Beruflich Dokumente
Kultur Dokumente
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
Thus as research topics trickle down into graduate-level teaching, first textbooks
written for new, cutting-edge courses may make their way into Universitext.
Distributions, Partial
Differential Equations,
and Harmonic Analysis
Second Edition
123
Dorina Mitrea
Department of Mathematics
University of Missouri
Columbia, MO, USA
Mathematics Subject Classification (2010): 35A08, 35A09, 35A20, 35B05, 35B53, 35B65, 35C15,
35D30, 35E05, 35G05, 35G35, 35H10, 35J05, 35J30, 35J47, 42B20, 42B37, 46E35
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.
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.
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
3.3 Additional Exercises for Chapter 3 . . . . . . . . . . . . . . . . . . . . . 114
xi
xii Contents
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
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:
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:
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
and, respectively, as
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.
⎪
⎪
⎪
(1.1.2)
⎪
⎪
⎪ u(·, 0) = ϕ in R,
⎪
⎪
⎪
0
⎪
⎪
⎩(∂2 u)(·, 0) = 0 in R.
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
(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
We emphasize that in general, there is no reason to expect that u has any pointwise
differentiability properties if ϕ0 is merely continuous.
4 1 Weak Derivatives
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)
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.
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 ).
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.
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
8 1 Weak Derivatives
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→∞
1.3 The Spaces E(Ω) and D(Ω) 9
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→∞
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
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→∞
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.
Note that ϕ ∈ C ∞ (R), supp ϕ = [0, 1], and ϕ > 0 in (0, 1). For each j ∈ N define
E(R)
Then ϕ j ∈ C ∞ (R), supp ϕ j = [1, j + 1], and ϕ j −−−−→ ϕ where
j→∞
∞
1
(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) 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.
t x0 : D(Rn ) −→ D(Rn )
(1.3.17)
t x0 (ϕ) := ϕ(· − x0 ), ∀ ϕ ∈ C0∞ (Rn ).
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).
12 1 Weak Derivatives
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).
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)
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
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 ).
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.
14 1 Weak Derivatives
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)
⎪
⎪
⎪
(1.4.2)
⎪
⎪
⎪u(·, 0) = ϕ0 in R,
⎪
⎪
⎪
⎪
⎪
⎩(∂2 u)(·, 0) = ϕ1 in R.
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
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
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
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
16 1 Weak Derivatives
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 1.41. 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
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.
Building on the idea emerging in (1.3.1), we now make the following definition that
is central to all subsequent considerations.
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.
|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,
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
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
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.
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.
20 2 The Space D (Ω) of Distributions
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
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.
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
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 .
The mapping in (2.1.22) is well defined, linear, and if K is an arbitrary compact set
in Ω, then
(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
2.1 The Definition of Distributions 23
that
Re Λu (ϕ) = ϕ dμ1 − ϕ dμ2 , ∀ ϕ ∈ C00 (Ω) real-valued. (2.1.26)
Ω Ω
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
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
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)
24 2 The Space D (Ω) of Distributions
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
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
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
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
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
26 2 The Space D (Ω) of Distributions
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.
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).
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
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
D (Rn )
This proves that φ j −−−−−→ δ.
j→∞
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
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 ).
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 ).
30 2 The Space D (Ω) of Distributions
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
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→∞
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).
(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
As a consequence,
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.
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
32 2 The Space D (Ω) of Distributions
thanks to (2.3.4) and (2.3.7), leading to the false conclusion that δ = 0 in D (R).
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.
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.
Proposition 2.39. For each α ∈ Nn0 and each u ∈ D (Ω) we have ∂α u ∈ D (Ω).
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).
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.
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,
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
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
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
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).
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
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
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
as claimed.
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-
38 2 The Space D (Ω) of Distributions
The next proposition shows that a distribution is uniquely determined by its local
behavior.
N
K⊂ ωj and u1 ω = u2 ω for j = 1, . . . , N. (2.5.6)
j j
j=1
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
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 .
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 .
40 2 The Space D (Ω) of Distributions
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:
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
Theorem 2.60. Let u ∈ D (Ω) and consider a relatively closed subset F of Ω satis-
fying supp u ⊆ F. Set
Before presenting the proof of this theorem, a few comments are in order.
u1 , ϕ =
u1 , ϕ0 +
u1 , ϕ1 = u, ϕ0 + 0 =
u2 , ϕ0
=
u2 , ϕ0 +
u2 , ϕ1 =
u2 , ϕ, (2.6.2)
Then, ϕ0 − ϕ0 = ϕ1 − ϕ1 , and since supp (ϕ1 − ϕ1 ) ∩ F = ∅, we also have
which in turn implies supp (ϕ0 − ϕ0 ) ⊆ Ω \ supp u. The latter condition entails
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.
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).
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
44 2 The Space D (Ω) of Distributions
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
∞
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.
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.
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
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
46 2 The Space D (Ω) of Distributions
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)
u, ϕ = u, ψϕ = u, (1 − ψ)ϕ + u, ϕ = u, ϕ, (2.6.19)
proving that
u is an extension of u.
2.6 Compactly Supported Distributions and the Space E (Ω) 47
u = ∂
∂α αu in D (Ω) (2.6.20)
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),
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
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!
∂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!
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
2.6 Compactly Supported Distributions and the Space E (Ω) 49
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
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).
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.
N
u= cα, j ∂α δa j , k j ∈ N0 , cα, j ∈ C. (2.6.32)
j=1 |α|≤k j
N
Hence, u = ψ j u in D (Rn ) which, given (2.6.33), proves (2.6.32).
j=1
u, ϕ = (x − a)(x − b)u , 1
(x−a)(x−b) ϕ = 0. (2.6.35)
N1
N2
u= c j δa( j) + d j δb( j) in D (R), (2.6.36)
j=0 j=0
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
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).
