(Ω), 8 partial, 139
S (Rn ), 101 Fréchet space, 546
convex set, 545 F -topology, 547
convolution of distributions, 65 function spaces
convolution of functions, 64 D(Ω), 10, 550
convolution, 125 Dk (Ω), 550
cross product in Rn , 568 E(Ω), 8, 549
cut-off functions, 560 S (Rn ), 99, 551
differential operators fundamental solutions
bi-Laplacian, 221, 249 definition of, 205
Cauchy–Riemann, 288 for higher order systems, 395
Dirac, 293, 295 for scalar second order operators, 306
596 Subject Index
for second order parabolic operators, 330 Lebesgue’s Dominated Convergence

for second order systems, 405 Theorem L p version, 554
for strongly elliptic operators, 314 Lebesgue’s Dominated Convergence
for the bi-Laplacian, 252 Theorem, 554
for the Cauchy operator, 289 Lebesgue Number Theorem, 565
for the Dirac operator, 296 Leibniz’s Formula, 553
for the heat operator, 323 Liouville’s theorem
for the Helmholtz operator, 269, 272 for bi-Laplacian, 254
for the iterated Helmholtz operator, 284 for general operators, 205
for the iterated perturbed Dirac operator, for Lamé, 378
305 for Laplacian, 243
for the Lamé operator, 369, 387 for systems, 359
for the Laplacian, 236 for the heat operator, 324
for the perturbed Dirac operator, 302 Lipschitz constant, 33
for the poly-harmonic operator, 257 Lipschitz domains
for the Schrödinger operator, 332 bounded, 451, 452, 456, 463, 464, 472,
for the Stokes operator, 382, 389 475, 477, 568
for the wave operator, 350 upper-graph, 451, 457, 466, 467, 469,
Gamma function, 566 472, 475, 569
generalized volume potential, 194 Lipschitz function, 83, 554, 556
Green’s formula, 576 definition, 33
Hahn–Banach Theorem, 548 compactly supported, 472
locally, 87
Hankel function, 266, 583
locally finite, 563
Hardy’s Inequality, 557
Malgrange–Ehrenpreis theorem, 209
harmonic function, 233, 241
mean value formulas
Heaviside function
for bi-Laplacian, 254
definition of, 5
for Lamé, 374
derivative, 34
for Laplacian, 242
Hilbert transform, 181, 186, 293
metric space
hypoelliptic operators, 216, 218, 221, 323 complete, 544
inductive limit topology, 546 definition of, 544
integral representation formula, 224, 407 Minkowski’s Inequality, 558
integration by parts formula, 576 Multinomial Theorem, 552
integration on surfaces, 568 multiplier, 182
interior estimates Neumann problem, 422
for elliptic homogeneous systems, 409 Newtonian potential, 189, 191, 193, 249
for hypoelliptic operators, 226 open neighborhood, 543
for the bi-Laplacian, 255 open set, 543
for the Lamé operator, 375, 377 order of a distribution, 19
for the Laplacian, 243 orthogonal transformation, 142
for the poly-harmonic operator, 261 parametrix, 219
invariant under orthogonal transforma- parametrization, 567
tions, 142 Parseval’s identity, 112
inverse Fourier transform, 109, 132 partition of unity
isometry, 558 for arbitrary open covers, 563
layer potential operators for compact sets, 562
double layer, 249, 418 with preservation of indexes, 565
harmonic double layer, 177 Plancherel’s identity, 113
single layer, 249, 417 Plemelj formula, 292
layer potential representation formula, Poisson equation, 206, 239, 378
245, 316 Poisson kernel, 177
Lebesgue’s Differentiation Poisson problem
Theorem, 554 for Lamé, 379
Subject Index 597
for the bi-Laplacian, 255 strongly elliptic operators, 308

for the Laplacian, 244 p-subaveraging, 412, 415
polar coordinates, 576, 577 support of an arbitrary function, 40
positive definite matrix, 309 support of a distribution, 39
positive homogeneous surface, 567
distribution, 143 Taylor’s Formula, 553
function, 142 tensor product
principal symbol, 203 of distributions, 58, 121
principal value distribution, 20, 148 of functions, 51
projection, 293, 299 the method of descent, 344
p-subaveraging, 411 topological space
Rademacher’s Theorem, 86, 554 definition of, 543
real-analytic function, 228–230, 243, 255, Hausdorff, 544
261, 377, 393, 409 metrizable, 544
restriction of a distribution, 37 separated, 544
reverse Hölder estimate, 414
topological vector space
Riesz’s Representation Theorem
definition, 545
for complex functionals, 559
dual, 546
for locally bounded functionals, 559
locally convex, 545
for positive functionals, 558
topology
Riesz transforms, 181, 183, 237, 239, 244,
coarser, 544
299
rigid transformation, 558 definition of, 543
seminorm, 545 finer, 544
separating seminorms, 545 induced, 544
sequentially continuous, 544 product, 544
sign function, 14 total symbol, 203
singular integral operator, 182, 187 Trace operator, 477
singular support, 216 transpose of an operator, 6
slowly increasing functions, 101 unique continuation property, 230
Sobolev spaces Urysohn’s Lemma, 559
global Sobolev spaces H s (Rn ), 425 Vitali’s Convergence Theorem, 559
intrinsic Sobolev spaces H m (Ω), 442 weak
on boundaries of Lipschitz domains derivative, 4
H 1/2 (Ω), 472 solution, 2
restriction Sobolev spaces H s (Ω), 438 weak∗-topology, 547
strongly elliptic matrix, 308 Young’s Inequality, 554
Symbol Index
α! defined on page, xxi elliptic, 314

|α| defined on page, xxi E A fundamental solution for LA
C parabolic, 330
Cauchy operator in C, 292 γ Euler’s constant, 164
Cauchy–Clifford operator, 298 Ex extension operator from H 1/2 to H 1 , 484
χE characteristic function of the set E xxii
f ∨ , 112 ! factorial, xxi
u∨ , 135 F Fourier transform, 97, 132
C k (Ω), xxii · Fourier transform, 97, 129
C k (Ω, xxii F −1 inverse Fourier transform, 109, 132
C0k (Ω), xxii F x partial Fourier transform, 139
C0∞ (Ω), xxii F x−1 inverse partial Fourier transform, 139
Cn Clifford algebra with n generators, 293 Fm,n fundamental solution for Δm in Rn , 257
Clifford algebra multiplication, 293
Γ gamma function, 566
∗ convolution of
S(Rn ) with S (Rn ), 125 H Heaviside function, 5
distributions, 65 H Hilbert transform, 181
functions, 64 H m (Ω) intrinsic Sobolev spaces, 442
v1 × v2 × · · · × vn−1 cross product in Rn , 568 H s (Ω) restriction Sobolev spaces, 438
H s (Rn ) global Sobolev spaces, 425
D Dirac
operator, 295 H 1/2 (∂Ω) Sobolev spaces on boundaries,
D = 1i ∂1 , . . . 1i ∂n , 99 472
Dk perturbed Dirac operator, 301
i complex imaginary unit, xxi
δ jk Kronecker symbol, xxii −A f dμ integral average, xxiii
Δ + k2 Helmholtz operator, 265
δ Dirac distribution, 39 Δ Laplace operator, 179
D(Ω) test functions, 10, 550 Δ2 bi-Laplace operator, 221
DK (Ω) test functions supported in K, 10, Δm poly-harmonic operator, 221
550 ∂t − Δ heat operator, 321
D (Ω) general distributions, 28, 551 ∂2t − Δ wave operator, 352
Lipcomp (E), 472
{e j }1≤ j≤n orthonormal basis in Rn , xxi Liploc (Ω), 87
p
E(Ω), 8, 549 Lcomp (Ω), xxiii
p
Em Calderón’s extension operator, 457 Lloc (Ω), xxiii
E (Ω) compactly supported distributions, LA
44, 549 elliptic, 307
E A fundamental solution for LA parabolic, 326
600 Symbol Index
Mn×m (R) matrices with entries from the (q j )t defined on page, 180
R, xxiii, 360
ring
Mn×m D (Ω) , 355 Re real part of a complex number, xxi
Mn×m E (Ω) , 357 Im imaginary part of a complex number, xxi
mΘ the Fourier transform of P.V. Θ, 154 u|ω restriction of the distribution u to the
open set ω, 37
N the Newtonian potential, 189 R j Riesz transform, 181
∂f
∂ν normal derivative of f , 236
∂νA conormal derivative, 316 S(Rn ) Schwartz functions, 99, 551
S (Rn ) tempered distributions, 119, 552
ωn−1 surface measure of unit ball in Rn , sgn x sign function, 14
xxiii, 566 sing supp u singular support of the distribu-
tion u, 216
∂α partial derivative of order α, xxi L(Rn ) slowly increasing functions, 101
ϕΔ defined on page, 65 supp f support of an arbitrary function f , 40
ΠΦ generalized volume potential, 194 supp u support of the distribution u, 39
P
harmonic Poisson kernel, 177 τt dilation, 141
projection, 293, 299 ⊗ tensor product
pt defined on page, 178 of distributions, 58
P harmonic double layer, 177 of functions, 51
P(D) linear constant coefficient partial of tempered distributions, 121
differential operator, 109, 203 Tr Trace operator, 477
P(x, ∂) linear partial differential operator, 6 t x0 translation by x0 map
P(ξ) total symbol of P(D), 203 of distributions, 71
P(ξ), 109 of functions, 11
Pm (ξ) principal symbol of P(D), 203 P transpose of the operator P, 6
P.V. Θ principal value distribution associ- T Θ defined on page, 182
ated with Θ, 148
P.V. 1x principal value distribution associ- u f distribution associated with f , 19, 118
ated with 1x , 20 u|ver
∂Rn
vertical limit of u to ∂Rn+ , 180
+
Q j defined on page, 180