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).
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
52 2 The Space D (Ω) of Distributions
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→∞
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
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
54 2 The Space D (Ω) of Distributions
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
(1 )
Define t1 := min : α ∈ N n
, |α| ≤ 1 .
α1 0
(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
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
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
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
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,
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)
(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).
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
∂ϕ
∂ j ψ(y) = u, (·, y) . (2.7.32)
∂y j
2.7 Tensor Product of Distributions 57
∂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
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
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
58 2 The Space D (Ω) of Distributions
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
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
2.7 Tensor Product of Distributions 59
u ⊗ v, ϕ1 ⊗ ϕ2 = u, ϕ1 v, ϕ2
(2.7.42)
∀ ϕ1 ∈ C0∞ (U), ∀ ϕ2 ∈ C0∞ (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.
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).
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)
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).
The uniqueness result from part (i) now implies u ⊗ v = (v ⊗ u) in D (U × 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.
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
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 )
= ∂αx u, ϕ1 ∂βy u, ϕ2
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 = 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)
(
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
2.7 Tensor Product of Distributions 63
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
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∗ )
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.
64 2 The Space D (Ω) of Distributions
= |g(y)| | f (x − y)| dx dy
Rn Rn
=
f
L1 (Rn )
g
L1 (Rn ) . (2.8.1)
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)
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
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)
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)
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
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
2.8 The Convolution of Distributions in Rn 67
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)
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.
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
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
is compact in Rn × Rn .
By part (a) in the theorem, we have supp (v ∗ w) ⊆ supp v + sup w. Thus
σ : Rn × Rn × Rn → Rn × Rn ,
(2.8.31)
σ(x, y, t) := (x, y + t) for every x, y, t ∈ Rn ,
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
70 2 The Space D (Ω) of Distributions
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)
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.
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
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).
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
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
2.8 The Convolution of Distributions in Rn 73
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,
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.
D (Rn ) ∗ C0∞ (Rn ) ⊆ C ∞ (Rn ) and E (Rn ) ∗ C0∞ (Rn ) ⊆ C0∞ (Rn ). (2.8.49)
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
74 2 The Space D (Ω) of Distributions
= 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
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)
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
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
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.
76 2 The Space D (Ω) of Distributions
The same type of result as in Theorem 2.106 actually holds in arbitrary open
subsets of the Euclidean ambient.
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
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,
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
lim w j , ϕ = lim u j0 +2 ∗ φ j , ϕ = u j0 +2 , ϕ = ψ j0 +2 u, ϕ
j→∞ j→∞
where DF −1 denotes the Jacobian matrix of F −1 . Then the following are true.
(1) u ◦ F ∈ D (Ω1 ) and the mapping
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
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 ).
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:
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
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
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→∞
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,
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
2.9 Distributions with Higher Order Gradients Continuous or Bounded 81
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
82 2 The Space D (Ω) of Distributions
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
∂ j v = ∂
j (ψu) = (∂ j ψ)u + ψ∂ j u ∈ C (R ).
0 n
(2.9.12)
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
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
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.
84 2 The Space D (Ω) of Distributions
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
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
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
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
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
86 2 The Space D (Ω) of Distributions
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
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.
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
2.9 Distributions with Higher Order Gradients Continuous or Bounded 87
pw
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
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
(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 2.120. 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
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
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
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
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
2.10 Additional Exercises for Chapter 2 89
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 2.125. 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}.
90 2 The Space D (Ω) of Distributions
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→∞
(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).
92 2 The Space D (Ω) of Distributions
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 ) .
then gδΣ ∈ D (R ). n
may be extended to ensure that f± ∈ C 0 (Ω± ). Prove that for each k in {1, 2 . . . , n}
the following equality holds:
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
(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).
94 2 The Space D (Ω) of Distributions
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 2.161. 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 2.163. 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 ).
D (Rn )
Prove that f j −−−−−→ δ.
j→∞
Exercise 2.166. Prove that the given convolutions exist and then compute them.
(a) H ∗ H
(c) x2 H ∗ (sin x H)
(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).
(−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
Exercise 2.176. Let {x j } j∈N be a sequence of points in R . Prove that {x j } j∈N is con-
n
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.
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.
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
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
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
Hence,
ϕ(ξ) ≤
sup ξβ Dαξ |Dβx (xα ϕ(x))| dx < ∞. (3.1.12)
ξ∈Rn Rn
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 .
100 3 The Schwartz Space and the Fourier Transform
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, ∞].
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.
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 ).
102 3 The Schwartz Space and the Fourier Transform
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.
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 ,
3.1 The Schwartz Space of Rapidly Decreasing Functions 103
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
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)
≤ 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→∞
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
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
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
106 3 The Schwartz Space and the Fourier Transform
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
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).
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
108 3 The Schwartz Space and the Fourier Transform
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
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)
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
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
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
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.
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.
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
(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
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
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
112 3 The Schwartz Space and the Fourier Transform
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
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
3.2 The Action of the Fourier Transform on the Schwartz Class 113
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)
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
n
f L2 (Rn ) = (2π) 2 f L2 (Rn ) . (3.2.28)
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)
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
(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 3.33. 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
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→∞
3.3 Additional Exercises for Chapter 3 115
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
2
sin(e−|x| )
(d) 1+|x|2
2
cos(e−|x| )
(e) (1+|x|2 )n
(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
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 3.44. 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.
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
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
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:
120 4 The Space of Tempered Distributions
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→∞
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
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)
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
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)
4.1 Definition and Properties of Tempered Distributions 123
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:
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
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
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}.
4.1 Definition and Properties of Tempered Distributions 125
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.
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
126 4 The Space of Tempered Distributions
≤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
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)
= u ∗ v, ϕ
. (4.1.50)
4.1 Definition and Properties of Tempered Distributions 127
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)
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
128 4 The Space of Tempered Distributions
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
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→∞
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).
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
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 ).
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)
130 4 The Space of Tempered Distributions
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
In particular, (4.2.8) also holds if f ∈ S(Rn ) (since S(Rn ) ⊂ L1 (Rn ) by Exercise 3.5).
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
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)
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
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
132 4 The Space of Tempered Distributions
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.
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
4.2 The Fourier Transform Acting on Tempered Distributions 133
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)
⊗ 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 ).
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→∞
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
4.2 The Fourier Transform Acting on Tempered Distributions 135
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)
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.
Theorem 4.35. The following statements are true:
(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).
136 4 The Space of Tempered Distributions
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)
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
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
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)
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
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
138 4 The Space of Tempered Distributions
= C1 δa + C2 δ
2π sin(−ax) = sin(ax) −a
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)
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
4.2 The Fourier Transform Acting on Tempered Distributions 139
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
√
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
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)
u in S (Rn+m ),
F x Fy u = Fy F x u = ∀ u ∈ S (Rn × Rm ). (4.2.62)
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
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
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:
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.
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.
4.3 Homogeneous Distributions 143
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)
144 4 The Space of Tempered Distributions
Given that vα Rn \{0} ∈ C ∞ (R \ {0}), the latter translates into
ξα
u ∈ C 0 (Rn ), ∀ α ∈ Nn0 with |α| > k + n. (4.3.15)
we obtain N!
|ξ|2N
u= u in S (Rn ).
ξ2α (4.3.17)
|α|=N
α!
|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.
(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)
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→∞
146 4 The Space of Tempered Distributions
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)
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
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).
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
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
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
|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)
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.
(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
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
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
proving (4.4.11).
The next example discusses a basic class of principal value tempered distribu-
tions arising naturally in applications.
∂ 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
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
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.
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
%% %%
%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
4.5 The Fourier Transform of Principal Value Distributions 155
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
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
156 4 The Space of Tempered Distributions
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
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
4.5 The Fourier Transform of Principal Value Distributions 157
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
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
158 4 The Space of Tempered Distributions
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)
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,
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
4.5 The Fourier Transform of Principal Value Distributions 159
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
160 4 The Space of Tempered Distributions
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
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
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
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
= wψ , ϕ , ∀ ϕ ∈ S(Rn ). (4.6.6)
)
*
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
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)
4.6 Tempered Distributions Associated with |x|−n 163
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)
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
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)
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
4.6 Tempered Distributions Associated with |x|−n 165
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)
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
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
166 4 The Space of Tempered Distributions
−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.
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
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
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
4.7 A General Jump-Formula in the Class of Tempered Distributions 169
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
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
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
170 4 The Space of Tempered Distributions
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
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)
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
172 4 The Space of Tempered Distributions
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
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
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
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
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
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 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
176 4 The Space of Tempered Distributions
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|
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 .
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)
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
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
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
so (4.8.6) gives
, 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)
u := Pϕ in Rn+ , (4.8.18)
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
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
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.
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
182 4 The Space of Tempered Distributions
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
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
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
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.
4.9 Singular Integral Operators 183
(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
(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
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
and
184 4 The Space of Tempered Distributions
n
ωn 2
R2k = − 2 I as operators on L2 (Rn ), (4.9.17)
k=1
This computation may be summarized by saying that, for each ϕ ∈ S(Rn ) fixed, the
linear functional
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
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
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
186 4 The Space of Tempered Distributions
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)
linear, and sequentially continuous.
(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)
4.9 Singular Integral Operators 187
(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
(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 f (ξ) = mΘ (ξ) f (ξ) a.e. in Rn , ∀ f ∈ L2 (Rn ), (4.9.42)
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
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
Of course, the same type of result is true for the Hilbert transform on the real line.
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
4.10 Derivatives of Volume Potentials 191
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
= 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.
Proof. Fix p ∈ (1, n), f ∈ L p (Rn ), and R ∈ (0, ∞). Then we write
192 4 The Space of Tempered Distributions
| 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)
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
4.10 Derivatives of Volume Potentials 193
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
D (Rn ) D (Rn )
T ∂ j Φ fk −−−−−→ T ∂ j Φ f and fk −−−−−→ f. (4.10.18)
k→∞ 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.
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).
∞
Then for each f ∈ Lcomp (Rn ) one has ΠΦ f ∈ C m+n
(Rn ) and
∂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)
∞
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,
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
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→∞
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,
196 4 The Space of Tempered Distributions
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
4.10 Derivatives of Volume Potentials 197
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
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
= (∂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
4.11 Additional Exercises for Chapter 4 199
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 4.108. Prove that |x|N ln |x| ∈ S (Rn ) if N is a real number satisfying
N > −n.
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 ;
200 4 The Space of Tempered Distributions
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 4.123. 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
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 4.130. 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:
202 4 The Space of Tempered Distributions
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.
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
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.
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
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
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
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.
E, P(−D)ϕ = ϕ(0), ∀ ϕ ∈ C0∞ (Rn ). (5.2.2)
206 5 The Concept of Fundamental Solution
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.
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
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.
208 5 The Concept of Fundamental Solution
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).
−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
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
210 5 The Concept of Fundamental Solution
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 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
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 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
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
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].
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 ).
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
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.
214 5 The Concept of Fundamental Solution
m−1
= 1
(m−1)! xm−1 H + ck xk : ck ∈ C, k = 1, . . . , m − 1 .
k=0
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
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
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
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)
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
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
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
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
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
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
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
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}.