x, 272

Dorina Mitrea PDF

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Dorina Mitrea PDF

Hochgeladen von

Copyright:

Verfügbare Formate

Universitext

Universitext is a series of textbooks that presents material from a wide variety of

More information about this series at http://www.springer.com/series/223

ISSN 0172-5939 ISSN 2191-6675 (electronic)

1st edition: © Springer Science+Business Media New York 2013

Columbia, MO, USA Dorina Mitrea

Columbia, MO, USA Dorina Mitrea

4 The Space of Tempered Distributions . . . . . . . . . . . . . . . . . . . . . . . 117

8 The Heat Operator and Related Versions . . . . . . . . . . . . . . . . . . . . 321

13 Solutions to Selected Exercises . . ............ . . . . . . . . . . . . . . 497

is systematically treated here for a variety of differential operators L, including the

where m denotes the order of the elliptic differential operator L.

introduced as a mean of extending the notion of solution to a more general setting,

The standard orthonormal basis of vectors in Rn is denoted by fej g1  j  n , where we

supp fi :¼ fj 2 f1; . . .; ng : fij 6¼ 0g; ð5Þ

C k ðXÞ :¼ fu : X ! C : @ fi u continuous 8 fi 2 Nn0 ; jfij  kg; ð9Þ

Ck ðXÞ :¼ fujX : u 2 Ck ðUÞ; U Rn open set containing X g; ð10Þ

C0k ðXÞ :¼ fu 2 C k ðXÞ : supp u compact subset of Xg: ð11Þ

As usual, for any Lebesgue-measurable (complex-valued) function f deﬁned on a

Lploc ðXÞ :¼ ff : X ! C : f Lebesgue measurable such that

Lpcomp ðXÞ :¼ f f 2 Lp ðXÞ : supp f compact subset of Xg; ð14Þ

where supp f is deﬁned in (2.5.12).