220 6 Hypoelliptic Operators
v = δ ∗ v = (P(D)F − w) ∗ v = F ∗ (P(D)v) − w ∗ v
Since sing supp F = {0}, the same argument as in the proof of (6.2.5) shows that
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)
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].
6.2 Hypoelliptic Operators with Constant Coefficients 221
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.
Hence, from (6.2.14) and (6.2.15) it follows that for every ξ ∈ Rn \ {0} we have
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)
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)
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
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
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
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,
224 6 Hypoelliptic Operators
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
(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,
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)
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)
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)
= 0 − 1 = −1. (6.3.6)
226 6 Hypoelliptic Operators
(−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
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)
6.3 Integral Representation Formulas and Interior Estimates 227
∞
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)
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)
228 6 Hypoelliptic Operators
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)
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)
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).
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
6.3 Integral Representation Formulas and Interior Estimates 229
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:
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!
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)
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].
6.4 Additional Exercises for Chapter 6 231
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 6.33. Determine sing supp (P.V. 1x ), where P.V. 1x is the distribution defined
in (2.1.13).
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.
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).
© Springer Nature Switzerland AG 2018 233
D. Mitrea, Distributions, Partial Differential Equations, and Harmonic Analysis,
Universitext, https://doi.org/10.1007/978-3-030-03296-8 7
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)
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
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
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)
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).
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 ).
∂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
∂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+
∂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
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
238 7 The Laplacian and Related Operators
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 ).
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)
Proposition 7.7. Let E be the fundamental solution for Δ from (7.1.12). Suppose
f ∈ C0∞ (Rn ). Then the function
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
240 7 The Laplacian and Related Operators
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
⎪
⎪ −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
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},
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
7.2 The Poisson Equation and Layer Potential Representation Formulas 241
= f (y) E, Δϕ(· + y) dy = f (y) ΔE, ϕ(· + y) dy
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},
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).
242 7 The Laplacian and Related Operators
(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)
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.
7.2 The Poisson Equation and Layer Potential Representation Formulas 243
(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).
Proof. This is an immediate consequence of Theorem 7.12 and Lemma 6.24 (or,
alternatively, Theorem 6.26).
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.
244 7 The Laplacian and Related Operators
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
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
7.2 The Poisson Equation and Layer Potential Representation Formulas 245
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
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).
∂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.
Δ
u, ϕ =
u, Δϕ = uΔϕ dx
Ω
∂ϕ ∂u
= ϕΔu dx + u −ϕ dσ, (7.2.24)
Ω ∂Ω ∂ν ∂ν
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
7.2 The Poisson Equation and Layer Potential Representation Formulas 247
∗ E)(x) =
(Δu E(x − y)Δu(y) dy, ∀ x ∈ Rn . (7.2.29)
Ω
∂u
Since ∂ν ∈ C 0 (∂Ω), by (7.2.25) one has
248 7 The Laplacian and Related Operators
∂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
∂
(uδ∂Ω ) ∗ E (x) (7.2.37)
∂ν
∂u
u(x) = E(x − y)Δu(y) dy − E(x − y) (y) dσ(y)
Ω ∂Ω ∂ν
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).
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
250 7 The Laplacian and Related Operators
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.
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.
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
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.
Δ |x|N ln |x| = |x|N−2 N(N + n − 2) ln |x| + 2N + n − 2 in S (Rn ). (7.3.6)
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.
⎧
⎪
⎪
⎪
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
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
∂|x|N
Since ∇(|x|N ) = N|x|N−2 x for every x 0, we have ∂ν = NεN−1 on ∂B(0, ε). Hence,
7.3 Fundamental Solutions for the Bi-Laplacian 253
N ∂ϕ ∂|x|N
− |x| (x) dσ(x) + ϕ(x) dσ(x) (7.3.13)
∂B(0,ε) ∂ν ∂B(0,ε) ∂ν
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
= (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
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}.
254 7 The Laplacian and Related Operators
Analogously to Theorem 7.10 we have the following regularity result for the bi-
Laplacian.
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).
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.
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
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)
y∈B(x,r/2) r|α| y∈B(x,r)
has a unique solution. Moreover, the solution u satisfies the following additional
properties.
256 7 The Laplacian and Related Operators
∂α u = cα f + T ∂α E f in D (Rn ), (7.4.11)
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).
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).
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
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
7.5 Fundamental Solutions for the Poly-harmonic Operator 259
(−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
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
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
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 .
260 7 The Laplacian and Related Operators
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)
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,
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 .
7.5 Fundamental Solutions for the Poly-harmonic Operator 261
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)
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.
C |α| |α|!
max |∂α u(y)| ≤ max |u(y)|, ∀ α ∈ Nn0 , (7.5.23)
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 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].
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
7.5 Fundamental Solutions for the Poly-harmonic Operator 263
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
264 7 The Laplacian and Related Operators
⎧
⎪
⎪
⎪ (−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
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 !
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
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) = 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)
266 7 The Laplacian and Related Operators
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
(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
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,
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.
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
268 7 The Laplacian and Related Operators
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
7.6 Fundamental Solutions for the Helmholtz Operator 269
(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
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 .
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
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
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)
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
7.6 Fundamental Solutions for the Helmholtz Operator 271
∂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
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)
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.
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| → ∞,
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| → ∞,
and
(∂α Φk )(x − y) = O |y|−(n−1)/2 as |y| → ∞,
(7.6.53)
uniformly for x in compact subsets of Rn .
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|
By invoking (14.10.17), the first term in the right-hand side of (7.6.56) becomes
274 7 The Laplacian and Related Operators
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|
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)
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)
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
276 7 The Laplacian and Related Operators
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)
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)
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 .
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
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
278 7 The Laplacian and Related Operators
⎧ 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
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
xj
lim+ Fλ (x) ϕ(x) dσ(x) = 0. (7.6.86)
ε→0 |x|=ε ε
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,
280 7 The Laplacian and Related Operators
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}:
7.7 Fundamental Solutions for the Iterated Helmholtz Operator 281
(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
(8) For each k ∈ (0, ∞) and each N ∈ N one has Φ(N) k ∈ S (R ).
n
where
282 7 The Laplacian and Related Operators
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,
(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
−2NkΦ(N+1)
k = −2(N + 1)k(Δ + k2 )Φ(N+2)
k + 2kΦ(N+1)
k (7.7.24)
7.7 Fundamental Solutions for the Iterated Helmholtz Operator 283
−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
284 7 The Laplacian and Related Operators
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
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-
7.7 Fundamental Solutions for the Iterated Helmholtz Operator 285
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)
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+
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
hence,
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)
7.7 Fundamental Solutions for the Iterated Helmholtz Operator 287
Φ(N)
k (x) = aN,k F (n−2N)/2 (x), ∀ x ∈ Rn \ {0}, (7.7.44)
∂ 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)
(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|)
(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 .
288 7 The Laplacian and Related Operators
∂ 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
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
Similarly,
∂ 2 1 1 1
ln |x| = (∂1 − i∂2 )[ln |x|] = · in S (R2 ). (7.8.9)
∂z π π π x1 + ix2
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
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
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
292 7 The Laplacian and Related Operators
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
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
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
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:
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.
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|=
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
Proposition 7.51. Let Ω ⊆ Rn be an open set. Then the Dirac operator D satisfies
which identifies vectors from Rn with elements in the Clifford algebra Cn . With this
identification in mind, show that
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
are fundamental solution for the Dirac operators D− and D+ , respectively, in Rn+1 .
298 7 The Laplacian and Related Operators
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
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
7.9 Fundamental Solutions for the Dirac Operator 299
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.
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))
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
300 7 The Laplacian and Related Operators
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
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
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.
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
n+1
u= uI eI with uI ∈ D (Ω) for each I, (7.10.2)
l=0 |I|=l
n
n+1
Du := ∂ j uI e j eI . (7.10.3)
j=1 l=0 |I|=l
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
and that
D2k = −(Δ + k2 ) in D (Ω, Cn+1 ), (7.10.7)
302 7 The Laplacian and Related Operators
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
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)
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
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).
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)
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)
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
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)
(−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
Then Θ(N)k belongs to Lloc (R , Cn+1 )∩S (R , Cn+1 ) and is a fundamental solution
1 n 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.
−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).
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).
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
7.12 Fundamental Solutions for General Second-Order Operators 307
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
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
308 7 The Laplacian and Related Operators
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
7.12 Fundamental Solutions for General Second-Order Operators 309
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
Consequently,
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
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
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
From Remark 7.64 and definitions we see that, given any A ∈ Mn×n (C), the
following implications hold:
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)
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
7.12 Fundamental Solutions for General Second-Order Operators 311
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.
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
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)
312 7 The Laplacian and Related Operators
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.
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
1 A−1 x
∇E A (x) = √ · −1 n ∀ x ∈ Rn \ {0} (7.12.45)
ωn−1 detA [(A x) · x] 2
314 7 The Laplacian and Related Operators
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
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
7.12 Fundamental Solutions for General Second-Order Operators 315
⎧
⎪
⎪
⎪
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
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 .
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.
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.
316 7 The Laplacian and Related Operators
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
⎧
⎪
⎨ 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).
(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
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
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)
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
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].
7.14 Additional Exercises for Chapter 7 319
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
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|
320 7 The Laplacian and Related Operators
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.
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
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).
© Springer Nature Switzerland AG 2018 321
D. Mitrea, Distributions, Partial Differential Equations, and Harmonic Analysis,
Universitext, https://doi.org/10.1007/978-3-030-03296-8 8
322 8 The Heat Operator and Related Versions
∂t + |ξ|2 E x = δ(t) in D (R). (8.1.3)
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
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
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
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
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)
324 8 The Heat Operator and Related Versions
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.
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)
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
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
Then
+ E ∗ f (x) ⊗ δ(t) exists, belongs to D+ (Rn+1 ),
u := E ∗ F (8.2.8)
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 .
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
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
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)
328 8 The Heat Operator and Related Versions
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
= √ 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
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
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
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
330 8 The Heat Operator and Related Versions
P(ξ) = 0 then
n
iξn+1 + a jk ξ j ξk = 0. (8.3.14)
j,k=1
(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
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
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.
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.
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|
we have
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.
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)
and ∞ ∞
π
lim+ IIε = ϕ(x, t) dx dt. (9.1.25)
ε→0 2 0 −t
9.1 Fundamental Solution for the Wave Operator 337
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
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
Remark 9.1.
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
338 9 The Wave Operator
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
9.1 Fundamental Solution for the Wave Operator 339
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)!
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
340 9 The Wave Operator
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
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)
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
342 9 The Wave Operator
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
∞
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
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
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)
Remark 9.2.(1) The distributions E+ and E− from (9.1.51) and (9.1.54), respec-
tively, satisfy
(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
Lemma 9.3. If h : [0, ∞) → R is continuous and bounded, then for each t in (0, ∞)
we have
344 9 The Wave Operator
∞
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.