If E is a Lebesgue-measurable subset of Rn , the Lebesgue measure of E is

1.1 The Cauchy Problem for a Vibrating Infinite String

The partial diﬀerential equation

∂21 u − ∂22 u = 0 in R2 (1.1.1)

was derived by Jean d’Alembert in 1747 to describe the displacement u(x1 , x2 ) of a

Thanks to the regularity assumption on ϕ0 , it may be checked without diﬃculty that

u(x1 , x2 ) := 12 [ϕ0 (x1 + x2 ) + ϕ0 (x1 − x2 )], for every (x1 , x2 ) ∈ R2 , (1.1.3)

© Springer Nature Switzerland AG 2018 1

is a solution of (1.1.2). This being said, the expression of u in (1.1.3) continues to

Definition 1.1. A function u ∈ C 0 (R2 ) is called a weak (generalized) solution

= ∂21 ϕ(x1 , x2 ) − ∂22 ϕ(x1 , x2 ) ϕ0 (x1 + x2 ) + ϕ0 (x1 − x2 ) dx1 dx2

= (∂21 ϕ) , ϕ0 (y1 ) + ϕ0 (y2 ) dy1 dy2

− (∂22 ϕ) , ϕ0 (y1 ) + ϕ0 (y2 ) dy1 dy2 ,

In summary, this proves that

for every ϕ0 ∈ C 0 (R) the function u defined as in (1.1.3)

1.2 Weak Derivatives

Convention. Unless otherwise specified, Ω denotes an arbitrary open subset of Rn .

This computation suggests the following definition.

φ ∈ C0∞ (Rn ), φ ≥ 0, supp φ ⊆ B(0, 1),

and for each ε > 0 define

φε ∈ C0∞ (Rn ), φε ≥ 0, supp φε ⊆ B(0, ε),

Example 1.4. Consider the function

Note that f is continuous on R but not diﬀerentiable at 0. Nonetheless, f has a weak

Indeed, if ϕ ∈ C0∞ (R), then integration by parts yields

which shows that

Note that H ∈ Lloc

Example 1.5. Does there exist a function g ∈ Lloc

Definition 1.6. Let u, f ∈ Lloc

From the comments in the preamble to Definition 1.6 we know that if u is a

1.3 The Spaces E(Ω) and D(Ω)

The functional in (1.3.1) has the following properties.

E(Ω) := C ∞ (Ω), τ , (1.3.4)

A standard way of constructing a sequence of smooth functions in Rn that con-

Now if we further set for each j ∈ N

then f j ∈ C ∞ (Rn ). Also, if K is an arbitrary compact set in Rn and α ∈ Nn0 , then

The previous approximation result may be further strengthened as indicated in

Exercise 1.13. Prove that C0∞ (Ω) is sequentially dense in E(Ω).

Hint: Consider the sequence of compacts

ϕ j (x) := ϕ(x − 1) + 12 ϕ(x − 2) + · · · + 1j ϕ(x − j), ∀ x ∈ R. (1.3.13)

ϕ(x) := jϕ x −j for each x ∈ R. (1.3.14)

Clearly this limit function does not have compact support.

The standard orthonormal basis of vectors in Rn is denoted by fej g1 j n , where we

C k ðXÞ :¼ fu : X ! C : @ fi u continuous 8 fi 2 Nn0 ; jfij kg; ð9Þ

form ϕ = ϕ0 a ϕ(t) dt + ψ , for some ψ ∈ C0∞ (a, b) .

|u, ϕ j | ≤ C sup |∂α ϕ j (x)| −−−−→ 0, (2.1.2)

|u, ϕ j | > j sup |∂α ϕ j (x)|. (2.1.3)

Λu (ϕ) := lim u, ϕ j . (2.1.36)

|u, ϕ j | ≤ C sup |ϕ j (x)| for each j ∈ N. (2.1.37)