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.
u j : D(Rn−1 ) → C, u j (ϕ) :=
u, ϕ ⊗ ψ j , ∀ ϕ ∈ C0∞ (Rn−1 ). (9.1.65)
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
We claim that
lim u , (P0 (−∂ )ϕ) ⊗ ψ j − P(−∂)(ϕ ⊗ ψ j ) = 0. (9.1.72)
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
9.1 Fundamental Solution for the Wave Operator 347
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
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)
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
9.1 Fundamental Solution for the Wave Operator 349
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
⎡ ⎤
∞ ⎢⎢⎢ 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:
⎡ ⎤
∞ ⎢⎢⎢ 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
H(t − |x|)
E+ (x, t) = & for every x ∈ R2 and every t ∈ R.
2π t2 − |x|2
H(−t − |x|)
E− (x, t) = & for every x ∈ R2 and every t ∈ R.
2π t2 − |x|2
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,
2
this expression was first found by Vito Volterra (cf. [79])
9.1 Fundamental Solution for the Wave Operator 351
∞
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,
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
⎡ ⎤
∞ ⎢⎢⎢ 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,
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
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
∞
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) .
) ) + 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.
9.3 Additional Exercises for Chapter 9 353
)
u =)
u∗δ =)
u ∗ ((∂2t − Δ x )E) = ((∂2t − Δ x ))
u) ∗ E = 0. (9.2.6)
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.
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
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
N
(LU) j := aαjk ∂α uk for all 1 ≤ j ≤ M, 1 ≤ ≤ K. (10.1.7)
|α|≤m k=1
N
(U ∗ V) j := u jk ∗ vk , 1 ≤ j ≤ M, 1 ≤ ≤ K. (10.1.9)
k=1
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)
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)
define
358 10 The Lamé and Stokes Operators
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
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
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.
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.
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:
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
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
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
362 10 The Lamé and Stokes Operators
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
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
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.
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).
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.
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}.
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.
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
366 10 The Lamé and Stokes Operators
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).
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
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
368 10 The Lamé and Stokes Operators
(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
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
10.3 Fundamental Solutions for the Lamé Operator 369
−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}.
370 10 The Lamé and Stokes Operators
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)
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:
10.3 Fundamental Solutions for the Lamé Operator 371
( (
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
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)
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
(μ + λ)(2 − n) δ jk
= 3μ + λ − (10.3.30)
n 2μ(2μ + λ)
and
372 10 The Lamé and Stokes Operators
n
x δ jk
(∂ j Ek )(x) dσ(x) = , (10.3.31)
=1
ε n(2μ + λ)
∂B(x,ε)
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)
ωn−1 rn+1 ∂B(x,r)
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
374 10 The Lamé and Stokes Operators
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
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)
ωn−1 rn+1 ∂B(x,r)
n
= (y j − x j )(y − x) · u(y) − r2 u j (y) dσ(y)
ωn−1 rn+1 ∂B(x,r)
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)
10.4 Mean Value Formulas and Interior Estimates for the Lamé Operator 375
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
376 10 The Lamé and Stokes Operators
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)
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.
Then any null solution of the Lamé system (10.4.1) has components that are real-
analytic in Ω.
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).
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|→∞
has a unique solution. Moreover, the solution u satisfies the following additional
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)
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).
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
Apply the Fourier transform to each of the equalities in (10.6.3) and (10.6.4) to
obtain
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
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
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
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
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
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).
384 10 The Lamé and Stokes Operators
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)
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 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].
10.7 Additional Exercises for Chapter 10 385
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 10.35. 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.
386 10 The Lamé and Stokes Operators
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
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
(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
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 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.
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
lim −(λ + μ)div[EkL ] = pk pointwise in Rn \ {0} and in S (Rn ). (11.2.14)
λ→∞
390 11 More on Fundamental Solutions for Systems
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
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.
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
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
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
392 11 More on Fundamental Solutions for Systems
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
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
= δI M×M , (11.3.11)
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)!
11.3 Fundamental Solutions for Higher Order Systems 393
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.
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
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)
394 11 More on Fundamental Solutions for Systems
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
11.3 Fundamental Solutions for Higher Order Systems 395
γ −1
∂
γ E(ξ) = (−1)m i|γ| ξ L(ξ) for every ξ ∈ Rn \ {0}. (11.3.31)
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)
396 11 More on Fundamental Solutions for Systems
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
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
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
Also,
R,λ x = R,x whenever x ∈ O and λ > 0, (11.3.53)
and the joint application
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
and for each given x ∈ O− define the linear map R−,x : Rn → Rn by
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
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
400 11 More on Fundamental Solutions for Systems
(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
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
11.3 Fundamental Solutions for Higher Order Systems 401
(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
(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
402 11 More on Fundamental Solutions for Systems
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
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
11.3 Fundamental Solutions for Higher Order Systems 403
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 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).
404 11 More on Fundamental Solutions for Systems
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)
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.
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
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 ).
406 11 More on Fundamental Solutions for Systems
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
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).
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.
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)
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)
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
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)
= 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
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)
410 11 More on Fundamental Solutions for Systems
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.
In particular,
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
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
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
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.
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)
412 11 More on Fundamental Solutions for Systems
for every x ∈ Ω and every r ∈ 0, dist (x, ∂Ω) .
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, ∞).
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)
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
414 11 More on Fundamental Solutions for Systems
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
≤ C < ∞. (11.5.13)
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.
11.5 Reverse Hölder Estimates for Null-Solutions of Systems 415
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|! 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
|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.
416 11 More on Fundamental Solutions 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
In particular, since L(e j ) = A j j for each j ∈ {1, . . . , n}, the ellipticity condition
(11.6.3) entails
since LE = δ I M×M in D (Rn ) which (upon recalling that E ∈ C ∞ (Rn \ {0})) forces
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.
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
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).
Proof. For each C M -valued function ϕ ∈ S(Rn−1 ), formula (11.6.21) follows using
Corollary 11.14 by writing
420 11 More on Fundamental Solutions for Systems
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
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)
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+
11.6 Layer Potentials and Jump Relations for Systems 421
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)
422 11 More on Fundamental Solutions for Systems
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))
11.6 Layer Potentials and Jump Relations for Systems 423
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)
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+
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-
424 11 More on Fundamental Solutions for Systems
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
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.
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)
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
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)
428 12 Sobolev Spaces
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
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
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
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
430 12 Sobolev Spaces
sup (∂ j ϕ)(ξ + te j − η) ≤
C
, ∀ η ∈ Rn . (12.1.24)
|t|≤1 (1 + |η|)m
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
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
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
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,
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
12.1 Global Sobolev Spaces H s (Rn ), s ∈ Rn 433
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
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
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
434 12 Sobolev Spaces
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
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
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)| < ∞
12.1 Global Sobolev Spaces H s (Rn ), s ∈ Rn 435
(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
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
436 12 Sobolev Spaces
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
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
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
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)
∨ ,
= (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)
438 12 Sobolev Spaces
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
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
defines a norm on H s (Ω). Moreover, when regarding H s (Ω) equipped with this norm,
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
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
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
440 12 Sobolev Spaces
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
(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)
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.
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,
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
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
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
(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
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
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
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
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
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).
446 12 Sobolev Spaces
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∗ .
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
"
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,
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
12.3 Intrinsic Sobolev Spaces H m (Ω), m ∈ N 447
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
#
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)
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
448 12 Sobolev Spaces
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
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
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
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
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
450 12 Sobolev Spaces
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
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
12.3 Intrinsic Sobolev Spaces H m (Ω), m ∈ N 451
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) .
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
Now fix ε ∈ (0, ∞). Proposition 2.102 ensures that u ∗ Θε ∈ C0∞ (Rn ), hence
uε := u ∗ Θε Ω ∈ C0∞ (Ω). (12.3.50)
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
# 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
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
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
find a sequence
H m (Rn )
{ϕ0k }k∈N ⊂ C0∞ (Rn ) such that #
ϕ0k −−−−−→ ψ0 u. (12.3.57)
k→∞
H m (Ω)
ϕ0k Ω −−−−−→ ψ0 u. (12.3.58)
k→∞
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
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
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
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→∞
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
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!
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:
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 )!
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.
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
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 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
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
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
12.3 Intrinsic Sobolev Spaces H m (Ω), m ∈ N 459
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
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)
= 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
460 12 Sobolev Spaces
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
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
12.3 Intrinsic Sobolev Spaces H m (Ω), m ∈ N 461
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)
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)
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).
462 12 Sobolev Spaces
To proceed, select u ∈ H m (Ω) and fix j ∈ {1, . . . , n}. Making use of Exer-
cise 12.18 we obtain
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
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 ) ,
Taking into account (12.3.114) and (12.3.116), the last condition in (12.3.119) fur-
ther yields
12.3 Intrinsic Sobolev Spaces H m (Ω), m ∈ N 463
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)
#
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)
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.
464 12 Sobolev Spaces
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.
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.
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
12.3 Intrinsic Sobolev Spaces H m (Ω), m ∈ N 465
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
H 1 (U)
um −−−−−→ u. (12.3.133)
m→∞
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→∞
L2 (O)
um ◦ Ψ −−−−→ u ◦ Ψ. (12.3.136)
m→∞
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.
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
and
v ∈ H 1 (Rn+ ) ⇐⇒ v ◦ Φ−1 ∈ H 1 (Ω), (12.3.141)
and the equivalences hold with naturally accompanying estimates.
12.3 Intrinsic Sobolev Spaces H m (Ω), m ∈ N 467
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).
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
468 12 Sobolev Spaces
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
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.
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
470 12 Sobolev Spaces
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
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
12.4 The Space H 1/2 (∂Ω) on Boundaries of Lipschitz Domains 471
∞
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
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
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
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
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 .
472 12 Sobolev Spaces
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
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
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
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).
and
474 12 Sobolev Spaces
and
v ∈ H 1/2 (∂Rn+ ) ⇐⇒ v ◦ Φ−1 ∈ H 1/2 (∂Ω), (12.4.21)
and the equivalences hold with naturally accompanying estimates.
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 .
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)
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
12.4 The Space H 1/2 (∂Ω) on Boundaries of Lipschitz Domains 475
|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
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.
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
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.
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
478 12 Sobolev Spaces
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].
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+
12.5 Traces and Extensions 479
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
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
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
12.5 Traces and Extensions 481
$ 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+
To this end, let ϕ : Rn−1 → R be the Lipschitz function with the property that
482 12 Sobolev Spaces
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
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
12.5 Traces and Extensions 483
|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
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 .
484 12 Sobolev Spaces
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
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)
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)
+
η ∈ C ∞ (Rn ), 0 ≤ η ≤ 1 on Rn ,
(12.5.45)
supp η ⊆ B(0, 4), η ≡ 1 on B(0, 2).
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,
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
12.5 Traces and Extensions 487
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 .
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
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
488 12 Sobolev Spaces
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.
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
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 )
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
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)
490 12 Sobolev Spaces
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)
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)
+
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)
12.5 Traces and Extensions 491
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
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)
+
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
492 12 Sobolev Spaces
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 .
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
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
| 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 )
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
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
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
N
Ex f := u j. (12.5.98)
j=1
= 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
12.6 Additional Exercises for Chapter 12 495
(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)
(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 12.52. 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.
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
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
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
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
13.1 Solutions to Exercises from Section 1.4 501
− 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
502 13 Solutions to Selected Exercises
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
13.1 Solutions to Exercises from Section 1.4 503
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
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
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
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
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 ).
13.2 Solutions to Exercises from Section 2.10 505
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
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
506 13 Solutions to Selected Exercises
ϕ(ε) − ϕ(−ε) ln ε ≤ 2ϕ L∞ (R) ε| ln ε| −−−−→ 0. (13.2.3)
ε→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
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
13.2 Solutions to Exercises from Section 2.10 507
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
508 13 Solutions to Selected Exercises
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
|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 ,
Combining all these facts with (13.2.11) it follows that for each j ∈ N,
13.2 Solutions to Exercises from Section 2.10 509
≤ C max θ(
) L∞ ([−1/2,1/2]) :
≤ k0 jk0 . (13.2.12)
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→∞
1
= ωn−1 R j −−−−→ ωn−1 . (13.2.15)
j→∞
=: 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
510 13 Solutions to Selected Exercises
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
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
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→∞
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
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.
13.2 Solutions to Exercises from Section 2.10 513
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
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],
514 13 Solutions to Selected Exercises
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
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
= 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.
516 13 Solutions to Selected Exercises
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
13.2 Solutions to Exercises from Section 2.10 517
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
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
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
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
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.
13.2 Solutions to Exercises from Section 2.10 519
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 .
1
≤ ϕ L∞ ([−1,1]) → 0 as j → ∞. (13.2.48)
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→∞
520 13 Solutions to Selected Exercises
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
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|)
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
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
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 ).
522 13 Solutions to Selected Exercises
(c) (1 + |x|2 )2 is a polynomial function and if (1 + |x|2 )−2 ∈ S(Rn ) then their
n n
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
" 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)
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 3.43 Use Exercise 3.42.
≤ sup |ϕ (x)| + 2 sup |xϕ(x)| (13.4.2)
x∈R 1 x2 x∈R
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
524 13 Solutions to Selected Exercises
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→∞
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
13.4 Solutions to Exercises from Section 4.11 525
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→∞
−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
|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
526 13 Solutions to Selected Exercises
⎧ 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)
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
* =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)
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
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
528 13 Solutions to Selected Exercises
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
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
530 13 Solutions to Selected Exercises
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
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
13.6 Solutions to Exercises from Section 6.4 531
for every ϕ1 ∈ C0∞ (Rn ) and every ϕ2 ∈ C0∞ (Rm ), and then use Proposition 2.81.
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.
532 13 Solutions to Selected Exercises
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)
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.
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
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
∂ ∂
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|
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.
13.7 Solutions to Exercises from Section 7.14 535
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
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
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
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+
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
538 13 Solutions to Selected Exercises
∞
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.
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)
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
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)),
13.11 Solutions to Exercises from Section 12.6 541
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
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
∞
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.
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
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)
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)
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
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
ιt : Y
−→ X
(14.1.14)
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.
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.
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
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).
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 → ∞.
14.1 Summary of Topological and Functional Analytic Results 551
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→∞
u j → u in D
(Ω) as j → ∞ ⇐⇒
(14.1.27)
u j (ϕ) → u(ϕ) as j → ∞, ∀ ϕ ∈ C0∞ (Ω).
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
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
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 ).
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.
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!
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
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!
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)
The same circle of ideas that are used to prove Theorem 14.15 also yields a
corresponding L p -version which is stated next.
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 .
| f (x − y)||g(y)| dy < ∞ for a.e. x ∈ Rn , (14.2.13)
Rn
14.2 Basic Results from Calculus, Measure Theory, and Topology 555
1 1
L p (Rn ) ∗ L p (Rn ) → L∞ (Rn ) if p, p
∈ [1, ∞], + = 1. (14.2.16)
p p
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
for every ε ∈ (0, 1) and if we set K := supp f + B(0, R) then K is compact and
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)
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
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
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
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.
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
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
Then f is of class C ∞ on R.
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 .
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.
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
Proof. Let x ∈ E be such that x∗ F j for every j ∈ I. Then there exists r > 0 with
∗
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
∞
Then Ω = Ω j and
j=1
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
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})
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.
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).
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.
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
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.
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
∂ 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.
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)
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
14.7 Lipschitz Domains 571
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)
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
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
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)
cos (θ/2)(|y
− x
|2 + (s − ϕ(x
))2 ) 2 + ϕ(x
) ≥ ϕ(y
).
1
(14.7.27)
|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+
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
14.7 Lipschitz Domains 573
%
f dσ = f x
, ϕ(x
) 1 + |∇ϕ(x
)| dx
for all f ∈ L1 (∂Ω). (14.7.31)
∂Ω Rn−1
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.
=
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
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
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.
∞
≤ (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
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)
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 ,
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
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)
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
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 Polar Coordinates and Integrals on Spheres 579
(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 ∞ − ∞.
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
Q(λ1 ζ, λ2 η) = λ21 λm
2 Q(ζ, η) for all λ1 , λ2 > 0, ζ, η ∈ Rn . (14.9.22)
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
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
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 .
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.
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
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 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
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].
where Γ is the gamma function from (14.5.1). Also, by Nλ we denote the irregular
Bessel function given by
584 14 Appendix
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
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.
H0(1) (r)
lim+ =1 (14.10.5)
r→0 2i
ln(r)
π
14.10 Hankel Functions 585
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:
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)
⎧
⎪
⎪
⎪ 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:
14.10 Hankel Functions 587
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)
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
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
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
b, c ∈ R fixed
b, c ∈ R fixed
√ - π−ξ2 π−ξ2
.
π
sin(x2 ) 2i ei 4 − e−i 4 Exercise 4.121
u F (u) Location
δ 1 Example 4.22
2 t
ωn−1 (t2 +|x
|2 ) n2 e−t|ξ | Proposition 4.90
2
ωn−1 (t2 +|x
|2 ) n2 −i |ξ
j | e−t|ξ | Proposition 4.92
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
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
+